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45 * are no longer valid.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 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
104 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
105 ceph_decode_need(p, end, info->pool_ns_len, bad);
106 *p += info->pool_ns_len;
108 info->pool_ns_len = 0;
117 * parse a normal reply, which may contain a (dir+)dentry and/or a
120 static int parse_reply_info_trace(void **p, void *end,
121 struct ceph_mds_reply_info_parsed *info,
126 if (info->head->is_dentry) {
127 err = parse_reply_info_in(p, end, &info->diri, features);
131 if (unlikely(*p + sizeof(*info->dirfrag) > end))
134 *p += sizeof(*info->dirfrag) +
135 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
136 if (unlikely(*p > end))
139 ceph_decode_32_safe(p, end, info->dname_len, bad);
140 ceph_decode_need(p, end, info->dname_len, bad);
142 *p += info->dname_len;
144 *p += sizeof(*info->dlease);
147 if (info->head->is_target) {
148 err = parse_reply_info_in(p, end, &info->targeti, features);
153 if (unlikely(*p != end))
160 pr_err("problem parsing mds trace %d\n", err);
165 * parse readdir results
167 static int parse_reply_info_dir(void **p, void *end,
168 struct ceph_mds_reply_info_parsed *info,
175 if (*p + sizeof(*info->dir_dir) > end)
177 *p += sizeof(*info->dir_dir) +
178 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
182 ceph_decode_need(p, end, sizeof(num) + 2, bad);
183 num = ceph_decode_32(p);
184 info->dir_end = ceph_decode_8(p);
185 info->dir_complete = ceph_decode_8(p);
189 BUG_ON(!info->dir_in);
190 info->dir_dname = (void *)(info->dir_in + num);
191 info->dir_dname_len = (void *)(info->dir_dname + num);
192 info->dir_dlease = (void *)(info->dir_dname_len + num);
193 if ((unsigned long)(info->dir_dlease + num) >
194 (unsigned long)info->dir_in + info->dir_buf_size) {
195 pr_err("dir contents are larger than expected\n");
203 ceph_decode_need(p, end, sizeof(u32)*2, bad);
204 info->dir_dname_len[i] = ceph_decode_32(p);
205 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
206 info->dir_dname[i] = *p;
207 *p += info->dir_dname_len[i];
208 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
210 info->dir_dlease[i] = *p;
211 *p += sizeof(struct ceph_mds_reply_lease);
214 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
229 pr_err("problem parsing dir contents %d\n", err);
234 * parse fcntl F_GETLK results
236 static int parse_reply_info_filelock(void **p, void *end,
237 struct ceph_mds_reply_info_parsed *info,
240 if (*p + sizeof(*info->filelock_reply) > end)
243 info->filelock_reply = *p;
244 *p += sizeof(*info->filelock_reply);
246 if (unlikely(*p != end))
255 * parse create results
257 static int parse_reply_info_create(void **p, void *end,
258 struct ceph_mds_reply_info_parsed *info,
261 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
263 info->has_create_ino = false;
265 info->has_create_ino = true;
266 info->ino = ceph_decode_64(p);
270 if (unlikely(*p != end))
279 * parse extra results
281 static int parse_reply_info_extra(void **p, void *end,
282 struct ceph_mds_reply_info_parsed *info,
285 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
286 return parse_reply_info_filelock(p, end, info, features);
287 else if (info->head->op == CEPH_MDS_OP_READDIR ||
288 info->head->op == CEPH_MDS_OP_LSSNAP)
289 return parse_reply_info_dir(p, end, info, features);
290 else if (info->head->op == CEPH_MDS_OP_CREATE)
291 return parse_reply_info_create(p, end, info, features);
297 * parse entire mds reply
299 static int parse_reply_info(struct ceph_msg *msg,
300 struct ceph_mds_reply_info_parsed *info,
307 info->head = msg->front.iov_base;
308 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
309 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
312 ceph_decode_32_safe(&p, end, len, bad);
314 ceph_decode_need(&p, end, len, bad);
315 err = parse_reply_info_trace(&p, p+len, info, features);
321 ceph_decode_32_safe(&p, end, len, bad);
323 ceph_decode_need(&p, end, len, bad);
324 err = parse_reply_info_extra(&p, p+len, info, features);
330 ceph_decode_32_safe(&p, end, len, bad);
331 info->snapblob_len = len;
342 pr_err("mds parse_reply err %d\n", err);
346 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
350 free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
357 const char *ceph_session_state_name(int s)
360 case CEPH_MDS_SESSION_NEW: return "new";
361 case CEPH_MDS_SESSION_OPENING: return "opening";
362 case CEPH_MDS_SESSION_OPEN: return "open";
363 case CEPH_MDS_SESSION_HUNG: return "hung";
364 case CEPH_MDS_SESSION_CLOSING: return "closing";
365 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
366 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
367 default: return "???";
371 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
373 if (atomic_inc_not_zero(&s->s_ref)) {
374 dout("mdsc get_session %p %d -> %d\n", s,
375 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
378 dout("mdsc get_session %p 0 -- FAIL", s);
383 void ceph_put_mds_session(struct ceph_mds_session *s)
385 dout("mdsc put_session %p %d -> %d\n", s,
386 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
387 if (atomic_dec_and_test(&s->s_ref)) {
388 if (s->s_auth.authorizer)
389 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
395 * called under mdsc->mutex
397 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
400 struct ceph_mds_session *session;
402 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
404 session = mdsc->sessions[mds];
405 dout("lookup_mds_session %p %d\n", session,
406 atomic_read(&session->s_ref));
407 get_session(session);
411 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
413 if (mds >= mdsc->max_sessions)
415 return mdsc->sessions[mds];
418 static int __verify_registered_session(struct ceph_mds_client *mdsc,
419 struct ceph_mds_session *s)
421 if (s->s_mds >= mdsc->max_sessions ||
422 mdsc->sessions[s->s_mds] != s)
428 * create+register a new session for given mds.
429 * called under mdsc->mutex.
431 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
434 struct ceph_mds_session *s;
436 if (mds >= mdsc->mdsmap->m_max_mds)
437 return ERR_PTR(-EINVAL);
439 s = kzalloc(sizeof(*s), GFP_NOFS);
441 return ERR_PTR(-ENOMEM);
444 s->s_state = CEPH_MDS_SESSION_NEW;
447 mutex_init(&s->s_mutex);
449 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
451 spin_lock_init(&s->s_gen_ttl_lock);
453 s->s_cap_ttl = jiffies - 1;
455 spin_lock_init(&s->s_cap_lock);
456 s->s_renew_requested = 0;
458 INIT_LIST_HEAD(&s->s_caps);
461 atomic_set(&s->s_ref, 1);
462 INIT_LIST_HEAD(&s->s_waiting);
463 INIT_LIST_HEAD(&s->s_unsafe);
464 s->s_num_cap_releases = 0;
465 s->s_cap_reconnect = 0;
466 s->s_cap_iterator = NULL;
467 INIT_LIST_HEAD(&s->s_cap_releases);
468 INIT_LIST_HEAD(&s->s_cap_flushing);
469 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
471 dout("register_session mds%d\n", mds);
472 if (mds >= mdsc->max_sessions) {
473 int newmax = 1 << get_count_order(mds+1);
474 struct ceph_mds_session **sa;
476 dout("register_session realloc to %d\n", newmax);
477 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
480 if (mdsc->sessions) {
481 memcpy(sa, mdsc->sessions,
482 mdsc->max_sessions * sizeof(void *));
483 kfree(mdsc->sessions);
486 mdsc->max_sessions = newmax;
488 mdsc->sessions[mds] = s;
489 atomic_inc(&mdsc->num_sessions);
490 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
492 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
493 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
499 return ERR_PTR(-ENOMEM);
503 * called under mdsc->mutex
505 static void __unregister_session(struct ceph_mds_client *mdsc,
506 struct ceph_mds_session *s)
508 dout("__unregister_session mds%d %p\n", s->s_mds, s);
509 BUG_ON(mdsc->sessions[s->s_mds] != s);
510 mdsc->sessions[s->s_mds] = NULL;
511 ceph_con_close(&s->s_con);
512 ceph_put_mds_session(s);
513 atomic_dec(&mdsc->num_sessions);
517 * drop session refs in request.
519 * should be last request ref, or hold mdsc->mutex
521 static void put_request_session(struct ceph_mds_request *req)
523 if (req->r_session) {
524 ceph_put_mds_session(req->r_session);
525 req->r_session = NULL;
529 void ceph_mdsc_release_request(struct kref *kref)
531 struct ceph_mds_request *req = container_of(kref,
532 struct ceph_mds_request,
534 destroy_reply_info(&req->r_reply_info);
536 ceph_msg_put(req->r_request);
538 ceph_msg_put(req->r_reply);
540 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
543 if (req->r_locked_dir)
544 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
545 iput(req->r_target_inode);
548 if (req->r_old_dentry)
549 dput(req->r_old_dentry);
550 if (req->r_old_dentry_dir) {
552 * track (and drop pins for) r_old_dentry_dir
553 * separately, since r_old_dentry's d_parent may have
554 * changed between the dir mutex being dropped and
555 * this request being freed.
557 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
559 iput(req->r_old_dentry_dir);
564 ceph_pagelist_release(req->r_pagelist);
565 put_request_session(req);
566 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
570 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
573 * lookup session, bump ref if found.
575 * called under mdsc->mutex.
577 static struct ceph_mds_request *
578 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
580 struct ceph_mds_request *req;
582 req = lookup_request(&mdsc->request_tree, tid);
584 ceph_mdsc_get_request(req);
590 * Register an in-flight request, and assign a tid. Link to directory
591 * are modifying (if any).
593 * Called under mdsc->mutex.
595 static void __register_request(struct ceph_mds_client *mdsc,
596 struct ceph_mds_request *req,
599 req->r_tid = ++mdsc->last_tid;
601 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
603 dout("__register_request %p tid %lld\n", req, req->r_tid);
604 ceph_mdsc_get_request(req);
605 insert_request(&mdsc->request_tree, req);
607 req->r_uid = current_fsuid();
608 req->r_gid = current_fsgid();
610 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
611 mdsc->oldest_tid = req->r_tid;
615 req->r_unsafe_dir = dir;
619 static void __unregister_request(struct ceph_mds_client *mdsc,
620 struct ceph_mds_request *req)
622 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
624 if (req->r_tid == mdsc->oldest_tid) {
625 struct rb_node *p = rb_next(&req->r_node);
626 mdsc->oldest_tid = 0;
628 struct ceph_mds_request *next_req =
629 rb_entry(p, struct ceph_mds_request, r_node);
630 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
631 mdsc->oldest_tid = next_req->r_tid;
638 erase_request(&mdsc->request_tree, req);
640 if (req->r_unsafe_dir && req->r_got_unsafe) {
641 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
642 spin_lock(&ci->i_unsafe_lock);
643 list_del_init(&req->r_unsafe_dir_item);
644 spin_unlock(&ci->i_unsafe_lock);
646 if (req->r_target_inode && req->r_got_unsafe) {
647 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
648 spin_lock(&ci->i_unsafe_lock);
649 list_del_init(&req->r_unsafe_target_item);
650 spin_unlock(&ci->i_unsafe_lock);
653 if (req->r_unsafe_dir) {
654 iput(req->r_unsafe_dir);
655 req->r_unsafe_dir = NULL;
658 complete_all(&req->r_safe_completion);
660 ceph_mdsc_put_request(req);
664 * Choose mds to send request to next. If there is a hint set in the
665 * request (e.g., due to a prior forward hint from the mds), use that.
666 * Otherwise, consult frag tree and/or caps to identify the
667 * appropriate mds. If all else fails, choose randomly.
669 * Called under mdsc->mutex.
671 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
674 * we don't need to worry about protecting the d_parent access
675 * here because we never renaming inside the snapped namespace
676 * except to resplice to another snapdir, and either the old or new
677 * result is a valid result.
679 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
680 dentry = dentry->d_parent;
684 static int __choose_mds(struct ceph_mds_client *mdsc,
685 struct ceph_mds_request *req)
688 struct ceph_inode_info *ci;
689 struct ceph_cap *cap;
690 int mode = req->r_direct_mode;
692 u32 hash = req->r_direct_hash;
693 bool is_hash = req->r_direct_is_hash;
696 * is there a specific mds we should try? ignore hint if we have
697 * no session and the mds is not up (active or recovering).
699 if (req->r_resend_mds >= 0 &&
700 (__have_session(mdsc, req->r_resend_mds) ||
701 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
702 dout("choose_mds using resend_mds mds%d\n",
704 return req->r_resend_mds;
707 if (mode == USE_RANDOM_MDS)
712 inode = req->r_inode;
713 } else if (req->r_dentry) {
714 /* ignore race with rename; old or new d_parent is okay */
715 struct dentry *parent = req->r_dentry->d_parent;
716 struct inode *dir = d_inode(parent);
718 if (dir->i_sb != mdsc->fsc->sb) {
720 inode = d_inode(req->r_dentry);
721 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
722 /* direct snapped/virtual snapdir requests
723 * based on parent dir inode */
724 struct dentry *dn = get_nonsnap_parent(parent);
726 dout("__choose_mds using nonsnap parent %p\n", inode);
729 inode = d_inode(req->r_dentry);
730 if (!inode || mode == USE_AUTH_MDS) {
733 hash = ceph_dentry_hash(dir, req->r_dentry);
739 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
743 ci = ceph_inode(inode);
745 if (is_hash && S_ISDIR(inode->i_mode)) {
746 struct ceph_inode_frag frag;
749 ceph_choose_frag(ci, hash, &frag, &found);
751 if (mode == USE_ANY_MDS && frag.ndist > 0) {
754 /* choose a random replica */
755 get_random_bytes(&r, 1);
758 dout("choose_mds %p %llx.%llx "
759 "frag %u mds%d (%d/%d)\n",
760 inode, ceph_vinop(inode),
763 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
764 CEPH_MDS_STATE_ACTIVE)
768 /* since this file/dir wasn't known to be
769 * replicated, then we want to look for the
770 * authoritative mds. */
773 /* choose auth mds */
775 dout("choose_mds %p %llx.%llx "
776 "frag %u mds%d (auth)\n",
777 inode, ceph_vinop(inode), frag.frag, mds);
778 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
779 CEPH_MDS_STATE_ACTIVE)
785 spin_lock(&ci->i_ceph_lock);
787 if (mode == USE_AUTH_MDS)
788 cap = ci->i_auth_cap;
789 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
790 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
792 spin_unlock(&ci->i_ceph_lock);
795 mds = cap->session->s_mds;
796 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
797 inode, ceph_vinop(inode), mds,
798 cap == ci->i_auth_cap ? "auth " : "", cap);
799 spin_unlock(&ci->i_ceph_lock);
803 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
804 dout("choose_mds chose random mds%d\n", mds);
812 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
814 struct ceph_msg *msg;
815 struct ceph_mds_session_head *h;
817 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
820 pr_err("create_session_msg ENOMEM creating msg\n");
823 h = msg->front.iov_base;
824 h->op = cpu_to_le32(op);
825 h->seq = cpu_to_le64(seq);
831 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
832 * to include additional client metadata fields.
834 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
836 struct ceph_msg *msg;
837 struct ceph_mds_session_head *h;
839 int metadata_bytes = 0;
840 int metadata_key_count = 0;
841 struct ceph_options *opt = mdsc->fsc->client->options;
844 const char* metadata[][2] = {
845 {"hostname", utsname()->nodename},
846 {"kernel_version", utsname()->release},
847 {"entity_id", opt->name ? opt->name : ""},
851 /* Calculate serialized length of metadata */
852 metadata_bytes = 4; /* map length */
853 for (i = 0; metadata[i][0] != NULL; ++i) {
854 metadata_bytes += 8 + strlen(metadata[i][0]) +
855 strlen(metadata[i][1]);
856 metadata_key_count++;
859 /* Allocate the message */
860 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
863 pr_err("create_session_msg ENOMEM creating msg\n");
866 h = msg->front.iov_base;
867 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
868 h->seq = cpu_to_le64(seq);
871 * Serialize client metadata into waiting buffer space, using
872 * the format that userspace expects for map<string, string>
874 * ClientSession messages with metadata are v2
876 msg->hdr.version = cpu_to_le16(2);
877 msg->hdr.compat_version = cpu_to_le16(1);
879 /* The write pointer, following the session_head structure */
880 p = msg->front.iov_base + sizeof(*h);
882 /* Number of entries in the map */
883 ceph_encode_32(&p, metadata_key_count);
885 /* Two length-prefixed strings for each entry in the map */
886 for (i = 0; metadata[i][0] != NULL; ++i) {
887 size_t const key_len = strlen(metadata[i][0]);
888 size_t const val_len = strlen(metadata[i][1]);
890 ceph_encode_32(&p, key_len);
891 memcpy(p, metadata[i][0], key_len);
893 ceph_encode_32(&p, val_len);
894 memcpy(p, metadata[i][1], val_len);
902 * send session open request.
904 * called under mdsc->mutex
906 static int __open_session(struct ceph_mds_client *mdsc,
907 struct ceph_mds_session *session)
909 struct ceph_msg *msg;
911 int mds = session->s_mds;
913 /* wait for mds to go active? */
914 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
915 dout("open_session to mds%d (%s)\n", mds,
916 ceph_mds_state_name(mstate));
917 session->s_state = CEPH_MDS_SESSION_OPENING;
918 session->s_renew_requested = jiffies;
920 /* send connect message */
921 msg = create_session_open_msg(mdsc, session->s_seq);
924 ceph_con_send(&session->s_con, msg);
929 * open sessions for any export targets for the given mds
931 * called under mdsc->mutex
933 static struct ceph_mds_session *
934 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
936 struct ceph_mds_session *session;
938 session = __ceph_lookup_mds_session(mdsc, target);
940 session = register_session(mdsc, target);
944 if (session->s_state == CEPH_MDS_SESSION_NEW ||
945 session->s_state == CEPH_MDS_SESSION_CLOSING)
946 __open_session(mdsc, session);
951 struct ceph_mds_session *
952 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
954 struct ceph_mds_session *session;
956 dout("open_export_target_session to mds%d\n", target);
958 mutex_lock(&mdsc->mutex);
959 session = __open_export_target_session(mdsc, target);
960 mutex_unlock(&mdsc->mutex);
965 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
966 struct ceph_mds_session *session)
968 struct ceph_mds_info *mi;
969 struct ceph_mds_session *ts;
970 int i, mds = session->s_mds;
972 if (mds >= mdsc->mdsmap->m_max_mds)
975 mi = &mdsc->mdsmap->m_info[mds];
976 dout("open_export_target_sessions for mds%d (%d targets)\n",
977 session->s_mds, mi->num_export_targets);
979 for (i = 0; i < mi->num_export_targets; i++) {
980 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
982 ceph_put_mds_session(ts);
986 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
987 struct ceph_mds_session *session)
989 mutex_lock(&mdsc->mutex);
990 __open_export_target_sessions(mdsc, session);
991 mutex_unlock(&mdsc->mutex);
998 /* caller holds s_cap_lock, we drop it */
999 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1000 struct ceph_mds_session *session)
1001 __releases(session->s_cap_lock)
1003 LIST_HEAD(tmp_list);
1004 list_splice_init(&session->s_cap_releases, &tmp_list);
1005 session->s_num_cap_releases = 0;
1006 spin_unlock(&session->s_cap_lock);
1008 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1009 while (!list_empty(&tmp_list)) {
1010 struct ceph_cap *cap;
1011 /* zero out the in-progress message */
1012 cap = list_first_entry(&tmp_list,
1013 struct ceph_cap, session_caps);
1014 list_del(&cap->session_caps);
1015 ceph_put_cap(mdsc, cap);
1019 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1020 struct ceph_mds_session *session)
1022 struct ceph_mds_request *req;
1025 dout("cleanup_session_requests mds%d\n", session->s_mds);
1026 mutex_lock(&mdsc->mutex);
1027 while (!list_empty(&session->s_unsafe)) {
1028 req = list_first_entry(&session->s_unsafe,
1029 struct ceph_mds_request, r_unsafe_item);
1030 list_del_init(&req->r_unsafe_item);
1031 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1033 __unregister_request(mdsc, req);
1035 /* zero r_attempts, so kick_requests() will re-send requests */
1036 p = rb_first(&mdsc->request_tree);
1038 req = rb_entry(p, struct ceph_mds_request, r_node);
1040 if (req->r_session &&
1041 req->r_session->s_mds == session->s_mds)
1042 req->r_attempts = 0;
1044 mutex_unlock(&mdsc->mutex);
1048 * Helper to safely iterate over all caps associated with a session, with
1049 * special care taken to handle a racing __ceph_remove_cap().
1051 * Caller must hold session s_mutex.
1053 static int iterate_session_caps(struct ceph_mds_session *session,
1054 int (*cb)(struct inode *, struct ceph_cap *,
1057 struct list_head *p;
1058 struct ceph_cap *cap;
1059 struct inode *inode, *last_inode = NULL;
1060 struct ceph_cap *old_cap = NULL;
1063 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1064 spin_lock(&session->s_cap_lock);
1065 p = session->s_caps.next;
1066 while (p != &session->s_caps) {
1067 cap = list_entry(p, struct ceph_cap, session_caps);
1068 inode = igrab(&cap->ci->vfs_inode);
1073 session->s_cap_iterator = cap;
1074 spin_unlock(&session->s_cap_lock);
1081 ceph_put_cap(session->s_mdsc, old_cap);
1085 ret = cb(inode, cap, arg);
1088 spin_lock(&session->s_cap_lock);
1090 if (cap->ci == NULL) {
1091 dout("iterate_session_caps finishing cap %p removal\n",
1093 BUG_ON(cap->session != session);
1094 cap->session = NULL;
1095 list_del_init(&cap->session_caps);
1096 session->s_nr_caps--;
1097 if (cap->queue_release) {
1098 list_add_tail(&cap->session_caps,
1099 &session->s_cap_releases);
1100 session->s_num_cap_releases++;
1102 old_cap = cap; /* put_cap it w/o locks held */
1110 session->s_cap_iterator = NULL;
1111 spin_unlock(&session->s_cap_lock);
1115 ceph_put_cap(session->s_mdsc, old_cap);
1120 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1123 struct ceph_inode_info *ci = ceph_inode(inode);
1124 LIST_HEAD(to_remove);
1127 dout("removing cap %p, ci is %p, inode is %p\n",
1128 cap, ci, &ci->vfs_inode);
1129 spin_lock(&ci->i_ceph_lock);
1130 __ceph_remove_cap(cap, false);
1131 if (!ci->i_auth_cap) {
1132 struct ceph_cap_flush *cf;
1133 struct ceph_mds_client *mdsc =
1134 ceph_sb_to_client(inode->i_sb)->mdsc;
1137 struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1140 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1141 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1142 list_add(&cf->list, &to_remove);
1145 spin_lock(&mdsc->cap_dirty_lock);
1147 list_for_each_entry(cf, &to_remove, list)
1148 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1150 if (!list_empty(&ci->i_dirty_item)) {
1151 pr_warn_ratelimited(
1152 " dropping dirty %s state for %p %lld\n",
1153 ceph_cap_string(ci->i_dirty_caps),
1154 inode, ceph_ino(inode));
1155 ci->i_dirty_caps = 0;
1156 list_del_init(&ci->i_dirty_item);
1159 if (!list_empty(&ci->i_flushing_item)) {
1160 pr_warn_ratelimited(
1161 " dropping dirty+flushing %s state for %p %lld\n",
1162 ceph_cap_string(ci->i_flushing_caps),
1163 inode, ceph_ino(inode));
1164 ci->i_flushing_caps = 0;
1165 list_del_init(&ci->i_flushing_item);
1166 mdsc->num_cap_flushing--;
1169 spin_unlock(&mdsc->cap_dirty_lock);
1171 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1172 list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1173 ci->i_prealloc_cap_flush = NULL;
1176 spin_unlock(&ci->i_ceph_lock);
1177 while (!list_empty(&to_remove)) {
1178 struct ceph_cap_flush *cf;
1179 cf = list_first_entry(&to_remove,
1180 struct ceph_cap_flush, list);
1181 list_del(&cf->list);
1182 ceph_free_cap_flush(cf);
1190 * caller must hold session s_mutex
1192 static void remove_session_caps(struct ceph_mds_session *session)
1194 dout("remove_session_caps on %p\n", session);
1195 iterate_session_caps(session, remove_session_caps_cb, NULL);
1197 spin_lock(&session->s_cap_lock);
1198 if (session->s_nr_caps > 0) {
1199 struct super_block *sb = session->s_mdsc->fsc->sb;
1200 struct inode *inode;
1201 struct ceph_cap *cap, *prev = NULL;
1202 struct ceph_vino vino;
1204 * iterate_session_caps() skips inodes that are being
1205 * deleted, we need to wait until deletions are complete.
1206 * __wait_on_freeing_inode() is designed for the job,
1207 * but it is not exported, so use lookup inode function
1210 while (!list_empty(&session->s_caps)) {
1211 cap = list_entry(session->s_caps.next,
1212 struct ceph_cap, session_caps);
1216 vino = cap->ci->i_vino;
1217 spin_unlock(&session->s_cap_lock);
1219 inode = ceph_find_inode(sb, vino);
1222 spin_lock(&session->s_cap_lock);
1226 // drop cap expires and unlock s_cap_lock
1227 cleanup_cap_releases(session->s_mdsc, session);
1229 BUG_ON(session->s_nr_caps > 0);
1230 BUG_ON(!list_empty(&session->s_cap_flushing));
1234 * wake up any threads waiting on this session's caps. if the cap is
1235 * old (didn't get renewed on the client reconnect), remove it now.
1237 * caller must hold s_mutex.
1239 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1242 struct ceph_inode_info *ci = ceph_inode(inode);
1244 wake_up_all(&ci->i_cap_wq);
1246 spin_lock(&ci->i_ceph_lock);
1247 ci->i_wanted_max_size = 0;
1248 ci->i_requested_max_size = 0;
1249 spin_unlock(&ci->i_ceph_lock);
1254 static void wake_up_session_caps(struct ceph_mds_session *session,
1257 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1258 iterate_session_caps(session, wake_up_session_cb,
1259 (void *)(unsigned long)reconnect);
1263 * Send periodic message to MDS renewing all currently held caps. The
1264 * ack will reset the expiration for all caps from this session.
1266 * caller holds s_mutex
1268 static int send_renew_caps(struct ceph_mds_client *mdsc,
1269 struct ceph_mds_session *session)
1271 struct ceph_msg *msg;
1274 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1275 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1276 pr_info("mds%d caps stale\n", session->s_mds);
1277 session->s_renew_requested = jiffies;
1279 /* do not try to renew caps until a recovering mds has reconnected
1280 * with its clients. */
1281 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1282 if (state < CEPH_MDS_STATE_RECONNECT) {
1283 dout("send_renew_caps ignoring mds%d (%s)\n",
1284 session->s_mds, ceph_mds_state_name(state));
1288 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1289 ceph_mds_state_name(state));
1290 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1291 ++session->s_renew_seq);
1294 ceph_con_send(&session->s_con, msg);
1298 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1299 struct ceph_mds_session *session, u64 seq)
1301 struct ceph_msg *msg;
1303 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1304 session->s_mds, ceph_session_state_name(session->s_state), seq);
1305 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1308 ceph_con_send(&session->s_con, msg);
1314 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1316 * Called under session->s_mutex
1318 static void renewed_caps(struct ceph_mds_client *mdsc,
1319 struct ceph_mds_session *session, int is_renew)
1324 spin_lock(&session->s_cap_lock);
1325 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1327 session->s_cap_ttl = session->s_renew_requested +
1328 mdsc->mdsmap->m_session_timeout*HZ;
1331 if (time_before(jiffies, session->s_cap_ttl)) {
1332 pr_info("mds%d caps renewed\n", session->s_mds);
1335 pr_info("mds%d caps still stale\n", session->s_mds);
1338 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1339 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1340 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1341 spin_unlock(&session->s_cap_lock);
1344 wake_up_session_caps(session, 0);
1348 * send a session close request
1350 static int request_close_session(struct ceph_mds_client *mdsc,
1351 struct ceph_mds_session *session)
1353 struct ceph_msg *msg;
1355 dout("request_close_session mds%d state %s seq %lld\n",
1356 session->s_mds, ceph_session_state_name(session->s_state),
1358 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1361 ceph_con_send(&session->s_con, msg);
1366 * Called with s_mutex held.
1368 static int __close_session(struct ceph_mds_client *mdsc,
1369 struct ceph_mds_session *session)
1371 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1373 session->s_state = CEPH_MDS_SESSION_CLOSING;
1374 return request_close_session(mdsc, session);
1378 * Trim old(er) caps.
1380 * Because we can't cache an inode without one or more caps, we do
1381 * this indirectly: if a cap is unused, we prune its aliases, at which
1382 * point the inode will hopefully get dropped to.
1384 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1385 * memory pressure from the MDS, though, so it needn't be perfect.
1387 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1389 struct ceph_mds_session *session = arg;
1390 struct ceph_inode_info *ci = ceph_inode(inode);
1391 int used, wanted, oissued, mine;
1393 if (session->s_trim_caps <= 0)
1396 spin_lock(&ci->i_ceph_lock);
1397 mine = cap->issued | cap->implemented;
1398 used = __ceph_caps_used(ci);
1399 wanted = __ceph_caps_file_wanted(ci);
1400 oissued = __ceph_caps_issued_other(ci, cap);
1402 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1403 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1404 ceph_cap_string(used), ceph_cap_string(wanted));
1405 if (cap == ci->i_auth_cap) {
1406 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1407 !list_empty(&ci->i_cap_snaps))
1409 if ((used | wanted) & CEPH_CAP_ANY_WR)
1412 /* The inode has cached pages, but it's no longer used.
1413 * we can safely drop it */
1414 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1415 !(oissued & CEPH_CAP_FILE_CACHE)) {
1419 if ((used | wanted) & ~oissued & mine)
1420 goto out; /* we need these caps */
1422 session->s_trim_caps--;
1424 /* we aren't the only cap.. just remove us */
1425 __ceph_remove_cap(cap, true);
1427 /* try dropping referring dentries */
1428 spin_unlock(&ci->i_ceph_lock);
1429 d_prune_aliases(inode);
1430 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1431 inode, cap, atomic_read(&inode->i_count));
1436 spin_unlock(&ci->i_ceph_lock);
1441 * Trim session cap count down to some max number.
1443 static int trim_caps(struct ceph_mds_client *mdsc,
1444 struct ceph_mds_session *session,
1447 int trim_caps = session->s_nr_caps - max_caps;
1449 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1450 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1451 if (trim_caps > 0) {
1452 session->s_trim_caps = trim_caps;
1453 iterate_session_caps(session, trim_caps_cb, session);
1454 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1455 session->s_mds, session->s_nr_caps, max_caps,
1456 trim_caps - session->s_trim_caps);
1457 session->s_trim_caps = 0;
1460 ceph_send_cap_releases(mdsc, session);
1464 static int check_capsnap_flush(struct ceph_inode_info *ci,
1468 spin_lock(&ci->i_ceph_lock);
1469 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1470 struct ceph_cap_snap *capsnap =
1471 list_first_entry(&ci->i_cap_snaps,
1472 struct ceph_cap_snap, ci_item);
1473 ret = capsnap->follows >= want_snap_seq;
1475 spin_unlock(&ci->i_ceph_lock);
1479 static int check_caps_flush(struct ceph_mds_client *mdsc,
1483 struct ceph_cap_flush *cf;
1486 spin_lock(&mdsc->cap_dirty_lock);
1487 n = rb_first(&mdsc->cap_flush_tree);
1488 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1489 if (cf && cf->tid <= want_flush_tid) {
1490 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1491 cf->tid, want_flush_tid);
1494 spin_unlock(&mdsc->cap_dirty_lock);
1499 * flush all dirty inode data to disk.
1501 * returns true if we've flushed through want_flush_tid
1503 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1504 u64 want_flush_tid, u64 want_snap_seq)
1508 dout("check_caps_flush want %llu snap want %llu\n",
1509 want_flush_tid, want_snap_seq);
1510 mutex_lock(&mdsc->mutex);
1511 for (mds = 0; mds < mdsc->max_sessions; ) {
1512 struct ceph_mds_session *session = mdsc->sessions[mds];
1513 struct inode *inode = NULL;
1519 get_session(session);
1520 mutex_unlock(&mdsc->mutex);
1522 mutex_lock(&session->s_mutex);
1523 if (!list_empty(&session->s_cap_snaps_flushing)) {
1524 struct ceph_cap_snap *capsnap =
1525 list_first_entry(&session->s_cap_snaps_flushing,
1526 struct ceph_cap_snap,
1528 struct ceph_inode_info *ci = capsnap->ci;
1529 if (!check_capsnap_flush(ci, want_snap_seq)) {
1530 dout("check_cap_flush still flushing snap %p "
1531 "follows %lld <= %lld to mds%d\n",
1532 &ci->vfs_inode, capsnap->follows,
1533 want_snap_seq, mds);
1534 inode = igrab(&ci->vfs_inode);
1537 mutex_unlock(&session->s_mutex);
1538 ceph_put_mds_session(session);
1541 wait_event(mdsc->cap_flushing_wq,
1542 check_capsnap_flush(ceph_inode(inode),
1549 mutex_lock(&mdsc->mutex);
1551 mutex_unlock(&mdsc->mutex);
1553 wait_event(mdsc->cap_flushing_wq,
1554 check_caps_flush(mdsc, want_flush_tid));
1556 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1560 * called under s_mutex
1562 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1563 struct ceph_mds_session *session)
1565 struct ceph_msg *msg = NULL;
1566 struct ceph_mds_cap_release *head;
1567 struct ceph_mds_cap_item *item;
1568 struct ceph_cap *cap;
1569 LIST_HEAD(tmp_list);
1570 int num_cap_releases;
1572 spin_lock(&session->s_cap_lock);
1574 list_splice_init(&session->s_cap_releases, &tmp_list);
1575 num_cap_releases = session->s_num_cap_releases;
1576 session->s_num_cap_releases = 0;
1577 spin_unlock(&session->s_cap_lock);
1579 while (!list_empty(&tmp_list)) {
1581 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1582 PAGE_SIZE, GFP_NOFS, false);
1585 head = msg->front.iov_base;
1586 head->num = cpu_to_le32(0);
1587 msg->front.iov_len = sizeof(*head);
1589 cap = list_first_entry(&tmp_list, struct ceph_cap,
1591 list_del(&cap->session_caps);
1594 head = msg->front.iov_base;
1595 le32_add_cpu(&head->num, 1);
1596 item = msg->front.iov_base + msg->front.iov_len;
1597 item->ino = cpu_to_le64(cap->cap_ino);
1598 item->cap_id = cpu_to_le64(cap->cap_id);
1599 item->migrate_seq = cpu_to_le32(cap->mseq);
1600 item->seq = cpu_to_le32(cap->issue_seq);
1601 msg->front.iov_len += sizeof(*item);
1603 ceph_put_cap(mdsc, cap);
1605 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1606 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1607 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1608 ceph_con_send(&session->s_con, msg);
1613 BUG_ON(num_cap_releases != 0);
1615 spin_lock(&session->s_cap_lock);
1616 if (!list_empty(&session->s_cap_releases))
1618 spin_unlock(&session->s_cap_lock);
1621 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1622 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1623 ceph_con_send(&session->s_con, msg);
1627 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1629 spin_lock(&session->s_cap_lock);
1630 list_splice(&tmp_list, &session->s_cap_releases);
1631 session->s_num_cap_releases += num_cap_releases;
1632 spin_unlock(&session->s_cap_lock);
1639 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1642 struct ceph_inode_info *ci = ceph_inode(dir);
1643 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1644 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1645 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1646 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1647 int order, num_entries;
1649 spin_lock(&ci->i_ceph_lock);
1650 num_entries = ci->i_files + ci->i_subdirs;
1651 spin_unlock(&ci->i_ceph_lock);
1652 num_entries = max(num_entries, 1);
1653 num_entries = min(num_entries, opt->max_readdir);
1655 order = get_order(size * num_entries);
1656 while (order >= 0) {
1657 rinfo->dir_in = (void*)__get_free_pages(GFP_KERNEL |
1667 num_entries = (PAGE_SIZE << order) / size;
1668 num_entries = min(num_entries, opt->max_readdir);
1670 rinfo->dir_buf_size = PAGE_SIZE << order;
1671 req->r_num_caps = num_entries + 1;
1672 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1673 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1678 * Create an mds request.
1680 struct ceph_mds_request *
1681 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1683 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1686 return ERR_PTR(-ENOMEM);
1688 mutex_init(&req->r_fill_mutex);
1690 req->r_started = jiffies;
1691 req->r_resend_mds = -1;
1692 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1693 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1695 kref_init(&req->r_kref);
1696 RB_CLEAR_NODE(&req->r_node);
1697 INIT_LIST_HEAD(&req->r_wait);
1698 init_completion(&req->r_completion);
1699 init_completion(&req->r_safe_completion);
1700 INIT_LIST_HEAD(&req->r_unsafe_item);
1702 req->r_stamp = current_fs_time(mdsc->fsc->sb);
1705 req->r_direct_mode = mode;
1710 * return oldest (lowest) request, tid in request tree, 0 if none.
1712 * called under mdsc->mutex.
1714 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1716 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1718 return rb_entry(rb_first(&mdsc->request_tree),
1719 struct ceph_mds_request, r_node);
1722 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1724 return mdsc->oldest_tid;
1728 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1729 * on build_path_from_dentry in fs/cifs/dir.c.
1731 * If @stop_on_nosnap, generate path relative to the first non-snapped
1734 * Encode hidden .snap dirs as a double /, i.e.
1735 * foo/.snap/bar -> foo//bar
1737 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1740 struct dentry *temp;
1746 return ERR_PTR(-EINVAL);
1750 seq = read_seqbegin(&rename_lock);
1752 for (temp = dentry; !IS_ROOT(temp);) {
1753 struct inode *inode = d_inode(temp);
1754 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1755 len++; /* slash only */
1756 else if (stop_on_nosnap && inode &&
1757 ceph_snap(inode) == CEPH_NOSNAP)
1760 len += 1 + temp->d_name.len;
1761 temp = temp->d_parent;
1765 len--; /* no leading '/' */
1767 path = kmalloc(len+1, GFP_NOFS);
1769 return ERR_PTR(-ENOMEM);
1771 path[pos] = 0; /* trailing null */
1773 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1774 struct inode *inode;
1776 spin_lock(&temp->d_lock);
1777 inode = d_inode(temp);
1778 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1779 dout("build_path path+%d: %p SNAPDIR\n",
1781 } else if (stop_on_nosnap && inode &&
1782 ceph_snap(inode) == CEPH_NOSNAP) {
1783 spin_unlock(&temp->d_lock);
1786 pos -= temp->d_name.len;
1788 spin_unlock(&temp->d_lock);
1791 strncpy(path + pos, temp->d_name.name,
1794 spin_unlock(&temp->d_lock);
1797 temp = temp->d_parent;
1800 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1801 pr_err("build_path did not end path lookup where "
1802 "expected, namelen is %d, pos is %d\n", len, pos);
1803 /* presumably this is only possible if racing with a
1804 rename of one of the parent directories (we can not
1805 lock the dentries above us to prevent this, but
1806 retrying should be harmless) */
1811 *base = ceph_ino(d_inode(temp));
1813 dout("build_path on %p %d built %llx '%.*s'\n",
1814 dentry, d_count(dentry), *base, len, path);
1818 static int build_dentry_path(struct dentry *dentry,
1819 const char **ppath, int *ppathlen, u64 *pino,
1824 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1825 *pino = ceph_ino(d_inode(dentry->d_parent));
1826 *ppath = dentry->d_name.name;
1827 *ppathlen = dentry->d_name.len;
1830 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1832 return PTR_ERR(path);
1838 static int build_inode_path(struct inode *inode,
1839 const char **ppath, int *ppathlen, u64 *pino,
1842 struct dentry *dentry;
1845 if (ceph_snap(inode) == CEPH_NOSNAP) {
1846 *pino = ceph_ino(inode);
1850 dentry = d_find_alias(inode);
1851 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1854 return PTR_ERR(path);
1861 * request arguments may be specified via an inode *, a dentry *, or
1862 * an explicit ino+path.
1864 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1865 const char *rpath, u64 rino,
1866 const char **ppath, int *pathlen,
1867 u64 *ino, int *freepath)
1872 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1873 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1875 } else if (rdentry) {
1876 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1877 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1879 } else if (rpath || rino) {
1882 *pathlen = rpath ? strlen(rpath) : 0;
1883 dout(" path %.*s\n", *pathlen, rpath);
1890 * called under mdsc->mutex
1892 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1893 struct ceph_mds_request *req,
1894 int mds, bool drop_cap_releases)
1896 struct ceph_msg *msg;
1897 struct ceph_mds_request_head *head;
1898 const char *path1 = NULL;
1899 const char *path2 = NULL;
1900 u64 ino1 = 0, ino2 = 0;
1901 int pathlen1 = 0, pathlen2 = 0;
1902 int freepath1 = 0, freepath2 = 0;
1908 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1909 req->r_path1, req->r_ino1.ino,
1910 &path1, &pathlen1, &ino1, &freepath1);
1916 ret = set_request_path_attr(NULL, req->r_old_dentry,
1917 req->r_path2, req->r_ino2.ino,
1918 &path2, &pathlen2, &ino2, &freepath2);
1924 len = sizeof(*head) +
1925 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1926 sizeof(struct ceph_timespec);
1928 /* calculate (max) length for cap releases */
1929 len += sizeof(struct ceph_mds_request_release) *
1930 (!!req->r_inode_drop + !!req->r_dentry_drop +
1931 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1932 if (req->r_dentry_drop)
1933 len += req->r_dentry->d_name.len;
1934 if (req->r_old_dentry_drop)
1935 len += req->r_old_dentry->d_name.len;
1937 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1939 msg = ERR_PTR(-ENOMEM);
1943 msg->hdr.version = cpu_to_le16(2);
1944 msg->hdr.tid = cpu_to_le64(req->r_tid);
1946 head = msg->front.iov_base;
1947 p = msg->front.iov_base + sizeof(*head);
1948 end = msg->front.iov_base + msg->front.iov_len;
1950 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1951 head->op = cpu_to_le32(req->r_op);
1952 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1953 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1954 head->args = req->r_args;
1956 ceph_encode_filepath(&p, end, ino1, path1);
1957 ceph_encode_filepath(&p, end, ino2, path2);
1959 /* make note of release offset, in case we need to replay */
1960 req->r_request_release_offset = p - msg->front.iov_base;
1964 if (req->r_inode_drop)
1965 releases += ceph_encode_inode_release(&p,
1966 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1967 mds, req->r_inode_drop, req->r_inode_unless, 0);
1968 if (req->r_dentry_drop)
1969 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1970 mds, req->r_dentry_drop, req->r_dentry_unless);
1971 if (req->r_old_dentry_drop)
1972 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1973 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1974 if (req->r_old_inode_drop)
1975 releases += ceph_encode_inode_release(&p,
1976 d_inode(req->r_old_dentry),
1977 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1979 if (drop_cap_releases) {
1981 p = msg->front.iov_base + req->r_request_release_offset;
1984 head->num_releases = cpu_to_le16(releases);
1988 struct ceph_timespec ts;
1989 ceph_encode_timespec(&ts, &req->r_stamp);
1990 ceph_encode_copy(&p, &ts, sizeof(ts));
1994 msg->front.iov_len = p - msg->front.iov_base;
1995 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1997 if (req->r_pagelist) {
1998 struct ceph_pagelist *pagelist = req->r_pagelist;
1999 atomic_inc(&pagelist->refcnt);
2000 ceph_msg_data_add_pagelist(msg, pagelist);
2001 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2003 msg->hdr.data_len = 0;
2006 msg->hdr.data_off = cpu_to_le16(0);
2010 kfree((char *)path2);
2013 kfree((char *)path1);
2019 * called under mdsc->mutex if error, under no mutex if
2022 static void complete_request(struct ceph_mds_client *mdsc,
2023 struct ceph_mds_request *req)
2025 if (req->r_callback)
2026 req->r_callback(mdsc, req);
2028 complete_all(&req->r_completion);
2032 * called under mdsc->mutex
2034 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2035 struct ceph_mds_request *req,
2036 int mds, bool drop_cap_releases)
2038 struct ceph_mds_request_head *rhead;
2039 struct ceph_msg *msg;
2044 struct ceph_cap *cap =
2045 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2048 req->r_sent_on_mseq = cap->mseq;
2050 req->r_sent_on_mseq = -1;
2052 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2053 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2055 if (req->r_got_unsafe) {
2058 * Replay. Do not regenerate message (and rebuild
2059 * paths, etc.); just use the original message.
2060 * Rebuilding paths will break for renames because
2061 * d_move mangles the src name.
2063 msg = req->r_request;
2064 rhead = msg->front.iov_base;
2066 flags = le32_to_cpu(rhead->flags);
2067 flags |= CEPH_MDS_FLAG_REPLAY;
2068 rhead->flags = cpu_to_le32(flags);
2070 if (req->r_target_inode)
2071 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2073 rhead->num_retry = req->r_attempts - 1;
2075 /* remove cap/dentry releases from message */
2076 rhead->num_releases = 0;
2079 p = msg->front.iov_base + req->r_request_release_offset;
2081 struct ceph_timespec ts;
2082 ceph_encode_timespec(&ts, &req->r_stamp);
2083 ceph_encode_copy(&p, &ts, sizeof(ts));
2086 msg->front.iov_len = p - msg->front.iov_base;
2087 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2091 if (req->r_request) {
2092 ceph_msg_put(req->r_request);
2093 req->r_request = NULL;
2095 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2097 req->r_err = PTR_ERR(msg);
2098 return PTR_ERR(msg);
2100 req->r_request = msg;
2102 rhead = msg->front.iov_base;
2103 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2104 if (req->r_got_unsafe)
2105 flags |= CEPH_MDS_FLAG_REPLAY;
2106 if (req->r_locked_dir)
2107 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2108 rhead->flags = cpu_to_le32(flags);
2109 rhead->num_fwd = req->r_num_fwd;
2110 rhead->num_retry = req->r_attempts - 1;
2113 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2118 * send request, or put it on the appropriate wait list.
2120 static int __do_request(struct ceph_mds_client *mdsc,
2121 struct ceph_mds_request *req)
2123 struct ceph_mds_session *session = NULL;
2127 if (req->r_err || req->r_got_result) {
2129 __unregister_request(mdsc, req);
2133 if (req->r_timeout &&
2134 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2135 dout("do_request timed out\n");
2139 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2140 dout("do_request forced umount\n");
2145 put_request_session(req);
2147 mds = __choose_mds(mdsc, req);
2149 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2150 dout("do_request no mds or not active, waiting for map\n");
2151 list_add(&req->r_wait, &mdsc->waiting_for_map);
2155 /* get, open session */
2156 session = __ceph_lookup_mds_session(mdsc, mds);
2158 session = register_session(mdsc, mds);
2159 if (IS_ERR(session)) {
2160 err = PTR_ERR(session);
2164 req->r_session = get_session(session);
2166 dout("do_request mds%d session %p state %s\n", mds, session,
2167 ceph_session_state_name(session->s_state));
2168 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2169 session->s_state != CEPH_MDS_SESSION_HUNG) {
2170 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2171 session->s_state == CEPH_MDS_SESSION_CLOSING)
2172 __open_session(mdsc, session);
2173 list_add(&req->r_wait, &session->s_waiting);
2178 req->r_resend_mds = -1; /* forget any previous mds hint */
2180 if (req->r_request_started == 0) /* note request start time */
2181 req->r_request_started = jiffies;
2183 err = __prepare_send_request(mdsc, req, mds, false);
2185 ceph_msg_get(req->r_request);
2186 ceph_con_send(&session->s_con, req->r_request);
2190 ceph_put_mds_session(session);
2193 dout("__do_request early error %d\n", err);
2195 complete_request(mdsc, req);
2196 __unregister_request(mdsc, req);
2203 * called under mdsc->mutex
2205 static void __wake_requests(struct ceph_mds_client *mdsc,
2206 struct list_head *head)
2208 struct ceph_mds_request *req;
2209 LIST_HEAD(tmp_list);
2211 list_splice_init(head, &tmp_list);
2213 while (!list_empty(&tmp_list)) {
2214 req = list_entry(tmp_list.next,
2215 struct ceph_mds_request, r_wait);
2216 list_del_init(&req->r_wait);
2217 dout(" wake request %p tid %llu\n", req, req->r_tid);
2218 __do_request(mdsc, req);
2223 * Wake up threads with requests pending for @mds, so that they can
2224 * resubmit their requests to a possibly different mds.
2226 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2228 struct ceph_mds_request *req;
2229 struct rb_node *p = rb_first(&mdsc->request_tree);
2231 dout("kick_requests mds%d\n", mds);
2233 req = rb_entry(p, struct ceph_mds_request, r_node);
2235 if (req->r_got_unsafe)
2237 if (req->r_attempts > 0)
2238 continue; /* only new requests */
2239 if (req->r_session &&
2240 req->r_session->s_mds == mds) {
2241 dout(" kicking tid %llu\n", req->r_tid);
2242 list_del_init(&req->r_wait);
2243 __do_request(mdsc, req);
2248 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2249 struct ceph_mds_request *req)
2251 dout("submit_request on %p\n", req);
2252 mutex_lock(&mdsc->mutex);
2253 __register_request(mdsc, req, NULL);
2254 __do_request(mdsc, req);
2255 mutex_unlock(&mdsc->mutex);
2259 * Synchrously perform an mds request. Take care of all of the
2260 * session setup, forwarding, retry details.
2262 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2264 struct ceph_mds_request *req)
2268 dout("do_request on %p\n", req);
2270 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2272 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2273 if (req->r_locked_dir)
2274 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2275 if (req->r_old_dentry_dir)
2276 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2279 /* deny access to directories with pool_ns layouts */
2280 if (req->r_inode && S_ISDIR(req->r_inode->i_mode) &&
2281 ceph_inode(req->r_inode)->i_pool_ns_len)
2283 if (req->r_locked_dir &&
2284 ceph_inode(req->r_locked_dir)->i_pool_ns_len)
2288 mutex_lock(&mdsc->mutex);
2289 __register_request(mdsc, req, dir);
2290 __do_request(mdsc, req);
2298 mutex_unlock(&mdsc->mutex);
2299 dout("do_request waiting\n");
2300 if (!req->r_timeout && req->r_wait_for_completion) {
2301 err = req->r_wait_for_completion(mdsc, req);
2303 long timeleft = wait_for_completion_killable_timeout(
2305 ceph_timeout_jiffies(req->r_timeout));
2309 err = -EIO; /* timed out */
2311 err = timeleft; /* killed */
2313 dout("do_request waited, got %d\n", err);
2314 mutex_lock(&mdsc->mutex);
2316 /* only abort if we didn't race with a real reply */
2317 if (req->r_got_result) {
2318 err = le32_to_cpu(req->r_reply_info.head->result);
2319 } else if (err < 0) {
2320 dout("aborted request %lld with %d\n", req->r_tid, err);
2323 * ensure we aren't running concurrently with
2324 * ceph_fill_trace or ceph_readdir_prepopulate, which
2325 * rely on locks (dir mutex) held by our caller.
2327 mutex_lock(&req->r_fill_mutex);
2329 req->r_aborted = true;
2330 mutex_unlock(&req->r_fill_mutex);
2332 if (req->r_locked_dir &&
2333 (req->r_op & CEPH_MDS_OP_WRITE))
2334 ceph_invalidate_dir_request(req);
2340 mutex_unlock(&mdsc->mutex);
2341 dout("do_request %p done, result %d\n", req, err);
2346 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2347 * namespace request.
2349 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2351 struct inode *inode = req->r_locked_dir;
2353 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2355 ceph_dir_clear_complete(inode);
2357 ceph_invalidate_dentry_lease(req->r_dentry);
2358 if (req->r_old_dentry)
2359 ceph_invalidate_dentry_lease(req->r_old_dentry);
2365 * We take the session mutex and parse and process the reply immediately.
2366 * This preserves the logical ordering of replies, capabilities, etc., sent
2367 * by the MDS as they are applied to our local cache.
2369 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2371 struct ceph_mds_client *mdsc = session->s_mdsc;
2372 struct ceph_mds_request *req;
2373 struct ceph_mds_reply_head *head = msg->front.iov_base;
2374 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2375 struct ceph_snap_realm *realm;
2378 int mds = session->s_mds;
2380 if (msg->front.iov_len < sizeof(*head)) {
2381 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2386 /* get request, session */
2387 tid = le64_to_cpu(msg->hdr.tid);
2388 mutex_lock(&mdsc->mutex);
2389 req = lookup_get_request(mdsc, tid);
2391 dout("handle_reply on unknown tid %llu\n", tid);
2392 mutex_unlock(&mdsc->mutex);
2395 dout("handle_reply %p\n", req);
2397 /* correct session? */
2398 if (req->r_session != session) {
2399 pr_err("mdsc_handle_reply got %llu on session mds%d"
2400 " not mds%d\n", tid, session->s_mds,
2401 req->r_session ? req->r_session->s_mds : -1);
2402 mutex_unlock(&mdsc->mutex);
2407 if ((req->r_got_unsafe && !head->safe) ||
2408 (req->r_got_safe && head->safe)) {
2409 pr_warn("got a dup %s reply on %llu from mds%d\n",
2410 head->safe ? "safe" : "unsafe", tid, mds);
2411 mutex_unlock(&mdsc->mutex);
2414 if (req->r_got_safe) {
2415 pr_warn("got unsafe after safe on %llu from mds%d\n",
2417 mutex_unlock(&mdsc->mutex);
2421 result = le32_to_cpu(head->result);
2425 * if we're not talking to the authority, send to them
2426 * if the authority has changed while we weren't looking,
2427 * send to new authority
2428 * Otherwise we just have to return an ESTALE
2430 if (result == -ESTALE) {
2431 dout("got ESTALE on request %llu", req->r_tid);
2432 req->r_resend_mds = -1;
2433 if (req->r_direct_mode != USE_AUTH_MDS) {
2434 dout("not using auth, setting for that now");
2435 req->r_direct_mode = USE_AUTH_MDS;
2436 __do_request(mdsc, req);
2437 mutex_unlock(&mdsc->mutex);
2440 int mds = __choose_mds(mdsc, req);
2441 if (mds >= 0 && mds != req->r_session->s_mds) {
2442 dout("but auth changed, so resending");
2443 __do_request(mdsc, req);
2444 mutex_unlock(&mdsc->mutex);
2448 dout("have to return ESTALE on request %llu", req->r_tid);
2453 req->r_got_safe = true;
2454 __unregister_request(mdsc, req);
2456 if (req->r_got_unsafe) {
2458 * We already handled the unsafe response, now do the
2459 * cleanup. No need to examine the response; the MDS
2460 * doesn't include any result info in the safe
2461 * response. And even if it did, there is nothing
2462 * useful we could do with a revised return value.
2464 dout("got safe reply %llu, mds%d\n", tid, mds);
2465 list_del_init(&req->r_unsafe_item);
2467 /* last unsafe request during umount? */
2468 if (mdsc->stopping && !__get_oldest_req(mdsc))
2469 complete_all(&mdsc->safe_umount_waiters);
2470 mutex_unlock(&mdsc->mutex);
2474 req->r_got_unsafe = true;
2475 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2476 if (req->r_unsafe_dir) {
2477 struct ceph_inode_info *ci =
2478 ceph_inode(req->r_unsafe_dir);
2479 spin_lock(&ci->i_unsafe_lock);
2480 list_add_tail(&req->r_unsafe_dir_item,
2481 &ci->i_unsafe_dirops);
2482 spin_unlock(&ci->i_unsafe_lock);
2486 dout("handle_reply tid %lld result %d\n", tid, result);
2487 rinfo = &req->r_reply_info;
2488 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2489 mutex_unlock(&mdsc->mutex);
2491 mutex_lock(&session->s_mutex);
2493 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2500 if (rinfo->snapblob_len) {
2501 down_write(&mdsc->snap_rwsem);
2502 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2503 rinfo->snapblob + rinfo->snapblob_len,
2504 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2506 downgrade_write(&mdsc->snap_rwsem);
2508 down_read(&mdsc->snap_rwsem);
2511 /* insert trace into our cache */
2512 mutex_lock(&req->r_fill_mutex);
2513 current->journal_info = req;
2514 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2516 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2517 req->r_op == CEPH_MDS_OP_LSSNAP))
2518 ceph_readdir_prepopulate(req, req->r_session);
2519 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2521 current->journal_info = NULL;
2522 mutex_unlock(&req->r_fill_mutex);
2524 up_read(&mdsc->snap_rwsem);
2526 ceph_put_snap_realm(mdsc, realm);
2528 if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2529 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2530 spin_lock(&ci->i_unsafe_lock);
2531 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2532 spin_unlock(&ci->i_unsafe_lock);
2535 mutex_lock(&mdsc->mutex);
2536 if (!req->r_aborted) {
2540 req->r_reply = ceph_msg_get(msg);
2541 req->r_got_result = true;
2544 dout("reply arrived after request %lld was aborted\n", tid);
2546 mutex_unlock(&mdsc->mutex);
2548 mutex_unlock(&session->s_mutex);
2550 /* kick calling process */
2551 complete_request(mdsc, req);
2553 ceph_mdsc_put_request(req);
2560 * handle mds notification that our request has been forwarded.
2562 static void handle_forward(struct ceph_mds_client *mdsc,
2563 struct ceph_mds_session *session,
2564 struct ceph_msg *msg)
2566 struct ceph_mds_request *req;
2567 u64 tid = le64_to_cpu(msg->hdr.tid);
2571 void *p = msg->front.iov_base;
2572 void *end = p + msg->front.iov_len;
2574 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2575 next_mds = ceph_decode_32(&p);
2576 fwd_seq = ceph_decode_32(&p);
2578 mutex_lock(&mdsc->mutex);
2579 req = lookup_get_request(mdsc, tid);
2581 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2582 goto out; /* dup reply? */
2585 if (req->r_aborted) {
2586 dout("forward tid %llu aborted, unregistering\n", tid);
2587 __unregister_request(mdsc, req);
2588 } else if (fwd_seq <= req->r_num_fwd) {
2589 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2590 tid, next_mds, req->r_num_fwd, fwd_seq);
2592 /* resend. forward race not possible; mds would drop */
2593 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2595 BUG_ON(req->r_got_result);
2596 req->r_attempts = 0;
2597 req->r_num_fwd = fwd_seq;
2598 req->r_resend_mds = next_mds;
2599 put_request_session(req);
2600 __do_request(mdsc, req);
2602 ceph_mdsc_put_request(req);
2604 mutex_unlock(&mdsc->mutex);
2608 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2612 * handle a mds session control message
2614 static void handle_session(struct ceph_mds_session *session,
2615 struct ceph_msg *msg)
2617 struct ceph_mds_client *mdsc = session->s_mdsc;
2620 int mds = session->s_mds;
2621 struct ceph_mds_session_head *h = msg->front.iov_base;
2625 if (msg->front.iov_len != sizeof(*h))
2627 op = le32_to_cpu(h->op);
2628 seq = le64_to_cpu(h->seq);
2630 mutex_lock(&mdsc->mutex);
2631 if (op == CEPH_SESSION_CLOSE)
2632 __unregister_session(mdsc, session);
2633 /* FIXME: this ttl calculation is generous */
2634 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2635 mutex_unlock(&mdsc->mutex);
2637 mutex_lock(&session->s_mutex);
2639 dout("handle_session mds%d %s %p state %s seq %llu\n",
2640 mds, ceph_session_op_name(op), session,
2641 ceph_session_state_name(session->s_state), seq);
2643 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2644 session->s_state = CEPH_MDS_SESSION_OPEN;
2645 pr_info("mds%d came back\n", session->s_mds);
2649 case CEPH_SESSION_OPEN:
2650 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2651 pr_info("mds%d reconnect success\n", session->s_mds);
2652 session->s_state = CEPH_MDS_SESSION_OPEN;
2653 renewed_caps(mdsc, session, 0);
2656 __close_session(mdsc, session);
2659 case CEPH_SESSION_RENEWCAPS:
2660 if (session->s_renew_seq == seq)
2661 renewed_caps(mdsc, session, 1);
2664 case CEPH_SESSION_CLOSE:
2665 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2666 pr_info("mds%d reconnect denied\n", session->s_mds);
2667 cleanup_session_requests(mdsc, session);
2668 remove_session_caps(session);
2669 wake = 2; /* for good measure */
2670 wake_up_all(&mdsc->session_close_wq);
2673 case CEPH_SESSION_STALE:
2674 pr_info("mds%d caps went stale, renewing\n",
2676 spin_lock(&session->s_gen_ttl_lock);
2677 session->s_cap_gen++;
2678 session->s_cap_ttl = jiffies - 1;
2679 spin_unlock(&session->s_gen_ttl_lock);
2680 send_renew_caps(mdsc, session);
2683 case CEPH_SESSION_RECALL_STATE:
2684 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2687 case CEPH_SESSION_FLUSHMSG:
2688 send_flushmsg_ack(mdsc, session, seq);
2691 case CEPH_SESSION_FORCE_RO:
2692 dout("force_session_readonly %p\n", session);
2693 spin_lock(&session->s_cap_lock);
2694 session->s_readonly = true;
2695 spin_unlock(&session->s_cap_lock);
2696 wake_up_session_caps(session, 0);
2700 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2704 mutex_unlock(&session->s_mutex);
2706 mutex_lock(&mdsc->mutex);
2707 __wake_requests(mdsc, &session->s_waiting);
2709 kick_requests(mdsc, mds);
2710 mutex_unlock(&mdsc->mutex);
2715 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2716 (int)msg->front.iov_len);
2723 * called under session->mutex.
2725 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2726 struct ceph_mds_session *session)
2728 struct ceph_mds_request *req, *nreq;
2732 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2734 mutex_lock(&mdsc->mutex);
2735 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2736 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2738 ceph_msg_get(req->r_request);
2739 ceph_con_send(&session->s_con, req->r_request);
2744 * also re-send old requests when MDS enters reconnect stage. So that MDS
2745 * can process completed request in clientreplay stage.
2747 p = rb_first(&mdsc->request_tree);
2749 req = rb_entry(p, struct ceph_mds_request, r_node);
2751 if (req->r_got_unsafe)
2753 if (req->r_attempts == 0)
2754 continue; /* only old requests */
2755 if (req->r_session &&
2756 req->r_session->s_mds == session->s_mds) {
2757 err = __prepare_send_request(mdsc, req,
2758 session->s_mds, true);
2760 ceph_msg_get(req->r_request);
2761 ceph_con_send(&session->s_con, req->r_request);
2765 mutex_unlock(&mdsc->mutex);
2769 * Encode information about a cap for a reconnect with the MDS.
2771 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2775 struct ceph_mds_cap_reconnect v2;
2776 struct ceph_mds_cap_reconnect_v1 v1;
2779 struct ceph_inode_info *ci;
2780 struct ceph_reconnect_state *recon_state = arg;
2781 struct ceph_pagelist *pagelist = recon_state->pagelist;
2785 struct dentry *dentry;
2789 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2790 inode, ceph_vinop(inode), cap, cap->cap_id,
2791 ceph_cap_string(cap->issued));
2792 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2796 dentry = d_find_alias(inode);
2798 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2800 err = PTR_ERR(path);
2807 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2811 spin_lock(&ci->i_ceph_lock);
2812 cap->seq = 0; /* reset cap seq */
2813 cap->issue_seq = 0; /* and issue_seq */
2814 cap->mseq = 0; /* and migrate_seq */
2815 cap->cap_gen = cap->session->s_cap_gen;
2817 if (recon_state->flock) {
2818 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2819 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2820 rec.v2.issued = cpu_to_le32(cap->issued);
2821 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2822 rec.v2.pathbase = cpu_to_le64(pathbase);
2823 rec.v2.flock_len = 0;
2824 reclen = sizeof(rec.v2);
2826 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2827 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2828 rec.v1.issued = cpu_to_le32(cap->issued);
2829 rec.v1.size = cpu_to_le64(inode->i_size);
2830 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2831 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2832 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2833 rec.v1.pathbase = cpu_to_le64(pathbase);
2834 reclen = sizeof(rec.v1);
2836 spin_unlock(&ci->i_ceph_lock);
2838 if (recon_state->flock) {
2839 int num_fcntl_locks, num_flock_locks;
2840 struct ceph_filelock *flocks;
2843 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2844 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2845 sizeof(struct ceph_filelock), GFP_NOFS);
2850 err = ceph_encode_locks_to_buffer(inode, flocks,
2860 * number of encoded locks is stable, so copy to pagelist
2862 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2863 (num_fcntl_locks+num_flock_locks) *
2864 sizeof(struct ceph_filelock));
2865 err = ceph_pagelist_append(pagelist, &rec, reclen);
2867 err = ceph_locks_to_pagelist(flocks, pagelist,
2872 err = ceph_pagelist_append(pagelist, &rec, reclen);
2875 recon_state->nr_caps++;
2885 * If an MDS fails and recovers, clients need to reconnect in order to
2886 * reestablish shared state. This includes all caps issued through
2887 * this session _and_ the snap_realm hierarchy. Because it's not
2888 * clear which snap realms the mds cares about, we send everything we
2889 * know about.. that ensures we'll then get any new info the
2890 * recovering MDS might have.
2892 * This is a relatively heavyweight operation, but it's rare.
2894 * called with mdsc->mutex held.
2896 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2897 struct ceph_mds_session *session)
2899 struct ceph_msg *reply;
2901 int mds = session->s_mds;
2904 struct ceph_pagelist *pagelist;
2905 struct ceph_reconnect_state recon_state;
2907 pr_info("mds%d reconnect start\n", mds);
2909 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2911 goto fail_nopagelist;
2912 ceph_pagelist_init(pagelist);
2914 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2918 mutex_lock(&session->s_mutex);
2919 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2922 dout("session %p state %s\n", session,
2923 ceph_session_state_name(session->s_state));
2925 spin_lock(&session->s_gen_ttl_lock);
2926 session->s_cap_gen++;
2927 spin_unlock(&session->s_gen_ttl_lock);
2929 spin_lock(&session->s_cap_lock);
2930 /* don't know if session is readonly */
2931 session->s_readonly = 0;
2933 * notify __ceph_remove_cap() that we are composing cap reconnect.
2934 * If a cap get released before being added to the cap reconnect,
2935 * __ceph_remove_cap() should skip queuing cap release.
2937 session->s_cap_reconnect = 1;
2938 /* drop old cap expires; we're about to reestablish that state */
2939 cleanup_cap_releases(mdsc, session);
2941 /* trim unused caps to reduce MDS's cache rejoin time */
2942 if (mdsc->fsc->sb->s_root)
2943 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2945 ceph_con_close(&session->s_con);
2946 ceph_con_open(&session->s_con,
2947 CEPH_ENTITY_TYPE_MDS, mds,
2948 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2950 /* replay unsafe requests */
2951 replay_unsafe_requests(mdsc, session);
2953 down_read(&mdsc->snap_rwsem);
2955 /* traverse this session's caps */
2956 s_nr_caps = session->s_nr_caps;
2957 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2961 recon_state.nr_caps = 0;
2962 recon_state.pagelist = pagelist;
2963 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2964 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2968 spin_lock(&session->s_cap_lock);
2969 session->s_cap_reconnect = 0;
2970 spin_unlock(&session->s_cap_lock);
2973 * snaprealms. we provide mds with the ino, seq (version), and
2974 * parent for all of our realms. If the mds has any newer info,
2977 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2978 struct ceph_snap_realm *realm =
2979 rb_entry(p, struct ceph_snap_realm, node);
2980 struct ceph_mds_snaprealm_reconnect sr_rec;
2982 dout(" adding snap realm %llx seq %lld parent %llx\n",
2983 realm->ino, realm->seq, realm->parent_ino);
2984 sr_rec.ino = cpu_to_le64(realm->ino);
2985 sr_rec.seq = cpu_to_le64(realm->seq);
2986 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2987 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2992 if (recon_state.flock)
2993 reply->hdr.version = cpu_to_le16(2);
2995 /* raced with cap release? */
2996 if (s_nr_caps != recon_state.nr_caps) {
2997 struct page *page = list_first_entry(&pagelist->head,
2999 __le32 *addr = kmap_atomic(page);
3000 *addr = cpu_to_le32(recon_state.nr_caps);
3001 kunmap_atomic(addr);
3004 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3005 ceph_msg_data_add_pagelist(reply, pagelist);
3007 ceph_early_kick_flushing_caps(mdsc, session);
3009 ceph_con_send(&session->s_con, reply);
3011 mutex_unlock(&session->s_mutex);
3013 mutex_lock(&mdsc->mutex);
3014 __wake_requests(mdsc, &session->s_waiting);
3015 mutex_unlock(&mdsc->mutex);
3017 up_read(&mdsc->snap_rwsem);
3021 ceph_msg_put(reply);
3022 up_read(&mdsc->snap_rwsem);
3023 mutex_unlock(&session->s_mutex);
3025 ceph_pagelist_release(pagelist);
3027 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3033 * compare old and new mdsmaps, kicking requests
3034 * and closing out old connections as necessary
3036 * called under mdsc->mutex.
3038 static void check_new_map(struct ceph_mds_client *mdsc,
3039 struct ceph_mdsmap *newmap,
3040 struct ceph_mdsmap *oldmap)
3043 int oldstate, newstate;
3044 struct ceph_mds_session *s;
3046 dout("check_new_map new %u old %u\n",
3047 newmap->m_epoch, oldmap->m_epoch);
3049 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3050 if (mdsc->sessions[i] == NULL)
3052 s = mdsc->sessions[i];
3053 oldstate = ceph_mdsmap_get_state(oldmap, i);
3054 newstate = ceph_mdsmap_get_state(newmap, i);
3056 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3057 i, ceph_mds_state_name(oldstate),
3058 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3059 ceph_mds_state_name(newstate),
3060 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3061 ceph_session_state_name(s->s_state));
3063 if (i >= newmap->m_max_mds ||
3064 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3065 ceph_mdsmap_get_addr(newmap, i),
3066 sizeof(struct ceph_entity_addr))) {
3067 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3068 /* the session never opened, just close it
3070 __wake_requests(mdsc, &s->s_waiting);
3071 __unregister_session(mdsc, s);
3074 mutex_unlock(&mdsc->mutex);
3075 mutex_lock(&s->s_mutex);
3076 mutex_lock(&mdsc->mutex);
3077 ceph_con_close(&s->s_con);
3078 mutex_unlock(&s->s_mutex);
3079 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3081 } else if (oldstate == newstate) {
3082 continue; /* nothing new with this mds */
3088 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3089 newstate >= CEPH_MDS_STATE_RECONNECT) {
3090 mutex_unlock(&mdsc->mutex);
3091 send_mds_reconnect(mdsc, s);
3092 mutex_lock(&mdsc->mutex);
3096 * kick request on any mds that has gone active.
3098 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3099 newstate >= CEPH_MDS_STATE_ACTIVE) {
3100 if (oldstate != CEPH_MDS_STATE_CREATING &&
3101 oldstate != CEPH_MDS_STATE_STARTING)
3102 pr_info("mds%d recovery completed\n", s->s_mds);
3103 kick_requests(mdsc, i);
3104 ceph_kick_flushing_caps(mdsc, s);
3105 wake_up_session_caps(s, 1);
3109 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3110 s = mdsc->sessions[i];
3113 if (!ceph_mdsmap_is_laggy(newmap, i))
3115 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3116 s->s_state == CEPH_MDS_SESSION_HUNG ||
3117 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3118 dout(" connecting to export targets of laggy mds%d\n",
3120 __open_export_target_sessions(mdsc, s);
3132 * caller must hold session s_mutex, dentry->d_lock
3134 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3136 struct ceph_dentry_info *di = ceph_dentry(dentry);
3138 ceph_put_mds_session(di->lease_session);
3139 di->lease_session = NULL;
3142 static void handle_lease(struct ceph_mds_client *mdsc,
3143 struct ceph_mds_session *session,
3144 struct ceph_msg *msg)
3146 struct super_block *sb = mdsc->fsc->sb;
3147 struct inode *inode;
3148 struct dentry *parent, *dentry;
3149 struct ceph_dentry_info *di;
3150 int mds = session->s_mds;
3151 struct ceph_mds_lease *h = msg->front.iov_base;
3153 struct ceph_vino vino;
3157 dout("handle_lease from mds%d\n", mds);
3160 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3162 vino.ino = le64_to_cpu(h->ino);
3163 vino.snap = CEPH_NOSNAP;
3164 seq = le32_to_cpu(h->seq);
3165 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3166 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3167 if (dname.len != get_unaligned_le32(h+1))
3171 inode = ceph_find_inode(sb, vino);
3172 dout("handle_lease %s, ino %llx %p %.*s\n",
3173 ceph_lease_op_name(h->action), vino.ino, inode,
3174 dname.len, dname.name);
3176 mutex_lock(&session->s_mutex);
3179 if (inode == NULL) {
3180 dout("handle_lease no inode %llx\n", vino.ino);
3185 parent = d_find_alias(inode);
3187 dout("no parent dentry on inode %p\n", inode);
3189 goto release; /* hrm... */
3191 dname.hash = full_name_hash(dname.name, dname.len);
3192 dentry = d_lookup(parent, &dname);
3197 spin_lock(&dentry->d_lock);
3198 di = ceph_dentry(dentry);
3199 switch (h->action) {
3200 case CEPH_MDS_LEASE_REVOKE:
3201 if (di->lease_session == session) {
3202 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3203 h->seq = cpu_to_le32(di->lease_seq);
3204 __ceph_mdsc_drop_dentry_lease(dentry);
3209 case CEPH_MDS_LEASE_RENEW:
3210 if (di->lease_session == session &&
3211 di->lease_gen == session->s_cap_gen &&
3212 di->lease_renew_from &&
3213 di->lease_renew_after == 0) {
3214 unsigned long duration =
3215 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3217 di->lease_seq = seq;
3218 dentry->d_time = di->lease_renew_from + duration;
3219 di->lease_renew_after = di->lease_renew_from +
3221 di->lease_renew_from = 0;
3225 spin_unlock(&dentry->d_lock);
3232 /* let's just reuse the same message */
3233 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3235 ceph_con_send(&session->s_con, msg);
3239 mutex_unlock(&session->s_mutex);
3243 pr_err("corrupt lease message\n");
3247 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3248 struct inode *inode,
3249 struct dentry *dentry, char action,
3252 struct ceph_msg *msg;
3253 struct ceph_mds_lease *lease;
3254 int len = sizeof(*lease) + sizeof(u32);
3257 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3258 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3259 dnamelen = dentry->d_name.len;
3262 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3265 lease = msg->front.iov_base;
3266 lease->action = action;
3267 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3268 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3269 lease->seq = cpu_to_le32(seq);
3270 put_unaligned_le32(dnamelen, lease + 1);
3271 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3274 * if this is a preemptive lease RELEASE, no need to
3275 * flush request stream, since the actual request will
3278 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3280 ceph_con_send(&session->s_con, msg);
3284 * Preemptively release a lease we expect to invalidate anyway.
3285 * Pass @inode always, @dentry is optional.
3287 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3288 struct dentry *dentry)
3290 struct ceph_dentry_info *di;
3291 struct ceph_mds_session *session;
3294 BUG_ON(inode == NULL);
3295 BUG_ON(dentry == NULL);
3297 /* is dentry lease valid? */
3298 spin_lock(&dentry->d_lock);
3299 di = ceph_dentry(dentry);
3300 if (!di || !di->lease_session ||
3301 di->lease_session->s_mds < 0 ||
3302 di->lease_gen != di->lease_session->s_cap_gen ||
3303 !time_before(jiffies, dentry->d_time)) {
3304 dout("lease_release inode %p dentry %p -- "
3307 spin_unlock(&dentry->d_lock);
3311 /* we do have a lease on this dentry; note mds and seq */
3312 session = ceph_get_mds_session(di->lease_session);
3313 seq = di->lease_seq;
3314 __ceph_mdsc_drop_dentry_lease(dentry);
3315 spin_unlock(&dentry->d_lock);
3317 dout("lease_release inode %p dentry %p to mds%d\n",
3318 inode, dentry, session->s_mds);
3319 ceph_mdsc_lease_send_msg(session, inode, dentry,
3320 CEPH_MDS_LEASE_RELEASE, seq);
3321 ceph_put_mds_session(session);
3325 * drop all leases (and dentry refs) in preparation for umount
3327 static void drop_leases(struct ceph_mds_client *mdsc)
3331 dout("drop_leases\n");
3332 mutex_lock(&mdsc->mutex);
3333 for (i = 0; i < mdsc->max_sessions; i++) {
3334 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3337 mutex_unlock(&mdsc->mutex);
3338 mutex_lock(&s->s_mutex);
3339 mutex_unlock(&s->s_mutex);
3340 ceph_put_mds_session(s);
3341 mutex_lock(&mdsc->mutex);
3343 mutex_unlock(&mdsc->mutex);
3349 * delayed work -- periodically trim expired leases, renew caps with mds
3351 static void schedule_delayed(struct ceph_mds_client *mdsc)
3354 unsigned hz = round_jiffies_relative(HZ * delay);
3355 schedule_delayed_work(&mdsc->delayed_work, hz);
3358 static void delayed_work(struct work_struct *work)
3361 struct ceph_mds_client *mdsc =
3362 container_of(work, struct ceph_mds_client, delayed_work.work);
3366 dout("mdsc delayed_work\n");
3367 ceph_check_delayed_caps(mdsc);
3369 mutex_lock(&mdsc->mutex);
3370 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3371 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3372 mdsc->last_renew_caps);
3374 mdsc->last_renew_caps = jiffies;
3376 for (i = 0; i < mdsc->max_sessions; i++) {
3377 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3380 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3381 dout("resending session close request for mds%d\n",
3383 request_close_session(mdsc, s);
3384 ceph_put_mds_session(s);
3387 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3388 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3389 s->s_state = CEPH_MDS_SESSION_HUNG;
3390 pr_info("mds%d hung\n", s->s_mds);
3393 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3394 /* this mds is failed or recovering, just wait */
3395 ceph_put_mds_session(s);
3398 mutex_unlock(&mdsc->mutex);
3400 mutex_lock(&s->s_mutex);
3402 send_renew_caps(mdsc, s);
3404 ceph_con_keepalive(&s->s_con);
3405 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3406 s->s_state == CEPH_MDS_SESSION_HUNG)
3407 ceph_send_cap_releases(mdsc, s);
3408 mutex_unlock(&s->s_mutex);
3409 ceph_put_mds_session(s);
3411 mutex_lock(&mdsc->mutex);
3413 mutex_unlock(&mdsc->mutex);
3415 schedule_delayed(mdsc);
3418 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3421 struct ceph_mds_client *mdsc;
3423 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3428 mutex_init(&mdsc->mutex);
3429 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3430 if (mdsc->mdsmap == NULL) {
3435 init_completion(&mdsc->safe_umount_waiters);
3436 init_waitqueue_head(&mdsc->session_close_wq);
3437 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3438 mdsc->sessions = NULL;
3439 atomic_set(&mdsc->num_sessions, 0);
3440 mdsc->max_sessions = 0;
3442 mdsc->last_snap_seq = 0;
3443 init_rwsem(&mdsc->snap_rwsem);
3444 mdsc->snap_realms = RB_ROOT;
3445 INIT_LIST_HEAD(&mdsc->snap_empty);
3446 spin_lock_init(&mdsc->snap_empty_lock);
3448 mdsc->oldest_tid = 0;
3449 mdsc->request_tree = RB_ROOT;
3450 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3451 mdsc->last_renew_caps = jiffies;
3452 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3453 spin_lock_init(&mdsc->cap_delay_lock);
3454 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3455 spin_lock_init(&mdsc->snap_flush_lock);
3456 mdsc->last_cap_flush_tid = 1;
3457 mdsc->cap_flush_tree = RB_ROOT;
3458 INIT_LIST_HEAD(&mdsc->cap_dirty);
3459 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3460 mdsc->num_cap_flushing = 0;
3461 spin_lock_init(&mdsc->cap_dirty_lock);
3462 init_waitqueue_head(&mdsc->cap_flushing_wq);
3463 spin_lock_init(&mdsc->dentry_lru_lock);
3464 INIT_LIST_HEAD(&mdsc->dentry_lru);
3466 ceph_caps_init(mdsc);
3467 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3469 init_rwsem(&mdsc->pool_perm_rwsem);
3470 mdsc->pool_perm_tree = RB_ROOT;
3476 * Wait for safe replies on open mds requests. If we time out, drop
3477 * all requests from the tree to avoid dangling dentry refs.
3479 static void wait_requests(struct ceph_mds_client *mdsc)
3481 struct ceph_options *opts = mdsc->fsc->client->options;
3482 struct ceph_mds_request *req;
3484 mutex_lock(&mdsc->mutex);
3485 if (__get_oldest_req(mdsc)) {
3486 mutex_unlock(&mdsc->mutex);
3488 dout("wait_requests waiting for requests\n");
3489 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3490 ceph_timeout_jiffies(opts->mount_timeout));
3492 /* tear down remaining requests */
3493 mutex_lock(&mdsc->mutex);
3494 while ((req = __get_oldest_req(mdsc))) {
3495 dout("wait_requests timed out on tid %llu\n",
3497 __unregister_request(mdsc, req);
3500 mutex_unlock(&mdsc->mutex);
3501 dout("wait_requests done\n");
3505 * called before mount is ro, and before dentries are torn down.
3506 * (hmm, does this still race with new lookups?)
3508 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3510 dout("pre_umount\n");
3514 ceph_flush_dirty_caps(mdsc);
3515 wait_requests(mdsc);
3518 * wait for reply handlers to drop their request refs and
3519 * their inode/dcache refs
3525 * wait for all write mds requests to flush.
3527 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3529 struct ceph_mds_request *req = NULL, *nextreq;
3532 mutex_lock(&mdsc->mutex);
3533 dout("wait_unsafe_requests want %lld\n", want_tid);
3535 req = __get_oldest_req(mdsc);
3536 while (req && req->r_tid <= want_tid) {
3537 /* find next request */
3538 n = rb_next(&req->r_node);
3540 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3543 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3544 (req->r_op & CEPH_MDS_OP_WRITE)) {
3546 ceph_mdsc_get_request(req);
3548 ceph_mdsc_get_request(nextreq);
3549 mutex_unlock(&mdsc->mutex);
3550 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3551 req->r_tid, want_tid);
3552 wait_for_completion(&req->r_safe_completion);
3553 mutex_lock(&mdsc->mutex);
3554 ceph_mdsc_put_request(req);
3556 break; /* next dne before, so we're done! */
3557 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3558 /* next request was removed from tree */
3559 ceph_mdsc_put_request(nextreq);
3562 ceph_mdsc_put_request(nextreq); /* won't go away */
3566 mutex_unlock(&mdsc->mutex);
3567 dout("wait_unsafe_requests done\n");
3570 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3572 u64 want_tid, want_flush, want_snap;
3574 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3578 mutex_lock(&mdsc->mutex);
3579 want_tid = mdsc->last_tid;
3580 mutex_unlock(&mdsc->mutex);
3582 ceph_flush_dirty_caps(mdsc);
3583 spin_lock(&mdsc->cap_dirty_lock);
3584 want_flush = mdsc->last_cap_flush_tid;
3585 spin_unlock(&mdsc->cap_dirty_lock);
3587 down_read(&mdsc->snap_rwsem);
3588 want_snap = mdsc->last_snap_seq;
3589 up_read(&mdsc->snap_rwsem);
3591 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3592 want_tid, want_flush, want_snap);
3594 wait_unsafe_requests(mdsc, want_tid);
3595 wait_caps_flush(mdsc, want_flush, want_snap);
3599 * true if all sessions are closed, or we force unmount
3601 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3603 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3605 return atomic_read(&mdsc->num_sessions) == 0;
3609 * called after sb is ro.
3611 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3613 struct ceph_options *opts = mdsc->fsc->client->options;
3614 struct ceph_mds_session *session;
3617 dout("close_sessions\n");
3619 /* close sessions */
3620 mutex_lock(&mdsc->mutex);
3621 for (i = 0; i < mdsc->max_sessions; i++) {
3622 session = __ceph_lookup_mds_session(mdsc, i);
3625 mutex_unlock(&mdsc->mutex);
3626 mutex_lock(&session->s_mutex);
3627 __close_session(mdsc, session);
3628 mutex_unlock(&session->s_mutex);
3629 ceph_put_mds_session(session);
3630 mutex_lock(&mdsc->mutex);
3632 mutex_unlock(&mdsc->mutex);
3634 dout("waiting for sessions to close\n");
3635 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3636 ceph_timeout_jiffies(opts->mount_timeout));
3638 /* tear down remaining sessions */
3639 mutex_lock(&mdsc->mutex);
3640 for (i = 0; i < mdsc->max_sessions; i++) {
3641 if (mdsc->sessions[i]) {
3642 session = get_session(mdsc->sessions[i]);
3643 __unregister_session(mdsc, session);
3644 mutex_unlock(&mdsc->mutex);
3645 mutex_lock(&session->s_mutex);
3646 remove_session_caps(session);
3647 mutex_unlock(&session->s_mutex);
3648 ceph_put_mds_session(session);
3649 mutex_lock(&mdsc->mutex);
3652 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3653 mutex_unlock(&mdsc->mutex);
3655 ceph_cleanup_empty_realms(mdsc);
3657 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3662 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3664 struct ceph_mds_session *session;
3667 dout("force umount\n");
3669 mutex_lock(&mdsc->mutex);
3670 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3671 session = __ceph_lookup_mds_session(mdsc, mds);
3674 mutex_unlock(&mdsc->mutex);
3675 mutex_lock(&session->s_mutex);
3676 __close_session(mdsc, session);
3677 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3678 cleanup_session_requests(mdsc, session);
3679 remove_session_caps(session);
3681 mutex_unlock(&session->s_mutex);
3682 ceph_put_mds_session(session);
3683 mutex_lock(&mdsc->mutex);
3684 kick_requests(mdsc, mds);
3686 __wake_requests(mdsc, &mdsc->waiting_for_map);
3687 mutex_unlock(&mdsc->mutex);
3690 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3693 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3695 ceph_mdsmap_destroy(mdsc->mdsmap);
3696 kfree(mdsc->sessions);
3697 ceph_caps_finalize(mdsc);
3698 ceph_pool_perm_destroy(mdsc);
3701 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3703 struct ceph_mds_client *mdsc = fsc->mdsc;
3705 dout("mdsc_destroy %p\n", mdsc);
3706 ceph_mdsc_stop(mdsc);
3708 /* flush out any connection work with references to us */
3713 dout("mdsc_destroy %p done\n", mdsc);
3718 * handle mds map update.
3720 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3724 void *p = msg->front.iov_base;
3725 void *end = p + msg->front.iov_len;
3726 struct ceph_mdsmap *newmap, *oldmap;
3727 struct ceph_fsid fsid;
3730 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3731 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3732 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3734 epoch = ceph_decode_32(&p);
3735 maplen = ceph_decode_32(&p);
3736 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3738 /* do we need it? */
3739 mutex_lock(&mdsc->mutex);
3740 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3741 dout("handle_map epoch %u <= our %u\n",
3742 epoch, mdsc->mdsmap->m_epoch);
3743 mutex_unlock(&mdsc->mutex);
3747 newmap = ceph_mdsmap_decode(&p, end);
3748 if (IS_ERR(newmap)) {
3749 err = PTR_ERR(newmap);
3753 /* swap into place */
3755 oldmap = mdsc->mdsmap;
3756 mdsc->mdsmap = newmap;
3757 check_new_map(mdsc, newmap, oldmap);
3758 ceph_mdsmap_destroy(oldmap);
3760 mdsc->mdsmap = newmap; /* first mds map */
3762 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3764 __wake_requests(mdsc, &mdsc->waiting_for_map);
3765 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3766 mdsc->mdsmap->m_epoch);
3768 mutex_unlock(&mdsc->mutex);
3769 schedule_delayed(mdsc);
3773 mutex_unlock(&mdsc->mutex);
3775 pr_err("error decoding mdsmap %d\n", err);
3779 static struct ceph_connection *con_get(struct ceph_connection *con)
3781 struct ceph_mds_session *s = con->private;
3783 if (get_session(s)) {
3784 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3787 dout("mdsc con_get %p FAIL\n", s);
3791 static void con_put(struct ceph_connection *con)
3793 struct ceph_mds_session *s = con->private;
3795 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3796 ceph_put_mds_session(s);
3800 * if the client is unresponsive for long enough, the mds will kill
3801 * the session entirely.
3803 static void peer_reset(struct ceph_connection *con)
3805 struct ceph_mds_session *s = con->private;
3806 struct ceph_mds_client *mdsc = s->s_mdsc;
3808 pr_warn("mds%d closed our session\n", s->s_mds);
3809 send_mds_reconnect(mdsc, s);
3812 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3814 struct ceph_mds_session *s = con->private;
3815 struct ceph_mds_client *mdsc = s->s_mdsc;
3816 int type = le16_to_cpu(msg->hdr.type);
3818 mutex_lock(&mdsc->mutex);
3819 if (__verify_registered_session(mdsc, s) < 0) {
3820 mutex_unlock(&mdsc->mutex);
3823 mutex_unlock(&mdsc->mutex);
3826 case CEPH_MSG_MDS_MAP:
3827 ceph_mdsc_handle_map(mdsc, msg);
3829 case CEPH_MSG_CLIENT_SESSION:
3830 handle_session(s, msg);
3832 case CEPH_MSG_CLIENT_REPLY:
3833 handle_reply(s, msg);
3835 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3836 handle_forward(mdsc, s, msg);
3838 case CEPH_MSG_CLIENT_CAPS:
3839 ceph_handle_caps(s, msg);
3841 case CEPH_MSG_CLIENT_SNAP:
3842 ceph_handle_snap(mdsc, s, msg);
3844 case CEPH_MSG_CLIENT_LEASE:
3845 handle_lease(mdsc, s, msg);
3849 pr_err("received unknown message type %d %s\n", type,
3850 ceph_msg_type_name(type));
3861 * Note: returned pointer is the address of a structure that's
3862 * managed separately. Caller must *not* attempt to free it.
3864 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3865 int *proto, int force_new)
3867 struct ceph_mds_session *s = con->private;
3868 struct ceph_mds_client *mdsc = s->s_mdsc;
3869 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3870 struct ceph_auth_handshake *auth = &s->s_auth;
3872 if (force_new && auth->authorizer) {
3873 ceph_auth_destroy_authorizer(auth->authorizer);
3874 auth->authorizer = NULL;
3876 if (!auth->authorizer) {
3877 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3880 return ERR_PTR(ret);
3882 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3885 return ERR_PTR(ret);
3887 *proto = ac->protocol;
3893 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3895 struct ceph_mds_session *s = con->private;
3896 struct ceph_mds_client *mdsc = s->s_mdsc;
3897 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3899 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3902 static int invalidate_authorizer(struct ceph_connection *con)
3904 struct ceph_mds_session *s = con->private;
3905 struct ceph_mds_client *mdsc = s->s_mdsc;
3906 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3908 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3910 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3913 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3914 struct ceph_msg_header *hdr, int *skip)
3916 struct ceph_msg *msg;
3917 int type = (int) le16_to_cpu(hdr->type);
3918 int front_len = (int) le32_to_cpu(hdr->front_len);
3924 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3926 pr_err("unable to allocate msg type %d len %d\n",
3934 static int mds_sign_message(struct ceph_msg *msg)
3936 struct ceph_mds_session *s = msg->con->private;
3937 struct ceph_auth_handshake *auth = &s->s_auth;
3939 return ceph_auth_sign_message(auth, msg);
3942 static int mds_check_message_signature(struct ceph_msg *msg)
3944 struct ceph_mds_session *s = msg->con->private;
3945 struct ceph_auth_handshake *auth = &s->s_auth;
3947 return ceph_auth_check_message_signature(auth, msg);
3950 static const struct ceph_connection_operations mds_con_ops = {
3953 .dispatch = dispatch,
3954 .get_authorizer = get_authorizer,
3955 .verify_authorizer_reply = verify_authorizer_reply,
3956 .invalidate_authorizer = invalidate_authorizer,
3957 .peer_reset = peer_reset,
3958 .alloc_msg = mds_alloc_msg,
3959 .sign_message = mds_sign_message,
3960 .check_message_signature = mds_check_message_signature,