1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.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45 * are no longer valid.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);
196 BUG_ON(!info->dir_entries);
197 if ((unsigned long)(info->dir_entries + num) >
198 (unsigned long)info->dir_entries + info->dir_buf_size) {
199 pr_err("dir contents are larger than expected\n");
206 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
208 ceph_decode_need(p, end, sizeof(u32)*2, bad);
209 rde->name_len = ceph_decode_32(p);
210 ceph_decode_need(p, end, rde->name_len, bad);
213 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
215 *p += sizeof(struct ceph_mds_reply_lease);
218 err = parse_reply_info_in(p, end, &rde->inode, features);
221 /* ceph_readdir_prepopulate() will update it */
235 pr_err("problem parsing dir contents %d\n", err);
240 * parse fcntl F_GETLK results
242 static int parse_reply_info_filelock(void **p, void *end,
243 struct ceph_mds_reply_info_parsed *info,
246 if (*p + sizeof(*info->filelock_reply) > end)
249 info->filelock_reply = *p;
250 *p += sizeof(*info->filelock_reply);
252 if (unlikely(*p != end))
261 * parse create results
263 static int parse_reply_info_create(void **p, void *end,
264 struct ceph_mds_reply_info_parsed *info,
267 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
269 info->has_create_ino = false;
271 info->has_create_ino = true;
272 info->ino = ceph_decode_64(p);
276 if (unlikely(*p != end))
285 * parse extra results
287 static int parse_reply_info_extra(void **p, void *end,
288 struct ceph_mds_reply_info_parsed *info,
291 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
292 return parse_reply_info_filelock(p, end, info, features);
293 else if (info->head->op == CEPH_MDS_OP_READDIR ||
294 info->head->op == CEPH_MDS_OP_LSSNAP)
295 return parse_reply_info_dir(p, end, info, features);
296 else if (info->head->op == CEPH_MDS_OP_CREATE)
297 return parse_reply_info_create(p, end, info, features);
303 * parse entire mds reply
305 static int parse_reply_info(struct ceph_msg *msg,
306 struct ceph_mds_reply_info_parsed *info,
313 info->head = msg->front.iov_base;
314 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
315 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
318 ceph_decode_32_safe(&p, end, len, bad);
320 ceph_decode_need(&p, end, len, bad);
321 err = parse_reply_info_trace(&p, p+len, info, features);
327 ceph_decode_32_safe(&p, end, len, bad);
329 ceph_decode_need(&p, end, len, bad);
330 err = parse_reply_info_extra(&p, p+len, info, features);
336 ceph_decode_32_safe(&p, end, len, bad);
337 info->snapblob_len = len;
348 pr_err("mds parse_reply err %d\n", err);
352 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
354 if (!info->dir_entries)
356 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
363 const char *ceph_session_state_name(int s)
366 case CEPH_MDS_SESSION_NEW: return "new";
367 case CEPH_MDS_SESSION_OPENING: return "opening";
368 case CEPH_MDS_SESSION_OPEN: return "open";
369 case CEPH_MDS_SESSION_HUNG: return "hung";
370 case CEPH_MDS_SESSION_CLOSING: return "closing";
371 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
372 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
373 case CEPH_MDS_SESSION_REJECTED: return "rejected";
374 default: return "???";
378 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
380 if (atomic_inc_not_zero(&s->s_ref)) {
381 dout("mdsc get_session %p %d -> %d\n", s,
382 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
385 dout("mdsc get_session %p 0 -- FAIL", s);
390 void ceph_put_mds_session(struct ceph_mds_session *s)
392 dout("mdsc put_session %p %d -> %d\n", s,
393 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
394 if (atomic_dec_and_test(&s->s_ref)) {
395 if (s->s_auth.authorizer)
396 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
402 * called under mdsc->mutex
404 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
407 struct ceph_mds_session *session;
409 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
411 session = mdsc->sessions[mds];
412 dout("lookup_mds_session %p %d\n", session,
413 atomic_read(&session->s_ref));
414 get_session(session);
418 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
420 if (mds >= mdsc->max_sessions)
422 return mdsc->sessions[mds];
425 static int __verify_registered_session(struct ceph_mds_client *mdsc,
426 struct ceph_mds_session *s)
428 if (s->s_mds >= mdsc->max_sessions ||
429 mdsc->sessions[s->s_mds] != s)
435 * create+register a new session for given mds.
436 * called under mdsc->mutex.
438 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
441 struct ceph_mds_session *s;
443 if (mds >= mdsc->mdsmap->m_max_mds)
444 return ERR_PTR(-EINVAL);
446 s = kzalloc(sizeof(*s), GFP_NOFS);
448 return ERR_PTR(-ENOMEM);
451 s->s_state = CEPH_MDS_SESSION_NEW;
454 mutex_init(&s->s_mutex);
456 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
458 spin_lock_init(&s->s_gen_ttl_lock);
460 s->s_cap_ttl = jiffies - 1;
462 spin_lock_init(&s->s_cap_lock);
463 s->s_renew_requested = 0;
465 INIT_LIST_HEAD(&s->s_caps);
468 atomic_set(&s->s_ref, 1);
469 INIT_LIST_HEAD(&s->s_waiting);
470 INIT_LIST_HEAD(&s->s_unsafe);
471 s->s_num_cap_releases = 0;
472 s->s_cap_reconnect = 0;
473 s->s_cap_iterator = NULL;
474 INIT_LIST_HEAD(&s->s_cap_releases);
475 INIT_LIST_HEAD(&s->s_cap_flushing);
477 dout("register_session mds%d\n", mds);
478 if (mds >= mdsc->max_sessions) {
479 int newmax = 1 << get_count_order(mds+1);
480 struct ceph_mds_session **sa;
482 dout("register_session realloc to %d\n", newmax);
483 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
486 if (mdsc->sessions) {
487 memcpy(sa, mdsc->sessions,
488 mdsc->max_sessions * sizeof(void *));
489 kfree(mdsc->sessions);
492 mdsc->max_sessions = newmax;
494 mdsc->sessions[mds] = s;
495 atomic_inc(&mdsc->num_sessions);
496 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
498 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
499 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
505 return ERR_PTR(-ENOMEM);
509 * called under mdsc->mutex
511 static void __unregister_session(struct ceph_mds_client *mdsc,
512 struct ceph_mds_session *s)
514 dout("__unregister_session mds%d %p\n", s->s_mds, s);
515 BUG_ON(mdsc->sessions[s->s_mds] != s);
516 mdsc->sessions[s->s_mds] = NULL;
517 ceph_con_close(&s->s_con);
518 ceph_put_mds_session(s);
519 atomic_dec(&mdsc->num_sessions);
523 * drop session refs in request.
525 * should be last request ref, or hold mdsc->mutex
527 static void put_request_session(struct ceph_mds_request *req)
529 if (req->r_session) {
530 ceph_put_mds_session(req->r_session);
531 req->r_session = NULL;
535 void ceph_mdsc_release_request(struct kref *kref)
537 struct ceph_mds_request *req = container_of(kref,
538 struct ceph_mds_request,
540 destroy_reply_info(&req->r_reply_info);
542 ceph_msg_put(req->r_request);
544 ceph_msg_put(req->r_reply);
546 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
549 if (req->r_locked_dir)
550 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
551 iput(req->r_target_inode);
554 if (req->r_old_dentry)
555 dput(req->r_old_dentry);
556 if (req->r_old_dentry_dir) {
558 * track (and drop pins for) r_old_dentry_dir
559 * separately, since r_old_dentry's d_parent may have
560 * changed between the dir mutex being dropped and
561 * this request being freed.
563 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
565 iput(req->r_old_dentry_dir);
570 ceph_pagelist_release(req->r_pagelist);
571 put_request_session(req);
572 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
576 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
579 * lookup session, bump ref if found.
581 * called under mdsc->mutex.
583 static struct ceph_mds_request *
584 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
586 struct ceph_mds_request *req;
588 req = lookup_request(&mdsc->request_tree, tid);
590 ceph_mdsc_get_request(req);
596 * Register an in-flight request, and assign a tid. Link to directory
597 * are modifying (if any).
599 * Called under mdsc->mutex.
601 static void __register_request(struct ceph_mds_client *mdsc,
602 struct ceph_mds_request *req,
605 req->r_tid = ++mdsc->last_tid;
607 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
609 dout("__register_request %p tid %lld\n", req, req->r_tid);
610 ceph_mdsc_get_request(req);
611 insert_request(&mdsc->request_tree, req);
613 req->r_uid = current_fsuid();
614 req->r_gid = current_fsgid();
616 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
617 mdsc->oldest_tid = req->r_tid;
621 req->r_unsafe_dir = dir;
625 static void __unregister_request(struct ceph_mds_client *mdsc,
626 struct ceph_mds_request *req)
628 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
630 if (req->r_tid == mdsc->oldest_tid) {
631 struct rb_node *p = rb_next(&req->r_node);
632 mdsc->oldest_tid = 0;
634 struct ceph_mds_request *next_req =
635 rb_entry(p, struct ceph_mds_request, r_node);
636 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
637 mdsc->oldest_tid = next_req->r_tid;
644 erase_request(&mdsc->request_tree, req);
646 if (req->r_unsafe_dir && req->r_got_unsafe) {
647 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
648 spin_lock(&ci->i_unsafe_lock);
649 list_del_init(&req->r_unsafe_dir_item);
650 spin_unlock(&ci->i_unsafe_lock);
652 if (req->r_target_inode && req->r_got_unsafe) {
653 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
654 spin_lock(&ci->i_unsafe_lock);
655 list_del_init(&req->r_unsafe_target_item);
656 spin_unlock(&ci->i_unsafe_lock);
659 if (req->r_unsafe_dir) {
660 iput(req->r_unsafe_dir);
661 req->r_unsafe_dir = NULL;
664 complete_all(&req->r_safe_completion);
666 ceph_mdsc_put_request(req);
670 * Choose mds to send request to next. If there is a hint set in the
671 * request (e.g., due to a prior forward hint from the mds), use that.
672 * Otherwise, consult frag tree and/or caps to identify the
673 * appropriate mds. If all else fails, choose randomly.
675 * Called under mdsc->mutex.
677 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
680 * we don't need to worry about protecting the d_parent access
681 * here because we never renaming inside the snapped namespace
682 * except to resplice to another snapdir, and either the old or new
683 * result is a valid result.
685 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
686 dentry = dentry->d_parent;
690 static int __choose_mds(struct ceph_mds_client *mdsc,
691 struct ceph_mds_request *req)
694 struct ceph_inode_info *ci;
695 struct ceph_cap *cap;
696 int mode = req->r_direct_mode;
698 u32 hash = req->r_direct_hash;
699 bool is_hash = req->r_direct_is_hash;
702 * is there a specific mds we should try? ignore hint if we have
703 * no session and the mds is not up (active or recovering).
705 if (req->r_resend_mds >= 0 &&
706 (__have_session(mdsc, req->r_resend_mds) ||
707 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
708 dout("choose_mds using resend_mds mds%d\n",
710 return req->r_resend_mds;
713 if (mode == USE_RANDOM_MDS)
718 inode = req->r_inode;
719 } else if (req->r_dentry) {
720 /* ignore race with rename; old or new d_parent is okay */
721 struct dentry *parent = req->r_dentry->d_parent;
722 struct inode *dir = d_inode(parent);
724 if (dir->i_sb != mdsc->fsc->sb) {
726 inode = d_inode(req->r_dentry);
727 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
728 /* direct snapped/virtual snapdir requests
729 * based on parent dir inode */
730 struct dentry *dn = get_nonsnap_parent(parent);
732 dout("__choose_mds using nonsnap parent %p\n", inode);
735 inode = d_inode(req->r_dentry);
736 if (!inode || mode == USE_AUTH_MDS) {
739 hash = ceph_dentry_hash(dir, req->r_dentry);
745 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
749 ci = ceph_inode(inode);
751 if (is_hash && S_ISDIR(inode->i_mode)) {
752 struct ceph_inode_frag frag;
755 ceph_choose_frag(ci, hash, &frag, &found);
757 if (mode == USE_ANY_MDS && frag.ndist > 0) {
760 /* choose a random replica */
761 get_random_bytes(&r, 1);
764 dout("choose_mds %p %llx.%llx "
765 "frag %u mds%d (%d/%d)\n",
766 inode, ceph_vinop(inode),
769 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
770 CEPH_MDS_STATE_ACTIVE)
774 /* since this file/dir wasn't known to be
775 * replicated, then we want to look for the
776 * authoritative mds. */
779 /* choose auth mds */
781 dout("choose_mds %p %llx.%llx "
782 "frag %u mds%d (auth)\n",
783 inode, ceph_vinop(inode), frag.frag, mds);
784 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
785 CEPH_MDS_STATE_ACTIVE)
791 spin_lock(&ci->i_ceph_lock);
793 if (mode == USE_AUTH_MDS)
794 cap = ci->i_auth_cap;
795 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
796 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
798 spin_unlock(&ci->i_ceph_lock);
801 mds = cap->session->s_mds;
802 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
803 inode, ceph_vinop(inode), mds,
804 cap == ci->i_auth_cap ? "auth " : "", cap);
805 spin_unlock(&ci->i_ceph_lock);
809 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
810 dout("choose_mds chose random mds%d\n", mds);
818 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
820 struct ceph_msg *msg;
821 struct ceph_mds_session_head *h;
823 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
826 pr_err("create_session_msg ENOMEM creating msg\n");
829 h = msg->front.iov_base;
830 h->op = cpu_to_le32(op);
831 h->seq = cpu_to_le64(seq);
837 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
838 * to include additional client metadata fields.
840 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
842 struct ceph_msg *msg;
843 struct ceph_mds_session_head *h;
845 int metadata_bytes = 0;
846 int metadata_key_count = 0;
847 struct ceph_options *opt = mdsc->fsc->client->options;
848 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
851 const char* metadata[][2] = {
852 {"hostname", utsname()->nodename},
853 {"kernel_version", utsname()->release},
854 {"entity_id", opt->name ? : ""},
855 {"root", fsopt->server_path ? : "/"},
859 /* Calculate serialized length of metadata */
860 metadata_bytes = 4; /* map length */
861 for (i = 0; metadata[i][0] != NULL; ++i) {
862 metadata_bytes += 8 + strlen(metadata[i][0]) +
863 strlen(metadata[i][1]);
864 metadata_key_count++;
867 /* Allocate the message */
868 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
871 pr_err("create_session_msg ENOMEM creating msg\n");
874 h = msg->front.iov_base;
875 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
876 h->seq = cpu_to_le64(seq);
879 * Serialize client metadata into waiting buffer space, using
880 * the format that userspace expects for map<string, string>
882 * ClientSession messages with metadata are v2
884 msg->hdr.version = cpu_to_le16(2);
885 msg->hdr.compat_version = cpu_to_le16(1);
887 /* The write pointer, following the session_head structure */
888 p = msg->front.iov_base + sizeof(*h);
890 /* Number of entries in the map */
891 ceph_encode_32(&p, metadata_key_count);
893 /* Two length-prefixed strings for each entry in the map */
894 for (i = 0; metadata[i][0] != NULL; ++i) {
895 size_t const key_len = strlen(metadata[i][0]);
896 size_t const val_len = strlen(metadata[i][1]);
898 ceph_encode_32(&p, key_len);
899 memcpy(p, metadata[i][0], key_len);
901 ceph_encode_32(&p, val_len);
902 memcpy(p, metadata[i][1], val_len);
910 * send session open request.
912 * called under mdsc->mutex
914 static int __open_session(struct ceph_mds_client *mdsc,
915 struct ceph_mds_session *session)
917 struct ceph_msg *msg;
919 int mds = session->s_mds;
921 /* wait for mds to go active? */
922 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
923 dout("open_session to mds%d (%s)\n", mds,
924 ceph_mds_state_name(mstate));
925 session->s_state = CEPH_MDS_SESSION_OPENING;
926 session->s_renew_requested = jiffies;
928 /* send connect message */
929 msg = create_session_open_msg(mdsc, session->s_seq);
932 ceph_con_send(&session->s_con, msg);
937 * open sessions for any export targets for the given mds
939 * called under mdsc->mutex
941 static struct ceph_mds_session *
942 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
944 struct ceph_mds_session *session;
946 session = __ceph_lookup_mds_session(mdsc, target);
948 session = register_session(mdsc, target);
952 if (session->s_state == CEPH_MDS_SESSION_NEW ||
953 session->s_state == CEPH_MDS_SESSION_CLOSING)
954 __open_session(mdsc, session);
959 struct ceph_mds_session *
960 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
962 struct ceph_mds_session *session;
964 dout("open_export_target_session to mds%d\n", target);
966 mutex_lock(&mdsc->mutex);
967 session = __open_export_target_session(mdsc, target);
968 mutex_unlock(&mdsc->mutex);
973 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
974 struct ceph_mds_session *session)
976 struct ceph_mds_info *mi;
977 struct ceph_mds_session *ts;
978 int i, mds = session->s_mds;
980 if (mds >= mdsc->mdsmap->m_max_mds)
983 mi = &mdsc->mdsmap->m_info[mds];
984 dout("open_export_target_sessions for mds%d (%d targets)\n",
985 session->s_mds, mi->num_export_targets);
987 for (i = 0; i < mi->num_export_targets; i++) {
988 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
990 ceph_put_mds_session(ts);
994 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
995 struct ceph_mds_session *session)
997 mutex_lock(&mdsc->mutex);
998 __open_export_target_sessions(mdsc, session);
999 mutex_unlock(&mdsc->mutex);
1006 /* caller holds s_cap_lock, we drop it */
1007 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1008 struct ceph_mds_session *session)
1009 __releases(session->s_cap_lock)
1011 LIST_HEAD(tmp_list);
1012 list_splice_init(&session->s_cap_releases, &tmp_list);
1013 session->s_num_cap_releases = 0;
1014 spin_unlock(&session->s_cap_lock);
1016 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1017 while (!list_empty(&tmp_list)) {
1018 struct ceph_cap *cap;
1019 /* zero out the in-progress message */
1020 cap = list_first_entry(&tmp_list,
1021 struct ceph_cap, session_caps);
1022 list_del(&cap->session_caps);
1023 ceph_put_cap(mdsc, cap);
1027 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1028 struct ceph_mds_session *session)
1030 struct ceph_mds_request *req;
1033 dout("cleanup_session_requests mds%d\n", session->s_mds);
1034 mutex_lock(&mdsc->mutex);
1035 while (!list_empty(&session->s_unsafe)) {
1036 req = list_first_entry(&session->s_unsafe,
1037 struct ceph_mds_request, r_unsafe_item);
1038 list_del_init(&req->r_unsafe_item);
1039 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1041 __unregister_request(mdsc, req);
1043 /* zero r_attempts, so kick_requests() will re-send requests */
1044 p = rb_first(&mdsc->request_tree);
1046 req = rb_entry(p, struct ceph_mds_request, r_node);
1048 if (req->r_session &&
1049 req->r_session->s_mds == session->s_mds)
1050 req->r_attempts = 0;
1052 mutex_unlock(&mdsc->mutex);
1056 * Helper to safely iterate over all caps associated with a session, with
1057 * special care taken to handle a racing __ceph_remove_cap().
1059 * Caller must hold session s_mutex.
1061 static int iterate_session_caps(struct ceph_mds_session *session,
1062 int (*cb)(struct inode *, struct ceph_cap *,
1065 struct list_head *p;
1066 struct ceph_cap *cap;
1067 struct inode *inode, *last_inode = NULL;
1068 struct ceph_cap *old_cap = NULL;
1071 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1072 spin_lock(&session->s_cap_lock);
1073 p = session->s_caps.next;
1074 while (p != &session->s_caps) {
1075 cap = list_entry(p, struct ceph_cap, session_caps);
1076 inode = igrab(&cap->ci->vfs_inode);
1081 session->s_cap_iterator = cap;
1082 spin_unlock(&session->s_cap_lock);
1089 ceph_put_cap(session->s_mdsc, old_cap);
1093 ret = cb(inode, cap, arg);
1096 spin_lock(&session->s_cap_lock);
1098 if (cap->ci == NULL) {
1099 dout("iterate_session_caps finishing cap %p removal\n",
1101 BUG_ON(cap->session != session);
1102 cap->session = NULL;
1103 list_del_init(&cap->session_caps);
1104 session->s_nr_caps--;
1105 if (cap->queue_release) {
1106 list_add_tail(&cap->session_caps,
1107 &session->s_cap_releases);
1108 session->s_num_cap_releases++;
1110 old_cap = cap; /* put_cap it w/o locks held */
1118 session->s_cap_iterator = NULL;
1119 spin_unlock(&session->s_cap_lock);
1123 ceph_put_cap(session->s_mdsc, old_cap);
1128 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1131 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1132 struct ceph_inode_info *ci = ceph_inode(inode);
1133 LIST_HEAD(to_remove);
1135 bool invalidate = false;
1137 dout("removing cap %p, ci is %p, inode is %p\n",
1138 cap, ci, &ci->vfs_inode);
1139 spin_lock(&ci->i_ceph_lock);
1140 __ceph_remove_cap(cap, false);
1141 if (!ci->i_auth_cap) {
1142 struct ceph_cap_flush *cf;
1143 struct ceph_mds_client *mdsc = fsc->mdsc;
1145 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1147 if (ci->i_wrbuffer_ref > 0 &&
1148 ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1151 while (!list_empty(&ci->i_cap_flush_list)) {
1152 cf = list_first_entry(&ci->i_cap_flush_list,
1153 struct ceph_cap_flush, i_list);
1154 list_move(&cf->i_list, &to_remove);
1157 spin_lock(&mdsc->cap_dirty_lock);
1159 list_for_each_entry(cf, &to_remove, i_list)
1160 list_del(&cf->g_list);
1162 if (!list_empty(&ci->i_dirty_item)) {
1163 pr_warn_ratelimited(
1164 " dropping dirty %s state for %p %lld\n",
1165 ceph_cap_string(ci->i_dirty_caps),
1166 inode, ceph_ino(inode));
1167 ci->i_dirty_caps = 0;
1168 list_del_init(&ci->i_dirty_item);
1171 if (!list_empty(&ci->i_flushing_item)) {
1172 pr_warn_ratelimited(
1173 " dropping dirty+flushing %s state for %p %lld\n",
1174 ceph_cap_string(ci->i_flushing_caps),
1175 inode, ceph_ino(inode));
1176 ci->i_flushing_caps = 0;
1177 list_del_init(&ci->i_flushing_item);
1178 mdsc->num_cap_flushing--;
1181 spin_unlock(&mdsc->cap_dirty_lock);
1183 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1184 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1185 ci->i_prealloc_cap_flush = NULL;
1188 spin_unlock(&ci->i_ceph_lock);
1189 while (!list_empty(&to_remove)) {
1190 struct ceph_cap_flush *cf;
1191 cf = list_first_entry(&to_remove,
1192 struct ceph_cap_flush, i_list);
1193 list_del(&cf->i_list);
1194 ceph_free_cap_flush(cf);
1197 wake_up_all(&ci->i_cap_wq);
1199 ceph_queue_invalidate(inode);
1206 * caller must hold session s_mutex
1208 static void remove_session_caps(struct ceph_mds_session *session)
1210 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1211 struct super_block *sb = fsc->sb;
1212 dout("remove_session_caps on %p\n", session);
1213 iterate_session_caps(session, remove_session_caps_cb, fsc);
1215 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1217 spin_lock(&session->s_cap_lock);
1218 if (session->s_nr_caps > 0) {
1219 struct inode *inode;
1220 struct ceph_cap *cap, *prev = NULL;
1221 struct ceph_vino vino;
1223 * iterate_session_caps() skips inodes that are being
1224 * deleted, we need to wait until deletions are complete.
1225 * __wait_on_freeing_inode() is designed for the job,
1226 * but it is not exported, so use lookup inode function
1229 while (!list_empty(&session->s_caps)) {
1230 cap = list_entry(session->s_caps.next,
1231 struct ceph_cap, session_caps);
1235 vino = cap->ci->i_vino;
1236 spin_unlock(&session->s_cap_lock);
1238 inode = ceph_find_inode(sb, vino);
1241 spin_lock(&session->s_cap_lock);
1245 // drop cap expires and unlock s_cap_lock
1246 cleanup_cap_releases(session->s_mdsc, session);
1248 BUG_ON(session->s_nr_caps > 0);
1249 BUG_ON(!list_empty(&session->s_cap_flushing));
1253 * wake up any threads waiting on this session's caps. if the cap is
1254 * old (didn't get renewed on the client reconnect), remove it now.
1256 * caller must hold s_mutex.
1258 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1261 struct ceph_inode_info *ci = ceph_inode(inode);
1264 spin_lock(&ci->i_ceph_lock);
1265 ci->i_wanted_max_size = 0;
1266 ci->i_requested_max_size = 0;
1267 spin_unlock(&ci->i_ceph_lock);
1269 wake_up_all(&ci->i_cap_wq);
1273 static void wake_up_session_caps(struct ceph_mds_session *session,
1276 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1277 iterate_session_caps(session, wake_up_session_cb,
1278 (void *)(unsigned long)reconnect);
1282 * Send periodic message to MDS renewing all currently held caps. The
1283 * ack will reset the expiration for all caps from this session.
1285 * caller holds s_mutex
1287 static int send_renew_caps(struct ceph_mds_client *mdsc,
1288 struct ceph_mds_session *session)
1290 struct ceph_msg *msg;
1293 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1294 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1295 pr_info("mds%d caps stale\n", session->s_mds);
1296 session->s_renew_requested = jiffies;
1298 /* do not try to renew caps until a recovering mds has reconnected
1299 * with its clients. */
1300 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1301 if (state < CEPH_MDS_STATE_RECONNECT) {
1302 dout("send_renew_caps ignoring mds%d (%s)\n",
1303 session->s_mds, ceph_mds_state_name(state));
1307 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1308 ceph_mds_state_name(state));
1309 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1310 ++session->s_renew_seq);
1313 ceph_con_send(&session->s_con, msg);
1317 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1318 struct ceph_mds_session *session, u64 seq)
1320 struct ceph_msg *msg;
1322 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1323 session->s_mds, ceph_session_state_name(session->s_state), seq);
1324 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1327 ceph_con_send(&session->s_con, msg);
1333 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1335 * Called under session->s_mutex
1337 static void renewed_caps(struct ceph_mds_client *mdsc,
1338 struct ceph_mds_session *session, int is_renew)
1343 spin_lock(&session->s_cap_lock);
1344 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1346 session->s_cap_ttl = session->s_renew_requested +
1347 mdsc->mdsmap->m_session_timeout*HZ;
1350 if (time_before(jiffies, session->s_cap_ttl)) {
1351 pr_info("mds%d caps renewed\n", session->s_mds);
1354 pr_info("mds%d caps still stale\n", session->s_mds);
1357 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1358 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1359 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1360 spin_unlock(&session->s_cap_lock);
1363 wake_up_session_caps(session, 0);
1367 * send a session close request
1369 static int request_close_session(struct ceph_mds_client *mdsc,
1370 struct ceph_mds_session *session)
1372 struct ceph_msg *msg;
1374 dout("request_close_session mds%d state %s seq %lld\n",
1375 session->s_mds, ceph_session_state_name(session->s_state),
1377 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1380 ceph_con_send(&session->s_con, msg);
1385 * Called with s_mutex held.
1387 static int __close_session(struct ceph_mds_client *mdsc,
1388 struct ceph_mds_session *session)
1390 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1392 session->s_state = CEPH_MDS_SESSION_CLOSING;
1393 return request_close_session(mdsc, session);
1397 * Trim old(er) caps.
1399 * Because we can't cache an inode without one or more caps, we do
1400 * this indirectly: if a cap is unused, we prune its aliases, at which
1401 * point the inode will hopefully get dropped to.
1403 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1404 * memory pressure from the MDS, though, so it needn't be perfect.
1406 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1408 struct ceph_mds_session *session = arg;
1409 struct ceph_inode_info *ci = ceph_inode(inode);
1410 int used, wanted, oissued, mine;
1412 if (session->s_trim_caps <= 0)
1415 spin_lock(&ci->i_ceph_lock);
1416 mine = cap->issued | cap->implemented;
1417 used = __ceph_caps_used(ci);
1418 wanted = __ceph_caps_file_wanted(ci);
1419 oissued = __ceph_caps_issued_other(ci, cap);
1421 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1422 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1423 ceph_cap_string(used), ceph_cap_string(wanted));
1424 if (cap == ci->i_auth_cap) {
1425 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1426 !list_empty(&ci->i_cap_snaps))
1428 if ((used | wanted) & CEPH_CAP_ANY_WR)
1431 /* The inode has cached pages, but it's no longer used.
1432 * we can safely drop it */
1433 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1434 !(oissued & CEPH_CAP_FILE_CACHE)) {
1438 if ((used | wanted) & ~oissued & mine)
1439 goto out; /* we need these caps */
1441 session->s_trim_caps--;
1443 /* we aren't the only cap.. just remove us */
1444 __ceph_remove_cap(cap, true);
1446 /* try dropping referring dentries */
1447 spin_unlock(&ci->i_ceph_lock);
1448 d_prune_aliases(inode);
1449 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1450 inode, cap, atomic_read(&inode->i_count));
1455 spin_unlock(&ci->i_ceph_lock);
1460 * Trim session cap count down to some max number.
1462 static int trim_caps(struct ceph_mds_client *mdsc,
1463 struct ceph_mds_session *session,
1466 int trim_caps = session->s_nr_caps - max_caps;
1468 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1469 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1470 if (trim_caps > 0) {
1471 session->s_trim_caps = trim_caps;
1472 iterate_session_caps(session, trim_caps_cb, session);
1473 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1474 session->s_mds, session->s_nr_caps, max_caps,
1475 trim_caps - session->s_trim_caps);
1476 session->s_trim_caps = 0;
1479 ceph_send_cap_releases(mdsc, session);
1483 static int check_caps_flush(struct ceph_mds_client *mdsc,
1488 spin_lock(&mdsc->cap_dirty_lock);
1489 if (!list_empty(&mdsc->cap_flush_list)) {
1490 struct ceph_cap_flush *cf =
1491 list_first_entry(&mdsc->cap_flush_list,
1492 struct ceph_cap_flush, g_list);
1493 if (cf->tid <= want_flush_tid) {
1494 dout("check_caps_flush still flushing tid "
1495 "%llu <= %llu\n", cf->tid, want_flush_tid);
1499 spin_unlock(&mdsc->cap_dirty_lock);
1504 * flush all dirty inode data to disk.
1506 * returns true if we've flushed through want_flush_tid
1508 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1511 dout("check_caps_flush want %llu\n", want_flush_tid);
1513 wait_event(mdsc->cap_flushing_wq,
1514 check_caps_flush(mdsc, want_flush_tid));
1516 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1520 * called under s_mutex
1522 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1523 struct ceph_mds_session *session)
1525 struct ceph_msg *msg = NULL;
1526 struct ceph_mds_cap_release *head;
1527 struct ceph_mds_cap_item *item;
1528 struct ceph_cap *cap;
1529 LIST_HEAD(tmp_list);
1530 int num_cap_releases;
1532 spin_lock(&session->s_cap_lock);
1534 list_splice_init(&session->s_cap_releases, &tmp_list);
1535 num_cap_releases = session->s_num_cap_releases;
1536 session->s_num_cap_releases = 0;
1537 spin_unlock(&session->s_cap_lock);
1539 while (!list_empty(&tmp_list)) {
1541 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1542 PAGE_SIZE, GFP_NOFS, false);
1545 head = msg->front.iov_base;
1546 head->num = cpu_to_le32(0);
1547 msg->front.iov_len = sizeof(*head);
1549 cap = list_first_entry(&tmp_list, struct ceph_cap,
1551 list_del(&cap->session_caps);
1554 head = msg->front.iov_base;
1555 le32_add_cpu(&head->num, 1);
1556 item = msg->front.iov_base + msg->front.iov_len;
1557 item->ino = cpu_to_le64(cap->cap_ino);
1558 item->cap_id = cpu_to_le64(cap->cap_id);
1559 item->migrate_seq = cpu_to_le32(cap->mseq);
1560 item->seq = cpu_to_le32(cap->issue_seq);
1561 msg->front.iov_len += sizeof(*item);
1563 ceph_put_cap(mdsc, cap);
1565 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1566 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1567 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1568 ceph_con_send(&session->s_con, msg);
1573 BUG_ON(num_cap_releases != 0);
1575 spin_lock(&session->s_cap_lock);
1576 if (!list_empty(&session->s_cap_releases))
1578 spin_unlock(&session->s_cap_lock);
1581 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1582 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1583 ceph_con_send(&session->s_con, msg);
1587 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1589 spin_lock(&session->s_cap_lock);
1590 list_splice(&tmp_list, &session->s_cap_releases);
1591 session->s_num_cap_releases += num_cap_releases;
1592 spin_unlock(&session->s_cap_lock);
1599 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1602 struct ceph_inode_info *ci = ceph_inode(dir);
1603 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1604 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1605 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1606 int order, num_entries;
1608 spin_lock(&ci->i_ceph_lock);
1609 num_entries = ci->i_files + ci->i_subdirs;
1610 spin_unlock(&ci->i_ceph_lock);
1611 num_entries = max(num_entries, 1);
1612 num_entries = min(num_entries, opt->max_readdir);
1614 order = get_order(size * num_entries);
1615 while (order >= 0) {
1616 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1619 if (rinfo->dir_entries)
1623 if (!rinfo->dir_entries)
1626 num_entries = (PAGE_SIZE << order) / size;
1627 num_entries = min(num_entries, opt->max_readdir);
1629 rinfo->dir_buf_size = PAGE_SIZE << order;
1630 req->r_num_caps = num_entries + 1;
1631 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1632 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1637 * Create an mds request.
1639 struct ceph_mds_request *
1640 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1642 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1645 return ERR_PTR(-ENOMEM);
1647 mutex_init(&req->r_fill_mutex);
1649 req->r_started = jiffies;
1650 req->r_resend_mds = -1;
1651 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1652 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1654 kref_init(&req->r_kref);
1655 RB_CLEAR_NODE(&req->r_node);
1656 INIT_LIST_HEAD(&req->r_wait);
1657 init_completion(&req->r_completion);
1658 init_completion(&req->r_safe_completion);
1659 INIT_LIST_HEAD(&req->r_unsafe_item);
1661 req->r_stamp = current_fs_time(mdsc->fsc->sb);
1664 req->r_direct_mode = mode;
1669 * return oldest (lowest) request, tid in request tree, 0 if none.
1671 * called under mdsc->mutex.
1673 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1675 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1677 return rb_entry(rb_first(&mdsc->request_tree),
1678 struct ceph_mds_request, r_node);
1681 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1683 return mdsc->oldest_tid;
1687 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1688 * on build_path_from_dentry in fs/cifs/dir.c.
1690 * If @stop_on_nosnap, generate path relative to the first non-snapped
1693 * Encode hidden .snap dirs as a double /, i.e.
1694 * foo/.snap/bar -> foo//bar
1696 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1699 struct dentry *temp;
1705 return ERR_PTR(-EINVAL);
1709 seq = read_seqbegin(&rename_lock);
1711 for (temp = dentry; !IS_ROOT(temp);) {
1712 struct inode *inode = d_inode(temp);
1713 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1714 len++; /* slash only */
1715 else if (stop_on_nosnap && inode &&
1716 ceph_snap(inode) == CEPH_NOSNAP)
1719 len += 1 + temp->d_name.len;
1720 temp = temp->d_parent;
1724 len--; /* no leading '/' */
1726 path = kmalloc(len+1, GFP_NOFS);
1728 return ERR_PTR(-ENOMEM);
1730 path[pos] = 0; /* trailing null */
1732 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1733 struct inode *inode;
1735 spin_lock(&temp->d_lock);
1736 inode = d_inode(temp);
1737 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1738 dout("build_path path+%d: %p SNAPDIR\n",
1740 } else if (stop_on_nosnap && inode &&
1741 ceph_snap(inode) == CEPH_NOSNAP) {
1742 spin_unlock(&temp->d_lock);
1745 pos -= temp->d_name.len;
1747 spin_unlock(&temp->d_lock);
1750 strncpy(path + pos, temp->d_name.name,
1753 spin_unlock(&temp->d_lock);
1756 temp = temp->d_parent;
1759 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1760 pr_err("build_path did not end path lookup where "
1761 "expected, namelen is %d, pos is %d\n", len, pos);
1762 /* presumably this is only possible if racing with a
1763 rename of one of the parent directories (we can not
1764 lock the dentries above us to prevent this, but
1765 retrying should be harmless) */
1770 *base = ceph_ino(d_inode(temp));
1772 dout("build_path on %p %d built %llx '%.*s'\n",
1773 dentry, d_count(dentry), *base, len, path);
1777 static int build_dentry_path(struct dentry *dentry,
1778 const char **ppath, int *ppathlen, u64 *pino,
1783 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1784 *pino = ceph_ino(d_inode(dentry->d_parent));
1785 *ppath = dentry->d_name.name;
1786 *ppathlen = dentry->d_name.len;
1789 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1791 return PTR_ERR(path);
1797 static int build_inode_path(struct inode *inode,
1798 const char **ppath, int *ppathlen, u64 *pino,
1801 struct dentry *dentry;
1804 if (ceph_snap(inode) == CEPH_NOSNAP) {
1805 *pino = ceph_ino(inode);
1809 dentry = d_find_alias(inode);
1810 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1813 return PTR_ERR(path);
1820 * request arguments may be specified via an inode *, a dentry *, or
1821 * an explicit ino+path.
1823 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1824 const char *rpath, u64 rino,
1825 const char **ppath, int *pathlen,
1826 u64 *ino, int *freepath)
1831 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1832 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1834 } else if (rdentry) {
1835 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1836 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1838 } else if (rpath || rino) {
1841 *pathlen = rpath ? strlen(rpath) : 0;
1842 dout(" path %.*s\n", *pathlen, rpath);
1849 * called under mdsc->mutex
1851 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1852 struct ceph_mds_request *req,
1853 int mds, bool drop_cap_releases)
1855 struct ceph_msg *msg;
1856 struct ceph_mds_request_head *head;
1857 const char *path1 = NULL;
1858 const char *path2 = NULL;
1859 u64 ino1 = 0, ino2 = 0;
1860 int pathlen1 = 0, pathlen2 = 0;
1861 int freepath1 = 0, freepath2 = 0;
1867 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1868 req->r_path1, req->r_ino1.ino,
1869 &path1, &pathlen1, &ino1, &freepath1);
1875 ret = set_request_path_attr(NULL, req->r_old_dentry,
1876 req->r_path2, req->r_ino2.ino,
1877 &path2, &pathlen2, &ino2, &freepath2);
1883 len = sizeof(*head) +
1884 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1885 sizeof(struct ceph_timespec);
1887 /* calculate (max) length for cap releases */
1888 len += sizeof(struct ceph_mds_request_release) *
1889 (!!req->r_inode_drop + !!req->r_dentry_drop +
1890 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1891 if (req->r_dentry_drop)
1892 len += req->r_dentry->d_name.len;
1893 if (req->r_old_dentry_drop)
1894 len += req->r_old_dentry->d_name.len;
1896 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1898 msg = ERR_PTR(-ENOMEM);
1902 msg->hdr.version = cpu_to_le16(2);
1903 msg->hdr.tid = cpu_to_le64(req->r_tid);
1905 head = msg->front.iov_base;
1906 p = msg->front.iov_base + sizeof(*head);
1907 end = msg->front.iov_base + msg->front.iov_len;
1909 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1910 head->op = cpu_to_le32(req->r_op);
1911 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1912 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1913 head->args = req->r_args;
1915 ceph_encode_filepath(&p, end, ino1, path1);
1916 ceph_encode_filepath(&p, end, ino2, path2);
1918 /* make note of release offset, in case we need to replay */
1919 req->r_request_release_offset = p - msg->front.iov_base;
1923 if (req->r_inode_drop)
1924 releases += ceph_encode_inode_release(&p,
1925 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1926 mds, req->r_inode_drop, req->r_inode_unless, 0);
1927 if (req->r_dentry_drop)
1928 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1929 mds, req->r_dentry_drop, req->r_dentry_unless);
1930 if (req->r_old_dentry_drop)
1931 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1932 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1933 if (req->r_old_inode_drop)
1934 releases += ceph_encode_inode_release(&p,
1935 d_inode(req->r_old_dentry),
1936 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1938 if (drop_cap_releases) {
1940 p = msg->front.iov_base + req->r_request_release_offset;
1943 head->num_releases = cpu_to_le16(releases);
1947 struct ceph_timespec ts;
1948 ceph_encode_timespec(&ts, &req->r_stamp);
1949 ceph_encode_copy(&p, &ts, sizeof(ts));
1953 msg->front.iov_len = p - msg->front.iov_base;
1954 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1956 if (req->r_pagelist) {
1957 struct ceph_pagelist *pagelist = req->r_pagelist;
1958 atomic_inc(&pagelist->refcnt);
1959 ceph_msg_data_add_pagelist(msg, pagelist);
1960 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1962 msg->hdr.data_len = 0;
1965 msg->hdr.data_off = cpu_to_le16(0);
1969 kfree((char *)path2);
1972 kfree((char *)path1);
1978 * called under mdsc->mutex if error, under no mutex if
1981 static void complete_request(struct ceph_mds_client *mdsc,
1982 struct ceph_mds_request *req)
1984 if (req->r_callback)
1985 req->r_callback(mdsc, req);
1987 complete_all(&req->r_completion);
1991 * called under mdsc->mutex
1993 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1994 struct ceph_mds_request *req,
1995 int mds, bool drop_cap_releases)
1997 struct ceph_mds_request_head *rhead;
1998 struct ceph_msg *msg;
2003 struct ceph_cap *cap =
2004 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2007 req->r_sent_on_mseq = cap->mseq;
2009 req->r_sent_on_mseq = -1;
2011 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2012 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2014 if (req->r_got_unsafe) {
2017 * Replay. Do not regenerate message (and rebuild
2018 * paths, etc.); just use the original message.
2019 * Rebuilding paths will break for renames because
2020 * d_move mangles the src name.
2022 msg = req->r_request;
2023 rhead = msg->front.iov_base;
2025 flags = le32_to_cpu(rhead->flags);
2026 flags |= CEPH_MDS_FLAG_REPLAY;
2027 rhead->flags = cpu_to_le32(flags);
2029 if (req->r_target_inode)
2030 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2032 rhead->num_retry = req->r_attempts - 1;
2034 /* remove cap/dentry releases from message */
2035 rhead->num_releases = 0;
2038 p = msg->front.iov_base + req->r_request_release_offset;
2040 struct ceph_timespec ts;
2041 ceph_encode_timespec(&ts, &req->r_stamp);
2042 ceph_encode_copy(&p, &ts, sizeof(ts));
2045 msg->front.iov_len = p - msg->front.iov_base;
2046 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2050 if (req->r_request) {
2051 ceph_msg_put(req->r_request);
2052 req->r_request = NULL;
2054 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2056 req->r_err = PTR_ERR(msg);
2057 return PTR_ERR(msg);
2059 req->r_request = msg;
2061 rhead = msg->front.iov_base;
2062 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2063 if (req->r_got_unsafe)
2064 flags |= CEPH_MDS_FLAG_REPLAY;
2065 if (req->r_locked_dir)
2066 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2067 rhead->flags = cpu_to_le32(flags);
2068 rhead->num_fwd = req->r_num_fwd;
2069 rhead->num_retry = req->r_attempts - 1;
2072 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2077 * send request, or put it on the appropriate wait list.
2079 static int __do_request(struct ceph_mds_client *mdsc,
2080 struct ceph_mds_request *req)
2082 struct ceph_mds_session *session = NULL;
2086 if (req->r_err || req->r_got_result) {
2088 __unregister_request(mdsc, req);
2092 if (req->r_timeout &&
2093 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2094 dout("do_request timed out\n");
2098 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2099 dout("do_request forced umount\n");
2104 put_request_session(req);
2106 mds = __choose_mds(mdsc, req);
2108 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2109 if (mdsc->mdsmap_err) {
2110 err = mdsc->mdsmap_err;
2111 dout("do_request mdsmap err %d\n", err);
2114 dout("do_request no mds or not active, waiting for map\n");
2115 list_add(&req->r_wait, &mdsc->waiting_for_map);
2119 /* get, open session */
2120 session = __ceph_lookup_mds_session(mdsc, mds);
2122 session = register_session(mdsc, mds);
2123 if (IS_ERR(session)) {
2124 err = PTR_ERR(session);
2128 req->r_session = get_session(session);
2130 dout("do_request mds%d session %p state %s\n", mds, session,
2131 ceph_session_state_name(session->s_state));
2132 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2133 session->s_state != CEPH_MDS_SESSION_HUNG) {
2134 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2138 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2139 session->s_state == CEPH_MDS_SESSION_CLOSING)
2140 __open_session(mdsc, session);
2141 list_add(&req->r_wait, &session->s_waiting);
2146 req->r_resend_mds = -1; /* forget any previous mds hint */
2148 if (req->r_request_started == 0) /* note request start time */
2149 req->r_request_started = jiffies;
2151 err = __prepare_send_request(mdsc, req, mds, false);
2153 ceph_msg_get(req->r_request);
2154 ceph_con_send(&session->s_con, req->r_request);
2158 ceph_put_mds_session(session);
2161 dout("__do_request early error %d\n", err);
2163 complete_request(mdsc, req);
2164 __unregister_request(mdsc, req);
2171 * called under mdsc->mutex
2173 static void __wake_requests(struct ceph_mds_client *mdsc,
2174 struct list_head *head)
2176 struct ceph_mds_request *req;
2177 LIST_HEAD(tmp_list);
2179 list_splice_init(head, &tmp_list);
2181 while (!list_empty(&tmp_list)) {
2182 req = list_entry(tmp_list.next,
2183 struct ceph_mds_request, r_wait);
2184 list_del_init(&req->r_wait);
2185 dout(" wake request %p tid %llu\n", req, req->r_tid);
2186 __do_request(mdsc, req);
2191 * Wake up threads with requests pending for @mds, so that they can
2192 * resubmit their requests to a possibly different mds.
2194 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2196 struct ceph_mds_request *req;
2197 struct rb_node *p = rb_first(&mdsc->request_tree);
2199 dout("kick_requests mds%d\n", mds);
2201 req = rb_entry(p, struct ceph_mds_request, r_node);
2203 if (req->r_got_unsafe)
2205 if (req->r_attempts > 0)
2206 continue; /* only new requests */
2207 if (req->r_session &&
2208 req->r_session->s_mds == mds) {
2209 dout(" kicking tid %llu\n", req->r_tid);
2210 list_del_init(&req->r_wait);
2211 __do_request(mdsc, req);
2216 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2217 struct ceph_mds_request *req)
2219 dout("submit_request on %p\n", req);
2220 mutex_lock(&mdsc->mutex);
2221 __register_request(mdsc, req, NULL);
2222 __do_request(mdsc, req);
2223 mutex_unlock(&mdsc->mutex);
2227 * Synchrously perform an mds request. Take care of all of the
2228 * session setup, forwarding, retry details.
2230 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2232 struct ceph_mds_request *req)
2236 dout("do_request on %p\n", req);
2238 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2240 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2241 if (req->r_locked_dir)
2242 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2243 if (req->r_old_dentry_dir)
2244 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2248 mutex_lock(&mdsc->mutex);
2249 __register_request(mdsc, req, dir);
2250 __do_request(mdsc, req);
2258 mutex_unlock(&mdsc->mutex);
2259 dout("do_request waiting\n");
2260 if (!req->r_timeout && req->r_wait_for_completion) {
2261 err = req->r_wait_for_completion(mdsc, req);
2263 long timeleft = wait_for_completion_killable_timeout(
2265 ceph_timeout_jiffies(req->r_timeout));
2269 err = -EIO; /* timed out */
2271 err = timeleft; /* killed */
2273 dout("do_request waited, got %d\n", err);
2274 mutex_lock(&mdsc->mutex);
2276 /* only abort if we didn't race with a real reply */
2277 if (req->r_got_result) {
2278 err = le32_to_cpu(req->r_reply_info.head->result);
2279 } else if (err < 0) {
2280 dout("aborted request %lld with %d\n", req->r_tid, err);
2283 * ensure we aren't running concurrently with
2284 * ceph_fill_trace or ceph_readdir_prepopulate, which
2285 * rely on locks (dir mutex) held by our caller.
2287 mutex_lock(&req->r_fill_mutex);
2289 req->r_aborted = true;
2290 mutex_unlock(&req->r_fill_mutex);
2292 if (req->r_locked_dir &&
2293 (req->r_op & CEPH_MDS_OP_WRITE))
2294 ceph_invalidate_dir_request(req);
2300 mutex_unlock(&mdsc->mutex);
2301 dout("do_request %p done, result %d\n", req, err);
2306 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2307 * namespace request.
2309 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2311 struct inode *inode = req->r_locked_dir;
2313 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2315 ceph_dir_clear_complete(inode);
2317 ceph_invalidate_dentry_lease(req->r_dentry);
2318 if (req->r_old_dentry)
2319 ceph_invalidate_dentry_lease(req->r_old_dentry);
2325 * We take the session mutex and parse and process the reply immediately.
2326 * This preserves the logical ordering of replies, capabilities, etc., sent
2327 * by the MDS as they are applied to our local cache.
2329 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2331 struct ceph_mds_client *mdsc = session->s_mdsc;
2332 struct ceph_mds_request *req;
2333 struct ceph_mds_reply_head *head = msg->front.iov_base;
2334 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2335 struct ceph_snap_realm *realm;
2338 int mds = session->s_mds;
2340 if (msg->front.iov_len < sizeof(*head)) {
2341 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2346 /* get request, session */
2347 tid = le64_to_cpu(msg->hdr.tid);
2348 mutex_lock(&mdsc->mutex);
2349 req = lookup_get_request(mdsc, tid);
2351 dout("handle_reply on unknown tid %llu\n", tid);
2352 mutex_unlock(&mdsc->mutex);
2355 dout("handle_reply %p\n", req);
2357 /* correct session? */
2358 if (req->r_session != session) {
2359 pr_err("mdsc_handle_reply got %llu on session mds%d"
2360 " not mds%d\n", tid, session->s_mds,
2361 req->r_session ? req->r_session->s_mds : -1);
2362 mutex_unlock(&mdsc->mutex);
2367 if ((req->r_got_unsafe && !head->safe) ||
2368 (req->r_got_safe && head->safe)) {
2369 pr_warn("got a dup %s reply on %llu from mds%d\n",
2370 head->safe ? "safe" : "unsafe", tid, mds);
2371 mutex_unlock(&mdsc->mutex);
2374 if (req->r_got_safe) {
2375 pr_warn("got unsafe after safe on %llu from mds%d\n",
2377 mutex_unlock(&mdsc->mutex);
2381 result = le32_to_cpu(head->result);
2385 * if we're not talking to the authority, send to them
2386 * if the authority has changed while we weren't looking,
2387 * send to new authority
2388 * Otherwise we just have to return an ESTALE
2390 if (result == -ESTALE) {
2391 dout("got ESTALE on request %llu", req->r_tid);
2392 req->r_resend_mds = -1;
2393 if (req->r_direct_mode != USE_AUTH_MDS) {
2394 dout("not using auth, setting for that now");
2395 req->r_direct_mode = USE_AUTH_MDS;
2396 __do_request(mdsc, req);
2397 mutex_unlock(&mdsc->mutex);
2400 int mds = __choose_mds(mdsc, req);
2401 if (mds >= 0 && mds != req->r_session->s_mds) {
2402 dout("but auth changed, so resending");
2403 __do_request(mdsc, req);
2404 mutex_unlock(&mdsc->mutex);
2408 dout("have to return ESTALE on request %llu", req->r_tid);
2413 req->r_got_safe = true;
2414 __unregister_request(mdsc, req);
2416 if (req->r_got_unsafe) {
2418 * We already handled the unsafe response, now do the
2419 * cleanup. No need to examine the response; the MDS
2420 * doesn't include any result info in the safe
2421 * response. And even if it did, there is nothing
2422 * useful we could do with a revised return value.
2424 dout("got safe reply %llu, mds%d\n", tid, mds);
2425 list_del_init(&req->r_unsafe_item);
2427 /* last unsafe request during umount? */
2428 if (mdsc->stopping && !__get_oldest_req(mdsc))
2429 complete_all(&mdsc->safe_umount_waiters);
2430 mutex_unlock(&mdsc->mutex);
2434 req->r_got_unsafe = true;
2435 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2436 if (req->r_unsafe_dir) {
2437 struct ceph_inode_info *ci =
2438 ceph_inode(req->r_unsafe_dir);
2439 spin_lock(&ci->i_unsafe_lock);
2440 list_add_tail(&req->r_unsafe_dir_item,
2441 &ci->i_unsafe_dirops);
2442 spin_unlock(&ci->i_unsafe_lock);
2446 dout("handle_reply tid %lld result %d\n", tid, result);
2447 rinfo = &req->r_reply_info;
2448 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2449 mutex_unlock(&mdsc->mutex);
2451 mutex_lock(&session->s_mutex);
2453 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2460 if (rinfo->snapblob_len) {
2461 down_write(&mdsc->snap_rwsem);
2462 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2463 rinfo->snapblob + rinfo->snapblob_len,
2464 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2466 downgrade_write(&mdsc->snap_rwsem);
2468 down_read(&mdsc->snap_rwsem);
2471 /* insert trace into our cache */
2472 mutex_lock(&req->r_fill_mutex);
2473 current->journal_info = req;
2474 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2476 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2477 req->r_op == CEPH_MDS_OP_LSSNAP))
2478 ceph_readdir_prepopulate(req, req->r_session);
2479 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2481 current->journal_info = NULL;
2482 mutex_unlock(&req->r_fill_mutex);
2484 up_read(&mdsc->snap_rwsem);
2486 ceph_put_snap_realm(mdsc, realm);
2488 if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2489 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2490 spin_lock(&ci->i_unsafe_lock);
2491 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2492 spin_unlock(&ci->i_unsafe_lock);
2495 mutex_lock(&mdsc->mutex);
2496 if (!req->r_aborted) {
2500 req->r_reply = ceph_msg_get(msg);
2501 req->r_got_result = true;
2504 dout("reply arrived after request %lld was aborted\n", tid);
2506 mutex_unlock(&mdsc->mutex);
2508 mutex_unlock(&session->s_mutex);
2510 /* kick calling process */
2511 complete_request(mdsc, req);
2513 ceph_mdsc_put_request(req);
2520 * handle mds notification that our request has been forwarded.
2522 static void handle_forward(struct ceph_mds_client *mdsc,
2523 struct ceph_mds_session *session,
2524 struct ceph_msg *msg)
2526 struct ceph_mds_request *req;
2527 u64 tid = le64_to_cpu(msg->hdr.tid);
2531 void *p = msg->front.iov_base;
2532 void *end = p + msg->front.iov_len;
2534 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2535 next_mds = ceph_decode_32(&p);
2536 fwd_seq = ceph_decode_32(&p);
2538 mutex_lock(&mdsc->mutex);
2539 req = lookup_get_request(mdsc, tid);
2541 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2542 goto out; /* dup reply? */
2545 if (req->r_aborted) {
2546 dout("forward tid %llu aborted, unregistering\n", tid);
2547 __unregister_request(mdsc, req);
2548 } else if (fwd_seq <= req->r_num_fwd) {
2549 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2550 tid, next_mds, req->r_num_fwd, fwd_seq);
2552 /* resend. forward race not possible; mds would drop */
2553 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2555 BUG_ON(req->r_got_result);
2556 req->r_attempts = 0;
2557 req->r_num_fwd = fwd_seq;
2558 req->r_resend_mds = next_mds;
2559 put_request_session(req);
2560 __do_request(mdsc, req);
2562 ceph_mdsc_put_request(req);
2564 mutex_unlock(&mdsc->mutex);
2568 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2572 * handle a mds session control message
2574 static void handle_session(struct ceph_mds_session *session,
2575 struct ceph_msg *msg)
2577 struct ceph_mds_client *mdsc = session->s_mdsc;
2580 int mds = session->s_mds;
2581 struct ceph_mds_session_head *h = msg->front.iov_base;
2585 if (msg->front.iov_len != sizeof(*h))
2587 op = le32_to_cpu(h->op);
2588 seq = le64_to_cpu(h->seq);
2590 mutex_lock(&mdsc->mutex);
2591 if (op == CEPH_SESSION_CLOSE)
2592 __unregister_session(mdsc, session);
2593 /* FIXME: this ttl calculation is generous */
2594 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2595 mutex_unlock(&mdsc->mutex);
2597 mutex_lock(&session->s_mutex);
2599 dout("handle_session mds%d %s %p state %s seq %llu\n",
2600 mds, ceph_session_op_name(op), session,
2601 ceph_session_state_name(session->s_state), seq);
2603 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2604 session->s_state = CEPH_MDS_SESSION_OPEN;
2605 pr_info("mds%d came back\n", session->s_mds);
2609 case CEPH_SESSION_OPEN:
2610 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2611 pr_info("mds%d reconnect success\n", session->s_mds);
2612 session->s_state = CEPH_MDS_SESSION_OPEN;
2613 renewed_caps(mdsc, session, 0);
2616 __close_session(mdsc, session);
2619 case CEPH_SESSION_RENEWCAPS:
2620 if (session->s_renew_seq == seq)
2621 renewed_caps(mdsc, session, 1);
2624 case CEPH_SESSION_CLOSE:
2625 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2626 pr_info("mds%d reconnect denied\n", session->s_mds);
2627 cleanup_session_requests(mdsc, session);
2628 remove_session_caps(session);
2629 wake = 2; /* for good measure */
2630 wake_up_all(&mdsc->session_close_wq);
2633 case CEPH_SESSION_STALE:
2634 pr_info("mds%d caps went stale, renewing\n",
2636 spin_lock(&session->s_gen_ttl_lock);
2637 session->s_cap_gen++;
2638 session->s_cap_ttl = jiffies - 1;
2639 spin_unlock(&session->s_gen_ttl_lock);
2640 send_renew_caps(mdsc, session);
2643 case CEPH_SESSION_RECALL_STATE:
2644 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2647 case CEPH_SESSION_FLUSHMSG:
2648 send_flushmsg_ack(mdsc, session, seq);
2651 case CEPH_SESSION_FORCE_RO:
2652 dout("force_session_readonly %p\n", session);
2653 spin_lock(&session->s_cap_lock);
2654 session->s_readonly = true;
2655 spin_unlock(&session->s_cap_lock);
2656 wake_up_session_caps(session, 0);
2659 case CEPH_SESSION_REJECT:
2660 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2661 pr_info("mds%d rejected session\n", session->s_mds);
2662 session->s_state = CEPH_MDS_SESSION_REJECTED;
2663 cleanup_session_requests(mdsc, session);
2664 remove_session_caps(session);
2665 wake = 2; /* for good measure */
2669 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2673 mutex_unlock(&session->s_mutex);
2675 mutex_lock(&mdsc->mutex);
2676 __wake_requests(mdsc, &session->s_waiting);
2678 kick_requests(mdsc, mds);
2679 mutex_unlock(&mdsc->mutex);
2684 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2685 (int)msg->front.iov_len);
2692 * called under session->mutex.
2694 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2695 struct ceph_mds_session *session)
2697 struct ceph_mds_request *req, *nreq;
2701 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2703 mutex_lock(&mdsc->mutex);
2704 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2705 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2707 ceph_msg_get(req->r_request);
2708 ceph_con_send(&session->s_con, req->r_request);
2713 * also re-send old requests when MDS enters reconnect stage. So that MDS
2714 * can process completed request in clientreplay stage.
2716 p = rb_first(&mdsc->request_tree);
2718 req = rb_entry(p, struct ceph_mds_request, r_node);
2720 if (req->r_got_unsafe)
2722 if (req->r_attempts == 0)
2723 continue; /* only old requests */
2724 if (req->r_session &&
2725 req->r_session->s_mds == session->s_mds) {
2726 err = __prepare_send_request(mdsc, req,
2727 session->s_mds, true);
2729 ceph_msg_get(req->r_request);
2730 ceph_con_send(&session->s_con, req->r_request);
2734 mutex_unlock(&mdsc->mutex);
2738 * Encode information about a cap for a reconnect with the MDS.
2740 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2744 struct ceph_mds_cap_reconnect v2;
2745 struct ceph_mds_cap_reconnect_v1 v1;
2747 struct ceph_inode_info *ci;
2748 struct ceph_reconnect_state *recon_state = arg;
2749 struct ceph_pagelist *pagelist = recon_state->pagelist;
2754 struct dentry *dentry;
2758 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2759 inode, ceph_vinop(inode), cap, cap->cap_id,
2760 ceph_cap_string(cap->issued));
2761 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2765 dentry = d_find_alias(inode);
2767 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2769 err = PTR_ERR(path);
2778 spin_lock(&ci->i_ceph_lock);
2779 cap->seq = 0; /* reset cap seq */
2780 cap->issue_seq = 0; /* and issue_seq */
2781 cap->mseq = 0; /* and migrate_seq */
2782 cap->cap_gen = cap->session->s_cap_gen;
2784 if (recon_state->msg_version >= 2) {
2785 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2786 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2787 rec.v2.issued = cpu_to_le32(cap->issued);
2788 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2789 rec.v2.pathbase = cpu_to_le64(pathbase);
2790 rec.v2.flock_len = 0;
2792 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2793 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2794 rec.v1.issued = cpu_to_le32(cap->issued);
2795 rec.v1.size = cpu_to_le64(inode->i_size);
2796 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2797 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2798 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2799 rec.v1.pathbase = cpu_to_le64(pathbase);
2802 if (list_empty(&ci->i_cap_snaps)) {
2805 struct ceph_cap_snap *capsnap =
2806 list_first_entry(&ci->i_cap_snaps,
2807 struct ceph_cap_snap, ci_item);
2808 snap_follows = capsnap->follows;
2810 spin_unlock(&ci->i_ceph_lock);
2812 if (recon_state->msg_version >= 2) {
2813 int num_fcntl_locks, num_flock_locks;
2814 struct ceph_filelock *flocks;
2815 size_t struct_len, total_len = 0;
2819 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2820 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2821 sizeof(struct ceph_filelock), GFP_NOFS);
2826 err = ceph_encode_locks_to_buffer(inode, flocks,
2836 if (recon_state->msg_version >= 3) {
2837 /* version, compat_version and struct_len */
2838 total_len = 2 * sizeof(u8) + sizeof(u32);
2842 * number of encoded locks is stable, so copy to pagelist
2844 struct_len = 2 * sizeof(u32) +
2845 (num_fcntl_locks + num_flock_locks) *
2846 sizeof(struct ceph_filelock);
2847 rec.v2.flock_len = cpu_to_le32(struct_len);
2849 struct_len += sizeof(rec.v2);
2850 struct_len += sizeof(u32) + pathlen;
2853 struct_len += sizeof(u64); /* snap_follows */
2855 total_len += struct_len;
2856 err = ceph_pagelist_reserve(pagelist, total_len);
2859 if (recon_state->msg_version >= 3) {
2860 ceph_pagelist_encode_8(pagelist, struct_v);
2861 ceph_pagelist_encode_8(pagelist, 1);
2862 ceph_pagelist_encode_32(pagelist, struct_len);
2864 ceph_pagelist_encode_string(pagelist, path, pathlen);
2865 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2866 ceph_locks_to_pagelist(flocks, pagelist,
2870 ceph_pagelist_encode_64(pagelist, snap_follows);
2874 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2875 err = ceph_pagelist_reserve(pagelist, size);
2877 ceph_pagelist_encode_string(pagelist, path, pathlen);
2878 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2882 recon_state->nr_caps++;
2892 * If an MDS fails and recovers, clients need to reconnect in order to
2893 * reestablish shared state. This includes all caps issued through
2894 * this session _and_ the snap_realm hierarchy. Because it's not
2895 * clear which snap realms the mds cares about, we send everything we
2896 * know about.. that ensures we'll then get any new info the
2897 * recovering MDS might have.
2899 * This is a relatively heavyweight operation, but it's rare.
2901 * called with mdsc->mutex held.
2903 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2904 struct ceph_mds_session *session)
2906 struct ceph_msg *reply;
2908 int mds = session->s_mds;
2911 struct ceph_pagelist *pagelist;
2912 struct ceph_reconnect_state recon_state;
2914 pr_info("mds%d reconnect start\n", mds);
2916 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2918 goto fail_nopagelist;
2919 ceph_pagelist_init(pagelist);
2921 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2925 mutex_lock(&session->s_mutex);
2926 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2929 dout("session %p state %s\n", session,
2930 ceph_session_state_name(session->s_state));
2932 spin_lock(&session->s_gen_ttl_lock);
2933 session->s_cap_gen++;
2934 spin_unlock(&session->s_gen_ttl_lock);
2936 spin_lock(&session->s_cap_lock);
2937 /* don't know if session is readonly */
2938 session->s_readonly = 0;
2940 * notify __ceph_remove_cap() that we are composing cap reconnect.
2941 * If a cap get released before being added to the cap reconnect,
2942 * __ceph_remove_cap() should skip queuing cap release.
2944 session->s_cap_reconnect = 1;
2945 /* drop old cap expires; we're about to reestablish that state */
2946 cleanup_cap_releases(mdsc, session);
2948 /* trim unused caps to reduce MDS's cache rejoin time */
2949 if (mdsc->fsc->sb->s_root)
2950 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2952 ceph_con_close(&session->s_con);
2953 ceph_con_open(&session->s_con,
2954 CEPH_ENTITY_TYPE_MDS, mds,
2955 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2957 /* replay unsafe requests */
2958 replay_unsafe_requests(mdsc, session);
2960 down_read(&mdsc->snap_rwsem);
2962 /* traverse this session's caps */
2963 s_nr_caps = session->s_nr_caps;
2964 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2968 recon_state.nr_caps = 0;
2969 recon_state.pagelist = pagelist;
2970 if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
2971 recon_state.msg_version = 3;
2972 else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
2973 recon_state.msg_version = 2;
2975 recon_state.msg_version = 1;
2976 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2980 spin_lock(&session->s_cap_lock);
2981 session->s_cap_reconnect = 0;
2982 spin_unlock(&session->s_cap_lock);
2985 * snaprealms. we provide mds with the ino, seq (version), and
2986 * parent for all of our realms. If the mds has any newer info,
2989 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2990 struct ceph_snap_realm *realm =
2991 rb_entry(p, struct ceph_snap_realm, node);
2992 struct ceph_mds_snaprealm_reconnect sr_rec;
2994 dout(" adding snap realm %llx seq %lld parent %llx\n",
2995 realm->ino, realm->seq, realm->parent_ino);
2996 sr_rec.ino = cpu_to_le64(realm->ino);
2997 sr_rec.seq = cpu_to_le64(realm->seq);
2998 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2999 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3004 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3006 /* raced with cap release? */
3007 if (s_nr_caps != recon_state.nr_caps) {
3008 struct page *page = list_first_entry(&pagelist->head,
3010 __le32 *addr = kmap_atomic(page);
3011 *addr = cpu_to_le32(recon_state.nr_caps);
3012 kunmap_atomic(addr);
3015 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3016 ceph_msg_data_add_pagelist(reply, pagelist);
3018 ceph_early_kick_flushing_caps(mdsc, session);
3020 ceph_con_send(&session->s_con, reply);
3022 mutex_unlock(&session->s_mutex);
3024 mutex_lock(&mdsc->mutex);
3025 __wake_requests(mdsc, &session->s_waiting);
3026 mutex_unlock(&mdsc->mutex);
3028 up_read(&mdsc->snap_rwsem);
3032 ceph_msg_put(reply);
3033 up_read(&mdsc->snap_rwsem);
3034 mutex_unlock(&session->s_mutex);
3036 ceph_pagelist_release(pagelist);
3038 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3044 * compare old and new mdsmaps, kicking requests
3045 * and closing out old connections as necessary
3047 * called under mdsc->mutex.
3049 static void check_new_map(struct ceph_mds_client *mdsc,
3050 struct ceph_mdsmap *newmap,
3051 struct ceph_mdsmap *oldmap)
3054 int oldstate, newstate;
3055 struct ceph_mds_session *s;
3057 dout("check_new_map new %u old %u\n",
3058 newmap->m_epoch, oldmap->m_epoch);
3060 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3061 if (mdsc->sessions[i] == NULL)
3063 s = mdsc->sessions[i];
3064 oldstate = ceph_mdsmap_get_state(oldmap, i);
3065 newstate = ceph_mdsmap_get_state(newmap, i);
3067 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3068 i, ceph_mds_state_name(oldstate),
3069 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3070 ceph_mds_state_name(newstate),
3071 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3072 ceph_session_state_name(s->s_state));
3074 if (i >= newmap->m_max_mds ||
3075 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3076 ceph_mdsmap_get_addr(newmap, i),
3077 sizeof(struct ceph_entity_addr))) {
3078 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3079 /* the session never opened, just close it
3081 __wake_requests(mdsc, &s->s_waiting);
3082 __unregister_session(mdsc, s);
3085 mutex_unlock(&mdsc->mutex);
3086 mutex_lock(&s->s_mutex);
3087 mutex_lock(&mdsc->mutex);
3088 ceph_con_close(&s->s_con);
3089 mutex_unlock(&s->s_mutex);
3090 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3092 } else if (oldstate == newstate) {
3093 continue; /* nothing new with this mds */
3099 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3100 newstate >= CEPH_MDS_STATE_RECONNECT) {
3101 mutex_unlock(&mdsc->mutex);
3102 send_mds_reconnect(mdsc, s);
3103 mutex_lock(&mdsc->mutex);
3107 * kick request on any mds that has gone active.
3109 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3110 newstate >= CEPH_MDS_STATE_ACTIVE) {
3111 if (oldstate != CEPH_MDS_STATE_CREATING &&
3112 oldstate != CEPH_MDS_STATE_STARTING)
3113 pr_info("mds%d recovery completed\n", s->s_mds);
3114 kick_requests(mdsc, i);
3115 ceph_kick_flushing_caps(mdsc, s);
3116 wake_up_session_caps(s, 1);
3120 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3121 s = mdsc->sessions[i];
3124 if (!ceph_mdsmap_is_laggy(newmap, i))
3126 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3127 s->s_state == CEPH_MDS_SESSION_HUNG ||
3128 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3129 dout(" connecting to export targets of laggy mds%d\n",
3131 __open_export_target_sessions(mdsc, s);
3143 * caller must hold session s_mutex, dentry->d_lock
3145 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3147 struct ceph_dentry_info *di = ceph_dentry(dentry);
3149 ceph_put_mds_session(di->lease_session);
3150 di->lease_session = NULL;
3153 static void handle_lease(struct ceph_mds_client *mdsc,
3154 struct ceph_mds_session *session,
3155 struct ceph_msg *msg)
3157 struct super_block *sb = mdsc->fsc->sb;
3158 struct inode *inode;
3159 struct dentry *parent, *dentry;
3160 struct ceph_dentry_info *di;
3161 int mds = session->s_mds;
3162 struct ceph_mds_lease *h = msg->front.iov_base;
3164 struct ceph_vino vino;
3168 dout("handle_lease from mds%d\n", mds);
3171 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3173 vino.ino = le64_to_cpu(h->ino);
3174 vino.snap = CEPH_NOSNAP;
3175 seq = le32_to_cpu(h->seq);
3176 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3177 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3178 if (dname.len != get_unaligned_le32(h+1))
3182 inode = ceph_find_inode(sb, vino);
3183 dout("handle_lease %s, ino %llx %p %.*s\n",
3184 ceph_lease_op_name(h->action), vino.ino, inode,
3185 dname.len, dname.name);
3187 mutex_lock(&session->s_mutex);
3190 if (inode == NULL) {
3191 dout("handle_lease no inode %llx\n", vino.ino);
3196 parent = d_find_alias(inode);
3198 dout("no parent dentry on inode %p\n", inode);
3200 goto release; /* hrm... */
3202 dname.hash = full_name_hash(parent, dname.name, dname.len);
3203 dentry = d_lookup(parent, &dname);
3208 spin_lock(&dentry->d_lock);
3209 di = ceph_dentry(dentry);
3210 switch (h->action) {
3211 case CEPH_MDS_LEASE_REVOKE:
3212 if (di->lease_session == session) {
3213 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3214 h->seq = cpu_to_le32(di->lease_seq);
3215 __ceph_mdsc_drop_dentry_lease(dentry);
3220 case CEPH_MDS_LEASE_RENEW:
3221 if (di->lease_session == session &&
3222 di->lease_gen == session->s_cap_gen &&
3223 di->lease_renew_from &&
3224 di->lease_renew_after == 0) {
3225 unsigned long duration =
3226 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3228 di->lease_seq = seq;
3229 di->time = di->lease_renew_from + duration;
3230 di->lease_renew_after = di->lease_renew_from +
3232 di->lease_renew_from = 0;
3236 spin_unlock(&dentry->d_lock);
3243 /* let's just reuse the same message */
3244 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3246 ceph_con_send(&session->s_con, msg);
3250 mutex_unlock(&session->s_mutex);
3254 pr_err("corrupt lease message\n");
3258 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3259 struct inode *inode,
3260 struct dentry *dentry, char action,
3263 struct ceph_msg *msg;
3264 struct ceph_mds_lease *lease;
3265 int len = sizeof(*lease) + sizeof(u32);
3268 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3269 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3270 dnamelen = dentry->d_name.len;
3273 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3276 lease = msg->front.iov_base;
3277 lease->action = action;
3278 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3279 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3280 lease->seq = cpu_to_le32(seq);
3281 put_unaligned_le32(dnamelen, lease + 1);
3282 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3285 * if this is a preemptive lease RELEASE, no need to
3286 * flush request stream, since the actual request will
3289 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3291 ceph_con_send(&session->s_con, msg);
3295 * drop all leases (and dentry refs) in preparation for umount
3297 static void drop_leases(struct ceph_mds_client *mdsc)
3301 dout("drop_leases\n");
3302 mutex_lock(&mdsc->mutex);
3303 for (i = 0; i < mdsc->max_sessions; i++) {
3304 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3307 mutex_unlock(&mdsc->mutex);
3308 mutex_lock(&s->s_mutex);
3309 mutex_unlock(&s->s_mutex);
3310 ceph_put_mds_session(s);
3311 mutex_lock(&mdsc->mutex);
3313 mutex_unlock(&mdsc->mutex);
3319 * delayed work -- periodically trim expired leases, renew caps with mds
3321 static void schedule_delayed(struct ceph_mds_client *mdsc)
3324 unsigned hz = round_jiffies_relative(HZ * delay);
3325 schedule_delayed_work(&mdsc->delayed_work, hz);
3328 static void delayed_work(struct work_struct *work)
3331 struct ceph_mds_client *mdsc =
3332 container_of(work, struct ceph_mds_client, delayed_work.work);
3336 dout("mdsc delayed_work\n");
3337 ceph_check_delayed_caps(mdsc);
3339 mutex_lock(&mdsc->mutex);
3340 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3341 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3342 mdsc->last_renew_caps);
3344 mdsc->last_renew_caps = jiffies;
3346 for (i = 0; i < mdsc->max_sessions; i++) {
3347 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3350 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3351 dout("resending session close request for mds%d\n",
3353 request_close_session(mdsc, s);
3354 ceph_put_mds_session(s);
3357 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3358 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3359 s->s_state = CEPH_MDS_SESSION_HUNG;
3360 pr_info("mds%d hung\n", s->s_mds);
3363 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3364 /* this mds is failed or recovering, just wait */
3365 ceph_put_mds_session(s);
3368 mutex_unlock(&mdsc->mutex);
3370 mutex_lock(&s->s_mutex);
3372 send_renew_caps(mdsc, s);
3374 ceph_con_keepalive(&s->s_con);
3375 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3376 s->s_state == CEPH_MDS_SESSION_HUNG)
3377 ceph_send_cap_releases(mdsc, s);
3378 mutex_unlock(&s->s_mutex);
3379 ceph_put_mds_session(s);
3381 mutex_lock(&mdsc->mutex);
3383 mutex_unlock(&mdsc->mutex);
3385 schedule_delayed(mdsc);
3388 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3391 struct ceph_mds_client *mdsc;
3393 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3398 mutex_init(&mdsc->mutex);
3399 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3400 if (mdsc->mdsmap == NULL) {
3405 init_completion(&mdsc->safe_umount_waiters);
3406 init_waitqueue_head(&mdsc->session_close_wq);
3407 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3408 mdsc->sessions = NULL;
3409 atomic_set(&mdsc->num_sessions, 0);
3410 mdsc->max_sessions = 0;
3412 mdsc->last_snap_seq = 0;
3413 init_rwsem(&mdsc->snap_rwsem);
3414 mdsc->snap_realms = RB_ROOT;
3415 INIT_LIST_HEAD(&mdsc->snap_empty);
3416 spin_lock_init(&mdsc->snap_empty_lock);
3418 mdsc->oldest_tid = 0;
3419 mdsc->request_tree = RB_ROOT;
3420 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3421 mdsc->last_renew_caps = jiffies;
3422 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3423 spin_lock_init(&mdsc->cap_delay_lock);
3424 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3425 spin_lock_init(&mdsc->snap_flush_lock);
3426 mdsc->last_cap_flush_tid = 1;
3427 INIT_LIST_HEAD(&mdsc->cap_flush_list);
3428 INIT_LIST_HEAD(&mdsc->cap_dirty);
3429 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3430 mdsc->num_cap_flushing = 0;
3431 spin_lock_init(&mdsc->cap_dirty_lock);
3432 init_waitqueue_head(&mdsc->cap_flushing_wq);
3433 spin_lock_init(&mdsc->dentry_lru_lock);
3434 INIT_LIST_HEAD(&mdsc->dentry_lru);
3436 ceph_caps_init(mdsc);
3437 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3439 init_rwsem(&mdsc->pool_perm_rwsem);
3440 mdsc->pool_perm_tree = RB_ROOT;
3446 * Wait for safe replies on open mds requests. If we time out, drop
3447 * all requests from the tree to avoid dangling dentry refs.
3449 static void wait_requests(struct ceph_mds_client *mdsc)
3451 struct ceph_options *opts = mdsc->fsc->client->options;
3452 struct ceph_mds_request *req;
3454 mutex_lock(&mdsc->mutex);
3455 if (__get_oldest_req(mdsc)) {
3456 mutex_unlock(&mdsc->mutex);
3458 dout("wait_requests waiting for requests\n");
3459 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3460 ceph_timeout_jiffies(opts->mount_timeout));
3462 /* tear down remaining requests */
3463 mutex_lock(&mdsc->mutex);
3464 while ((req = __get_oldest_req(mdsc))) {
3465 dout("wait_requests timed out on tid %llu\n",
3467 __unregister_request(mdsc, req);
3470 mutex_unlock(&mdsc->mutex);
3471 dout("wait_requests done\n");
3475 * called before mount is ro, and before dentries are torn down.
3476 * (hmm, does this still race with new lookups?)
3478 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3480 dout("pre_umount\n");
3484 ceph_flush_dirty_caps(mdsc);
3485 wait_requests(mdsc);
3488 * wait for reply handlers to drop their request refs and
3489 * their inode/dcache refs
3495 * wait for all write mds requests to flush.
3497 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3499 struct ceph_mds_request *req = NULL, *nextreq;
3502 mutex_lock(&mdsc->mutex);
3503 dout("wait_unsafe_requests want %lld\n", want_tid);
3505 req = __get_oldest_req(mdsc);
3506 while (req && req->r_tid <= want_tid) {
3507 /* find next request */
3508 n = rb_next(&req->r_node);
3510 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3513 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3514 (req->r_op & CEPH_MDS_OP_WRITE)) {
3516 ceph_mdsc_get_request(req);
3518 ceph_mdsc_get_request(nextreq);
3519 mutex_unlock(&mdsc->mutex);
3520 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3521 req->r_tid, want_tid);
3522 wait_for_completion(&req->r_safe_completion);
3523 mutex_lock(&mdsc->mutex);
3524 ceph_mdsc_put_request(req);
3526 break; /* next dne before, so we're done! */
3527 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3528 /* next request was removed from tree */
3529 ceph_mdsc_put_request(nextreq);
3532 ceph_mdsc_put_request(nextreq); /* won't go away */
3536 mutex_unlock(&mdsc->mutex);
3537 dout("wait_unsafe_requests done\n");
3540 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3542 u64 want_tid, want_flush;
3544 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3548 mutex_lock(&mdsc->mutex);
3549 want_tid = mdsc->last_tid;
3550 mutex_unlock(&mdsc->mutex);
3552 ceph_flush_dirty_caps(mdsc);
3553 spin_lock(&mdsc->cap_dirty_lock);
3554 want_flush = mdsc->last_cap_flush_tid;
3555 if (!list_empty(&mdsc->cap_flush_list)) {
3556 struct ceph_cap_flush *cf =
3557 list_last_entry(&mdsc->cap_flush_list,
3558 struct ceph_cap_flush, g_list);
3561 spin_unlock(&mdsc->cap_dirty_lock);
3563 dout("sync want tid %lld flush_seq %lld\n",
3564 want_tid, want_flush);
3566 wait_unsafe_requests(mdsc, want_tid);
3567 wait_caps_flush(mdsc, want_flush);
3571 * true if all sessions are closed, or we force unmount
3573 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3575 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3577 return atomic_read(&mdsc->num_sessions) <= skipped;
3581 * called after sb is ro.
3583 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3585 struct ceph_options *opts = mdsc->fsc->client->options;
3586 struct ceph_mds_session *session;
3590 dout("close_sessions\n");
3592 /* close sessions */
3593 mutex_lock(&mdsc->mutex);
3594 for (i = 0; i < mdsc->max_sessions; i++) {
3595 session = __ceph_lookup_mds_session(mdsc, i);
3598 mutex_unlock(&mdsc->mutex);
3599 mutex_lock(&session->s_mutex);
3600 if (__close_session(mdsc, session) <= 0)
3602 mutex_unlock(&session->s_mutex);
3603 ceph_put_mds_session(session);
3604 mutex_lock(&mdsc->mutex);
3606 mutex_unlock(&mdsc->mutex);
3608 dout("waiting for sessions to close\n");
3609 wait_event_timeout(mdsc->session_close_wq,
3610 done_closing_sessions(mdsc, skipped),
3611 ceph_timeout_jiffies(opts->mount_timeout));
3613 /* tear down remaining sessions */
3614 mutex_lock(&mdsc->mutex);
3615 for (i = 0; i < mdsc->max_sessions; i++) {
3616 if (mdsc->sessions[i]) {
3617 session = get_session(mdsc->sessions[i]);
3618 __unregister_session(mdsc, session);
3619 mutex_unlock(&mdsc->mutex);
3620 mutex_lock(&session->s_mutex);
3621 remove_session_caps(session);
3622 mutex_unlock(&session->s_mutex);
3623 ceph_put_mds_session(session);
3624 mutex_lock(&mdsc->mutex);
3627 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3628 mutex_unlock(&mdsc->mutex);
3630 ceph_cleanup_empty_realms(mdsc);
3632 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3637 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3639 struct ceph_mds_session *session;
3642 dout("force umount\n");
3644 mutex_lock(&mdsc->mutex);
3645 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3646 session = __ceph_lookup_mds_session(mdsc, mds);
3649 mutex_unlock(&mdsc->mutex);
3650 mutex_lock(&session->s_mutex);
3651 __close_session(mdsc, session);
3652 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3653 cleanup_session_requests(mdsc, session);
3654 remove_session_caps(session);
3656 mutex_unlock(&session->s_mutex);
3657 ceph_put_mds_session(session);
3658 mutex_lock(&mdsc->mutex);
3659 kick_requests(mdsc, mds);
3661 __wake_requests(mdsc, &mdsc->waiting_for_map);
3662 mutex_unlock(&mdsc->mutex);
3665 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3668 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3670 ceph_mdsmap_destroy(mdsc->mdsmap);
3671 kfree(mdsc->sessions);
3672 ceph_caps_finalize(mdsc);
3673 ceph_pool_perm_destroy(mdsc);
3676 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3678 struct ceph_mds_client *mdsc = fsc->mdsc;
3680 dout("mdsc_destroy %p\n", mdsc);
3681 ceph_mdsc_stop(mdsc);
3683 /* flush out any connection work with references to us */
3688 dout("mdsc_destroy %p done\n", mdsc);
3691 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3693 struct ceph_fs_client *fsc = mdsc->fsc;
3694 const char *mds_namespace = fsc->mount_options->mds_namespace;
3695 void *p = msg->front.iov_base;
3696 void *end = p + msg->front.iov_len;
3700 u32 mount_fscid = (u32)-1;
3701 u8 struct_v, struct_cv;
3704 ceph_decode_need(&p, end, sizeof(u32), bad);
3705 epoch = ceph_decode_32(&p);
3707 dout("handle_fsmap epoch %u\n", epoch);
3709 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3710 struct_v = ceph_decode_8(&p);
3711 struct_cv = ceph_decode_8(&p);
3712 map_len = ceph_decode_32(&p);
3714 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3715 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3717 num_fs = ceph_decode_32(&p);
3718 while (num_fs-- > 0) {
3719 void *info_p, *info_end;
3724 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3725 info_v = ceph_decode_8(&p);
3726 info_cv = ceph_decode_8(&p);
3727 info_len = ceph_decode_32(&p);
3728 ceph_decode_need(&p, end, info_len, bad);
3730 info_end = p + info_len;
3733 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3734 fscid = ceph_decode_32(&info_p);
3735 namelen = ceph_decode_32(&info_p);
3736 ceph_decode_need(&info_p, info_end, namelen, bad);
3738 if (mds_namespace &&
3739 strlen(mds_namespace) == namelen &&
3740 !strncmp(mds_namespace, (char *)info_p, namelen)) {
3741 mount_fscid = fscid;
3746 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3747 if (mount_fscid != (u32)-1) {
3748 fsc->client->monc.fs_cluster_id = mount_fscid;
3749 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3751 ceph_monc_renew_subs(&fsc->client->monc);
3758 pr_err("error decoding fsmap\n");
3760 mutex_lock(&mdsc->mutex);
3761 mdsc->mdsmap_err = -ENOENT;
3762 __wake_requests(mdsc, &mdsc->waiting_for_map);
3763 mutex_unlock(&mdsc->mutex);
3768 * handle mds map update.
3770 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3774 void *p = msg->front.iov_base;
3775 void *end = p + msg->front.iov_len;
3776 struct ceph_mdsmap *newmap, *oldmap;
3777 struct ceph_fsid fsid;
3780 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3781 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3782 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3784 epoch = ceph_decode_32(&p);
3785 maplen = ceph_decode_32(&p);
3786 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3788 /* do we need it? */
3789 mutex_lock(&mdsc->mutex);
3790 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3791 dout("handle_map epoch %u <= our %u\n",
3792 epoch, mdsc->mdsmap->m_epoch);
3793 mutex_unlock(&mdsc->mutex);
3797 newmap = ceph_mdsmap_decode(&p, end);
3798 if (IS_ERR(newmap)) {
3799 err = PTR_ERR(newmap);
3803 /* swap into place */
3805 oldmap = mdsc->mdsmap;
3806 mdsc->mdsmap = newmap;
3807 check_new_map(mdsc, newmap, oldmap);
3808 ceph_mdsmap_destroy(oldmap);
3810 mdsc->mdsmap = newmap; /* first mds map */
3812 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3814 __wake_requests(mdsc, &mdsc->waiting_for_map);
3815 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3816 mdsc->mdsmap->m_epoch);
3818 mutex_unlock(&mdsc->mutex);
3819 schedule_delayed(mdsc);
3823 mutex_unlock(&mdsc->mutex);
3825 pr_err("error decoding mdsmap %d\n", err);
3829 static struct ceph_connection *con_get(struct ceph_connection *con)
3831 struct ceph_mds_session *s = con->private;
3833 if (get_session(s)) {
3834 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3837 dout("mdsc con_get %p FAIL\n", s);
3841 static void con_put(struct ceph_connection *con)
3843 struct ceph_mds_session *s = con->private;
3845 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3846 ceph_put_mds_session(s);
3850 * if the client is unresponsive for long enough, the mds will kill
3851 * the session entirely.
3853 static void peer_reset(struct ceph_connection *con)
3855 struct ceph_mds_session *s = con->private;
3856 struct ceph_mds_client *mdsc = s->s_mdsc;
3858 pr_warn("mds%d closed our session\n", s->s_mds);
3859 send_mds_reconnect(mdsc, s);
3862 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3864 struct ceph_mds_session *s = con->private;
3865 struct ceph_mds_client *mdsc = s->s_mdsc;
3866 int type = le16_to_cpu(msg->hdr.type);
3868 mutex_lock(&mdsc->mutex);
3869 if (__verify_registered_session(mdsc, s) < 0) {
3870 mutex_unlock(&mdsc->mutex);
3873 mutex_unlock(&mdsc->mutex);
3876 case CEPH_MSG_MDS_MAP:
3877 ceph_mdsc_handle_mdsmap(mdsc, msg);
3879 case CEPH_MSG_FS_MAP_USER:
3880 ceph_mdsc_handle_fsmap(mdsc, msg);
3882 case CEPH_MSG_CLIENT_SESSION:
3883 handle_session(s, msg);
3885 case CEPH_MSG_CLIENT_REPLY:
3886 handle_reply(s, msg);
3888 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3889 handle_forward(mdsc, s, msg);
3891 case CEPH_MSG_CLIENT_CAPS:
3892 ceph_handle_caps(s, msg);
3894 case CEPH_MSG_CLIENT_SNAP:
3895 ceph_handle_snap(mdsc, s, msg);
3897 case CEPH_MSG_CLIENT_LEASE:
3898 handle_lease(mdsc, s, msg);
3902 pr_err("received unknown message type %d %s\n", type,
3903 ceph_msg_type_name(type));
3914 * Note: returned pointer is the address of a structure that's
3915 * managed separately. Caller must *not* attempt to free it.
3917 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3918 int *proto, int force_new)
3920 struct ceph_mds_session *s = con->private;
3921 struct ceph_mds_client *mdsc = s->s_mdsc;
3922 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3923 struct ceph_auth_handshake *auth = &s->s_auth;
3925 if (force_new && auth->authorizer) {
3926 ceph_auth_destroy_authorizer(auth->authorizer);
3927 auth->authorizer = NULL;
3929 if (!auth->authorizer) {
3930 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3933 return ERR_PTR(ret);
3935 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3938 return ERR_PTR(ret);
3940 *proto = ac->protocol;
3946 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3948 struct ceph_mds_session *s = con->private;
3949 struct ceph_mds_client *mdsc = s->s_mdsc;
3950 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3952 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3955 static int invalidate_authorizer(struct ceph_connection *con)
3957 struct ceph_mds_session *s = con->private;
3958 struct ceph_mds_client *mdsc = s->s_mdsc;
3959 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3961 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3963 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3966 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3967 struct ceph_msg_header *hdr, int *skip)
3969 struct ceph_msg *msg;
3970 int type = (int) le16_to_cpu(hdr->type);
3971 int front_len = (int) le32_to_cpu(hdr->front_len);
3977 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3979 pr_err("unable to allocate msg type %d len %d\n",
3987 static int mds_sign_message(struct ceph_msg *msg)
3989 struct ceph_mds_session *s = msg->con->private;
3990 struct ceph_auth_handshake *auth = &s->s_auth;
3992 return ceph_auth_sign_message(auth, msg);
3995 static int mds_check_message_signature(struct ceph_msg *msg)
3997 struct ceph_mds_session *s = msg->con->private;
3998 struct ceph_auth_handshake *auth = &s->s_auth;
4000 return ceph_auth_check_message_signature(auth, msg);
4003 static const struct ceph_connection_operations mds_con_ops = {
4006 .dispatch = dispatch,
4007 .get_authorizer = get_authorizer,
4008 .verify_authorizer_reply = verify_authorizer_reply,
4009 .invalidate_authorizer = invalidate_authorizer,
4010 .peer_reset = peer_reset,
4011 .alloc_msg = mds_alloc_msg,
4012 .sign_message = mds_sign_message,
4013 .check_message_signature = mds_check_message_signature,