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>
13 #include "mds_client.h"
15 #include <linux/ceph/ceph_features.h>
16 #include <linux/ceph/messenger.h>
17 #include <linux/ceph/decode.h>
18 #include <linux/ceph/pagelist.h>
19 #include <linux/ceph/auth.h>
20 #include <linux/ceph/debugfs.h>
23 * A cluster of MDS (metadata server) daemons is responsible for
24 * managing the file system namespace (the directory hierarchy and
25 * inodes) and for coordinating shared access to storage. Metadata is
26 * partitioning hierarchically across a number of servers, and that
27 * partition varies over time as the cluster adjusts the distribution
28 * in order to balance load.
30 * The MDS client is primarily responsible to managing synchronous
31 * metadata requests for operations like open, unlink, and so forth.
32 * If there is a MDS failure, we find out about it when we (possibly
33 * request and) receive a new MDS map, and can resubmit affected
36 * For the most part, though, we take advantage of a lossless
37 * communications channel to the MDS, and do not need to worry about
38 * timing out or resubmitting requests.
40 * We maintain a stateful "session" with each MDS we interact with.
41 * Within each session, we sent periodic heartbeat messages to ensure
42 * any capabilities or leases we have been issues remain valid. If
43 * the session times out and goes stale, our leases and capabilities
44 * are no longer valid.
47 struct ceph_reconnect_state {
49 struct ceph_pagelist *pagelist;
53 static void __wake_requests(struct ceph_mds_client *mdsc,
54 struct list_head *head);
56 static const struct ceph_connection_operations mds_con_ops;
64 * parse individual inode info
66 static int parse_reply_info_in(void **p, void *end,
67 struct ceph_mds_reply_info_in *info,
73 *p += sizeof(struct ceph_mds_reply_inode) +
74 sizeof(*info->in->fragtree.splits) *
75 le32_to_cpu(info->in->fragtree.nsplits);
77 ceph_decode_32_safe(p, end, info->symlink_len, bad);
78 ceph_decode_need(p, end, info->symlink_len, bad);
80 *p += info->symlink_len;
82 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
83 ceph_decode_copy_safe(p, end, &info->dir_layout,
84 sizeof(info->dir_layout), bad);
86 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
88 ceph_decode_32_safe(p, end, info->xattr_len, bad);
89 ceph_decode_need(p, end, info->xattr_len, bad);
90 info->xattr_data = *p;
91 *p += info->xattr_len;
93 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
94 ceph_decode_64_safe(p, end, info->inline_version, bad);
95 ceph_decode_32_safe(p, end, info->inline_len, bad);
96 ceph_decode_need(p, end, info->inline_len, bad);
97 info->inline_data = *p;
98 *p += info->inline_len;
100 info->inline_version = CEPH_INLINE_NONE;
108 * parse a normal reply, which may contain a (dir+)dentry and/or a
111 static int parse_reply_info_trace(void **p, void *end,
112 struct ceph_mds_reply_info_parsed *info,
117 if (info->head->is_dentry) {
118 err = parse_reply_info_in(p, end, &info->diri, features);
122 if (unlikely(*p + sizeof(*info->dirfrag) > end))
125 *p += sizeof(*info->dirfrag) +
126 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
127 if (unlikely(*p > end))
130 ceph_decode_32_safe(p, end, info->dname_len, bad);
131 ceph_decode_need(p, end, info->dname_len, bad);
133 *p += info->dname_len;
135 *p += sizeof(*info->dlease);
138 if (info->head->is_target) {
139 err = parse_reply_info_in(p, end, &info->targeti, features);
144 if (unlikely(*p != end))
151 pr_err("problem parsing mds trace %d\n", err);
156 * parse readdir results
158 static int parse_reply_info_dir(void **p, void *end,
159 struct ceph_mds_reply_info_parsed *info,
166 if (*p + sizeof(*info->dir_dir) > end)
168 *p += sizeof(*info->dir_dir) +
169 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
173 ceph_decode_need(p, end, sizeof(num) + 2, bad);
174 num = ceph_decode_32(p);
175 info->dir_end = ceph_decode_8(p);
176 info->dir_complete = ceph_decode_8(p);
180 BUG_ON(!info->dir_in);
181 info->dir_dname = (void *)(info->dir_in + num);
182 info->dir_dname_len = (void *)(info->dir_dname + num);
183 info->dir_dlease = (void *)(info->dir_dname_len + num);
184 if ((unsigned long)(info->dir_dlease + num) >
185 (unsigned long)info->dir_in + info->dir_buf_size) {
186 pr_err("dir contents are larger than expected\n");
194 ceph_decode_need(p, end, sizeof(u32)*2, bad);
195 info->dir_dname_len[i] = ceph_decode_32(p);
196 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
197 info->dir_dname[i] = *p;
198 *p += info->dir_dname_len[i];
199 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
201 info->dir_dlease[i] = *p;
202 *p += sizeof(struct ceph_mds_reply_lease);
205 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
220 pr_err("problem parsing dir contents %d\n", err);
225 * parse fcntl F_GETLK results
227 static int parse_reply_info_filelock(void **p, void *end,
228 struct ceph_mds_reply_info_parsed *info,
231 if (*p + sizeof(*info->filelock_reply) > end)
234 info->filelock_reply = *p;
235 *p += sizeof(*info->filelock_reply);
237 if (unlikely(*p != end))
246 * parse create results
248 static int parse_reply_info_create(void **p, void *end,
249 struct ceph_mds_reply_info_parsed *info,
252 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
254 info->has_create_ino = false;
256 info->has_create_ino = true;
257 info->ino = ceph_decode_64(p);
261 if (unlikely(*p != end))
270 * parse extra results
272 static int parse_reply_info_extra(void **p, void *end,
273 struct ceph_mds_reply_info_parsed *info,
276 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
277 return parse_reply_info_filelock(p, end, info, features);
278 else if (info->head->op == CEPH_MDS_OP_READDIR ||
279 info->head->op == CEPH_MDS_OP_LSSNAP)
280 return parse_reply_info_dir(p, end, info, features);
281 else if (info->head->op == CEPH_MDS_OP_CREATE)
282 return parse_reply_info_create(p, end, info, features);
288 * parse entire mds reply
290 static int parse_reply_info(struct ceph_msg *msg,
291 struct ceph_mds_reply_info_parsed *info,
298 info->head = msg->front.iov_base;
299 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
300 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
303 ceph_decode_32_safe(&p, end, len, bad);
305 ceph_decode_need(&p, end, len, bad);
306 err = parse_reply_info_trace(&p, p+len, info, features);
312 ceph_decode_32_safe(&p, end, len, bad);
314 ceph_decode_need(&p, end, len, bad);
315 err = parse_reply_info_extra(&p, p+len, info, features);
321 ceph_decode_32_safe(&p, end, len, bad);
322 info->snapblob_len = len;
333 pr_err("mds parse_reply err %d\n", err);
337 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
341 free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
348 const char *ceph_session_state_name(int s)
351 case CEPH_MDS_SESSION_NEW: return "new";
352 case CEPH_MDS_SESSION_OPENING: return "opening";
353 case CEPH_MDS_SESSION_OPEN: return "open";
354 case CEPH_MDS_SESSION_HUNG: return "hung";
355 case CEPH_MDS_SESSION_CLOSING: return "closing";
356 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
357 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
358 default: return "???";
362 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
364 if (atomic_inc_not_zero(&s->s_ref)) {
365 dout("mdsc get_session %p %d -> %d\n", s,
366 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
369 dout("mdsc get_session %p 0 -- FAIL", s);
374 void ceph_put_mds_session(struct ceph_mds_session *s)
376 dout("mdsc put_session %p %d -> %d\n", s,
377 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
378 if (atomic_dec_and_test(&s->s_ref)) {
379 if (s->s_auth.authorizer)
380 ceph_auth_destroy_authorizer(
381 s->s_mdsc->fsc->client->monc.auth,
382 s->s_auth.authorizer);
388 * called under mdsc->mutex
390 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
393 struct ceph_mds_session *session;
395 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
397 session = mdsc->sessions[mds];
398 dout("lookup_mds_session %p %d\n", session,
399 atomic_read(&session->s_ref));
400 get_session(session);
404 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
406 if (mds >= mdsc->max_sessions)
408 return mdsc->sessions[mds];
411 static int __verify_registered_session(struct ceph_mds_client *mdsc,
412 struct ceph_mds_session *s)
414 if (s->s_mds >= mdsc->max_sessions ||
415 mdsc->sessions[s->s_mds] != s)
421 * create+register a new session for given mds.
422 * called under mdsc->mutex.
424 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
427 struct ceph_mds_session *s;
429 if (mds >= mdsc->mdsmap->m_max_mds)
430 return ERR_PTR(-EINVAL);
432 s = kzalloc(sizeof(*s), GFP_NOFS);
434 return ERR_PTR(-ENOMEM);
437 s->s_state = CEPH_MDS_SESSION_NEW;
440 mutex_init(&s->s_mutex);
442 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
444 spin_lock_init(&s->s_gen_ttl_lock);
446 s->s_cap_ttl = jiffies - 1;
448 spin_lock_init(&s->s_cap_lock);
449 s->s_renew_requested = 0;
451 INIT_LIST_HEAD(&s->s_caps);
454 atomic_set(&s->s_ref, 1);
455 INIT_LIST_HEAD(&s->s_waiting);
456 INIT_LIST_HEAD(&s->s_unsafe);
457 s->s_num_cap_releases = 0;
458 s->s_cap_reconnect = 0;
459 s->s_cap_iterator = NULL;
460 INIT_LIST_HEAD(&s->s_cap_releases);
461 INIT_LIST_HEAD(&s->s_cap_flushing);
462 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
464 dout("register_session mds%d\n", mds);
465 if (mds >= mdsc->max_sessions) {
466 int newmax = 1 << get_count_order(mds+1);
467 struct ceph_mds_session **sa;
469 dout("register_session realloc to %d\n", newmax);
470 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
473 if (mdsc->sessions) {
474 memcpy(sa, mdsc->sessions,
475 mdsc->max_sessions * sizeof(void *));
476 kfree(mdsc->sessions);
479 mdsc->max_sessions = newmax;
481 mdsc->sessions[mds] = s;
482 atomic_inc(&mdsc->num_sessions);
483 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
485 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
486 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
492 return ERR_PTR(-ENOMEM);
496 * called under mdsc->mutex
498 static void __unregister_session(struct ceph_mds_client *mdsc,
499 struct ceph_mds_session *s)
501 dout("__unregister_session mds%d %p\n", s->s_mds, s);
502 BUG_ON(mdsc->sessions[s->s_mds] != s);
503 mdsc->sessions[s->s_mds] = NULL;
504 ceph_con_close(&s->s_con);
505 ceph_put_mds_session(s);
506 atomic_dec(&mdsc->num_sessions);
510 * drop session refs in request.
512 * should be last request ref, or hold mdsc->mutex
514 static void put_request_session(struct ceph_mds_request *req)
516 if (req->r_session) {
517 ceph_put_mds_session(req->r_session);
518 req->r_session = NULL;
522 void ceph_mdsc_release_request(struct kref *kref)
524 struct ceph_mds_request *req = container_of(kref,
525 struct ceph_mds_request,
527 destroy_reply_info(&req->r_reply_info);
529 ceph_msg_put(req->r_request);
531 ceph_msg_put(req->r_reply);
533 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
536 if (req->r_locked_dir)
537 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
538 iput(req->r_target_inode);
541 if (req->r_old_dentry)
542 dput(req->r_old_dentry);
543 if (req->r_old_dentry_dir) {
545 * track (and drop pins for) r_old_dentry_dir
546 * separately, since r_old_dentry's d_parent may have
547 * changed between the dir mutex being dropped and
548 * this request being freed.
550 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
552 iput(req->r_old_dentry_dir);
557 ceph_pagelist_release(req->r_pagelist);
558 put_request_session(req);
559 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
564 * lookup session, bump ref if found.
566 * called under mdsc->mutex.
568 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
571 struct ceph_mds_request *req;
572 struct rb_node *n = mdsc->request_tree.rb_node;
575 req = rb_entry(n, struct ceph_mds_request, r_node);
576 if (tid < req->r_tid)
578 else if (tid > req->r_tid)
581 ceph_mdsc_get_request(req);
588 static void __insert_request(struct ceph_mds_client *mdsc,
589 struct ceph_mds_request *new)
591 struct rb_node **p = &mdsc->request_tree.rb_node;
592 struct rb_node *parent = NULL;
593 struct ceph_mds_request *req = NULL;
597 req = rb_entry(parent, struct ceph_mds_request, r_node);
598 if (new->r_tid < req->r_tid)
600 else if (new->r_tid > req->r_tid)
606 rb_link_node(&new->r_node, parent, p);
607 rb_insert_color(&new->r_node, &mdsc->request_tree);
611 * Register an in-flight request, and assign a tid. Link to directory
612 * are modifying (if any).
614 * Called under mdsc->mutex.
616 static void __register_request(struct ceph_mds_client *mdsc,
617 struct ceph_mds_request *req,
620 req->r_tid = ++mdsc->last_tid;
622 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
624 dout("__register_request %p tid %lld\n", req, req->r_tid);
625 ceph_mdsc_get_request(req);
626 __insert_request(mdsc, req);
628 req->r_uid = current_fsuid();
629 req->r_gid = current_fsgid();
631 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
632 mdsc->oldest_tid = req->r_tid;
635 struct ceph_inode_info *ci = ceph_inode(dir);
638 spin_lock(&ci->i_unsafe_lock);
639 req->r_unsafe_dir = dir;
640 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
641 spin_unlock(&ci->i_unsafe_lock);
645 static void __unregister_request(struct ceph_mds_client *mdsc,
646 struct ceph_mds_request *req)
648 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
650 if (req->r_tid == mdsc->oldest_tid) {
651 struct rb_node *p = rb_next(&req->r_node);
652 mdsc->oldest_tid = 0;
654 struct ceph_mds_request *next_req =
655 rb_entry(p, struct ceph_mds_request, r_node);
656 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
657 mdsc->oldest_tid = next_req->r_tid;
664 rb_erase(&req->r_node, &mdsc->request_tree);
665 RB_CLEAR_NODE(&req->r_node);
667 if (req->r_unsafe_dir) {
668 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
670 spin_lock(&ci->i_unsafe_lock);
671 list_del_init(&req->r_unsafe_dir_item);
672 spin_unlock(&ci->i_unsafe_lock);
674 iput(req->r_unsafe_dir);
675 req->r_unsafe_dir = NULL;
678 complete_all(&req->r_safe_completion);
680 ceph_mdsc_put_request(req);
684 * Choose mds to send request to next. If there is a hint set in the
685 * request (e.g., due to a prior forward hint from the mds), use that.
686 * Otherwise, consult frag tree and/or caps to identify the
687 * appropriate mds. If all else fails, choose randomly.
689 * Called under mdsc->mutex.
691 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
694 * we don't need to worry about protecting the d_parent access
695 * here because we never renaming inside the snapped namespace
696 * except to resplice to another snapdir, and either the old or new
697 * result is a valid result.
699 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
700 dentry = dentry->d_parent;
704 static int __choose_mds(struct ceph_mds_client *mdsc,
705 struct ceph_mds_request *req)
708 struct ceph_inode_info *ci;
709 struct ceph_cap *cap;
710 int mode = req->r_direct_mode;
712 u32 hash = req->r_direct_hash;
713 bool is_hash = req->r_direct_is_hash;
716 * is there a specific mds we should try? ignore hint if we have
717 * no session and the mds is not up (active or recovering).
719 if (req->r_resend_mds >= 0 &&
720 (__have_session(mdsc, req->r_resend_mds) ||
721 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
722 dout("choose_mds using resend_mds mds%d\n",
724 return req->r_resend_mds;
727 if (mode == USE_RANDOM_MDS)
732 inode = req->r_inode;
733 } else if (req->r_dentry) {
734 /* ignore race with rename; old or new d_parent is okay */
735 struct dentry *parent = req->r_dentry->d_parent;
736 struct inode *dir = d_inode(parent);
738 if (dir->i_sb != mdsc->fsc->sb) {
740 inode = d_inode(req->r_dentry);
741 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
742 /* direct snapped/virtual snapdir requests
743 * based on parent dir inode */
744 struct dentry *dn = get_nonsnap_parent(parent);
746 dout("__choose_mds using nonsnap parent %p\n", inode);
749 inode = d_inode(req->r_dentry);
750 if (!inode || mode == USE_AUTH_MDS) {
753 hash = ceph_dentry_hash(dir, req->r_dentry);
759 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
763 ci = ceph_inode(inode);
765 if (is_hash && S_ISDIR(inode->i_mode)) {
766 struct ceph_inode_frag frag;
769 ceph_choose_frag(ci, hash, &frag, &found);
771 if (mode == USE_ANY_MDS && frag.ndist > 0) {
774 /* choose a random replica */
775 get_random_bytes(&r, 1);
778 dout("choose_mds %p %llx.%llx "
779 "frag %u mds%d (%d/%d)\n",
780 inode, ceph_vinop(inode),
783 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
784 CEPH_MDS_STATE_ACTIVE)
788 /* since this file/dir wasn't known to be
789 * replicated, then we want to look for the
790 * authoritative mds. */
793 /* choose auth mds */
795 dout("choose_mds %p %llx.%llx "
796 "frag %u mds%d (auth)\n",
797 inode, ceph_vinop(inode), frag.frag, mds);
798 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
799 CEPH_MDS_STATE_ACTIVE)
805 spin_lock(&ci->i_ceph_lock);
807 if (mode == USE_AUTH_MDS)
808 cap = ci->i_auth_cap;
809 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
810 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
812 spin_unlock(&ci->i_ceph_lock);
815 mds = cap->session->s_mds;
816 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
817 inode, ceph_vinop(inode), mds,
818 cap == ci->i_auth_cap ? "auth " : "", cap);
819 spin_unlock(&ci->i_ceph_lock);
823 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
824 dout("choose_mds chose random mds%d\n", mds);
832 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
834 struct ceph_msg *msg;
835 struct ceph_mds_session_head *h;
837 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
840 pr_err("create_session_msg ENOMEM creating msg\n");
843 h = msg->front.iov_base;
844 h->op = cpu_to_le32(op);
845 h->seq = cpu_to_le64(seq);
851 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
852 * to include additional client metadata fields.
854 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
856 struct ceph_msg *msg;
857 struct ceph_mds_session_head *h;
859 int metadata_bytes = 0;
860 int metadata_key_count = 0;
861 struct ceph_options *opt = mdsc->fsc->client->options;
864 const char* metadata[][2] = {
865 {"hostname", utsname()->nodename},
866 {"kernel_version", utsname()->release},
867 {"entity_id", opt->name ? opt->name : ""},
871 /* Calculate serialized length of metadata */
872 metadata_bytes = 4; /* map length */
873 for (i = 0; metadata[i][0] != NULL; ++i) {
874 metadata_bytes += 8 + strlen(metadata[i][0]) +
875 strlen(metadata[i][1]);
876 metadata_key_count++;
879 /* Allocate the message */
880 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
883 pr_err("create_session_msg ENOMEM creating msg\n");
886 h = msg->front.iov_base;
887 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
888 h->seq = cpu_to_le64(seq);
891 * Serialize client metadata into waiting buffer space, using
892 * the format that userspace expects for map<string, string>
894 * ClientSession messages with metadata are v2
896 msg->hdr.version = cpu_to_le16(2);
897 msg->hdr.compat_version = cpu_to_le16(1);
899 /* The write pointer, following the session_head structure */
900 p = msg->front.iov_base + sizeof(*h);
902 /* Number of entries in the map */
903 ceph_encode_32(&p, metadata_key_count);
905 /* Two length-prefixed strings for each entry in the map */
906 for (i = 0; metadata[i][0] != NULL; ++i) {
907 size_t const key_len = strlen(metadata[i][0]);
908 size_t const val_len = strlen(metadata[i][1]);
910 ceph_encode_32(&p, key_len);
911 memcpy(p, metadata[i][0], key_len);
913 ceph_encode_32(&p, val_len);
914 memcpy(p, metadata[i][1], val_len);
922 * send session open request.
924 * called under mdsc->mutex
926 static int __open_session(struct ceph_mds_client *mdsc,
927 struct ceph_mds_session *session)
929 struct ceph_msg *msg;
931 int mds = session->s_mds;
933 /* wait for mds to go active? */
934 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
935 dout("open_session to mds%d (%s)\n", mds,
936 ceph_mds_state_name(mstate));
937 session->s_state = CEPH_MDS_SESSION_OPENING;
938 session->s_renew_requested = jiffies;
940 /* send connect message */
941 msg = create_session_open_msg(mdsc, session->s_seq);
944 ceph_con_send(&session->s_con, msg);
949 * open sessions for any export targets for the given mds
951 * called under mdsc->mutex
953 static struct ceph_mds_session *
954 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
956 struct ceph_mds_session *session;
958 session = __ceph_lookup_mds_session(mdsc, target);
960 session = register_session(mdsc, target);
964 if (session->s_state == CEPH_MDS_SESSION_NEW ||
965 session->s_state == CEPH_MDS_SESSION_CLOSING)
966 __open_session(mdsc, session);
971 struct ceph_mds_session *
972 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
974 struct ceph_mds_session *session;
976 dout("open_export_target_session to mds%d\n", target);
978 mutex_lock(&mdsc->mutex);
979 session = __open_export_target_session(mdsc, target);
980 mutex_unlock(&mdsc->mutex);
985 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
986 struct ceph_mds_session *session)
988 struct ceph_mds_info *mi;
989 struct ceph_mds_session *ts;
990 int i, mds = session->s_mds;
992 if (mds >= mdsc->mdsmap->m_max_mds)
995 mi = &mdsc->mdsmap->m_info[mds];
996 dout("open_export_target_sessions for mds%d (%d targets)\n",
997 session->s_mds, mi->num_export_targets);
999 for (i = 0; i < mi->num_export_targets; i++) {
1000 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1002 ceph_put_mds_session(ts);
1006 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1007 struct ceph_mds_session *session)
1009 mutex_lock(&mdsc->mutex);
1010 __open_export_target_sessions(mdsc, session);
1011 mutex_unlock(&mdsc->mutex);
1018 /* caller holds s_cap_lock, we drop it */
1019 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1020 struct ceph_mds_session *session)
1021 __releases(session->s_cap_lock)
1023 LIST_HEAD(tmp_list);
1024 list_splice_init(&session->s_cap_releases, &tmp_list);
1025 session->s_num_cap_releases = 0;
1026 spin_unlock(&session->s_cap_lock);
1028 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1029 while (!list_empty(&tmp_list)) {
1030 struct ceph_cap *cap;
1031 /* zero out the in-progress message */
1032 cap = list_first_entry(&tmp_list,
1033 struct ceph_cap, session_caps);
1034 list_del(&cap->session_caps);
1035 ceph_put_cap(mdsc, cap);
1039 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1040 struct ceph_mds_session *session)
1042 struct ceph_mds_request *req;
1045 dout("cleanup_session_requests mds%d\n", session->s_mds);
1046 mutex_lock(&mdsc->mutex);
1047 while (!list_empty(&session->s_unsafe)) {
1048 req = list_first_entry(&session->s_unsafe,
1049 struct ceph_mds_request, r_unsafe_item);
1050 list_del_init(&req->r_unsafe_item);
1051 pr_info(" dropping unsafe request %llu\n", req->r_tid);
1052 __unregister_request(mdsc, req);
1054 /* zero r_attempts, so kick_requests() will re-send requests */
1055 p = rb_first(&mdsc->request_tree);
1057 req = rb_entry(p, struct ceph_mds_request, r_node);
1059 if (req->r_session &&
1060 req->r_session->s_mds == session->s_mds)
1061 req->r_attempts = 0;
1063 mutex_unlock(&mdsc->mutex);
1067 * Helper to safely iterate over all caps associated with a session, with
1068 * special care taken to handle a racing __ceph_remove_cap().
1070 * Caller must hold session s_mutex.
1072 static int iterate_session_caps(struct ceph_mds_session *session,
1073 int (*cb)(struct inode *, struct ceph_cap *,
1076 struct list_head *p;
1077 struct ceph_cap *cap;
1078 struct inode *inode, *last_inode = NULL;
1079 struct ceph_cap *old_cap = NULL;
1082 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1083 spin_lock(&session->s_cap_lock);
1084 p = session->s_caps.next;
1085 while (p != &session->s_caps) {
1086 cap = list_entry(p, struct ceph_cap, session_caps);
1087 inode = igrab(&cap->ci->vfs_inode);
1092 session->s_cap_iterator = cap;
1093 spin_unlock(&session->s_cap_lock);
1100 ceph_put_cap(session->s_mdsc, old_cap);
1104 ret = cb(inode, cap, arg);
1107 spin_lock(&session->s_cap_lock);
1109 if (cap->ci == NULL) {
1110 dout("iterate_session_caps finishing cap %p removal\n",
1112 BUG_ON(cap->session != session);
1113 cap->session = NULL;
1114 list_del_init(&cap->session_caps);
1115 session->s_nr_caps--;
1116 if (cap->queue_release) {
1117 list_add_tail(&cap->session_caps,
1118 &session->s_cap_releases);
1119 session->s_num_cap_releases++;
1121 old_cap = cap; /* put_cap it w/o locks held */
1129 session->s_cap_iterator = NULL;
1130 spin_unlock(&session->s_cap_lock);
1134 ceph_put_cap(session->s_mdsc, old_cap);
1139 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1142 struct ceph_inode_info *ci = ceph_inode(inode);
1145 dout("removing cap %p, ci is %p, inode is %p\n",
1146 cap, ci, &ci->vfs_inode);
1147 spin_lock(&ci->i_ceph_lock);
1148 __ceph_remove_cap(cap, false);
1149 if (!ci->i_auth_cap) {
1150 struct ceph_mds_client *mdsc =
1151 ceph_sb_to_client(inode->i_sb)->mdsc;
1153 spin_lock(&mdsc->cap_dirty_lock);
1154 if (!list_empty(&ci->i_dirty_item)) {
1155 pr_info(" dropping dirty %s state for %p %lld\n",
1156 ceph_cap_string(ci->i_dirty_caps),
1157 inode, ceph_ino(inode));
1158 ci->i_dirty_caps = 0;
1159 list_del_init(&ci->i_dirty_item);
1162 if (!list_empty(&ci->i_flushing_item)) {
1163 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1164 ceph_cap_string(ci->i_flushing_caps),
1165 inode, ceph_ino(inode));
1166 ci->i_flushing_caps = 0;
1167 list_del_init(&ci->i_flushing_item);
1168 mdsc->num_cap_flushing--;
1171 spin_unlock(&mdsc->cap_dirty_lock);
1173 spin_unlock(&ci->i_ceph_lock);
1180 * caller must hold session s_mutex
1182 static void remove_session_caps(struct ceph_mds_session *session)
1184 dout("remove_session_caps on %p\n", session);
1185 iterate_session_caps(session, remove_session_caps_cb, NULL);
1187 spin_lock(&session->s_cap_lock);
1188 if (session->s_nr_caps > 0) {
1189 struct super_block *sb = session->s_mdsc->fsc->sb;
1190 struct inode *inode;
1191 struct ceph_cap *cap, *prev = NULL;
1192 struct ceph_vino vino;
1194 * iterate_session_caps() skips inodes that are being
1195 * deleted, we need to wait until deletions are complete.
1196 * __wait_on_freeing_inode() is designed for the job,
1197 * but it is not exported, so use lookup inode function
1200 while (!list_empty(&session->s_caps)) {
1201 cap = list_entry(session->s_caps.next,
1202 struct ceph_cap, session_caps);
1206 vino = cap->ci->i_vino;
1207 spin_unlock(&session->s_cap_lock);
1209 inode = ceph_find_inode(sb, vino);
1212 spin_lock(&session->s_cap_lock);
1216 // drop cap expires and unlock s_cap_lock
1217 cleanup_cap_releases(session->s_mdsc, session);
1219 BUG_ON(session->s_nr_caps > 0);
1220 BUG_ON(!list_empty(&session->s_cap_flushing));
1224 * wake up any threads waiting on this session's caps. if the cap is
1225 * old (didn't get renewed on the client reconnect), remove it now.
1227 * caller must hold s_mutex.
1229 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1232 struct ceph_inode_info *ci = ceph_inode(inode);
1234 wake_up_all(&ci->i_cap_wq);
1236 spin_lock(&ci->i_ceph_lock);
1237 ci->i_wanted_max_size = 0;
1238 ci->i_requested_max_size = 0;
1239 spin_unlock(&ci->i_ceph_lock);
1244 static void wake_up_session_caps(struct ceph_mds_session *session,
1247 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1248 iterate_session_caps(session, wake_up_session_cb,
1249 (void *)(unsigned long)reconnect);
1253 * Send periodic message to MDS renewing all currently held caps. The
1254 * ack will reset the expiration for all caps from this session.
1256 * caller holds s_mutex
1258 static int send_renew_caps(struct ceph_mds_client *mdsc,
1259 struct ceph_mds_session *session)
1261 struct ceph_msg *msg;
1264 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1265 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1266 pr_info("mds%d caps stale\n", session->s_mds);
1267 session->s_renew_requested = jiffies;
1269 /* do not try to renew caps until a recovering mds has reconnected
1270 * with its clients. */
1271 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1272 if (state < CEPH_MDS_STATE_RECONNECT) {
1273 dout("send_renew_caps ignoring mds%d (%s)\n",
1274 session->s_mds, ceph_mds_state_name(state));
1278 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1279 ceph_mds_state_name(state));
1280 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1281 ++session->s_renew_seq);
1284 ceph_con_send(&session->s_con, msg);
1288 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1289 struct ceph_mds_session *session, u64 seq)
1291 struct ceph_msg *msg;
1293 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1294 session->s_mds, ceph_session_state_name(session->s_state), seq);
1295 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1298 ceph_con_send(&session->s_con, msg);
1304 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1306 * Called under session->s_mutex
1308 static void renewed_caps(struct ceph_mds_client *mdsc,
1309 struct ceph_mds_session *session, int is_renew)
1314 spin_lock(&session->s_cap_lock);
1315 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1317 session->s_cap_ttl = session->s_renew_requested +
1318 mdsc->mdsmap->m_session_timeout*HZ;
1321 if (time_before(jiffies, session->s_cap_ttl)) {
1322 pr_info("mds%d caps renewed\n", session->s_mds);
1325 pr_info("mds%d caps still stale\n", session->s_mds);
1328 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1329 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1330 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1331 spin_unlock(&session->s_cap_lock);
1334 wake_up_session_caps(session, 0);
1338 * send a session close request
1340 static int request_close_session(struct ceph_mds_client *mdsc,
1341 struct ceph_mds_session *session)
1343 struct ceph_msg *msg;
1345 dout("request_close_session mds%d state %s seq %lld\n",
1346 session->s_mds, ceph_session_state_name(session->s_state),
1348 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1351 ceph_con_send(&session->s_con, msg);
1356 * Called with s_mutex held.
1358 static int __close_session(struct ceph_mds_client *mdsc,
1359 struct ceph_mds_session *session)
1361 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1363 session->s_state = CEPH_MDS_SESSION_CLOSING;
1364 return request_close_session(mdsc, session);
1368 * Trim old(er) caps.
1370 * Because we can't cache an inode without one or more caps, we do
1371 * this indirectly: if a cap is unused, we prune its aliases, at which
1372 * point the inode will hopefully get dropped to.
1374 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1375 * memory pressure from the MDS, though, so it needn't be perfect.
1377 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1379 struct ceph_mds_session *session = arg;
1380 struct ceph_inode_info *ci = ceph_inode(inode);
1381 int used, wanted, oissued, mine;
1383 if (session->s_trim_caps <= 0)
1386 spin_lock(&ci->i_ceph_lock);
1387 mine = cap->issued | cap->implemented;
1388 used = __ceph_caps_used(ci);
1389 wanted = __ceph_caps_file_wanted(ci);
1390 oissued = __ceph_caps_issued_other(ci, cap);
1392 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1393 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1394 ceph_cap_string(used), ceph_cap_string(wanted));
1395 if (cap == ci->i_auth_cap) {
1396 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1397 !list_empty(&ci->i_cap_snaps))
1399 if ((used | wanted) & CEPH_CAP_ANY_WR)
1402 if ((used | wanted) & ~oissued & mine)
1403 goto out; /* we need these caps */
1405 session->s_trim_caps--;
1407 /* we aren't the only cap.. just remove us */
1408 __ceph_remove_cap(cap, true);
1410 /* try to drop referring dentries */
1411 spin_unlock(&ci->i_ceph_lock);
1412 d_prune_aliases(inode);
1413 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1414 inode, cap, atomic_read(&inode->i_count));
1419 spin_unlock(&ci->i_ceph_lock);
1424 * Trim session cap count down to some max number.
1426 static int trim_caps(struct ceph_mds_client *mdsc,
1427 struct ceph_mds_session *session,
1430 int trim_caps = session->s_nr_caps - max_caps;
1432 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1433 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1434 if (trim_caps > 0) {
1435 session->s_trim_caps = trim_caps;
1436 iterate_session_caps(session, trim_caps_cb, session);
1437 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1438 session->s_mds, session->s_nr_caps, max_caps,
1439 trim_caps - session->s_trim_caps);
1440 session->s_trim_caps = 0;
1443 ceph_send_cap_releases(mdsc, session);
1447 static int check_cap_flush(struct ceph_inode_info *ci,
1448 u64 want_flush_seq, u64 want_snap_seq)
1450 int ret1 = 1, ret2 = 1;
1451 spin_lock(&ci->i_ceph_lock);
1452 if (want_flush_seq > 0 && ci->i_flushing_caps)
1453 ret1 = ci->i_cap_flush_seq >= want_flush_seq;
1455 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1456 struct ceph_cap_snap *capsnap =
1457 list_first_entry(&ci->i_cap_snaps,
1458 struct ceph_cap_snap, ci_item);
1459 ret2 = capsnap->follows >= want_snap_seq;
1461 spin_unlock(&ci->i_ceph_lock);
1462 return ret1 && ret2;
1466 * flush all dirty inode data to disk.
1468 * returns true if we've flushed through want_flush_seq
1470 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1471 u64 want_flush_seq, u64 want_snap_seq)
1475 dout("check_cap_flush want %lld\n", want_flush_seq);
1476 mutex_lock(&mdsc->mutex);
1477 for (mds = 0; mds < mdsc->max_sessions; ) {
1478 struct ceph_mds_session *session = mdsc->sessions[mds];
1479 struct inode *inode1 = NULL, *inode2 = NULL;
1485 get_session(session);
1486 mutex_unlock(&mdsc->mutex);
1488 mutex_lock(&session->s_mutex);
1489 if (!list_empty(&session->s_cap_flushing)) {
1490 struct ceph_inode_info *ci =
1491 list_first_entry(&session->s_cap_flushing,
1492 struct ceph_inode_info,
1495 if (!check_cap_flush(ci, want_flush_seq, 0)) {
1496 dout("check_cap_flush still flushing %p "
1497 "seq %lld <= %lld to mds%d\n",
1498 &ci->vfs_inode, ci->i_cap_flush_seq,
1499 want_flush_seq, mds);
1500 inode1 = igrab(&ci->vfs_inode);
1503 if (!list_empty(&session->s_cap_snaps_flushing)) {
1504 struct ceph_cap_snap *capsnap =
1505 list_first_entry(&session->s_cap_snaps_flushing,
1506 struct ceph_cap_snap,
1508 struct ceph_inode_info *ci = capsnap->ci;
1509 if (!check_cap_flush(ci, 0, want_snap_seq)) {
1510 dout("check_cap_flush still flushing snap %p "
1511 "follows %lld <= %lld to mds%d\n",
1512 &ci->vfs_inode, capsnap->follows,
1513 want_snap_seq, mds);
1514 inode2 = igrab(&ci->vfs_inode);
1517 mutex_unlock(&session->s_mutex);
1518 ceph_put_mds_session(session);
1521 wait_event(mdsc->cap_flushing_wq,
1522 check_cap_flush(ceph_inode(inode1),
1523 want_flush_seq, 0));
1527 wait_event(mdsc->cap_flushing_wq,
1528 check_cap_flush(ceph_inode(inode2),
1533 if (!inode1 && !inode2)
1536 mutex_lock(&mdsc->mutex);
1539 mutex_unlock(&mdsc->mutex);
1540 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1544 * called under s_mutex
1546 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1547 struct ceph_mds_session *session)
1549 struct ceph_msg *msg = NULL;
1550 struct ceph_mds_cap_release *head;
1551 struct ceph_mds_cap_item *item;
1552 struct ceph_cap *cap;
1553 LIST_HEAD(tmp_list);
1554 int num_cap_releases;
1556 spin_lock(&session->s_cap_lock);
1558 list_splice_init(&session->s_cap_releases, &tmp_list);
1559 num_cap_releases = session->s_num_cap_releases;
1560 session->s_num_cap_releases = 0;
1561 spin_unlock(&session->s_cap_lock);
1563 while (!list_empty(&tmp_list)) {
1565 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1566 PAGE_CACHE_SIZE, GFP_NOFS, false);
1569 head = msg->front.iov_base;
1570 head->num = cpu_to_le32(0);
1571 msg->front.iov_len = sizeof(*head);
1573 cap = list_first_entry(&tmp_list, struct ceph_cap,
1575 list_del(&cap->session_caps);
1578 head = msg->front.iov_base;
1579 le32_add_cpu(&head->num, 1);
1580 item = msg->front.iov_base + msg->front.iov_len;
1581 item->ino = cpu_to_le64(cap->cap_ino);
1582 item->cap_id = cpu_to_le64(cap->cap_id);
1583 item->migrate_seq = cpu_to_le32(cap->mseq);
1584 item->seq = cpu_to_le32(cap->issue_seq);
1585 msg->front.iov_len += sizeof(*item);
1587 ceph_put_cap(mdsc, cap);
1589 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1590 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1591 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1592 ceph_con_send(&session->s_con, msg);
1597 BUG_ON(num_cap_releases != 0);
1599 spin_lock(&session->s_cap_lock);
1600 if (!list_empty(&session->s_cap_releases))
1602 spin_unlock(&session->s_cap_lock);
1605 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1606 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1607 ceph_con_send(&session->s_con, msg);
1611 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1613 spin_lock(&session->s_cap_lock);
1614 list_splice(&tmp_list, &session->s_cap_releases);
1615 session->s_num_cap_releases += num_cap_releases;
1616 spin_unlock(&session->s_cap_lock);
1623 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1626 struct ceph_inode_info *ci = ceph_inode(dir);
1627 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1628 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1629 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1630 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1631 int order, num_entries;
1633 spin_lock(&ci->i_ceph_lock);
1634 num_entries = ci->i_files + ci->i_subdirs;
1635 spin_unlock(&ci->i_ceph_lock);
1636 num_entries = max(num_entries, 1);
1637 num_entries = min(num_entries, opt->max_readdir);
1639 order = get_order(size * num_entries);
1640 while (order >= 0) {
1641 rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN,
1650 num_entries = (PAGE_SIZE << order) / size;
1651 num_entries = min(num_entries, opt->max_readdir);
1653 rinfo->dir_buf_size = PAGE_SIZE << order;
1654 req->r_num_caps = num_entries + 1;
1655 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1656 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1661 * Create an mds request.
1663 struct ceph_mds_request *
1664 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1666 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1669 return ERR_PTR(-ENOMEM);
1671 mutex_init(&req->r_fill_mutex);
1673 req->r_started = jiffies;
1674 req->r_resend_mds = -1;
1675 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1677 kref_init(&req->r_kref);
1678 INIT_LIST_HEAD(&req->r_wait);
1679 init_completion(&req->r_completion);
1680 init_completion(&req->r_safe_completion);
1681 INIT_LIST_HEAD(&req->r_unsafe_item);
1683 req->r_stamp = CURRENT_TIME;
1686 req->r_direct_mode = mode;
1691 * return oldest (lowest) request, tid in request tree, 0 if none.
1693 * called under mdsc->mutex.
1695 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1697 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1699 return rb_entry(rb_first(&mdsc->request_tree),
1700 struct ceph_mds_request, r_node);
1703 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1705 return mdsc->oldest_tid;
1709 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1710 * on build_path_from_dentry in fs/cifs/dir.c.
1712 * If @stop_on_nosnap, generate path relative to the first non-snapped
1715 * Encode hidden .snap dirs as a double /, i.e.
1716 * foo/.snap/bar -> foo//bar
1718 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1721 struct dentry *temp;
1727 return ERR_PTR(-EINVAL);
1731 seq = read_seqbegin(&rename_lock);
1733 for (temp = dentry; !IS_ROOT(temp);) {
1734 struct inode *inode = d_inode(temp);
1735 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1736 len++; /* slash only */
1737 else if (stop_on_nosnap && inode &&
1738 ceph_snap(inode) == CEPH_NOSNAP)
1741 len += 1 + temp->d_name.len;
1742 temp = temp->d_parent;
1746 len--; /* no leading '/' */
1748 path = kmalloc(len+1, GFP_NOFS);
1750 return ERR_PTR(-ENOMEM);
1752 path[pos] = 0; /* trailing null */
1754 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1755 struct inode *inode;
1757 spin_lock(&temp->d_lock);
1758 inode = d_inode(temp);
1759 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1760 dout("build_path path+%d: %p SNAPDIR\n",
1762 } else if (stop_on_nosnap && inode &&
1763 ceph_snap(inode) == CEPH_NOSNAP) {
1764 spin_unlock(&temp->d_lock);
1767 pos -= temp->d_name.len;
1769 spin_unlock(&temp->d_lock);
1772 strncpy(path + pos, temp->d_name.name,
1775 spin_unlock(&temp->d_lock);
1778 temp = temp->d_parent;
1781 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1782 pr_err("build_path did not end path lookup where "
1783 "expected, namelen is %d, pos is %d\n", len, pos);
1784 /* presumably this is only possible if racing with a
1785 rename of one of the parent directories (we can not
1786 lock the dentries above us to prevent this, but
1787 retrying should be harmless) */
1792 *base = ceph_ino(d_inode(temp));
1794 dout("build_path on %p %d built %llx '%.*s'\n",
1795 dentry, d_count(dentry), *base, len, path);
1799 static int build_dentry_path(struct dentry *dentry,
1800 const char **ppath, int *ppathlen, u64 *pino,
1805 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1806 *pino = ceph_ino(d_inode(dentry->d_parent));
1807 *ppath = dentry->d_name.name;
1808 *ppathlen = dentry->d_name.len;
1811 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1813 return PTR_ERR(path);
1819 static int build_inode_path(struct inode *inode,
1820 const char **ppath, int *ppathlen, u64 *pino,
1823 struct dentry *dentry;
1826 if (ceph_snap(inode) == CEPH_NOSNAP) {
1827 *pino = ceph_ino(inode);
1831 dentry = d_find_alias(inode);
1832 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1835 return PTR_ERR(path);
1842 * request arguments may be specified via an inode *, a dentry *, or
1843 * an explicit ino+path.
1845 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1846 const char *rpath, u64 rino,
1847 const char **ppath, int *pathlen,
1848 u64 *ino, int *freepath)
1853 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1854 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1856 } else if (rdentry) {
1857 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1858 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1860 } else if (rpath || rino) {
1863 *pathlen = rpath ? strlen(rpath) : 0;
1864 dout(" path %.*s\n", *pathlen, rpath);
1871 * called under mdsc->mutex
1873 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1874 struct ceph_mds_request *req,
1875 int mds, bool drop_cap_releases)
1877 struct ceph_msg *msg;
1878 struct ceph_mds_request_head *head;
1879 const char *path1 = NULL;
1880 const char *path2 = NULL;
1881 u64 ino1 = 0, ino2 = 0;
1882 int pathlen1 = 0, pathlen2 = 0;
1883 int freepath1 = 0, freepath2 = 0;
1889 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1890 req->r_path1, req->r_ino1.ino,
1891 &path1, &pathlen1, &ino1, &freepath1);
1897 ret = set_request_path_attr(NULL, req->r_old_dentry,
1898 req->r_path2, req->r_ino2.ino,
1899 &path2, &pathlen2, &ino2, &freepath2);
1905 len = sizeof(*head) +
1906 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1907 sizeof(struct timespec);
1909 /* calculate (max) length for cap releases */
1910 len += sizeof(struct ceph_mds_request_release) *
1911 (!!req->r_inode_drop + !!req->r_dentry_drop +
1912 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1913 if (req->r_dentry_drop)
1914 len += req->r_dentry->d_name.len;
1915 if (req->r_old_dentry_drop)
1916 len += req->r_old_dentry->d_name.len;
1918 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1920 msg = ERR_PTR(-ENOMEM);
1924 msg->hdr.version = cpu_to_le16(2);
1925 msg->hdr.tid = cpu_to_le64(req->r_tid);
1927 head = msg->front.iov_base;
1928 p = msg->front.iov_base + sizeof(*head);
1929 end = msg->front.iov_base + msg->front.iov_len;
1931 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1932 head->op = cpu_to_le32(req->r_op);
1933 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1934 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1935 head->args = req->r_args;
1937 ceph_encode_filepath(&p, end, ino1, path1);
1938 ceph_encode_filepath(&p, end, ino2, path2);
1940 /* make note of release offset, in case we need to replay */
1941 req->r_request_release_offset = p - msg->front.iov_base;
1945 if (req->r_inode_drop)
1946 releases += ceph_encode_inode_release(&p,
1947 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1948 mds, req->r_inode_drop, req->r_inode_unless, 0);
1949 if (req->r_dentry_drop)
1950 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1951 mds, req->r_dentry_drop, req->r_dentry_unless);
1952 if (req->r_old_dentry_drop)
1953 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1954 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1955 if (req->r_old_inode_drop)
1956 releases += ceph_encode_inode_release(&p,
1957 d_inode(req->r_old_dentry),
1958 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1960 if (drop_cap_releases) {
1962 p = msg->front.iov_base + req->r_request_release_offset;
1965 head->num_releases = cpu_to_le16(releases);
1969 struct ceph_timespec ts;
1970 ceph_encode_timespec(&ts, &req->r_stamp);
1971 ceph_encode_copy(&p, &ts, sizeof(ts));
1975 msg->front.iov_len = p - msg->front.iov_base;
1976 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1978 if (req->r_pagelist) {
1979 struct ceph_pagelist *pagelist = req->r_pagelist;
1980 atomic_inc(&pagelist->refcnt);
1981 ceph_msg_data_add_pagelist(msg, pagelist);
1982 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1984 msg->hdr.data_len = 0;
1987 msg->hdr.data_off = cpu_to_le16(0);
1991 kfree((char *)path2);
1994 kfree((char *)path1);
2000 * called under mdsc->mutex if error, under no mutex if
2003 static void complete_request(struct ceph_mds_client *mdsc,
2004 struct ceph_mds_request *req)
2006 if (req->r_callback)
2007 req->r_callback(mdsc, req);
2009 complete_all(&req->r_completion);
2013 * called under mdsc->mutex
2015 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2016 struct ceph_mds_request *req,
2017 int mds, bool drop_cap_releases)
2019 struct ceph_mds_request_head *rhead;
2020 struct ceph_msg *msg;
2025 struct ceph_cap *cap =
2026 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2029 req->r_sent_on_mseq = cap->mseq;
2031 req->r_sent_on_mseq = -1;
2033 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2034 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2036 if (req->r_got_unsafe) {
2039 * Replay. Do not regenerate message (and rebuild
2040 * paths, etc.); just use the original message.
2041 * Rebuilding paths will break for renames because
2042 * d_move mangles the src name.
2044 msg = req->r_request;
2045 rhead = msg->front.iov_base;
2047 flags = le32_to_cpu(rhead->flags);
2048 flags |= CEPH_MDS_FLAG_REPLAY;
2049 rhead->flags = cpu_to_le32(flags);
2051 if (req->r_target_inode)
2052 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2054 rhead->num_retry = req->r_attempts - 1;
2056 /* remove cap/dentry releases from message */
2057 rhead->num_releases = 0;
2060 p = msg->front.iov_base + req->r_request_release_offset;
2062 struct ceph_timespec ts;
2063 ceph_encode_timespec(&ts, &req->r_stamp);
2064 ceph_encode_copy(&p, &ts, sizeof(ts));
2067 msg->front.iov_len = p - msg->front.iov_base;
2068 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2072 if (req->r_request) {
2073 ceph_msg_put(req->r_request);
2074 req->r_request = NULL;
2076 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2078 req->r_err = PTR_ERR(msg);
2079 complete_request(mdsc, req);
2080 return PTR_ERR(msg);
2082 req->r_request = msg;
2084 rhead = msg->front.iov_base;
2085 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2086 if (req->r_got_unsafe)
2087 flags |= CEPH_MDS_FLAG_REPLAY;
2088 if (req->r_locked_dir)
2089 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2090 rhead->flags = cpu_to_le32(flags);
2091 rhead->num_fwd = req->r_num_fwd;
2092 rhead->num_retry = req->r_attempts - 1;
2095 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2100 * send request, or put it on the appropriate wait list.
2102 static int __do_request(struct ceph_mds_client *mdsc,
2103 struct ceph_mds_request *req)
2105 struct ceph_mds_session *session = NULL;
2109 if (req->r_err || req->r_got_result) {
2111 __unregister_request(mdsc, req);
2115 if (req->r_timeout &&
2116 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2117 dout("do_request timed out\n");
2122 put_request_session(req);
2124 mds = __choose_mds(mdsc, req);
2126 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2127 dout("do_request no mds or not active, waiting for map\n");
2128 list_add(&req->r_wait, &mdsc->waiting_for_map);
2132 /* get, open session */
2133 session = __ceph_lookup_mds_session(mdsc, mds);
2135 session = register_session(mdsc, mds);
2136 if (IS_ERR(session)) {
2137 err = PTR_ERR(session);
2141 req->r_session = get_session(session);
2143 dout("do_request mds%d session %p state %s\n", mds, session,
2144 ceph_session_state_name(session->s_state));
2145 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2146 session->s_state != CEPH_MDS_SESSION_HUNG) {
2147 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2148 session->s_state == CEPH_MDS_SESSION_CLOSING)
2149 __open_session(mdsc, session);
2150 list_add(&req->r_wait, &session->s_waiting);
2155 req->r_resend_mds = -1; /* forget any previous mds hint */
2157 if (req->r_request_started == 0) /* note request start time */
2158 req->r_request_started = jiffies;
2160 err = __prepare_send_request(mdsc, req, mds, false);
2162 ceph_msg_get(req->r_request);
2163 ceph_con_send(&session->s_con, req->r_request);
2167 ceph_put_mds_session(session);
2173 complete_request(mdsc, req);
2178 * called under mdsc->mutex
2180 static void __wake_requests(struct ceph_mds_client *mdsc,
2181 struct list_head *head)
2183 struct ceph_mds_request *req;
2184 LIST_HEAD(tmp_list);
2186 list_splice_init(head, &tmp_list);
2188 while (!list_empty(&tmp_list)) {
2189 req = list_entry(tmp_list.next,
2190 struct ceph_mds_request, r_wait);
2191 list_del_init(&req->r_wait);
2192 dout(" wake request %p tid %llu\n", req, req->r_tid);
2193 __do_request(mdsc, req);
2198 * Wake up threads with requests pending for @mds, so that they can
2199 * resubmit their requests to a possibly different mds.
2201 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2203 struct ceph_mds_request *req;
2204 struct rb_node *p = rb_first(&mdsc->request_tree);
2206 dout("kick_requests mds%d\n", mds);
2208 req = rb_entry(p, struct ceph_mds_request, r_node);
2210 if (req->r_got_unsafe)
2212 if (req->r_attempts > 0)
2213 continue; /* only new requests */
2214 if (req->r_session &&
2215 req->r_session->s_mds == mds) {
2216 dout(" kicking tid %llu\n", req->r_tid);
2217 list_del_init(&req->r_wait);
2218 __do_request(mdsc, req);
2223 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2224 struct ceph_mds_request *req)
2226 dout("submit_request on %p\n", req);
2227 mutex_lock(&mdsc->mutex);
2228 __register_request(mdsc, req, NULL);
2229 __do_request(mdsc, req);
2230 mutex_unlock(&mdsc->mutex);
2234 * Synchrously perform an mds request. Take care of all of the
2235 * session setup, forwarding, retry details.
2237 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2239 struct ceph_mds_request *req)
2243 dout("do_request on %p\n", req);
2245 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2247 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2248 if (req->r_locked_dir)
2249 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2250 if (req->r_old_dentry_dir)
2251 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2255 mutex_lock(&mdsc->mutex);
2256 __register_request(mdsc, req, dir);
2257 __do_request(mdsc, req);
2261 __unregister_request(mdsc, req);
2262 dout("do_request early error %d\n", err);
2267 mutex_unlock(&mdsc->mutex);
2268 dout("do_request waiting\n");
2269 if (!req->r_timeout && req->r_wait_for_completion) {
2270 err = req->r_wait_for_completion(mdsc, req);
2272 long timeleft = wait_for_completion_killable_timeout(
2274 ceph_timeout_jiffies(req->r_timeout));
2278 err = -EIO; /* timed out */
2280 err = timeleft; /* killed */
2282 dout("do_request waited, got %d\n", err);
2283 mutex_lock(&mdsc->mutex);
2285 /* only abort if we didn't race with a real reply */
2286 if (req->r_got_result) {
2287 err = le32_to_cpu(req->r_reply_info.head->result);
2288 } else if (err < 0) {
2289 dout("aborted request %lld with %d\n", req->r_tid, err);
2292 * ensure we aren't running concurrently with
2293 * ceph_fill_trace or ceph_readdir_prepopulate, which
2294 * rely on locks (dir mutex) held by our caller.
2296 mutex_lock(&req->r_fill_mutex);
2298 req->r_aborted = true;
2299 mutex_unlock(&req->r_fill_mutex);
2301 if (req->r_locked_dir &&
2302 (req->r_op & CEPH_MDS_OP_WRITE))
2303 ceph_invalidate_dir_request(req);
2309 mutex_unlock(&mdsc->mutex);
2310 dout("do_request %p done, result %d\n", req, err);
2315 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2316 * namespace request.
2318 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2320 struct inode *inode = req->r_locked_dir;
2322 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2324 ceph_dir_clear_complete(inode);
2326 ceph_invalidate_dentry_lease(req->r_dentry);
2327 if (req->r_old_dentry)
2328 ceph_invalidate_dentry_lease(req->r_old_dentry);
2334 * We take the session mutex and parse and process the reply immediately.
2335 * This preserves the logical ordering of replies, capabilities, etc., sent
2336 * by the MDS as they are applied to our local cache.
2338 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2340 struct ceph_mds_client *mdsc = session->s_mdsc;
2341 struct ceph_mds_request *req;
2342 struct ceph_mds_reply_head *head = msg->front.iov_base;
2343 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2344 struct ceph_snap_realm *realm;
2347 int mds = session->s_mds;
2349 if (msg->front.iov_len < sizeof(*head)) {
2350 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2355 /* get request, session */
2356 tid = le64_to_cpu(msg->hdr.tid);
2357 mutex_lock(&mdsc->mutex);
2358 req = __lookup_request(mdsc, tid);
2360 dout("handle_reply on unknown tid %llu\n", tid);
2361 mutex_unlock(&mdsc->mutex);
2364 dout("handle_reply %p\n", req);
2366 /* correct session? */
2367 if (req->r_session != session) {
2368 pr_err("mdsc_handle_reply got %llu on session mds%d"
2369 " not mds%d\n", tid, session->s_mds,
2370 req->r_session ? req->r_session->s_mds : -1);
2371 mutex_unlock(&mdsc->mutex);
2376 if ((req->r_got_unsafe && !head->safe) ||
2377 (req->r_got_safe && head->safe)) {
2378 pr_warn("got a dup %s reply on %llu from mds%d\n",
2379 head->safe ? "safe" : "unsafe", tid, mds);
2380 mutex_unlock(&mdsc->mutex);
2383 if (req->r_got_safe && !head->safe) {
2384 pr_warn("got unsafe after safe on %llu from mds%d\n",
2386 mutex_unlock(&mdsc->mutex);
2390 result = le32_to_cpu(head->result);
2394 * if we're not talking to the authority, send to them
2395 * if the authority has changed while we weren't looking,
2396 * send to new authority
2397 * Otherwise we just have to return an ESTALE
2399 if (result == -ESTALE) {
2400 dout("got ESTALE on request %llu", req->r_tid);
2401 req->r_resend_mds = -1;
2402 if (req->r_direct_mode != USE_AUTH_MDS) {
2403 dout("not using auth, setting for that now");
2404 req->r_direct_mode = USE_AUTH_MDS;
2405 __do_request(mdsc, req);
2406 mutex_unlock(&mdsc->mutex);
2409 int mds = __choose_mds(mdsc, req);
2410 if (mds >= 0 && mds != req->r_session->s_mds) {
2411 dout("but auth changed, so resending");
2412 __do_request(mdsc, req);
2413 mutex_unlock(&mdsc->mutex);
2417 dout("have to return ESTALE on request %llu", req->r_tid);
2422 req->r_got_safe = true;
2423 __unregister_request(mdsc, req);
2425 if (req->r_got_unsafe) {
2427 * We already handled the unsafe response, now do the
2428 * cleanup. No need to examine the response; the MDS
2429 * doesn't include any result info in the safe
2430 * response. And even if it did, there is nothing
2431 * useful we could do with a revised return value.
2433 dout("got safe reply %llu, mds%d\n", tid, mds);
2434 list_del_init(&req->r_unsafe_item);
2436 /* last unsafe request during umount? */
2437 if (mdsc->stopping && !__get_oldest_req(mdsc))
2438 complete_all(&mdsc->safe_umount_waiters);
2439 mutex_unlock(&mdsc->mutex);
2443 req->r_got_unsafe = true;
2444 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2447 dout("handle_reply tid %lld result %d\n", tid, result);
2448 rinfo = &req->r_reply_info;
2449 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2450 mutex_unlock(&mdsc->mutex);
2452 mutex_lock(&session->s_mutex);
2454 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2461 if (rinfo->snapblob_len) {
2462 down_write(&mdsc->snap_rwsem);
2463 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2464 rinfo->snapblob + rinfo->snapblob_len,
2465 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2467 downgrade_write(&mdsc->snap_rwsem);
2469 down_read(&mdsc->snap_rwsem);
2472 /* insert trace into our cache */
2473 mutex_lock(&req->r_fill_mutex);
2474 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2476 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2477 req->r_op == CEPH_MDS_OP_LSSNAP))
2478 ceph_readdir_prepopulate(req, req->r_session);
2479 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2481 mutex_unlock(&req->r_fill_mutex);
2483 up_read(&mdsc->snap_rwsem);
2485 ceph_put_snap_realm(mdsc, realm);
2487 mutex_lock(&mdsc->mutex);
2488 if (!req->r_aborted) {
2494 req->r_got_result = true;
2497 dout("reply arrived after request %lld was aborted\n", tid);
2499 mutex_unlock(&mdsc->mutex);
2501 mutex_unlock(&session->s_mutex);
2503 /* kick calling process */
2504 complete_request(mdsc, req);
2506 ceph_mdsc_put_request(req);
2513 * handle mds notification that our request has been forwarded.
2515 static void handle_forward(struct ceph_mds_client *mdsc,
2516 struct ceph_mds_session *session,
2517 struct ceph_msg *msg)
2519 struct ceph_mds_request *req;
2520 u64 tid = le64_to_cpu(msg->hdr.tid);
2524 void *p = msg->front.iov_base;
2525 void *end = p + msg->front.iov_len;
2527 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2528 next_mds = ceph_decode_32(&p);
2529 fwd_seq = ceph_decode_32(&p);
2531 mutex_lock(&mdsc->mutex);
2532 req = __lookup_request(mdsc, tid);
2534 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2535 goto out; /* dup reply? */
2538 if (req->r_aborted) {
2539 dout("forward tid %llu aborted, unregistering\n", tid);
2540 __unregister_request(mdsc, req);
2541 } else if (fwd_seq <= req->r_num_fwd) {
2542 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2543 tid, next_mds, req->r_num_fwd, fwd_seq);
2545 /* resend. forward race not possible; mds would drop */
2546 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2548 BUG_ON(req->r_got_result);
2549 req->r_attempts = 0;
2550 req->r_num_fwd = fwd_seq;
2551 req->r_resend_mds = next_mds;
2552 put_request_session(req);
2553 __do_request(mdsc, req);
2555 ceph_mdsc_put_request(req);
2557 mutex_unlock(&mdsc->mutex);
2561 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2565 * handle a mds session control message
2567 static void handle_session(struct ceph_mds_session *session,
2568 struct ceph_msg *msg)
2570 struct ceph_mds_client *mdsc = session->s_mdsc;
2573 int mds = session->s_mds;
2574 struct ceph_mds_session_head *h = msg->front.iov_base;
2578 if (msg->front.iov_len != sizeof(*h))
2580 op = le32_to_cpu(h->op);
2581 seq = le64_to_cpu(h->seq);
2583 mutex_lock(&mdsc->mutex);
2584 if (op == CEPH_SESSION_CLOSE)
2585 __unregister_session(mdsc, session);
2586 /* FIXME: this ttl calculation is generous */
2587 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2588 mutex_unlock(&mdsc->mutex);
2590 mutex_lock(&session->s_mutex);
2592 dout("handle_session mds%d %s %p state %s seq %llu\n",
2593 mds, ceph_session_op_name(op), session,
2594 ceph_session_state_name(session->s_state), seq);
2596 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2597 session->s_state = CEPH_MDS_SESSION_OPEN;
2598 pr_info("mds%d came back\n", session->s_mds);
2602 case CEPH_SESSION_OPEN:
2603 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2604 pr_info("mds%d reconnect success\n", session->s_mds);
2605 session->s_state = CEPH_MDS_SESSION_OPEN;
2606 renewed_caps(mdsc, session, 0);
2609 __close_session(mdsc, session);
2612 case CEPH_SESSION_RENEWCAPS:
2613 if (session->s_renew_seq == seq)
2614 renewed_caps(mdsc, session, 1);
2617 case CEPH_SESSION_CLOSE:
2618 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2619 pr_info("mds%d reconnect denied\n", session->s_mds);
2620 cleanup_session_requests(mdsc, session);
2621 remove_session_caps(session);
2622 wake = 2; /* for good measure */
2623 wake_up_all(&mdsc->session_close_wq);
2626 case CEPH_SESSION_STALE:
2627 pr_info("mds%d caps went stale, renewing\n",
2629 spin_lock(&session->s_gen_ttl_lock);
2630 session->s_cap_gen++;
2631 session->s_cap_ttl = jiffies - 1;
2632 spin_unlock(&session->s_gen_ttl_lock);
2633 send_renew_caps(mdsc, session);
2636 case CEPH_SESSION_RECALL_STATE:
2637 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2640 case CEPH_SESSION_FLUSHMSG:
2641 send_flushmsg_ack(mdsc, session, seq);
2644 case CEPH_SESSION_FORCE_RO:
2645 dout("force_session_readonly %p\n", session);
2646 spin_lock(&session->s_cap_lock);
2647 session->s_readonly = true;
2648 spin_unlock(&session->s_cap_lock);
2649 wake_up_session_caps(session, 0);
2653 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2657 mutex_unlock(&session->s_mutex);
2659 mutex_lock(&mdsc->mutex);
2660 __wake_requests(mdsc, &session->s_waiting);
2662 kick_requests(mdsc, mds);
2663 mutex_unlock(&mdsc->mutex);
2668 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2669 (int)msg->front.iov_len);
2676 * called under session->mutex.
2678 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2679 struct ceph_mds_session *session)
2681 struct ceph_mds_request *req, *nreq;
2685 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2687 mutex_lock(&mdsc->mutex);
2688 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2689 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2691 ceph_msg_get(req->r_request);
2692 ceph_con_send(&session->s_con, req->r_request);
2697 * also re-send old requests when MDS enters reconnect stage. So that MDS
2698 * can process completed request in clientreplay stage.
2700 p = rb_first(&mdsc->request_tree);
2702 req = rb_entry(p, struct ceph_mds_request, r_node);
2704 if (req->r_got_unsafe)
2706 if (req->r_attempts == 0)
2707 continue; /* only old requests */
2708 if (req->r_session &&
2709 req->r_session->s_mds == session->s_mds) {
2710 err = __prepare_send_request(mdsc, req,
2711 session->s_mds, true);
2713 ceph_msg_get(req->r_request);
2714 ceph_con_send(&session->s_con, req->r_request);
2718 mutex_unlock(&mdsc->mutex);
2722 * Encode information about a cap for a reconnect with the MDS.
2724 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2728 struct ceph_mds_cap_reconnect v2;
2729 struct ceph_mds_cap_reconnect_v1 v1;
2732 struct ceph_inode_info *ci;
2733 struct ceph_reconnect_state *recon_state = arg;
2734 struct ceph_pagelist *pagelist = recon_state->pagelist;
2738 struct dentry *dentry;
2742 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2743 inode, ceph_vinop(inode), cap, cap->cap_id,
2744 ceph_cap_string(cap->issued));
2745 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2749 dentry = d_find_alias(inode);
2751 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2753 err = PTR_ERR(path);
2760 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2764 spin_lock(&ci->i_ceph_lock);
2765 cap->seq = 0; /* reset cap seq */
2766 cap->issue_seq = 0; /* and issue_seq */
2767 cap->mseq = 0; /* and migrate_seq */
2768 cap->cap_gen = cap->session->s_cap_gen;
2770 if (recon_state->flock) {
2771 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2772 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2773 rec.v2.issued = cpu_to_le32(cap->issued);
2774 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2775 rec.v2.pathbase = cpu_to_le64(pathbase);
2776 rec.v2.flock_len = 0;
2777 reclen = sizeof(rec.v2);
2779 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2780 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2781 rec.v1.issued = cpu_to_le32(cap->issued);
2782 rec.v1.size = cpu_to_le64(inode->i_size);
2783 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2784 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2785 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2786 rec.v1.pathbase = cpu_to_le64(pathbase);
2787 reclen = sizeof(rec.v1);
2789 spin_unlock(&ci->i_ceph_lock);
2791 if (recon_state->flock) {
2792 int num_fcntl_locks, num_flock_locks;
2793 struct ceph_filelock *flocks;
2796 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2797 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2798 sizeof(struct ceph_filelock), GFP_NOFS);
2803 err = ceph_encode_locks_to_buffer(inode, flocks,
2813 * number of encoded locks is stable, so copy to pagelist
2815 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2816 (num_fcntl_locks+num_flock_locks) *
2817 sizeof(struct ceph_filelock));
2818 err = ceph_pagelist_append(pagelist, &rec, reclen);
2820 err = ceph_locks_to_pagelist(flocks, pagelist,
2825 err = ceph_pagelist_append(pagelist, &rec, reclen);
2828 recon_state->nr_caps++;
2838 * If an MDS fails and recovers, clients need to reconnect in order to
2839 * reestablish shared state. This includes all caps issued through
2840 * this session _and_ the snap_realm hierarchy. Because it's not
2841 * clear which snap realms the mds cares about, we send everything we
2842 * know about.. that ensures we'll then get any new info the
2843 * recovering MDS might have.
2845 * This is a relatively heavyweight operation, but it's rare.
2847 * called with mdsc->mutex held.
2849 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2850 struct ceph_mds_session *session)
2852 struct ceph_msg *reply;
2854 int mds = session->s_mds;
2857 struct ceph_pagelist *pagelist;
2858 struct ceph_reconnect_state recon_state;
2860 pr_info("mds%d reconnect start\n", mds);
2862 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2864 goto fail_nopagelist;
2865 ceph_pagelist_init(pagelist);
2867 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2871 mutex_lock(&session->s_mutex);
2872 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2875 dout("session %p state %s\n", session,
2876 ceph_session_state_name(session->s_state));
2878 spin_lock(&session->s_gen_ttl_lock);
2879 session->s_cap_gen++;
2880 spin_unlock(&session->s_gen_ttl_lock);
2882 spin_lock(&session->s_cap_lock);
2883 /* don't know if session is readonly */
2884 session->s_readonly = 0;
2886 * notify __ceph_remove_cap() that we are composing cap reconnect.
2887 * If a cap get released before being added to the cap reconnect,
2888 * __ceph_remove_cap() should skip queuing cap release.
2890 session->s_cap_reconnect = 1;
2891 /* drop old cap expires; we're about to reestablish that state */
2892 cleanup_cap_releases(mdsc, session);
2894 /* trim unused caps to reduce MDS's cache rejoin time */
2895 if (mdsc->fsc->sb->s_root)
2896 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2898 ceph_con_close(&session->s_con);
2899 ceph_con_open(&session->s_con,
2900 CEPH_ENTITY_TYPE_MDS, mds,
2901 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2903 /* replay unsafe requests */
2904 replay_unsafe_requests(mdsc, session);
2906 down_read(&mdsc->snap_rwsem);
2908 /* traverse this session's caps */
2909 s_nr_caps = session->s_nr_caps;
2910 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2914 recon_state.nr_caps = 0;
2915 recon_state.pagelist = pagelist;
2916 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2917 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2921 spin_lock(&session->s_cap_lock);
2922 session->s_cap_reconnect = 0;
2923 spin_unlock(&session->s_cap_lock);
2926 * snaprealms. we provide mds with the ino, seq (version), and
2927 * parent for all of our realms. If the mds has any newer info,
2930 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2931 struct ceph_snap_realm *realm =
2932 rb_entry(p, struct ceph_snap_realm, node);
2933 struct ceph_mds_snaprealm_reconnect sr_rec;
2935 dout(" adding snap realm %llx seq %lld parent %llx\n",
2936 realm->ino, realm->seq, realm->parent_ino);
2937 sr_rec.ino = cpu_to_le64(realm->ino);
2938 sr_rec.seq = cpu_to_le64(realm->seq);
2939 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2940 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2945 if (recon_state.flock)
2946 reply->hdr.version = cpu_to_le16(2);
2948 /* raced with cap release? */
2949 if (s_nr_caps != recon_state.nr_caps) {
2950 struct page *page = list_first_entry(&pagelist->head,
2952 __le32 *addr = kmap_atomic(page);
2953 *addr = cpu_to_le32(recon_state.nr_caps);
2954 kunmap_atomic(addr);
2957 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2958 ceph_msg_data_add_pagelist(reply, pagelist);
2959 ceph_con_send(&session->s_con, reply);
2961 mutex_unlock(&session->s_mutex);
2963 mutex_lock(&mdsc->mutex);
2964 __wake_requests(mdsc, &session->s_waiting);
2965 mutex_unlock(&mdsc->mutex);
2967 up_read(&mdsc->snap_rwsem);
2971 ceph_msg_put(reply);
2972 up_read(&mdsc->snap_rwsem);
2973 mutex_unlock(&session->s_mutex);
2975 ceph_pagelist_release(pagelist);
2977 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2983 * compare old and new mdsmaps, kicking requests
2984 * and closing out old connections as necessary
2986 * called under mdsc->mutex.
2988 static void check_new_map(struct ceph_mds_client *mdsc,
2989 struct ceph_mdsmap *newmap,
2990 struct ceph_mdsmap *oldmap)
2993 int oldstate, newstate;
2994 struct ceph_mds_session *s;
2996 dout("check_new_map new %u old %u\n",
2997 newmap->m_epoch, oldmap->m_epoch);
2999 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3000 if (mdsc->sessions[i] == NULL)
3002 s = mdsc->sessions[i];
3003 oldstate = ceph_mdsmap_get_state(oldmap, i);
3004 newstate = ceph_mdsmap_get_state(newmap, i);
3006 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3007 i, ceph_mds_state_name(oldstate),
3008 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3009 ceph_mds_state_name(newstate),
3010 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3011 ceph_session_state_name(s->s_state));
3013 if (i >= newmap->m_max_mds ||
3014 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3015 ceph_mdsmap_get_addr(newmap, i),
3016 sizeof(struct ceph_entity_addr))) {
3017 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3018 /* the session never opened, just close it
3020 __wake_requests(mdsc, &s->s_waiting);
3021 __unregister_session(mdsc, s);
3024 mutex_unlock(&mdsc->mutex);
3025 mutex_lock(&s->s_mutex);
3026 mutex_lock(&mdsc->mutex);
3027 ceph_con_close(&s->s_con);
3028 mutex_unlock(&s->s_mutex);
3029 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3031 } else if (oldstate == newstate) {
3032 continue; /* nothing new with this mds */
3038 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3039 newstate >= CEPH_MDS_STATE_RECONNECT) {
3040 mutex_unlock(&mdsc->mutex);
3041 send_mds_reconnect(mdsc, s);
3042 mutex_lock(&mdsc->mutex);
3046 * kick request on any mds that has gone active.
3048 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3049 newstate >= CEPH_MDS_STATE_ACTIVE) {
3050 if (oldstate != CEPH_MDS_STATE_CREATING &&
3051 oldstate != CEPH_MDS_STATE_STARTING)
3052 pr_info("mds%d recovery completed\n", s->s_mds);
3053 kick_requests(mdsc, i);
3054 ceph_kick_flushing_caps(mdsc, s);
3055 wake_up_session_caps(s, 1);
3059 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3060 s = mdsc->sessions[i];
3063 if (!ceph_mdsmap_is_laggy(newmap, i))
3065 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3066 s->s_state == CEPH_MDS_SESSION_HUNG ||
3067 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3068 dout(" connecting to export targets of laggy mds%d\n",
3070 __open_export_target_sessions(mdsc, s);
3082 * caller must hold session s_mutex, dentry->d_lock
3084 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3086 struct ceph_dentry_info *di = ceph_dentry(dentry);
3088 ceph_put_mds_session(di->lease_session);
3089 di->lease_session = NULL;
3092 static void handle_lease(struct ceph_mds_client *mdsc,
3093 struct ceph_mds_session *session,
3094 struct ceph_msg *msg)
3096 struct super_block *sb = mdsc->fsc->sb;
3097 struct inode *inode;
3098 struct dentry *parent, *dentry;
3099 struct ceph_dentry_info *di;
3100 int mds = session->s_mds;
3101 struct ceph_mds_lease *h = msg->front.iov_base;
3103 struct ceph_vino vino;
3107 dout("handle_lease from mds%d\n", mds);
3110 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3112 vino.ino = le64_to_cpu(h->ino);
3113 vino.snap = CEPH_NOSNAP;
3114 seq = le32_to_cpu(h->seq);
3115 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3116 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3117 if (dname.len != get_unaligned_le32(h+1))
3121 inode = ceph_find_inode(sb, vino);
3122 dout("handle_lease %s, ino %llx %p %.*s\n",
3123 ceph_lease_op_name(h->action), vino.ino, inode,
3124 dname.len, dname.name);
3126 mutex_lock(&session->s_mutex);
3129 if (inode == NULL) {
3130 dout("handle_lease no inode %llx\n", vino.ino);
3135 parent = d_find_alias(inode);
3137 dout("no parent dentry on inode %p\n", inode);
3139 goto release; /* hrm... */
3141 dname.hash = full_name_hash(dname.name, dname.len);
3142 dentry = d_lookup(parent, &dname);
3147 spin_lock(&dentry->d_lock);
3148 di = ceph_dentry(dentry);
3149 switch (h->action) {
3150 case CEPH_MDS_LEASE_REVOKE:
3151 if (di->lease_session == session) {
3152 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3153 h->seq = cpu_to_le32(di->lease_seq);
3154 __ceph_mdsc_drop_dentry_lease(dentry);
3159 case CEPH_MDS_LEASE_RENEW:
3160 if (di->lease_session == session &&
3161 di->lease_gen == session->s_cap_gen &&
3162 di->lease_renew_from &&
3163 di->lease_renew_after == 0) {
3164 unsigned long duration =
3165 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3167 di->lease_seq = seq;
3168 dentry->d_time = di->lease_renew_from + duration;
3169 di->lease_renew_after = di->lease_renew_from +
3171 di->lease_renew_from = 0;
3175 spin_unlock(&dentry->d_lock);
3182 /* let's just reuse the same message */
3183 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3185 ceph_con_send(&session->s_con, msg);
3189 mutex_unlock(&session->s_mutex);
3193 pr_err("corrupt lease message\n");
3197 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3198 struct inode *inode,
3199 struct dentry *dentry, char action,
3202 struct ceph_msg *msg;
3203 struct ceph_mds_lease *lease;
3204 int len = sizeof(*lease) + sizeof(u32);
3207 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3208 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3209 dnamelen = dentry->d_name.len;
3212 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3215 lease = msg->front.iov_base;
3216 lease->action = action;
3217 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3218 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3219 lease->seq = cpu_to_le32(seq);
3220 put_unaligned_le32(dnamelen, lease + 1);
3221 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3224 * if this is a preemptive lease RELEASE, no need to
3225 * flush request stream, since the actual request will
3228 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3230 ceph_con_send(&session->s_con, msg);
3234 * Preemptively release a lease we expect to invalidate anyway.
3235 * Pass @inode always, @dentry is optional.
3237 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3238 struct dentry *dentry)
3240 struct ceph_dentry_info *di;
3241 struct ceph_mds_session *session;
3244 BUG_ON(inode == NULL);
3245 BUG_ON(dentry == NULL);
3247 /* is dentry lease valid? */
3248 spin_lock(&dentry->d_lock);
3249 di = ceph_dentry(dentry);
3250 if (!di || !di->lease_session ||
3251 di->lease_session->s_mds < 0 ||
3252 di->lease_gen != di->lease_session->s_cap_gen ||
3253 !time_before(jiffies, dentry->d_time)) {
3254 dout("lease_release inode %p dentry %p -- "
3257 spin_unlock(&dentry->d_lock);
3261 /* we do have a lease on this dentry; note mds and seq */
3262 session = ceph_get_mds_session(di->lease_session);
3263 seq = di->lease_seq;
3264 __ceph_mdsc_drop_dentry_lease(dentry);
3265 spin_unlock(&dentry->d_lock);
3267 dout("lease_release inode %p dentry %p to mds%d\n",
3268 inode, dentry, session->s_mds);
3269 ceph_mdsc_lease_send_msg(session, inode, dentry,
3270 CEPH_MDS_LEASE_RELEASE, seq);
3271 ceph_put_mds_session(session);
3275 * drop all leases (and dentry refs) in preparation for umount
3277 static void drop_leases(struct ceph_mds_client *mdsc)
3281 dout("drop_leases\n");
3282 mutex_lock(&mdsc->mutex);
3283 for (i = 0; i < mdsc->max_sessions; i++) {
3284 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3287 mutex_unlock(&mdsc->mutex);
3288 mutex_lock(&s->s_mutex);
3289 mutex_unlock(&s->s_mutex);
3290 ceph_put_mds_session(s);
3291 mutex_lock(&mdsc->mutex);
3293 mutex_unlock(&mdsc->mutex);
3299 * delayed work -- periodically trim expired leases, renew caps with mds
3301 static void schedule_delayed(struct ceph_mds_client *mdsc)
3304 unsigned hz = round_jiffies_relative(HZ * delay);
3305 schedule_delayed_work(&mdsc->delayed_work, hz);
3308 static void delayed_work(struct work_struct *work)
3311 struct ceph_mds_client *mdsc =
3312 container_of(work, struct ceph_mds_client, delayed_work.work);
3316 dout("mdsc delayed_work\n");
3317 ceph_check_delayed_caps(mdsc);
3319 mutex_lock(&mdsc->mutex);
3320 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3321 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3322 mdsc->last_renew_caps);
3324 mdsc->last_renew_caps = jiffies;
3326 for (i = 0; i < mdsc->max_sessions; i++) {
3327 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3330 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3331 dout("resending session close request for mds%d\n",
3333 request_close_session(mdsc, s);
3334 ceph_put_mds_session(s);
3337 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3338 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3339 s->s_state = CEPH_MDS_SESSION_HUNG;
3340 pr_info("mds%d hung\n", s->s_mds);
3343 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3344 /* this mds is failed or recovering, just wait */
3345 ceph_put_mds_session(s);
3348 mutex_unlock(&mdsc->mutex);
3350 mutex_lock(&s->s_mutex);
3352 send_renew_caps(mdsc, s);
3354 ceph_con_keepalive(&s->s_con);
3355 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3356 s->s_state == CEPH_MDS_SESSION_HUNG)
3357 ceph_send_cap_releases(mdsc, s);
3358 mutex_unlock(&s->s_mutex);
3359 ceph_put_mds_session(s);
3361 mutex_lock(&mdsc->mutex);
3363 mutex_unlock(&mdsc->mutex);
3365 schedule_delayed(mdsc);
3368 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3371 struct ceph_mds_client *mdsc;
3373 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3378 mutex_init(&mdsc->mutex);
3379 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3380 if (mdsc->mdsmap == NULL) {
3385 init_completion(&mdsc->safe_umount_waiters);
3386 init_waitqueue_head(&mdsc->session_close_wq);
3387 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3388 mdsc->sessions = NULL;
3389 atomic_set(&mdsc->num_sessions, 0);
3390 mdsc->max_sessions = 0;
3392 mdsc->last_snap_seq = 0;
3393 init_rwsem(&mdsc->snap_rwsem);
3394 mdsc->snap_realms = RB_ROOT;
3395 INIT_LIST_HEAD(&mdsc->snap_empty);
3396 spin_lock_init(&mdsc->snap_empty_lock);
3398 mdsc->oldest_tid = 0;
3399 mdsc->request_tree = RB_ROOT;
3400 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3401 mdsc->last_renew_caps = jiffies;
3402 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3403 spin_lock_init(&mdsc->cap_delay_lock);
3404 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3405 spin_lock_init(&mdsc->snap_flush_lock);
3406 mdsc->cap_flush_seq = 0;
3407 INIT_LIST_HEAD(&mdsc->cap_dirty);
3408 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3409 mdsc->num_cap_flushing = 0;
3410 spin_lock_init(&mdsc->cap_dirty_lock);
3411 init_waitqueue_head(&mdsc->cap_flushing_wq);
3412 spin_lock_init(&mdsc->dentry_lru_lock);
3413 INIT_LIST_HEAD(&mdsc->dentry_lru);
3415 ceph_caps_init(mdsc);
3416 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3418 init_rwsem(&mdsc->pool_perm_rwsem);
3419 mdsc->pool_perm_tree = RB_ROOT;
3425 * Wait for safe replies on open mds requests. If we time out, drop
3426 * all requests from the tree to avoid dangling dentry refs.
3428 static void wait_requests(struct ceph_mds_client *mdsc)
3430 struct ceph_options *opts = mdsc->fsc->client->options;
3431 struct ceph_mds_request *req;
3433 mutex_lock(&mdsc->mutex);
3434 if (__get_oldest_req(mdsc)) {
3435 mutex_unlock(&mdsc->mutex);
3437 dout("wait_requests waiting for requests\n");
3438 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3439 ceph_timeout_jiffies(opts->mount_timeout));
3441 /* tear down remaining requests */
3442 mutex_lock(&mdsc->mutex);
3443 while ((req = __get_oldest_req(mdsc))) {
3444 dout("wait_requests timed out on tid %llu\n",
3446 __unregister_request(mdsc, req);
3449 mutex_unlock(&mdsc->mutex);
3450 dout("wait_requests done\n");
3454 * called before mount is ro, and before dentries are torn down.
3455 * (hmm, does this still race with new lookups?)
3457 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3459 dout("pre_umount\n");
3463 ceph_flush_dirty_caps(mdsc);
3464 wait_requests(mdsc);
3467 * wait for reply handlers to drop their request refs and
3468 * their inode/dcache refs
3474 * wait for all write mds requests to flush.
3476 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3478 struct ceph_mds_request *req = NULL, *nextreq;
3481 mutex_lock(&mdsc->mutex);
3482 dout("wait_unsafe_requests want %lld\n", want_tid);
3484 req = __get_oldest_req(mdsc);
3485 while (req && req->r_tid <= want_tid) {
3486 /* find next request */
3487 n = rb_next(&req->r_node);
3489 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3492 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3493 (req->r_op & CEPH_MDS_OP_WRITE)) {
3495 ceph_mdsc_get_request(req);
3497 ceph_mdsc_get_request(nextreq);
3498 mutex_unlock(&mdsc->mutex);
3499 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3500 req->r_tid, want_tid);
3501 wait_for_completion(&req->r_safe_completion);
3502 mutex_lock(&mdsc->mutex);
3503 ceph_mdsc_put_request(req);
3505 break; /* next dne before, so we're done! */
3506 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3507 /* next request was removed from tree */
3508 ceph_mdsc_put_request(nextreq);
3511 ceph_mdsc_put_request(nextreq); /* won't go away */
3515 mutex_unlock(&mdsc->mutex);
3516 dout("wait_unsafe_requests done\n");
3519 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3521 u64 want_tid, want_flush, want_snap;
3523 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3527 mutex_lock(&mdsc->mutex);
3528 want_tid = mdsc->last_tid;
3529 mutex_unlock(&mdsc->mutex);
3531 ceph_flush_dirty_caps(mdsc);
3532 spin_lock(&mdsc->cap_dirty_lock);
3533 want_flush = mdsc->cap_flush_seq;
3534 spin_unlock(&mdsc->cap_dirty_lock);
3536 down_read(&mdsc->snap_rwsem);
3537 want_snap = mdsc->last_snap_seq;
3538 up_read(&mdsc->snap_rwsem);
3540 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3541 want_tid, want_flush, want_snap);
3543 wait_unsafe_requests(mdsc, want_tid);
3544 wait_caps_flush(mdsc, want_flush, want_snap);
3548 * true if all sessions are closed, or we force unmount
3550 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3552 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3554 return atomic_read(&mdsc->num_sessions) == 0;
3558 * called after sb is ro.
3560 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3562 struct ceph_options *opts = mdsc->fsc->client->options;
3563 struct ceph_mds_session *session;
3566 dout("close_sessions\n");
3568 /* close sessions */
3569 mutex_lock(&mdsc->mutex);
3570 for (i = 0; i < mdsc->max_sessions; i++) {
3571 session = __ceph_lookup_mds_session(mdsc, i);
3574 mutex_unlock(&mdsc->mutex);
3575 mutex_lock(&session->s_mutex);
3576 __close_session(mdsc, session);
3577 mutex_unlock(&session->s_mutex);
3578 ceph_put_mds_session(session);
3579 mutex_lock(&mdsc->mutex);
3581 mutex_unlock(&mdsc->mutex);
3583 dout("waiting for sessions to close\n");
3584 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3585 ceph_timeout_jiffies(opts->mount_timeout));
3587 /* tear down remaining sessions */
3588 mutex_lock(&mdsc->mutex);
3589 for (i = 0; i < mdsc->max_sessions; i++) {
3590 if (mdsc->sessions[i]) {
3591 session = get_session(mdsc->sessions[i]);
3592 __unregister_session(mdsc, session);
3593 mutex_unlock(&mdsc->mutex);
3594 mutex_lock(&session->s_mutex);
3595 remove_session_caps(session);
3596 mutex_unlock(&session->s_mutex);
3597 ceph_put_mds_session(session);
3598 mutex_lock(&mdsc->mutex);
3601 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3602 mutex_unlock(&mdsc->mutex);
3604 ceph_cleanup_empty_realms(mdsc);
3606 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3611 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3614 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3616 ceph_mdsmap_destroy(mdsc->mdsmap);
3617 kfree(mdsc->sessions);
3618 ceph_caps_finalize(mdsc);
3619 ceph_pool_perm_destroy(mdsc);
3622 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3624 struct ceph_mds_client *mdsc = fsc->mdsc;
3626 dout("mdsc_destroy %p\n", mdsc);
3627 ceph_mdsc_stop(mdsc);
3629 /* flush out any connection work with references to us */
3634 dout("mdsc_destroy %p done\n", mdsc);
3639 * handle mds map update.
3641 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3645 void *p = msg->front.iov_base;
3646 void *end = p + msg->front.iov_len;
3647 struct ceph_mdsmap *newmap, *oldmap;
3648 struct ceph_fsid fsid;
3651 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3652 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3653 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3655 epoch = ceph_decode_32(&p);
3656 maplen = ceph_decode_32(&p);
3657 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3659 /* do we need it? */
3660 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3661 mutex_lock(&mdsc->mutex);
3662 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3663 dout("handle_map epoch %u <= our %u\n",
3664 epoch, mdsc->mdsmap->m_epoch);
3665 mutex_unlock(&mdsc->mutex);
3669 newmap = ceph_mdsmap_decode(&p, end);
3670 if (IS_ERR(newmap)) {
3671 err = PTR_ERR(newmap);
3675 /* swap into place */
3677 oldmap = mdsc->mdsmap;
3678 mdsc->mdsmap = newmap;
3679 check_new_map(mdsc, newmap, oldmap);
3680 ceph_mdsmap_destroy(oldmap);
3682 mdsc->mdsmap = newmap; /* first mds map */
3684 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3686 __wake_requests(mdsc, &mdsc->waiting_for_map);
3688 mutex_unlock(&mdsc->mutex);
3689 schedule_delayed(mdsc);
3693 mutex_unlock(&mdsc->mutex);
3695 pr_err("error decoding mdsmap %d\n", err);
3699 static struct ceph_connection *con_get(struct ceph_connection *con)
3701 struct ceph_mds_session *s = con->private;
3703 if (get_session(s)) {
3704 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3707 dout("mdsc con_get %p FAIL\n", s);
3711 static void con_put(struct ceph_connection *con)
3713 struct ceph_mds_session *s = con->private;
3715 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3716 ceph_put_mds_session(s);
3720 * if the client is unresponsive for long enough, the mds will kill
3721 * the session entirely.
3723 static void peer_reset(struct ceph_connection *con)
3725 struct ceph_mds_session *s = con->private;
3726 struct ceph_mds_client *mdsc = s->s_mdsc;
3728 pr_warn("mds%d closed our session\n", s->s_mds);
3729 send_mds_reconnect(mdsc, s);
3732 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3734 struct ceph_mds_session *s = con->private;
3735 struct ceph_mds_client *mdsc = s->s_mdsc;
3736 int type = le16_to_cpu(msg->hdr.type);
3738 mutex_lock(&mdsc->mutex);
3739 if (__verify_registered_session(mdsc, s) < 0) {
3740 mutex_unlock(&mdsc->mutex);
3743 mutex_unlock(&mdsc->mutex);
3746 case CEPH_MSG_MDS_MAP:
3747 ceph_mdsc_handle_map(mdsc, msg);
3749 case CEPH_MSG_CLIENT_SESSION:
3750 handle_session(s, msg);
3752 case CEPH_MSG_CLIENT_REPLY:
3753 handle_reply(s, msg);
3755 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3756 handle_forward(mdsc, s, msg);
3758 case CEPH_MSG_CLIENT_CAPS:
3759 ceph_handle_caps(s, msg);
3761 case CEPH_MSG_CLIENT_SNAP:
3762 ceph_handle_snap(mdsc, s, msg);
3764 case CEPH_MSG_CLIENT_LEASE:
3765 handle_lease(mdsc, s, msg);
3769 pr_err("received unknown message type %d %s\n", type,
3770 ceph_msg_type_name(type));
3781 * Note: returned pointer is the address of a structure that's
3782 * managed separately. Caller must *not* attempt to free it.
3784 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3785 int *proto, int force_new)
3787 struct ceph_mds_session *s = con->private;
3788 struct ceph_mds_client *mdsc = s->s_mdsc;
3789 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3790 struct ceph_auth_handshake *auth = &s->s_auth;
3792 if (force_new && auth->authorizer) {
3793 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3794 auth->authorizer = NULL;
3796 if (!auth->authorizer) {
3797 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3800 return ERR_PTR(ret);
3802 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3805 return ERR_PTR(ret);
3807 *proto = ac->protocol;
3813 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3815 struct ceph_mds_session *s = con->private;
3816 struct ceph_mds_client *mdsc = s->s_mdsc;
3817 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3819 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3822 static int invalidate_authorizer(struct ceph_connection *con)
3824 struct ceph_mds_session *s = con->private;
3825 struct ceph_mds_client *mdsc = s->s_mdsc;
3826 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3828 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3830 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3833 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3834 struct ceph_msg_header *hdr, int *skip)
3836 struct ceph_msg *msg;
3837 int type = (int) le16_to_cpu(hdr->type);
3838 int front_len = (int) le32_to_cpu(hdr->front_len);
3844 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3846 pr_err("unable to allocate msg type %d len %d\n",
3854 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3856 struct ceph_mds_session *s = con->private;
3857 struct ceph_auth_handshake *auth = &s->s_auth;
3858 return ceph_auth_sign_message(auth, msg);
3861 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3863 struct ceph_mds_session *s = con->private;
3864 struct ceph_auth_handshake *auth = &s->s_auth;
3865 return ceph_auth_check_message_signature(auth, msg);
3868 static const struct ceph_connection_operations mds_con_ops = {
3871 .dispatch = dispatch,
3872 .get_authorizer = get_authorizer,
3873 .verify_authorizer_reply = verify_authorizer_reply,
3874 .invalidate_authorizer = invalidate_authorizer,
3875 .peer_reset = peer_reset,
3876 .alloc_msg = mds_alloc_msg,
3877 .sign_message = sign_message,
3878 .check_message_signature = check_message_signature,