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>
14 #include "mds_client.h"
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
23 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
26 * A cluster of MDS (metadata server) daemons is responsible for
27 * managing the file system namespace (the directory hierarchy and
28 * inodes) and for coordinating shared access to storage. Metadata is
29 * partitioning hierarchically across a number of servers, and that
30 * partition varies over time as the cluster adjusts the distribution
31 * in order to balance load.
33 * The MDS client is primarily responsible to managing synchronous
34 * metadata requests for operations like open, unlink, and so forth.
35 * If there is a MDS failure, we find out about it when we (possibly
36 * request and) receive a new MDS map, and can resubmit affected
39 * For the most part, though, we take advantage of a lossless
40 * communications channel to the MDS, and do not need to worry about
41 * timing out or resubmitting requests.
43 * We maintain a stateful "session" with each MDS we interact with.
44 * Within each session, we sent periodic heartbeat messages to ensure
45 * any capabilities or leases we have been issues remain valid. If
46 * the session times out and goes stale, our leases and capabilities
47 * are no longer valid.
50 struct ceph_reconnect_state {
51 struct ceph_mds_session *session;
52 int nr_caps, nr_realms;
53 struct ceph_pagelist *pagelist;
58 static void __wake_requests(struct ceph_mds_client *mdsc,
59 struct list_head *head);
61 static const struct ceph_connection_operations mds_con_ops;
69 * parse individual inode info
71 static int parse_reply_info_in(void **p, void *end,
72 struct ceph_mds_reply_info_in *info,
78 *p += sizeof(struct ceph_mds_reply_inode) +
79 sizeof(*info->in->fragtree.splits) *
80 le32_to_cpu(info->in->fragtree.nsplits);
82 ceph_decode_32_safe(p, end, info->symlink_len, bad);
83 ceph_decode_need(p, end, info->symlink_len, bad);
85 *p += info->symlink_len;
87 ceph_decode_copy_safe(p, end, &info->dir_layout,
88 sizeof(info->dir_layout), bad);
89 ceph_decode_32_safe(p, end, info->xattr_len, bad);
90 ceph_decode_need(p, end, info->xattr_len, bad);
91 info->xattr_data = *p;
92 *p += info->xattr_len;
94 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
95 ceph_decode_64_safe(p, end, info->inline_version, bad);
96 ceph_decode_32_safe(p, end, info->inline_len, bad);
97 ceph_decode_need(p, end, info->inline_len, bad);
98 info->inline_data = *p;
99 *p += info->inline_len;
101 info->inline_version = CEPH_INLINE_NONE;
103 if (features & CEPH_FEATURE_MDS_QUOTA) {
104 u8 struct_v, struct_compat;
108 * both struct_v and struct_compat are expected to be >= 1
110 ceph_decode_8_safe(p, end, struct_v, bad);
111 ceph_decode_8_safe(p, end, struct_compat, bad);
112 if (!struct_v || !struct_compat)
114 ceph_decode_32_safe(p, end, struct_len, bad);
115 ceph_decode_need(p, end, struct_len, bad);
116 ceph_decode_64_safe(p, end, info->max_bytes, bad);
117 ceph_decode_64_safe(p, end, info->max_files, bad);
123 info->pool_ns_len = 0;
124 info->pool_ns_data = NULL;
125 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
126 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
127 if (info->pool_ns_len > 0) {
128 ceph_decode_need(p, end, info->pool_ns_len, bad);
129 info->pool_ns_data = *p;
130 *p += info->pool_ns_len;
140 * parse a normal reply, which may contain a (dir+)dentry and/or a
143 static int parse_reply_info_trace(void **p, void *end,
144 struct ceph_mds_reply_info_parsed *info,
149 if (info->head->is_dentry) {
150 err = parse_reply_info_in(p, end, &info->diri, features);
154 if (unlikely(*p + sizeof(*info->dirfrag) > end))
157 *p += sizeof(*info->dirfrag) +
158 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
159 if (unlikely(*p > end))
162 ceph_decode_32_safe(p, end, info->dname_len, bad);
163 ceph_decode_need(p, end, info->dname_len, bad);
165 *p += info->dname_len;
167 *p += sizeof(*info->dlease);
170 if (info->head->is_target) {
171 err = parse_reply_info_in(p, end, &info->targeti, features);
176 if (unlikely(*p != end))
183 pr_err("problem parsing mds trace %d\n", err);
188 * parse readdir results
190 static int parse_reply_info_dir(void **p, void *end,
191 struct ceph_mds_reply_info_parsed *info,
198 if (*p + sizeof(*info->dir_dir) > end)
200 *p += sizeof(*info->dir_dir) +
201 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
205 ceph_decode_need(p, end, sizeof(num) + 2, bad);
206 num = ceph_decode_32(p);
208 u16 flags = ceph_decode_16(p);
209 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
210 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
211 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
212 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
217 BUG_ON(!info->dir_entries);
218 if ((unsigned long)(info->dir_entries + num) >
219 (unsigned long)info->dir_entries + info->dir_buf_size) {
220 pr_err("dir contents are larger than expected\n");
227 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
229 ceph_decode_need(p, end, sizeof(u32)*2, bad);
230 rde->name_len = ceph_decode_32(p);
231 ceph_decode_need(p, end, rde->name_len, bad);
234 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
236 *p += sizeof(struct ceph_mds_reply_lease);
239 err = parse_reply_info_in(p, end, &rde->inode, features);
242 /* ceph_readdir_prepopulate() will update it */
256 pr_err("problem parsing dir contents %d\n", err);
261 * parse fcntl F_GETLK results
263 static int parse_reply_info_filelock(void **p, void *end,
264 struct ceph_mds_reply_info_parsed *info,
267 if (*p + sizeof(*info->filelock_reply) > end)
270 info->filelock_reply = *p;
271 *p += sizeof(*info->filelock_reply);
273 if (unlikely(*p != end))
282 * parse create results
284 static int parse_reply_info_create(void **p, void *end,
285 struct ceph_mds_reply_info_parsed *info,
288 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
290 info->has_create_ino = false;
292 info->has_create_ino = true;
293 info->ino = ceph_decode_64(p);
297 if (unlikely(*p != end))
306 * parse extra results
308 static int parse_reply_info_extra(void **p, void *end,
309 struct ceph_mds_reply_info_parsed *info,
312 u32 op = le32_to_cpu(info->head->op);
314 if (op == CEPH_MDS_OP_GETFILELOCK)
315 return parse_reply_info_filelock(p, end, info, features);
316 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
317 return parse_reply_info_dir(p, end, info, features);
318 else if (op == CEPH_MDS_OP_CREATE)
319 return parse_reply_info_create(p, end, info, features);
325 * parse entire mds reply
327 static int parse_reply_info(struct ceph_msg *msg,
328 struct ceph_mds_reply_info_parsed *info,
335 info->head = msg->front.iov_base;
336 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
337 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
340 ceph_decode_32_safe(&p, end, len, bad);
342 ceph_decode_need(&p, end, len, bad);
343 err = parse_reply_info_trace(&p, p+len, info, features);
349 ceph_decode_32_safe(&p, end, len, bad);
351 ceph_decode_need(&p, end, len, bad);
352 err = parse_reply_info_extra(&p, p+len, info, features);
358 ceph_decode_32_safe(&p, end, len, bad);
359 info->snapblob_len = len;
370 pr_err("mds parse_reply err %d\n", err);
374 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
376 if (!info->dir_entries)
378 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
385 const char *ceph_session_state_name(int s)
388 case CEPH_MDS_SESSION_NEW: return "new";
389 case CEPH_MDS_SESSION_OPENING: return "opening";
390 case CEPH_MDS_SESSION_OPEN: return "open";
391 case CEPH_MDS_SESSION_HUNG: return "hung";
392 case CEPH_MDS_SESSION_CLOSING: return "closing";
393 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
394 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
395 case CEPH_MDS_SESSION_REJECTED: return "rejected";
396 default: return "???";
400 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
402 if (refcount_inc_not_zero(&s->s_ref)) {
403 dout("mdsc get_session %p %d -> %d\n", s,
404 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
407 dout("mdsc get_session %p 0 -- FAIL\n", s);
412 void ceph_put_mds_session(struct ceph_mds_session *s)
414 dout("mdsc put_session %p %d -> %d\n", s,
415 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
416 if (refcount_dec_and_test(&s->s_ref)) {
417 if (s->s_auth.authorizer)
418 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
424 * called under mdsc->mutex
426 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
429 struct ceph_mds_session *session;
431 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
433 session = mdsc->sessions[mds];
434 dout("lookup_mds_session %p %d\n", session,
435 refcount_read(&session->s_ref));
436 get_session(session);
440 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
442 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
448 static int __verify_registered_session(struct ceph_mds_client *mdsc,
449 struct ceph_mds_session *s)
451 if (s->s_mds >= mdsc->max_sessions ||
452 mdsc->sessions[s->s_mds] != s)
458 * create+register a new session for given mds.
459 * called under mdsc->mutex.
461 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
464 struct ceph_mds_session *s;
466 if (mds >= mdsc->mdsmap->m_num_mds)
467 return ERR_PTR(-EINVAL);
469 s = kzalloc(sizeof(*s), GFP_NOFS);
471 return ERR_PTR(-ENOMEM);
473 if (mds >= mdsc->max_sessions) {
474 int newmax = 1 << get_count_order(mds + 1);
475 struct ceph_mds_session **sa;
477 dout("%s: realloc to %d\n", __func__, newmax);
478 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
481 if (mdsc->sessions) {
482 memcpy(sa, mdsc->sessions,
483 mdsc->max_sessions * sizeof(void *));
484 kfree(mdsc->sessions);
487 mdsc->max_sessions = newmax;
490 dout("%s: mds%d\n", __func__, mds);
493 s->s_state = CEPH_MDS_SESSION_NEW;
496 mutex_init(&s->s_mutex);
498 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
500 spin_lock_init(&s->s_gen_ttl_lock);
502 s->s_cap_ttl = jiffies - 1;
504 spin_lock_init(&s->s_cap_lock);
505 s->s_renew_requested = 0;
507 INIT_LIST_HEAD(&s->s_caps);
510 refcount_set(&s->s_ref, 1);
511 INIT_LIST_HEAD(&s->s_waiting);
512 INIT_LIST_HEAD(&s->s_unsafe);
513 s->s_num_cap_releases = 0;
514 s->s_cap_reconnect = 0;
515 s->s_cap_iterator = NULL;
516 INIT_LIST_HEAD(&s->s_cap_releases);
517 INIT_LIST_HEAD(&s->s_cap_flushing);
519 mdsc->sessions[mds] = s;
520 atomic_inc(&mdsc->num_sessions);
521 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
523 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
524 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
530 return ERR_PTR(-ENOMEM);
534 * called under mdsc->mutex
536 static void __unregister_session(struct ceph_mds_client *mdsc,
537 struct ceph_mds_session *s)
539 dout("__unregister_session mds%d %p\n", s->s_mds, s);
540 BUG_ON(mdsc->sessions[s->s_mds] != s);
541 mdsc->sessions[s->s_mds] = NULL;
542 ceph_con_close(&s->s_con);
543 ceph_put_mds_session(s);
544 atomic_dec(&mdsc->num_sessions);
548 * drop session refs in request.
550 * should be last request ref, or hold mdsc->mutex
552 static void put_request_session(struct ceph_mds_request *req)
554 if (req->r_session) {
555 ceph_put_mds_session(req->r_session);
556 req->r_session = NULL;
560 void ceph_mdsc_release_request(struct kref *kref)
562 struct ceph_mds_request *req = container_of(kref,
563 struct ceph_mds_request,
565 destroy_reply_info(&req->r_reply_info);
567 ceph_msg_put(req->r_request);
569 ceph_msg_put(req->r_reply);
571 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
575 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
576 iput(req->r_target_inode);
579 if (req->r_old_dentry)
580 dput(req->r_old_dentry);
581 if (req->r_old_dentry_dir) {
583 * track (and drop pins for) r_old_dentry_dir
584 * separately, since r_old_dentry's d_parent may have
585 * changed between the dir mutex being dropped and
586 * this request being freed.
588 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
590 iput(req->r_old_dentry_dir);
595 ceph_pagelist_release(req->r_pagelist);
596 put_request_session(req);
597 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
601 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
604 * lookup session, bump ref if found.
606 * called under mdsc->mutex.
608 static struct ceph_mds_request *
609 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
611 struct ceph_mds_request *req;
613 req = lookup_request(&mdsc->request_tree, tid);
615 ceph_mdsc_get_request(req);
621 * Register an in-flight request, and assign a tid. Link to directory
622 * are modifying (if any).
624 * Called under mdsc->mutex.
626 static void __register_request(struct ceph_mds_client *mdsc,
627 struct ceph_mds_request *req,
632 req->r_tid = ++mdsc->last_tid;
633 if (req->r_num_caps) {
634 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
637 pr_err("__register_request %p "
638 "failed to reserve caps: %d\n", req, ret);
639 /* set req->r_err to fail early from __do_request */
644 dout("__register_request %p tid %lld\n", req, req->r_tid);
645 ceph_mdsc_get_request(req);
646 insert_request(&mdsc->request_tree, req);
648 req->r_uid = current_fsuid();
649 req->r_gid = current_fsgid();
651 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
652 mdsc->oldest_tid = req->r_tid;
656 req->r_unsafe_dir = dir;
660 static void __unregister_request(struct ceph_mds_client *mdsc,
661 struct ceph_mds_request *req)
663 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
665 /* Never leave an unregistered request on an unsafe list! */
666 list_del_init(&req->r_unsafe_item);
668 if (req->r_tid == mdsc->oldest_tid) {
669 struct rb_node *p = rb_next(&req->r_node);
670 mdsc->oldest_tid = 0;
672 struct ceph_mds_request *next_req =
673 rb_entry(p, struct ceph_mds_request, r_node);
674 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
675 mdsc->oldest_tid = next_req->r_tid;
682 erase_request(&mdsc->request_tree, req);
684 if (req->r_unsafe_dir &&
685 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
686 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
687 spin_lock(&ci->i_unsafe_lock);
688 list_del_init(&req->r_unsafe_dir_item);
689 spin_unlock(&ci->i_unsafe_lock);
691 if (req->r_target_inode &&
692 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
693 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
694 spin_lock(&ci->i_unsafe_lock);
695 list_del_init(&req->r_unsafe_target_item);
696 spin_unlock(&ci->i_unsafe_lock);
699 if (req->r_unsafe_dir) {
700 iput(req->r_unsafe_dir);
701 req->r_unsafe_dir = NULL;
704 complete_all(&req->r_safe_completion);
706 ceph_mdsc_put_request(req);
710 * Walk back up the dentry tree until we hit a dentry representing a
711 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
712 * when calling this) to ensure that the objects won't disappear while we're
713 * working with them. Once we hit a candidate dentry, we attempt to take a
714 * reference to it, and return that as the result.
716 static struct inode *get_nonsnap_parent(struct dentry *dentry)
718 struct inode *inode = NULL;
720 while (dentry && !IS_ROOT(dentry)) {
721 inode = d_inode_rcu(dentry);
722 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
724 dentry = dentry->d_parent;
727 inode = igrab(inode);
732 * Choose mds to send request to next. If there is a hint set in the
733 * request (e.g., due to a prior forward hint from the mds), use that.
734 * Otherwise, consult frag tree and/or caps to identify the
735 * appropriate mds. If all else fails, choose randomly.
737 * Called under mdsc->mutex.
739 static int __choose_mds(struct ceph_mds_client *mdsc,
740 struct ceph_mds_request *req)
743 struct ceph_inode_info *ci;
744 struct ceph_cap *cap;
745 int mode = req->r_direct_mode;
747 u32 hash = req->r_direct_hash;
748 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
751 * is there a specific mds we should try? ignore hint if we have
752 * no session and the mds is not up (active or recovering).
754 if (req->r_resend_mds >= 0 &&
755 (__have_session(mdsc, req->r_resend_mds) ||
756 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
757 dout("choose_mds using resend_mds mds%d\n",
759 return req->r_resend_mds;
762 if (mode == USE_RANDOM_MDS)
767 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
768 inode = req->r_inode;
771 /* req->r_dentry is non-null for LSSNAP request */
773 inode = get_nonsnap_parent(req->r_dentry);
775 dout("__choose_mds using snapdir's parent %p\n", inode);
777 } else if (req->r_dentry) {
778 /* ignore race with rename; old or new d_parent is okay */
779 struct dentry *parent;
783 parent = req->r_dentry->d_parent;
784 dir = req->r_parent ? : d_inode_rcu(parent);
786 if (!dir || dir->i_sb != mdsc->fsc->sb) {
787 /* not this fs or parent went negative */
788 inode = d_inode(req->r_dentry);
791 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
792 /* direct snapped/virtual snapdir requests
793 * based on parent dir inode */
794 inode = get_nonsnap_parent(parent);
795 dout("__choose_mds using nonsnap parent %p\n", inode);
798 inode = d_inode(req->r_dentry);
799 if (!inode || mode == USE_AUTH_MDS) {
802 hash = ceph_dentry_hash(dir, req->r_dentry);
811 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
815 ci = ceph_inode(inode);
817 if (is_hash && S_ISDIR(inode->i_mode)) {
818 struct ceph_inode_frag frag;
821 ceph_choose_frag(ci, hash, &frag, &found);
823 if (mode == USE_ANY_MDS && frag.ndist > 0) {
826 /* choose a random replica */
827 get_random_bytes(&r, 1);
830 dout("choose_mds %p %llx.%llx "
831 "frag %u mds%d (%d/%d)\n",
832 inode, ceph_vinop(inode),
835 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
836 CEPH_MDS_STATE_ACTIVE)
840 /* since this file/dir wasn't known to be
841 * replicated, then we want to look for the
842 * authoritative mds. */
845 /* choose auth mds */
847 dout("choose_mds %p %llx.%llx "
848 "frag %u mds%d (auth)\n",
849 inode, ceph_vinop(inode), frag.frag, mds);
850 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
851 CEPH_MDS_STATE_ACTIVE)
857 spin_lock(&ci->i_ceph_lock);
859 if (mode == USE_AUTH_MDS)
860 cap = ci->i_auth_cap;
861 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
862 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
864 spin_unlock(&ci->i_ceph_lock);
868 mds = cap->session->s_mds;
869 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
870 inode, ceph_vinop(inode), mds,
871 cap == ci->i_auth_cap ? "auth " : "", cap);
872 spin_unlock(&ci->i_ceph_lock);
878 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
879 dout("choose_mds chose random mds%d\n", mds);
887 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
889 struct ceph_msg *msg;
890 struct ceph_mds_session_head *h;
892 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
895 pr_err("create_session_msg ENOMEM creating msg\n");
898 h = msg->front.iov_base;
899 h->op = cpu_to_le32(op);
900 h->seq = cpu_to_le64(seq);
905 static void encode_supported_features(void **p, void *end)
907 static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
908 static const size_t count = ARRAY_SIZE(bits);
912 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
914 BUG_ON(*p + 4 + size > end);
915 ceph_encode_32(p, size);
917 for (i = 0; i < count; i++)
918 ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
921 BUG_ON(*p + 4 > end);
922 ceph_encode_32(p, 0);
927 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
928 * to include additional client metadata fields.
930 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
932 struct ceph_msg *msg;
933 struct ceph_mds_session_head *h;
936 int metadata_key_count = 0;
937 struct ceph_options *opt = mdsc->fsc->client->options;
938 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
941 const char* metadata[][2] = {
942 {"hostname", mdsc->nodename},
943 {"kernel_version", init_utsname()->release},
944 {"entity_id", opt->name ? : ""},
945 {"root", fsopt->server_path ? : "/"},
949 /* Calculate serialized length of metadata */
950 extra_bytes = 4; /* map length */
951 for (i = 0; metadata[i][0]; ++i) {
952 extra_bytes += 8 + strlen(metadata[i][0]) +
953 strlen(metadata[i][1]);
954 metadata_key_count++;
956 /* supported feature */
957 extra_bytes += 4 + 8;
959 /* Allocate the message */
960 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
963 pr_err("create_session_msg ENOMEM creating msg\n");
966 p = msg->front.iov_base;
967 end = p + msg->front.iov_len;
970 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
971 h->seq = cpu_to_le64(seq);
974 * Serialize client metadata into waiting buffer space, using
975 * the format that userspace expects for map<string, string>
977 * ClientSession messages with metadata are v2
979 msg->hdr.version = cpu_to_le16(3);
980 msg->hdr.compat_version = cpu_to_le16(1);
982 /* The write pointer, following the session_head structure */
985 /* Number of entries in the map */
986 ceph_encode_32(&p, metadata_key_count);
988 /* Two length-prefixed strings for each entry in the map */
989 for (i = 0; metadata[i][0]; ++i) {
990 size_t const key_len = strlen(metadata[i][0]);
991 size_t const val_len = strlen(metadata[i][1]);
993 ceph_encode_32(&p, key_len);
994 memcpy(p, metadata[i][0], key_len);
996 ceph_encode_32(&p, val_len);
997 memcpy(p, metadata[i][1], val_len);
1001 encode_supported_features(&p, end);
1002 msg->front.iov_len = p - msg->front.iov_base;
1003 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1009 * send session open request.
1011 * called under mdsc->mutex
1013 static int __open_session(struct ceph_mds_client *mdsc,
1014 struct ceph_mds_session *session)
1016 struct ceph_msg *msg;
1018 int mds = session->s_mds;
1020 /* wait for mds to go active? */
1021 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1022 dout("open_session to mds%d (%s)\n", mds,
1023 ceph_mds_state_name(mstate));
1024 session->s_state = CEPH_MDS_SESSION_OPENING;
1025 session->s_renew_requested = jiffies;
1027 /* send connect message */
1028 msg = create_session_open_msg(mdsc, session->s_seq);
1031 ceph_con_send(&session->s_con, msg);
1036 * open sessions for any export targets for the given mds
1038 * called under mdsc->mutex
1040 static struct ceph_mds_session *
1041 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1043 struct ceph_mds_session *session;
1045 session = __ceph_lookup_mds_session(mdsc, target);
1047 session = register_session(mdsc, target);
1048 if (IS_ERR(session))
1051 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1052 session->s_state == CEPH_MDS_SESSION_CLOSING)
1053 __open_session(mdsc, session);
1058 struct ceph_mds_session *
1059 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1061 struct ceph_mds_session *session;
1063 dout("open_export_target_session to mds%d\n", target);
1065 mutex_lock(&mdsc->mutex);
1066 session = __open_export_target_session(mdsc, target);
1067 mutex_unlock(&mdsc->mutex);
1072 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1073 struct ceph_mds_session *session)
1075 struct ceph_mds_info *mi;
1076 struct ceph_mds_session *ts;
1077 int i, mds = session->s_mds;
1079 if (mds >= mdsc->mdsmap->m_num_mds)
1082 mi = &mdsc->mdsmap->m_info[mds];
1083 dout("open_export_target_sessions for mds%d (%d targets)\n",
1084 session->s_mds, mi->num_export_targets);
1086 for (i = 0; i < mi->num_export_targets; i++) {
1087 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1089 ceph_put_mds_session(ts);
1093 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1094 struct ceph_mds_session *session)
1096 mutex_lock(&mdsc->mutex);
1097 __open_export_target_sessions(mdsc, session);
1098 mutex_unlock(&mdsc->mutex);
1105 static void detach_cap_releases(struct ceph_mds_session *session,
1106 struct list_head *target)
1108 lockdep_assert_held(&session->s_cap_lock);
1110 list_splice_init(&session->s_cap_releases, target);
1111 session->s_num_cap_releases = 0;
1112 dout("dispose_cap_releases mds%d\n", session->s_mds);
1115 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1116 struct list_head *dispose)
1118 while (!list_empty(dispose)) {
1119 struct ceph_cap *cap;
1120 /* zero out the in-progress message */
1121 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1122 list_del(&cap->session_caps);
1123 ceph_put_cap(mdsc, cap);
1127 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1128 struct ceph_mds_session *session)
1130 struct ceph_mds_request *req;
1133 dout("cleanup_session_requests mds%d\n", session->s_mds);
1134 mutex_lock(&mdsc->mutex);
1135 while (!list_empty(&session->s_unsafe)) {
1136 req = list_first_entry(&session->s_unsafe,
1137 struct ceph_mds_request, r_unsafe_item);
1138 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1140 __unregister_request(mdsc, req);
1142 /* zero r_attempts, so kick_requests() will re-send requests */
1143 p = rb_first(&mdsc->request_tree);
1145 req = rb_entry(p, struct ceph_mds_request, r_node);
1147 if (req->r_session &&
1148 req->r_session->s_mds == session->s_mds)
1149 req->r_attempts = 0;
1151 mutex_unlock(&mdsc->mutex);
1155 * Helper to safely iterate over all caps associated with a session, with
1156 * special care taken to handle a racing __ceph_remove_cap().
1158 * Caller must hold session s_mutex.
1160 static int iterate_session_caps(struct ceph_mds_session *session,
1161 int (*cb)(struct inode *, struct ceph_cap *,
1164 struct list_head *p;
1165 struct ceph_cap *cap;
1166 struct inode *inode, *last_inode = NULL;
1167 struct ceph_cap *old_cap = NULL;
1170 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1171 spin_lock(&session->s_cap_lock);
1172 p = session->s_caps.next;
1173 while (p != &session->s_caps) {
1174 cap = list_entry(p, struct ceph_cap, session_caps);
1175 inode = igrab(&cap->ci->vfs_inode);
1180 session->s_cap_iterator = cap;
1181 spin_unlock(&session->s_cap_lock);
1188 ceph_put_cap(session->s_mdsc, old_cap);
1192 ret = cb(inode, cap, arg);
1195 spin_lock(&session->s_cap_lock);
1198 dout("iterate_session_caps finishing cap %p removal\n",
1200 BUG_ON(cap->session != session);
1201 cap->session = NULL;
1202 list_del_init(&cap->session_caps);
1203 session->s_nr_caps--;
1204 if (cap->queue_release) {
1205 list_add_tail(&cap->session_caps,
1206 &session->s_cap_releases);
1207 session->s_num_cap_releases++;
1209 old_cap = cap; /* put_cap it w/o locks held */
1217 session->s_cap_iterator = NULL;
1218 spin_unlock(&session->s_cap_lock);
1222 ceph_put_cap(session->s_mdsc, old_cap);
1227 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1230 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1231 struct ceph_inode_info *ci = ceph_inode(inode);
1232 LIST_HEAD(to_remove);
1234 bool invalidate = false;
1236 dout("removing cap %p, ci is %p, inode is %p\n",
1237 cap, ci, &ci->vfs_inode);
1238 spin_lock(&ci->i_ceph_lock);
1239 if (cap->mds_wanted | cap->issued)
1240 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1241 __ceph_remove_cap(cap, false);
1242 if (!ci->i_auth_cap) {
1243 struct ceph_cap_flush *cf;
1244 struct ceph_mds_client *mdsc = fsc->mdsc;
1246 if (ci->i_wrbuffer_ref > 0 &&
1247 READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1250 while (!list_empty(&ci->i_cap_flush_list)) {
1251 cf = list_first_entry(&ci->i_cap_flush_list,
1252 struct ceph_cap_flush, i_list);
1253 list_move(&cf->i_list, &to_remove);
1256 spin_lock(&mdsc->cap_dirty_lock);
1258 list_for_each_entry(cf, &to_remove, i_list)
1259 list_del(&cf->g_list);
1261 if (!list_empty(&ci->i_dirty_item)) {
1262 pr_warn_ratelimited(
1263 " dropping dirty %s state for %p %lld\n",
1264 ceph_cap_string(ci->i_dirty_caps),
1265 inode, ceph_ino(inode));
1266 ci->i_dirty_caps = 0;
1267 list_del_init(&ci->i_dirty_item);
1270 if (!list_empty(&ci->i_flushing_item)) {
1271 pr_warn_ratelimited(
1272 " dropping dirty+flushing %s state for %p %lld\n",
1273 ceph_cap_string(ci->i_flushing_caps),
1274 inode, ceph_ino(inode));
1275 ci->i_flushing_caps = 0;
1276 list_del_init(&ci->i_flushing_item);
1277 mdsc->num_cap_flushing--;
1280 spin_unlock(&mdsc->cap_dirty_lock);
1282 if (atomic_read(&ci->i_filelock_ref) > 0) {
1283 /* make further file lock syscall return -EIO */
1284 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1285 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1286 inode, ceph_ino(inode));
1289 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1290 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1291 ci->i_prealloc_cap_flush = NULL;
1294 spin_unlock(&ci->i_ceph_lock);
1295 while (!list_empty(&to_remove)) {
1296 struct ceph_cap_flush *cf;
1297 cf = list_first_entry(&to_remove,
1298 struct ceph_cap_flush, i_list);
1299 list_del(&cf->i_list);
1300 ceph_free_cap_flush(cf);
1303 wake_up_all(&ci->i_cap_wq);
1305 ceph_queue_invalidate(inode);
1312 * caller must hold session s_mutex
1314 static void remove_session_caps(struct ceph_mds_session *session)
1316 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1317 struct super_block *sb = fsc->sb;
1320 dout("remove_session_caps on %p\n", session);
1321 iterate_session_caps(session, remove_session_caps_cb, fsc);
1323 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1325 spin_lock(&session->s_cap_lock);
1326 if (session->s_nr_caps > 0) {
1327 struct inode *inode;
1328 struct ceph_cap *cap, *prev = NULL;
1329 struct ceph_vino vino;
1331 * iterate_session_caps() skips inodes that are being
1332 * deleted, we need to wait until deletions are complete.
1333 * __wait_on_freeing_inode() is designed for the job,
1334 * but it is not exported, so use lookup inode function
1337 while (!list_empty(&session->s_caps)) {
1338 cap = list_entry(session->s_caps.next,
1339 struct ceph_cap, session_caps);
1343 vino = cap->ci->i_vino;
1344 spin_unlock(&session->s_cap_lock);
1346 inode = ceph_find_inode(sb, vino);
1349 spin_lock(&session->s_cap_lock);
1353 // drop cap expires and unlock s_cap_lock
1354 detach_cap_releases(session, &dispose);
1356 BUG_ON(session->s_nr_caps > 0);
1357 BUG_ON(!list_empty(&session->s_cap_flushing));
1358 spin_unlock(&session->s_cap_lock);
1359 dispose_cap_releases(session->s_mdsc, &dispose);
1369 * wake up any threads waiting on this session's caps. if the cap is
1370 * old (didn't get renewed on the client reconnect), remove it now.
1372 * caller must hold s_mutex.
1374 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1377 struct ceph_inode_info *ci = ceph_inode(inode);
1378 unsigned long ev = (unsigned long)arg;
1380 if (ev == RECONNECT) {
1381 spin_lock(&ci->i_ceph_lock);
1382 ci->i_wanted_max_size = 0;
1383 ci->i_requested_max_size = 0;
1384 spin_unlock(&ci->i_ceph_lock);
1385 } else if (ev == RENEWCAPS) {
1386 if (cap->cap_gen < cap->session->s_cap_gen) {
1387 /* mds did not re-issue stale cap */
1388 spin_lock(&ci->i_ceph_lock);
1389 cap->issued = cap->implemented = CEPH_CAP_PIN;
1390 /* make sure mds knows what we want */
1391 if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1392 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1393 spin_unlock(&ci->i_ceph_lock);
1395 } else if (ev == FORCE_RO) {
1397 wake_up_all(&ci->i_cap_wq);
1401 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1403 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1404 iterate_session_caps(session, wake_up_session_cb,
1405 (void *)(unsigned long)ev);
1409 * Send periodic message to MDS renewing all currently held caps. The
1410 * ack will reset the expiration for all caps from this session.
1412 * caller holds s_mutex
1414 static int send_renew_caps(struct ceph_mds_client *mdsc,
1415 struct ceph_mds_session *session)
1417 struct ceph_msg *msg;
1420 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1421 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1422 pr_info("mds%d caps stale\n", session->s_mds);
1423 session->s_renew_requested = jiffies;
1425 /* do not try to renew caps until a recovering mds has reconnected
1426 * with its clients. */
1427 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1428 if (state < CEPH_MDS_STATE_RECONNECT) {
1429 dout("send_renew_caps ignoring mds%d (%s)\n",
1430 session->s_mds, ceph_mds_state_name(state));
1434 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1435 ceph_mds_state_name(state));
1436 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1437 ++session->s_renew_seq);
1440 ceph_con_send(&session->s_con, msg);
1444 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1445 struct ceph_mds_session *session, u64 seq)
1447 struct ceph_msg *msg;
1449 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1450 session->s_mds, ceph_session_state_name(session->s_state), seq);
1451 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1454 ceph_con_send(&session->s_con, msg);
1460 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1462 * Called under session->s_mutex
1464 static void renewed_caps(struct ceph_mds_client *mdsc,
1465 struct ceph_mds_session *session, int is_renew)
1470 spin_lock(&session->s_cap_lock);
1471 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1473 session->s_cap_ttl = session->s_renew_requested +
1474 mdsc->mdsmap->m_session_timeout*HZ;
1477 if (time_before(jiffies, session->s_cap_ttl)) {
1478 pr_info("mds%d caps renewed\n", session->s_mds);
1481 pr_info("mds%d caps still stale\n", session->s_mds);
1484 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1485 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1486 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1487 spin_unlock(&session->s_cap_lock);
1490 wake_up_session_caps(session, RENEWCAPS);
1494 * send a session close request
1496 static int request_close_session(struct ceph_mds_client *mdsc,
1497 struct ceph_mds_session *session)
1499 struct ceph_msg *msg;
1501 dout("request_close_session mds%d state %s seq %lld\n",
1502 session->s_mds, ceph_session_state_name(session->s_state),
1504 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1507 ceph_con_send(&session->s_con, msg);
1512 * Called with s_mutex held.
1514 static int __close_session(struct ceph_mds_client *mdsc,
1515 struct ceph_mds_session *session)
1517 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1519 session->s_state = CEPH_MDS_SESSION_CLOSING;
1520 return request_close_session(mdsc, session);
1523 static bool drop_negative_children(struct dentry *dentry)
1525 struct dentry *child;
1526 bool all_negative = true;
1528 if (!d_is_dir(dentry))
1531 spin_lock(&dentry->d_lock);
1532 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1533 if (d_really_is_positive(child)) {
1534 all_negative = false;
1538 spin_unlock(&dentry->d_lock);
1541 shrink_dcache_parent(dentry);
1543 return all_negative;
1547 * Trim old(er) caps.
1549 * Because we can't cache an inode without one or more caps, we do
1550 * this indirectly: if a cap is unused, we prune its aliases, at which
1551 * point the inode will hopefully get dropped to.
1553 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1554 * memory pressure from the MDS, though, so it needn't be perfect.
1556 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1558 struct ceph_mds_session *session = arg;
1559 struct ceph_inode_info *ci = ceph_inode(inode);
1560 int used, wanted, oissued, mine;
1562 if (session->s_trim_caps <= 0)
1565 spin_lock(&ci->i_ceph_lock);
1566 mine = cap->issued | cap->implemented;
1567 used = __ceph_caps_used(ci);
1568 wanted = __ceph_caps_file_wanted(ci);
1569 oissued = __ceph_caps_issued_other(ci, cap);
1571 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1572 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1573 ceph_cap_string(used), ceph_cap_string(wanted));
1574 if (cap == ci->i_auth_cap) {
1575 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1576 !list_empty(&ci->i_cap_snaps))
1578 if ((used | wanted) & CEPH_CAP_ANY_WR)
1580 /* Note: it's possible that i_filelock_ref becomes non-zero
1581 * after dropping auth caps. It doesn't hurt because reply
1582 * of lock mds request will re-add auth caps. */
1583 if (atomic_read(&ci->i_filelock_ref) > 0)
1586 /* The inode has cached pages, but it's no longer used.
1587 * we can safely drop it */
1588 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1589 !(oissued & CEPH_CAP_FILE_CACHE)) {
1593 if ((used | wanted) & ~oissued & mine)
1594 goto out; /* we need these caps */
1597 /* we aren't the only cap.. just remove us */
1598 __ceph_remove_cap(cap, true);
1599 session->s_trim_caps--;
1601 struct dentry *dentry;
1602 /* try dropping referring dentries */
1603 spin_unlock(&ci->i_ceph_lock);
1604 dentry = d_find_any_alias(inode);
1605 if (dentry && drop_negative_children(dentry)) {
1608 d_prune_aliases(inode);
1609 count = atomic_read(&inode->i_count);
1611 session->s_trim_caps--;
1612 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1621 spin_unlock(&ci->i_ceph_lock);
1626 * Trim session cap count down to some max number.
1628 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1629 struct ceph_mds_session *session,
1632 int trim_caps = session->s_nr_caps - max_caps;
1634 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1635 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1636 if (trim_caps > 0) {
1637 session->s_trim_caps = trim_caps;
1638 iterate_session_caps(session, trim_caps_cb, session);
1639 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1640 session->s_mds, session->s_nr_caps, max_caps,
1641 trim_caps - session->s_trim_caps);
1642 session->s_trim_caps = 0;
1645 ceph_send_cap_releases(mdsc, session);
1649 static int check_caps_flush(struct ceph_mds_client *mdsc,
1654 spin_lock(&mdsc->cap_dirty_lock);
1655 if (!list_empty(&mdsc->cap_flush_list)) {
1656 struct ceph_cap_flush *cf =
1657 list_first_entry(&mdsc->cap_flush_list,
1658 struct ceph_cap_flush, g_list);
1659 if (cf->tid <= want_flush_tid) {
1660 dout("check_caps_flush still flushing tid "
1661 "%llu <= %llu\n", cf->tid, want_flush_tid);
1665 spin_unlock(&mdsc->cap_dirty_lock);
1670 * flush all dirty inode data to disk.
1672 * returns true if we've flushed through want_flush_tid
1674 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1677 dout("check_caps_flush want %llu\n", want_flush_tid);
1679 wait_event(mdsc->cap_flushing_wq,
1680 check_caps_flush(mdsc, want_flush_tid));
1682 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1686 * called under s_mutex
1688 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1689 struct ceph_mds_session *session)
1691 struct ceph_msg *msg = NULL;
1692 struct ceph_mds_cap_release *head;
1693 struct ceph_mds_cap_item *item;
1694 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1695 struct ceph_cap *cap;
1696 LIST_HEAD(tmp_list);
1697 int num_cap_releases;
1698 __le32 barrier, *cap_barrier;
1700 down_read(&osdc->lock);
1701 barrier = cpu_to_le32(osdc->epoch_barrier);
1702 up_read(&osdc->lock);
1704 spin_lock(&session->s_cap_lock);
1706 list_splice_init(&session->s_cap_releases, &tmp_list);
1707 num_cap_releases = session->s_num_cap_releases;
1708 session->s_num_cap_releases = 0;
1709 spin_unlock(&session->s_cap_lock);
1711 while (!list_empty(&tmp_list)) {
1713 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1714 PAGE_SIZE, GFP_NOFS, false);
1717 head = msg->front.iov_base;
1718 head->num = cpu_to_le32(0);
1719 msg->front.iov_len = sizeof(*head);
1721 msg->hdr.version = cpu_to_le16(2);
1722 msg->hdr.compat_version = cpu_to_le16(1);
1725 cap = list_first_entry(&tmp_list, struct ceph_cap,
1727 list_del(&cap->session_caps);
1730 head = msg->front.iov_base;
1731 le32_add_cpu(&head->num, 1);
1732 item = msg->front.iov_base + msg->front.iov_len;
1733 item->ino = cpu_to_le64(cap->cap_ino);
1734 item->cap_id = cpu_to_le64(cap->cap_id);
1735 item->migrate_seq = cpu_to_le32(cap->mseq);
1736 item->seq = cpu_to_le32(cap->issue_seq);
1737 msg->front.iov_len += sizeof(*item);
1739 ceph_put_cap(mdsc, cap);
1741 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1742 // Append cap_barrier field
1743 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1744 *cap_barrier = barrier;
1745 msg->front.iov_len += sizeof(*cap_barrier);
1747 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1748 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1749 ceph_con_send(&session->s_con, msg);
1754 BUG_ON(num_cap_releases != 0);
1756 spin_lock(&session->s_cap_lock);
1757 if (!list_empty(&session->s_cap_releases))
1759 spin_unlock(&session->s_cap_lock);
1762 // Append cap_barrier field
1763 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1764 *cap_barrier = barrier;
1765 msg->front.iov_len += sizeof(*cap_barrier);
1767 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1768 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1769 ceph_con_send(&session->s_con, msg);
1773 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1775 spin_lock(&session->s_cap_lock);
1776 list_splice(&tmp_list, &session->s_cap_releases);
1777 session->s_num_cap_releases += num_cap_releases;
1778 spin_unlock(&session->s_cap_lock);
1785 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1788 struct ceph_inode_info *ci = ceph_inode(dir);
1789 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1790 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1791 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1792 int order, num_entries;
1794 spin_lock(&ci->i_ceph_lock);
1795 num_entries = ci->i_files + ci->i_subdirs;
1796 spin_unlock(&ci->i_ceph_lock);
1797 num_entries = max(num_entries, 1);
1798 num_entries = min(num_entries, opt->max_readdir);
1800 order = get_order(size * num_entries);
1801 while (order >= 0) {
1802 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1805 if (rinfo->dir_entries)
1809 if (!rinfo->dir_entries)
1812 num_entries = (PAGE_SIZE << order) / size;
1813 num_entries = min(num_entries, opt->max_readdir);
1815 rinfo->dir_buf_size = PAGE_SIZE << order;
1816 req->r_num_caps = num_entries + 1;
1817 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1818 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1823 * Create an mds request.
1825 struct ceph_mds_request *
1826 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1828 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1829 struct timespec64 ts;
1832 return ERR_PTR(-ENOMEM);
1834 mutex_init(&req->r_fill_mutex);
1836 req->r_started = jiffies;
1837 req->r_resend_mds = -1;
1838 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1839 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1841 kref_init(&req->r_kref);
1842 RB_CLEAR_NODE(&req->r_node);
1843 INIT_LIST_HEAD(&req->r_wait);
1844 init_completion(&req->r_completion);
1845 init_completion(&req->r_safe_completion);
1846 INIT_LIST_HEAD(&req->r_unsafe_item);
1848 ktime_get_coarse_real_ts64(&ts);
1849 req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
1852 req->r_direct_mode = mode;
1857 * return oldest (lowest) request, tid in request tree, 0 if none.
1859 * called under mdsc->mutex.
1861 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1863 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1865 return rb_entry(rb_first(&mdsc->request_tree),
1866 struct ceph_mds_request, r_node);
1869 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1871 return mdsc->oldest_tid;
1875 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1876 * on build_path_from_dentry in fs/cifs/dir.c.
1878 * If @stop_on_nosnap, generate path relative to the first non-snapped
1881 * Encode hidden .snap dirs as a double /, i.e.
1882 * foo/.snap/bar -> foo//bar
1884 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1887 struct dentry *temp;
1893 return ERR_PTR(-EINVAL);
1897 seq = read_seqbegin(&rename_lock);
1899 for (temp = dentry; !IS_ROOT(temp);) {
1900 struct inode *inode = d_inode(temp);
1901 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1902 len++; /* slash only */
1903 else if (stop_on_nosnap && inode &&
1904 ceph_snap(inode) == CEPH_NOSNAP)
1907 len += 1 + temp->d_name.len;
1908 temp = temp->d_parent;
1912 len--; /* no leading '/' */
1914 path = kmalloc(len+1, GFP_NOFS);
1916 return ERR_PTR(-ENOMEM);
1918 path[pos] = 0; /* trailing null */
1920 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1921 struct inode *inode;
1923 spin_lock(&temp->d_lock);
1924 inode = d_inode(temp);
1925 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1926 dout("build_path path+%d: %p SNAPDIR\n",
1928 } else if (stop_on_nosnap && inode &&
1929 ceph_snap(inode) == CEPH_NOSNAP) {
1930 spin_unlock(&temp->d_lock);
1933 pos -= temp->d_name.len;
1935 spin_unlock(&temp->d_lock);
1938 strncpy(path + pos, temp->d_name.name,
1941 spin_unlock(&temp->d_lock);
1944 temp = temp->d_parent;
1947 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1948 pr_err("build_path did not end path lookup where "
1949 "expected, namelen is %d, pos is %d\n", len, pos);
1950 /* presumably this is only possible if racing with a
1951 rename of one of the parent directories (we can not
1952 lock the dentries above us to prevent this, but
1953 retrying should be harmless) */
1958 *base = ceph_ino(d_inode(temp));
1960 dout("build_path on %p %d built %llx '%.*s'\n",
1961 dentry, d_count(dentry), *base, len, path);
1965 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1966 const char **ppath, int *ppathlen, u64 *pino,
1973 dir = d_inode_rcu(dentry->d_parent);
1974 if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1975 *pino = ceph_ino(dir);
1977 *ppath = dentry->d_name.name;
1978 *ppathlen = dentry->d_name.len;
1982 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1984 return PTR_ERR(path);
1990 static int build_inode_path(struct inode *inode,
1991 const char **ppath, int *ppathlen, u64 *pino,
1994 struct dentry *dentry;
1997 if (ceph_snap(inode) == CEPH_NOSNAP) {
1998 *pino = ceph_ino(inode);
2002 dentry = d_find_alias(inode);
2003 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2006 return PTR_ERR(path);
2013 * request arguments may be specified via an inode *, a dentry *, or
2014 * an explicit ino+path.
2016 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2017 struct inode *rdiri, const char *rpath,
2018 u64 rino, const char **ppath, int *pathlen,
2019 u64 *ino, int *freepath)
2024 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2025 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2027 } else if (rdentry) {
2028 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2030 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2032 } else if (rpath || rino) {
2035 *pathlen = rpath ? strlen(rpath) : 0;
2036 dout(" path %.*s\n", *pathlen, rpath);
2043 * called under mdsc->mutex
2045 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2046 struct ceph_mds_request *req,
2047 int mds, bool drop_cap_releases)
2049 struct ceph_msg *msg;
2050 struct ceph_mds_request_head *head;
2051 const char *path1 = NULL;
2052 const char *path2 = NULL;
2053 u64 ino1 = 0, ino2 = 0;
2054 int pathlen1 = 0, pathlen2 = 0;
2055 int freepath1 = 0, freepath2 = 0;
2061 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2062 req->r_parent, req->r_path1, req->r_ino1.ino,
2063 &path1, &pathlen1, &ino1, &freepath1);
2069 ret = set_request_path_attr(NULL, req->r_old_dentry,
2070 req->r_old_dentry_dir,
2071 req->r_path2, req->r_ino2.ino,
2072 &path2, &pathlen2, &ino2, &freepath2);
2078 len = sizeof(*head) +
2079 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2080 sizeof(struct ceph_timespec);
2082 /* calculate (max) length for cap releases */
2083 len += sizeof(struct ceph_mds_request_release) *
2084 (!!req->r_inode_drop + !!req->r_dentry_drop +
2085 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2086 if (req->r_dentry_drop)
2087 len += req->r_dentry->d_name.len;
2088 if (req->r_old_dentry_drop)
2089 len += req->r_old_dentry->d_name.len;
2091 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2093 msg = ERR_PTR(-ENOMEM);
2097 msg->hdr.version = cpu_to_le16(2);
2098 msg->hdr.tid = cpu_to_le64(req->r_tid);
2100 head = msg->front.iov_base;
2101 p = msg->front.iov_base + sizeof(*head);
2102 end = msg->front.iov_base + msg->front.iov_len;
2104 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2105 head->op = cpu_to_le32(req->r_op);
2106 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2107 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2108 head->args = req->r_args;
2110 ceph_encode_filepath(&p, end, ino1, path1);
2111 ceph_encode_filepath(&p, end, ino2, path2);
2113 /* make note of release offset, in case we need to replay */
2114 req->r_request_release_offset = p - msg->front.iov_base;
2118 if (req->r_inode_drop)
2119 releases += ceph_encode_inode_release(&p,
2120 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2121 mds, req->r_inode_drop, req->r_inode_unless, 0);
2122 if (req->r_dentry_drop)
2123 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2124 req->r_parent, mds, req->r_dentry_drop,
2125 req->r_dentry_unless);
2126 if (req->r_old_dentry_drop)
2127 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2128 req->r_old_dentry_dir, mds,
2129 req->r_old_dentry_drop,
2130 req->r_old_dentry_unless);
2131 if (req->r_old_inode_drop)
2132 releases += ceph_encode_inode_release(&p,
2133 d_inode(req->r_old_dentry),
2134 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2136 if (drop_cap_releases) {
2138 p = msg->front.iov_base + req->r_request_release_offset;
2141 head->num_releases = cpu_to_le16(releases);
2145 struct ceph_timespec ts;
2146 ceph_encode_timespec64(&ts, &req->r_stamp);
2147 ceph_encode_copy(&p, &ts, sizeof(ts));
2151 msg->front.iov_len = p - msg->front.iov_base;
2152 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2154 if (req->r_pagelist) {
2155 struct ceph_pagelist *pagelist = req->r_pagelist;
2156 ceph_msg_data_add_pagelist(msg, pagelist);
2157 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2159 msg->hdr.data_len = 0;
2162 msg->hdr.data_off = cpu_to_le16(0);
2166 kfree((char *)path2);
2169 kfree((char *)path1);
2175 * called under mdsc->mutex if error, under no mutex if
2178 static void complete_request(struct ceph_mds_client *mdsc,
2179 struct ceph_mds_request *req)
2181 if (req->r_callback)
2182 req->r_callback(mdsc, req);
2184 complete_all(&req->r_completion);
2188 * called under mdsc->mutex
2190 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2191 struct ceph_mds_request *req,
2192 int mds, bool drop_cap_releases)
2194 struct ceph_mds_request_head *rhead;
2195 struct ceph_msg *msg;
2200 struct ceph_cap *cap =
2201 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2204 req->r_sent_on_mseq = cap->mseq;
2206 req->r_sent_on_mseq = -1;
2208 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2209 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2211 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2214 * Replay. Do not regenerate message (and rebuild
2215 * paths, etc.); just use the original message.
2216 * Rebuilding paths will break for renames because
2217 * d_move mangles the src name.
2219 msg = req->r_request;
2220 rhead = msg->front.iov_base;
2222 flags = le32_to_cpu(rhead->flags);
2223 flags |= CEPH_MDS_FLAG_REPLAY;
2224 rhead->flags = cpu_to_le32(flags);
2226 if (req->r_target_inode)
2227 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2229 rhead->num_retry = req->r_attempts - 1;
2231 /* remove cap/dentry releases from message */
2232 rhead->num_releases = 0;
2235 p = msg->front.iov_base + req->r_request_release_offset;
2237 struct ceph_timespec ts;
2238 ceph_encode_timespec64(&ts, &req->r_stamp);
2239 ceph_encode_copy(&p, &ts, sizeof(ts));
2242 msg->front.iov_len = p - msg->front.iov_base;
2243 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2247 if (req->r_request) {
2248 ceph_msg_put(req->r_request);
2249 req->r_request = NULL;
2251 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2253 req->r_err = PTR_ERR(msg);
2254 return PTR_ERR(msg);
2256 req->r_request = msg;
2258 rhead = msg->front.iov_base;
2259 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2260 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2261 flags |= CEPH_MDS_FLAG_REPLAY;
2263 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2264 rhead->flags = cpu_to_le32(flags);
2265 rhead->num_fwd = req->r_num_fwd;
2266 rhead->num_retry = req->r_attempts - 1;
2269 dout(" r_parent = %p\n", req->r_parent);
2274 * send request, or put it on the appropriate wait list.
2276 static void __do_request(struct ceph_mds_client *mdsc,
2277 struct ceph_mds_request *req)
2279 struct ceph_mds_session *session = NULL;
2283 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2284 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2285 __unregister_request(mdsc, req);
2289 if (req->r_timeout &&
2290 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2291 dout("do_request timed out\n");
2295 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2296 dout("do_request forced umount\n");
2300 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2301 if (mdsc->mdsmap_err) {
2302 err = mdsc->mdsmap_err;
2303 dout("do_request mdsmap err %d\n", err);
2306 if (mdsc->mdsmap->m_epoch == 0) {
2307 dout("do_request no mdsmap, waiting for map\n");
2308 list_add(&req->r_wait, &mdsc->waiting_for_map);
2311 if (!(mdsc->fsc->mount_options->flags &
2312 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2313 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2315 pr_info("probably no mds server is up\n");
2320 put_request_session(req);
2322 mds = __choose_mds(mdsc, req);
2324 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2325 dout("do_request no mds or not active, waiting for map\n");
2326 list_add(&req->r_wait, &mdsc->waiting_for_map);
2330 /* get, open session */
2331 session = __ceph_lookup_mds_session(mdsc, mds);
2333 session = register_session(mdsc, mds);
2334 if (IS_ERR(session)) {
2335 err = PTR_ERR(session);
2339 req->r_session = get_session(session);
2341 dout("do_request mds%d session %p state %s\n", mds, session,
2342 ceph_session_state_name(session->s_state));
2343 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2344 session->s_state != CEPH_MDS_SESSION_HUNG) {
2345 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2349 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2350 session->s_state == CEPH_MDS_SESSION_CLOSING)
2351 __open_session(mdsc, session);
2352 list_add(&req->r_wait, &session->s_waiting);
2357 req->r_resend_mds = -1; /* forget any previous mds hint */
2359 if (req->r_request_started == 0) /* note request start time */
2360 req->r_request_started = jiffies;
2362 err = __prepare_send_request(mdsc, req, mds, false);
2364 ceph_msg_get(req->r_request);
2365 ceph_con_send(&session->s_con, req->r_request);
2369 ceph_put_mds_session(session);
2372 dout("__do_request early error %d\n", err);
2374 complete_request(mdsc, req);
2375 __unregister_request(mdsc, req);
2381 * called under mdsc->mutex
2383 static void __wake_requests(struct ceph_mds_client *mdsc,
2384 struct list_head *head)
2386 struct ceph_mds_request *req;
2387 LIST_HEAD(tmp_list);
2389 list_splice_init(head, &tmp_list);
2391 while (!list_empty(&tmp_list)) {
2392 req = list_entry(tmp_list.next,
2393 struct ceph_mds_request, r_wait);
2394 list_del_init(&req->r_wait);
2395 dout(" wake request %p tid %llu\n", req, req->r_tid);
2396 __do_request(mdsc, req);
2401 * Wake up threads with requests pending for @mds, so that they can
2402 * resubmit their requests to a possibly different mds.
2404 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2406 struct ceph_mds_request *req;
2407 struct rb_node *p = rb_first(&mdsc->request_tree);
2409 dout("kick_requests mds%d\n", mds);
2411 req = rb_entry(p, struct ceph_mds_request, r_node);
2413 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2415 if (req->r_attempts > 0)
2416 continue; /* only new requests */
2417 if (req->r_session &&
2418 req->r_session->s_mds == mds) {
2419 dout(" kicking tid %llu\n", req->r_tid);
2420 list_del_init(&req->r_wait);
2421 __do_request(mdsc, req);
2426 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2427 struct ceph_mds_request *req)
2429 dout("submit_request on %p\n", req);
2430 mutex_lock(&mdsc->mutex);
2431 __register_request(mdsc, req, NULL);
2432 __do_request(mdsc, req);
2433 mutex_unlock(&mdsc->mutex);
2437 * Synchrously perform an mds request. Take care of all of the
2438 * session setup, forwarding, retry details.
2440 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2442 struct ceph_mds_request *req)
2446 dout("do_request on %p\n", req);
2448 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2450 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2452 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2453 if (req->r_old_dentry_dir)
2454 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2458 mutex_lock(&mdsc->mutex);
2459 __register_request(mdsc, req, dir);
2460 __do_request(mdsc, req);
2468 mutex_unlock(&mdsc->mutex);
2469 dout("do_request waiting\n");
2470 if (!req->r_timeout && req->r_wait_for_completion) {
2471 err = req->r_wait_for_completion(mdsc, req);
2473 long timeleft = wait_for_completion_killable_timeout(
2475 ceph_timeout_jiffies(req->r_timeout));
2479 err = -EIO; /* timed out */
2481 err = timeleft; /* killed */
2483 dout("do_request waited, got %d\n", err);
2484 mutex_lock(&mdsc->mutex);
2486 /* only abort if we didn't race with a real reply */
2487 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2488 err = le32_to_cpu(req->r_reply_info.head->result);
2489 } else if (err < 0) {
2490 dout("aborted request %lld with %d\n", req->r_tid, err);
2493 * ensure we aren't running concurrently with
2494 * ceph_fill_trace or ceph_readdir_prepopulate, which
2495 * rely on locks (dir mutex) held by our caller.
2497 mutex_lock(&req->r_fill_mutex);
2499 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2500 mutex_unlock(&req->r_fill_mutex);
2502 if (req->r_parent &&
2503 (req->r_op & CEPH_MDS_OP_WRITE))
2504 ceph_invalidate_dir_request(req);
2510 mutex_unlock(&mdsc->mutex);
2511 dout("do_request %p done, result %d\n", req, err);
2516 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2517 * namespace request.
2519 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2521 struct inode *dir = req->r_parent;
2522 struct inode *old_dir = req->r_old_dentry_dir;
2524 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2526 ceph_dir_clear_complete(dir);
2528 ceph_dir_clear_complete(old_dir);
2530 ceph_invalidate_dentry_lease(req->r_dentry);
2531 if (req->r_old_dentry)
2532 ceph_invalidate_dentry_lease(req->r_old_dentry);
2538 * We take the session mutex and parse and process the reply immediately.
2539 * This preserves the logical ordering of replies, capabilities, etc., sent
2540 * by the MDS as they are applied to our local cache.
2542 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2544 struct ceph_mds_client *mdsc = session->s_mdsc;
2545 struct ceph_mds_request *req;
2546 struct ceph_mds_reply_head *head = msg->front.iov_base;
2547 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2548 struct ceph_snap_realm *realm;
2551 int mds = session->s_mds;
2553 if (msg->front.iov_len < sizeof(*head)) {
2554 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2559 /* get request, session */
2560 tid = le64_to_cpu(msg->hdr.tid);
2561 mutex_lock(&mdsc->mutex);
2562 req = lookup_get_request(mdsc, tid);
2564 dout("handle_reply on unknown tid %llu\n", tid);
2565 mutex_unlock(&mdsc->mutex);
2568 dout("handle_reply %p\n", req);
2570 /* correct session? */
2571 if (req->r_session != session) {
2572 pr_err("mdsc_handle_reply got %llu on session mds%d"
2573 " not mds%d\n", tid, session->s_mds,
2574 req->r_session ? req->r_session->s_mds : -1);
2575 mutex_unlock(&mdsc->mutex);
2580 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2581 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2582 pr_warn("got a dup %s reply on %llu from mds%d\n",
2583 head->safe ? "safe" : "unsafe", tid, mds);
2584 mutex_unlock(&mdsc->mutex);
2587 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2588 pr_warn("got unsafe after safe on %llu from mds%d\n",
2590 mutex_unlock(&mdsc->mutex);
2594 result = le32_to_cpu(head->result);
2598 * if we're not talking to the authority, send to them
2599 * if the authority has changed while we weren't looking,
2600 * send to new authority
2601 * Otherwise we just have to return an ESTALE
2603 if (result == -ESTALE) {
2604 dout("got ESTALE on request %llu\n", req->r_tid);
2605 req->r_resend_mds = -1;
2606 if (req->r_direct_mode != USE_AUTH_MDS) {
2607 dout("not using auth, setting for that now\n");
2608 req->r_direct_mode = USE_AUTH_MDS;
2609 __do_request(mdsc, req);
2610 mutex_unlock(&mdsc->mutex);
2613 int mds = __choose_mds(mdsc, req);
2614 if (mds >= 0 && mds != req->r_session->s_mds) {
2615 dout("but auth changed, so resending\n");
2616 __do_request(mdsc, req);
2617 mutex_unlock(&mdsc->mutex);
2621 dout("have to return ESTALE on request %llu\n", req->r_tid);
2626 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2627 __unregister_request(mdsc, req);
2629 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2631 * We already handled the unsafe response, now do the
2632 * cleanup. No need to examine the response; the MDS
2633 * doesn't include any result info in the safe
2634 * response. And even if it did, there is nothing
2635 * useful we could do with a revised return value.
2637 dout("got safe reply %llu, mds%d\n", tid, mds);
2639 /* last unsafe request during umount? */
2640 if (mdsc->stopping && !__get_oldest_req(mdsc))
2641 complete_all(&mdsc->safe_umount_waiters);
2642 mutex_unlock(&mdsc->mutex);
2646 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2647 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2648 if (req->r_unsafe_dir) {
2649 struct ceph_inode_info *ci =
2650 ceph_inode(req->r_unsafe_dir);
2651 spin_lock(&ci->i_unsafe_lock);
2652 list_add_tail(&req->r_unsafe_dir_item,
2653 &ci->i_unsafe_dirops);
2654 spin_unlock(&ci->i_unsafe_lock);
2658 dout("handle_reply tid %lld result %d\n", tid, result);
2659 rinfo = &req->r_reply_info;
2660 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2661 mutex_unlock(&mdsc->mutex);
2663 mutex_lock(&session->s_mutex);
2665 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2672 if (rinfo->snapblob_len) {
2673 down_write(&mdsc->snap_rwsem);
2674 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2675 rinfo->snapblob + rinfo->snapblob_len,
2676 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2678 downgrade_write(&mdsc->snap_rwsem);
2680 down_read(&mdsc->snap_rwsem);
2683 /* insert trace into our cache */
2684 mutex_lock(&req->r_fill_mutex);
2685 current->journal_info = req;
2686 err = ceph_fill_trace(mdsc->fsc->sb, req);
2688 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2689 req->r_op == CEPH_MDS_OP_LSSNAP))
2690 ceph_readdir_prepopulate(req, req->r_session);
2691 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2693 current->journal_info = NULL;
2694 mutex_unlock(&req->r_fill_mutex);
2696 up_read(&mdsc->snap_rwsem);
2698 ceph_put_snap_realm(mdsc, realm);
2700 if (err == 0 && req->r_target_inode &&
2701 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2702 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2703 spin_lock(&ci->i_unsafe_lock);
2704 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2705 spin_unlock(&ci->i_unsafe_lock);
2708 mutex_lock(&mdsc->mutex);
2709 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2713 req->r_reply = ceph_msg_get(msg);
2714 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2717 dout("reply arrived after request %lld was aborted\n", tid);
2719 mutex_unlock(&mdsc->mutex);
2721 mutex_unlock(&session->s_mutex);
2723 /* kick calling process */
2724 complete_request(mdsc, req);
2726 ceph_mdsc_put_request(req);
2733 * handle mds notification that our request has been forwarded.
2735 static void handle_forward(struct ceph_mds_client *mdsc,
2736 struct ceph_mds_session *session,
2737 struct ceph_msg *msg)
2739 struct ceph_mds_request *req;
2740 u64 tid = le64_to_cpu(msg->hdr.tid);
2744 void *p = msg->front.iov_base;
2745 void *end = p + msg->front.iov_len;
2747 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2748 next_mds = ceph_decode_32(&p);
2749 fwd_seq = ceph_decode_32(&p);
2751 mutex_lock(&mdsc->mutex);
2752 req = lookup_get_request(mdsc, tid);
2754 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2755 goto out; /* dup reply? */
2758 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2759 dout("forward tid %llu aborted, unregistering\n", tid);
2760 __unregister_request(mdsc, req);
2761 } else if (fwd_seq <= req->r_num_fwd) {
2762 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2763 tid, next_mds, req->r_num_fwd, fwd_seq);
2765 /* resend. forward race not possible; mds would drop */
2766 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2768 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2769 req->r_attempts = 0;
2770 req->r_num_fwd = fwd_seq;
2771 req->r_resend_mds = next_mds;
2772 put_request_session(req);
2773 __do_request(mdsc, req);
2775 ceph_mdsc_put_request(req);
2777 mutex_unlock(&mdsc->mutex);
2781 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2784 static int __decode_and_drop_session_metadata(void **p, void *end)
2786 /* map<string,string> */
2788 ceph_decode_32_safe(p, end, n, bad);
2791 ceph_decode_32_safe(p, end, len, bad);
2792 ceph_decode_need(p, end, len, bad);
2794 ceph_decode_32_safe(p, end, len, bad);
2795 ceph_decode_need(p, end, len, bad);
2804 * handle a mds session control message
2806 static void handle_session(struct ceph_mds_session *session,
2807 struct ceph_msg *msg)
2809 struct ceph_mds_client *mdsc = session->s_mdsc;
2810 int mds = session->s_mds;
2811 int msg_version = le16_to_cpu(msg->hdr.version);
2812 void *p = msg->front.iov_base;
2813 void *end = p + msg->front.iov_len;
2814 struct ceph_mds_session_head *h;
2817 unsigned long features = 0;
2821 ceph_decode_need(&p, end, sizeof(*h), bad);
2825 op = le32_to_cpu(h->op);
2826 seq = le64_to_cpu(h->seq);
2828 if (msg_version >= 3) {
2830 /* version >= 2, metadata */
2831 if (__decode_and_drop_session_metadata(&p, end) < 0)
2833 /* version >= 3, feature bits */
2834 ceph_decode_32_safe(&p, end, len, bad);
2835 ceph_decode_need(&p, end, len, bad);
2836 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
2840 mutex_lock(&mdsc->mutex);
2841 if (op == CEPH_SESSION_CLOSE) {
2842 get_session(session);
2843 __unregister_session(mdsc, session);
2845 /* FIXME: this ttl calculation is generous */
2846 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2847 mutex_unlock(&mdsc->mutex);
2849 mutex_lock(&session->s_mutex);
2851 dout("handle_session mds%d %s %p state %s seq %llu\n",
2852 mds, ceph_session_op_name(op), session,
2853 ceph_session_state_name(session->s_state), seq);
2855 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2856 session->s_state = CEPH_MDS_SESSION_OPEN;
2857 pr_info("mds%d came back\n", session->s_mds);
2861 case CEPH_SESSION_OPEN:
2862 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2863 pr_info("mds%d reconnect success\n", session->s_mds);
2864 session->s_state = CEPH_MDS_SESSION_OPEN;
2865 session->s_features = features;
2866 renewed_caps(mdsc, session, 0);
2869 __close_session(mdsc, session);
2872 case CEPH_SESSION_RENEWCAPS:
2873 if (session->s_renew_seq == seq)
2874 renewed_caps(mdsc, session, 1);
2877 case CEPH_SESSION_CLOSE:
2878 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2879 pr_info("mds%d reconnect denied\n", session->s_mds);
2880 cleanup_session_requests(mdsc, session);
2881 remove_session_caps(session);
2882 wake = 2; /* for good measure */
2883 wake_up_all(&mdsc->session_close_wq);
2886 case CEPH_SESSION_STALE:
2887 pr_info("mds%d caps went stale, renewing\n",
2889 spin_lock(&session->s_gen_ttl_lock);
2890 session->s_cap_gen++;
2891 session->s_cap_ttl = jiffies - 1;
2892 spin_unlock(&session->s_gen_ttl_lock);
2893 send_renew_caps(mdsc, session);
2896 case CEPH_SESSION_RECALL_STATE:
2897 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2900 case CEPH_SESSION_FLUSHMSG:
2901 send_flushmsg_ack(mdsc, session, seq);
2904 case CEPH_SESSION_FORCE_RO:
2905 dout("force_session_readonly %p\n", session);
2906 spin_lock(&session->s_cap_lock);
2907 session->s_readonly = true;
2908 spin_unlock(&session->s_cap_lock);
2909 wake_up_session_caps(session, FORCE_RO);
2912 case CEPH_SESSION_REJECT:
2913 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2914 pr_info("mds%d rejected session\n", session->s_mds);
2915 session->s_state = CEPH_MDS_SESSION_REJECTED;
2916 cleanup_session_requests(mdsc, session);
2917 remove_session_caps(session);
2918 wake = 2; /* for good measure */
2922 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2926 mutex_unlock(&session->s_mutex);
2928 mutex_lock(&mdsc->mutex);
2929 __wake_requests(mdsc, &session->s_waiting);
2931 kick_requests(mdsc, mds);
2932 mutex_unlock(&mdsc->mutex);
2934 if (op == CEPH_SESSION_CLOSE)
2935 ceph_put_mds_session(session);
2939 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2940 (int)msg->front.iov_len);
2947 * called under session->mutex.
2949 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2950 struct ceph_mds_session *session)
2952 struct ceph_mds_request *req, *nreq;
2956 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2958 mutex_lock(&mdsc->mutex);
2959 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2960 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2962 ceph_msg_get(req->r_request);
2963 ceph_con_send(&session->s_con, req->r_request);
2968 * also re-send old requests when MDS enters reconnect stage. So that MDS
2969 * can process completed request in clientreplay stage.
2971 p = rb_first(&mdsc->request_tree);
2973 req = rb_entry(p, struct ceph_mds_request, r_node);
2975 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2977 if (req->r_attempts == 0)
2978 continue; /* only old requests */
2979 if (req->r_session &&
2980 req->r_session->s_mds == session->s_mds) {
2981 err = __prepare_send_request(mdsc, req,
2982 session->s_mds, true);
2984 ceph_msg_get(req->r_request);
2985 ceph_con_send(&session->s_con, req->r_request);
2989 mutex_unlock(&mdsc->mutex);
2992 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
2994 struct ceph_msg *reply;
2995 struct ceph_pagelist *_pagelist;
3000 if (!recon_state->allow_multi)
3003 /* can't handle message that contains both caps and realm */
3004 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3006 /* pre-allocate new pagelist */
3007 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3011 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3015 /* placeholder for nr_caps */
3016 err = ceph_pagelist_encode_32(_pagelist, 0);
3020 if (recon_state->nr_caps) {
3021 /* currently encoding caps */
3022 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3026 /* placeholder for nr_realms (currently encoding relams) */
3027 err = ceph_pagelist_encode_32(_pagelist, 0);
3032 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3036 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3037 addr = kmap_atomic(page);
3038 if (recon_state->nr_caps) {
3039 /* currently encoding caps */
3040 *addr = cpu_to_le32(recon_state->nr_caps);
3042 /* currently encoding relams */
3043 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3045 kunmap_atomic(addr);
3047 reply->hdr.version = cpu_to_le16(5);
3048 reply->hdr.compat_version = cpu_to_le16(4);
3050 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3051 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3053 ceph_con_send(&recon_state->session->s_con, reply);
3054 ceph_pagelist_release(recon_state->pagelist);
3056 recon_state->pagelist = _pagelist;
3057 recon_state->nr_caps = 0;
3058 recon_state->nr_realms = 0;
3059 recon_state->msg_version = 5;
3062 ceph_msg_put(reply);
3064 ceph_pagelist_release(_pagelist);
3069 * Encode information about a cap for a reconnect with the MDS.
3071 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3075 struct ceph_mds_cap_reconnect v2;
3076 struct ceph_mds_cap_reconnect_v1 v1;
3078 struct ceph_inode_info *ci = cap->ci;
3079 struct ceph_reconnect_state *recon_state = arg;
3080 struct ceph_pagelist *pagelist = recon_state->pagelist;
3084 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3085 inode, ceph_vinop(inode), cap, cap->cap_id,
3086 ceph_cap_string(cap->issued));
3088 spin_lock(&ci->i_ceph_lock);
3089 cap->seq = 0; /* reset cap seq */
3090 cap->issue_seq = 0; /* and issue_seq */
3091 cap->mseq = 0; /* and migrate_seq */
3092 cap->cap_gen = cap->session->s_cap_gen;
3094 if (recon_state->msg_version >= 2) {
3095 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3096 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3097 rec.v2.issued = cpu_to_le32(cap->issued);
3098 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3099 rec.v2.pathbase = 0;
3100 rec.v2.flock_len = (__force __le32)
3101 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3103 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3104 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3105 rec.v1.issued = cpu_to_le32(cap->issued);
3106 rec.v1.size = cpu_to_le64(inode->i_size);
3107 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3108 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3109 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3110 rec.v1.pathbase = 0;
3113 if (list_empty(&ci->i_cap_snaps)) {
3114 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3116 struct ceph_cap_snap *capsnap =
3117 list_first_entry(&ci->i_cap_snaps,
3118 struct ceph_cap_snap, ci_item);
3119 snap_follows = capsnap->follows;
3121 spin_unlock(&ci->i_ceph_lock);
3123 if (recon_state->msg_version >= 2) {
3124 int num_fcntl_locks, num_flock_locks;
3125 struct ceph_filelock *flocks = NULL;
3126 size_t struct_len, total_len = sizeof(u64);
3130 if (rec.v2.flock_len) {
3131 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3133 num_fcntl_locks = 0;
3134 num_flock_locks = 0;
3136 if (num_fcntl_locks + num_flock_locks > 0) {
3137 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3138 sizeof(struct ceph_filelock),
3144 err = ceph_encode_locks_to_buffer(inode, flocks,
3159 if (recon_state->msg_version >= 3) {
3160 /* version, compat_version and struct_len */
3161 total_len += 2 * sizeof(u8) + sizeof(u32);
3165 * number of encoded locks is stable, so copy to pagelist
3167 struct_len = 2 * sizeof(u32) +
3168 (num_fcntl_locks + num_flock_locks) *
3169 sizeof(struct ceph_filelock);
3170 rec.v2.flock_len = cpu_to_le32(struct_len);
3172 struct_len += sizeof(u32) + sizeof(rec.v2);
3175 struct_len += sizeof(u64); /* snap_follows */
3177 total_len += struct_len;
3179 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3180 err = send_reconnect_partial(recon_state);
3182 goto out_freeflocks;
3183 pagelist = recon_state->pagelist;
3186 err = ceph_pagelist_reserve(pagelist, total_len);
3188 goto out_freeflocks;
3190 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3191 if (recon_state->msg_version >= 3) {
3192 ceph_pagelist_encode_8(pagelist, struct_v);
3193 ceph_pagelist_encode_8(pagelist, 1);
3194 ceph_pagelist_encode_32(pagelist, struct_len);
3196 ceph_pagelist_encode_string(pagelist, NULL, 0);
3197 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3198 ceph_locks_to_pagelist(flocks, pagelist,
3199 num_fcntl_locks, num_flock_locks);
3201 ceph_pagelist_encode_64(pagelist, snap_follows);
3208 struct dentry *dentry;
3210 dentry = d_find_alias(inode);
3212 path = ceph_mdsc_build_path(dentry,
3213 &pathlen, &pathbase, 0);
3216 err = PTR_ERR(path);
3219 rec.v1.pathbase = cpu_to_le64(pathbase);
3222 err = ceph_pagelist_reserve(pagelist,
3223 sizeof(u64) + sizeof(u32) +
3224 pathlen + sizeof(rec.v1));
3229 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3230 ceph_pagelist_encode_string(pagelist, path, pathlen);
3231 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3238 recon_state->nr_caps++;
3242 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3243 struct ceph_reconnect_state *recon_state)
3246 struct ceph_pagelist *pagelist = recon_state->pagelist;
3249 if (recon_state->msg_version >= 4) {
3250 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3256 * snaprealms. we provide mds with the ino, seq (version), and
3257 * parent for all of our realms. If the mds has any newer info,
3260 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3261 struct ceph_snap_realm *realm =
3262 rb_entry(p, struct ceph_snap_realm, node);
3263 struct ceph_mds_snaprealm_reconnect sr_rec;
3265 if (recon_state->msg_version >= 4) {
3266 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3269 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3270 err = send_reconnect_partial(recon_state);
3273 pagelist = recon_state->pagelist;
3276 err = ceph_pagelist_reserve(pagelist, need);
3280 ceph_pagelist_encode_8(pagelist, 1);
3281 ceph_pagelist_encode_8(pagelist, 1);
3282 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3285 dout(" adding snap realm %llx seq %lld parent %llx\n",
3286 realm->ino, realm->seq, realm->parent_ino);
3287 sr_rec.ino = cpu_to_le64(realm->ino);
3288 sr_rec.seq = cpu_to_le64(realm->seq);
3289 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3291 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3295 recon_state->nr_realms++;
3303 * If an MDS fails and recovers, clients need to reconnect in order to
3304 * reestablish shared state. This includes all caps issued through
3305 * this session _and_ the snap_realm hierarchy. Because it's not
3306 * clear which snap realms the mds cares about, we send everything we
3307 * know about.. that ensures we'll then get any new info the
3308 * recovering MDS might have.
3310 * This is a relatively heavyweight operation, but it's rare.
3312 * called with mdsc->mutex held.
3314 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3315 struct ceph_mds_session *session)
3317 struct ceph_msg *reply;
3318 int mds = session->s_mds;
3320 struct ceph_reconnect_state recon_state = {
3325 pr_info("mds%d reconnect start\n", mds);
3327 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3328 if (!recon_state.pagelist)
3329 goto fail_nopagelist;
3331 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3335 mutex_lock(&session->s_mutex);
3336 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3339 dout("session %p state %s\n", session,
3340 ceph_session_state_name(session->s_state));
3342 spin_lock(&session->s_gen_ttl_lock);
3343 session->s_cap_gen++;
3344 spin_unlock(&session->s_gen_ttl_lock);
3346 spin_lock(&session->s_cap_lock);
3347 /* don't know if session is readonly */
3348 session->s_readonly = 0;
3350 * notify __ceph_remove_cap() that we are composing cap reconnect.
3351 * If a cap get released before being added to the cap reconnect,
3352 * __ceph_remove_cap() should skip queuing cap release.
3354 session->s_cap_reconnect = 1;
3355 /* drop old cap expires; we're about to reestablish that state */
3356 detach_cap_releases(session, &dispose);
3357 spin_unlock(&session->s_cap_lock);
3358 dispose_cap_releases(mdsc, &dispose);
3360 /* trim unused caps to reduce MDS's cache rejoin time */
3361 if (mdsc->fsc->sb->s_root)
3362 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3364 ceph_con_close(&session->s_con);
3365 ceph_con_open(&session->s_con,
3366 CEPH_ENTITY_TYPE_MDS, mds,
3367 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3369 /* replay unsafe requests */
3370 replay_unsafe_requests(mdsc, session);
3372 ceph_early_kick_flushing_caps(mdsc, session);
3374 down_read(&mdsc->snap_rwsem);
3376 /* placeholder for nr_caps */
3377 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3381 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3382 recon_state.msg_version = 3;
3383 recon_state.allow_multi = true;
3384 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3385 recon_state.msg_version = 3;
3387 recon_state.msg_version = 2;
3389 /* trsaverse this session's caps */
3390 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3392 spin_lock(&session->s_cap_lock);
3393 session->s_cap_reconnect = 0;
3394 spin_unlock(&session->s_cap_lock);
3399 /* check if all realms can be encoded into current message */
3400 if (mdsc->num_snap_realms) {
3402 recon_state.pagelist->length +
3403 mdsc->num_snap_realms *
3404 sizeof(struct ceph_mds_snaprealm_reconnect);
3405 if (recon_state.msg_version >= 4) {
3406 /* number of realms */
3407 total_len += sizeof(u32);
3408 /* version, compat_version and struct_len */
3409 total_len += mdsc->num_snap_realms *
3410 (2 * sizeof(u8) + sizeof(u32));
3412 if (total_len > RECONNECT_MAX_SIZE) {
3413 if (!recon_state.allow_multi) {
3417 if (recon_state.nr_caps) {
3418 err = send_reconnect_partial(&recon_state);
3422 recon_state.msg_version = 5;
3426 err = encode_snap_realms(mdsc, &recon_state);
3430 if (recon_state.msg_version >= 5) {
3431 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3436 if (recon_state.nr_caps || recon_state.nr_realms) {
3438 list_first_entry(&recon_state.pagelist->head,
3440 __le32 *addr = kmap_atomic(page);
3441 if (recon_state.nr_caps) {
3442 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3443 *addr = cpu_to_le32(recon_state.nr_caps);
3444 } else if (recon_state.msg_version >= 4) {
3445 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3447 kunmap_atomic(addr);
3450 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3451 if (recon_state.msg_version >= 4)
3452 reply->hdr.compat_version = cpu_to_le16(4);
3454 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3455 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3457 ceph_con_send(&session->s_con, reply);
3459 mutex_unlock(&session->s_mutex);
3461 mutex_lock(&mdsc->mutex);
3462 __wake_requests(mdsc, &session->s_waiting);
3463 mutex_unlock(&mdsc->mutex);
3465 up_read(&mdsc->snap_rwsem);
3466 ceph_pagelist_release(recon_state.pagelist);
3470 ceph_msg_put(reply);
3471 up_read(&mdsc->snap_rwsem);
3472 mutex_unlock(&session->s_mutex);
3474 ceph_pagelist_release(recon_state.pagelist);
3476 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3482 * compare old and new mdsmaps, kicking requests
3483 * and closing out old connections as necessary
3485 * called under mdsc->mutex.
3487 static void check_new_map(struct ceph_mds_client *mdsc,
3488 struct ceph_mdsmap *newmap,
3489 struct ceph_mdsmap *oldmap)
3492 int oldstate, newstate;
3493 struct ceph_mds_session *s;
3495 dout("check_new_map new %u old %u\n",
3496 newmap->m_epoch, oldmap->m_epoch);
3498 for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3499 if (!mdsc->sessions[i])
3501 s = mdsc->sessions[i];
3502 oldstate = ceph_mdsmap_get_state(oldmap, i);
3503 newstate = ceph_mdsmap_get_state(newmap, i);
3505 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3506 i, ceph_mds_state_name(oldstate),
3507 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3508 ceph_mds_state_name(newstate),
3509 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3510 ceph_session_state_name(s->s_state));
3512 if (i >= newmap->m_num_mds ||
3513 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3514 ceph_mdsmap_get_addr(newmap, i),
3515 sizeof(struct ceph_entity_addr))) {
3516 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3517 /* the session never opened, just close it
3520 __unregister_session(mdsc, s);
3521 __wake_requests(mdsc, &s->s_waiting);
3522 ceph_put_mds_session(s);
3523 } else if (i >= newmap->m_num_mds) {
3524 /* force close session for stopped mds */
3526 __unregister_session(mdsc, s);
3527 __wake_requests(mdsc, &s->s_waiting);
3528 kick_requests(mdsc, i);
3529 mutex_unlock(&mdsc->mutex);
3531 mutex_lock(&s->s_mutex);
3532 cleanup_session_requests(mdsc, s);
3533 remove_session_caps(s);
3534 mutex_unlock(&s->s_mutex);
3536 ceph_put_mds_session(s);
3538 mutex_lock(&mdsc->mutex);
3541 mutex_unlock(&mdsc->mutex);
3542 mutex_lock(&s->s_mutex);
3543 mutex_lock(&mdsc->mutex);
3544 ceph_con_close(&s->s_con);
3545 mutex_unlock(&s->s_mutex);
3546 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3548 } else if (oldstate == newstate) {
3549 continue; /* nothing new with this mds */
3555 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3556 newstate >= CEPH_MDS_STATE_RECONNECT) {
3557 mutex_unlock(&mdsc->mutex);
3558 send_mds_reconnect(mdsc, s);
3559 mutex_lock(&mdsc->mutex);
3563 * kick request on any mds that has gone active.
3565 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3566 newstate >= CEPH_MDS_STATE_ACTIVE) {
3567 if (oldstate != CEPH_MDS_STATE_CREATING &&
3568 oldstate != CEPH_MDS_STATE_STARTING)
3569 pr_info("mds%d recovery completed\n", s->s_mds);
3570 kick_requests(mdsc, i);
3571 ceph_kick_flushing_caps(mdsc, s);
3572 wake_up_session_caps(s, RECONNECT);
3576 for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3577 s = mdsc->sessions[i];
3580 if (!ceph_mdsmap_is_laggy(newmap, i))
3582 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3583 s->s_state == CEPH_MDS_SESSION_HUNG ||
3584 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3585 dout(" connecting to export targets of laggy mds%d\n",
3587 __open_export_target_sessions(mdsc, s);
3599 * caller must hold session s_mutex, dentry->d_lock
3601 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3603 struct ceph_dentry_info *di = ceph_dentry(dentry);
3605 ceph_put_mds_session(di->lease_session);
3606 di->lease_session = NULL;
3609 static void handle_lease(struct ceph_mds_client *mdsc,
3610 struct ceph_mds_session *session,
3611 struct ceph_msg *msg)
3613 struct super_block *sb = mdsc->fsc->sb;
3614 struct inode *inode;
3615 struct dentry *parent, *dentry;
3616 struct ceph_dentry_info *di;
3617 int mds = session->s_mds;
3618 struct ceph_mds_lease *h = msg->front.iov_base;
3620 struct ceph_vino vino;
3624 dout("handle_lease from mds%d\n", mds);
3627 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3629 vino.ino = le64_to_cpu(h->ino);
3630 vino.snap = CEPH_NOSNAP;
3631 seq = le32_to_cpu(h->seq);
3632 dname.len = get_unaligned_le32(h + 1);
3633 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3635 dname.name = (void *)(h + 1) + sizeof(u32);
3638 inode = ceph_find_inode(sb, vino);
3639 dout("handle_lease %s, ino %llx %p %.*s\n",
3640 ceph_lease_op_name(h->action), vino.ino, inode,
3641 dname.len, dname.name);
3643 mutex_lock(&session->s_mutex);
3647 dout("handle_lease no inode %llx\n", vino.ino);
3652 parent = d_find_alias(inode);
3654 dout("no parent dentry on inode %p\n", inode);
3656 goto release; /* hrm... */
3658 dname.hash = full_name_hash(parent, dname.name, dname.len);
3659 dentry = d_lookup(parent, &dname);
3664 spin_lock(&dentry->d_lock);
3665 di = ceph_dentry(dentry);
3666 switch (h->action) {
3667 case CEPH_MDS_LEASE_REVOKE:
3668 if (di->lease_session == session) {
3669 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3670 h->seq = cpu_to_le32(di->lease_seq);
3671 __ceph_mdsc_drop_dentry_lease(dentry);
3676 case CEPH_MDS_LEASE_RENEW:
3677 if (di->lease_session == session &&
3678 di->lease_gen == session->s_cap_gen &&
3679 di->lease_renew_from &&
3680 di->lease_renew_after == 0) {
3681 unsigned long duration =
3682 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3684 di->lease_seq = seq;
3685 di->time = di->lease_renew_from + duration;
3686 di->lease_renew_after = di->lease_renew_from +
3688 di->lease_renew_from = 0;
3692 spin_unlock(&dentry->d_lock);
3699 /* let's just reuse the same message */
3700 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3702 ceph_con_send(&session->s_con, msg);
3706 mutex_unlock(&session->s_mutex);
3710 pr_err("corrupt lease message\n");
3714 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3715 struct inode *inode,
3716 struct dentry *dentry, char action,
3719 struct ceph_msg *msg;
3720 struct ceph_mds_lease *lease;
3721 int len = sizeof(*lease) + sizeof(u32);
3724 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3725 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3726 dnamelen = dentry->d_name.len;
3729 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3732 lease = msg->front.iov_base;
3733 lease->action = action;
3734 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3735 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3736 lease->seq = cpu_to_le32(seq);
3737 put_unaligned_le32(dnamelen, lease + 1);
3738 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3741 * if this is a preemptive lease RELEASE, no need to
3742 * flush request stream, since the actual request will
3745 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3747 ceph_con_send(&session->s_con, msg);
3751 * lock unlock sessions, to wait ongoing session activities
3753 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
3757 mutex_lock(&mdsc->mutex);
3758 for (i = 0; i < mdsc->max_sessions; i++) {
3759 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3762 mutex_unlock(&mdsc->mutex);
3763 mutex_lock(&s->s_mutex);
3764 mutex_unlock(&s->s_mutex);
3765 ceph_put_mds_session(s);
3766 mutex_lock(&mdsc->mutex);
3768 mutex_unlock(&mdsc->mutex);
3774 * delayed work -- periodically trim expired leases, renew caps with mds
3776 static void schedule_delayed(struct ceph_mds_client *mdsc)
3779 unsigned hz = round_jiffies_relative(HZ * delay);
3780 schedule_delayed_work(&mdsc->delayed_work, hz);
3783 static void delayed_work(struct work_struct *work)
3786 struct ceph_mds_client *mdsc =
3787 container_of(work, struct ceph_mds_client, delayed_work.work);
3791 dout("mdsc delayed_work\n");
3792 ceph_check_delayed_caps(mdsc);
3794 ceph_trim_snapid_map(mdsc);
3796 mutex_lock(&mdsc->mutex);
3797 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3798 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3799 mdsc->last_renew_caps);
3801 mdsc->last_renew_caps = jiffies;
3803 for (i = 0; i < mdsc->max_sessions; i++) {
3804 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3807 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3808 dout("resending session close request for mds%d\n",
3810 request_close_session(mdsc, s);
3811 ceph_put_mds_session(s);
3814 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3815 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3816 s->s_state = CEPH_MDS_SESSION_HUNG;
3817 pr_info("mds%d hung\n", s->s_mds);
3820 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3821 /* this mds is failed or recovering, just wait */
3822 ceph_put_mds_session(s);
3825 mutex_unlock(&mdsc->mutex);
3827 mutex_lock(&s->s_mutex);
3829 send_renew_caps(mdsc, s);
3831 ceph_con_keepalive(&s->s_con);
3832 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3833 s->s_state == CEPH_MDS_SESSION_HUNG)
3834 ceph_send_cap_releases(mdsc, s);
3835 mutex_unlock(&s->s_mutex);
3836 ceph_put_mds_session(s);
3838 mutex_lock(&mdsc->mutex);
3840 mutex_unlock(&mdsc->mutex);
3842 schedule_delayed(mdsc);
3845 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3848 struct ceph_mds_client *mdsc;
3850 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3854 mutex_init(&mdsc->mutex);
3855 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3856 if (!mdsc->mdsmap) {
3862 init_completion(&mdsc->safe_umount_waiters);
3863 init_waitqueue_head(&mdsc->session_close_wq);
3864 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3865 mdsc->sessions = NULL;
3866 atomic_set(&mdsc->num_sessions, 0);
3867 mdsc->max_sessions = 0;
3869 atomic64_set(&mdsc->quotarealms_count, 0);
3870 mdsc->last_snap_seq = 0;
3871 init_rwsem(&mdsc->snap_rwsem);
3872 mdsc->snap_realms = RB_ROOT;
3873 INIT_LIST_HEAD(&mdsc->snap_empty);
3874 mdsc->num_snap_realms = 0;
3875 spin_lock_init(&mdsc->snap_empty_lock);
3877 mdsc->oldest_tid = 0;
3878 mdsc->request_tree = RB_ROOT;
3879 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3880 mdsc->last_renew_caps = jiffies;
3881 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3882 spin_lock_init(&mdsc->cap_delay_lock);
3883 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3884 spin_lock_init(&mdsc->snap_flush_lock);
3885 mdsc->last_cap_flush_tid = 1;
3886 INIT_LIST_HEAD(&mdsc->cap_flush_list);
3887 INIT_LIST_HEAD(&mdsc->cap_dirty);
3888 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3889 mdsc->num_cap_flushing = 0;
3890 spin_lock_init(&mdsc->cap_dirty_lock);
3891 init_waitqueue_head(&mdsc->cap_flushing_wq);
3892 spin_lock_init(&mdsc->dentry_lru_lock);
3893 INIT_LIST_HEAD(&mdsc->dentry_lru);
3895 ceph_caps_init(mdsc);
3896 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3898 spin_lock_init(&mdsc->snapid_map_lock);
3899 mdsc->snapid_map_tree = RB_ROOT;
3900 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
3902 init_rwsem(&mdsc->pool_perm_rwsem);
3903 mdsc->pool_perm_tree = RB_ROOT;
3905 strscpy(mdsc->nodename, utsname()->nodename,
3906 sizeof(mdsc->nodename));
3911 * Wait for safe replies on open mds requests. If we time out, drop
3912 * all requests from the tree to avoid dangling dentry refs.
3914 static void wait_requests(struct ceph_mds_client *mdsc)
3916 struct ceph_options *opts = mdsc->fsc->client->options;
3917 struct ceph_mds_request *req;
3919 mutex_lock(&mdsc->mutex);
3920 if (__get_oldest_req(mdsc)) {
3921 mutex_unlock(&mdsc->mutex);
3923 dout("wait_requests waiting for requests\n");
3924 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3925 ceph_timeout_jiffies(opts->mount_timeout));
3927 /* tear down remaining requests */
3928 mutex_lock(&mdsc->mutex);
3929 while ((req = __get_oldest_req(mdsc))) {
3930 dout("wait_requests timed out on tid %llu\n",
3932 __unregister_request(mdsc, req);
3935 mutex_unlock(&mdsc->mutex);
3936 dout("wait_requests done\n");
3940 * called before mount is ro, and before dentries are torn down.
3941 * (hmm, does this still race with new lookups?)
3943 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3945 dout("pre_umount\n");
3948 lock_unlock_sessions(mdsc);
3949 ceph_flush_dirty_caps(mdsc);
3950 wait_requests(mdsc);
3953 * wait for reply handlers to drop their request refs and
3954 * their inode/dcache refs
3960 * wait for all write mds requests to flush.
3962 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3964 struct ceph_mds_request *req = NULL, *nextreq;
3967 mutex_lock(&mdsc->mutex);
3968 dout("wait_unsafe_requests want %lld\n", want_tid);
3970 req = __get_oldest_req(mdsc);
3971 while (req && req->r_tid <= want_tid) {
3972 /* find next request */
3973 n = rb_next(&req->r_node);
3975 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3978 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3979 (req->r_op & CEPH_MDS_OP_WRITE)) {
3981 ceph_mdsc_get_request(req);
3983 ceph_mdsc_get_request(nextreq);
3984 mutex_unlock(&mdsc->mutex);
3985 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3986 req->r_tid, want_tid);
3987 wait_for_completion(&req->r_safe_completion);
3988 mutex_lock(&mdsc->mutex);
3989 ceph_mdsc_put_request(req);
3991 break; /* next dne before, so we're done! */
3992 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3993 /* next request was removed from tree */
3994 ceph_mdsc_put_request(nextreq);
3997 ceph_mdsc_put_request(nextreq); /* won't go away */
4001 mutex_unlock(&mdsc->mutex);
4002 dout("wait_unsafe_requests done\n");
4005 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4007 u64 want_tid, want_flush;
4009 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4013 mutex_lock(&mdsc->mutex);
4014 want_tid = mdsc->last_tid;
4015 mutex_unlock(&mdsc->mutex);
4017 ceph_flush_dirty_caps(mdsc);
4018 spin_lock(&mdsc->cap_dirty_lock);
4019 want_flush = mdsc->last_cap_flush_tid;
4020 if (!list_empty(&mdsc->cap_flush_list)) {
4021 struct ceph_cap_flush *cf =
4022 list_last_entry(&mdsc->cap_flush_list,
4023 struct ceph_cap_flush, g_list);
4026 spin_unlock(&mdsc->cap_dirty_lock);
4028 dout("sync want tid %lld flush_seq %lld\n",
4029 want_tid, want_flush);
4031 wait_unsafe_requests(mdsc, want_tid);
4032 wait_caps_flush(mdsc, want_flush);
4036 * true if all sessions are closed, or we force unmount
4038 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4040 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4042 return atomic_read(&mdsc->num_sessions) <= skipped;
4046 * called after sb is ro.
4048 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4050 struct ceph_options *opts = mdsc->fsc->client->options;
4051 struct ceph_mds_session *session;
4055 dout("close_sessions\n");
4057 /* close sessions */
4058 mutex_lock(&mdsc->mutex);
4059 for (i = 0; i < mdsc->max_sessions; i++) {
4060 session = __ceph_lookup_mds_session(mdsc, i);
4063 mutex_unlock(&mdsc->mutex);
4064 mutex_lock(&session->s_mutex);
4065 if (__close_session(mdsc, session) <= 0)
4067 mutex_unlock(&session->s_mutex);
4068 ceph_put_mds_session(session);
4069 mutex_lock(&mdsc->mutex);
4071 mutex_unlock(&mdsc->mutex);
4073 dout("waiting for sessions to close\n");
4074 wait_event_timeout(mdsc->session_close_wq,
4075 done_closing_sessions(mdsc, skipped),
4076 ceph_timeout_jiffies(opts->mount_timeout));
4078 /* tear down remaining sessions */
4079 mutex_lock(&mdsc->mutex);
4080 for (i = 0; i < mdsc->max_sessions; i++) {
4081 if (mdsc->sessions[i]) {
4082 session = get_session(mdsc->sessions[i]);
4083 __unregister_session(mdsc, session);
4084 mutex_unlock(&mdsc->mutex);
4085 mutex_lock(&session->s_mutex);
4086 remove_session_caps(session);
4087 mutex_unlock(&session->s_mutex);
4088 ceph_put_mds_session(session);
4089 mutex_lock(&mdsc->mutex);
4092 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4093 mutex_unlock(&mdsc->mutex);
4095 ceph_cleanup_snapid_map(mdsc);
4097 ceph_cleanup_empty_realms(mdsc);
4099 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4104 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4106 struct ceph_mds_session *session;
4109 dout("force umount\n");
4111 mutex_lock(&mdsc->mutex);
4112 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4113 session = __ceph_lookup_mds_session(mdsc, mds);
4116 mutex_unlock(&mdsc->mutex);
4117 mutex_lock(&session->s_mutex);
4118 __close_session(mdsc, session);
4119 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4120 cleanup_session_requests(mdsc, session);
4121 remove_session_caps(session);
4123 mutex_unlock(&session->s_mutex);
4124 ceph_put_mds_session(session);
4125 mutex_lock(&mdsc->mutex);
4126 kick_requests(mdsc, mds);
4128 __wake_requests(mdsc, &mdsc->waiting_for_map);
4129 mutex_unlock(&mdsc->mutex);
4132 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4135 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4137 ceph_mdsmap_destroy(mdsc->mdsmap);
4138 kfree(mdsc->sessions);
4139 ceph_caps_finalize(mdsc);
4140 ceph_pool_perm_destroy(mdsc);
4143 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4145 struct ceph_mds_client *mdsc = fsc->mdsc;
4146 dout("mdsc_destroy %p\n", mdsc);
4151 /* flush out any connection work with references to us */
4154 ceph_mdsc_stop(mdsc);
4158 dout("mdsc_destroy %p done\n", mdsc);
4161 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4163 struct ceph_fs_client *fsc = mdsc->fsc;
4164 const char *mds_namespace = fsc->mount_options->mds_namespace;
4165 void *p = msg->front.iov_base;
4166 void *end = p + msg->front.iov_len;
4170 u32 mount_fscid = (u32)-1;
4171 u8 struct_v, struct_cv;
4174 ceph_decode_need(&p, end, sizeof(u32), bad);
4175 epoch = ceph_decode_32(&p);
4177 dout("handle_fsmap epoch %u\n", epoch);
4179 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4180 struct_v = ceph_decode_8(&p);
4181 struct_cv = ceph_decode_8(&p);
4182 map_len = ceph_decode_32(&p);
4184 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4185 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4187 num_fs = ceph_decode_32(&p);
4188 while (num_fs-- > 0) {
4189 void *info_p, *info_end;
4194 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4195 info_v = ceph_decode_8(&p);
4196 info_cv = ceph_decode_8(&p);
4197 info_len = ceph_decode_32(&p);
4198 ceph_decode_need(&p, end, info_len, bad);
4200 info_end = p + info_len;
4203 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4204 fscid = ceph_decode_32(&info_p);
4205 namelen = ceph_decode_32(&info_p);
4206 ceph_decode_need(&info_p, info_end, namelen, bad);
4208 if (mds_namespace &&
4209 strlen(mds_namespace) == namelen &&
4210 !strncmp(mds_namespace, (char *)info_p, namelen)) {
4211 mount_fscid = fscid;
4216 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4217 if (mount_fscid != (u32)-1) {
4218 fsc->client->monc.fs_cluster_id = mount_fscid;
4219 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4221 ceph_monc_renew_subs(&fsc->client->monc);
4229 pr_err("error decoding fsmap\n");
4231 mutex_lock(&mdsc->mutex);
4232 mdsc->mdsmap_err = err;
4233 __wake_requests(mdsc, &mdsc->waiting_for_map);
4234 mutex_unlock(&mdsc->mutex);
4238 * handle mds map update.
4240 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4244 void *p = msg->front.iov_base;
4245 void *end = p + msg->front.iov_len;
4246 struct ceph_mdsmap *newmap, *oldmap;
4247 struct ceph_fsid fsid;
4250 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4251 ceph_decode_copy(&p, &fsid, sizeof(fsid));
4252 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4254 epoch = ceph_decode_32(&p);
4255 maplen = ceph_decode_32(&p);
4256 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4258 /* do we need it? */
4259 mutex_lock(&mdsc->mutex);
4260 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4261 dout("handle_map epoch %u <= our %u\n",
4262 epoch, mdsc->mdsmap->m_epoch);
4263 mutex_unlock(&mdsc->mutex);
4267 newmap = ceph_mdsmap_decode(&p, end);
4268 if (IS_ERR(newmap)) {
4269 err = PTR_ERR(newmap);
4273 /* swap into place */
4275 oldmap = mdsc->mdsmap;
4276 mdsc->mdsmap = newmap;
4277 check_new_map(mdsc, newmap, oldmap);
4278 ceph_mdsmap_destroy(oldmap);
4280 mdsc->mdsmap = newmap; /* first mds map */
4282 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4285 __wake_requests(mdsc, &mdsc->waiting_for_map);
4286 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4287 mdsc->mdsmap->m_epoch);
4289 mutex_unlock(&mdsc->mutex);
4290 schedule_delayed(mdsc);
4294 mutex_unlock(&mdsc->mutex);
4296 pr_err("error decoding mdsmap %d\n", err);
4300 static struct ceph_connection *con_get(struct ceph_connection *con)
4302 struct ceph_mds_session *s = con->private;
4304 if (get_session(s)) {
4305 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4308 dout("mdsc con_get %p FAIL\n", s);
4312 static void con_put(struct ceph_connection *con)
4314 struct ceph_mds_session *s = con->private;
4316 dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4317 ceph_put_mds_session(s);
4321 * if the client is unresponsive for long enough, the mds will kill
4322 * the session entirely.
4324 static void peer_reset(struct ceph_connection *con)
4326 struct ceph_mds_session *s = con->private;
4327 struct ceph_mds_client *mdsc = s->s_mdsc;
4329 pr_warn("mds%d closed our session\n", s->s_mds);
4330 send_mds_reconnect(mdsc, s);
4333 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4335 struct ceph_mds_session *s = con->private;
4336 struct ceph_mds_client *mdsc = s->s_mdsc;
4337 int type = le16_to_cpu(msg->hdr.type);
4339 mutex_lock(&mdsc->mutex);
4340 if (__verify_registered_session(mdsc, s) < 0) {
4341 mutex_unlock(&mdsc->mutex);
4344 mutex_unlock(&mdsc->mutex);
4347 case CEPH_MSG_MDS_MAP:
4348 ceph_mdsc_handle_mdsmap(mdsc, msg);
4350 case CEPH_MSG_FS_MAP_USER:
4351 ceph_mdsc_handle_fsmap(mdsc, msg);
4353 case CEPH_MSG_CLIENT_SESSION:
4354 handle_session(s, msg);
4356 case CEPH_MSG_CLIENT_REPLY:
4357 handle_reply(s, msg);
4359 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4360 handle_forward(mdsc, s, msg);
4362 case CEPH_MSG_CLIENT_CAPS:
4363 ceph_handle_caps(s, msg);
4365 case CEPH_MSG_CLIENT_SNAP:
4366 ceph_handle_snap(mdsc, s, msg);
4368 case CEPH_MSG_CLIENT_LEASE:
4369 handle_lease(mdsc, s, msg);
4371 case CEPH_MSG_CLIENT_QUOTA:
4372 ceph_handle_quota(mdsc, s, msg);
4376 pr_err("received unknown message type %d %s\n", type,
4377 ceph_msg_type_name(type));
4388 * Note: returned pointer is the address of a structure that's
4389 * managed separately. Caller must *not* attempt to free it.
4391 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4392 int *proto, int force_new)
4394 struct ceph_mds_session *s = con->private;
4395 struct ceph_mds_client *mdsc = s->s_mdsc;
4396 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4397 struct ceph_auth_handshake *auth = &s->s_auth;
4399 if (force_new && auth->authorizer) {
4400 ceph_auth_destroy_authorizer(auth->authorizer);
4401 auth->authorizer = NULL;
4403 if (!auth->authorizer) {
4404 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4407 return ERR_PTR(ret);
4409 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4412 return ERR_PTR(ret);
4414 *proto = ac->protocol;
4419 static int add_authorizer_challenge(struct ceph_connection *con,
4420 void *challenge_buf, int challenge_buf_len)
4422 struct ceph_mds_session *s = con->private;
4423 struct ceph_mds_client *mdsc = s->s_mdsc;
4424 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4426 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4427 challenge_buf, challenge_buf_len);
4430 static int verify_authorizer_reply(struct ceph_connection *con)
4432 struct ceph_mds_session *s = con->private;
4433 struct ceph_mds_client *mdsc = s->s_mdsc;
4434 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4436 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4439 static int invalidate_authorizer(struct ceph_connection *con)
4441 struct ceph_mds_session *s = con->private;
4442 struct ceph_mds_client *mdsc = s->s_mdsc;
4443 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4445 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4447 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4450 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4451 struct ceph_msg_header *hdr, int *skip)
4453 struct ceph_msg *msg;
4454 int type = (int) le16_to_cpu(hdr->type);
4455 int front_len = (int) le32_to_cpu(hdr->front_len);
4461 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4463 pr_err("unable to allocate msg type %d len %d\n",
4471 static int mds_sign_message(struct ceph_msg *msg)
4473 struct ceph_mds_session *s = msg->con->private;
4474 struct ceph_auth_handshake *auth = &s->s_auth;
4476 return ceph_auth_sign_message(auth, msg);
4479 static int mds_check_message_signature(struct ceph_msg *msg)
4481 struct ceph_mds_session *s = msg->con->private;
4482 struct ceph_auth_handshake *auth = &s->s_auth;
4484 return ceph_auth_check_message_signature(auth, msg);
4487 static const struct ceph_connection_operations mds_con_ops = {
4490 .dispatch = dispatch,
4491 .get_authorizer = get_authorizer,
4492 .add_authorizer_challenge = add_authorizer_challenge,
4493 .verify_authorizer_reply = verify_authorizer_reply,
4494 .invalidate_authorizer = invalidate_authorizer,
4495 .peer_reset = peer_reset,
4496 .alloc_msg = mds_alloc_msg,
4497 .sign_message = mds_sign_message,
4498 .check_message_signature = mds_check_message_signature,