1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11 #include <linux/ratelimit.h>
14 #include "mds_client.h"
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
48 struct ceph_reconnect_state {
50 struct ceph_pagelist *pagelist;
54 static void __wake_requests(struct ceph_mds_client *mdsc,
55 struct list_head *head);
57 static const struct ceph_connection_operations mds_con_ops;
65 * parse individual inode info
67 static int parse_reply_info_in(void **p, void *end,
68 struct ceph_mds_reply_info_in *info,
74 *p += sizeof(struct ceph_mds_reply_inode) +
75 sizeof(*info->in->fragtree.splits) *
76 le32_to_cpu(info->in->fragtree.nsplits);
78 ceph_decode_32_safe(p, end, info->symlink_len, bad);
79 ceph_decode_need(p, end, info->symlink_len, bad);
81 *p += info->symlink_len;
83 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
84 ceph_decode_copy_safe(p, end, &info->dir_layout,
85 sizeof(info->dir_layout), bad);
87 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
89 ceph_decode_32_safe(p, end, info->xattr_len, bad);
90 ceph_decode_need(p, end, info->xattr_len, bad);
91 info->xattr_data = *p;
92 *p += info->xattr_len;
94 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
95 ceph_decode_64_safe(p, end, info->inline_version, bad);
96 ceph_decode_32_safe(p, end, info->inline_len, bad);
97 ceph_decode_need(p, end, info->inline_len, bad);
98 info->inline_data = *p;
99 *p += info->inline_len;
101 info->inline_version = CEPH_INLINE_NONE;
103 info->pool_ns_len = 0;
104 info->pool_ns_data = NULL;
105 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
106 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
107 if (info->pool_ns_len > 0) {
108 ceph_decode_need(p, end, info->pool_ns_len, bad);
109 info->pool_ns_data = *p;
110 *p += info->pool_ns_len;
120 * parse a normal reply, which may contain a (dir+)dentry and/or a
123 static int parse_reply_info_trace(void **p, void *end,
124 struct ceph_mds_reply_info_parsed *info,
129 if (info->head->is_dentry) {
130 err = parse_reply_info_in(p, end, &info->diri, features);
134 if (unlikely(*p + sizeof(*info->dirfrag) > end))
137 *p += sizeof(*info->dirfrag) +
138 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
139 if (unlikely(*p > end))
142 ceph_decode_32_safe(p, end, info->dname_len, bad);
143 ceph_decode_need(p, end, info->dname_len, bad);
145 *p += info->dname_len;
147 *p += sizeof(*info->dlease);
150 if (info->head->is_target) {
151 err = parse_reply_info_in(p, end, &info->targeti, features);
156 if (unlikely(*p != end))
163 pr_err("problem parsing mds trace %d\n", err);
168 * parse readdir results
170 static int parse_reply_info_dir(void **p, void *end,
171 struct ceph_mds_reply_info_parsed *info,
178 if (*p + sizeof(*info->dir_dir) > end)
180 *p += sizeof(*info->dir_dir) +
181 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
185 ceph_decode_need(p, end, sizeof(num) + 2, bad);
186 num = ceph_decode_32(p);
188 u16 flags = ceph_decode_16(p);
189 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
190 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
191 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
192 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
197 BUG_ON(!info->dir_entries);
198 if ((unsigned long)(info->dir_entries + num) >
199 (unsigned long)info->dir_entries + info->dir_buf_size) {
200 pr_err("dir contents are larger than expected\n");
207 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
209 ceph_decode_need(p, end, sizeof(u32)*2, bad);
210 rde->name_len = ceph_decode_32(p);
211 ceph_decode_need(p, end, rde->name_len, bad);
214 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
216 *p += sizeof(struct ceph_mds_reply_lease);
219 err = parse_reply_info_in(p, end, &rde->inode, features);
222 /* ceph_readdir_prepopulate() will update it */
236 pr_err("problem parsing dir contents %d\n", err);
241 * parse fcntl F_GETLK results
243 static int parse_reply_info_filelock(void **p, void *end,
244 struct ceph_mds_reply_info_parsed *info,
247 if (*p + sizeof(*info->filelock_reply) > end)
250 info->filelock_reply = *p;
251 *p += sizeof(*info->filelock_reply);
253 if (unlikely(*p != end))
262 * parse create results
264 static int parse_reply_info_create(void **p, void *end,
265 struct ceph_mds_reply_info_parsed *info,
268 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
270 info->has_create_ino = false;
272 info->has_create_ino = true;
273 info->ino = ceph_decode_64(p);
277 if (unlikely(*p != end))
286 * parse extra results
288 static int parse_reply_info_extra(void **p, void *end,
289 struct ceph_mds_reply_info_parsed *info,
292 u32 op = le32_to_cpu(info->head->op);
294 if (op == CEPH_MDS_OP_GETFILELOCK)
295 return parse_reply_info_filelock(p, end, info, features);
296 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
297 return parse_reply_info_dir(p, end, info, features);
298 else if (op == CEPH_MDS_OP_CREATE)
299 return parse_reply_info_create(p, end, info, features);
305 * parse entire mds reply
307 static int parse_reply_info(struct ceph_msg *msg,
308 struct ceph_mds_reply_info_parsed *info,
315 info->head = msg->front.iov_base;
316 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
317 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
320 ceph_decode_32_safe(&p, end, len, bad);
322 ceph_decode_need(&p, end, len, bad);
323 err = parse_reply_info_trace(&p, p+len, info, features);
329 ceph_decode_32_safe(&p, end, len, bad);
331 ceph_decode_need(&p, end, len, bad);
332 err = parse_reply_info_extra(&p, p+len, info, features);
338 ceph_decode_32_safe(&p, end, len, bad);
339 info->snapblob_len = len;
350 pr_err("mds parse_reply err %d\n", err);
354 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
356 if (!info->dir_entries)
358 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
365 const char *ceph_session_state_name(int s)
368 case CEPH_MDS_SESSION_NEW: return "new";
369 case CEPH_MDS_SESSION_OPENING: return "opening";
370 case CEPH_MDS_SESSION_OPEN: return "open";
371 case CEPH_MDS_SESSION_HUNG: return "hung";
372 case CEPH_MDS_SESSION_CLOSING: return "closing";
373 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
374 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
375 case CEPH_MDS_SESSION_REJECTED: return "rejected";
376 default: return "???";
380 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
382 if (refcount_inc_not_zero(&s->s_ref)) {
383 dout("mdsc get_session %p %d -> %d\n", s,
384 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
387 dout("mdsc get_session %p 0 -- FAIL", s);
392 void ceph_put_mds_session(struct ceph_mds_session *s)
394 dout("mdsc put_session %p %d -> %d\n", s,
395 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
396 if (refcount_dec_and_test(&s->s_ref)) {
397 if (s->s_auth.authorizer)
398 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
404 * called under mdsc->mutex
406 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
409 struct ceph_mds_session *session;
411 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
413 session = mdsc->sessions[mds];
414 dout("lookup_mds_session %p %d\n", session,
415 refcount_read(&session->s_ref));
416 get_session(session);
420 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
422 if (mds >= mdsc->max_sessions)
424 return mdsc->sessions[mds];
427 static int __verify_registered_session(struct ceph_mds_client *mdsc,
428 struct ceph_mds_session *s)
430 if (s->s_mds >= mdsc->max_sessions ||
431 mdsc->sessions[s->s_mds] != s)
437 * create+register a new session for given mds.
438 * called under mdsc->mutex.
440 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
443 struct ceph_mds_session *s;
445 if (mds >= mdsc->mdsmap->m_num_mds)
446 return ERR_PTR(-EINVAL);
448 s = kzalloc(sizeof(*s), GFP_NOFS);
450 return ERR_PTR(-ENOMEM);
453 s->s_state = CEPH_MDS_SESSION_NEW;
456 mutex_init(&s->s_mutex);
458 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
460 spin_lock_init(&s->s_gen_ttl_lock);
462 s->s_cap_ttl = jiffies - 1;
464 spin_lock_init(&s->s_cap_lock);
465 s->s_renew_requested = 0;
467 INIT_LIST_HEAD(&s->s_caps);
470 refcount_set(&s->s_ref, 1);
471 INIT_LIST_HEAD(&s->s_waiting);
472 INIT_LIST_HEAD(&s->s_unsafe);
473 s->s_num_cap_releases = 0;
474 s->s_cap_reconnect = 0;
475 s->s_cap_iterator = NULL;
476 INIT_LIST_HEAD(&s->s_cap_releases);
477 INIT_LIST_HEAD(&s->s_cap_flushing);
479 dout("register_session mds%d\n", mds);
480 if (mds >= mdsc->max_sessions) {
481 int newmax = 1 << get_count_order(mds+1);
482 struct ceph_mds_session **sa;
484 dout("register_session realloc to %d\n", newmax);
485 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
488 if (mdsc->sessions) {
489 memcpy(sa, mdsc->sessions,
490 mdsc->max_sessions * sizeof(void *));
491 kfree(mdsc->sessions);
494 mdsc->max_sessions = newmax;
496 mdsc->sessions[mds] = s;
497 atomic_inc(&mdsc->num_sessions);
498 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
500 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
501 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
507 return ERR_PTR(-ENOMEM);
511 * called under mdsc->mutex
513 static void __unregister_session(struct ceph_mds_client *mdsc,
514 struct ceph_mds_session *s)
516 dout("__unregister_session mds%d %p\n", s->s_mds, s);
517 BUG_ON(mdsc->sessions[s->s_mds] != s);
518 mdsc->sessions[s->s_mds] = NULL;
519 ceph_con_close(&s->s_con);
520 ceph_put_mds_session(s);
521 atomic_dec(&mdsc->num_sessions);
525 * drop session refs in request.
527 * should be last request ref, or hold mdsc->mutex
529 static void put_request_session(struct ceph_mds_request *req)
531 if (req->r_session) {
532 ceph_put_mds_session(req->r_session);
533 req->r_session = NULL;
537 void ceph_mdsc_release_request(struct kref *kref)
539 struct ceph_mds_request *req = container_of(kref,
540 struct ceph_mds_request,
542 destroy_reply_info(&req->r_reply_info);
544 ceph_msg_put(req->r_request);
546 ceph_msg_put(req->r_reply);
548 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
552 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
553 iput(req->r_target_inode);
556 if (req->r_old_dentry)
557 dput(req->r_old_dentry);
558 if (req->r_old_dentry_dir) {
560 * track (and drop pins for) r_old_dentry_dir
561 * separately, since r_old_dentry's d_parent may have
562 * changed between the dir mutex being dropped and
563 * this request being freed.
565 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
567 iput(req->r_old_dentry_dir);
572 ceph_pagelist_release(req->r_pagelist);
573 put_request_session(req);
574 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
578 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
581 * lookup session, bump ref if found.
583 * called under mdsc->mutex.
585 static struct ceph_mds_request *
586 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
588 struct ceph_mds_request *req;
590 req = lookup_request(&mdsc->request_tree, tid);
592 ceph_mdsc_get_request(req);
598 * Register an in-flight request, and assign a tid. Link to directory
599 * are modifying (if any).
601 * Called under mdsc->mutex.
603 static void __register_request(struct ceph_mds_client *mdsc,
604 struct ceph_mds_request *req,
607 req->r_tid = ++mdsc->last_tid;
609 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
611 dout("__register_request %p tid %lld\n", req, req->r_tid);
612 ceph_mdsc_get_request(req);
613 insert_request(&mdsc->request_tree, req);
615 req->r_uid = current_fsuid();
616 req->r_gid = current_fsgid();
618 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
619 mdsc->oldest_tid = req->r_tid;
623 req->r_unsafe_dir = dir;
627 static void __unregister_request(struct ceph_mds_client *mdsc,
628 struct ceph_mds_request *req)
630 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
632 /* Never leave an unregistered request on an unsafe list! */
633 list_del_init(&req->r_unsafe_item);
635 if (req->r_tid == mdsc->oldest_tid) {
636 struct rb_node *p = rb_next(&req->r_node);
637 mdsc->oldest_tid = 0;
639 struct ceph_mds_request *next_req =
640 rb_entry(p, struct ceph_mds_request, r_node);
641 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
642 mdsc->oldest_tid = next_req->r_tid;
649 erase_request(&mdsc->request_tree, req);
651 if (req->r_unsafe_dir &&
652 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
653 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
654 spin_lock(&ci->i_unsafe_lock);
655 list_del_init(&req->r_unsafe_dir_item);
656 spin_unlock(&ci->i_unsafe_lock);
658 if (req->r_target_inode &&
659 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
660 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
661 spin_lock(&ci->i_unsafe_lock);
662 list_del_init(&req->r_unsafe_target_item);
663 spin_unlock(&ci->i_unsafe_lock);
666 if (req->r_unsafe_dir) {
667 iput(req->r_unsafe_dir);
668 req->r_unsafe_dir = NULL;
671 complete_all(&req->r_safe_completion);
673 ceph_mdsc_put_request(req);
677 * Walk back up the dentry tree until we hit a dentry representing a
678 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
679 * when calling this) to ensure that the objects won't disappear while we're
680 * working with them. Once we hit a candidate dentry, we attempt to take a
681 * reference to it, and return that as the result.
683 static struct inode *get_nonsnap_parent(struct dentry *dentry)
685 struct inode *inode = NULL;
687 while (dentry && !IS_ROOT(dentry)) {
688 inode = d_inode_rcu(dentry);
689 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
691 dentry = dentry->d_parent;
694 inode = igrab(inode);
699 * Choose mds to send request to next. If there is a hint set in the
700 * request (e.g., due to a prior forward hint from the mds), use that.
701 * Otherwise, consult frag tree and/or caps to identify the
702 * appropriate mds. If all else fails, choose randomly.
704 * Called under mdsc->mutex.
706 static int __choose_mds(struct ceph_mds_client *mdsc,
707 struct ceph_mds_request *req)
710 struct ceph_inode_info *ci;
711 struct ceph_cap *cap;
712 int mode = req->r_direct_mode;
714 u32 hash = req->r_direct_hash;
715 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
718 * is there a specific mds we should try? ignore hint if we have
719 * no session and the mds is not up (active or recovering).
721 if (req->r_resend_mds >= 0 &&
722 (__have_session(mdsc, req->r_resend_mds) ||
723 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
724 dout("choose_mds using resend_mds mds%d\n",
726 return req->r_resend_mds;
729 if (mode == USE_RANDOM_MDS)
734 inode = req->r_inode;
736 } else if (req->r_dentry) {
737 /* ignore race with rename; old or new d_parent is okay */
738 struct dentry *parent;
742 parent = req->r_dentry->d_parent;
743 dir = req->r_parent ? : d_inode_rcu(parent);
745 if (!dir || dir->i_sb != mdsc->fsc->sb) {
746 /* not this fs or parent went negative */
747 inode = d_inode(req->r_dentry);
750 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
751 /* direct snapped/virtual snapdir requests
752 * based on parent dir inode */
753 inode = get_nonsnap_parent(parent);
754 dout("__choose_mds using nonsnap parent %p\n", inode);
757 inode = d_inode(req->r_dentry);
758 if (!inode || mode == USE_AUTH_MDS) {
761 hash = ceph_dentry_hash(dir, req->r_dentry);
770 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
774 ci = ceph_inode(inode);
776 if (is_hash && S_ISDIR(inode->i_mode)) {
777 struct ceph_inode_frag frag;
780 ceph_choose_frag(ci, hash, &frag, &found);
782 if (mode == USE_ANY_MDS && frag.ndist > 0) {
785 /* choose a random replica */
786 get_random_bytes(&r, 1);
789 dout("choose_mds %p %llx.%llx "
790 "frag %u mds%d (%d/%d)\n",
791 inode, ceph_vinop(inode),
794 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
795 CEPH_MDS_STATE_ACTIVE)
799 /* since this file/dir wasn't known to be
800 * replicated, then we want to look for the
801 * authoritative mds. */
804 /* choose auth mds */
806 dout("choose_mds %p %llx.%llx "
807 "frag %u mds%d (auth)\n",
808 inode, ceph_vinop(inode), frag.frag, mds);
809 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
810 CEPH_MDS_STATE_ACTIVE)
816 spin_lock(&ci->i_ceph_lock);
818 if (mode == USE_AUTH_MDS)
819 cap = ci->i_auth_cap;
820 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
821 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
823 spin_unlock(&ci->i_ceph_lock);
827 mds = cap->session->s_mds;
828 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
829 inode, ceph_vinop(inode), mds,
830 cap == ci->i_auth_cap ? "auth " : "", cap);
831 spin_unlock(&ci->i_ceph_lock);
837 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
838 dout("choose_mds chose random mds%d\n", mds);
846 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
848 struct ceph_msg *msg;
849 struct ceph_mds_session_head *h;
851 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
854 pr_err("create_session_msg ENOMEM creating msg\n");
857 h = msg->front.iov_base;
858 h->op = cpu_to_le32(op);
859 h->seq = cpu_to_le64(seq);
865 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
866 * to include additional client metadata fields.
868 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
870 struct ceph_msg *msg;
871 struct ceph_mds_session_head *h;
873 int metadata_bytes = 0;
874 int metadata_key_count = 0;
875 struct ceph_options *opt = mdsc->fsc->client->options;
876 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
879 const char* metadata[][2] = {
880 {"hostname", utsname()->nodename},
881 {"kernel_version", utsname()->release},
882 {"entity_id", opt->name ? : ""},
883 {"root", fsopt->server_path ? : "/"},
887 /* Calculate serialized length of metadata */
888 metadata_bytes = 4; /* map length */
889 for (i = 0; metadata[i][0] != NULL; ++i) {
890 metadata_bytes += 8 + strlen(metadata[i][0]) +
891 strlen(metadata[i][1]);
892 metadata_key_count++;
895 /* Allocate the message */
896 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
899 pr_err("create_session_msg ENOMEM creating msg\n");
902 h = msg->front.iov_base;
903 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
904 h->seq = cpu_to_le64(seq);
907 * Serialize client metadata into waiting buffer space, using
908 * the format that userspace expects for map<string, string>
910 * ClientSession messages with metadata are v2
912 msg->hdr.version = cpu_to_le16(2);
913 msg->hdr.compat_version = cpu_to_le16(1);
915 /* The write pointer, following the session_head structure */
916 p = msg->front.iov_base + sizeof(*h);
918 /* Number of entries in the map */
919 ceph_encode_32(&p, metadata_key_count);
921 /* Two length-prefixed strings for each entry in the map */
922 for (i = 0; metadata[i][0] != NULL; ++i) {
923 size_t const key_len = strlen(metadata[i][0]);
924 size_t const val_len = strlen(metadata[i][1]);
926 ceph_encode_32(&p, key_len);
927 memcpy(p, metadata[i][0], key_len);
929 ceph_encode_32(&p, val_len);
930 memcpy(p, metadata[i][1], val_len);
938 * send session open request.
940 * called under mdsc->mutex
942 static int __open_session(struct ceph_mds_client *mdsc,
943 struct ceph_mds_session *session)
945 struct ceph_msg *msg;
947 int mds = session->s_mds;
949 /* wait for mds to go active? */
950 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
951 dout("open_session to mds%d (%s)\n", mds,
952 ceph_mds_state_name(mstate));
953 session->s_state = CEPH_MDS_SESSION_OPENING;
954 session->s_renew_requested = jiffies;
956 /* send connect message */
957 msg = create_session_open_msg(mdsc, session->s_seq);
960 ceph_con_send(&session->s_con, msg);
965 * open sessions for any export targets for the given mds
967 * called under mdsc->mutex
969 static struct ceph_mds_session *
970 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
972 struct ceph_mds_session *session;
974 session = __ceph_lookup_mds_session(mdsc, target);
976 session = register_session(mdsc, target);
980 if (session->s_state == CEPH_MDS_SESSION_NEW ||
981 session->s_state == CEPH_MDS_SESSION_CLOSING)
982 __open_session(mdsc, session);
987 struct ceph_mds_session *
988 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
990 struct ceph_mds_session *session;
992 dout("open_export_target_session to mds%d\n", target);
994 mutex_lock(&mdsc->mutex);
995 session = __open_export_target_session(mdsc, target);
996 mutex_unlock(&mdsc->mutex);
1001 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1002 struct ceph_mds_session *session)
1004 struct ceph_mds_info *mi;
1005 struct ceph_mds_session *ts;
1006 int i, mds = session->s_mds;
1008 if (mds >= mdsc->mdsmap->m_num_mds)
1011 mi = &mdsc->mdsmap->m_info[mds];
1012 dout("open_export_target_sessions for mds%d (%d targets)\n",
1013 session->s_mds, mi->num_export_targets);
1015 for (i = 0; i < mi->num_export_targets; i++) {
1016 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1018 ceph_put_mds_session(ts);
1022 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1023 struct ceph_mds_session *session)
1025 mutex_lock(&mdsc->mutex);
1026 __open_export_target_sessions(mdsc, session);
1027 mutex_unlock(&mdsc->mutex);
1034 /* caller holds s_cap_lock, we drop it */
1035 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1036 struct ceph_mds_session *session)
1037 __releases(session->s_cap_lock)
1039 LIST_HEAD(tmp_list);
1040 list_splice_init(&session->s_cap_releases, &tmp_list);
1041 session->s_num_cap_releases = 0;
1042 spin_unlock(&session->s_cap_lock);
1044 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1045 while (!list_empty(&tmp_list)) {
1046 struct ceph_cap *cap;
1047 /* zero out the in-progress message */
1048 cap = list_first_entry(&tmp_list,
1049 struct ceph_cap, session_caps);
1050 list_del(&cap->session_caps);
1051 ceph_put_cap(mdsc, cap);
1055 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1056 struct ceph_mds_session *session)
1058 struct ceph_mds_request *req;
1061 dout("cleanup_session_requests mds%d\n", session->s_mds);
1062 mutex_lock(&mdsc->mutex);
1063 while (!list_empty(&session->s_unsafe)) {
1064 req = list_first_entry(&session->s_unsafe,
1065 struct ceph_mds_request, r_unsafe_item);
1066 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1068 __unregister_request(mdsc, req);
1070 /* zero r_attempts, so kick_requests() will re-send requests */
1071 p = rb_first(&mdsc->request_tree);
1073 req = rb_entry(p, struct ceph_mds_request, r_node);
1075 if (req->r_session &&
1076 req->r_session->s_mds == session->s_mds)
1077 req->r_attempts = 0;
1079 mutex_unlock(&mdsc->mutex);
1083 * Helper to safely iterate over all caps associated with a session, with
1084 * special care taken to handle a racing __ceph_remove_cap().
1086 * Caller must hold session s_mutex.
1088 static int iterate_session_caps(struct ceph_mds_session *session,
1089 int (*cb)(struct inode *, struct ceph_cap *,
1092 struct list_head *p;
1093 struct ceph_cap *cap;
1094 struct inode *inode, *last_inode = NULL;
1095 struct ceph_cap *old_cap = NULL;
1098 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1099 spin_lock(&session->s_cap_lock);
1100 p = session->s_caps.next;
1101 while (p != &session->s_caps) {
1102 cap = list_entry(p, struct ceph_cap, session_caps);
1103 inode = igrab(&cap->ci->vfs_inode);
1108 session->s_cap_iterator = cap;
1109 spin_unlock(&session->s_cap_lock);
1116 ceph_put_cap(session->s_mdsc, old_cap);
1120 ret = cb(inode, cap, arg);
1123 spin_lock(&session->s_cap_lock);
1125 if (cap->ci == NULL) {
1126 dout("iterate_session_caps finishing cap %p removal\n",
1128 BUG_ON(cap->session != session);
1129 cap->session = NULL;
1130 list_del_init(&cap->session_caps);
1131 session->s_nr_caps--;
1132 if (cap->queue_release) {
1133 list_add_tail(&cap->session_caps,
1134 &session->s_cap_releases);
1135 session->s_num_cap_releases++;
1137 old_cap = cap; /* put_cap it w/o locks held */
1145 session->s_cap_iterator = NULL;
1146 spin_unlock(&session->s_cap_lock);
1150 ceph_put_cap(session->s_mdsc, old_cap);
1155 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1158 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1159 struct ceph_inode_info *ci = ceph_inode(inode);
1160 LIST_HEAD(to_remove);
1162 bool invalidate = false;
1164 dout("removing cap %p, ci is %p, inode is %p\n",
1165 cap, ci, &ci->vfs_inode);
1166 spin_lock(&ci->i_ceph_lock);
1167 __ceph_remove_cap(cap, false);
1168 if (!ci->i_auth_cap) {
1169 struct ceph_cap_flush *cf;
1170 struct ceph_mds_client *mdsc = fsc->mdsc;
1172 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1174 if (ci->i_wrbuffer_ref > 0 &&
1175 READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1178 while (!list_empty(&ci->i_cap_flush_list)) {
1179 cf = list_first_entry(&ci->i_cap_flush_list,
1180 struct ceph_cap_flush, i_list);
1181 list_move(&cf->i_list, &to_remove);
1184 spin_lock(&mdsc->cap_dirty_lock);
1186 list_for_each_entry(cf, &to_remove, i_list)
1187 list_del(&cf->g_list);
1189 if (!list_empty(&ci->i_dirty_item)) {
1190 pr_warn_ratelimited(
1191 " dropping dirty %s state for %p %lld\n",
1192 ceph_cap_string(ci->i_dirty_caps),
1193 inode, ceph_ino(inode));
1194 ci->i_dirty_caps = 0;
1195 list_del_init(&ci->i_dirty_item);
1198 if (!list_empty(&ci->i_flushing_item)) {
1199 pr_warn_ratelimited(
1200 " dropping dirty+flushing %s state for %p %lld\n",
1201 ceph_cap_string(ci->i_flushing_caps),
1202 inode, ceph_ino(inode));
1203 ci->i_flushing_caps = 0;
1204 list_del_init(&ci->i_flushing_item);
1205 mdsc->num_cap_flushing--;
1208 spin_unlock(&mdsc->cap_dirty_lock);
1210 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1211 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1212 ci->i_prealloc_cap_flush = NULL;
1215 spin_unlock(&ci->i_ceph_lock);
1216 while (!list_empty(&to_remove)) {
1217 struct ceph_cap_flush *cf;
1218 cf = list_first_entry(&to_remove,
1219 struct ceph_cap_flush, i_list);
1220 list_del(&cf->i_list);
1221 ceph_free_cap_flush(cf);
1224 wake_up_all(&ci->i_cap_wq);
1226 ceph_queue_invalidate(inode);
1233 * caller must hold session s_mutex
1235 static void remove_session_caps(struct ceph_mds_session *session)
1237 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1238 struct super_block *sb = fsc->sb;
1239 dout("remove_session_caps on %p\n", session);
1240 iterate_session_caps(session, remove_session_caps_cb, fsc);
1242 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1244 spin_lock(&session->s_cap_lock);
1245 if (session->s_nr_caps > 0) {
1246 struct inode *inode;
1247 struct ceph_cap *cap, *prev = NULL;
1248 struct ceph_vino vino;
1250 * iterate_session_caps() skips inodes that are being
1251 * deleted, we need to wait until deletions are complete.
1252 * __wait_on_freeing_inode() is designed for the job,
1253 * but it is not exported, so use lookup inode function
1256 while (!list_empty(&session->s_caps)) {
1257 cap = list_entry(session->s_caps.next,
1258 struct ceph_cap, session_caps);
1262 vino = cap->ci->i_vino;
1263 spin_unlock(&session->s_cap_lock);
1265 inode = ceph_find_inode(sb, vino);
1268 spin_lock(&session->s_cap_lock);
1272 // drop cap expires and unlock s_cap_lock
1273 cleanup_cap_releases(session->s_mdsc, session);
1275 BUG_ON(session->s_nr_caps > 0);
1276 BUG_ON(!list_empty(&session->s_cap_flushing));
1280 * wake up any threads waiting on this session's caps. if the cap is
1281 * old (didn't get renewed on the client reconnect), remove it now.
1283 * caller must hold s_mutex.
1285 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1288 struct ceph_inode_info *ci = ceph_inode(inode);
1291 spin_lock(&ci->i_ceph_lock);
1292 ci->i_wanted_max_size = 0;
1293 ci->i_requested_max_size = 0;
1294 spin_unlock(&ci->i_ceph_lock);
1296 wake_up_all(&ci->i_cap_wq);
1300 static void wake_up_session_caps(struct ceph_mds_session *session,
1303 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1304 iterate_session_caps(session, wake_up_session_cb,
1305 (void *)(unsigned long)reconnect);
1309 * Send periodic message to MDS renewing all currently held caps. The
1310 * ack will reset the expiration for all caps from this session.
1312 * caller holds s_mutex
1314 static int send_renew_caps(struct ceph_mds_client *mdsc,
1315 struct ceph_mds_session *session)
1317 struct ceph_msg *msg;
1320 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1321 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1322 pr_info("mds%d caps stale\n", session->s_mds);
1323 session->s_renew_requested = jiffies;
1325 /* do not try to renew caps until a recovering mds has reconnected
1326 * with its clients. */
1327 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1328 if (state < CEPH_MDS_STATE_RECONNECT) {
1329 dout("send_renew_caps ignoring mds%d (%s)\n",
1330 session->s_mds, ceph_mds_state_name(state));
1334 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1335 ceph_mds_state_name(state));
1336 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1337 ++session->s_renew_seq);
1340 ceph_con_send(&session->s_con, msg);
1344 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1345 struct ceph_mds_session *session, u64 seq)
1347 struct ceph_msg *msg;
1349 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1350 session->s_mds, ceph_session_state_name(session->s_state), seq);
1351 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1354 ceph_con_send(&session->s_con, msg);
1360 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1362 * Called under session->s_mutex
1364 static void renewed_caps(struct ceph_mds_client *mdsc,
1365 struct ceph_mds_session *session, int is_renew)
1370 spin_lock(&session->s_cap_lock);
1371 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1373 session->s_cap_ttl = session->s_renew_requested +
1374 mdsc->mdsmap->m_session_timeout*HZ;
1377 if (time_before(jiffies, session->s_cap_ttl)) {
1378 pr_info("mds%d caps renewed\n", session->s_mds);
1381 pr_info("mds%d caps still stale\n", session->s_mds);
1384 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1385 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1386 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1387 spin_unlock(&session->s_cap_lock);
1390 wake_up_session_caps(session, 0);
1394 * send a session close request
1396 static int request_close_session(struct ceph_mds_client *mdsc,
1397 struct ceph_mds_session *session)
1399 struct ceph_msg *msg;
1401 dout("request_close_session mds%d state %s seq %lld\n",
1402 session->s_mds, ceph_session_state_name(session->s_state),
1404 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1407 ceph_con_send(&session->s_con, msg);
1412 * Called with s_mutex held.
1414 static int __close_session(struct ceph_mds_client *mdsc,
1415 struct ceph_mds_session *session)
1417 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1419 session->s_state = CEPH_MDS_SESSION_CLOSING;
1420 return request_close_session(mdsc, session);
1424 * Trim old(er) caps.
1426 * Because we can't cache an inode without one or more caps, we do
1427 * this indirectly: if a cap is unused, we prune its aliases, at which
1428 * point the inode will hopefully get dropped to.
1430 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1431 * memory pressure from the MDS, though, so it needn't be perfect.
1433 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1435 struct ceph_mds_session *session = arg;
1436 struct ceph_inode_info *ci = ceph_inode(inode);
1437 int used, wanted, oissued, mine;
1439 if (session->s_trim_caps <= 0)
1442 spin_lock(&ci->i_ceph_lock);
1443 mine = cap->issued | cap->implemented;
1444 used = __ceph_caps_used(ci);
1445 wanted = __ceph_caps_file_wanted(ci);
1446 oissued = __ceph_caps_issued_other(ci, cap);
1448 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1449 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1450 ceph_cap_string(used), ceph_cap_string(wanted));
1451 if (cap == ci->i_auth_cap) {
1452 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1453 !list_empty(&ci->i_cap_snaps))
1455 if ((used | wanted) & CEPH_CAP_ANY_WR)
1458 /* The inode has cached pages, but it's no longer used.
1459 * we can safely drop it */
1460 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1461 !(oissued & CEPH_CAP_FILE_CACHE)) {
1465 if ((used | wanted) & ~oissued & mine)
1466 goto out; /* we need these caps */
1468 session->s_trim_caps--;
1470 /* we aren't the only cap.. just remove us */
1471 __ceph_remove_cap(cap, true);
1473 /* try dropping referring dentries */
1474 spin_unlock(&ci->i_ceph_lock);
1475 d_prune_aliases(inode);
1476 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1477 inode, cap, atomic_read(&inode->i_count));
1482 spin_unlock(&ci->i_ceph_lock);
1487 * Trim session cap count down to some max number.
1489 static int trim_caps(struct ceph_mds_client *mdsc,
1490 struct ceph_mds_session *session,
1493 int trim_caps = session->s_nr_caps - max_caps;
1495 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1496 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1497 if (trim_caps > 0) {
1498 session->s_trim_caps = trim_caps;
1499 iterate_session_caps(session, trim_caps_cb, session);
1500 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1501 session->s_mds, session->s_nr_caps, max_caps,
1502 trim_caps - session->s_trim_caps);
1503 session->s_trim_caps = 0;
1506 ceph_send_cap_releases(mdsc, session);
1510 static int check_caps_flush(struct ceph_mds_client *mdsc,
1515 spin_lock(&mdsc->cap_dirty_lock);
1516 if (!list_empty(&mdsc->cap_flush_list)) {
1517 struct ceph_cap_flush *cf =
1518 list_first_entry(&mdsc->cap_flush_list,
1519 struct ceph_cap_flush, g_list);
1520 if (cf->tid <= want_flush_tid) {
1521 dout("check_caps_flush still flushing tid "
1522 "%llu <= %llu\n", cf->tid, want_flush_tid);
1526 spin_unlock(&mdsc->cap_dirty_lock);
1531 * flush all dirty inode data to disk.
1533 * returns true if we've flushed through want_flush_tid
1535 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1538 dout("check_caps_flush want %llu\n", want_flush_tid);
1540 wait_event(mdsc->cap_flushing_wq,
1541 check_caps_flush(mdsc, want_flush_tid));
1543 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1547 * called under s_mutex
1549 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1550 struct ceph_mds_session *session)
1552 struct ceph_msg *msg = NULL;
1553 struct ceph_mds_cap_release *head;
1554 struct ceph_mds_cap_item *item;
1555 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1556 struct ceph_cap *cap;
1557 LIST_HEAD(tmp_list);
1558 int num_cap_releases;
1559 __le32 barrier, *cap_barrier;
1561 down_read(&osdc->lock);
1562 barrier = cpu_to_le32(osdc->epoch_barrier);
1563 up_read(&osdc->lock);
1565 spin_lock(&session->s_cap_lock);
1567 list_splice_init(&session->s_cap_releases, &tmp_list);
1568 num_cap_releases = session->s_num_cap_releases;
1569 session->s_num_cap_releases = 0;
1570 spin_unlock(&session->s_cap_lock);
1572 while (!list_empty(&tmp_list)) {
1574 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1575 PAGE_SIZE, GFP_NOFS, false);
1578 head = msg->front.iov_base;
1579 head->num = cpu_to_le32(0);
1580 msg->front.iov_len = sizeof(*head);
1582 msg->hdr.version = cpu_to_le16(2);
1583 msg->hdr.compat_version = cpu_to_le16(1);
1586 cap = list_first_entry(&tmp_list, struct ceph_cap,
1588 list_del(&cap->session_caps);
1591 head = msg->front.iov_base;
1592 le32_add_cpu(&head->num, 1);
1593 item = msg->front.iov_base + msg->front.iov_len;
1594 item->ino = cpu_to_le64(cap->cap_ino);
1595 item->cap_id = cpu_to_le64(cap->cap_id);
1596 item->migrate_seq = cpu_to_le32(cap->mseq);
1597 item->seq = cpu_to_le32(cap->issue_seq);
1598 msg->front.iov_len += sizeof(*item);
1600 ceph_put_cap(mdsc, cap);
1602 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1603 // Append cap_barrier field
1604 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1605 *cap_barrier = barrier;
1606 msg->front.iov_len += sizeof(*cap_barrier);
1608 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1609 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1610 ceph_con_send(&session->s_con, msg);
1615 BUG_ON(num_cap_releases != 0);
1617 spin_lock(&session->s_cap_lock);
1618 if (!list_empty(&session->s_cap_releases))
1620 spin_unlock(&session->s_cap_lock);
1623 // Append cap_barrier field
1624 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1625 *cap_barrier = barrier;
1626 msg->front.iov_len += sizeof(*cap_barrier);
1628 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1629 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1630 ceph_con_send(&session->s_con, msg);
1634 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1636 spin_lock(&session->s_cap_lock);
1637 list_splice(&tmp_list, &session->s_cap_releases);
1638 session->s_num_cap_releases += num_cap_releases;
1639 spin_unlock(&session->s_cap_lock);
1646 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1649 struct ceph_inode_info *ci = ceph_inode(dir);
1650 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1651 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1652 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1653 int order, num_entries;
1655 spin_lock(&ci->i_ceph_lock);
1656 num_entries = ci->i_files + ci->i_subdirs;
1657 spin_unlock(&ci->i_ceph_lock);
1658 num_entries = max(num_entries, 1);
1659 num_entries = min(num_entries, opt->max_readdir);
1661 order = get_order(size * num_entries);
1662 while (order >= 0) {
1663 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1666 if (rinfo->dir_entries)
1670 if (!rinfo->dir_entries)
1673 num_entries = (PAGE_SIZE << order) / size;
1674 num_entries = min(num_entries, opt->max_readdir);
1676 rinfo->dir_buf_size = PAGE_SIZE << order;
1677 req->r_num_caps = num_entries + 1;
1678 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1679 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1684 * Create an mds request.
1686 struct ceph_mds_request *
1687 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1689 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1692 return ERR_PTR(-ENOMEM);
1694 mutex_init(&req->r_fill_mutex);
1696 req->r_started = jiffies;
1697 req->r_resend_mds = -1;
1698 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1699 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1701 kref_init(&req->r_kref);
1702 RB_CLEAR_NODE(&req->r_node);
1703 INIT_LIST_HEAD(&req->r_wait);
1704 init_completion(&req->r_completion);
1705 init_completion(&req->r_safe_completion);
1706 INIT_LIST_HEAD(&req->r_unsafe_item);
1708 req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1711 req->r_direct_mode = mode;
1716 * return oldest (lowest) request, tid in request tree, 0 if none.
1718 * called under mdsc->mutex.
1720 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1722 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1724 return rb_entry(rb_first(&mdsc->request_tree),
1725 struct ceph_mds_request, r_node);
1728 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1730 return mdsc->oldest_tid;
1734 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1735 * on build_path_from_dentry in fs/cifs/dir.c.
1737 * If @stop_on_nosnap, generate path relative to the first non-snapped
1740 * Encode hidden .snap dirs as a double /, i.e.
1741 * foo/.snap/bar -> foo//bar
1743 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1746 struct dentry *temp;
1752 return ERR_PTR(-EINVAL);
1756 seq = read_seqbegin(&rename_lock);
1758 for (temp = dentry; !IS_ROOT(temp);) {
1759 struct inode *inode = d_inode(temp);
1760 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1761 len++; /* slash only */
1762 else if (stop_on_nosnap && inode &&
1763 ceph_snap(inode) == CEPH_NOSNAP)
1766 len += 1 + temp->d_name.len;
1767 temp = temp->d_parent;
1771 len--; /* no leading '/' */
1773 path = kmalloc(len+1, GFP_NOFS);
1775 return ERR_PTR(-ENOMEM);
1777 path[pos] = 0; /* trailing null */
1779 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1780 struct inode *inode;
1782 spin_lock(&temp->d_lock);
1783 inode = d_inode(temp);
1784 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1785 dout("build_path path+%d: %p SNAPDIR\n",
1787 } else if (stop_on_nosnap && inode &&
1788 ceph_snap(inode) == CEPH_NOSNAP) {
1789 spin_unlock(&temp->d_lock);
1792 pos -= temp->d_name.len;
1794 spin_unlock(&temp->d_lock);
1797 strncpy(path + pos, temp->d_name.name,
1800 spin_unlock(&temp->d_lock);
1803 temp = temp->d_parent;
1806 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1807 pr_err("build_path did not end path lookup where "
1808 "expected, namelen is %d, pos is %d\n", len, pos);
1809 /* presumably this is only possible if racing with a
1810 rename of one of the parent directories (we can not
1811 lock the dentries above us to prevent this, but
1812 retrying should be harmless) */
1817 *base = ceph_ino(d_inode(temp));
1819 dout("build_path on %p %d built %llx '%.*s'\n",
1820 dentry, d_count(dentry), *base, len, path);
1824 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1825 const char **ppath, int *ppathlen, u64 *pino,
1832 dir = d_inode_rcu(dentry->d_parent);
1833 if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1834 *pino = ceph_ino(dir);
1836 *ppath = dentry->d_name.name;
1837 *ppathlen = dentry->d_name.len;
1841 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1843 return PTR_ERR(path);
1849 static int build_inode_path(struct inode *inode,
1850 const char **ppath, int *ppathlen, u64 *pino,
1853 struct dentry *dentry;
1856 if (ceph_snap(inode) == CEPH_NOSNAP) {
1857 *pino = ceph_ino(inode);
1861 dentry = d_find_alias(inode);
1862 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1865 return PTR_ERR(path);
1872 * request arguments may be specified via an inode *, a dentry *, or
1873 * an explicit ino+path.
1875 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1876 struct inode *rdiri, const char *rpath,
1877 u64 rino, const char **ppath, int *pathlen,
1878 u64 *ino, int *freepath)
1883 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1884 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1886 } else if (rdentry) {
1887 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1889 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1891 } else if (rpath || rino) {
1894 *pathlen = rpath ? strlen(rpath) : 0;
1895 dout(" path %.*s\n", *pathlen, rpath);
1902 * called under mdsc->mutex
1904 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1905 struct ceph_mds_request *req,
1906 int mds, bool drop_cap_releases)
1908 struct ceph_msg *msg;
1909 struct ceph_mds_request_head *head;
1910 const char *path1 = NULL;
1911 const char *path2 = NULL;
1912 u64 ino1 = 0, ino2 = 0;
1913 int pathlen1 = 0, pathlen2 = 0;
1914 int freepath1 = 0, freepath2 = 0;
1920 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1921 req->r_parent, req->r_path1, req->r_ino1.ino,
1922 &path1, &pathlen1, &ino1, &freepath1);
1928 ret = set_request_path_attr(NULL, req->r_old_dentry,
1929 req->r_old_dentry_dir,
1930 req->r_path2, req->r_ino2.ino,
1931 &path2, &pathlen2, &ino2, &freepath2);
1937 len = sizeof(*head) +
1938 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1939 sizeof(struct ceph_timespec);
1941 /* calculate (max) length for cap releases */
1942 len += sizeof(struct ceph_mds_request_release) *
1943 (!!req->r_inode_drop + !!req->r_dentry_drop +
1944 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1945 if (req->r_dentry_drop)
1946 len += req->r_dentry->d_name.len;
1947 if (req->r_old_dentry_drop)
1948 len += req->r_old_dentry->d_name.len;
1950 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1952 msg = ERR_PTR(-ENOMEM);
1956 msg->hdr.version = cpu_to_le16(2);
1957 msg->hdr.tid = cpu_to_le64(req->r_tid);
1959 head = msg->front.iov_base;
1960 p = msg->front.iov_base + sizeof(*head);
1961 end = msg->front.iov_base + msg->front.iov_len;
1963 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1964 head->op = cpu_to_le32(req->r_op);
1965 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1966 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1967 head->args = req->r_args;
1969 ceph_encode_filepath(&p, end, ino1, path1);
1970 ceph_encode_filepath(&p, end, ino2, path2);
1972 /* make note of release offset, in case we need to replay */
1973 req->r_request_release_offset = p - msg->front.iov_base;
1977 if (req->r_inode_drop)
1978 releases += ceph_encode_inode_release(&p,
1979 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1980 mds, req->r_inode_drop, req->r_inode_unless, 0);
1981 if (req->r_dentry_drop)
1982 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1983 req->r_parent, mds, req->r_dentry_drop,
1984 req->r_dentry_unless);
1985 if (req->r_old_dentry_drop)
1986 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1987 req->r_old_dentry_dir, mds,
1988 req->r_old_dentry_drop,
1989 req->r_old_dentry_unless);
1990 if (req->r_old_inode_drop)
1991 releases += ceph_encode_inode_release(&p,
1992 d_inode(req->r_old_dentry),
1993 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1995 if (drop_cap_releases) {
1997 p = msg->front.iov_base + req->r_request_release_offset;
2000 head->num_releases = cpu_to_le16(releases);
2004 struct ceph_timespec ts;
2005 ceph_encode_timespec(&ts, &req->r_stamp);
2006 ceph_encode_copy(&p, &ts, sizeof(ts));
2010 msg->front.iov_len = p - msg->front.iov_base;
2011 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2013 if (req->r_pagelist) {
2014 struct ceph_pagelist *pagelist = req->r_pagelist;
2015 refcount_inc(&pagelist->refcnt);
2016 ceph_msg_data_add_pagelist(msg, pagelist);
2017 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2019 msg->hdr.data_len = 0;
2022 msg->hdr.data_off = cpu_to_le16(0);
2026 kfree((char *)path2);
2029 kfree((char *)path1);
2035 * called under mdsc->mutex if error, under no mutex if
2038 static void complete_request(struct ceph_mds_client *mdsc,
2039 struct ceph_mds_request *req)
2041 if (req->r_callback)
2042 req->r_callback(mdsc, req);
2044 complete_all(&req->r_completion);
2048 * called under mdsc->mutex
2050 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2051 struct ceph_mds_request *req,
2052 int mds, bool drop_cap_releases)
2054 struct ceph_mds_request_head *rhead;
2055 struct ceph_msg *msg;
2060 struct ceph_cap *cap =
2061 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2064 req->r_sent_on_mseq = cap->mseq;
2066 req->r_sent_on_mseq = -1;
2068 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2069 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2071 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2074 * Replay. Do not regenerate message (and rebuild
2075 * paths, etc.); just use the original message.
2076 * Rebuilding paths will break for renames because
2077 * d_move mangles the src name.
2079 msg = req->r_request;
2080 rhead = msg->front.iov_base;
2082 flags = le32_to_cpu(rhead->flags);
2083 flags |= CEPH_MDS_FLAG_REPLAY;
2084 rhead->flags = cpu_to_le32(flags);
2086 if (req->r_target_inode)
2087 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2089 rhead->num_retry = req->r_attempts - 1;
2091 /* remove cap/dentry releases from message */
2092 rhead->num_releases = 0;
2095 p = msg->front.iov_base + req->r_request_release_offset;
2097 struct ceph_timespec ts;
2098 ceph_encode_timespec(&ts, &req->r_stamp);
2099 ceph_encode_copy(&p, &ts, sizeof(ts));
2102 msg->front.iov_len = p - msg->front.iov_base;
2103 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2107 if (req->r_request) {
2108 ceph_msg_put(req->r_request);
2109 req->r_request = NULL;
2111 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2113 req->r_err = PTR_ERR(msg);
2114 return PTR_ERR(msg);
2116 req->r_request = msg;
2118 rhead = msg->front.iov_base;
2119 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2120 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2121 flags |= CEPH_MDS_FLAG_REPLAY;
2123 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2124 rhead->flags = cpu_to_le32(flags);
2125 rhead->num_fwd = req->r_num_fwd;
2126 rhead->num_retry = req->r_attempts - 1;
2129 dout(" r_parent = %p\n", req->r_parent);
2134 * send request, or put it on the appropriate wait list.
2136 static int __do_request(struct ceph_mds_client *mdsc,
2137 struct ceph_mds_request *req)
2139 struct ceph_mds_session *session = NULL;
2143 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2144 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2145 __unregister_request(mdsc, req);
2149 if (req->r_timeout &&
2150 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2151 dout("do_request timed out\n");
2155 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2156 dout("do_request forced umount\n");
2160 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2161 if (mdsc->mdsmap_err) {
2162 err = mdsc->mdsmap_err;
2163 dout("do_request mdsmap err %d\n", err);
2166 if (mdsc->mdsmap->m_epoch == 0) {
2167 dout("do_request no mdsmap, waiting for map\n");
2168 list_add(&req->r_wait, &mdsc->waiting_for_map);
2171 if (!(mdsc->fsc->mount_options->flags &
2172 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2173 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2175 pr_info("probably no mds server is up\n");
2180 put_request_session(req);
2182 mds = __choose_mds(mdsc, req);
2184 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2185 dout("do_request no mds or not active, waiting for map\n");
2186 list_add(&req->r_wait, &mdsc->waiting_for_map);
2190 /* get, open session */
2191 session = __ceph_lookup_mds_session(mdsc, mds);
2193 session = register_session(mdsc, mds);
2194 if (IS_ERR(session)) {
2195 err = PTR_ERR(session);
2199 req->r_session = get_session(session);
2201 dout("do_request mds%d session %p state %s\n", mds, session,
2202 ceph_session_state_name(session->s_state));
2203 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2204 session->s_state != CEPH_MDS_SESSION_HUNG) {
2205 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2209 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2210 session->s_state == CEPH_MDS_SESSION_CLOSING)
2211 __open_session(mdsc, session);
2212 list_add(&req->r_wait, &session->s_waiting);
2217 req->r_resend_mds = -1; /* forget any previous mds hint */
2219 if (req->r_request_started == 0) /* note request start time */
2220 req->r_request_started = jiffies;
2222 err = __prepare_send_request(mdsc, req, mds, false);
2224 ceph_msg_get(req->r_request);
2225 ceph_con_send(&session->s_con, req->r_request);
2229 ceph_put_mds_session(session);
2232 dout("__do_request early error %d\n", err);
2234 complete_request(mdsc, req);
2235 __unregister_request(mdsc, req);
2242 * called under mdsc->mutex
2244 static void __wake_requests(struct ceph_mds_client *mdsc,
2245 struct list_head *head)
2247 struct ceph_mds_request *req;
2248 LIST_HEAD(tmp_list);
2250 list_splice_init(head, &tmp_list);
2252 while (!list_empty(&tmp_list)) {
2253 req = list_entry(tmp_list.next,
2254 struct ceph_mds_request, r_wait);
2255 list_del_init(&req->r_wait);
2256 dout(" wake request %p tid %llu\n", req, req->r_tid);
2257 __do_request(mdsc, req);
2262 * Wake up threads with requests pending for @mds, so that they can
2263 * resubmit their requests to a possibly different mds.
2265 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2267 struct ceph_mds_request *req;
2268 struct rb_node *p = rb_first(&mdsc->request_tree);
2270 dout("kick_requests mds%d\n", mds);
2272 req = rb_entry(p, struct ceph_mds_request, r_node);
2274 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2276 if (req->r_attempts > 0)
2277 continue; /* only new requests */
2278 if (req->r_session &&
2279 req->r_session->s_mds == mds) {
2280 dout(" kicking tid %llu\n", req->r_tid);
2281 list_del_init(&req->r_wait);
2282 __do_request(mdsc, req);
2287 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2288 struct ceph_mds_request *req)
2290 dout("submit_request on %p\n", req);
2291 mutex_lock(&mdsc->mutex);
2292 __register_request(mdsc, req, NULL);
2293 __do_request(mdsc, req);
2294 mutex_unlock(&mdsc->mutex);
2298 * Synchrously perform an mds request. Take care of all of the
2299 * session setup, forwarding, retry details.
2301 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2303 struct ceph_mds_request *req)
2307 dout("do_request on %p\n", req);
2309 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2311 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2313 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2314 if (req->r_old_dentry_dir)
2315 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2319 mutex_lock(&mdsc->mutex);
2320 __register_request(mdsc, req, dir);
2321 __do_request(mdsc, req);
2329 mutex_unlock(&mdsc->mutex);
2330 dout("do_request waiting\n");
2331 if (!req->r_timeout && req->r_wait_for_completion) {
2332 err = req->r_wait_for_completion(mdsc, req);
2334 long timeleft = wait_for_completion_killable_timeout(
2336 ceph_timeout_jiffies(req->r_timeout));
2340 err = -EIO; /* timed out */
2342 err = timeleft; /* killed */
2344 dout("do_request waited, got %d\n", err);
2345 mutex_lock(&mdsc->mutex);
2347 /* only abort if we didn't race with a real reply */
2348 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2349 err = le32_to_cpu(req->r_reply_info.head->result);
2350 } else if (err < 0) {
2351 dout("aborted request %lld with %d\n", req->r_tid, err);
2354 * ensure we aren't running concurrently with
2355 * ceph_fill_trace or ceph_readdir_prepopulate, which
2356 * rely on locks (dir mutex) held by our caller.
2358 mutex_lock(&req->r_fill_mutex);
2360 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2361 mutex_unlock(&req->r_fill_mutex);
2363 if (req->r_parent &&
2364 (req->r_op & CEPH_MDS_OP_WRITE))
2365 ceph_invalidate_dir_request(req);
2371 mutex_unlock(&mdsc->mutex);
2372 dout("do_request %p done, result %d\n", req, err);
2377 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2378 * namespace request.
2380 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2382 struct inode *inode = req->r_parent;
2384 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2386 ceph_dir_clear_complete(inode);
2388 ceph_invalidate_dentry_lease(req->r_dentry);
2389 if (req->r_old_dentry)
2390 ceph_invalidate_dentry_lease(req->r_old_dentry);
2396 * We take the session mutex and parse and process the reply immediately.
2397 * This preserves the logical ordering of replies, capabilities, etc., sent
2398 * by the MDS as they are applied to our local cache.
2400 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2402 struct ceph_mds_client *mdsc = session->s_mdsc;
2403 struct ceph_mds_request *req;
2404 struct ceph_mds_reply_head *head = msg->front.iov_base;
2405 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2406 struct ceph_snap_realm *realm;
2409 int mds = session->s_mds;
2411 if (msg->front.iov_len < sizeof(*head)) {
2412 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2417 /* get request, session */
2418 tid = le64_to_cpu(msg->hdr.tid);
2419 mutex_lock(&mdsc->mutex);
2420 req = lookup_get_request(mdsc, tid);
2422 dout("handle_reply on unknown tid %llu\n", tid);
2423 mutex_unlock(&mdsc->mutex);
2426 dout("handle_reply %p\n", req);
2428 /* correct session? */
2429 if (req->r_session != session) {
2430 pr_err("mdsc_handle_reply got %llu on session mds%d"
2431 " not mds%d\n", tid, session->s_mds,
2432 req->r_session ? req->r_session->s_mds : -1);
2433 mutex_unlock(&mdsc->mutex);
2438 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2439 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2440 pr_warn("got a dup %s reply on %llu from mds%d\n",
2441 head->safe ? "safe" : "unsafe", tid, mds);
2442 mutex_unlock(&mdsc->mutex);
2445 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2446 pr_warn("got unsafe after safe on %llu from mds%d\n",
2448 mutex_unlock(&mdsc->mutex);
2452 result = le32_to_cpu(head->result);
2456 * if we're not talking to the authority, send to them
2457 * if the authority has changed while we weren't looking,
2458 * send to new authority
2459 * Otherwise we just have to return an ESTALE
2461 if (result == -ESTALE) {
2462 dout("got ESTALE on request %llu", req->r_tid);
2463 req->r_resend_mds = -1;
2464 if (req->r_direct_mode != USE_AUTH_MDS) {
2465 dout("not using auth, setting for that now");
2466 req->r_direct_mode = USE_AUTH_MDS;
2467 __do_request(mdsc, req);
2468 mutex_unlock(&mdsc->mutex);
2471 int mds = __choose_mds(mdsc, req);
2472 if (mds >= 0 && mds != req->r_session->s_mds) {
2473 dout("but auth changed, so resending");
2474 __do_request(mdsc, req);
2475 mutex_unlock(&mdsc->mutex);
2479 dout("have to return ESTALE on request %llu", req->r_tid);
2484 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2485 __unregister_request(mdsc, req);
2487 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2489 * We already handled the unsafe response, now do the
2490 * cleanup. No need to examine the response; the MDS
2491 * doesn't include any result info in the safe
2492 * response. And even if it did, there is nothing
2493 * useful we could do with a revised return value.
2495 dout("got safe reply %llu, mds%d\n", tid, mds);
2497 /* last unsafe request during umount? */
2498 if (mdsc->stopping && !__get_oldest_req(mdsc))
2499 complete_all(&mdsc->safe_umount_waiters);
2500 mutex_unlock(&mdsc->mutex);
2504 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2505 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2506 if (req->r_unsafe_dir) {
2507 struct ceph_inode_info *ci =
2508 ceph_inode(req->r_unsafe_dir);
2509 spin_lock(&ci->i_unsafe_lock);
2510 list_add_tail(&req->r_unsafe_dir_item,
2511 &ci->i_unsafe_dirops);
2512 spin_unlock(&ci->i_unsafe_lock);
2516 dout("handle_reply tid %lld result %d\n", tid, result);
2517 rinfo = &req->r_reply_info;
2518 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2519 mutex_unlock(&mdsc->mutex);
2521 mutex_lock(&session->s_mutex);
2523 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2530 if (rinfo->snapblob_len) {
2531 down_write(&mdsc->snap_rwsem);
2532 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2533 rinfo->snapblob + rinfo->snapblob_len,
2534 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2536 downgrade_write(&mdsc->snap_rwsem);
2538 down_read(&mdsc->snap_rwsem);
2541 /* insert trace into our cache */
2542 mutex_lock(&req->r_fill_mutex);
2543 current->journal_info = req;
2544 err = ceph_fill_trace(mdsc->fsc->sb, req);
2546 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2547 req->r_op == CEPH_MDS_OP_LSSNAP))
2548 ceph_readdir_prepopulate(req, req->r_session);
2549 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2551 current->journal_info = NULL;
2552 mutex_unlock(&req->r_fill_mutex);
2554 up_read(&mdsc->snap_rwsem);
2556 ceph_put_snap_realm(mdsc, realm);
2558 if (err == 0 && req->r_target_inode &&
2559 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2560 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2561 spin_lock(&ci->i_unsafe_lock);
2562 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2563 spin_unlock(&ci->i_unsafe_lock);
2566 mutex_lock(&mdsc->mutex);
2567 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2571 req->r_reply = ceph_msg_get(msg);
2572 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2575 dout("reply arrived after request %lld was aborted\n", tid);
2577 mutex_unlock(&mdsc->mutex);
2579 mutex_unlock(&session->s_mutex);
2581 /* kick calling process */
2582 complete_request(mdsc, req);
2584 ceph_mdsc_put_request(req);
2591 * handle mds notification that our request has been forwarded.
2593 static void handle_forward(struct ceph_mds_client *mdsc,
2594 struct ceph_mds_session *session,
2595 struct ceph_msg *msg)
2597 struct ceph_mds_request *req;
2598 u64 tid = le64_to_cpu(msg->hdr.tid);
2602 void *p = msg->front.iov_base;
2603 void *end = p + msg->front.iov_len;
2605 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2606 next_mds = ceph_decode_32(&p);
2607 fwd_seq = ceph_decode_32(&p);
2609 mutex_lock(&mdsc->mutex);
2610 req = lookup_get_request(mdsc, tid);
2612 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2613 goto out; /* dup reply? */
2616 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2617 dout("forward tid %llu aborted, unregistering\n", tid);
2618 __unregister_request(mdsc, req);
2619 } else if (fwd_seq <= req->r_num_fwd) {
2620 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2621 tid, next_mds, req->r_num_fwd, fwd_seq);
2623 /* resend. forward race not possible; mds would drop */
2624 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2626 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2627 req->r_attempts = 0;
2628 req->r_num_fwd = fwd_seq;
2629 req->r_resend_mds = next_mds;
2630 put_request_session(req);
2631 __do_request(mdsc, req);
2633 ceph_mdsc_put_request(req);
2635 mutex_unlock(&mdsc->mutex);
2639 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2643 * handle a mds session control message
2645 static void handle_session(struct ceph_mds_session *session,
2646 struct ceph_msg *msg)
2648 struct ceph_mds_client *mdsc = session->s_mdsc;
2651 int mds = session->s_mds;
2652 struct ceph_mds_session_head *h = msg->front.iov_base;
2656 if (msg->front.iov_len != sizeof(*h))
2658 op = le32_to_cpu(h->op);
2659 seq = le64_to_cpu(h->seq);
2661 mutex_lock(&mdsc->mutex);
2662 if (op == CEPH_SESSION_CLOSE) {
2663 get_session(session);
2664 __unregister_session(mdsc, session);
2666 /* FIXME: this ttl calculation is generous */
2667 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2668 mutex_unlock(&mdsc->mutex);
2670 mutex_lock(&session->s_mutex);
2672 dout("handle_session mds%d %s %p state %s seq %llu\n",
2673 mds, ceph_session_op_name(op), session,
2674 ceph_session_state_name(session->s_state), seq);
2676 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2677 session->s_state = CEPH_MDS_SESSION_OPEN;
2678 pr_info("mds%d came back\n", session->s_mds);
2682 case CEPH_SESSION_OPEN:
2683 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2684 pr_info("mds%d reconnect success\n", session->s_mds);
2685 session->s_state = CEPH_MDS_SESSION_OPEN;
2686 renewed_caps(mdsc, session, 0);
2689 __close_session(mdsc, session);
2692 case CEPH_SESSION_RENEWCAPS:
2693 if (session->s_renew_seq == seq)
2694 renewed_caps(mdsc, session, 1);
2697 case CEPH_SESSION_CLOSE:
2698 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2699 pr_info("mds%d reconnect denied\n", session->s_mds);
2700 cleanup_session_requests(mdsc, session);
2701 remove_session_caps(session);
2702 wake = 2; /* for good measure */
2703 wake_up_all(&mdsc->session_close_wq);
2706 case CEPH_SESSION_STALE:
2707 pr_info("mds%d caps went stale, renewing\n",
2709 spin_lock(&session->s_gen_ttl_lock);
2710 session->s_cap_gen++;
2711 session->s_cap_ttl = jiffies - 1;
2712 spin_unlock(&session->s_gen_ttl_lock);
2713 send_renew_caps(mdsc, session);
2716 case CEPH_SESSION_RECALL_STATE:
2717 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2720 case CEPH_SESSION_FLUSHMSG:
2721 send_flushmsg_ack(mdsc, session, seq);
2724 case CEPH_SESSION_FORCE_RO:
2725 dout("force_session_readonly %p\n", session);
2726 spin_lock(&session->s_cap_lock);
2727 session->s_readonly = true;
2728 spin_unlock(&session->s_cap_lock);
2729 wake_up_session_caps(session, 0);
2732 case CEPH_SESSION_REJECT:
2733 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2734 pr_info("mds%d rejected session\n", session->s_mds);
2735 session->s_state = CEPH_MDS_SESSION_REJECTED;
2736 cleanup_session_requests(mdsc, session);
2737 remove_session_caps(session);
2738 wake = 2; /* for good measure */
2742 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2746 mutex_unlock(&session->s_mutex);
2748 mutex_lock(&mdsc->mutex);
2749 __wake_requests(mdsc, &session->s_waiting);
2751 kick_requests(mdsc, mds);
2752 mutex_unlock(&mdsc->mutex);
2754 if (op == CEPH_SESSION_CLOSE)
2755 ceph_put_mds_session(session);
2759 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2760 (int)msg->front.iov_len);
2767 * called under session->mutex.
2769 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2770 struct ceph_mds_session *session)
2772 struct ceph_mds_request *req, *nreq;
2776 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2778 mutex_lock(&mdsc->mutex);
2779 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2780 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2782 ceph_msg_get(req->r_request);
2783 ceph_con_send(&session->s_con, req->r_request);
2788 * also re-send old requests when MDS enters reconnect stage. So that MDS
2789 * can process completed request in clientreplay stage.
2791 p = rb_first(&mdsc->request_tree);
2793 req = rb_entry(p, struct ceph_mds_request, r_node);
2795 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2797 if (req->r_attempts == 0)
2798 continue; /* only old requests */
2799 if (req->r_session &&
2800 req->r_session->s_mds == session->s_mds) {
2801 err = __prepare_send_request(mdsc, req,
2802 session->s_mds, true);
2804 ceph_msg_get(req->r_request);
2805 ceph_con_send(&session->s_con, req->r_request);
2809 mutex_unlock(&mdsc->mutex);
2813 * Encode information about a cap for a reconnect with the MDS.
2815 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2819 struct ceph_mds_cap_reconnect v2;
2820 struct ceph_mds_cap_reconnect_v1 v1;
2822 struct ceph_inode_info *ci;
2823 struct ceph_reconnect_state *recon_state = arg;
2824 struct ceph_pagelist *pagelist = recon_state->pagelist;
2829 struct dentry *dentry;
2833 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2834 inode, ceph_vinop(inode), cap, cap->cap_id,
2835 ceph_cap_string(cap->issued));
2836 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2840 dentry = d_find_alias(inode);
2842 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2844 err = PTR_ERR(path);
2853 spin_lock(&ci->i_ceph_lock);
2854 cap->seq = 0; /* reset cap seq */
2855 cap->issue_seq = 0; /* and issue_seq */
2856 cap->mseq = 0; /* and migrate_seq */
2857 cap->cap_gen = cap->session->s_cap_gen;
2859 if (recon_state->msg_version >= 2) {
2860 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2861 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2862 rec.v2.issued = cpu_to_le32(cap->issued);
2863 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2864 rec.v2.pathbase = cpu_to_le64(pathbase);
2865 rec.v2.flock_len = 0;
2867 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2868 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2869 rec.v1.issued = cpu_to_le32(cap->issued);
2870 rec.v1.size = cpu_to_le64(inode->i_size);
2871 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2872 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2873 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2874 rec.v1.pathbase = cpu_to_le64(pathbase);
2877 if (list_empty(&ci->i_cap_snaps)) {
2880 struct ceph_cap_snap *capsnap =
2881 list_first_entry(&ci->i_cap_snaps,
2882 struct ceph_cap_snap, ci_item);
2883 snap_follows = capsnap->follows;
2885 spin_unlock(&ci->i_ceph_lock);
2887 if (recon_state->msg_version >= 2) {
2888 int num_fcntl_locks, num_flock_locks;
2889 struct ceph_filelock *flocks;
2890 size_t struct_len, total_len = 0;
2894 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2895 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2896 sizeof(struct ceph_filelock), GFP_NOFS);
2901 err = ceph_encode_locks_to_buffer(inode, flocks,
2911 if (recon_state->msg_version >= 3) {
2912 /* version, compat_version and struct_len */
2913 total_len = 2 * sizeof(u8) + sizeof(u32);
2917 * number of encoded locks is stable, so copy to pagelist
2919 struct_len = 2 * sizeof(u32) +
2920 (num_fcntl_locks + num_flock_locks) *
2921 sizeof(struct ceph_filelock);
2922 rec.v2.flock_len = cpu_to_le32(struct_len);
2924 struct_len += sizeof(rec.v2);
2925 struct_len += sizeof(u32) + pathlen;
2928 struct_len += sizeof(u64); /* snap_follows */
2930 total_len += struct_len;
2931 err = ceph_pagelist_reserve(pagelist, total_len);
2934 if (recon_state->msg_version >= 3) {
2935 ceph_pagelist_encode_8(pagelist, struct_v);
2936 ceph_pagelist_encode_8(pagelist, 1);
2937 ceph_pagelist_encode_32(pagelist, struct_len);
2939 ceph_pagelist_encode_string(pagelist, path, pathlen);
2940 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2941 ceph_locks_to_pagelist(flocks, pagelist,
2945 ceph_pagelist_encode_64(pagelist, snap_follows);
2949 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2950 err = ceph_pagelist_reserve(pagelist, size);
2952 ceph_pagelist_encode_string(pagelist, path, pathlen);
2953 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2957 recon_state->nr_caps++;
2967 * If an MDS fails and recovers, clients need to reconnect in order to
2968 * reestablish shared state. This includes all caps issued through
2969 * this session _and_ the snap_realm hierarchy. Because it's not
2970 * clear which snap realms the mds cares about, we send everything we
2971 * know about.. that ensures we'll then get any new info the
2972 * recovering MDS might have.
2974 * This is a relatively heavyweight operation, but it's rare.
2976 * called with mdsc->mutex held.
2978 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2979 struct ceph_mds_session *session)
2981 struct ceph_msg *reply;
2983 int mds = session->s_mds;
2986 struct ceph_pagelist *pagelist;
2987 struct ceph_reconnect_state recon_state;
2989 pr_info("mds%d reconnect start\n", mds);
2991 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2993 goto fail_nopagelist;
2994 ceph_pagelist_init(pagelist);
2996 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3000 mutex_lock(&session->s_mutex);
3001 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3004 dout("session %p state %s\n", session,
3005 ceph_session_state_name(session->s_state));
3007 spin_lock(&session->s_gen_ttl_lock);
3008 session->s_cap_gen++;
3009 spin_unlock(&session->s_gen_ttl_lock);
3011 spin_lock(&session->s_cap_lock);
3012 /* don't know if session is readonly */
3013 session->s_readonly = 0;
3015 * notify __ceph_remove_cap() that we are composing cap reconnect.
3016 * If a cap get released before being added to the cap reconnect,
3017 * __ceph_remove_cap() should skip queuing cap release.
3019 session->s_cap_reconnect = 1;
3020 /* drop old cap expires; we're about to reestablish that state */
3021 cleanup_cap_releases(mdsc, session);
3023 /* trim unused caps to reduce MDS's cache rejoin time */
3024 if (mdsc->fsc->sb->s_root)
3025 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3027 ceph_con_close(&session->s_con);
3028 ceph_con_open(&session->s_con,
3029 CEPH_ENTITY_TYPE_MDS, mds,
3030 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3032 /* replay unsafe requests */
3033 replay_unsafe_requests(mdsc, session);
3035 down_read(&mdsc->snap_rwsem);
3037 /* traverse this session's caps */
3038 s_nr_caps = session->s_nr_caps;
3039 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3043 recon_state.nr_caps = 0;
3044 recon_state.pagelist = pagelist;
3045 if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3046 recon_state.msg_version = 3;
3047 else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3048 recon_state.msg_version = 2;
3050 recon_state.msg_version = 1;
3051 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3055 spin_lock(&session->s_cap_lock);
3056 session->s_cap_reconnect = 0;
3057 spin_unlock(&session->s_cap_lock);
3060 * snaprealms. we provide mds with the ino, seq (version), and
3061 * parent for all of our realms. If the mds has any newer info,
3064 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3065 struct ceph_snap_realm *realm =
3066 rb_entry(p, struct ceph_snap_realm, node);
3067 struct ceph_mds_snaprealm_reconnect sr_rec;
3069 dout(" adding snap realm %llx seq %lld parent %llx\n",
3070 realm->ino, realm->seq, realm->parent_ino);
3071 sr_rec.ino = cpu_to_le64(realm->ino);
3072 sr_rec.seq = cpu_to_le64(realm->seq);
3073 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3074 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3079 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3081 /* raced with cap release? */
3082 if (s_nr_caps != recon_state.nr_caps) {
3083 struct page *page = list_first_entry(&pagelist->head,
3085 __le32 *addr = kmap_atomic(page);
3086 *addr = cpu_to_le32(recon_state.nr_caps);
3087 kunmap_atomic(addr);
3090 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3091 ceph_msg_data_add_pagelist(reply, pagelist);
3093 ceph_early_kick_flushing_caps(mdsc, session);
3095 ceph_con_send(&session->s_con, reply);
3097 mutex_unlock(&session->s_mutex);
3099 mutex_lock(&mdsc->mutex);
3100 __wake_requests(mdsc, &session->s_waiting);
3101 mutex_unlock(&mdsc->mutex);
3103 up_read(&mdsc->snap_rwsem);
3107 ceph_msg_put(reply);
3108 up_read(&mdsc->snap_rwsem);
3109 mutex_unlock(&session->s_mutex);
3111 ceph_pagelist_release(pagelist);
3113 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3119 * compare old and new mdsmaps, kicking requests
3120 * and closing out old connections as necessary
3122 * called under mdsc->mutex.
3124 static void check_new_map(struct ceph_mds_client *mdsc,
3125 struct ceph_mdsmap *newmap,
3126 struct ceph_mdsmap *oldmap)
3129 int oldstate, newstate;
3130 struct ceph_mds_session *s;
3132 dout("check_new_map new %u old %u\n",
3133 newmap->m_epoch, oldmap->m_epoch);
3135 for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3136 if (mdsc->sessions[i] == NULL)
3138 s = mdsc->sessions[i];
3139 oldstate = ceph_mdsmap_get_state(oldmap, i);
3140 newstate = ceph_mdsmap_get_state(newmap, i);
3142 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3143 i, ceph_mds_state_name(oldstate),
3144 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3145 ceph_mds_state_name(newstate),
3146 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3147 ceph_session_state_name(s->s_state));
3149 if (i >= newmap->m_num_mds ||
3150 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3151 ceph_mdsmap_get_addr(newmap, i),
3152 sizeof(struct ceph_entity_addr))) {
3153 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3154 /* the session never opened, just close it
3157 __unregister_session(mdsc, s);
3158 __wake_requests(mdsc, &s->s_waiting);
3159 ceph_put_mds_session(s);
3160 } else if (i >= newmap->m_num_mds) {
3161 /* force close session for stopped mds */
3163 __unregister_session(mdsc, s);
3164 __wake_requests(mdsc, &s->s_waiting);
3165 kick_requests(mdsc, i);
3166 mutex_unlock(&mdsc->mutex);
3168 mutex_lock(&s->s_mutex);
3169 cleanup_session_requests(mdsc, s);
3170 remove_session_caps(s);
3171 mutex_unlock(&s->s_mutex);
3173 ceph_put_mds_session(s);
3175 mutex_lock(&mdsc->mutex);
3178 mutex_unlock(&mdsc->mutex);
3179 mutex_lock(&s->s_mutex);
3180 mutex_lock(&mdsc->mutex);
3181 ceph_con_close(&s->s_con);
3182 mutex_unlock(&s->s_mutex);
3183 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3185 } else if (oldstate == newstate) {
3186 continue; /* nothing new with this mds */
3192 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3193 newstate >= CEPH_MDS_STATE_RECONNECT) {
3194 mutex_unlock(&mdsc->mutex);
3195 send_mds_reconnect(mdsc, s);
3196 mutex_lock(&mdsc->mutex);
3200 * kick request on any mds that has gone active.
3202 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3203 newstate >= CEPH_MDS_STATE_ACTIVE) {
3204 if (oldstate != CEPH_MDS_STATE_CREATING &&
3205 oldstate != CEPH_MDS_STATE_STARTING)
3206 pr_info("mds%d recovery completed\n", s->s_mds);
3207 kick_requests(mdsc, i);
3208 ceph_kick_flushing_caps(mdsc, s);
3209 wake_up_session_caps(s, 1);
3213 for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3214 s = mdsc->sessions[i];
3217 if (!ceph_mdsmap_is_laggy(newmap, i))
3219 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3220 s->s_state == CEPH_MDS_SESSION_HUNG ||
3221 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3222 dout(" connecting to export targets of laggy mds%d\n",
3224 __open_export_target_sessions(mdsc, s);
3236 * caller must hold session s_mutex, dentry->d_lock
3238 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3240 struct ceph_dentry_info *di = ceph_dentry(dentry);
3242 ceph_put_mds_session(di->lease_session);
3243 di->lease_session = NULL;
3246 static void handle_lease(struct ceph_mds_client *mdsc,
3247 struct ceph_mds_session *session,
3248 struct ceph_msg *msg)
3250 struct super_block *sb = mdsc->fsc->sb;
3251 struct inode *inode;
3252 struct dentry *parent, *dentry;
3253 struct ceph_dentry_info *di;
3254 int mds = session->s_mds;
3255 struct ceph_mds_lease *h = msg->front.iov_base;
3257 struct ceph_vino vino;
3261 dout("handle_lease from mds%d\n", mds);
3264 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3266 vino.ino = le64_to_cpu(h->ino);
3267 vino.snap = CEPH_NOSNAP;
3268 seq = le32_to_cpu(h->seq);
3269 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3270 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3271 if (dname.len != get_unaligned_le32(h+1))
3275 inode = ceph_find_inode(sb, vino);
3276 dout("handle_lease %s, ino %llx %p %.*s\n",
3277 ceph_lease_op_name(h->action), vino.ino, inode,
3278 dname.len, dname.name);
3280 mutex_lock(&session->s_mutex);
3283 if (inode == NULL) {
3284 dout("handle_lease no inode %llx\n", vino.ino);
3289 parent = d_find_alias(inode);
3291 dout("no parent dentry on inode %p\n", inode);
3293 goto release; /* hrm... */
3295 dname.hash = full_name_hash(parent, dname.name, dname.len);
3296 dentry = d_lookup(parent, &dname);
3301 spin_lock(&dentry->d_lock);
3302 di = ceph_dentry(dentry);
3303 switch (h->action) {
3304 case CEPH_MDS_LEASE_REVOKE:
3305 if (di->lease_session == session) {
3306 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3307 h->seq = cpu_to_le32(di->lease_seq);
3308 __ceph_mdsc_drop_dentry_lease(dentry);
3313 case CEPH_MDS_LEASE_RENEW:
3314 if (di->lease_session == session &&
3315 di->lease_gen == session->s_cap_gen &&
3316 di->lease_renew_from &&
3317 di->lease_renew_after == 0) {
3318 unsigned long duration =
3319 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3321 di->lease_seq = seq;
3322 di->time = di->lease_renew_from + duration;
3323 di->lease_renew_after = di->lease_renew_from +
3325 di->lease_renew_from = 0;
3329 spin_unlock(&dentry->d_lock);
3336 /* let's just reuse the same message */
3337 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3339 ceph_con_send(&session->s_con, msg);
3343 mutex_unlock(&session->s_mutex);
3347 pr_err("corrupt lease message\n");
3351 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3352 struct inode *inode,
3353 struct dentry *dentry, char action,
3356 struct ceph_msg *msg;
3357 struct ceph_mds_lease *lease;
3358 int len = sizeof(*lease) + sizeof(u32);
3361 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3362 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3363 dnamelen = dentry->d_name.len;
3366 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3369 lease = msg->front.iov_base;
3370 lease->action = action;
3371 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3372 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3373 lease->seq = cpu_to_le32(seq);
3374 put_unaligned_le32(dnamelen, lease + 1);
3375 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3378 * if this is a preemptive lease RELEASE, no need to
3379 * flush request stream, since the actual request will
3382 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3384 ceph_con_send(&session->s_con, msg);
3388 * drop all leases (and dentry refs) in preparation for umount
3390 static void drop_leases(struct ceph_mds_client *mdsc)
3394 dout("drop_leases\n");
3395 mutex_lock(&mdsc->mutex);
3396 for (i = 0; i < mdsc->max_sessions; i++) {
3397 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3400 mutex_unlock(&mdsc->mutex);
3401 mutex_lock(&s->s_mutex);
3402 mutex_unlock(&s->s_mutex);
3403 ceph_put_mds_session(s);
3404 mutex_lock(&mdsc->mutex);
3406 mutex_unlock(&mdsc->mutex);
3412 * delayed work -- periodically trim expired leases, renew caps with mds
3414 static void schedule_delayed(struct ceph_mds_client *mdsc)
3417 unsigned hz = round_jiffies_relative(HZ * delay);
3418 schedule_delayed_work(&mdsc->delayed_work, hz);
3421 static void delayed_work(struct work_struct *work)
3424 struct ceph_mds_client *mdsc =
3425 container_of(work, struct ceph_mds_client, delayed_work.work);
3429 dout("mdsc delayed_work\n");
3430 ceph_check_delayed_caps(mdsc);
3432 mutex_lock(&mdsc->mutex);
3433 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3434 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3435 mdsc->last_renew_caps);
3437 mdsc->last_renew_caps = jiffies;
3439 for (i = 0; i < mdsc->max_sessions; i++) {
3440 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3443 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3444 dout("resending session close request for mds%d\n",
3446 request_close_session(mdsc, s);
3447 ceph_put_mds_session(s);
3450 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3451 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3452 s->s_state = CEPH_MDS_SESSION_HUNG;
3453 pr_info("mds%d hung\n", s->s_mds);
3456 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3457 /* this mds is failed or recovering, just wait */
3458 ceph_put_mds_session(s);
3461 mutex_unlock(&mdsc->mutex);
3463 mutex_lock(&s->s_mutex);
3465 send_renew_caps(mdsc, s);
3467 ceph_con_keepalive(&s->s_con);
3468 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3469 s->s_state == CEPH_MDS_SESSION_HUNG)
3470 ceph_send_cap_releases(mdsc, s);
3471 mutex_unlock(&s->s_mutex);
3472 ceph_put_mds_session(s);
3474 mutex_lock(&mdsc->mutex);
3476 mutex_unlock(&mdsc->mutex);
3478 schedule_delayed(mdsc);
3481 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3484 struct ceph_mds_client *mdsc;
3486 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3491 mutex_init(&mdsc->mutex);
3492 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3493 if (mdsc->mdsmap == NULL) {
3498 init_completion(&mdsc->safe_umount_waiters);
3499 init_waitqueue_head(&mdsc->session_close_wq);
3500 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3501 mdsc->sessions = NULL;
3502 atomic_set(&mdsc->num_sessions, 0);
3503 mdsc->max_sessions = 0;
3505 mdsc->last_snap_seq = 0;
3506 init_rwsem(&mdsc->snap_rwsem);
3507 mdsc->snap_realms = RB_ROOT;
3508 INIT_LIST_HEAD(&mdsc->snap_empty);
3509 spin_lock_init(&mdsc->snap_empty_lock);
3511 mdsc->oldest_tid = 0;
3512 mdsc->request_tree = RB_ROOT;
3513 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3514 mdsc->last_renew_caps = jiffies;
3515 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3516 spin_lock_init(&mdsc->cap_delay_lock);
3517 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3518 spin_lock_init(&mdsc->snap_flush_lock);
3519 mdsc->last_cap_flush_tid = 1;
3520 INIT_LIST_HEAD(&mdsc->cap_flush_list);
3521 INIT_LIST_HEAD(&mdsc->cap_dirty);
3522 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3523 mdsc->num_cap_flushing = 0;
3524 spin_lock_init(&mdsc->cap_dirty_lock);
3525 init_waitqueue_head(&mdsc->cap_flushing_wq);
3526 spin_lock_init(&mdsc->dentry_lru_lock);
3527 INIT_LIST_HEAD(&mdsc->dentry_lru);
3529 ceph_caps_init(mdsc);
3530 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3532 init_rwsem(&mdsc->pool_perm_rwsem);
3533 mdsc->pool_perm_tree = RB_ROOT;
3539 * Wait for safe replies on open mds requests. If we time out, drop
3540 * all requests from the tree to avoid dangling dentry refs.
3542 static void wait_requests(struct ceph_mds_client *mdsc)
3544 struct ceph_options *opts = mdsc->fsc->client->options;
3545 struct ceph_mds_request *req;
3547 mutex_lock(&mdsc->mutex);
3548 if (__get_oldest_req(mdsc)) {
3549 mutex_unlock(&mdsc->mutex);
3551 dout("wait_requests waiting for requests\n");
3552 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3553 ceph_timeout_jiffies(opts->mount_timeout));
3555 /* tear down remaining requests */
3556 mutex_lock(&mdsc->mutex);
3557 while ((req = __get_oldest_req(mdsc))) {
3558 dout("wait_requests timed out on tid %llu\n",
3560 __unregister_request(mdsc, req);
3563 mutex_unlock(&mdsc->mutex);
3564 dout("wait_requests done\n");
3568 * called before mount is ro, and before dentries are torn down.
3569 * (hmm, does this still race with new lookups?)
3571 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3573 dout("pre_umount\n");
3577 ceph_flush_dirty_caps(mdsc);
3578 wait_requests(mdsc);
3581 * wait for reply handlers to drop their request refs and
3582 * their inode/dcache refs
3588 * wait for all write mds requests to flush.
3590 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3592 struct ceph_mds_request *req = NULL, *nextreq;
3595 mutex_lock(&mdsc->mutex);
3596 dout("wait_unsafe_requests want %lld\n", want_tid);
3598 req = __get_oldest_req(mdsc);
3599 while (req && req->r_tid <= want_tid) {
3600 /* find next request */
3601 n = rb_next(&req->r_node);
3603 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3606 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3607 (req->r_op & CEPH_MDS_OP_WRITE)) {
3609 ceph_mdsc_get_request(req);
3611 ceph_mdsc_get_request(nextreq);
3612 mutex_unlock(&mdsc->mutex);
3613 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3614 req->r_tid, want_tid);
3615 wait_for_completion(&req->r_safe_completion);
3616 mutex_lock(&mdsc->mutex);
3617 ceph_mdsc_put_request(req);
3619 break; /* next dne before, so we're done! */
3620 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3621 /* next request was removed from tree */
3622 ceph_mdsc_put_request(nextreq);
3625 ceph_mdsc_put_request(nextreq); /* won't go away */
3629 mutex_unlock(&mdsc->mutex);
3630 dout("wait_unsafe_requests done\n");
3633 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3635 u64 want_tid, want_flush;
3637 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3641 mutex_lock(&mdsc->mutex);
3642 want_tid = mdsc->last_tid;
3643 mutex_unlock(&mdsc->mutex);
3645 ceph_flush_dirty_caps(mdsc);
3646 spin_lock(&mdsc->cap_dirty_lock);
3647 want_flush = mdsc->last_cap_flush_tid;
3648 if (!list_empty(&mdsc->cap_flush_list)) {
3649 struct ceph_cap_flush *cf =
3650 list_last_entry(&mdsc->cap_flush_list,
3651 struct ceph_cap_flush, g_list);
3654 spin_unlock(&mdsc->cap_dirty_lock);
3656 dout("sync want tid %lld flush_seq %lld\n",
3657 want_tid, want_flush);
3659 wait_unsafe_requests(mdsc, want_tid);
3660 wait_caps_flush(mdsc, want_flush);
3664 * true if all sessions are closed, or we force unmount
3666 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3668 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3670 return atomic_read(&mdsc->num_sessions) <= skipped;
3674 * called after sb is ro.
3676 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3678 struct ceph_options *opts = mdsc->fsc->client->options;
3679 struct ceph_mds_session *session;
3683 dout("close_sessions\n");
3685 /* close sessions */
3686 mutex_lock(&mdsc->mutex);
3687 for (i = 0; i < mdsc->max_sessions; i++) {
3688 session = __ceph_lookup_mds_session(mdsc, i);
3691 mutex_unlock(&mdsc->mutex);
3692 mutex_lock(&session->s_mutex);
3693 if (__close_session(mdsc, session) <= 0)
3695 mutex_unlock(&session->s_mutex);
3696 ceph_put_mds_session(session);
3697 mutex_lock(&mdsc->mutex);
3699 mutex_unlock(&mdsc->mutex);
3701 dout("waiting for sessions to close\n");
3702 wait_event_timeout(mdsc->session_close_wq,
3703 done_closing_sessions(mdsc, skipped),
3704 ceph_timeout_jiffies(opts->mount_timeout));
3706 /* tear down remaining sessions */
3707 mutex_lock(&mdsc->mutex);
3708 for (i = 0; i < mdsc->max_sessions; i++) {
3709 if (mdsc->sessions[i]) {
3710 session = get_session(mdsc->sessions[i]);
3711 __unregister_session(mdsc, session);
3712 mutex_unlock(&mdsc->mutex);
3713 mutex_lock(&session->s_mutex);
3714 remove_session_caps(session);
3715 mutex_unlock(&session->s_mutex);
3716 ceph_put_mds_session(session);
3717 mutex_lock(&mdsc->mutex);
3720 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3721 mutex_unlock(&mdsc->mutex);
3723 ceph_cleanup_empty_realms(mdsc);
3725 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3730 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3732 struct ceph_mds_session *session;
3735 dout("force umount\n");
3737 mutex_lock(&mdsc->mutex);
3738 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3739 session = __ceph_lookup_mds_session(mdsc, mds);
3742 mutex_unlock(&mdsc->mutex);
3743 mutex_lock(&session->s_mutex);
3744 __close_session(mdsc, session);
3745 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3746 cleanup_session_requests(mdsc, session);
3747 remove_session_caps(session);
3749 mutex_unlock(&session->s_mutex);
3750 ceph_put_mds_session(session);
3751 mutex_lock(&mdsc->mutex);
3752 kick_requests(mdsc, mds);
3754 __wake_requests(mdsc, &mdsc->waiting_for_map);
3755 mutex_unlock(&mdsc->mutex);
3758 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3761 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3763 ceph_mdsmap_destroy(mdsc->mdsmap);
3764 kfree(mdsc->sessions);
3765 ceph_caps_finalize(mdsc);
3766 ceph_pool_perm_destroy(mdsc);
3769 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3771 struct ceph_mds_client *mdsc = fsc->mdsc;
3772 dout("mdsc_destroy %p\n", mdsc);
3774 /* flush out any connection work with references to us */
3777 ceph_mdsc_stop(mdsc);
3781 dout("mdsc_destroy %p done\n", mdsc);
3784 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3786 struct ceph_fs_client *fsc = mdsc->fsc;
3787 const char *mds_namespace = fsc->mount_options->mds_namespace;
3788 void *p = msg->front.iov_base;
3789 void *end = p + msg->front.iov_len;
3793 u32 mount_fscid = (u32)-1;
3794 u8 struct_v, struct_cv;
3797 ceph_decode_need(&p, end, sizeof(u32), bad);
3798 epoch = ceph_decode_32(&p);
3800 dout("handle_fsmap epoch %u\n", epoch);
3802 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3803 struct_v = ceph_decode_8(&p);
3804 struct_cv = ceph_decode_8(&p);
3805 map_len = ceph_decode_32(&p);
3807 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3808 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3810 num_fs = ceph_decode_32(&p);
3811 while (num_fs-- > 0) {
3812 void *info_p, *info_end;
3817 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3818 info_v = ceph_decode_8(&p);
3819 info_cv = ceph_decode_8(&p);
3820 info_len = ceph_decode_32(&p);
3821 ceph_decode_need(&p, end, info_len, bad);
3823 info_end = p + info_len;
3826 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3827 fscid = ceph_decode_32(&info_p);
3828 namelen = ceph_decode_32(&info_p);
3829 ceph_decode_need(&info_p, info_end, namelen, bad);
3831 if (mds_namespace &&
3832 strlen(mds_namespace) == namelen &&
3833 !strncmp(mds_namespace, (char *)info_p, namelen)) {
3834 mount_fscid = fscid;
3839 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3840 if (mount_fscid != (u32)-1) {
3841 fsc->client->monc.fs_cluster_id = mount_fscid;
3842 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3844 ceph_monc_renew_subs(&fsc->client->monc);
3851 pr_err("error decoding fsmap\n");
3853 mutex_lock(&mdsc->mutex);
3854 mdsc->mdsmap_err = -ENOENT;
3855 __wake_requests(mdsc, &mdsc->waiting_for_map);
3856 mutex_unlock(&mdsc->mutex);
3861 * handle mds map update.
3863 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3867 void *p = msg->front.iov_base;
3868 void *end = p + msg->front.iov_len;
3869 struct ceph_mdsmap *newmap, *oldmap;
3870 struct ceph_fsid fsid;
3873 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3874 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3875 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3877 epoch = ceph_decode_32(&p);
3878 maplen = ceph_decode_32(&p);
3879 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3881 /* do we need it? */
3882 mutex_lock(&mdsc->mutex);
3883 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3884 dout("handle_map epoch %u <= our %u\n",
3885 epoch, mdsc->mdsmap->m_epoch);
3886 mutex_unlock(&mdsc->mutex);
3890 newmap = ceph_mdsmap_decode(&p, end);
3891 if (IS_ERR(newmap)) {
3892 err = PTR_ERR(newmap);
3896 /* swap into place */
3898 oldmap = mdsc->mdsmap;
3899 mdsc->mdsmap = newmap;
3900 check_new_map(mdsc, newmap, oldmap);
3901 ceph_mdsmap_destroy(oldmap);
3903 mdsc->mdsmap = newmap; /* first mds map */
3905 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3907 __wake_requests(mdsc, &mdsc->waiting_for_map);
3908 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3909 mdsc->mdsmap->m_epoch);
3911 mutex_unlock(&mdsc->mutex);
3912 schedule_delayed(mdsc);
3916 mutex_unlock(&mdsc->mutex);
3918 pr_err("error decoding mdsmap %d\n", err);
3922 static struct ceph_connection *con_get(struct ceph_connection *con)
3924 struct ceph_mds_session *s = con->private;
3926 if (get_session(s)) {
3927 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
3930 dout("mdsc con_get %p FAIL\n", s);
3934 static void con_put(struct ceph_connection *con)
3936 struct ceph_mds_session *s = con->private;
3938 dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
3939 ceph_put_mds_session(s);
3943 * if the client is unresponsive for long enough, the mds will kill
3944 * the session entirely.
3946 static void peer_reset(struct ceph_connection *con)
3948 struct ceph_mds_session *s = con->private;
3949 struct ceph_mds_client *mdsc = s->s_mdsc;
3951 pr_warn("mds%d closed our session\n", s->s_mds);
3952 send_mds_reconnect(mdsc, s);
3955 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3957 struct ceph_mds_session *s = con->private;
3958 struct ceph_mds_client *mdsc = s->s_mdsc;
3959 int type = le16_to_cpu(msg->hdr.type);
3961 mutex_lock(&mdsc->mutex);
3962 if (__verify_registered_session(mdsc, s) < 0) {
3963 mutex_unlock(&mdsc->mutex);
3966 mutex_unlock(&mdsc->mutex);
3969 case CEPH_MSG_MDS_MAP:
3970 ceph_mdsc_handle_mdsmap(mdsc, msg);
3972 case CEPH_MSG_FS_MAP_USER:
3973 ceph_mdsc_handle_fsmap(mdsc, msg);
3975 case CEPH_MSG_CLIENT_SESSION:
3976 handle_session(s, msg);
3978 case CEPH_MSG_CLIENT_REPLY:
3979 handle_reply(s, msg);
3981 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3982 handle_forward(mdsc, s, msg);
3984 case CEPH_MSG_CLIENT_CAPS:
3985 ceph_handle_caps(s, msg);
3987 case CEPH_MSG_CLIENT_SNAP:
3988 ceph_handle_snap(mdsc, s, msg);
3990 case CEPH_MSG_CLIENT_LEASE:
3991 handle_lease(mdsc, s, msg);
3995 pr_err("received unknown message type %d %s\n", type,
3996 ceph_msg_type_name(type));
4007 * Note: returned pointer is the address of a structure that's
4008 * managed separately. Caller must *not* attempt to free it.
4010 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4011 int *proto, int force_new)
4013 struct ceph_mds_session *s = con->private;
4014 struct ceph_mds_client *mdsc = s->s_mdsc;
4015 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4016 struct ceph_auth_handshake *auth = &s->s_auth;
4018 if (force_new && auth->authorizer) {
4019 ceph_auth_destroy_authorizer(auth->authorizer);
4020 auth->authorizer = NULL;
4022 if (!auth->authorizer) {
4023 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4026 return ERR_PTR(ret);
4028 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4031 return ERR_PTR(ret);
4033 *proto = ac->protocol;
4039 static int verify_authorizer_reply(struct ceph_connection *con)
4041 struct ceph_mds_session *s = con->private;
4042 struct ceph_mds_client *mdsc = s->s_mdsc;
4043 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4045 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4048 static int invalidate_authorizer(struct ceph_connection *con)
4050 struct ceph_mds_session *s = con->private;
4051 struct ceph_mds_client *mdsc = s->s_mdsc;
4052 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4054 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4056 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4059 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4060 struct ceph_msg_header *hdr, int *skip)
4062 struct ceph_msg *msg;
4063 int type = (int) le16_to_cpu(hdr->type);
4064 int front_len = (int) le32_to_cpu(hdr->front_len);
4070 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4072 pr_err("unable to allocate msg type %d len %d\n",
4080 static int mds_sign_message(struct ceph_msg *msg)
4082 struct ceph_mds_session *s = msg->con->private;
4083 struct ceph_auth_handshake *auth = &s->s_auth;
4085 return ceph_auth_sign_message(auth, msg);
4088 static int mds_check_message_signature(struct ceph_msg *msg)
4090 struct ceph_mds_session *s = msg->con->private;
4091 struct ceph_auth_handshake *auth = &s->s_auth;
4093 return ceph_auth_check_message_signature(auth, msg);
4096 static const struct ceph_connection_operations mds_con_ops = {
4099 .dispatch = dispatch,
4100 .get_authorizer = get_authorizer,
4101 .verify_authorizer_reply = verify_authorizer_reply,
4102 .invalidate_authorizer = invalidate_authorizer,
4103 .peer_reset = peer_reset,
4104 .alloc_msg = mds_alloc_msg,
4105 .sign_message = mds_sign_message,
4106 .check_message_signature = mds_check_message_signature,