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>
16 #include "mds_client.h"
18 #include <linux/ceph/ceph_features.h>
19 #include <linux/ceph/messenger.h>
20 #include <linux/ceph/decode.h>
21 #include <linux/ceph/pagelist.h>
22 #include <linux/ceph/auth.h>
23 #include <linux/ceph/debugfs.h>
25 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
28 * A cluster of MDS (metadata server) daemons is responsible for
29 * managing the file system namespace (the directory hierarchy and
30 * inodes) and for coordinating shared access to storage. Metadata is
31 * partitioning hierarchically across a number of servers, and that
32 * partition varies over time as the cluster adjusts the distribution
33 * in order to balance load.
35 * The MDS client is primarily responsible to managing synchronous
36 * metadata requests for operations like open, unlink, and so forth.
37 * If there is a MDS failure, we find out about it when we (possibly
38 * request and) receive a new MDS map, and can resubmit affected
41 * For the most part, though, we take advantage of a lossless
42 * communications channel to the MDS, and do not need to worry about
43 * timing out or resubmitting requests.
45 * We maintain a stateful "session" with each MDS we interact with.
46 * Within each session, we sent periodic heartbeat messages to ensure
47 * any capabilities or leases we have been issues remain valid. If
48 * the session times out and goes stale, our leases and capabilities
49 * are no longer valid.
52 struct ceph_reconnect_state {
53 struct ceph_mds_session *session;
54 int nr_caps, nr_realms;
55 struct ceph_pagelist *pagelist;
60 static void __wake_requests(struct ceph_mds_client *mdsc,
61 struct list_head *head);
62 static void ceph_cap_release_work(struct work_struct *work);
63 static void ceph_cap_reclaim_work(struct work_struct *work);
65 static const struct ceph_connection_operations mds_con_ops;
72 static int parse_reply_info_quota(void **p, void *end,
73 struct ceph_mds_reply_info_in *info)
75 u8 struct_v, struct_compat;
78 ceph_decode_8_safe(p, end, struct_v, bad);
79 ceph_decode_8_safe(p, end, struct_compat, bad);
80 /* struct_v is expected to be >= 1. we only
81 * understand encoding with struct_compat == 1. */
82 if (!struct_v || struct_compat != 1)
84 ceph_decode_32_safe(p, end, struct_len, bad);
85 ceph_decode_need(p, end, struct_len, bad);
86 end = *p + struct_len;
87 ceph_decode_64_safe(p, end, info->max_bytes, bad);
88 ceph_decode_64_safe(p, end, info->max_files, bad);
96 * parse individual inode info
98 static int parse_reply_info_in(void **p, void *end,
99 struct ceph_mds_reply_info_in *info,
105 if (features == (u64)-1) {
108 ceph_decode_8_safe(p, end, struct_v, bad);
109 ceph_decode_8_safe(p, end, struct_compat, bad);
110 /* struct_v is expected to be >= 1. we only understand
111 * encoding with struct_compat == 1. */
112 if (!struct_v || struct_compat != 1)
114 ceph_decode_32_safe(p, end, struct_len, bad);
115 ceph_decode_need(p, end, struct_len, bad);
116 end = *p + struct_len;
119 ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
121 *p += sizeof(struct ceph_mds_reply_inode) +
122 sizeof(*info->in->fragtree.splits) *
123 le32_to_cpu(info->in->fragtree.nsplits);
125 ceph_decode_32_safe(p, end, info->symlink_len, bad);
126 ceph_decode_need(p, end, info->symlink_len, bad);
128 *p += info->symlink_len;
130 ceph_decode_copy_safe(p, end, &info->dir_layout,
131 sizeof(info->dir_layout), bad);
132 ceph_decode_32_safe(p, end, info->xattr_len, bad);
133 ceph_decode_need(p, end, info->xattr_len, bad);
134 info->xattr_data = *p;
135 *p += info->xattr_len;
137 if (features == (u64)-1) {
139 ceph_decode_64_safe(p, end, info->inline_version, bad);
140 ceph_decode_32_safe(p, end, info->inline_len, bad);
141 ceph_decode_need(p, end, info->inline_len, bad);
142 info->inline_data = *p;
143 *p += info->inline_len;
145 err = parse_reply_info_quota(p, end, info);
149 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
150 if (info->pool_ns_len > 0) {
151 ceph_decode_need(p, end, info->pool_ns_len, bad);
152 info->pool_ns_data = *p;
153 *p += info->pool_ns_len;
157 ceph_decode_need(p, end, sizeof(info->btime), bad);
158 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
160 /* change attribute */
161 ceph_decode_64_safe(p, end, info->change_attr, bad);
165 ceph_decode_32_safe(p, end, info->dir_pin, bad);
167 info->dir_pin = -ENODATA;
170 /* snapshot birth time, remains zero for v<=2 */
172 ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
173 ceph_decode_copy(p, &info->snap_btime,
174 sizeof(info->snap_btime));
176 memset(&info->snap_btime, 0, sizeof(info->snap_btime));
181 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
182 ceph_decode_64_safe(p, end, info->inline_version, bad);
183 ceph_decode_32_safe(p, end, info->inline_len, bad);
184 ceph_decode_need(p, end, info->inline_len, bad);
185 info->inline_data = *p;
186 *p += info->inline_len;
188 info->inline_version = CEPH_INLINE_NONE;
190 if (features & CEPH_FEATURE_MDS_QUOTA) {
191 err = parse_reply_info_quota(p, end, info);
199 info->pool_ns_len = 0;
200 info->pool_ns_data = NULL;
201 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
202 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
203 if (info->pool_ns_len > 0) {
204 ceph_decode_need(p, end, info->pool_ns_len, bad);
205 info->pool_ns_data = *p;
206 *p += info->pool_ns_len;
210 if (features & CEPH_FEATURE_FS_BTIME) {
211 ceph_decode_need(p, end, sizeof(info->btime), bad);
212 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
213 ceph_decode_64_safe(p, end, info->change_attr, bad);
216 info->dir_pin = -ENODATA;
217 /* info->snap_btime remains zero */
226 static int parse_reply_info_dir(void **p, void *end,
227 struct ceph_mds_reply_dirfrag **dirfrag,
230 if (features == (u64)-1) {
231 u8 struct_v, struct_compat;
233 ceph_decode_8_safe(p, end, struct_v, bad);
234 ceph_decode_8_safe(p, end, struct_compat, bad);
235 /* struct_v is expected to be >= 1. we only understand
236 * encoding whose struct_compat == 1. */
237 if (!struct_v || struct_compat != 1)
239 ceph_decode_32_safe(p, end, struct_len, bad);
240 ceph_decode_need(p, end, struct_len, bad);
241 end = *p + struct_len;
244 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
246 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
247 if (unlikely(*p > end))
249 if (features == (u64)-1)
256 static int parse_reply_info_lease(void **p, void *end,
257 struct ceph_mds_reply_lease **lease,
260 if (features == (u64)-1) {
261 u8 struct_v, struct_compat;
263 ceph_decode_8_safe(p, end, struct_v, bad);
264 ceph_decode_8_safe(p, end, struct_compat, bad);
265 /* struct_v is expected to be >= 1. we only understand
266 * encoding whose struct_compat == 1. */
267 if (!struct_v || struct_compat != 1)
269 ceph_decode_32_safe(p, end, struct_len, bad);
270 ceph_decode_need(p, end, struct_len, bad);
271 end = *p + struct_len;
274 ceph_decode_need(p, end, sizeof(**lease), bad);
276 *p += sizeof(**lease);
277 if (features == (u64)-1)
285 * parse a normal reply, which may contain a (dir+)dentry and/or a
288 static int parse_reply_info_trace(void **p, void *end,
289 struct ceph_mds_reply_info_parsed *info,
294 if (info->head->is_dentry) {
295 err = parse_reply_info_in(p, end, &info->diri, features);
299 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
303 ceph_decode_32_safe(p, end, info->dname_len, bad);
304 ceph_decode_need(p, end, info->dname_len, bad);
306 *p += info->dname_len;
308 err = parse_reply_info_lease(p, end, &info->dlease, features);
313 if (info->head->is_target) {
314 err = parse_reply_info_in(p, end, &info->targeti, features);
319 if (unlikely(*p != end))
326 pr_err("problem parsing mds trace %d\n", err);
331 * parse readdir results
333 static int parse_reply_info_readdir(void **p, void *end,
334 struct ceph_mds_reply_info_parsed *info,
340 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
344 ceph_decode_need(p, end, sizeof(num) + 2, bad);
345 num = ceph_decode_32(p);
347 u16 flags = ceph_decode_16(p);
348 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
349 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
350 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
351 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
356 BUG_ON(!info->dir_entries);
357 if ((unsigned long)(info->dir_entries + num) >
358 (unsigned long)info->dir_entries + info->dir_buf_size) {
359 pr_err("dir contents are larger than expected\n");
366 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
368 ceph_decode_32_safe(p, end, rde->name_len, bad);
369 ceph_decode_need(p, end, rde->name_len, bad);
372 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
375 err = parse_reply_info_lease(p, end, &rde->lease, features);
379 err = parse_reply_info_in(p, end, &rde->inode, features);
382 /* ceph_readdir_prepopulate() will update it */
389 /* Skip over any unrecognized fields */
396 pr_err("problem parsing dir contents %d\n", err);
401 * parse fcntl F_GETLK results
403 static int parse_reply_info_filelock(void **p, void *end,
404 struct ceph_mds_reply_info_parsed *info,
407 if (*p + sizeof(*info->filelock_reply) > end)
410 info->filelock_reply = *p;
412 /* Skip over any unrecognized fields */
420 #if BITS_PER_LONG == 64
422 #define DELEGATED_INO_AVAILABLE xa_mk_value(1)
424 static int ceph_parse_deleg_inos(void **p, void *end,
425 struct ceph_mds_session *s)
429 ceph_decode_32_safe(p, end, sets, bad);
430 dout("got %u sets of delegated inodes\n", sets);
434 ceph_decode_64_safe(p, end, start, bad);
435 ceph_decode_64_safe(p, end, len, bad);
437 /* Don't accept a delegation of system inodes */
438 if (start < CEPH_INO_SYSTEM_BASE) {
439 pr_warn_ratelimited("ceph: ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
444 int err = xa_insert(&s->s_delegated_inos, ino = start++,
445 DELEGATED_INO_AVAILABLE,
448 dout("added delegated inode 0x%llx\n",
450 } else if (err == -EBUSY) {
451 pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
463 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
468 xa_for_each(&s->s_delegated_inos, ino, val) {
469 val = xa_erase(&s->s_delegated_inos, ino);
470 if (val == DELEGATED_INO_AVAILABLE)
476 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
478 return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
481 #else /* BITS_PER_LONG == 64 */
483 * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
484 * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
487 static int ceph_parse_deleg_inos(void **p, void *end,
488 struct ceph_mds_session *s)
492 ceph_decode_32_safe(p, end, sets, bad);
494 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
500 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
505 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
509 #endif /* BITS_PER_LONG == 64 */
512 * parse create results
514 static int parse_reply_info_create(void **p, void *end,
515 struct ceph_mds_reply_info_parsed *info,
516 u64 features, struct ceph_mds_session *s)
520 if (features == (u64)-1 ||
521 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
523 /* Malformed reply? */
524 info->has_create_ino = false;
525 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
526 u8 struct_v, struct_compat;
529 info->has_create_ino = true;
530 ceph_decode_8_safe(p, end, struct_v, bad);
531 ceph_decode_8_safe(p, end, struct_compat, bad);
532 ceph_decode_32_safe(p, end, len, bad);
533 ceph_decode_64_safe(p, end, info->ino, bad);
534 ret = ceph_parse_deleg_inos(p, end, s);
539 ceph_decode_64_safe(p, end, info->ino, bad);
540 info->has_create_ino = true;
547 /* Skip over any unrecognized fields */
555 * parse extra results
557 static int parse_reply_info_extra(void **p, void *end,
558 struct ceph_mds_reply_info_parsed *info,
559 u64 features, struct ceph_mds_session *s)
561 u32 op = le32_to_cpu(info->head->op);
563 if (op == CEPH_MDS_OP_GETFILELOCK)
564 return parse_reply_info_filelock(p, end, info, features);
565 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
566 return parse_reply_info_readdir(p, end, info, features);
567 else if (op == CEPH_MDS_OP_CREATE)
568 return parse_reply_info_create(p, end, info, features, s);
574 * parse entire mds reply
576 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
577 struct ceph_mds_reply_info_parsed *info,
584 info->head = msg->front.iov_base;
585 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
586 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
589 ceph_decode_32_safe(&p, end, len, bad);
591 ceph_decode_need(&p, end, len, bad);
592 err = parse_reply_info_trace(&p, p+len, info, features);
598 ceph_decode_32_safe(&p, end, len, bad);
600 ceph_decode_need(&p, end, len, bad);
601 err = parse_reply_info_extra(&p, p+len, info, features, s);
607 ceph_decode_32_safe(&p, end, len, bad);
608 info->snapblob_len = len;
619 pr_err("mds parse_reply err %d\n", err);
623 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
625 if (!info->dir_entries)
627 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
634 const char *ceph_session_state_name(int s)
637 case CEPH_MDS_SESSION_NEW: return "new";
638 case CEPH_MDS_SESSION_OPENING: return "opening";
639 case CEPH_MDS_SESSION_OPEN: return "open";
640 case CEPH_MDS_SESSION_HUNG: return "hung";
641 case CEPH_MDS_SESSION_CLOSING: return "closing";
642 case CEPH_MDS_SESSION_CLOSED: return "closed";
643 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
644 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
645 case CEPH_MDS_SESSION_REJECTED: return "rejected";
646 default: return "???";
650 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
652 if (refcount_inc_not_zero(&s->s_ref)) {
653 dout("mdsc get_session %p %d -> %d\n", s,
654 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
657 dout("mdsc get_session %p 0 -- FAIL\n", s);
662 void ceph_put_mds_session(struct ceph_mds_session *s)
664 if (IS_ERR_OR_NULL(s))
667 dout("mdsc put_session %p %d -> %d\n", s,
668 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
669 if (refcount_dec_and_test(&s->s_ref)) {
670 if (s->s_auth.authorizer)
671 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
672 WARN_ON(mutex_is_locked(&s->s_mutex));
673 xa_destroy(&s->s_delegated_inos);
679 * called under mdsc->mutex
681 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
684 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
686 return ceph_get_mds_session(mdsc->sessions[mds]);
689 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
691 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
697 static int __verify_registered_session(struct ceph_mds_client *mdsc,
698 struct ceph_mds_session *s)
700 if (s->s_mds >= mdsc->max_sessions ||
701 mdsc->sessions[s->s_mds] != s)
707 * create+register a new session for given mds.
708 * called under mdsc->mutex.
710 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
713 struct ceph_mds_session *s;
715 if (mds >= mdsc->mdsmap->possible_max_rank)
716 return ERR_PTR(-EINVAL);
718 s = kzalloc(sizeof(*s), GFP_NOFS);
720 return ERR_PTR(-ENOMEM);
722 if (mds >= mdsc->max_sessions) {
723 int newmax = 1 << get_count_order(mds + 1);
724 struct ceph_mds_session **sa;
726 dout("%s: realloc to %d\n", __func__, newmax);
727 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
730 if (mdsc->sessions) {
731 memcpy(sa, mdsc->sessions,
732 mdsc->max_sessions * sizeof(void *));
733 kfree(mdsc->sessions);
736 mdsc->max_sessions = newmax;
739 dout("%s: mds%d\n", __func__, mds);
742 s->s_state = CEPH_MDS_SESSION_NEW;
745 mutex_init(&s->s_mutex);
747 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
749 spin_lock_init(&s->s_gen_ttl_lock);
751 s->s_cap_ttl = jiffies - 1;
753 spin_lock_init(&s->s_cap_lock);
754 s->s_renew_requested = 0;
756 INIT_LIST_HEAD(&s->s_caps);
758 refcount_set(&s->s_ref, 1);
759 INIT_LIST_HEAD(&s->s_waiting);
760 INIT_LIST_HEAD(&s->s_unsafe);
761 xa_init(&s->s_delegated_inos);
762 s->s_num_cap_releases = 0;
763 s->s_cap_reconnect = 0;
764 s->s_cap_iterator = NULL;
765 INIT_LIST_HEAD(&s->s_cap_releases);
766 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
768 INIT_LIST_HEAD(&s->s_cap_dirty);
769 INIT_LIST_HEAD(&s->s_cap_flushing);
771 mdsc->sessions[mds] = s;
772 atomic_inc(&mdsc->num_sessions);
773 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
775 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
776 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
782 return ERR_PTR(-ENOMEM);
786 * called under mdsc->mutex
788 static void __unregister_session(struct ceph_mds_client *mdsc,
789 struct ceph_mds_session *s)
791 dout("__unregister_session mds%d %p\n", s->s_mds, s);
792 BUG_ON(mdsc->sessions[s->s_mds] != s);
793 mdsc->sessions[s->s_mds] = NULL;
794 ceph_con_close(&s->s_con);
795 ceph_put_mds_session(s);
796 atomic_dec(&mdsc->num_sessions);
800 * drop session refs in request.
802 * should be last request ref, or hold mdsc->mutex
804 static void put_request_session(struct ceph_mds_request *req)
806 if (req->r_session) {
807 ceph_put_mds_session(req->r_session);
808 req->r_session = NULL;
812 void ceph_mdsc_release_request(struct kref *kref)
814 struct ceph_mds_request *req = container_of(kref,
815 struct ceph_mds_request,
817 ceph_mdsc_release_dir_caps_no_check(req);
818 destroy_reply_info(&req->r_reply_info);
820 ceph_msg_put(req->r_request);
822 ceph_msg_put(req->r_reply);
824 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
825 /* avoid calling iput_final() in mds dispatch threads */
826 ceph_async_iput(req->r_inode);
829 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
830 ceph_async_iput(req->r_parent);
832 ceph_async_iput(req->r_target_inode);
835 if (req->r_old_dentry)
836 dput(req->r_old_dentry);
837 if (req->r_old_dentry_dir) {
839 * track (and drop pins for) r_old_dentry_dir
840 * separately, since r_old_dentry's d_parent may have
841 * changed between the dir mutex being dropped and
842 * this request being freed.
844 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
846 ceph_async_iput(req->r_old_dentry_dir);
851 ceph_pagelist_release(req->r_pagelist);
852 put_request_session(req);
853 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
854 WARN_ON_ONCE(!list_empty(&req->r_wait));
855 kmem_cache_free(ceph_mds_request_cachep, req);
858 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
861 * lookup session, bump ref if found.
863 * called under mdsc->mutex.
865 static struct ceph_mds_request *
866 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
868 struct ceph_mds_request *req;
870 req = lookup_request(&mdsc->request_tree, tid);
872 ceph_mdsc_get_request(req);
878 * Register an in-flight request, and assign a tid. Link to directory
879 * are modifying (if any).
881 * Called under mdsc->mutex.
883 static void __register_request(struct ceph_mds_client *mdsc,
884 struct ceph_mds_request *req,
889 req->r_tid = ++mdsc->last_tid;
890 if (req->r_num_caps) {
891 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
894 pr_err("__register_request %p "
895 "failed to reserve caps: %d\n", req, ret);
896 /* set req->r_err to fail early from __do_request */
901 dout("__register_request %p tid %lld\n", req, req->r_tid);
902 ceph_mdsc_get_request(req);
903 insert_request(&mdsc->request_tree, req);
905 req->r_uid = current_fsuid();
906 req->r_gid = current_fsgid();
908 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
909 mdsc->oldest_tid = req->r_tid;
912 struct ceph_inode_info *ci = ceph_inode(dir);
915 req->r_unsafe_dir = dir;
916 spin_lock(&ci->i_unsafe_lock);
917 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
918 spin_unlock(&ci->i_unsafe_lock);
922 static void __unregister_request(struct ceph_mds_client *mdsc,
923 struct ceph_mds_request *req)
925 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
927 /* Never leave an unregistered request on an unsafe list! */
928 list_del_init(&req->r_unsafe_item);
930 if (req->r_tid == mdsc->oldest_tid) {
931 struct rb_node *p = rb_next(&req->r_node);
932 mdsc->oldest_tid = 0;
934 struct ceph_mds_request *next_req =
935 rb_entry(p, struct ceph_mds_request, r_node);
936 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
937 mdsc->oldest_tid = next_req->r_tid;
944 erase_request(&mdsc->request_tree, req);
946 if (req->r_unsafe_dir) {
947 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
948 spin_lock(&ci->i_unsafe_lock);
949 list_del_init(&req->r_unsafe_dir_item);
950 spin_unlock(&ci->i_unsafe_lock);
952 if (req->r_target_inode &&
953 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
954 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
955 spin_lock(&ci->i_unsafe_lock);
956 list_del_init(&req->r_unsafe_target_item);
957 spin_unlock(&ci->i_unsafe_lock);
960 if (req->r_unsafe_dir) {
961 /* avoid calling iput_final() in mds dispatch threads */
962 ceph_async_iput(req->r_unsafe_dir);
963 req->r_unsafe_dir = NULL;
966 complete_all(&req->r_safe_completion);
968 ceph_mdsc_put_request(req);
972 * Walk back up the dentry tree until we hit a dentry representing a
973 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
974 * when calling this) to ensure that the objects won't disappear while we're
975 * working with them. Once we hit a candidate dentry, we attempt to take a
976 * reference to it, and return that as the result.
978 static struct inode *get_nonsnap_parent(struct dentry *dentry)
980 struct inode *inode = NULL;
982 while (dentry && !IS_ROOT(dentry)) {
983 inode = d_inode_rcu(dentry);
984 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
986 dentry = dentry->d_parent;
989 inode = igrab(inode);
994 * Choose mds to send request to next. If there is a hint set in the
995 * request (e.g., due to a prior forward hint from the mds), use that.
996 * Otherwise, consult frag tree and/or caps to identify the
997 * appropriate mds. If all else fails, choose randomly.
999 * Called under mdsc->mutex.
1001 static int __choose_mds(struct ceph_mds_client *mdsc,
1002 struct ceph_mds_request *req,
1005 struct inode *inode;
1006 struct ceph_inode_info *ci;
1007 struct ceph_cap *cap;
1008 int mode = req->r_direct_mode;
1010 u32 hash = req->r_direct_hash;
1011 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1017 * is there a specific mds we should try? ignore hint if we have
1018 * no session and the mds is not up (active or recovering).
1020 if (req->r_resend_mds >= 0 &&
1021 (__have_session(mdsc, req->r_resend_mds) ||
1022 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1023 dout("%s using resend_mds mds%d\n", __func__,
1025 return req->r_resend_mds;
1028 if (mode == USE_RANDOM_MDS)
1033 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1034 inode = req->r_inode;
1037 /* req->r_dentry is non-null for LSSNAP request */
1039 inode = get_nonsnap_parent(req->r_dentry);
1041 dout("%s using snapdir's parent %p\n", __func__, inode);
1043 } else if (req->r_dentry) {
1044 /* ignore race with rename; old or new d_parent is okay */
1045 struct dentry *parent;
1049 parent = READ_ONCE(req->r_dentry->d_parent);
1050 dir = req->r_parent ? : d_inode_rcu(parent);
1052 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1053 /* not this fs or parent went negative */
1054 inode = d_inode(req->r_dentry);
1057 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1058 /* direct snapped/virtual snapdir requests
1059 * based on parent dir inode */
1060 inode = get_nonsnap_parent(parent);
1061 dout("%s using nonsnap parent %p\n", __func__, inode);
1064 inode = d_inode(req->r_dentry);
1065 if (!inode || mode == USE_AUTH_MDS) {
1068 hash = ceph_dentry_hash(dir, req->r_dentry);
1077 dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1081 ci = ceph_inode(inode);
1083 if (is_hash && S_ISDIR(inode->i_mode)) {
1084 struct ceph_inode_frag frag;
1087 ceph_choose_frag(ci, hash, &frag, &found);
1089 if (mode == USE_ANY_MDS && frag.ndist > 0) {
1092 /* choose a random replica */
1093 get_random_bytes(&r, 1);
1096 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1097 __func__, inode, ceph_vinop(inode),
1098 frag.frag, mds, (int)r, frag.ndist);
1099 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1100 CEPH_MDS_STATE_ACTIVE &&
1101 !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1105 /* since this file/dir wasn't known to be
1106 * replicated, then we want to look for the
1107 * authoritative mds. */
1108 if (frag.mds >= 0) {
1109 /* choose auth mds */
1111 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1112 __func__, inode, ceph_vinop(inode),
1114 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1115 CEPH_MDS_STATE_ACTIVE) {
1116 if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1121 mode = USE_AUTH_MDS;
1125 spin_lock(&ci->i_ceph_lock);
1127 if (mode == USE_AUTH_MDS)
1128 cap = ci->i_auth_cap;
1129 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1130 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1132 spin_unlock(&ci->i_ceph_lock);
1133 ceph_async_iput(inode);
1136 mds = cap->session->s_mds;
1137 dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1138 inode, ceph_vinop(inode), mds,
1139 cap == ci->i_auth_cap ? "auth " : "", cap);
1140 spin_unlock(&ci->i_ceph_lock);
1142 /* avoid calling iput_final() while holding mdsc->mutex or
1143 * in mds dispatch threads */
1144 ceph_async_iput(inode);
1151 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1152 dout("%s chose random mds%d\n", __func__, mds);
1160 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1162 struct ceph_msg *msg;
1163 struct ceph_mds_session_head *h;
1165 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1168 pr_err("create_session_msg ENOMEM creating msg\n");
1171 h = msg->front.iov_base;
1172 h->op = cpu_to_le32(op);
1173 h->seq = cpu_to_le64(seq);
1178 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1179 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1180 static int encode_supported_features(void **p, void *end)
1182 static const size_t count = ARRAY_SIZE(feature_bits);
1186 size_t size = FEATURE_BYTES(count);
1189 if (WARN_ON_ONCE(*p + 4 + size > end))
1192 ceph_encode_32(p, size);
1193 memset(*p, 0, size);
1194 for (i = 0; i < count; i++) {
1195 bit = feature_bits[i];
1196 ((unsigned char *)(*p))[bit / 8] |= BIT(bit % 8);
1200 if (WARN_ON_ONCE(*p + 4 > end))
1203 ceph_encode_32(p, 0);
1209 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1210 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1211 static int encode_metric_spec(void **p, void *end)
1213 static const size_t count = ARRAY_SIZE(metric_bits);
1216 if (WARN_ON_ONCE(*p + 2 > end))
1219 ceph_encode_8(p, 1); /* version */
1220 ceph_encode_8(p, 1); /* compat */
1224 size_t size = METRIC_BYTES(count);
1226 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1229 /* metric spec info length */
1230 ceph_encode_32(p, 4 + size);
1233 ceph_encode_32(p, size);
1234 memset(*p, 0, size);
1235 for (i = 0; i < count; i++)
1236 ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1239 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1242 /* metric spec info length */
1243 ceph_encode_32(p, 4);
1245 ceph_encode_32(p, 0);
1252 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1253 * to include additional client metadata fields.
1255 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1257 struct ceph_msg *msg;
1258 struct ceph_mds_session_head *h;
1260 int extra_bytes = 0;
1261 int metadata_key_count = 0;
1262 struct ceph_options *opt = mdsc->fsc->client->options;
1263 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1268 const char* metadata[][2] = {
1269 {"hostname", mdsc->nodename},
1270 {"kernel_version", init_utsname()->release},
1271 {"entity_id", opt->name ? : ""},
1272 {"root", fsopt->server_path ? : "/"},
1276 /* Calculate serialized length of metadata */
1277 extra_bytes = 4; /* map length */
1278 for (i = 0; metadata[i][0]; ++i) {
1279 extra_bytes += 8 + strlen(metadata[i][0]) +
1280 strlen(metadata[i][1]);
1281 metadata_key_count++;
1284 /* supported feature */
1286 count = ARRAY_SIZE(feature_bits);
1288 size = FEATURE_BYTES(count);
1289 extra_bytes += 4 + size;
1293 count = ARRAY_SIZE(metric_bits);
1295 size = METRIC_BYTES(count);
1296 extra_bytes += 2 + 4 + 4 + size;
1298 /* Allocate the message */
1299 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1302 pr_err("create_session_msg ENOMEM creating msg\n");
1303 return ERR_PTR(-ENOMEM);
1305 p = msg->front.iov_base;
1306 end = p + msg->front.iov_len;
1309 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1310 h->seq = cpu_to_le64(seq);
1313 * Serialize client metadata into waiting buffer space, using
1314 * the format that userspace expects for map<string, string>
1316 * ClientSession messages with metadata are v4
1318 msg->hdr.version = cpu_to_le16(4);
1319 msg->hdr.compat_version = cpu_to_le16(1);
1321 /* The write pointer, following the session_head structure */
1324 /* Number of entries in the map */
1325 ceph_encode_32(&p, metadata_key_count);
1327 /* Two length-prefixed strings for each entry in the map */
1328 for (i = 0; metadata[i][0]; ++i) {
1329 size_t const key_len = strlen(metadata[i][0]);
1330 size_t const val_len = strlen(metadata[i][1]);
1332 ceph_encode_32(&p, key_len);
1333 memcpy(p, metadata[i][0], key_len);
1335 ceph_encode_32(&p, val_len);
1336 memcpy(p, metadata[i][1], val_len);
1340 ret = encode_supported_features(&p, end);
1342 pr_err("encode_supported_features failed!\n");
1344 return ERR_PTR(ret);
1347 ret = encode_metric_spec(&p, end);
1349 pr_err("encode_metric_spec failed!\n");
1351 return ERR_PTR(ret);
1354 msg->front.iov_len = p - msg->front.iov_base;
1355 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1361 * send session open request.
1363 * called under mdsc->mutex
1365 static int __open_session(struct ceph_mds_client *mdsc,
1366 struct ceph_mds_session *session)
1368 struct ceph_msg *msg;
1370 int mds = session->s_mds;
1372 /* wait for mds to go active? */
1373 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1374 dout("open_session to mds%d (%s)\n", mds,
1375 ceph_mds_state_name(mstate));
1376 session->s_state = CEPH_MDS_SESSION_OPENING;
1377 session->s_renew_requested = jiffies;
1379 /* send connect message */
1380 msg = create_session_open_msg(mdsc, session->s_seq);
1382 return PTR_ERR(msg);
1383 ceph_con_send(&session->s_con, msg);
1388 * open sessions for any export targets for the given mds
1390 * called under mdsc->mutex
1392 static struct ceph_mds_session *
1393 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1395 struct ceph_mds_session *session;
1398 session = __ceph_lookup_mds_session(mdsc, target);
1400 session = register_session(mdsc, target);
1401 if (IS_ERR(session))
1404 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1405 session->s_state == CEPH_MDS_SESSION_CLOSING) {
1406 ret = __open_session(mdsc, session);
1408 return ERR_PTR(ret);
1414 struct ceph_mds_session *
1415 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1417 struct ceph_mds_session *session;
1419 dout("open_export_target_session to mds%d\n", target);
1421 mutex_lock(&mdsc->mutex);
1422 session = __open_export_target_session(mdsc, target);
1423 mutex_unlock(&mdsc->mutex);
1428 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1429 struct ceph_mds_session *session)
1431 struct ceph_mds_info *mi;
1432 struct ceph_mds_session *ts;
1433 int i, mds = session->s_mds;
1435 if (mds >= mdsc->mdsmap->possible_max_rank)
1438 mi = &mdsc->mdsmap->m_info[mds];
1439 dout("open_export_target_sessions for mds%d (%d targets)\n",
1440 session->s_mds, mi->num_export_targets);
1442 for (i = 0; i < mi->num_export_targets; i++) {
1443 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1444 ceph_put_mds_session(ts);
1448 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1449 struct ceph_mds_session *session)
1451 mutex_lock(&mdsc->mutex);
1452 __open_export_target_sessions(mdsc, session);
1453 mutex_unlock(&mdsc->mutex);
1460 static void detach_cap_releases(struct ceph_mds_session *session,
1461 struct list_head *target)
1463 lockdep_assert_held(&session->s_cap_lock);
1465 list_splice_init(&session->s_cap_releases, target);
1466 session->s_num_cap_releases = 0;
1467 dout("dispose_cap_releases mds%d\n", session->s_mds);
1470 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1471 struct list_head *dispose)
1473 while (!list_empty(dispose)) {
1474 struct ceph_cap *cap;
1475 /* zero out the in-progress message */
1476 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1477 list_del(&cap->session_caps);
1478 ceph_put_cap(mdsc, cap);
1482 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1483 struct ceph_mds_session *session)
1485 struct ceph_mds_request *req;
1488 dout("cleanup_session_requests mds%d\n", session->s_mds);
1489 mutex_lock(&mdsc->mutex);
1490 while (!list_empty(&session->s_unsafe)) {
1491 req = list_first_entry(&session->s_unsafe,
1492 struct ceph_mds_request, r_unsafe_item);
1493 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1495 if (req->r_target_inode)
1496 mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1497 if (req->r_unsafe_dir)
1498 mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1499 __unregister_request(mdsc, req);
1501 /* zero r_attempts, so kick_requests() will re-send requests */
1502 p = rb_first(&mdsc->request_tree);
1504 req = rb_entry(p, struct ceph_mds_request, r_node);
1506 if (req->r_session &&
1507 req->r_session->s_mds == session->s_mds)
1508 req->r_attempts = 0;
1510 mutex_unlock(&mdsc->mutex);
1514 * Helper to safely iterate over all caps associated with a session, with
1515 * special care taken to handle a racing __ceph_remove_cap().
1517 * Caller must hold session s_mutex.
1519 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1520 int (*cb)(struct inode *, struct ceph_cap *,
1523 struct list_head *p;
1524 struct ceph_cap *cap;
1525 struct inode *inode, *last_inode = NULL;
1526 struct ceph_cap *old_cap = NULL;
1529 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1530 spin_lock(&session->s_cap_lock);
1531 p = session->s_caps.next;
1532 while (p != &session->s_caps) {
1533 cap = list_entry(p, struct ceph_cap, session_caps);
1534 inode = igrab(&cap->ci->vfs_inode);
1539 session->s_cap_iterator = cap;
1540 spin_unlock(&session->s_cap_lock);
1543 /* avoid calling iput_final() while holding
1544 * s_mutex or in mds dispatch threads */
1545 ceph_async_iput(last_inode);
1549 ceph_put_cap(session->s_mdsc, old_cap);
1553 ret = cb(inode, cap, arg);
1556 spin_lock(&session->s_cap_lock);
1559 dout("iterate_session_caps finishing cap %p removal\n",
1561 BUG_ON(cap->session != session);
1562 cap->session = NULL;
1563 list_del_init(&cap->session_caps);
1564 session->s_nr_caps--;
1565 atomic64_dec(&session->s_mdsc->metric.total_caps);
1566 if (cap->queue_release)
1567 __ceph_queue_cap_release(session, cap);
1569 old_cap = cap; /* put_cap it w/o locks held */
1576 session->s_cap_iterator = NULL;
1577 spin_unlock(&session->s_cap_lock);
1579 ceph_async_iput(last_inode);
1581 ceph_put_cap(session->s_mdsc, old_cap);
1586 static int remove_capsnaps(struct ceph_mds_client *mdsc, struct inode *inode)
1588 struct ceph_inode_info *ci = ceph_inode(inode);
1589 struct ceph_cap_snap *capsnap;
1590 int capsnap_release = 0;
1592 lockdep_assert_held(&ci->i_ceph_lock);
1594 dout("removing capsnaps, ci is %p, inode is %p\n", ci, inode);
1596 while (!list_empty(&ci->i_cap_snaps)) {
1597 capsnap = list_first_entry(&ci->i_cap_snaps,
1598 struct ceph_cap_snap, ci_item);
1599 __ceph_remove_capsnap(inode, capsnap, NULL, NULL);
1600 ceph_put_snap_context(capsnap->context);
1601 ceph_put_cap_snap(capsnap);
1604 wake_up_all(&ci->i_cap_wq);
1605 wake_up_all(&mdsc->cap_flushing_wq);
1606 return capsnap_release;
1609 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1612 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1613 struct ceph_mds_client *mdsc = fsc->mdsc;
1614 struct ceph_inode_info *ci = ceph_inode(inode);
1615 LIST_HEAD(to_remove);
1616 bool dirty_dropped = false;
1617 bool invalidate = false;
1618 int capsnap_release = 0;
1620 dout("removing cap %p, ci is %p, inode is %p\n",
1621 cap, ci, &ci->vfs_inode);
1622 spin_lock(&ci->i_ceph_lock);
1623 __ceph_remove_cap(cap, false);
1624 if (!ci->i_auth_cap) {
1625 struct ceph_cap_flush *cf;
1627 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1628 if (inode->i_data.nrpages > 0)
1630 if (ci->i_wrbuffer_ref > 0)
1631 mapping_set_error(&inode->i_data, -EIO);
1634 while (!list_empty(&ci->i_cap_flush_list)) {
1635 cf = list_first_entry(&ci->i_cap_flush_list,
1636 struct ceph_cap_flush, i_list);
1637 list_move(&cf->i_list, &to_remove);
1640 spin_lock(&mdsc->cap_dirty_lock);
1642 list_for_each_entry(cf, &to_remove, i_list)
1643 list_del_init(&cf->g_list);
1645 if (!list_empty(&ci->i_dirty_item)) {
1646 pr_warn_ratelimited(
1647 " dropping dirty %s state for %p %lld\n",
1648 ceph_cap_string(ci->i_dirty_caps),
1649 inode, ceph_ino(inode));
1650 ci->i_dirty_caps = 0;
1651 list_del_init(&ci->i_dirty_item);
1652 dirty_dropped = true;
1654 if (!list_empty(&ci->i_flushing_item)) {
1655 pr_warn_ratelimited(
1656 " dropping dirty+flushing %s state for %p %lld\n",
1657 ceph_cap_string(ci->i_flushing_caps),
1658 inode, ceph_ino(inode));
1659 ci->i_flushing_caps = 0;
1660 list_del_init(&ci->i_flushing_item);
1661 mdsc->num_cap_flushing--;
1662 dirty_dropped = true;
1664 spin_unlock(&mdsc->cap_dirty_lock);
1666 if (dirty_dropped) {
1667 mapping_set_error(inode->i_mapping, -EIO);
1669 if (ci->i_wrbuffer_ref_head == 0 &&
1670 ci->i_wr_ref == 0 &&
1671 ci->i_dirty_caps == 0 &&
1672 ci->i_flushing_caps == 0) {
1673 ceph_put_snap_context(ci->i_head_snapc);
1674 ci->i_head_snapc = NULL;
1678 if (atomic_read(&ci->i_filelock_ref) > 0) {
1679 /* make further file lock syscall return -EIO */
1680 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1681 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1682 inode, ceph_ino(inode));
1685 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1686 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1687 ci->i_prealloc_cap_flush = NULL;
1690 if (!list_empty(&ci->i_cap_snaps))
1691 capsnap_release = remove_capsnaps(mdsc, inode);
1693 spin_unlock(&ci->i_ceph_lock);
1694 while (!list_empty(&to_remove)) {
1695 struct ceph_cap_flush *cf;
1696 cf = list_first_entry(&to_remove,
1697 struct ceph_cap_flush, i_list);
1698 list_del_init(&cf->i_list);
1699 if (!cf->is_capsnap)
1700 ceph_free_cap_flush(cf);
1703 wake_up_all(&ci->i_cap_wq);
1705 ceph_queue_invalidate(inode);
1708 while (capsnap_release--)
1714 * caller must hold session s_mutex
1716 static void remove_session_caps(struct ceph_mds_session *session)
1718 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1719 struct super_block *sb = fsc->sb;
1722 dout("remove_session_caps on %p\n", session);
1723 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1725 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1727 spin_lock(&session->s_cap_lock);
1728 if (session->s_nr_caps > 0) {
1729 struct inode *inode;
1730 struct ceph_cap *cap, *prev = NULL;
1731 struct ceph_vino vino;
1733 * iterate_session_caps() skips inodes that are being
1734 * deleted, we need to wait until deletions are complete.
1735 * __wait_on_freeing_inode() is designed for the job,
1736 * but it is not exported, so use lookup inode function
1739 while (!list_empty(&session->s_caps)) {
1740 cap = list_entry(session->s_caps.next,
1741 struct ceph_cap, session_caps);
1745 vino = cap->ci->i_vino;
1746 spin_unlock(&session->s_cap_lock);
1748 inode = ceph_find_inode(sb, vino);
1749 /* avoid calling iput_final() while holding s_mutex */
1750 ceph_async_iput(inode);
1752 spin_lock(&session->s_cap_lock);
1756 // drop cap expires and unlock s_cap_lock
1757 detach_cap_releases(session, &dispose);
1759 BUG_ON(session->s_nr_caps > 0);
1760 BUG_ON(!list_empty(&session->s_cap_flushing));
1761 spin_unlock(&session->s_cap_lock);
1762 dispose_cap_releases(session->s_mdsc, &dispose);
1772 * wake up any threads waiting on this session's caps. if the cap is
1773 * old (didn't get renewed on the client reconnect), remove it now.
1775 * caller must hold s_mutex.
1777 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1780 struct ceph_inode_info *ci = ceph_inode(inode);
1781 unsigned long ev = (unsigned long)arg;
1783 if (ev == RECONNECT) {
1784 spin_lock(&ci->i_ceph_lock);
1785 ci->i_wanted_max_size = 0;
1786 ci->i_requested_max_size = 0;
1787 spin_unlock(&ci->i_ceph_lock);
1788 } else if (ev == RENEWCAPS) {
1789 if (cap->cap_gen < cap->session->s_cap_gen) {
1790 /* mds did not re-issue stale cap */
1791 spin_lock(&ci->i_ceph_lock);
1792 cap->issued = cap->implemented = CEPH_CAP_PIN;
1793 spin_unlock(&ci->i_ceph_lock);
1795 } else if (ev == FORCE_RO) {
1797 wake_up_all(&ci->i_cap_wq);
1801 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1803 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1804 ceph_iterate_session_caps(session, wake_up_session_cb,
1805 (void *)(unsigned long)ev);
1809 * Send periodic message to MDS renewing all currently held caps. The
1810 * ack will reset the expiration for all caps from this session.
1812 * caller holds s_mutex
1814 static int send_renew_caps(struct ceph_mds_client *mdsc,
1815 struct ceph_mds_session *session)
1817 struct ceph_msg *msg;
1820 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1821 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1822 pr_info("mds%d caps stale\n", session->s_mds);
1823 session->s_renew_requested = jiffies;
1825 /* do not try to renew caps until a recovering mds has reconnected
1826 * with its clients. */
1827 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1828 if (state < CEPH_MDS_STATE_RECONNECT) {
1829 dout("send_renew_caps ignoring mds%d (%s)\n",
1830 session->s_mds, ceph_mds_state_name(state));
1834 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1835 ceph_mds_state_name(state));
1836 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1837 ++session->s_renew_seq);
1840 ceph_con_send(&session->s_con, msg);
1844 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1845 struct ceph_mds_session *session, u64 seq)
1847 struct ceph_msg *msg;
1849 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1850 session->s_mds, ceph_session_state_name(session->s_state), seq);
1851 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1854 ceph_con_send(&session->s_con, msg);
1860 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1862 * Called under session->s_mutex
1864 static void renewed_caps(struct ceph_mds_client *mdsc,
1865 struct ceph_mds_session *session, int is_renew)
1870 spin_lock(&session->s_cap_lock);
1871 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1873 session->s_cap_ttl = session->s_renew_requested +
1874 mdsc->mdsmap->m_session_timeout*HZ;
1877 if (time_before(jiffies, session->s_cap_ttl)) {
1878 pr_info("mds%d caps renewed\n", session->s_mds);
1881 pr_info("mds%d caps still stale\n", session->s_mds);
1884 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1885 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1886 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1887 spin_unlock(&session->s_cap_lock);
1890 wake_up_session_caps(session, RENEWCAPS);
1894 * send a session close request
1896 static int request_close_session(struct ceph_mds_session *session)
1898 struct ceph_msg *msg;
1900 dout("request_close_session mds%d state %s seq %lld\n",
1901 session->s_mds, ceph_session_state_name(session->s_state),
1903 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1906 ceph_con_send(&session->s_con, msg);
1911 * Called with s_mutex held.
1913 static int __close_session(struct ceph_mds_client *mdsc,
1914 struct ceph_mds_session *session)
1916 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1918 session->s_state = CEPH_MDS_SESSION_CLOSING;
1919 return request_close_session(session);
1922 static bool drop_negative_children(struct dentry *dentry)
1924 struct dentry *child;
1925 bool all_negative = true;
1927 if (!d_is_dir(dentry))
1930 spin_lock(&dentry->d_lock);
1931 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1932 if (d_really_is_positive(child)) {
1933 all_negative = false;
1937 spin_unlock(&dentry->d_lock);
1940 shrink_dcache_parent(dentry);
1942 return all_negative;
1946 * Trim old(er) caps.
1948 * Because we can't cache an inode without one or more caps, we do
1949 * this indirectly: if a cap is unused, we prune its aliases, at which
1950 * point the inode will hopefully get dropped to.
1952 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1953 * memory pressure from the MDS, though, so it needn't be perfect.
1955 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1957 int *remaining = arg;
1958 struct ceph_inode_info *ci = ceph_inode(inode);
1959 int used, wanted, oissued, mine;
1961 if (*remaining <= 0)
1964 spin_lock(&ci->i_ceph_lock);
1965 mine = cap->issued | cap->implemented;
1966 used = __ceph_caps_used(ci);
1967 wanted = __ceph_caps_file_wanted(ci);
1968 oissued = __ceph_caps_issued_other(ci, cap);
1970 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1971 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1972 ceph_cap_string(used), ceph_cap_string(wanted));
1973 if (cap == ci->i_auth_cap) {
1974 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1975 !list_empty(&ci->i_cap_snaps))
1977 if ((used | wanted) & CEPH_CAP_ANY_WR)
1979 /* Note: it's possible that i_filelock_ref becomes non-zero
1980 * after dropping auth caps. It doesn't hurt because reply
1981 * of lock mds request will re-add auth caps. */
1982 if (atomic_read(&ci->i_filelock_ref) > 0)
1985 /* The inode has cached pages, but it's no longer used.
1986 * we can safely drop it */
1987 if (S_ISREG(inode->i_mode) &&
1988 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1989 !(oissued & CEPH_CAP_FILE_CACHE)) {
1993 if ((used | wanted) & ~oissued & mine)
1994 goto out; /* we need these caps */
1997 /* we aren't the only cap.. just remove us */
1998 __ceph_remove_cap(cap, true);
2001 struct dentry *dentry;
2002 /* try dropping referring dentries */
2003 spin_unlock(&ci->i_ceph_lock);
2004 dentry = d_find_any_alias(inode);
2005 if (dentry && drop_negative_children(dentry)) {
2008 d_prune_aliases(inode);
2009 count = atomic_read(&inode->i_count);
2012 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
2021 spin_unlock(&ci->i_ceph_lock);
2026 * Trim session cap count down to some max number.
2028 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2029 struct ceph_mds_session *session,
2032 int trim_caps = session->s_nr_caps - max_caps;
2034 dout("trim_caps mds%d start: %d / %d, trim %d\n",
2035 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
2036 if (trim_caps > 0) {
2037 int remaining = trim_caps;
2039 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2040 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2041 session->s_mds, session->s_nr_caps, max_caps,
2042 trim_caps - remaining);
2045 ceph_flush_cap_releases(mdsc, session);
2049 static int check_caps_flush(struct ceph_mds_client *mdsc,
2054 spin_lock(&mdsc->cap_dirty_lock);
2055 if (!list_empty(&mdsc->cap_flush_list)) {
2056 struct ceph_cap_flush *cf =
2057 list_first_entry(&mdsc->cap_flush_list,
2058 struct ceph_cap_flush, g_list);
2059 if (cf->tid <= want_flush_tid) {
2060 dout("check_caps_flush still flushing tid "
2061 "%llu <= %llu\n", cf->tid, want_flush_tid);
2065 spin_unlock(&mdsc->cap_dirty_lock);
2070 * flush all dirty inode data to disk.
2072 * returns true if we've flushed through want_flush_tid
2074 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2077 dout("check_caps_flush want %llu\n", want_flush_tid);
2079 wait_event(mdsc->cap_flushing_wq,
2080 check_caps_flush(mdsc, want_flush_tid));
2082 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2086 * called under s_mutex
2088 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2089 struct ceph_mds_session *session)
2091 struct ceph_msg *msg = NULL;
2092 struct ceph_mds_cap_release *head;
2093 struct ceph_mds_cap_item *item;
2094 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2095 struct ceph_cap *cap;
2096 LIST_HEAD(tmp_list);
2097 int num_cap_releases;
2098 __le32 barrier, *cap_barrier;
2100 down_read(&osdc->lock);
2101 barrier = cpu_to_le32(osdc->epoch_barrier);
2102 up_read(&osdc->lock);
2104 spin_lock(&session->s_cap_lock);
2106 list_splice_init(&session->s_cap_releases, &tmp_list);
2107 num_cap_releases = session->s_num_cap_releases;
2108 session->s_num_cap_releases = 0;
2109 spin_unlock(&session->s_cap_lock);
2111 while (!list_empty(&tmp_list)) {
2113 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2114 PAGE_SIZE, GFP_NOFS, false);
2117 head = msg->front.iov_base;
2118 head->num = cpu_to_le32(0);
2119 msg->front.iov_len = sizeof(*head);
2121 msg->hdr.version = cpu_to_le16(2);
2122 msg->hdr.compat_version = cpu_to_le16(1);
2125 cap = list_first_entry(&tmp_list, struct ceph_cap,
2127 list_del(&cap->session_caps);
2130 head = msg->front.iov_base;
2131 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2133 item = msg->front.iov_base + msg->front.iov_len;
2134 item->ino = cpu_to_le64(cap->cap_ino);
2135 item->cap_id = cpu_to_le64(cap->cap_id);
2136 item->migrate_seq = cpu_to_le32(cap->mseq);
2137 item->seq = cpu_to_le32(cap->issue_seq);
2138 msg->front.iov_len += sizeof(*item);
2140 ceph_put_cap(mdsc, cap);
2142 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2143 // Append cap_barrier field
2144 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2145 *cap_barrier = barrier;
2146 msg->front.iov_len += sizeof(*cap_barrier);
2148 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2149 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2150 ceph_con_send(&session->s_con, msg);
2155 BUG_ON(num_cap_releases != 0);
2157 spin_lock(&session->s_cap_lock);
2158 if (!list_empty(&session->s_cap_releases))
2160 spin_unlock(&session->s_cap_lock);
2163 // Append cap_barrier field
2164 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2165 *cap_barrier = barrier;
2166 msg->front.iov_len += sizeof(*cap_barrier);
2168 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2169 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2170 ceph_con_send(&session->s_con, msg);
2174 pr_err("send_cap_releases mds%d, failed to allocate message\n",
2176 spin_lock(&session->s_cap_lock);
2177 list_splice(&tmp_list, &session->s_cap_releases);
2178 session->s_num_cap_releases += num_cap_releases;
2179 spin_unlock(&session->s_cap_lock);
2182 static void ceph_cap_release_work(struct work_struct *work)
2184 struct ceph_mds_session *session =
2185 container_of(work, struct ceph_mds_session, s_cap_release_work);
2187 mutex_lock(&session->s_mutex);
2188 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2189 session->s_state == CEPH_MDS_SESSION_HUNG)
2190 ceph_send_cap_releases(session->s_mdsc, session);
2191 mutex_unlock(&session->s_mutex);
2192 ceph_put_mds_session(session);
2195 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2196 struct ceph_mds_session *session)
2201 ceph_get_mds_session(session);
2202 if (queue_work(mdsc->fsc->cap_wq,
2203 &session->s_cap_release_work)) {
2204 dout("cap release work queued\n");
2206 ceph_put_mds_session(session);
2207 dout("failed to queue cap release work\n");
2212 * caller holds session->s_cap_lock
2214 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2215 struct ceph_cap *cap)
2217 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2218 session->s_num_cap_releases++;
2220 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2221 ceph_flush_cap_releases(session->s_mdsc, session);
2224 static void ceph_cap_reclaim_work(struct work_struct *work)
2226 struct ceph_mds_client *mdsc =
2227 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2228 int ret = ceph_trim_dentries(mdsc);
2230 ceph_queue_cap_reclaim_work(mdsc);
2233 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2238 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2239 dout("caps reclaim work queued\n");
2241 dout("failed to queue caps release work\n");
2245 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2250 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2251 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2252 atomic_set(&mdsc->cap_reclaim_pending, 0);
2253 ceph_queue_cap_reclaim_work(mdsc);
2261 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2264 struct ceph_inode_info *ci = ceph_inode(dir);
2265 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2266 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2267 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2268 unsigned int num_entries;
2271 spin_lock(&ci->i_ceph_lock);
2272 num_entries = ci->i_files + ci->i_subdirs;
2273 spin_unlock(&ci->i_ceph_lock);
2274 num_entries = max(num_entries, 1U);
2275 num_entries = min(num_entries, opt->max_readdir);
2277 order = get_order(size * num_entries);
2278 while (order >= 0) {
2279 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2282 if (rinfo->dir_entries)
2286 if (!rinfo->dir_entries)
2289 num_entries = (PAGE_SIZE << order) / size;
2290 num_entries = min(num_entries, opt->max_readdir);
2292 rinfo->dir_buf_size = PAGE_SIZE << order;
2293 req->r_num_caps = num_entries + 1;
2294 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2295 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2300 * Create an mds request.
2302 struct ceph_mds_request *
2303 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2305 struct ceph_mds_request *req;
2307 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2309 return ERR_PTR(-ENOMEM);
2311 mutex_init(&req->r_fill_mutex);
2313 req->r_started = jiffies;
2314 req->r_start_latency = ktime_get();
2315 req->r_resend_mds = -1;
2316 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2317 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2319 kref_init(&req->r_kref);
2320 RB_CLEAR_NODE(&req->r_node);
2321 INIT_LIST_HEAD(&req->r_wait);
2322 init_completion(&req->r_completion);
2323 init_completion(&req->r_safe_completion);
2324 INIT_LIST_HEAD(&req->r_unsafe_item);
2326 ktime_get_coarse_real_ts64(&req->r_stamp);
2329 req->r_direct_mode = mode;
2334 * return oldest (lowest) request, tid in request tree, 0 if none.
2336 * called under mdsc->mutex.
2338 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2340 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2342 return rb_entry(rb_first(&mdsc->request_tree),
2343 struct ceph_mds_request, r_node);
2346 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2348 return mdsc->oldest_tid;
2352 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2353 * on build_path_from_dentry in fs/cifs/dir.c.
2355 * If @stop_on_nosnap, generate path relative to the first non-snapped
2358 * Encode hidden .snap dirs as a double /, i.e.
2359 * foo/.snap/bar -> foo//bar
2361 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2364 struct dentry *temp;
2371 return ERR_PTR(-EINVAL);
2375 return ERR_PTR(-ENOMEM);
2380 seq = read_seqbegin(&rename_lock);
2384 struct inode *inode;
2386 spin_lock(&temp->d_lock);
2387 inode = d_inode(temp);
2388 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2389 dout("build_path path+%d: %p SNAPDIR\n",
2391 } else if (stop_on_nosnap && inode && dentry != temp &&
2392 ceph_snap(inode) == CEPH_NOSNAP) {
2393 spin_unlock(&temp->d_lock);
2394 pos++; /* get rid of any prepended '/' */
2397 pos -= temp->d_name.len;
2399 spin_unlock(&temp->d_lock);
2402 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2404 spin_unlock(&temp->d_lock);
2405 temp = READ_ONCE(temp->d_parent);
2407 /* Are we at the root? */
2411 /* Are we out of buffer? */
2417 base = ceph_ino(d_inode(temp));
2420 if (read_seqretry(&rename_lock, seq))
2425 * A rename didn't occur, but somehow we didn't end up where
2426 * we thought we would. Throw a warning and try again.
2428 pr_warn("build_path did not end path lookup where "
2429 "expected, pos is %d\n", pos);
2434 *plen = PATH_MAX - 1 - pos;
2435 dout("build_path on %p %d built %llx '%.*s'\n",
2436 dentry, d_count(dentry), base, *plen, path + pos);
2440 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2441 const char **ppath, int *ppathlen, u64 *pino,
2442 bool *pfreepath, bool parent_locked)
2448 dir = d_inode_rcu(dentry->d_parent);
2449 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2450 *pino = ceph_ino(dir);
2452 *ppath = dentry->d_name.name;
2453 *ppathlen = dentry->d_name.len;
2457 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2459 return PTR_ERR(path);
2465 static int build_inode_path(struct inode *inode,
2466 const char **ppath, int *ppathlen, u64 *pino,
2469 struct dentry *dentry;
2472 if (ceph_snap(inode) == CEPH_NOSNAP) {
2473 *pino = ceph_ino(inode);
2477 dentry = d_find_alias(inode);
2478 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2481 return PTR_ERR(path);
2488 * request arguments may be specified via an inode *, a dentry *, or
2489 * an explicit ino+path.
2491 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2492 struct inode *rdiri, const char *rpath,
2493 u64 rino, const char **ppath, int *pathlen,
2494 u64 *ino, bool *freepath, bool parent_locked)
2499 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2500 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2502 } else if (rdentry) {
2503 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2504 freepath, parent_locked);
2505 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2507 } else if (rpath || rino) {
2510 *pathlen = rpath ? strlen(rpath) : 0;
2511 dout(" path %.*s\n", *pathlen, rpath);
2518 * called under mdsc->mutex
2520 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2521 struct ceph_mds_request *req,
2522 int mds, bool drop_cap_releases)
2524 struct ceph_msg *msg;
2525 struct ceph_mds_request_head *head;
2526 const char *path1 = NULL;
2527 const char *path2 = NULL;
2528 u64 ino1 = 0, ino2 = 0;
2529 int pathlen1 = 0, pathlen2 = 0;
2530 bool freepath1 = false, freepath2 = false;
2536 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2537 req->r_parent, req->r_path1, req->r_ino1.ino,
2538 &path1, &pathlen1, &ino1, &freepath1,
2539 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2540 &req->r_req_flags));
2546 /* If r_old_dentry is set, then assume that its parent is locked */
2547 ret = set_request_path_attr(NULL, req->r_old_dentry,
2548 req->r_old_dentry_dir,
2549 req->r_path2, req->r_ino2.ino,
2550 &path2, &pathlen2, &ino2, &freepath2, true);
2556 len = sizeof(*head) +
2557 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2558 sizeof(struct ceph_timespec);
2560 /* calculate (max) length for cap releases */
2561 len += sizeof(struct ceph_mds_request_release) *
2562 (!!req->r_inode_drop + !!req->r_dentry_drop +
2563 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2564 if (req->r_dentry_drop)
2566 if (req->r_old_dentry_drop)
2569 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2571 msg = ERR_PTR(-ENOMEM);
2575 msg->hdr.version = cpu_to_le16(2);
2576 msg->hdr.tid = cpu_to_le64(req->r_tid);
2578 head = msg->front.iov_base;
2579 p = msg->front.iov_base + sizeof(*head);
2580 end = msg->front.iov_base + msg->front.iov_len;
2582 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2583 head->op = cpu_to_le32(req->r_op);
2584 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2585 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2586 head->ino = cpu_to_le64(req->r_deleg_ino);
2587 head->args = req->r_args;
2589 ceph_encode_filepath(&p, end, ino1, path1);
2590 ceph_encode_filepath(&p, end, ino2, path2);
2592 /* make note of release offset, in case we need to replay */
2593 req->r_request_release_offset = p - msg->front.iov_base;
2597 if (req->r_inode_drop)
2598 releases += ceph_encode_inode_release(&p,
2599 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2600 mds, req->r_inode_drop, req->r_inode_unless,
2601 req->r_op == CEPH_MDS_OP_READDIR);
2602 if (req->r_dentry_drop)
2603 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2604 req->r_parent, mds, req->r_dentry_drop,
2605 req->r_dentry_unless);
2606 if (req->r_old_dentry_drop)
2607 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2608 req->r_old_dentry_dir, mds,
2609 req->r_old_dentry_drop,
2610 req->r_old_dentry_unless);
2611 if (req->r_old_inode_drop)
2612 releases += ceph_encode_inode_release(&p,
2613 d_inode(req->r_old_dentry),
2614 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2616 if (drop_cap_releases) {
2618 p = msg->front.iov_base + req->r_request_release_offset;
2621 head->num_releases = cpu_to_le16(releases);
2625 struct ceph_timespec ts;
2626 ceph_encode_timespec64(&ts, &req->r_stamp);
2627 ceph_encode_copy(&p, &ts, sizeof(ts));
2630 if (WARN_ON_ONCE(p > end)) {
2632 msg = ERR_PTR(-ERANGE);
2636 msg->front.iov_len = p - msg->front.iov_base;
2637 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2639 if (req->r_pagelist) {
2640 struct ceph_pagelist *pagelist = req->r_pagelist;
2641 ceph_msg_data_add_pagelist(msg, pagelist);
2642 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2644 msg->hdr.data_len = 0;
2647 msg->hdr.data_off = cpu_to_le16(0);
2651 ceph_mdsc_free_path((char *)path2, pathlen2);
2654 ceph_mdsc_free_path((char *)path1, pathlen1);
2660 * called under mdsc->mutex if error, under no mutex if
2663 static void complete_request(struct ceph_mds_client *mdsc,
2664 struct ceph_mds_request *req)
2666 req->r_end_latency = ktime_get();
2668 if (req->r_callback)
2669 req->r_callback(mdsc, req);
2670 complete_all(&req->r_completion);
2674 * called under mdsc->mutex
2676 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2677 struct ceph_mds_request *req,
2678 int mds, bool drop_cap_releases)
2680 struct ceph_mds_request_head *rhead;
2681 struct ceph_msg *msg;
2686 struct ceph_cap *cap =
2687 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2690 req->r_sent_on_mseq = cap->mseq;
2692 req->r_sent_on_mseq = -1;
2694 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2695 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2697 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2700 * Replay. Do not regenerate message (and rebuild
2701 * paths, etc.); just use the original message.
2702 * Rebuilding paths will break for renames because
2703 * d_move mangles the src name.
2705 msg = req->r_request;
2706 rhead = msg->front.iov_base;
2708 flags = le32_to_cpu(rhead->flags);
2709 flags |= CEPH_MDS_FLAG_REPLAY;
2710 rhead->flags = cpu_to_le32(flags);
2712 if (req->r_target_inode)
2713 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2715 rhead->num_retry = req->r_attempts - 1;
2717 /* remove cap/dentry releases from message */
2718 rhead->num_releases = 0;
2721 p = msg->front.iov_base + req->r_request_release_offset;
2723 struct ceph_timespec ts;
2724 ceph_encode_timespec64(&ts, &req->r_stamp);
2725 ceph_encode_copy(&p, &ts, sizeof(ts));
2728 msg->front.iov_len = p - msg->front.iov_base;
2729 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2733 if (req->r_request) {
2734 ceph_msg_put(req->r_request);
2735 req->r_request = NULL;
2737 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2739 req->r_err = PTR_ERR(msg);
2740 return PTR_ERR(msg);
2742 req->r_request = msg;
2744 rhead = msg->front.iov_base;
2745 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2746 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2747 flags |= CEPH_MDS_FLAG_REPLAY;
2748 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2749 flags |= CEPH_MDS_FLAG_ASYNC;
2751 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2752 rhead->flags = cpu_to_le32(flags);
2753 rhead->num_fwd = req->r_num_fwd;
2754 rhead->num_retry = req->r_attempts - 1;
2756 dout(" r_parent = %p\n", req->r_parent);
2761 * called under mdsc->mutex
2763 static int __send_request(struct ceph_mds_client *mdsc,
2764 struct ceph_mds_session *session,
2765 struct ceph_mds_request *req,
2766 bool drop_cap_releases)
2770 err = __prepare_send_request(mdsc, req, session->s_mds,
2773 ceph_msg_get(req->r_request);
2774 ceph_con_send(&session->s_con, req->r_request);
2781 * send request, or put it on the appropriate wait list.
2783 static void __do_request(struct ceph_mds_client *mdsc,
2784 struct ceph_mds_request *req)
2786 struct ceph_mds_session *session = NULL;
2791 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2792 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2793 __unregister_request(mdsc, req);
2797 if (req->r_timeout &&
2798 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2799 dout("do_request timed out\n");
2803 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2804 dout("do_request forced umount\n");
2808 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2809 if (mdsc->mdsmap_err) {
2810 err = mdsc->mdsmap_err;
2811 dout("do_request mdsmap err %d\n", err);
2814 if (mdsc->mdsmap->m_epoch == 0) {
2815 dout("do_request no mdsmap, waiting for map\n");
2816 list_add(&req->r_wait, &mdsc->waiting_for_map);
2819 if (!(mdsc->fsc->mount_options->flags &
2820 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2821 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2822 err = -EHOSTUNREACH;
2827 put_request_session(req);
2829 mds = __choose_mds(mdsc, req, &random);
2831 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2832 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2836 dout("do_request no mds or not active, waiting for map\n");
2837 list_add(&req->r_wait, &mdsc->waiting_for_map);
2841 /* get, open session */
2842 session = __ceph_lookup_mds_session(mdsc, mds);
2844 session = register_session(mdsc, mds);
2845 if (IS_ERR(session)) {
2846 err = PTR_ERR(session);
2850 req->r_session = ceph_get_mds_session(session);
2852 dout("do_request mds%d session %p state %s\n", mds, session,
2853 ceph_session_state_name(session->s_state));
2854 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2855 session->s_state != CEPH_MDS_SESSION_HUNG) {
2856 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2861 * We cannot queue async requests since the caps and delegated
2862 * inodes are bound to the session. Just return -EJUKEBOX and
2863 * let the caller retry a sync request in that case.
2865 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2869 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2870 session->s_state == CEPH_MDS_SESSION_CLOSING) {
2871 err = __open_session(mdsc, session);
2874 /* retry the same mds later */
2876 req->r_resend_mds = mds;
2878 list_add(&req->r_wait, &session->s_waiting);
2883 req->r_resend_mds = -1; /* forget any previous mds hint */
2885 if (req->r_request_started == 0) /* note request start time */
2886 req->r_request_started = jiffies;
2888 err = __send_request(mdsc, session, req, false);
2891 ceph_put_mds_session(session);
2894 dout("__do_request early error %d\n", err);
2896 complete_request(mdsc, req);
2897 __unregister_request(mdsc, req);
2903 * called under mdsc->mutex
2905 static void __wake_requests(struct ceph_mds_client *mdsc,
2906 struct list_head *head)
2908 struct ceph_mds_request *req;
2909 LIST_HEAD(tmp_list);
2911 list_splice_init(head, &tmp_list);
2913 while (!list_empty(&tmp_list)) {
2914 req = list_entry(tmp_list.next,
2915 struct ceph_mds_request, r_wait);
2916 list_del_init(&req->r_wait);
2917 dout(" wake request %p tid %llu\n", req, req->r_tid);
2918 __do_request(mdsc, req);
2923 * Wake up threads with requests pending for @mds, so that they can
2924 * resubmit their requests to a possibly different mds.
2926 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2928 struct ceph_mds_request *req;
2929 struct rb_node *p = rb_first(&mdsc->request_tree);
2931 dout("kick_requests mds%d\n", mds);
2933 req = rb_entry(p, struct ceph_mds_request, r_node);
2935 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2937 if (req->r_attempts > 0)
2938 continue; /* only new requests */
2939 if (req->r_session &&
2940 req->r_session->s_mds == mds) {
2941 dout(" kicking tid %llu\n", req->r_tid);
2942 list_del_init(&req->r_wait);
2943 __do_request(mdsc, req);
2948 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2949 struct ceph_mds_request *req)
2953 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2955 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2956 if (req->r_parent) {
2957 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2958 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2959 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2960 spin_lock(&ci->i_ceph_lock);
2961 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2962 __ceph_touch_fmode(ci, mdsc, fmode);
2963 spin_unlock(&ci->i_ceph_lock);
2964 ihold(req->r_parent);
2966 if (req->r_old_dentry_dir)
2967 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2971 err = ceph_wait_on_async_create(req->r_inode);
2973 dout("%s: wait for async create returned: %d\n",
2979 if (!err && req->r_old_inode) {
2980 err = ceph_wait_on_async_create(req->r_old_inode);
2982 dout("%s: wait for async create returned: %d\n",
2988 dout("submit_request on %p for inode %p\n", req, dir);
2989 mutex_lock(&mdsc->mutex);
2990 __register_request(mdsc, req, dir);
2991 __do_request(mdsc, req);
2993 mutex_unlock(&mdsc->mutex);
2997 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2998 struct ceph_mds_request *req)
3003 dout("do_request waiting\n");
3004 if (!req->r_timeout && req->r_wait_for_completion) {
3005 err = req->r_wait_for_completion(mdsc, req);
3007 long timeleft = wait_for_completion_killable_timeout(
3009 ceph_timeout_jiffies(req->r_timeout));
3013 err = -ETIMEDOUT; /* timed out */
3015 err = timeleft; /* killed */
3017 dout("do_request waited, got %d\n", err);
3018 mutex_lock(&mdsc->mutex);
3020 /* only abort if we didn't race with a real reply */
3021 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3022 err = le32_to_cpu(req->r_reply_info.head->result);
3023 } else if (err < 0) {
3024 dout("aborted request %lld with %d\n", req->r_tid, err);
3027 * ensure we aren't running concurrently with
3028 * ceph_fill_trace or ceph_readdir_prepopulate, which
3029 * rely on locks (dir mutex) held by our caller.
3031 mutex_lock(&req->r_fill_mutex);
3033 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3034 mutex_unlock(&req->r_fill_mutex);
3036 if (req->r_parent &&
3037 (req->r_op & CEPH_MDS_OP_WRITE))
3038 ceph_invalidate_dir_request(req);
3043 mutex_unlock(&mdsc->mutex);
3048 * Synchrously perform an mds request. Take care of all of the
3049 * session setup, forwarding, retry details.
3051 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3053 struct ceph_mds_request *req)
3057 dout("do_request on %p\n", req);
3060 err = ceph_mdsc_submit_request(mdsc, dir, req);
3062 err = ceph_mdsc_wait_request(mdsc, req);
3063 dout("do_request %p done, result %d\n", req, err);
3068 * Invalidate dir's completeness, dentry lease state on an aborted MDS
3069 * namespace request.
3071 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3073 struct inode *dir = req->r_parent;
3074 struct inode *old_dir = req->r_old_dentry_dir;
3076 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3078 ceph_dir_clear_complete(dir);
3080 ceph_dir_clear_complete(old_dir);
3082 ceph_invalidate_dentry_lease(req->r_dentry);
3083 if (req->r_old_dentry)
3084 ceph_invalidate_dentry_lease(req->r_old_dentry);
3090 * We take the session mutex and parse and process the reply immediately.
3091 * This preserves the logical ordering of replies, capabilities, etc., sent
3092 * by the MDS as they are applied to our local cache.
3094 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3096 struct ceph_mds_client *mdsc = session->s_mdsc;
3097 struct ceph_mds_request *req;
3098 struct ceph_mds_reply_head *head = msg->front.iov_base;
3099 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
3100 struct ceph_snap_realm *realm;
3103 int mds = session->s_mds;
3105 if (msg->front.iov_len < sizeof(*head)) {
3106 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3111 /* get request, session */
3112 tid = le64_to_cpu(msg->hdr.tid);
3113 mutex_lock(&mdsc->mutex);
3114 req = lookup_get_request(mdsc, tid);
3116 dout("handle_reply on unknown tid %llu\n", tid);
3117 mutex_unlock(&mdsc->mutex);
3120 dout("handle_reply %p\n", req);
3122 /* correct session? */
3123 if (req->r_session != session) {
3124 pr_err("mdsc_handle_reply got %llu on session mds%d"
3125 " not mds%d\n", tid, session->s_mds,
3126 req->r_session ? req->r_session->s_mds : -1);
3127 mutex_unlock(&mdsc->mutex);
3132 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3133 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3134 pr_warn("got a dup %s reply on %llu from mds%d\n",
3135 head->safe ? "safe" : "unsafe", tid, mds);
3136 mutex_unlock(&mdsc->mutex);
3139 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3140 pr_warn("got unsafe after safe on %llu from mds%d\n",
3142 mutex_unlock(&mdsc->mutex);
3146 result = le32_to_cpu(head->result);
3150 * if we're not talking to the authority, send to them
3151 * if the authority has changed while we weren't looking,
3152 * send to new authority
3153 * Otherwise we just have to return an ESTALE
3155 if (result == -ESTALE) {
3156 dout("got ESTALE on request %llu\n", req->r_tid);
3157 req->r_resend_mds = -1;
3158 if (req->r_direct_mode != USE_AUTH_MDS) {
3159 dout("not using auth, setting for that now\n");
3160 req->r_direct_mode = USE_AUTH_MDS;
3161 __do_request(mdsc, req);
3162 mutex_unlock(&mdsc->mutex);
3165 int mds = __choose_mds(mdsc, req, NULL);
3166 if (mds >= 0 && mds != req->r_session->s_mds) {
3167 dout("but auth changed, so resending\n");
3168 __do_request(mdsc, req);
3169 mutex_unlock(&mdsc->mutex);
3173 dout("have to return ESTALE on request %llu\n", req->r_tid);
3178 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3179 __unregister_request(mdsc, req);
3181 /* last request during umount? */
3182 if (mdsc->stopping && !__get_oldest_req(mdsc))
3183 complete_all(&mdsc->safe_umount_waiters);
3185 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3187 * We already handled the unsafe response, now do the
3188 * cleanup. No need to examine the response; the MDS
3189 * doesn't include any result info in the safe
3190 * response. And even if it did, there is nothing
3191 * useful we could do with a revised return value.
3193 dout("got safe reply %llu, mds%d\n", tid, mds);
3195 mutex_unlock(&mdsc->mutex);
3199 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3200 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3203 dout("handle_reply tid %lld result %d\n", tid, result);
3204 rinfo = &req->r_reply_info;
3205 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3206 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3208 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3209 mutex_unlock(&mdsc->mutex);
3211 mutex_lock(&session->s_mutex);
3213 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3220 if (rinfo->snapblob_len) {
3221 down_write(&mdsc->snap_rwsem);
3222 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3223 rinfo->snapblob + rinfo->snapblob_len,
3224 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3226 downgrade_write(&mdsc->snap_rwsem);
3228 down_read(&mdsc->snap_rwsem);
3231 /* insert trace into our cache */
3232 mutex_lock(&req->r_fill_mutex);
3233 current->journal_info = req;
3234 err = ceph_fill_trace(mdsc->fsc->sb, req);
3236 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3237 req->r_op == CEPH_MDS_OP_LSSNAP))
3238 ceph_readdir_prepopulate(req, req->r_session);
3240 current->journal_info = NULL;
3241 mutex_unlock(&req->r_fill_mutex);
3243 up_read(&mdsc->snap_rwsem);
3245 ceph_put_snap_realm(mdsc, realm);
3248 if (req->r_target_inode &&
3249 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3250 struct ceph_inode_info *ci =
3251 ceph_inode(req->r_target_inode);
3252 spin_lock(&ci->i_unsafe_lock);
3253 list_add_tail(&req->r_unsafe_target_item,
3254 &ci->i_unsafe_iops);
3255 spin_unlock(&ci->i_unsafe_lock);
3258 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3261 mutex_lock(&mdsc->mutex);
3262 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3266 req->r_reply = ceph_msg_get(msg);
3267 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3270 dout("reply arrived after request %lld was aborted\n", tid);
3272 mutex_unlock(&mdsc->mutex);
3274 mutex_unlock(&session->s_mutex);
3276 /* kick calling process */
3277 complete_request(mdsc, req);
3279 ceph_update_metadata_latency(&mdsc->metric, req->r_start_latency,
3280 req->r_end_latency, err);
3282 ceph_mdsc_put_request(req);
3289 * handle mds notification that our request has been forwarded.
3291 static void handle_forward(struct ceph_mds_client *mdsc,
3292 struct ceph_mds_session *session,
3293 struct ceph_msg *msg)
3295 struct ceph_mds_request *req;
3296 u64 tid = le64_to_cpu(msg->hdr.tid);
3300 void *p = msg->front.iov_base;
3301 void *end = p + msg->front.iov_len;
3303 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3304 next_mds = ceph_decode_32(&p);
3305 fwd_seq = ceph_decode_32(&p);
3307 mutex_lock(&mdsc->mutex);
3308 req = lookup_get_request(mdsc, tid);
3310 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3311 goto out; /* dup reply? */
3314 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3315 dout("forward tid %llu aborted, unregistering\n", tid);
3316 __unregister_request(mdsc, req);
3317 } else if (fwd_seq <= req->r_num_fwd) {
3318 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3319 tid, next_mds, req->r_num_fwd, fwd_seq);
3321 /* resend. forward race not possible; mds would drop */
3322 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3324 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3325 req->r_attempts = 0;
3326 req->r_num_fwd = fwd_seq;
3327 req->r_resend_mds = next_mds;
3328 put_request_session(req);
3329 __do_request(mdsc, req);
3331 ceph_mdsc_put_request(req);
3333 mutex_unlock(&mdsc->mutex);
3337 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3340 static int __decode_session_metadata(void **p, void *end,
3343 /* map<string,string> */
3346 ceph_decode_32_safe(p, end, n, bad);
3349 ceph_decode_32_safe(p, end, len, bad);
3350 ceph_decode_need(p, end, len, bad);
3351 err_str = !strncmp(*p, "error_string", len);
3353 ceph_decode_32_safe(p, end, len, bad);
3354 ceph_decode_need(p, end, len, bad);
3356 * Match "blocklisted (blacklisted)" from newer MDSes,
3357 * or "blacklisted" from older MDSes.
3359 if (err_str && strnstr(*p, "blacklisted", len))
3360 *blocklisted = true;
3369 * handle a mds session control message
3371 static void handle_session(struct ceph_mds_session *session,
3372 struct ceph_msg *msg)
3374 struct ceph_mds_client *mdsc = session->s_mdsc;
3375 int mds = session->s_mds;
3376 int msg_version = le16_to_cpu(msg->hdr.version);
3377 void *p = msg->front.iov_base;
3378 void *end = p + msg->front.iov_len;
3379 struct ceph_mds_session_head *h;
3381 u64 seq, features = 0;
3383 bool blocklisted = false;
3386 ceph_decode_need(&p, end, sizeof(*h), bad);
3390 op = le32_to_cpu(h->op);
3391 seq = le64_to_cpu(h->seq);
3393 if (msg_version >= 3) {
3395 /* version >= 2, metadata */
3396 if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3398 /* version >= 3, feature bits */
3399 ceph_decode_32_safe(&p, end, len, bad);
3401 ceph_decode_64_safe(&p, end, features, bad);
3402 p += len - sizeof(features);
3406 mutex_lock(&mdsc->mutex);
3407 if (op == CEPH_SESSION_CLOSE) {
3408 ceph_get_mds_session(session);
3409 __unregister_session(mdsc, session);
3411 /* FIXME: this ttl calculation is generous */
3412 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3413 mutex_unlock(&mdsc->mutex);
3415 mutex_lock(&session->s_mutex);
3417 dout("handle_session mds%d %s %p state %s seq %llu\n",
3418 mds, ceph_session_op_name(op), session,
3419 ceph_session_state_name(session->s_state), seq);
3421 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3422 session->s_state = CEPH_MDS_SESSION_OPEN;
3423 pr_info("mds%d came back\n", session->s_mds);
3427 case CEPH_SESSION_OPEN:
3428 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3429 pr_info("mds%d reconnect success\n", session->s_mds);
3430 session->s_state = CEPH_MDS_SESSION_OPEN;
3431 session->s_features = features;
3432 renewed_caps(mdsc, session, 0);
3433 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3434 metric_schedule_delayed(&mdsc->metric);
3437 __close_session(mdsc, session);
3440 case CEPH_SESSION_RENEWCAPS:
3441 if (session->s_renew_seq == seq)
3442 renewed_caps(mdsc, session, 1);
3445 case CEPH_SESSION_CLOSE:
3446 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3447 pr_info("mds%d reconnect denied\n", session->s_mds);
3448 session->s_state = CEPH_MDS_SESSION_CLOSED;
3449 cleanup_session_requests(mdsc, session);
3450 remove_session_caps(session);
3451 wake = 2; /* for good measure */
3452 wake_up_all(&mdsc->session_close_wq);
3455 case CEPH_SESSION_STALE:
3456 pr_info("mds%d caps went stale, renewing\n",
3458 spin_lock(&session->s_gen_ttl_lock);
3459 session->s_cap_gen++;
3460 session->s_cap_ttl = jiffies - 1;
3461 spin_unlock(&session->s_gen_ttl_lock);
3462 send_renew_caps(mdsc, session);
3465 case CEPH_SESSION_RECALL_STATE:
3466 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3469 case CEPH_SESSION_FLUSHMSG:
3470 send_flushmsg_ack(mdsc, session, seq);
3473 case CEPH_SESSION_FORCE_RO:
3474 dout("force_session_readonly %p\n", session);
3475 spin_lock(&session->s_cap_lock);
3476 session->s_readonly = true;
3477 spin_unlock(&session->s_cap_lock);
3478 wake_up_session_caps(session, FORCE_RO);
3481 case CEPH_SESSION_REJECT:
3482 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3483 pr_info("mds%d rejected session\n", session->s_mds);
3484 session->s_state = CEPH_MDS_SESSION_REJECTED;
3485 cleanup_session_requests(mdsc, session);
3486 remove_session_caps(session);
3488 mdsc->fsc->blocklisted = true;
3489 wake = 2; /* for good measure */
3493 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3497 mutex_unlock(&session->s_mutex);
3499 mutex_lock(&mdsc->mutex);
3500 __wake_requests(mdsc, &session->s_waiting);
3502 kick_requests(mdsc, mds);
3503 mutex_unlock(&mdsc->mutex);
3505 if (op == CEPH_SESSION_CLOSE)
3506 ceph_put_mds_session(session);
3510 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3511 (int)msg->front.iov_len);
3516 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3520 dcaps = xchg(&req->r_dir_caps, 0);
3522 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3523 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3527 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3531 dcaps = xchg(&req->r_dir_caps, 0);
3533 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3534 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3540 * called under session->mutex.
3542 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3543 struct ceph_mds_session *session)
3545 struct ceph_mds_request *req, *nreq;
3548 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3550 mutex_lock(&mdsc->mutex);
3551 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3552 __send_request(mdsc, session, req, true);
3555 * also re-send old requests when MDS enters reconnect stage. So that MDS
3556 * can process completed request in clientreplay stage.
3558 p = rb_first(&mdsc->request_tree);
3560 req = rb_entry(p, struct ceph_mds_request, r_node);
3562 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3564 if (req->r_attempts == 0)
3565 continue; /* only old requests */
3566 if (!req->r_session)
3568 if (req->r_session->s_mds != session->s_mds)
3571 ceph_mdsc_release_dir_caps_no_check(req);
3573 __send_request(mdsc, session, req, true);
3575 mutex_unlock(&mdsc->mutex);
3578 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3580 struct ceph_msg *reply;
3581 struct ceph_pagelist *_pagelist;
3586 if (!recon_state->allow_multi)
3589 /* can't handle message that contains both caps and realm */
3590 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3592 /* pre-allocate new pagelist */
3593 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3597 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3601 /* placeholder for nr_caps */
3602 err = ceph_pagelist_encode_32(_pagelist, 0);
3606 if (recon_state->nr_caps) {
3607 /* currently encoding caps */
3608 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3612 /* placeholder for nr_realms (currently encoding relams) */
3613 err = ceph_pagelist_encode_32(_pagelist, 0);
3618 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3622 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3623 addr = kmap_atomic(page);
3624 if (recon_state->nr_caps) {
3625 /* currently encoding caps */
3626 *addr = cpu_to_le32(recon_state->nr_caps);
3628 /* currently encoding relams */
3629 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3631 kunmap_atomic(addr);
3633 reply->hdr.version = cpu_to_le16(5);
3634 reply->hdr.compat_version = cpu_to_le16(4);
3636 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3637 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3639 ceph_con_send(&recon_state->session->s_con, reply);
3640 ceph_pagelist_release(recon_state->pagelist);
3642 recon_state->pagelist = _pagelist;
3643 recon_state->nr_caps = 0;
3644 recon_state->nr_realms = 0;
3645 recon_state->msg_version = 5;
3648 ceph_msg_put(reply);
3650 ceph_pagelist_release(_pagelist);
3654 static struct dentry* d_find_primary(struct inode *inode)
3656 struct dentry *alias, *dn = NULL;
3658 if (hlist_empty(&inode->i_dentry))
3661 spin_lock(&inode->i_lock);
3662 if (hlist_empty(&inode->i_dentry))
3665 if (S_ISDIR(inode->i_mode)) {
3666 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3667 if (!IS_ROOT(alias))
3672 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3673 spin_lock(&alias->d_lock);
3674 if (!d_unhashed(alias) &&
3675 (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3676 dn = dget_dlock(alias);
3678 spin_unlock(&alias->d_lock);
3683 spin_unlock(&inode->i_lock);
3688 * Encode information about a cap for a reconnect with the MDS.
3690 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3694 struct ceph_mds_cap_reconnect v2;
3695 struct ceph_mds_cap_reconnect_v1 v1;
3697 struct ceph_inode_info *ci = cap->ci;
3698 struct ceph_reconnect_state *recon_state = arg;
3699 struct ceph_pagelist *pagelist = recon_state->pagelist;
3700 struct dentry *dentry;
3702 int pathlen = 0, err;
3706 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3707 inode, ceph_vinop(inode), cap, cap->cap_id,
3708 ceph_cap_string(cap->issued));
3710 dentry = d_find_primary(inode);
3712 /* set pathbase to parent dir when msg_version >= 2 */
3713 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3714 recon_state->msg_version >= 2);
3717 err = PTR_ERR(path);
3725 spin_lock(&ci->i_ceph_lock);
3726 cap->seq = 0; /* reset cap seq */
3727 cap->issue_seq = 0; /* and issue_seq */
3728 cap->mseq = 0; /* and migrate_seq */
3729 cap->cap_gen = cap->session->s_cap_gen;
3731 /* These are lost when the session goes away */
3732 if (S_ISDIR(inode->i_mode)) {
3733 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3734 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3735 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3737 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3740 if (recon_state->msg_version >= 2) {
3741 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3742 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3743 rec.v2.issued = cpu_to_le32(cap->issued);
3744 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3745 rec.v2.pathbase = cpu_to_le64(pathbase);
3746 rec.v2.flock_len = (__force __le32)
3747 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3749 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3750 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3751 rec.v1.issued = cpu_to_le32(cap->issued);
3752 rec.v1.size = cpu_to_le64(inode->i_size);
3753 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3754 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3755 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3756 rec.v1.pathbase = cpu_to_le64(pathbase);
3759 if (list_empty(&ci->i_cap_snaps)) {
3760 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3762 struct ceph_cap_snap *capsnap =
3763 list_first_entry(&ci->i_cap_snaps,
3764 struct ceph_cap_snap, ci_item);
3765 snap_follows = capsnap->follows;
3767 spin_unlock(&ci->i_ceph_lock);
3769 if (recon_state->msg_version >= 2) {
3770 int num_fcntl_locks, num_flock_locks;
3771 struct ceph_filelock *flocks = NULL;
3772 size_t struct_len, total_len = sizeof(u64);
3776 if (rec.v2.flock_len) {
3777 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3779 num_fcntl_locks = 0;
3780 num_flock_locks = 0;
3782 if (num_fcntl_locks + num_flock_locks > 0) {
3783 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3784 sizeof(struct ceph_filelock),
3790 err = ceph_encode_locks_to_buffer(inode, flocks,
3805 if (recon_state->msg_version >= 3) {
3806 /* version, compat_version and struct_len */
3807 total_len += 2 * sizeof(u8) + sizeof(u32);
3811 * number of encoded locks is stable, so copy to pagelist
3813 struct_len = 2 * sizeof(u32) +
3814 (num_fcntl_locks + num_flock_locks) *
3815 sizeof(struct ceph_filelock);
3816 rec.v2.flock_len = cpu_to_le32(struct_len);
3818 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3821 struct_len += sizeof(u64); /* snap_follows */
3823 total_len += struct_len;
3825 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3826 err = send_reconnect_partial(recon_state);
3828 goto out_freeflocks;
3829 pagelist = recon_state->pagelist;
3832 err = ceph_pagelist_reserve(pagelist, total_len);
3834 goto out_freeflocks;
3836 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3837 if (recon_state->msg_version >= 3) {
3838 ceph_pagelist_encode_8(pagelist, struct_v);
3839 ceph_pagelist_encode_8(pagelist, 1);
3840 ceph_pagelist_encode_32(pagelist, struct_len);
3842 ceph_pagelist_encode_string(pagelist, path, pathlen);
3843 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3844 ceph_locks_to_pagelist(flocks, pagelist,
3845 num_fcntl_locks, num_flock_locks);
3847 ceph_pagelist_encode_64(pagelist, snap_follows);
3851 err = ceph_pagelist_reserve(pagelist,
3852 sizeof(u64) + sizeof(u32) +
3853 pathlen + sizeof(rec.v1));
3857 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3858 ceph_pagelist_encode_string(pagelist, path, pathlen);
3859 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3863 ceph_mdsc_free_path(path, pathlen);
3865 recon_state->nr_caps++;
3869 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3870 struct ceph_reconnect_state *recon_state)
3873 struct ceph_pagelist *pagelist = recon_state->pagelist;
3876 if (recon_state->msg_version >= 4) {
3877 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3883 * snaprealms. we provide mds with the ino, seq (version), and
3884 * parent for all of our realms. If the mds has any newer info,
3887 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3888 struct ceph_snap_realm *realm =
3889 rb_entry(p, struct ceph_snap_realm, node);
3890 struct ceph_mds_snaprealm_reconnect sr_rec;
3892 if (recon_state->msg_version >= 4) {
3893 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3896 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3897 err = send_reconnect_partial(recon_state);
3900 pagelist = recon_state->pagelist;
3903 err = ceph_pagelist_reserve(pagelist, need);
3907 ceph_pagelist_encode_8(pagelist, 1);
3908 ceph_pagelist_encode_8(pagelist, 1);
3909 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3912 dout(" adding snap realm %llx seq %lld parent %llx\n",
3913 realm->ino, realm->seq, realm->parent_ino);
3914 sr_rec.ino = cpu_to_le64(realm->ino);
3915 sr_rec.seq = cpu_to_le64(realm->seq);
3916 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3918 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3922 recon_state->nr_realms++;
3930 * If an MDS fails and recovers, clients need to reconnect in order to
3931 * reestablish shared state. This includes all caps issued through
3932 * this session _and_ the snap_realm hierarchy. Because it's not
3933 * clear which snap realms the mds cares about, we send everything we
3934 * know about.. that ensures we'll then get any new info the
3935 * recovering MDS might have.
3937 * This is a relatively heavyweight operation, but it's rare.
3939 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3940 struct ceph_mds_session *session)
3942 struct ceph_msg *reply;
3943 int mds = session->s_mds;
3945 struct ceph_reconnect_state recon_state = {
3950 pr_info("mds%d reconnect start\n", mds);
3952 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3953 if (!recon_state.pagelist)
3954 goto fail_nopagelist;
3956 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3960 xa_destroy(&session->s_delegated_inos);
3962 mutex_lock(&session->s_mutex);
3963 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3966 dout("session %p state %s\n", session,
3967 ceph_session_state_name(session->s_state));
3969 spin_lock(&session->s_gen_ttl_lock);
3970 session->s_cap_gen++;
3971 spin_unlock(&session->s_gen_ttl_lock);
3973 spin_lock(&session->s_cap_lock);
3974 /* don't know if session is readonly */
3975 session->s_readonly = 0;
3977 * notify __ceph_remove_cap() that we are composing cap reconnect.
3978 * If a cap get released before being added to the cap reconnect,
3979 * __ceph_remove_cap() should skip queuing cap release.
3981 session->s_cap_reconnect = 1;
3982 /* drop old cap expires; we're about to reestablish that state */
3983 detach_cap_releases(session, &dispose);
3984 spin_unlock(&session->s_cap_lock);
3985 dispose_cap_releases(mdsc, &dispose);
3987 /* trim unused caps to reduce MDS's cache rejoin time */
3988 if (mdsc->fsc->sb->s_root)
3989 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3991 ceph_con_close(&session->s_con);
3992 ceph_con_open(&session->s_con,
3993 CEPH_ENTITY_TYPE_MDS, mds,
3994 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3996 /* replay unsafe requests */
3997 replay_unsafe_requests(mdsc, session);
3999 ceph_early_kick_flushing_caps(mdsc, session);
4001 down_read(&mdsc->snap_rwsem);
4003 /* placeholder for nr_caps */
4004 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4008 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4009 recon_state.msg_version = 3;
4010 recon_state.allow_multi = true;
4011 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4012 recon_state.msg_version = 3;
4014 recon_state.msg_version = 2;
4016 /* trsaverse this session's caps */
4017 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4019 spin_lock(&session->s_cap_lock);
4020 session->s_cap_reconnect = 0;
4021 spin_unlock(&session->s_cap_lock);
4026 /* check if all realms can be encoded into current message */
4027 if (mdsc->num_snap_realms) {
4029 recon_state.pagelist->length +
4030 mdsc->num_snap_realms *
4031 sizeof(struct ceph_mds_snaprealm_reconnect);
4032 if (recon_state.msg_version >= 4) {
4033 /* number of realms */
4034 total_len += sizeof(u32);
4035 /* version, compat_version and struct_len */
4036 total_len += mdsc->num_snap_realms *
4037 (2 * sizeof(u8) + sizeof(u32));
4039 if (total_len > RECONNECT_MAX_SIZE) {
4040 if (!recon_state.allow_multi) {
4044 if (recon_state.nr_caps) {
4045 err = send_reconnect_partial(&recon_state);
4049 recon_state.msg_version = 5;
4053 err = encode_snap_realms(mdsc, &recon_state);
4057 if (recon_state.msg_version >= 5) {
4058 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4063 if (recon_state.nr_caps || recon_state.nr_realms) {
4065 list_first_entry(&recon_state.pagelist->head,
4067 __le32 *addr = kmap_atomic(page);
4068 if (recon_state.nr_caps) {
4069 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4070 *addr = cpu_to_le32(recon_state.nr_caps);
4071 } else if (recon_state.msg_version >= 4) {
4072 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4074 kunmap_atomic(addr);
4077 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4078 if (recon_state.msg_version >= 4)
4079 reply->hdr.compat_version = cpu_to_le16(4);
4081 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4082 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4084 ceph_con_send(&session->s_con, reply);
4086 mutex_unlock(&session->s_mutex);
4088 mutex_lock(&mdsc->mutex);
4089 __wake_requests(mdsc, &session->s_waiting);
4090 mutex_unlock(&mdsc->mutex);
4092 up_read(&mdsc->snap_rwsem);
4093 ceph_pagelist_release(recon_state.pagelist);
4097 ceph_msg_put(reply);
4098 up_read(&mdsc->snap_rwsem);
4099 mutex_unlock(&session->s_mutex);
4101 ceph_pagelist_release(recon_state.pagelist);
4103 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4109 * compare old and new mdsmaps, kicking requests
4110 * and closing out old connections as necessary
4112 * called under mdsc->mutex.
4114 static void check_new_map(struct ceph_mds_client *mdsc,
4115 struct ceph_mdsmap *newmap,
4116 struct ceph_mdsmap *oldmap)
4119 int oldstate, newstate;
4120 struct ceph_mds_session *s;
4122 dout("check_new_map new %u old %u\n",
4123 newmap->m_epoch, oldmap->m_epoch);
4125 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4126 if (!mdsc->sessions[i])
4128 s = mdsc->sessions[i];
4129 oldstate = ceph_mdsmap_get_state(oldmap, i);
4130 newstate = ceph_mdsmap_get_state(newmap, i);
4132 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4133 i, ceph_mds_state_name(oldstate),
4134 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4135 ceph_mds_state_name(newstate),
4136 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4137 ceph_session_state_name(s->s_state));
4139 if (i >= newmap->possible_max_rank) {
4140 /* force close session for stopped mds */
4141 ceph_get_mds_session(s);
4142 __unregister_session(mdsc, s);
4143 __wake_requests(mdsc, &s->s_waiting);
4144 mutex_unlock(&mdsc->mutex);
4146 mutex_lock(&s->s_mutex);
4147 cleanup_session_requests(mdsc, s);
4148 remove_session_caps(s);
4149 mutex_unlock(&s->s_mutex);
4151 ceph_put_mds_session(s);
4153 mutex_lock(&mdsc->mutex);
4154 kick_requests(mdsc, i);
4158 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4159 ceph_mdsmap_get_addr(newmap, i),
4160 sizeof(struct ceph_entity_addr))) {
4162 mutex_unlock(&mdsc->mutex);
4163 mutex_lock(&s->s_mutex);
4164 mutex_lock(&mdsc->mutex);
4165 ceph_con_close(&s->s_con);
4166 mutex_unlock(&s->s_mutex);
4167 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4168 } else if (oldstate == newstate) {
4169 continue; /* nothing new with this mds */
4175 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4176 newstate >= CEPH_MDS_STATE_RECONNECT) {
4177 mutex_unlock(&mdsc->mutex);
4178 send_mds_reconnect(mdsc, s);
4179 mutex_lock(&mdsc->mutex);
4183 * kick request on any mds that has gone active.
4185 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4186 newstate >= CEPH_MDS_STATE_ACTIVE) {
4187 if (oldstate != CEPH_MDS_STATE_CREATING &&
4188 oldstate != CEPH_MDS_STATE_STARTING)
4189 pr_info("mds%d recovery completed\n", s->s_mds);
4190 kick_requests(mdsc, i);
4191 mutex_unlock(&mdsc->mutex);
4192 mutex_lock(&s->s_mutex);
4193 mutex_lock(&mdsc->mutex);
4194 ceph_kick_flushing_caps(mdsc, s);
4195 mutex_unlock(&s->s_mutex);
4196 wake_up_session_caps(s, RECONNECT);
4200 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4201 s = mdsc->sessions[i];
4204 if (!ceph_mdsmap_is_laggy(newmap, i))
4206 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4207 s->s_state == CEPH_MDS_SESSION_HUNG ||
4208 s->s_state == CEPH_MDS_SESSION_CLOSING) {
4209 dout(" connecting to export targets of laggy mds%d\n",
4211 __open_export_target_sessions(mdsc, s);
4223 * caller must hold session s_mutex, dentry->d_lock
4225 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4227 struct ceph_dentry_info *di = ceph_dentry(dentry);
4229 ceph_put_mds_session(di->lease_session);
4230 di->lease_session = NULL;
4233 static void handle_lease(struct ceph_mds_client *mdsc,
4234 struct ceph_mds_session *session,
4235 struct ceph_msg *msg)
4237 struct super_block *sb = mdsc->fsc->sb;
4238 struct inode *inode;
4239 struct dentry *parent, *dentry;
4240 struct ceph_dentry_info *di;
4241 int mds = session->s_mds;
4242 struct ceph_mds_lease *h = msg->front.iov_base;
4244 struct ceph_vino vino;
4248 dout("handle_lease from mds%d\n", mds);
4251 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4253 vino.ino = le64_to_cpu(h->ino);
4254 vino.snap = CEPH_NOSNAP;
4255 seq = le32_to_cpu(h->seq);
4256 dname.len = get_unaligned_le32(h + 1);
4257 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4259 dname.name = (void *)(h + 1) + sizeof(u32);
4262 inode = ceph_find_inode(sb, vino);
4263 dout("handle_lease %s, ino %llx %p %.*s\n",
4264 ceph_lease_op_name(h->action), vino.ino, inode,
4265 dname.len, dname.name);
4267 mutex_lock(&session->s_mutex);
4268 inc_session_sequence(session);
4271 dout("handle_lease no inode %llx\n", vino.ino);
4276 parent = d_find_alias(inode);
4278 dout("no parent dentry on inode %p\n", inode);
4280 goto release; /* hrm... */
4282 dname.hash = full_name_hash(parent, dname.name, dname.len);
4283 dentry = d_lookup(parent, &dname);
4288 spin_lock(&dentry->d_lock);
4289 di = ceph_dentry(dentry);
4290 switch (h->action) {
4291 case CEPH_MDS_LEASE_REVOKE:
4292 if (di->lease_session == session) {
4293 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4294 h->seq = cpu_to_le32(di->lease_seq);
4295 __ceph_mdsc_drop_dentry_lease(dentry);
4300 case CEPH_MDS_LEASE_RENEW:
4301 if (di->lease_session == session &&
4302 di->lease_gen == session->s_cap_gen &&
4303 di->lease_renew_from &&
4304 di->lease_renew_after == 0) {
4305 unsigned long duration =
4306 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4308 di->lease_seq = seq;
4309 di->time = di->lease_renew_from + duration;
4310 di->lease_renew_after = di->lease_renew_from +
4312 di->lease_renew_from = 0;
4316 spin_unlock(&dentry->d_lock);
4323 /* let's just reuse the same message */
4324 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4326 ceph_con_send(&session->s_con, msg);
4329 mutex_unlock(&session->s_mutex);
4330 /* avoid calling iput_final() in mds dispatch threads */
4331 ceph_async_iput(inode);
4335 pr_err("corrupt lease message\n");
4339 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4340 struct dentry *dentry, char action,
4343 struct ceph_msg *msg;
4344 struct ceph_mds_lease *lease;
4346 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4348 dout("lease_send_msg identry %p %s to mds%d\n",
4349 dentry, ceph_lease_op_name(action), session->s_mds);
4351 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4354 lease = msg->front.iov_base;
4355 lease->action = action;
4356 lease->seq = cpu_to_le32(seq);
4358 spin_lock(&dentry->d_lock);
4359 dir = d_inode(dentry->d_parent);
4360 lease->ino = cpu_to_le64(ceph_ino(dir));
4361 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4363 put_unaligned_le32(dentry->d_name.len, lease + 1);
4364 memcpy((void *)(lease + 1) + 4,
4365 dentry->d_name.name, dentry->d_name.len);
4366 spin_unlock(&dentry->d_lock);
4368 * if this is a preemptive lease RELEASE, no need to
4369 * flush request stream, since the actual request will
4372 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4374 ceph_con_send(&session->s_con, msg);
4378 * lock unlock sessions, to wait ongoing session activities
4380 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4384 mutex_lock(&mdsc->mutex);
4385 for (i = 0; i < mdsc->max_sessions; i++) {
4386 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4389 mutex_unlock(&mdsc->mutex);
4390 mutex_lock(&s->s_mutex);
4391 mutex_unlock(&s->s_mutex);
4392 ceph_put_mds_session(s);
4393 mutex_lock(&mdsc->mutex);
4395 mutex_unlock(&mdsc->mutex);
4398 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4400 struct ceph_fs_client *fsc = mdsc->fsc;
4402 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4405 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4408 if (!READ_ONCE(fsc->blocklisted))
4411 if (fsc->last_auto_reconnect &&
4412 time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4415 pr_info("auto reconnect after blocklisted\n");
4416 fsc->last_auto_reconnect = jiffies;
4417 ceph_force_reconnect(fsc->sb);
4420 bool check_session_state(struct ceph_mds_session *s)
4422 switch (s->s_state) {
4423 case CEPH_MDS_SESSION_OPEN:
4424 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4425 s->s_state = CEPH_MDS_SESSION_HUNG;
4426 pr_info("mds%d hung\n", s->s_mds);
4429 case CEPH_MDS_SESSION_CLOSING:
4430 /* Should never reach this when we're unmounting */
4431 WARN_ON_ONCE(s->s_ttl);
4433 case CEPH_MDS_SESSION_NEW:
4434 case CEPH_MDS_SESSION_RESTARTING:
4435 case CEPH_MDS_SESSION_CLOSED:
4436 case CEPH_MDS_SESSION_REJECTED:
4444 * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4445 * then we need to retransmit that request.
4447 void inc_session_sequence(struct ceph_mds_session *s)
4449 lockdep_assert_held(&s->s_mutex);
4453 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4456 dout("resending session close request for mds%d\n", s->s_mds);
4457 ret = request_close_session(s);
4459 pr_err("unable to close session to mds%d: %d\n",
4465 * delayed work -- periodically trim expired leases, renew caps with mds. If
4466 * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4467 * workqueue delay value of 5 secs will be used.
4469 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4471 unsigned long max_delay = HZ * 5;
4473 /* 5 secs default delay */
4474 if (!delay || (delay > max_delay))
4476 schedule_delayed_work(&mdsc->delayed_work,
4477 round_jiffies_relative(delay));
4480 static void delayed_work(struct work_struct *work)
4482 struct ceph_mds_client *mdsc =
4483 container_of(work, struct ceph_mds_client, delayed_work.work);
4484 unsigned long delay;
4489 dout("mdsc delayed_work\n");
4494 mutex_lock(&mdsc->mutex);
4495 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4496 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4497 mdsc->last_renew_caps);
4499 mdsc->last_renew_caps = jiffies;
4501 for (i = 0; i < mdsc->max_sessions; i++) {
4502 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4506 if (!check_session_state(s)) {
4507 ceph_put_mds_session(s);
4510 mutex_unlock(&mdsc->mutex);
4512 mutex_lock(&s->s_mutex);
4514 send_renew_caps(mdsc, s);
4516 ceph_con_keepalive(&s->s_con);
4517 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4518 s->s_state == CEPH_MDS_SESSION_HUNG)
4519 ceph_send_cap_releases(mdsc, s);
4520 mutex_unlock(&s->s_mutex);
4521 ceph_put_mds_session(s);
4523 mutex_lock(&mdsc->mutex);
4525 mutex_unlock(&mdsc->mutex);
4527 delay = ceph_check_delayed_caps(mdsc);
4529 ceph_queue_cap_reclaim_work(mdsc);
4531 ceph_trim_snapid_map(mdsc);
4533 maybe_recover_session(mdsc);
4535 schedule_delayed(mdsc, delay);
4538 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4541 struct ceph_mds_client *mdsc;
4544 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4548 mutex_init(&mdsc->mutex);
4549 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4550 if (!mdsc->mdsmap) {
4555 init_completion(&mdsc->safe_umount_waiters);
4556 init_waitqueue_head(&mdsc->session_close_wq);
4557 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4558 mdsc->sessions = NULL;
4559 atomic_set(&mdsc->num_sessions, 0);
4560 mdsc->max_sessions = 0;
4562 atomic64_set(&mdsc->quotarealms_count, 0);
4563 mdsc->quotarealms_inodes = RB_ROOT;
4564 mutex_init(&mdsc->quotarealms_inodes_mutex);
4565 mdsc->last_snap_seq = 0;
4566 init_rwsem(&mdsc->snap_rwsem);
4567 mdsc->snap_realms = RB_ROOT;
4568 INIT_LIST_HEAD(&mdsc->snap_empty);
4569 mdsc->num_snap_realms = 0;
4570 spin_lock_init(&mdsc->snap_empty_lock);
4572 mdsc->oldest_tid = 0;
4573 mdsc->request_tree = RB_ROOT;
4574 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4575 mdsc->last_renew_caps = jiffies;
4576 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4577 INIT_LIST_HEAD(&mdsc->cap_wait_list);
4578 spin_lock_init(&mdsc->cap_delay_lock);
4579 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4580 spin_lock_init(&mdsc->snap_flush_lock);
4581 mdsc->last_cap_flush_tid = 1;
4582 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4583 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4584 mdsc->num_cap_flushing = 0;
4585 spin_lock_init(&mdsc->cap_dirty_lock);
4586 init_waitqueue_head(&mdsc->cap_flushing_wq);
4587 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4588 atomic_set(&mdsc->cap_reclaim_pending, 0);
4589 err = ceph_metric_init(&mdsc->metric);
4593 spin_lock_init(&mdsc->dentry_list_lock);
4594 INIT_LIST_HEAD(&mdsc->dentry_leases);
4595 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4597 ceph_caps_init(mdsc);
4598 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4600 spin_lock_init(&mdsc->snapid_map_lock);
4601 mdsc->snapid_map_tree = RB_ROOT;
4602 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4604 init_rwsem(&mdsc->pool_perm_rwsem);
4605 mdsc->pool_perm_tree = RB_ROOT;
4607 strscpy(mdsc->nodename, utsname()->nodename,
4608 sizeof(mdsc->nodename));
4614 kfree(mdsc->mdsmap);
4621 * Wait for safe replies on open mds requests. If we time out, drop
4622 * all requests from the tree to avoid dangling dentry refs.
4624 static void wait_requests(struct ceph_mds_client *mdsc)
4626 struct ceph_options *opts = mdsc->fsc->client->options;
4627 struct ceph_mds_request *req;
4629 mutex_lock(&mdsc->mutex);
4630 if (__get_oldest_req(mdsc)) {
4631 mutex_unlock(&mdsc->mutex);
4633 dout("wait_requests waiting for requests\n");
4634 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4635 ceph_timeout_jiffies(opts->mount_timeout));
4637 /* tear down remaining requests */
4638 mutex_lock(&mdsc->mutex);
4639 while ((req = __get_oldest_req(mdsc))) {
4640 dout("wait_requests timed out on tid %llu\n",
4642 list_del_init(&req->r_wait);
4643 __unregister_request(mdsc, req);
4646 mutex_unlock(&mdsc->mutex);
4647 dout("wait_requests done\n");
4651 * called before mount is ro, and before dentries are torn down.
4652 * (hmm, does this still race with new lookups?)
4654 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4656 dout("pre_umount\n");
4659 lock_unlock_sessions(mdsc);
4660 ceph_flush_dirty_caps(mdsc);
4661 wait_requests(mdsc);
4664 * wait for reply handlers to drop their request refs and
4665 * their inode/dcache refs
4669 ceph_cleanup_quotarealms_inodes(mdsc);
4673 * wait for all write mds requests to flush.
4675 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4677 struct ceph_mds_request *req = NULL, *nextreq;
4680 mutex_lock(&mdsc->mutex);
4681 dout("wait_unsafe_requests want %lld\n", want_tid);
4683 req = __get_oldest_req(mdsc);
4684 while (req && req->r_tid <= want_tid) {
4685 /* find next request */
4686 n = rb_next(&req->r_node);
4688 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4691 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4692 (req->r_op & CEPH_MDS_OP_WRITE)) {
4694 ceph_mdsc_get_request(req);
4696 ceph_mdsc_get_request(nextreq);
4697 mutex_unlock(&mdsc->mutex);
4698 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
4699 req->r_tid, want_tid);
4700 wait_for_completion(&req->r_safe_completion);
4701 mutex_lock(&mdsc->mutex);
4702 ceph_mdsc_put_request(req);
4704 break; /* next dne before, so we're done! */
4705 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4706 /* next request was removed from tree */
4707 ceph_mdsc_put_request(nextreq);
4710 ceph_mdsc_put_request(nextreq); /* won't go away */
4714 mutex_unlock(&mdsc->mutex);
4715 dout("wait_unsafe_requests done\n");
4718 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4720 u64 want_tid, want_flush;
4722 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4726 mutex_lock(&mdsc->mutex);
4727 want_tid = mdsc->last_tid;
4728 mutex_unlock(&mdsc->mutex);
4730 ceph_flush_dirty_caps(mdsc);
4731 spin_lock(&mdsc->cap_dirty_lock);
4732 want_flush = mdsc->last_cap_flush_tid;
4733 if (!list_empty(&mdsc->cap_flush_list)) {
4734 struct ceph_cap_flush *cf =
4735 list_last_entry(&mdsc->cap_flush_list,
4736 struct ceph_cap_flush, g_list);
4739 spin_unlock(&mdsc->cap_dirty_lock);
4741 dout("sync want tid %lld flush_seq %lld\n",
4742 want_tid, want_flush);
4744 wait_unsafe_requests(mdsc, want_tid);
4745 wait_caps_flush(mdsc, want_flush);
4749 * true if all sessions are closed, or we force unmount
4751 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4753 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4755 return atomic_read(&mdsc->num_sessions) <= skipped;
4759 * called after sb is ro.
4761 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4763 struct ceph_options *opts = mdsc->fsc->client->options;
4764 struct ceph_mds_session *session;
4768 dout("close_sessions\n");
4770 /* close sessions */
4771 mutex_lock(&mdsc->mutex);
4772 for (i = 0; i < mdsc->max_sessions; i++) {
4773 session = __ceph_lookup_mds_session(mdsc, i);
4776 mutex_unlock(&mdsc->mutex);
4777 mutex_lock(&session->s_mutex);
4778 if (__close_session(mdsc, session) <= 0)
4780 mutex_unlock(&session->s_mutex);
4781 ceph_put_mds_session(session);
4782 mutex_lock(&mdsc->mutex);
4784 mutex_unlock(&mdsc->mutex);
4786 dout("waiting for sessions to close\n");
4787 wait_event_timeout(mdsc->session_close_wq,
4788 done_closing_sessions(mdsc, skipped),
4789 ceph_timeout_jiffies(opts->mount_timeout));
4791 /* tear down remaining sessions */
4792 mutex_lock(&mdsc->mutex);
4793 for (i = 0; i < mdsc->max_sessions; i++) {
4794 if (mdsc->sessions[i]) {
4795 session = ceph_get_mds_session(mdsc->sessions[i]);
4796 __unregister_session(mdsc, session);
4797 mutex_unlock(&mdsc->mutex);
4798 mutex_lock(&session->s_mutex);
4799 remove_session_caps(session);
4800 mutex_unlock(&session->s_mutex);
4801 ceph_put_mds_session(session);
4802 mutex_lock(&mdsc->mutex);
4805 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4806 mutex_unlock(&mdsc->mutex);
4808 ceph_cleanup_snapid_map(mdsc);
4809 ceph_cleanup_empty_realms(mdsc);
4811 cancel_work_sync(&mdsc->cap_reclaim_work);
4812 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4817 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4819 struct ceph_mds_session *session;
4822 dout("force umount\n");
4824 mutex_lock(&mdsc->mutex);
4825 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4826 session = __ceph_lookup_mds_session(mdsc, mds);
4830 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4831 __unregister_session(mdsc, session);
4832 __wake_requests(mdsc, &session->s_waiting);
4833 mutex_unlock(&mdsc->mutex);
4835 mutex_lock(&session->s_mutex);
4836 __close_session(mdsc, session);
4837 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4838 cleanup_session_requests(mdsc, session);
4839 remove_session_caps(session);
4841 mutex_unlock(&session->s_mutex);
4842 ceph_put_mds_session(session);
4844 mutex_lock(&mdsc->mutex);
4845 kick_requests(mdsc, mds);
4847 __wake_requests(mdsc, &mdsc->waiting_for_map);
4848 mutex_unlock(&mdsc->mutex);
4851 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4855 * Make sure the delayed work stopped before releasing
4858 * Because the cancel_delayed_work_sync() will only
4859 * guarantee that the work finishes executing. But the
4860 * delayed work will re-arm itself again after that.
4862 flush_delayed_work(&mdsc->delayed_work);
4865 ceph_mdsmap_destroy(mdsc->mdsmap);
4866 kfree(mdsc->sessions);
4867 ceph_caps_finalize(mdsc);
4868 ceph_pool_perm_destroy(mdsc);
4871 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4873 struct ceph_mds_client *mdsc = fsc->mdsc;
4874 dout("mdsc_destroy %p\n", mdsc);
4879 /* flush out any connection work with references to us */
4882 ceph_mdsc_stop(mdsc);
4884 ceph_metric_destroy(&mdsc->metric);
4888 dout("mdsc_destroy %p done\n", mdsc);
4891 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4893 struct ceph_fs_client *fsc = mdsc->fsc;
4894 const char *mds_namespace = fsc->mount_options->mds_namespace;
4895 void *p = msg->front.iov_base;
4896 void *end = p + msg->front.iov_len;
4900 u32 mount_fscid = (u32)-1;
4901 u8 struct_v, struct_cv;
4904 ceph_decode_need(&p, end, sizeof(u32), bad);
4905 epoch = ceph_decode_32(&p);
4907 dout("handle_fsmap epoch %u\n", epoch);
4909 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4910 struct_v = ceph_decode_8(&p);
4911 struct_cv = ceph_decode_8(&p);
4912 map_len = ceph_decode_32(&p);
4914 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4915 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4917 num_fs = ceph_decode_32(&p);
4918 while (num_fs-- > 0) {
4919 void *info_p, *info_end;
4924 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4925 info_v = ceph_decode_8(&p);
4926 info_cv = ceph_decode_8(&p);
4927 info_len = ceph_decode_32(&p);
4928 ceph_decode_need(&p, end, info_len, bad);
4930 info_end = p + info_len;
4933 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4934 fscid = ceph_decode_32(&info_p);
4935 namelen = ceph_decode_32(&info_p);
4936 ceph_decode_need(&info_p, info_end, namelen, bad);
4938 if (mds_namespace &&
4939 strlen(mds_namespace) == namelen &&
4940 !strncmp(mds_namespace, (char *)info_p, namelen)) {
4941 mount_fscid = fscid;
4946 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4947 if (mount_fscid != (u32)-1) {
4948 fsc->client->monc.fs_cluster_id = mount_fscid;
4949 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4951 ceph_monc_renew_subs(&fsc->client->monc);
4959 pr_err("error decoding fsmap\n");
4961 mutex_lock(&mdsc->mutex);
4962 mdsc->mdsmap_err = err;
4963 __wake_requests(mdsc, &mdsc->waiting_for_map);
4964 mutex_unlock(&mdsc->mutex);
4968 * handle mds map update.
4970 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4974 void *p = msg->front.iov_base;
4975 void *end = p + msg->front.iov_len;
4976 struct ceph_mdsmap *newmap, *oldmap;
4977 struct ceph_fsid fsid;
4980 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4981 ceph_decode_copy(&p, &fsid, sizeof(fsid));
4982 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4984 epoch = ceph_decode_32(&p);
4985 maplen = ceph_decode_32(&p);
4986 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4988 /* do we need it? */
4989 mutex_lock(&mdsc->mutex);
4990 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4991 dout("handle_map epoch %u <= our %u\n",
4992 epoch, mdsc->mdsmap->m_epoch);
4993 mutex_unlock(&mdsc->mutex);
4997 newmap = ceph_mdsmap_decode(&p, end);
4998 if (IS_ERR(newmap)) {
4999 err = PTR_ERR(newmap);
5003 /* swap into place */
5005 oldmap = mdsc->mdsmap;
5006 mdsc->mdsmap = newmap;
5007 check_new_map(mdsc, newmap, oldmap);
5008 ceph_mdsmap_destroy(oldmap);
5010 mdsc->mdsmap = newmap; /* first mds map */
5012 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5015 __wake_requests(mdsc, &mdsc->waiting_for_map);
5016 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5017 mdsc->mdsmap->m_epoch);
5019 mutex_unlock(&mdsc->mutex);
5020 schedule_delayed(mdsc, 0);
5024 mutex_unlock(&mdsc->mutex);
5026 pr_err("error decoding mdsmap %d\n", err);
5030 static struct ceph_connection *con_get(struct ceph_connection *con)
5032 struct ceph_mds_session *s = con->private;
5034 if (ceph_get_mds_session(s))
5039 static void con_put(struct ceph_connection *con)
5041 struct ceph_mds_session *s = con->private;
5043 ceph_put_mds_session(s);
5047 * if the client is unresponsive for long enough, the mds will kill
5048 * the session entirely.
5050 static void peer_reset(struct ceph_connection *con)
5052 struct ceph_mds_session *s = con->private;
5053 struct ceph_mds_client *mdsc = s->s_mdsc;
5055 pr_warn("mds%d closed our session\n", s->s_mds);
5056 send_mds_reconnect(mdsc, s);
5059 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5061 struct ceph_mds_session *s = con->private;
5062 struct ceph_mds_client *mdsc = s->s_mdsc;
5063 int type = le16_to_cpu(msg->hdr.type);
5065 mutex_lock(&mdsc->mutex);
5066 if (__verify_registered_session(mdsc, s) < 0) {
5067 mutex_unlock(&mdsc->mutex);
5070 mutex_unlock(&mdsc->mutex);
5073 case CEPH_MSG_MDS_MAP:
5074 ceph_mdsc_handle_mdsmap(mdsc, msg);
5076 case CEPH_MSG_FS_MAP_USER:
5077 ceph_mdsc_handle_fsmap(mdsc, msg);
5079 case CEPH_MSG_CLIENT_SESSION:
5080 handle_session(s, msg);
5082 case CEPH_MSG_CLIENT_REPLY:
5083 handle_reply(s, msg);
5085 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5086 handle_forward(mdsc, s, msg);
5088 case CEPH_MSG_CLIENT_CAPS:
5089 ceph_handle_caps(s, msg);
5091 case CEPH_MSG_CLIENT_SNAP:
5092 ceph_handle_snap(mdsc, s, msg);
5094 case CEPH_MSG_CLIENT_LEASE:
5095 handle_lease(mdsc, s, msg);
5097 case CEPH_MSG_CLIENT_QUOTA:
5098 ceph_handle_quota(mdsc, s, msg);
5102 pr_err("received unknown message type %d %s\n", type,
5103 ceph_msg_type_name(type));
5114 * Note: returned pointer is the address of a structure that's
5115 * managed separately. Caller must *not* attempt to free it.
5117 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
5118 int *proto, int force_new)
5120 struct ceph_mds_session *s = con->private;
5121 struct ceph_mds_client *mdsc = s->s_mdsc;
5122 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5123 struct ceph_auth_handshake *auth = &s->s_auth;
5125 if (force_new && auth->authorizer) {
5126 ceph_auth_destroy_authorizer(auth->authorizer);
5127 auth->authorizer = NULL;
5129 if (!auth->authorizer) {
5130 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
5133 return ERR_PTR(ret);
5135 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
5138 return ERR_PTR(ret);
5140 *proto = ac->protocol;
5145 static int add_authorizer_challenge(struct ceph_connection *con,
5146 void *challenge_buf, int challenge_buf_len)
5148 struct ceph_mds_session *s = con->private;
5149 struct ceph_mds_client *mdsc = s->s_mdsc;
5150 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5152 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5153 challenge_buf, challenge_buf_len);
5156 static int verify_authorizer_reply(struct ceph_connection *con)
5158 struct ceph_mds_session *s = con->private;
5159 struct ceph_mds_client *mdsc = s->s_mdsc;
5160 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5162 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
5165 static int invalidate_authorizer(struct ceph_connection *con)
5167 struct ceph_mds_session *s = con->private;
5168 struct ceph_mds_client *mdsc = s->s_mdsc;
5169 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5171 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5173 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5176 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5177 struct ceph_msg_header *hdr, int *skip)
5179 struct ceph_msg *msg;
5180 int type = (int) le16_to_cpu(hdr->type);
5181 int front_len = (int) le32_to_cpu(hdr->front_len);
5187 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5189 pr_err("unable to allocate msg type %d len %d\n",
5197 static int mds_sign_message(struct ceph_msg *msg)
5199 struct ceph_mds_session *s = msg->con->private;
5200 struct ceph_auth_handshake *auth = &s->s_auth;
5202 return ceph_auth_sign_message(auth, msg);
5205 static int mds_check_message_signature(struct ceph_msg *msg)
5207 struct ceph_mds_session *s = msg->con->private;
5208 struct ceph_auth_handshake *auth = &s->s_auth;
5210 return ceph_auth_check_message_signature(auth, msg);
5213 static const struct ceph_connection_operations mds_con_ops = {
5216 .dispatch = dispatch,
5217 .get_authorizer = get_authorizer,
5218 .add_authorizer_challenge = add_authorizer_challenge,
5219 .verify_authorizer_reply = verify_authorizer_reply,
5220 .invalidate_authorizer = invalidate_authorizer,
5221 .peer_reset = peer_reset,
5222 .alloc_msg = mds_alloc_msg,
5223 .sign_message = mds_sign_message,
5224 .check_message_signature = mds_check_message_signature,