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>
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
48 struct ceph_reconnect_state {
50 struct ceph_pagelist *pagelist;
54 static void __wake_requests(struct ceph_mds_client *mdsc,
55 struct list_head *head);
57 static const struct ceph_connection_operations mds_con_ops;
65 * parse individual inode info
67 static int parse_reply_info_in(void **p, void *end,
68 struct ceph_mds_reply_info_in *info,
74 *p += sizeof(struct ceph_mds_reply_inode) +
75 sizeof(*info->in->fragtree.splits) *
76 le32_to_cpu(info->in->fragtree.nsplits);
78 ceph_decode_32_safe(p, end, info->symlink_len, bad);
79 ceph_decode_need(p, end, info->symlink_len, bad);
81 *p += info->symlink_len;
83 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
84 ceph_decode_copy_safe(p, end, &info->dir_layout,
85 sizeof(info->dir_layout), bad);
87 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
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 info->pool_ns_len = 0;
104 info->pool_ns_data = NULL;
105 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
106 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
107 if (info->pool_ns_len > 0) {
108 ceph_decode_need(p, end, info->pool_ns_len, bad);
109 info->pool_ns_data = *p;
110 *p += info->pool_ns_len;
120 * parse a normal reply, which may contain a (dir+)dentry and/or a
123 static int parse_reply_info_trace(void **p, void *end,
124 struct ceph_mds_reply_info_parsed *info,
129 if (info->head->is_dentry) {
130 err = parse_reply_info_in(p, end, &info->diri, features);
134 if (unlikely(*p + sizeof(*info->dirfrag) > end))
137 *p += sizeof(*info->dirfrag) +
138 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
139 if (unlikely(*p > end))
142 ceph_decode_32_safe(p, end, info->dname_len, bad);
143 ceph_decode_need(p, end, info->dname_len, bad);
145 *p += info->dname_len;
147 *p += sizeof(*info->dlease);
150 if (info->head->is_target) {
151 err = parse_reply_info_in(p, end, &info->targeti, features);
156 if (unlikely(*p != end))
163 pr_err("problem parsing mds trace %d\n", err);
168 * parse readdir results
170 static int parse_reply_info_dir(void **p, void *end,
171 struct ceph_mds_reply_info_parsed *info,
178 if (*p + sizeof(*info->dir_dir) > end)
180 *p += sizeof(*info->dir_dir) +
181 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
185 ceph_decode_need(p, end, sizeof(num) + 2, bad);
186 num = ceph_decode_32(p);
188 u16 flags = ceph_decode_16(p);
189 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
190 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
191 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
192 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
197 BUG_ON(!info->dir_entries);
198 if ((unsigned long)(info->dir_entries + num) >
199 (unsigned long)info->dir_entries + info->dir_buf_size) {
200 pr_err("dir contents are larger than expected\n");
207 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
209 ceph_decode_need(p, end, sizeof(u32)*2, bad);
210 rde->name_len = ceph_decode_32(p);
211 ceph_decode_need(p, end, rde->name_len, bad);
214 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
216 *p += sizeof(struct ceph_mds_reply_lease);
219 err = parse_reply_info_in(p, end, &rde->inode, features);
222 /* ceph_readdir_prepopulate() will update it */
236 pr_err("problem parsing dir contents %d\n", err);
241 * parse fcntl F_GETLK results
243 static int parse_reply_info_filelock(void **p, void *end,
244 struct ceph_mds_reply_info_parsed *info,
247 if (*p + sizeof(*info->filelock_reply) > end)
250 info->filelock_reply = *p;
251 *p += sizeof(*info->filelock_reply);
253 if (unlikely(*p != end))
262 * parse create results
264 static int parse_reply_info_create(void **p, void *end,
265 struct ceph_mds_reply_info_parsed *info,
268 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
270 info->has_create_ino = false;
272 info->has_create_ino = true;
273 info->ino = ceph_decode_64(p);
277 if (unlikely(*p != end))
286 * parse extra results
288 static int parse_reply_info_extra(void **p, void *end,
289 struct ceph_mds_reply_info_parsed *info,
292 u32 op = le32_to_cpu(info->head->op);
294 if (op == CEPH_MDS_OP_GETFILELOCK)
295 return parse_reply_info_filelock(p, end, info, features);
296 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
297 return parse_reply_info_dir(p, end, info, features);
298 else if (op == CEPH_MDS_OP_CREATE)
299 return parse_reply_info_create(p, end, info, features);
305 * parse entire mds reply
307 static int parse_reply_info(struct ceph_msg *msg,
308 struct ceph_mds_reply_info_parsed *info,
315 info->head = msg->front.iov_base;
316 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
317 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
320 ceph_decode_32_safe(&p, end, len, bad);
322 ceph_decode_need(&p, end, len, bad);
323 err = parse_reply_info_trace(&p, p+len, info, features);
329 ceph_decode_32_safe(&p, end, len, bad);
331 ceph_decode_need(&p, end, len, bad);
332 err = parse_reply_info_extra(&p, p+len, info, features);
338 ceph_decode_32_safe(&p, end, len, bad);
339 info->snapblob_len = len;
350 pr_err("mds parse_reply err %d\n", err);
354 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
356 if (!info->dir_entries)
358 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
365 const char *ceph_session_state_name(int s)
368 case CEPH_MDS_SESSION_NEW: return "new";
369 case CEPH_MDS_SESSION_OPENING: return "opening";
370 case CEPH_MDS_SESSION_OPEN: return "open";
371 case CEPH_MDS_SESSION_HUNG: return "hung";
372 case CEPH_MDS_SESSION_CLOSING: return "closing";
373 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
374 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
375 case CEPH_MDS_SESSION_REJECTED: return "rejected";
376 default: return "???";
380 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
382 if (refcount_inc_not_zero(&s->s_ref)) {
383 dout("mdsc get_session %p %d -> %d\n", s,
384 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
387 dout("mdsc get_session %p 0 -- FAIL", s);
392 void ceph_put_mds_session(struct ceph_mds_session *s)
394 dout("mdsc put_session %p %d -> %d\n", s,
395 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
396 if (refcount_dec_and_test(&s->s_ref)) {
397 if (s->s_auth.authorizer)
398 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
404 * called under mdsc->mutex
406 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
409 struct ceph_mds_session *session;
411 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
413 session = mdsc->sessions[mds];
414 dout("lookup_mds_session %p %d\n", session,
415 refcount_read(&session->s_ref));
416 get_session(session);
420 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
422 if (mds >= mdsc->max_sessions)
424 return mdsc->sessions[mds];
427 static int __verify_registered_session(struct ceph_mds_client *mdsc,
428 struct ceph_mds_session *s)
430 if (s->s_mds >= mdsc->max_sessions ||
431 mdsc->sessions[s->s_mds] != s)
437 * create+register a new session for given mds.
438 * called under mdsc->mutex.
440 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
443 struct ceph_mds_session *s;
445 if (mds >= mdsc->mdsmap->m_num_mds)
446 return ERR_PTR(-EINVAL);
448 s = kzalloc(sizeof(*s), GFP_NOFS);
450 return ERR_PTR(-ENOMEM);
453 s->s_state = CEPH_MDS_SESSION_NEW;
456 mutex_init(&s->s_mutex);
458 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
460 spin_lock_init(&s->s_gen_ttl_lock);
462 s->s_cap_ttl = jiffies - 1;
464 spin_lock_init(&s->s_cap_lock);
465 s->s_renew_requested = 0;
467 INIT_LIST_HEAD(&s->s_caps);
470 refcount_set(&s->s_ref, 1);
471 INIT_LIST_HEAD(&s->s_waiting);
472 INIT_LIST_HEAD(&s->s_unsafe);
473 s->s_num_cap_releases = 0;
474 s->s_cap_reconnect = 0;
475 s->s_cap_iterator = NULL;
476 INIT_LIST_HEAD(&s->s_cap_releases);
477 INIT_LIST_HEAD(&s->s_cap_flushing);
479 dout("register_session mds%d\n", mds);
480 if (mds >= mdsc->max_sessions) {
481 int newmax = 1 << get_count_order(mds+1);
482 struct ceph_mds_session **sa;
484 dout("register_session realloc to %d\n", newmax);
485 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
488 if (mdsc->sessions) {
489 memcpy(sa, mdsc->sessions,
490 mdsc->max_sessions * sizeof(void *));
491 kfree(mdsc->sessions);
494 mdsc->max_sessions = newmax;
496 mdsc->sessions[mds] = s;
497 atomic_inc(&mdsc->num_sessions);
498 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
500 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
501 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
507 return ERR_PTR(-ENOMEM);
511 * called under mdsc->mutex
513 static void __unregister_session(struct ceph_mds_client *mdsc,
514 struct ceph_mds_session *s)
516 dout("__unregister_session mds%d %p\n", s->s_mds, s);
517 BUG_ON(mdsc->sessions[s->s_mds] != s);
518 mdsc->sessions[s->s_mds] = NULL;
519 ceph_con_close(&s->s_con);
520 ceph_put_mds_session(s);
521 atomic_dec(&mdsc->num_sessions);
525 * drop session refs in request.
527 * should be last request ref, or hold mdsc->mutex
529 static void put_request_session(struct ceph_mds_request *req)
531 if (req->r_session) {
532 ceph_put_mds_session(req->r_session);
533 req->r_session = NULL;
537 void ceph_mdsc_release_request(struct kref *kref)
539 struct ceph_mds_request *req = container_of(kref,
540 struct ceph_mds_request,
542 destroy_reply_info(&req->r_reply_info);
544 ceph_msg_put(req->r_request);
546 ceph_msg_put(req->r_reply);
548 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
552 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
553 iput(req->r_target_inode);
556 if (req->r_old_dentry)
557 dput(req->r_old_dentry);
558 if (req->r_old_dentry_dir) {
560 * track (and drop pins for) r_old_dentry_dir
561 * separately, since r_old_dentry's d_parent may have
562 * changed between the dir mutex being dropped and
563 * this request being freed.
565 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
567 iput(req->r_old_dentry_dir);
572 ceph_pagelist_release(req->r_pagelist);
573 put_request_session(req);
574 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
578 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
581 * lookup session, bump ref if found.
583 * called under mdsc->mutex.
585 static struct ceph_mds_request *
586 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
588 struct ceph_mds_request *req;
590 req = lookup_request(&mdsc->request_tree, tid);
592 ceph_mdsc_get_request(req);
598 * Register an in-flight request, and assign a tid. Link to directory
599 * are modifying (if any).
601 * Called under mdsc->mutex.
603 static void __register_request(struct ceph_mds_client *mdsc,
604 struct ceph_mds_request *req,
607 req->r_tid = ++mdsc->last_tid;
609 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
611 dout("__register_request %p tid %lld\n", req, req->r_tid);
612 ceph_mdsc_get_request(req);
613 insert_request(&mdsc->request_tree, req);
615 req->r_uid = current_fsuid();
616 req->r_gid = current_fsgid();
618 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
619 mdsc->oldest_tid = req->r_tid;
623 req->r_unsafe_dir = dir;
627 static void __unregister_request(struct ceph_mds_client *mdsc,
628 struct ceph_mds_request *req)
630 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
632 /* Never leave an unregistered request on an unsafe list! */
633 list_del_init(&req->r_unsafe_item);
635 if (req->r_tid == mdsc->oldest_tid) {
636 struct rb_node *p = rb_next(&req->r_node);
637 mdsc->oldest_tid = 0;
639 struct ceph_mds_request *next_req =
640 rb_entry(p, struct ceph_mds_request, r_node);
641 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
642 mdsc->oldest_tid = next_req->r_tid;
649 erase_request(&mdsc->request_tree, req);
651 if (req->r_unsafe_dir &&
652 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
653 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
654 spin_lock(&ci->i_unsafe_lock);
655 list_del_init(&req->r_unsafe_dir_item);
656 spin_unlock(&ci->i_unsafe_lock);
658 if (req->r_target_inode &&
659 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
660 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
661 spin_lock(&ci->i_unsafe_lock);
662 list_del_init(&req->r_unsafe_target_item);
663 spin_unlock(&ci->i_unsafe_lock);
666 if (req->r_unsafe_dir) {
667 iput(req->r_unsafe_dir);
668 req->r_unsafe_dir = NULL;
671 complete_all(&req->r_safe_completion);
673 ceph_mdsc_put_request(req);
677 * Walk back up the dentry tree until we hit a dentry representing a
678 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
679 * when calling this) to ensure that the objects won't disappear while we're
680 * working with them. Once we hit a candidate dentry, we attempt to take a
681 * reference to it, and return that as the result.
683 static struct inode *get_nonsnap_parent(struct dentry *dentry)
685 struct inode *inode = NULL;
687 while (dentry && !IS_ROOT(dentry)) {
688 inode = d_inode_rcu(dentry);
689 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
691 dentry = dentry->d_parent;
694 inode = igrab(inode);
699 * Choose mds to send request to next. If there is a hint set in the
700 * request (e.g., due to a prior forward hint from the mds), use that.
701 * Otherwise, consult frag tree and/or caps to identify the
702 * appropriate mds. If all else fails, choose randomly.
704 * Called under mdsc->mutex.
706 static int __choose_mds(struct ceph_mds_client *mdsc,
707 struct ceph_mds_request *req)
710 struct ceph_inode_info *ci;
711 struct ceph_cap *cap;
712 int mode = req->r_direct_mode;
714 u32 hash = req->r_direct_hash;
715 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
718 * is there a specific mds we should try? ignore hint if we have
719 * no session and the mds is not up (active or recovering).
721 if (req->r_resend_mds >= 0 &&
722 (__have_session(mdsc, req->r_resend_mds) ||
723 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
724 dout("choose_mds using resend_mds mds%d\n",
726 return req->r_resend_mds;
729 if (mode == USE_RANDOM_MDS)
734 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
735 inode = req->r_inode;
738 /* req->r_dentry is non-null for LSSNAP request */
740 inode = get_nonsnap_parent(req->r_dentry);
742 dout("__choose_mds using snapdir's parent %p\n", inode);
744 } else if (req->r_dentry) {
745 /* ignore race with rename; old or new d_parent is okay */
746 struct dentry *parent;
750 parent = req->r_dentry->d_parent;
751 dir = req->r_parent ? : d_inode_rcu(parent);
753 if (!dir || dir->i_sb != mdsc->fsc->sb) {
754 /* not this fs or parent went negative */
755 inode = d_inode(req->r_dentry);
758 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
759 /* direct snapped/virtual snapdir requests
760 * based on parent dir inode */
761 inode = get_nonsnap_parent(parent);
762 dout("__choose_mds using nonsnap parent %p\n", inode);
765 inode = d_inode(req->r_dentry);
766 if (!inode || mode == USE_AUTH_MDS) {
769 hash = ceph_dentry_hash(dir, req->r_dentry);
778 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
782 ci = ceph_inode(inode);
784 if (is_hash && S_ISDIR(inode->i_mode)) {
785 struct ceph_inode_frag frag;
788 ceph_choose_frag(ci, hash, &frag, &found);
790 if (mode == USE_ANY_MDS && frag.ndist > 0) {
793 /* choose a random replica */
794 get_random_bytes(&r, 1);
797 dout("choose_mds %p %llx.%llx "
798 "frag %u mds%d (%d/%d)\n",
799 inode, ceph_vinop(inode),
802 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
803 CEPH_MDS_STATE_ACTIVE)
807 /* since this file/dir wasn't known to be
808 * replicated, then we want to look for the
809 * authoritative mds. */
812 /* choose auth mds */
814 dout("choose_mds %p %llx.%llx "
815 "frag %u mds%d (auth)\n",
816 inode, ceph_vinop(inode), frag.frag, mds);
817 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
818 CEPH_MDS_STATE_ACTIVE)
824 spin_lock(&ci->i_ceph_lock);
826 if (mode == USE_AUTH_MDS)
827 cap = ci->i_auth_cap;
828 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
829 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
831 spin_unlock(&ci->i_ceph_lock);
835 mds = cap->session->s_mds;
836 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
837 inode, ceph_vinop(inode), mds,
838 cap == ci->i_auth_cap ? "auth " : "", cap);
839 spin_unlock(&ci->i_ceph_lock);
845 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
846 dout("choose_mds chose random mds%d\n", mds);
854 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
856 struct ceph_msg *msg;
857 struct ceph_mds_session_head *h;
859 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
862 pr_err("create_session_msg ENOMEM creating msg\n");
865 h = msg->front.iov_base;
866 h->op = cpu_to_le32(op);
867 h->seq = cpu_to_le64(seq);
873 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
874 * to include additional client metadata fields.
876 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
878 struct ceph_msg *msg;
879 struct ceph_mds_session_head *h;
881 int metadata_bytes = 0;
882 int metadata_key_count = 0;
883 struct ceph_options *opt = mdsc->fsc->client->options;
884 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
887 const char* metadata[][2] = {
888 {"hostname", mdsc->nodename},
889 {"kernel_version", init_utsname()->release},
890 {"entity_id", opt->name ? : ""},
891 {"root", fsopt->server_path ? : "/"},
895 /* Calculate serialized length of metadata */
896 metadata_bytes = 4; /* map length */
897 for (i = 0; metadata[i][0]; ++i) {
898 metadata_bytes += 8 + strlen(metadata[i][0]) +
899 strlen(metadata[i][1]);
900 metadata_key_count++;
903 /* Allocate the message */
904 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
907 pr_err("create_session_msg ENOMEM creating msg\n");
910 h = msg->front.iov_base;
911 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
912 h->seq = cpu_to_le64(seq);
915 * Serialize client metadata into waiting buffer space, using
916 * the format that userspace expects for map<string, string>
918 * ClientSession messages with metadata are v2
920 msg->hdr.version = cpu_to_le16(2);
921 msg->hdr.compat_version = cpu_to_le16(1);
923 /* The write pointer, following the session_head structure */
924 p = msg->front.iov_base + sizeof(*h);
926 /* Number of entries in the map */
927 ceph_encode_32(&p, metadata_key_count);
929 /* Two length-prefixed strings for each entry in the map */
930 for (i = 0; metadata[i][0]; ++i) {
931 size_t const key_len = strlen(metadata[i][0]);
932 size_t const val_len = strlen(metadata[i][1]);
934 ceph_encode_32(&p, key_len);
935 memcpy(p, metadata[i][0], key_len);
937 ceph_encode_32(&p, val_len);
938 memcpy(p, metadata[i][1], val_len);
946 * send session open request.
948 * called under mdsc->mutex
950 static int __open_session(struct ceph_mds_client *mdsc,
951 struct ceph_mds_session *session)
953 struct ceph_msg *msg;
955 int mds = session->s_mds;
957 /* wait for mds to go active? */
958 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
959 dout("open_session to mds%d (%s)\n", mds,
960 ceph_mds_state_name(mstate));
961 session->s_state = CEPH_MDS_SESSION_OPENING;
962 session->s_renew_requested = jiffies;
964 /* send connect message */
965 msg = create_session_open_msg(mdsc, session->s_seq);
968 ceph_con_send(&session->s_con, msg);
973 * open sessions for any export targets for the given mds
975 * called under mdsc->mutex
977 static struct ceph_mds_session *
978 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
980 struct ceph_mds_session *session;
982 session = __ceph_lookup_mds_session(mdsc, target);
984 session = register_session(mdsc, target);
988 if (session->s_state == CEPH_MDS_SESSION_NEW ||
989 session->s_state == CEPH_MDS_SESSION_CLOSING)
990 __open_session(mdsc, session);
995 struct ceph_mds_session *
996 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
998 struct ceph_mds_session *session;
1000 dout("open_export_target_session to mds%d\n", target);
1002 mutex_lock(&mdsc->mutex);
1003 session = __open_export_target_session(mdsc, target);
1004 mutex_unlock(&mdsc->mutex);
1009 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1010 struct ceph_mds_session *session)
1012 struct ceph_mds_info *mi;
1013 struct ceph_mds_session *ts;
1014 int i, mds = session->s_mds;
1016 if (mds >= mdsc->mdsmap->m_num_mds)
1019 mi = &mdsc->mdsmap->m_info[mds];
1020 dout("open_export_target_sessions for mds%d (%d targets)\n",
1021 session->s_mds, mi->num_export_targets);
1023 for (i = 0; i < mi->num_export_targets; i++) {
1024 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1026 ceph_put_mds_session(ts);
1030 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1031 struct ceph_mds_session *session)
1033 mutex_lock(&mdsc->mutex);
1034 __open_export_target_sessions(mdsc, session);
1035 mutex_unlock(&mdsc->mutex);
1042 /* caller holds s_cap_lock, we drop it */
1043 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1044 struct ceph_mds_session *session)
1045 __releases(session->s_cap_lock)
1047 LIST_HEAD(tmp_list);
1048 list_splice_init(&session->s_cap_releases, &tmp_list);
1049 session->s_num_cap_releases = 0;
1050 spin_unlock(&session->s_cap_lock);
1052 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1053 while (!list_empty(&tmp_list)) {
1054 struct ceph_cap *cap;
1055 /* zero out the in-progress message */
1056 cap = list_first_entry(&tmp_list,
1057 struct ceph_cap, session_caps);
1058 list_del(&cap->session_caps);
1059 ceph_put_cap(mdsc, cap);
1063 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1064 struct ceph_mds_session *session)
1066 struct ceph_mds_request *req;
1069 dout("cleanup_session_requests mds%d\n", session->s_mds);
1070 mutex_lock(&mdsc->mutex);
1071 while (!list_empty(&session->s_unsafe)) {
1072 req = list_first_entry(&session->s_unsafe,
1073 struct ceph_mds_request, r_unsafe_item);
1074 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1076 __unregister_request(mdsc, req);
1078 /* zero r_attempts, so kick_requests() will re-send requests */
1079 p = rb_first(&mdsc->request_tree);
1081 req = rb_entry(p, struct ceph_mds_request, r_node);
1083 if (req->r_session &&
1084 req->r_session->s_mds == session->s_mds)
1085 req->r_attempts = 0;
1087 mutex_unlock(&mdsc->mutex);
1091 * Helper to safely iterate over all caps associated with a session, with
1092 * special care taken to handle a racing __ceph_remove_cap().
1094 * Caller must hold session s_mutex.
1096 static int iterate_session_caps(struct ceph_mds_session *session,
1097 int (*cb)(struct inode *, struct ceph_cap *,
1100 struct list_head *p;
1101 struct ceph_cap *cap;
1102 struct inode *inode, *last_inode = NULL;
1103 struct ceph_cap *old_cap = NULL;
1106 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1107 spin_lock(&session->s_cap_lock);
1108 p = session->s_caps.next;
1109 while (p != &session->s_caps) {
1110 cap = list_entry(p, struct ceph_cap, session_caps);
1111 inode = igrab(&cap->ci->vfs_inode);
1116 session->s_cap_iterator = cap;
1117 spin_unlock(&session->s_cap_lock);
1124 ceph_put_cap(session->s_mdsc, old_cap);
1128 ret = cb(inode, cap, arg);
1131 spin_lock(&session->s_cap_lock);
1134 dout("iterate_session_caps finishing cap %p removal\n",
1136 BUG_ON(cap->session != session);
1137 cap->session = NULL;
1138 list_del_init(&cap->session_caps);
1139 session->s_nr_caps--;
1140 if (cap->queue_release) {
1141 list_add_tail(&cap->session_caps,
1142 &session->s_cap_releases);
1143 session->s_num_cap_releases++;
1145 old_cap = cap; /* put_cap it w/o locks held */
1153 session->s_cap_iterator = NULL;
1154 spin_unlock(&session->s_cap_lock);
1158 ceph_put_cap(session->s_mdsc, old_cap);
1163 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1166 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1167 struct ceph_inode_info *ci = ceph_inode(inode);
1168 LIST_HEAD(to_remove);
1170 bool invalidate = false;
1172 dout("removing cap %p, ci is %p, inode is %p\n",
1173 cap, ci, &ci->vfs_inode);
1174 spin_lock(&ci->i_ceph_lock);
1175 __ceph_remove_cap(cap, false);
1176 if (!ci->i_auth_cap) {
1177 struct ceph_cap_flush *cf;
1178 struct ceph_mds_client *mdsc = fsc->mdsc;
1180 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1182 if (ci->i_wrbuffer_ref > 0 &&
1183 READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1186 while (!list_empty(&ci->i_cap_flush_list)) {
1187 cf = list_first_entry(&ci->i_cap_flush_list,
1188 struct ceph_cap_flush, i_list);
1189 list_move(&cf->i_list, &to_remove);
1192 spin_lock(&mdsc->cap_dirty_lock);
1194 list_for_each_entry(cf, &to_remove, i_list)
1195 list_del(&cf->g_list);
1197 if (!list_empty(&ci->i_dirty_item)) {
1198 pr_warn_ratelimited(
1199 " dropping dirty %s state for %p %lld\n",
1200 ceph_cap_string(ci->i_dirty_caps),
1201 inode, ceph_ino(inode));
1202 ci->i_dirty_caps = 0;
1203 list_del_init(&ci->i_dirty_item);
1206 if (!list_empty(&ci->i_flushing_item)) {
1207 pr_warn_ratelimited(
1208 " dropping dirty+flushing %s state for %p %lld\n",
1209 ceph_cap_string(ci->i_flushing_caps),
1210 inode, ceph_ino(inode));
1211 ci->i_flushing_caps = 0;
1212 list_del_init(&ci->i_flushing_item);
1213 mdsc->num_cap_flushing--;
1216 spin_unlock(&mdsc->cap_dirty_lock);
1218 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1219 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1220 ci->i_prealloc_cap_flush = NULL;
1224 ci->i_wrbuffer_ref_head == 0 &&
1225 ci->i_wr_ref == 0 &&
1226 ci->i_dirty_caps == 0 &&
1227 ci->i_flushing_caps == 0) {
1228 ceph_put_snap_context(ci->i_head_snapc);
1229 ci->i_head_snapc = NULL;
1232 spin_unlock(&ci->i_ceph_lock);
1233 while (!list_empty(&to_remove)) {
1234 struct ceph_cap_flush *cf;
1235 cf = list_first_entry(&to_remove,
1236 struct ceph_cap_flush, i_list);
1237 list_del(&cf->i_list);
1238 ceph_free_cap_flush(cf);
1241 wake_up_all(&ci->i_cap_wq);
1243 ceph_queue_invalidate(inode);
1250 * caller must hold session s_mutex
1252 static void remove_session_caps(struct ceph_mds_session *session)
1254 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1255 struct super_block *sb = fsc->sb;
1256 dout("remove_session_caps on %p\n", session);
1257 iterate_session_caps(session, remove_session_caps_cb, fsc);
1259 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1261 spin_lock(&session->s_cap_lock);
1262 if (session->s_nr_caps > 0) {
1263 struct inode *inode;
1264 struct ceph_cap *cap, *prev = NULL;
1265 struct ceph_vino vino;
1267 * iterate_session_caps() skips inodes that are being
1268 * deleted, we need to wait until deletions are complete.
1269 * __wait_on_freeing_inode() is designed for the job,
1270 * but it is not exported, so use lookup inode function
1273 while (!list_empty(&session->s_caps)) {
1274 cap = list_entry(session->s_caps.next,
1275 struct ceph_cap, session_caps);
1279 vino = cap->ci->i_vino;
1280 spin_unlock(&session->s_cap_lock);
1282 inode = ceph_find_inode(sb, vino);
1285 spin_lock(&session->s_cap_lock);
1289 // drop cap expires and unlock s_cap_lock
1290 cleanup_cap_releases(session->s_mdsc, session);
1292 BUG_ON(session->s_nr_caps > 0);
1293 BUG_ON(!list_empty(&session->s_cap_flushing));
1297 * wake up any threads waiting on this session's caps. if the cap is
1298 * old (didn't get renewed on the client reconnect), remove it now.
1300 * caller must hold s_mutex.
1302 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1305 struct ceph_inode_info *ci = ceph_inode(inode);
1308 spin_lock(&ci->i_ceph_lock);
1309 ci->i_wanted_max_size = 0;
1310 ci->i_requested_max_size = 0;
1311 spin_unlock(&ci->i_ceph_lock);
1313 wake_up_all(&ci->i_cap_wq);
1317 static void wake_up_session_caps(struct ceph_mds_session *session,
1320 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1321 iterate_session_caps(session, wake_up_session_cb,
1322 (void *)(unsigned long)reconnect);
1326 * Send periodic message to MDS renewing all currently held caps. The
1327 * ack will reset the expiration for all caps from this session.
1329 * caller holds s_mutex
1331 static int send_renew_caps(struct ceph_mds_client *mdsc,
1332 struct ceph_mds_session *session)
1334 struct ceph_msg *msg;
1337 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1338 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1339 pr_info("mds%d caps stale\n", session->s_mds);
1340 session->s_renew_requested = jiffies;
1342 /* do not try to renew caps until a recovering mds has reconnected
1343 * with its clients. */
1344 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1345 if (state < CEPH_MDS_STATE_RECONNECT) {
1346 dout("send_renew_caps ignoring mds%d (%s)\n",
1347 session->s_mds, ceph_mds_state_name(state));
1351 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1352 ceph_mds_state_name(state));
1353 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1354 ++session->s_renew_seq);
1357 ceph_con_send(&session->s_con, msg);
1361 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1362 struct ceph_mds_session *session, u64 seq)
1364 struct ceph_msg *msg;
1366 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1367 session->s_mds, ceph_session_state_name(session->s_state), seq);
1368 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1371 ceph_con_send(&session->s_con, msg);
1377 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1379 * Called under session->s_mutex
1381 static void renewed_caps(struct ceph_mds_client *mdsc,
1382 struct ceph_mds_session *session, int is_renew)
1387 spin_lock(&session->s_cap_lock);
1388 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1390 session->s_cap_ttl = session->s_renew_requested +
1391 mdsc->mdsmap->m_session_timeout*HZ;
1394 if (time_before(jiffies, session->s_cap_ttl)) {
1395 pr_info("mds%d caps renewed\n", session->s_mds);
1398 pr_info("mds%d caps still stale\n", session->s_mds);
1401 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1402 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1403 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1404 spin_unlock(&session->s_cap_lock);
1407 wake_up_session_caps(session, 0);
1411 * send a session close request
1413 static int request_close_session(struct ceph_mds_client *mdsc,
1414 struct ceph_mds_session *session)
1416 struct ceph_msg *msg;
1418 dout("request_close_session mds%d state %s seq %lld\n",
1419 session->s_mds, ceph_session_state_name(session->s_state),
1421 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1424 ceph_con_send(&session->s_con, msg);
1429 * Called with s_mutex held.
1431 static int __close_session(struct ceph_mds_client *mdsc,
1432 struct ceph_mds_session *session)
1434 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1436 session->s_state = CEPH_MDS_SESSION_CLOSING;
1437 return request_close_session(mdsc, session);
1440 static bool drop_negative_children(struct dentry *dentry)
1442 struct dentry *child;
1443 bool all_negative = true;
1445 if (!d_is_dir(dentry))
1448 spin_lock(&dentry->d_lock);
1449 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1450 if (d_really_is_positive(child)) {
1451 all_negative = false;
1455 spin_unlock(&dentry->d_lock);
1458 shrink_dcache_parent(dentry);
1460 return all_negative;
1464 * Trim old(er) caps.
1466 * Because we can't cache an inode without one or more caps, we do
1467 * this indirectly: if a cap is unused, we prune its aliases, at which
1468 * point the inode will hopefully get dropped to.
1470 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1471 * memory pressure from the MDS, though, so it needn't be perfect.
1473 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1475 struct ceph_mds_session *session = arg;
1476 struct ceph_inode_info *ci = ceph_inode(inode);
1477 int used, wanted, oissued, mine;
1479 if (session->s_trim_caps <= 0)
1482 spin_lock(&ci->i_ceph_lock);
1483 mine = cap->issued | cap->implemented;
1484 used = __ceph_caps_used(ci);
1485 wanted = __ceph_caps_file_wanted(ci);
1486 oissued = __ceph_caps_issued_other(ci, cap);
1488 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1489 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1490 ceph_cap_string(used), ceph_cap_string(wanted));
1491 if (cap == ci->i_auth_cap) {
1492 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1493 !list_empty(&ci->i_cap_snaps))
1495 if ((used | wanted) & CEPH_CAP_ANY_WR)
1498 /* The inode has cached pages, but it's no longer used.
1499 * we can safely drop it */
1500 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1501 !(oissued & CEPH_CAP_FILE_CACHE)) {
1505 if ((used | wanted) & ~oissued & mine)
1506 goto out; /* we need these caps */
1509 /* we aren't the only cap.. just remove us */
1510 __ceph_remove_cap(cap, true);
1511 session->s_trim_caps--;
1513 struct dentry *dentry;
1514 /* try dropping referring dentries */
1515 spin_unlock(&ci->i_ceph_lock);
1516 dentry = d_find_any_alias(inode);
1517 if (dentry && drop_negative_children(dentry)) {
1520 d_prune_aliases(inode);
1521 count = atomic_read(&inode->i_count);
1523 session->s_trim_caps--;
1524 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1533 spin_unlock(&ci->i_ceph_lock);
1538 * Trim session cap count down to some max number.
1540 static int trim_caps(struct ceph_mds_client *mdsc,
1541 struct ceph_mds_session *session,
1544 int trim_caps = session->s_nr_caps - max_caps;
1546 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1547 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1548 if (trim_caps > 0) {
1549 session->s_trim_caps = trim_caps;
1550 iterate_session_caps(session, trim_caps_cb, session);
1551 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1552 session->s_mds, session->s_nr_caps, max_caps,
1553 trim_caps - session->s_trim_caps);
1554 session->s_trim_caps = 0;
1557 ceph_send_cap_releases(mdsc, session);
1561 static int check_caps_flush(struct ceph_mds_client *mdsc,
1566 spin_lock(&mdsc->cap_dirty_lock);
1567 if (!list_empty(&mdsc->cap_flush_list)) {
1568 struct ceph_cap_flush *cf =
1569 list_first_entry(&mdsc->cap_flush_list,
1570 struct ceph_cap_flush, g_list);
1571 if (cf->tid <= want_flush_tid) {
1572 dout("check_caps_flush still flushing tid "
1573 "%llu <= %llu\n", cf->tid, want_flush_tid);
1577 spin_unlock(&mdsc->cap_dirty_lock);
1582 * flush all dirty inode data to disk.
1584 * returns true if we've flushed through want_flush_tid
1586 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1589 dout("check_caps_flush want %llu\n", want_flush_tid);
1591 wait_event(mdsc->cap_flushing_wq,
1592 check_caps_flush(mdsc, want_flush_tid));
1594 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1598 * called under s_mutex
1600 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1601 struct ceph_mds_session *session)
1603 struct ceph_msg *msg = NULL;
1604 struct ceph_mds_cap_release *head;
1605 struct ceph_mds_cap_item *item;
1606 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1607 struct ceph_cap *cap;
1608 LIST_HEAD(tmp_list);
1609 int num_cap_releases;
1610 __le32 barrier, *cap_barrier;
1612 down_read(&osdc->lock);
1613 barrier = cpu_to_le32(osdc->epoch_barrier);
1614 up_read(&osdc->lock);
1616 spin_lock(&session->s_cap_lock);
1618 list_splice_init(&session->s_cap_releases, &tmp_list);
1619 num_cap_releases = session->s_num_cap_releases;
1620 session->s_num_cap_releases = 0;
1621 spin_unlock(&session->s_cap_lock);
1623 while (!list_empty(&tmp_list)) {
1625 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1626 PAGE_SIZE, GFP_NOFS, false);
1629 head = msg->front.iov_base;
1630 head->num = cpu_to_le32(0);
1631 msg->front.iov_len = sizeof(*head);
1633 msg->hdr.version = cpu_to_le16(2);
1634 msg->hdr.compat_version = cpu_to_le16(1);
1637 cap = list_first_entry(&tmp_list, struct ceph_cap,
1639 list_del(&cap->session_caps);
1642 head = msg->front.iov_base;
1643 le32_add_cpu(&head->num, 1);
1644 item = msg->front.iov_base + msg->front.iov_len;
1645 item->ino = cpu_to_le64(cap->cap_ino);
1646 item->cap_id = cpu_to_le64(cap->cap_id);
1647 item->migrate_seq = cpu_to_le32(cap->mseq);
1648 item->seq = cpu_to_le32(cap->issue_seq);
1649 msg->front.iov_len += sizeof(*item);
1651 ceph_put_cap(mdsc, cap);
1653 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1654 // Append cap_barrier field
1655 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1656 *cap_barrier = barrier;
1657 msg->front.iov_len += sizeof(*cap_barrier);
1659 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1660 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1661 ceph_con_send(&session->s_con, msg);
1666 BUG_ON(num_cap_releases != 0);
1668 spin_lock(&session->s_cap_lock);
1669 if (!list_empty(&session->s_cap_releases))
1671 spin_unlock(&session->s_cap_lock);
1674 // Append cap_barrier field
1675 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1676 *cap_barrier = barrier;
1677 msg->front.iov_len += sizeof(*cap_barrier);
1679 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1680 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1681 ceph_con_send(&session->s_con, msg);
1685 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1687 spin_lock(&session->s_cap_lock);
1688 list_splice(&tmp_list, &session->s_cap_releases);
1689 session->s_num_cap_releases += num_cap_releases;
1690 spin_unlock(&session->s_cap_lock);
1697 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1700 struct ceph_inode_info *ci = ceph_inode(dir);
1701 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1702 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1703 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1704 int order, num_entries;
1706 spin_lock(&ci->i_ceph_lock);
1707 num_entries = ci->i_files + ci->i_subdirs;
1708 spin_unlock(&ci->i_ceph_lock);
1709 num_entries = max(num_entries, 1);
1710 num_entries = min(num_entries, opt->max_readdir);
1712 order = get_order(size * num_entries);
1713 while (order >= 0) {
1714 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1717 if (rinfo->dir_entries)
1721 if (!rinfo->dir_entries)
1724 num_entries = (PAGE_SIZE << order) / size;
1725 num_entries = min(num_entries, opt->max_readdir);
1727 rinfo->dir_buf_size = PAGE_SIZE << order;
1728 req->r_num_caps = num_entries + 1;
1729 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1730 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1735 * Create an mds request.
1737 struct ceph_mds_request *
1738 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1740 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1743 return ERR_PTR(-ENOMEM);
1745 mutex_init(&req->r_fill_mutex);
1747 req->r_started = jiffies;
1748 req->r_resend_mds = -1;
1749 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1750 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1752 kref_init(&req->r_kref);
1753 RB_CLEAR_NODE(&req->r_node);
1754 INIT_LIST_HEAD(&req->r_wait);
1755 init_completion(&req->r_completion);
1756 init_completion(&req->r_safe_completion);
1757 INIT_LIST_HEAD(&req->r_unsafe_item);
1759 req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1762 req->r_direct_mode = mode;
1767 * return oldest (lowest) request, tid in request tree, 0 if none.
1769 * called under mdsc->mutex.
1771 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1773 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1775 return rb_entry(rb_first(&mdsc->request_tree),
1776 struct ceph_mds_request, r_node);
1779 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1781 return mdsc->oldest_tid;
1785 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1786 * on build_path_from_dentry in fs/cifs/dir.c.
1788 * If @stop_on_nosnap, generate path relative to the first non-snapped
1791 * Encode hidden .snap dirs as a double /, i.e.
1792 * foo/.snap/bar -> foo//bar
1794 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1797 struct dentry *temp;
1803 return ERR_PTR(-EINVAL);
1807 seq = read_seqbegin(&rename_lock);
1809 for (temp = dentry; !IS_ROOT(temp);) {
1810 struct inode *inode = d_inode(temp);
1811 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1812 len++; /* slash only */
1813 else if (stop_on_nosnap && inode &&
1814 ceph_snap(inode) == CEPH_NOSNAP)
1817 len += 1 + temp->d_name.len;
1818 temp = temp->d_parent;
1822 len--; /* no leading '/' */
1824 path = kmalloc(len+1, GFP_NOFS);
1826 return ERR_PTR(-ENOMEM);
1828 path[pos] = 0; /* trailing null */
1830 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1831 struct inode *inode;
1833 spin_lock(&temp->d_lock);
1834 inode = d_inode(temp);
1835 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1836 dout("build_path path+%d: %p SNAPDIR\n",
1838 } else if (stop_on_nosnap && inode &&
1839 ceph_snap(inode) == CEPH_NOSNAP) {
1840 spin_unlock(&temp->d_lock);
1843 pos -= temp->d_name.len;
1845 spin_unlock(&temp->d_lock);
1848 strncpy(path + pos, temp->d_name.name,
1851 spin_unlock(&temp->d_lock);
1854 temp = temp->d_parent;
1857 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1858 pr_err("build_path did not end path lookup where "
1859 "expected, namelen is %d, pos is %d\n", len, pos);
1860 /* presumably this is only possible if racing with a
1861 rename of one of the parent directories (we can not
1862 lock the dentries above us to prevent this, but
1863 retrying should be harmless) */
1868 *base = ceph_ino(d_inode(temp));
1870 dout("build_path on %p %d built %llx '%.*s'\n",
1871 dentry, d_count(dentry), *base, len, path);
1875 /* Duplicate the dentry->d_name.name safely */
1876 static int clone_dentry_name(struct dentry *dentry, const char **ppath,
1883 len = READ_ONCE(dentry->d_name.len);
1884 name = kmalloc(len + 1, GFP_NOFS);
1888 spin_lock(&dentry->d_lock);
1889 if (dentry->d_name.len != len) {
1890 spin_unlock(&dentry->d_lock);
1894 memcpy(name, dentry->d_name.name, len);
1895 spin_unlock(&dentry->d_lock);
1903 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1904 const char **ppath, int *ppathlen, u64 *pino,
1905 bool *pfreepath, bool parent_locked)
1912 dir = d_inode_rcu(dentry->d_parent);
1913 if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1914 *pino = ceph_ino(dir);
1916 if (parent_locked) {
1917 *ppath = dentry->d_name.name;
1918 *ppathlen = dentry->d_name.len;
1920 ret = clone_dentry_name(dentry, ppath, ppathlen);
1928 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1930 return PTR_ERR(path);
1936 static int build_inode_path(struct inode *inode,
1937 const char **ppath, int *ppathlen, u64 *pino,
1940 struct dentry *dentry;
1943 if (ceph_snap(inode) == CEPH_NOSNAP) {
1944 *pino = ceph_ino(inode);
1948 dentry = d_find_alias(inode);
1949 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1952 return PTR_ERR(path);
1959 * request arguments may be specified via an inode *, a dentry *, or
1960 * an explicit ino+path.
1962 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1963 struct inode *rdiri, const char *rpath,
1964 u64 rino, const char **ppath, int *pathlen,
1965 u64 *ino, bool *freepath, bool parent_locked)
1970 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1971 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1973 } else if (rdentry) {
1974 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1975 freepath, parent_locked);
1976 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1978 } else if (rpath || rino) {
1981 *pathlen = rpath ? strlen(rpath) : 0;
1982 dout(" path %.*s\n", *pathlen, rpath);
1989 * called under mdsc->mutex
1991 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1992 struct ceph_mds_request *req,
1993 int mds, bool drop_cap_releases)
1995 struct ceph_msg *msg;
1996 struct ceph_mds_request_head *head;
1997 const char *path1 = NULL;
1998 const char *path2 = NULL;
1999 u64 ino1 = 0, ino2 = 0;
2000 int pathlen1 = 0, pathlen2 = 0;
2001 bool freepath1 = false, freepath2 = false;
2007 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2008 req->r_parent, req->r_path1, req->r_ino1.ino,
2009 &path1, &pathlen1, &ino1, &freepath1,
2010 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2011 &req->r_req_flags));
2017 /* If r_old_dentry is set, then assume that its parent is locked */
2018 ret = set_request_path_attr(NULL, req->r_old_dentry,
2019 req->r_old_dentry_dir,
2020 req->r_path2, req->r_ino2.ino,
2021 &path2, &pathlen2, &ino2, &freepath2, true);
2027 len = sizeof(*head) +
2028 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2029 sizeof(struct ceph_timespec);
2031 /* calculate (max) length for cap releases */
2032 len += sizeof(struct ceph_mds_request_release) *
2033 (!!req->r_inode_drop + !!req->r_dentry_drop +
2034 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2035 if (req->r_dentry_drop)
2036 len += req->r_dentry->d_name.len;
2037 if (req->r_old_dentry_drop)
2038 len += req->r_old_dentry->d_name.len;
2040 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
2042 msg = ERR_PTR(-ENOMEM);
2046 msg->hdr.version = cpu_to_le16(2);
2047 msg->hdr.tid = cpu_to_le64(req->r_tid);
2049 head = msg->front.iov_base;
2050 p = msg->front.iov_base + sizeof(*head);
2051 end = msg->front.iov_base + msg->front.iov_len;
2053 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2054 head->op = cpu_to_le32(req->r_op);
2055 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2056 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2057 head->args = req->r_args;
2059 ceph_encode_filepath(&p, end, ino1, path1);
2060 ceph_encode_filepath(&p, end, ino2, path2);
2062 /* make note of release offset, in case we need to replay */
2063 req->r_request_release_offset = p - msg->front.iov_base;
2067 if (req->r_inode_drop)
2068 releases += ceph_encode_inode_release(&p,
2069 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2070 mds, req->r_inode_drop, req->r_inode_unless, 0);
2071 if (req->r_dentry_drop)
2072 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2073 req->r_parent, mds, req->r_dentry_drop,
2074 req->r_dentry_unless);
2075 if (req->r_old_dentry_drop)
2076 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2077 req->r_old_dentry_dir, mds,
2078 req->r_old_dentry_drop,
2079 req->r_old_dentry_unless);
2080 if (req->r_old_inode_drop)
2081 releases += ceph_encode_inode_release(&p,
2082 d_inode(req->r_old_dentry),
2083 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2085 if (drop_cap_releases) {
2087 p = msg->front.iov_base + req->r_request_release_offset;
2090 head->num_releases = cpu_to_le16(releases);
2094 struct ceph_timespec ts;
2095 ceph_encode_timespec(&ts, &req->r_stamp);
2096 ceph_encode_copy(&p, &ts, sizeof(ts));
2100 msg->front.iov_len = p - msg->front.iov_base;
2101 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2103 if (req->r_pagelist) {
2104 struct ceph_pagelist *pagelist = req->r_pagelist;
2105 refcount_inc(&pagelist->refcnt);
2106 ceph_msg_data_add_pagelist(msg, pagelist);
2107 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2109 msg->hdr.data_len = 0;
2112 msg->hdr.data_off = cpu_to_le16(0);
2116 kfree((char *)path2);
2119 kfree((char *)path1);
2125 * called under mdsc->mutex if error, under no mutex if
2128 static void complete_request(struct ceph_mds_client *mdsc,
2129 struct ceph_mds_request *req)
2131 if (req->r_callback)
2132 req->r_callback(mdsc, req);
2134 complete_all(&req->r_completion);
2138 * called under mdsc->mutex
2140 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2141 struct ceph_mds_request *req,
2142 int mds, bool drop_cap_releases)
2144 struct ceph_mds_request_head *rhead;
2145 struct ceph_msg *msg;
2150 struct ceph_cap *cap =
2151 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2154 req->r_sent_on_mseq = cap->mseq;
2156 req->r_sent_on_mseq = -1;
2158 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2159 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2161 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2164 * Replay. Do not regenerate message (and rebuild
2165 * paths, etc.); just use the original message.
2166 * Rebuilding paths will break for renames because
2167 * d_move mangles the src name.
2169 msg = req->r_request;
2170 rhead = msg->front.iov_base;
2172 flags = le32_to_cpu(rhead->flags);
2173 flags |= CEPH_MDS_FLAG_REPLAY;
2174 rhead->flags = cpu_to_le32(flags);
2176 if (req->r_target_inode)
2177 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2179 rhead->num_retry = req->r_attempts - 1;
2181 /* remove cap/dentry releases from message */
2182 rhead->num_releases = 0;
2185 p = msg->front.iov_base + req->r_request_release_offset;
2187 struct ceph_timespec ts;
2188 ceph_encode_timespec(&ts, &req->r_stamp);
2189 ceph_encode_copy(&p, &ts, sizeof(ts));
2192 msg->front.iov_len = p - msg->front.iov_base;
2193 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2197 if (req->r_request) {
2198 ceph_msg_put(req->r_request);
2199 req->r_request = NULL;
2201 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2203 req->r_err = PTR_ERR(msg);
2204 return PTR_ERR(msg);
2206 req->r_request = msg;
2208 rhead = msg->front.iov_base;
2209 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2210 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2211 flags |= CEPH_MDS_FLAG_REPLAY;
2213 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2214 rhead->flags = cpu_to_le32(flags);
2215 rhead->num_fwd = req->r_num_fwd;
2216 rhead->num_retry = req->r_attempts - 1;
2219 dout(" r_parent = %p\n", req->r_parent);
2224 * send request, or put it on the appropriate wait list.
2226 static int __do_request(struct ceph_mds_client *mdsc,
2227 struct ceph_mds_request *req)
2229 struct ceph_mds_session *session = NULL;
2233 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2234 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2235 __unregister_request(mdsc, req);
2239 if (req->r_timeout &&
2240 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2241 dout("do_request timed out\n");
2245 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2246 dout("do_request forced umount\n");
2250 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2251 if (mdsc->mdsmap_err) {
2252 err = mdsc->mdsmap_err;
2253 dout("do_request mdsmap err %d\n", err);
2256 if (mdsc->mdsmap->m_epoch == 0) {
2257 dout("do_request no mdsmap, waiting for map\n");
2258 list_add(&req->r_wait, &mdsc->waiting_for_map);
2261 if (!(mdsc->fsc->mount_options->flags &
2262 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2263 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2264 err = -EHOSTUNREACH;
2269 put_request_session(req);
2271 mds = __choose_mds(mdsc, req);
2273 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2274 dout("do_request no mds or not active, waiting for map\n");
2275 list_add(&req->r_wait, &mdsc->waiting_for_map);
2279 /* get, open session */
2280 session = __ceph_lookup_mds_session(mdsc, mds);
2282 session = register_session(mdsc, mds);
2283 if (IS_ERR(session)) {
2284 err = PTR_ERR(session);
2288 req->r_session = get_session(session);
2290 dout("do_request mds%d session %p state %s\n", mds, session,
2291 ceph_session_state_name(session->s_state));
2292 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2293 session->s_state != CEPH_MDS_SESSION_HUNG) {
2294 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2298 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2299 session->s_state == CEPH_MDS_SESSION_CLOSING)
2300 __open_session(mdsc, session);
2301 list_add(&req->r_wait, &session->s_waiting);
2306 req->r_resend_mds = -1; /* forget any previous mds hint */
2308 if (req->r_request_started == 0) /* note request start time */
2309 req->r_request_started = jiffies;
2311 err = __prepare_send_request(mdsc, req, mds, false);
2313 ceph_msg_get(req->r_request);
2314 ceph_con_send(&session->s_con, req->r_request);
2318 ceph_put_mds_session(session);
2321 dout("__do_request early error %d\n", err);
2323 complete_request(mdsc, req);
2324 __unregister_request(mdsc, req);
2331 * called under mdsc->mutex
2333 static void __wake_requests(struct ceph_mds_client *mdsc,
2334 struct list_head *head)
2336 struct ceph_mds_request *req;
2337 LIST_HEAD(tmp_list);
2339 list_splice_init(head, &tmp_list);
2341 while (!list_empty(&tmp_list)) {
2342 req = list_entry(tmp_list.next,
2343 struct ceph_mds_request, r_wait);
2344 list_del_init(&req->r_wait);
2345 dout(" wake request %p tid %llu\n", req, req->r_tid);
2346 __do_request(mdsc, req);
2351 * Wake up threads with requests pending for @mds, so that they can
2352 * resubmit their requests to a possibly different mds.
2354 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2356 struct ceph_mds_request *req;
2357 struct rb_node *p = rb_first(&mdsc->request_tree);
2359 dout("kick_requests mds%d\n", mds);
2361 req = rb_entry(p, struct ceph_mds_request, r_node);
2363 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2365 if (req->r_attempts > 0)
2366 continue; /* only new requests */
2367 if (req->r_session &&
2368 req->r_session->s_mds == mds) {
2369 dout(" kicking tid %llu\n", req->r_tid);
2370 list_del_init(&req->r_wait);
2371 __do_request(mdsc, req);
2376 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2377 struct ceph_mds_request *req)
2379 dout("submit_request on %p\n", req);
2380 mutex_lock(&mdsc->mutex);
2381 __register_request(mdsc, req, NULL);
2382 __do_request(mdsc, req);
2383 mutex_unlock(&mdsc->mutex);
2387 * Synchrously perform an mds request. Take care of all of the
2388 * session setup, forwarding, retry details.
2390 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2392 struct ceph_mds_request *req)
2396 dout("do_request on %p\n", req);
2398 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2400 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2402 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2403 if (req->r_old_dentry_dir)
2404 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2408 mutex_lock(&mdsc->mutex);
2409 __register_request(mdsc, req, dir);
2410 __do_request(mdsc, req);
2418 mutex_unlock(&mdsc->mutex);
2419 dout("do_request waiting\n");
2420 if (!req->r_timeout && req->r_wait_for_completion) {
2421 err = req->r_wait_for_completion(mdsc, req);
2423 long timeleft = wait_for_completion_killable_timeout(
2425 ceph_timeout_jiffies(req->r_timeout));
2429 err = -EIO; /* timed out */
2431 err = timeleft; /* killed */
2433 dout("do_request waited, got %d\n", err);
2434 mutex_lock(&mdsc->mutex);
2436 /* only abort if we didn't race with a real reply */
2437 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2438 err = le32_to_cpu(req->r_reply_info.head->result);
2439 } else if (err < 0) {
2440 dout("aborted request %lld with %d\n", req->r_tid, err);
2443 * ensure we aren't running concurrently with
2444 * ceph_fill_trace or ceph_readdir_prepopulate, which
2445 * rely on locks (dir mutex) held by our caller.
2447 mutex_lock(&req->r_fill_mutex);
2449 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2450 mutex_unlock(&req->r_fill_mutex);
2452 if (req->r_parent &&
2453 (req->r_op & CEPH_MDS_OP_WRITE))
2454 ceph_invalidate_dir_request(req);
2460 mutex_unlock(&mdsc->mutex);
2461 dout("do_request %p done, result %d\n", req, err);
2466 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2467 * namespace request.
2469 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2471 struct inode *inode = req->r_parent;
2473 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2475 ceph_dir_clear_complete(inode);
2477 ceph_invalidate_dentry_lease(req->r_dentry);
2478 if (req->r_old_dentry)
2479 ceph_invalidate_dentry_lease(req->r_old_dentry);
2485 * We take the session mutex and parse and process the reply immediately.
2486 * This preserves the logical ordering of replies, capabilities, etc., sent
2487 * by the MDS as they are applied to our local cache.
2489 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2491 struct ceph_mds_client *mdsc = session->s_mdsc;
2492 struct ceph_mds_request *req;
2493 struct ceph_mds_reply_head *head = msg->front.iov_base;
2494 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2495 struct ceph_snap_realm *realm;
2498 int mds = session->s_mds;
2500 if (msg->front.iov_len < sizeof(*head)) {
2501 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2506 /* get request, session */
2507 tid = le64_to_cpu(msg->hdr.tid);
2508 mutex_lock(&mdsc->mutex);
2509 req = lookup_get_request(mdsc, tid);
2511 dout("handle_reply on unknown tid %llu\n", tid);
2512 mutex_unlock(&mdsc->mutex);
2515 dout("handle_reply %p\n", req);
2517 /* correct session? */
2518 if (req->r_session != session) {
2519 pr_err("mdsc_handle_reply got %llu on session mds%d"
2520 " not mds%d\n", tid, session->s_mds,
2521 req->r_session ? req->r_session->s_mds : -1);
2522 mutex_unlock(&mdsc->mutex);
2527 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2528 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2529 pr_warn("got a dup %s reply on %llu from mds%d\n",
2530 head->safe ? "safe" : "unsafe", tid, mds);
2531 mutex_unlock(&mdsc->mutex);
2534 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2535 pr_warn("got unsafe after safe on %llu from mds%d\n",
2537 mutex_unlock(&mdsc->mutex);
2541 result = le32_to_cpu(head->result);
2545 * if we're not talking to the authority, send to them
2546 * if the authority has changed while we weren't looking,
2547 * send to new authority
2548 * Otherwise we just have to return an ESTALE
2550 if (result == -ESTALE) {
2551 dout("got ESTALE on request %llu", req->r_tid);
2552 req->r_resend_mds = -1;
2553 if (req->r_direct_mode != USE_AUTH_MDS) {
2554 dout("not using auth, setting for that now");
2555 req->r_direct_mode = USE_AUTH_MDS;
2556 __do_request(mdsc, req);
2557 mutex_unlock(&mdsc->mutex);
2560 int mds = __choose_mds(mdsc, req);
2561 if (mds >= 0 && mds != req->r_session->s_mds) {
2562 dout("but auth changed, so resending");
2563 __do_request(mdsc, req);
2564 mutex_unlock(&mdsc->mutex);
2568 dout("have to return ESTALE on request %llu", req->r_tid);
2573 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2574 __unregister_request(mdsc, req);
2576 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2578 * We already handled the unsafe response, now do the
2579 * cleanup. No need to examine the response; the MDS
2580 * doesn't include any result info in the safe
2581 * response. And even if it did, there is nothing
2582 * useful we could do with a revised return value.
2584 dout("got safe reply %llu, mds%d\n", tid, mds);
2586 /* last unsafe request during umount? */
2587 if (mdsc->stopping && !__get_oldest_req(mdsc))
2588 complete_all(&mdsc->safe_umount_waiters);
2589 mutex_unlock(&mdsc->mutex);
2593 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2594 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2595 if (req->r_unsafe_dir) {
2596 struct ceph_inode_info *ci =
2597 ceph_inode(req->r_unsafe_dir);
2598 spin_lock(&ci->i_unsafe_lock);
2599 list_add_tail(&req->r_unsafe_dir_item,
2600 &ci->i_unsafe_dirops);
2601 spin_unlock(&ci->i_unsafe_lock);
2605 dout("handle_reply tid %lld result %d\n", tid, result);
2606 rinfo = &req->r_reply_info;
2607 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2608 mutex_unlock(&mdsc->mutex);
2610 mutex_lock(&session->s_mutex);
2612 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2619 if (rinfo->snapblob_len) {
2620 down_write(&mdsc->snap_rwsem);
2621 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2622 rinfo->snapblob + rinfo->snapblob_len,
2623 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2625 downgrade_write(&mdsc->snap_rwsem);
2627 down_read(&mdsc->snap_rwsem);
2630 /* insert trace into our cache */
2631 mutex_lock(&req->r_fill_mutex);
2632 current->journal_info = req;
2633 err = ceph_fill_trace(mdsc->fsc->sb, req);
2635 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2636 req->r_op == CEPH_MDS_OP_LSSNAP))
2637 ceph_readdir_prepopulate(req, req->r_session);
2638 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2640 current->journal_info = NULL;
2641 mutex_unlock(&req->r_fill_mutex);
2643 up_read(&mdsc->snap_rwsem);
2645 ceph_put_snap_realm(mdsc, realm);
2647 if (err == 0 && req->r_target_inode &&
2648 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2649 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2650 spin_lock(&ci->i_unsafe_lock);
2651 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2652 spin_unlock(&ci->i_unsafe_lock);
2655 mutex_lock(&mdsc->mutex);
2656 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2660 req->r_reply = ceph_msg_get(msg);
2661 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2664 dout("reply arrived after request %lld was aborted\n", tid);
2666 mutex_unlock(&mdsc->mutex);
2668 mutex_unlock(&session->s_mutex);
2670 /* kick calling process */
2671 complete_request(mdsc, req);
2673 ceph_mdsc_put_request(req);
2680 * handle mds notification that our request has been forwarded.
2682 static void handle_forward(struct ceph_mds_client *mdsc,
2683 struct ceph_mds_session *session,
2684 struct ceph_msg *msg)
2686 struct ceph_mds_request *req;
2687 u64 tid = le64_to_cpu(msg->hdr.tid);
2691 void *p = msg->front.iov_base;
2692 void *end = p + msg->front.iov_len;
2694 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2695 next_mds = ceph_decode_32(&p);
2696 fwd_seq = ceph_decode_32(&p);
2698 mutex_lock(&mdsc->mutex);
2699 req = lookup_get_request(mdsc, tid);
2701 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2702 goto out; /* dup reply? */
2705 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2706 dout("forward tid %llu aborted, unregistering\n", tid);
2707 __unregister_request(mdsc, req);
2708 } else if (fwd_seq <= req->r_num_fwd) {
2709 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2710 tid, next_mds, req->r_num_fwd, fwd_seq);
2712 /* resend. forward race not possible; mds would drop */
2713 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2715 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2716 req->r_attempts = 0;
2717 req->r_num_fwd = fwd_seq;
2718 req->r_resend_mds = next_mds;
2719 put_request_session(req);
2720 __do_request(mdsc, req);
2722 ceph_mdsc_put_request(req);
2724 mutex_unlock(&mdsc->mutex);
2728 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2732 * handle a mds session control message
2734 static void handle_session(struct ceph_mds_session *session,
2735 struct ceph_msg *msg)
2737 struct ceph_mds_client *mdsc = session->s_mdsc;
2740 int mds = session->s_mds;
2741 struct ceph_mds_session_head *h = msg->front.iov_base;
2745 if (msg->front.iov_len != sizeof(*h))
2747 op = le32_to_cpu(h->op);
2748 seq = le64_to_cpu(h->seq);
2750 mutex_lock(&mdsc->mutex);
2751 if (op == CEPH_SESSION_CLOSE) {
2752 get_session(session);
2753 __unregister_session(mdsc, session);
2755 /* FIXME: this ttl calculation is generous */
2756 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2757 mutex_unlock(&mdsc->mutex);
2759 mutex_lock(&session->s_mutex);
2761 dout("handle_session mds%d %s %p state %s seq %llu\n",
2762 mds, ceph_session_op_name(op), session,
2763 ceph_session_state_name(session->s_state), seq);
2765 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2766 session->s_state = CEPH_MDS_SESSION_OPEN;
2767 pr_info("mds%d came back\n", session->s_mds);
2771 case CEPH_SESSION_OPEN:
2772 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2773 pr_info("mds%d reconnect success\n", session->s_mds);
2774 session->s_state = CEPH_MDS_SESSION_OPEN;
2775 renewed_caps(mdsc, session, 0);
2778 __close_session(mdsc, session);
2781 case CEPH_SESSION_RENEWCAPS:
2782 if (session->s_renew_seq == seq)
2783 renewed_caps(mdsc, session, 1);
2786 case CEPH_SESSION_CLOSE:
2787 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2788 pr_info("mds%d reconnect denied\n", session->s_mds);
2789 cleanup_session_requests(mdsc, session);
2790 remove_session_caps(session);
2791 wake = 2; /* for good measure */
2792 wake_up_all(&mdsc->session_close_wq);
2795 case CEPH_SESSION_STALE:
2796 pr_info("mds%d caps went stale, renewing\n",
2798 spin_lock(&session->s_gen_ttl_lock);
2799 session->s_cap_gen++;
2800 session->s_cap_ttl = jiffies - 1;
2801 spin_unlock(&session->s_gen_ttl_lock);
2802 send_renew_caps(mdsc, session);
2805 case CEPH_SESSION_RECALL_STATE:
2806 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2809 case CEPH_SESSION_FLUSHMSG:
2810 send_flushmsg_ack(mdsc, session, seq);
2813 case CEPH_SESSION_FORCE_RO:
2814 dout("force_session_readonly %p\n", session);
2815 spin_lock(&session->s_cap_lock);
2816 session->s_readonly = true;
2817 spin_unlock(&session->s_cap_lock);
2818 wake_up_session_caps(session, 0);
2821 case CEPH_SESSION_REJECT:
2822 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2823 pr_info("mds%d rejected session\n", session->s_mds);
2824 session->s_state = CEPH_MDS_SESSION_REJECTED;
2825 cleanup_session_requests(mdsc, session);
2826 remove_session_caps(session);
2827 wake = 2; /* for good measure */
2831 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2835 mutex_unlock(&session->s_mutex);
2837 mutex_lock(&mdsc->mutex);
2838 __wake_requests(mdsc, &session->s_waiting);
2840 kick_requests(mdsc, mds);
2841 mutex_unlock(&mdsc->mutex);
2843 if (op == CEPH_SESSION_CLOSE)
2844 ceph_put_mds_session(session);
2848 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2849 (int)msg->front.iov_len);
2856 * called under session->mutex.
2858 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2859 struct ceph_mds_session *session)
2861 struct ceph_mds_request *req, *nreq;
2865 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2867 mutex_lock(&mdsc->mutex);
2868 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2869 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2871 ceph_msg_get(req->r_request);
2872 ceph_con_send(&session->s_con, req->r_request);
2877 * also re-send old requests when MDS enters reconnect stage. So that MDS
2878 * can process completed request in clientreplay stage.
2880 p = rb_first(&mdsc->request_tree);
2882 req = rb_entry(p, struct ceph_mds_request, r_node);
2884 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2886 if (req->r_attempts == 0)
2887 continue; /* only old requests */
2888 if (req->r_session &&
2889 req->r_session->s_mds == session->s_mds) {
2890 err = __prepare_send_request(mdsc, req,
2891 session->s_mds, true);
2893 ceph_msg_get(req->r_request);
2894 ceph_con_send(&session->s_con, req->r_request);
2898 mutex_unlock(&mdsc->mutex);
2902 * Encode information about a cap for a reconnect with the MDS.
2904 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2908 struct ceph_mds_cap_reconnect v2;
2909 struct ceph_mds_cap_reconnect_v1 v1;
2911 struct ceph_inode_info *ci;
2912 struct ceph_reconnect_state *recon_state = arg;
2913 struct ceph_pagelist *pagelist = recon_state->pagelist;
2918 struct dentry *dentry;
2922 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2923 inode, ceph_vinop(inode), cap, cap->cap_id,
2924 ceph_cap_string(cap->issued));
2925 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2929 dentry = d_find_alias(inode);
2931 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2933 err = PTR_ERR(path);
2942 spin_lock(&ci->i_ceph_lock);
2943 cap->seq = 0; /* reset cap seq */
2944 cap->issue_seq = 0; /* and issue_seq */
2945 cap->mseq = 0; /* and migrate_seq */
2946 cap->cap_gen = cap->session->s_cap_gen;
2948 if (recon_state->msg_version >= 2) {
2949 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2950 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2951 rec.v2.issued = cpu_to_le32(cap->issued);
2952 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2953 rec.v2.pathbase = cpu_to_le64(pathbase);
2954 rec.v2.flock_len = 0;
2956 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2957 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2958 rec.v1.issued = cpu_to_le32(cap->issued);
2959 rec.v1.size = cpu_to_le64(inode->i_size);
2960 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2961 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2962 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2963 rec.v1.pathbase = cpu_to_le64(pathbase);
2966 if (list_empty(&ci->i_cap_snaps)) {
2967 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2969 struct ceph_cap_snap *capsnap =
2970 list_first_entry(&ci->i_cap_snaps,
2971 struct ceph_cap_snap, ci_item);
2972 snap_follows = capsnap->follows;
2974 spin_unlock(&ci->i_ceph_lock);
2976 if (recon_state->msg_version >= 2) {
2977 int num_fcntl_locks, num_flock_locks;
2978 struct ceph_filelock *flocks;
2979 size_t struct_len, total_len = 0;
2983 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2984 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2985 sizeof(struct ceph_filelock), GFP_NOFS);
2990 err = ceph_encode_locks_to_buffer(inode, flocks,
3000 if (recon_state->msg_version >= 3) {
3001 /* version, compat_version and struct_len */
3002 total_len = 2 * sizeof(u8) + sizeof(u32);
3006 * number of encoded locks is stable, so copy to pagelist
3008 struct_len = 2 * sizeof(u32) +
3009 (num_fcntl_locks + num_flock_locks) *
3010 sizeof(struct ceph_filelock);
3011 rec.v2.flock_len = cpu_to_le32(struct_len);
3013 struct_len += sizeof(rec.v2);
3014 struct_len += sizeof(u32) + pathlen;
3017 struct_len += sizeof(u64); /* snap_follows */
3019 total_len += struct_len;
3020 err = ceph_pagelist_reserve(pagelist, total_len);
3023 if (recon_state->msg_version >= 3) {
3024 ceph_pagelist_encode_8(pagelist, struct_v);
3025 ceph_pagelist_encode_8(pagelist, 1);
3026 ceph_pagelist_encode_32(pagelist, struct_len);
3028 ceph_pagelist_encode_string(pagelist, path, pathlen);
3029 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3030 ceph_locks_to_pagelist(flocks, pagelist,
3034 ceph_pagelist_encode_64(pagelist, snap_follows);
3038 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
3039 err = ceph_pagelist_reserve(pagelist, size);
3041 ceph_pagelist_encode_string(pagelist, path, pathlen);
3042 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3046 recon_state->nr_caps++;
3056 * If an MDS fails and recovers, clients need to reconnect in order to
3057 * reestablish shared state. This includes all caps issued through
3058 * this session _and_ the snap_realm hierarchy. Because it's not
3059 * clear which snap realms the mds cares about, we send everything we
3060 * know about.. that ensures we'll then get any new info the
3061 * recovering MDS might have.
3063 * This is a relatively heavyweight operation, but it's rare.
3065 * called with mdsc->mutex held.
3067 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3068 struct ceph_mds_session *session)
3070 struct ceph_msg *reply;
3072 int mds = session->s_mds;
3075 struct ceph_pagelist *pagelist;
3076 struct ceph_reconnect_state recon_state;
3078 pr_info("mds%d reconnect start\n", mds);
3080 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
3082 goto fail_nopagelist;
3083 ceph_pagelist_init(pagelist);
3085 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3089 mutex_lock(&session->s_mutex);
3090 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3093 dout("session %p state %s\n", session,
3094 ceph_session_state_name(session->s_state));
3096 spin_lock(&session->s_gen_ttl_lock);
3097 session->s_cap_gen++;
3098 spin_unlock(&session->s_gen_ttl_lock);
3100 spin_lock(&session->s_cap_lock);
3101 /* don't know if session is readonly */
3102 session->s_readonly = 0;
3104 * notify __ceph_remove_cap() that we are composing cap reconnect.
3105 * If a cap get released before being added to the cap reconnect,
3106 * __ceph_remove_cap() should skip queuing cap release.
3108 session->s_cap_reconnect = 1;
3109 /* drop old cap expires; we're about to reestablish that state */
3110 cleanup_cap_releases(mdsc, session);
3112 /* trim unused caps to reduce MDS's cache rejoin time */
3113 if (mdsc->fsc->sb->s_root)
3114 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3116 ceph_con_close(&session->s_con);
3117 ceph_con_open(&session->s_con,
3118 CEPH_ENTITY_TYPE_MDS, mds,
3119 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3121 /* replay unsafe requests */
3122 replay_unsafe_requests(mdsc, session);
3124 down_read(&mdsc->snap_rwsem);
3126 /* traverse this session's caps */
3127 s_nr_caps = session->s_nr_caps;
3128 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3132 recon_state.nr_caps = 0;
3133 recon_state.pagelist = pagelist;
3134 if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3135 recon_state.msg_version = 3;
3136 else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3137 recon_state.msg_version = 2;
3139 recon_state.msg_version = 1;
3140 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3144 spin_lock(&session->s_cap_lock);
3145 session->s_cap_reconnect = 0;
3146 spin_unlock(&session->s_cap_lock);
3149 * snaprealms. we provide mds with the ino, seq (version), and
3150 * parent for all of our realms. If the mds has any newer info,
3153 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3154 struct ceph_snap_realm *realm =
3155 rb_entry(p, struct ceph_snap_realm, node);
3156 struct ceph_mds_snaprealm_reconnect sr_rec;
3158 dout(" adding snap realm %llx seq %lld parent %llx\n",
3159 realm->ino, realm->seq, realm->parent_ino);
3160 sr_rec.ino = cpu_to_le64(realm->ino);
3161 sr_rec.seq = cpu_to_le64(realm->seq);
3162 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3163 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3168 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3170 /* raced with cap release? */
3171 if (s_nr_caps != recon_state.nr_caps) {
3172 struct page *page = list_first_entry(&pagelist->head,
3174 __le32 *addr = kmap_atomic(page);
3175 *addr = cpu_to_le32(recon_state.nr_caps);
3176 kunmap_atomic(addr);
3179 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3180 ceph_msg_data_add_pagelist(reply, pagelist);
3182 ceph_early_kick_flushing_caps(mdsc, session);
3184 ceph_con_send(&session->s_con, reply);
3186 mutex_unlock(&session->s_mutex);
3188 mutex_lock(&mdsc->mutex);
3189 __wake_requests(mdsc, &session->s_waiting);
3190 mutex_unlock(&mdsc->mutex);
3192 up_read(&mdsc->snap_rwsem);
3196 ceph_msg_put(reply);
3197 up_read(&mdsc->snap_rwsem);
3198 mutex_unlock(&session->s_mutex);
3200 ceph_pagelist_release(pagelist);
3202 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3208 * compare old and new mdsmaps, kicking requests
3209 * and closing out old connections as necessary
3211 * called under mdsc->mutex.
3213 static void check_new_map(struct ceph_mds_client *mdsc,
3214 struct ceph_mdsmap *newmap,
3215 struct ceph_mdsmap *oldmap)
3218 int oldstate, newstate;
3219 struct ceph_mds_session *s;
3221 dout("check_new_map new %u old %u\n",
3222 newmap->m_epoch, oldmap->m_epoch);
3224 for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3225 if (!mdsc->sessions[i])
3227 s = mdsc->sessions[i];
3228 oldstate = ceph_mdsmap_get_state(oldmap, i);
3229 newstate = ceph_mdsmap_get_state(newmap, i);
3231 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3232 i, ceph_mds_state_name(oldstate),
3233 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3234 ceph_mds_state_name(newstate),
3235 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3236 ceph_session_state_name(s->s_state));
3238 if (i >= newmap->m_num_mds ||
3239 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3240 ceph_mdsmap_get_addr(newmap, i),
3241 sizeof(struct ceph_entity_addr))) {
3242 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3243 /* the session never opened, just close it
3246 __unregister_session(mdsc, s);
3247 __wake_requests(mdsc, &s->s_waiting);
3248 ceph_put_mds_session(s);
3249 } else if (i >= newmap->m_num_mds) {
3250 /* force close session for stopped mds */
3252 __unregister_session(mdsc, s);
3253 __wake_requests(mdsc, &s->s_waiting);
3254 kick_requests(mdsc, i);
3255 mutex_unlock(&mdsc->mutex);
3257 mutex_lock(&s->s_mutex);
3258 cleanup_session_requests(mdsc, s);
3259 remove_session_caps(s);
3260 mutex_unlock(&s->s_mutex);
3262 ceph_put_mds_session(s);
3264 mutex_lock(&mdsc->mutex);
3267 mutex_unlock(&mdsc->mutex);
3268 mutex_lock(&s->s_mutex);
3269 mutex_lock(&mdsc->mutex);
3270 ceph_con_close(&s->s_con);
3271 mutex_unlock(&s->s_mutex);
3272 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3274 } else if (oldstate == newstate) {
3275 continue; /* nothing new with this mds */
3281 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3282 newstate >= CEPH_MDS_STATE_RECONNECT) {
3283 mutex_unlock(&mdsc->mutex);
3284 send_mds_reconnect(mdsc, s);
3285 mutex_lock(&mdsc->mutex);
3289 * kick request on any mds that has gone active.
3291 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3292 newstate >= CEPH_MDS_STATE_ACTIVE) {
3293 if (oldstate != CEPH_MDS_STATE_CREATING &&
3294 oldstate != CEPH_MDS_STATE_STARTING)
3295 pr_info("mds%d recovery completed\n", s->s_mds);
3296 kick_requests(mdsc, i);
3297 ceph_kick_flushing_caps(mdsc, s);
3298 wake_up_session_caps(s, 1);
3302 for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3303 s = mdsc->sessions[i];
3306 if (!ceph_mdsmap_is_laggy(newmap, i))
3308 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3309 s->s_state == CEPH_MDS_SESSION_HUNG ||
3310 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3311 dout(" connecting to export targets of laggy mds%d\n",
3313 __open_export_target_sessions(mdsc, s);
3325 * caller must hold session s_mutex, dentry->d_lock
3327 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3329 struct ceph_dentry_info *di = ceph_dentry(dentry);
3331 ceph_put_mds_session(di->lease_session);
3332 di->lease_session = NULL;
3335 static void handle_lease(struct ceph_mds_client *mdsc,
3336 struct ceph_mds_session *session,
3337 struct ceph_msg *msg)
3339 struct super_block *sb = mdsc->fsc->sb;
3340 struct inode *inode;
3341 struct dentry *parent, *dentry;
3342 struct ceph_dentry_info *di;
3343 int mds = session->s_mds;
3344 struct ceph_mds_lease *h = msg->front.iov_base;
3346 struct ceph_vino vino;
3350 dout("handle_lease from mds%d\n", mds);
3353 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3355 vino.ino = le64_to_cpu(h->ino);
3356 vino.snap = CEPH_NOSNAP;
3357 seq = le32_to_cpu(h->seq);
3358 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3359 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3360 if (dname.len != get_unaligned_le32(h+1))
3364 inode = ceph_find_inode(sb, vino);
3365 dout("handle_lease %s, ino %llx %p %.*s\n",
3366 ceph_lease_op_name(h->action), vino.ino, inode,
3367 dname.len, dname.name);
3369 mutex_lock(&session->s_mutex);
3373 dout("handle_lease no inode %llx\n", vino.ino);
3378 parent = d_find_alias(inode);
3380 dout("no parent dentry on inode %p\n", inode);
3382 goto release; /* hrm... */
3384 dname.hash = full_name_hash(parent, dname.name, dname.len);
3385 dentry = d_lookup(parent, &dname);
3390 spin_lock(&dentry->d_lock);
3391 di = ceph_dentry(dentry);
3392 switch (h->action) {
3393 case CEPH_MDS_LEASE_REVOKE:
3394 if (di->lease_session == session) {
3395 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3396 h->seq = cpu_to_le32(di->lease_seq);
3397 __ceph_mdsc_drop_dentry_lease(dentry);
3402 case CEPH_MDS_LEASE_RENEW:
3403 if (di->lease_session == session &&
3404 di->lease_gen == session->s_cap_gen &&
3405 di->lease_renew_from &&
3406 di->lease_renew_after == 0) {
3407 unsigned long duration =
3408 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3410 di->lease_seq = seq;
3411 di->time = di->lease_renew_from + duration;
3412 di->lease_renew_after = di->lease_renew_from +
3414 di->lease_renew_from = 0;
3418 spin_unlock(&dentry->d_lock);
3425 /* let's just reuse the same message */
3426 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3428 ceph_con_send(&session->s_con, msg);
3432 mutex_unlock(&session->s_mutex);
3436 pr_err("corrupt lease message\n");
3440 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3441 struct inode *inode,
3442 struct dentry *dentry, char action,
3445 struct ceph_msg *msg;
3446 struct ceph_mds_lease *lease;
3447 int len = sizeof(*lease) + sizeof(u32);
3450 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3451 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3452 dnamelen = dentry->d_name.len;
3455 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3458 lease = msg->front.iov_base;
3459 lease->action = action;
3460 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3461 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3462 lease->seq = cpu_to_le32(seq);
3463 put_unaligned_le32(dnamelen, lease + 1);
3464 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3467 * if this is a preemptive lease RELEASE, no need to
3468 * flush request stream, since the actual request will
3471 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3473 ceph_con_send(&session->s_con, msg);
3477 * drop all leases (and dentry refs) in preparation for umount
3479 static void drop_leases(struct ceph_mds_client *mdsc)
3483 dout("drop_leases\n");
3484 mutex_lock(&mdsc->mutex);
3485 for (i = 0; i < mdsc->max_sessions; i++) {
3486 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3489 mutex_unlock(&mdsc->mutex);
3490 mutex_lock(&s->s_mutex);
3491 mutex_unlock(&s->s_mutex);
3492 ceph_put_mds_session(s);
3493 mutex_lock(&mdsc->mutex);
3495 mutex_unlock(&mdsc->mutex);
3501 * delayed work -- periodically trim expired leases, renew caps with mds
3503 static void schedule_delayed(struct ceph_mds_client *mdsc)
3506 unsigned hz = round_jiffies_relative(HZ * delay);
3507 schedule_delayed_work(&mdsc->delayed_work, hz);
3510 static void delayed_work(struct work_struct *work)
3513 struct ceph_mds_client *mdsc =
3514 container_of(work, struct ceph_mds_client, delayed_work.work);
3518 dout("mdsc delayed_work\n");
3519 ceph_check_delayed_caps(mdsc);
3524 mutex_lock(&mdsc->mutex);
3525 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3526 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3527 mdsc->last_renew_caps);
3529 mdsc->last_renew_caps = jiffies;
3531 for (i = 0; i < mdsc->max_sessions; i++) {
3532 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3535 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3536 dout("resending session close request for mds%d\n",
3538 request_close_session(mdsc, s);
3539 ceph_put_mds_session(s);
3542 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3543 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3544 s->s_state = CEPH_MDS_SESSION_HUNG;
3545 pr_info("mds%d hung\n", s->s_mds);
3548 if (s->s_state == CEPH_MDS_SESSION_NEW ||
3549 s->s_state == CEPH_MDS_SESSION_RESTARTING ||
3550 s->s_state == CEPH_MDS_SESSION_REJECTED) {
3551 /* this mds is failed or recovering, just wait */
3552 ceph_put_mds_session(s);
3555 mutex_unlock(&mdsc->mutex);
3557 mutex_lock(&s->s_mutex);
3559 send_renew_caps(mdsc, s);
3561 ceph_con_keepalive(&s->s_con);
3562 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3563 s->s_state == CEPH_MDS_SESSION_HUNG)
3564 ceph_send_cap_releases(mdsc, s);
3565 mutex_unlock(&s->s_mutex);
3566 ceph_put_mds_session(s);
3568 mutex_lock(&mdsc->mutex);
3570 mutex_unlock(&mdsc->mutex);
3572 schedule_delayed(mdsc);
3575 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3578 struct ceph_mds_client *mdsc;
3580 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3585 mutex_init(&mdsc->mutex);
3586 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3587 if (!mdsc->mdsmap) {
3592 init_completion(&mdsc->safe_umount_waiters);
3593 init_waitqueue_head(&mdsc->session_close_wq);
3594 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3595 mdsc->sessions = NULL;
3596 atomic_set(&mdsc->num_sessions, 0);
3597 mdsc->max_sessions = 0;
3599 mdsc->last_snap_seq = 0;
3600 init_rwsem(&mdsc->snap_rwsem);
3601 mdsc->snap_realms = RB_ROOT;
3602 INIT_LIST_HEAD(&mdsc->snap_empty);
3603 spin_lock_init(&mdsc->snap_empty_lock);
3605 mdsc->oldest_tid = 0;
3606 mdsc->request_tree = RB_ROOT;
3607 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3608 mdsc->last_renew_caps = jiffies;
3609 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3610 spin_lock_init(&mdsc->cap_delay_lock);
3611 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3612 spin_lock_init(&mdsc->snap_flush_lock);
3613 mdsc->last_cap_flush_tid = 1;
3614 INIT_LIST_HEAD(&mdsc->cap_flush_list);
3615 INIT_LIST_HEAD(&mdsc->cap_dirty);
3616 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3617 mdsc->num_cap_flushing = 0;
3618 spin_lock_init(&mdsc->cap_dirty_lock);
3619 init_waitqueue_head(&mdsc->cap_flushing_wq);
3620 spin_lock_init(&mdsc->dentry_lru_lock);
3621 INIT_LIST_HEAD(&mdsc->dentry_lru);
3623 ceph_caps_init(mdsc);
3624 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3626 init_rwsem(&mdsc->pool_perm_rwsem);
3627 mdsc->pool_perm_tree = RB_ROOT;
3629 strncpy(mdsc->nodename, utsname()->nodename,
3630 sizeof(mdsc->nodename) - 1);
3635 * Wait for safe replies on open mds requests. If we time out, drop
3636 * all requests from the tree to avoid dangling dentry refs.
3638 static void wait_requests(struct ceph_mds_client *mdsc)
3640 struct ceph_options *opts = mdsc->fsc->client->options;
3641 struct ceph_mds_request *req;
3643 mutex_lock(&mdsc->mutex);
3644 if (__get_oldest_req(mdsc)) {
3645 mutex_unlock(&mdsc->mutex);
3647 dout("wait_requests waiting for requests\n");
3648 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3649 ceph_timeout_jiffies(opts->mount_timeout));
3651 /* tear down remaining requests */
3652 mutex_lock(&mdsc->mutex);
3653 while ((req = __get_oldest_req(mdsc))) {
3654 dout("wait_requests timed out on tid %llu\n",
3656 __unregister_request(mdsc, req);
3659 mutex_unlock(&mdsc->mutex);
3660 dout("wait_requests done\n");
3664 * called before mount is ro, and before dentries are torn down.
3665 * (hmm, does this still race with new lookups?)
3667 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3669 dout("pre_umount\n");
3673 ceph_flush_dirty_caps(mdsc);
3674 wait_requests(mdsc);
3677 * wait for reply handlers to drop their request refs and
3678 * their inode/dcache refs
3684 * wait for all write mds requests to flush.
3686 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3688 struct ceph_mds_request *req = NULL, *nextreq;
3691 mutex_lock(&mdsc->mutex);
3692 dout("wait_unsafe_requests want %lld\n", want_tid);
3694 req = __get_oldest_req(mdsc);
3695 while (req && req->r_tid <= want_tid) {
3696 /* find next request */
3697 n = rb_next(&req->r_node);
3699 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3702 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3703 (req->r_op & CEPH_MDS_OP_WRITE)) {
3705 ceph_mdsc_get_request(req);
3707 ceph_mdsc_get_request(nextreq);
3708 mutex_unlock(&mdsc->mutex);
3709 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3710 req->r_tid, want_tid);
3711 wait_for_completion(&req->r_safe_completion);
3712 mutex_lock(&mdsc->mutex);
3713 ceph_mdsc_put_request(req);
3715 break; /* next dne before, so we're done! */
3716 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3717 /* next request was removed from tree */
3718 ceph_mdsc_put_request(nextreq);
3721 ceph_mdsc_put_request(nextreq); /* won't go away */
3725 mutex_unlock(&mdsc->mutex);
3726 dout("wait_unsafe_requests done\n");
3729 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3731 u64 want_tid, want_flush;
3733 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3737 mutex_lock(&mdsc->mutex);
3738 want_tid = mdsc->last_tid;
3739 mutex_unlock(&mdsc->mutex);
3741 ceph_flush_dirty_caps(mdsc);
3742 spin_lock(&mdsc->cap_dirty_lock);
3743 want_flush = mdsc->last_cap_flush_tid;
3744 if (!list_empty(&mdsc->cap_flush_list)) {
3745 struct ceph_cap_flush *cf =
3746 list_last_entry(&mdsc->cap_flush_list,
3747 struct ceph_cap_flush, g_list);
3750 spin_unlock(&mdsc->cap_dirty_lock);
3752 dout("sync want tid %lld flush_seq %lld\n",
3753 want_tid, want_flush);
3755 wait_unsafe_requests(mdsc, want_tid);
3756 wait_caps_flush(mdsc, want_flush);
3760 * true if all sessions are closed, or we force unmount
3762 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3764 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3766 return atomic_read(&mdsc->num_sessions) <= skipped;
3770 * called after sb is ro.
3772 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3774 struct ceph_options *opts = mdsc->fsc->client->options;
3775 struct ceph_mds_session *session;
3779 dout("close_sessions\n");
3781 /* close sessions */
3782 mutex_lock(&mdsc->mutex);
3783 for (i = 0; i < mdsc->max_sessions; i++) {
3784 session = __ceph_lookup_mds_session(mdsc, i);
3787 mutex_unlock(&mdsc->mutex);
3788 mutex_lock(&session->s_mutex);
3789 if (__close_session(mdsc, session) <= 0)
3791 mutex_unlock(&session->s_mutex);
3792 ceph_put_mds_session(session);
3793 mutex_lock(&mdsc->mutex);
3795 mutex_unlock(&mdsc->mutex);
3797 dout("waiting for sessions to close\n");
3798 wait_event_timeout(mdsc->session_close_wq,
3799 done_closing_sessions(mdsc, skipped),
3800 ceph_timeout_jiffies(opts->mount_timeout));
3802 /* tear down remaining sessions */
3803 mutex_lock(&mdsc->mutex);
3804 for (i = 0; i < mdsc->max_sessions; i++) {
3805 if (mdsc->sessions[i]) {
3806 session = get_session(mdsc->sessions[i]);
3807 __unregister_session(mdsc, session);
3808 mutex_unlock(&mdsc->mutex);
3809 mutex_lock(&session->s_mutex);
3810 remove_session_caps(session);
3811 mutex_unlock(&session->s_mutex);
3812 ceph_put_mds_session(session);
3813 mutex_lock(&mdsc->mutex);
3816 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3817 mutex_unlock(&mdsc->mutex);
3819 ceph_cleanup_empty_realms(mdsc);
3821 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3826 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3828 struct ceph_mds_session *session;
3831 dout("force umount\n");
3833 mutex_lock(&mdsc->mutex);
3834 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3835 session = __ceph_lookup_mds_session(mdsc, mds);
3838 mutex_unlock(&mdsc->mutex);
3839 mutex_lock(&session->s_mutex);
3840 __close_session(mdsc, session);
3841 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3842 cleanup_session_requests(mdsc, session);
3843 remove_session_caps(session);
3845 mutex_unlock(&session->s_mutex);
3846 ceph_put_mds_session(session);
3847 mutex_lock(&mdsc->mutex);
3848 kick_requests(mdsc, mds);
3850 __wake_requests(mdsc, &mdsc->waiting_for_map);
3851 mutex_unlock(&mdsc->mutex);
3854 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3858 * Make sure the delayed work stopped before releasing
3861 * Because the cancel_delayed_work_sync() will only
3862 * guarantee that the work finishes executing. But the
3863 * delayed work will re-arm itself again after that.
3865 flush_delayed_work(&mdsc->delayed_work);
3868 ceph_mdsmap_destroy(mdsc->mdsmap);
3869 kfree(mdsc->sessions);
3870 ceph_caps_finalize(mdsc);
3871 ceph_pool_perm_destroy(mdsc);
3874 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3876 struct ceph_mds_client *mdsc = fsc->mdsc;
3877 dout("mdsc_destroy %p\n", mdsc);
3879 /* flush out any connection work with references to us */
3882 ceph_mdsc_stop(mdsc);
3886 dout("mdsc_destroy %p done\n", mdsc);
3889 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3891 struct ceph_fs_client *fsc = mdsc->fsc;
3892 const char *mds_namespace = fsc->mount_options->mds_namespace;
3893 void *p = msg->front.iov_base;
3894 void *end = p + msg->front.iov_len;
3898 u32 mount_fscid = (u32)-1;
3899 u8 struct_v, struct_cv;
3902 ceph_decode_need(&p, end, sizeof(u32), bad);
3903 epoch = ceph_decode_32(&p);
3905 dout("handle_fsmap epoch %u\n", epoch);
3907 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3908 struct_v = ceph_decode_8(&p);
3909 struct_cv = ceph_decode_8(&p);
3910 map_len = ceph_decode_32(&p);
3912 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3913 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3915 num_fs = ceph_decode_32(&p);
3916 while (num_fs-- > 0) {
3917 void *info_p, *info_end;
3922 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3923 info_v = ceph_decode_8(&p);
3924 info_cv = ceph_decode_8(&p);
3925 info_len = ceph_decode_32(&p);
3926 ceph_decode_need(&p, end, info_len, bad);
3928 info_end = p + info_len;
3931 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3932 fscid = ceph_decode_32(&info_p);
3933 namelen = ceph_decode_32(&info_p);
3934 ceph_decode_need(&info_p, info_end, namelen, bad);
3936 if (mds_namespace &&
3937 strlen(mds_namespace) == namelen &&
3938 !strncmp(mds_namespace, (char *)info_p, namelen)) {
3939 mount_fscid = fscid;
3944 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3945 if (mount_fscid != (u32)-1) {
3946 fsc->client->monc.fs_cluster_id = mount_fscid;
3947 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3949 ceph_monc_renew_subs(&fsc->client->monc);
3956 pr_err("error decoding fsmap\n");
3958 mutex_lock(&mdsc->mutex);
3959 mdsc->mdsmap_err = -ENOENT;
3960 __wake_requests(mdsc, &mdsc->waiting_for_map);
3961 mutex_unlock(&mdsc->mutex);
3966 * handle mds map update.
3968 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3972 void *p = msg->front.iov_base;
3973 void *end = p + msg->front.iov_len;
3974 struct ceph_mdsmap *newmap, *oldmap;
3975 struct ceph_fsid fsid;
3978 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3979 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3980 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3982 epoch = ceph_decode_32(&p);
3983 maplen = ceph_decode_32(&p);
3984 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3986 /* do we need it? */
3987 mutex_lock(&mdsc->mutex);
3988 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3989 dout("handle_map epoch %u <= our %u\n",
3990 epoch, mdsc->mdsmap->m_epoch);
3991 mutex_unlock(&mdsc->mutex);
3995 newmap = ceph_mdsmap_decode(&p, end);
3996 if (IS_ERR(newmap)) {
3997 err = PTR_ERR(newmap);
4001 /* swap into place */
4003 oldmap = mdsc->mdsmap;
4004 mdsc->mdsmap = newmap;
4005 check_new_map(mdsc, newmap, oldmap);
4006 ceph_mdsmap_destroy(oldmap);
4008 mdsc->mdsmap = newmap; /* first mds map */
4010 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
4012 __wake_requests(mdsc, &mdsc->waiting_for_map);
4013 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4014 mdsc->mdsmap->m_epoch);
4016 mutex_unlock(&mdsc->mutex);
4017 schedule_delayed(mdsc);
4021 mutex_unlock(&mdsc->mutex);
4023 pr_err("error decoding mdsmap %d\n", err);
4027 static struct ceph_connection *con_get(struct ceph_connection *con)
4029 struct ceph_mds_session *s = con->private;
4031 if (get_session(s)) {
4032 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4035 dout("mdsc con_get %p FAIL\n", s);
4039 static void con_put(struct ceph_connection *con)
4041 struct ceph_mds_session *s = con->private;
4043 dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4044 ceph_put_mds_session(s);
4048 * if the client is unresponsive for long enough, the mds will kill
4049 * the session entirely.
4051 static void peer_reset(struct ceph_connection *con)
4053 struct ceph_mds_session *s = con->private;
4054 struct ceph_mds_client *mdsc = s->s_mdsc;
4056 pr_warn("mds%d closed our session\n", s->s_mds);
4057 send_mds_reconnect(mdsc, s);
4060 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4062 struct ceph_mds_session *s = con->private;
4063 struct ceph_mds_client *mdsc = s->s_mdsc;
4064 int type = le16_to_cpu(msg->hdr.type);
4066 mutex_lock(&mdsc->mutex);
4067 if (__verify_registered_session(mdsc, s) < 0) {
4068 mutex_unlock(&mdsc->mutex);
4071 mutex_unlock(&mdsc->mutex);
4074 case CEPH_MSG_MDS_MAP:
4075 ceph_mdsc_handle_mdsmap(mdsc, msg);
4077 case CEPH_MSG_FS_MAP_USER:
4078 ceph_mdsc_handle_fsmap(mdsc, msg);
4080 case CEPH_MSG_CLIENT_SESSION:
4081 handle_session(s, msg);
4083 case CEPH_MSG_CLIENT_REPLY:
4084 handle_reply(s, msg);
4086 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4087 handle_forward(mdsc, s, msg);
4089 case CEPH_MSG_CLIENT_CAPS:
4090 ceph_handle_caps(s, msg);
4092 case CEPH_MSG_CLIENT_SNAP:
4093 ceph_handle_snap(mdsc, s, msg);
4095 case CEPH_MSG_CLIENT_LEASE:
4096 handle_lease(mdsc, s, msg);
4100 pr_err("received unknown message type %d %s\n", type,
4101 ceph_msg_type_name(type));
4112 * Note: returned pointer is the address of a structure that's
4113 * managed separately. Caller must *not* attempt to free it.
4115 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4116 int *proto, int force_new)
4118 struct ceph_mds_session *s = con->private;
4119 struct ceph_mds_client *mdsc = s->s_mdsc;
4120 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4121 struct ceph_auth_handshake *auth = &s->s_auth;
4123 if (force_new && auth->authorizer) {
4124 ceph_auth_destroy_authorizer(auth->authorizer);
4125 auth->authorizer = NULL;
4127 if (!auth->authorizer) {
4128 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4131 return ERR_PTR(ret);
4133 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4136 return ERR_PTR(ret);
4138 *proto = ac->protocol;
4143 static int add_authorizer_challenge(struct ceph_connection *con,
4144 void *challenge_buf, int challenge_buf_len)
4146 struct ceph_mds_session *s = con->private;
4147 struct ceph_mds_client *mdsc = s->s_mdsc;
4148 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4150 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4151 challenge_buf, challenge_buf_len);
4154 static int verify_authorizer_reply(struct ceph_connection *con)
4156 struct ceph_mds_session *s = con->private;
4157 struct ceph_mds_client *mdsc = s->s_mdsc;
4158 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4160 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4163 static int invalidate_authorizer(struct ceph_connection *con)
4165 struct ceph_mds_session *s = con->private;
4166 struct ceph_mds_client *mdsc = s->s_mdsc;
4167 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4169 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4171 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4174 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4175 struct ceph_msg_header *hdr, int *skip)
4177 struct ceph_msg *msg;
4178 int type = (int) le16_to_cpu(hdr->type);
4179 int front_len = (int) le32_to_cpu(hdr->front_len);
4185 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4187 pr_err("unable to allocate msg type %d len %d\n",
4195 static int mds_sign_message(struct ceph_msg *msg)
4197 struct ceph_mds_session *s = msg->con->private;
4198 struct ceph_auth_handshake *auth = &s->s_auth;
4200 return ceph_auth_sign_message(auth, msg);
4203 static int mds_check_message_signature(struct ceph_msg *msg)
4205 struct ceph_mds_session *s = msg->con->private;
4206 struct ceph_auth_handshake *auth = &s->s_auth;
4208 return ceph_auth_check_message_signature(auth, msg);
4211 static const struct ceph_connection_operations mds_con_ops = {
4214 .dispatch = dispatch,
4215 .get_authorizer = get_authorizer,
4216 .add_authorizer_challenge = add_authorizer_challenge,
4217 .verify_authorizer_reply = verify_authorizer_reply,
4218 .invalidate_authorizer = invalidate_authorizer,
4219 .peer_reset = peer_reset,
4220 .alloc_msg = mds_alloc_msg,
4221 .sign_message = mds_sign_message,
4222 .check_message_signature = mds_check_message_signature,