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45 * are no longer valid.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 default: return "???";
377 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
379 if (atomic_inc_not_zero(&s->s_ref)) {
380 dout("mdsc get_session %p %d -> %d\n", s,
381 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
384 dout("mdsc get_session %p 0 -- FAIL", s);
389 void ceph_put_mds_session(struct ceph_mds_session *s)
391 dout("mdsc put_session %p %d -> %d\n", s,
392 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
393 if (atomic_dec_and_test(&s->s_ref)) {
394 if (s->s_auth.authorizer)
395 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
401 * called under mdsc->mutex
403 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
406 struct ceph_mds_session *session;
408 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
410 session = mdsc->sessions[mds];
411 dout("lookup_mds_session %p %d\n", session,
412 atomic_read(&session->s_ref));
413 get_session(session);
417 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
419 if (mds >= mdsc->max_sessions)
421 return mdsc->sessions[mds];
424 static int __verify_registered_session(struct ceph_mds_client *mdsc,
425 struct ceph_mds_session *s)
427 if (s->s_mds >= mdsc->max_sessions ||
428 mdsc->sessions[s->s_mds] != s)
434 * create+register a new session for given mds.
435 * called under mdsc->mutex.
437 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
440 struct ceph_mds_session *s;
442 if (mds >= mdsc->mdsmap->m_max_mds)
443 return ERR_PTR(-EINVAL);
445 s = kzalloc(sizeof(*s), GFP_NOFS);
447 return ERR_PTR(-ENOMEM);
450 s->s_state = CEPH_MDS_SESSION_NEW;
453 mutex_init(&s->s_mutex);
455 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
457 spin_lock_init(&s->s_gen_ttl_lock);
459 s->s_cap_ttl = jiffies - 1;
461 spin_lock_init(&s->s_cap_lock);
462 s->s_renew_requested = 0;
464 INIT_LIST_HEAD(&s->s_caps);
467 atomic_set(&s->s_ref, 1);
468 INIT_LIST_HEAD(&s->s_waiting);
469 INIT_LIST_HEAD(&s->s_unsafe);
470 s->s_num_cap_releases = 0;
471 s->s_cap_reconnect = 0;
472 s->s_cap_iterator = NULL;
473 INIT_LIST_HEAD(&s->s_cap_releases);
474 INIT_LIST_HEAD(&s->s_cap_flushing);
475 INIT_LIST_HEAD(&s->s_cap_snaps_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)
1152 struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1155 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1156 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1157 list_add(&cf->list, &to_remove);
1160 spin_lock(&mdsc->cap_dirty_lock);
1162 list_for_each_entry(cf, &to_remove, list)
1163 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1165 if (!list_empty(&ci->i_dirty_item)) {
1166 pr_warn_ratelimited(
1167 " dropping dirty %s state for %p %lld\n",
1168 ceph_cap_string(ci->i_dirty_caps),
1169 inode, ceph_ino(inode));
1170 ci->i_dirty_caps = 0;
1171 list_del_init(&ci->i_dirty_item);
1174 if (!list_empty(&ci->i_flushing_item)) {
1175 pr_warn_ratelimited(
1176 " dropping dirty+flushing %s state for %p %lld\n",
1177 ceph_cap_string(ci->i_flushing_caps),
1178 inode, ceph_ino(inode));
1179 ci->i_flushing_caps = 0;
1180 list_del_init(&ci->i_flushing_item);
1181 mdsc->num_cap_flushing--;
1184 spin_unlock(&mdsc->cap_dirty_lock);
1186 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1187 list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1188 ci->i_prealloc_cap_flush = NULL;
1191 spin_unlock(&ci->i_ceph_lock);
1192 while (!list_empty(&to_remove)) {
1193 struct ceph_cap_flush *cf;
1194 cf = list_first_entry(&to_remove,
1195 struct ceph_cap_flush, list);
1196 list_del(&cf->list);
1197 ceph_free_cap_flush(cf);
1200 wake_up_all(&ci->i_cap_wq);
1202 ceph_queue_invalidate(inode);
1209 * caller must hold session s_mutex
1211 static void remove_session_caps(struct ceph_mds_session *session)
1213 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1214 struct super_block *sb = fsc->sb;
1215 dout("remove_session_caps on %p\n", session);
1216 iterate_session_caps(session, remove_session_caps_cb, fsc);
1218 spin_lock(&session->s_cap_lock);
1219 if (session->s_nr_caps > 0) {
1220 struct inode *inode;
1221 struct ceph_cap *cap, *prev = NULL;
1222 struct ceph_vino vino;
1224 * iterate_session_caps() skips inodes that are being
1225 * deleted, we need to wait until deletions are complete.
1226 * __wait_on_freeing_inode() is designed for the job,
1227 * but it is not exported, so use lookup inode function
1230 while (!list_empty(&session->s_caps)) {
1231 cap = list_entry(session->s_caps.next,
1232 struct ceph_cap, session_caps);
1236 vino = cap->ci->i_vino;
1237 spin_unlock(&session->s_cap_lock);
1239 inode = ceph_find_inode(sb, vino);
1242 spin_lock(&session->s_cap_lock);
1246 // drop cap expires and unlock s_cap_lock
1247 cleanup_cap_releases(session->s_mdsc, session);
1249 BUG_ON(session->s_nr_caps > 0);
1250 BUG_ON(!list_empty(&session->s_cap_flushing));
1254 * wake up any threads waiting on this session's caps. if the cap is
1255 * old (didn't get renewed on the client reconnect), remove it now.
1257 * caller must hold s_mutex.
1259 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1262 struct ceph_inode_info *ci = ceph_inode(inode);
1265 spin_lock(&ci->i_ceph_lock);
1266 ci->i_wanted_max_size = 0;
1267 ci->i_requested_max_size = 0;
1268 spin_unlock(&ci->i_ceph_lock);
1270 wake_up_all(&ci->i_cap_wq);
1274 static void wake_up_session_caps(struct ceph_mds_session *session,
1277 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1278 iterate_session_caps(session, wake_up_session_cb,
1279 (void *)(unsigned long)reconnect);
1283 * Send periodic message to MDS renewing all currently held caps. The
1284 * ack will reset the expiration for all caps from this session.
1286 * caller holds s_mutex
1288 static int send_renew_caps(struct ceph_mds_client *mdsc,
1289 struct ceph_mds_session *session)
1291 struct ceph_msg *msg;
1294 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1295 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1296 pr_info("mds%d caps stale\n", session->s_mds);
1297 session->s_renew_requested = jiffies;
1299 /* do not try to renew caps until a recovering mds has reconnected
1300 * with its clients. */
1301 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1302 if (state < CEPH_MDS_STATE_RECONNECT) {
1303 dout("send_renew_caps ignoring mds%d (%s)\n",
1304 session->s_mds, ceph_mds_state_name(state));
1308 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1309 ceph_mds_state_name(state));
1310 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1311 ++session->s_renew_seq);
1314 ceph_con_send(&session->s_con, msg);
1318 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1319 struct ceph_mds_session *session, u64 seq)
1321 struct ceph_msg *msg;
1323 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1324 session->s_mds, ceph_session_state_name(session->s_state), seq);
1325 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1328 ceph_con_send(&session->s_con, msg);
1334 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1336 * Called under session->s_mutex
1338 static void renewed_caps(struct ceph_mds_client *mdsc,
1339 struct ceph_mds_session *session, int is_renew)
1344 spin_lock(&session->s_cap_lock);
1345 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1347 session->s_cap_ttl = session->s_renew_requested +
1348 mdsc->mdsmap->m_session_timeout*HZ;
1351 if (time_before(jiffies, session->s_cap_ttl)) {
1352 pr_info("mds%d caps renewed\n", session->s_mds);
1355 pr_info("mds%d caps still stale\n", session->s_mds);
1358 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1359 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1360 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1361 spin_unlock(&session->s_cap_lock);
1364 wake_up_session_caps(session, 0);
1368 * send a session close request
1370 static int request_close_session(struct ceph_mds_client *mdsc,
1371 struct ceph_mds_session *session)
1373 struct ceph_msg *msg;
1375 dout("request_close_session mds%d state %s seq %lld\n",
1376 session->s_mds, ceph_session_state_name(session->s_state),
1378 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1381 ceph_con_send(&session->s_con, msg);
1386 * Called with s_mutex held.
1388 static int __close_session(struct ceph_mds_client *mdsc,
1389 struct ceph_mds_session *session)
1391 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1393 session->s_state = CEPH_MDS_SESSION_CLOSING;
1394 return request_close_session(mdsc, session);
1398 * Trim old(er) caps.
1400 * Because we can't cache an inode without one or more caps, we do
1401 * this indirectly: if a cap is unused, we prune its aliases, at which
1402 * point the inode will hopefully get dropped to.
1404 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1405 * memory pressure from the MDS, though, so it needn't be perfect.
1407 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1409 struct ceph_mds_session *session = arg;
1410 struct ceph_inode_info *ci = ceph_inode(inode);
1411 int used, wanted, oissued, mine;
1413 if (session->s_trim_caps <= 0)
1416 spin_lock(&ci->i_ceph_lock);
1417 mine = cap->issued | cap->implemented;
1418 used = __ceph_caps_used(ci);
1419 wanted = __ceph_caps_file_wanted(ci);
1420 oissued = __ceph_caps_issued_other(ci, cap);
1422 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1423 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1424 ceph_cap_string(used), ceph_cap_string(wanted));
1425 if (cap == ci->i_auth_cap) {
1426 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1427 !list_empty(&ci->i_cap_snaps))
1429 if ((used | wanted) & CEPH_CAP_ANY_WR)
1432 /* The inode has cached pages, but it's no longer used.
1433 * we can safely drop it */
1434 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1435 !(oissued & CEPH_CAP_FILE_CACHE)) {
1439 if ((used | wanted) & ~oissued & mine)
1440 goto out; /* we need these caps */
1442 session->s_trim_caps--;
1444 /* we aren't the only cap.. just remove us */
1445 __ceph_remove_cap(cap, true);
1447 /* try dropping referring dentries */
1448 spin_unlock(&ci->i_ceph_lock);
1449 d_prune_aliases(inode);
1450 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1451 inode, cap, atomic_read(&inode->i_count));
1456 spin_unlock(&ci->i_ceph_lock);
1461 * Trim session cap count down to some max number.
1463 static int trim_caps(struct ceph_mds_client *mdsc,
1464 struct ceph_mds_session *session,
1467 int trim_caps = session->s_nr_caps - max_caps;
1469 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1470 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1471 if (trim_caps > 0) {
1472 session->s_trim_caps = trim_caps;
1473 iterate_session_caps(session, trim_caps_cb, session);
1474 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1475 session->s_mds, session->s_nr_caps, max_caps,
1476 trim_caps - session->s_trim_caps);
1477 session->s_trim_caps = 0;
1480 ceph_send_cap_releases(mdsc, session);
1484 static int check_capsnap_flush(struct ceph_inode_info *ci,
1488 spin_lock(&ci->i_ceph_lock);
1489 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1490 struct ceph_cap_snap *capsnap =
1491 list_first_entry(&ci->i_cap_snaps,
1492 struct ceph_cap_snap, ci_item);
1493 ret = capsnap->follows >= want_snap_seq;
1495 spin_unlock(&ci->i_ceph_lock);
1499 static int check_caps_flush(struct ceph_mds_client *mdsc,
1503 struct ceph_cap_flush *cf;
1506 spin_lock(&mdsc->cap_dirty_lock);
1507 n = rb_first(&mdsc->cap_flush_tree);
1508 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1509 if (cf && cf->tid <= want_flush_tid) {
1510 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1511 cf->tid, want_flush_tid);
1514 spin_unlock(&mdsc->cap_dirty_lock);
1519 * flush all dirty inode data to disk.
1521 * returns true if we've flushed through want_flush_tid
1523 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1524 u64 want_flush_tid, u64 want_snap_seq)
1528 dout("check_caps_flush want %llu snap want %llu\n",
1529 want_flush_tid, want_snap_seq);
1530 mutex_lock(&mdsc->mutex);
1531 for (mds = 0; mds < mdsc->max_sessions; ) {
1532 struct ceph_mds_session *session = mdsc->sessions[mds];
1533 struct inode *inode = NULL;
1539 get_session(session);
1540 mutex_unlock(&mdsc->mutex);
1542 mutex_lock(&session->s_mutex);
1543 if (!list_empty(&session->s_cap_snaps_flushing)) {
1544 struct ceph_cap_snap *capsnap =
1545 list_first_entry(&session->s_cap_snaps_flushing,
1546 struct ceph_cap_snap,
1548 struct ceph_inode_info *ci = capsnap->ci;
1549 if (!check_capsnap_flush(ci, want_snap_seq)) {
1550 dout("check_cap_flush still flushing snap %p "
1551 "follows %lld <= %lld to mds%d\n",
1552 &ci->vfs_inode, capsnap->follows,
1553 want_snap_seq, mds);
1554 inode = igrab(&ci->vfs_inode);
1557 mutex_unlock(&session->s_mutex);
1558 ceph_put_mds_session(session);
1561 wait_event(mdsc->cap_flushing_wq,
1562 check_capsnap_flush(ceph_inode(inode),
1569 mutex_lock(&mdsc->mutex);
1571 mutex_unlock(&mdsc->mutex);
1573 wait_event(mdsc->cap_flushing_wq,
1574 check_caps_flush(mdsc, want_flush_tid));
1576 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1580 * called under s_mutex
1582 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1583 struct ceph_mds_session *session)
1585 struct ceph_msg *msg = NULL;
1586 struct ceph_mds_cap_release *head;
1587 struct ceph_mds_cap_item *item;
1588 struct ceph_cap *cap;
1589 LIST_HEAD(tmp_list);
1590 int num_cap_releases;
1592 spin_lock(&session->s_cap_lock);
1594 list_splice_init(&session->s_cap_releases, &tmp_list);
1595 num_cap_releases = session->s_num_cap_releases;
1596 session->s_num_cap_releases = 0;
1597 spin_unlock(&session->s_cap_lock);
1599 while (!list_empty(&tmp_list)) {
1601 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1602 PAGE_SIZE, GFP_NOFS, false);
1605 head = msg->front.iov_base;
1606 head->num = cpu_to_le32(0);
1607 msg->front.iov_len = sizeof(*head);
1609 cap = list_first_entry(&tmp_list, struct ceph_cap,
1611 list_del(&cap->session_caps);
1614 head = msg->front.iov_base;
1615 le32_add_cpu(&head->num, 1);
1616 item = msg->front.iov_base + msg->front.iov_len;
1617 item->ino = cpu_to_le64(cap->cap_ino);
1618 item->cap_id = cpu_to_le64(cap->cap_id);
1619 item->migrate_seq = cpu_to_le32(cap->mseq);
1620 item->seq = cpu_to_le32(cap->issue_seq);
1621 msg->front.iov_len += sizeof(*item);
1623 ceph_put_cap(mdsc, cap);
1625 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1626 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1627 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1628 ceph_con_send(&session->s_con, msg);
1633 BUG_ON(num_cap_releases != 0);
1635 spin_lock(&session->s_cap_lock);
1636 if (!list_empty(&session->s_cap_releases))
1638 spin_unlock(&session->s_cap_lock);
1641 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1642 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1643 ceph_con_send(&session->s_con, msg);
1647 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1649 spin_lock(&session->s_cap_lock);
1650 list_splice(&tmp_list, &session->s_cap_releases);
1651 session->s_num_cap_releases += num_cap_releases;
1652 spin_unlock(&session->s_cap_lock);
1659 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1662 struct ceph_inode_info *ci = ceph_inode(dir);
1663 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1664 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1665 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1666 int order, num_entries;
1668 spin_lock(&ci->i_ceph_lock);
1669 num_entries = ci->i_files + ci->i_subdirs;
1670 spin_unlock(&ci->i_ceph_lock);
1671 num_entries = max(num_entries, 1);
1672 num_entries = min(num_entries, opt->max_readdir);
1674 order = get_order(size * num_entries);
1675 while (order >= 0) {
1676 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1679 if (rinfo->dir_entries)
1683 if (!rinfo->dir_entries)
1686 num_entries = (PAGE_SIZE << order) / size;
1687 num_entries = min(num_entries, opt->max_readdir);
1689 rinfo->dir_buf_size = PAGE_SIZE << order;
1690 req->r_num_caps = num_entries + 1;
1691 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1692 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1697 * Create an mds request.
1699 struct ceph_mds_request *
1700 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1702 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1705 return ERR_PTR(-ENOMEM);
1707 mutex_init(&req->r_fill_mutex);
1709 req->r_started = jiffies;
1710 req->r_resend_mds = -1;
1711 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1712 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1714 kref_init(&req->r_kref);
1715 RB_CLEAR_NODE(&req->r_node);
1716 INIT_LIST_HEAD(&req->r_wait);
1717 init_completion(&req->r_completion);
1718 init_completion(&req->r_safe_completion);
1719 INIT_LIST_HEAD(&req->r_unsafe_item);
1721 req->r_stamp = current_fs_time(mdsc->fsc->sb);
1724 req->r_direct_mode = mode;
1729 * return oldest (lowest) request, tid in request tree, 0 if none.
1731 * called under mdsc->mutex.
1733 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1735 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1737 return rb_entry(rb_first(&mdsc->request_tree),
1738 struct ceph_mds_request, r_node);
1741 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1743 return mdsc->oldest_tid;
1747 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1748 * on build_path_from_dentry in fs/cifs/dir.c.
1750 * If @stop_on_nosnap, generate path relative to the first non-snapped
1753 * Encode hidden .snap dirs as a double /, i.e.
1754 * foo/.snap/bar -> foo//bar
1756 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1759 struct dentry *temp;
1765 return ERR_PTR(-EINVAL);
1769 seq = read_seqbegin(&rename_lock);
1771 for (temp = dentry; !IS_ROOT(temp);) {
1772 struct inode *inode = d_inode(temp);
1773 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1774 len++; /* slash only */
1775 else if (stop_on_nosnap && inode &&
1776 ceph_snap(inode) == CEPH_NOSNAP)
1779 len += 1 + temp->d_name.len;
1780 temp = temp->d_parent;
1784 len--; /* no leading '/' */
1786 path = kmalloc(len+1, GFP_NOFS);
1788 return ERR_PTR(-ENOMEM);
1790 path[pos] = 0; /* trailing null */
1792 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1793 struct inode *inode;
1795 spin_lock(&temp->d_lock);
1796 inode = d_inode(temp);
1797 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1798 dout("build_path path+%d: %p SNAPDIR\n",
1800 } else if (stop_on_nosnap && inode &&
1801 ceph_snap(inode) == CEPH_NOSNAP) {
1802 spin_unlock(&temp->d_lock);
1805 pos -= temp->d_name.len;
1807 spin_unlock(&temp->d_lock);
1810 strncpy(path + pos, temp->d_name.name,
1813 spin_unlock(&temp->d_lock);
1816 temp = temp->d_parent;
1819 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1820 pr_err("build_path did not end path lookup where "
1821 "expected, namelen is %d, pos is %d\n", len, pos);
1822 /* presumably this is only possible if racing with a
1823 rename of one of the parent directories (we can not
1824 lock the dentries above us to prevent this, but
1825 retrying should be harmless) */
1830 *base = ceph_ino(d_inode(temp));
1832 dout("build_path on %p %d built %llx '%.*s'\n",
1833 dentry, d_count(dentry), *base, len, path);
1837 static int build_dentry_path(struct dentry *dentry,
1838 const char **ppath, int *ppathlen, u64 *pino,
1843 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1844 *pino = ceph_ino(d_inode(dentry->d_parent));
1845 *ppath = dentry->d_name.name;
1846 *ppathlen = dentry->d_name.len;
1849 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1851 return PTR_ERR(path);
1857 static int build_inode_path(struct inode *inode,
1858 const char **ppath, int *ppathlen, u64 *pino,
1861 struct dentry *dentry;
1864 if (ceph_snap(inode) == CEPH_NOSNAP) {
1865 *pino = ceph_ino(inode);
1869 dentry = d_find_alias(inode);
1870 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1873 return PTR_ERR(path);
1880 * request arguments may be specified via an inode *, a dentry *, or
1881 * an explicit ino+path.
1883 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1884 const char *rpath, u64 rino,
1885 const char **ppath, int *pathlen,
1886 u64 *ino, int *freepath)
1891 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1892 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1894 } else if (rdentry) {
1895 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1896 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1898 } else if (rpath || rino) {
1901 *pathlen = rpath ? strlen(rpath) : 0;
1902 dout(" path %.*s\n", *pathlen, rpath);
1909 * called under mdsc->mutex
1911 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1912 struct ceph_mds_request *req,
1913 int mds, bool drop_cap_releases)
1915 struct ceph_msg *msg;
1916 struct ceph_mds_request_head *head;
1917 const char *path1 = NULL;
1918 const char *path2 = NULL;
1919 u64 ino1 = 0, ino2 = 0;
1920 int pathlen1 = 0, pathlen2 = 0;
1921 int freepath1 = 0, freepath2 = 0;
1927 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1928 req->r_path1, req->r_ino1.ino,
1929 &path1, &pathlen1, &ino1, &freepath1);
1935 ret = set_request_path_attr(NULL, req->r_old_dentry,
1936 req->r_path2, req->r_ino2.ino,
1937 &path2, &pathlen2, &ino2, &freepath2);
1943 len = sizeof(*head) +
1944 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1945 sizeof(struct ceph_timespec);
1947 /* calculate (max) length for cap releases */
1948 len += sizeof(struct ceph_mds_request_release) *
1949 (!!req->r_inode_drop + !!req->r_dentry_drop +
1950 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1951 if (req->r_dentry_drop)
1952 len += req->r_dentry->d_name.len;
1953 if (req->r_old_dentry_drop)
1954 len += req->r_old_dentry->d_name.len;
1956 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1958 msg = ERR_PTR(-ENOMEM);
1962 msg->hdr.version = cpu_to_le16(2);
1963 msg->hdr.tid = cpu_to_le64(req->r_tid);
1965 head = msg->front.iov_base;
1966 p = msg->front.iov_base + sizeof(*head);
1967 end = msg->front.iov_base + msg->front.iov_len;
1969 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1970 head->op = cpu_to_le32(req->r_op);
1971 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1972 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1973 head->args = req->r_args;
1975 ceph_encode_filepath(&p, end, ino1, path1);
1976 ceph_encode_filepath(&p, end, ino2, path2);
1978 /* make note of release offset, in case we need to replay */
1979 req->r_request_release_offset = p - msg->front.iov_base;
1983 if (req->r_inode_drop)
1984 releases += ceph_encode_inode_release(&p,
1985 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1986 mds, req->r_inode_drop, req->r_inode_unless, 0);
1987 if (req->r_dentry_drop)
1988 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1989 mds, req->r_dentry_drop, req->r_dentry_unless);
1990 if (req->r_old_dentry_drop)
1991 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1992 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1993 if (req->r_old_inode_drop)
1994 releases += ceph_encode_inode_release(&p,
1995 d_inode(req->r_old_dentry),
1996 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1998 if (drop_cap_releases) {
2000 p = msg->front.iov_base + req->r_request_release_offset;
2003 head->num_releases = cpu_to_le16(releases);
2007 struct ceph_timespec ts;
2008 ceph_encode_timespec(&ts, &req->r_stamp);
2009 ceph_encode_copy(&p, &ts, sizeof(ts));
2013 msg->front.iov_len = p - msg->front.iov_base;
2014 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2016 if (req->r_pagelist) {
2017 struct ceph_pagelist *pagelist = req->r_pagelist;
2018 atomic_inc(&pagelist->refcnt);
2019 ceph_msg_data_add_pagelist(msg, pagelist);
2020 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2022 msg->hdr.data_len = 0;
2025 msg->hdr.data_off = cpu_to_le16(0);
2029 kfree((char *)path2);
2032 kfree((char *)path1);
2038 * called under mdsc->mutex if error, under no mutex if
2041 static void complete_request(struct ceph_mds_client *mdsc,
2042 struct ceph_mds_request *req)
2044 if (req->r_callback)
2045 req->r_callback(mdsc, req);
2047 complete_all(&req->r_completion);
2051 * called under mdsc->mutex
2053 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2054 struct ceph_mds_request *req,
2055 int mds, bool drop_cap_releases)
2057 struct ceph_mds_request_head *rhead;
2058 struct ceph_msg *msg;
2063 struct ceph_cap *cap =
2064 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2067 req->r_sent_on_mseq = cap->mseq;
2069 req->r_sent_on_mseq = -1;
2071 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2072 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2074 if (req->r_got_unsafe) {
2077 * Replay. Do not regenerate message (and rebuild
2078 * paths, etc.); just use the original message.
2079 * Rebuilding paths will break for renames because
2080 * d_move mangles the src name.
2082 msg = req->r_request;
2083 rhead = msg->front.iov_base;
2085 flags = le32_to_cpu(rhead->flags);
2086 flags |= CEPH_MDS_FLAG_REPLAY;
2087 rhead->flags = cpu_to_le32(flags);
2089 if (req->r_target_inode)
2090 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2092 rhead->num_retry = req->r_attempts - 1;
2094 /* remove cap/dentry releases from message */
2095 rhead->num_releases = 0;
2098 p = msg->front.iov_base + req->r_request_release_offset;
2100 struct ceph_timespec ts;
2101 ceph_encode_timespec(&ts, &req->r_stamp);
2102 ceph_encode_copy(&p, &ts, sizeof(ts));
2105 msg->front.iov_len = p - msg->front.iov_base;
2106 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2110 if (req->r_request) {
2111 ceph_msg_put(req->r_request);
2112 req->r_request = NULL;
2114 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2116 req->r_err = PTR_ERR(msg);
2117 return PTR_ERR(msg);
2119 req->r_request = msg;
2121 rhead = msg->front.iov_base;
2122 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2123 if (req->r_got_unsafe)
2124 flags |= CEPH_MDS_FLAG_REPLAY;
2125 if (req->r_locked_dir)
2126 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2127 rhead->flags = cpu_to_le32(flags);
2128 rhead->num_fwd = req->r_num_fwd;
2129 rhead->num_retry = req->r_attempts - 1;
2132 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2137 * send request, or put it on the appropriate wait list.
2139 static int __do_request(struct ceph_mds_client *mdsc,
2140 struct ceph_mds_request *req)
2142 struct ceph_mds_session *session = NULL;
2146 if (req->r_err || req->r_got_result) {
2148 __unregister_request(mdsc, req);
2152 if (req->r_timeout &&
2153 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2154 dout("do_request timed out\n");
2158 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2159 dout("do_request forced umount\n");
2164 put_request_session(req);
2166 mds = __choose_mds(mdsc, req);
2168 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2169 dout("do_request no mds or not active, waiting for map\n");
2170 list_add(&req->r_wait, &mdsc->waiting_for_map);
2174 /* get, open session */
2175 session = __ceph_lookup_mds_session(mdsc, mds);
2177 session = register_session(mdsc, mds);
2178 if (IS_ERR(session)) {
2179 err = PTR_ERR(session);
2183 req->r_session = get_session(session);
2185 dout("do_request mds%d session %p state %s\n", mds, session,
2186 ceph_session_state_name(session->s_state));
2187 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2188 session->s_state != CEPH_MDS_SESSION_HUNG) {
2189 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2190 session->s_state == CEPH_MDS_SESSION_CLOSING)
2191 __open_session(mdsc, session);
2192 list_add(&req->r_wait, &session->s_waiting);
2197 req->r_resend_mds = -1; /* forget any previous mds hint */
2199 if (req->r_request_started == 0) /* note request start time */
2200 req->r_request_started = jiffies;
2202 err = __prepare_send_request(mdsc, req, mds, false);
2204 ceph_msg_get(req->r_request);
2205 ceph_con_send(&session->s_con, req->r_request);
2209 ceph_put_mds_session(session);
2212 dout("__do_request early error %d\n", err);
2214 complete_request(mdsc, req);
2215 __unregister_request(mdsc, req);
2222 * called under mdsc->mutex
2224 static void __wake_requests(struct ceph_mds_client *mdsc,
2225 struct list_head *head)
2227 struct ceph_mds_request *req;
2228 LIST_HEAD(tmp_list);
2230 list_splice_init(head, &tmp_list);
2232 while (!list_empty(&tmp_list)) {
2233 req = list_entry(tmp_list.next,
2234 struct ceph_mds_request, r_wait);
2235 list_del_init(&req->r_wait);
2236 dout(" wake request %p tid %llu\n", req, req->r_tid);
2237 __do_request(mdsc, req);
2242 * Wake up threads with requests pending for @mds, so that they can
2243 * resubmit their requests to a possibly different mds.
2245 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2247 struct ceph_mds_request *req;
2248 struct rb_node *p = rb_first(&mdsc->request_tree);
2250 dout("kick_requests mds%d\n", mds);
2252 req = rb_entry(p, struct ceph_mds_request, r_node);
2254 if (req->r_got_unsafe)
2256 if (req->r_attempts > 0)
2257 continue; /* only new requests */
2258 if (req->r_session &&
2259 req->r_session->s_mds == mds) {
2260 dout(" kicking tid %llu\n", req->r_tid);
2261 list_del_init(&req->r_wait);
2262 __do_request(mdsc, req);
2267 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2268 struct ceph_mds_request *req)
2270 dout("submit_request on %p\n", req);
2271 mutex_lock(&mdsc->mutex);
2272 __register_request(mdsc, req, NULL);
2273 __do_request(mdsc, req);
2274 mutex_unlock(&mdsc->mutex);
2278 * Synchrously perform an mds request. Take care of all of the
2279 * session setup, forwarding, retry details.
2281 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2283 struct ceph_mds_request *req)
2287 dout("do_request on %p\n", req);
2289 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2291 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2292 if (req->r_locked_dir)
2293 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2294 if (req->r_old_dentry_dir)
2295 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2299 mutex_lock(&mdsc->mutex);
2300 __register_request(mdsc, req, dir);
2301 __do_request(mdsc, req);
2309 mutex_unlock(&mdsc->mutex);
2310 dout("do_request waiting\n");
2311 if (!req->r_timeout && req->r_wait_for_completion) {
2312 err = req->r_wait_for_completion(mdsc, req);
2314 long timeleft = wait_for_completion_killable_timeout(
2316 ceph_timeout_jiffies(req->r_timeout));
2320 err = -EIO; /* timed out */
2322 err = timeleft; /* killed */
2324 dout("do_request waited, got %d\n", err);
2325 mutex_lock(&mdsc->mutex);
2327 /* only abort if we didn't race with a real reply */
2328 if (req->r_got_result) {
2329 err = le32_to_cpu(req->r_reply_info.head->result);
2330 } else if (err < 0) {
2331 dout("aborted request %lld with %d\n", req->r_tid, err);
2334 * ensure we aren't running concurrently with
2335 * ceph_fill_trace or ceph_readdir_prepopulate, which
2336 * rely on locks (dir mutex) held by our caller.
2338 mutex_lock(&req->r_fill_mutex);
2340 req->r_aborted = true;
2341 mutex_unlock(&req->r_fill_mutex);
2343 if (req->r_locked_dir &&
2344 (req->r_op & CEPH_MDS_OP_WRITE))
2345 ceph_invalidate_dir_request(req);
2351 mutex_unlock(&mdsc->mutex);
2352 dout("do_request %p done, result %d\n", req, err);
2357 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2358 * namespace request.
2360 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2362 struct inode *inode = req->r_locked_dir;
2364 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2366 ceph_dir_clear_complete(inode);
2368 ceph_invalidate_dentry_lease(req->r_dentry);
2369 if (req->r_old_dentry)
2370 ceph_invalidate_dentry_lease(req->r_old_dentry);
2376 * We take the session mutex and parse and process the reply immediately.
2377 * This preserves the logical ordering of replies, capabilities, etc., sent
2378 * by the MDS as they are applied to our local cache.
2380 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2382 struct ceph_mds_client *mdsc = session->s_mdsc;
2383 struct ceph_mds_request *req;
2384 struct ceph_mds_reply_head *head = msg->front.iov_base;
2385 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2386 struct ceph_snap_realm *realm;
2389 int mds = session->s_mds;
2391 if (msg->front.iov_len < sizeof(*head)) {
2392 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2397 /* get request, session */
2398 tid = le64_to_cpu(msg->hdr.tid);
2399 mutex_lock(&mdsc->mutex);
2400 req = lookup_get_request(mdsc, tid);
2402 dout("handle_reply on unknown tid %llu\n", tid);
2403 mutex_unlock(&mdsc->mutex);
2406 dout("handle_reply %p\n", req);
2408 /* correct session? */
2409 if (req->r_session != session) {
2410 pr_err("mdsc_handle_reply got %llu on session mds%d"
2411 " not mds%d\n", tid, session->s_mds,
2412 req->r_session ? req->r_session->s_mds : -1);
2413 mutex_unlock(&mdsc->mutex);
2418 if ((req->r_got_unsafe && !head->safe) ||
2419 (req->r_got_safe && head->safe)) {
2420 pr_warn("got a dup %s reply on %llu from mds%d\n",
2421 head->safe ? "safe" : "unsafe", tid, mds);
2422 mutex_unlock(&mdsc->mutex);
2425 if (req->r_got_safe) {
2426 pr_warn("got unsafe after safe on %llu from mds%d\n",
2428 mutex_unlock(&mdsc->mutex);
2432 result = le32_to_cpu(head->result);
2436 * if we're not talking to the authority, send to them
2437 * if the authority has changed while we weren't looking,
2438 * send to new authority
2439 * Otherwise we just have to return an ESTALE
2441 if (result == -ESTALE) {
2442 dout("got ESTALE on request %llu", req->r_tid);
2443 req->r_resend_mds = -1;
2444 if (req->r_direct_mode != USE_AUTH_MDS) {
2445 dout("not using auth, setting for that now");
2446 req->r_direct_mode = USE_AUTH_MDS;
2447 __do_request(mdsc, req);
2448 mutex_unlock(&mdsc->mutex);
2451 int mds = __choose_mds(mdsc, req);
2452 if (mds >= 0 && mds != req->r_session->s_mds) {
2453 dout("but auth changed, so resending");
2454 __do_request(mdsc, req);
2455 mutex_unlock(&mdsc->mutex);
2459 dout("have to return ESTALE on request %llu", req->r_tid);
2464 req->r_got_safe = true;
2465 __unregister_request(mdsc, req);
2467 if (req->r_got_unsafe) {
2469 * We already handled the unsafe response, now do the
2470 * cleanup. No need to examine the response; the MDS
2471 * doesn't include any result info in the safe
2472 * response. And even if it did, there is nothing
2473 * useful we could do with a revised return value.
2475 dout("got safe reply %llu, mds%d\n", tid, mds);
2476 list_del_init(&req->r_unsafe_item);
2478 /* last unsafe request during umount? */
2479 if (mdsc->stopping && !__get_oldest_req(mdsc))
2480 complete_all(&mdsc->safe_umount_waiters);
2481 mutex_unlock(&mdsc->mutex);
2485 req->r_got_unsafe = true;
2486 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2487 if (req->r_unsafe_dir) {
2488 struct ceph_inode_info *ci =
2489 ceph_inode(req->r_unsafe_dir);
2490 spin_lock(&ci->i_unsafe_lock);
2491 list_add_tail(&req->r_unsafe_dir_item,
2492 &ci->i_unsafe_dirops);
2493 spin_unlock(&ci->i_unsafe_lock);
2497 dout("handle_reply tid %lld result %d\n", tid, result);
2498 rinfo = &req->r_reply_info;
2499 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2500 mutex_unlock(&mdsc->mutex);
2502 mutex_lock(&session->s_mutex);
2504 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2511 if (rinfo->snapblob_len) {
2512 down_write(&mdsc->snap_rwsem);
2513 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2514 rinfo->snapblob + rinfo->snapblob_len,
2515 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2517 downgrade_write(&mdsc->snap_rwsem);
2519 down_read(&mdsc->snap_rwsem);
2522 /* insert trace into our cache */
2523 mutex_lock(&req->r_fill_mutex);
2524 current->journal_info = req;
2525 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2527 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2528 req->r_op == CEPH_MDS_OP_LSSNAP))
2529 ceph_readdir_prepopulate(req, req->r_session);
2530 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2532 current->journal_info = NULL;
2533 mutex_unlock(&req->r_fill_mutex);
2535 up_read(&mdsc->snap_rwsem);
2537 ceph_put_snap_realm(mdsc, realm);
2539 if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2540 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2541 spin_lock(&ci->i_unsafe_lock);
2542 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2543 spin_unlock(&ci->i_unsafe_lock);
2546 mutex_lock(&mdsc->mutex);
2547 if (!req->r_aborted) {
2551 req->r_reply = ceph_msg_get(msg);
2552 req->r_got_result = true;
2555 dout("reply arrived after request %lld was aborted\n", tid);
2557 mutex_unlock(&mdsc->mutex);
2559 mutex_unlock(&session->s_mutex);
2561 /* kick calling process */
2562 complete_request(mdsc, req);
2564 ceph_mdsc_put_request(req);
2571 * handle mds notification that our request has been forwarded.
2573 static void handle_forward(struct ceph_mds_client *mdsc,
2574 struct ceph_mds_session *session,
2575 struct ceph_msg *msg)
2577 struct ceph_mds_request *req;
2578 u64 tid = le64_to_cpu(msg->hdr.tid);
2582 void *p = msg->front.iov_base;
2583 void *end = p + msg->front.iov_len;
2585 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2586 next_mds = ceph_decode_32(&p);
2587 fwd_seq = ceph_decode_32(&p);
2589 mutex_lock(&mdsc->mutex);
2590 req = lookup_get_request(mdsc, tid);
2592 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2593 goto out; /* dup reply? */
2596 if (req->r_aborted) {
2597 dout("forward tid %llu aborted, unregistering\n", tid);
2598 __unregister_request(mdsc, req);
2599 } else if (fwd_seq <= req->r_num_fwd) {
2600 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2601 tid, next_mds, req->r_num_fwd, fwd_seq);
2603 /* resend. forward race not possible; mds would drop */
2604 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2606 BUG_ON(req->r_got_result);
2607 req->r_attempts = 0;
2608 req->r_num_fwd = fwd_seq;
2609 req->r_resend_mds = next_mds;
2610 put_request_session(req);
2611 __do_request(mdsc, req);
2613 ceph_mdsc_put_request(req);
2615 mutex_unlock(&mdsc->mutex);
2619 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2623 * handle a mds session control message
2625 static void handle_session(struct ceph_mds_session *session,
2626 struct ceph_msg *msg)
2628 struct ceph_mds_client *mdsc = session->s_mdsc;
2631 int mds = session->s_mds;
2632 struct ceph_mds_session_head *h = msg->front.iov_base;
2636 if (msg->front.iov_len != sizeof(*h))
2638 op = le32_to_cpu(h->op);
2639 seq = le64_to_cpu(h->seq);
2641 mutex_lock(&mdsc->mutex);
2642 if (op == CEPH_SESSION_CLOSE)
2643 __unregister_session(mdsc, session);
2644 /* FIXME: this ttl calculation is generous */
2645 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2646 mutex_unlock(&mdsc->mutex);
2648 mutex_lock(&session->s_mutex);
2650 dout("handle_session mds%d %s %p state %s seq %llu\n",
2651 mds, ceph_session_op_name(op), session,
2652 ceph_session_state_name(session->s_state), seq);
2654 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2655 session->s_state = CEPH_MDS_SESSION_OPEN;
2656 pr_info("mds%d came back\n", session->s_mds);
2660 case CEPH_SESSION_OPEN:
2661 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2662 pr_info("mds%d reconnect success\n", session->s_mds);
2663 session->s_state = CEPH_MDS_SESSION_OPEN;
2664 renewed_caps(mdsc, session, 0);
2667 __close_session(mdsc, session);
2670 case CEPH_SESSION_RENEWCAPS:
2671 if (session->s_renew_seq == seq)
2672 renewed_caps(mdsc, session, 1);
2675 case CEPH_SESSION_CLOSE:
2676 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2677 pr_info("mds%d reconnect denied\n", session->s_mds);
2678 cleanup_session_requests(mdsc, session);
2679 remove_session_caps(session);
2680 wake = 2; /* for good measure */
2681 wake_up_all(&mdsc->session_close_wq);
2684 case CEPH_SESSION_STALE:
2685 pr_info("mds%d caps went stale, renewing\n",
2687 spin_lock(&session->s_gen_ttl_lock);
2688 session->s_cap_gen++;
2689 session->s_cap_ttl = jiffies - 1;
2690 spin_unlock(&session->s_gen_ttl_lock);
2691 send_renew_caps(mdsc, session);
2694 case CEPH_SESSION_RECALL_STATE:
2695 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2698 case CEPH_SESSION_FLUSHMSG:
2699 send_flushmsg_ack(mdsc, session, seq);
2702 case CEPH_SESSION_FORCE_RO:
2703 dout("force_session_readonly %p\n", session);
2704 spin_lock(&session->s_cap_lock);
2705 session->s_readonly = true;
2706 spin_unlock(&session->s_cap_lock);
2707 wake_up_session_caps(session, 0);
2711 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2715 mutex_unlock(&session->s_mutex);
2717 mutex_lock(&mdsc->mutex);
2718 __wake_requests(mdsc, &session->s_waiting);
2720 kick_requests(mdsc, mds);
2721 mutex_unlock(&mdsc->mutex);
2726 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2727 (int)msg->front.iov_len);
2734 * called under session->mutex.
2736 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2737 struct ceph_mds_session *session)
2739 struct ceph_mds_request *req, *nreq;
2743 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2745 mutex_lock(&mdsc->mutex);
2746 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2747 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2749 ceph_msg_get(req->r_request);
2750 ceph_con_send(&session->s_con, req->r_request);
2755 * also re-send old requests when MDS enters reconnect stage. So that MDS
2756 * can process completed request in clientreplay stage.
2758 p = rb_first(&mdsc->request_tree);
2760 req = rb_entry(p, struct ceph_mds_request, r_node);
2762 if (req->r_got_unsafe)
2764 if (req->r_attempts == 0)
2765 continue; /* only old requests */
2766 if (req->r_session &&
2767 req->r_session->s_mds == session->s_mds) {
2768 err = __prepare_send_request(mdsc, req,
2769 session->s_mds, true);
2771 ceph_msg_get(req->r_request);
2772 ceph_con_send(&session->s_con, req->r_request);
2776 mutex_unlock(&mdsc->mutex);
2780 * Encode information about a cap for a reconnect with the MDS.
2782 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2786 struct ceph_mds_cap_reconnect v2;
2787 struct ceph_mds_cap_reconnect_v1 v1;
2790 struct ceph_inode_info *ci;
2791 struct ceph_reconnect_state *recon_state = arg;
2792 struct ceph_pagelist *pagelist = recon_state->pagelist;
2796 struct dentry *dentry;
2800 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2801 inode, ceph_vinop(inode), cap, cap->cap_id,
2802 ceph_cap_string(cap->issued));
2803 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2807 dentry = d_find_alias(inode);
2809 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2811 err = PTR_ERR(path);
2818 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2822 spin_lock(&ci->i_ceph_lock);
2823 cap->seq = 0; /* reset cap seq */
2824 cap->issue_seq = 0; /* and issue_seq */
2825 cap->mseq = 0; /* and migrate_seq */
2826 cap->cap_gen = cap->session->s_cap_gen;
2828 if (recon_state->flock) {
2829 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2830 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2831 rec.v2.issued = cpu_to_le32(cap->issued);
2832 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2833 rec.v2.pathbase = cpu_to_le64(pathbase);
2834 rec.v2.flock_len = 0;
2835 reclen = sizeof(rec.v2);
2837 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2838 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2839 rec.v1.issued = cpu_to_le32(cap->issued);
2840 rec.v1.size = cpu_to_le64(inode->i_size);
2841 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2842 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2843 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2844 rec.v1.pathbase = cpu_to_le64(pathbase);
2845 reclen = sizeof(rec.v1);
2847 spin_unlock(&ci->i_ceph_lock);
2849 if (recon_state->flock) {
2850 int num_fcntl_locks, num_flock_locks;
2851 struct ceph_filelock *flocks;
2854 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2855 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2856 sizeof(struct ceph_filelock), GFP_NOFS);
2861 err = ceph_encode_locks_to_buffer(inode, flocks,
2871 * number of encoded locks is stable, so copy to pagelist
2873 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2874 (num_fcntl_locks+num_flock_locks) *
2875 sizeof(struct ceph_filelock));
2876 err = ceph_pagelist_append(pagelist, &rec, reclen);
2878 err = ceph_locks_to_pagelist(flocks, pagelist,
2883 err = ceph_pagelist_append(pagelist, &rec, reclen);
2886 recon_state->nr_caps++;
2896 * If an MDS fails and recovers, clients need to reconnect in order to
2897 * reestablish shared state. This includes all caps issued through
2898 * this session _and_ the snap_realm hierarchy. Because it's not
2899 * clear which snap realms the mds cares about, we send everything we
2900 * know about.. that ensures we'll then get any new info the
2901 * recovering MDS might have.
2903 * This is a relatively heavyweight operation, but it's rare.
2905 * called with mdsc->mutex held.
2907 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2908 struct ceph_mds_session *session)
2910 struct ceph_msg *reply;
2912 int mds = session->s_mds;
2915 struct ceph_pagelist *pagelist;
2916 struct ceph_reconnect_state recon_state;
2918 pr_info("mds%d reconnect start\n", mds);
2920 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2922 goto fail_nopagelist;
2923 ceph_pagelist_init(pagelist);
2925 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2929 mutex_lock(&session->s_mutex);
2930 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2933 dout("session %p state %s\n", session,
2934 ceph_session_state_name(session->s_state));
2936 spin_lock(&session->s_gen_ttl_lock);
2937 session->s_cap_gen++;
2938 spin_unlock(&session->s_gen_ttl_lock);
2940 spin_lock(&session->s_cap_lock);
2941 /* don't know if session is readonly */
2942 session->s_readonly = 0;
2944 * notify __ceph_remove_cap() that we are composing cap reconnect.
2945 * If a cap get released before being added to the cap reconnect,
2946 * __ceph_remove_cap() should skip queuing cap release.
2948 session->s_cap_reconnect = 1;
2949 /* drop old cap expires; we're about to reestablish that state */
2950 cleanup_cap_releases(mdsc, session);
2952 /* trim unused caps to reduce MDS's cache rejoin time */
2953 if (mdsc->fsc->sb->s_root)
2954 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2956 ceph_con_close(&session->s_con);
2957 ceph_con_open(&session->s_con,
2958 CEPH_ENTITY_TYPE_MDS, mds,
2959 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2961 /* replay unsafe requests */
2962 replay_unsafe_requests(mdsc, session);
2964 down_read(&mdsc->snap_rwsem);
2966 /* traverse this session's caps */
2967 s_nr_caps = session->s_nr_caps;
2968 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2972 recon_state.nr_caps = 0;
2973 recon_state.pagelist = pagelist;
2974 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2975 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2979 spin_lock(&session->s_cap_lock);
2980 session->s_cap_reconnect = 0;
2981 spin_unlock(&session->s_cap_lock);
2984 * snaprealms. we provide mds with the ino, seq (version), and
2985 * parent for all of our realms. If the mds has any newer info,
2988 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2989 struct ceph_snap_realm *realm =
2990 rb_entry(p, struct ceph_snap_realm, node);
2991 struct ceph_mds_snaprealm_reconnect sr_rec;
2993 dout(" adding snap realm %llx seq %lld parent %llx\n",
2994 realm->ino, realm->seq, realm->parent_ino);
2995 sr_rec.ino = cpu_to_le64(realm->ino);
2996 sr_rec.seq = cpu_to_le64(realm->seq);
2997 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2998 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3003 if (recon_state.flock)
3004 reply->hdr.version = cpu_to_le16(2);
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(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 * Preemptively release a lease we expect to invalidate anyway.
3296 * Pass @inode always, @dentry is optional.
3298 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3299 struct dentry *dentry)
3301 struct ceph_dentry_info *di;
3302 struct ceph_mds_session *session;
3305 BUG_ON(inode == NULL);
3306 BUG_ON(dentry == NULL);
3308 /* is dentry lease valid? */
3309 spin_lock(&dentry->d_lock);
3310 di = ceph_dentry(dentry);
3311 if (!di || !di->lease_session ||
3312 di->lease_session->s_mds < 0 ||
3313 di->lease_gen != di->lease_session->s_cap_gen ||
3314 !time_before(jiffies, di->time)) {
3315 dout("lease_release inode %p dentry %p -- "
3318 spin_unlock(&dentry->d_lock);
3322 /* we do have a lease on this dentry; note mds and seq */
3323 session = ceph_get_mds_session(di->lease_session);
3324 seq = di->lease_seq;
3325 __ceph_mdsc_drop_dentry_lease(dentry);
3326 spin_unlock(&dentry->d_lock);
3328 dout("lease_release inode %p dentry %p to mds%d\n",
3329 inode, dentry, session->s_mds);
3330 ceph_mdsc_lease_send_msg(session, inode, dentry,
3331 CEPH_MDS_LEASE_RELEASE, seq);
3332 ceph_put_mds_session(session);
3336 * drop all leases (and dentry refs) in preparation for umount
3338 static void drop_leases(struct ceph_mds_client *mdsc)
3342 dout("drop_leases\n");
3343 mutex_lock(&mdsc->mutex);
3344 for (i = 0; i < mdsc->max_sessions; i++) {
3345 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3348 mutex_unlock(&mdsc->mutex);
3349 mutex_lock(&s->s_mutex);
3350 mutex_unlock(&s->s_mutex);
3351 ceph_put_mds_session(s);
3352 mutex_lock(&mdsc->mutex);
3354 mutex_unlock(&mdsc->mutex);
3360 * delayed work -- periodically trim expired leases, renew caps with mds
3362 static void schedule_delayed(struct ceph_mds_client *mdsc)
3365 unsigned hz = round_jiffies_relative(HZ * delay);
3366 schedule_delayed_work(&mdsc->delayed_work, hz);
3369 static void delayed_work(struct work_struct *work)
3372 struct ceph_mds_client *mdsc =
3373 container_of(work, struct ceph_mds_client, delayed_work.work);
3377 dout("mdsc delayed_work\n");
3378 ceph_check_delayed_caps(mdsc);
3380 mutex_lock(&mdsc->mutex);
3381 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3382 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3383 mdsc->last_renew_caps);
3385 mdsc->last_renew_caps = jiffies;
3387 for (i = 0; i < mdsc->max_sessions; i++) {
3388 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3391 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3392 dout("resending session close request for mds%d\n",
3394 request_close_session(mdsc, s);
3395 ceph_put_mds_session(s);
3398 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3399 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3400 s->s_state = CEPH_MDS_SESSION_HUNG;
3401 pr_info("mds%d hung\n", s->s_mds);
3404 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3405 /* this mds is failed or recovering, just wait */
3406 ceph_put_mds_session(s);
3409 mutex_unlock(&mdsc->mutex);
3411 mutex_lock(&s->s_mutex);
3413 send_renew_caps(mdsc, s);
3415 ceph_con_keepalive(&s->s_con);
3416 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3417 s->s_state == CEPH_MDS_SESSION_HUNG)
3418 ceph_send_cap_releases(mdsc, s);
3419 mutex_unlock(&s->s_mutex);
3420 ceph_put_mds_session(s);
3422 mutex_lock(&mdsc->mutex);
3424 mutex_unlock(&mdsc->mutex);
3426 schedule_delayed(mdsc);
3429 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3432 struct ceph_mds_client *mdsc;
3434 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3439 mutex_init(&mdsc->mutex);
3440 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3441 if (mdsc->mdsmap == NULL) {
3446 init_completion(&mdsc->safe_umount_waiters);
3447 init_waitqueue_head(&mdsc->session_close_wq);
3448 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3449 mdsc->sessions = NULL;
3450 atomic_set(&mdsc->num_sessions, 0);
3451 mdsc->max_sessions = 0;
3453 mdsc->last_snap_seq = 0;
3454 init_rwsem(&mdsc->snap_rwsem);
3455 mdsc->snap_realms = RB_ROOT;
3456 INIT_LIST_HEAD(&mdsc->snap_empty);
3457 spin_lock_init(&mdsc->snap_empty_lock);
3459 mdsc->oldest_tid = 0;
3460 mdsc->request_tree = RB_ROOT;
3461 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3462 mdsc->last_renew_caps = jiffies;
3463 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3464 spin_lock_init(&mdsc->cap_delay_lock);
3465 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3466 spin_lock_init(&mdsc->snap_flush_lock);
3467 mdsc->last_cap_flush_tid = 1;
3468 mdsc->cap_flush_tree = RB_ROOT;
3469 INIT_LIST_HEAD(&mdsc->cap_dirty);
3470 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3471 mdsc->num_cap_flushing = 0;
3472 spin_lock_init(&mdsc->cap_dirty_lock);
3473 init_waitqueue_head(&mdsc->cap_flushing_wq);
3474 spin_lock_init(&mdsc->dentry_lru_lock);
3475 INIT_LIST_HEAD(&mdsc->dentry_lru);
3477 ceph_caps_init(mdsc);
3478 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3480 init_rwsem(&mdsc->pool_perm_rwsem);
3481 mdsc->pool_perm_tree = RB_ROOT;
3487 * Wait for safe replies on open mds requests. If we time out, drop
3488 * all requests from the tree to avoid dangling dentry refs.
3490 static void wait_requests(struct ceph_mds_client *mdsc)
3492 struct ceph_options *opts = mdsc->fsc->client->options;
3493 struct ceph_mds_request *req;
3495 mutex_lock(&mdsc->mutex);
3496 if (__get_oldest_req(mdsc)) {
3497 mutex_unlock(&mdsc->mutex);
3499 dout("wait_requests waiting for requests\n");
3500 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3501 ceph_timeout_jiffies(opts->mount_timeout));
3503 /* tear down remaining requests */
3504 mutex_lock(&mdsc->mutex);
3505 while ((req = __get_oldest_req(mdsc))) {
3506 dout("wait_requests timed out on tid %llu\n",
3508 __unregister_request(mdsc, req);
3511 mutex_unlock(&mdsc->mutex);
3512 dout("wait_requests done\n");
3516 * called before mount is ro, and before dentries are torn down.
3517 * (hmm, does this still race with new lookups?)
3519 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3521 dout("pre_umount\n");
3525 ceph_flush_dirty_caps(mdsc);
3526 wait_requests(mdsc);
3529 * wait for reply handlers to drop their request refs and
3530 * their inode/dcache refs
3536 * wait for all write mds requests to flush.
3538 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3540 struct ceph_mds_request *req = NULL, *nextreq;
3543 mutex_lock(&mdsc->mutex);
3544 dout("wait_unsafe_requests want %lld\n", want_tid);
3546 req = __get_oldest_req(mdsc);
3547 while (req && req->r_tid <= want_tid) {
3548 /* find next request */
3549 n = rb_next(&req->r_node);
3551 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3554 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3555 (req->r_op & CEPH_MDS_OP_WRITE)) {
3557 ceph_mdsc_get_request(req);
3559 ceph_mdsc_get_request(nextreq);
3560 mutex_unlock(&mdsc->mutex);
3561 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3562 req->r_tid, want_tid);
3563 wait_for_completion(&req->r_safe_completion);
3564 mutex_lock(&mdsc->mutex);
3565 ceph_mdsc_put_request(req);
3567 break; /* next dne before, so we're done! */
3568 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3569 /* next request was removed from tree */
3570 ceph_mdsc_put_request(nextreq);
3573 ceph_mdsc_put_request(nextreq); /* won't go away */
3577 mutex_unlock(&mdsc->mutex);
3578 dout("wait_unsafe_requests done\n");
3581 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3583 u64 want_tid, want_flush, want_snap;
3585 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3589 mutex_lock(&mdsc->mutex);
3590 want_tid = mdsc->last_tid;
3591 mutex_unlock(&mdsc->mutex);
3593 ceph_flush_dirty_caps(mdsc);
3594 spin_lock(&mdsc->cap_dirty_lock);
3595 want_flush = mdsc->last_cap_flush_tid;
3596 spin_unlock(&mdsc->cap_dirty_lock);
3598 down_read(&mdsc->snap_rwsem);
3599 want_snap = mdsc->last_snap_seq;
3600 up_read(&mdsc->snap_rwsem);
3602 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3603 want_tid, want_flush, want_snap);
3605 wait_unsafe_requests(mdsc, want_tid);
3606 wait_caps_flush(mdsc, want_flush, want_snap);
3610 * true if all sessions are closed, or we force unmount
3612 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3614 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3616 return atomic_read(&mdsc->num_sessions) == 0;
3620 * called after sb is ro.
3622 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3624 struct ceph_options *opts = mdsc->fsc->client->options;
3625 struct ceph_mds_session *session;
3628 dout("close_sessions\n");
3630 /* close sessions */
3631 mutex_lock(&mdsc->mutex);
3632 for (i = 0; i < mdsc->max_sessions; i++) {
3633 session = __ceph_lookup_mds_session(mdsc, i);
3636 mutex_unlock(&mdsc->mutex);
3637 mutex_lock(&session->s_mutex);
3638 __close_session(mdsc, session);
3639 mutex_unlock(&session->s_mutex);
3640 ceph_put_mds_session(session);
3641 mutex_lock(&mdsc->mutex);
3643 mutex_unlock(&mdsc->mutex);
3645 dout("waiting for sessions to close\n");
3646 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3647 ceph_timeout_jiffies(opts->mount_timeout));
3649 /* tear down remaining sessions */
3650 mutex_lock(&mdsc->mutex);
3651 for (i = 0; i < mdsc->max_sessions; i++) {
3652 if (mdsc->sessions[i]) {
3653 session = get_session(mdsc->sessions[i]);
3654 __unregister_session(mdsc, session);
3655 mutex_unlock(&mdsc->mutex);
3656 mutex_lock(&session->s_mutex);
3657 remove_session_caps(session);
3658 mutex_unlock(&session->s_mutex);
3659 ceph_put_mds_session(session);
3660 mutex_lock(&mdsc->mutex);
3663 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3664 mutex_unlock(&mdsc->mutex);
3666 ceph_cleanup_empty_realms(mdsc);
3668 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3673 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3675 struct ceph_mds_session *session;
3678 dout("force umount\n");
3680 mutex_lock(&mdsc->mutex);
3681 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3682 session = __ceph_lookup_mds_session(mdsc, mds);
3685 mutex_unlock(&mdsc->mutex);
3686 mutex_lock(&session->s_mutex);
3687 __close_session(mdsc, session);
3688 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3689 cleanup_session_requests(mdsc, session);
3690 remove_session_caps(session);
3692 mutex_unlock(&session->s_mutex);
3693 ceph_put_mds_session(session);
3694 mutex_lock(&mdsc->mutex);
3695 kick_requests(mdsc, mds);
3697 __wake_requests(mdsc, &mdsc->waiting_for_map);
3698 mutex_unlock(&mdsc->mutex);
3701 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3704 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3706 ceph_mdsmap_destroy(mdsc->mdsmap);
3707 kfree(mdsc->sessions);
3708 ceph_caps_finalize(mdsc);
3709 ceph_pool_perm_destroy(mdsc);
3712 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3714 struct ceph_mds_client *mdsc = fsc->mdsc;
3716 dout("mdsc_destroy %p\n", mdsc);
3717 ceph_mdsc_stop(mdsc);
3719 /* flush out any connection work with references to us */
3724 dout("mdsc_destroy %p done\n", mdsc);
3729 * handle mds map update.
3731 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3735 void *p = msg->front.iov_base;
3736 void *end = p + msg->front.iov_len;
3737 struct ceph_mdsmap *newmap, *oldmap;
3738 struct ceph_fsid fsid;
3741 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3742 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3743 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3745 epoch = ceph_decode_32(&p);
3746 maplen = ceph_decode_32(&p);
3747 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3749 /* do we need it? */
3750 mutex_lock(&mdsc->mutex);
3751 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3752 dout("handle_map epoch %u <= our %u\n",
3753 epoch, mdsc->mdsmap->m_epoch);
3754 mutex_unlock(&mdsc->mutex);
3758 newmap = ceph_mdsmap_decode(&p, end);
3759 if (IS_ERR(newmap)) {
3760 err = PTR_ERR(newmap);
3764 /* swap into place */
3766 oldmap = mdsc->mdsmap;
3767 mdsc->mdsmap = newmap;
3768 check_new_map(mdsc, newmap, oldmap);
3769 ceph_mdsmap_destroy(oldmap);
3771 mdsc->mdsmap = newmap; /* first mds map */
3773 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3775 __wake_requests(mdsc, &mdsc->waiting_for_map);
3776 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3777 mdsc->mdsmap->m_epoch);
3779 mutex_unlock(&mdsc->mutex);
3780 schedule_delayed(mdsc);
3784 mutex_unlock(&mdsc->mutex);
3786 pr_err("error decoding mdsmap %d\n", err);
3790 static struct ceph_connection *con_get(struct ceph_connection *con)
3792 struct ceph_mds_session *s = con->private;
3794 if (get_session(s)) {
3795 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3798 dout("mdsc con_get %p FAIL\n", s);
3802 static void con_put(struct ceph_connection *con)
3804 struct ceph_mds_session *s = con->private;
3806 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3807 ceph_put_mds_session(s);
3811 * if the client is unresponsive for long enough, the mds will kill
3812 * the session entirely.
3814 static void peer_reset(struct ceph_connection *con)
3816 struct ceph_mds_session *s = con->private;
3817 struct ceph_mds_client *mdsc = s->s_mdsc;
3819 pr_warn("mds%d closed our session\n", s->s_mds);
3820 send_mds_reconnect(mdsc, s);
3823 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3825 struct ceph_mds_session *s = con->private;
3826 struct ceph_mds_client *mdsc = s->s_mdsc;
3827 int type = le16_to_cpu(msg->hdr.type);
3829 mutex_lock(&mdsc->mutex);
3830 if (__verify_registered_session(mdsc, s) < 0) {
3831 mutex_unlock(&mdsc->mutex);
3834 mutex_unlock(&mdsc->mutex);
3837 case CEPH_MSG_MDS_MAP:
3838 ceph_mdsc_handle_map(mdsc, msg);
3840 case CEPH_MSG_CLIENT_SESSION:
3841 handle_session(s, msg);
3843 case CEPH_MSG_CLIENT_REPLY:
3844 handle_reply(s, msg);
3846 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3847 handle_forward(mdsc, s, msg);
3849 case CEPH_MSG_CLIENT_CAPS:
3850 ceph_handle_caps(s, msg);
3852 case CEPH_MSG_CLIENT_SNAP:
3853 ceph_handle_snap(mdsc, s, msg);
3855 case CEPH_MSG_CLIENT_LEASE:
3856 handle_lease(mdsc, s, msg);
3860 pr_err("received unknown message type %d %s\n", type,
3861 ceph_msg_type_name(type));
3872 * Note: returned pointer is the address of a structure that's
3873 * managed separately. Caller must *not* attempt to free it.
3875 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3876 int *proto, int force_new)
3878 struct ceph_mds_session *s = con->private;
3879 struct ceph_mds_client *mdsc = s->s_mdsc;
3880 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3881 struct ceph_auth_handshake *auth = &s->s_auth;
3883 if (force_new && auth->authorizer) {
3884 ceph_auth_destroy_authorizer(auth->authorizer);
3885 auth->authorizer = NULL;
3887 if (!auth->authorizer) {
3888 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3891 return ERR_PTR(ret);
3893 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3896 return ERR_PTR(ret);
3898 *proto = ac->protocol;
3904 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3906 struct ceph_mds_session *s = con->private;
3907 struct ceph_mds_client *mdsc = s->s_mdsc;
3908 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3910 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3913 static int invalidate_authorizer(struct ceph_connection *con)
3915 struct ceph_mds_session *s = con->private;
3916 struct ceph_mds_client *mdsc = s->s_mdsc;
3917 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3919 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3921 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3924 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3925 struct ceph_msg_header *hdr, int *skip)
3927 struct ceph_msg *msg;
3928 int type = (int) le16_to_cpu(hdr->type);
3929 int front_len = (int) le32_to_cpu(hdr->front_len);
3935 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3937 pr_err("unable to allocate msg type %d len %d\n",
3945 static int mds_sign_message(struct ceph_msg *msg)
3947 struct ceph_mds_session *s = msg->con->private;
3948 struct ceph_auth_handshake *auth = &s->s_auth;
3950 return ceph_auth_sign_message(auth, msg);
3953 static int mds_check_message_signature(struct ceph_msg *msg)
3955 struct ceph_mds_session *s = msg->con->private;
3956 struct ceph_auth_handshake *auth = &s->s_auth;
3958 return ceph_auth_check_message_signature(auth, msg);
3961 static const struct ceph_connection_operations mds_con_ops = {
3964 .dispatch = dispatch,
3965 .get_authorizer = get_authorizer,
3966 .verify_authorizer_reply = verify_authorizer_reply,
3967 .invalidate_authorizer = invalidate_authorizer,
3968 .peer_reset = peer_reset,
3969 .alloc_msg = mds_alloc_msg,
3970 .sign_message = mds_sign_message,
3971 .check_message_signature = mds_check_message_signature,