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45 * are no longer valid.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;
109 * parse a normal reply, which may contain a (dir+)dentry and/or a
112 static int parse_reply_info_trace(void **p, void *end,
113 struct ceph_mds_reply_info_parsed *info,
118 if (info->head->is_dentry) {
119 err = parse_reply_info_in(p, end, &info->diri, features);
123 if (unlikely(*p + sizeof(*info->dirfrag) > end))
126 *p += sizeof(*info->dirfrag) +
127 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
128 if (unlikely(*p > end))
131 ceph_decode_32_safe(p, end, info->dname_len, bad);
132 ceph_decode_need(p, end, info->dname_len, bad);
134 *p += info->dname_len;
136 *p += sizeof(*info->dlease);
139 if (info->head->is_target) {
140 err = parse_reply_info_in(p, end, &info->targeti, features);
145 if (unlikely(*p != end))
152 pr_err("problem parsing mds trace %d\n", err);
157 * parse readdir results
159 static int parse_reply_info_dir(void **p, void *end,
160 struct ceph_mds_reply_info_parsed *info,
167 if (*p + sizeof(*info->dir_dir) > end)
169 *p += sizeof(*info->dir_dir) +
170 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
174 ceph_decode_need(p, end, sizeof(num) + 2, bad);
175 num = ceph_decode_32(p);
176 info->dir_end = ceph_decode_8(p);
177 info->dir_complete = ceph_decode_8(p);
181 BUG_ON(!info->dir_in);
182 info->dir_dname = (void *)(info->dir_in + num);
183 info->dir_dname_len = (void *)(info->dir_dname + num);
184 info->dir_dlease = (void *)(info->dir_dname_len + num);
185 if ((unsigned long)(info->dir_dlease + num) >
186 (unsigned long)info->dir_in + info->dir_buf_size) {
187 pr_err("dir contents are larger than expected\n");
195 ceph_decode_need(p, end, sizeof(u32)*2, bad);
196 info->dir_dname_len[i] = ceph_decode_32(p);
197 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
198 info->dir_dname[i] = *p;
199 *p += info->dir_dname_len[i];
200 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
202 info->dir_dlease[i] = *p;
203 *p += sizeof(struct ceph_mds_reply_lease);
206 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
221 pr_err("problem parsing dir contents %d\n", err);
226 * parse fcntl F_GETLK results
228 static int parse_reply_info_filelock(void **p, void *end,
229 struct ceph_mds_reply_info_parsed *info,
232 if (*p + sizeof(*info->filelock_reply) > end)
235 info->filelock_reply = *p;
236 *p += sizeof(*info->filelock_reply);
238 if (unlikely(*p != end))
247 * parse create results
249 static int parse_reply_info_create(void **p, void *end,
250 struct ceph_mds_reply_info_parsed *info,
253 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
255 info->has_create_ino = false;
257 info->has_create_ino = true;
258 info->ino = ceph_decode_64(p);
262 if (unlikely(*p != end))
271 * parse extra results
273 static int parse_reply_info_extra(void **p, void *end,
274 struct ceph_mds_reply_info_parsed *info,
277 u32 op = le32_to_cpu(info->head->op);
279 if (op == CEPH_MDS_OP_GETFILELOCK)
280 return parse_reply_info_filelock(p, end, info, features);
281 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
282 return parse_reply_info_dir(p, end, info, features);
283 else if (op == CEPH_MDS_OP_CREATE)
284 return parse_reply_info_create(p, end, info, features);
290 * parse entire mds reply
292 static int parse_reply_info(struct ceph_msg *msg,
293 struct ceph_mds_reply_info_parsed *info,
300 info->head = msg->front.iov_base;
301 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
302 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
305 ceph_decode_32_safe(&p, end, len, bad);
307 ceph_decode_need(&p, end, len, bad);
308 err = parse_reply_info_trace(&p, p+len, info, features);
314 ceph_decode_32_safe(&p, end, len, bad);
316 ceph_decode_need(&p, end, len, bad);
317 err = parse_reply_info_extra(&p, p+len, info, features);
323 ceph_decode_32_safe(&p, end, len, bad);
324 info->snapblob_len = len;
335 pr_err("mds parse_reply err %d\n", err);
339 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
343 free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
350 const char *ceph_session_state_name(int s)
353 case CEPH_MDS_SESSION_NEW: return "new";
354 case CEPH_MDS_SESSION_OPENING: return "opening";
355 case CEPH_MDS_SESSION_OPEN: return "open";
356 case CEPH_MDS_SESSION_HUNG: return "hung";
357 case CEPH_MDS_SESSION_CLOSING: return "closing";
358 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
359 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
360 default: return "???";
364 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
366 if (atomic_inc_not_zero(&s->s_ref)) {
367 dout("mdsc get_session %p %d -> %d\n", s,
368 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
371 dout("mdsc get_session %p 0 -- FAIL", s);
376 void ceph_put_mds_session(struct ceph_mds_session *s)
378 dout("mdsc put_session %p %d -> %d\n", s,
379 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
380 if (atomic_dec_and_test(&s->s_ref)) {
381 if (s->s_auth.authorizer)
382 ceph_auth_destroy_authorizer(
383 s->s_mdsc->fsc->client->monc.auth,
384 s->s_auth.authorizer);
390 * called under mdsc->mutex
392 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
395 struct ceph_mds_session *session;
397 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
399 session = mdsc->sessions[mds];
400 dout("lookup_mds_session %p %d\n", session,
401 atomic_read(&session->s_ref));
402 get_session(session);
406 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
408 if (mds >= mdsc->max_sessions)
410 return mdsc->sessions[mds];
413 static int __verify_registered_session(struct ceph_mds_client *mdsc,
414 struct ceph_mds_session *s)
416 if (s->s_mds >= mdsc->max_sessions ||
417 mdsc->sessions[s->s_mds] != s)
423 * create+register a new session for given mds.
424 * called under mdsc->mutex.
426 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
429 struct ceph_mds_session *s;
431 if (mds >= mdsc->mdsmap->m_max_mds)
432 return ERR_PTR(-EINVAL);
434 s = kzalloc(sizeof(*s), GFP_NOFS);
436 return ERR_PTR(-ENOMEM);
439 s->s_state = CEPH_MDS_SESSION_NEW;
442 mutex_init(&s->s_mutex);
444 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
446 spin_lock_init(&s->s_gen_ttl_lock);
448 s->s_cap_ttl = jiffies - 1;
450 spin_lock_init(&s->s_cap_lock);
451 s->s_renew_requested = 0;
453 INIT_LIST_HEAD(&s->s_caps);
456 atomic_set(&s->s_ref, 1);
457 INIT_LIST_HEAD(&s->s_waiting);
458 INIT_LIST_HEAD(&s->s_unsafe);
459 s->s_num_cap_releases = 0;
460 s->s_cap_reconnect = 0;
461 s->s_cap_iterator = NULL;
462 INIT_LIST_HEAD(&s->s_cap_releases);
463 INIT_LIST_HEAD(&s->s_cap_flushing);
464 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
466 dout("register_session mds%d\n", mds);
467 if (mds >= mdsc->max_sessions) {
468 int newmax = 1 << get_count_order(mds+1);
469 struct ceph_mds_session **sa;
471 dout("register_session realloc to %d\n", newmax);
472 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
475 if (mdsc->sessions) {
476 memcpy(sa, mdsc->sessions,
477 mdsc->max_sessions * sizeof(void *));
478 kfree(mdsc->sessions);
481 mdsc->max_sessions = newmax;
483 mdsc->sessions[mds] = s;
484 atomic_inc(&mdsc->num_sessions);
485 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
487 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
488 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
494 return ERR_PTR(-ENOMEM);
498 * called under mdsc->mutex
500 static void __unregister_session(struct ceph_mds_client *mdsc,
501 struct ceph_mds_session *s)
503 dout("__unregister_session mds%d %p\n", s->s_mds, s);
504 BUG_ON(mdsc->sessions[s->s_mds] != s);
505 mdsc->sessions[s->s_mds] = NULL;
506 ceph_con_close(&s->s_con);
507 ceph_put_mds_session(s);
508 atomic_dec(&mdsc->num_sessions);
512 * drop session refs in request.
514 * should be last request ref, or hold mdsc->mutex
516 static void put_request_session(struct ceph_mds_request *req)
518 if (req->r_session) {
519 ceph_put_mds_session(req->r_session);
520 req->r_session = NULL;
524 void ceph_mdsc_release_request(struct kref *kref)
526 struct ceph_mds_request *req = container_of(kref,
527 struct ceph_mds_request,
529 destroy_reply_info(&req->r_reply_info);
531 ceph_msg_put(req->r_request);
533 ceph_msg_put(req->r_reply);
535 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
538 if (req->r_locked_dir)
539 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
540 iput(req->r_target_inode);
543 if (req->r_old_dentry)
544 dput(req->r_old_dentry);
545 if (req->r_old_dentry_dir) {
547 * track (and drop pins for) r_old_dentry_dir
548 * separately, since r_old_dentry's d_parent may have
549 * changed between the dir mutex being dropped and
550 * this request being freed.
552 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
554 iput(req->r_old_dentry_dir);
559 ceph_pagelist_release(req->r_pagelist);
560 put_request_session(req);
561 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
566 * lookup session, bump ref if found.
568 * called under mdsc->mutex.
570 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
573 struct ceph_mds_request *req;
574 struct rb_node *n = mdsc->request_tree.rb_node;
577 req = rb_entry(n, struct ceph_mds_request, r_node);
578 if (tid < req->r_tid)
580 else if (tid > req->r_tid)
583 ceph_mdsc_get_request(req);
590 static void __insert_request(struct ceph_mds_client *mdsc,
591 struct ceph_mds_request *new)
593 struct rb_node **p = &mdsc->request_tree.rb_node;
594 struct rb_node *parent = NULL;
595 struct ceph_mds_request *req = NULL;
599 req = rb_entry(parent, struct ceph_mds_request, r_node);
600 if (new->r_tid < req->r_tid)
602 else if (new->r_tid > req->r_tid)
608 rb_link_node(&new->r_node, parent, p);
609 rb_insert_color(&new->r_node, &mdsc->request_tree);
613 * Register an in-flight request, and assign a tid. Link to directory
614 * are modifying (if any).
616 * Called under mdsc->mutex.
618 static void __register_request(struct ceph_mds_client *mdsc,
619 struct ceph_mds_request *req,
622 req->r_tid = ++mdsc->last_tid;
624 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
626 dout("__register_request %p tid %lld\n", req, req->r_tid);
627 ceph_mdsc_get_request(req);
628 __insert_request(mdsc, req);
630 req->r_uid = current_fsuid();
631 req->r_gid = current_fsgid();
633 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
634 mdsc->oldest_tid = req->r_tid;
638 req->r_unsafe_dir = dir;
642 static void __unregister_request(struct ceph_mds_client *mdsc,
643 struct ceph_mds_request *req)
645 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
647 if (req->r_tid == mdsc->oldest_tid) {
648 struct rb_node *p = rb_next(&req->r_node);
649 mdsc->oldest_tid = 0;
651 struct ceph_mds_request *next_req =
652 rb_entry(p, struct ceph_mds_request, r_node);
653 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
654 mdsc->oldest_tid = next_req->r_tid;
661 rb_erase(&req->r_node, &mdsc->request_tree);
662 RB_CLEAR_NODE(&req->r_node);
664 if (req->r_unsafe_dir && req->r_got_unsafe) {
665 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
666 spin_lock(&ci->i_unsafe_lock);
667 list_del_init(&req->r_unsafe_dir_item);
668 spin_unlock(&ci->i_unsafe_lock);
670 if (req->r_target_inode && req->r_got_unsafe) {
671 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
672 spin_lock(&ci->i_unsafe_lock);
673 list_del_init(&req->r_unsafe_target_item);
674 spin_unlock(&ci->i_unsafe_lock);
677 if (req->r_unsafe_dir) {
678 iput(req->r_unsafe_dir);
679 req->r_unsafe_dir = NULL;
682 complete_all(&req->r_safe_completion);
684 ceph_mdsc_put_request(req);
688 * Choose mds to send request to next. If there is a hint set in the
689 * request (e.g., due to a prior forward hint from the mds), use that.
690 * Otherwise, consult frag tree and/or caps to identify the
691 * appropriate mds. If all else fails, choose randomly.
693 * Called under mdsc->mutex.
695 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
698 * we don't need to worry about protecting the d_parent access
699 * here because we never renaming inside the snapped namespace
700 * except to resplice to another snapdir, and either the old or new
701 * result is a valid result.
703 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
704 dentry = dentry->d_parent;
708 static int __choose_mds(struct ceph_mds_client *mdsc,
709 struct ceph_mds_request *req)
712 struct ceph_inode_info *ci;
713 struct ceph_cap *cap;
714 int mode = req->r_direct_mode;
716 u32 hash = req->r_direct_hash;
717 bool is_hash = req->r_direct_is_hash;
720 * is there a specific mds we should try? ignore hint if we have
721 * no session and the mds is not up (active or recovering).
723 if (req->r_resend_mds >= 0 &&
724 (__have_session(mdsc, req->r_resend_mds) ||
725 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
726 dout("choose_mds using resend_mds mds%d\n",
728 return req->r_resend_mds;
731 if (mode == USE_RANDOM_MDS)
736 inode = req->r_inode;
737 } else if (req->r_dentry) {
738 /* ignore race with rename; old or new d_parent is okay */
739 struct dentry *parent = req->r_dentry->d_parent;
740 struct inode *dir = d_inode(parent);
742 if (dir->i_sb != mdsc->fsc->sb) {
744 inode = d_inode(req->r_dentry);
745 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
746 /* direct snapped/virtual snapdir requests
747 * based on parent dir inode */
748 struct dentry *dn = get_nonsnap_parent(parent);
750 dout("__choose_mds using nonsnap parent %p\n", inode);
753 inode = d_inode(req->r_dentry);
754 if (!inode || mode == USE_AUTH_MDS) {
757 hash = ceph_dentry_hash(dir, req->r_dentry);
763 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
767 ci = ceph_inode(inode);
769 if (is_hash && S_ISDIR(inode->i_mode)) {
770 struct ceph_inode_frag frag;
773 ceph_choose_frag(ci, hash, &frag, &found);
775 if (mode == USE_ANY_MDS && frag.ndist > 0) {
778 /* choose a random replica */
779 get_random_bytes(&r, 1);
782 dout("choose_mds %p %llx.%llx "
783 "frag %u mds%d (%d/%d)\n",
784 inode, ceph_vinop(inode),
787 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
788 CEPH_MDS_STATE_ACTIVE)
792 /* since this file/dir wasn't known to be
793 * replicated, then we want to look for the
794 * authoritative mds. */
797 /* choose auth mds */
799 dout("choose_mds %p %llx.%llx "
800 "frag %u mds%d (auth)\n",
801 inode, ceph_vinop(inode), frag.frag, mds);
802 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
803 CEPH_MDS_STATE_ACTIVE)
809 spin_lock(&ci->i_ceph_lock);
811 if (mode == USE_AUTH_MDS)
812 cap = ci->i_auth_cap;
813 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
814 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
816 spin_unlock(&ci->i_ceph_lock);
819 mds = cap->session->s_mds;
820 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
821 inode, ceph_vinop(inode), mds,
822 cap == ci->i_auth_cap ? "auth " : "", cap);
823 spin_unlock(&ci->i_ceph_lock);
827 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
828 dout("choose_mds chose random mds%d\n", mds);
836 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
838 struct ceph_msg *msg;
839 struct ceph_mds_session_head *h;
841 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
844 pr_err("create_session_msg ENOMEM creating msg\n");
847 h = msg->front.iov_base;
848 h->op = cpu_to_le32(op);
849 h->seq = cpu_to_le64(seq);
855 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
856 * to include additional client metadata fields.
858 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
860 struct ceph_msg *msg;
861 struct ceph_mds_session_head *h;
863 int metadata_bytes = 0;
864 int metadata_key_count = 0;
865 struct ceph_options *opt = mdsc->fsc->client->options;
868 const char* metadata[][2] = {
869 {"hostname", utsname()->nodename},
870 {"kernel_version", utsname()->release},
871 {"entity_id", opt->name ? opt->name : ""},
875 /* Calculate serialized length of metadata */
876 metadata_bytes = 4; /* map length */
877 for (i = 0; metadata[i][0] != NULL; ++i) {
878 metadata_bytes += 8 + strlen(metadata[i][0]) +
879 strlen(metadata[i][1]);
880 metadata_key_count++;
883 /* Allocate the message */
884 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
887 pr_err("create_session_msg ENOMEM creating msg\n");
890 h = msg->front.iov_base;
891 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
892 h->seq = cpu_to_le64(seq);
895 * Serialize client metadata into waiting buffer space, using
896 * the format that userspace expects for map<string, string>
898 * ClientSession messages with metadata are v2
900 msg->hdr.version = cpu_to_le16(2);
901 msg->hdr.compat_version = cpu_to_le16(1);
903 /* The write pointer, following the session_head structure */
904 p = msg->front.iov_base + sizeof(*h);
906 /* Number of entries in the map */
907 ceph_encode_32(&p, metadata_key_count);
909 /* Two length-prefixed strings for each entry in the map */
910 for (i = 0; metadata[i][0] != NULL; ++i) {
911 size_t const key_len = strlen(metadata[i][0]);
912 size_t const val_len = strlen(metadata[i][1]);
914 ceph_encode_32(&p, key_len);
915 memcpy(p, metadata[i][0], key_len);
917 ceph_encode_32(&p, val_len);
918 memcpy(p, metadata[i][1], val_len);
926 * send session open request.
928 * called under mdsc->mutex
930 static int __open_session(struct ceph_mds_client *mdsc,
931 struct ceph_mds_session *session)
933 struct ceph_msg *msg;
935 int mds = session->s_mds;
937 /* wait for mds to go active? */
938 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
939 dout("open_session to mds%d (%s)\n", mds,
940 ceph_mds_state_name(mstate));
941 session->s_state = CEPH_MDS_SESSION_OPENING;
942 session->s_renew_requested = jiffies;
944 /* send connect message */
945 msg = create_session_open_msg(mdsc, session->s_seq);
948 ceph_con_send(&session->s_con, msg);
953 * open sessions for any export targets for the given mds
955 * called under mdsc->mutex
957 static struct ceph_mds_session *
958 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
960 struct ceph_mds_session *session;
962 session = __ceph_lookup_mds_session(mdsc, target);
964 session = register_session(mdsc, target);
968 if (session->s_state == CEPH_MDS_SESSION_NEW ||
969 session->s_state == CEPH_MDS_SESSION_CLOSING)
970 __open_session(mdsc, session);
975 struct ceph_mds_session *
976 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
978 struct ceph_mds_session *session;
980 dout("open_export_target_session to mds%d\n", target);
982 mutex_lock(&mdsc->mutex);
983 session = __open_export_target_session(mdsc, target);
984 mutex_unlock(&mdsc->mutex);
989 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
990 struct ceph_mds_session *session)
992 struct ceph_mds_info *mi;
993 struct ceph_mds_session *ts;
994 int i, mds = session->s_mds;
996 if (mds >= mdsc->mdsmap->m_max_mds)
999 mi = &mdsc->mdsmap->m_info[mds];
1000 dout("open_export_target_sessions for mds%d (%d targets)\n",
1001 session->s_mds, mi->num_export_targets);
1003 for (i = 0; i < mi->num_export_targets; i++) {
1004 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1006 ceph_put_mds_session(ts);
1010 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1011 struct ceph_mds_session *session)
1013 mutex_lock(&mdsc->mutex);
1014 __open_export_target_sessions(mdsc, session);
1015 mutex_unlock(&mdsc->mutex);
1022 /* caller holds s_cap_lock, we drop it */
1023 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1024 struct ceph_mds_session *session)
1025 __releases(session->s_cap_lock)
1027 LIST_HEAD(tmp_list);
1028 list_splice_init(&session->s_cap_releases, &tmp_list);
1029 session->s_num_cap_releases = 0;
1030 spin_unlock(&session->s_cap_lock);
1032 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1033 while (!list_empty(&tmp_list)) {
1034 struct ceph_cap *cap;
1035 /* zero out the in-progress message */
1036 cap = list_first_entry(&tmp_list,
1037 struct ceph_cap, session_caps);
1038 list_del(&cap->session_caps);
1039 ceph_put_cap(mdsc, cap);
1043 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1044 struct ceph_mds_session *session)
1046 struct ceph_mds_request *req;
1049 dout("cleanup_session_requests mds%d\n", session->s_mds);
1050 mutex_lock(&mdsc->mutex);
1051 while (!list_empty(&session->s_unsafe)) {
1052 req = list_first_entry(&session->s_unsafe,
1053 struct ceph_mds_request, r_unsafe_item);
1054 list_del_init(&req->r_unsafe_item);
1055 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1057 __unregister_request(mdsc, req);
1059 /* zero r_attempts, so kick_requests() will re-send requests */
1060 p = rb_first(&mdsc->request_tree);
1062 req = rb_entry(p, struct ceph_mds_request, r_node);
1064 if (req->r_session &&
1065 req->r_session->s_mds == session->s_mds)
1066 req->r_attempts = 0;
1068 mutex_unlock(&mdsc->mutex);
1072 * Helper to safely iterate over all caps associated with a session, with
1073 * special care taken to handle a racing __ceph_remove_cap().
1075 * Caller must hold session s_mutex.
1077 static int iterate_session_caps(struct ceph_mds_session *session,
1078 int (*cb)(struct inode *, struct ceph_cap *,
1081 struct list_head *p;
1082 struct ceph_cap *cap;
1083 struct inode *inode, *last_inode = NULL;
1084 struct ceph_cap *old_cap = NULL;
1087 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1088 spin_lock(&session->s_cap_lock);
1089 p = session->s_caps.next;
1090 while (p != &session->s_caps) {
1091 cap = list_entry(p, struct ceph_cap, session_caps);
1092 inode = igrab(&cap->ci->vfs_inode);
1097 session->s_cap_iterator = cap;
1098 spin_unlock(&session->s_cap_lock);
1105 ceph_put_cap(session->s_mdsc, old_cap);
1109 ret = cb(inode, cap, arg);
1112 spin_lock(&session->s_cap_lock);
1114 if (cap->ci == NULL) {
1115 dout("iterate_session_caps finishing cap %p removal\n",
1117 BUG_ON(cap->session != session);
1118 cap->session = NULL;
1119 list_del_init(&cap->session_caps);
1120 session->s_nr_caps--;
1121 if (cap->queue_release) {
1122 list_add_tail(&cap->session_caps,
1123 &session->s_cap_releases);
1124 session->s_num_cap_releases++;
1126 old_cap = cap; /* put_cap it w/o locks held */
1134 session->s_cap_iterator = NULL;
1135 spin_unlock(&session->s_cap_lock);
1139 ceph_put_cap(session->s_mdsc, old_cap);
1144 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1147 struct ceph_inode_info *ci = ceph_inode(inode);
1148 LIST_HEAD(to_remove);
1151 dout("removing cap %p, ci is %p, inode is %p\n",
1152 cap, ci, &ci->vfs_inode);
1153 spin_lock(&ci->i_ceph_lock);
1154 __ceph_remove_cap(cap, false);
1155 if (!ci->i_auth_cap) {
1156 struct ceph_cap_flush *cf;
1157 struct ceph_mds_client *mdsc =
1158 ceph_sb_to_client(inode->i_sb)->mdsc;
1161 struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1164 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1165 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1166 list_add(&cf->list, &to_remove);
1169 spin_lock(&mdsc->cap_dirty_lock);
1171 list_for_each_entry(cf, &to_remove, list)
1172 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1174 if (!list_empty(&ci->i_dirty_item)) {
1175 pr_warn_ratelimited(
1176 " dropping dirty %s state for %p %lld\n",
1177 ceph_cap_string(ci->i_dirty_caps),
1178 inode, ceph_ino(inode));
1179 ci->i_dirty_caps = 0;
1180 list_del_init(&ci->i_dirty_item);
1183 if (!list_empty(&ci->i_flushing_item)) {
1184 pr_warn_ratelimited(
1185 " dropping dirty+flushing %s state for %p %lld\n",
1186 ceph_cap_string(ci->i_flushing_caps),
1187 inode, ceph_ino(inode));
1188 ci->i_flushing_caps = 0;
1189 list_del_init(&ci->i_flushing_item);
1190 mdsc->num_cap_flushing--;
1193 spin_unlock(&mdsc->cap_dirty_lock);
1195 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1196 list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1197 ci->i_prealloc_cap_flush = NULL;
1200 spin_unlock(&ci->i_ceph_lock);
1201 while (!list_empty(&to_remove)) {
1202 struct ceph_cap_flush *cf;
1203 cf = list_first_entry(&to_remove,
1204 struct ceph_cap_flush, list);
1205 list_del(&cf->list);
1206 ceph_free_cap_flush(cf);
1214 * caller must hold session s_mutex
1216 static void remove_session_caps(struct ceph_mds_session *session)
1218 dout("remove_session_caps on %p\n", session);
1219 iterate_session_caps(session, remove_session_caps_cb, NULL);
1221 spin_lock(&session->s_cap_lock);
1222 if (session->s_nr_caps > 0) {
1223 struct super_block *sb = session->s_mdsc->fsc->sb;
1224 struct inode *inode;
1225 struct ceph_cap *cap, *prev = NULL;
1226 struct ceph_vino vino;
1228 * iterate_session_caps() skips inodes that are being
1229 * deleted, we need to wait until deletions are complete.
1230 * __wait_on_freeing_inode() is designed for the job,
1231 * but it is not exported, so use lookup inode function
1234 while (!list_empty(&session->s_caps)) {
1235 cap = list_entry(session->s_caps.next,
1236 struct ceph_cap, session_caps);
1240 vino = cap->ci->i_vino;
1241 spin_unlock(&session->s_cap_lock);
1243 inode = ceph_find_inode(sb, vino);
1246 spin_lock(&session->s_cap_lock);
1250 // drop cap expires and unlock s_cap_lock
1251 cleanup_cap_releases(session->s_mdsc, session);
1253 BUG_ON(session->s_nr_caps > 0);
1254 BUG_ON(!list_empty(&session->s_cap_flushing));
1258 * wake up any threads waiting on this session's caps. if the cap is
1259 * old (didn't get renewed on the client reconnect), remove it now.
1261 * caller must hold s_mutex.
1263 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1266 struct ceph_inode_info *ci = ceph_inode(inode);
1268 wake_up_all(&ci->i_cap_wq);
1270 spin_lock(&ci->i_ceph_lock);
1271 ci->i_wanted_max_size = 0;
1272 ci->i_requested_max_size = 0;
1273 spin_unlock(&ci->i_ceph_lock);
1278 static void wake_up_session_caps(struct ceph_mds_session *session,
1281 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1282 iterate_session_caps(session, wake_up_session_cb,
1283 (void *)(unsigned long)reconnect);
1287 * Send periodic message to MDS renewing all currently held caps. The
1288 * ack will reset the expiration for all caps from this session.
1290 * caller holds s_mutex
1292 static int send_renew_caps(struct ceph_mds_client *mdsc,
1293 struct ceph_mds_session *session)
1295 struct ceph_msg *msg;
1298 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1299 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1300 pr_info("mds%d caps stale\n", session->s_mds);
1301 session->s_renew_requested = jiffies;
1303 /* do not try to renew caps until a recovering mds has reconnected
1304 * with its clients. */
1305 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1306 if (state < CEPH_MDS_STATE_RECONNECT) {
1307 dout("send_renew_caps ignoring mds%d (%s)\n",
1308 session->s_mds, ceph_mds_state_name(state));
1312 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1313 ceph_mds_state_name(state));
1314 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1315 ++session->s_renew_seq);
1318 ceph_con_send(&session->s_con, msg);
1322 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1323 struct ceph_mds_session *session, u64 seq)
1325 struct ceph_msg *msg;
1327 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1328 session->s_mds, ceph_session_state_name(session->s_state), seq);
1329 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1332 ceph_con_send(&session->s_con, msg);
1338 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1340 * Called under session->s_mutex
1342 static void renewed_caps(struct ceph_mds_client *mdsc,
1343 struct ceph_mds_session *session, int is_renew)
1348 spin_lock(&session->s_cap_lock);
1349 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1351 session->s_cap_ttl = session->s_renew_requested +
1352 mdsc->mdsmap->m_session_timeout*HZ;
1355 if (time_before(jiffies, session->s_cap_ttl)) {
1356 pr_info("mds%d caps renewed\n", session->s_mds);
1359 pr_info("mds%d caps still stale\n", session->s_mds);
1362 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1363 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1364 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1365 spin_unlock(&session->s_cap_lock);
1368 wake_up_session_caps(session, 0);
1372 * send a session close request
1374 static int request_close_session(struct ceph_mds_client *mdsc,
1375 struct ceph_mds_session *session)
1377 struct ceph_msg *msg;
1379 dout("request_close_session mds%d state %s seq %lld\n",
1380 session->s_mds, ceph_session_state_name(session->s_state),
1382 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1385 ceph_con_send(&session->s_con, msg);
1390 * Called with s_mutex held.
1392 static int __close_session(struct ceph_mds_client *mdsc,
1393 struct ceph_mds_session *session)
1395 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1397 session->s_state = CEPH_MDS_SESSION_CLOSING;
1398 return request_close_session(mdsc, session);
1402 * Trim old(er) caps.
1404 * Because we can't cache an inode without one or more caps, we do
1405 * this indirectly: if a cap is unused, we prune its aliases, at which
1406 * point the inode will hopefully get dropped to.
1408 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1409 * memory pressure from the MDS, though, so it needn't be perfect.
1411 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1413 struct ceph_mds_session *session = arg;
1414 struct ceph_inode_info *ci = ceph_inode(inode);
1415 int used, wanted, oissued, mine;
1417 if (session->s_trim_caps <= 0)
1420 spin_lock(&ci->i_ceph_lock);
1421 mine = cap->issued | cap->implemented;
1422 used = __ceph_caps_used(ci);
1423 wanted = __ceph_caps_file_wanted(ci);
1424 oissued = __ceph_caps_issued_other(ci, cap);
1426 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1427 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1428 ceph_cap_string(used), ceph_cap_string(wanted));
1429 if (cap == ci->i_auth_cap) {
1430 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1431 !list_empty(&ci->i_cap_snaps))
1433 if ((used | wanted) & CEPH_CAP_ANY_WR)
1436 /* The inode has cached pages, but it's no longer used.
1437 * we can safely drop it */
1438 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1439 !(oissued & CEPH_CAP_FILE_CACHE)) {
1443 if ((used | wanted) & ~oissued & mine)
1444 goto out; /* we need these caps */
1446 session->s_trim_caps--;
1448 /* we aren't the only cap.. just remove us */
1449 __ceph_remove_cap(cap, true);
1451 /* try dropping referring dentries */
1452 spin_unlock(&ci->i_ceph_lock);
1453 d_prune_aliases(inode);
1454 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1455 inode, cap, atomic_read(&inode->i_count));
1460 spin_unlock(&ci->i_ceph_lock);
1465 * Trim session cap count down to some max number.
1467 static int trim_caps(struct ceph_mds_client *mdsc,
1468 struct ceph_mds_session *session,
1471 int trim_caps = session->s_nr_caps - max_caps;
1473 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1474 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1475 if (trim_caps > 0) {
1476 session->s_trim_caps = trim_caps;
1477 iterate_session_caps(session, trim_caps_cb, session);
1478 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1479 session->s_mds, session->s_nr_caps, max_caps,
1480 trim_caps - session->s_trim_caps);
1481 session->s_trim_caps = 0;
1484 ceph_send_cap_releases(mdsc, session);
1488 static int check_capsnap_flush(struct ceph_inode_info *ci,
1492 spin_lock(&ci->i_ceph_lock);
1493 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1494 struct ceph_cap_snap *capsnap =
1495 list_first_entry(&ci->i_cap_snaps,
1496 struct ceph_cap_snap, ci_item);
1497 ret = capsnap->follows >= want_snap_seq;
1499 spin_unlock(&ci->i_ceph_lock);
1503 static int check_caps_flush(struct ceph_mds_client *mdsc,
1507 struct ceph_cap_flush *cf;
1510 spin_lock(&mdsc->cap_dirty_lock);
1511 n = rb_first(&mdsc->cap_flush_tree);
1512 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1513 if (cf && cf->tid <= want_flush_tid) {
1514 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1515 cf->tid, want_flush_tid);
1518 spin_unlock(&mdsc->cap_dirty_lock);
1523 * flush all dirty inode data to disk.
1525 * returns true if we've flushed through want_flush_tid
1527 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1528 u64 want_flush_tid, u64 want_snap_seq)
1532 dout("check_caps_flush want %llu snap want %llu\n",
1533 want_flush_tid, want_snap_seq);
1534 mutex_lock(&mdsc->mutex);
1535 for (mds = 0; mds < mdsc->max_sessions; ) {
1536 struct ceph_mds_session *session = mdsc->sessions[mds];
1537 struct inode *inode = NULL;
1543 get_session(session);
1544 mutex_unlock(&mdsc->mutex);
1546 mutex_lock(&session->s_mutex);
1547 if (!list_empty(&session->s_cap_snaps_flushing)) {
1548 struct ceph_cap_snap *capsnap =
1549 list_first_entry(&session->s_cap_snaps_flushing,
1550 struct ceph_cap_snap,
1552 struct ceph_inode_info *ci = capsnap->ci;
1553 if (!check_capsnap_flush(ci, want_snap_seq)) {
1554 dout("check_cap_flush still flushing snap %p "
1555 "follows %lld <= %lld to mds%d\n",
1556 &ci->vfs_inode, capsnap->follows,
1557 want_snap_seq, mds);
1558 inode = igrab(&ci->vfs_inode);
1561 mutex_unlock(&session->s_mutex);
1562 ceph_put_mds_session(session);
1565 wait_event(mdsc->cap_flushing_wq,
1566 check_capsnap_flush(ceph_inode(inode),
1573 mutex_lock(&mdsc->mutex);
1575 mutex_unlock(&mdsc->mutex);
1577 wait_event(mdsc->cap_flushing_wq,
1578 check_caps_flush(mdsc, want_flush_tid));
1580 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1584 * called under s_mutex
1586 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1587 struct ceph_mds_session *session)
1589 struct ceph_msg *msg = NULL;
1590 struct ceph_mds_cap_release *head;
1591 struct ceph_mds_cap_item *item;
1592 struct ceph_cap *cap;
1593 LIST_HEAD(tmp_list);
1594 int num_cap_releases;
1596 spin_lock(&session->s_cap_lock);
1598 list_splice_init(&session->s_cap_releases, &tmp_list);
1599 num_cap_releases = session->s_num_cap_releases;
1600 session->s_num_cap_releases = 0;
1601 spin_unlock(&session->s_cap_lock);
1603 while (!list_empty(&tmp_list)) {
1605 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1606 PAGE_CACHE_SIZE, GFP_NOFS, false);
1609 head = msg->front.iov_base;
1610 head->num = cpu_to_le32(0);
1611 msg->front.iov_len = sizeof(*head);
1613 cap = list_first_entry(&tmp_list, struct ceph_cap,
1615 list_del(&cap->session_caps);
1618 head = msg->front.iov_base;
1619 le32_add_cpu(&head->num, 1);
1620 item = msg->front.iov_base + msg->front.iov_len;
1621 item->ino = cpu_to_le64(cap->cap_ino);
1622 item->cap_id = cpu_to_le64(cap->cap_id);
1623 item->migrate_seq = cpu_to_le32(cap->mseq);
1624 item->seq = cpu_to_le32(cap->issue_seq);
1625 msg->front.iov_len += sizeof(*item);
1627 ceph_put_cap(mdsc, cap);
1629 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1630 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1631 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1632 ceph_con_send(&session->s_con, msg);
1637 BUG_ON(num_cap_releases != 0);
1639 spin_lock(&session->s_cap_lock);
1640 if (!list_empty(&session->s_cap_releases))
1642 spin_unlock(&session->s_cap_lock);
1645 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1646 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1647 ceph_con_send(&session->s_con, msg);
1651 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1653 spin_lock(&session->s_cap_lock);
1654 list_splice(&tmp_list, &session->s_cap_releases);
1655 session->s_num_cap_releases += num_cap_releases;
1656 spin_unlock(&session->s_cap_lock);
1663 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1666 struct ceph_inode_info *ci = ceph_inode(dir);
1667 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1668 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1669 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1670 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1671 int order, num_entries;
1673 spin_lock(&ci->i_ceph_lock);
1674 num_entries = ci->i_files + ci->i_subdirs;
1675 spin_unlock(&ci->i_ceph_lock);
1676 num_entries = max(num_entries, 1);
1677 num_entries = min(num_entries, opt->max_readdir);
1679 order = get_order(size * num_entries);
1680 while (order >= 0) {
1681 rinfo->dir_in = (void*)__get_free_pages(GFP_KERNEL |
1691 num_entries = (PAGE_SIZE << order) / size;
1692 num_entries = min(num_entries, opt->max_readdir);
1694 rinfo->dir_buf_size = PAGE_SIZE << order;
1695 req->r_num_caps = num_entries + 1;
1696 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1697 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1702 * Create an mds request.
1704 struct ceph_mds_request *
1705 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1707 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1710 return ERR_PTR(-ENOMEM);
1712 mutex_init(&req->r_fill_mutex);
1714 req->r_started = jiffies;
1715 req->r_resend_mds = -1;
1716 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1717 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1719 kref_init(&req->r_kref);
1720 INIT_LIST_HEAD(&req->r_wait);
1721 init_completion(&req->r_completion);
1722 init_completion(&req->r_safe_completion);
1723 INIT_LIST_HEAD(&req->r_unsafe_item);
1725 req->r_stamp = CURRENT_TIME;
1728 req->r_direct_mode = mode;
1733 * return oldest (lowest) request, tid in request tree, 0 if none.
1735 * called under mdsc->mutex.
1737 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1739 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1741 return rb_entry(rb_first(&mdsc->request_tree),
1742 struct ceph_mds_request, r_node);
1745 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1747 return mdsc->oldest_tid;
1751 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1752 * on build_path_from_dentry in fs/cifs/dir.c.
1754 * If @stop_on_nosnap, generate path relative to the first non-snapped
1757 * Encode hidden .snap dirs as a double /, i.e.
1758 * foo/.snap/bar -> foo//bar
1760 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1763 struct dentry *temp;
1769 return ERR_PTR(-EINVAL);
1773 seq = read_seqbegin(&rename_lock);
1775 for (temp = dentry; !IS_ROOT(temp);) {
1776 struct inode *inode = d_inode(temp);
1777 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1778 len++; /* slash only */
1779 else if (stop_on_nosnap && inode &&
1780 ceph_snap(inode) == CEPH_NOSNAP)
1783 len += 1 + temp->d_name.len;
1784 temp = temp->d_parent;
1788 len--; /* no leading '/' */
1790 path = kmalloc(len+1, GFP_NOFS);
1792 return ERR_PTR(-ENOMEM);
1794 path[pos] = 0; /* trailing null */
1796 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1797 struct inode *inode;
1799 spin_lock(&temp->d_lock);
1800 inode = d_inode(temp);
1801 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1802 dout("build_path path+%d: %p SNAPDIR\n",
1804 } else if (stop_on_nosnap && inode &&
1805 ceph_snap(inode) == CEPH_NOSNAP) {
1806 spin_unlock(&temp->d_lock);
1809 pos -= temp->d_name.len;
1811 spin_unlock(&temp->d_lock);
1814 strncpy(path + pos, temp->d_name.name,
1817 spin_unlock(&temp->d_lock);
1820 temp = temp->d_parent;
1823 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1824 pr_err("build_path did not end path lookup where "
1825 "expected, namelen is %d, pos is %d\n", len, pos);
1826 /* presumably this is only possible if racing with a
1827 rename of one of the parent directories (we can not
1828 lock the dentries above us to prevent this, but
1829 retrying should be harmless) */
1834 *base = ceph_ino(d_inode(temp));
1836 dout("build_path on %p %d built %llx '%.*s'\n",
1837 dentry, d_count(dentry), *base, len, path);
1841 static int build_dentry_path(struct dentry *dentry,
1842 const char **ppath, int *ppathlen, u64 *pino,
1847 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1848 *pino = ceph_ino(d_inode(dentry->d_parent));
1849 *ppath = dentry->d_name.name;
1850 *ppathlen = dentry->d_name.len;
1853 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1855 return PTR_ERR(path);
1861 static int build_inode_path(struct inode *inode,
1862 const char **ppath, int *ppathlen, u64 *pino,
1865 struct dentry *dentry;
1868 if (ceph_snap(inode) == CEPH_NOSNAP) {
1869 *pino = ceph_ino(inode);
1873 dentry = d_find_alias(inode);
1874 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1877 return PTR_ERR(path);
1884 * request arguments may be specified via an inode *, a dentry *, or
1885 * an explicit ino+path.
1887 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1888 const char *rpath, u64 rino,
1889 const char **ppath, int *pathlen,
1890 u64 *ino, int *freepath)
1895 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1896 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1898 } else if (rdentry) {
1899 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1900 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1902 } else if (rpath || rino) {
1905 *pathlen = rpath ? strlen(rpath) : 0;
1906 dout(" path %.*s\n", *pathlen, rpath);
1913 * called under mdsc->mutex
1915 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1916 struct ceph_mds_request *req,
1917 int mds, bool drop_cap_releases)
1919 struct ceph_msg *msg;
1920 struct ceph_mds_request_head *head;
1921 const char *path1 = NULL;
1922 const char *path2 = NULL;
1923 u64 ino1 = 0, ino2 = 0;
1924 int pathlen1 = 0, pathlen2 = 0;
1925 int freepath1 = 0, freepath2 = 0;
1931 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1932 req->r_path1, req->r_ino1.ino,
1933 &path1, &pathlen1, &ino1, &freepath1);
1939 ret = set_request_path_attr(NULL, req->r_old_dentry,
1940 req->r_path2, req->r_ino2.ino,
1941 &path2, &pathlen2, &ino2, &freepath2);
1947 len = sizeof(*head) +
1948 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1949 sizeof(struct ceph_timespec);
1951 /* calculate (max) length for cap releases */
1952 len += sizeof(struct ceph_mds_request_release) *
1953 (!!req->r_inode_drop + !!req->r_dentry_drop +
1954 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1955 if (req->r_dentry_drop)
1956 len += req->r_dentry->d_name.len;
1957 if (req->r_old_dentry_drop)
1958 len += req->r_old_dentry->d_name.len;
1960 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1962 msg = ERR_PTR(-ENOMEM);
1966 msg->hdr.version = cpu_to_le16(2);
1967 msg->hdr.tid = cpu_to_le64(req->r_tid);
1969 head = msg->front.iov_base;
1970 p = msg->front.iov_base + sizeof(*head);
1971 end = msg->front.iov_base + msg->front.iov_len;
1973 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1974 head->op = cpu_to_le32(req->r_op);
1975 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1976 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1977 head->args = req->r_args;
1979 ceph_encode_filepath(&p, end, ino1, path1);
1980 ceph_encode_filepath(&p, end, ino2, path2);
1982 /* make note of release offset, in case we need to replay */
1983 req->r_request_release_offset = p - msg->front.iov_base;
1987 if (req->r_inode_drop)
1988 releases += ceph_encode_inode_release(&p,
1989 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1990 mds, req->r_inode_drop, req->r_inode_unless, 0);
1991 if (req->r_dentry_drop)
1992 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1993 mds, req->r_dentry_drop, req->r_dentry_unless);
1994 if (req->r_old_dentry_drop)
1995 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1996 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1997 if (req->r_old_inode_drop)
1998 releases += ceph_encode_inode_release(&p,
1999 d_inode(req->r_old_dentry),
2000 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2002 if (drop_cap_releases) {
2004 p = msg->front.iov_base + req->r_request_release_offset;
2007 head->num_releases = cpu_to_le16(releases);
2011 struct ceph_timespec ts;
2012 ceph_encode_timespec(&ts, &req->r_stamp);
2013 ceph_encode_copy(&p, &ts, sizeof(ts));
2017 msg->front.iov_len = p - msg->front.iov_base;
2018 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2020 if (req->r_pagelist) {
2021 struct ceph_pagelist *pagelist = req->r_pagelist;
2022 atomic_inc(&pagelist->refcnt);
2023 ceph_msg_data_add_pagelist(msg, pagelist);
2024 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2026 msg->hdr.data_len = 0;
2029 msg->hdr.data_off = cpu_to_le16(0);
2033 kfree((char *)path2);
2036 kfree((char *)path1);
2042 * called under mdsc->mutex if error, under no mutex if
2045 static void complete_request(struct ceph_mds_client *mdsc,
2046 struct ceph_mds_request *req)
2048 if (req->r_callback)
2049 req->r_callback(mdsc, req);
2051 complete_all(&req->r_completion);
2055 * called under mdsc->mutex
2057 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2058 struct ceph_mds_request *req,
2059 int mds, bool drop_cap_releases)
2061 struct ceph_mds_request_head *rhead;
2062 struct ceph_msg *msg;
2067 struct ceph_cap *cap =
2068 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2071 req->r_sent_on_mseq = cap->mseq;
2073 req->r_sent_on_mseq = -1;
2075 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2076 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2078 if (req->r_got_unsafe) {
2081 * Replay. Do not regenerate message (and rebuild
2082 * paths, etc.); just use the original message.
2083 * Rebuilding paths will break for renames because
2084 * d_move mangles the src name.
2086 msg = req->r_request;
2087 rhead = msg->front.iov_base;
2089 flags = le32_to_cpu(rhead->flags);
2090 flags |= CEPH_MDS_FLAG_REPLAY;
2091 rhead->flags = cpu_to_le32(flags);
2093 if (req->r_target_inode)
2094 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2096 rhead->num_retry = req->r_attempts - 1;
2098 /* remove cap/dentry releases from message */
2099 rhead->num_releases = 0;
2102 p = msg->front.iov_base + req->r_request_release_offset;
2104 struct ceph_timespec ts;
2105 ceph_encode_timespec(&ts, &req->r_stamp);
2106 ceph_encode_copy(&p, &ts, sizeof(ts));
2109 msg->front.iov_len = p - msg->front.iov_base;
2110 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2114 if (req->r_request) {
2115 ceph_msg_put(req->r_request);
2116 req->r_request = NULL;
2118 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2120 req->r_err = PTR_ERR(msg);
2121 return PTR_ERR(msg);
2123 req->r_request = msg;
2125 rhead = msg->front.iov_base;
2126 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2127 if (req->r_got_unsafe)
2128 flags |= CEPH_MDS_FLAG_REPLAY;
2129 if (req->r_locked_dir)
2130 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2131 rhead->flags = cpu_to_le32(flags);
2132 rhead->num_fwd = req->r_num_fwd;
2133 rhead->num_retry = req->r_attempts - 1;
2136 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2141 * send request, or put it on the appropriate wait list.
2143 static int __do_request(struct ceph_mds_client *mdsc,
2144 struct ceph_mds_request *req)
2146 struct ceph_mds_session *session = NULL;
2150 if (req->r_err || req->r_got_result) {
2152 __unregister_request(mdsc, req);
2156 if (req->r_timeout &&
2157 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2158 dout("do_request timed out\n");
2162 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2163 dout("do_request forced umount\n");
2168 put_request_session(req);
2170 mds = __choose_mds(mdsc, req);
2172 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2173 dout("do_request no mds or not active, waiting for map\n");
2174 list_add(&req->r_wait, &mdsc->waiting_for_map);
2178 /* get, open session */
2179 session = __ceph_lookup_mds_session(mdsc, mds);
2181 session = register_session(mdsc, mds);
2182 if (IS_ERR(session)) {
2183 err = PTR_ERR(session);
2187 req->r_session = get_session(session);
2189 dout("do_request mds%d session %p state %s\n", mds, session,
2190 ceph_session_state_name(session->s_state));
2191 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2192 session->s_state != CEPH_MDS_SESSION_HUNG) {
2193 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2194 session->s_state == CEPH_MDS_SESSION_CLOSING)
2195 __open_session(mdsc, session);
2196 list_add(&req->r_wait, &session->s_waiting);
2201 req->r_resend_mds = -1; /* forget any previous mds hint */
2203 if (req->r_request_started == 0) /* note request start time */
2204 req->r_request_started = jiffies;
2206 err = __prepare_send_request(mdsc, req, mds, false);
2208 ceph_msg_get(req->r_request);
2209 ceph_con_send(&session->s_con, req->r_request);
2213 ceph_put_mds_session(session);
2216 dout("__do_request early error %d\n", err);
2218 complete_request(mdsc, req);
2219 __unregister_request(mdsc, req);
2226 * called under mdsc->mutex
2228 static void __wake_requests(struct ceph_mds_client *mdsc,
2229 struct list_head *head)
2231 struct ceph_mds_request *req;
2232 LIST_HEAD(tmp_list);
2234 list_splice_init(head, &tmp_list);
2236 while (!list_empty(&tmp_list)) {
2237 req = list_entry(tmp_list.next,
2238 struct ceph_mds_request, r_wait);
2239 list_del_init(&req->r_wait);
2240 dout(" wake request %p tid %llu\n", req, req->r_tid);
2241 __do_request(mdsc, req);
2246 * Wake up threads with requests pending for @mds, so that they can
2247 * resubmit their requests to a possibly different mds.
2249 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2251 struct ceph_mds_request *req;
2252 struct rb_node *p = rb_first(&mdsc->request_tree);
2254 dout("kick_requests mds%d\n", mds);
2256 req = rb_entry(p, struct ceph_mds_request, r_node);
2258 if (req->r_got_unsafe)
2260 if (req->r_attempts > 0)
2261 continue; /* only new requests */
2262 if (req->r_session &&
2263 req->r_session->s_mds == mds) {
2264 dout(" kicking tid %llu\n", req->r_tid);
2265 list_del_init(&req->r_wait);
2266 __do_request(mdsc, req);
2271 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2272 struct ceph_mds_request *req)
2274 dout("submit_request on %p\n", req);
2275 mutex_lock(&mdsc->mutex);
2276 __register_request(mdsc, req, NULL);
2277 __do_request(mdsc, req);
2278 mutex_unlock(&mdsc->mutex);
2282 * Synchrously perform an mds request. Take care of all of the
2283 * session setup, forwarding, retry details.
2285 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2287 struct ceph_mds_request *req)
2291 dout("do_request on %p\n", req);
2293 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2295 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2296 if (req->r_locked_dir)
2297 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2298 if (req->r_old_dentry_dir)
2299 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2303 mutex_lock(&mdsc->mutex);
2304 __register_request(mdsc, req, dir);
2305 __do_request(mdsc, req);
2313 mutex_unlock(&mdsc->mutex);
2314 dout("do_request waiting\n");
2315 if (!req->r_timeout && req->r_wait_for_completion) {
2316 err = req->r_wait_for_completion(mdsc, req);
2318 long timeleft = wait_for_completion_killable_timeout(
2320 ceph_timeout_jiffies(req->r_timeout));
2324 err = -EIO; /* timed out */
2326 err = timeleft; /* killed */
2328 dout("do_request waited, got %d\n", err);
2329 mutex_lock(&mdsc->mutex);
2331 /* only abort if we didn't race with a real reply */
2332 if (req->r_got_result) {
2333 err = le32_to_cpu(req->r_reply_info.head->result);
2334 } else if (err < 0) {
2335 dout("aborted request %lld with %d\n", req->r_tid, err);
2338 * ensure we aren't running concurrently with
2339 * ceph_fill_trace or ceph_readdir_prepopulate, which
2340 * rely on locks (dir mutex) held by our caller.
2342 mutex_lock(&req->r_fill_mutex);
2344 req->r_aborted = true;
2345 mutex_unlock(&req->r_fill_mutex);
2347 if (req->r_locked_dir &&
2348 (req->r_op & CEPH_MDS_OP_WRITE))
2349 ceph_invalidate_dir_request(req);
2355 mutex_unlock(&mdsc->mutex);
2356 dout("do_request %p done, result %d\n", req, err);
2361 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2362 * namespace request.
2364 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2366 struct inode *inode = req->r_locked_dir;
2368 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2370 ceph_dir_clear_complete(inode);
2372 ceph_invalidate_dentry_lease(req->r_dentry);
2373 if (req->r_old_dentry)
2374 ceph_invalidate_dentry_lease(req->r_old_dentry);
2380 * We take the session mutex and parse and process the reply immediately.
2381 * This preserves the logical ordering of replies, capabilities, etc., sent
2382 * by the MDS as they are applied to our local cache.
2384 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2386 struct ceph_mds_client *mdsc = session->s_mdsc;
2387 struct ceph_mds_request *req;
2388 struct ceph_mds_reply_head *head = msg->front.iov_base;
2389 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2390 struct ceph_snap_realm *realm;
2393 int mds = session->s_mds;
2395 if (msg->front.iov_len < sizeof(*head)) {
2396 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2401 /* get request, session */
2402 tid = le64_to_cpu(msg->hdr.tid);
2403 mutex_lock(&mdsc->mutex);
2404 req = __lookup_request(mdsc, tid);
2406 dout("handle_reply on unknown tid %llu\n", tid);
2407 mutex_unlock(&mdsc->mutex);
2410 dout("handle_reply %p\n", req);
2412 /* correct session? */
2413 if (req->r_session != session) {
2414 pr_err("mdsc_handle_reply got %llu on session mds%d"
2415 " not mds%d\n", tid, session->s_mds,
2416 req->r_session ? req->r_session->s_mds : -1);
2417 mutex_unlock(&mdsc->mutex);
2422 if ((req->r_got_unsafe && !head->safe) ||
2423 (req->r_got_safe && head->safe)) {
2424 pr_warn("got a dup %s reply on %llu from mds%d\n",
2425 head->safe ? "safe" : "unsafe", tid, mds);
2426 mutex_unlock(&mdsc->mutex);
2429 if (req->r_got_safe) {
2430 pr_warn("got unsafe after safe on %llu from mds%d\n",
2432 mutex_unlock(&mdsc->mutex);
2436 result = le32_to_cpu(head->result);
2440 * if we're not talking to the authority, send to them
2441 * if the authority has changed while we weren't looking,
2442 * send to new authority
2443 * Otherwise we just have to return an ESTALE
2445 if (result == -ESTALE) {
2446 dout("got ESTALE on request %llu", req->r_tid);
2447 req->r_resend_mds = -1;
2448 if (req->r_direct_mode != USE_AUTH_MDS) {
2449 dout("not using auth, setting for that now");
2450 req->r_direct_mode = USE_AUTH_MDS;
2451 __do_request(mdsc, req);
2452 mutex_unlock(&mdsc->mutex);
2455 int mds = __choose_mds(mdsc, req);
2456 if (mds >= 0 && mds != req->r_session->s_mds) {
2457 dout("but auth changed, so resending");
2458 __do_request(mdsc, req);
2459 mutex_unlock(&mdsc->mutex);
2463 dout("have to return ESTALE on request %llu", req->r_tid);
2468 req->r_got_safe = true;
2469 __unregister_request(mdsc, req);
2471 if (req->r_got_unsafe) {
2473 * We already handled the unsafe response, now do the
2474 * cleanup. No need to examine the response; the MDS
2475 * doesn't include any result info in the safe
2476 * response. And even if it did, there is nothing
2477 * useful we could do with a revised return value.
2479 dout("got safe reply %llu, mds%d\n", tid, mds);
2480 list_del_init(&req->r_unsafe_item);
2482 /* last unsafe request during umount? */
2483 if (mdsc->stopping && !__get_oldest_req(mdsc))
2484 complete_all(&mdsc->safe_umount_waiters);
2485 mutex_unlock(&mdsc->mutex);
2489 req->r_got_unsafe = true;
2490 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2491 if (req->r_unsafe_dir) {
2492 struct ceph_inode_info *ci =
2493 ceph_inode(req->r_unsafe_dir);
2494 spin_lock(&ci->i_unsafe_lock);
2495 list_add_tail(&req->r_unsafe_dir_item,
2496 &ci->i_unsafe_dirops);
2497 spin_unlock(&ci->i_unsafe_lock);
2501 dout("handle_reply tid %lld result %d\n", tid, result);
2502 rinfo = &req->r_reply_info;
2503 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2504 mutex_unlock(&mdsc->mutex);
2506 mutex_lock(&session->s_mutex);
2508 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2515 if (rinfo->snapblob_len) {
2516 down_write(&mdsc->snap_rwsem);
2517 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2518 rinfo->snapblob + rinfo->snapblob_len,
2519 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2521 downgrade_write(&mdsc->snap_rwsem);
2523 down_read(&mdsc->snap_rwsem);
2526 /* insert trace into our cache */
2527 mutex_lock(&req->r_fill_mutex);
2528 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2530 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2531 req->r_op == CEPH_MDS_OP_LSSNAP))
2532 ceph_readdir_prepopulate(req, req->r_session);
2533 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2535 mutex_unlock(&req->r_fill_mutex);
2537 up_read(&mdsc->snap_rwsem);
2539 ceph_put_snap_realm(mdsc, realm);
2541 if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2542 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2543 spin_lock(&ci->i_unsafe_lock);
2544 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2545 spin_unlock(&ci->i_unsafe_lock);
2548 mutex_lock(&mdsc->mutex);
2549 if (!req->r_aborted) {
2553 req->r_reply = ceph_msg_get(msg);
2554 req->r_got_result = true;
2557 dout("reply arrived after request %lld was aborted\n", tid);
2559 mutex_unlock(&mdsc->mutex);
2561 mutex_unlock(&session->s_mutex);
2563 /* kick calling process */
2564 complete_request(mdsc, req);
2566 ceph_mdsc_put_request(req);
2573 * handle mds notification that our request has been forwarded.
2575 static void handle_forward(struct ceph_mds_client *mdsc,
2576 struct ceph_mds_session *session,
2577 struct ceph_msg *msg)
2579 struct ceph_mds_request *req;
2580 u64 tid = le64_to_cpu(msg->hdr.tid);
2584 void *p = msg->front.iov_base;
2585 void *end = p + msg->front.iov_len;
2587 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2588 next_mds = ceph_decode_32(&p);
2589 fwd_seq = ceph_decode_32(&p);
2591 mutex_lock(&mdsc->mutex);
2592 req = __lookup_request(mdsc, tid);
2594 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2595 goto out; /* dup reply? */
2598 if (req->r_aborted) {
2599 dout("forward tid %llu aborted, unregistering\n", tid);
2600 __unregister_request(mdsc, req);
2601 } else if (fwd_seq <= req->r_num_fwd) {
2602 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2603 tid, next_mds, req->r_num_fwd, fwd_seq);
2605 /* resend. forward race not possible; mds would drop */
2606 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2608 BUG_ON(req->r_got_result);
2609 req->r_attempts = 0;
2610 req->r_num_fwd = fwd_seq;
2611 req->r_resend_mds = next_mds;
2612 put_request_session(req);
2613 __do_request(mdsc, req);
2615 ceph_mdsc_put_request(req);
2617 mutex_unlock(&mdsc->mutex);
2621 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2625 * handle a mds session control message
2627 static void handle_session(struct ceph_mds_session *session,
2628 struct ceph_msg *msg)
2630 struct ceph_mds_client *mdsc = session->s_mdsc;
2633 int mds = session->s_mds;
2634 struct ceph_mds_session_head *h = msg->front.iov_base;
2638 if (msg->front.iov_len != sizeof(*h))
2640 op = le32_to_cpu(h->op);
2641 seq = le64_to_cpu(h->seq);
2643 mutex_lock(&mdsc->mutex);
2644 if (op == CEPH_SESSION_CLOSE)
2645 __unregister_session(mdsc, session);
2646 /* FIXME: this ttl calculation is generous */
2647 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2648 mutex_unlock(&mdsc->mutex);
2650 mutex_lock(&session->s_mutex);
2652 dout("handle_session mds%d %s %p state %s seq %llu\n",
2653 mds, ceph_session_op_name(op), session,
2654 ceph_session_state_name(session->s_state), seq);
2656 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2657 session->s_state = CEPH_MDS_SESSION_OPEN;
2658 pr_info("mds%d came back\n", session->s_mds);
2662 case CEPH_SESSION_OPEN:
2663 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2664 pr_info("mds%d reconnect success\n", session->s_mds);
2665 session->s_state = CEPH_MDS_SESSION_OPEN;
2666 renewed_caps(mdsc, session, 0);
2669 __close_session(mdsc, session);
2672 case CEPH_SESSION_RENEWCAPS:
2673 if (session->s_renew_seq == seq)
2674 renewed_caps(mdsc, session, 1);
2677 case CEPH_SESSION_CLOSE:
2678 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2679 pr_info("mds%d reconnect denied\n", session->s_mds);
2680 cleanup_session_requests(mdsc, session);
2681 remove_session_caps(session);
2682 wake = 2; /* for good measure */
2683 wake_up_all(&mdsc->session_close_wq);
2686 case CEPH_SESSION_STALE:
2687 pr_info("mds%d caps went stale, renewing\n",
2689 spin_lock(&session->s_gen_ttl_lock);
2690 session->s_cap_gen++;
2691 session->s_cap_ttl = jiffies - 1;
2692 spin_unlock(&session->s_gen_ttl_lock);
2693 send_renew_caps(mdsc, session);
2696 case CEPH_SESSION_RECALL_STATE:
2697 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2700 case CEPH_SESSION_FLUSHMSG:
2701 send_flushmsg_ack(mdsc, session, seq);
2704 case CEPH_SESSION_FORCE_RO:
2705 dout("force_session_readonly %p\n", session);
2706 spin_lock(&session->s_cap_lock);
2707 session->s_readonly = true;
2708 spin_unlock(&session->s_cap_lock);
2709 wake_up_session_caps(session, 0);
2713 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2717 mutex_unlock(&session->s_mutex);
2719 mutex_lock(&mdsc->mutex);
2720 __wake_requests(mdsc, &session->s_waiting);
2722 kick_requests(mdsc, mds);
2723 mutex_unlock(&mdsc->mutex);
2728 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2729 (int)msg->front.iov_len);
2736 * called under session->mutex.
2738 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2739 struct ceph_mds_session *session)
2741 struct ceph_mds_request *req, *nreq;
2745 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2747 mutex_lock(&mdsc->mutex);
2748 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2749 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2751 ceph_msg_get(req->r_request);
2752 ceph_con_send(&session->s_con, req->r_request);
2757 * also re-send old requests when MDS enters reconnect stage. So that MDS
2758 * can process completed request in clientreplay stage.
2760 p = rb_first(&mdsc->request_tree);
2762 req = rb_entry(p, struct ceph_mds_request, r_node);
2764 if (req->r_got_unsafe)
2766 if (req->r_attempts == 0)
2767 continue; /* only old requests */
2768 if (req->r_session &&
2769 req->r_session->s_mds == session->s_mds) {
2770 err = __prepare_send_request(mdsc, req,
2771 session->s_mds, true);
2773 ceph_msg_get(req->r_request);
2774 ceph_con_send(&session->s_con, req->r_request);
2778 mutex_unlock(&mdsc->mutex);
2782 * Encode information about a cap for a reconnect with the MDS.
2784 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2788 struct ceph_mds_cap_reconnect v2;
2789 struct ceph_mds_cap_reconnect_v1 v1;
2792 struct ceph_inode_info *ci;
2793 struct ceph_reconnect_state *recon_state = arg;
2794 struct ceph_pagelist *pagelist = recon_state->pagelist;
2798 struct dentry *dentry;
2802 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2803 inode, ceph_vinop(inode), cap, cap->cap_id,
2804 ceph_cap_string(cap->issued));
2805 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2809 dentry = d_find_alias(inode);
2811 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2813 err = PTR_ERR(path);
2820 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2824 spin_lock(&ci->i_ceph_lock);
2825 cap->seq = 0; /* reset cap seq */
2826 cap->issue_seq = 0; /* and issue_seq */
2827 cap->mseq = 0; /* and migrate_seq */
2828 cap->cap_gen = cap->session->s_cap_gen;
2830 if (recon_state->flock) {
2831 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2832 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2833 rec.v2.issued = cpu_to_le32(cap->issued);
2834 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2835 rec.v2.pathbase = cpu_to_le64(pathbase);
2836 rec.v2.flock_len = 0;
2837 reclen = sizeof(rec.v2);
2839 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2840 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2841 rec.v1.issued = cpu_to_le32(cap->issued);
2842 rec.v1.size = cpu_to_le64(inode->i_size);
2843 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2844 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2845 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2846 rec.v1.pathbase = cpu_to_le64(pathbase);
2847 reclen = sizeof(rec.v1);
2849 spin_unlock(&ci->i_ceph_lock);
2851 if (recon_state->flock) {
2852 int num_fcntl_locks, num_flock_locks;
2853 struct ceph_filelock *flocks;
2856 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2857 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2858 sizeof(struct ceph_filelock), GFP_NOFS);
2863 err = ceph_encode_locks_to_buffer(inode, flocks,
2873 * number of encoded locks is stable, so copy to pagelist
2875 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2876 (num_fcntl_locks+num_flock_locks) *
2877 sizeof(struct ceph_filelock));
2878 err = ceph_pagelist_append(pagelist, &rec, reclen);
2880 err = ceph_locks_to_pagelist(flocks, pagelist,
2885 err = ceph_pagelist_append(pagelist, &rec, reclen);
2888 recon_state->nr_caps++;
2898 * If an MDS fails and recovers, clients need to reconnect in order to
2899 * reestablish shared state. This includes all caps issued through
2900 * this session _and_ the snap_realm hierarchy. Because it's not
2901 * clear which snap realms the mds cares about, we send everything we
2902 * know about.. that ensures we'll then get any new info the
2903 * recovering MDS might have.
2905 * This is a relatively heavyweight operation, but it's rare.
2907 * called with mdsc->mutex held.
2909 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2910 struct ceph_mds_session *session)
2912 struct ceph_msg *reply;
2914 int mds = session->s_mds;
2917 struct ceph_pagelist *pagelist;
2918 struct ceph_reconnect_state recon_state;
2920 pr_info("mds%d reconnect start\n", mds);
2922 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2924 goto fail_nopagelist;
2925 ceph_pagelist_init(pagelist);
2927 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2931 mutex_lock(&session->s_mutex);
2932 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2935 dout("session %p state %s\n", session,
2936 ceph_session_state_name(session->s_state));
2938 spin_lock(&session->s_gen_ttl_lock);
2939 session->s_cap_gen++;
2940 spin_unlock(&session->s_gen_ttl_lock);
2942 spin_lock(&session->s_cap_lock);
2943 /* don't know if session is readonly */
2944 session->s_readonly = 0;
2946 * notify __ceph_remove_cap() that we are composing cap reconnect.
2947 * If a cap get released before being added to the cap reconnect,
2948 * __ceph_remove_cap() should skip queuing cap release.
2950 session->s_cap_reconnect = 1;
2951 /* drop old cap expires; we're about to reestablish that state */
2952 cleanup_cap_releases(mdsc, session);
2954 /* trim unused caps to reduce MDS's cache rejoin time */
2955 if (mdsc->fsc->sb->s_root)
2956 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2958 ceph_con_close(&session->s_con);
2959 ceph_con_open(&session->s_con,
2960 CEPH_ENTITY_TYPE_MDS, mds,
2961 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2963 /* replay unsafe requests */
2964 replay_unsafe_requests(mdsc, session);
2966 down_read(&mdsc->snap_rwsem);
2968 /* traverse this session's caps */
2969 s_nr_caps = session->s_nr_caps;
2970 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2974 recon_state.nr_caps = 0;
2975 recon_state.pagelist = pagelist;
2976 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2977 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2981 spin_lock(&session->s_cap_lock);
2982 session->s_cap_reconnect = 0;
2983 spin_unlock(&session->s_cap_lock);
2986 * snaprealms. we provide mds with the ino, seq (version), and
2987 * parent for all of our realms. If the mds has any newer info,
2990 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2991 struct ceph_snap_realm *realm =
2992 rb_entry(p, struct ceph_snap_realm, node);
2993 struct ceph_mds_snaprealm_reconnect sr_rec;
2995 dout(" adding snap realm %llx seq %lld parent %llx\n",
2996 realm->ino, realm->seq, realm->parent_ino);
2997 sr_rec.ino = cpu_to_le64(realm->ino);
2998 sr_rec.seq = cpu_to_le64(realm->seq);
2999 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3000 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3005 if (recon_state.flock)
3006 reply->hdr.version = cpu_to_le16(2);
3008 /* raced with cap release? */
3009 if (s_nr_caps != recon_state.nr_caps) {
3010 struct page *page = list_first_entry(&pagelist->head,
3012 __le32 *addr = kmap_atomic(page);
3013 *addr = cpu_to_le32(recon_state.nr_caps);
3014 kunmap_atomic(addr);
3017 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3018 ceph_msg_data_add_pagelist(reply, pagelist);
3020 ceph_early_kick_flushing_caps(mdsc, session);
3022 ceph_con_send(&session->s_con, reply);
3024 mutex_unlock(&session->s_mutex);
3026 mutex_lock(&mdsc->mutex);
3027 __wake_requests(mdsc, &session->s_waiting);
3028 mutex_unlock(&mdsc->mutex);
3030 up_read(&mdsc->snap_rwsem);
3034 ceph_msg_put(reply);
3035 up_read(&mdsc->snap_rwsem);
3036 mutex_unlock(&session->s_mutex);
3038 ceph_pagelist_release(pagelist);
3040 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3046 * compare old and new mdsmaps, kicking requests
3047 * and closing out old connections as necessary
3049 * called under mdsc->mutex.
3051 static void check_new_map(struct ceph_mds_client *mdsc,
3052 struct ceph_mdsmap *newmap,
3053 struct ceph_mdsmap *oldmap)
3056 int oldstate, newstate;
3057 struct ceph_mds_session *s;
3059 dout("check_new_map new %u old %u\n",
3060 newmap->m_epoch, oldmap->m_epoch);
3062 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3063 if (mdsc->sessions[i] == NULL)
3065 s = mdsc->sessions[i];
3066 oldstate = ceph_mdsmap_get_state(oldmap, i);
3067 newstate = ceph_mdsmap_get_state(newmap, i);
3069 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3070 i, ceph_mds_state_name(oldstate),
3071 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3072 ceph_mds_state_name(newstate),
3073 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3074 ceph_session_state_name(s->s_state));
3076 if (i >= newmap->m_max_mds ||
3077 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3078 ceph_mdsmap_get_addr(newmap, i),
3079 sizeof(struct ceph_entity_addr))) {
3080 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3081 /* the session never opened, just close it
3083 __wake_requests(mdsc, &s->s_waiting);
3084 __unregister_session(mdsc, s);
3087 mutex_unlock(&mdsc->mutex);
3088 mutex_lock(&s->s_mutex);
3089 mutex_lock(&mdsc->mutex);
3090 ceph_con_close(&s->s_con);
3091 mutex_unlock(&s->s_mutex);
3092 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3094 } else if (oldstate == newstate) {
3095 continue; /* nothing new with this mds */
3101 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3102 newstate >= CEPH_MDS_STATE_RECONNECT) {
3103 mutex_unlock(&mdsc->mutex);
3104 send_mds_reconnect(mdsc, s);
3105 mutex_lock(&mdsc->mutex);
3109 * kick request on any mds that has gone active.
3111 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3112 newstate >= CEPH_MDS_STATE_ACTIVE) {
3113 if (oldstate != CEPH_MDS_STATE_CREATING &&
3114 oldstate != CEPH_MDS_STATE_STARTING)
3115 pr_info("mds%d recovery completed\n", s->s_mds);
3116 kick_requests(mdsc, i);
3117 ceph_kick_flushing_caps(mdsc, s);
3118 wake_up_session_caps(s, 1);
3122 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3123 s = mdsc->sessions[i];
3126 if (!ceph_mdsmap_is_laggy(newmap, i))
3128 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3129 s->s_state == CEPH_MDS_SESSION_HUNG ||
3130 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3131 dout(" connecting to export targets of laggy mds%d\n",
3133 __open_export_target_sessions(mdsc, s);
3145 * caller must hold session s_mutex, dentry->d_lock
3147 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3149 struct ceph_dentry_info *di = ceph_dentry(dentry);
3151 ceph_put_mds_session(di->lease_session);
3152 di->lease_session = NULL;
3155 static void handle_lease(struct ceph_mds_client *mdsc,
3156 struct ceph_mds_session *session,
3157 struct ceph_msg *msg)
3159 struct super_block *sb = mdsc->fsc->sb;
3160 struct inode *inode;
3161 struct dentry *parent, *dentry;
3162 struct ceph_dentry_info *di;
3163 int mds = session->s_mds;
3164 struct ceph_mds_lease *h = msg->front.iov_base;
3166 struct ceph_vino vino;
3170 dout("handle_lease from mds%d\n", mds);
3173 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3175 vino.ino = le64_to_cpu(h->ino);
3176 vino.snap = CEPH_NOSNAP;
3177 seq = le32_to_cpu(h->seq);
3178 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3179 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3180 if (dname.len != get_unaligned_le32(h+1))
3184 inode = ceph_find_inode(sb, vino);
3185 dout("handle_lease %s, ino %llx %p %.*s\n",
3186 ceph_lease_op_name(h->action), vino.ino, inode,
3187 dname.len, dname.name);
3189 mutex_lock(&session->s_mutex);
3192 if (inode == NULL) {
3193 dout("handle_lease no inode %llx\n", vino.ino);
3198 parent = d_find_alias(inode);
3200 dout("no parent dentry on inode %p\n", inode);
3202 goto release; /* hrm... */
3204 dname.hash = full_name_hash(dname.name, dname.len);
3205 dentry = d_lookup(parent, &dname);
3210 spin_lock(&dentry->d_lock);
3211 di = ceph_dentry(dentry);
3212 switch (h->action) {
3213 case CEPH_MDS_LEASE_REVOKE:
3214 if (di->lease_session == session) {
3215 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3216 h->seq = cpu_to_le32(di->lease_seq);
3217 __ceph_mdsc_drop_dentry_lease(dentry);
3222 case CEPH_MDS_LEASE_RENEW:
3223 if (di->lease_session == session &&
3224 di->lease_gen == session->s_cap_gen &&
3225 di->lease_renew_from &&
3226 di->lease_renew_after == 0) {
3227 unsigned long duration =
3228 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3230 di->lease_seq = seq;
3231 dentry->d_time = di->lease_renew_from + duration;
3232 di->lease_renew_after = di->lease_renew_from +
3234 di->lease_renew_from = 0;
3238 spin_unlock(&dentry->d_lock);
3245 /* let's just reuse the same message */
3246 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3248 ceph_con_send(&session->s_con, msg);
3252 mutex_unlock(&session->s_mutex);
3256 pr_err("corrupt lease message\n");
3260 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3261 struct inode *inode,
3262 struct dentry *dentry, char action,
3265 struct ceph_msg *msg;
3266 struct ceph_mds_lease *lease;
3267 int len = sizeof(*lease) + sizeof(u32);
3270 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3271 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3272 dnamelen = dentry->d_name.len;
3275 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3278 lease = msg->front.iov_base;
3279 lease->action = action;
3280 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3281 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3282 lease->seq = cpu_to_le32(seq);
3283 put_unaligned_le32(dnamelen, lease + 1);
3284 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3287 * if this is a preemptive lease RELEASE, no need to
3288 * flush request stream, since the actual request will
3291 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3293 ceph_con_send(&session->s_con, msg);
3297 * Preemptively release a lease we expect to invalidate anyway.
3298 * Pass @inode always, @dentry is optional.
3300 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3301 struct dentry *dentry)
3303 struct ceph_dentry_info *di;
3304 struct ceph_mds_session *session;
3307 BUG_ON(inode == NULL);
3308 BUG_ON(dentry == NULL);
3310 /* is dentry lease valid? */
3311 spin_lock(&dentry->d_lock);
3312 di = ceph_dentry(dentry);
3313 if (!di || !di->lease_session ||
3314 di->lease_session->s_mds < 0 ||
3315 di->lease_gen != di->lease_session->s_cap_gen ||
3316 !time_before(jiffies, dentry->d_time)) {
3317 dout("lease_release inode %p dentry %p -- "
3320 spin_unlock(&dentry->d_lock);
3324 /* we do have a lease on this dentry; note mds and seq */
3325 session = ceph_get_mds_session(di->lease_session);
3326 seq = di->lease_seq;
3327 __ceph_mdsc_drop_dentry_lease(dentry);
3328 spin_unlock(&dentry->d_lock);
3330 dout("lease_release inode %p dentry %p to mds%d\n",
3331 inode, dentry, session->s_mds);
3332 ceph_mdsc_lease_send_msg(session, inode, dentry,
3333 CEPH_MDS_LEASE_RELEASE, seq);
3334 ceph_put_mds_session(session);
3338 * drop all leases (and dentry refs) in preparation for umount
3340 static void drop_leases(struct ceph_mds_client *mdsc)
3344 dout("drop_leases\n");
3345 mutex_lock(&mdsc->mutex);
3346 for (i = 0; i < mdsc->max_sessions; i++) {
3347 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3350 mutex_unlock(&mdsc->mutex);
3351 mutex_lock(&s->s_mutex);
3352 mutex_unlock(&s->s_mutex);
3353 ceph_put_mds_session(s);
3354 mutex_lock(&mdsc->mutex);
3356 mutex_unlock(&mdsc->mutex);
3362 * delayed work -- periodically trim expired leases, renew caps with mds
3364 static void schedule_delayed(struct ceph_mds_client *mdsc)
3367 unsigned hz = round_jiffies_relative(HZ * delay);
3368 schedule_delayed_work(&mdsc->delayed_work, hz);
3371 static void delayed_work(struct work_struct *work)
3374 struct ceph_mds_client *mdsc =
3375 container_of(work, struct ceph_mds_client, delayed_work.work);
3379 dout("mdsc delayed_work\n");
3380 ceph_check_delayed_caps(mdsc);
3382 mutex_lock(&mdsc->mutex);
3383 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3384 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3385 mdsc->last_renew_caps);
3387 mdsc->last_renew_caps = jiffies;
3389 for (i = 0; i < mdsc->max_sessions; i++) {
3390 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3393 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3394 dout("resending session close request for mds%d\n",
3396 request_close_session(mdsc, s);
3397 ceph_put_mds_session(s);
3400 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3401 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3402 s->s_state = CEPH_MDS_SESSION_HUNG;
3403 pr_info("mds%d hung\n", s->s_mds);
3406 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3407 /* this mds is failed or recovering, just wait */
3408 ceph_put_mds_session(s);
3411 mutex_unlock(&mdsc->mutex);
3413 mutex_lock(&s->s_mutex);
3415 send_renew_caps(mdsc, s);
3417 ceph_con_keepalive(&s->s_con);
3418 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3419 s->s_state == CEPH_MDS_SESSION_HUNG)
3420 ceph_send_cap_releases(mdsc, s);
3421 mutex_unlock(&s->s_mutex);
3422 ceph_put_mds_session(s);
3424 mutex_lock(&mdsc->mutex);
3426 mutex_unlock(&mdsc->mutex);
3428 schedule_delayed(mdsc);
3431 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3434 struct ceph_mds_client *mdsc;
3436 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3441 mutex_init(&mdsc->mutex);
3442 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3443 if (mdsc->mdsmap == NULL) {
3448 init_completion(&mdsc->safe_umount_waiters);
3449 init_waitqueue_head(&mdsc->session_close_wq);
3450 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3451 mdsc->sessions = NULL;
3452 atomic_set(&mdsc->num_sessions, 0);
3453 mdsc->max_sessions = 0;
3455 mdsc->last_snap_seq = 0;
3456 init_rwsem(&mdsc->snap_rwsem);
3457 mdsc->snap_realms = RB_ROOT;
3458 INIT_LIST_HEAD(&mdsc->snap_empty);
3459 spin_lock_init(&mdsc->snap_empty_lock);
3461 mdsc->oldest_tid = 0;
3462 mdsc->request_tree = RB_ROOT;
3463 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3464 mdsc->last_renew_caps = jiffies;
3465 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3466 spin_lock_init(&mdsc->cap_delay_lock);
3467 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3468 spin_lock_init(&mdsc->snap_flush_lock);
3469 mdsc->last_cap_flush_tid = 1;
3470 mdsc->cap_flush_tree = RB_ROOT;
3471 INIT_LIST_HEAD(&mdsc->cap_dirty);
3472 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3473 mdsc->num_cap_flushing = 0;
3474 spin_lock_init(&mdsc->cap_dirty_lock);
3475 init_waitqueue_head(&mdsc->cap_flushing_wq);
3476 spin_lock_init(&mdsc->dentry_lru_lock);
3477 INIT_LIST_HEAD(&mdsc->dentry_lru);
3479 ceph_caps_init(mdsc);
3480 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3482 init_rwsem(&mdsc->pool_perm_rwsem);
3483 mdsc->pool_perm_tree = RB_ROOT;
3489 * Wait for safe replies on open mds requests. If we time out, drop
3490 * all requests from the tree to avoid dangling dentry refs.
3492 static void wait_requests(struct ceph_mds_client *mdsc)
3494 struct ceph_options *opts = mdsc->fsc->client->options;
3495 struct ceph_mds_request *req;
3497 mutex_lock(&mdsc->mutex);
3498 if (__get_oldest_req(mdsc)) {
3499 mutex_unlock(&mdsc->mutex);
3501 dout("wait_requests waiting for requests\n");
3502 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3503 ceph_timeout_jiffies(opts->mount_timeout));
3505 /* tear down remaining requests */
3506 mutex_lock(&mdsc->mutex);
3507 while ((req = __get_oldest_req(mdsc))) {
3508 dout("wait_requests timed out on tid %llu\n",
3510 __unregister_request(mdsc, req);
3513 mutex_unlock(&mdsc->mutex);
3514 dout("wait_requests done\n");
3518 * called before mount is ro, and before dentries are torn down.
3519 * (hmm, does this still race with new lookups?)
3521 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3523 dout("pre_umount\n");
3527 ceph_flush_dirty_caps(mdsc);
3528 wait_requests(mdsc);
3531 * wait for reply handlers to drop their request refs and
3532 * their inode/dcache refs
3538 * wait for all write mds requests to flush.
3540 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3542 struct ceph_mds_request *req = NULL, *nextreq;
3545 mutex_lock(&mdsc->mutex);
3546 dout("wait_unsafe_requests want %lld\n", want_tid);
3548 req = __get_oldest_req(mdsc);
3549 while (req && req->r_tid <= want_tid) {
3550 /* find next request */
3551 n = rb_next(&req->r_node);
3553 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3556 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3557 (req->r_op & CEPH_MDS_OP_WRITE)) {
3559 ceph_mdsc_get_request(req);
3561 ceph_mdsc_get_request(nextreq);
3562 mutex_unlock(&mdsc->mutex);
3563 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3564 req->r_tid, want_tid);
3565 wait_for_completion(&req->r_safe_completion);
3566 mutex_lock(&mdsc->mutex);
3567 ceph_mdsc_put_request(req);
3569 break; /* next dne before, so we're done! */
3570 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3571 /* next request was removed from tree */
3572 ceph_mdsc_put_request(nextreq);
3575 ceph_mdsc_put_request(nextreq); /* won't go away */
3579 mutex_unlock(&mdsc->mutex);
3580 dout("wait_unsafe_requests done\n");
3583 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3585 u64 want_tid, want_flush, want_snap;
3587 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3591 mutex_lock(&mdsc->mutex);
3592 want_tid = mdsc->last_tid;
3593 mutex_unlock(&mdsc->mutex);
3595 ceph_flush_dirty_caps(mdsc);
3596 spin_lock(&mdsc->cap_dirty_lock);
3597 want_flush = mdsc->last_cap_flush_tid;
3598 spin_unlock(&mdsc->cap_dirty_lock);
3600 down_read(&mdsc->snap_rwsem);
3601 want_snap = mdsc->last_snap_seq;
3602 up_read(&mdsc->snap_rwsem);
3604 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3605 want_tid, want_flush, want_snap);
3607 wait_unsafe_requests(mdsc, want_tid);
3608 wait_caps_flush(mdsc, want_flush, want_snap);
3612 * true if all sessions are closed, or we force unmount
3614 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3616 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3618 return atomic_read(&mdsc->num_sessions) == 0;
3622 * called after sb is ro.
3624 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3626 struct ceph_options *opts = mdsc->fsc->client->options;
3627 struct ceph_mds_session *session;
3630 dout("close_sessions\n");
3632 /* close sessions */
3633 mutex_lock(&mdsc->mutex);
3634 for (i = 0; i < mdsc->max_sessions; i++) {
3635 session = __ceph_lookup_mds_session(mdsc, i);
3638 mutex_unlock(&mdsc->mutex);
3639 mutex_lock(&session->s_mutex);
3640 __close_session(mdsc, session);
3641 mutex_unlock(&session->s_mutex);
3642 ceph_put_mds_session(session);
3643 mutex_lock(&mdsc->mutex);
3645 mutex_unlock(&mdsc->mutex);
3647 dout("waiting for sessions to close\n");
3648 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3649 ceph_timeout_jiffies(opts->mount_timeout));
3651 /* tear down remaining sessions */
3652 mutex_lock(&mdsc->mutex);
3653 for (i = 0; i < mdsc->max_sessions; i++) {
3654 if (mdsc->sessions[i]) {
3655 session = get_session(mdsc->sessions[i]);
3656 __unregister_session(mdsc, session);
3657 mutex_unlock(&mdsc->mutex);
3658 mutex_lock(&session->s_mutex);
3659 remove_session_caps(session);
3660 mutex_unlock(&session->s_mutex);
3661 ceph_put_mds_session(session);
3662 mutex_lock(&mdsc->mutex);
3665 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3666 mutex_unlock(&mdsc->mutex);
3668 ceph_cleanup_empty_realms(mdsc);
3670 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3675 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3677 struct ceph_mds_session *session;
3680 dout("force umount\n");
3682 mutex_lock(&mdsc->mutex);
3683 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3684 session = __ceph_lookup_mds_session(mdsc, mds);
3687 mutex_unlock(&mdsc->mutex);
3688 mutex_lock(&session->s_mutex);
3689 __close_session(mdsc, session);
3690 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3691 cleanup_session_requests(mdsc, session);
3692 remove_session_caps(session);
3694 mutex_unlock(&session->s_mutex);
3695 ceph_put_mds_session(session);
3696 mutex_lock(&mdsc->mutex);
3697 kick_requests(mdsc, mds);
3699 __wake_requests(mdsc, &mdsc->waiting_for_map);
3700 mutex_unlock(&mdsc->mutex);
3703 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3706 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3708 ceph_mdsmap_destroy(mdsc->mdsmap);
3709 kfree(mdsc->sessions);
3710 ceph_caps_finalize(mdsc);
3711 ceph_pool_perm_destroy(mdsc);
3714 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3716 struct ceph_mds_client *mdsc = fsc->mdsc;
3718 dout("mdsc_destroy %p\n", mdsc);
3719 ceph_mdsc_stop(mdsc);
3721 /* flush out any connection work with references to us */
3726 dout("mdsc_destroy %p done\n", mdsc);
3731 * handle mds map update.
3733 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3737 void *p = msg->front.iov_base;
3738 void *end = p + msg->front.iov_len;
3739 struct ceph_mdsmap *newmap, *oldmap;
3740 struct ceph_fsid fsid;
3743 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3744 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3745 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3747 epoch = ceph_decode_32(&p);
3748 maplen = ceph_decode_32(&p);
3749 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3751 /* do we need it? */
3752 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3753 mutex_lock(&mdsc->mutex);
3754 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3755 dout("handle_map epoch %u <= our %u\n",
3756 epoch, mdsc->mdsmap->m_epoch);
3757 mutex_unlock(&mdsc->mutex);
3761 newmap = ceph_mdsmap_decode(&p, end);
3762 if (IS_ERR(newmap)) {
3763 err = PTR_ERR(newmap);
3767 /* swap into place */
3769 oldmap = mdsc->mdsmap;
3770 mdsc->mdsmap = newmap;
3771 check_new_map(mdsc, newmap, oldmap);
3772 ceph_mdsmap_destroy(oldmap);
3774 mdsc->mdsmap = newmap; /* first mds map */
3776 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3778 __wake_requests(mdsc, &mdsc->waiting_for_map);
3780 mutex_unlock(&mdsc->mutex);
3781 schedule_delayed(mdsc);
3785 mutex_unlock(&mdsc->mutex);
3787 pr_err("error decoding mdsmap %d\n", err);
3791 static struct ceph_connection *con_get(struct ceph_connection *con)
3793 struct ceph_mds_session *s = con->private;
3795 if (get_session(s)) {
3796 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3799 dout("mdsc con_get %p FAIL\n", s);
3803 static void con_put(struct ceph_connection *con)
3805 struct ceph_mds_session *s = con->private;
3807 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3808 ceph_put_mds_session(s);
3812 * if the client is unresponsive for long enough, the mds will kill
3813 * the session entirely.
3815 static void peer_reset(struct ceph_connection *con)
3817 struct ceph_mds_session *s = con->private;
3818 struct ceph_mds_client *mdsc = s->s_mdsc;
3820 pr_warn("mds%d closed our session\n", s->s_mds);
3821 send_mds_reconnect(mdsc, s);
3824 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3826 struct ceph_mds_session *s = con->private;
3827 struct ceph_mds_client *mdsc = s->s_mdsc;
3828 int type = le16_to_cpu(msg->hdr.type);
3830 mutex_lock(&mdsc->mutex);
3831 if (__verify_registered_session(mdsc, s) < 0) {
3832 mutex_unlock(&mdsc->mutex);
3835 mutex_unlock(&mdsc->mutex);
3838 case CEPH_MSG_MDS_MAP:
3839 ceph_mdsc_handle_map(mdsc, msg);
3841 case CEPH_MSG_CLIENT_SESSION:
3842 handle_session(s, msg);
3844 case CEPH_MSG_CLIENT_REPLY:
3845 handle_reply(s, msg);
3847 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3848 handle_forward(mdsc, s, msg);
3850 case CEPH_MSG_CLIENT_CAPS:
3851 ceph_handle_caps(s, msg);
3853 case CEPH_MSG_CLIENT_SNAP:
3854 ceph_handle_snap(mdsc, s, msg);
3856 case CEPH_MSG_CLIENT_LEASE:
3857 handle_lease(mdsc, s, msg);
3861 pr_err("received unknown message type %d %s\n", type,
3862 ceph_msg_type_name(type));
3873 * Note: returned pointer is the address of a structure that's
3874 * managed separately. Caller must *not* attempt to free it.
3876 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3877 int *proto, int force_new)
3879 struct ceph_mds_session *s = con->private;
3880 struct ceph_mds_client *mdsc = s->s_mdsc;
3881 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3882 struct ceph_auth_handshake *auth = &s->s_auth;
3884 if (force_new && auth->authorizer) {
3885 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3886 auth->authorizer = NULL;
3888 if (!auth->authorizer) {
3889 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3892 return ERR_PTR(ret);
3894 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3897 return ERR_PTR(ret);
3899 *proto = ac->protocol;
3905 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3907 struct ceph_mds_session *s = con->private;
3908 struct ceph_mds_client *mdsc = s->s_mdsc;
3909 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3911 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3914 static int invalidate_authorizer(struct ceph_connection *con)
3916 struct ceph_mds_session *s = con->private;
3917 struct ceph_mds_client *mdsc = s->s_mdsc;
3918 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3920 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3922 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3925 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3926 struct ceph_msg_header *hdr, int *skip)
3928 struct ceph_msg *msg;
3929 int type = (int) le16_to_cpu(hdr->type);
3930 int front_len = (int) le32_to_cpu(hdr->front_len);
3936 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3938 pr_err("unable to allocate msg type %d len %d\n",
3946 static int mds_sign_message(struct ceph_msg *msg)
3948 struct ceph_mds_session *s = msg->con->private;
3949 struct ceph_auth_handshake *auth = &s->s_auth;
3951 return ceph_auth_sign_message(auth, msg);
3954 static int mds_check_message_signature(struct ceph_msg *msg)
3956 struct ceph_mds_session *s = msg->con->private;
3957 struct ceph_auth_handshake *auth = &s->s_auth;
3959 return ceph_auth_check_message_signature(auth, msg);
3962 static const struct ceph_connection_operations mds_con_ops = {
3965 .dispatch = dispatch,
3966 .get_authorizer = get_authorizer,
3967 .verify_authorizer_reply = verify_authorizer_reply,
3968 .invalidate_authorizer = invalidate_authorizer,
3969 .peer_reset = peer_reset,
3970 .alloc_msg = mds_alloc_msg,
3971 .sign_message = mds_sign_message,
3972 .check_message_signature = mds_check_message_signature,