1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11 #include <linux/ratelimit.h>
14 #include "mds_client.h"
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
48 struct ceph_reconnect_state {
50 struct ceph_pagelist *pagelist;
54 static void __wake_requests(struct ceph_mds_client *mdsc,
55 struct list_head *head);
57 static const struct ceph_connection_operations mds_con_ops;
65 * parse individual inode info
67 static int parse_reply_info_in(void **p, void *end,
68 struct ceph_mds_reply_info_in *info,
74 *p += sizeof(struct ceph_mds_reply_inode) +
75 sizeof(*info->in->fragtree.splits) *
76 le32_to_cpu(info->in->fragtree.nsplits);
78 ceph_decode_32_safe(p, end, info->symlink_len, bad);
79 ceph_decode_need(p, end, info->symlink_len, bad);
81 *p += info->symlink_len;
83 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
84 ceph_decode_copy_safe(p, end, &info->dir_layout,
85 sizeof(info->dir_layout), bad);
87 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
89 ceph_decode_32_safe(p, end, info->xattr_len, bad);
90 ceph_decode_need(p, end, info->xattr_len, bad);
91 info->xattr_data = *p;
92 *p += info->xattr_len;
94 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
95 ceph_decode_64_safe(p, end, info->inline_version, bad);
96 ceph_decode_32_safe(p, end, info->inline_len, bad);
97 ceph_decode_need(p, end, info->inline_len, bad);
98 info->inline_data = *p;
99 *p += info->inline_len;
101 info->inline_version = CEPH_INLINE_NONE;
109 * parse a normal reply, which may contain a (dir+)dentry and/or a
112 static int parse_reply_info_trace(void **p, void *end,
113 struct ceph_mds_reply_info_parsed *info,
118 if (info->head->is_dentry) {
119 err = parse_reply_info_in(p, end, &info->diri, features);
123 if (unlikely(*p + sizeof(*info->dirfrag) > end))
126 *p += sizeof(*info->dirfrag) +
127 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
128 if (unlikely(*p > end))
131 ceph_decode_32_safe(p, end, info->dname_len, bad);
132 ceph_decode_need(p, end, info->dname_len, bad);
134 *p += info->dname_len;
136 *p += sizeof(*info->dlease);
139 if (info->head->is_target) {
140 err = parse_reply_info_in(p, end, &info->targeti, features);
145 if (unlikely(*p != end))
152 pr_err("problem parsing mds trace %d\n", err);
157 * parse readdir results
159 static int parse_reply_info_dir(void **p, void *end,
160 struct ceph_mds_reply_info_parsed *info,
167 if (*p + sizeof(*info->dir_dir) > end)
169 *p += sizeof(*info->dir_dir) +
170 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
174 ceph_decode_need(p, end, sizeof(num) + 2, bad);
175 num = ceph_decode_32(p);
176 info->dir_end = ceph_decode_8(p);
177 info->dir_complete = ceph_decode_8(p);
181 BUG_ON(!info->dir_in);
182 info->dir_dname = (void *)(info->dir_in + num);
183 info->dir_dname_len = (void *)(info->dir_dname + num);
184 info->dir_dlease = (void *)(info->dir_dname_len + num);
185 if ((unsigned long)(info->dir_dlease + num) >
186 (unsigned long)info->dir_in + info->dir_buf_size) {
187 pr_err("dir contents are larger than expected\n");
195 ceph_decode_need(p, end, sizeof(u32)*2, bad);
196 info->dir_dname_len[i] = ceph_decode_32(p);
197 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
198 info->dir_dname[i] = *p;
199 *p += info->dir_dname_len[i];
200 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
202 info->dir_dlease[i] = *p;
203 *p += sizeof(struct ceph_mds_reply_lease);
206 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
221 pr_err("problem parsing dir contents %d\n", err);
226 * parse fcntl F_GETLK results
228 static int parse_reply_info_filelock(void **p, void *end,
229 struct ceph_mds_reply_info_parsed *info,
232 if (*p + sizeof(*info->filelock_reply) > end)
235 info->filelock_reply = *p;
236 *p += sizeof(*info->filelock_reply);
238 if (unlikely(*p != end))
247 * parse create results
249 static int parse_reply_info_create(void **p, void *end,
250 struct ceph_mds_reply_info_parsed *info,
253 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
255 info->has_create_ino = false;
257 info->has_create_ino = true;
258 info->ino = ceph_decode_64(p);
262 if (unlikely(*p != end))
271 * parse extra results
273 static int parse_reply_info_extra(void **p, void *end,
274 struct ceph_mds_reply_info_parsed *info,
277 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
278 return parse_reply_info_filelock(p, end, info, features);
279 else if (info->head->op == CEPH_MDS_OP_READDIR ||
280 info->head->op == CEPH_MDS_OP_LSSNAP)
281 return parse_reply_info_dir(p, end, info, features);
282 else if (info->head->op == CEPH_MDS_OP_CREATE)
283 return parse_reply_info_create(p, end, info, features);
289 * parse entire mds reply
291 static int parse_reply_info(struct ceph_msg *msg,
292 struct ceph_mds_reply_info_parsed *info,
299 info->head = msg->front.iov_base;
300 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
301 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
304 ceph_decode_32_safe(&p, end, len, bad);
306 ceph_decode_need(&p, end, len, bad);
307 err = parse_reply_info_trace(&p, p+len, info, features);
313 ceph_decode_32_safe(&p, end, len, bad);
315 ceph_decode_need(&p, end, len, bad);
316 err = parse_reply_info_extra(&p, p+len, info, features);
322 ceph_decode_32_safe(&p, end, len, bad);
323 info->snapblob_len = len;
334 pr_err("mds parse_reply err %d\n", err);
338 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
342 free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
349 const char *ceph_session_state_name(int s)
352 case CEPH_MDS_SESSION_NEW: return "new";
353 case CEPH_MDS_SESSION_OPENING: return "opening";
354 case CEPH_MDS_SESSION_OPEN: return "open";
355 case CEPH_MDS_SESSION_HUNG: return "hung";
356 case CEPH_MDS_SESSION_CLOSING: return "closing";
357 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
358 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
359 default: return "???";
363 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
365 if (atomic_inc_not_zero(&s->s_ref)) {
366 dout("mdsc get_session %p %d -> %d\n", s,
367 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
370 dout("mdsc get_session %p 0 -- FAIL", s);
375 void ceph_put_mds_session(struct ceph_mds_session *s)
377 dout("mdsc put_session %p %d -> %d\n", s,
378 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
379 if (atomic_dec_and_test(&s->s_ref)) {
380 if (s->s_auth.authorizer)
381 ceph_auth_destroy_authorizer(
382 s->s_mdsc->fsc->client->monc.auth,
383 s->s_auth.authorizer);
389 * called under mdsc->mutex
391 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
394 struct ceph_mds_session *session;
396 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
398 session = mdsc->sessions[mds];
399 dout("lookup_mds_session %p %d\n", session,
400 atomic_read(&session->s_ref));
401 get_session(session);
405 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
407 if (mds >= mdsc->max_sessions)
409 return mdsc->sessions[mds];
412 static int __verify_registered_session(struct ceph_mds_client *mdsc,
413 struct ceph_mds_session *s)
415 if (s->s_mds >= mdsc->max_sessions ||
416 mdsc->sessions[s->s_mds] != s)
422 * create+register a new session for given mds.
423 * called under mdsc->mutex.
425 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
428 struct ceph_mds_session *s;
430 if (mds >= mdsc->mdsmap->m_max_mds)
431 return ERR_PTR(-EINVAL);
433 s = kzalloc(sizeof(*s), GFP_NOFS);
435 return ERR_PTR(-ENOMEM);
438 s->s_state = CEPH_MDS_SESSION_NEW;
441 mutex_init(&s->s_mutex);
443 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
445 spin_lock_init(&s->s_gen_ttl_lock);
447 s->s_cap_ttl = jiffies - 1;
449 spin_lock_init(&s->s_cap_lock);
450 s->s_renew_requested = 0;
452 INIT_LIST_HEAD(&s->s_caps);
455 atomic_set(&s->s_ref, 1);
456 INIT_LIST_HEAD(&s->s_waiting);
457 INIT_LIST_HEAD(&s->s_unsafe);
458 s->s_num_cap_releases = 0;
459 s->s_cap_reconnect = 0;
460 s->s_cap_iterator = NULL;
461 INIT_LIST_HEAD(&s->s_cap_releases);
462 INIT_LIST_HEAD(&s->s_cap_flushing);
463 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
465 dout("register_session mds%d\n", mds);
466 if (mds >= mdsc->max_sessions) {
467 int newmax = 1 << get_count_order(mds+1);
468 struct ceph_mds_session **sa;
470 dout("register_session realloc to %d\n", newmax);
471 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
474 if (mdsc->sessions) {
475 memcpy(sa, mdsc->sessions,
476 mdsc->max_sessions * sizeof(void *));
477 kfree(mdsc->sessions);
480 mdsc->max_sessions = newmax;
482 mdsc->sessions[mds] = s;
483 atomic_inc(&mdsc->num_sessions);
484 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
486 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
487 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
493 return ERR_PTR(-ENOMEM);
497 * called under mdsc->mutex
499 static void __unregister_session(struct ceph_mds_client *mdsc,
500 struct ceph_mds_session *s)
502 dout("__unregister_session mds%d %p\n", s->s_mds, s);
503 BUG_ON(mdsc->sessions[s->s_mds] != s);
504 mdsc->sessions[s->s_mds] = NULL;
505 ceph_con_close(&s->s_con);
506 ceph_put_mds_session(s);
507 atomic_dec(&mdsc->num_sessions);
511 * drop session refs in request.
513 * should be last request ref, or hold mdsc->mutex
515 static void put_request_session(struct ceph_mds_request *req)
517 if (req->r_session) {
518 ceph_put_mds_session(req->r_session);
519 req->r_session = NULL;
523 void ceph_mdsc_release_request(struct kref *kref)
525 struct ceph_mds_request *req = container_of(kref,
526 struct ceph_mds_request,
528 destroy_reply_info(&req->r_reply_info);
530 ceph_msg_put(req->r_request);
532 ceph_msg_put(req->r_reply);
534 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
537 if (req->r_locked_dir)
538 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
539 iput(req->r_target_inode);
542 if (req->r_old_dentry)
543 dput(req->r_old_dentry);
544 if (req->r_old_dentry_dir) {
546 * track (and drop pins for) r_old_dentry_dir
547 * separately, since r_old_dentry's d_parent may have
548 * changed between the dir mutex being dropped and
549 * this request being freed.
551 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
553 iput(req->r_old_dentry_dir);
558 ceph_pagelist_release(req->r_pagelist);
559 put_request_session(req);
560 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
565 * lookup session, bump ref if found.
567 * called under mdsc->mutex.
569 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
572 struct ceph_mds_request *req;
573 struct rb_node *n = mdsc->request_tree.rb_node;
576 req = rb_entry(n, struct ceph_mds_request, r_node);
577 if (tid < req->r_tid)
579 else if (tid > req->r_tid)
582 ceph_mdsc_get_request(req);
589 static void __insert_request(struct ceph_mds_client *mdsc,
590 struct ceph_mds_request *new)
592 struct rb_node **p = &mdsc->request_tree.rb_node;
593 struct rb_node *parent = NULL;
594 struct ceph_mds_request *req = NULL;
598 req = rb_entry(parent, struct ceph_mds_request, r_node);
599 if (new->r_tid < req->r_tid)
601 else if (new->r_tid > req->r_tid)
607 rb_link_node(&new->r_node, parent, p);
608 rb_insert_color(&new->r_node, &mdsc->request_tree);
612 * Register an in-flight request, and assign a tid. Link to directory
613 * are modifying (if any).
615 * Called under mdsc->mutex.
617 static void __register_request(struct ceph_mds_client *mdsc,
618 struct ceph_mds_request *req,
621 req->r_tid = ++mdsc->last_tid;
623 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
625 dout("__register_request %p tid %lld\n", req, req->r_tid);
626 ceph_mdsc_get_request(req);
627 __insert_request(mdsc, req);
629 req->r_uid = current_fsuid();
630 req->r_gid = current_fsgid();
632 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
633 mdsc->oldest_tid = req->r_tid;
636 struct ceph_inode_info *ci = ceph_inode(dir);
639 spin_lock(&ci->i_unsafe_lock);
640 req->r_unsafe_dir = dir;
641 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
642 spin_unlock(&ci->i_unsafe_lock);
646 static void __unregister_request(struct ceph_mds_client *mdsc,
647 struct ceph_mds_request *req)
649 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
651 if (req->r_tid == mdsc->oldest_tid) {
652 struct rb_node *p = rb_next(&req->r_node);
653 mdsc->oldest_tid = 0;
655 struct ceph_mds_request *next_req =
656 rb_entry(p, struct ceph_mds_request, r_node);
657 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
658 mdsc->oldest_tid = next_req->r_tid;
665 rb_erase(&req->r_node, &mdsc->request_tree);
666 RB_CLEAR_NODE(&req->r_node);
668 if (req->r_unsafe_dir) {
669 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
671 spin_lock(&ci->i_unsafe_lock);
672 list_del_init(&req->r_unsafe_dir_item);
673 spin_unlock(&ci->i_unsafe_lock);
675 iput(req->r_unsafe_dir);
676 req->r_unsafe_dir = NULL;
679 complete_all(&req->r_safe_completion);
681 ceph_mdsc_put_request(req);
685 * Choose mds to send request to next. If there is a hint set in the
686 * request (e.g., due to a prior forward hint from the mds), use that.
687 * Otherwise, consult frag tree and/or caps to identify the
688 * appropriate mds. If all else fails, choose randomly.
690 * Called under mdsc->mutex.
692 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
695 * we don't need to worry about protecting the d_parent access
696 * here because we never renaming inside the snapped namespace
697 * except to resplice to another snapdir, and either the old or new
698 * result is a valid result.
700 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
701 dentry = dentry->d_parent;
705 static int __choose_mds(struct ceph_mds_client *mdsc,
706 struct ceph_mds_request *req)
709 struct ceph_inode_info *ci;
710 struct ceph_cap *cap;
711 int mode = req->r_direct_mode;
713 u32 hash = req->r_direct_hash;
714 bool is_hash = req->r_direct_is_hash;
717 * is there a specific mds we should try? ignore hint if we have
718 * no session and the mds is not up (active or recovering).
720 if (req->r_resend_mds >= 0 &&
721 (__have_session(mdsc, req->r_resend_mds) ||
722 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
723 dout("choose_mds using resend_mds mds%d\n",
725 return req->r_resend_mds;
728 if (mode == USE_RANDOM_MDS)
733 inode = req->r_inode;
734 } else if (req->r_dentry) {
735 /* ignore race with rename; old or new d_parent is okay */
736 struct dentry *parent = req->r_dentry->d_parent;
737 struct inode *dir = d_inode(parent);
739 if (dir->i_sb != mdsc->fsc->sb) {
741 inode = d_inode(req->r_dentry);
742 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
743 /* direct snapped/virtual snapdir requests
744 * based on parent dir inode */
745 struct dentry *dn = get_nonsnap_parent(parent);
747 dout("__choose_mds using nonsnap parent %p\n", inode);
750 inode = d_inode(req->r_dentry);
751 if (!inode || mode == USE_AUTH_MDS) {
754 hash = ceph_dentry_hash(dir, req->r_dentry);
760 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
764 ci = ceph_inode(inode);
766 if (is_hash && S_ISDIR(inode->i_mode)) {
767 struct ceph_inode_frag frag;
770 ceph_choose_frag(ci, hash, &frag, &found);
772 if (mode == USE_ANY_MDS && frag.ndist > 0) {
775 /* choose a random replica */
776 get_random_bytes(&r, 1);
779 dout("choose_mds %p %llx.%llx "
780 "frag %u mds%d (%d/%d)\n",
781 inode, ceph_vinop(inode),
784 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
785 CEPH_MDS_STATE_ACTIVE)
789 /* since this file/dir wasn't known to be
790 * replicated, then we want to look for the
791 * authoritative mds. */
794 /* choose auth mds */
796 dout("choose_mds %p %llx.%llx "
797 "frag %u mds%d (auth)\n",
798 inode, ceph_vinop(inode), frag.frag, mds);
799 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
800 CEPH_MDS_STATE_ACTIVE)
806 spin_lock(&ci->i_ceph_lock);
808 if (mode == USE_AUTH_MDS)
809 cap = ci->i_auth_cap;
810 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
811 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
813 spin_unlock(&ci->i_ceph_lock);
816 mds = cap->session->s_mds;
817 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
818 inode, ceph_vinop(inode), mds,
819 cap == ci->i_auth_cap ? "auth " : "", cap);
820 spin_unlock(&ci->i_ceph_lock);
824 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
825 dout("choose_mds chose random mds%d\n", mds);
833 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
835 struct ceph_msg *msg;
836 struct ceph_mds_session_head *h;
838 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
841 pr_err("create_session_msg ENOMEM creating msg\n");
844 h = msg->front.iov_base;
845 h->op = cpu_to_le32(op);
846 h->seq = cpu_to_le64(seq);
852 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
853 * to include additional client metadata fields.
855 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
857 struct ceph_msg *msg;
858 struct ceph_mds_session_head *h;
860 int metadata_bytes = 0;
861 int metadata_key_count = 0;
862 struct ceph_options *opt = mdsc->fsc->client->options;
865 const char* metadata[][2] = {
866 {"hostname", utsname()->nodename},
867 {"kernel_version", utsname()->release},
868 {"entity_id", opt->name ? opt->name : ""},
872 /* Calculate serialized length of metadata */
873 metadata_bytes = 4; /* map length */
874 for (i = 0; metadata[i][0] != NULL; ++i) {
875 metadata_bytes += 8 + strlen(metadata[i][0]) +
876 strlen(metadata[i][1]);
877 metadata_key_count++;
880 /* Allocate the message */
881 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
884 pr_err("create_session_msg ENOMEM creating msg\n");
887 h = msg->front.iov_base;
888 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
889 h->seq = cpu_to_le64(seq);
892 * Serialize client metadata into waiting buffer space, using
893 * the format that userspace expects for map<string, string>
895 * ClientSession messages with metadata are v2
897 msg->hdr.version = cpu_to_le16(2);
898 msg->hdr.compat_version = cpu_to_le16(1);
900 /* The write pointer, following the session_head structure */
901 p = msg->front.iov_base + sizeof(*h);
903 /* Number of entries in the map */
904 ceph_encode_32(&p, metadata_key_count);
906 /* Two length-prefixed strings for each entry in the map */
907 for (i = 0; metadata[i][0] != NULL; ++i) {
908 size_t const key_len = strlen(metadata[i][0]);
909 size_t const val_len = strlen(metadata[i][1]);
911 ceph_encode_32(&p, key_len);
912 memcpy(p, metadata[i][0], key_len);
914 ceph_encode_32(&p, val_len);
915 memcpy(p, metadata[i][1], val_len);
923 * send session open request.
925 * called under mdsc->mutex
927 static int __open_session(struct ceph_mds_client *mdsc,
928 struct ceph_mds_session *session)
930 struct ceph_msg *msg;
932 int mds = session->s_mds;
934 /* wait for mds to go active? */
935 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
936 dout("open_session to mds%d (%s)\n", mds,
937 ceph_mds_state_name(mstate));
938 session->s_state = CEPH_MDS_SESSION_OPENING;
939 session->s_renew_requested = jiffies;
941 /* send connect message */
942 msg = create_session_open_msg(mdsc, session->s_seq);
945 ceph_con_send(&session->s_con, msg);
950 * open sessions for any export targets for the given mds
952 * called under mdsc->mutex
954 static struct ceph_mds_session *
955 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
957 struct ceph_mds_session *session;
959 session = __ceph_lookup_mds_session(mdsc, target);
961 session = register_session(mdsc, target);
965 if (session->s_state == CEPH_MDS_SESSION_NEW ||
966 session->s_state == CEPH_MDS_SESSION_CLOSING)
967 __open_session(mdsc, session);
972 struct ceph_mds_session *
973 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
975 struct ceph_mds_session *session;
977 dout("open_export_target_session to mds%d\n", target);
979 mutex_lock(&mdsc->mutex);
980 session = __open_export_target_session(mdsc, target);
981 mutex_unlock(&mdsc->mutex);
986 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
987 struct ceph_mds_session *session)
989 struct ceph_mds_info *mi;
990 struct ceph_mds_session *ts;
991 int i, mds = session->s_mds;
993 if (mds >= mdsc->mdsmap->m_max_mds)
996 mi = &mdsc->mdsmap->m_info[mds];
997 dout("open_export_target_sessions for mds%d (%d targets)\n",
998 session->s_mds, mi->num_export_targets);
1000 for (i = 0; i < mi->num_export_targets; i++) {
1001 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1003 ceph_put_mds_session(ts);
1007 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1008 struct ceph_mds_session *session)
1010 mutex_lock(&mdsc->mutex);
1011 __open_export_target_sessions(mdsc, session);
1012 mutex_unlock(&mdsc->mutex);
1019 /* caller holds s_cap_lock, we drop it */
1020 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1021 struct ceph_mds_session *session)
1022 __releases(session->s_cap_lock)
1024 LIST_HEAD(tmp_list);
1025 list_splice_init(&session->s_cap_releases, &tmp_list);
1026 session->s_num_cap_releases = 0;
1027 spin_unlock(&session->s_cap_lock);
1029 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1030 while (!list_empty(&tmp_list)) {
1031 struct ceph_cap *cap;
1032 /* zero out the in-progress message */
1033 cap = list_first_entry(&tmp_list,
1034 struct ceph_cap, session_caps);
1035 list_del(&cap->session_caps);
1036 ceph_put_cap(mdsc, cap);
1040 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1041 struct ceph_mds_session *session)
1043 struct ceph_mds_request *req;
1046 dout("cleanup_session_requests mds%d\n", session->s_mds);
1047 mutex_lock(&mdsc->mutex);
1048 while (!list_empty(&session->s_unsafe)) {
1049 req = list_first_entry(&session->s_unsafe,
1050 struct ceph_mds_request, r_unsafe_item);
1051 list_del_init(&req->r_unsafe_item);
1052 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1054 __unregister_request(mdsc, req);
1056 /* zero r_attempts, so kick_requests() will re-send requests */
1057 p = rb_first(&mdsc->request_tree);
1059 req = rb_entry(p, struct ceph_mds_request, r_node);
1061 if (req->r_session &&
1062 req->r_session->s_mds == session->s_mds)
1063 req->r_attempts = 0;
1065 mutex_unlock(&mdsc->mutex);
1069 * Helper to safely iterate over all caps associated with a session, with
1070 * special care taken to handle a racing __ceph_remove_cap().
1072 * Caller must hold session s_mutex.
1074 static int iterate_session_caps(struct ceph_mds_session *session,
1075 int (*cb)(struct inode *, struct ceph_cap *,
1078 struct list_head *p;
1079 struct ceph_cap *cap;
1080 struct inode *inode, *last_inode = NULL;
1081 struct ceph_cap *old_cap = NULL;
1084 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1085 spin_lock(&session->s_cap_lock);
1086 p = session->s_caps.next;
1087 while (p != &session->s_caps) {
1088 cap = list_entry(p, struct ceph_cap, session_caps);
1089 inode = igrab(&cap->ci->vfs_inode);
1094 session->s_cap_iterator = cap;
1095 spin_unlock(&session->s_cap_lock);
1102 ceph_put_cap(session->s_mdsc, old_cap);
1106 ret = cb(inode, cap, arg);
1109 spin_lock(&session->s_cap_lock);
1111 if (cap->ci == NULL) {
1112 dout("iterate_session_caps finishing cap %p removal\n",
1114 BUG_ON(cap->session != session);
1115 cap->session = NULL;
1116 list_del_init(&cap->session_caps);
1117 session->s_nr_caps--;
1118 if (cap->queue_release) {
1119 list_add_tail(&cap->session_caps,
1120 &session->s_cap_releases);
1121 session->s_num_cap_releases++;
1123 old_cap = cap; /* put_cap it w/o locks held */
1131 session->s_cap_iterator = NULL;
1132 spin_unlock(&session->s_cap_lock);
1136 ceph_put_cap(session->s_mdsc, old_cap);
1141 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1144 struct ceph_inode_info *ci = ceph_inode(inode);
1145 LIST_HEAD(to_remove);
1148 dout("removing cap %p, ci is %p, inode is %p\n",
1149 cap, ci, &ci->vfs_inode);
1150 spin_lock(&ci->i_ceph_lock);
1151 __ceph_remove_cap(cap, false);
1152 if (!ci->i_auth_cap) {
1153 struct ceph_cap_flush *cf;
1154 struct ceph_mds_client *mdsc =
1155 ceph_sb_to_client(inode->i_sb)->mdsc;
1158 struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1161 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1162 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1163 list_add(&cf->list, &to_remove);
1166 spin_lock(&mdsc->cap_dirty_lock);
1168 list_for_each_entry(cf, &to_remove, list)
1169 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1171 if (!list_empty(&ci->i_dirty_item)) {
1172 pr_warn_ratelimited(
1173 " dropping dirty %s state for %p %lld\n",
1174 ceph_cap_string(ci->i_dirty_caps),
1175 inode, ceph_ino(inode));
1176 ci->i_dirty_caps = 0;
1177 list_del_init(&ci->i_dirty_item);
1180 if (!list_empty(&ci->i_flushing_item)) {
1181 pr_warn_ratelimited(
1182 " dropping dirty+flushing %s state for %p %lld\n",
1183 ceph_cap_string(ci->i_flushing_caps),
1184 inode, ceph_ino(inode));
1185 ci->i_flushing_caps = 0;
1186 list_del_init(&ci->i_flushing_item);
1187 mdsc->num_cap_flushing--;
1190 spin_unlock(&mdsc->cap_dirty_lock);
1192 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1193 list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1194 ci->i_prealloc_cap_flush = NULL;
1197 spin_unlock(&ci->i_ceph_lock);
1198 while (!list_empty(&to_remove)) {
1199 struct ceph_cap_flush *cf;
1200 cf = list_first_entry(&to_remove,
1201 struct ceph_cap_flush, list);
1202 list_del(&cf->list);
1203 ceph_free_cap_flush(cf);
1211 * caller must hold session s_mutex
1213 static void remove_session_caps(struct ceph_mds_session *session)
1215 dout("remove_session_caps on %p\n", session);
1216 iterate_session_caps(session, remove_session_caps_cb, NULL);
1218 spin_lock(&session->s_cap_lock);
1219 if (session->s_nr_caps > 0) {
1220 struct super_block *sb = session->s_mdsc->fsc->sb;
1221 struct inode *inode;
1222 struct ceph_cap *cap, *prev = NULL;
1223 struct ceph_vino vino;
1225 * iterate_session_caps() skips inodes that are being
1226 * deleted, we need to wait until deletions are complete.
1227 * __wait_on_freeing_inode() is designed for the job,
1228 * but it is not exported, so use lookup inode function
1231 while (!list_empty(&session->s_caps)) {
1232 cap = list_entry(session->s_caps.next,
1233 struct ceph_cap, session_caps);
1237 vino = cap->ci->i_vino;
1238 spin_unlock(&session->s_cap_lock);
1240 inode = ceph_find_inode(sb, vino);
1243 spin_lock(&session->s_cap_lock);
1247 // drop cap expires and unlock s_cap_lock
1248 cleanup_cap_releases(session->s_mdsc, session);
1250 BUG_ON(session->s_nr_caps > 0);
1251 BUG_ON(!list_empty(&session->s_cap_flushing));
1255 * wake up any threads waiting on this session's caps. if the cap is
1256 * old (didn't get renewed on the client reconnect), remove it now.
1258 * caller must hold s_mutex.
1260 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1263 struct ceph_inode_info *ci = ceph_inode(inode);
1265 wake_up_all(&ci->i_cap_wq);
1267 spin_lock(&ci->i_ceph_lock);
1268 ci->i_wanted_max_size = 0;
1269 ci->i_requested_max_size = 0;
1270 spin_unlock(&ci->i_ceph_lock);
1275 static void wake_up_session_caps(struct ceph_mds_session *session,
1278 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1279 iterate_session_caps(session, wake_up_session_cb,
1280 (void *)(unsigned long)reconnect);
1284 * Send periodic message to MDS renewing all currently held caps. The
1285 * ack will reset the expiration for all caps from this session.
1287 * caller holds s_mutex
1289 static int send_renew_caps(struct ceph_mds_client *mdsc,
1290 struct ceph_mds_session *session)
1292 struct ceph_msg *msg;
1295 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1296 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1297 pr_info("mds%d caps stale\n", session->s_mds);
1298 session->s_renew_requested = jiffies;
1300 /* do not try to renew caps until a recovering mds has reconnected
1301 * with its clients. */
1302 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1303 if (state < CEPH_MDS_STATE_RECONNECT) {
1304 dout("send_renew_caps ignoring mds%d (%s)\n",
1305 session->s_mds, ceph_mds_state_name(state));
1309 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1310 ceph_mds_state_name(state));
1311 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1312 ++session->s_renew_seq);
1315 ceph_con_send(&session->s_con, msg);
1319 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1320 struct ceph_mds_session *session, u64 seq)
1322 struct ceph_msg *msg;
1324 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1325 session->s_mds, ceph_session_state_name(session->s_state), seq);
1326 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1329 ceph_con_send(&session->s_con, msg);
1335 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1337 * Called under session->s_mutex
1339 static void renewed_caps(struct ceph_mds_client *mdsc,
1340 struct ceph_mds_session *session, int is_renew)
1345 spin_lock(&session->s_cap_lock);
1346 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1348 session->s_cap_ttl = session->s_renew_requested +
1349 mdsc->mdsmap->m_session_timeout*HZ;
1352 if (time_before(jiffies, session->s_cap_ttl)) {
1353 pr_info("mds%d caps renewed\n", session->s_mds);
1356 pr_info("mds%d caps still stale\n", session->s_mds);
1359 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1360 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1361 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1362 spin_unlock(&session->s_cap_lock);
1365 wake_up_session_caps(session, 0);
1369 * send a session close request
1371 static int request_close_session(struct ceph_mds_client *mdsc,
1372 struct ceph_mds_session *session)
1374 struct ceph_msg *msg;
1376 dout("request_close_session mds%d state %s seq %lld\n",
1377 session->s_mds, ceph_session_state_name(session->s_state),
1379 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1382 ceph_con_send(&session->s_con, msg);
1387 * Called with s_mutex held.
1389 static int __close_session(struct ceph_mds_client *mdsc,
1390 struct ceph_mds_session *session)
1392 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1394 session->s_state = CEPH_MDS_SESSION_CLOSING;
1395 return request_close_session(mdsc, session);
1399 * Trim old(er) caps.
1401 * Because we can't cache an inode without one or more caps, we do
1402 * this indirectly: if a cap is unused, we prune its aliases, at which
1403 * point the inode will hopefully get dropped to.
1405 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1406 * memory pressure from the MDS, though, so it needn't be perfect.
1408 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1410 struct ceph_mds_session *session = arg;
1411 struct ceph_inode_info *ci = ceph_inode(inode);
1412 int used, wanted, oissued, mine;
1414 if (session->s_trim_caps <= 0)
1417 spin_lock(&ci->i_ceph_lock);
1418 mine = cap->issued | cap->implemented;
1419 used = __ceph_caps_used(ci);
1420 wanted = __ceph_caps_file_wanted(ci);
1421 oissued = __ceph_caps_issued_other(ci, cap);
1423 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1424 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1425 ceph_cap_string(used), ceph_cap_string(wanted));
1426 if (cap == ci->i_auth_cap) {
1427 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1428 !list_empty(&ci->i_cap_snaps))
1430 if ((used | wanted) & CEPH_CAP_ANY_WR)
1433 /* The inode has cached pages, but it's no longer used.
1434 * we can safely drop it */
1435 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1436 !(oissued & CEPH_CAP_FILE_CACHE)) {
1440 if ((used | wanted) & ~oissued & mine)
1441 goto out; /* we need these caps */
1443 session->s_trim_caps--;
1445 /* we aren't the only cap.. just remove us */
1446 __ceph_remove_cap(cap, true);
1448 /* try dropping referring dentries */
1449 spin_unlock(&ci->i_ceph_lock);
1450 d_prune_aliases(inode);
1451 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1452 inode, cap, atomic_read(&inode->i_count));
1457 spin_unlock(&ci->i_ceph_lock);
1462 * Trim session cap count down to some max number.
1464 static int trim_caps(struct ceph_mds_client *mdsc,
1465 struct ceph_mds_session *session,
1468 int trim_caps = session->s_nr_caps - max_caps;
1470 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1471 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1472 if (trim_caps > 0) {
1473 session->s_trim_caps = trim_caps;
1474 iterate_session_caps(session, trim_caps_cb, session);
1475 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1476 session->s_mds, session->s_nr_caps, max_caps,
1477 trim_caps - session->s_trim_caps);
1478 session->s_trim_caps = 0;
1481 ceph_send_cap_releases(mdsc, session);
1485 static int check_capsnap_flush(struct ceph_inode_info *ci,
1489 spin_lock(&ci->i_ceph_lock);
1490 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1491 struct ceph_cap_snap *capsnap =
1492 list_first_entry(&ci->i_cap_snaps,
1493 struct ceph_cap_snap, ci_item);
1494 ret = capsnap->follows >= want_snap_seq;
1496 spin_unlock(&ci->i_ceph_lock);
1500 static int check_caps_flush(struct ceph_mds_client *mdsc,
1504 struct ceph_cap_flush *cf;
1507 spin_lock(&mdsc->cap_dirty_lock);
1508 n = rb_first(&mdsc->cap_flush_tree);
1509 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1510 if (cf && cf->tid <= want_flush_tid) {
1511 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1512 cf->tid, want_flush_tid);
1515 spin_unlock(&mdsc->cap_dirty_lock);
1520 * flush all dirty inode data to disk.
1522 * returns true if we've flushed through want_flush_tid
1524 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1525 u64 want_flush_tid, u64 want_snap_seq)
1529 dout("check_caps_flush want %llu snap want %llu\n",
1530 want_flush_tid, want_snap_seq);
1531 mutex_lock(&mdsc->mutex);
1532 for (mds = 0; mds < mdsc->max_sessions; ) {
1533 struct ceph_mds_session *session = mdsc->sessions[mds];
1534 struct inode *inode = NULL;
1540 get_session(session);
1541 mutex_unlock(&mdsc->mutex);
1543 mutex_lock(&session->s_mutex);
1544 if (!list_empty(&session->s_cap_snaps_flushing)) {
1545 struct ceph_cap_snap *capsnap =
1546 list_first_entry(&session->s_cap_snaps_flushing,
1547 struct ceph_cap_snap,
1549 struct ceph_inode_info *ci = capsnap->ci;
1550 if (!check_capsnap_flush(ci, want_snap_seq)) {
1551 dout("check_cap_flush still flushing snap %p "
1552 "follows %lld <= %lld to mds%d\n",
1553 &ci->vfs_inode, capsnap->follows,
1554 want_snap_seq, mds);
1555 inode = igrab(&ci->vfs_inode);
1558 mutex_unlock(&session->s_mutex);
1559 ceph_put_mds_session(session);
1562 wait_event(mdsc->cap_flushing_wq,
1563 check_capsnap_flush(ceph_inode(inode),
1570 mutex_lock(&mdsc->mutex);
1572 mutex_unlock(&mdsc->mutex);
1574 wait_event(mdsc->cap_flushing_wq,
1575 check_caps_flush(mdsc, want_flush_tid));
1577 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1581 * called under s_mutex
1583 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1584 struct ceph_mds_session *session)
1586 struct ceph_msg *msg = NULL;
1587 struct ceph_mds_cap_release *head;
1588 struct ceph_mds_cap_item *item;
1589 struct ceph_cap *cap;
1590 LIST_HEAD(tmp_list);
1591 int num_cap_releases;
1593 spin_lock(&session->s_cap_lock);
1595 list_splice_init(&session->s_cap_releases, &tmp_list);
1596 num_cap_releases = session->s_num_cap_releases;
1597 session->s_num_cap_releases = 0;
1598 spin_unlock(&session->s_cap_lock);
1600 while (!list_empty(&tmp_list)) {
1602 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1603 PAGE_CACHE_SIZE, GFP_NOFS, false);
1606 head = msg->front.iov_base;
1607 head->num = cpu_to_le32(0);
1608 msg->front.iov_len = sizeof(*head);
1610 cap = list_first_entry(&tmp_list, struct ceph_cap,
1612 list_del(&cap->session_caps);
1615 head = msg->front.iov_base;
1616 le32_add_cpu(&head->num, 1);
1617 item = msg->front.iov_base + msg->front.iov_len;
1618 item->ino = cpu_to_le64(cap->cap_ino);
1619 item->cap_id = cpu_to_le64(cap->cap_id);
1620 item->migrate_seq = cpu_to_le32(cap->mseq);
1621 item->seq = cpu_to_le32(cap->issue_seq);
1622 msg->front.iov_len += sizeof(*item);
1624 ceph_put_cap(mdsc, cap);
1626 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1627 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1628 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1629 ceph_con_send(&session->s_con, msg);
1634 BUG_ON(num_cap_releases != 0);
1636 spin_lock(&session->s_cap_lock);
1637 if (!list_empty(&session->s_cap_releases))
1639 spin_unlock(&session->s_cap_lock);
1642 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1643 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1644 ceph_con_send(&session->s_con, msg);
1648 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1650 spin_lock(&session->s_cap_lock);
1651 list_splice(&tmp_list, &session->s_cap_releases);
1652 session->s_num_cap_releases += num_cap_releases;
1653 spin_unlock(&session->s_cap_lock);
1660 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1663 struct ceph_inode_info *ci = ceph_inode(dir);
1664 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1665 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1666 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1667 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1668 int order, num_entries;
1670 spin_lock(&ci->i_ceph_lock);
1671 num_entries = ci->i_files + ci->i_subdirs;
1672 spin_unlock(&ci->i_ceph_lock);
1673 num_entries = max(num_entries, 1);
1674 num_entries = min(num_entries, opt->max_readdir);
1676 order = get_order(size * num_entries);
1677 while (order >= 0) {
1678 rinfo->dir_in = (void*)__get_free_pages(GFP_KERNEL |
1688 num_entries = (PAGE_SIZE << order) / size;
1689 num_entries = min(num_entries, opt->max_readdir);
1691 rinfo->dir_buf_size = PAGE_SIZE << order;
1692 req->r_num_caps = num_entries + 1;
1693 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1694 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1699 * Create an mds request.
1701 struct ceph_mds_request *
1702 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1704 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1707 return ERR_PTR(-ENOMEM);
1709 mutex_init(&req->r_fill_mutex);
1711 req->r_started = jiffies;
1712 req->r_resend_mds = -1;
1713 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1715 kref_init(&req->r_kref);
1716 INIT_LIST_HEAD(&req->r_wait);
1717 init_completion(&req->r_completion);
1718 init_completion(&req->r_safe_completion);
1719 INIT_LIST_HEAD(&req->r_unsafe_item);
1721 req->r_stamp = CURRENT_TIME;
1724 req->r_direct_mode = mode;
1729 * return oldest (lowest) request, tid in request tree, 0 if none.
1731 * called under mdsc->mutex.
1733 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1735 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1737 return rb_entry(rb_first(&mdsc->request_tree),
1738 struct ceph_mds_request, r_node);
1741 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1743 return mdsc->oldest_tid;
1747 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1748 * on build_path_from_dentry in fs/cifs/dir.c.
1750 * If @stop_on_nosnap, generate path relative to the first non-snapped
1753 * Encode hidden .snap dirs as a double /, i.e.
1754 * foo/.snap/bar -> foo//bar
1756 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1759 struct dentry *temp;
1765 return ERR_PTR(-EINVAL);
1769 seq = read_seqbegin(&rename_lock);
1771 for (temp = dentry; !IS_ROOT(temp);) {
1772 struct inode *inode = d_inode(temp);
1773 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1774 len++; /* slash only */
1775 else if (stop_on_nosnap && inode &&
1776 ceph_snap(inode) == CEPH_NOSNAP)
1779 len += 1 + temp->d_name.len;
1780 temp = temp->d_parent;
1784 len--; /* no leading '/' */
1786 path = kmalloc(len+1, GFP_NOFS);
1788 return ERR_PTR(-ENOMEM);
1790 path[pos] = 0; /* trailing null */
1792 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1793 struct inode *inode;
1795 spin_lock(&temp->d_lock);
1796 inode = d_inode(temp);
1797 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1798 dout("build_path path+%d: %p SNAPDIR\n",
1800 } else if (stop_on_nosnap && inode &&
1801 ceph_snap(inode) == CEPH_NOSNAP) {
1802 spin_unlock(&temp->d_lock);
1805 pos -= temp->d_name.len;
1807 spin_unlock(&temp->d_lock);
1810 strncpy(path + pos, temp->d_name.name,
1813 spin_unlock(&temp->d_lock);
1816 temp = temp->d_parent;
1819 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1820 pr_err("build_path did not end path lookup where "
1821 "expected, namelen is %d, pos is %d\n", len, pos);
1822 /* presumably this is only possible if racing with a
1823 rename of one of the parent directories (we can not
1824 lock the dentries above us to prevent this, but
1825 retrying should be harmless) */
1830 *base = ceph_ino(d_inode(temp));
1832 dout("build_path on %p %d built %llx '%.*s'\n",
1833 dentry, d_count(dentry), *base, len, path);
1837 static int build_dentry_path(struct dentry *dentry,
1838 const char **ppath, int *ppathlen, u64 *pino,
1843 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1844 *pino = ceph_ino(d_inode(dentry->d_parent));
1845 *ppath = dentry->d_name.name;
1846 *ppathlen = dentry->d_name.len;
1849 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1851 return PTR_ERR(path);
1857 static int build_inode_path(struct inode *inode,
1858 const char **ppath, int *ppathlen, u64 *pino,
1861 struct dentry *dentry;
1864 if (ceph_snap(inode) == CEPH_NOSNAP) {
1865 *pino = ceph_ino(inode);
1869 dentry = d_find_alias(inode);
1870 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1873 return PTR_ERR(path);
1880 * request arguments may be specified via an inode *, a dentry *, or
1881 * an explicit ino+path.
1883 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1884 const char *rpath, u64 rino,
1885 const char **ppath, int *pathlen,
1886 u64 *ino, int *freepath)
1891 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1892 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1894 } else if (rdentry) {
1895 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1896 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1898 } else if (rpath || rino) {
1901 *pathlen = rpath ? strlen(rpath) : 0;
1902 dout(" path %.*s\n", *pathlen, rpath);
1909 * called under mdsc->mutex
1911 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1912 struct ceph_mds_request *req,
1913 int mds, bool drop_cap_releases)
1915 struct ceph_msg *msg;
1916 struct ceph_mds_request_head *head;
1917 const char *path1 = NULL;
1918 const char *path2 = NULL;
1919 u64 ino1 = 0, ino2 = 0;
1920 int pathlen1 = 0, pathlen2 = 0;
1921 int freepath1 = 0, freepath2 = 0;
1927 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1928 req->r_path1, req->r_ino1.ino,
1929 &path1, &pathlen1, &ino1, &freepath1);
1935 ret = set_request_path_attr(NULL, req->r_old_dentry,
1936 req->r_path2, req->r_ino2.ino,
1937 &path2, &pathlen2, &ino2, &freepath2);
1943 len = sizeof(*head) +
1944 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1945 sizeof(struct ceph_timespec);
1947 /* calculate (max) length for cap releases */
1948 len += sizeof(struct ceph_mds_request_release) *
1949 (!!req->r_inode_drop + !!req->r_dentry_drop +
1950 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1951 if (req->r_dentry_drop)
1952 len += req->r_dentry->d_name.len;
1953 if (req->r_old_dentry_drop)
1954 len += req->r_old_dentry->d_name.len;
1956 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1958 msg = ERR_PTR(-ENOMEM);
1962 msg->hdr.version = cpu_to_le16(2);
1963 msg->hdr.tid = cpu_to_le64(req->r_tid);
1965 head = msg->front.iov_base;
1966 p = msg->front.iov_base + sizeof(*head);
1967 end = msg->front.iov_base + msg->front.iov_len;
1969 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1970 head->op = cpu_to_le32(req->r_op);
1971 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1972 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1973 head->args = req->r_args;
1975 ceph_encode_filepath(&p, end, ino1, path1);
1976 ceph_encode_filepath(&p, end, ino2, path2);
1978 /* make note of release offset, in case we need to replay */
1979 req->r_request_release_offset = p - msg->front.iov_base;
1983 if (req->r_inode_drop)
1984 releases += ceph_encode_inode_release(&p,
1985 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1986 mds, req->r_inode_drop, req->r_inode_unless, 0);
1987 if (req->r_dentry_drop)
1988 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1989 mds, req->r_dentry_drop, req->r_dentry_unless);
1990 if (req->r_old_dentry_drop)
1991 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1992 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1993 if (req->r_old_inode_drop)
1994 releases += ceph_encode_inode_release(&p,
1995 d_inode(req->r_old_dentry),
1996 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1998 if (drop_cap_releases) {
2000 p = msg->front.iov_base + req->r_request_release_offset;
2003 head->num_releases = cpu_to_le16(releases);
2007 struct ceph_timespec ts;
2008 ceph_encode_timespec(&ts, &req->r_stamp);
2009 ceph_encode_copy(&p, &ts, sizeof(ts));
2013 msg->front.iov_len = p - msg->front.iov_base;
2014 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2016 if (req->r_pagelist) {
2017 struct ceph_pagelist *pagelist = req->r_pagelist;
2018 atomic_inc(&pagelist->refcnt);
2019 ceph_msg_data_add_pagelist(msg, pagelist);
2020 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2022 msg->hdr.data_len = 0;
2025 msg->hdr.data_off = cpu_to_le16(0);
2029 kfree((char *)path2);
2032 kfree((char *)path1);
2038 * called under mdsc->mutex if error, under no mutex if
2041 static void complete_request(struct ceph_mds_client *mdsc,
2042 struct ceph_mds_request *req)
2044 if (req->r_callback)
2045 req->r_callback(mdsc, req);
2047 complete_all(&req->r_completion);
2051 * called under mdsc->mutex
2053 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2054 struct ceph_mds_request *req,
2055 int mds, bool drop_cap_releases)
2057 struct ceph_mds_request_head *rhead;
2058 struct ceph_msg *msg;
2063 struct ceph_cap *cap =
2064 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2067 req->r_sent_on_mseq = cap->mseq;
2069 req->r_sent_on_mseq = -1;
2071 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2072 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2074 if (req->r_got_unsafe) {
2077 * Replay. Do not regenerate message (and rebuild
2078 * paths, etc.); just use the original message.
2079 * Rebuilding paths will break for renames because
2080 * d_move mangles the src name.
2082 msg = req->r_request;
2083 rhead = msg->front.iov_base;
2085 flags = le32_to_cpu(rhead->flags);
2086 flags |= CEPH_MDS_FLAG_REPLAY;
2087 rhead->flags = cpu_to_le32(flags);
2089 if (req->r_target_inode)
2090 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2092 rhead->num_retry = req->r_attempts - 1;
2094 /* remove cap/dentry releases from message */
2095 rhead->num_releases = 0;
2098 p = msg->front.iov_base + req->r_request_release_offset;
2100 struct ceph_timespec ts;
2101 ceph_encode_timespec(&ts, &req->r_stamp);
2102 ceph_encode_copy(&p, &ts, sizeof(ts));
2105 msg->front.iov_len = p - msg->front.iov_base;
2106 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2110 if (req->r_request) {
2111 ceph_msg_put(req->r_request);
2112 req->r_request = NULL;
2114 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2116 req->r_err = PTR_ERR(msg);
2117 return PTR_ERR(msg);
2119 req->r_request = msg;
2121 rhead = msg->front.iov_base;
2122 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2123 if (req->r_got_unsafe)
2124 flags |= CEPH_MDS_FLAG_REPLAY;
2125 if (req->r_locked_dir)
2126 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2127 rhead->flags = cpu_to_le32(flags);
2128 rhead->num_fwd = req->r_num_fwd;
2129 rhead->num_retry = req->r_attempts - 1;
2132 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2137 * send request, or put it on the appropriate wait list.
2139 static int __do_request(struct ceph_mds_client *mdsc,
2140 struct ceph_mds_request *req)
2142 struct ceph_mds_session *session = NULL;
2146 if (req->r_err || req->r_got_result) {
2148 __unregister_request(mdsc, req);
2152 if (req->r_timeout &&
2153 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2154 dout("do_request timed out\n");
2158 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2159 dout("do_request forced umount\n");
2164 put_request_session(req);
2166 mds = __choose_mds(mdsc, req);
2168 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2169 dout("do_request no mds or not active, waiting for map\n");
2170 list_add(&req->r_wait, &mdsc->waiting_for_map);
2174 /* get, open session */
2175 session = __ceph_lookup_mds_session(mdsc, mds);
2177 session = register_session(mdsc, mds);
2178 if (IS_ERR(session)) {
2179 err = PTR_ERR(session);
2183 req->r_session = get_session(session);
2185 dout("do_request mds%d session %p state %s\n", mds, session,
2186 ceph_session_state_name(session->s_state));
2187 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2188 session->s_state != CEPH_MDS_SESSION_HUNG) {
2189 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2190 session->s_state == CEPH_MDS_SESSION_CLOSING)
2191 __open_session(mdsc, session);
2192 list_add(&req->r_wait, &session->s_waiting);
2197 req->r_resend_mds = -1; /* forget any previous mds hint */
2199 if (req->r_request_started == 0) /* note request start time */
2200 req->r_request_started = jiffies;
2202 err = __prepare_send_request(mdsc, req, mds, false);
2204 ceph_msg_get(req->r_request);
2205 ceph_con_send(&session->s_con, req->r_request);
2209 ceph_put_mds_session(session);
2212 dout("__do_request early error %d\n", err);
2214 complete_request(mdsc, req);
2215 __unregister_request(mdsc, req);
2222 * called under mdsc->mutex
2224 static void __wake_requests(struct ceph_mds_client *mdsc,
2225 struct list_head *head)
2227 struct ceph_mds_request *req;
2228 LIST_HEAD(tmp_list);
2230 list_splice_init(head, &tmp_list);
2232 while (!list_empty(&tmp_list)) {
2233 req = list_entry(tmp_list.next,
2234 struct ceph_mds_request, r_wait);
2235 list_del_init(&req->r_wait);
2236 dout(" wake request %p tid %llu\n", req, req->r_tid);
2237 __do_request(mdsc, req);
2242 * Wake up threads with requests pending for @mds, so that they can
2243 * resubmit their requests to a possibly different mds.
2245 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2247 struct ceph_mds_request *req;
2248 struct rb_node *p = rb_first(&mdsc->request_tree);
2250 dout("kick_requests mds%d\n", mds);
2252 req = rb_entry(p, struct ceph_mds_request, r_node);
2254 if (req->r_got_unsafe)
2256 if (req->r_attempts > 0)
2257 continue; /* only new requests */
2258 if (req->r_session &&
2259 req->r_session->s_mds == mds) {
2260 dout(" kicking tid %llu\n", req->r_tid);
2261 list_del_init(&req->r_wait);
2262 __do_request(mdsc, req);
2267 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2268 struct ceph_mds_request *req)
2270 dout("submit_request on %p\n", req);
2271 mutex_lock(&mdsc->mutex);
2272 __register_request(mdsc, req, NULL);
2273 __do_request(mdsc, req);
2274 mutex_unlock(&mdsc->mutex);
2278 * Synchrously perform an mds request. Take care of all of the
2279 * session setup, forwarding, retry details.
2281 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2283 struct ceph_mds_request *req)
2287 dout("do_request on %p\n", req);
2289 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2291 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2292 if (req->r_locked_dir)
2293 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2294 if (req->r_old_dentry_dir)
2295 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2299 mutex_lock(&mdsc->mutex);
2300 __register_request(mdsc, req, dir);
2301 __do_request(mdsc, req);
2309 mutex_unlock(&mdsc->mutex);
2310 dout("do_request waiting\n");
2311 if (!req->r_timeout && req->r_wait_for_completion) {
2312 err = req->r_wait_for_completion(mdsc, req);
2314 long timeleft = wait_for_completion_killable_timeout(
2316 ceph_timeout_jiffies(req->r_timeout));
2320 err = -EIO; /* timed out */
2322 err = timeleft; /* killed */
2324 dout("do_request waited, got %d\n", err);
2325 mutex_lock(&mdsc->mutex);
2327 /* only abort if we didn't race with a real reply */
2328 if (req->r_got_result) {
2329 err = le32_to_cpu(req->r_reply_info.head->result);
2330 } else if (err < 0) {
2331 dout("aborted request %lld with %d\n", req->r_tid, err);
2334 * ensure we aren't running concurrently with
2335 * ceph_fill_trace or ceph_readdir_prepopulate, which
2336 * rely on locks (dir mutex) held by our caller.
2338 mutex_lock(&req->r_fill_mutex);
2340 req->r_aborted = true;
2341 mutex_unlock(&req->r_fill_mutex);
2343 if (req->r_locked_dir &&
2344 (req->r_op & CEPH_MDS_OP_WRITE))
2345 ceph_invalidate_dir_request(req);
2351 mutex_unlock(&mdsc->mutex);
2352 dout("do_request %p done, result %d\n", req, err);
2357 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2358 * namespace request.
2360 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2362 struct inode *inode = req->r_locked_dir;
2364 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2366 ceph_dir_clear_complete(inode);
2368 ceph_invalidate_dentry_lease(req->r_dentry);
2369 if (req->r_old_dentry)
2370 ceph_invalidate_dentry_lease(req->r_old_dentry);
2376 * We take the session mutex and parse and process the reply immediately.
2377 * This preserves the logical ordering of replies, capabilities, etc., sent
2378 * by the MDS as they are applied to our local cache.
2380 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2382 struct ceph_mds_client *mdsc = session->s_mdsc;
2383 struct ceph_mds_request *req;
2384 struct ceph_mds_reply_head *head = msg->front.iov_base;
2385 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2386 struct ceph_snap_realm *realm;
2389 int mds = session->s_mds;
2391 if (msg->front.iov_len < sizeof(*head)) {
2392 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2397 /* get request, session */
2398 tid = le64_to_cpu(msg->hdr.tid);
2399 mutex_lock(&mdsc->mutex);
2400 req = __lookup_request(mdsc, tid);
2402 dout("handle_reply on unknown tid %llu\n", tid);
2403 mutex_unlock(&mdsc->mutex);
2406 dout("handle_reply %p\n", req);
2408 /* correct session? */
2409 if (req->r_session != session) {
2410 pr_err("mdsc_handle_reply got %llu on session mds%d"
2411 " not mds%d\n", tid, session->s_mds,
2412 req->r_session ? req->r_session->s_mds : -1);
2413 mutex_unlock(&mdsc->mutex);
2418 if ((req->r_got_unsafe && !head->safe) ||
2419 (req->r_got_safe && head->safe)) {
2420 pr_warn("got a dup %s reply on %llu from mds%d\n",
2421 head->safe ? "safe" : "unsafe", tid, mds);
2422 mutex_unlock(&mdsc->mutex);
2425 if (req->r_got_safe) {
2426 pr_warn("got unsafe after safe on %llu from mds%d\n",
2428 mutex_unlock(&mdsc->mutex);
2432 result = le32_to_cpu(head->result);
2436 * if we're not talking to the authority, send to them
2437 * if the authority has changed while we weren't looking,
2438 * send to new authority
2439 * Otherwise we just have to return an ESTALE
2441 if (result == -ESTALE) {
2442 dout("got ESTALE on request %llu", req->r_tid);
2443 req->r_resend_mds = -1;
2444 if (req->r_direct_mode != USE_AUTH_MDS) {
2445 dout("not using auth, setting for that now");
2446 req->r_direct_mode = USE_AUTH_MDS;
2447 __do_request(mdsc, req);
2448 mutex_unlock(&mdsc->mutex);
2451 int mds = __choose_mds(mdsc, req);
2452 if (mds >= 0 && mds != req->r_session->s_mds) {
2453 dout("but auth changed, so resending");
2454 __do_request(mdsc, req);
2455 mutex_unlock(&mdsc->mutex);
2459 dout("have to return ESTALE on request %llu", req->r_tid);
2464 req->r_got_safe = true;
2465 __unregister_request(mdsc, req);
2467 if (req->r_got_unsafe) {
2469 * We already handled the unsafe response, now do the
2470 * cleanup. No need to examine the response; the MDS
2471 * doesn't include any result info in the safe
2472 * response. And even if it did, there is nothing
2473 * useful we could do with a revised return value.
2475 dout("got safe reply %llu, mds%d\n", tid, mds);
2476 list_del_init(&req->r_unsafe_item);
2478 /* last unsafe request during umount? */
2479 if (mdsc->stopping && !__get_oldest_req(mdsc))
2480 complete_all(&mdsc->safe_umount_waiters);
2481 mutex_unlock(&mdsc->mutex);
2485 req->r_got_unsafe = true;
2486 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2489 dout("handle_reply tid %lld result %d\n", tid, result);
2490 rinfo = &req->r_reply_info;
2491 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2492 mutex_unlock(&mdsc->mutex);
2494 mutex_lock(&session->s_mutex);
2496 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2503 if (rinfo->snapblob_len) {
2504 down_write(&mdsc->snap_rwsem);
2505 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2506 rinfo->snapblob + rinfo->snapblob_len,
2507 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2509 downgrade_write(&mdsc->snap_rwsem);
2511 down_read(&mdsc->snap_rwsem);
2514 /* insert trace into our cache */
2515 mutex_lock(&req->r_fill_mutex);
2516 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2518 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2519 req->r_op == CEPH_MDS_OP_LSSNAP))
2520 ceph_readdir_prepopulate(req, req->r_session);
2521 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2523 mutex_unlock(&req->r_fill_mutex);
2525 up_read(&mdsc->snap_rwsem);
2527 ceph_put_snap_realm(mdsc, realm);
2529 mutex_lock(&mdsc->mutex);
2530 if (!req->r_aborted) {
2534 req->r_reply = ceph_msg_get(msg);
2535 req->r_got_result = true;
2538 dout("reply arrived after request %lld was aborted\n", tid);
2540 mutex_unlock(&mdsc->mutex);
2542 mutex_unlock(&session->s_mutex);
2544 /* kick calling process */
2545 complete_request(mdsc, req);
2547 ceph_mdsc_put_request(req);
2554 * handle mds notification that our request has been forwarded.
2556 static void handle_forward(struct ceph_mds_client *mdsc,
2557 struct ceph_mds_session *session,
2558 struct ceph_msg *msg)
2560 struct ceph_mds_request *req;
2561 u64 tid = le64_to_cpu(msg->hdr.tid);
2565 void *p = msg->front.iov_base;
2566 void *end = p + msg->front.iov_len;
2568 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2569 next_mds = ceph_decode_32(&p);
2570 fwd_seq = ceph_decode_32(&p);
2572 mutex_lock(&mdsc->mutex);
2573 req = __lookup_request(mdsc, tid);
2575 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2576 goto out; /* dup reply? */
2579 if (req->r_aborted) {
2580 dout("forward tid %llu aborted, unregistering\n", tid);
2581 __unregister_request(mdsc, req);
2582 } else if (fwd_seq <= req->r_num_fwd) {
2583 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2584 tid, next_mds, req->r_num_fwd, fwd_seq);
2586 /* resend. forward race not possible; mds would drop */
2587 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2589 BUG_ON(req->r_got_result);
2590 req->r_attempts = 0;
2591 req->r_num_fwd = fwd_seq;
2592 req->r_resend_mds = next_mds;
2593 put_request_session(req);
2594 __do_request(mdsc, req);
2596 ceph_mdsc_put_request(req);
2598 mutex_unlock(&mdsc->mutex);
2602 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2606 * handle a mds session control message
2608 static void handle_session(struct ceph_mds_session *session,
2609 struct ceph_msg *msg)
2611 struct ceph_mds_client *mdsc = session->s_mdsc;
2614 int mds = session->s_mds;
2615 struct ceph_mds_session_head *h = msg->front.iov_base;
2619 if (msg->front.iov_len != sizeof(*h))
2621 op = le32_to_cpu(h->op);
2622 seq = le64_to_cpu(h->seq);
2624 mutex_lock(&mdsc->mutex);
2625 if (op == CEPH_SESSION_CLOSE)
2626 __unregister_session(mdsc, session);
2627 /* FIXME: this ttl calculation is generous */
2628 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2629 mutex_unlock(&mdsc->mutex);
2631 mutex_lock(&session->s_mutex);
2633 dout("handle_session mds%d %s %p state %s seq %llu\n",
2634 mds, ceph_session_op_name(op), session,
2635 ceph_session_state_name(session->s_state), seq);
2637 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2638 session->s_state = CEPH_MDS_SESSION_OPEN;
2639 pr_info("mds%d came back\n", session->s_mds);
2643 case CEPH_SESSION_OPEN:
2644 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2645 pr_info("mds%d reconnect success\n", session->s_mds);
2646 session->s_state = CEPH_MDS_SESSION_OPEN;
2647 renewed_caps(mdsc, session, 0);
2650 __close_session(mdsc, session);
2653 case CEPH_SESSION_RENEWCAPS:
2654 if (session->s_renew_seq == seq)
2655 renewed_caps(mdsc, session, 1);
2658 case CEPH_SESSION_CLOSE:
2659 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2660 pr_info("mds%d reconnect denied\n", session->s_mds);
2661 cleanup_session_requests(mdsc, session);
2662 remove_session_caps(session);
2663 wake = 2; /* for good measure */
2664 wake_up_all(&mdsc->session_close_wq);
2667 case CEPH_SESSION_STALE:
2668 pr_info("mds%d caps went stale, renewing\n",
2670 spin_lock(&session->s_gen_ttl_lock);
2671 session->s_cap_gen++;
2672 session->s_cap_ttl = jiffies - 1;
2673 spin_unlock(&session->s_gen_ttl_lock);
2674 send_renew_caps(mdsc, session);
2677 case CEPH_SESSION_RECALL_STATE:
2678 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2681 case CEPH_SESSION_FLUSHMSG:
2682 send_flushmsg_ack(mdsc, session, seq);
2685 case CEPH_SESSION_FORCE_RO:
2686 dout("force_session_readonly %p\n", session);
2687 spin_lock(&session->s_cap_lock);
2688 session->s_readonly = true;
2689 spin_unlock(&session->s_cap_lock);
2690 wake_up_session_caps(session, 0);
2694 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2698 mutex_unlock(&session->s_mutex);
2700 mutex_lock(&mdsc->mutex);
2701 __wake_requests(mdsc, &session->s_waiting);
2703 kick_requests(mdsc, mds);
2704 mutex_unlock(&mdsc->mutex);
2709 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2710 (int)msg->front.iov_len);
2717 * called under session->mutex.
2719 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2720 struct ceph_mds_session *session)
2722 struct ceph_mds_request *req, *nreq;
2726 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2728 mutex_lock(&mdsc->mutex);
2729 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2730 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2732 ceph_msg_get(req->r_request);
2733 ceph_con_send(&session->s_con, req->r_request);
2738 * also re-send old requests when MDS enters reconnect stage. So that MDS
2739 * can process completed request in clientreplay stage.
2741 p = rb_first(&mdsc->request_tree);
2743 req = rb_entry(p, struct ceph_mds_request, r_node);
2745 if (req->r_got_unsafe)
2747 if (req->r_attempts == 0)
2748 continue; /* only old requests */
2749 if (req->r_session &&
2750 req->r_session->s_mds == session->s_mds) {
2751 err = __prepare_send_request(mdsc, req,
2752 session->s_mds, true);
2754 ceph_msg_get(req->r_request);
2755 ceph_con_send(&session->s_con, req->r_request);
2759 mutex_unlock(&mdsc->mutex);
2763 * Encode information about a cap for a reconnect with the MDS.
2765 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2769 struct ceph_mds_cap_reconnect v2;
2770 struct ceph_mds_cap_reconnect_v1 v1;
2773 struct ceph_inode_info *ci;
2774 struct ceph_reconnect_state *recon_state = arg;
2775 struct ceph_pagelist *pagelist = recon_state->pagelist;
2779 struct dentry *dentry;
2783 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2784 inode, ceph_vinop(inode), cap, cap->cap_id,
2785 ceph_cap_string(cap->issued));
2786 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2790 dentry = d_find_alias(inode);
2792 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2794 err = PTR_ERR(path);
2801 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2805 spin_lock(&ci->i_ceph_lock);
2806 cap->seq = 0; /* reset cap seq */
2807 cap->issue_seq = 0; /* and issue_seq */
2808 cap->mseq = 0; /* and migrate_seq */
2809 cap->cap_gen = cap->session->s_cap_gen;
2811 if (recon_state->flock) {
2812 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2813 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2814 rec.v2.issued = cpu_to_le32(cap->issued);
2815 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2816 rec.v2.pathbase = cpu_to_le64(pathbase);
2817 rec.v2.flock_len = 0;
2818 reclen = sizeof(rec.v2);
2820 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2821 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2822 rec.v1.issued = cpu_to_le32(cap->issued);
2823 rec.v1.size = cpu_to_le64(inode->i_size);
2824 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2825 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2826 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2827 rec.v1.pathbase = cpu_to_le64(pathbase);
2828 reclen = sizeof(rec.v1);
2830 spin_unlock(&ci->i_ceph_lock);
2832 if (recon_state->flock) {
2833 int num_fcntl_locks, num_flock_locks;
2834 struct ceph_filelock *flocks;
2837 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2838 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2839 sizeof(struct ceph_filelock), GFP_NOFS);
2844 err = ceph_encode_locks_to_buffer(inode, flocks,
2854 * number of encoded locks is stable, so copy to pagelist
2856 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2857 (num_fcntl_locks+num_flock_locks) *
2858 sizeof(struct ceph_filelock));
2859 err = ceph_pagelist_append(pagelist, &rec, reclen);
2861 err = ceph_locks_to_pagelist(flocks, pagelist,
2866 err = ceph_pagelist_append(pagelist, &rec, reclen);
2869 recon_state->nr_caps++;
2879 * If an MDS fails and recovers, clients need to reconnect in order to
2880 * reestablish shared state. This includes all caps issued through
2881 * this session _and_ the snap_realm hierarchy. Because it's not
2882 * clear which snap realms the mds cares about, we send everything we
2883 * know about.. that ensures we'll then get any new info the
2884 * recovering MDS might have.
2886 * This is a relatively heavyweight operation, but it's rare.
2888 * called with mdsc->mutex held.
2890 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2891 struct ceph_mds_session *session)
2893 struct ceph_msg *reply;
2895 int mds = session->s_mds;
2898 struct ceph_pagelist *pagelist;
2899 struct ceph_reconnect_state recon_state;
2901 pr_info("mds%d reconnect start\n", mds);
2903 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2905 goto fail_nopagelist;
2906 ceph_pagelist_init(pagelist);
2908 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2912 mutex_lock(&session->s_mutex);
2913 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2916 dout("session %p state %s\n", session,
2917 ceph_session_state_name(session->s_state));
2919 spin_lock(&session->s_gen_ttl_lock);
2920 session->s_cap_gen++;
2921 spin_unlock(&session->s_gen_ttl_lock);
2923 spin_lock(&session->s_cap_lock);
2924 /* don't know if session is readonly */
2925 session->s_readonly = 0;
2927 * notify __ceph_remove_cap() that we are composing cap reconnect.
2928 * If a cap get released before being added to the cap reconnect,
2929 * __ceph_remove_cap() should skip queuing cap release.
2931 session->s_cap_reconnect = 1;
2932 /* drop old cap expires; we're about to reestablish that state */
2933 cleanup_cap_releases(mdsc, session);
2935 /* trim unused caps to reduce MDS's cache rejoin time */
2936 if (mdsc->fsc->sb->s_root)
2937 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2939 ceph_con_close(&session->s_con);
2940 ceph_con_open(&session->s_con,
2941 CEPH_ENTITY_TYPE_MDS, mds,
2942 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2944 /* replay unsafe requests */
2945 replay_unsafe_requests(mdsc, session);
2947 down_read(&mdsc->snap_rwsem);
2949 /* traverse this session's caps */
2950 s_nr_caps = session->s_nr_caps;
2951 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2955 recon_state.nr_caps = 0;
2956 recon_state.pagelist = pagelist;
2957 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2958 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2962 spin_lock(&session->s_cap_lock);
2963 session->s_cap_reconnect = 0;
2964 spin_unlock(&session->s_cap_lock);
2967 * snaprealms. we provide mds with the ino, seq (version), and
2968 * parent for all of our realms. If the mds has any newer info,
2971 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2972 struct ceph_snap_realm *realm =
2973 rb_entry(p, struct ceph_snap_realm, node);
2974 struct ceph_mds_snaprealm_reconnect sr_rec;
2976 dout(" adding snap realm %llx seq %lld parent %llx\n",
2977 realm->ino, realm->seq, realm->parent_ino);
2978 sr_rec.ino = cpu_to_le64(realm->ino);
2979 sr_rec.seq = cpu_to_le64(realm->seq);
2980 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2981 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2986 if (recon_state.flock)
2987 reply->hdr.version = cpu_to_le16(2);
2989 /* raced with cap release? */
2990 if (s_nr_caps != recon_state.nr_caps) {
2991 struct page *page = list_first_entry(&pagelist->head,
2993 __le32 *addr = kmap_atomic(page);
2994 *addr = cpu_to_le32(recon_state.nr_caps);
2995 kunmap_atomic(addr);
2998 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2999 ceph_msg_data_add_pagelist(reply, pagelist);
3001 ceph_early_kick_flushing_caps(mdsc, session);
3003 ceph_con_send(&session->s_con, reply);
3005 mutex_unlock(&session->s_mutex);
3007 mutex_lock(&mdsc->mutex);
3008 __wake_requests(mdsc, &session->s_waiting);
3009 mutex_unlock(&mdsc->mutex);
3011 up_read(&mdsc->snap_rwsem);
3015 ceph_msg_put(reply);
3016 up_read(&mdsc->snap_rwsem);
3017 mutex_unlock(&session->s_mutex);
3019 ceph_pagelist_release(pagelist);
3021 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3027 * compare old and new mdsmaps, kicking requests
3028 * and closing out old connections as necessary
3030 * called under mdsc->mutex.
3032 static void check_new_map(struct ceph_mds_client *mdsc,
3033 struct ceph_mdsmap *newmap,
3034 struct ceph_mdsmap *oldmap)
3037 int oldstate, newstate;
3038 struct ceph_mds_session *s;
3040 dout("check_new_map new %u old %u\n",
3041 newmap->m_epoch, oldmap->m_epoch);
3043 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3044 if (mdsc->sessions[i] == NULL)
3046 s = mdsc->sessions[i];
3047 oldstate = ceph_mdsmap_get_state(oldmap, i);
3048 newstate = ceph_mdsmap_get_state(newmap, i);
3050 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3051 i, ceph_mds_state_name(oldstate),
3052 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3053 ceph_mds_state_name(newstate),
3054 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3055 ceph_session_state_name(s->s_state));
3057 if (i >= newmap->m_max_mds ||
3058 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3059 ceph_mdsmap_get_addr(newmap, i),
3060 sizeof(struct ceph_entity_addr))) {
3061 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3062 /* the session never opened, just close it
3064 __wake_requests(mdsc, &s->s_waiting);
3065 __unregister_session(mdsc, s);
3068 mutex_unlock(&mdsc->mutex);
3069 mutex_lock(&s->s_mutex);
3070 mutex_lock(&mdsc->mutex);
3071 ceph_con_close(&s->s_con);
3072 mutex_unlock(&s->s_mutex);
3073 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3075 } else if (oldstate == newstate) {
3076 continue; /* nothing new with this mds */
3082 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3083 newstate >= CEPH_MDS_STATE_RECONNECT) {
3084 mutex_unlock(&mdsc->mutex);
3085 send_mds_reconnect(mdsc, s);
3086 mutex_lock(&mdsc->mutex);
3090 * kick request on any mds that has gone active.
3092 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3093 newstate >= CEPH_MDS_STATE_ACTIVE) {
3094 if (oldstate != CEPH_MDS_STATE_CREATING &&
3095 oldstate != CEPH_MDS_STATE_STARTING)
3096 pr_info("mds%d recovery completed\n", s->s_mds);
3097 kick_requests(mdsc, i);
3098 ceph_kick_flushing_caps(mdsc, s);
3099 wake_up_session_caps(s, 1);
3103 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3104 s = mdsc->sessions[i];
3107 if (!ceph_mdsmap_is_laggy(newmap, i))
3109 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3110 s->s_state == CEPH_MDS_SESSION_HUNG ||
3111 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3112 dout(" connecting to export targets of laggy mds%d\n",
3114 __open_export_target_sessions(mdsc, s);
3126 * caller must hold session s_mutex, dentry->d_lock
3128 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3130 struct ceph_dentry_info *di = ceph_dentry(dentry);
3132 ceph_put_mds_session(di->lease_session);
3133 di->lease_session = NULL;
3136 static void handle_lease(struct ceph_mds_client *mdsc,
3137 struct ceph_mds_session *session,
3138 struct ceph_msg *msg)
3140 struct super_block *sb = mdsc->fsc->sb;
3141 struct inode *inode;
3142 struct dentry *parent, *dentry;
3143 struct ceph_dentry_info *di;
3144 int mds = session->s_mds;
3145 struct ceph_mds_lease *h = msg->front.iov_base;
3147 struct ceph_vino vino;
3151 dout("handle_lease from mds%d\n", mds);
3154 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3156 vino.ino = le64_to_cpu(h->ino);
3157 vino.snap = CEPH_NOSNAP;
3158 seq = le32_to_cpu(h->seq);
3159 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3160 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3161 if (dname.len != get_unaligned_le32(h+1))
3165 inode = ceph_find_inode(sb, vino);
3166 dout("handle_lease %s, ino %llx %p %.*s\n",
3167 ceph_lease_op_name(h->action), vino.ino, inode,
3168 dname.len, dname.name);
3170 mutex_lock(&session->s_mutex);
3173 if (inode == NULL) {
3174 dout("handle_lease no inode %llx\n", vino.ino);
3179 parent = d_find_alias(inode);
3181 dout("no parent dentry on inode %p\n", inode);
3183 goto release; /* hrm... */
3185 dname.hash = full_name_hash(dname.name, dname.len);
3186 dentry = d_lookup(parent, &dname);
3191 spin_lock(&dentry->d_lock);
3192 di = ceph_dentry(dentry);
3193 switch (h->action) {
3194 case CEPH_MDS_LEASE_REVOKE:
3195 if (di->lease_session == session) {
3196 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3197 h->seq = cpu_to_le32(di->lease_seq);
3198 __ceph_mdsc_drop_dentry_lease(dentry);
3203 case CEPH_MDS_LEASE_RENEW:
3204 if (di->lease_session == session &&
3205 di->lease_gen == session->s_cap_gen &&
3206 di->lease_renew_from &&
3207 di->lease_renew_after == 0) {
3208 unsigned long duration =
3209 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3211 di->lease_seq = seq;
3212 dentry->d_time = di->lease_renew_from + duration;
3213 di->lease_renew_after = di->lease_renew_from +
3215 di->lease_renew_from = 0;
3219 spin_unlock(&dentry->d_lock);
3226 /* let's just reuse the same message */
3227 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3229 ceph_con_send(&session->s_con, msg);
3233 mutex_unlock(&session->s_mutex);
3237 pr_err("corrupt lease message\n");
3241 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3242 struct inode *inode,
3243 struct dentry *dentry, char action,
3246 struct ceph_msg *msg;
3247 struct ceph_mds_lease *lease;
3248 int len = sizeof(*lease) + sizeof(u32);
3251 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3252 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3253 dnamelen = dentry->d_name.len;
3256 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3259 lease = msg->front.iov_base;
3260 lease->action = action;
3261 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3262 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3263 lease->seq = cpu_to_le32(seq);
3264 put_unaligned_le32(dnamelen, lease + 1);
3265 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3268 * if this is a preemptive lease RELEASE, no need to
3269 * flush request stream, since the actual request will
3272 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3274 ceph_con_send(&session->s_con, msg);
3278 * Preemptively release a lease we expect to invalidate anyway.
3279 * Pass @inode always, @dentry is optional.
3281 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3282 struct dentry *dentry)
3284 struct ceph_dentry_info *di;
3285 struct ceph_mds_session *session;
3288 BUG_ON(inode == NULL);
3289 BUG_ON(dentry == NULL);
3291 /* is dentry lease valid? */
3292 spin_lock(&dentry->d_lock);
3293 di = ceph_dentry(dentry);
3294 if (!di || !di->lease_session ||
3295 di->lease_session->s_mds < 0 ||
3296 di->lease_gen != di->lease_session->s_cap_gen ||
3297 !time_before(jiffies, dentry->d_time)) {
3298 dout("lease_release inode %p dentry %p -- "
3301 spin_unlock(&dentry->d_lock);
3305 /* we do have a lease on this dentry; note mds and seq */
3306 session = ceph_get_mds_session(di->lease_session);
3307 seq = di->lease_seq;
3308 __ceph_mdsc_drop_dentry_lease(dentry);
3309 spin_unlock(&dentry->d_lock);
3311 dout("lease_release inode %p dentry %p to mds%d\n",
3312 inode, dentry, session->s_mds);
3313 ceph_mdsc_lease_send_msg(session, inode, dentry,
3314 CEPH_MDS_LEASE_RELEASE, seq);
3315 ceph_put_mds_session(session);
3319 * drop all leases (and dentry refs) in preparation for umount
3321 static void drop_leases(struct ceph_mds_client *mdsc)
3325 dout("drop_leases\n");
3326 mutex_lock(&mdsc->mutex);
3327 for (i = 0; i < mdsc->max_sessions; i++) {
3328 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3331 mutex_unlock(&mdsc->mutex);
3332 mutex_lock(&s->s_mutex);
3333 mutex_unlock(&s->s_mutex);
3334 ceph_put_mds_session(s);
3335 mutex_lock(&mdsc->mutex);
3337 mutex_unlock(&mdsc->mutex);
3343 * delayed work -- periodically trim expired leases, renew caps with mds
3345 static void schedule_delayed(struct ceph_mds_client *mdsc)
3348 unsigned hz = round_jiffies_relative(HZ * delay);
3349 schedule_delayed_work(&mdsc->delayed_work, hz);
3352 static void delayed_work(struct work_struct *work)
3355 struct ceph_mds_client *mdsc =
3356 container_of(work, struct ceph_mds_client, delayed_work.work);
3360 dout("mdsc delayed_work\n");
3361 ceph_check_delayed_caps(mdsc);
3363 mutex_lock(&mdsc->mutex);
3364 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3365 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3366 mdsc->last_renew_caps);
3368 mdsc->last_renew_caps = jiffies;
3370 for (i = 0; i < mdsc->max_sessions; i++) {
3371 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3374 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3375 dout("resending session close request for mds%d\n",
3377 request_close_session(mdsc, s);
3378 ceph_put_mds_session(s);
3381 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3382 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3383 s->s_state = CEPH_MDS_SESSION_HUNG;
3384 pr_info("mds%d hung\n", s->s_mds);
3387 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3388 /* this mds is failed or recovering, just wait */
3389 ceph_put_mds_session(s);
3392 mutex_unlock(&mdsc->mutex);
3394 mutex_lock(&s->s_mutex);
3396 send_renew_caps(mdsc, s);
3398 ceph_con_keepalive(&s->s_con);
3399 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3400 s->s_state == CEPH_MDS_SESSION_HUNG)
3401 ceph_send_cap_releases(mdsc, s);
3402 mutex_unlock(&s->s_mutex);
3403 ceph_put_mds_session(s);
3405 mutex_lock(&mdsc->mutex);
3407 mutex_unlock(&mdsc->mutex);
3409 schedule_delayed(mdsc);
3412 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3415 struct ceph_mds_client *mdsc;
3417 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3422 mutex_init(&mdsc->mutex);
3423 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3424 if (mdsc->mdsmap == NULL) {
3429 init_completion(&mdsc->safe_umount_waiters);
3430 init_waitqueue_head(&mdsc->session_close_wq);
3431 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3432 mdsc->sessions = NULL;
3433 atomic_set(&mdsc->num_sessions, 0);
3434 mdsc->max_sessions = 0;
3436 mdsc->last_snap_seq = 0;
3437 init_rwsem(&mdsc->snap_rwsem);
3438 mdsc->snap_realms = RB_ROOT;
3439 INIT_LIST_HEAD(&mdsc->snap_empty);
3440 spin_lock_init(&mdsc->snap_empty_lock);
3442 mdsc->oldest_tid = 0;
3443 mdsc->request_tree = RB_ROOT;
3444 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3445 mdsc->last_renew_caps = jiffies;
3446 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3447 spin_lock_init(&mdsc->cap_delay_lock);
3448 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3449 spin_lock_init(&mdsc->snap_flush_lock);
3450 mdsc->last_cap_flush_tid = 1;
3451 mdsc->cap_flush_tree = RB_ROOT;
3452 INIT_LIST_HEAD(&mdsc->cap_dirty);
3453 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3454 mdsc->num_cap_flushing = 0;
3455 spin_lock_init(&mdsc->cap_dirty_lock);
3456 init_waitqueue_head(&mdsc->cap_flushing_wq);
3457 spin_lock_init(&mdsc->dentry_lru_lock);
3458 INIT_LIST_HEAD(&mdsc->dentry_lru);
3460 ceph_caps_init(mdsc);
3461 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3463 init_rwsem(&mdsc->pool_perm_rwsem);
3464 mdsc->pool_perm_tree = RB_ROOT;
3470 * Wait for safe replies on open mds requests. If we time out, drop
3471 * all requests from the tree to avoid dangling dentry refs.
3473 static void wait_requests(struct ceph_mds_client *mdsc)
3475 struct ceph_options *opts = mdsc->fsc->client->options;
3476 struct ceph_mds_request *req;
3478 mutex_lock(&mdsc->mutex);
3479 if (__get_oldest_req(mdsc)) {
3480 mutex_unlock(&mdsc->mutex);
3482 dout("wait_requests waiting for requests\n");
3483 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3484 ceph_timeout_jiffies(opts->mount_timeout));
3486 /* tear down remaining requests */
3487 mutex_lock(&mdsc->mutex);
3488 while ((req = __get_oldest_req(mdsc))) {
3489 dout("wait_requests timed out on tid %llu\n",
3491 __unregister_request(mdsc, req);
3494 mutex_unlock(&mdsc->mutex);
3495 dout("wait_requests done\n");
3499 * called before mount is ro, and before dentries are torn down.
3500 * (hmm, does this still race with new lookups?)
3502 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3504 dout("pre_umount\n");
3508 ceph_flush_dirty_caps(mdsc);
3509 wait_requests(mdsc);
3512 * wait for reply handlers to drop their request refs and
3513 * their inode/dcache refs
3519 * wait for all write mds requests to flush.
3521 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3523 struct ceph_mds_request *req = NULL, *nextreq;
3526 mutex_lock(&mdsc->mutex);
3527 dout("wait_unsafe_requests want %lld\n", want_tid);
3529 req = __get_oldest_req(mdsc);
3530 while (req && req->r_tid <= want_tid) {
3531 /* find next request */
3532 n = rb_next(&req->r_node);
3534 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3537 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3538 (req->r_op & CEPH_MDS_OP_WRITE)) {
3540 ceph_mdsc_get_request(req);
3542 ceph_mdsc_get_request(nextreq);
3543 mutex_unlock(&mdsc->mutex);
3544 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3545 req->r_tid, want_tid);
3546 wait_for_completion(&req->r_safe_completion);
3547 mutex_lock(&mdsc->mutex);
3548 ceph_mdsc_put_request(req);
3550 break; /* next dne before, so we're done! */
3551 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3552 /* next request was removed from tree */
3553 ceph_mdsc_put_request(nextreq);
3556 ceph_mdsc_put_request(nextreq); /* won't go away */
3560 mutex_unlock(&mdsc->mutex);
3561 dout("wait_unsafe_requests done\n");
3564 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3566 u64 want_tid, want_flush, want_snap;
3568 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3572 mutex_lock(&mdsc->mutex);
3573 want_tid = mdsc->last_tid;
3574 mutex_unlock(&mdsc->mutex);
3576 ceph_flush_dirty_caps(mdsc);
3577 spin_lock(&mdsc->cap_dirty_lock);
3578 want_flush = mdsc->last_cap_flush_tid;
3579 spin_unlock(&mdsc->cap_dirty_lock);
3581 down_read(&mdsc->snap_rwsem);
3582 want_snap = mdsc->last_snap_seq;
3583 up_read(&mdsc->snap_rwsem);
3585 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3586 want_tid, want_flush, want_snap);
3588 wait_unsafe_requests(mdsc, want_tid);
3589 wait_caps_flush(mdsc, want_flush, want_snap);
3593 * true if all sessions are closed, or we force unmount
3595 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3597 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3599 return atomic_read(&mdsc->num_sessions) == 0;
3603 * called after sb is ro.
3605 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3607 struct ceph_options *opts = mdsc->fsc->client->options;
3608 struct ceph_mds_session *session;
3611 dout("close_sessions\n");
3613 /* close sessions */
3614 mutex_lock(&mdsc->mutex);
3615 for (i = 0; i < mdsc->max_sessions; i++) {
3616 session = __ceph_lookup_mds_session(mdsc, i);
3619 mutex_unlock(&mdsc->mutex);
3620 mutex_lock(&session->s_mutex);
3621 __close_session(mdsc, session);
3622 mutex_unlock(&session->s_mutex);
3623 ceph_put_mds_session(session);
3624 mutex_lock(&mdsc->mutex);
3626 mutex_unlock(&mdsc->mutex);
3628 dout("waiting for sessions to close\n");
3629 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3630 ceph_timeout_jiffies(opts->mount_timeout));
3632 /* tear down remaining sessions */
3633 mutex_lock(&mdsc->mutex);
3634 for (i = 0; i < mdsc->max_sessions; i++) {
3635 if (mdsc->sessions[i]) {
3636 session = get_session(mdsc->sessions[i]);
3637 __unregister_session(mdsc, session);
3638 mutex_unlock(&mdsc->mutex);
3639 mutex_lock(&session->s_mutex);
3640 remove_session_caps(session);
3641 mutex_unlock(&session->s_mutex);
3642 ceph_put_mds_session(session);
3643 mutex_lock(&mdsc->mutex);
3646 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3647 mutex_unlock(&mdsc->mutex);
3649 ceph_cleanup_empty_realms(mdsc);
3651 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3656 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3658 struct ceph_mds_session *session;
3661 dout("force umount\n");
3663 mutex_lock(&mdsc->mutex);
3664 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3665 session = __ceph_lookup_mds_session(mdsc, mds);
3668 mutex_unlock(&mdsc->mutex);
3669 mutex_lock(&session->s_mutex);
3670 __close_session(mdsc, session);
3671 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3672 cleanup_session_requests(mdsc, session);
3673 remove_session_caps(session);
3675 mutex_unlock(&session->s_mutex);
3676 ceph_put_mds_session(session);
3677 mutex_lock(&mdsc->mutex);
3678 kick_requests(mdsc, mds);
3680 __wake_requests(mdsc, &mdsc->waiting_for_map);
3681 mutex_unlock(&mdsc->mutex);
3684 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3687 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3689 ceph_mdsmap_destroy(mdsc->mdsmap);
3690 kfree(mdsc->sessions);
3691 ceph_caps_finalize(mdsc);
3692 ceph_pool_perm_destroy(mdsc);
3695 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3697 struct ceph_mds_client *mdsc = fsc->mdsc;
3699 dout("mdsc_destroy %p\n", mdsc);
3700 ceph_mdsc_stop(mdsc);
3702 /* flush out any connection work with references to us */
3707 dout("mdsc_destroy %p done\n", mdsc);
3712 * handle mds map update.
3714 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3718 void *p = msg->front.iov_base;
3719 void *end = p + msg->front.iov_len;
3720 struct ceph_mdsmap *newmap, *oldmap;
3721 struct ceph_fsid fsid;
3724 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3725 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3726 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3728 epoch = ceph_decode_32(&p);
3729 maplen = ceph_decode_32(&p);
3730 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3732 /* do we need it? */
3733 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3734 mutex_lock(&mdsc->mutex);
3735 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3736 dout("handle_map epoch %u <= our %u\n",
3737 epoch, mdsc->mdsmap->m_epoch);
3738 mutex_unlock(&mdsc->mutex);
3742 newmap = ceph_mdsmap_decode(&p, end);
3743 if (IS_ERR(newmap)) {
3744 err = PTR_ERR(newmap);
3748 /* swap into place */
3750 oldmap = mdsc->mdsmap;
3751 mdsc->mdsmap = newmap;
3752 check_new_map(mdsc, newmap, oldmap);
3753 ceph_mdsmap_destroy(oldmap);
3755 mdsc->mdsmap = newmap; /* first mds map */
3757 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3759 __wake_requests(mdsc, &mdsc->waiting_for_map);
3761 mutex_unlock(&mdsc->mutex);
3762 schedule_delayed(mdsc);
3766 mutex_unlock(&mdsc->mutex);
3768 pr_err("error decoding mdsmap %d\n", err);
3772 static struct ceph_connection *con_get(struct ceph_connection *con)
3774 struct ceph_mds_session *s = con->private;
3776 if (get_session(s)) {
3777 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3780 dout("mdsc con_get %p FAIL\n", s);
3784 static void con_put(struct ceph_connection *con)
3786 struct ceph_mds_session *s = con->private;
3788 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3789 ceph_put_mds_session(s);
3793 * if the client is unresponsive for long enough, the mds will kill
3794 * the session entirely.
3796 static void peer_reset(struct ceph_connection *con)
3798 struct ceph_mds_session *s = con->private;
3799 struct ceph_mds_client *mdsc = s->s_mdsc;
3801 pr_warn("mds%d closed our session\n", s->s_mds);
3802 send_mds_reconnect(mdsc, s);
3805 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3807 struct ceph_mds_session *s = con->private;
3808 struct ceph_mds_client *mdsc = s->s_mdsc;
3809 int type = le16_to_cpu(msg->hdr.type);
3811 mutex_lock(&mdsc->mutex);
3812 if (__verify_registered_session(mdsc, s) < 0) {
3813 mutex_unlock(&mdsc->mutex);
3816 mutex_unlock(&mdsc->mutex);
3819 case CEPH_MSG_MDS_MAP:
3820 ceph_mdsc_handle_map(mdsc, msg);
3822 case CEPH_MSG_CLIENT_SESSION:
3823 handle_session(s, msg);
3825 case CEPH_MSG_CLIENT_REPLY:
3826 handle_reply(s, msg);
3828 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3829 handle_forward(mdsc, s, msg);
3831 case CEPH_MSG_CLIENT_CAPS:
3832 ceph_handle_caps(s, msg);
3834 case CEPH_MSG_CLIENT_SNAP:
3835 ceph_handle_snap(mdsc, s, msg);
3837 case CEPH_MSG_CLIENT_LEASE:
3838 handle_lease(mdsc, s, msg);
3842 pr_err("received unknown message type %d %s\n", type,
3843 ceph_msg_type_name(type));
3854 * Note: returned pointer is the address of a structure that's
3855 * managed separately. Caller must *not* attempt to free it.
3857 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3858 int *proto, int force_new)
3860 struct ceph_mds_session *s = con->private;
3861 struct ceph_mds_client *mdsc = s->s_mdsc;
3862 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3863 struct ceph_auth_handshake *auth = &s->s_auth;
3865 if (force_new && auth->authorizer) {
3866 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3867 auth->authorizer = NULL;
3869 if (!auth->authorizer) {
3870 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3873 return ERR_PTR(ret);
3875 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3878 return ERR_PTR(ret);
3880 *proto = ac->protocol;
3886 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3888 struct ceph_mds_session *s = con->private;
3889 struct ceph_mds_client *mdsc = s->s_mdsc;
3890 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3892 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3895 static int invalidate_authorizer(struct ceph_connection *con)
3897 struct ceph_mds_session *s = con->private;
3898 struct ceph_mds_client *mdsc = s->s_mdsc;
3899 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3901 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3903 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3906 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3907 struct ceph_msg_header *hdr, int *skip)
3909 struct ceph_msg *msg;
3910 int type = (int) le16_to_cpu(hdr->type);
3911 int front_len = (int) le32_to_cpu(hdr->front_len);
3917 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3919 pr_err("unable to allocate msg type %d len %d\n",
3927 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3929 struct ceph_mds_session *s = con->private;
3930 struct ceph_auth_handshake *auth = &s->s_auth;
3931 return ceph_auth_sign_message(auth, msg);
3934 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3936 struct ceph_mds_session *s = con->private;
3937 struct ceph_auth_handshake *auth = &s->s_auth;
3938 return ceph_auth_check_message_signature(auth, msg);
3941 static const struct ceph_connection_operations mds_con_ops = {
3944 .dispatch = dispatch,
3945 .get_authorizer = get_authorizer,
3946 .verify_authorizer_reply = verify_authorizer_reply,
3947 .invalidate_authorizer = invalidate_authorizer,
3948 .peer_reset = peer_reset,
3949 .alloc_msg = mds_alloc_msg,
3950 .sign_message = sign_message,
3951 .check_message_signature = check_message_signature,