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45 * are no longer valid.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 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
278 return parse_reply_info_filelock(p, end, info, features);
279 else if (info->head->op == CEPH_MDS_OP_READDIR ||
280 info->head->op == CEPH_MDS_OP_LSSNAP)
281 return parse_reply_info_dir(p, end, info, features);
282 else if (info->head->op == CEPH_MDS_OP_CREATE)
283 return parse_reply_info_create(p, end, info, features);
289 * parse entire mds reply
291 static int parse_reply_info(struct ceph_msg *msg,
292 struct ceph_mds_reply_info_parsed *info,
299 info->head = msg->front.iov_base;
300 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
301 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
304 ceph_decode_32_safe(&p, end, len, bad);
306 ceph_decode_need(&p, end, len, bad);
307 err = parse_reply_info_trace(&p, p+len, info, features);
313 ceph_decode_32_safe(&p, end, len, bad);
315 ceph_decode_need(&p, end, len, bad);
316 err = parse_reply_info_extra(&p, p+len, info, features);
322 ceph_decode_32_safe(&p, end, len, bad);
323 info->snapblob_len = len;
334 pr_err("mds parse_reply err %d\n", err);
338 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
342 free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
349 const char *ceph_session_state_name(int s)
352 case CEPH_MDS_SESSION_NEW: return "new";
353 case CEPH_MDS_SESSION_OPENING: return "opening";
354 case CEPH_MDS_SESSION_OPEN: return "open";
355 case CEPH_MDS_SESSION_HUNG: return "hung";
356 case CEPH_MDS_SESSION_CLOSING: return "closing";
357 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
358 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
359 default: return "???";
363 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
365 if (atomic_inc_not_zero(&s->s_ref)) {
366 dout("mdsc get_session %p %d -> %d\n", s,
367 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
370 dout("mdsc get_session %p 0 -- FAIL", s);
375 void ceph_put_mds_session(struct ceph_mds_session *s)
377 dout("mdsc put_session %p %d -> %d\n", s,
378 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
379 if (atomic_dec_and_test(&s->s_ref)) {
380 if (s->s_auth.authorizer)
381 ceph_auth_destroy_authorizer(
382 s->s_mdsc->fsc->client->monc.auth,
383 s->s_auth.authorizer);
389 * called under mdsc->mutex
391 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
394 struct ceph_mds_session *session;
396 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
398 session = mdsc->sessions[mds];
399 dout("lookup_mds_session %p %d\n", session,
400 atomic_read(&session->s_ref));
401 get_session(session);
405 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
407 if (mds >= mdsc->max_sessions)
409 return mdsc->sessions[mds];
412 static int __verify_registered_session(struct ceph_mds_client *mdsc,
413 struct ceph_mds_session *s)
415 if (s->s_mds >= mdsc->max_sessions ||
416 mdsc->sessions[s->s_mds] != s)
422 * create+register a new session for given mds.
423 * called under mdsc->mutex.
425 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
428 struct ceph_mds_session *s;
430 if (mds >= mdsc->mdsmap->m_max_mds)
431 return ERR_PTR(-EINVAL);
433 s = kzalloc(sizeof(*s), GFP_NOFS);
435 return ERR_PTR(-ENOMEM);
438 s->s_state = CEPH_MDS_SESSION_NEW;
441 mutex_init(&s->s_mutex);
443 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
445 spin_lock_init(&s->s_gen_ttl_lock);
447 s->s_cap_ttl = jiffies - 1;
449 spin_lock_init(&s->s_cap_lock);
450 s->s_renew_requested = 0;
452 INIT_LIST_HEAD(&s->s_caps);
455 atomic_set(&s->s_ref, 1);
456 INIT_LIST_HEAD(&s->s_waiting);
457 INIT_LIST_HEAD(&s->s_unsafe);
458 s->s_num_cap_releases = 0;
459 s->s_cap_reconnect = 0;
460 s->s_cap_iterator = NULL;
461 INIT_LIST_HEAD(&s->s_cap_releases);
462 INIT_LIST_HEAD(&s->s_cap_flushing);
463 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
465 dout("register_session mds%d\n", mds);
466 if (mds >= mdsc->max_sessions) {
467 int newmax = 1 << get_count_order(mds+1);
468 struct ceph_mds_session **sa;
470 dout("register_session realloc to %d\n", newmax);
471 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
474 if (mdsc->sessions) {
475 memcpy(sa, mdsc->sessions,
476 mdsc->max_sessions * sizeof(void *));
477 kfree(mdsc->sessions);
480 mdsc->max_sessions = newmax;
482 mdsc->sessions[mds] = s;
483 atomic_inc(&mdsc->num_sessions);
484 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
486 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
487 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
493 return ERR_PTR(-ENOMEM);
497 * called under mdsc->mutex
499 static void __unregister_session(struct ceph_mds_client *mdsc,
500 struct ceph_mds_session *s)
502 dout("__unregister_session mds%d %p\n", s->s_mds, s);
503 BUG_ON(mdsc->sessions[s->s_mds] != s);
504 mdsc->sessions[s->s_mds] = NULL;
505 ceph_con_close(&s->s_con);
506 ceph_put_mds_session(s);
507 atomic_dec(&mdsc->num_sessions);
511 * drop session refs in request.
513 * should be last request ref, or hold mdsc->mutex
515 static void put_request_session(struct ceph_mds_request *req)
517 if (req->r_session) {
518 ceph_put_mds_session(req->r_session);
519 req->r_session = NULL;
523 void ceph_mdsc_release_request(struct kref *kref)
525 struct ceph_mds_request *req = container_of(kref,
526 struct ceph_mds_request,
528 destroy_reply_info(&req->r_reply_info);
530 ceph_msg_put(req->r_request);
532 ceph_msg_put(req->r_reply);
534 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
537 if (req->r_locked_dir)
538 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
539 iput(req->r_target_inode);
542 if (req->r_old_dentry)
543 dput(req->r_old_dentry);
544 if (req->r_old_dentry_dir) {
546 * track (and drop pins for) r_old_dentry_dir
547 * separately, since r_old_dentry's d_parent may have
548 * changed between the dir mutex being dropped and
549 * this request being freed.
551 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
553 iput(req->r_old_dentry_dir);
558 ceph_pagelist_release(req->r_pagelist);
559 put_request_session(req);
560 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
565 * lookup session, bump ref if found.
567 * called under mdsc->mutex.
569 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
572 struct ceph_mds_request *req;
573 struct rb_node *n = mdsc->request_tree.rb_node;
576 req = rb_entry(n, struct ceph_mds_request, r_node);
577 if (tid < req->r_tid)
579 else if (tid > req->r_tid)
582 ceph_mdsc_get_request(req);
589 static void __insert_request(struct ceph_mds_client *mdsc,
590 struct ceph_mds_request *new)
592 struct rb_node **p = &mdsc->request_tree.rb_node;
593 struct rb_node *parent = NULL;
594 struct ceph_mds_request *req = NULL;
598 req = rb_entry(parent, struct ceph_mds_request, r_node);
599 if (new->r_tid < req->r_tid)
601 else if (new->r_tid > req->r_tid)
607 rb_link_node(&new->r_node, parent, p);
608 rb_insert_color(&new->r_node, &mdsc->request_tree);
612 * Register an in-flight request, and assign a tid. Link to directory
613 * are modifying (if any).
615 * Called under mdsc->mutex.
617 static void __register_request(struct ceph_mds_client *mdsc,
618 struct ceph_mds_request *req,
621 req->r_tid = ++mdsc->last_tid;
623 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
625 dout("__register_request %p tid %lld\n", req, req->r_tid);
626 ceph_mdsc_get_request(req);
627 __insert_request(mdsc, req);
629 req->r_uid = current_fsuid();
630 req->r_gid = current_fsgid();
632 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
633 mdsc->oldest_tid = req->r_tid;
636 struct ceph_inode_info *ci = ceph_inode(dir);
639 spin_lock(&ci->i_unsafe_lock);
640 req->r_unsafe_dir = dir;
641 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
642 spin_unlock(&ci->i_unsafe_lock);
646 static void __unregister_request(struct ceph_mds_client *mdsc,
647 struct ceph_mds_request *req)
649 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
651 if (req->r_tid == mdsc->oldest_tid) {
652 struct rb_node *p = rb_next(&req->r_node);
653 mdsc->oldest_tid = 0;
655 struct ceph_mds_request *next_req =
656 rb_entry(p, struct ceph_mds_request, r_node);
657 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
658 mdsc->oldest_tid = next_req->r_tid;
665 rb_erase(&req->r_node, &mdsc->request_tree);
666 RB_CLEAR_NODE(&req->r_node);
668 if (req->r_unsafe_dir) {
669 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
671 spin_lock(&ci->i_unsafe_lock);
672 list_del_init(&req->r_unsafe_dir_item);
673 spin_unlock(&ci->i_unsafe_lock);
675 iput(req->r_unsafe_dir);
676 req->r_unsafe_dir = NULL;
679 complete_all(&req->r_safe_completion);
681 ceph_mdsc_put_request(req);
685 * Choose mds to send request to next. If there is a hint set in the
686 * request (e.g., due to a prior forward hint from the mds), use that.
687 * Otherwise, consult frag tree and/or caps to identify the
688 * appropriate mds. If all else fails, choose randomly.
690 * Called under mdsc->mutex.
692 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
695 * we don't need to worry about protecting the d_parent access
696 * here because we never renaming inside the snapped namespace
697 * except to resplice to another snapdir, and either the old or new
698 * result is a valid result.
700 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
701 dentry = dentry->d_parent;
705 static int __choose_mds(struct ceph_mds_client *mdsc,
706 struct ceph_mds_request *req)
709 struct ceph_inode_info *ci;
710 struct ceph_cap *cap;
711 int mode = req->r_direct_mode;
713 u32 hash = req->r_direct_hash;
714 bool is_hash = req->r_direct_is_hash;
717 * is there a specific mds we should try? ignore hint if we have
718 * no session and the mds is not up (active or recovering).
720 if (req->r_resend_mds >= 0 &&
721 (__have_session(mdsc, req->r_resend_mds) ||
722 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
723 dout("choose_mds using resend_mds mds%d\n",
725 return req->r_resend_mds;
728 if (mode == USE_RANDOM_MDS)
733 inode = req->r_inode;
734 } else if (req->r_dentry) {
735 /* ignore race with rename; old or new d_parent is okay */
736 struct dentry *parent = req->r_dentry->d_parent;
737 struct inode *dir = d_inode(parent);
739 if (dir->i_sb != mdsc->fsc->sb) {
741 inode = d_inode(req->r_dentry);
742 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
743 /* direct snapped/virtual snapdir requests
744 * based on parent dir inode */
745 struct dentry *dn = get_nonsnap_parent(parent);
747 dout("__choose_mds using nonsnap parent %p\n", inode);
750 inode = d_inode(req->r_dentry);
751 if (!inode || mode == USE_AUTH_MDS) {
754 hash = ceph_dentry_hash(dir, req->r_dentry);
760 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
764 ci = ceph_inode(inode);
766 if (is_hash && S_ISDIR(inode->i_mode)) {
767 struct ceph_inode_frag frag;
770 ceph_choose_frag(ci, hash, &frag, &found);
772 if (mode == USE_ANY_MDS && frag.ndist > 0) {
775 /* choose a random replica */
776 get_random_bytes(&r, 1);
779 dout("choose_mds %p %llx.%llx "
780 "frag %u mds%d (%d/%d)\n",
781 inode, ceph_vinop(inode),
784 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
785 CEPH_MDS_STATE_ACTIVE)
789 /* since this file/dir wasn't known to be
790 * replicated, then we want to look for the
791 * authoritative mds. */
794 /* choose auth mds */
796 dout("choose_mds %p %llx.%llx "
797 "frag %u mds%d (auth)\n",
798 inode, ceph_vinop(inode), frag.frag, mds);
799 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
800 CEPH_MDS_STATE_ACTIVE)
806 spin_lock(&ci->i_ceph_lock);
808 if (mode == USE_AUTH_MDS)
809 cap = ci->i_auth_cap;
810 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
811 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
813 spin_unlock(&ci->i_ceph_lock);
816 mds = cap->session->s_mds;
817 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
818 inode, ceph_vinop(inode), mds,
819 cap == ci->i_auth_cap ? "auth " : "", cap);
820 spin_unlock(&ci->i_ceph_lock);
824 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
825 dout("choose_mds chose random mds%d\n", mds);
833 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
835 struct ceph_msg *msg;
836 struct ceph_mds_session_head *h;
838 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
841 pr_err("create_session_msg ENOMEM creating msg\n");
844 h = msg->front.iov_base;
845 h->op = cpu_to_le32(op);
846 h->seq = cpu_to_le64(seq);
852 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
853 * to include additional client metadata fields.
855 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
857 struct ceph_msg *msg;
858 struct ceph_mds_session_head *h;
860 int metadata_bytes = 0;
861 int metadata_key_count = 0;
862 struct ceph_options *opt = mdsc->fsc->client->options;
865 const char* metadata[][2] = {
866 {"hostname", utsname()->nodename},
867 {"kernel_version", utsname()->release},
868 {"entity_id", opt->name ? opt->name : ""},
872 /* Calculate serialized length of metadata */
873 metadata_bytes = 4; /* map length */
874 for (i = 0; metadata[i][0] != NULL; ++i) {
875 metadata_bytes += 8 + strlen(metadata[i][0]) +
876 strlen(metadata[i][1]);
877 metadata_key_count++;
880 /* Allocate the message */
881 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
884 pr_err("create_session_msg ENOMEM creating msg\n");
887 h = msg->front.iov_base;
888 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
889 h->seq = cpu_to_le64(seq);
892 * Serialize client metadata into waiting buffer space, using
893 * the format that userspace expects for map<string, string>
895 * ClientSession messages with metadata are v2
897 msg->hdr.version = cpu_to_le16(2);
898 msg->hdr.compat_version = cpu_to_le16(1);
900 /* The write pointer, following the session_head structure */
901 p = msg->front.iov_base + sizeof(*h);
903 /* Number of entries in the map */
904 ceph_encode_32(&p, metadata_key_count);
906 /* Two length-prefixed strings for each entry in the map */
907 for (i = 0; metadata[i][0] != NULL; ++i) {
908 size_t const key_len = strlen(metadata[i][0]);
909 size_t const val_len = strlen(metadata[i][1]);
911 ceph_encode_32(&p, key_len);
912 memcpy(p, metadata[i][0], key_len);
914 ceph_encode_32(&p, val_len);
915 memcpy(p, metadata[i][1], val_len);
923 * send session open request.
925 * called under mdsc->mutex
927 static int __open_session(struct ceph_mds_client *mdsc,
928 struct ceph_mds_session *session)
930 struct ceph_msg *msg;
932 int mds = session->s_mds;
934 /* wait for mds to go active? */
935 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
936 dout("open_session to mds%d (%s)\n", mds,
937 ceph_mds_state_name(mstate));
938 session->s_state = CEPH_MDS_SESSION_OPENING;
939 session->s_renew_requested = jiffies;
941 /* send connect message */
942 msg = create_session_open_msg(mdsc, session->s_seq);
945 ceph_con_send(&session->s_con, msg);
950 * open sessions for any export targets for the given mds
952 * called under mdsc->mutex
954 static struct ceph_mds_session *
955 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
957 struct ceph_mds_session *session;
959 session = __ceph_lookup_mds_session(mdsc, target);
961 session = register_session(mdsc, target);
965 if (session->s_state == CEPH_MDS_SESSION_NEW ||
966 session->s_state == CEPH_MDS_SESSION_CLOSING)
967 __open_session(mdsc, session);
972 struct ceph_mds_session *
973 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
975 struct ceph_mds_session *session;
977 dout("open_export_target_session to mds%d\n", target);
979 mutex_lock(&mdsc->mutex);
980 session = __open_export_target_session(mdsc, target);
981 mutex_unlock(&mdsc->mutex);
986 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
987 struct ceph_mds_session *session)
989 struct ceph_mds_info *mi;
990 struct ceph_mds_session *ts;
991 int i, mds = session->s_mds;
993 if (mds >= mdsc->mdsmap->m_max_mds)
996 mi = &mdsc->mdsmap->m_info[mds];
997 dout("open_export_target_sessions for mds%d (%d targets)\n",
998 session->s_mds, mi->num_export_targets);
1000 for (i = 0; i < mi->num_export_targets; i++) {
1001 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1003 ceph_put_mds_session(ts);
1007 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1008 struct ceph_mds_session *session)
1010 mutex_lock(&mdsc->mutex);
1011 __open_export_target_sessions(mdsc, session);
1012 mutex_unlock(&mdsc->mutex);
1019 /* caller holds s_cap_lock, we drop it */
1020 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1021 struct ceph_mds_session *session)
1022 __releases(session->s_cap_lock)
1024 LIST_HEAD(tmp_list);
1025 list_splice_init(&session->s_cap_releases, &tmp_list);
1026 session->s_num_cap_releases = 0;
1027 spin_unlock(&session->s_cap_lock);
1029 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1030 while (!list_empty(&tmp_list)) {
1031 struct ceph_cap *cap;
1032 /* zero out the in-progress message */
1033 cap = list_first_entry(&tmp_list,
1034 struct ceph_cap, session_caps);
1035 list_del(&cap->session_caps);
1036 ceph_put_cap(mdsc, cap);
1040 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1041 struct ceph_mds_session *session)
1043 struct ceph_mds_request *req;
1046 dout("cleanup_session_requests mds%d\n", session->s_mds);
1047 mutex_lock(&mdsc->mutex);
1048 while (!list_empty(&session->s_unsafe)) {
1049 req = list_first_entry(&session->s_unsafe,
1050 struct ceph_mds_request, r_unsafe_item);
1051 list_del_init(&req->r_unsafe_item);
1052 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1054 __unregister_request(mdsc, req);
1056 /* zero r_attempts, so kick_requests() will re-send requests */
1057 p = rb_first(&mdsc->request_tree);
1059 req = rb_entry(p, struct ceph_mds_request, r_node);
1061 if (req->r_session &&
1062 req->r_session->s_mds == session->s_mds)
1063 req->r_attempts = 0;
1065 mutex_unlock(&mdsc->mutex);
1069 * Helper to safely iterate over all caps associated with a session, with
1070 * special care taken to handle a racing __ceph_remove_cap().
1072 * Caller must hold session s_mutex.
1074 static int iterate_session_caps(struct ceph_mds_session *session,
1075 int (*cb)(struct inode *, struct ceph_cap *,
1078 struct list_head *p;
1079 struct ceph_cap *cap;
1080 struct inode *inode, *last_inode = NULL;
1081 struct ceph_cap *old_cap = NULL;
1084 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1085 spin_lock(&session->s_cap_lock);
1086 p = session->s_caps.next;
1087 while (p != &session->s_caps) {
1088 cap = list_entry(p, struct ceph_cap, session_caps);
1089 inode = igrab(&cap->ci->vfs_inode);
1094 session->s_cap_iterator = cap;
1095 spin_unlock(&session->s_cap_lock);
1102 ceph_put_cap(session->s_mdsc, old_cap);
1106 ret = cb(inode, cap, arg);
1109 spin_lock(&session->s_cap_lock);
1111 if (cap->ci == NULL) {
1112 dout("iterate_session_caps finishing cap %p removal\n",
1114 BUG_ON(cap->session != session);
1115 cap->session = NULL;
1116 list_del_init(&cap->session_caps);
1117 session->s_nr_caps--;
1118 if (cap->queue_release) {
1119 list_add_tail(&cap->session_caps,
1120 &session->s_cap_releases);
1121 session->s_num_cap_releases++;
1123 old_cap = cap; /* put_cap it w/o locks held */
1131 session->s_cap_iterator = NULL;
1132 spin_unlock(&session->s_cap_lock);
1136 ceph_put_cap(session->s_mdsc, old_cap);
1141 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1144 struct ceph_inode_info *ci = ceph_inode(inode);
1145 LIST_HEAD(to_remove);
1148 dout("removing cap %p, ci is %p, inode is %p\n",
1149 cap, ci, &ci->vfs_inode);
1150 spin_lock(&ci->i_ceph_lock);
1151 __ceph_remove_cap(cap, false);
1152 if (!ci->i_auth_cap) {
1153 struct ceph_cap_flush *cf;
1154 struct ceph_mds_client *mdsc =
1155 ceph_sb_to_client(inode->i_sb)->mdsc;
1158 struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1161 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1162 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1163 list_add(&cf->list, &to_remove);
1166 spin_lock(&mdsc->cap_dirty_lock);
1168 list_for_each_entry(cf, &to_remove, list)
1169 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1171 if (!list_empty(&ci->i_dirty_item)) {
1172 pr_warn_ratelimited(
1173 " dropping dirty %s state for %p %lld\n",
1174 ceph_cap_string(ci->i_dirty_caps),
1175 inode, ceph_ino(inode));
1176 ci->i_dirty_caps = 0;
1177 list_del_init(&ci->i_dirty_item);
1180 if (!list_empty(&ci->i_flushing_item)) {
1181 pr_warn_ratelimited(
1182 " dropping dirty+flushing %s state for %p %lld\n",
1183 ceph_cap_string(ci->i_flushing_caps),
1184 inode, ceph_ino(inode));
1185 ci->i_flushing_caps = 0;
1186 list_del_init(&ci->i_flushing_item);
1187 mdsc->num_cap_flushing--;
1190 spin_unlock(&mdsc->cap_dirty_lock);
1193 spin_unlock(&ci->i_ceph_lock);
1194 while (!list_empty(&to_remove)) {
1195 struct ceph_cap_flush *cf;
1196 cf = list_first_entry(&to_remove,
1197 struct ceph_cap_flush, list);
1198 list_del(&cf->list);
1207 * caller must hold session s_mutex
1209 static void remove_session_caps(struct ceph_mds_session *session)
1211 dout("remove_session_caps on %p\n", session);
1212 iterate_session_caps(session, remove_session_caps_cb, NULL);
1214 spin_lock(&session->s_cap_lock);
1215 if (session->s_nr_caps > 0) {
1216 struct super_block *sb = session->s_mdsc->fsc->sb;
1217 struct inode *inode;
1218 struct ceph_cap *cap, *prev = NULL;
1219 struct ceph_vino vino;
1221 * iterate_session_caps() skips inodes that are being
1222 * deleted, we need to wait until deletions are complete.
1223 * __wait_on_freeing_inode() is designed for the job,
1224 * but it is not exported, so use lookup inode function
1227 while (!list_empty(&session->s_caps)) {
1228 cap = list_entry(session->s_caps.next,
1229 struct ceph_cap, session_caps);
1233 vino = cap->ci->i_vino;
1234 spin_unlock(&session->s_cap_lock);
1236 inode = ceph_find_inode(sb, vino);
1239 spin_lock(&session->s_cap_lock);
1243 // drop cap expires and unlock s_cap_lock
1244 cleanup_cap_releases(session->s_mdsc, session);
1246 BUG_ON(session->s_nr_caps > 0);
1247 BUG_ON(!list_empty(&session->s_cap_flushing));
1251 * wake up any threads waiting on this session's caps. if the cap is
1252 * old (didn't get renewed on the client reconnect), remove it now.
1254 * caller must hold s_mutex.
1256 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1259 struct ceph_inode_info *ci = ceph_inode(inode);
1261 wake_up_all(&ci->i_cap_wq);
1263 spin_lock(&ci->i_ceph_lock);
1264 ci->i_wanted_max_size = 0;
1265 ci->i_requested_max_size = 0;
1266 spin_unlock(&ci->i_ceph_lock);
1271 static void wake_up_session_caps(struct ceph_mds_session *session,
1274 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1275 iterate_session_caps(session, wake_up_session_cb,
1276 (void *)(unsigned long)reconnect);
1280 * Send periodic message to MDS renewing all currently held caps. The
1281 * ack will reset the expiration for all caps from this session.
1283 * caller holds s_mutex
1285 static int send_renew_caps(struct ceph_mds_client *mdsc,
1286 struct ceph_mds_session *session)
1288 struct ceph_msg *msg;
1291 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1292 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1293 pr_info("mds%d caps stale\n", session->s_mds);
1294 session->s_renew_requested = jiffies;
1296 /* do not try to renew caps until a recovering mds has reconnected
1297 * with its clients. */
1298 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1299 if (state < CEPH_MDS_STATE_RECONNECT) {
1300 dout("send_renew_caps ignoring mds%d (%s)\n",
1301 session->s_mds, ceph_mds_state_name(state));
1305 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1306 ceph_mds_state_name(state));
1307 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1308 ++session->s_renew_seq);
1311 ceph_con_send(&session->s_con, msg);
1315 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1316 struct ceph_mds_session *session, u64 seq)
1318 struct ceph_msg *msg;
1320 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1321 session->s_mds, ceph_session_state_name(session->s_state), seq);
1322 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1325 ceph_con_send(&session->s_con, msg);
1331 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1333 * Called under session->s_mutex
1335 static void renewed_caps(struct ceph_mds_client *mdsc,
1336 struct ceph_mds_session *session, int is_renew)
1341 spin_lock(&session->s_cap_lock);
1342 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1344 session->s_cap_ttl = session->s_renew_requested +
1345 mdsc->mdsmap->m_session_timeout*HZ;
1348 if (time_before(jiffies, session->s_cap_ttl)) {
1349 pr_info("mds%d caps renewed\n", session->s_mds);
1352 pr_info("mds%d caps still stale\n", session->s_mds);
1355 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1356 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1357 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1358 spin_unlock(&session->s_cap_lock);
1361 wake_up_session_caps(session, 0);
1365 * send a session close request
1367 static int request_close_session(struct ceph_mds_client *mdsc,
1368 struct ceph_mds_session *session)
1370 struct ceph_msg *msg;
1372 dout("request_close_session mds%d state %s seq %lld\n",
1373 session->s_mds, ceph_session_state_name(session->s_state),
1375 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1378 ceph_con_send(&session->s_con, msg);
1383 * Called with s_mutex held.
1385 static int __close_session(struct ceph_mds_client *mdsc,
1386 struct ceph_mds_session *session)
1388 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1390 session->s_state = CEPH_MDS_SESSION_CLOSING;
1391 return request_close_session(mdsc, session);
1395 * Trim old(er) caps.
1397 * Because we can't cache an inode without one or more caps, we do
1398 * this indirectly: if a cap is unused, we prune its aliases, at which
1399 * point the inode will hopefully get dropped to.
1401 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1402 * memory pressure from the MDS, though, so it needn't be perfect.
1404 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1406 struct ceph_mds_session *session = arg;
1407 struct ceph_inode_info *ci = ceph_inode(inode);
1408 int used, wanted, oissued, mine;
1410 if (session->s_trim_caps <= 0)
1413 spin_lock(&ci->i_ceph_lock);
1414 mine = cap->issued | cap->implemented;
1415 used = __ceph_caps_used(ci);
1416 wanted = __ceph_caps_file_wanted(ci);
1417 oissued = __ceph_caps_issued_other(ci, cap);
1419 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1420 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1421 ceph_cap_string(used), ceph_cap_string(wanted));
1422 if (cap == ci->i_auth_cap) {
1423 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1424 !list_empty(&ci->i_cap_snaps))
1426 if ((used | wanted) & CEPH_CAP_ANY_WR)
1429 if ((used | wanted) & ~oissued & mine)
1430 goto out; /* we need these caps */
1432 session->s_trim_caps--;
1434 /* we aren't the only cap.. just remove us */
1435 __ceph_remove_cap(cap, true);
1437 /* try to drop referring dentries */
1438 spin_unlock(&ci->i_ceph_lock);
1439 d_prune_aliases(inode);
1440 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1441 inode, cap, atomic_read(&inode->i_count));
1446 spin_unlock(&ci->i_ceph_lock);
1451 * Trim session cap count down to some max number.
1453 static int trim_caps(struct ceph_mds_client *mdsc,
1454 struct ceph_mds_session *session,
1457 int trim_caps = session->s_nr_caps - max_caps;
1459 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1460 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1461 if (trim_caps > 0) {
1462 session->s_trim_caps = trim_caps;
1463 iterate_session_caps(session, trim_caps_cb, session);
1464 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1465 session->s_mds, session->s_nr_caps, max_caps,
1466 trim_caps - session->s_trim_caps);
1467 session->s_trim_caps = 0;
1470 ceph_send_cap_releases(mdsc, session);
1474 static int check_capsnap_flush(struct ceph_inode_info *ci,
1478 spin_lock(&ci->i_ceph_lock);
1479 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1480 struct ceph_cap_snap *capsnap =
1481 list_first_entry(&ci->i_cap_snaps,
1482 struct ceph_cap_snap, ci_item);
1483 ret = capsnap->follows >= want_snap_seq;
1485 spin_unlock(&ci->i_ceph_lock);
1489 static int check_caps_flush(struct ceph_mds_client *mdsc,
1493 struct ceph_cap_flush *cf;
1496 spin_lock(&mdsc->cap_dirty_lock);
1497 n = rb_first(&mdsc->cap_flush_tree);
1498 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1499 if (cf && cf->tid <= want_flush_tid) {
1500 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1501 cf->tid, want_flush_tid);
1504 spin_unlock(&mdsc->cap_dirty_lock);
1509 * flush all dirty inode data to disk.
1511 * returns true if we've flushed through want_flush_tid
1513 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1514 u64 want_flush_tid, u64 want_snap_seq)
1518 dout("check_caps_flush want %llu snap want %llu\n",
1519 want_flush_tid, want_snap_seq);
1520 mutex_lock(&mdsc->mutex);
1521 for (mds = 0; mds < mdsc->max_sessions; ) {
1522 struct ceph_mds_session *session = mdsc->sessions[mds];
1523 struct inode *inode = NULL;
1529 get_session(session);
1530 mutex_unlock(&mdsc->mutex);
1532 mutex_lock(&session->s_mutex);
1533 if (!list_empty(&session->s_cap_snaps_flushing)) {
1534 struct ceph_cap_snap *capsnap =
1535 list_first_entry(&session->s_cap_snaps_flushing,
1536 struct ceph_cap_snap,
1538 struct ceph_inode_info *ci = capsnap->ci;
1539 if (!check_capsnap_flush(ci, want_snap_seq)) {
1540 dout("check_cap_flush still flushing snap %p "
1541 "follows %lld <= %lld to mds%d\n",
1542 &ci->vfs_inode, capsnap->follows,
1543 want_snap_seq, mds);
1544 inode = igrab(&ci->vfs_inode);
1547 mutex_unlock(&session->s_mutex);
1548 ceph_put_mds_session(session);
1551 wait_event(mdsc->cap_flushing_wq,
1552 check_capsnap_flush(ceph_inode(inode),
1559 mutex_lock(&mdsc->mutex);
1561 mutex_unlock(&mdsc->mutex);
1563 wait_event(mdsc->cap_flushing_wq,
1564 check_caps_flush(mdsc, want_flush_tid));
1566 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1570 * called under s_mutex
1572 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1573 struct ceph_mds_session *session)
1575 struct ceph_msg *msg = NULL;
1576 struct ceph_mds_cap_release *head;
1577 struct ceph_mds_cap_item *item;
1578 struct ceph_cap *cap;
1579 LIST_HEAD(tmp_list);
1580 int num_cap_releases;
1582 spin_lock(&session->s_cap_lock);
1584 list_splice_init(&session->s_cap_releases, &tmp_list);
1585 num_cap_releases = session->s_num_cap_releases;
1586 session->s_num_cap_releases = 0;
1587 spin_unlock(&session->s_cap_lock);
1589 while (!list_empty(&tmp_list)) {
1591 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1592 PAGE_CACHE_SIZE, GFP_NOFS, false);
1595 head = msg->front.iov_base;
1596 head->num = cpu_to_le32(0);
1597 msg->front.iov_len = sizeof(*head);
1599 cap = list_first_entry(&tmp_list, struct ceph_cap,
1601 list_del(&cap->session_caps);
1604 head = msg->front.iov_base;
1605 le32_add_cpu(&head->num, 1);
1606 item = msg->front.iov_base + msg->front.iov_len;
1607 item->ino = cpu_to_le64(cap->cap_ino);
1608 item->cap_id = cpu_to_le64(cap->cap_id);
1609 item->migrate_seq = cpu_to_le32(cap->mseq);
1610 item->seq = cpu_to_le32(cap->issue_seq);
1611 msg->front.iov_len += sizeof(*item);
1613 ceph_put_cap(mdsc, cap);
1615 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1616 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1617 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1618 ceph_con_send(&session->s_con, msg);
1623 BUG_ON(num_cap_releases != 0);
1625 spin_lock(&session->s_cap_lock);
1626 if (!list_empty(&session->s_cap_releases))
1628 spin_unlock(&session->s_cap_lock);
1631 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1632 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1633 ceph_con_send(&session->s_con, msg);
1637 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1639 spin_lock(&session->s_cap_lock);
1640 list_splice(&tmp_list, &session->s_cap_releases);
1641 session->s_num_cap_releases += num_cap_releases;
1642 spin_unlock(&session->s_cap_lock);
1649 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1652 struct ceph_inode_info *ci = ceph_inode(dir);
1653 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1654 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1655 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1656 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1657 int order, num_entries;
1659 spin_lock(&ci->i_ceph_lock);
1660 num_entries = ci->i_files + ci->i_subdirs;
1661 spin_unlock(&ci->i_ceph_lock);
1662 num_entries = max(num_entries, 1);
1663 num_entries = min(num_entries, opt->max_readdir);
1665 order = get_order(size * num_entries);
1666 while (order >= 0) {
1667 rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN,
1676 num_entries = (PAGE_SIZE << order) / size;
1677 num_entries = min(num_entries, opt->max_readdir);
1679 rinfo->dir_buf_size = PAGE_SIZE << order;
1680 req->r_num_caps = num_entries + 1;
1681 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1682 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1687 * Create an mds request.
1689 struct ceph_mds_request *
1690 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1692 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1695 return ERR_PTR(-ENOMEM);
1697 mutex_init(&req->r_fill_mutex);
1699 req->r_started = jiffies;
1700 req->r_resend_mds = -1;
1701 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1703 kref_init(&req->r_kref);
1704 INIT_LIST_HEAD(&req->r_wait);
1705 init_completion(&req->r_completion);
1706 init_completion(&req->r_safe_completion);
1707 INIT_LIST_HEAD(&req->r_unsafe_item);
1709 req->r_stamp = CURRENT_TIME;
1712 req->r_direct_mode = mode;
1717 * return oldest (lowest) request, tid in request tree, 0 if none.
1719 * called under mdsc->mutex.
1721 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1723 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1725 return rb_entry(rb_first(&mdsc->request_tree),
1726 struct ceph_mds_request, r_node);
1729 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1731 return mdsc->oldest_tid;
1735 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1736 * on build_path_from_dentry in fs/cifs/dir.c.
1738 * If @stop_on_nosnap, generate path relative to the first non-snapped
1741 * Encode hidden .snap dirs as a double /, i.e.
1742 * foo/.snap/bar -> foo//bar
1744 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1747 struct dentry *temp;
1753 return ERR_PTR(-EINVAL);
1757 seq = read_seqbegin(&rename_lock);
1759 for (temp = dentry; !IS_ROOT(temp);) {
1760 struct inode *inode = d_inode(temp);
1761 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1762 len++; /* slash only */
1763 else if (stop_on_nosnap && inode &&
1764 ceph_snap(inode) == CEPH_NOSNAP)
1767 len += 1 + temp->d_name.len;
1768 temp = temp->d_parent;
1772 len--; /* no leading '/' */
1774 path = kmalloc(len+1, GFP_NOFS);
1776 return ERR_PTR(-ENOMEM);
1778 path[pos] = 0; /* trailing null */
1780 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1781 struct inode *inode;
1783 spin_lock(&temp->d_lock);
1784 inode = d_inode(temp);
1785 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1786 dout("build_path path+%d: %p SNAPDIR\n",
1788 } else if (stop_on_nosnap && inode &&
1789 ceph_snap(inode) == CEPH_NOSNAP) {
1790 spin_unlock(&temp->d_lock);
1793 pos -= temp->d_name.len;
1795 spin_unlock(&temp->d_lock);
1798 strncpy(path + pos, temp->d_name.name,
1801 spin_unlock(&temp->d_lock);
1804 temp = temp->d_parent;
1807 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1808 pr_err("build_path did not end path lookup where "
1809 "expected, namelen is %d, pos is %d\n", len, pos);
1810 /* presumably this is only possible if racing with a
1811 rename of one of the parent directories (we can not
1812 lock the dentries above us to prevent this, but
1813 retrying should be harmless) */
1818 *base = ceph_ino(d_inode(temp));
1820 dout("build_path on %p %d built %llx '%.*s'\n",
1821 dentry, d_count(dentry), *base, len, path);
1825 static int build_dentry_path(struct dentry *dentry,
1826 const char **ppath, int *ppathlen, u64 *pino,
1831 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1832 *pino = ceph_ino(d_inode(dentry->d_parent));
1833 *ppath = dentry->d_name.name;
1834 *ppathlen = dentry->d_name.len;
1837 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1839 return PTR_ERR(path);
1845 static int build_inode_path(struct inode *inode,
1846 const char **ppath, int *ppathlen, u64 *pino,
1849 struct dentry *dentry;
1852 if (ceph_snap(inode) == CEPH_NOSNAP) {
1853 *pino = ceph_ino(inode);
1857 dentry = d_find_alias(inode);
1858 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1861 return PTR_ERR(path);
1868 * request arguments may be specified via an inode *, a dentry *, or
1869 * an explicit ino+path.
1871 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1872 const char *rpath, u64 rino,
1873 const char **ppath, int *pathlen,
1874 u64 *ino, int *freepath)
1879 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1880 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1882 } else if (rdentry) {
1883 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1884 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1886 } else if (rpath || rino) {
1889 *pathlen = rpath ? strlen(rpath) : 0;
1890 dout(" path %.*s\n", *pathlen, rpath);
1897 * called under mdsc->mutex
1899 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1900 struct ceph_mds_request *req,
1901 int mds, bool drop_cap_releases)
1903 struct ceph_msg *msg;
1904 struct ceph_mds_request_head *head;
1905 const char *path1 = NULL;
1906 const char *path2 = NULL;
1907 u64 ino1 = 0, ino2 = 0;
1908 int pathlen1 = 0, pathlen2 = 0;
1909 int freepath1 = 0, freepath2 = 0;
1915 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1916 req->r_path1, req->r_ino1.ino,
1917 &path1, &pathlen1, &ino1, &freepath1);
1923 ret = set_request_path_attr(NULL, req->r_old_dentry,
1924 req->r_path2, req->r_ino2.ino,
1925 &path2, &pathlen2, &ino2, &freepath2);
1931 len = sizeof(*head) +
1932 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1933 sizeof(struct timespec);
1935 /* calculate (max) length for cap releases */
1936 len += sizeof(struct ceph_mds_request_release) *
1937 (!!req->r_inode_drop + !!req->r_dentry_drop +
1938 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1939 if (req->r_dentry_drop)
1940 len += req->r_dentry->d_name.len;
1941 if (req->r_old_dentry_drop)
1942 len += req->r_old_dentry->d_name.len;
1944 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1946 msg = ERR_PTR(-ENOMEM);
1950 msg->hdr.version = cpu_to_le16(2);
1951 msg->hdr.tid = cpu_to_le64(req->r_tid);
1953 head = msg->front.iov_base;
1954 p = msg->front.iov_base + sizeof(*head);
1955 end = msg->front.iov_base + msg->front.iov_len;
1957 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1958 head->op = cpu_to_le32(req->r_op);
1959 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1960 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1961 head->args = req->r_args;
1963 ceph_encode_filepath(&p, end, ino1, path1);
1964 ceph_encode_filepath(&p, end, ino2, path2);
1966 /* make note of release offset, in case we need to replay */
1967 req->r_request_release_offset = p - msg->front.iov_base;
1971 if (req->r_inode_drop)
1972 releases += ceph_encode_inode_release(&p,
1973 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1974 mds, req->r_inode_drop, req->r_inode_unless, 0);
1975 if (req->r_dentry_drop)
1976 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1977 mds, req->r_dentry_drop, req->r_dentry_unless);
1978 if (req->r_old_dentry_drop)
1979 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1980 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1981 if (req->r_old_inode_drop)
1982 releases += ceph_encode_inode_release(&p,
1983 d_inode(req->r_old_dentry),
1984 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1986 if (drop_cap_releases) {
1988 p = msg->front.iov_base + req->r_request_release_offset;
1991 head->num_releases = cpu_to_le16(releases);
1995 struct ceph_timespec ts;
1996 ceph_encode_timespec(&ts, &req->r_stamp);
1997 ceph_encode_copy(&p, &ts, sizeof(ts));
2001 msg->front.iov_len = p - msg->front.iov_base;
2002 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2004 if (req->r_pagelist) {
2005 struct ceph_pagelist *pagelist = req->r_pagelist;
2006 atomic_inc(&pagelist->refcnt);
2007 ceph_msg_data_add_pagelist(msg, pagelist);
2008 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2010 msg->hdr.data_len = 0;
2013 msg->hdr.data_off = cpu_to_le16(0);
2017 kfree((char *)path2);
2020 kfree((char *)path1);
2026 * called under mdsc->mutex if error, under no mutex if
2029 static void complete_request(struct ceph_mds_client *mdsc,
2030 struct ceph_mds_request *req)
2032 if (req->r_callback)
2033 req->r_callback(mdsc, req);
2035 complete_all(&req->r_completion);
2039 * called under mdsc->mutex
2041 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2042 struct ceph_mds_request *req,
2043 int mds, bool drop_cap_releases)
2045 struct ceph_mds_request_head *rhead;
2046 struct ceph_msg *msg;
2051 struct ceph_cap *cap =
2052 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2055 req->r_sent_on_mseq = cap->mseq;
2057 req->r_sent_on_mseq = -1;
2059 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2060 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2062 if (req->r_got_unsafe) {
2065 * Replay. Do not regenerate message (and rebuild
2066 * paths, etc.); just use the original message.
2067 * Rebuilding paths will break for renames because
2068 * d_move mangles the src name.
2070 msg = req->r_request;
2071 rhead = msg->front.iov_base;
2073 flags = le32_to_cpu(rhead->flags);
2074 flags |= CEPH_MDS_FLAG_REPLAY;
2075 rhead->flags = cpu_to_le32(flags);
2077 if (req->r_target_inode)
2078 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2080 rhead->num_retry = req->r_attempts - 1;
2082 /* remove cap/dentry releases from message */
2083 rhead->num_releases = 0;
2086 p = msg->front.iov_base + req->r_request_release_offset;
2088 struct ceph_timespec ts;
2089 ceph_encode_timespec(&ts, &req->r_stamp);
2090 ceph_encode_copy(&p, &ts, sizeof(ts));
2093 msg->front.iov_len = p - msg->front.iov_base;
2094 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2098 if (req->r_request) {
2099 ceph_msg_put(req->r_request);
2100 req->r_request = NULL;
2102 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2104 req->r_err = PTR_ERR(msg);
2105 complete_request(mdsc, req);
2106 return PTR_ERR(msg);
2108 req->r_request = msg;
2110 rhead = msg->front.iov_base;
2111 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2112 if (req->r_got_unsafe)
2113 flags |= CEPH_MDS_FLAG_REPLAY;
2114 if (req->r_locked_dir)
2115 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2116 rhead->flags = cpu_to_le32(flags);
2117 rhead->num_fwd = req->r_num_fwd;
2118 rhead->num_retry = req->r_attempts - 1;
2121 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2126 * send request, or put it on the appropriate wait list.
2128 static int __do_request(struct ceph_mds_client *mdsc,
2129 struct ceph_mds_request *req)
2131 struct ceph_mds_session *session = NULL;
2135 if (req->r_err || req->r_got_result) {
2137 __unregister_request(mdsc, req);
2141 if (req->r_timeout &&
2142 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2143 dout("do_request timed out\n");
2148 put_request_session(req);
2150 mds = __choose_mds(mdsc, req);
2152 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2153 dout("do_request no mds or not active, waiting for map\n");
2154 list_add(&req->r_wait, &mdsc->waiting_for_map);
2158 /* get, open session */
2159 session = __ceph_lookup_mds_session(mdsc, mds);
2161 session = register_session(mdsc, mds);
2162 if (IS_ERR(session)) {
2163 err = PTR_ERR(session);
2167 req->r_session = get_session(session);
2169 dout("do_request mds%d session %p state %s\n", mds, session,
2170 ceph_session_state_name(session->s_state));
2171 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2172 session->s_state != CEPH_MDS_SESSION_HUNG) {
2173 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2174 session->s_state == CEPH_MDS_SESSION_CLOSING)
2175 __open_session(mdsc, session);
2176 list_add(&req->r_wait, &session->s_waiting);
2181 req->r_resend_mds = -1; /* forget any previous mds hint */
2183 if (req->r_request_started == 0) /* note request start time */
2184 req->r_request_started = jiffies;
2186 err = __prepare_send_request(mdsc, req, mds, false);
2188 ceph_msg_get(req->r_request);
2189 ceph_con_send(&session->s_con, req->r_request);
2193 ceph_put_mds_session(session);
2199 complete_request(mdsc, req);
2204 * called under mdsc->mutex
2206 static void __wake_requests(struct ceph_mds_client *mdsc,
2207 struct list_head *head)
2209 struct ceph_mds_request *req;
2210 LIST_HEAD(tmp_list);
2212 list_splice_init(head, &tmp_list);
2214 while (!list_empty(&tmp_list)) {
2215 req = list_entry(tmp_list.next,
2216 struct ceph_mds_request, r_wait);
2217 list_del_init(&req->r_wait);
2218 dout(" wake request %p tid %llu\n", req, req->r_tid);
2219 __do_request(mdsc, req);
2224 * Wake up threads with requests pending for @mds, so that they can
2225 * resubmit their requests to a possibly different mds.
2227 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2229 struct ceph_mds_request *req;
2230 struct rb_node *p = rb_first(&mdsc->request_tree);
2232 dout("kick_requests mds%d\n", mds);
2234 req = rb_entry(p, struct ceph_mds_request, r_node);
2236 if (req->r_got_unsafe)
2238 if (req->r_attempts > 0)
2239 continue; /* only new requests */
2240 if (req->r_session &&
2241 req->r_session->s_mds == mds) {
2242 dout(" kicking tid %llu\n", req->r_tid);
2243 list_del_init(&req->r_wait);
2244 __do_request(mdsc, req);
2249 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2250 struct ceph_mds_request *req)
2252 dout("submit_request on %p\n", req);
2253 mutex_lock(&mdsc->mutex);
2254 __register_request(mdsc, req, NULL);
2255 __do_request(mdsc, req);
2256 mutex_unlock(&mdsc->mutex);
2260 * Synchrously perform an mds request. Take care of all of the
2261 * session setup, forwarding, retry details.
2263 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2265 struct ceph_mds_request *req)
2269 dout("do_request on %p\n", req);
2271 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2273 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2274 if (req->r_locked_dir)
2275 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2276 if (req->r_old_dentry_dir)
2277 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2281 mutex_lock(&mdsc->mutex);
2282 __register_request(mdsc, req, dir);
2283 __do_request(mdsc, req);
2287 __unregister_request(mdsc, req);
2288 dout("do_request early error %d\n", err);
2293 mutex_unlock(&mdsc->mutex);
2294 dout("do_request waiting\n");
2295 if (!req->r_timeout && req->r_wait_for_completion) {
2296 err = req->r_wait_for_completion(mdsc, req);
2298 long timeleft = wait_for_completion_killable_timeout(
2300 ceph_timeout_jiffies(req->r_timeout));
2304 err = -EIO; /* timed out */
2306 err = timeleft; /* killed */
2308 dout("do_request waited, got %d\n", err);
2309 mutex_lock(&mdsc->mutex);
2311 /* only abort if we didn't race with a real reply */
2312 if (req->r_got_result) {
2313 err = le32_to_cpu(req->r_reply_info.head->result);
2314 } else if (err < 0) {
2315 dout("aborted request %lld with %d\n", req->r_tid, err);
2318 * ensure we aren't running concurrently with
2319 * ceph_fill_trace or ceph_readdir_prepopulate, which
2320 * rely on locks (dir mutex) held by our caller.
2322 mutex_lock(&req->r_fill_mutex);
2324 req->r_aborted = true;
2325 mutex_unlock(&req->r_fill_mutex);
2327 if (req->r_locked_dir &&
2328 (req->r_op & CEPH_MDS_OP_WRITE))
2329 ceph_invalidate_dir_request(req);
2335 mutex_unlock(&mdsc->mutex);
2336 dout("do_request %p done, result %d\n", req, err);
2341 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2342 * namespace request.
2344 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2346 struct inode *inode = req->r_locked_dir;
2348 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2350 ceph_dir_clear_complete(inode);
2352 ceph_invalidate_dentry_lease(req->r_dentry);
2353 if (req->r_old_dentry)
2354 ceph_invalidate_dentry_lease(req->r_old_dentry);
2360 * We take the session mutex and parse and process the reply immediately.
2361 * This preserves the logical ordering of replies, capabilities, etc., sent
2362 * by the MDS as they are applied to our local cache.
2364 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2366 struct ceph_mds_client *mdsc = session->s_mdsc;
2367 struct ceph_mds_request *req;
2368 struct ceph_mds_reply_head *head = msg->front.iov_base;
2369 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2370 struct ceph_snap_realm *realm;
2373 int mds = session->s_mds;
2375 if (msg->front.iov_len < sizeof(*head)) {
2376 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2381 /* get request, session */
2382 tid = le64_to_cpu(msg->hdr.tid);
2383 mutex_lock(&mdsc->mutex);
2384 req = __lookup_request(mdsc, tid);
2386 dout("handle_reply on unknown tid %llu\n", tid);
2387 mutex_unlock(&mdsc->mutex);
2390 dout("handle_reply %p\n", req);
2392 /* correct session? */
2393 if (req->r_session != session) {
2394 pr_err("mdsc_handle_reply got %llu on session mds%d"
2395 " not mds%d\n", tid, session->s_mds,
2396 req->r_session ? req->r_session->s_mds : -1);
2397 mutex_unlock(&mdsc->mutex);
2402 if ((req->r_got_unsafe && !head->safe) ||
2403 (req->r_got_safe && head->safe)) {
2404 pr_warn("got a dup %s reply on %llu from mds%d\n",
2405 head->safe ? "safe" : "unsafe", tid, mds);
2406 mutex_unlock(&mdsc->mutex);
2409 if (req->r_got_safe && !head->safe) {
2410 pr_warn("got unsafe after safe on %llu from mds%d\n",
2412 mutex_unlock(&mdsc->mutex);
2416 result = le32_to_cpu(head->result);
2420 * if we're not talking to the authority, send to them
2421 * if the authority has changed while we weren't looking,
2422 * send to new authority
2423 * Otherwise we just have to return an ESTALE
2425 if (result == -ESTALE) {
2426 dout("got ESTALE on request %llu", req->r_tid);
2427 req->r_resend_mds = -1;
2428 if (req->r_direct_mode != USE_AUTH_MDS) {
2429 dout("not using auth, setting for that now");
2430 req->r_direct_mode = USE_AUTH_MDS;
2431 __do_request(mdsc, req);
2432 mutex_unlock(&mdsc->mutex);
2435 int mds = __choose_mds(mdsc, req);
2436 if (mds >= 0 && mds != req->r_session->s_mds) {
2437 dout("but auth changed, so resending");
2438 __do_request(mdsc, req);
2439 mutex_unlock(&mdsc->mutex);
2443 dout("have to return ESTALE on request %llu", req->r_tid);
2448 req->r_got_safe = true;
2449 __unregister_request(mdsc, req);
2451 if (req->r_got_unsafe) {
2453 * We already handled the unsafe response, now do the
2454 * cleanup. No need to examine the response; the MDS
2455 * doesn't include any result info in the safe
2456 * response. And even if it did, there is nothing
2457 * useful we could do with a revised return value.
2459 dout("got safe reply %llu, mds%d\n", tid, mds);
2460 list_del_init(&req->r_unsafe_item);
2462 /* last unsafe request during umount? */
2463 if (mdsc->stopping && !__get_oldest_req(mdsc))
2464 complete_all(&mdsc->safe_umount_waiters);
2465 mutex_unlock(&mdsc->mutex);
2469 req->r_got_unsafe = true;
2470 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2473 dout("handle_reply tid %lld result %d\n", tid, result);
2474 rinfo = &req->r_reply_info;
2475 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2476 mutex_unlock(&mdsc->mutex);
2478 mutex_lock(&session->s_mutex);
2480 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2487 if (rinfo->snapblob_len) {
2488 down_write(&mdsc->snap_rwsem);
2489 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2490 rinfo->snapblob + rinfo->snapblob_len,
2491 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2493 downgrade_write(&mdsc->snap_rwsem);
2495 down_read(&mdsc->snap_rwsem);
2498 /* insert trace into our cache */
2499 mutex_lock(&req->r_fill_mutex);
2500 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2502 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2503 req->r_op == CEPH_MDS_OP_LSSNAP))
2504 ceph_readdir_prepopulate(req, req->r_session);
2505 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2507 mutex_unlock(&req->r_fill_mutex);
2509 up_read(&mdsc->snap_rwsem);
2511 ceph_put_snap_realm(mdsc, realm);
2513 mutex_lock(&mdsc->mutex);
2514 if (!req->r_aborted) {
2520 req->r_got_result = true;
2523 dout("reply arrived after request %lld was aborted\n", tid);
2525 mutex_unlock(&mdsc->mutex);
2527 mutex_unlock(&session->s_mutex);
2529 /* kick calling process */
2530 complete_request(mdsc, req);
2532 ceph_mdsc_put_request(req);
2539 * handle mds notification that our request has been forwarded.
2541 static void handle_forward(struct ceph_mds_client *mdsc,
2542 struct ceph_mds_session *session,
2543 struct ceph_msg *msg)
2545 struct ceph_mds_request *req;
2546 u64 tid = le64_to_cpu(msg->hdr.tid);
2550 void *p = msg->front.iov_base;
2551 void *end = p + msg->front.iov_len;
2553 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2554 next_mds = ceph_decode_32(&p);
2555 fwd_seq = ceph_decode_32(&p);
2557 mutex_lock(&mdsc->mutex);
2558 req = __lookup_request(mdsc, tid);
2560 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2561 goto out; /* dup reply? */
2564 if (req->r_aborted) {
2565 dout("forward tid %llu aborted, unregistering\n", tid);
2566 __unregister_request(mdsc, req);
2567 } else if (fwd_seq <= req->r_num_fwd) {
2568 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2569 tid, next_mds, req->r_num_fwd, fwd_seq);
2571 /* resend. forward race not possible; mds would drop */
2572 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2574 BUG_ON(req->r_got_result);
2575 req->r_attempts = 0;
2576 req->r_num_fwd = fwd_seq;
2577 req->r_resend_mds = next_mds;
2578 put_request_session(req);
2579 __do_request(mdsc, req);
2581 ceph_mdsc_put_request(req);
2583 mutex_unlock(&mdsc->mutex);
2587 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2591 * handle a mds session control message
2593 static void handle_session(struct ceph_mds_session *session,
2594 struct ceph_msg *msg)
2596 struct ceph_mds_client *mdsc = session->s_mdsc;
2599 int mds = session->s_mds;
2600 struct ceph_mds_session_head *h = msg->front.iov_base;
2604 if (msg->front.iov_len != sizeof(*h))
2606 op = le32_to_cpu(h->op);
2607 seq = le64_to_cpu(h->seq);
2609 mutex_lock(&mdsc->mutex);
2610 if (op == CEPH_SESSION_CLOSE)
2611 __unregister_session(mdsc, session);
2612 /* FIXME: this ttl calculation is generous */
2613 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2614 mutex_unlock(&mdsc->mutex);
2616 mutex_lock(&session->s_mutex);
2618 dout("handle_session mds%d %s %p state %s seq %llu\n",
2619 mds, ceph_session_op_name(op), session,
2620 ceph_session_state_name(session->s_state), seq);
2622 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2623 session->s_state = CEPH_MDS_SESSION_OPEN;
2624 pr_info("mds%d came back\n", session->s_mds);
2628 case CEPH_SESSION_OPEN:
2629 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2630 pr_info("mds%d reconnect success\n", session->s_mds);
2631 session->s_state = CEPH_MDS_SESSION_OPEN;
2632 renewed_caps(mdsc, session, 0);
2635 __close_session(mdsc, session);
2638 case CEPH_SESSION_RENEWCAPS:
2639 if (session->s_renew_seq == seq)
2640 renewed_caps(mdsc, session, 1);
2643 case CEPH_SESSION_CLOSE:
2644 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2645 pr_info("mds%d reconnect denied\n", session->s_mds);
2646 cleanup_session_requests(mdsc, session);
2647 remove_session_caps(session);
2648 wake = 2; /* for good measure */
2649 wake_up_all(&mdsc->session_close_wq);
2652 case CEPH_SESSION_STALE:
2653 pr_info("mds%d caps went stale, renewing\n",
2655 spin_lock(&session->s_gen_ttl_lock);
2656 session->s_cap_gen++;
2657 session->s_cap_ttl = jiffies - 1;
2658 spin_unlock(&session->s_gen_ttl_lock);
2659 send_renew_caps(mdsc, session);
2662 case CEPH_SESSION_RECALL_STATE:
2663 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2666 case CEPH_SESSION_FLUSHMSG:
2667 send_flushmsg_ack(mdsc, session, seq);
2670 case CEPH_SESSION_FORCE_RO:
2671 dout("force_session_readonly %p\n", session);
2672 spin_lock(&session->s_cap_lock);
2673 session->s_readonly = true;
2674 spin_unlock(&session->s_cap_lock);
2675 wake_up_session_caps(session, 0);
2679 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2683 mutex_unlock(&session->s_mutex);
2685 mutex_lock(&mdsc->mutex);
2686 __wake_requests(mdsc, &session->s_waiting);
2688 kick_requests(mdsc, mds);
2689 mutex_unlock(&mdsc->mutex);
2694 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2695 (int)msg->front.iov_len);
2702 * called under session->mutex.
2704 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2705 struct ceph_mds_session *session)
2707 struct ceph_mds_request *req, *nreq;
2711 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2713 mutex_lock(&mdsc->mutex);
2714 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2715 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2717 ceph_msg_get(req->r_request);
2718 ceph_con_send(&session->s_con, req->r_request);
2723 * also re-send old requests when MDS enters reconnect stage. So that MDS
2724 * can process completed request in clientreplay stage.
2726 p = rb_first(&mdsc->request_tree);
2728 req = rb_entry(p, struct ceph_mds_request, r_node);
2730 if (req->r_got_unsafe)
2732 if (req->r_attempts == 0)
2733 continue; /* only old requests */
2734 if (req->r_session &&
2735 req->r_session->s_mds == session->s_mds) {
2736 err = __prepare_send_request(mdsc, req,
2737 session->s_mds, true);
2739 ceph_msg_get(req->r_request);
2740 ceph_con_send(&session->s_con, req->r_request);
2744 mutex_unlock(&mdsc->mutex);
2748 * Encode information about a cap for a reconnect with the MDS.
2750 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2754 struct ceph_mds_cap_reconnect v2;
2755 struct ceph_mds_cap_reconnect_v1 v1;
2758 struct ceph_inode_info *ci;
2759 struct ceph_reconnect_state *recon_state = arg;
2760 struct ceph_pagelist *pagelist = recon_state->pagelist;
2764 struct dentry *dentry;
2768 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2769 inode, ceph_vinop(inode), cap, cap->cap_id,
2770 ceph_cap_string(cap->issued));
2771 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2775 dentry = d_find_alias(inode);
2777 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2779 err = PTR_ERR(path);
2786 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2790 spin_lock(&ci->i_ceph_lock);
2791 cap->seq = 0; /* reset cap seq */
2792 cap->issue_seq = 0; /* and issue_seq */
2793 cap->mseq = 0; /* and migrate_seq */
2794 cap->cap_gen = cap->session->s_cap_gen;
2796 if (recon_state->flock) {
2797 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2798 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2799 rec.v2.issued = cpu_to_le32(cap->issued);
2800 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2801 rec.v2.pathbase = cpu_to_le64(pathbase);
2802 rec.v2.flock_len = 0;
2803 reclen = sizeof(rec.v2);
2805 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2806 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2807 rec.v1.issued = cpu_to_le32(cap->issued);
2808 rec.v1.size = cpu_to_le64(inode->i_size);
2809 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2810 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2811 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2812 rec.v1.pathbase = cpu_to_le64(pathbase);
2813 reclen = sizeof(rec.v1);
2815 spin_unlock(&ci->i_ceph_lock);
2817 if (recon_state->flock) {
2818 int num_fcntl_locks, num_flock_locks;
2819 struct ceph_filelock *flocks;
2822 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2823 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2824 sizeof(struct ceph_filelock), GFP_NOFS);
2829 err = ceph_encode_locks_to_buffer(inode, flocks,
2839 * number of encoded locks is stable, so copy to pagelist
2841 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2842 (num_fcntl_locks+num_flock_locks) *
2843 sizeof(struct ceph_filelock));
2844 err = ceph_pagelist_append(pagelist, &rec, reclen);
2846 err = ceph_locks_to_pagelist(flocks, pagelist,
2851 err = ceph_pagelist_append(pagelist, &rec, reclen);
2854 recon_state->nr_caps++;
2864 * If an MDS fails and recovers, clients need to reconnect in order to
2865 * reestablish shared state. This includes all caps issued through
2866 * this session _and_ the snap_realm hierarchy. Because it's not
2867 * clear which snap realms the mds cares about, we send everything we
2868 * know about.. that ensures we'll then get any new info the
2869 * recovering MDS might have.
2871 * This is a relatively heavyweight operation, but it's rare.
2873 * called with mdsc->mutex held.
2875 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2876 struct ceph_mds_session *session)
2878 struct ceph_msg *reply;
2880 int mds = session->s_mds;
2883 struct ceph_pagelist *pagelist;
2884 struct ceph_reconnect_state recon_state;
2886 pr_info("mds%d reconnect start\n", mds);
2888 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2890 goto fail_nopagelist;
2891 ceph_pagelist_init(pagelist);
2893 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2897 mutex_lock(&session->s_mutex);
2898 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2901 dout("session %p state %s\n", session,
2902 ceph_session_state_name(session->s_state));
2904 spin_lock(&session->s_gen_ttl_lock);
2905 session->s_cap_gen++;
2906 spin_unlock(&session->s_gen_ttl_lock);
2908 spin_lock(&session->s_cap_lock);
2909 /* don't know if session is readonly */
2910 session->s_readonly = 0;
2912 * notify __ceph_remove_cap() that we are composing cap reconnect.
2913 * If a cap get released before being added to the cap reconnect,
2914 * __ceph_remove_cap() should skip queuing cap release.
2916 session->s_cap_reconnect = 1;
2917 /* drop old cap expires; we're about to reestablish that state */
2918 cleanup_cap_releases(mdsc, session);
2920 /* trim unused caps to reduce MDS's cache rejoin time */
2921 if (mdsc->fsc->sb->s_root)
2922 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2924 ceph_con_close(&session->s_con);
2925 ceph_con_open(&session->s_con,
2926 CEPH_ENTITY_TYPE_MDS, mds,
2927 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2929 /* replay unsafe requests */
2930 replay_unsafe_requests(mdsc, session);
2932 down_read(&mdsc->snap_rwsem);
2934 /* traverse this session's caps */
2935 s_nr_caps = session->s_nr_caps;
2936 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2940 recon_state.nr_caps = 0;
2941 recon_state.pagelist = pagelist;
2942 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2943 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2947 spin_lock(&session->s_cap_lock);
2948 session->s_cap_reconnect = 0;
2949 spin_unlock(&session->s_cap_lock);
2952 * snaprealms. we provide mds with the ino, seq (version), and
2953 * parent for all of our realms. If the mds has any newer info,
2956 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2957 struct ceph_snap_realm *realm =
2958 rb_entry(p, struct ceph_snap_realm, node);
2959 struct ceph_mds_snaprealm_reconnect sr_rec;
2961 dout(" adding snap realm %llx seq %lld parent %llx\n",
2962 realm->ino, realm->seq, realm->parent_ino);
2963 sr_rec.ino = cpu_to_le64(realm->ino);
2964 sr_rec.seq = cpu_to_le64(realm->seq);
2965 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2966 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2971 if (recon_state.flock)
2972 reply->hdr.version = cpu_to_le16(2);
2974 /* raced with cap release? */
2975 if (s_nr_caps != recon_state.nr_caps) {
2976 struct page *page = list_first_entry(&pagelist->head,
2978 __le32 *addr = kmap_atomic(page);
2979 *addr = cpu_to_le32(recon_state.nr_caps);
2980 kunmap_atomic(addr);
2983 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2984 ceph_msg_data_add_pagelist(reply, pagelist);
2986 ceph_early_kick_flushing_caps(mdsc, session);
2988 ceph_con_send(&session->s_con, reply);
2990 mutex_unlock(&session->s_mutex);
2992 mutex_lock(&mdsc->mutex);
2993 __wake_requests(mdsc, &session->s_waiting);
2994 mutex_unlock(&mdsc->mutex);
2996 up_read(&mdsc->snap_rwsem);
3000 ceph_msg_put(reply);
3001 up_read(&mdsc->snap_rwsem);
3002 mutex_unlock(&session->s_mutex);
3004 ceph_pagelist_release(pagelist);
3006 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3012 * compare old and new mdsmaps, kicking requests
3013 * and closing out old connections as necessary
3015 * called under mdsc->mutex.
3017 static void check_new_map(struct ceph_mds_client *mdsc,
3018 struct ceph_mdsmap *newmap,
3019 struct ceph_mdsmap *oldmap)
3022 int oldstate, newstate;
3023 struct ceph_mds_session *s;
3025 dout("check_new_map new %u old %u\n",
3026 newmap->m_epoch, oldmap->m_epoch);
3028 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3029 if (mdsc->sessions[i] == NULL)
3031 s = mdsc->sessions[i];
3032 oldstate = ceph_mdsmap_get_state(oldmap, i);
3033 newstate = ceph_mdsmap_get_state(newmap, i);
3035 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3036 i, ceph_mds_state_name(oldstate),
3037 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3038 ceph_mds_state_name(newstate),
3039 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3040 ceph_session_state_name(s->s_state));
3042 if (i >= newmap->m_max_mds ||
3043 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3044 ceph_mdsmap_get_addr(newmap, i),
3045 sizeof(struct ceph_entity_addr))) {
3046 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3047 /* the session never opened, just close it
3049 __wake_requests(mdsc, &s->s_waiting);
3050 __unregister_session(mdsc, s);
3053 mutex_unlock(&mdsc->mutex);
3054 mutex_lock(&s->s_mutex);
3055 mutex_lock(&mdsc->mutex);
3056 ceph_con_close(&s->s_con);
3057 mutex_unlock(&s->s_mutex);
3058 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3060 } else if (oldstate == newstate) {
3061 continue; /* nothing new with this mds */
3067 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3068 newstate >= CEPH_MDS_STATE_RECONNECT) {
3069 mutex_unlock(&mdsc->mutex);
3070 send_mds_reconnect(mdsc, s);
3071 mutex_lock(&mdsc->mutex);
3075 * kick request on any mds that has gone active.
3077 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3078 newstate >= CEPH_MDS_STATE_ACTIVE) {
3079 if (oldstate != CEPH_MDS_STATE_CREATING &&
3080 oldstate != CEPH_MDS_STATE_STARTING)
3081 pr_info("mds%d recovery completed\n", s->s_mds);
3082 kick_requests(mdsc, i);
3083 ceph_kick_flushing_caps(mdsc, s);
3084 wake_up_session_caps(s, 1);
3088 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3089 s = mdsc->sessions[i];
3092 if (!ceph_mdsmap_is_laggy(newmap, i))
3094 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3095 s->s_state == CEPH_MDS_SESSION_HUNG ||
3096 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3097 dout(" connecting to export targets of laggy mds%d\n",
3099 __open_export_target_sessions(mdsc, s);
3111 * caller must hold session s_mutex, dentry->d_lock
3113 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3115 struct ceph_dentry_info *di = ceph_dentry(dentry);
3117 ceph_put_mds_session(di->lease_session);
3118 di->lease_session = NULL;
3121 static void handle_lease(struct ceph_mds_client *mdsc,
3122 struct ceph_mds_session *session,
3123 struct ceph_msg *msg)
3125 struct super_block *sb = mdsc->fsc->sb;
3126 struct inode *inode;
3127 struct dentry *parent, *dentry;
3128 struct ceph_dentry_info *di;
3129 int mds = session->s_mds;
3130 struct ceph_mds_lease *h = msg->front.iov_base;
3132 struct ceph_vino vino;
3136 dout("handle_lease from mds%d\n", mds);
3139 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3141 vino.ino = le64_to_cpu(h->ino);
3142 vino.snap = CEPH_NOSNAP;
3143 seq = le32_to_cpu(h->seq);
3144 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3145 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3146 if (dname.len != get_unaligned_le32(h+1))
3150 inode = ceph_find_inode(sb, vino);
3151 dout("handle_lease %s, ino %llx %p %.*s\n",
3152 ceph_lease_op_name(h->action), vino.ino, inode,
3153 dname.len, dname.name);
3155 mutex_lock(&session->s_mutex);
3158 if (inode == NULL) {
3159 dout("handle_lease no inode %llx\n", vino.ino);
3164 parent = d_find_alias(inode);
3166 dout("no parent dentry on inode %p\n", inode);
3168 goto release; /* hrm... */
3170 dname.hash = full_name_hash(dname.name, dname.len);
3171 dentry = d_lookup(parent, &dname);
3176 spin_lock(&dentry->d_lock);
3177 di = ceph_dentry(dentry);
3178 switch (h->action) {
3179 case CEPH_MDS_LEASE_REVOKE:
3180 if (di->lease_session == session) {
3181 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3182 h->seq = cpu_to_le32(di->lease_seq);
3183 __ceph_mdsc_drop_dentry_lease(dentry);
3188 case CEPH_MDS_LEASE_RENEW:
3189 if (di->lease_session == session &&
3190 di->lease_gen == session->s_cap_gen &&
3191 di->lease_renew_from &&
3192 di->lease_renew_after == 0) {
3193 unsigned long duration =
3194 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3196 di->lease_seq = seq;
3197 dentry->d_time = di->lease_renew_from + duration;
3198 di->lease_renew_after = di->lease_renew_from +
3200 di->lease_renew_from = 0;
3204 spin_unlock(&dentry->d_lock);
3211 /* let's just reuse the same message */
3212 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3214 ceph_con_send(&session->s_con, msg);
3218 mutex_unlock(&session->s_mutex);
3222 pr_err("corrupt lease message\n");
3226 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3227 struct inode *inode,
3228 struct dentry *dentry, char action,
3231 struct ceph_msg *msg;
3232 struct ceph_mds_lease *lease;
3233 int len = sizeof(*lease) + sizeof(u32);
3236 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3237 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3238 dnamelen = dentry->d_name.len;
3241 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3244 lease = msg->front.iov_base;
3245 lease->action = action;
3246 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3247 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3248 lease->seq = cpu_to_le32(seq);
3249 put_unaligned_le32(dnamelen, lease + 1);
3250 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3253 * if this is a preemptive lease RELEASE, no need to
3254 * flush request stream, since the actual request will
3257 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3259 ceph_con_send(&session->s_con, msg);
3263 * Preemptively release a lease we expect to invalidate anyway.
3264 * Pass @inode always, @dentry is optional.
3266 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3267 struct dentry *dentry)
3269 struct ceph_dentry_info *di;
3270 struct ceph_mds_session *session;
3273 BUG_ON(inode == NULL);
3274 BUG_ON(dentry == NULL);
3276 /* is dentry lease valid? */
3277 spin_lock(&dentry->d_lock);
3278 di = ceph_dentry(dentry);
3279 if (!di || !di->lease_session ||
3280 di->lease_session->s_mds < 0 ||
3281 di->lease_gen != di->lease_session->s_cap_gen ||
3282 !time_before(jiffies, dentry->d_time)) {
3283 dout("lease_release inode %p dentry %p -- "
3286 spin_unlock(&dentry->d_lock);
3290 /* we do have a lease on this dentry; note mds and seq */
3291 session = ceph_get_mds_session(di->lease_session);
3292 seq = di->lease_seq;
3293 __ceph_mdsc_drop_dentry_lease(dentry);
3294 spin_unlock(&dentry->d_lock);
3296 dout("lease_release inode %p dentry %p to mds%d\n",
3297 inode, dentry, session->s_mds);
3298 ceph_mdsc_lease_send_msg(session, inode, dentry,
3299 CEPH_MDS_LEASE_RELEASE, seq);
3300 ceph_put_mds_session(session);
3304 * drop all leases (and dentry refs) in preparation for umount
3306 static void drop_leases(struct ceph_mds_client *mdsc)
3310 dout("drop_leases\n");
3311 mutex_lock(&mdsc->mutex);
3312 for (i = 0; i < mdsc->max_sessions; i++) {
3313 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3316 mutex_unlock(&mdsc->mutex);
3317 mutex_lock(&s->s_mutex);
3318 mutex_unlock(&s->s_mutex);
3319 ceph_put_mds_session(s);
3320 mutex_lock(&mdsc->mutex);
3322 mutex_unlock(&mdsc->mutex);
3328 * delayed work -- periodically trim expired leases, renew caps with mds
3330 static void schedule_delayed(struct ceph_mds_client *mdsc)
3333 unsigned hz = round_jiffies_relative(HZ * delay);
3334 schedule_delayed_work(&mdsc->delayed_work, hz);
3337 static void delayed_work(struct work_struct *work)
3340 struct ceph_mds_client *mdsc =
3341 container_of(work, struct ceph_mds_client, delayed_work.work);
3345 dout("mdsc delayed_work\n");
3346 ceph_check_delayed_caps(mdsc);
3348 mutex_lock(&mdsc->mutex);
3349 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3350 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3351 mdsc->last_renew_caps);
3353 mdsc->last_renew_caps = jiffies;
3355 for (i = 0; i < mdsc->max_sessions; i++) {
3356 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3359 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3360 dout("resending session close request for mds%d\n",
3362 request_close_session(mdsc, s);
3363 ceph_put_mds_session(s);
3366 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3367 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3368 s->s_state = CEPH_MDS_SESSION_HUNG;
3369 pr_info("mds%d hung\n", s->s_mds);
3372 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3373 /* this mds is failed or recovering, just wait */
3374 ceph_put_mds_session(s);
3377 mutex_unlock(&mdsc->mutex);
3379 mutex_lock(&s->s_mutex);
3381 send_renew_caps(mdsc, s);
3383 ceph_con_keepalive(&s->s_con);
3384 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3385 s->s_state == CEPH_MDS_SESSION_HUNG)
3386 ceph_send_cap_releases(mdsc, s);
3387 mutex_unlock(&s->s_mutex);
3388 ceph_put_mds_session(s);
3390 mutex_lock(&mdsc->mutex);
3392 mutex_unlock(&mdsc->mutex);
3394 schedule_delayed(mdsc);
3397 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3400 struct ceph_mds_client *mdsc;
3402 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3407 mutex_init(&mdsc->mutex);
3408 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3409 if (mdsc->mdsmap == NULL) {
3414 init_completion(&mdsc->safe_umount_waiters);
3415 init_waitqueue_head(&mdsc->session_close_wq);
3416 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3417 mdsc->sessions = NULL;
3418 atomic_set(&mdsc->num_sessions, 0);
3419 mdsc->max_sessions = 0;
3421 mdsc->last_snap_seq = 0;
3422 init_rwsem(&mdsc->snap_rwsem);
3423 mdsc->snap_realms = RB_ROOT;
3424 INIT_LIST_HEAD(&mdsc->snap_empty);
3425 spin_lock_init(&mdsc->snap_empty_lock);
3427 mdsc->oldest_tid = 0;
3428 mdsc->request_tree = RB_ROOT;
3429 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3430 mdsc->last_renew_caps = jiffies;
3431 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3432 spin_lock_init(&mdsc->cap_delay_lock);
3433 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3434 spin_lock_init(&mdsc->snap_flush_lock);
3435 mdsc->last_cap_flush_tid = 1;
3436 mdsc->cap_flush_tree = RB_ROOT;
3437 INIT_LIST_HEAD(&mdsc->cap_dirty);
3438 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3439 mdsc->num_cap_flushing = 0;
3440 spin_lock_init(&mdsc->cap_dirty_lock);
3441 init_waitqueue_head(&mdsc->cap_flushing_wq);
3442 spin_lock_init(&mdsc->dentry_lru_lock);
3443 INIT_LIST_HEAD(&mdsc->dentry_lru);
3445 ceph_caps_init(mdsc);
3446 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3448 init_rwsem(&mdsc->pool_perm_rwsem);
3449 mdsc->pool_perm_tree = RB_ROOT;
3455 * Wait for safe replies on open mds requests. If we time out, drop
3456 * all requests from the tree to avoid dangling dentry refs.
3458 static void wait_requests(struct ceph_mds_client *mdsc)
3460 struct ceph_options *opts = mdsc->fsc->client->options;
3461 struct ceph_mds_request *req;
3463 mutex_lock(&mdsc->mutex);
3464 if (__get_oldest_req(mdsc)) {
3465 mutex_unlock(&mdsc->mutex);
3467 dout("wait_requests waiting for requests\n");
3468 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3469 ceph_timeout_jiffies(opts->mount_timeout));
3471 /* tear down remaining requests */
3472 mutex_lock(&mdsc->mutex);
3473 while ((req = __get_oldest_req(mdsc))) {
3474 dout("wait_requests timed out on tid %llu\n",
3476 __unregister_request(mdsc, req);
3479 mutex_unlock(&mdsc->mutex);
3480 dout("wait_requests done\n");
3484 * called before mount is ro, and before dentries are torn down.
3485 * (hmm, does this still race with new lookups?)
3487 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3489 dout("pre_umount\n");
3493 ceph_flush_dirty_caps(mdsc);
3494 wait_requests(mdsc);
3497 * wait for reply handlers to drop their request refs and
3498 * their inode/dcache refs
3504 * wait for all write mds requests to flush.
3506 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3508 struct ceph_mds_request *req = NULL, *nextreq;
3511 mutex_lock(&mdsc->mutex);
3512 dout("wait_unsafe_requests want %lld\n", want_tid);
3514 req = __get_oldest_req(mdsc);
3515 while (req && req->r_tid <= want_tid) {
3516 /* find next request */
3517 n = rb_next(&req->r_node);
3519 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3522 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3523 (req->r_op & CEPH_MDS_OP_WRITE)) {
3525 ceph_mdsc_get_request(req);
3527 ceph_mdsc_get_request(nextreq);
3528 mutex_unlock(&mdsc->mutex);
3529 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3530 req->r_tid, want_tid);
3531 wait_for_completion(&req->r_safe_completion);
3532 mutex_lock(&mdsc->mutex);
3533 ceph_mdsc_put_request(req);
3535 break; /* next dne before, so we're done! */
3536 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3537 /* next request was removed from tree */
3538 ceph_mdsc_put_request(nextreq);
3541 ceph_mdsc_put_request(nextreq); /* won't go away */
3545 mutex_unlock(&mdsc->mutex);
3546 dout("wait_unsafe_requests done\n");
3549 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3551 u64 want_tid, want_flush, want_snap;
3553 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3557 mutex_lock(&mdsc->mutex);
3558 want_tid = mdsc->last_tid;
3559 mutex_unlock(&mdsc->mutex);
3561 ceph_flush_dirty_caps(mdsc);
3562 spin_lock(&mdsc->cap_dirty_lock);
3563 want_flush = mdsc->last_cap_flush_tid;
3564 spin_unlock(&mdsc->cap_dirty_lock);
3566 down_read(&mdsc->snap_rwsem);
3567 want_snap = mdsc->last_snap_seq;
3568 up_read(&mdsc->snap_rwsem);
3570 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3571 want_tid, want_flush, want_snap);
3573 wait_unsafe_requests(mdsc, want_tid);
3574 wait_caps_flush(mdsc, want_flush, want_snap);
3578 * true if all sessions are closed, or we force unmount
3580 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3582 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3584 return atomic_read(&mdsc->num_sessions) == 0;
3588 * called after sb is ro.
3590 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3592 struct ceph_options *opts = mdsc->fsc->client->options;
3593 struct ceph_mds_session *session;
3596 dout("close_sessions\n");
3598 /* close sessions */
3599 mutex_lock(&mdsc->mutex);
3600 for (i = 0; i < mdsc->max_sessions; i++) {
3601 session = __ceph_lookup_mds_session(mdsc, i);
3604 mutex_unlock(&mdsc->mutex);
3605 mutex_lock(&session->s_mutex);
3606 __close_session(mdsc, session);
3607 mutex_unlock(&session->s_mutex);
3608 ceph_put_mds_session(session);
3609 mutex_lock(&mdsc->mutex);
3611 mutex_unlock(&mdsc->mutex);
3613 dout("waiting for sessions to close\n");
3614 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3615 ceph_timeout_jiffies(opts->mount_timeout));
3617 /* tear down remaining sessions */
3618 mutex_lock(&mdsc->mutex);
3619 for (i = 0; i < mdsc->max_sessions; i++) {
3620 if (mdsc->sessions[i]) {
3621 session = get_session(mdsc->sessions[i]);
3622 __unregister_session(mdsc, session);
3623 mutex_unlock(&mdsc->mutex);
3624 mutex_lock(&session->s_mutex);
3625 remove_session_caps(session);
3626 mutex_unlock(&session->s_mutex);
3627 ceph_put_mds_session(session);
3628 mutex_lock(&mdsc->mutex);
3631 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3632 mutex_unlock(&mdsc->mutex);
3634 ceph_cleanup_empty_realms(mdsc);
3636 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3641 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3644 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3646 ceph_mdsmap_destroy(mdsc->mdsmap);
3647 kfree(mdsc->sessions);
3648 ceph_caps_finalize(mdsc);
3649 ceph_pool_perm_destroy(mdsc);
3652 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3654 struct ceph_mds_client *mdsc = fsc->mdsc;
3656 dout("mdsc_destroy %p\n", mdsc);
3657 ceph_mdsc_stop(mdsc);
3659 /* flush out any connection work with references to us */
3664 dout("mdsc_destroy %p done\n", mdsc);
3669 * handle mds map update.
3671 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3675 void *p = msg->front.iov_base;
3676 void *end = p + msg->front.iov_len;
3677 struct ceph_mdsmap *newmap, *oldmap;
3678 struct ceph_fsid fsid;
3681 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3682 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3683 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3685 epoch = ceph_decode_32(&p);
3686 maplen = ceph_decode_32(&p);
3687 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3689 /* do we need it? */
3690 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3691 mutex_lock(&mdsc->mutex);
3692 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3693 dout("handle_map epoch %u <= our %u\n",
3694 epoch, mdsc->mdsmap->m_epoch);
3695 mutex_unlock(&mdsc->mutex);
3699 newmap = ceph_mdsmap_decode(&p, end);
3700 if (IS_ERR(newmap)) {
3701 err = PTR_ERR(newmap);
3705 /* swap into place */
3707 oldmap = mdsc->mdsmap;
3708 mdsc->mdsmap = newmap;
3709 check_new_map(mdsc, newmap, oldmap);
3710 ceph_mdsmap_destroy(oldmap);
3712 mdsc->mdsmap = newmap; /* first mds map */
3714 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3716 __wake_requests(mdsc, &mdsc->waiting_for_map);
3718 mutex_unlock(&mdsc->mutex);
3719 schedule_delayed(mdsc);
3723 mutex_unlock(&mdsc->mutex);
3725 pr_err("error decoding mdsmap %d\n", err);
3729 static struct ceph_connection *con_get(struct ceph_connection *con)
3731 struct ceph_mds_session *s = con->private;
3733 if (get_session(s)) {
3734 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3737 dout("mdsc con_get %p FAIL\n", s);
3741 static void con_put(struct ceph_connection *con)
3743 struct ceph_mds_session *s = con->private;
3745 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3746 ceph_put_mds_session(s);
3750 * if the client is unresponsive for long enough, the mds will kill
3751 * the session entirely.
3753 static void peer_reset(struct ceph_connection *con)
3755 struct ceph_mds_session *s = con->private;
3756 struct ceph_mds_client *mdsc = s->s_mdsc;
3758 pr_warn("mds%d closed our session\n", s->s_mds);
3759 send_mds_reconnect(mdsc, s);
3762 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3764 struct ceph_mds_session *s = con->private;
3765 struct ceph_mds_client *mdsc = s->s_mdsc;
3766 int type = le16_to_cpu(msg->hdr.type);
3768 mutex_lock(&mdsc->mutex);
3769 if (__verify_registered_session(mdsc, s) < 0) {
3770 mutex_unlock(&mdsc->mutex);
3773 mutex_unlock(&mdsc->mutex);
3776 case CEPH_MSG_MDS_MAP:
3777 ceph_mdsc_handle_map(mdsc, msg);
3779 case CEPH_MSG_CLIENT_SESSION:
3780 handle_session(s, msg);
3782 case CEPH_MSG_CLIENT_REPLY:
3783 handle_reply(s, msg);
3785 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3786 handle_forward(mdsc, s, msg);
3788 case CEPH_MSG_CLIENT_CAPS:
3789 ceph_handle_caps(s, msg);
3791 case CEPH_MSG_CLIENT_SNAP:
3792 ceph_handle_snap(mdsc, s, msg);
3794 case CEPH_MSG_CLIENT_LEASE:
3795 handle_lease(mdsc, s, msg);
3799 pr_err("received unknown message type %d %s\n", type,
3800 ceph_msg_type_name(type));
3811 * Note: returned pointer is the address of a structure that's
3812 * managed separately. Caller must *not* attempt to free it.
3814 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3815 int *proto, int force_new)
3817 struct ceph_mds_session *s = con->private;
3818 struct ceph_mds_client *mdsc = s->s_mdsc;
3819 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3820 struct ceph_auth_handshake *auth = &s->s_auth;
3822 if (force_new && auth->authorizer) {
3823 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3824 auth->authorizer = NULL;
3826 if (!auth->authorizer) {
3827 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3830 return ERR_PTR(ret);
3832 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3835 return ERR_PTR(ret);
3837 *proto = ac->protocol;
3843 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3845 struct ceph_mds_session *s = con->private;
3846 struct ceph_mds_client *mdsc = s->s_mdsc;
3847 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3849 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3852 static int invalidate_authorizer(struct ceph_connection *con)
3854 struct ceph_mds_session *s = con->private;
3855 struct ceph_mds_client *mdsc = s->s_mdsc;
3856 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3858 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3860 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3863 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3864 struct ceph_msg_header *hdr, int *skip)
3866 struct ceph_msg *msg;
3867 int type = (int) le16_to_cpu(hdr->type);
3868 int front_len = (int) le32_to_cpu(hdr->front_len);
3874 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3876 pr_err("unable to allocate msg type %d len %d\n",
3884 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3886 struct ceph_mds_session *s = con->private;
3887 struct ceph_auth_handshake *auth = &s->s_auth;
3888 return ceph_auth_sign_message(auth, msg);
3891 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3893 struct ceph_mds_session *s = con->private;
3894 struct ceph_auth_handshake *auth = &s->s_auth;
3895 return ceph_auth_check_message_signature(auth, msg);
3898 static const struct ceph_connection_operations mds_con_ops = {
3901 .dispatch = dispatch,
3902 .get_authorizer = get_authorizer,
3903 .verify_authorizer_reply = verify_authorizer_reply,
3904 .invalidate_authorizer = invalidate_authorizer,
3905 .peer_reset = peer_reset,
3906 .alloc_msg = mds_alloc_msg,
3907 .sign_message = sign_message,
3908 .check_message_signature = check_message_signature,