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49 * are no longer valid.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));
179 /* snapshot count, remains zero for v<=3 */
181 ceph_decode_64_safe(p, end, info->rsnaps, bad);
188 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
189 ceph_decode_64_safe(p, end, info->inline_version, bad);
190 ceph_decode_32_safe(p, end, info->inline_len, bad);
191 ceph_decode_need(p, end, info->inline_len, bad);
192 info->inline_data = *p;
193 *p += info->inline_len;
195 info->inline_version = CEPH_INLINE_NONE;
197 if (features & CEPH_FEATURE_MDS_QUOTA) {
198 err = parse_reply_info_quota(p, end, info);
206 info->pool_ns_len = 0;
207 info->pool_ns_data = NULL;
208 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
209 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
210 if (info->pool_ns_len > 0) {
211 ceph_decode_need(p, end, info->pool_ns_len, bad);
212 info->pool_ns_data = *p;
213 *p += info->pool_ns_len;
217 if (features & CEPH_FEATURE_FS_BTIME) {
218 ceph_decode_need(p, end, sizeof(info->btime), bad);
219 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
220 ceph_decode_64_safe(p, end, info->change_attr, bad);
223 info->dir_pin = -ENODATA;
224 /* info->snap_btime and info->rsnaps remain zero */
233 static int parse_reply_info_dir(void **p, void *end,
234 struct ceph_mds_reply_dirfrag **dirfrag,
237 if (features == (u64)-1) {
238 u8 struct_v, struct_compat;
240 ceph_decode_8_safe(p, end, struct_v, bad);
241 ceph_decode_8_safe(p, end, struct_compat, bad);
242 /* struct_v is expected to be >= 1. we only understand
243 * encoding whose struct_compat == 1. */
244 if (!struct_v || struct_compat != 1)
246 ceph_decode_32_safe(p, end, struct_len, bad);
247 ceph_decode_need(p, end, struct_len, bad);
248 end = *p + struct_len;
251 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
253 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
254 if (unlikely(*p > end))
256 if (features == (u64)-1)
263 static int parse_reply_info_lease(void **p, void *end,
264 struct ceph_mds_reply_lease **lease,
267 if (features == (u64)-1) {
268 u8 struct_v, struct_compat;
270 ceph_decode_8_safe(p, end, struct_v, bad);
271 ceph_decode_8_safe(p, end, struct_compat, bad);
272 /* struct_v is expected to be >= 1. we only understand
273 * encoding whose struct_compat == 1. */
274 if (!struct_v || struct_compat != 1)
276 ceph_decode_32_safe(p, end, struct_len, bad);
277 ceph_decode_need(p, end, struct_len, bad);
278 end = *p + struct_len;
281 ceph_decode_need(p, end, sizeof(**lease), bad);
283 *p += sizeof(**lease);
284 if (features == (u64)-1)
292 * parse a normal reply, which may contain a (dir+)dentry and/or a
295 static int parse_reply_info_trace(void **p, void *end,
296 struct ceph_mds_reply_info_parsed *info,
301 if (info->head->is_dentry) {
302 err = parse_reply_info_in(p, end, &info->diri, features);
306 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
310 ceph_decode_32_safe(p, end, info->dname_len, bad);
311 ceph_decode_need(p, end, info->dname_len, bad);
313 *p += info->dname_len;
315 err = parse_reply_info_lease(p, end, &info->dlease, features);
320 if (info->head->is_target) {
321 err = parse_reply_info_in(p, end, &info->targeti, features);
326 if (unlikely(*p != end))
333 pr_err("problem parsing mds trace %d\n", err);
338 * parse readdir results
340 static int parse_reply_info_readdir(void **p, void *end,
341 struct ceph_mds_reply_info_parsed *info,
347 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
351 ceph_decode_need(p, end, sizeof(num) + 2, bad);
352 num = ceph_decode_32(p);
354 u16 flags = ceph_decode_16(p);
355 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
356 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
357 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
358 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
363 BUG_ON(!info->dir_entries);
364 if ((unsigned long)(info->dir_entries + num) >
365 (unsigned long)info->dir_entries + info->dir_buf_size) {
366 pr_err("dir contents are larger than expected\n");
373 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
375 ceph_decode_32_safe(p, end, rde->name_len, bad);
376 ceph_decode_need(p, end, rde->name_len, bad);
379 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
382 err = parse_reply_info_lease(p, end, &rde->lease, features);
386 err = parse_reply_info_in(p, end, &rde->inode, features);
389 /* ceph_readdir_prepopulate() will update it */
396 /* Skip over any unrecognized fields */
403 pr_err("problem parsing dir contents %d\n", err);
408 * parse fcntl F_GETLK results
410 static int parse_reply_info_filelock(void **p, void *end,
411 struct ceph_mds_reply_info_parsed *info,
414 if (*p + sizeof(*info->filelock_reply) > end)
417 info->filelock_reply = *p;
419 /* Skip over any unrecognized fields */
427 #if BITS_PER_LONG == 64
429 #define DELEGATED_INO_AVAILABLE xa_mk_value(1)
431 static int ceph_parse_deleg_inos(void **p, void *end,
432 struct ceph_mds_session *s)
436 ceph_decode_32_safe(p, end, sets, bad);
437 dout("got %u sets of delegated inodes\n", sets);
441 ceph_decode_64_safe(p, end, start, bad);
442 ceph_decode_64_safe(p, end, len, bad);
444 /* Don't accept a delegation of system inodes */
445 if (start < CEPH_INO_SYSTEM_BASE) {
446 pr_warn_ratelimited("ceph: ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
451 int err = xa_insert(&s->s_delegated_inos, ino = start++,
452 DELEGATED_INO_AVAILABLE,
455 dout("added delegated inode 0x%llx\n",
457 } else if (err == -EBUSY) {
458 pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
470 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
475 xa_for_each(&s->s_delegated_inos, ino, val) {
476 val = xa_erase(&s->s_delegated_inos, ino);
477 if (val == DELEGATED_INO_AVAILABLE)
483 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
485 return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
488 #else /* BITS_PER_LONG == 64 */
490 * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
491 * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
494 static int ceph_parse_deleg_inos(void **p, void *end,
495 struct ceph_mds_session *s)
499 ceph_decode_32_safe(p, end, sets, bad);
501 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
507 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
512 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
516 #endif /* BITS_PER_LONG == 64 */
519 * parse create results
521 static int parse_reply_info_create(void **p, void *end,
522 struct ceph_mds_reply_info_parsed *info,
523 u64 features, struct ceph_mds_session *s)
527 if (features == (u64)-1 ||
528 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
530 /* Malformed reply? */
531 info->has_create_ino = false;
532 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
533 info->has_create_ino = true;
534 /* struct_v, struct_compat, and len */
535 ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
536 ceph_decode_64_safe(p, end, info->ino, bad);
537 ret = ceph_parse_deleg_inos(p, end, s);
542 ceph_decode_64_safe(p, end, info->ino, bad);
543 info->has_create_ino = true;
550 /* Skip over any unrecognized fields */
558 * parse extra results
560 static int parse_reply_info_extra(void **p, void *end,
561 struct ceph_mds_reply_info_parsed *info,
562 u64 features, struct ceph_mds_session *s)
564 u32 op = le32_to_cpu(info->head->op);
566 if (op == CEPH_MDS_OP_GETFILELOCK)
567 return parse_reply_info_filelock(p, end, info, features);
568 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
569 return parse_reply_info_readdir(p, end, info, features);
570 else if (op == CEPH_MDS_OP_CREATE)
571 return parse_reply_info_create(p, end, info, features, s);
577 * parse entire mds reply
579 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
580 struct ceph_mds_reply_info_parsed *info,
587 info->head = msg->front.iov_base;
588 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
589 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
592 ceph_decode_32_safe(&p, end, len, bad);
594 ceph_decode_need(&p, end, len, bad);
595 err = parse_reply_info_trace(&p, p+len, info, features);
601 ceph_decode_32_safe(&p, end, len, bad);
603 ceph_decode_need(&p, end, len, bad);
604 err = parse_reply_info_extra(&p, p+len, info, features, s);
610 ceph_decode_32_safe(&p, end, len, bad);
611 info->snapblob_len = len;
622 pr_err("mds parse_reply err %d\n", err);
626 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
628 if (!info->dir_entries)
630 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
637 const char *ceph_session_state_name(int s)
640 case CEPH_MDS_SESSION_NEW: return "new";
641 case CEPH_MDS_SESSION_OPENING: return "opening";
642 case CEPH_MDS_SESSION_OPEN: return "open";
643 case CEPH_MDS_SESSION_HUNG: return "hung";
644 case CEPH_MDS_SESSION_CLOSING: return "closing";
645 case CEPH_MDS_SESSION_CLOSED: return "closed";
646 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
647 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
648 case CEPH_MDS_SESSION_REJECTED: return "rejected";
649 default: return "???";
653 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
655 if (refcount_inc_not_zero(&s->s_ref)) {
656 dout("mdsc get_session %p %d -> %d\n", s,
657 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
660 dout("mdsc get_session %p 0 -- FAIL\n", s);
665 void ceph_put_mds_session(struct ceph_mds_session *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);
850 put_cred(req->r_cred);
852 ceph_pagelist_release(req->r_pagelist);
853 put_request_session(req);
854 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
855 WARN_ON_ONCE(!list_empty(&req->r_wait));
856 kmem_cache_free(ceph_mds_request_cachep, req);
859 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
862 * lookup session, bump ref if found.
864 * called under mdsc->mutex.
866 static struct ceph_mds_request *
867 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
869 struct ceph_mds_request *req;
871 req = lookup_request(&mdsc->request_tree, tid);
873 ceph_mdsc_get_request(req);
879 * Register an in-flight request, and assign a tid. Link to directory
880 * are modifying (if any).
882 * Called under mdsc->mutex.
884 static void __register_request(struct ceph_mds_client *mdsc,
885 struct ceph_mds_request *req,
890 req->r_tid = ++mdsc->last_tid;
891 if (req->r_num_caps) {
892 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
895 pr_err("__register_request %p "
896 "failed to reserve caps: %d\n", req, ret);
897 /* set req->r_err to fail early from __do_request */
902 dout("__register_request %p tid %lld\n", req, req->r_tid);
903 ceph_mdsc_get_request(req);
904 insert_request(&mdsc->request_tree, req);
906 req->r_cred = get_current_cred();
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);
1188 if (WARN_ON_ONCE(*p + 4 + size > end))
1191 ceph_encode_32(p, size);
1192 memset(*p, 0, size);
1193 for (i = 0; i < count; i++)
1194 ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1197 if (WARN_ON_ONCE(*p + 4 > end))
1200 ceph_encode_32(p, 0);
1206 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1207 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1208 static int encode_metric_spec(void **p, void *end)
1210 static const size_t count = ARRAY_SIZE(metric_bits);
1213 if (WARN_ON_ONCE(*p + 2 > end))
1216 ceph_encode_8(p, 1); /* version */
1217 ceph_encode_8(p, 1); /* compat */
1221 size_t size = METRIC_BYTES(count);
1223 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1226 /* metric spec info length */
1227 ceph_encode_32(p, 4 + size);
1230 ceph_encode_32(p, size);
1231 memset(*p, 0, size);
1232 for (i = 0; i < count; i++)
1233 ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1236 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1239 /* metric spec info length */
1240 ceph_encode_32(p, 4);
1242 ceph_encode_32(p, 0);
1249 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1250 * to include additional client metadata fields.
1252 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1254 struct ceph_msg *msg;
1255 struct ceph_mds_session_head *h;
1257 int extra_bytes = 0;
1258 int metadata_key_count = 0;
1259 struct ceph_options *opt = mdsc->fsc->client->options;
1260 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1265 const char* metadata[][2] = {
1266 {"hostname", mdsc->nodename},
1267 {"kernel_version", init_utsname()->release},
1268 {"entity_id", opt->name ? : ""},
1269 {"root", fsopt->server_path ? : "/"},
1273 /* Calculate serialized length of metadata */
1274 extra_bytes = 4; /* map length */
1275 for (i = 0; metadata[i][0]; ++i) {
1276 extra_bytes += 8 + strlen(metadata[i][0]) +
1277 strlen(metadata[i][1]);
1278 metadata_key_count++;
1281 /* supported feature */
1283 count = ARRAY_SIZE(feature_bits);
1285 size = FEATURE_BYTES(count);
1286 extra_bytes += 4 + size;
1290 count = ARRAY_SIZE(metric_bits);
1292 size = METRIC_BYTES(count);
1293 extra_bytes += 2 + 4 + 4 + size;
1295 /* Allocate the message */
1296 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1299 pr_err("create_session_msg ENOMEM creating msg\n");
1300 return ERR_PTR(-ENOMEM);
1302 p = msg->front.iov_base;
1303 end = p + msg->front.iov_len;
1306 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1307 h->seq = cpu_to_le64(seq);
1310 * Serialize client metadata into waiting buffer space, using
1311 * the format that userspace expects for map<string, string>
1313 * ClientSession messages with metadata are v4
1315 msg->hdr.version = cpu_to_le16(4);
1316 msg->hdr.compat_version = cpu_to_le16(1);
1318 /* The write pointer, following the session_head structure */
1321 /* Number of entries in the map */
1322 ceph_encode_32(&p, metadata_key_count);
1324 /* Two length-prefixed strings for each entry in the map */
1325 for (i = 0; metadata[i][0]; ++i) {
1326 size_t const key_len = strlen(metadata[i][0]);
1327 size_t const val_len = strlen(metadata[i][1]);
1329 ceph_encode_32(&p, key_len);
1330 memcpy(p, metadata[i][0], key_len);
1332 ceph_encode_32(&p, val_len);
1333 memcpy(p, metadata[i][1], val_len);
1337 ret = encode_supported_features(&p, end);
1339 pr_err("encode_supported_features failed!\n");
1341 return ERR_PTR(ret);
1344 ret = encode_metric_spec(&p, end);
1346 pr_err("encode_metric_spec failed!\n");
1348 return ERR_PTR(ret);
1351 msg->front.iov_len = p - msg->front.iov_base;
1352 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1358 * send session open request.
1360 * called under mdsc->mutex
1362 static int __open_session(struct ceph_mds_client *mdsc,
1363 struct ceph_mds_session *session)
1365 struct ceph_msg *msg;
1367 int mds = session->s_mds;
1369 /* wait for mds to go active? */
1370 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1371 dout("open_session to mds%d (%s)\n", mds,
1372 ceph_mds_state_name(mstate));
1373 session->s_state = CEPH_MDS_SESSION_OPENING;
1374 session->s_renew_requested = jiffies;
1376 /* send connect message */
1377 msg = create_session_open_msg(mdsc, session->s_seq);
1379 return PTR_ERR(msg);
1380 ceph_con_send(&session->s_con, msg);
1385 * open sessions for any export targets for the given mds
1387 * called under mdsc->mutex
1389 static struct ceph_mds_session *
1390 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1392 struct ceph_mds_session *session;
1395 session = __ceph_lookup_mds_session(mdsc, target);
1397 session = register_session(mdsc, target);
1398 if (IS_ERR(session))
1401 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1402 session->s_state == CEPH_MDS_SESSION_CLOSING) {
1403 ret = __open_session(mdsc, session);
1405 return ERR_PTR(ret);
1411 struct ceph_mds_session *
1412 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1414 struct ceph_mds_session *session;
1416 dout("open_export_target_session to mds%d\n", target);
1418 mutex_lock(&mdsc->mutex);
1419 session = __open_export_target_session(mdsc, target);
1420 mutex_unlock(&mdsc->mutex);
1425 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1426 struct ceph_mds_session *session)
1428 struct ceph_mds_info *mi;
1429 struct ceph_mds_session *ts;
1430 int i, mds = session->s_mds;
1432 if (mds >= mdsc->mdsmap->possible_max_rank)
1435 mi = &mdsc->mdsmap->m_info[mds];
1436 dout("open_export_target_sessions for mds%d (%d targets)\n",
1437 session->s_mds, mi->num_export_targets);
1439 for (i = 0; i < mi->num_export_targets; i++) {
1440 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1442 ceph_put_mds_session(ts);
1446 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1447 struct ceph_mds_session *session)
1449 mutex_lock(&mdsc->mutex);
1450 __open_export_target_sessions(mdsc, session);
1451 mutex_unlock(&mdsc->mutex);
1458 static void detach_cap_releases(struct ceph_mds_session *session,
1459 struct list_head *target)
1461 lockdep_assert_held(&session->s_cap_lock);
1463 list_splice_init(&session->s_cap_releases, target);
1464 session->s_num_cap_releases = 0;
1465 dout("dispose_cap_releases mds%d\n", session->s_mds);
1468 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1469 struct list_head *dispose)
1471 while (!list_empty(dispose)) {
1472 struct ceph_cap *cap;
1473 /* zero out the in-progress message */
1474 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1475 list_del(&cap->session_caps);
1476 ceph_put_cap(mdsc, cap);
1480 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1481 struct ceph_mds_session *session)
1483 struct ceph_mds_request *req;
1485 struct ceph_inode_info *ci;
1487 dout("cleanup_session_requests mds%d\n", session->s_mds);
1488 mutex_lock(&mdsc->mutex);
1489 while (!list_empty(&session->s_unsafe)) {
1490 req = list_first_entry(&session->s_unsafe,
1491 struct ceph_mds_request, r_unsafe_item);
1492 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1494 if (req->r_target_inode) {
1495 /* dropping unsafe change of inode's attributes */
1496 ci = ceph_inode(req->r_target_inode);
1497 errseq_set(&ci->i_meta_err, -EIO);
1499 if (req->r_unsafe_dir) {
1500 /* dropping unsafe directory operation */
1501 ci = ceph_inode(req->r_unsafe_dir);
1502 errseq_set(&ci->i_meta_err, -EIO);
1504 __unregister_request(mdsc, req);
1506 /* zero r_attempts, so kick_requests() will re-send requests */
1507 p = rb_first(&mdsc->request_tree);
1509 req = rb_entry(p, struct ceph_mds_request, r_node);
1511 if (req->r_session &&
1512 req->r_session->s_mds == session->s_mds)
1513 req->r_attempts = 0;
1515 mutex_unlock(&mdsc->mutex);
1519 * Helper to safely iterate over all caps associated with a session, with
1520 * special care taken to handle a racing __ceph_remove_cap().
1522 * Caller must hold session s_mutex.
1524 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1525 int (*cb)(struct inode *, struct ceph_cap *,
1528 struct list_head *p;
1529 struct ceph_cap *cap;
1530 struct inode *inode, *last_inode = NULL;
1531 struct ceph_cap *old_cap = NULL;
1534 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1535 spin_lock(&session->s_cap_lock);
1536 p = session->s_caps.next;
1537 while (p != &session->s_caps) {
1538 cap = list_entry(p, struct ceph_cap, session_caps);
1539 inode = igrab(&cap->ci->vfs_inode);
1544 session->s_cap_iterator = cap;
1545 spin_unlock(&session->s_cap_lock);
1548 /* avoid calling iput_final() while holding
1549 * s_mutex or in mds dispatch threads */
1550 ceph_async_iput(last_inode);
1554 ceph_put_cap(session->s_mdsc, old_cap);
1558 ret = cb(inode, cap, arg);
1561 spin_lock(&session->s_cap_lock);
1564 dout("iterate_session_caps finishing cap %p removal\n",
1566 BUG_ON(cap->session != session);
1567 cap->session = NULL;
1568 list_del_init(&cap->session_caps);
1569 session->s_nr_caps--;
1570 atomic64_dec(&session->s_mdsc->metric.total_caps);
1571 if (cap->queue_release)
1572 __ceph_queue_cap_release(session, cap);
1574 old_cap = cap; /* put_cap it w/o locks held */
1581 session->s_cap_iterator = NULL;
1582 spin_unlock(&session->s_cap_lock);
1584 ceph_async_iput(last_inode);
1586 ceph_put_cap(session->s_mdsc, old_cap);
1591 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1594 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1595 struct ceph_inode_info *ci = ceph_inode(inode);
1596 LIST_HEAD(to_remove);
1597 bool dirty_dropped = false;
1598 bool invalidate = false;
1600 dout("removing cap %p, ci is %p, inode is %p\n",
1601 cap, ci, &ci->vfs_inode);
1602 spin_lock(&ci->i_ceph_lock);
1603 __ceph_remove_cap(cap, false);
1604 if (!ci->i_auth_cap) {
1605 struct ceph_cap_flush *cf;
1606 struct ceph_mds_client *mdsc = fsc->mdsc;
1608 if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1609 if (inode->i_data.nrpages > 0)
1611 if (ci->i_wrbuffer_ref > 0)
1612 mapping_set_error(&inode->i_data, -EIO);
1615 while (!list_empty(&ci->i_cap_flush_list)) {
1616 cf = list_first_entry(&ci->i_cap_flush_list,
1617 struct ceph_cap_flush, i_list);
1618 list_move(&cf->i_list, &to_remove);
1621 spin_lock(&mdsc->cap_dirty_lock);
1623 list_for_each_entry(cf, &to_remove, i_list)
1624 list_del_init(&cf->g_list);
1626 if (!list_empty(&ci->i_dirty_item)) {
1627 pr_warn_ratelimited(
1628 " dropping dirty %s state for %p %lld\n",
1629 ceph_cap_string(ci->i_dirty_caps),
1630 inode, ceph_ino(inode));
1631 ci->i_dirty_caps = 0;
1632 list_del_init(&ci->i_dirty_item);
1633 dirty_dropped = true;
1635 if (!list_empty(&ci->i_flushing_item)) {
1636 pr_warn_ratelimited(
1637 " dropping dirty+flushing %s state for %p %lld\n",
1638 ceph_cap_string(ci->i_flushing_caps),
1639 inode, ceph_ino(inode));
1640 ci->i_flushing_caps = 0;
1641 list_del_init(&ci->i_flushing_item);
1642 mdsc->num_cap_flushing--;
1643 dirty_dropped = true;
1645 spin_unlock(&mdsc->cap_dirty_lock);
1647 if (dirty_dropped) {
1648 errseq_set(&ci->i_meta_err, -EIO);
1650 if (ci->i_wrbuffer_ref_head == 0 &&
1651 ci->i_wr_ref == 0 &&
1652 ci->i_dirty_caps == 0 &&
1653 ci->i_flushing_caps == 0) {
1654 ceph_put_snap_context(ci->i_head_snapc);
1655 ci->i_head_snapc = NULL;
1659 if (atomic_read(&ci->i_filelock_ref) > 0) {
1660 /* make further file lock syscall return -EIO */
1661 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1662 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1663 inode, ceph_ino(inode));
1666 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1667 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1668 ci->i_prealloc_cap_flush = NULL;
1671 spin_unlock(&ci->i_ceph_lock);
1672 while (!list_empty(&to_remove)) {
1673 struct ceph_cap_flush *cf;
1674 cf = list_first_entry(&to_remove,
1675 struct ceph_cap_flush, i_list);
1676 list_del_init(&cf->i_list);
1677 if (!cf->is_capsnap)
1678 ceph_free_cap_flush(cf);
1681 wake_up_all(&ci->i_cap_wq);
1683 ceph_queue_invalidate(inode);
1690 * caller must hold session s_mutex
1692 static void remove_session_caps(struct ceph_mds_session *session)
1694 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1695 struct super_block *sb = fsc->sb;
1698 dout("remove_session_caps on %p\n", session);
1699 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1701 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1703 spin_lock(&session->s_cap_lock);
1704 if (session->s_nr_caps > 0) {
1705 struct inode *inode;
1706 struct ceph_cap *cap, *prev = NULL;
1707 struct ceph_vino vino;
1709 * iterate_session_caps() skips inodes that are being
1710 * deleted, we need to wait until deletions are complete.
1711 * __wait_on_freeing_inode() is designed for the job,
1712 * but it is not exported, so use lookup inode function
1715 while (!list_empty(&session->s_caps)) {
1716 cap = list_entry(session->s_caps.next,
1717 struct ceph_cap, session_caps);
1721 vino = cap->ci->i_vino;
1722 spin_unlock(&session->s_cap_lock);
1724 inode = ceph_find_inode(sb, vino);
1725 /* avoid calling iput_final() while holding s_mutex */
1726 ceph_async_iput(inode);
1728 spin_lock(&session->s_cap_lock);
1732 // drop cap expires and unlock s_cap_lock
1733 detach_cap_releases(session, &dispose);
1735 BUG_ON(session->s_nr_caps > 0);
1736 BUG_ON(!list_empty(&session->s_cap_flushing));
1737 spin_unlock(&session->s_cap_lock);
1738 dispose_cap_releases(session->s_mdsc, &dispose);
1748 * wake up any threads waiting on this session's caps. if the cap is
1749 * old (didn't get renewed on the client reconnect), remove it now.
1751 * caller must hold s_mutex.
1753 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1756 struct ceph_inode_info *ci = ceph_inode(inode);
1757 unsigned long ev = (unsigned long)arg;
1759 if (ev == RECONNECT) {
1760 spin_lock(&ci->i_ceph_lock);
1761 ci->i_wanted_max_size = 0;
1762 ci->i_requested_max_size = 0;
1763 spin_unlock(&ci->i_ceph_lock);
1764 } else if (ev == RENEWCAPS) {
1765 if (cap->cap_gen < cap->session->s_cap_gen) {
1766 /* mds did not re-issue stale cap */
1767 spin_lock(&ci->i_ceph_lock);
1768 cap->issued = cap->implemented = CEPH_CAP_PIN;
1769 spin_unlock(&ci->i_ceph_lock);
1771 } else if (ev == FORCE_RO) {
1773 wake_up_all(&ci->i_cap_wq);
1777 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1779 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1780 ceph_iterate_session_caps(session, wake_up_session_cb,
1781 (void *)(unsigned long)ev);
1785 * Send periodic message to MDS renewing all currently held caps. The
1786 * ack will reset the expiration for all caps from this session.
1788 * caller holds s_mutex
1790 static int send_renew_caps(struct ceph_mds_client *mdsc,
1791 struct ceph_mds_session *session)
1793 struct ceph_msg *msg;
1796 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1797 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1798 pr_info("mds%d caps stale\n", session->s_mds);
1799 session->s_renew_requested = jiffies;
1801 /* do not try to renew caps until a recovering mds has reconnected
1802 * with its clients. */
1803 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1804 if (state < CEPH_MDS_STATE_RECONNECT) {
1805 dout("send_renew_caps ignoring mds%d (%s)\n",
1806 session->s_mds, ceph_mds_state_name(state));
1810 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1811 ceph_mds_state_name(state));
1812 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1813 ++session->s_renew_seq);
1816 ceph_con_send(&session->s_con, msg);
1820 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1821 struct ceph_mds_session *session, u64 seq)
1823 struct ceph_msg *msg;
1825 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1826 session->s_mds, ceph_session_state_name(session->s_state), seq);
1827 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1830 ceph_con_send(&session->s_con, msg);
1836 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1838 * Called under session->s_mutex
1840 static void renewed_caps(struct ceph_mds_client *mdsc,
1841 struct ceph_mds_session *session, int is_renew)
1846 spin_lock(&session->s_cap_lock);
1847 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1849 session->s_cap_ttl = session->s_renew_requested +
1850 mdsc->mdsmap->m_session_timeout*HZ;
1853 if (time_before(jiffies, session->s_cap_ttl)) {
1854 pr_info("mds%d caps renewed\n", session->s_mds);
1857 pr_info("mds%d caps still stale\n", session->s_mds);
1860 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1861 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1862 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1863 spin_unlock(&session->s_cap_lock);
1866 wake_up_session_caps(session, RENEWCAPS);
1870 * send a session close request
1872 static int request_close_session(struct ceph_mds_session *session)
1874 struct ceph_msg *msg;
1876 dout("request_close_session mds%d state %s seq %lld\n",
1877 session->s_mds, ceph_session_state_name(session->s_state),
1879 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1882 ceph_con_send(&session->s_con, msg);
1887 * Called with s_mutex held.
1889 static int __close_session(struct ceph_mds_client *mdsc,
1890 struct ceph_mds_session *session)
1892 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1894 session->s_state = CEPH_MDS_SESSION_CLOSING;
1895 return request_close_session(session);
1898 static bool drop_negative_children(struct dentry *dentry)
1900 struct dentry *child;
1901 bool all_negative = true;
1903 if (!d_is_dir(dentry))
1906 spin_lock(&dentry->d_lock);
1907 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1908 if (d_really_is_positive(child)) {
1909 all_negative = false;
1913 spin_unlock(&dentry->d_lock);
1916 shrink_dcache_parent(dentry);
1918 return all_negative;
1922 * Trim old(er) caps.
1924 * Because we can't cache an inode without one or more caps, we do
1925 * this indirectly: if a cap is unused, we prune its aliases, at which
1926 * point the inode will hopefully get dropped to.
1928 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1929 * memory pressure from the MDS, though, so it needn't be perfect.
1931 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1933 int *remaining = arg;
1934 struct ceph_inode_info *ci = ceph_inode(inode);
1935 int used, wanted, oissued, mine;
1937 if (*remaining <= 0)
1940 spin_lock(&ci->i_ceph_lock);
1941 mine = cap->issued | cap->implemented;
1942 used = __ceph_caps_used(ci);
1943 wanted = __ceph_caps_file_wanted(ci);
1944 oissued = __ceph_caps_issued_other(ci, cap);
1946 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1947 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1948 ceph_cap_string(used), ceph_cap_string(wanted));
1949 if (cap == ci->i_auth_cap) {
1950 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1951 !list_empty(&ci->i_cap_snaps))
1953 if ((used | wanted) & CEPH_CAP_ANY_WR)
1955 /* Note: it's possible that i_filelock_ref becomes non-zero
1956 * after dropping auth caps. It doesn't hurt because reply
1957 * of lock mds request will re-add auth caps. */
1958 if (atomic_read(&ci->i_filelock_ref) > 0)
1961 /* The inode has cached pages, but it's no longer used.
1962 * we can safely drop it */
1963 if (S_ISREG(inode->i_mode) &&
1964 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1965 !(oissued & CEPH_CAP_FILE_CACHE)) {
1969 if ((used | wanted) & ~oissued & mine)
1970 goto out; /* we need these caps */
1973 /* we aren't the only cap.. just remove us */
1974 __ceph_remove_cap(cap, true);
1977 struct dentry *dentry;
1978 /* try dropping referring dentries */
1979 spin_unlock(&ci->i_ceph_lock);
1980 dentry = d_find_any_alias(inode);
1981 if (dentry && drop_negative_children(dentry)) {
1984 d_prune_aliases(inode);
1985 count = atomic_read(&inode->i_count);
1988 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1997 spin_unlock(&ci->i_ceph_lock);
2002 * Trim session cap count down to some max number.
2004 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2005 struct ceph_mds_session *session,
2008 int trim_caps = session->s_nr_caps - max_caps;
2010 dout("trim_caps mds%d start: %d / %d, trim %d\n",
2011 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
2012 if (trim_caps > 0) {
2013 int remaining = trim_caps;
2015 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2016 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2017 session->s_mds, session->s_nr_caps, max_caps,
2018 trim_caps - remaining);
2021 ceph_flush_cap_releases(mdsc, session);
2025 static int check_caps_flush(struct ceph_mds_client *mdsc,
2030 spin_lock(&mdsc->cap_dirty_lock);
2031 if (!list_empty(&mdsc->cap_flush_list)) {
2032 struct ceph_cap_flush *cf =
2033 list_first_entry(&mdsc->cap_flush_list,
2034 struct ceph_cap_flush, g_list);
2035 if (cf->tid <= want_flush_tid) {
2036 dout("check_caps_flush still flushing tid "
2037 "%llu <= %llu\n", cf->tid, want_flush_tid);
2041 spin_unlock(&mdsc->cap_dirty_lock);
2046 * flush all dirty inode data to disk.
2048 * returns true if we've flushed through want_flush_tid
2050 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2053 dout("check_caps_flush want %llu\n", want_flush_tid);
2055 wait_event(mdsc->cap_flushing_wq,
2056 check_caps_flush(mdsc, want_flush_tid));
2058 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2062 * called under s_mutex
2064 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2065 struct ceph_mds_session *session)
2067 struct ceph_msg *msg = NULL;
2068 struct ceph_mds_cap_release *head;
2069 struct ceph_mds_cap_item *item;
2070 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2071 struct ceph_cap *cap;
2072 LIST_HEAD(tmp_list);
2073 int num_cap_releases;
2074 __le32 barrier, *cap_barrier;
2076 down_read(&osdc->lock);
2077 barrier = cpu_to_le32(osdc->epoch_barrier);
2078 up_read(&osdc->lock);
2080 spin_lock(&session->s_cap_lock);
2082 list_splice_init(&session->s_cap_releases, &tmp_list);
2083 num_cap_releases = session->s_num_cap_releases;
2084 session->s_num_cap_releases = 0;
2085 spin_unlock(&session->s_cap_lock);
2087 while (!list_empty(&tmp_list)) {
2089 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2090 PAGE_SIZE, GFP_NOFS, false);
2093 head = msg->front.iov_base;
2094 head->num = cpu_to_le32(0);
2095 msg->front.iov_len = sizeof(*head);
2097 msg->hdr.version = cpu_to_le16(2);
2098 msg->hdr.compat_version = cpu_to_le16(1);
2101 cap = list_first_entry(&tmp_list, struct ceph_cap,
2103 list_del(&cap->session_caps);
2106 head = msg->front.iov_base;
2107 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2109 item = msg->front.iov_base + msg->front.iov_len;
2110 item->ino = cpu_to_le64(cap->cap_ino);
2111 item->cap_id = cpu_to_le64(cap->cap_id);
2112 item->migrate_seq = cpu_to_le32(cap->mseq);
2113 item->seq = cpu_to_le32(cap->issue_seq);
2114 msg->front.iov_len += sizeof(*item);
2116 ceph_put_cap(mdsc, cap);
2118 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2119 // Append cap_barrier field
2120 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2121 *cap_barrier = barrier;
2122 msg->front.iov_len += sizeof(*cap_barrier);
2124 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2125 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2126 ceph_con_send(&session->s_con, msg);
2131 BUG_ON(num_cap_releases != 0);
2133 spin_lock(&session->s_cap_lock);
2134 if (!list_empty(&session->s_cap_releases))
2136 spin_unlock(&session->s_cap_lock);
2139 // Append cap_barrier field
2140 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2141 *cap_barrier = barrier;
2142 msg->front.iov_len += sizeof(*cap_barrier);
2144 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2145 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2146 ceph_con_send(&session->s_con, msg);
2150 pr_err("send_cap_releases mds%d, failed to allocate message\n",
2152 spin_lock(&session->s_cap_lock);
2153 list_splice(&tmp_list, &session->s_cap_releases);
2154 session->s_num_cap_releases += num_cap_releases;
2155 spin_unlock(&session->s_cap_lock);
2158 static void ceph_cap_release_work(struct work_struct *work)
2160 struct ceph_mds_session *session =
2161 container_of(work, struct ceph_mds_session, s_cap_release_work);
2163 mutex_lock(&session->s_mutex);
2164 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2165 session->s_state == CEPH_MDS_SESSION_HUNG)
2166 ceph_send_cap_releases(session->s_mdsc, session);
2167 mutex_unlock(&session->s_mutex);
2168 ceph_put_mds_session(session);
2171 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2172 struct ceph_mds_session *session)
2177 ceph_get_mds_session(session);
2178 if (queue_work(mdsc->fsc->cap_wq,
2179 &session->s_cap_release_work)) {
2180 dout("cap release work queued\n");
2182 ceph_put_mds_session(session);
2183 dout("failed to queue cap release work\n");
2188 * caller holds session->s_cap_lock
2190 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2191 struct ceph_cap *cap)
2193 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2194 session->s_num_cap_releases++;
2196 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2197 ceph_flush_cap_releases(session->s_mdsc, session);
2200 static void ceph_cap_reclaim_work(struct work_struct *work)
2202 struct ceph_mds_client *mdsc =
2203 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2204 int ret = ceph_trim_dentries(mdsc);
2206 ceph_queue_cap_reclaim_work(mdsc);
2209 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2214 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2215 dout("caps reclaim work queued\n");
2217 dout("failed to queue caps release work\n");
2221 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2226 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2227 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2228 atomic_set(&mdsc->cap_reclaim_pending, 0);
2229 ceph_queue_cap_reclaim_work(mdsc);
2237 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2240 struct ceph_inode_info *ci = ceph_inode(dir);
2241 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2242 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2243 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2244 unsigned int num_entries;
2247 spin_lock(&ci->i_ceph_lock);
2248 num_entries = ci->i_files + ci->i_subdirs;
2249 spin_unlock(&ci->i_ceph_lock);
2250 num_entries = max(num_entries, 1U);
2251 num_entries = min(num_entries, opt->max_readdir);
2253 order = get_order(size * num_entries);
2254 while (order >= 0) {
2255 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2258 if (rinfo->dir_entries)
2262 if (!rinfo->dir_entries)
2265 num_entries = (PAGE_SIZE << order) / size;
2266 num_entries = min(num_entries, opt->max_readdir);
2268 rinfo->dir_buf_size = PAGE_SIZE << order;
2269 req->r_num_caps = num_entries + 1;
2270 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2271 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2276 * Create an mds request.
2278 struct ceph_mds_request *
2279 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2281 struct ceph_mds_request *req;
2283 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2285 return ERR_PTR(-ENOMEM);
2287 mutex_init(&req->r_fill_mutex);
2289 req->r_started = jiffies;
2290 req->r_start_latency = ktime_get();
2291 req->r_resend_mds = -1;
2292 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2293 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2295 kref_init(&req->r_kref);
2296 RB_CLEAR_NODE(&req->r_node);
2297 INIT_LIST_HEAD(&req->r_wait);
2298 init_completion(&req->r_completion);
2299 init_completion(&req->r_safe_completion);
2300 INIT_LIST_HEAD(&req->r_unsafe_item);
2302 ktime_get_coarse_real_ts64(&req->r_stamp);
2305 req->r_direct_mode = mode;
2310 * return oldest (lowest) request, tid in request tree, 0 if none.
2312 * called under mdsc->mutex.
2314 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2316 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2318 return rb_entry(rb_first(&mdsc->request_tree),
2319 struct ceph_mds_request, r_node);
2322 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2324 return mdsc->oldest_tid;
2328 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2329 * on build_path_from_dentry in fs/cifs/dir.c.
2331 * If @stop_on_nosnap, generate path relative to the first non-snapped
2334 * Encode hidden .snap dirs as a double /, i.e.
2335 * foo/.snap/bar -> foo//bar
2337 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2340 struct dentry *temp;
2347 return ERR_PTR(-EINVAL);
2351 return ERR_PTR(-ENOMEM);
2356 seq = read_seqbegin(&rename_lock);
2360 struct inode *inode;
2362 spin_lock(&temp->d_lock);
2363 inode = d_inode(temp);
2364 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2365 dout("build_path path+%d: %p SNAPDIR\n",
2367 } else if (stop_on_nosnap && inode && dentry != temp &&
2368 ceph_snap(inode) == CEPH_NOSNAP) {
2369 spin_unlock(&temp->d_lock);
2370 pos++; /* get rid of any prepended '/' */
2373 pos -= temp->d_name.len;
2375 spin_unlock(&temp->d_lock);
2378 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2380 spin_unlock(&temp->d_lock);
2381 temp = READ_ONCE(temp->d_parent);
2383 /* Are we at the root? */
2387 /* Are we out of buffer? */
2393 base = ceph_ino(d_inode(temp));
2396 if (read_seqretry(&rename_lock, seq))
2401 * A rename didn't occur, but somehow we didn't end up where
2402 * we thought we would. Throw a warning and try again.
2404 pr_warn("build_path did not end path lookup where "
2405 "expected, pos is %d\n", pos);
2410 *plen = PATH_MAX - 1 - pos;
2411 dout("build_path on %p %d built %llx '%.*s'\n",
2412 dentry, d_count(dentry), base, *plen, path + pos);
2416 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2417 const char **ppath, int *ppathlen, u64 *pino,
2418 bool *pfreepath, bool parent_locked)
2424 dir = d_inode_rcu(dentry->d_parent);
2425 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2426 *pino = ceph_ino(dir);
2428 *ppath = dentry->d_name.name;
2429 *ppathlen = dentry->d_name.len;
2433 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2435 return PTR_ERR(path);
2441 static int build_inode_path(struct inode *inode,
2442 const char **ppath, int *ppathlen, u64 *pino,
2445 struct dentry *dentry;
2448 if (ceph_snap(inode) == CEPH_NOSNAP) {
2449 *pino = ceph_ino(inode);
2453 dentry = d_find_alias(inode);
2454 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2457 return PTR_ERR(path);
2464 * request arguments may be specified via an inode *, a dentry *, or
2465 * an explicit ino+path.
2467 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2468 struct inode *rdiri, const char *rpath,
2469 u64 rino, const char **ppath, int *pathlen,
2470 u64 *ino, bool *freepath, bool parent_locked)
2475 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2476 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2478 } else if (rdentry) {
2479 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2480 freepath, parent_locked);
2481 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2483 } else if (rpath || rino) {
2486 *pathlen = rpath ? strlen(rpath) : 0;
2487 dout(" path %.*s\n", *pathlen, rpath);
2493 static void encode_timestamp_and_gids(void **p,
2494 const struct ceph_mds_request *req)
2496 struct ceph_timespec ts;
2499 ceph_encode_timespec64(&ts, &req->r_stamp);
2500 ceph_encode_copy(p, &ts, sizeof(ts));
2503 ceph_encode_32(p, req->r_cred->group_info->ngroups);
2504 for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2505 ceph_encode_64(p, from_kgid(&init_user_ns,
2506 req->r_cred->group_info->gid[i]));
2510 * called under mdsc->mutex
2512 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2513 struct ceph_mds_request *req,
2514 bool drop_cap_releases)
2516 int mds = session->s_mds;
2517 struct ceph_mds_client *mdsc = session->s_mdsc;
2518 struct ceph_msg *msg;
2519 struct ceph_mds_request_head_old *head;
2520 const char *path1 = NULL;
2521 const char *path2 = NULL;
2522 u64 ino1 = 0, ino2 = 0;
2523 int pathlen1 = 0, pathlen2 = 0;
2524 bool freepath1 = false, freepath2 = false;
2529 bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2531 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2532 req->r_parent, req->r_path1, req->r_ino1.ino,
2533 &path1, &pathlen1, &ino1, &freepath1,
2534 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2535 &req->r_req_flags));
2541 /* If r_old_dentry is set, then assume that its parent is locked */
2542 ret = set_request_path_attr(NULL, req->r_old_dentry,
2543 req->r_old_dentry_dir,
2544 req->r_path2, req->r_ino2.ino,
2545 &path2, &pathlen2, &ino2, &freepath2, true);
2551 len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2552 len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2553 sizeof(struct ceph_timespec);
2554 len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2556 /* calculate (max) length for cap releases */
2557 len += sizeof(struct ceph_mds_request_release) *
2558 (!!req->r_inode_drop + !!req->r_dentry_drop +
2559 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2561 if (req->r_dentry_drop)
2563 if (req->r_old_dentry_drop)
2566 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2568 msg = ERR_PTR(-ENOMEM);
2572 msg->hdr.tid = cpu_to_le64(req->r_tid);
2575 * The old ceph_mds_request_head didn't contain a version field, and
2576 * one was added when we moved the message version from 3->4.
2579 msg->hdr.version = cpu_to_le16(3);
2580 head = msg->front.iov_base;
2581 p = msg->front.iov_base + sizeof(*head);
2583 struct ceph_mds_request_head *new_head = msg->front.iov_base;
2585 msg->hdr.version = cpu_to_le16(4);
2586 new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2587 head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2588 p = msg->front.iov_base + sizeof(*new_head);
2591 end = msg->front.iov_base + msg->front.iov_len;
2593 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2594 head->op = cpu_to_le32(req->r_op);
2595 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2596 req->r_cred->fsuid));
2597 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2598 req->r_cred->fsgid));
2599 head->ino = cpu_to_le64(req->r_deleg_ino);
2600 head->args = req->r_args;
2602 ceph_encode_filepath(&p, end, ino1, path1);
2603 ceph_encode_filepath(&p, end, ino2, path2);
2605 /* make note of release offset, in case we need to replay */
2606 req->r_request_release_offset = p - msg->front.iov_base;
2610 if (req->r_inode_drop)
2611 releases += ceph_encode_inode_release(&p,
2612 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2613 mds, req->r_inode_drop, req->r_inode_unless,
2614 req->r_op == CEPH_MDS_OP_READDIR);
2615 if (req->r_dentry_drop)
2616 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2617 req->r_parent, mds, req->r_dentry_drop,
2618 req->r_dentry_unless);
2619 if (req->r_old_dentry_drop)
2620 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2621 req->r_old_dentry_dir, mds,
2622 req->r_old_dentry_drop,
2623 req->r_old_dentry_unless);
2624 if (req->r_old_inode_drop)
2625 releases += ceph_encode_inode_release(&p,
2626 d_inode(req->r_old_dentry),
2627 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2629 if (drop_cap_releases) {
2631 p = msg->front.iov_base + req->r_request_release_offset;
2634 head->num_releases = cpu_to_le16(releases);
2636 encode_timestamp_and_gids(&p, req);
2638 if (WARN_ON_ONCE(p > end)) {
2640 msg = ERR_PTR(-ERANGE);
2644 msg->front.iov_len = p - msg->front.iov_base;
2645 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2647 if (req->r_pagelist) {
2648 struct ceph_pagelist *pagelist = req->r_pagelist;
2649 ceph_msg_data_add_pagelist(msg, pagelist);
2650 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2652 msg->hdr.data_len = 0;
2655 msg->hdr.data_off = cpu_to_le16(0);
2659 ceph_mdsc_free_path((char *)path2, pathlen2);
2662 ceph_mdsc_free_path((char *)path1, pathlen1);
2668 * called under mdsc->mutex if error, under no mutex if
2671 static void complete_request(struct ceph_mds_client *mdsc,
2672 struct ceph_mds_request *req)
2674 req->r_end_latency = ktime_get();
2676 if (req->r_callback)
2677 req->r_callback(mdsc, req);
2678 complete_all(&req->r_completion);
2681 static struct ceph_mds_request_head_old *
2682 find_old_request_head(void *p, u64 features)
2684 bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2685 struct ceph_mds_request_head *new_head;
2688 return (struct ceph_mds_request_head_old *)p;
2689 new_head = (struct ceph_mds_request_head *)p;
2690 return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2694 * called under mdsc->mutex
2696 static int __prepare_send_request(struct ceph_mds_session *session,
2697 struct ceph_mds_request *req,
2698 bool drop_cap_releases)
2700 int mds = session->s_mds;
2701 struct ceph_mds_client *mdsc = session->s_mdsc;
2702 struct ceph_mds_request_head_old *rhead;
2703 struct ceph_msg *msg;
2708 struct ceph_cap *cap =
2709 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2712 req->r_sent_on_mseq = cap->mseq;
2714 req->r_sent_on_mseq = -1;
2716 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2717 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2719 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2723 * Replay. Do not regenerate message (and rebuild
2724 * paths, etc.); just use the original message.
2725 * Rebuilding paths will break for renames because
2726 * d_move mangles the src name.
2728 msg = req->r_request;
2729 rhead = find_old_request_head(msg->front.iov_base,
2730 session->s_con.peer_features);
2732 flags = le32_to_cpu(rhead->flags);
2733 flags |= CEPH_MDS_FLAG_REPLAY;
2734 rhead->flags = cpu_to_le32(flags);
2736 if (req->r_target_inode)
2737 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2739 rhead->num_retry = req->r_attempts - 1;
2741 /* remove cap/dentry releases from message */
2742 rhead->num_releases = 0;
2744 p = msg->front.iov_base + req->r_request_release_offset;
2745 encode_timestamp_and_gids(&p, req);
2747 msg->front.iov_len = p - msg->front.iov_base;
2748 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2752 if (req->r_request) {
2753 ceph_msg_put(req->r_request);
2754 req->r_request = NULL;
2756 msg = create_request_message(session, req, drop_cap_releases);
2758 req->r_err = PTR_ERR(msg);
2759 return PTR_ERR(msg);
2761 req->r_request = msg;
2763 rhead = find_old_request_head(msg->front.iov_base,
2764 session->s_con.peer_features);
2765 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2766 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2767 flags |= CEPH_MDS_FLAG_REPLAY;
2768 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2769 flags |= CEPH_MDS_FLAG_ASYNC;
2771 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2772 rhead->flags = cpu_to_le32(flags);
2773 rhead->num_fwd = req->r_num_fwd;
2774 rhead->num_retry = req->r_attempts - 1;
2776 dout(" r_parent = %p\n", req->r_parent);
2781 * called under mdsc->mutex
2783 static int __send_request(struct ceph_mds_session *session,
2784 struct ceph_mds_request *req,
2785 bool drop_cap_releases)
2789 err = __prepare_send_request(session, req, drop_cap_releases);
2791 ceph_msg_get(req->r_request);
2792 ceph_con_send(&session->s_con, req->r_request);
2799 * send request, or put it on the appropriate wait list.
2801 static void __do_request(struct ceph_mds_client *mdsc,
2802 struct ceph_mds_request *req)
2804 struct ceph_mds_session *session = NULL;
2809 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2810 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2811 __unregister_request(mdsc, req);
2815 if (req->r_timeout &&
2816 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2817 dout("do_request timed out\n");
2821 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2822 dout("do_request forced umount\n");
2826 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2827 if (mdsc->mdsmap_err) {
2828 err = mdsc->mdsmap_err;
2829 dout("do_request mdsmap err %d\n", err);
2832 if (mdsc->mdsmap->m_epoch == 0) {
2833 dout("do_request no mdsmap, waiting for map\n");
2834 list_add(&req->r_wait, &mdsc->waiting_for_map);
2837 if (!(mdsc->fsc->mount_options->flags &
2838 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2839 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2840 err = -EHOSTUNREACH;
2845 put_request_session(req);
2847 mds = __choose_mds(mdsc, req, &random);
2849 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2850 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2854 dout("do_request no mds or not active, waiting for map\n");
2855 list_add(&req->r_wait, &mdsc->waiting_for_map);
2859 /* get, open session */
2860 session = __ceph_lookup_mds_session(mdsc, mds);
2862 session = register_session(mdsc, mds);
2863 if (IS_ERR(session)) {
2864 err = PTR_ERR(session);
2868 req->r_session = ceph_get_mds_session(session);
2870 dout("do_request mds%d session %p state %s\n", mds, session,
2871 ceph_session_state_name(session->s_state));
2872 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2873 session->s_state != CEPH_MDS_SESSION_HUNG) {
2875 * We cannot queue async requests since the caps and delegated
2876 * inodes are bound to the session. Just return -EJUKEBOX and
2877 * let the caller retry a sync request in that case.
2879 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2885 * If the session has been REJECTED, then return a hard error,
2886 * unless it's a CLEANRECOVER mount, in which case we'll queue
2887 * it to the mdsc queue.
2889 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2890 if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2891 list_add(&req->r_wait, &mdsc->waiting_for_map);
2897 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2898 session->s_state == CEPH_MDS_SESSION_CLOSING) {
2899 err = __open_session(mdsc, session);
2902 /* retry the same mds later */
2904 req->r_resend_mds = mds;
2906 list_add(&req->r_wait, &session->s_waiting);
2911 req->r_resend_mds = -1; /* forget any previous mds hint */
2913 if (req->r_request_started == 0) /* note request start time */
2914 req->r_request_started = jiffies;
2916 err = __send_request(session, req, false);
2919 ceph_put_mds_session(session);
2922 dout("__do_request early error %d\n", err);
2924 complete_request(mdsc, req);
2925 __unregister_request(mdsc, req);
2931 * called under mdsc->mutex
2933 static void __wake_requests(struct ceph_mds_client *mdsc,
2934 struct list_head *head)
2936 struct ceph_mds_request *req;
2937 LIST_HEAD(tmp_list);
2939 list_splice_init(head, &tmp_list);
2941 while (!list_empty(&tmp_list)) {
2942 req = list_entry(tmp_list.next,
2943 struct ceph_mds_request, r_wait);
2944 list_del_init(&req->r_wait);
2945 dout(" wake request %p tid %llu\n", req, req->r_tid);
2946 __do_request(mdsc, req);
2951 * Wake up threads with requests pending for @mds, so that they can
2952 * resubmit their requests to a possibly different mds.
2954 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2956 struct ceph_mds_request *req;
2957 struct rb_node *p = rb_first(&mdsc->request_tree);
2959 dout("kick_requests mds%d\n", mds);
2961 req = rb_entry(p, struct ceph_mds_request, r_node);
2963 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2965 if (req->r_attempts > 0)
2966 continue; /* only new requests */
2967 if (req->r_session &&
2968 req->r_session->s_mds == mds) {
2969 dout(" kicking tid %llu\n", req->r_tid);
2970 list_del_init(&req->r_wait);
2971 __do_request(mdsc, req);
2976 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2977 struct ceph_mds_request *req)
2981 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2983 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2984 if (req->r_parent) {
2985 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2986 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2987 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2988 spin_lock(&ci->i_ceph_lock);
2989 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2990 __ceph_touch_fmode(ci, mdsc, fmode);
2991 spin_unlock(&ci->i_ceph_lock);
2992 ihold(req->r_parent);
2994 if (req->r_old_dentry_dir)
2995 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2999 err = ceph_wait_on_async_create(req->r_inode);
3001 dout("%s: wait for async create returned: %d\n",
3007 if (!err && req->r_old_inode) {
3008 err = ceph_wait_on_async_create(req->r_old_inode);
3010 dout("%s: wait for async create returned: %d\n",
3016 dout("submit_request on %p for inode %p\n", req, dir);
3017 mutex_lock(&mdsc->mutex);
3018 __register_request(mdsc, req, dir);
3019 __do_request(mdsc, req);
3021 mutex_unlock(&mdsc->mutex);
3025 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3026 struct ceph_mds_request *req)
3031 dout("do_request waiting\n");
3032 if (!req->r_timeout && req->r_wait_for_completion) {
3033 err = req->r_wait_for_completion(mdsc, req);
3035 long timeleft = wait_for_completion_killable_timeout(
3037 ceph_timeout_jiffies(req->r_timeout));
3041 err = -ETIMEDOUT; /* timed out */
3043 err = timeleft; /* killed */
3045 dout("do_request waited, got %d\n", err);
3046 mutex_lock(&mdsc->mutex);
3048 /* only abort if we didn't race with a real reply */
3049 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3050 err = le32_to_cpu(req->r_reply_info.head->result);
3051 } else if (err < 0) {
3052 dout("aborted request %lld with %d\n", req->r_tid, err);
3055 * ensure we aren't running concurrently with
3056 * ceph_fill_trace or ceph_readdir_prepopulate, which
3057 * rely on locks (dir mutex) held by our caller.
3059 mutex_lock(&req->r_fill_mutex);
3061 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3062 mutex_unlock(&req->r_fill_mutex);
3064 if (req->r_parent &&
3065 (req->r_op & CEPH_MDS_OP_WRITE))
3066 ceph_invalidate_dir_request(req);
3071 mutex_unlock(&mdsc->mutex);
3076 * Synchrously perform an mds request. Take care of all of the
3077 * session setup, forwarding, retry details.
3079 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3081 struct ceph_mds_request *req)
3085 dout("do_request on %p\n", req);
3088 err = ceph_mdsc_submit_request(mdsc, dir, req);
3090 err = ceph_mdsc_wait_request(mdsc, req);
3091 dout("do_request %p done, result %d\n", req, err);
3096 * Invalidate dir's completeness, dentry lease state on an aborted MDS
3097 * namespace request.
3099 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3101 struct inode *dir = req->r_parent;
3102 struct inode *old_dir = req->r_old_dentry_dir;
3104 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3106 ceph_dir_clear_complete(dir);
3108 ceph_dir_clear_complete(old_dir);
3110 ceph_invalidate_dentry_lease(req->r_dentry);
3111 if (req->r_old_dentry)
3112 ceph_invalidate_dentry_lease(req->r_old_dentry);
3118 * We take the session mutex and parse and process the reply immediately.
3119 * This preserves the logical ordering of replies, capabilities, etc., sent
3120 * by the MDS as they are applied to our local cache.
3122 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3124 struct ceph_mds_client *mdsc = session->s_mdsc;
3125 struct ceph_mds_request *req;
3126 struct ceph_mds_reply_head *head = msg->front.iov_base;
3127 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
3128 struct ceph_snap_realm *realm;
3131 int mds = session->s_mds;
3133 if (msg->front.iov_len < sizeof(*head)) {
3134 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3139 /* get request, session */
3140 tid = le64_to_cpu(msg->hdr.tid);
3141 mutex_lock(&mdsc->mutex);
3142 req = lookup_get_request(mdsc, tid);
3144 dout("handle_reply on unknown tid %llu\n", tid);
3145 mutex_unlock(&mdsc->mutex);
3148 dout("handle_reply %p\n", req);
3150 /* correct session? */
3151 if (req->r_session != session) {
3152 pr_err("mdsc_handle_reply got %llu on session mds%d"
3153 " not mds%d\n", tid, session->s_mds,
3154 req->r_session ? req->r_session->s_mds : -1);
3155 mutex_unlock(&mdsc->mutex);
3160 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3161 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3162 pr_warn("got a dup %s reply on %llu from mds%d\n",
3163 head->safe ? "safe" : "unsafe", tid, mds);
3164 mutex_unlock(&mdsc->mutex);
3167 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3168 pr_warn("got unsafe after safe on %llu from mds%d\n",
3170 mutex_unlock(&mdsc->mutex);
3174 result = le32_to_cpu(head->result);
3178 * if we're not talking to the authority, send to them
3179 * if the authority has changed while we weren't looking,
3180 * send to new authority
3181 * Otherwise we just have to return an ESTALE
3183 if (result == -ESTALE) {
3184 dout("got ESTALE on request %llu\n", req->r_tid);
3185 req->r_resend_mds = -1;
3186 if (req->r_direct_mode != USE_AUTH_MDS) {
3187 dout("not using auth, setting for that now\n");
3188 req->r_direct_mode = USE_AUTH_MDS;
3189 __do_request(mdsc, req);
3190 mutex_unlock(&mdsc->mutex);
3193 int mds = __choose_mds(mdsc, req, NULL);
3194 if (mds >= 0 && mds != req->r_session->s_mds) {
3195 dout("but auth changed, so resending\n");
3196 __do_request(mdsc, req);
3197 mutex_unlock(&mdsc->mutex);
3201 dout("have to return ESTALE on request %llu\n", req->r_tid);
3206 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3207 __unregister_request(mdsc, req);
3209 /* last request during umount? */
3210 if (mdsc->stopping && !__get_oldest_req(mdsc))
3211 complete_all(&mdsc->safe_umount_waiters);
3213 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3215 * We already handled the unsafe response, now do the
3216 * cleanup. No need to examine the response; the MDS
3217 * doesn't include any result info in the safe
3218 * response. And even if it did, there is nothing
3219 * useful we could do with a revised return value.
3221 dout("got safe reply %llu, mds%d\n", tid, mds);
3223 mutex_unlock(&mdsc->mutex);
3227 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3228 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3231 dout("handle_reply tid %lld result %d\n", tid, result);
3232 rinfo = &req->r_reply_info;
3233 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3234 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3236 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3237 mutex_unlock(&mdsc->mutex);
3239 /* Must find target inode outside of mutexes to avoid deadlocks */
3240 if ((err >= 0) && rinfo->head->is_target) {
3242 struct ceph_vino tvino = {
3243 .ino = le64_to_cpu(rinfo->targeti.in->ino),
3244 .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3247 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3250 mutex_lock(&session->s_mutex);
3253 req->r_target_inode = in;
3256 mutex_lock(&session->s_mutex);
3258 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3265 if (rinfo->snapblob_len) {
3266 down_write(&mdsc->snap_rwsem);
3267 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3268 rinfo->snapblob + rinfo->snapblob_len,
3269 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3271 downgrade_write(&mdsc->snap_rwsem);
3273 down_read(&mdsc->snap_rwsem);
3276 /* insert trace into our cache */
3277 mutex_lock(&req->r_fill_mutex);
3278 current->journal_info = req;
3279 err = ceph_fill_trace(mdsc->fsc->sb, req);
3281 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3282 req->r_op == CEPH_MDS_OP_LSSNAP))
3283 ceph_readdir_prepopulate(req, req->r_session);
3285 current->journal_info = NULL;
3286 mutex_unlock(&req->r_fill_mutex);
3288 up_read(&mdsc->snap_rwsem);
3290 ceph_put_snap_realm(mdsc, realm);
3293 if (req->r_target_inode &&
3294 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3295 struct ceph_inode_info *ci =
3296 ceph_inode(req->r_target_inode);
3297 spin_lock(&ci->i_unsafe_lock);
3298 list_add_tail(&req->r_unsafe_target_item,
3299 &ci->i_unsafe_iops);
3300 spin_unlock(&ci->i_unsafe_lock);
3303 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3306 mutex_lock(&mdsc->mutex);
3307 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3311 req->r_reply = ceph_msg_get(msg);
3312 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3315 dout("reply arrived after request %lld was aborted\n", tid);
3317 mutex_unlock(&mdsc->mutex);
3319 mutex_unlock(&session->s_mutex);
3321 /* kick calling process */
3322 complete_request(mdsc, req);
3324 ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3325 req->r_end_latency, err);
3327 ceph_mdsc_put_request(req);
3334 * handle mds notification that our request has been forwarded.
3336 static void handle_forward(struct ceph_mds_client *mdsc,
3337 struct ceph_mds_session *session,
3338 struct ceph_msg *msg)
3340 struct ceph_mds_request *req;
3341 u64 tid = le64_to_cpu(msg->hdr.tid);
3345 void *p = msg->front.iov_base;
3346 void *end = p + msg->front.iov_len;
3348 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3349 next_mds = ceph_decode_32(&p);
3350 fwd_seq = ceph_decode_32(&p);
3352 mutex_lock(&mdsc->mutex);
3353 req = lookup_get_request(mdsc, tid);
3355 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3356 goto out; /* dup reply? */
3359 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3360 dout("forward tid %llu aborted, unregistering\n", tid);
3361 __unregister_request(mdsc, req);
3362 } else if (fwd_seq <= req->r_num_fwd) {
3363 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3364 tid, next_mds, req->r_num_fwd, fwd_seq);
3366 /* resend. forward race not possible; mds would drop */
3367 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3369 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3370 req->r_attempts = 0;
3371 req->r_num_fwd = fwd_seq;
3372 req->r_resend_mds = next_mds;
3373 put_request_session(req);
3374 __do_request(mdsc, req);
3376 ceph_mdsc_put_request(req);
3378 mutex_unlock(&mdsc->mutex);
3382 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3385 static int __decode_session_metadata(void **p, void *end,
3388 /* map<string,string> */
3391 ceph_decode_32_safe(p, end, n, bad);
3394 ceph_decode_32_safe(p, end, len, bad);
3395 ceph_decode_need(p, end, len, bad);
3396 err_str = !strncmp(*p, "error_string", len);
3398 ceph_decode_32_safe(p, end, len, bad);
3399 ceph_decode_need(p, end, len, bad);
3401 * Match "blocklisted (blacklisted)" from newer MDSes,
3402 * or "blacklisted" from older MDSes.
3404 if (err_str && strnstr(*p, "blacklisted", len))
3405 *blocklisted = true;
3414 * handle a mds session control message
3416 static void handle_session(struct ceph_mds_session *session,
3417 struct ceph_msg *msg)
3419 struct ceph_mds_client *mdsc = session->s_mdsc;
3420 int mds = session->s_mds;
3421 int msg_version = le16_to_cpu(msg->hdr.version);
3422 void *p = msg->front.iov_base;
3423 void *end = p + msg->front.iov_len;
3424 struct ceph_mds_session_head *h;
3426 u64 seq, features = 0;
3428 bool blocklisted = false;
3431 ceph_decode_need(&p, end, sizeof(*h), bad);
3435 op = le32_to_cpu(h->op);
3436 seq = le64_to_cpu(h->seq);
3438 if (msg_version >= 3) {
3440 /* version >= 2, metadata */
3441 if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3443 /* version >= 3, feature bits */
3444 ceph_decode_32_safe(&p, end, len, bad);
3446 ceph_decode_64_safe(&p, end, features, bad);
3447 p += len - sizeof(features);
3451 mutex_lock(&mdsc->mutex);
3452 if (op == CEPH_SESSION_CLOSE) {
3453 ceph_get_mds_session(session);
3454 __unregister_session(mdsc, session);
3456 /* FIXME: this ttl calculation is generous */
3457 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3458 mutex_unlock(&mdsc->mutex);
3460 mutex_lock(&session->s_mutex);
3462 dout("handle_session mds%d %s %p state %s seq %llu\n",
3463 mds, ceph_session_op_name(op), session,
3464 ceph_session_state_name(session->s_state), seq);
3466 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3467 session->s_state = CEPH_MDS_SESSION_OPEN;
3468 pr_info("mds%d came back\n", session->s_mds);
3472 case CEPH_SESSION_OPEN:
3473 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3474 pr_info("mds%d reconnect success\n", session->s_mds);
3475 session->s_state = CEPH_MDS_SESSION_OPEN;
3476 session->s_features = features;
3477 renewed_caps(mdsc, session, 0);
3478 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3479 metric_schedule_delayed(&mdsc->metric);
3482 __close_session(mdsc, session);
3485 case CEPH_SESSION_RENEWCAPS:
3486 if (session->s_renew_seq == seq)
3487 renewed_caps(mdsc, session, 1);
3490 case CEPH_SESSION_CLOSE:
3491 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3492 pr_info("mds%d reconnect denied\n", session->s_mds);
3493 session->s_state = CEPH_MDS_SESSION_CLOSED;
3494 cleanup_session_requests(mdsc, session);
3495 remove_session_caps(session);
3496 wake = 2; /* for good measure */
3497 wake_up_all(&mdsc->session_close_wq);
3500 case CEPH_SESSION_STALE:
3501 pr_info("mds%d caps went stale, renewing\n",
3503 spin_lock(&session->s_gen_ttl_lock);
3504 session->s_cap_gen++;
3505 session->s_cap_ttl = jiffies - 1;
3506 spin_unlock(&session->s_gen_ttl_lock);
3507 send_renew_caps(mdsc, session);
3510 case CEPH_SESSION_RECALL_STATE:
3511 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3514 case CEPH_SESSION_FLUSHMSG:
3515 send_flushmsg_ack(mdsc, session, seq);
3518 case CEPH_SESSION_FORCE_RO:
3519 dout("force_session_readonly %p\n", session);
3520 spin_lock(&session->s_cap_lock);
3521 session->s_readonly = true;
3522 spin_unlock(&session->s_cap_lock);
3523 wake_up_session_caps(session, FORCE_RO);
3526 case CEPH_SESSION_REJECT:
3527 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3528 pr_info("mds%d rejected session\n", session->s_mds);
3529 session->s_state = CEPH_MDS_SESSION_REJECTED;
3530 cleanup_session_requests(mdsc, session);
3531 remove_session_caps(session);
3533 mdsc->fsc->blocklisted = true;
3534 wake = 2; /* for good measure */
3538 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3542 mutex_unlock(&session->s_mutex);
3544 mutex_lock(&mdsc->mutex);
3545 __wake_requests(mdsc, &session->s_waiting);
3547 kick_requests(mdsc, mds);
3548 mutex_unlock(&mdsc->mutex);
3550 if (op == CEPH_SESSION_CLOSE)
3551 ceph_put_mds_session(session);
3555 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3556 (int)msg->front.iov_len);
3561 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3565 dcaps = xchg(&req->r_dir_caps, 0);
3567 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3568 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3572 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3576 dcaps = xchg(&req->r_dir_caps, 0);
3578 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3579 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3585 * called under session->mutex.
3587 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3588 struct ceph_mds_session *session)
3590 struct ceph_mds_request *req, *nreq;
3593 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3595 mutex_lock(&mdsc->mutex);
3596 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3597 __send_request(session, req, true);
3600 * also re-send old requests when MDS enters reconnect stage. So that MDS
3601 * can process completed request in clientreplay stage.
3603 p = rb_first(&mdsc->request_tree);
3605 req = rb_entry(p, struct ceph_mds_request, r_node);
3607 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3609 if (req->r_attempts == 0)
3610 continue; /* only old requests */
3611 if (!req->r_session)
3613 if (req->r_session->s_mds != session->s_mds)
3616 ceph_mdsc_release_dir_caps_no_check(req);
3618 __send_request(session, req, true);
3620 mutex_unlock(&mdsc->mutex);
3623 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3625 struct ceph_msg *reply;
3626 struct ceph_pagelist *_pagelist;
3631 if (!recon_state->allow_multi)
3634 /* can't handle message that contains both caps and realm */
3635 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3637 /* pre-allocate new pagelist */
3638 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3642 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3646 /* placeholder for nr_caps */
3647 err = ceph_pagelist_encode_32(_pagelist, 0);
3651 if (recon_state->nr_caps) {
3652 /* currently encoding caps */
3653 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3657 /* placeholder for nr_realms (currently encoding relams) */
3658 err = ceph_pagelist_encode_32(_pagelist, 0);
3663 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3667 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3668 addr = kmap_atomic(page);
3669 if (recon_state->nr_caps) {
3670 /* currently encoding caps */
3671 *addr = cpu_to_le32(recon_state->nr_caps);
3673 /* currently encoding relams */
3674 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3676 kunmap_atomic(addr);
3678 reply->hdr.version = cpu_to_le16(5);
3679 reply->hdr.compat_version = cpu_to_le16(4);
3681 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3682 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3684 ceph_con_send(&recon_state->session->s_con, reply);
3685 ceph_pagelist_release(recon_state->pagelist);
3687 recon_state->pagelist = _pagelist;
3688 recon_state->nr_caps = 0;
3689 recon_state->nr_realms = 0;
3690 recon_state->msg_version = 5;
3693 ceph_msg_put(reply);
3695 ceph_pagelist_release(_pagelist);
3699 static struct dentry* d_find_primary(struct inode *inode)
3701 struct dentry *alias, *dn = NULL;
3703 if (hlist_empty(&inode->i_dentry))
3706 spin_lock(&inode->i_lock);
3707 if (hlist_empty(&inode->i_dentry))
3710 if (S_ISDIR(inode->i_mode)) {
3711 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3712 if (!IS_ROOT(alias))
3717 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3718 spin_lock(&alias->d_lock);
3719 if (!d_unhashed(alias) &&
3720 (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3721 dn = dget_dlock(alias);
3723 spin_unlock(&alias->d_lock);
3728 spin_unlock(&inode->i_lock);
3733 * Encode information about a cap for a reconnect with the MDS.
3735 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3739 struct ceph_mds_cap_reconnect v2;
3740 struct ceph_mds_cap_reconnect_v1 v1;
3742 struct ceph_inode_info *ci = cap->ci;
3743 struct ceph_reconnect_state *recon_state = arg;
3744 struct ceph_pagelist *pagelist = recon_state->pagelist;
3745 struct dentry *dentry;
3751 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3752 inode, ceph_vinop(inode), cap, cap->cap_id,
3753 ceph_cap_string(cap->issued));
3755 dentry = d_find_primary(inode);
3757 /* set pathbase to parent dir when msg_version >= 2 */
3758 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3759 recon_state->msg_version >= 2);
3762 err = PTR_ERR(path);
3771 spin_lock(&ci->i_ceph_lock);
3772 cap->seq = 0; /* reset cap seq */
3773 cap->issue_seq = 0; /* and issue_seq */
3774 cap->mseq = 0; /* and migrate_seq */
3775 cap->cap_gen = cap->session->s_cap_gen;
3777 /* These are lost when the session goes away */
3778 if (S_ISDIR(inode->i_mode)) {
3779 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3780 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3781 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3783 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3786 if (recon_state->msg_version >= 2) {
3787 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3788 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3789 rec.v2.issued = cpu_to_le32(cap->issued);
3790 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3791 rec.v2.pathbase = cpu_to_le64(pathbase);
3792 rec.v2.flock_len = (__force __le32)
3793 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3795 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3796 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3797 rec.v1.issued = cpu_to_le32(cap->issued);
3798 rec.v1.size = cpu_to_le64(i_size_read(inode));
3799 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3800 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3801 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3802 rec.v1.pathbase = cpu_to_le64(pathbase);
3805 if (list_empty(&ci->i_cap_snaps)) {
3806 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3808 struct ceph_cap_snap *capsnap =
3809 list_first_entry(&ci->i_cap_snaps,
3810 struct ceph_cap_snap, ci_item);
3811 snap_follows = capsnap->follows;
3813 spin_unlock(&ci->i_ceph_lock);
3815 if (recon_state->msg_version >= 2) {
3816 int num_fcntl_locks, num_flock_locks;
3817 struct ceph_filelock *flocks = NULL;
3818 size_t struct_len, total_len = sizeof(u64);
3822 if (rec.v2.flock_len) {
3823 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3825 num_fcntl_locks = 0;
3826 num_flock_locks = 0;
3828 if (num_fcntl_locks + num_flock_locks > 0) {
3829 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3830 sizeof(struct ceph_filelock),
3836 err = ceph_encode_locks_to_buffer(inode, flocks,
3851 if (recon_state->msg_version >= 3) {
3852 /* version, compat_version and struct_len */
3853 total_len += 2 * sizeof(u8) + sizeof(u32);
3857 * number of encoded locks is stable, so copy to pagelist
3859 struct_len = 2 * sizeof(u32) +
3860 (num_fcntl_locks + num_flock_locks) *
3861 sizeof(struct ceph_filelock);
3862 rec.v2.flock_len = cpu_to_le32(struct_len);
3864 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3867 struct_len += sizeof(u64); /* snap_follows */
3869 total_len += struct_len;
3871 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3872 err = send_reconnect_partial(recon_state);
3874 goto out_freeflocks;
3875 pagelist = recon_state->pagelist;
3878 err = ceph_pagelist_reserve(pagelist, total_len);
3880 goto out_freeflocks;
3882 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3883 if (recon_state->msg_version >= 3) {
3884 ceph_pagelist_encode_8(pagelist, struct_v);
3885 ceph_pagelist_encode_8(pagelist, 1);
3886 ceph_pagelist_encode_32(pagelist, struct_len);
3888 ceph_pagelist_encode_string(pagelist, path, pathlen);
3889 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3890 ceph_locks_to_pagelist(flocks, pagelist,
3891 num_fcntl_locks, num_flock_locks);
3893 ceph_pagelist_encode_64(pagelist, snap_follows);
3897 err = ceph_pagelist_reserve(pagelist,
3898 sizeof(u64) + sizeof(u32) +
3899 pathlen + sizeof(rec.v1));
3903 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3904 ceph_pagelist_encode_string(pagelist, path, pathlen);
3905 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3909 ceph_mdsc_free_path(path, pathlen);
3911 recon_state->nr_caps++;
3915 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3916 struct ceph_reconnect_state *recon_state)
3919 struct ceph_pagelist *pagelist = recon_state->pagelist;
3922 if (recon_state->msg_version >= 4) {
3923 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3929 * snaprealms. we provide mds with the ino, seq (version), and
3930 * parent for all of our realms. If the mds has any newer info,
3933 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3934 struct ceph_snap_realm *realm =
3935 rb_entry(p, struct ceph_snap_realm, node);
3936 struct ceph_mds_snaprealm_reconnect sr_rec;
3938 if (recon_state->msg_version >= 4) {
3939 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3942 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3943 err = send_reconnect_partial(recon_state);
3946 pagelist = recon_state->pagelist;
3949 err = ceph_pagelist_reserve(pagelist, need);
3953 ceph_pagelist_encode_8(pagelist, 1);
3954 ceph_pagelist_encode_8(pagelist, 1);
3955 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3958 dout(" adding snap realm %llx seq %lld parent %llx\n",
3959 realm->ino, realm->seq, realm->parent_ino);
3960 sr_rec.ino = cpu_to_le64(realm->ino);
3961 sr_rec.seq = cpu_to_le64(realm->seq);
3962 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3964 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3968 recon_state->nr_realms++;
3976 * If an MDS fails and recovers, clients need to reconnect in order to
3977 * reestablish shared state. This includes all caps issued through
3978 * this session _and_ the snap_realm hierarchy. Because it's not
3979 * clear which snap realms the mds cares about, we send everything we
3980 * know about.. that ensures we'll then get any new info the
3981 * recovering MDS might have.
3983 * This is a relatively heavyweight operation, but it's rare.
3985 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3986 struct ceph_mds_session *session)
3988 struct ceph_msg *reply;
3989 int mds = session->s_mds;
3991 struct ceph_reconnect_state recon_state = {
3996 pr_info("mds%d reconnect start\n", mds);
3998 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3999 if (!recon_state.pagelist)
4000 goto fail_nopagelist;
4002 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4006 xa_destroy(&session->s_delegated_inos);
4008 mutex_lock(&session->s_mutex);
4009 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4012 dout("session %p state %s\n", session,
4013 ceph_session_state_name(session->s_state));
4015 spin_lock(&session->s_gen_ttl_lock);
4016 session->s_cap_gen++;
4017 spin_unlock(&session->s_gen_ttl_lock);
4019 spin_lock(&session->s_cap_lock);
4020 /* don't know if session is readonly */
4021 session->s_readonly = 0;
4023 * notify __ceph_remove_cap() that we are composing cap reconnect.
4024 * If a cap get released before being added to the cap reconnect,
4025 * __ceph_remove_cap() should skip queuing cap release.
4027 session->s_cap_reconnect = 1;
4028 /* drop old cap expires; we're about to reestablish that state */
4029 detach_cap_releases(session, &dispose);
4030 spin_unlock(&session->s_cap_lock);
4031 dispose_cap_releases(mdsc, &dispose);
4033 /* trim unused caps to reduce MDS's cache rejoin time */
4034 if (mdsc->fsc->sb->s_root)
4035 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4037 ceph_con_close(&session->s_con);
4038 ceph_con_open(&session->s_con,
4039 CEPH_ENTITY_TYPE_MDS, mds,
4040 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4042 /* replay unsafe requests */
4043 replay_unsafe_requests(mdsc, session);
4045 ceph_early_kick_flushing_caps(mdsc, session);
4047 down_read(&mdsc->snap_rwsem);
4049 /* placeholder for nr_caps */
4050 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4054 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4055 recon_state.msg_version = 3;
4056 recon_state.allow_multi = true;
4057 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4058 recon_state.msg_version = 3;
4060 recon_state.msg_version = 2;
4062 /* trsaverse this session's caps */
4063 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4065 spin_lock(&session->s_cap_lock);
4066 session->s_cap_reconnect = 0;
4067 spin_unlock(&session->s_cap_lock);
4072 /* check if all realms can be encoded into current message */
4073 if (mdsc->num_snap_realms) {
4075 recon_state.pagelist->length +
4076 mdsc->num_snap_realms *
4077 sizeof(struct ceph_mds_snaprealm_reconnect);
4078 if (recon_state.msg_version >= 4) {
4079 /* number of realms */
4080 total_len += sizeof(u32);
4081 /* version, compat_version and struct_len */
4082 total_len += mdsc->num_snap_realms *
4083 (2 * sizeof(u8) + sizeof(u32));
4085 if (total_len > RECONNECT_MAX_SIZE) {
4086 if (!recon_state.allow_multi) {
4090 if (recon_state.nr_caps) {
4091 err = send_reconnect_partial(&recon_state);
4095 recon_state.msg_version = 5;
4099 err = encode_snap_realms(mdsc, &recon_state);
4103 if (recon_state.msg_version >= 5) {
4104 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4109 if (recon_state.nr_caps || recon_state.nr_realms) {
4111 list_first_entry(&recon_state.pagelist->head,
4113 __le32 *addr = kmap_atomic(page);
4114 if (recon_state.nr_caps) {
4115 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4116 *addr = cpu_to_le32(recon_state.nr_caps);
4117 } else if (recon_state.msg_version >= 4) {
4118 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4120 kunmap_atomic(addr);
4123 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4124 if (recon_state.msg_version >= 4)
4125 reply->hdr.compat_version = cpu_to_le16(4);
4127 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4128 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4130 ceph_con_send(&session->s_con, reply);
4132 mutex_unlock(&session->s_mutex);
4134 mutex_lock(&mdsc->mutex);
4135 __wake_requests(mdsc, &session->s_waiting);
4136 mutex_unlock(&mdsc->mutex);
4138 up_read(&mdsc->snap_rwsem);
4139 ceph_pagelist_release(recon_state.pagelist);
4143 ceph_msg_put(reply);
4144 up_read(&mdsc->snap_rwsem);
4145 mutex_unlock(&session->s_mutex);
4147 ceph_pagelist_release(recon_state.pagelist);
4149 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4155 * compare old and new mdsmaps, kicking requests
4156 * and closing out old connections as necessary
4158 * called under mdsc->mutex.
4160 static void check_new_map(struct ceph_mds_client *mdsc,
4161 struct ceph_mdsmap *newmap,
4162 struct ceph_mdsmap *oldmap)
4165 int oldstate, newstate;
4166 struct ceph_mds_session *s;
4168 dout("check_new_map new %u old %u\n",
4169 newmap->m_epoch, oldmap->m_epoch);
4171 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4172 if (!mdsc->sessions[i])
4174 s = mdsc->sessions[i];
4175 oldstate = ceph_mdsmap_get_state(oldmap, i);
4176 newstate = ceph_mdsmap_get_state(newmap, i);
4178 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4179 i, ceph_mds_state_name(oldstate),
4180 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4181 ceph_mds_state_name(newstate),
4182 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4183 ceph_session_state_name(s->s_state));
4185 if (i >= newmap->possible_max_rank) {
4186 /* force close session for stopped mds */
4187 ceph_get_mds_session(s);
4188 __unregister_session(mdsc, s);
4189 __wake_requests(mdsc, &s->s_waiting);
4190 mutex_unlock(&mdsc->mutex);
4192 mutex_lock(&s->s_mutex);
4193 cleanup_session_requests(mdsc, s);
4194 remove_session_caps(s);
4195 mutex_unlock(&s->s_mutex);
4197 ceph_put_mds_session(s);
4199 mutex_lock(&mdsc->mutex);
4200 kick_requests(mdsc, i);
4204 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4205 ceph_mdsmap_get_addr(newmap, i),
4206 sizeof(struct ceph_entity_addr))) {
4208 mutex_unlock(&mdsc->mutex);
4209 mutex_lock(&s->s_mutex);
4210 mutex_lock(&mdsc->mutex);
4211 ceph_con_close(&s->s_con);
4212 mutex_unlock(&s->s_mutex);
4213 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4214 } else if (oldstate == newstate) {
4215 continue; /* nothing new with this mds */
4221 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4222 newstate >= CEPH_MDS_STATE_RECONNECT) {
4223 mutex_unlock(&mdsc->mutex);
4224 send_mds_reconnect(mdsc, s);
4225 mutex_lock(&mdsc->mutex);
4229 * kick request on any mds that has gone active.
4231 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4232 newstate >= CEPH_MDS_STATE_ACTIVE) {
4233 if (oldstate != CEPH_MDS_STATE_CREATING &&
4234 oldstate != CEPH_MDS_STATE_STARTING)
4235 pr_info("mds%d recovery completed\n", s->s_mds);
4236 kick_requests(mdsc, i);
4237 mutex_unlock(&mdsc->mutex);
4238 mutex_lock(&s->s_mutex);
4239 mutex_lock(&mdsc->mutex);
4240 ceph_kick_flushing_caps(mdsc, s);
4241 mutex_unlock(&s->s_mutex);
4242 wake_up_session_caps(s, RECONNECT);
4246 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4247 s = mdsc->sessions[i];
4250 if (!ceph_mdsmap_is_laggy(newmap, i))
4252 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4253 s->s_state == CEPH_MDS_SESSION_HUNG ||
4254 s->s_state == CEPH_MDS_SESSION_CLOSING) {
4255 dout(" connecting to export targets of laggy mds%d\n",
4257 __open_export_target_sessions(mdsc, s);
4269 * caller must hold session s_mutex, dentry->d_lock
4271 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4273 struct ceph_dentry_info *di = ceph_dentry(dentry);
4275 ceph_put_mds_session(di->lease_session);
4276 di->lease_session = NULL;
4279 static void handle_lease(struct ceph_mds_client *mdsc,
4280 struct ceph_mds_session *session,
4281 struct ceph_msg *msg)
4283 struct super_block *sb = mdsc->fsc->sb;
4284 struct inode *inode;
4285 struct dentry *parent, *dentry;
4286 struct ceph_dentry_info *di;
4287 int mds = session->s_mds;
4288 struct ceph_mds_lease *h = msg->front.iov_base;
4290 struct ceph_vino vino;
4294 dout("handle_lease from mds%d\n", mds);
4297 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4299 vino.ino = le64_to_cpu(h->ino);
4300 vino.snap = CEPH_NOSNAP;
4301 seq = le32_to_cpu(h->seq);
4302 dname.len = get_unaligned_le32(h + 1);
4303 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4305 dname.name = (void *)(h + 1) + sizeof(u32);
4308 inode = ceph_find_inode(sb, vino);
4309 dout("handle_lease %s, ino %llx %p %.*s\n",
4310 ceph_lease_op_name(h->action), vino.ino, inode,
4311 dname.len, dname.name);
4313 mutex_lock(&session->s_mutex);
4314 inc_session_sequence(session);
4317 dout("handle_lease no inode %llx\n", vino.ino);
4322 parent = d_find_alias(inode);
4324 dout("no parent dentry on inode %p\n", inode);
4326 goto release; /* hrm... */
4328 dname.hash = full_name_hash(parent, dname.name, dname.len);
4329 dentry = d_lookup(parent, &dname);
4334 spin_lock(&dentry->d_lock);
4335 di = ceph_dentry(dentry);
4336 switch (h->action) {
4337 case CEPH_MDS_LEASE_REVOKE:
4338 if (di->lease_session == session) {
4339 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4340 h->seq = cpu_to_le32(di->lease_seq);
4341 __ceph_mdsc_drop_dentry_lease(dentry);
4346 case CEPH_MDS_LEASE_RENEW:
4347 if (di->lease_session == session &&
4348 di->lease_gen == session->s_cap_gen &&
4349 di->lease_renew_from &&
4350 di->lease_renew_after == 0) {
4351 unsigned long duration =
4352 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4354 di->lease_seq = seq;
4355 di->time = di->lease_renew_from + duration;
4356 di->lease_renew_after = di->lease_renew_from +
4358 di->lease_renew_from = 0;
4362 spin_unlock(&dentry->d_lock);
4369 /* let's just reuse the same message */
4370 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4372 ceph_con_send(&session->s_con, msg);
4375 mutex_unlock(&session->s_mutex);
4376 /* avoid calling iput_final() in mds dispatch threads */
4377 ceph_async_iput(inode);
4381 pr_err("corrupt lease message\n");
4385 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4386 struct dentry *dentry, char action,
4389 struct ceph_msg *msg;
4390 struct ceph_mds_lease *lease;
4392 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4394 dout("lease_send_msg identry %p %s to mds%d\n",
4395 dentry, ceph_lease_op_name(action), session->s_mds);
4397 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4400 lease = msg->front.iov_base;
4401 lease->action = action;
4402 lease->seq = cpu_to_le32(seq);
4404 spin_lock(&dentry->d_lock);
4405 dir = d_inode(dentry->d_parent);
4406 lease->ino = cpu_to_le64(ceph_ino(dir));
4407 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4409 put_unaligned_le32(dentry->d_name.len, lease + 1);
4410 memcpy((void *)(lease + 1) + 4,
4411 dentry->d_name.name, dentry->d_name.len);
4412 spin_unlock(&dentry->d_lock);
4414 * if this is a preemptive lease RELEASE, no need to
4415 * flush request stream, since the actual request will
4418 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4420 ceph_con_send(&session->s_con, msg);
4424 * lock unlock sessions, to wait ongoing session activities
4426 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4430 mutex_lock(&mdsc->mutex);
4431 for (i = 0; i < mdsc->max_sessions; i++) {
4432 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4435 mutex_unlock(&mdsc->mutex);
4436 mutex_lock(&s->s_mutex);
4437 mutex_unlock(&s->s_mutex);
4438 ceph_put_mds_session(s);
4439 mutex_lock(&mdsc->mutex);
4441 mutex_unlock(&mdsc->mutex);
4444 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4446 struct ceph_fs_client *fsc = mdsc->fsc;
4448 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4451 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4454 if (!READ_ONCE(fsc->blocklisted))
4457 pr_info("auto reconnect after blocklisted\n");
4458 ceph_force_reconnect(fsc->sb);
4461 bool check_session_state(struct ceph_mds_session *s)
4463 switch (s->s_state) {
4464 case CEPH_MDS_SESSION_OPEN:
4465 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4466 s->s_state = CEPH_MDS_SESSION_HUNG;
4467 pr_info("mds%d hung\n", s->s_mds);
4470 case CEPH_MDS_SESSION_CLOSING:
4471 /* Should never reach this when we're unmounting */
4472 WARN_ON_ONCE(s->s_ttl);
4474 case CEPH_MDS_SESSION_NEW:
4475 case CEPH_MDS_SESSION_RESTARTING:
4476 case CEPH_MDS_SESSION_CLOSED:
4477 case CEPH_MDS_SESSION_REJECTED:
4485 * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4486 * then we need to retransmit that request.
4488 void inc_session_sequence(struct ceph_mds_session *s)
4490 lockdep_assert_held(&s->s_mutex);
4494 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4497 dout("resending session close request for mds%d\n", s->s_mds);
4498 ret = request_close_session(s);
4500 pr_err("unable to close session to mds%d: %d\n",
4506 * delayed work -- periodically trim expired leases, renew caps with mds. If
4507 * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4508 * workqueue delay value of 5 secs will be used.
4510 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4512 unsigned long max_delay = HZ * 5;
4514 /* 5 secs default delay */
4515 if (!delay || (delay > max_delay))
4517 schedule_delayed_work(&mdsc->delayed_work,
4518 round_jiffies_relative(delay));
4521 static void delayed_work(struct work_struct *work)
4523 struct ceph_mds_client *mdsc =
4524 container_of(work, struct ceph_mds_client, delayed_work.work);
4525 unsigned long delay;
4530 dout("mdsc delayed_work\n");
4535 mutex_lock(&mdsc->mutex);
4536 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4537 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4538 mdsc->last_renew_caps);
4540 mdsc->last_renew_caps = jiffies;
4542 for (i = 0; i < mdsc->max_sessions; i++) {
4543 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4547 if (!check_session_state(s)) {
4548 ceph_put_mds_session(s);
4551 mutex_unlock(&mdsc->mutex);
4553 mutex_lock(&s->s_mutex);
4555 send_renew_caps(mdsc, s);
4557 ceph_con_keepalive(&s->s_con);
4558 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4559 s->s_state == CEPH_MDS_SESSION_HUNG)
4560 ceph_send_cap_releases(mdsc, s);
4561 mutex_unlock(&s->s_mutex);
4562 ceph_put_mds_session(s);
4564 mutex_lock(&mdsc->mutex);
4566 mutex_unlock(&mdsc->mutex);
4568 delay = ceph_check_delayed_caps(mdsc);
4570 ceph_queue_cap_reclaim_work(mdsc);
4572 ceph_trim_snapid_map(mdsc);
4574 maybe_recover_session(mdsc);
4576 schedule_delayed(mdsc, delay);
4579 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4582 struct ceph_mds_client *mdsc;
4585 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4589 mutex_init(&mdsc->mutex);
4590 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4591 if (!mdsc->mdsmap) {
4596 init_completion(&mdsc->safe_umount_waiters);
4597 init_waitqueue_head(&mdsc->session_close_wq);
4598 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4599 mdsc->sessions = NULL;
4600 atomic_set(&mdsc->num_sessions, 0);
4601 mdsc->max_sessions = 0;
4603 atomic64_set(&mdsc->quotarealms_count, 0);
4604 mdsc->quotarealms_inodes = RB_ROOT;
4605 mutex_init(&mdsc->quotarealms_inodes_mutex);
4606 mdsc->last_snap_seq = 0;
4607 init_rwsem(&mdsc->snap_rwsem);
4608 mdsc->snap_realms = RB_ROOT;
4609 INIT_LIST_HEAD(&mdsc->snap_empty);
4610 mdsc->num_snap_realms = 0;
4611 spin_lock_init(&mdsc->snap_empty_lock);
4613 mdsc->oldest_tid = 0;
4614 mdsc->request_tree = RB_ROOT;
4615 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4616 mdsc->last_renew_caps = jiffies;
4617 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4618 INIT_LIST_HEAD(&mdsc->cap_wait_list);
4619 spin_lock_init(&mdsc->cap_delay_lock);
4620 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4621 spin_lock_init(&mdsc->snap_flush_lock);
4622 mdsc->last_cap_flush_tid = 1;
4623 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4624 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4625 mdsc->num_cap_flushing = 0;
4626 spin_lock_init(&mdsc->cap_dirty_lock);
4627 init_waitqueue_head(&mdsc->cap_flushing_wq);
4628 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4629 atomic_set(&mdsc->cap_reclaim_pending, 0);
4630 err = ceph_metric_init(&mdsc->metric);
4634 spin_lock_init(&mdsc->dentry_list_lock);
4635 INIT_LIST_HEAD(&mdsc->dentry_leases);
4636 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4638 ceph_caps_init(mdsc);
4639 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4641 spin_lock_init(&mdsc->snapid_map_lock);
4642 mdsc->snapid_map_tree = RB_ROOT;
4643 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4645 init_rwsem(&mdsc->pool_perm_rwsem);
4646 mdsc->pool_perm_tree = RB_ROOT;
4648 strscpy(mdsc->nodename, utsname()->nodename,
4649 sizeof(mdsc->nodename));
4655 kfree(mdsc->mdsmap);
4662 * Wait for safe replies on open mds requests. If we time out, drop
4663 * all requests from the tree to avoid dangling dentry refs.
4665 static void wait_requests(struct ceph_mds_client *mdsc)
4667 struct ceph_options *opts = mdsc->fsc->client->options;
4668 struct ceph_mds_request *req;
4670 mutex_lock(&mdsc->mutex);
4671 if (__get_oldest_req(mdsc)) {
4672 mutex_unlock(&mdsc->mutex);
4674 dout("wait_requests waiting for requests\n");
4675 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4676 ceph_timeout_jiffies(opts->mount_timeout));
4678 /* tear down remaining requests */
4679 mutex_lock(&mdsc->mutex);
4680 while ((req = __get_oldest_req(mdsc))) {
4681 dout("wait_requests timed out on tid %llu\n",
4683 list_del_init(&req->r_wait);
4684 __unregister_request(mdsc, req);
4687 mutex_unlock(&mdsc->mutex);
4688 dout("wait_requests done\n");
4692 * called before mount is ro, and before dentries are torn down.
4693 * (hmm, does this still race with new lookups?)
4695 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4697 dout("pre_umount\n");
4700 lock_unlock_sessions(mdsc);
4701 ceph_flush_dirty_caps(mdsc);
4702 wait_requests(mdsc);
4705 * wait for reply handlers to drop their request refs and
4706 * their inode/dcache refs
4710 ceph_cleanup_quotarealms_inodes(mdsc);
4714 * wait for all write mds requests to flush.
4716 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4718 struct ceph_mds_request *req = NULL, *nextreq;
4721 mutex_lock(&mdsc->mutex);
4722 dout("wait_unsafe_requests want %lld\n", want_tid);
4724 req = __get_oldest_req(mdsc);
4725 while (req && req->r_tid <= want_tid) {
4726 /* find next request */
4727 n = rb_next(&req->r_node);
4729 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4732 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4733 (req->r_op & CEPH_MDS_OP_WRITE)) {
4735 ceph_mdsc_get_request(req);
4737 ceph_mdsc_get_request(nextreq);
4738 mutex_unlock(&mdsc->mutex);
4739 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
4740 req->r_tid, want_tid);
4741 wait_for_completion(&req->r_safe_completion);
4742 mutex_lock(&mdsc->mutex);
4743 ceph_mdsc_put_request(req);
4745 break; /* next dne before, so we're done! */
4746 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4747 /* next request was removed from tree */
4748 ceph_mdsc_put_request(nextreq);
4751 ceph_mdsc_put_request(nextreq); /* won't go away */
4755 mutex_unlock(&mdsc->mutex);
4756 dout("wait_unsafe_requests done\n");
4759 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4761 u64 want_tid, want_flush;
4763 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4767 mutex_lock(&mdsc->mutex);
4768 want_tid = mdsc->last_tid;
4769 mutex_unlock(&mdsc->mutex);
4771 ceph_flush_dirty_caps(mdsc);
4772 spin_lock(&mdsc->cap_dirty_lock);
4773 want_flush = mdsc->last_cap_flush_tid;
4774 if (!list_empty(&mdsc->cap_flush_list)) {
4775 struct ceph_cap_flush *cf =
4776 list_last_entry(&mdsc->cap_flush_list,
4777 struct ceph_cap_flush, g_list);
4780 spin_unlock(&mdsc->cap_dirty_lock);
4782 dout("sync want tid %lld flush_seq %lld\n",
4783 want_tid, want_flush);
4785 wait_unsafe_requests(mdsc, want_tid);
4786 wait_caps_flush(mdsc, want_flush);
4790 * true if all sessions are closed, or we force unmount
4792 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4794 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4796 return atomic_read(&mdsc->num_sessions) <= skipped;
4800 * called after sb is ro.
4802 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4804 struct ceph_options *opts = mdsc->fsc->client->options;
4805 struct ceph_mds_session *session;
4809 dout("close_sessions\n");
4811 /* close sessions */
4812 mutex_lock(&mdsc->mutex);
4813 for (i = 0; i < mdsc->max_sessions; i++) {
4814 session = __ceph_lookup_mds_session(mdsc, i);
4817 mutex_unlock(&mdsc->mutex);
4818 mutex_lock(&session->s_mutex);
4819 if (__close_session(mdsc, session) <= 0)
4821 mutex_unlock(&session->s_mutex);
4822 ceph_put_mds_session(session);
4823 mutex_lock(&mdsc->mutex);
4825 mutex_unlock(&mdsc->mutex);
4827 dout("waiting for sessions to close\n");
4828 wait_event_timeout(mdsc->session_close_wq,
4829 done_closing_sessions(mdsc, skipped),
4830 ceph_timeout_jiffies(opts->mount_timeout));
4832 /* tear down remaining sessions */
4833 mutex_lock(&mdsc->mutex);
4834 for (i = 0; i < mdsc->max_sessions; i++) {
4835 if (mdsc->sessions[i]) {
4836 session = ceph_get_mds_session(mdsc->sessions[i]);
4837 __unregister_session(mdsc, session);
4838 mutex_unlock(&mdsc->mutex);
4839 mutex_lock(&session->s_mutex);
4840 remove_session_caps(session);
4841 mutex_unlock(&session->s_mutex);
4842 ceph_put_mds_session(session);
4843 mutex_lock(&mdsc->mutex);
4846 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4847 mutex_unlock(&mdsc->mutex);
4849 ceph_cleanup_snapid_map(mdsc);
4850 ceph_cleanup_empty_realms(mdsc);
4852 cancel_work_sync(&mdsc->cap_reclaim_work);
4853 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4858 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4860 struct ceph_mds_session *session;
4863 dout("force umount\n");
4865 mutex_lock(&mdsc->mutex);
4866 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4867 session = __ceph_lookup_mds_session(mdsc, mds);
4871 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4872 __unregister_session(mdsc, session);
4873 __wake_requests(mdsc, &session->s_waiting);
4874 mutex_unlock(&mdsc->mutex);
4876 mutex_lock(&session->s_mutex);
4877 __close_session(mdsc, session);
4878 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4879 cleanup_session_requests(mdsc, session);
4880 remove_session_caps(session);
4882 mutex_unlock(&session->s_mutex);
4883 ceph_put_mds_session(session);
4885 mutex_lock(&mdsc->mutex);
4886 kick_requests(mdsc, mds);
4888 __wake_requests(mdsc, &mdsc->waiting_for_map);
4889 mutex_unlock(&mdsc->mutex);
4892 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4896 * Make sure the delayed work stopped before releasing
4899 * Because the cancel_delayed_work_sync() will only
4900 * guarantee that the work finishes executing. But the
4901 * delayed work will re-arm itself again after that.
4903 flush_delayed_work(&mdsc->delayed_work);
4906 ceph_mdsmap_destroy(mdsc->mdsmap);
4907 kfree(mdsc->sessions);
4908 ceph_caps_finalize(mdsc);
4909 ceph_pool_perm_destroy(mdsc);
4912 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4914 struct ceph_mds_client *mdsc = fsc->mdsc;
4915 dout("mdsc_destroy %p\n", mdsc);
4920 /* flush out any connection work with references to us */
4923 ceph_mdsc_stop(mdsc);
4925 ceph_metric_destroy(&mdsc->metric);
4927 flush_delayed_work(&mdsc->metric.delayed_work);
4930 dout("mdsc_destroy %p done\n", mdsc);
4933 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4935 struct ceph_fs_client *fsc = mdsc->fsc;
4936 const char *mds_namespace = fsc->mount_options->mds_namespace;
4937 void *p = msg->front.iov_base;
4938 void *end = p + msg->front.iov_len;
4941 u32 mount_fscid = (u32)-1;
4944 ceph_decode_need(&p, end, sizeof(u32), bad);
4945 epoch = ceph_decode_32(&p);
4947 dout("handle_fsmap epoch %u\n", epoch);
4949 /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
4950 ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
4952 ceph_decode_32_safe(&p, end, num_fs, bad);
4953 while (num_fs-- > 0) {
4954 void *info_p, *info_end;
4958 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4959 p += 2; // info_v, info_cv
4960 info_len = ceph_decode_32(&p);
4961 ceph_decode_need(&p, end, info_len, bad);
4963 info_end = p + info_len;
4966 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4967 fscid = ceph_decode_32(&info_p);
4968 namelen = ceph_decode_32(&info_p);
4969 ceph_decode_need(&info_p, info_end, namelen, bad);
4971 if (mds_namespace &&
4972 strlen(mds_namespace) == namelen &&
4973 !strncmp(mds_namespace, (char *)info_p, namelen)) {
4974 mount_fscid = fscid;
4979 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4980 if (mount_fscid != (u32)-1) {
4981 fsc->client->monc.fs_cluster_id = mount_fscid;
4982 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4984 ceph_monc_renew_subs(&fsc->client->monc);
4992 pr_err("error decoding fsmap\n");
4994 mutex_lock(&mdsc->mutex);
4995 mdsc->mdsmap_err = err;
4996 __wake_requests(mdsc, &mdsc->waiting_for_map);
4997 mutex_unlock(&mdsc->mutex);
5001 * handle mds map update.
5003 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5007 void *p = msg->front.iov_base;
5008 void *end = p + msg->front.iov_len;
5009 struct ceph_mdsmap *newmap, *oldmap;
5010 struct ceph_fsid fsid;
5013 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5014 ceph_decode_copy(&p, &fsid, sizeof(fsid));
5015 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5017 epoch = ceph_decode_32(&p);
5018 maplen = ceph_decode_32(&p);
5019 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5021 /* do we need it? */
5022 mutex_lock(&mdsc->mutex);
5023 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5024 dout("handle_map epoch %u <= our %u\n",
5025 epoch, mdsc->mdsmap->m_epoch);
5026 mutex_unlock(&mdsc->mutex);
5030 newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5031 if (IS_ERR(newmap)) {
5032 err = PTR_ERR(newmap);
5036 /* swap into place */
5038 oldmap = mdsc->mdsmap;
5039 mdsc->mdsmap = newmap;
5040 check_new_map(mdsc, newmap, oldmap);
5041 ceph_mdsmap_destroy(oldmap);
5043 mdsc->mdsmap = newmap; /* first mds map */
5045 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5048 __wake_requests(mdsc, &mdsc->waiting_for_map);
5049 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5050 mdsc->mdsmap->m_epoch);
5052 mutex_unlock(&mdsc->mutex);
5053 schedule_delayed(mdsc, 0);
5057 mutex_unlock(&mdsc->mutex);
5059 pr_err("error decoding mdsmap %d\n", err);
5063 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5065 struct ceph_mds_session *s = con->private;
5067 if (ceph_get_mds_session(s))
5072 static void mds_put_con(struct ceph_connection *con)
5074 struct ceph_mds_session *s = con->private;
5076 ceph_put_mds_session(s);
5080 * if the client is unresponsive for long enough, the mds will kill
5081 * the session entirely.
5083 static void mds_peer_reset(struct ceph_connection *con)
5085 struct ceph_mds_session *s = con->private;
5086 struct ceph_mds_client *mdsc = s->s_mdsc;
5088 pr_warn("mds%d closed our session\n", s->s_mds);
5089 send_mds_reconnect(mdsc, s);
5092 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5094 struct ceph_mds_session *s = con->private;
5095 struct ceph_mds_client *mdsc = s->s_mdsc;
5096 int type = le16_to_cpu(msg->hdr.type);
5098 mutex_lock(&mdsc->mutex);
5099 if (__verify_registered_session(mdsc, s) < 0) {
5100 mutex_unlock(&mdsc->mutex);
5103 mutex_unlock(&mdsc->mutex);
5106 case CEPH_MSG_MDS_MAP:
5107 ceph_mdsc_handle_mdsmap(mdsc, msg);
5109 case CEPH_MSG_FS_MAP_USER:
5110 ceph_mdsc_handle_fsmap(mdsc, msg);
5112 case CEPH_MSG_CLIENT_SESSION:
5113 handle_session(s, msg);
5115 case CEPH_MSG_CLIENT_REPLY:
5116 handle_reply(s, msg);
5118 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5119 handle_forward(mdsc, s, msg);
5121 case CEPH_MSG_CLIENT_CAPS:
5122 ceph_handle_caps(s, msg);
5124 case CEPH_MSG_CLIENT_SNAP:
5125 ceph_handle_snap(mdsc, s, msg);
5127 case CEPH_MSG_CLIENT_LEASE:
5128 handle_lease(mdsc, s, msg);
5130 case CEPH_MSG_CLIENT_QUOTA:
5131 ceph_handle_quota(mdsc, s, msg);
5135 pr_err("received unknown message type %d %s\n", type,
5136 ceph_msg_type_name(type));
5147 * Note: returned pointer is the address of a structure that's
5148 * managed separately. Caller must *not* attempt to free it.
5150 static struct ceph_auth_handshake *
5151 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5153 struct ceph_mds_session *s = con->private;
5154 struct ceph_mds_client *mdsc = s->s_mdsc;
5155 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5156 struct ceph_auth_handshake *auth = &s->s_auth;
5159 ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5160 force_new, proto, NULL, NULL);
5162 return ERR_PTR(ret);
5167 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5168 void *challenge_buf, int challenge_buf_len)
5170 struct ceph_mds_session *s = con->private;
5171 struct ceph_mds_client *mdsc = s->s_mdsc;
5172 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5174 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5175 challenge_buf, challenge_buf_len);
5178 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5180 struct ceph_mds_session *s = con->private;
5181 struct ceph_mds_client *mdsc = s->s_mdsc;
5182 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5183 struct ceph_auth_handshake *auth = &s->s_auth;
5185 return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5186 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5187 NULL, NULL, NULL, NULL);
5190 static int mds_invalidate_authorizer(struct ceph_connection *con)
5192 struct ceph_mds_session *s = con->private;
5193 struct ceph_mds_client *mdsc = s->s_mdsc;
5194 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5196 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5198 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5201 static int mds_get_auth_request(struct ceph_connection *con,
5202 void *buf, int *buf_len,
5203 void **authorizer, int *authorizer_len)
5205 struct ceph_mds_session *s = con->private;
5206 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5207 struct ceph_auth_handshake *auth = &s->s_auth;
5210 ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5215 *authorizer = auth->authorizer_buf;
5216 *authorizer_len = auth->authorizer_buf_len;
5220 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5221 void *reply, int reply_len,
5222 void *buf, int *buf_len,
5223 void **authorizer, int *authorizer_len)
5225 struct ceph_mds_session *s = con->private;
5226 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5227 struct ceph_auth_handshake *auth = &s->s_auth;
5230 ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5235 *authorizer = auth->authorizer_buf;
5236 *authorizer_len = auth->authorizer_buf_len;
5240 static int mds_handle_auth_done(struct ceph_connection *con,
5241 u64 global_id, void *reply, int reply_len,
5242 u8 *session_key, int *session_key_len,
5243 u8 *con_secret, int *con_secret_len)
5245 struct ceph_mds_session *s = con->private;
5246 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5247 struct ceph_auth_handshake *auth = &s->s_auth;
5249 return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5250 session_key, session_key_len,
5251 con_secret, con_secret_len);
5254 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5255 int used_proto, int result,
5256 const int *allowed_protos, int proto_cnt,
5257 const int *allowed_modes, int mode_cnt)
5259 struct ceph_mds_session *s = con->private;
5260 struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5263 if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5265 allowed_protos, proto_cnt,
5266 allowed_modes, mode_cnt)) {
5267 ret = ceph_monc_validate_auth(monc);
5275 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5276 struct ceph_msg_header *hdr, int *skip)
5278 struct ceph_msg *msg;
5279 int type = (int) le16_to_cpu(hdr->type);
5280 int front_len = (int) le32_to_cpu(hdr->front_len);
5286 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5288 pr_err("unable to allocate msg type %d len %d\n",
5296 static int mds_sign_message(struct ceph_msg *msg)
5298 struct ceph_mds_session *s = msg->con->private;
5299 struct ceph_auth_handshake *auth = &s->s_auth;
5301 return ceph_auth_sign_message(auth, msg);
5304 static int mds_check_message_signature(struct ceph_msg *msg)
5306 struct ceph_mds_session *s = msg->con->private;
5307 struct ceph_auth_handshake *auth = &s->s_auth;
5309 return ceph_auth_check_message_signature(auth, msg);
5312 static const struct ceph_connection_operations mds_con_ops = {
5315 .alloc_msg = mds_alloc_msg,
5316 .dispatch = mds_dispatch,
5317 .peer_reset = mds_peer_reset,
5318 .get_authorizer = mds_get_authorizer,
5319 .add_authorizer_challenge = mds_add_authorizer_challenge,
5320 .verify_authorizer_reply = mds_verify_authorizer_reply,
5321 .invalidate_authorizer = mds_invalidate_authorizer,
5322 .sign_message = mds_sign_message,
5323 .check_message_signature = mds_check_message_signature,
5324 .get_auth_request = mds_get_auth_request,
5325 .handle_auth_reply_more = mds_handle_auth_reply_more,
5326 .handle_auth_done = mds_handle_auth_done,
5327 .handle_auth_bad_method = mds_handle_auth_bad_method,