1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/wait.h>
6 #include <linux/slab.h>
8 #include <linux/sched.h>
9 #include <linux/debugfs.h>
10 #include <linux/seq_file.h>
11 #include <linux/ratelimit.h>
12 #include <linux/bits.h>
13 #include <linux/ktime.h>
16 #include "mds_client.h"
18 #include <linux/ceph/ceph_features.h>
19 #include <linux/ceph/messenger.h>
20 #include <linux/ceph/decode.h>
21 #include <linux/ceph/pagelist.h>
22 #include <linux/ceph/auth.h>
23 #include <linux/ceph/debugfs.h>
25 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
28 * A cluster of MDS (metadata server) daemons is responsible for
29 * managing the file system namespace (the directory hierarchy and
30 * inodes) and for coordinating shared access to storage. Metadata is
31 * partitioning hierarchically across a number of servers, and that
32 * partition varies over time as the cluster adjusts the distribution
33 * in order to balance load.
35 * The MDS client is primarily responsible to managing synchronous
36 * metadata requests for operations like open, unlink, and so forth.
37 * If there is a MDS failure, we find out about it when we (possibly
38 * request and) receive a new MDS map, and can resubmit affected
41 * For the most part, though, we take advantage of a lossless
42 * communications channel to the MDS, and do not need to worry about
43 * timing out or resubmitting requests.
45 * We maintain a stateful "session" with each MDS we interact with.
46 * Within each session, we sent periodic heartbeat messages to ensure
47 * any capabilities or leases we have been issues remain valid. If
48 * the session times out and goes stale, our leases and capabilities
49 * are no longer valid.
52 struct ceph_reconnect_state {
53 struct ceph_mds_session *session;
54 int nr_caps, nr_realms;
55 struct ceph_pagelist *pagelist;
60 static void __wake_requests(struct ceph_mds_client *mdsc,
61 struct list_head *head);
62 static void ceph_cap_release_work(struct work_struct *work);
63 static void ceph_cap_reclaim_work(struct work_struct *work);
65 static const struct ceph_connection_operations mds_con_ops;
72 static int parse_reply_info_quota(void **p, void *end,
73 struct ceph_mds_reply_info_in *info)
75 u8 struct_v, struct_compat;
78 ceph_decode_8_safe(p, end, struct_v, bad);
79 ceph_decode_8_safe(p, end, struct_compat, bad);
80 /* struct_v is expected to be >= 1. we only
81 * understand encoding with struct_compat == 1. */
82 if (!struct_v || struct_compat != 1)
84 ceph_decode_32_safe(p, end, struct_len, bad);
85 ceph_decode_need(p, end, struct_len, bad);
86 end = *p + struct_len;
87 ceph_decode_64_safe(p, end, info->max_bytes, bad);
88 ceph_decode_64_safe(p, end, info->max_files, bad);
96 * parse individual inode info
98 static int parse_reply_info_in(void **p, void *end,
99 struct ceph_mds_reply_info_in *info,
105 if (features == (u64)-1) {
108 ceph_decode_8_safe(p, end, struct_v, bad);
109 ceph_decode_8_safe(p, end, struct_compat, bad);
110 /* struct_v is expected to be >= 1. we only understand
111 * encoding with struct_compat == 1. */
112 if (!struct_v || struct_compat != 1)
114 ceph_decode_32_safe(p, end, struct_len, bad);
115 ceph_decode_need(p, end, struct_len, bad);
116 end = *p + struct_len;
119 ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
121 *p += sizeof(struct ceph_mds_reply_inode) +
122 sizeof(*info->in->fragtree.splits) *
123 le32_to_cpu(info->in->fragtree.nsplits);
125 ceph_decode_32_safe(p, end, info->symlink_len, bad);
126 ceph_decode_need(p, end, info->symlink_len, bad);
128 *p += info->symlink_len;
130 ceph_decode_copy_safe(p, end, &info->dir_layout,
131 sizeof(info->dir_layout), bad);
132 ceph_decode_32_safe(p, end, info->xattr_len, bad);
133 ceph_decode_need(p, end, info->xattr_len, bad);
134 info->xattr_data = *p;
135 *p += info->xattr_len;
137 if (features == (u64)-1) {
139 ceph_decode_64_safe(p, end, info->inline_version, bad);
140 ceph_decode_32_safe(p, end, info->inline_len, bad);
141 ceph_decode_need(p, end, info->inline_len, bad);
142 info->inline_data = *p;
143 *p += info->inline_len;
145 err = parse_reply_info_quota(p, end, info);
149 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
150 if (info->pool_ns_len > 0) {
151 ceph_decode_need(p, end, info->pool_ns_len, bad);
152 info->pool_ns_data = *p;
153 *p += info->pool_ns_len;
157 ceph_decode_need(p, end, sizeof(info->btime), bad);
158 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
160 /* change attribute */
161 ceph_decode_64_safe(p, end, info->change_attr, bad);
165 ceph_decode_32_safe(p, end, info->dir_pin, bad);
167 info->dir_pin = -ENODATA;
170 /* snapshot birth time, remains zero for v<=2 */
172 ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
173 ceph_decode_copy(p, &info->snap_btime,
174 sizeof(info->snap_btime));
176 memset(&info->snap_btime, 0, sizeof(info->snap_btime));
181 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
182 ceph_decode_64_safe(p, end, info->inline_version, bad);
183 ceph_decode_32_safe(p, end, info->inline_len, bad);
184 ceph_decode_need(p, end, info->inline_len, bad);
185 info->inline_data = *p;
186 *p += info->inline_len;
188 info->inline_version = CEPH_INLINE_NONE;
190 if (features & CEPH_FEATURE_MDS_QUOTA) {
191 err = parse_reply_info_quota(p, end, info);
199 info->pool_ns_len = 0;
200 info->pool_ns_data = NULL;
201 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
202 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
203 if (info->pool_ns_len > 0) {
204 ceph_decode_need(p, end, info->pool_ns_len, bad);
205 info->pool_ns_data = *p;
206 *p += info->pool_ns_len;
210 if (features & CEPH_FEATURE_FS_BTIME) {
211 ceph_decode_need(p, end, sizeof(info->btime), bad);
212 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
213 ceph_decode_64_safe(p, end, info->change_attr, bad);
216 info->dir_pin = -ENODATA;
217 /* info->snap_btime remains zero */
226 static int parse_reply_info_dir(void **p, void *end,
227 struct ceph_mds_reply_dirfrag **dirfrag,
230 if (features == (u64)-1) {
231 u8 struct_v, struct_compat;
233 ceph_decode_8_safe(p, end, struct_v, bad);
234 ceph_decode_8_safe(p, end, struct_compat, bad);
235 /* struct_v is expected to be >= 1. we only understand
236 * encoding whose struct_compat == 1. */
237 if (!struct_v || struct_compat != 1)
239 ceph_decode_32_safe(p, end, struct_len, bad);
240 ceph_decode_need(p, end, struct_len, bad);
241 end = *p + struct_len;
244 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
246 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
247 if (unlikely(*p > end))
249 if (features == (u64)-1)
256 static int parse_reply_info_lease(void **p, void *end,
257 struct ceph_mds_reply_lease **lease,
260 if (features == (u64)-1) {
261 u8 struct_v, struct_compat;
263 ceph_decode_8_safe(p, end, struct_v, bad);
264 ceph_decode_8_safe(p, end, struct_compat, bad);
265 /* struct_v is expected to be >= 1. we only understand
266 * encoding whose struct_compat == 1. */
267 if (!struct_v || struct_compat != 1)
269 ceph_decode_32_safe(p, end, struct_len, bad);
270 ceph_decode_need(p, end, struct_len, bad);
271 end = *p + struct_len;
274 ceph_decode_need(p, end, sizeof(**lease), bad);
276 *p += sizeof(**lease);
277 if (features == (u64)-1)
285 * parse a normal reply, which may contain a (dir+)dentry and/or a
288 static int parse_reply_info_trace(void **p, void *end,
289 struct ceph_mds_reply_info_parsed *info,
294 if (info->head->is_dentry) {
295 err = parse_reply_info_in(p, end, &info->diri, features);
299 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
303 ceph_decode_32_safe(p, end, info->dname_len, bad);
304 ceph_decode_need(p, end, info->dname_len, bad);
306 *p += info->dname_len;
308 err = parse_reply_info_lease(p, end, &info->dlease, features);
313 if (info->head->is_target) {
314 err = parse_reply_info_in(p, end, &info->targeti, features);
319 if (unlikely(*p != end))
326 pr_err("problem parsing mds trace %d\n", err);
331 * parse readdir results
333 static int parse_reply_info_readdir(void **p, void *end,
334 struct ceph_mds_reply_info_parsed *info,
340 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
344 ceph_decode_need(p, end, sizeof(num) + 2, bad);
345 num = ceph_decode_32(p);
347 u16 flags = ceph_decode_16(p);
348 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
349 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
350 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
351 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
356 BUG_ON(!info->dir_entries);
357 if ((unsigned long)(info->dir_entries + num) >
358 (unsigned long)info->dir_entries + info->dir_buf_size) {
359 pr_err("dir contents are larger than expected\n");
366 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
368 ceph_decode_32_safe(p, end, rde->name_len, bad);
369 ceph_decode_need(p, end, rde->name_len, bad);
372 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
375 err = parse_reply_info_lease(p, end, &rde->lease, features);
379 err = parse_reply_info_in(p, end, &rde->inode, features);
382 /* ceph_readdir_prepopulate() will update it */
389 /* Skip over any unrecognized fields */
396 pr_err("problem parsing dir contents %d\n", err);
401 * parse fcntl F_GETLK results
403 static int parse_reply_info_filelock(void **p, void *end,
404 struct ceph_mds_reply_info_parsed *info,
407 if (*p + sizeof(*info->filelock_reply) > end)
410 info->filelock_reply = *p;
412 /* Skip over any unrecognized fields */
420 #if BITS_PER_LONG == 64
422 #define DELEGATED_INO_AVAILABLE xa_mk_value(1)
424 static int ceph_parse_deleg_inos(void **p, void *end,
425 struct ceph_mds_session *s)
429 ceph_decode_32_safe(p, end, sets, bad);
430 dout("got %u sets of delegated inodes\n", sets);
434 ceph_decode_64_safe(p, end, start, bad);
435 ceph_decode_64_safe(p, end, len, bad);
437 int err = xa_insert(&s->s_delegated_inos, ino = start++,
438 DELEGATED_INO_AVAILABLE,
441 dout("added delegated inode 0x%llx\n",
443 } else if (err == -EBUSY) {
444 pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
456 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
461 xa_for_each(&s->s_delegated_inos, ino, val) {
462 val = xa_erase(&s->s_delegated_inos, ino);
463 if (val == DELEGATED_INO_AVAILABLE)
469 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
471 return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
474 #else /* BITS_PER_LONG == 64 */
476 * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
477 * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
480 static int ceph_parse_deleg_inos(void **p, void *end,
481 struct ceph_mds_session *s)
485 ceph_decode_32_safe(p, end, sets, bad);
487 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
493 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
498 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
502 #endif /* BITS_PER_LONG == 64 */
505 * parse create results
507 static int parse_reply_info_create(void **p, void *end,
508 struct ceph_mds_reply_info_parsed *info,
509 u64 features, struct ceph_mds_session *s)
513 if (features == (u64)-1 ||
514 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
516 /* Malformed reply? */
517 info->has_create_ino = false;
518 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
519 u8 struct_v, struct_compat;
522 info->has_create_ino = true;
523 ceph_decode_8_safe(p, end, struct_v, bad);
524 ceph_decode_8_safe(p, end, struct_compat, bad);
525 ceph_decode_32_safe(p, end, len, bad);
526 ceph_decode_64_safe(p, end, info->ino, bad);
527 ret = ceph_parse_deleg_inos(p, end, s);
532 ceph_decode_64_safe(p, end, info->ino, bad);
533 info->has_create_ino = true;
540 /* Skip over any unrecognized fields */
548 * parse extra results
550 static int parse_reply_info_extra(void **p, void *end,
551 struct ceph_mds_reply_info_parsed *info,
552 u64 features, struct ceph_mds_session *s)
554 u32 op = le32_to_cpu(info->head->op);
556 if (op == CEPH_MDS_OP_GETFILELOCK)
557 return parse_reply_info_filelock(p, end, info, features);
558 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
559 return parse_reply_info_readdir(p, end, info, features);
560 else if (op == CEPH_MDS_OP_CREATE)
561 return parse_reply_info_create(p, end, info, features, s);
567 * parse entire mds reply
569 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
570 struct ceph_mds_reply_info_parsed *info,
577 info->head = msg->front.iov_base;
578 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
579 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
582 ceph_decode_32_safe(&p, end, len, bad);
584 ceph_decode_need(&p, end, len, bad);
585 err = parse_reply_info_trace(&p, p+len, info, features);
591 ceph_decode_32_safe(&p, end, len, bad);
593 ceph_decode_need(&p, end, len, bad);
594 err = parse_reply_info_extra(&p, p+len, info, features, s);
600 ceph_decode_32_safe(&p, end, len, bad);
601 info->snapblob_len = len;
612 pr_err("mds parse_reply err %d\n", err);
616 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
618 if (!info->dir_entries)
620 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
627 const char *ceph_session_state_name(int s)
630 case CEPH_MDS_SESSION_NEW: return "new";
631 case CEPH_MDS_SESSION_OPENING: return "opening";
632 case CEPH_MDS_SESSION_OPEN: return "open";
633 case CEPH_MDS_SESSION_HUNG: return "hung";
634 case CEPH_MDS_SESSION_CLOSING: return "closing";
635 case CEPH_MDS_SESSION_CLOSED: return "closed";
636 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
637 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
638 case CEPH_MDS_SESSION_REJECTED: return "rejected";
639 default: return "???";
643 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
645 if (refcount_inc_not_zero(&s->s_ref)) {
646 dout("mdsc get_session %p %d -> %d\n", s,
647 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
650 dout("mdsc get_session %p 0 -- FAIL\n", s);
655 void ceph_put_mds_session(struct ceph_mds_session *s)
657 dout("mdsc put_session %p %d -> %d\n", s,
658 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
659 if (refcount_dec_and_test(&s->s_ref)) {
660 if (s->s_auth.authorizer)
661 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
662 WARN_ON(mutex_is_locked(&s->s_mutex));
663 xa_destroy(&s->s_delegated_inos);
669 * called under mdsc->mutex
671 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
674 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
676 return ceph_get_mds_session(mdsc->sessions[mds]);
679 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
681 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
687 static int __verify_registered_session(struct ceph_mds_client *mdsc,
688 struct ceph_mds_session *s)
690 if (s->s_mds >= mdsc->max_sessions ||
691 mdsc->sessions[s->s_mds] != s)
697 * create+register a new session for given mds.
698 * called under mdsc->mutex.
700 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
703 struct ceph_mds_session *s;
705 if (mds >= mdsc->mdsmap->possible_max_rank)
706 return ERR_PTR(-EINVAL);
708 s = kzalloc(sizeof(*s), GFP_NOFS);
710 return ERR_PTR(-ENOMEM);
712 if (mds >= mdsc->max_sessions) {
713 int newmax = 1 << get_count_order(mds + 1);
714 struct ceph_mds_session **sa;
716 dout("%s: realloc to %d\n", __func__, newmax);
717 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
720 if (mdsc->sessions) {
721 memcpy(sa, mdsc->sessions,
722 mdsc->max_sessions * sizeof(void *));
723 kfree(mdsc->sessions);
726 mdsc->max_sessions = newmax;
729 dout("%s: mds%d\n", __func__, mds);
732 s->s_state = CEPH_MDS_SESSION_NEW;
735 mutex_init(&s->s_mutex);
737 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
739 spin_lock_init(&s->s_gen_ttl_lock);
741 s->s_cap_ttl = jiffies - 1;
743 spin_lock_init(&s->s_cap_lock);
744 s->s_renew_requested = 0;
746 INIT_LIST_HEAD(&s->s_caps);
748 refcount_set(&s->s_ref, 1);
749 INIT_LIST_HEAD(&s->s_waiting);
750 INIT_LIST_HEAD(&s->s_unsafe);
751 xa_init(&s->s_delegated_inos);
752 s->s_num_cap_releases = 0;
753 s->s_cap_reconnect = 0;
754 s->s_cap_iterator = NULL;
755 INIT_LIST_HEAD(&s->s_cap_releases);
756 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
758 INIT_LIST_HEAD(&s->s_cap_dirty);
759 INIT_LIST_HEAD(&s->s_cap_flushing);
761 mdsc->sessions[mds] = s;
762 atomic_inc(&mdsc->num_sessions);
763 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
765 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
766 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
772 return ERR_PTR(-ENOMEM);
776 * called under mdsc->mutex
778 static void __unregister_session(struct ceph_mds_client *mdsc,
779 struct ceph_mds_session *s)
781 dout("__unregister_session mds%d %p\n", s->s_mds, s);
782 BUG_ON(mdsc->sessions[s->s_mds] != s);
783 mdsc->sessions[s->s_mds] = NULL;
784 ceph_con_close(&s->s_con);
785 ceph_put_mds_session(s);
786 atomic_dec(&mdsc->num_sessions);
790 * drop session refs in request.
792 * should be last request ref, or hold mdsc->mutex
794 static void put_request_session(struct ceph_mds_request *req)
796 if (req->r_session) {
797 ceph_put_mds_session(req->r_session);
798 req->r_session = NULL;
802 void ceph_mdsc_release_request(struct kref *kref)
804 struct ceph_mds_request *req = container_of(kref,
805 struct ceph_mds_request,
807 ceph_mdsc_release_dir_caps_no_check(req);
808 destroy_reply_info(&req->r_reply_info);
810 ceph_msg_put(req->r_request);
812 ceph_msg_put(req->r_reply);
814 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
815 /* avoid calling iput_final() in mds dispatch threads */
816 ceph_async_iput(req->r_inode);
819 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
820 ceph_async_iput(req->r_parent);
822 ceph_async_iput(req->r_target_inode);
825 if (req->r_old_dentry)
826 dput(req->r_old_dentry);
827 if (req->r_old_dentry_dir) {
829 * track (and drop pins for) r_old_dentry_dir
830 * separately, since r_old_dentry's d_parent may have
831 * changed between the dir mutex being dropped and
832 * this request being freed.
834 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
836 ceph_async_iput(req->r_old_dentry_dir);
841 ceph_pagelist_release(req->r_pagelist);
842 put_request_session(req);
843 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
844 WARN_ON_ONCE(!list_empty(&req->r_wait));
845 kmem_cache_free(ceph_mds_request_cachep, req);
848 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
851 * lookup session, bump ref if found.
853 * called under mdsc->mutex.
855 static struct ceph_mds_request *
856 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
858 struct ceph_mds_request *req;
860 req = lookup_request(&mdsc->request_tree, tid);
862 ceph_mdsc_get_request(req);
868 * Register an in-flight request, and assign a tid. Link to directory
869 * are modifying (if any).
871 * Called under mdsc->mutex.
873 static void __register_request(struct ceph_mds_client *mdsc,
874 struct ceph_mds_request *req,
879 req->r_tid = ++mdsc->last_tid;
880 if (req->r_num_caps) {
881 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
884 pr_err("__register_request %p "
885 "failed to reserve caps: %d\n", req, ret);
886 /* set req->r_err to fail early from __do_request */
891 dout("__register_request %p tid %lld\n", req, req->r_tid);
892 ceph_mdsc_get_request(req);
893 insert_request(&mdsc->request_tree, req);
895 req->r_uid = current_fsuid();
896 req->r_gid = current_fsgid();
898 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
899 mdsc->oldest_tid = req->r_tid;
902 struct ceph_inode_info *ci = ceph_inode(dir);
905 req->r_unsafe_dir = dir;
906 spin_lock(&ci->i_unsafe_lock);
907 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
908 spin_unlock(&ci->i_unsafe_lock);
912 static void __unregister_request(struct ceph_mds_client *mdsc,
913 struct ceph_mds_request *req)
915 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
917 /* Never leave an unregistered request on an unsafe list! */
918 list_del_init(&req->r_unsafe_item);
920 if (req->r_tid == mdsc->oldest_tid) {
921 struct rb_node *p = rb_next(&req->r_node);
922 mdsc->oldest_tid = 0;
924 struct ceph_mds_request *next_req =
925 rb_entry(p, struct ceph_mds_request, r_node);
926 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
927 mdsc->oldest_tid = next_req->r_tid;
934 erase_request(&mdsc->request_tree, req);
936 if (req->r_unsafe_dir) {
937 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
938 spin_lock(&ci->i_unsafe_lock);
939 list_del_init(&req->r_unsafe_dir_item);
940 spin_unlock(&ci->i_unsafe_lock);
942 if (req->r_target_inode &&
943 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
944 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
945 spin_lock(&ci->i_unsafe_lock);
946 list_del_init(&req->r_unsafe_target_item);
947 spin_unlock(&ci->i_unsafe_lock);
950 if (req->r_unsafe_dir) {
951 /* avoid calling iput_final() in mds dispatch threads */
952 ceph_async_iput(req->r_unsafe_dir);
953 req->r_unsafe_dir = NULL;
956 complete_all(&req->r_safe_completion);
958 ceph_mdsc_put_request(req);
962 * Walk back up the dentry tree until we hit a dentry representing a
963 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
964 * when calling this) to ensure that the objects won't disappear while we're
965 * working with them. Once we hit a candidate dentry, we attempt to take a
966 * reference to it, and return that as the result.
968 static struct inode *get_nonsnap_parent(struct dentry *dentry)
970 struct inode *inode = NULL;
972 while (dentry && !IS_ROOT(dentry)) {
973 inode = d_inode_rcu(dentry);
974 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
976 dentry = dentry->d_parent;
979 inode = igrab(inode);
984 * Choose mds to send request to next. If there is a hint set in the
985 * request (e.g., due to a prior forward hint from the mds), use that.
986 * Otherwise, consult frag tree and/or caps to identify the
987 * appropriate mds. If all else fails, choose randomly.
989 * Called under mdsc->mutex.
991 static int __choose_mds(struct ceph_mds_client *mdsc,
992 struct ceph_mds_request *req,
996 struct ceph_inode_info *ci;
997 struct ceph_cap *cap;
998 int mode = req->r_direct_mode;
1000 u32 hash = req->r_direct_hash;
1001 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1007 * is there a specific mds we should try? ignore hint if we have
1008 * no session and the mds is not up (active or recovering).
1010 if (req->r_resend_mds >= 0 &&
1011 (__have_session(mdsc, req->r_resend_mds) ||
1012 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1013 dout("%s using resend_mds mds%d\n", __func__,
1015 return req->r_resend_mds;
1018 if (mode == USE_RANDOM_MDS)
1023 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1024 inode = req->r_inode;
1027 /* req->r_dentry is non-null for LSSNAP request */
1029 inode = get_nonsnap_parent(req->r_dentry);
1031 dout("%s using snapdir's parent %p\n", __func__, inode);
1033 } else if (req->r_dentry) {
1034 /* ignore race with rename; old or new d_parent is okay */
1035 struct dentry *parent;
1039 parent = READ_ONCE(req->r_dentry->d_parent);
1040 dir = req->r_parent ? : d_inode_rcu(parent);
1042 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1043 /* not this fs or parent went negative */
1044 inode = d_inode(req->r_dentry);
1047 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1048 /* direct snapped/virtual snapdir requests
1049 * based on parent dir inode */
1050 inode = get_nonsnap_parent(parent);
1051 dout("%s using nonsnap parent %p\n", __func__, inode);
1054 inode = d_inode(req->r_dentry);
1055 if (!inode || mode == USE_AUTH_MDS) {
1058 hash = ceph_dentry_hash(dir, req->r_dentry);
1067 dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1071 ci = ceph_inode(inode);
1073 if (is_hash && S_ISDIR(inode->i_mode)) {
1074 struct ceph_inode_frag frag;
1077 ceph_choose_frag(ci, hash, &frag, &found);
1079 if (mode == USE_ANY_MDS && frag.ndist > 0) {
1082 /* choose a random replica */
1083 get_random_bytes(&r, 1);
1086 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1087 __func__, inode, ceph_vinop(inode),
1088 frag.frag, mds, (int)r, frag.ndist);
1089 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1090 CEPH_MDS_STATE_ACTIVE &&
1091 !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1095 /* since this file/dir wasn't known to be
1096 * replicated, then we want to look for the
1097 * authoritative mds. */
1098 if (frag.mds >= 0) {
1099 /* choose auth mds */
1101 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1102 __func__, inode, ceph_vinop(inode),
1104 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1105 CEPH_MDS_STATE_ACTIVE) {
1106 if (mode == USE_ANY_MDS &&
1107 !ceph_mdsmap_is_laggy(mdsc->mdsmap,
1112 mode = USE_AUTH_MDS;
1116 spin_lock(&ci->i_ceph_lock);
1118 if (mode == USE_AUTH_MDS)
1119 cap = ci->i_auth_cap;
1120 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1121 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1123 spin_unlock(&ci->i_ceph_lock);
1124 ceph_async_iput(inode);
1127 mds = cap->session->s_mds;
1128 dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1129 inode, ceph_vinop(inode), mds,
1130 cap == ci->i_auth_cap ? "auth " : "", cap);
1131 spin_unlock(&ci->i_ceph_lock);
1133 /* avoid calling iput_final() while holding mdsc->mutex or
1134 * in mds dispatch threads */
1135 ceph_async_iput(inode);
1142 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1143 dout("%s chose random mds%d\n", __func__, mds);
1151 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1153 struct ceph_msg *msg;
1154 struct ceph_mds_session_head *h;
1156 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1159 pr_err("create_session_msg ENOMEM creating msg\n");
1162 h = msg->front.iov_base;
1163 h->op = cpu_to_le32(op);
1164 h->seq = cpu_to_le64(seq);
1169 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1170 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1171 static void encode_supported_features(void **p, void *end)
1173 static const size_t count = ARRAY_SIZE(feature_bits);
1177 size_t size = FEATURE_BYTES(count);
1179 BUG_ON(*p + 4 + size > end);
1180 ceph_encode_32(p, size);
1181 memset(*p, 0, size);
1182 for (i = 0; i < count; i++)
1183 ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1186 BUG_ON(*p + 4 > end);
1187 ceph_encode_32(p, 0);
1192 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1193 * to include additional client metadata fields.
1195 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1197 struct ceph_msg *msg;
1198 struct ceph_mds_session_head *h;
1200 int extra_bytes = 0;
1201 int metadata_key_count = 0;
1202 struct ceph_options *opt = mdsc->fsc->client->options;
1203 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1207 const char* metadata[][2] = {
1208 {"hostname", mdsc->nodename},
1209 {"kernel_version", init_utsname()->release},
1210 {"entity_id", opt->name ? : ""},
1211 {"root", fsopt->server_path ? : "/"},
1215 /* Calculate serialized length of metadata */
1216 extra_bytes = 4; /* map length */
1217 for (i = 0; metadata[i][0]; ++i) {
1218 extra_bytes += 8 + strlen(metadata[i][0]) +
1219 strlen(metadata[i][1]);
1220 metadata_key_count++;
1223 /* supported feature */
1225 count = ARRAY_SIZE(feature_bits);
1227 size = FEATURE_BYTES(count);
1228 extra_bytes += 4 + size;
1230 /* Allocate the message */
1231 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1234 pr_err("create_session_msg ENOMEM creating msg\n");
1237 p = msg->front.iov_base;
1238 end = p + msg->front.iov_len;
1241 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1242 h->seq = cpu_to_le64(seq);
1245 * Serialize client metadata into waiting buffer space, using
1246 * the format that userspace expects for map<string, string>
1248 * ClientSession messages with metadata are v3
1250 msg->hdr.version = cpu_to_le16(3);
1251 msg->hdr.compat_version = cpu_to_le16(1);
1253 /* The write pointer, following the session_head structure */
1256 /* Number of entries in the map */
1257 ceph_encode_32(&p, metadata_key_count);
1259 /* Two length-prefixed strings for each entry in the map */
1260 for (i = 0; metadata[i][0]; ++i) {
1261 size_t const key_len = strlen(metadata[i][0]);
1262 size_t const val_len = strlen(metadata[i][1]);
1264 ceph_encode_32(&p, key_len);
1265 memcpy(p, metadata[i][0], key_len);
1267 ceph_encode_32(&p, val_len);
1268 memcpy(p, metadata[i][1], val_len);
1272 encode_supported_features(&p, end);
1273 msg->front.iov_len = p - msg->front.iov_base;
1274 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1280 * send session open request.
1282 * called under mdsc->mutex
1284 static int __open_session(struct ceph_mds_client *mdsc,
1285 struct ceph_mds_session *session)
1287 struct ceph_msg *msg;
1289 int mds = session->s_mds;
1291 /* wait for mds to go active? */
1292 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1293 dout("open_session to mds%d (%s)\n", mds,
1294 ceph_mds_state_name(mstate));
1295 session->s_state = CEPH_MDS_SESSION_OPENING;
1296 session->s_renew_requested = jiffies;
1298 /* send connect message */
1299 msg = create_session_open_msg(mdsc, session->s_seq);
1302 ceph_con_send(&session->s_con, msg);
1307 * open sessions for any export targets for the given mds
1309 * called under mdsc->mutex
1311 static struct ceph_mds_session *
1312 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1314 struct ceph_mds_session *session;
1316 session = __ceph_lookup_mds_session(mdsc, target);
1318 session = register_session(mdsc, target);
1319 if (IS_ERR(session))
1322 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1323 session->s_state == CEPH_MDS_SESSION_CLOSING)
1324 __open_session(mdsc, session);
1329 struct ceph_mds_session *
1330 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1332 struct ceph_mds_session *session;
1334 dout("open_export_target_session to mds%d\n", target);
1336 mutex_lock(&mdsc->mutex);
1337 session = __open_export_target_session(mdsc, target);
1338 mutex_unlock(&mdsc->mutex);
1343 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1344 struct ceph_mds_session *session)
1346 struct ceph_mds_info *mi;
1347 struct ceph_mds_session *ts;
1348 int i, mds = session->s_mds;
1350 if (mds >= mdsc->mdsmap->possible_max_rank)
1353 mi = &mdsc->mdsmap->m_info[mds];
1354 dout("open_export_target_sessions for mds%d (%d targets)\n",
1355 session->s_mds, mi->num_export_targets);
1357 for (i = 0; i < mi->num_export_targets; i++) {
1358 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1360 ceph_put_mds_session(ts);
1364 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1365 struct ceph_mds_session *session)
1367 mutex_lock(&mdsc->mutex);
1368 __open_export_target_sessions(mdsc, session);
1369 mutex_unlock(&mdsc->mutex);
1376 static void detach_cap_releases(struct ceph_mds_session *session,
1377 struct list_head *target)
1379 lockdep_assert_held(&session->s_cap_lock);
1381 list_splice_init(&session->s_cap_releases, target);
1382 session->s_num_cap_releases = 0;
1383 dout("dispose_cap_releases mds%d\n", session->s_mds);
1386 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1387 struct list_head *dispose)
1389 while (!list_empty(dispose)) {
1390 struct ceph_cap *cap;
1391 /* zero out the in-progress message */
1392 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1393 list_del(&cap->session_caps);
1394 ceph_put_cap(mdsc, cap);
1398 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1399 struct ceph_mds_session *session)
1401 struct ceph_mds_request *req;
1403 struct ceph_inode_info *ci;
1405 dout("cleanup_session_requests mds%d\n", session->s_mds);
1406 mutex_lock(&mdsc->mutex);
1407 while (!list_empty(&session->s_unsafe)) {
1408 req = list_first_entry(&session->s_unsafe,
1409 struct ceph_mds_request, r_unsafe_item);
1410 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1412 if (req->r_target_inode) {
1413 /* dropping unsafe change of inode's attributes */
1414 ci = ceph_inode(req->r_target_inode);
1415 errseq_set(&ci->i_meta_err, -EIO);
1417 if (req->r_unsafe_dir) {
1418 /* dropping unsafe directory operation */
1419 ci = ceph_inode(req->r_unsafe_dir);
1420 errseq_set(&ci->i_meta_err, -EIO);
1422 __unregister_request(mdsc, req);
1424 /* zero r_attempts, so kick_requests() will re-send requests */
1425 p = rb_first(&mdsc->request_tree);
1427 req = rb_entry(p, struct ceph_mds_request, r_node);
1429 if (req->r_session &&
1430 req->r_session->s_mds == session->s_mds)
1431 req->r_attempts = 0;
1433 mutex_unlock(&mdsc->mutex);
1437 * Helper to safely iterate over all caps associated with a session, with
1438 * special care taken to handle a racing __ceph_remove_cap().
1440 * Caller must hold session s_mutex.
1442 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1443 int (*cb)(struct inode *, struct ceph_cap *,
1446 struct list_head *p;
1447 struct ceph_cap *cap;
1448 struct inode *inode, *last_inode = NULL;
1449 struct ceph_cap *old_cap = NULL;
1452 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1453 spin_lock(&session->s_cap_lock);
1454 p = session->s_caps.next;
1455 while (p != &session->s_caps) {
1456 cap = list_entry(p, struct ceph_cap, session_caps);
1457 inode = igrab(&cap->ci->vfs_inode);
1462 session->s_cap_iterator = cap;
1463 spin_unlock(&session->s_cap_lock);
1466 /* avoid calling iput_final() while holding
1467 * s_mutex or in mds dispatch threads */
1468 ceph_async_iput(last_inode);
1472 ceph_put_cap(session->s_mdsc, old_cap);
1476 ret = cb(inode, cap, arg);
1479 spin_lock(&session->s_cap_lock);
1482 dout("iterate_session_caps finishing cap %p removal\n",
1484 BUG_ON(cap->session != session);
1485 cap->session = NULL;
1486 list_del_init(&cap->session_caps);
1487 session->s_nr_caps--;
1488 if (cap->queue_release)
1489 __ceph_queue_cap_release(session, cap);
1491 old_cap = cap; /* put_cap it w/o locks held */
1498 session->s_cap_iterator = NULL;
1499 spin_unlock(&session->s_cap_lock);
1501 ceph_async_iput(last_inode);
1503 ceph_put_cap(session->s_mdsc, old_cap);
1508 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1511 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1512 struct ceph_inode_info *ci = ceph_inode(inode);
1513 LIST_HEAD(to_remove);
1514 bool dirty_dropped = false;
1515 bool invalidate = false;
1517 dout("removing cap %p, ci is %p, inode is %p\n",
1518 cap, ci, &ci->vfs_inode);
1519 spin_lock(&ci->i_ceph_lock);
1520 __ceph_remove_cap(cap, false);
1521 if (!ci->i_auth_cap) {
1522 struct ceph_cap_flush *cf;
1523 struct ceph_mds_client *mdsc = fsc->mdsc;
1525 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1526 if (inode->i_data.nrpages > 0)
1528 if (ci->i_wrbuffer_ref > 0)
1529 mapping_set_error(&inode->i_data, -EIO);
1532 while (!list_empty(&ci->i_cap_flush_list)) {
1533 cf = list_first_entry(&ci->i_cap_flush_list,
1534 struct ceph_cap_flush, i_list);
1535 list_move(&cf->i_list, &to_remove);
1538 spin_lock(&mdsc->cap_dirty_lock);
1540 list_for_each_entry(cf, &to_remove, i_list)
1541 list_del(&cf->g_list);
1543 if (!list_empty(&ci->i_dirty_item)) {
1544 pr_warn_ratelimited(
1545 " dropping dirty %s state for %p %lld\n",
1546 ceph_cap_string(ci->i_dirty_caps),
1547 inode, ceph_ino(inode));
1548 ci->i_dirty_caps = 0;
1549 list_del_init(&ci->i_dirty_item);
1550 dirty_dropped = true;
1552 if (!list_empty(&ci->i_flushing_item)) {
1553 pr_warn_ratelimited(
1554 " dropping dirty+flushing %s state for %p %lld\n",
1555 ceph_cap_string(ci->i_flushing_caps),
1556 inode, ceph_ino(inode));
1557 ci->i_flushing_caps = 0;
1558 list_del_init(&ci->i_flushing_item);
1559 mdsc->num_cap_flushing--;
1560 dirty_dropped = true;
1562 spin_unlock(&mdsc->cap_dirty_lock);
1564 if (dirty_dropped) {
1565 errseq_set(&ci->i_meta_err, -EIO);
1567 if (ci->i_wrbuffer_ref_head == 0 &&
1568 ci->i_wr_ref == 0 &&
1569 ci->i_dirty_caps == 0 &&
1570 ci->i_flushing_caps == 0) {
1571 ceph_put_snap_context(ci->i_head_snapc);
1572 ci->i_head_snapc = NULL;
1576 if (atomic_read(&ci->i_filelock_ref) > 0) {
1577 /* make further file lock syscall return -EIO */
1578 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1579 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1580 inode, ceph_ino(inode));
1583 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1584 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1585 ci->i_prealloc_cap_flush = NULL;
1588 spin_unlock(&ci->i_ceph_lock);
1589 while (!list_empty(&to_remove)) {
1590 struct ceph_cap_flush *cf;
1591 cf = list_first_entry(&to_remove,
1592 struct ceph_cap_flush, i_list);
1593 list_del(&cf->i_list);
1594 ceph_free_cap_flush(cf);
1597 wake_up_all(&ci->i_cap_wq);
1599 ceph_queue_invalidate(inode);
1606 * caller must hold session s_mutex
1608 static void remove_session_caps(struct ceph_mds_session *session)
1610 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1611 struct super_block *sb = fsc->sb;
1614 dout("remove_session_caps on %p\n", session);
1615 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1617 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1619 spin_lock(&session->s_cap_lock);
1620 if (session->s_nr_caps > 0) {
1621 struct inode *inode;
1622 struct ceph_cap *cap, *prev = NULL;
1623 struct ceph_vino vino;
1625 * iterate_session_caps() skips inodes that are being
1626 * deleted, we need to wait until deletions are complete.
1627 * __wait_on_freeing_inode() is designed for the job,
1628 * but it is not exported, so use lookup inode function
1631 while (!list_empty(&session->s_caps)) {
1632 cap = list_entry(session->s_caps.next,
1633 struct ceph_cap, session_caps);
1637 vino = cap->ci->i_vino;
1638 spin_unlock(&session->s_cap_lock);
1640 inode = ceph_find_inode(sb, vino);
1641 /* avoid calling iput_final() while holding s_mutex */
1642 ceph_async_iput(inode);
1644 spin_lock(&session->s_cap_lock);
1648 // drop cap expires and unlock s_cap_lock
1649 detach_cap_releases(session, &dispose);
1651 BUG_ON(session->s_nr_caps > 0);
1652 BUG_ON(!list_empty(&session->s_cap_flushing));
1653 spin_unlock(&session->s_cap_lock);
1654 dispose_cap_releases(session->s_mdsc, &dispose);
1664 * wake up any threads waiting on this session's caps. if the cap is
1665 * old (didn't get renewed on the client reconnect), remove it now.
1667 * caller must hold s_mutex.
1669 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1672 struct ceph_inode_info *ci = ceph_inode(inode);
1673 unsigned long ev = (unsigned long)arg;
1675 if (ev == RECONNECT) {
1676 spin_lock(&ci->i_ceph_lock);
1677 ci->i_wanted_max_size = 0;
1678 ci->i_requested_max_size = 0;
1679 spin_unlock(&ci->i_ceph_lock);
1680 } else if (ev == RENEWCAPS) {
1681 if (cap->cap_gen < cap->session->s_cap_gen) {
1682 /* mds did not re-issue stale cap */
1683 spin_lock(&ci->i_ceph_lock);
1684 cap->issued = cap->implemented = CEPH_CAP_PIN;
1685 spin_unlock(&ci->i_ceph_lock);
1687 } else if (ev == FORCE_RO) {
1689 wake_up_all(&ci->i_cap_wq);
1693 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1695 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1696 ceph_iterate_session_caps(session, wake_up_session_cb,
1697 (void *)(unsigned long)ev);
1701 * Send periodic message to MDS renewing all currently held caps. The
1702 * ack will reset the expiration for all caps from this session.
1704 * caller holds s_mutex
1706 static int send_renew_caps(struct ceph_mds_client *mdsc,
1707 struct ceph_mds_session *session)
1709 struct ceph_msg *msg;
1712 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1713 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1714 pr_info("mds%d caps stale\n", session->s_mds);
1715 session->s_renew_requested = jiffies;
1717 /* do not try to renew caps until a recovering mds has reconnected
1718 * with its clients. */
1719 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1720 if (state < CEPH_MDS_STATE_RECONNECT) {
1721 dout("send_renew_caps ignoring mds%d (%s)\n",
1722 session->s_mds, ceph_mds_state_name(state));
1726 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1727 ceph_mds_state_name(state));
1728 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1729 ++session->s_renew_seq);
1732 ceph_con_send(&session->s_con, msg);
1736 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1737 struct ceph_mds_session *session, u64 seq)
1739 struct ceph_msg *msg;
1741 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1742 session->s_mds, ceph_session_state_name(session->s_state), seq);
1743 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1746 ceph_con_send(&session->s_con, msg);
1752 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1754 * Called under session->s_mutex
1756 static void renewed_caps(struct ceph_mds_client *mdsc,
1757 struct ceph_mds_session *session, int is_renew)
1762 spin_lock(&session->s_cap_lock);
1763 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1765 session->s_cap_ttl = session->s_renew_requested +
1766 mdsc->mdsmap->m_session_timeout*HZ;
1769 if (time_before(jiffies, session->s_cap_ttl)) {
1770 pr_info("mds%d caps renewed\n", session->s_mds);
1773 pr_info("mds%d caps still stale\n", session->s_mds);
1776 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1777 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1778 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1779 spin_unlock(&session->s_cap_lock);
1782 wake_up_session_caps(session, RENEWCAPS);
1786 * send a session close request
1788 static int request_close_session(struct ceph_mds_client *mdsc,
1789 struct ceph_mds_session *session)
1791 struct ceph_msg *msg;
1793 dout("request_close_session mds%d state %s seq %lld\n",
1794 session->s_mds, ceph_session_state_name(session->s_state),
1796 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1799 ceph_con_send(&session->s_con, msg);
1804 * Called with s_mutex held.
1806 static int __close_session(struct ceph_mds_client *mdsc,
1807 struct ceph_mds_session *session)
1809 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1811 session->s_state = CEPH_MDS_SESSION_CLOSING;
1812 return request_close_session(mdsc, session);
1815 static bool drop_negative_children(struct dentry *dentry)
1817 struct dentry *child;
1818 bool all_negative = true;
1820 if (!d_is_dir(dentry))
1823 spin_lock(&dentry->d_lock);
1824 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1825 if (d_really_is_positive(child)) {
1826 all_negative = false;
1830 spin_unlock(&dentry->d_lock);
1833 shrink_dcache_parent(dentry);
1835 return all_negative;
1839 * Trim old(er) caps.
1841 * Because we can't cache an inode without one or more caps, we do
1842 * this indirectly: if a cap is unused, we prune its aliases, at which
1843 * point the inode will hopefully get dropped to.
1845 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1846 * memory pressure from the MDS, though, so it needn't be perfect.
1848 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1850 int *remaining = arg;
1851 struct ceph_inode_info *ci = ceph_inode(inode);
1852 int used, wanted, oissued, mine;
1854 if (*remaining <= 0)
1857 spin_lock(&ci->i_ceph_lock);
1858 mine = cap->issued | cap->implemented;
1859 used = __ceph_caps_used(ci);
1860 wanted = __ceph_caps_file_wanted(ci);
1861 oissued = __ceph_caps_issued_other(ci, cap);
1863 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1864 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1865 ceph_cap_string(used), ceph_cap_string(wanted));
1866 if (cap == ci->i_auth_cap) {
1867 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1868 !list_empty(&ci->i_cap_snaps))
1870 if ((used | wanted) & CEPH_CAP_ANY_WR)
1872 /* Note: it's possible that i_filelock_ref becomes non-zero
1873 * after dropping auth caps. It doesn't hurt because reply
1874 * of lock mds request will re-add auth caps. */
1875 if (atomic_read(&ci->i_filelock_ref) > 0)
1878 /* The inode has cached pages, but it's no longer used.
1879 * we can safely drop it */
1880 if (S_ISREG(inode->i_mode) &&
1881 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1882 !(oissued & CEPH_CAP_FILE_CACHE)) {
1886 if ((used | wanted) & ~oissued & mine)
1887 goto out; /* we need these caps */
1890 /* we aren't the only cap.. just remove us */
1891 __ceph_remove_cap(cap, true);
1894 struct dentry *dentry;
1895 /* try dropping referring dentries */
1896 spin_unlock(&ci->i_ceph_lock);
1897 dentry = d_find_any_alias(inode);
1898 if (dentry && drop_negative_children(dentry)) {
1901 d_prune_aliases(inode);
1902 count = atomic_read(&inode->i_count);
1905 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1914 spin_unlock(&ci->i_ceph_lock);
1919 * Trim session cap count down to some max number.
1921 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1922 struct ceph_mds_session *session,
1925 int trim_caps = session->s_nr_caps - max_caps;
1927 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1928 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1929 if (trim_caps > 0) {
1930 int remaining = trim_caps;
1932 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1933 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1934 session->s_mds, session->s_nr_caps, max_caps,
1935 trim_caps - remaining);
1938 ceph_flush_cap_releases(mdsc, session);
1942 static int check_caps_flush(struct ceph_mds_client *mdsc,
1947 spin_lock(&mdsc->cap_dirty_lock);
1948 if (!list_empty(&mdsc->cap_flush_list)) {
1949 struct ceph_cap_flush *cf =
1950 list_first_entry(&mdsc->cap_flush_list,
1951 struct ceph_cap_flush, g_list);
1952 if (cf->tid <= want_flush_tid) {
1953 dout("check_caps_flush still flushing tid "
1954 "%llu <= %llu\n", cf->tid, want_flush_tid);
1958 spin_unlock(&mdsc->cap_dirty_lock);
1963 * flush all dirty inode data to disk.
1965 * returns true if we've flushed through want_flush_tid
1967 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1970 dout("check_caps_flush want %llu\n", want_flush_tid);
1972 wait_event(mdsc->cap_flushing_wq,
1973 check_caps_flush(mdsc, want_flush_tid));
1975 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1979 * called under s_mutex
1981 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1982 struct ceph_mds_session *session)
1984 struct ceph_msg *msg = NULL;
1985 struct ceph_mds_cap_release *head;
1986 struct ceph_mds_cap_item *item;
1987 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1988 struct ceph_cap *cap;
1989 LIST_HEAD(tmp_list);
1990 int num_cap_releases;
1991 __le32 barrier, *cap_barrier;
1993 down_read(&osdc->lock);
1994 barrier = cpu_to_le32(osdc->epoch_barrier);
1995 up_read(&osdc->lock);
1997 spin_lock(&session->s_cap_lock);
1999 list_splice_init(&session->s_cap_releases, &tmp_list);
2000 num_cap_releases = session->s_num_cap_releases;
2001 session->s_num_cap_releases = 0;
2002 spin_unlock(&session->s_cap_lock);
2004 while (!list_empty(&tmp_list)) {
2006 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2007 PAGE_SIZE, GFP_NOFS, false);
2010 head = msg->front.iov_base;
2011 head->num = cpu_to_le32(0);
2012 msg->front.iov_len = sizeof(*head);
2014 msg->hdr.version = cpu_to_le16(2);
2015 msg->hdr.compat_version = cpu_to_le16(1);
2018 cap = list_first_entry(&tmp_list, struct ceph_cap,
2020 list_del(&cap->session_caps);
2023 head = msg->front.iov_base;
2024 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2026 item = msg->front.iov_base + msg->front.iov_len;
2027 item->ino = cpu_to_le64(cap->cap_ino);
2028 item->cap_id = cpu_to_le64(cap->cap_id);
2029 item->migrate_seq = cpu_to_le32(cap->mseq);
2030 item->seq = cpu_to_le32(cap->issue_seq);
2031 msg->front.iov_len += sizeof(*item);
2033 ceph_put_cap(mdsc, cap);
2035 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2036 // Append cap_barrier field
2037 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2038 *cap_barrier = barrier;
2039 msg->front.iov_len += sizeof(*cap_barrier);
2041 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2042 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2043 ceph_con_send(&session->s_con, msg);
2048 BUG_ON(num_cap_releases != 0);
2050 spin_lock(&session->s_cap_lock);
2051 if (!list_empty(&session->s_cap_releases))
2053 spin_unlock(&session->s_cap_lock);
2056 // Append cap_barrier field
2057 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2058 *cap_barrier = barrier;
2059 msg->front.iov_len += sizeof(*cap_barrier);
2061 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2062 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2063 ceph_con_send(&session->s_con, msg);
2067 pr_err("send_cap_releases mds%d, failed to allocate message\n",
2069 spin_lock(&session->s_cap_lock);
2070 list_splice(&tmp_list, &session->s_cap_releases);
2071 session->s_num_cap_releases += num_cap_releases;
2072 spin_unlock(&session->s_cap_lock);
2075 static void ceph_cap_release_work(struct work_struct *work)
2077 struct ceph_mds_session *session =
2078 container_of(work, struct ceph_mds_session, s_cap_release_work);
2080 mutex_lock(&session->s_mutex);
2081 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2082 session->s_state == CEPH_MDS_SESSION_HUNG)
2083 ceph_send_cap_releases(session->s_mdsc, session);
2084 mutex_unlock(&session->s_mutex);
2085 ceph_put_mds_session(session);
2088 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2089 struct ceph_mds_session *session)
2094 ceph_get_mds_session(session);
2095 if (queue_work(mdsc->fsc->cap_wq,
2096 &session->s_cap_release_work)) {
2097 dout("cap release work queued\n");
2099 ceph_put_mds_session(session);
2100 dout("failed to queue cap release work\n");
2105 * caller holds session->s_cap_lock
2107 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2108 struct ceph_cap *cap)
2110 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2111 session->s_num_cap_releases++;
2113 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2114 ceph_flush_cap_releases(session->s_mdsc, session);
2117 static void ceph_cap_reclaim_work(struct work_struct *work)
2119 struct ceph_mds_client *mdsc =
2120 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2121 int ret = ceph_trim_dentries(mdsc);
2123 ceph_queue_cap_reclaim_work(mdsc);
2126 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2131 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2132 dout("caps reclaim work queued\n");
2134 dout("failed to queue caps release work\n");
2138 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2143 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2144 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2145 atomic_set(&mdsc->cap_reclaim_pending, 0);
2146 ceph_queue_cap_reclaim_work(mdsc);
2154 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2157 struct ceph_inode_info *ci = ceph_inode(dir);
2158 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2159 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2160 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2161 unsigned int num_entries;
2164 spin_lock(&ci->i_ceph_lock);
2165 num_entries = ci->i_files + ci->i_subdirs;
2166 spin_unlock(&ci->i_ceph_lock);
2167 num_entries = max(num_entries, 1U);
2168 num_entries = min(num_entries, opt->max_readdir);
2170 order = get_order(size * num_entries);
2171 while (order >= 0) {
2172 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2175 if (rinfo->dir_entries)
2179 if (!rinfo->dir_entries)
2182 num_entries = (PAGE_SIZE << order) / size;
2183 num_entries = min(num_entries, opt->max_readdir);
2185 rinfo->dir_buf_size = PAGE_SIZE << order;
2186 req->r_num_caps = num_entries + 1;
2187 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2188 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2193 * Create an mds request.
2195 struct ceph_mds_request *
2196 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2198 struct ceph_mds_request *req;
2200 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2202 return ERR_PTR(-ENOMEM);
2204 mutex_init(&req->r_fill_mutex);
2206 req->r_started = jiffies;
2207 req->r_start_latency = ktime_get();
2208 req->r_resend_mds = -1;
2209 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2210 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2212 kref_init(&req->r_kref);
2213 RB_CLEAR_NODE(&req->r_node);
2214 INIT_LIST_HEAD(&req->r_wait);
2215 init_completion(&req->r_completion);
2216 init_completion(&req->r_safe_completion);
2217 INIT_LIST_HEAD(&req->r_unsafe_item);
2219 ktime_get_coarse_real_ts64(&req->r_stamp);
2222 req->r_direct_mode = mode;
2227 * return oldest (lowest) request, tid in request tree, 0 if none.
2229 * called under mdsc->mutex.
2231 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2233 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2235 return rb_entry(rb_first(&mdsc->request_tree),
2236 struct ceph_mds_request, r_node);
2239 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2241 return mdsc->oldest_tid;
2245 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2246 * on build_path_from_dentry in fs/cifs/dir.c.
2248 * If @stop_on_nosnap, generate path relative to the first non-snapped
2251 * Encode hidden .snap dirs as a double /, i.e.
2252 * foo/.snap/bar -> foo//bar
2254 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2257 struct dentry *temp;
2264 return ERR_PTR(-EINVAL);
2268 return ERR_PTR(-ENOMEM);
2273 seq = read_seqbegin(&rename_lock);
2277 struct inode *inode;
2279 spin_lock(&temp->d_lock);
2280 inode = d_inode(temp);
2281 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2282 dout("build_path path+%d: %p SNAPDIR\n",
2284 } else if (stop_on_nosnap && inode && dentry != temp &&
2285 ceph_snap(inode) == CEPH_NOSNAP) {
2286 spin_unlock(&temp->d_lock);
2287 pos++; /* get rid of any prepended '/' */
2290 pos -= temp->d_name.len;
2292 spin_unlock(&temp->d_lock);
2295 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2297 spin_unlock(&temp->d_lock);
2298 temp = READ_ONCE(temp->d_parent);
2300 /* Are we at the root? */
2304 /* Are we out of buffer? */
2310 base = ceph_ino(d_inode(temp));
2313 if (read_seqretry(&rename_lock, seq))
2318 * A rename didn't occur, but somehow we didn't end up where
2319 * we thought we would. Throw a warning and try again.
2321 pr_warn("build_path did not end path lookup where "
2322 "expected, pos is %d\n", pos);
2327 *plen = PATH_MAX - 1 - pos;
2328 dout("build_path on %p %d built %llx '%.*s'\n",
2329 dentry, d_count(dentry), base, *plen, path + pos);
2333 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2334 const char **ppath, int *ppathlen, u64 *pino,
2335 bool *pfreepath, bool parent_locked)
2341 dir = d_inode_rcu(dentry->d_parent);
2342 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2343 *pino = ceph_ino(dir);
2345 *ppath = dentry->d_name.name;
2346 *ppathlen = dentry->d_name.len;
2350 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2352 return PTR_ERR(path);
2358 static int build_inode_path(struct inode *inode,
2359 const char **ppath, int *ppathlen, u64 *pino,
2362 struct dentry *dentry;
2365 if (ceph_snap(inode) == CEPH_NOSNAP) {
2366 *pino = ceph_ino(inode);
2370 dentry = d_find_alias(inode);
2371 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2374 return PTR_ERR(path);
2381 * request arguments may be specified via an inode *, a dentry *, or
2382 * an explicit ino+path.
2384 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2385 struct inode *rdiri, const char *rpath,
2386 u64 rino, const char **ppath, int *pathlen,
2387 u64 *ino, bool *freepath, bool parent_locked)
2392 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2393 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2395 } else if (rdentry) {
2396 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2397 freepath, parent_locked);
2398 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2400 } else if (rpath || rino) {
2403 *pathlen = rpath ? strlen(rpath) : 0;
2404 dout(" path %.*s\n", *pathlen, rpath);
2411 * called under mdsc->mutex
2413 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2414 struct ceph_mds_request *req,
2415 int mds, bool drop_cap_releases)
2417 struct ceph_msg *msg;
2418 struct ceph_mds_request_head *head;
2419 const char *path1 = NULL;
2420 const char *path2 = NULL;
2421 u64 ino1 = 0, ino2 = 0;
2422 int pathlen1 = 0, pathlen2 = 0;
2423 bool freepath1 = false, freepath2 = false;
2429 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2430 req->r_parent, req->r_path1, req->r_ino1.ino,
2431 &path1, &pathlen1, &ino1, &freepath1,
2432 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2433 &req->r_req_flags));
2439 /* If r_old_dentry is set, then assume that its parent is locked */
2440 ret = set_request_path_attr(NULL, req->r_old_dentry,
2441 req->r_old_dentry_dir,
2442 req->r_path2, req->r_ino2.ino,
2443 &path2, &pathlen2, &ino2, &freepath2, true);
2449 len = sizeof(*head) +
2450 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2451 sizeof(struct ceph_timespec);
2453 /* calculate (max) length for cap releases */
2454 len += sizeof(struct ceph_mds_request_release) *
2455 (!!req->r_inode_drop + !!req->r_dentry_drop +
2456 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2457 if (req->r_dentry_drop)
2459 if (req->r_old_dentry_drop)
2462 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2464 msg = ERR_PTR(-ENOMEM);
2468 msg->hdr.version = cpu_to_le16(2);
2469 msg->hdr.tid = cpu_to_le64(req->r_tid);
2471 head = msg->front.iov_base;
2472 p = msg->front.iov_base + sizeof(*head);
2473 end = msg->front.iov_base + msg->front.iov_len;
2475 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2476 head->op = cpu_to_le32(req->r_op);
2477 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2478 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2479 head->ino = cpu_to_le64(req->r_deleg_ino);
2480 head->args = req->r_args;
2482 ceph_encode_filepath(&p, end, ino1, path1);
2483 ceph_encode_filepath(&p, end, ino2, path2);
2485 /* make note of release offset, in case we need to replay */
2486 req->r_request_release_offset = p - msg->front.iov_base;
2490 if (req->r_inode_drop)
2491 releases += ceph_encode_inode_release(&p,
2492 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2493 mds, req->r_inode_drop, req->r_inode_unless,
2494 req->r_op == CEPH_MDS_OP_READDIR);
2495 if (req->r_dentry_drop)
2496 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2497 req->r_parent, mds, req->r_dentry_drop,
2498 req->r_dentry_unless);
2499 if (req->r_old_dentry_drop)
2500 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2501 req->r_old_dentry_dir, mds,
2502 req->r_old_dentry_drop,
2503 req->r_old_dentry_unless);
2504 if (req->r_old_inode_drop)
2505 releases += ceph_encode_inode_release(&p,
2506 d_inode(req->r_old_dentry),
2507 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2509 if (drop_cap_releases) {
2511 p = msg->front.iov_base + req->r_request_release_offset;
2514 head->num_releases = cpu_to_le16(releases);
2518 struct ceph_timespec ts;
2519 ceph_encode_timespec64(&ts, &req->r_stamp);
2520 ceph_encode_copy(&p, &ts, sizeof(ts));
2524 msg->front.iov_len = p - msg->front.iov_base;
2525 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2527 if (req->r_pagelist) {
2528 struct ceph_pagelist *pagelist = req->r_pagelist;
2529 ceph_msg_data_add_pagelist(msg, pagelist);
2530 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2532 msg->hdr.data_len = 0;
2535 msg->hdr.data_off = cpu_to_le16(0);
2539 ceph_mdsc_free_path((char *)path2, pathlen2);
2542 ceph_mdsc_free_path((char *)path1, pathlen1);
2548 * called under mdsc->mutex if error, under no mutex if
2551 static void complete_request(struct ceph_mds_client *mdsc,
2552 struct ceph_mds_request *req)
2554 req->r_end_latency = ktime_get();
2556 if (req->r_callback)
2557 req->r_callback(mdsc, req);
2558 complete_all(&req->r_completion);
2562 * called under mdsc->mutex
2564 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2565 struct ceph_mds_request *req,
2566 int mds, bool drop_cap_releases)
2568 struct ceph_mds_request_head *rhead;
2569 struct ceph_msg *msg;
2574 struct ceph_cap *cap =
2575 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2578 req->r_sent_on_mseq = cap->mseq;
2580 req->r_sent_on_mseq = -1;
2582 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2583 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2585 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2588 * Replay. Do not regenerate message (and rebuild
2589 * paths, etc.); just use the original message.
2590 * Rebuilding paths will break for renames because
2591 * d_move mangles the src name.
2593 msg = req->r_request;
2594 rhead = msg->front.iov_base;
2596 flags = le32_to_cpu(rhead->flags);
2597 flags |= CEPH_MDS_FLAG_REPLAY;
2598 rhead->flags = cpu_to_le32(flags);
2600 if (req->r_target_inode)
2601 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2603 rhead->num_retry = req->r_attempts - 1;
2605 /* remove cap/dentry releases from message */
2606 rhead->num_releases = 0;
2609 p = msg->front.iov_base + req->r_request_release_offset;
2611 struct ceph_timespec ts;
2612 ceph_encode_timespec64(&ts, &req->r_stamp);
2613 ceph_encode_copy(&p, &ts, sizeof(ts));
2616 msg->front.iov_len = p - msg->front.iov_base;
2617 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2621 if (req->r_request) {
2622 ceph_msg_put(req->r_request);
2623 req->r_request = NULL;
2625 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2627 req->r_err = PTR_ERR(msg);
2628 return PTR_ERR(msg);
2630 req->r_request = msg;
2632 rhead = msg->front.iov_base;
2633 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2634 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2635 flags |= CEPH_MDS_FLAG_REPLAY;
2636 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2637 flags |= CEPH_MDS_FLAG_ASYNC;
2639 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2640 rhead->flags = cpu_to_le32(flags);
2641 rhead->num_fwd = req->r_num_fwd;
2642 rhead->num_retry = req->r_attempts - 1;
2644 dout(" r_parent = %p\n", req->r_parent);
2649 * called under mdsc->mutex
2651 static int __send_request(struct ceph_mds_client *mdsc,
2652 struct ceph_mds_session *session,
2653 struct ceph_mds_request *req,
2654 bool drop_cap_releases)
2658 err = __prepare_send_request(mdsc, req, session->s_mds,
2661 ceph_msg_get(req->r_request);
2662 ceph_con_send(&session->s_con, req->r_request);
2669 * send request, or put it on the appropriate wait list.
2671 static void __do_request(struct ceph_mds_client *mdsc,
2672 struct ceph_mds_request *req)
2674 struct ceph_mds_session *session = NULL;
2679 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2680 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2681 __unregister_request(mdsc, req);
2685 if (req->r_timeout &&
2686 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2687 dout("do_request timed out\n");
2691 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2692 dout("do_request forced umount\n");
2696 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2697 if (mdsc->mdsmap_err) {
2698 err = mdsc->mdsmap_err;
2699 dout("do_request mdsmap err %d\n", err);
2702 if (mdsc->mdsmap->m_epoch == 0) {
2703 dout("do_request no mdsmap, waiting for map\n");
2704 list_add(&req->r_wait, &mdsc->waiting_for_map);
2707 if (!(mdsc->fsc->mount_options->flags &
2708 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2709 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2710 err = -EHOSTUNREACH;
2715 put_request_session(req);
2717 mds = __choose_mds(mdsc, req, &random);
2719 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2720 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2724 dout("do_request no mds or not active, waiting for map\n");
2725 list_add(&req->r_wait, &mdsc->waiting_for_map);
2729 /* get, open session */
2730 session = __ceph_lookup_mds_session(mdsc, mds);
2732 session = register_session(mdsc, mds);
2733 if (IS_ERR(session)) {
2734 err = PTR_ERR(session);
2738 req->r_session = ceph_get_mds_session(session);
2740 dout("do_request mds%d session %p state %s\n", mds, session,
2741 ceph_session_state_name(session->s_state));
2742 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2743 session->s_state != CEPH_MDS_SESSION_HUNG) {
2744 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2749 * We cannot queue async requests since the caps and delegated
2750 * inodes are bound to the session. Just return -EJUKEBOX and
2751 * let the caller retry a sync request in that case.
2753 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2757 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2758 session->s_state == CEPH_MDS_SESSION_CLOSING) {
2759 __open_session(mdsc, session);
2760 /* retry the same mds later */
2762 req->r_resend_mds = mds;
2764 list_add(&req->r_wait, &session->s_waiting);
2769 req->r_resend_mds = -1; /* forget any previous mds hint */
2771 if (req->r_request_started == 0) /* note request start time */
2772 req->r_request_started = jiffies;
2774 err = __send_request(mdsc, session, req, false);
2777 ceph_put_mds_session(session);
2780 dout("__do_request early error %d\n", err);
2782 complete_request(mdsc, req);
2783 __unregister_request(mdsc, req);
2789 * called under mdsc->mutex
2791 static void __wake_requests(struct ceph_mds_client *mdsc,
2792 struct list_head *head)
2794 struct ceph_mds_request *req;
2795 LIST_HEAD(tmp_list);
2797 list_splice_init(head, &tmp_list);
2799 while (!list_empty(&tmp_list)) {
2800 req = list_entry(tmp_list.next,
2801 struct ceph_mds_request, r_wait);
2802 list_del_init(&req->r_wait);
2803 dout(" wake request %p tid %llu\n", req, req->r_tid);
2804 __do_request(mdsc, req);
2809 * Wake up threads with requests pending for @mds, so that they can
2810 * resubmit their requests to a possibly different mds.
2812 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2814 struct ceph_mds_request *req;
2815 struct rb_node *p = rb_first(&mdsc->request_tree);
2817 dout("kick_requests mds%d\n", mds);
2819 req = rb_entry(p, struct ceph_mds_request, r_node);
2821 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2823 if (req->r_attempts > 0)
2824 continue; /* only new requests */
2825 if (req->r_session &&
2826 req->r_session->s_mds == mds) {
2827 dout(" kicking tid %llu\n", req->r_tid);
2828 list_del_init(&req->r_wait);
2829 __do_request(mdsc, req);
2834 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2835 struct ceph_mds_request *req)
2839 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2841 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2842 if (req->r_parent) {
2843 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2844 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2845 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2846 spin_lock(&ci->i_ceph_lock);
2847 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2848 __ceph_touch_fmode(ci, mdsc, fmode);
2849 spin_unlock(&ci->i_ceph_lock);
2850 ihold(req->r_parent);
2852 if (req->r_old_dentry_dir)
2853 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2857 err = ceph_wait_on_async_create(req->r_inode);
2859 dout("%s: wait for async create returned: %d\n",
2865 if (!err && req->r_old_inode) {
2866 err = ceph_wait_on_async_create(req->r_old_inode);
2868 dout("%s: wait for async create returned: %d\n",
2874 dout("submit_request on %p for inode %p\n", req, dir);
2875 mutex_lock(&mdsc->mutex);
2876 __register_request(mdsc, req, dir);
2877 __do_request(mdsc, req);
2879 mutex_unlock(&mdsc->mutex);
2883 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2884 struct ceph_mds_request *req)
2889 dout("do_request waiting\n");
2890 if (!req->r_timeout && req->r_wait_for_completion) {
2891 err = req->r_wait_for_completion(mdsc, req);
2893 long timeleft = wait_for_completion_killable_timeout(
2895 ceph_timeout_jiffies(req->r_timeout));
2899 err = -ETIMEDOUT; /* timed out */
2901 err = timeleft; /* killed */
2903 dout("do_request waited, got %d\n", err);
2904 mutex_lock(&mdsc->mutex);
2906 /* only abort if we didn't race with a real reply */
2907 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2908 err = le32_to_cpu(req->r_reply_info.head->result);
2909 } else if (err < 0) {
2910 dout("aborted request %lld with %d\n", req->r_tid, err);
2913 * ensure we aren't running concurrently with
2914 * ceph_fill_trace or ceph_readdir_prepopulate, which
2915 * rely on locks (dir mutex) held by our caller.
2917 mutex_lock(&req->r_fill_mutex);
2919 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2920 mutex_unlock(&req->r_fill_mutex);
2922 if (req->r_parent &&
2923 (req->r_op & CEPH_MDS_OP_WRITE))
2924 ceph_invalidate_dir_request(req);
2929 mutex_unlock(&mdsc->mutex);
2934 * Synchrously perform an mds request. Take care of all of the
2935 * session setup, forwarding, retry details.
2937 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2939 struct ceph_mds_request *req)
2943 dout("do_request on %p\n", req);
2946 err = ceph_mdsc_submit_request(mdsc, dir, req);
2948 err = ceph_mdsc_wait_request(mdsc, req);
2949 dout("do_request %p done, result %d\n", req, err);
2954 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2955 * namespace request.
2957 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2959 struct inode *dir = req->r_parent;
2960 struct inode *old_dir = req->r_old_dentry_dir;
2962 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2964 ceph_dir_clear_complete(dir);
2966 ceph_dir_clear_complete(old_dir);
2968 ceph_invalidate_dentry_lease(req->r_dentry);
2969 if (req->r_old_dentry)
2970 ceph_invalidate_dentry_lease(req->r_old_dentry);
2976 * We take the session mutex and parse and process the reply immediately.
2977 * This preserves the logical ordering of replies, capabilities, etc., sent
2978 * by the MDS as they are applied to our local cache.
2980 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2982 struct ceph_mds_client *mdsc = session->s_mdsc;
2983 struct ceph_mds_request *req;
2984 struct ceph_mds_reply_head *head = msg->front.iov_base;
2985 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2986 struct ceph_snap_realm *realm;
2989 int mds = session->s_mds;
2991 if (msg->front.iov_len < sizeof(*head)) {
2992 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2997 /* get request, session */
2998 tid = le64_to_cpu(msg->hdr.tid);
2999 mutex_lock(&mdsc->mutex);
3000 req = lookup_get_request(mdsc, tid);
3002 dout("handle_reply on unknown tid %llu\n", tid);
3003 mutex_unlock(&mdsc->mutex);
3006 dout("handle_reply %p\n", req);
3008 /* correct session? */
3009 if (req->r_session != session) {
3010 pr_err("mdsc_handle_reply got %llu on session mds%d"
3011 " not mds%d\n", tid, session->s_mds,
3012 req->r_session ? req->r_session->s_mds : -1);
3013 mutex_unlock(&mdsc->mutex);
3018 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3019 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3020 pr_warn("got a dup %s reply on %llu from mds%d\n",
3021 head->safe ? "safe" : "unsafe", tid, mds);
3022 mutex_unlock(&mdsc->mutex);
3025 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3026 pr_warn("got unsafe after safe on %llu from mds%d\n",
3028 mutex_unlock(&mdsc->mutex);
3032 result = le32_to_cpu(head->result);
3036 * if we're not talking to the authority, send to them
3037 * if the authority has changed while we weren't looking,
3038 * send to new authority
3039 * Otherwise we just have to return an ESTALE
3041 if (result == -ESTALE) {
3042 dout("got ESTALE on request %llu\n", req->r_tid);
3043 req->r_resend_mds = -1;
3044 if (req->r_direct_mode != USE_AUTH_MDS) {
3045 dout("not using auth, setting for that now\n");
3046 req->r_direct_mode = USE_AUTH_MDS;
3047 __do_request(mdsc, req);
3048 mutex_unlock(&mdsc->mutex);
3051 int mds = __choose_mds(mdsc, req, NULL);
3052 if (mds >= 0 && mds != req->r_session->s_mds) {
3053 dout("but auth changed, so resending\n");
3054 __do_request(mdsc, req);
3055 mutex_unlock(&mdsc->mutex);
3059 dout("have to return ESTALE on request %llu\n", req->r_tid);
3064 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3065 __unregister_request(mdsc, req);
3067 /* last request during umount? */
3068 if (mdsc->stopping && !__get_oldest_req(mdsc))
3069 complete_all(&mdsc->safe_umount_waiters);
3071 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3073 * We already handled the unsafe response, now do the
3074 * cleanup. No need to examine the response; the MDS
3075 * doesn't include any result info in the safe
3076 * response. And even if it did, there is nothing
3077 * useful we could do with a revised return value.
3079 dout("got safe reply %llu, mds%d\n", tid, mds);
3081 mutex_unlock(&mdsc->mutex);
3085 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3086 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3089 dout("handle_reply tid %lld result %d\n", tid, result);
3090 rinfo = &req->r_reply_info;
3091 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3092 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3094 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3095 mutex_unlock(&mdsc->mutex);
3097 mutex_lock(&session->s_mutex);
3099 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3106 if (rinfo->snapblob_len) {
3107 down_write(&mdsc->snap_rwsem);
3108 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3109 rinfo->snapblob + rinfo->snapblob_len,
3110 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3112 downgrade_write(&mdsc->snap_rwsem);
3114 down_read(&mdsc->snap_rwsem);
3117 /* insert trace into our cache */
3118 mutex_lock(&req->r_fill_mutex);
3119 current->journal_info = req;
3120 err = ceph_fill_trace(mdsc->fsc->sb, req);
3122 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3123 req->r_op == CEPH_MDS_OP_LSSNAP))
3124 ceph_readdir_prepopulate(req, req->r_session);
3126 current->journal_info = NULL;
3127 mutex_unlock(&req->r_fill_mutex);
3129 up_read(&mdsc->snap_rwsem);
3131 ceph_put_snap_realm(mdsc, realm);
3134 if (req->r_target_inode &&
3135 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3136 struct ceph_inode_info *ci =
3137 ceph_inode(req->r_target_inode);
3138 spin_lock(&ci->i_unsafe_lock);
3139 list_add_tail(&req->r_unsafe_target_item,
3140 &ci->i_unsafe_iops);
3141 spin_unlock(&ci->i_unsafe_lock);
3144 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3147 mutex_lock(&mdsc->mutex);
3148 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3152 req->r_reply = ceph_msg_get(msg);
3153 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3156 dout("reply arrived after request %lld was aborted\n", tid);
3158 mutex_unlock(&mdsc->mutex);
3160 mutex_unlock(&session->s_mutex);
3162 /* kick calling process */
3163 complete_request(mdsc, req);
3165 ceph_update_metadata_latency(&mdsc->metric, req->r_start_latency,
3166 req->r_end_latency, err);
3168 ceph_mdsc_put_request(req);
3175 * handle mds notification that our request has been forwarded.
3177 static void handle_forward(struct ceph_mds_client *mdsc,
3178 struct ceph_mds_session *session,
3179 struct ceph_msg *msg)
3181 struct ceph_mds_request *req;
3182 u64 tid = le64_to_cpu(msg->hdr.tid);
3186 void *p = msg->front.iov_base;
3187 void *end = p + msg->front.iov_len;
3189 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3190 next_mds = ceph_decode_32(&p);
3191 fwd_seq = ceph_decode_32(&p);
3193 mutex_lock(&mdsc->mutex);
3194 req = lookup_get_request(mdsc, tid);
3196 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3197 goto out; /* dup reply? */
3200 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3201 dout("forward tid %llu aborted, unregistering\n", tid);
3202 __unregister_request(mdsc, req);
3203 } else if (fwd_seq <= req->r_num_fwd) {
3204 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3205 tid, next_mds, req->r_num_fwd, fwd_seq);
3207 /* resend. forward race not possible; mds would drop */
3208 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3210 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3211 req->r_attempts = 0;
3212 req->r_num_fwd = fwd_seq;
3213 req->r_resend_mds = next_mds;
3214 put_request_session(req);
3215 __do_request(mdsc, req);
3217 ceph_mdsc_put_request(req);
3219 mutex_unlock(&mdsc->mutex);
3223 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3226 static int __decode_session_metadata(void **p, void *end,
3229 /* map<string,string> */
3232 ceph_decode_32_safe(p, end, n, bad);
3235 ceph_decode_32_safe(p, end, len, bad);
3236 ceph_decode_need(p, end, len, bad);
3237 err_str = !strncmp(*p, "error_string", len);
3239 ceph_decode_32_safe(p, end, len, bad);
3240 ceph_decode_need(p, end, len, bad);
3241 if (err_str && strnstr(*p, "blacklisted", len))
3242 *blacklisted = true;
3251 * handle a mds session control message
3253 static void handle_session(struct ceph_mds_session *session,
3254 struct ceph_msg *msg)
3256 struct ceph_mds_client *mdsc = session->s_mdsc;
3257 int mds = session->s_mds;
3258 int msg_version = le16_to_cpu(msg->hdr.version);
3259 void *p = msg->front.iov_base;
3260 void *end = p + msg->front.iov_len;
3261 struct ceph_mds_session_head *h;
3263 u64 seq, features = 0;
3265 bool blacklisted = false;
3268 ceph_decode_need(&p, end, sizeof(*h), bad);
3272 op = le32_to_cpu(h->op);
3273 seq = le64_to_cpu(h->seq);
3275 if (msg_version >= 3) {
3277 /* version >= 2, metadata */
3278 if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3280 /* version >= 3, feature bits */
3281 ceph_decode_32_safe(&p, end, len, bad);
3282 ceph_decode_64_safe(&p, end, features, bad);
3283 p += len - sizeof(features);
3286 mutex_lock(&mdsc->mutex);
3287 if (op == CEPH_SESSION_CLOSE) {
3288 ceph_get_mds_session(session);
3289 __unregister_session(mdsc, session);
3291 /* FIXME: this ttl calculation is generous */
3292 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3293 mutex_unlock(&mdsc->mutex);
3295 mutex_lock(&session->s_mutex);
3297 dout("handle_session mds%d %s %p state %s seq %llu\n",
3298 mds, ceph_session_op_name(op), session,
3299 ceph_session_state_name(session->s_state), seq);
3301 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3302 session->s_state = CEPH_MDS_SESSION_OPEN;
3303 pr_info("mds%d came back\n", session->s_mds);
3307 case CEPH_SESSION_OPEN:
3308 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3309 pr_info("mds%d reconnect success\n", session->s_mds);
3310 session->s_state = CEPH_MDS_SESSION_OPEN;
3311 session->s_features = features;
3312 renewed_caps(mdsc, session, 0);
3315 __close_session(mdsc, session);
3318 case CEPH_SESSION_RENEWCAPS:
3319 if (session->s_renew_seq == seq)
3320 renewed_caps(mdsc, session, 1);
3323 case CEPH_SESSION_CLOSE:
3324 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3325 pr_info("mds%d reconnect denied\n", session->s_mds);
3326 session->s_state = CEPH_MDS_SESSION_CLOSED;
3327 cleanup_session_requests(mdsc, session);
3328 remove_session_caps(session);
3329 wake = 2; /* for good measure */
3330 wake_up_all(&mdsc->session_close_wq);
3333 case CEPH_SESSION_STALE:
3334 pr_info("mds%d caps went stale, renewing\n",
3336 spin_lock(&session->s_gen_ttl_lock);
3337 session->s_cap_gen++;
3338 session->s_cap_ttl = jiffies - 1;
3339 spin_unlock(&session->s_gen_ttl_lock);
3340 send_renew_caps(mdsc, session);
3343 case CEPH_SESSION_RECALL_STATE:
3344 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3347 case CEPH_SESSION_FLUSHMSG:
3348 send_flushmsg_ack(mdsc, session, seq);
3351 case CEPH_SESSION_FORCE_RO:
3352 dout("force_session_readonly %p\n", session);
3353 spin_lock(&session->s_cap_lock);
3354 session->s_readonly = true;
3355 spin_unlock(&session->s_cap_lock);
3356 wake_up_session_caps(session, FORCE_RO);
3359 case CEPH_SESSION_REJECT:
3360 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3361 pr_info("mds%d rejected session\n", session->s_mds);
3362 session->s_state = CEPH_MDS_SESSION_REJECTED;
3363 cleanup_session_requests(mdsc, session);
3364 remove_session_caps(session);
3366 mdsc->fsc->blacklisted = true;
3367 wake = 2; /* for good measure */
3371 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3375 mutex_unlock(&session->s_mutex);
3377 mutex_lock(&mdsc->mutex);
3378 __wake_requests(mdsc, &session->s_waiting);
3380 kick_requests(mdsc, mds);
3381 mutex_unlock(&mdsc->mutex);
3383 if (op == CEPH_SESSION_CLOSE)
3384 ceph_put_mds_session(session);
3388 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3389 (int)msg->front.iov_len);
3394 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3398 dcaps = xchg(&req->r_dir_caps, 0);
3400 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3401 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3405 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3409 dcaps = xchg(&req->r_dir_caps, 0);
3411 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3412 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3418 * called under session->mutex.
3420 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3421 struct ceph_mds_session *session)
3423 struct ceph_mds_request *req, *nreq;
3426 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3428 mutex_lock(&mdsc->mutex);
3429 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3430 __send_request(mdsc, session, req, true);
3433 * also re-send old requests when MDS enters reconnect stage. So that MDS
3434 * can process completed request in clientreplay stage.
3436 p = rb_first(&mdsc->request_tree);
3438 req = rb_entry(p, struct ceph_mds_request, r_node);
3440 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3442 if (req->r_attempts == 0)
3443 continue; /* only old requests */
3444 if (!req->r_session)
3446 if (req->r_session->s_mds != session->s_mds)
3449 ceph_mdsc_release_dir_caps_no_check(req);
3451 __send_request(mdsc, session, req, true);
3453 mutex_unlock(&mdsc->mutex);
3456 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3458 struct ceph_msg *reply;
3459 struct ceph_pagelist *_pagelist;
3464 if (!recon_state->allow_multi)
3467 /* can't handle message that contains both caps and realm */
3468 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3470 /* pre-allocate new pagelist */
3471 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3475 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3479 /* placeholder for nr_caps */
3480 err = ceph_pagelist_encode_32(_pagelist, 0);
3484 if (recon_state->nr_caps) {
3485 /* currently encoding caps */
3486 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3490 /* placeholder for nr_realms (currently encoding relams) */
3491 err = ceph_pagelist_encode_32(_pagelist, 0);
3496 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3500 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3501 addr = kmap_atomic(page);
3502 if (recon_state->nr_caps) {
3503 /* currently encoding caps */
3504 *addr = cpu_to_le32(recon_state->nr_caps);
3506 /* currently encoding relams */
3507 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3509 kunmap_atomic(addr);
3511 reply->hdr.version = cpu_to_le16(5);
3512 reply->hdr.compat_version = cpu_to_le16(4);
3514 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3515 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3517 ceph_con_send(&recon_state->session->s_con, reply);
3518 ceph_pagelist_release(recon_state->pagelist);
3520 recon_state->pagelist = _pagelist;
3521 recon_state->nr_caps = 0;
3522 recon_state->nr_realms = 0;
3523 recon_state->msg_version = 5;
3526 ceph_msg_put(reply);
3528 ceph_pagelist_release(_pagelist);
3533 * Encode information about a cap for a reconnect with the MDS.
3535 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3539 struct ceph_mds_cap_reconnect v2;
3540 struct ceph_mds_cap_reconnect_v1 v1;
3542 struct ceph_inode_info *ci = cap->ci;
3543 struct ceph_reconnect_state *recon_state = arg;
3544 struct ceph_pagelist *pagelist = recon_state->pagelist;
3548 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3549 inode, ceph_vinop(inode), cap, cap->cap_id,
3550 ceph_cap_string(cap->issued));
3552 spin_lock(&ci->i_ceph_lock);
3553 cap->seq = 0; /* reset cap seq */
3554 cap->issue_seq = 0; /* and issue_seq */
3555 cap->mseq = 0; /* and migrate_seq */
3556 cap->cap_gen = cap->session->s_cap_gen;
3558 /* These are lost when the session goes away */
3559 if (S_ISDIR(inode->i_mode)) {
3560 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3561 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3562 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3564 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3567 if (recon_state->msg_version >= 2) {
3568 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3569 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3570 rec.v2.issued = cpu_to_le32(cap->issued);
3571 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3572 rec.v2.pathbase = 0;
3573 rec.v2.flock_len = (__force __le32)
3574 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3576 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3577 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3578 rec.v1.issued = cpu_to_le32(cap->issued);
3579 rec.v1.size = cpu_to_le64(inode->i_size);
3580 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3581 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3582 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3583 rec.v1.pathbase = 0;
3586 if (list_empty(&ci->i_cap_snaps)) {
3587 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3589 struct ceph_cap_snap *capsnap =
3590 list_first_entry(&ci->i_cap_snaps,
3591 struct ceph_cap_snap, ci_item);
3592 snap_follows = capsnap->follows;
3594 spin_unlock(&ci->i_ceph_lock);
3596 if (recon_state->msg_version >= 2) {
3597 int num_fcntl_locks, num_flock_locks;
3598 struct ceph_filelock *flocks = NULL;
3599 size_t struct_len, total_len = sizeof(u64);
3603 if (rec.v2.flock_len) {
3604 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3606 num_fcntl_locks = 0;
3607 num_flock_locks = 0;
3609 if (num_fcntl_locks + num_flock_locks > 0) {
3610 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3611 sizeof(struct ceph_filelock),
3617 err = ceph_encode_locks_to_buffer(inode, flocks,
3632 if (recon_state->msg_version >= 3) {
3633 /* version, compat_version and struct_len */
3634 total_len += 2 * sizeof(u8) + sizeof(u32);
3638 * number of encoded locks is stable, so copy to pagelist
3640 struct_len = 2 * sizeof(u32) +
3641 (num_fcntl_locks + num_flock_locks) *
3642 sizeof(struct ceph_filelock);
3643 rec.v2.flock_len = cpu_to_le32(struct_len);
3645 struct_len += sizeof(u32) + sizeof(rec.v2);
3648 struct_len += sizeof(u64); /* snap_follows */
3650 total_len += struct_len;
3652 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3653 err = send_reconnect_partial(recon_state);
3655 goto out_freeflocks;
3656 pagelist = recon_state->pagelist;
3659 err = ceph_pagelist_reserve(pagelist, total_len);
3661 goto out_freeflocks;
3663 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3664 if (recon_state->msg_version >= 3) {
3665 ceph_pagelist_encode_8(pagelist, struct_v);
3666 ceph_pagelist_encode_8(pagelist, 1);
3667 ceph_pagelist_encode_32(pagelist, struct_len);
3669 ceph_pagelist_encode_string(pagelist, NULL, 0);
3670 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3671 ceph_locks_to_pagelist(flocks, pagelist,
3672 num_fcntl_locks, num_flock_locks);
3674 ceph_pagelist_encode_64(pagelist, snap_follows);
3681 struct dentry *dentry;
3683 dentry = d_find_alias(inode);
3685 path = ceph_mdsc_build_path(dentry,
3686 &pathlen, &pathbase, 0);
3689 err = PTR_ERR(path);
3692 rec.v1.pathbase = cpu_to_le64(pathbase);
3695 err = ceph_pagelist_reserve(pagelist,
3696 sizeof(u64) + sizeof(u32) +
3697 pathlen + sizeof(rec.v1));
3702 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3703 ceph_pagelist_encode_string(pagelist, path, pathlen);
3704 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3706 ceph_mdsc_free_path(path, pathlen);
3711 recon_state->nr_caps++;
3715 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3716 struct ceph_reconnect_state *recon_state)
3719 struct ceph_pagelist *pagelist = recon_state->pagelist;
3722 if (recon_state->msg_version >= 4) {
3723 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3729 * snaprealms. we provide mds with the ino, seq (version), and
3730 * parent for all of our realms. If the mds has any newer info,
3733 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3734 struct ceph_snap_realm *realm =
3735 rb_entry(p, struct ceph_snap_realm, node);
3736 struct ceph_mds_snaprealm_reconnect sr_rec;
3738 if (recon_state->msg_version >= 4) {
3739 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3742 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3743 err = send_reconnect_partial(recon_state);
3746 pagelist = recon_state->pagelist;
3749 err = ceph_pagelist_reserve(pagelist, need);
3753 ceph_pagelist_encode_8(pagelist, 1);
3754 ceph_pagelist_encode_8(pagelist, 1);
3755 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3758 dout(" adding snap realm %llx seq %lld parent %llx\n",
3759 realm->ino, realm->seq, realm->parent_ino);
3760 sr_rec.ino = cpu_to_le64(realm->ino);
3761 sr_rec.seq = cpu_to_le64(realm->seq);
3762 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3764 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3768 recon_state->nr_realms++;
3776 * If an MDS fails and recovers, clients need to reconnect in order to
3777 * reestablish shared state. This includes all caps issued through
3778 * this session _and_ the snap_realm hierarchy. Because it's not
3779 * clear which snap realms the mds cares about, we send everything we
3780 * know about.. that ensures we'll then get any new info the
3781 * recovering MDS might have.
3783 * This is a relatively heavyweight operation, but it's rare.
3785 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3786 struct ceph_mds_session *session)
3788 struct ceph_msg *reply;
3789 int mds = session->s_mds;
3791 struct ceph_reconnect_state recon_state = {
3796 pr_info("mds%d reconnect start\n", mds);
3798 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3799 if (!recon_state.pagelist)
3800 goto fail_nopagelist;
3802 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3806 xa_destroy(&session->s_delegated_inos);
3808 mutex_lock(&session->s_mutex);
3809 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3812 dout("session %p state %s\n", session,
3813 ceph_session_state_name(session->s_state));
3815 spin_lock(&session->s_gen_ttl_lock);
3816 session->s_cap_gen++;
3817 spin_unlock(&session->s_gen_ttl_lock);
3819 spin_lock(&session->s_cap_lock);
3820 /* don't know if session is readonly */
3821 session->s_readonly = 0;
3823 * notify __ceph_remove_cap() that we are composing cap reconnect.
3824 * If a cap get released before being added to the cap reconnect,
3825 * __ceph_remove_cap() should skip queuing cap release.
3827 session->s_cap_reconnect = 1;
3828 /* drop old cap expires; we're about to reestablish that state */
3829 detach_cap_releases(session, &dispose);
3830 spin_unlock(&session->s_cap_lock);
3831 dispose_cap_releases(mdsc, &dispose);
3833 /* trim unused caps to reduce MDS's cache rejoin time */
3834 if (mdsc->fsc->sb->s_root)
3835 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3837 ceph_con_close(&session->s_con);
3838 ceph_con_open(&session->s_con,
3839 CEPH_ENTITY_TYPE_MDS, mds,
3840 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3842 /* replay unsafe requests */
3843 replay_unsafe_requests(mdsc, session);
3845 ceph_early_kick_flushing_caps(mdsc, session);
3847 down_read(&mdsc->snap_rwsem);
3849 /* placeholder for nr_caps */
3850 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3854 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3855 recon_state.msg_version = 3;
3856 recon_state.allow_multi = true;
3857 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3858 recon_state.msg_version = 3;
3860 recon_state.msg_version = 2;
3862 /* trsaverse this session's caps */
3863 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
3865 spin_lock(&session->s_cap_lock);
3866 session->s_cap_reconnect = 0;
3867 spin_unlock(&session->s_cap_lock);
3872 /* check if all realms can be encoded into current message */
3873 if (mdsc->num_snap_realms) {
3875 recon_state.pagelist->length +
3876 mdsc->num_snap_realms *
3877 sizeof(struct ceph_mds_snaprealm_reconnect);
3878 if (recon_state.msg_version >= 4) {
3879 /* number of realms */
3880 total_len += sizeof(u32);
3881 /* version, compat_version and struct_len */
3882 total_len += mdsc->num_snap_realms *
3883 (2 * sizeof(u8) + sizeof(u32));
3885 if (total_len > RECONNECT_MAX_SIZE) {
3886 if (!recon_state.allow_multi) {
3890 if (recon_state.nr_caps) {
3891 err = send_reconnect_partial(&recon_state);
3895 recon_state.msg_version = 5;
3899 err = encode_snap_realms(mdsc, &recon_state);
3903 if (recon_state.msg_version >= 5) {
3904 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3909 if (recon_state.nr_caps || recon_state.nr_realms) {
3911 list_first_entry(&recon_state.pagelist->head,
3913 __le32 *addr = kmap_atomic(page);
3914 if (recon_state.nr_caps) {
3915 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3916 *addr = cpu_to_le32(recon_state.nr_caps);
3917 } else if (recon_state.msg_version >= 4) {
3918 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3920 kunmap_atomic(addr);
3923 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3924 if (recon_state.msg_version >= 4)
3925 reply->hdr.compat_version = cpu_to_le16(4);
3927 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3928 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3930 ceph_con_send(&session->s_con, reply);
3932 mutex_unlock(&session->s_mutex);
3934 mutex_lock(&mdsc->mutex);
3935 __wake_requests(mdsc, &session->s_waiting);
3936 mutex_unlock(&mdsc->mutex);
3938 up_read(&mdsc->snap_rwsem);
3939 ceph_pagelist_release(recon_state.pagelist);
3943 ceph_msg_put(reply);
3944 up_read(&mdsc->snap_rwsem);
3945 mutex_unlock(&session->s_mutex);
3947 ceph_pagelist_release(recon_state.pagelist);
3949 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3955 * compare old and new mdsmaps, kicking requests
3956 * and closing out old connections as necessary
3958 * called under mdsc->mutex.
3960 static void check_new_map(struct ceph_mds_client *mdsc,
3961 struct ceph_mdsmap *newmap,
3962 struct ceph_mdsmap *oldmap)
3965 int oldstate, newstate;
3966 struct ceph_mds_session *s;
3968 dout("check_new_map new %u old %u\n",
3969 newmap->m_epoch, oldmap->m_epoch);
3971 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
3972 if (!mdsc->sessions[i])
3974 s = mdsc->sessions[i];
3975 oldstate = ceph_mdsmap_get_state(oldmap, i);
3976 newstate = ceph_mdsmap_get_state(newmap, i);
3978 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3979 i, ceph_mds_state_name(oldstate),
3980 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3981 ceph_mds_state_name(newstate),
3982 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3983 ceph_session_state_name(s->s_state));
3985 if (i >= newmap->possible_max_rank) {
3986 /* force close session for stopped mds */
3987 ceph_get_mds_session(s);
3988 __unregister_session(mdsc, s);
3989 __wake_requests(mdsc, &s->s_waiting);
3990 mutex_unlock(&mdsc->mutex);
3992 mutex_lock(&s->s_mutex);
3993 cleanup_session_requests(mdsc, s);
3994 remove_session_caps(s);
3995 mutex_unlock(&s->s_mutex);
3997 ceph_put_mds_session(s);
3999 mutex_lock(&mdsc->mutex);
4000 kick_requests(mdsc, i);
4004 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4005 ceph_mdsmap_get_addr(newmap, i),
4006 sizeof(struct ceph_entity_addr))) {
4008 mutex_unlock(&mdsc->mutex);
4009 mutex_lock(&s->s_mutex);
4010 mutex_lock(&mdsc->mutex);
4011 ceph_con_close(&s->s_con);
4012 mutex_unlock(&s->s_mutex);
4013 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4014 } else if (oldstate == newstate) {
4015 continue; /* nothing new with this mds */
4021 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4022 newstate >= CEPH_MDS_STATE_RECONNECT) {
4023 mutex_unlock(&mdsc->mutex);
4024 send_mds_reconnect(mdsc, s);
4025 mutex_lock(&mdsc->mutex);
4029 * kick request on any mds that has gone active.
4031 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4032 newstate >= CEPH_MDS_STATE_ACTIVE) {
4033 if (oldstate != CEPH_MDS_STATE_CREATING &&
4034 oldstate != CEPH_MDS_STATE_STARTING)
4035 pr_info("mds%d recovery completed\n", s->s_mds);
4036 kick_requests(mdsc, i);
4037 mutex_unlock(&mdsc->mutex);
4038 mutex_lock(&s->s_mutex);
4039 mutex_lock(&mdsc->mutex);
4040 ceph_kick_flushing_caps(mdsc, s);
4041 mutex_unlock(&s->s_mutex);
4042 wake_up_session_caps(s, RECONNECT);
4046 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4047 s = mdsc->sessions[i];
4050 if (!ceph_mdsmap_is_laggy(newmap, i))
4052 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4053 s->s_state == CEPH_MDS_SESSION_HUNG ||
4054 s->s_state == CEPH_MDS_SESSION_CLOSING) {
4055 dout(" connecting to export targets of laggy mds%d\n",
4057 __open_export_target_sessions(mdsc, s);
4069 * caller must hold session s_mutex, dentry->d_lock
4071 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4073 struct ceph_dentry_info *di = ceph_dentry(dentry);
4075 ceph_put_mds_session(di->lease_session);
4076 di->lease_session = NULL;
4079 static void handle_lease(struct ceph_mds_client *mdsc,
4080 struct ceph_mds_session *session,
4081 struct ceph_msg *msg)
4083 struct super_block *sb = mdsc->fsc->sb;
4084 struct inode *inode;
4085 struct dentry *parent, *dentry;
4086 struct ceph_dentry_info *di;
4087 int mds = session->s_mds;
4088 struct ceph_mds_lease *h = msg->front.iov_base;
4090 struct ceph_vino vino;
4094 dout("handle_lease from mds%d\n", mds);
4097 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4099 vino.ino = le64_to_cpu(h->ino);
4100 vino.snap = CEPH_NOSNAP;
4101 seq = le32_to_cpu(h->seq);
4102 dname.len = get_unaligned_le32(h + 1);
4103 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4105 dname.name = (void *)(h + 1) + sizeof(u32);
4108 inode = ceph_find_inode(sb, vino);
4109 dout("handle_lease %s, ino %llx %p %.*s\n",
4110 ceph_lease_op_name(h->action), vino.ino, inode,
4111 dname.len, dname.name);
4113 mutex_lock(&session->s_mutex);
4117 dout("handle_lease no inode %llx\n", vino.ino);
4122 parent = d_find_alias(inode);
4124 dout("no parent dentry on inode %p\n", inode);
4126 goto release; /* hrm... */
4128 dname.hash = full_name_hash(parent, dname.name, dname.len);
4129 dentry = d_lookup(parent, &dname);
4134 spin_lock(&dentry->d_lock);
4135 di = ceph_dentry(dentry);
4136 switch (h->action) {
4137 case CEPH_MDS_LEASE_REVOKE:
4138 if (di->lease_session == session) {
4139 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4140 h->seq = cpu_to_le32(di->lease_seq);
4141 __ceph_mdsc_drop_dentry_lease(dentry);
4146 case CEPH_MDS_LEASE_RENEW:
4147 if (di->lease_session == session &&
4148 di->lease_gen == session->s_cap_gen &&
4149 di->lease_renew_from &&
4150 di->lease_renew_after == 0) {
4151 unsigned long duration =
4152 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4154 di->lease_seq = seq;
4155 di->time = di->lease_renew_from + duration;
4156 di->lease_renew_after = di->lease_renew_from +
4158 di->lease_renew_from = 0;
4162 spin_unlock(&dentry->d_lock);
4169 /* let's just reuse the same message */
4170 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4172 ceph_con_send(&session->s_con, msg);
4175 mutex_unlock(&session->s_mutex);
4176 /* avoid calling iput_final() in mds dispatch threads */
4177 ceph_async_iput(inode);
4181 pr_err("corrupt lease message\n");
4185 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4186 struct dentry *dentry, char action,
4189 struct ceph_msg *msg;
4190 struct ceph_mds_lease *lease;
4192 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4194 dout("lease_send_msg identry %p %s to mds%d\n",
4195 dentry, ceph_lease_op_name(action), session->s_mds);
4197 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4200 lease = msg->front.iov_base;
4201 lease->action = action;
4202 lease->seq = cpu_to_le32(seq);
4204 spin_lock(&dentry->d_lock);
4205 dir = d_inode(dentry->d_parent);
4206 lease->ino = cpu_to_le64(ceph_ino(dir));
4207 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4209 put_unaligned_le32(dentry->d_name.len, lease + 1);
4210 memcpy((void *)(lease + 1) + 4,
4211 dentry->d_name.name, dentry->d_name.len);
4212 spin_unlock(&dentry->d_lock);
4214 * if this is a preemptive lease RELEASE, no need to
4215 * flush request stream, since the actual request will
4218 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4220 ceph_con_send(&session->s_con, msg);
4224 * lock unlock sessions, to wait ongoing session activities
4226 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4230 mutex_lock(&mdsc->mutex);
4231 for (i = 0; i < mdsc->max_sessions; i++) {
4232 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4235 mutex_unlock(&mdsc->mutex);
4236 mutex_lock(&s->s_mutex);
4237 mutex_unlock(&s->s_mutex);
4238 ceph_put_mds_session(s);
4239 mutex_lock(&mdsc->mutex);
4241 mutex_unlock(&mdsc->mutex);
4244 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4246 struct ceph_fs_client *fsc = mdsc->fsc;
4248 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4251 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4254 if (!READ_ONCE(fsc->blacklisted))
4257 if (fsc->last_auto_reconnect &&
4258 time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4261 pr_info("auto reconnect after blacklisted\n");
4262 fsc->last_auto_reconnect = jiffies;
4263 ceph_force_reconnect(fsc->sb);
4267 * delayed work -- periodically trim expired leases, renew caps with mds
4269 static void schedule_delayed(struct ceph_mds_client *mdsc)
4272 unsigned hz = round_jiffies_relative(HZ * delay);
4273 schedule_delayed_work(&mdsc->delayed_work, hz);
4276 static void delayed_work(struct work_struct *work)
4279 struct ceph_mds_client *mdsc =
4280 container_of(work, struct ceph_mds_client, delayed_work.work);
4284 dout("mdsc delayed_work\n");
4286 mutex_lock(&mdsc->mutex);
4287 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4288 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4289 mdsc->last_renew_caps);
4291 mdsc->last_renew_caps = jiffies;
4293 for (i = 0; i < mdsc->max_sessions; i++) {
4294 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4297 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4298 dout("resending session close request for mds%d\n",
4300 request_close_session(mdsc, s);
4301 ceph_put_mds_session(s);
4304 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4305 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4306 s->s_state = CEPH_MDS_SESSION_HUNG;
4307 pr_info("mds%d hung\n", s->s_mds);
4310 if (s->s_state == CEPH_MDS_SESSION_NEW ||
4311 s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4312 s->s_state == CEPH_MDS_SESSION_REJECTED) {
4313 /* this mds is failed or recovering, just wait */
4314 ceph_put_mds_session(s);
4317 mutex_unlock(&mdsc->mutex);
4319 mutex_lock(&s->s_mutex);
4321 send_renew_caps(mdsc, s);
4323 ceph_con_keepalive(&s->s_con);
4324 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4325 s->s_state == CEPH_MDS_SESSION_HUNG)
4326 ceph_send_cap_releases(mdsc, s);
4327 mutex_unlock(&s->s_mutex);
4328 ceph_put_mds_session(s);
4330 mutex_lock(&mdsc->mutex);
4332 mutex_unlock(&mdsc->mutex);
4334 ceph_check_delayed_caps(mdsc);
4336 ceph_queue_cap_reclaim_work(mdsc);
4338 ceph_trim_snapid_map(mdsc);
4340 maybe_recover_session(mdsc);
4342 schedule_delayed(mdsc);
4345 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4348 struct ceph_mds_client *mdsc;
4351 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4355 mutex_init(&mdsc->mutex);
4356 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4357 if (!mdsc->mdsmap) {
4363 init_completion(&mdsc->safe_umount_waiters);
4364 init_waitqueue_head(&mdsc->session_close_wq);
4365 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4366 mdsc->sessions = NULL;
4367 atomic_set(&mdsc->num_sessions, 0);
4368 mdsc->max_sessions = 0;
4370 atomic64_set(&mdsc->quotarealms_count, 0);
4371 mdsc->quotarealms_inodes = RB_ROOT;
4372 mutex_init(&mdsc->quotarealms_inodes_mutex);
4373 mdsc->last_snap_seq = 0;
4374 init_rwsem(&mdsc->snap_rwsem);
4375 mdsc->snap_realms = RB_ROOT;
4376 INIT_LIST_HEAD(&mdsc->snap_empty);
4377 mdsc->num_snap_realms = 0;
4378 spin_lock_init(&mdsc->snap_empty_lock);
4380 mdsc->oldest_tid = 0;
4381 mdsc->request_tree = RB_ROOT;
4382 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4383 mdsc->last_renew_caps = jiffies;
4384 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4385 INIT_LIST_HEAD(&mdsc->cap_wait_list);
4386 spin_lock_init(&mdsc->cap_delay_lock);
4387 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4388 spin_lock_init(&mdsc->snap_flush_lock);
4389 mdsc->last_cap_flush_tid = 1;
4390 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4391 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4392 mdsc->num_cap_flushing = 0;
4393 spin_lock_init(&mdsc->cap_dirty_lock);
4394 init_waitqueue_head(&mdsc->cap_flushing_wq);
4395 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4396 atomic_set(&mdsc->cap_reclaim_pending, 0);
4397 err = ceph_metric_init(&mdsc->metric);
4401 spin_lock_init(&mdsc->dentry_list_lock);
4402 INIT_LIST_HEAD(&mdsc->dentry_leases);
4403 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4405 ceph_caps_init(mdsc);
4406 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4408 spin_lock_init(&mdsc->snapid_map_lock);
4409 mdsc->snapid_map_tree = RB_ROOT;
4410 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4412 init_rwsem(&mdsc->pool_perm_rwsem);
4413 mdsc->pool_perm_tree = RB_ROOT;
4415 strscpy(mdsc->nodename, utsname()->nodename,
4416 sizeof(mdsc->nodename));
4420 kfree(mdsc->mdsmap);
4427 * Wait for safe replies on open mds requests. If we time out, drop
4428 * all requests from the tree to avoid dangling dentry refs.
4430 static void wait_requests(struct ceph_mds_client *mdsc)
4432 struct ceph_options *opts = mdsc->fsc->client->options;
4433 struct ceph_mds_request *req;
4435 mutex_lock(&mdsc->mutex);
4436 if (__get_oldest_req(mdsc)) {
4437 mutex_unlock(&mdsc->mutex);
4439 dout("wait_requests waiting for requests\n");
4440 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4441 ceph_timeout_jiffies(opts->mount_timeout));
4443 /* tear down remaining requests */
4444 mutex_lock(&mdsc->mutex);
4445 while ((req = __get_oldest_req(mdsc))) {
4446 dout("wait_requests timed out on tid %llu\n",
4448 list_del_init(&req->r_wait);
4449 __unregister_request(mdsc, req);
4452 mutex_unlock(&mdsc->mutex);
4453 dout("wait_requests done\n");
4457 * called before mount is ro, and before dentries are torn down.
4458 * (hmm, does this still race with new lookups?)
4460 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4462 dout("pre_umount\n");
4465 lock_unlock_sessions(mdsc);
4466 ceph_flush_dirty_caps(mdsc);
4467 wait_requests(mdsc);
4470 * wait for reply handlers to drop their request refs and
4471 * their inode/dcache refs
4475 ceph_cleanup_quotarealms_inodes(mdsc);
4479 * wait for all write mds requests to flush.
4481 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4483 struct ceph_mds_request *req = NULL, *nextreq;
4486 mutex_lock(&mdsc->mutex);
4487 dout("wait_unsafe_requests want %lld\n", want_tid);
4489 req = __get_oldest_req(mdsc);
4490 while (req && req->r_tid <= want_tid) {
4491 /* find next request */
4492 n = rb_next(&req->r_node);
4494 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4497 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4498 (req->r_op & CEPH_MDS_OP_WRITE)) {
4500 ceph_mdsc_get_request(req);
4502 ceph_mdsc_get_request(nextreq);
4503 mutex_unlock(&mdsc->mutex);
4504 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
4505 req->r_tid, want_tid);
4506 wait_for_completion(&req->r_safe_completion);
4507 mutex_lock(&mdsc->mutex);
4508 ceph_mdsc_put_request(req);
4510 break; /* next dne before, so we're done! */
4511 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4512 /* next request was removed from tree */
4513 ceph_mdsc_put_request(nextreq);
4516 ceph_mdsc_put_request(nextreq); /* won't go away */
4520 mutex_unlock(&mdsc->mutex);
4521 dout("wait_unsafe_requests done\n");
4524 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4526 u64 want_tid, want_flush;
4528 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4532 mutex_lock(&mdsc->mutex);
4533 want_tid = mdsc->last_tid;
4534 mutex_unlock(&mdsc->mutex);
4536 ceph_flush_dirty_caps(mdsc);
4537 spin_lock(&mdsc->cap_dirty_lock);
4538 want_flush = mdsc->last_cap_flush_tid;
4539 if (!list_empty(&mdsc->cap_flush_list)) {
4540 struct ceph_cap_flush *cf =
4541 list_last_entry(&mdsc->cap_flush_list,
4542 struct ceph_cap_flush, g_list);
4545 spin_unlock(&mdsc->cap_dirty_lock);
4547 dout("sync want tid %lld flush_seq %lld\n",
4548 want_tid, want_flush);
4550 wait_unsafe_requests(mdsc, want_tid);
4551 wait_caps_flush(mdsc, want_flush);
4555 * true if all sessions are closed, or we force unmount
4557 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4559 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4561 return atomic_read(&mdsc->num_sessions) <= skipped;
4565 * called after sb is ro.
4567 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4569 struct ceph_options *opts = mdsc->fsc->client->options;
4570 struct ceph_mds_session *session;
4574 dout("close_sessions\n");
4576 /* close sessions */
4577 mutex_lock(&mdsc->mutex);
4578 for (i = 0; i < mdsc->max_sessions; i++) {
4579 session = __ceph_lookup_mds_session(mdsc, i);
4582 mutex_unlock(&mdsc->mutex);
4583 mutex_lock(&session->s_mutex);
4584 if (__close_session(mdsc, session) <= 0)
4586 mutex_unlock(&session->s_mutex);
4587 ceph_put_mds_session(session);
4588 mutex_lock(&mdsc->mutex);
4590 mutex_unlock(&mdsc->mutex);
4592 dout("waiting for sessions to close\n");
4593 wait_event_timeout(mdsc->session_close_wq,
4594 done_closing_sessions(mdsc, skipped),
4595 ceph_timeout_jiffies(opts->mount_timeout));
4597 /* tear down remaining sessions */
4598 mutex_lock(&mdsc->mutex);
4599 for (i = 0; i < mdsc->max_sessions; i++) {
4600 if (mdsc->sessions[i]) {
4601 session = ceph_get_mds_session(mdsc->sessions[i]);
4602 __unregister_session(mdsc, session);
4603 mutex_unlock(&mdsc->mutex);
4604 mutex_lock(&session->s_mutex);
4605 remove_session_caps(session);
4606 mutex_unlock(&session->s_mutex);
4607 ceph_put_mds_session(session);
4608 mutex_lock(&mdsc->mutex);
4611 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4612 mutex_unlock(&mdsc->mutex);
4614 ceph_cleanup_snapid_map(mdsc);
4615 ceph_cleanup_empty_realms(mdsc);
4617 cancel_work_sync(&mdsc->cap_reclaim_work);
4618 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4623 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4625 struct ceph_mds_session *session;
4628 dout("force umount\n");
4630 mutex_lock(&mdsc->mutex);
4631 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4632 session = __ceph_lookup_mds_session(mdsc, mds);
4636 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4637 __unregister_session(mdsc, session);
4638 __wake_requests(mdsc, &session->s_waiting);
4639 mutex_unlock(&mdsc->mutex);
4641 mutex_lock(&session->s_mutex);
4642 __close_session(mdsc, session);
4643 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4644 cleanup_session_requests(mdsc, session);
4645 remove_session_caps(session);
4647 mutex_unlock(&session->s_mutex);
4648 ceph_put_mds_session(session);
4650 mutex_lock(&mdsc->mutex);
4651 kick_requests(mdsc, mds);
4653 __wake_requests(mdsc, &mdsc->waiting_for_map);
4654 mutex_unlock(&mdsc->mutex);
4657 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4660 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4662 ceph_mdsmap_destroy(mdsc->mdsmap);
4663 kfree(mdsc->sessions);
4664 ceph_caps_finalize(mdsc);
4665 ceph_pool_perm_destroy(mdsc);
4668 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4670 struct ceph_mds_client *mdsc = fsc->mdsc;
4671 dout("mdsc_destroy %p\n", mdsc);
4676 /* flush out any connection work with references to us */
4679 ceph_mdsc_stop(mdsc);
4681 ceph_metric_destroy(&mdsc->metric);
4685 dout("mdsc_destroy %p done\n", mdsc);
4688 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4690 struct ceph_fs_client *fsc = mdsc->fsc;
4691 const char *mds_namespace = fsc->mount_options->mds_namespace;
4692 void *p = msg->front.iov_base;
4693 void *end = p + msg->front.iov_len;
4697 u32 mount_fscid = (u32)-1;
4698 u8 struct_v, struct_cv;
4701 ceph_decode_need(&p, end, sizeof(u32), bad);
4702 epoch = ceph_decode_32(&p);
4704 dout("handle_fsmap epoch %u\n", epoch);
4706 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4707 struct_v = ceph_decode_8(&p);
4708 struct_cv = ceph_decode_8(&p);
4709 map_len = ceph_decode_32(&p);
4711 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4712 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4714 num_fs = ceph_decode_32(&p);
4715 while (num_fs-- > 0) {
4716 void *info_p, *info_end;
4721 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4722 info_v = ceph_decode_8(&p);
4723 info_cv = ceph_decode_8(&p);
4724 info_len = ceph_decode_32(&p);
4725 ceph_decode_need(&p, end, info_len, bad);
4727 info_end = p + info_len;
4730 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4731 fscid = ceph_decode_32(&info_p);
4732 namelen = ceph_decode_32(&info_p);
4733 ceph_decode_need(&info_p, info_end, namelen, bad);
4735 if (mds_namespace &&
4736 strlen(mds_namespace) == namelen &&
4737 !strncmp(mds_namespace, (char *)info_p, namelen)) {
4738 mount_fscid = fscid;
4743 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4744 if (mount_fscid != (u32)-1) {
4745 fsc->client->monc.fs_cluster_id = mount_fscid;
4746 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4748 ceph_monc_renew_subs(&fsc->client->monc);
4756 pr_err("error decoding fsmap\n");
4758 mutex_lock(&mdsc->mutex);
4759 mdsc->mdsmap_err = err;
4760 __wake_requests(mdsc, &mdsc->waiting_for_map);
4761 mutex_unlock(&mdsc->mutex);
4765 * handle mds map update.
4767 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4771 void *p = msg->front.iov_base;
4772 void *end = p + msg->front.iov_len;
4773 struct ceph_mdsmap *newmap, *oldmap;
4774 struct ceph_fsid fsid;
4777 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4778 ceph_decode_copy(&p, &fsid, sizeof(fsid));
4779 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4781 epoch = ceph_decode_32(&p);
4782 maplen = ceph_decode_32(&p);
4783 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4785 /* do we need it? */
4786 mutex_lock(&mdsc->mutex);
4787 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4788 dout("handle_map epoch %u <= our %u\n",
4789 epoch, mdsc->mdsmap->m_epoch);
4790 mutex_unlock(&mdsc->mutex);
4794 newmap = ceph_mdsmap_decode(&p, end);
4795 if (IS_ERR(newmap)) {
4796 err = PTR_ERR(newmap);
4800 /* swap into place */
4802 oldmap = mdsc->mdsmap;
4803 mdsc->mdsmap = newmap;
4804 check_new_map(mdsc, newmap, oldmap);
4805 ceph_mdsmap_destroy(oldmap);
4807 mdsc->mdsmap = newmap; /* first mds map */
4809 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4812 __wake_requests(mdsc, &mdsc->waiting_for_map);
4813 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4814 mdsc->mdsmap->m_epoch);
4816 mutex_unlock(&mdsc->mutex);
4817 schedule_delayed(mdsc);
4821 mutex_unlock(&mdsc->mutex);
4823 pr_err("error decoding mdsmap %d\n", err);
4827 static struct ceph_connection *con_get(struct ceph_connection *con)
4829 struct ceph_mds_session *s = con->private;
4831 if (ceph_get_mds_session(s))
4836 static void con_put(struct ceph_connection *con)
4838 struct ceph_mds_session *s = con->private;
4840 ceph_put_mds_session(s);
4844 * if the client is unresponsive for long enough, the mds will kill
4845 * the session entirely.
4847 static void peer_reset(struct ceph_connection *con)
4849 struct ceph_mds_session *s = con->private;
4850 struct ceph_mds_client *mdsc = s->s_mdsc;
4852 pr_warn("mds%d closed our session\n", s->s_mds);
4853 send_mds_reconnect(mdsc, s);
4856 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4858 struct ceph_mds_session *s = con->private;
4859 struct ceph_mds_client *mdsc = s->s_mdsc;
4860 int type = le16_to_cpu(msg->hdr.type);
4862 mutex_lock(&mdsc->mutex);
4863 if (__verify_registered_session(mdsc, s) < 0) {
4864 mutex_unlock(&mdsc->mutex);
4867 mutex_unlock(&mdsc->mutex);
4870 case CEPH_MSG_MDS_MAP:
4871 ceph_mdsc_handle_mdsmap(mdsc, msg);
4873 case CEPH_MSG_FS_MAP_USER:
4874 ceph_mdsc_handle_fsmap(mdsc, msg);
4876 case CEPH_MSG_CLIENT_SESSION:
4877 handle_session(s, msg);
4879 case CEPH_MSG_CLIENT_REPLY:
4880 handle_reply(s, msg);
4882 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4883 handle_forward(mdsc, s, msg);
4885 case CEPH_MSG_CLIENT_CAPS:
4886 ceph_handle_caps(s, msg);
4888 case CEPH_MSG_CLIENT_SNAP:
4889 ceph_handle_snap(mdsc, s, msg);
4891 case CEPH_MSG_CLIENT_LEASE:
4892 handle_lease(mdsc, s, msg);
4894 case CEPH_MSG_CLIENT_QUOTA:
4895 ceph_handle_quota(mdsc, s, msg);
4899 pr_err("received unknown message type %d %s\n", type,
4900 ceph_msg_type_name(type));
4911 * Note: returned pointer is the address of a structure that's
4912 * managed separately. Caller must *not* attempt to free it.
4914 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4915 int *proto, int force_new)
4917 struct ceph_mds_session *s = con->private;
4918 struct ceph_mds_client *mdsc = s->s_mdsc;
4919 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4920 struct ceph_auth_handshake *auth = &s->s_auth;
4922 if (force_new && auth->authorizer) {
4923 ceph_auth_destroy_authorizer(auth->authorizer);
4924 auth->authorizer = NULL;
4926 if (!auth->authorizer) {
4927 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4930 return ERR_PTR(ret);
4932 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4935 return ERR_PTR(ret);
4937 *proto = ac->protocol;
4942 static int add_authorizer_challenge(struct ceph_connection *con,
4943 void *challenge_buf, int challenge_buf_len)
4945 struct ceph_mds_session *s = con->private;
4946 struct ceph_mds_client *mdsc = s->s_mdsc;
4947 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4949 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4950 challenge_buf, challenge_buf_len);
4953 static int verify_authorizer_reply(struct ceph_connection *con)
4955 struct ceph_mds_session *s = con->private;
4956 struct ceph_mds_client *mdsc = s->s_mdsc;
4957 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4959 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4962 static int invalidate_authorizer(struct ceph_connection *con)
4964 struct ceph_mds_session *s = con->private;
4965 struct ceph_mds_client *mdsc = s->s_mdsc;
4966 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4968 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4970 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4973 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4974 struct ceph_msg_header *hdr, int *skip)
4976 struct ceph_msg *msg;
4977 int type = (int) le16_to_cpu(hdr->type);
4978 int front_len = (int) le32_to_cpu(hdr->front_len);
4984 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4986 pr_err("unable to allocate msg type %d len %d\n",
4994 static int mds_sign_message(struct ceph_msg *msg)
4996 struct ceph_mds_session *s = msg->con->private;
4997 struct ceph_auth_handshake *auth = &s->s_auth;
4999 return ceph_auth_sign_message(auth, msg);
5002 static int mds_check_message_signature(struct ceph_msg *msg)
5004 struct ceph_mds_session *s = msg->con->private;
5005 struct ceph_auth_handshake *auth = &s->s_auth;
5007 return ceph_auth_check_message_signature(auth, msg);
5010 static const struct ceph_connection_operations mds_con_ops = {
5013 .dispatch = dispatch,
5014 .get_authorizer = get_authorizer,
5015 .add_authorizer_challenge = add_authorizer_challenge,
5016 .verify_authorizer_reply = verify_authorizer_reply,
5017 .invalidate_authorizer = invalidate_authorizer,
5018 .peer_reset = peer_reset,
5019 .alloc_msg = mds_alloc_msg,
5020 .sign_message = mds_sign_message,
5021 .check_message_signature = mds_check_message_signature,