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45 * are no longer valid.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);
1167 if (!list_empty(&ci->i_dirty_item)) {
1168 pr_warn_ratelimited(
1169 " dropping dirty %s state for %p %lld\n",
1170 ceph_cap_string(ci->i_dirty_caps),
1171 inode, ceph_ino(inode));
1172 ci->i_dirty_caps = 0;
1173 list_del_init(&ci->i_dirty_item);
1176 if (!list_empty(&ci->i_flushing_item)) {
1177 pr_warn_ratelimited(
1178 " dropping dirty+flushing %s state for %p %lld\n",
1179 ceph_cap_string(ci->i_flushing_caps),
1180 inode, ceph_ino(inode));
1181 ci->i_flushing_caps = 0;
1182 list_del_init(&ci->i_flushing_item);
1183 mdsc->num_cap_flushing--;
1186 spin_unlock(&mdsc->cap_dirty_lock);
1189 spin_unlock(&ci->i_ceph_lock);
1190 while (!list_empty(&to_remove)) {
1191 struct ceph_cap_flush *cf;
1192 cf = list_first_entry(&to_remove,
1193 struct ceph_cap_flush, list);
1194 list_del(&cf->list);
1203 * caller must hold session s_mutex
1205 static void remove_session_caps(struct ceph_mds_session *session)
1207 dout("remove_session_caps on %p\n", session);
1208 iterate_session_caps(session, remove_session_caps_cb, NULL);
1210 spin_lock(&session->s_cap_lock);
1211 if (session->s_nr_caps > 0) {
1212 struct super_block *sb = session->s_mdsc->fsc->sb;
1213 struct inode *inode;
1214 struct ceph_cap *cap, *prev = NULL;
1215 struct ceph_vino vino;
1217 * iterate_session_caps() skips inodes that are being
1218 * deleted, we need to wait until deletions are complete.
1219 * __wait_on_freeing_inode() is designed for the job,
1220 * but it is not exported, so use lookup inode function
1223 while (!list_empty(&session->s_caps)) {
1224 cap = list_entry(session->s_caps.next,
1225 struct ceph_cap, session_caps);
1229 vino = cap->ci->i_vino;
1230 spin_unlock(&session->s_cap_lock);
1232 inode = ceph_find_inode(sb, vino);
1235 spin_lock(&session->s_cap_lock);
1239 // drop cap expires and unlock s_cap_lock
1240 cleanup_cap_releases(session->s_mdsc, session);
1242 BUG_ON(session->s_nr_caps > 0);
1243 BUG_ON(!list_empty(&session->s_cap_flushing));
1247 * wake up any threads waiting on this session's caps. if the cap is
1248 * old (didn't get renewed on the client reconnect), remove it now.
1250 * caller must hold s_mutex.
1252 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1255 struct ceph_inode_info *ci = ceph_inode(inode);
1257 wake_up_all(&ci->i_cap_wq);
1259 spin_lock(&ci->i_ceph_lock);
1260 ci->i_wanted_max_size = 0;
1261 ci->i_requested_max_size = 0;
1262 spin_unlock(&ci->i_ceph_lock);
1267 static void wake_up_session_caps(struct ceph_mds_session *session,
1270 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1271 iterate_session_caps(session, wake_up_session_cb,
1272 (void *)(unsigned long)reconnect);
1276 * Send periodic message to MDS renewing all currently held caps. The
1277 * ack will reset the expiration for all caps from this session.
1279 * caller holds s_mutex
1281 static int send_renew_caps(struct ceph_mds_client *mdsc,
1282 struct ceph_mds_session *session)
1284 struct ceph_msg *msg;
1287 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1288 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1289 pr_info("mds%d caps stale\n", session->s_mds);
1290 session->s_renew_requested = jiffies;
1292 /* do not try to renew caps until a recovering mds has reconnected
1293 * with its clients. */
1294 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1295 if (state < CEPH_MDS_STATE_RECONNECT) {
1296 dout("send_renew_caps ignoring mds%d (%s)\n",
1297 session->s_mds, ceph_mds_state_name(state));
1301 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1302 ceph_mds_state_name(state));
1303 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1304 ++session->s_renew_seq);
1307 ceph_con_send(&session->s_con, msg);
1311 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1312 struct ceph_mds_session *session, u64 seq)
1314 struct ceph_msg *msg;
1316 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1317 session->s_mds, ceph_session_state_name(session->s_state), seq);
1318 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1321 ceph_con_send(&session->s_con, msg);
1327 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1329 * Called under session->s_mutex
1331 static void renewed_caps(struct ceph_mds_client *mdsc,
1332 struct ceph_mds_session *session, int is_renew)
1337 spin_lock(&session->s_cap_lock);
1338 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1340 session->s_cap_ttl = session->s_renew_requested +
1341 mdsc->mdsmap->m_session_timeout*HZ;
1344 if (time_before(jiffies, session->s_cap_ttl)) {
1345 pr_info("mds%d caps renewed\n", session->s_mds);
1348 pr_info("mds%d caps still stale\n", session->s_mds);
1351 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1352 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1353 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1354 spin_unlock(&session->s_cap_lock);
1357 wake_up_session_caps(session, 0);
1361 * send a session close request
1363 static int request_close_session(struct ceph_mds_client *mdsc,
1364 struct ceph_mds_session *session)
1366 struct ceph_msg *msg;
1368 dout("request_close_session mds%d state %s seq %lld\n",
1369 session->s_mds, ceph_session_state_name(session->s_state),
1371 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1374 ceph_con_send(&session->s_con, msg);
1379 * Called with s_mutex held.
1381 static int __close_session(struct ceph_mds_client *mdsc,
1382 struct ceph_mds_session *session)
1384 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1386 session->s_state = CEPH_MDS_SESSION_CLOSING;
1387 return request_close_session(mdsc, session);
1391 * Trim old(er) caps.
1393 * Because we can't cache an inode without one or more caps, we do
1394 * this indirectly: if a cap is unused, we prune its aliases, at which
1395 * point the inode will hopefully get dropped to.
1397 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1398 * memory pressure from the MDS, though, so it needn't be perfect.
1400 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1402 struct ceph_mds_session *session = arg;
1403 struct ceph_inode_info *ci = ceph_inode(inode);
1404 int used, wanted, oissued, mine;
1406 if (session->s_trim_caps <= 0)
1409 spin_lock(&ci->i_ceph_lock);
1410 mine = cap->issued | cap->implemented;
1411 used = __ceph_caps_used(ci);
1412 wanted = __ceph_caps_file_wanted(ci);
1413 oissued = __ceph_caps_issued_other(ci, cap);
1415 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1416 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1417 ceph_cap_string(used), ceph_cap_string(wanted));
1418 if (cap == ci->i_auth_cap) {
1419 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1420 !list_empty(&ci->i_cap_snaps))
1422 if ((used | wanted) & CEPH_CAP_ANY_WR)
1425 if ((used | wanted) & ~oissued & mine)
1426 goto out; /* we need these caps */
1428 session->s_trim_caps--;
1430 /* we aren't the only cap.. just remove us */
1431 __ceph_remove_cap(cap, true);
1433 /* try to drop referring dentries */
1434 spin_unlock(&ci->i_ceph_lock);
1435 d_prune_aliases(inode);
1436 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1437 inode, cap, atomic_read(&inode->i_count));
1442 spin_unlock(&ci->i_ceph_lock);
1447 * Trim session cap count down to some max number.
1449 static int trim_caps(struct ceph_mds_client *mdsc,
1450 struct ceph_mds_session *session,
1453 int trim_caps = session->s_nr_caps - max_caps;
1455 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1456 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1457 if (trim_caps > 0) {
1458 session->s_trim_caps = trim_caps;
1459 iterate_session_caps(session, trim_caps_cb, session);
1460 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1461 session->s_mds, session->s_nr_caps, max_caps,
1462 trim_caps - session->s_trim_caps);
1463 session->s_trim_caps = 0;
1466 ceph_send_cap_releases(mdsc, session);
1470 static int check_cap_flush(struct ceph_inode_info *ci,
1471 u64 want_flush_seq, u64 want_snap_seq)
1473 int ret1 = 1, ret2 = 1;
1474 spin_lock(&ci->i_ceph_lock);
1475 if (want_flush_seq > 0 && ci->i_flushing_caps)
1476 ret1 = ci->i_cap_flush_seq >= want_flush_seq;
1478 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1479 struct ceph_cap_snap *capsnap =
1480 list_first_entry(&ci->i_cap_snaps,
1481 struct ceph_cap_snap, ci_item);
1482 ret2 = capsnap->follows >= want_snap_seq;
1484 spin_unlock(&ci->i_ceph_lock);
1485 return ret1 && ret2;
1489 * flush all dirty inode data to disk.
1491 * returns true if we've flushed through want_flush_seq
1493 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1494 u64 want_flush_seq, u64 want_snap_seq)
1498 dout("check_cap_flush want %lld\n", want_flush_seq);
1499 mutex_lock(&mdsc->mutex);
1500 for (mds = 0; mds < mdsc->max_sessions; ) {
1501 struct ceph_mds_session *session = mdsc->sessions[mds];
1502 struct inode *inode1 = NULL, *inode2 = NULL;
1508 get_session(session);
1509 mutex_unlock(&mdsc->mutex);
1511 mutex_lock(&session->s_mutex);
1512 if (!list_empty(&session->s_cap_flushing)) {
1513 struct ceph_inode_info *ci =
1514 list_first_entry(&session->s_cap_flushing,
1515 struct ceph_inode_info,
1518 if (!check_cap_flush(ci, want_flush_seq, 0)) {
1519 dout("check_cap_flush still flushing %p "
1520 "seq %lld <= %lld to mds%d\n",
1521 &ci->vfs_inode, ci->i_cap_flush_seq,
1522 want_flush_seq, mds);
1523 inode1 = igrab(&ci->vfs_inode);
1526 if (!list_empty(&session->s_cap_snaps_flushing)) {
1527 struct ceph_cap_snap *capsnap =
1528 list_first_entry(&session->s_cap_snaps_flushing,
1529 struct ceph_cap_snap,
1531 struct ceph_inode_info *ci = capsnap->ci;
1532 if (!check_cap_flush(ci, 0, want_snap_seq)) {
1533 dout("check_cap_flush still flushing snap %p "
1534 "follows %lld <= %lld to mds%d\n",
1535 &ci->vfs_inode, capsnap->follows,
1536 want_snap_seq, mds);
1537 inode2 = igrab(&ci->vfs_inode);
1540 mutex_unlock(&session->s_mutex);
1541 ceph_put_mds_session(session);
1544 wait_event(mdsc->cap_flushing_wq,
1545 check_cap_flush(ceph_inode(inode1),
1546 want_flush_seq, 0));
1550 wait_event(mdsc->cap_flushing_wq,
1551 check_cap_flush(ceph_inode(inode2),
1556 if (!inode1 && !inode2)
1559 mutex_lock(&mdsc->mutex);
1562 mutex_unlock(&mdsc->mutex);
1563 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1567 * called under s_mutex
1569 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1570 struct ceph_mds_session *session)
1572 struct ceph_msg *msg = NULL;
1573 struct ceph_mds_cap_release *head;
1574 struct ceph_mds_cap_item *item;
1575 struct ceph_cap *cap;
1576 LIST_HEAD(tmp_list);
1577 int num_cap_releases;
1579 spin_lock(&session->s_cap_lock);
1581 list_splice_init(&session->s_cap_releases, &tmp_list);
1582 num_cap_releases = session->s_num_cap_releases;
1583 session->s_num_cap_releases = 0;
1584 spin_unlock(&session->s_cap_lock);
1586 while (!list_empty(&tmp_list)) {
1588 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1589 PAGE_CACHE_SIZE, GFP_NOFS, false);
1592 head = msg->front.iov_base;
1593 head->num = cpu_to_le32(0);
1594 msg->front.iov_len = sizeof(*head);
1596 cap = list_first_entry(&tmp_list, struct ceph_cap,
1598 list_del(&cap->session_caps);
1601 head = msg->front.iov_base;
1602 le32_add_cpu(&head->num, 1);
1603 item = msg->front.iov_base + msg->front.iov_len;
1604 item->ino = cpu_to_le64(cap->cap_ino);
1605 item->cap_id = cpu_to_le64(cap->cap_id);
1606 item->migrate_seq = cpu_to_le32(cap->mseq);
1607 item->seq = cpu_to_le32(cap->issue_seq);
1608 msg->front.iov_len += sizeof(*item);
1610 ceph_put_cap(mdsc, cap);
1612 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1613 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1614 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1615 ceph_con_send(&session->s_con, msg);
1620 BUG_ON(num_cap_releases != 0);
1622 spin_lock(&session->s_cap_lock);
1623 if (!list_empty(&session->s_cap_releases))
1625 spin_unlock(&session->s_cap_lock);
1628 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1629 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1630 ceph_con_send(&session->s_con, msg);
1634 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1636 spin_lock(&session->s_cap_lock);
1637 list_splice(&tmp_list, &session->s_cap_releases);
1638 session->s_num_cap_releases += num_cap_releases;
1639 spin_unlock(&session->s_cap_lock);
1646 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1649 struct ceph_inode_info *ci = ceph_inode(dir);
1650 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1651 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1652 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1653 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1654 int order, num_entries;
1656 spin_lock(&ci->i_ceph_lock);
1657 num_entries = ci->i_files + ci->i_subdirs;
1658 spin_unlock(&ci->i_ceph_lock);
1659 num_entries = max(num_entries, 1);
1660 num_entries = min(num_entries, opt->max_readdir);
1662 order = get_order(size * num_entries);
1663 while (order >= 0) {
1664 rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN,
1673 num_entries = (PAGE_SIZE << order) / size;
1674 num_entries = min(num_entries, opt->max_readdir);
1676 rinfo->dir_buf_size = PAGE_SIZE << order;
1677 req->r_num_caps = num_entries + 1;
1678 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1679 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1684 * Create an mds request.
1686 struct ceph_mds_request *
1687 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1689 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1692 return ERR_PTR(-ENOMEM);
1694 mutex_init(&req->r_fill_mutex);
1696 req->r_started = jiffies;
1697 req->r_resend_mds = -1;
1698 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1700 kref_init(&req->r_kref);
1701 INIT_LIST_HEAD(&req->r_wait);
1702 init_completion(&req->r_completion);
1703 init_completion(&req->r_safe_completion);
1704 INIT_LIST_HEAD(&req->r_unsafe_item);
1706 req->r_stamp = CURRENT_TIME;
1709 req->r_direct_mode = mode;
1714 * return oldest (lowest) request, tid in request tree, 0 if none.
1716 * called under mdsc->mutex.
1718 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1720 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1722 return rb_entry(rb_first(&mdsc->request_tree),
1723 struct ceph_mds_request, r_node);
1726 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1728 return mdsc->oldest_tid;
1732 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1733 * on build_path_from_dentry in fs/cifs/dir.c.
1735 * If @stop_on_nosnap, generate path relative to the first non-snapped
1738 * Encode hidden .snap dirs as a double /, i.e.
1739 * foo/.snap/bar -> foo//bar
1741 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1744 struct dentry *temp;
1750 return ERR_PTR(-EINVAL);
1754 seq = read_seqbegin(&rename_lock);
1756 for (temp = dentry; !IS_ROOT(temp);) {
1757 struct inode *inode = d_inode(temp);
1758 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1759 len++; /* slash only */
1760 else if (stop_on_nosnap && inode &&
1761 ceph_snap(inode) == CEPH_NOSNAP)
1764 len += 1 + temp->d_name.len;
1765 temp = temp->d_parent;
1769 len--; /* no leading '/' */
1771 path = kmalloc(len+1, GFP_NOFS);
1773 return ERR_PTR(-ENOMEM);
1775 path[pos] = 0; /* trailing null */
1777 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1778 struct inode *inode;
1780 spin_lock(&temp->d_lock);
1781 inode = d_inode(temp);
1782 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1783 dout("build_path path+%d: %p SNAPDIR\n",
1785 } else if (stop_on_nosnap && inode &&
1786 ceph_snap(inode) == CEPH_NOSNAP) {
1787 spin_unlock(&temp->d_lock);
1790 pos -= temp->d_name.len;
1792 spin_unlock(&temp->d_lock);
1795 strncpy(path + pos, temp->d_name.name,
1798 spin_unlock(&temp->d_lock);
1801 temp = temp->d_parent;
1804 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1805 pr_err("build_path did not end path lookup where "
1806 "expected, namelen is %d, pos is %d\n", len, pos);
1807 /* presumably this is only possible if racing with a
1808 rename of one of the parent directories (we can not
1809 lock the dentries above us to prevent this, but
1810 retrying should be harmless) */
1815 *base = ceph_ino(d_inode(temp));
1817 dout("build_path on %p %d built %llx '%.*s'\n",
1818 dentry, d_count(dentry), *base, len, path);
1822 static int build_dentry_path(struct dentry *dentry,
1823 const char **ppath, int *ppathlen, u64 *pino,
1828 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1829 *pino = ceph_ino(d_inode(dentry->d_parent));
1830 *ppath = dentry->d_name.name;
1831 *ppathlen = dentry->d_name.len;
1834 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1836 return PTR_ERR(path);
1842 static int build_inode_path(struct inode *inode,
1843 const char **ppath, int *ppathlen, u64 *pino,
1846 struct dentry *dentry;
1849 if (ceph_snap(inode) == CEPH_NOSNAP) {
1850 *pino = ceph_ino(inode);
1854 dentry = d_find_alias(inode);
1855 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1858 return PTR_ERR(path);
1865 * request arguments may be specified via an inode *, a dentry *, or
1866 * an explicit ino+path.
1868 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1869 const char *rpath, u64 rino,
1870 const char **ppath, int *pathlen,
1871 u64 *ino, int *freepath)
1876 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1877 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1879 } else if (rdentry) {
1880 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1881 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1883 } else if (rpath || rino) {
1886 *pathlen = rpath ? strlen(rpath) : 0;
1887 dout(" path %.*s\n", *pathlen, rpath);
1894 * called under mdsc->mutex
1896 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1897 struct ceph_mds_request *req,
1898 int mds, bool drop_cap_releases)
1900 struct ceph_msg *msg;
1901 struct ceph_mds_request_head *head;
1902 const char *path1 = NULL;
1903 const char *path2 = NULL;
1904 u64 ino1 = 0, ino2 = 0;
1905 int pathlen1 = 0, pathlen2 = 0;
1906 int freepath1 = 0, freepath2 = 0;
1912 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1913 req->r_path1, req->r_ino1.ino,
1914 &path1, &pathlen1, &ino1, &freepath1);
1920 ret = set_request_path_attr(NULL, req->r_old_dentry,
1921 req->r_path2, req->r_ino2.ino,
1922 &path2, &pathlen2, &ino2, &freepath2);
1928 len = sizeof(*head) +
1929 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1930 sizeof(struct timespec);
1932 /* calculate (max) length for cap releases */
1933 len += sizeof(struct ceph_mds_request_release) *
1934 (!!req->r_inode_drop + !!req->r_dentry_drop +
1935 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1936 if (req->r_dentry_drop)
1937 len += req->r_dentry->d_name.len;
1938 if (req->r_old_dentry_drop)
1939 len += req->r_old_dentry->d_name.len;
1941 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1943 msg = ERR_PTR(-ENOMEM);
1947 msg->hdr.version = cpu_to_le16(2);
1948 msg->hdr.tid = cpu_to_le64(req->r_tid);
1950 head = msg->front.iov_base;
1951 p = msg->front.iov_base + sizeof(*head);
1952 end = msg->front.iov_base + msg->front.iov_len;
1954 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1955 head->op = cpu_to_le32(req->r_op);
1956 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1957 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1958 head->args = req->r_args;
1960 ceph_encode_filepath(&p, end, ino1, path1);
1961 ceph_encode_filepath(&p, end, ino2, path2);
1963 /* make note of release offset, in case we need to replay */
1964 req->r_request_release_offset = p - msg->front.iov_base;
1968 if (req->r_inode_drop)
1969 releases += ceph_encode_inode_release(&p,
1970 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1971 mds, req->r_inode_drop, req->r_inode_unless, 0);
1972 if (req->r_dentry_drop)
1973 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1974 mds, req->r_dentry_drop, req->r_dentry_unless);
1975 if (req->r_old_dentry_drop)
1976 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1977 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1978 if (req->r_old_inode_drop)
1979 releases += ceph_encode_inode_release(&p,
1980 d_inode(req->r_old_dentry),
1981 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1983 if (drop_cap_releases) {
1985 p = msg->front.iov_base + req->r_request_release_offset;
1988 head->num_releases = cpu_to_le16(releases);
1992 struct ceph_timespec ts;
1993 ceph_encode_timespec(&ts, &req->r_stamp);
1994 ceph_encode_copy(&p, &ts, sizeof(ts));
1998 msg->front.iov_len = p - msg->front.iov_base;
1999 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2001 if (req->r_pagelist) {
2002 struct ceph_pagelist *pagelist = req->r_pagelist;
2003 atomic_inc(&pagelist->refcnt);
2004 ceph_msg_data_add_pagelist(msg, pagelist);
2005 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2007 msg->hdr.data_len = 0;
2010 msg->hdr.data_off = cpu_to_le16(0);
2014 kfree((char *)path2);
2017 kfree((char *)path1);
2023 * called under mdsc->mutex if error, under no mutex if
2026 static void complete_request(struct ceph_mds_client *mdsc,
2027 struct ceph_mds_request *req)
2029 if (req->r_callback)
2030 req->r_callback(mdsc, req);
2032 complete_all(&req->r_completion);
2036 * called under mdsc->mutex
2038 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2039 struct ceph_mds_request *req,
2040 int mds, bool drop_cap_releases)
2042 struct ceph_mds_request_head *rhead;
2043 struct ceph_msg *msg;
2048 struct ceph_cap *cap =
2049 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2052 req->r_sent_on_mseq = cap->mseq;
2054 req->r_sent_on_mseq = -1;
2056 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2057 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2059 if (req->r_got_unsafe) {
2062 * Replay. Do not regenerate message (and rebuild
2063 * paths, etc.); just use the original message.
2064 * Rebuilding paths will break for renames because
2065 * d_move mangles the src name.
2067 msg = req->r_request;
2068 rhead = msg->front.iov_base;
2070 flags = le32_to_cpu(rhead->flags);
2071 flags |= CEPH_MDS_FLAG_REPLAY;
2072 rhead->flags = cpu_to_le32(flags);
2074 if (req->r_target_inode)
2075 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2077 rhead->num_retry = req->r_attempts - 1;
2079 /* remove cap/dentry releases from message */
2080 rhead->num_releases = 0;
2083 p = msg->front.iov_base + req->r_request_release_offset;
2085 struct ceph_timespec ts;
2086 ceph_encode_timespec(&ts, &req->r_stamp);
2087 ceph_encode_copy(&p, &ts, sizeof(ts));
2090 msg->front.iov_len = p - msg->front.iov_base;
2091 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2095 if (req->r_request) {
2096 ceph_msg_put(req->r_request);
2097 req->r_request = NULL;
2099 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2101 req->r_err = PTR_ERR(msg);
2102 complete_request(mdsc, req);
2103 return PTR_ERR(msg);
2105 req->r_request = msg;
2107 rhead = msg->front.iov_base;
2108 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2109 if (req->r_got_unsafe)
2110 flags |= CEPH_MDS_FLAG_REPLAY;
2111 if (req->r_locked_dir)
2112 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2113 rhead->flags = cpu_to_le32(flags);
2114 rhead->num_fwd = req->r_num_fwd;
2115 rhead->num_retry = req->r_attempts - 1;
2118 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2123 * send request, or put it on the appropriate wait list.
2125 static int __do_request(struct ceph_mds_client *mdsc,
2126 struct ceph_mds_request *req)
2128 struct ceph_mds_session *session = NULL;
2132 if (req->r_err || req->r_got_result) {
2134 __unregister_request(mdsc, req);
2138 if (req->r_timeout &&
2139 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2140 dout("do_request timed out\n");
2145 put_request_session(req);
2147 mds = __choose_mds(mdsc, req);
2149 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2150 dout("do_request no mds or not active, waiting for map\n");
2151 list_add(&req->r_wait, &mdsc->waiting_for_map);
2155 /* get, open session */
2156 session = __ceph_lookup_mds_session(mdsc, mds);
2158 session = register_session(mdsc, mds);
2159 if (IS_ERR(session)) {
2160 err = PTR_ERR(session);
2164 req->r_session = get_session(session);
2166 dout("do_request mds%d session %p state %s\n", mds, session,
2167 ceph_session_state_name(session->s_state));
2168 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2169 session->s_state != CEPH_MDS_SESSION_HUNG) {
2170 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2171 session->s_state == CEPH_MDS_SESSION_CLOSING)
2172 __open_session(mdsc, session);
2173 list_add(&req->r_wait, &session->s_waiting);
2178 req->r_resend_mds = -1; /* forget any previous mds hint */
2180 if (req->r_request_started == 0) /* note request start time */
2181 req->r_request_started = jiffies;
2183 err = __prepare_send_request(mdsc, req, mds, false);
2185 ceph_msg_get(req->r_request);
2186 ceph_con_send(&session->s_con, req->r_request);
2190 ceph_put_mds_session(session);
2196 complete_request(mdsc, req);
2201 * called under mdsc->mutex
2203 static void __wake_requests(struct ceph_mds_client *mdsc,
2204 struct list_head *head)
2206 struct ceph_mds_request *req;
2207 LIST_HEAD(tmp_list);
2209 list_splice_init(head, &tmp_list);
2211 while (!list_empty(&tmp_list)) {
2212 req = list_entry(tmp_list.next,
2213 struct ceph_mds_request, r_wait);
2214 list_del_init(&req->r_wait);
2215 dout(" wake request %p tid %llu\n", req, req->r_tid);
2216 __do_request(mdsc, req);
2221 * Wake up threads with requests pending for @mds, so that they can
2222 * resubmit their requests to a possibly different mds.
2224 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2226 struct ceph_mds_request *req;
2227 struct rb_node *p = rb_first(&mdsc->request_tree);
2229 dout("kick_requests mds%d\n", mds);
2231 req = rb_entry(p, struct ceph_mds_request, r_node);
2233 if (req->r_got_unsafe)
2235 if (req->r_attempts > 0)
2236 continue; /* only new requests */
2237 if (req->r_session &&
2238 req->r_session->s_mds == mds) {
2239 dout(" kicking tid %llu\n", req->r_tid);
2240 list_del_init(&req->r_wait);
2241 __do_request(mdsc, req);
2246 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2247 struct ceph_mds_request *req)
2249 dout("submit_request on %p\n", req);
2250 mutex_lock(&mdsc->mutex);
2251 __register_request(mdsc, req, NULL);
2252 __do_request(mdsc, req);
2253 mutex_unlock(&mdsc->mutex);
2257 * Synchrously perform an mds request. Take care of all of the
2258 * session setup, forwarding, retry details.
2260 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2262 struct ceph_mds_request *req)
2266 dout("do_request on %p\n", req);
2268 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2270 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2271 if (req->r_locked_dir)
2272 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2273 if (req->r_old_dentry_dir)
2274 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2278 mutex_lock(&mdsc->mutex);
2279 __register_request(mdsc, req, dir);
2280 __do_request(mdsc, req);
2284 __unregister_request(mdsc, req);
2285 dout("do_request early error %d\n", err);
2290 mutex_unlock(&mdsc->mutex);
2291 dout("do_request waiting\n");
2292 if (!req->r_timeout && req->r_wait_for_completion) {
2293 err = req->r_wait_for_completion(mdsc, req);
2295 long timeleft = wait_for_completion_killable_timeout(
2297 ceph_timeout_jiffies(req->r_timeout));
2301 err = -EIO; /* timed out */
2303 err = timeleft; /* killed */
2305 dout("do_request waited, got %d\n", err);
2306 mutex_lock(&mdsc->mutex);
2308 /* only abort if we didn't race with a real reply */
2309 if (req->r_got_result) {
2310 err = le32_to_cpu(req->r_reply_info.head->result);
2311 } else if (err < 0) {
2312 dout("aborted request %lld with %d\n", req->r_tid, err);
2315 * ensure we aren't running concurrently with
2316 * ceph_fill_trace or ceph_readdir_prepopulate, which
2317 * rely on locks (dir mutex) held by our caller.
2319 mutex_lock(&req->r_fill_mutex);
2321 req->r_aborted = true;
2322 mutex_unlock(&req->r_fill_mutex);
2324 if (req->r_locked_dir &&
2325 (req->r_op & CEPH_MDS_OP_WRITE))
2326 ceph_invalidate_dir_request(req);
2332 mutex_unlock(&mdsc->mutex);
2333 dout("do_request %p done, result %d\n", req, err);
2338 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2339 * namespace request.
2341 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2343 struct inode *inode = req->r_locked_dir;
2345 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2347 ceph_dir_clear_complete(inode);
2349 ceph_invalidate_dentry_lease(req->r_dentry);
2350 if (req->r_old_dentry)
2351 ceph_invalidate_dentry_lease(req->r_old_dentry);
2357 * We take the session mutex and parse and process the reply immediately.
2358 * This preserves the logical ordering of replies, capabilities, etc., sent
2359 * by the MDS as they are applied to our local cache.
2361 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2363 struct ceph_mds_client *mdsc = session->s_mdsc;
2364 struct ceph_mds_request *req;
2365 struct ceph_mds_reply_head *head = msg->front.iov_base;
2366 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2367 struct ceph_snap_realm *realm;
2370 int mds = session->s_mds;
2372 if (msg->front.iov_len < sizeof(*head)) {
2373 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2378 /* get request, session */
2379 tid = le64_to_cpu(msg->hdr.tid);
2380 mutex_lock(&mdsc->mutex);
2381 req = __lookup_request(mdsc, tid);
2383 dout("handle_reply on unknown tid %llu\n", tid);
2384 mutex_unlock(&mdsc->mutex);
2387 dout("handle_reply %p\n", req);
2389 /* correct session? */
2390 if (req->r_session != session) {
2391 pr_err("mdsc_handle_reply got %llu on session mds%d"
2392 " not mds%d\n", tid, session->s_mds,
2393 req->r_session ? req->r_session->s_mds : -1);
2394 mutex_unlock(&mdsc->mutex);
2399 if ((req->r_got_unsafe && !head->safe) ||
2400 (req->r_got_safe && head->safe)) {
2401 pr_warn("got a dup %s reply on %llu from mds%d\n",
2402 head->safe ? "safe" : "unsafe", tid, mds);
2403 mutex_unlock(&mdsc->mutex);
2406 if (req->r_got_safe && !head->safe) {
2407 pr_warn("got unsafe after safe on %llu from mds%d\n",
2409 mutex_unlock(&mdsc->mutex);
2413 result = le32_to_cpu(head->result);
2417 * if we're not talking to the authority, send to them
2418 * if the authority has changed while we weren't looking,
2419 * send to new authority
2420 * Otherwise we just have to return an ESTALE
2422 if (result == -ESTALE) {
2423 dout("got ESTALE on request %llu", req->r_tid);
2424 req->r_resend_mds = -1;
2425 if (req->r_direct_mode != USE_AUTH_MDS) {
2426 dout("not using auth, setting for that now");
2427 req->r_direct_mode = USE_AUTH_MDS;
2428 __do_request(mdsc, req);
2429 mutex_unlock(&mdsc->mutex);
2432 int mds = __choose_mds(mdsc, req);
2433 if (mds >= 0 && mds != req->r_session->s_mds) {
2434 dout("but auth changed, so resending");
2435 __do_request(mdsc, req);
2436 mutex_unlock(&mdsc->mutex);
2440 dout("have to return ESTALE on request %llu", req->r_tid);
2445 req->r_got_safe = true;
2446 __unregister_request(mdsc, req);
2448 if (req->r_got_unsafe) {
2450 * We already handled the unsafe response, now do the
2451 * cleanup. No need to examine the response; the MDS
2452 * doesn't include any result info in the safe
2453 * response. And even if it did, there is nothing
2454 * useful we could do with a revised return value.
2456 dout("got safe reply %llu, mds%d\n", tid, mds);
2457 list_del_init(&req->r_unsafe_item);
2459 /* last unsafe request during umount? */
2460 if (mdsc->stopping && !__get_oldest_req(mdsc))
2461 complete_all(&mdsc->safe_umount_waiters);
2462 mutex_unlock(&mdsc->mutex);
2466 req->r_got_unsafe = true;
2467 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2470 dout("handle_reply tid %lld result %d\n", tid, result);
2471 rinfo = &req->r_reply_info;
2472 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2473 mutex_unlock(&mdsc->mutex);
2475 mutex_lock(&session->s_mutex);
2477 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2484 if (rinfo->snapblob_len) {
2485 down_write(&mdsc->snap_rwsem);
2486 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2487 rinfo->snapblob + rinfo->snapblob_len,
2488 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2490 downgrade_write(&mdsc->snap_rwsem);
2492 down_read(&mdsc->snap_rwsem);
2495 /* insert trace into our cache */
2496 mutex_lock(&req->r_fill_mutex);
2497 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2499 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2500 req->r_op == CEPH_MDS_OP_LSSNAP))
2501 ceph_readdir_prepopulate(req, req->r_session);
2502 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2504 mutex_unlock(&req->r_fill_mutex);
2506 up_read(&mdsc->snap_rwsem);
2508 ceph_put_snap_realm(mdsc, realm);
2510 mutex_lock(&mdsc->mutex);
2511 if (!req->r_aborted) {
2517 req->r_got_result = true;
2520 dout("reply arrived after request %lld was aborted\n", tid);
2522 mutex_unlock(&mdsc->mutex);
2524 mutex_unlock(&session->s_mutex);
2526 /* kick calling process */
2527 complete_request(mdsc, req);
2529 ceph_mdsc_put_request(req);
2536 * handle mds notification that our request has been forwarded.
2538 static void handle_forward(struct ceph_mds_client *mdsc,
2539 struct ceph_mds_session *session,
2540 struct ceph_msg *msg)
2542 struct ceph_mds_request *req;
2543 u64 tid = le64_to_cpu(msg->hdr.tid);
2547 void *p = msg->front.iov_base;
2548 void *end = p + msg->front.iov_len;
2550 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2551 next_mds = ceph_decode_32(&p);
2552 fwd_seq = ceph_decode_32(&p);
2554 mutex_lock(&mdsc->mutex);
2555 req = __lookup_request(mdsc, tid);
2557 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2558 goto out; /* dup reply? */
2561 if (req->r_aborted) {
2562 dout("forward tid %llu aborted, unregistering\n", tid);
2563 __unregister_request(mdsc, req);
2564 } else if (fwd_seq <= req->r_num_fwd) {
2565 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2566 tid, next_mds, req->r_num_fwd, fwd_seq);
2568 /* resend. forward race not possible; mds would drop */
2569 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2571 BUG_ON(req->r_got_result);
2572 req->r_attempts = 0;
2573 req->r_num_fwd = fwd_seq;
2574 req->r_resend_mds = next_mds;
2575 put_request_session(req);
2576 __do_request(mdsc, req);
2578 ceph_mdsc_put_request(req);
2580 mutex_unlock(&mdsc->mutex);
2584 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2588 * handle a mds session control message
2590 static void handle_session(struct ceph_mds_session *session,
2591 struct ceph_msg *msg)
2593 struct ceph_mds_client *mdsc = session->s_mdsc;
2596 int mds = session->s_mds;
2597 struct ceph_mds_session_head *h = msg->front.iov_base;
2601 if (msg->front.iov_len != sizeof(*h))
2603 op = le32_to_cpu(h->op);
2604 seq = le64_to_cpu(h->seq);
2606 mutex_lock(&mdsc->mutex);
2607 if (op == CEPH_SESSION_CLOSE)
2608 __unregister_session(mdsc, session);
2609 /* FIXME: this ttl calculation is generous */
2610 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2611 mutex_unlock(&mdsc->mutex);
2613 mutex_lock(&session->s_mutex);
2615 dout("handle_session mds%d %s %p state %s seq %llu\n",
2616 mds, ceph_session_op_name(op), session,
2617 ceph_session_state_name(session->s_state), seq);
2619 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2620 session->s_state = CEPH_MDS_SESSION_OPEN;
2621 pr_info("mds%d came back\n", session->s_mds);
2625 case CEPH_SESSION_OPEN:
2626 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2627 pr_info("mds%d reconnect success\n", session->s_mds);
2628 session->s_state = CEPH_MDS_SESSION_OPEN;
2629 renewed_caps(mdsc, session, 0);
2632 __close_session(mdsc, session);
2635 case CEPH_SESSION_RENEWCAPS:
2636 if (session->s_renew_seq == seq)
2637 renewed_caps(mdsc, session, 1);
2640 case CEPH_SESSION_CLOSE:
2641 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2642 pr_info("mds%d reconnect denied\n", session->s_mds);
2643 cleanup_session_requests(mdsc, session);
2644 remove_session_caps(session);
2645 wake = 2; /* for good measure */
2646 wake_up_all(&mdsc->session_close_wq);
2649 case CEPH_SESSION_STALE:
2650 pr_info("mds%d caps went stale, renewing\n",
2652 spin_lock(&session->s_gen_ttl_lock);
2653 session->s_cap_gen++;
2654 session->s_cap_ttl = jiffies - 1;
2655 spin_unlock(&session->s_gen_ttl_lock);
2656 send_renew_caps(mdsc, session);
2659 case CEPH_SESSION_RECALL_STATE:
2660 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2663 case CEPH_SESSION_FLUSHMSG:
2664 send_flushmsg_ack(mdsc, session, seq);
2667 case CEPH_SESSION_FORCE_RO:
2668 dout("force_session_readonly %p\n", session);
2669 spin_lock(&session->s_cap_lock);
2670 session->s_readonly = true;
2671 spin_unlock(&session->s_cap_lock);
2672 wake_up_session_caps(session, 0);
2676 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2680 mutex_unlock(&session->s_mutex);
2682 mutex_lock(&mdsc->mutex);
2683 __wake_requests(mdsc, &session->s_waiting);
2685 kick_requests(mdsc, mds);
2686 mutex_unlock(&mdsc->mutex);
2691 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2692 (int)msg->front.iov_len);
2699 * called under session->mutex.
2701 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2702 struct ceph_mds_session *session)
2704 struct ceph_mds_request *req, *nreq;
2708 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2710 mutex_lock(&mdsc->mutex);
2711 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2712 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2714 ceph_msg_get(req->r_request);
2715 ceph_con_send(&session->s_con, req->r_request);
2720 * also re-send old requests when MDS enters reconnect stage. So that MDS
2721 * can process completed request in clientreplay stage.
2723 p = rb_first(&mdsc->request_tree);
2725 req = rb_entry(p, struct ceph_mds_request, r_node);
2727 if (req->r_got_unsafe)
2729 if (req->r_attempts == 0)
2730 continue; /* only old requests */
2731 if (req->r_session &&
2732 req->r_session->s_mds == session->s_mds) {
2733 err = __prepare_send_request(mdsc, req,
2734 session->s_mds, true);
2736 ceph_msg_get(req->r_request);
2737 ceph_con_send(&session->s_con, req->r_request);
2741 mutex_unlock(&mdsc->mutex);
2745 * Encode information about a cap for a reconnect with the MDS.
2747 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2751 struct ceph_mds_cap_reconnect v2;
2752 struct ceph_mds_cap_reconnect_v1 v1;
2755 struct ceph_inode_info *ci;
2756 struct ceph_reconnect_state *recon_state = arg;
2757 struct ceph_pagelist *pagelist = recon_state->pagelist;
2761 struct dentry *dentry;
2765 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2766 inode, ceph_vinop(inode), cap, cap->cap_id,
2767 ceph_cap_string(cap->issued));
2768 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2772 dentry = d_find_alias(inode);
2774 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2776 err = PTR_ERR(path);
2783 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2787 spin_lock(&ci->i_ceph_lock);
2788 cap->seq = 0; /* reset cap seq */
2789 cap->issue_seq = 0; /* and issue_seq */
2790 cap->mseq = 0; /* and migrate_seq */
2791 cap->cap_gen = cap->session->s_cap_gen;
2793 if (recon_state->flock) {
2794 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2795 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2796 rec.v2.issued = cpu_to_le32(cap->issued);
2797 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2798 rec.v2.pathbase = cpu_to_le64(pathbase);
2799 rec.v2.flock_len = 0;
2800 reclen = sizeof(rec.v2);
2802 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2803 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2804 rec.v1.issued = cpu_to_le32(cap->issued);
2805 rec.v1.size = cpu_to_le64(inode->i_size);
2806 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2807 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2808 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2809 rec.v1.pathbase = cpu_to_le64(pathbase);
2810 reclen = sizeof(rec.v1);
2812 spin_unlock(&ci->i_ceph_lock);
2814 if (recon_state->flock) {
2815 int num_fcntl_locks, num_flock_locks;
2816 struct ceph_filelock *flocks;
2819 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2820 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2821 sizeof(struct ceph_filelock), GFP_NOFS);
2826 err = ceph_encode_locks_to_buffer(inode, flocks,
2836 * number of encoded locks is stable, so copy to pagelist
2838 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2839 (num_fcntl_locks+num_flock_locks) *
2840 sizeof(struct ceph_filelock));
2841 err = ceph_pagelist_append(pagelist, &rec, reclen);
2843 err = ceph_locks_to_pagelist(flocks, pagelist,
2848 err = ceph_pagelist_append(pagelist, &rec, reclen);
2851 recon_state->nr_caps++;
2861 * If an MDS fails and recovers, clients need to reconnect in order to
2862 * reestablish shared state. This includes all caps issued through
2863 * this session _and_ the snap_realm hierarchy. Because it's not
2864 * clear which snap realms the mds cares about, we send everything we
2865 * know about.. that ensures we'll then get any new info the
2866 * recovering MDS might have.
2868 * This is a relatively heavyweight operation, but it's rare.
2870 * called with mdsc->mutex held.
2872 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2873 struct ceph_mds_session *session)
2875 struct ceph_msg *reply;
2877 int mds = session->s_mds;
2880 struct ceph_pagelist *pagelist;
2881 struct ceph_reconnect_state recon_state;
2883 pr_info("mds%d reconnect start\n", mds);
2885 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2887 goto fail_nopagelist;
2888 ceph_pagelist_init(pagelist);
2890 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2894 mutex_lock(&session->s_mutex);
2895 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2898 dout("session %p state %s\n", session,
2899 ceph_session_state_name(session->s_state));
2901 spin_lock(&session->s_gen_ttl_lock);
2902 session->s_cap_gen++;
2903 spin_unlock(&session->s_gen_ttl_lock);
2905 spin_lock(&session->s_cap_lock);
2906 /* don't know if session is readonly */
2907 session->s_readonly = 0;
2909 * notify __ceph_remove_cap() that we are composing cap reconnect.
2910 * If a cap get released before being added to the cap reconnect,
2911 * __ceph_remove_cap() should skip queuing cap release.
2913 session->s_cap_reconnect = 1;
2914 /* drop old cap expires; we're about to reestablish that state */
2915 cleanup_cap_releases(mdsc, session);
2917 /* trim unused caps to reduce MDS's cache rejoin time */
2918 if (mdsc->fsc->sb->s_root)
2919 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2921 ceph_con_close(&session->s_con);
2922 ceph_con_open(&session->s_con,
2923 CEPH_ENTITY_TYPE_MDS, mds,
2924 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2926 /* replay unsafe requests */
2927 replay_unsafe_requests(mdsc, session);
2929 down_read(&mdsc->snap_rwsem);
2931 /* traverse this session's caps */
2932 s_nr_caps = session->s_nr_caps;
2933 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2937 recon_state.nr_caps = 0;
2938 recon_state.pagelist = pagelist;
2939 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2940 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2944 spin_lock(&session->s_cap_lock);
2945 session->s_cap_reconnect = 0;
2946 spin_unlock(&session->s_cap_lock);
2949 * snaprealms. we provide mds with the ino, seq (version), and
2950 * parent for all of our realms. If the mds has any newer info,
2953 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2954 struct ceph_snap_realm *realm =
2955 rb_entry(p, struct ceph_snap_realm, node);
2956 struct ceph_mds_snaprealm_reconnect sr_rec;
2958 dout(" adding snap realm %llx seq %lld parent %llx\n",
2959 realm->ino, realm->seq, realm->parent_ino);
2960 sr_rec.ino = cpu_to_le64(realm->ino);
2961 sr_rec.seq = cpu_to_le64(realm->seq);
2962 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2963 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2968 if (recon_state.flock)
2969 reply->hdr.version = cpu_to_le16(2);
2971 /* raced with cap release? */
2972 if (s_nr_caps != recon_state.nr_caps) {
2973 struct page *page = list_first_entry(&pagelist->head,
2975 __le32 *addr = kmap_atomic(page);
2976 *addr = cpu_to_le32(recon_state.nr_caps);
2977 kunmap_atomic(addr);
2980 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2981 ceph_msg_data_add_pagelist(reply, pagelist);
2982 ceph_con_send(&session->s_con, reply);
2984 mutex_unlock(&session->s_mutex);
2986 mutex_lock(&mdsc->mutex);
2987 __wake_requests(mdsc, &session->s_waiting);
2988 mutex_unlock(&mdsc->mutex);
2990 up_read(&mdsc->snap_rwsem);
2994 ceph_msg_put(reply);
2995 up_read(&mdsc->snap_rwsem);
2996 mutex_unlock(&session->s_mutex);
2998 ceph_pagelist_release(pagelist);
3000 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3006 * compare old and new mdsmaps, kicking requests
3007 * and closing out old connections as necessary
3009 * called under mdsc->mutex.
3011 static void check_new_map(struct ceph_mds_client *mdsc,
3012 struct ceph_mdsmap *newmap,
3013 struct ceph_mdsmap *oldmap)
3016 int oldstate, newstate;
3017 struct ceph_mds_session *s;
3019 dout("check_new_map new %u old %u\n",
3020 newmap->m_epoch, oldmap->m_epoch);
3022 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3023 if (mdsc->sessions[i] == NULL)
3025 s = mdsc->sessions[i];
3026 oldstate = ceph_mdsmap_get_state(oldmap, i);
3027 newstate = ceph_mdsmap_get_state(newmap, i);
3029 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3030 i, ceph_mds_state_name(oldstate),
3031 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3032 ceph_mds_state_name(newstate),
3033 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3034 ceph_session_state_name(s->s_state));
3036 if (i >= newmap->m_max_mds ||
3037 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3038 ceph_mdsmap_get_addr(newmap, i),
3039 sizeof(struct ceph_entity_addr))) {
3040 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3041 /* the session never opened, just close it
3043 __wake_requests(mdsc, &s->s_waiting);
3044 __unregister_session(mdsc, s);
3047 mutex_unlock(&mdsc->mutex);
3048 mutex_lock(&s->s_mutex);
3049 mutex_lock(&mdsc->mutex);
3050 ceph_con_close(&s->s_con);
3051 mutex_unlock(&s->s_mutex);
3052 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3054 } else if (oldstate == newstate) {
3055 continue; /* nothing new with this mds */
3061 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3062 newstate >= CEPH_MDS_STATE_RECONNECT) {
3063 mutex_unlock(&mdsc->mutex);
3064 send_mds_reconnect(mdsc, s);
3065 mutex_lock(&mdsc->mutex);
3069 * kick request on any mds that has gone active.
3071 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3072 newstate >= CEPH_MDS_STATE_ACTIVE) {
3073 if (oldstate != CEPH_MDS_STATE_CREATING &&
3074 oldstate != CEPH_MDS_STATE_STARTING)
3075 pr_info("mds%d recovery completed\n", s->s_mds);
3076 kick_requests(mdsc, i);
3077 ceph_kick_flushing_caps(mdsc, s);
3078 wake_up_session_caps(s, 1);
3082 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3083 s = mdsc->sessions[i];
3086 if (!ceph_mdsmap_is_laggy(newmap, i))
3088 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3089 s->s_state == CEPH_MDS_SESSION_HUNG ||
3090 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3091 dout(" connecting to export targets of laggy mds%d\n",
3093 __open_export_target_sessions(mdsc, s);
3105 * caller must hold session s_mutex, dentry->d_lock
3107 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3109 struct ceph_dentry_info *di = ceph_dentry(dentry);
3111 ceph_put_mds_session(di->lease_session);
3112 di->lease_session = NULL;
3115 static void handle_lease(struct ceph_mds_client *mdsc,
3116 struct ceph_mds_session *session,
3117 struct ceph_msg *msg)
3119 struct super_block *sb = mdsc->fsc->sb;
3120 struct inode *inode;
3121 struct dentry *parent, *dentry;
3122 struct ceph_dentry_info *di;
3123 int mds = session->s_mds;
3124 struct ceph_mds_lease *h = msg->front.iov_base;
3126 struct ceph_vino vino;
3130 dout("handle_lease from mds%d\n", mds);
3133 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3135 vino.ino = le64_to_cpu(h->ino);
3136 vino.snap = CEPH_NOSNAP;
3137 seq = le32_to_cpu(h->seq);
3138 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3139 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3140 if (dname.len != get_unaligned_le32(h+1))
3144 inode = ceph_find_inode(sb, vino);
3145 dout("handle_lease %s, ino %llx %p %.*s\n",
3146 ceph_lease_op_name(h->action), vino.ino, inode,
3147 dname.len, dname.name);
3149 mutex_lock(&session->s_mutex);
3152 if (inode == NULL) {
3153 dout("handle_lease no inode %llx\n", vino.ino);
3158 parent = d_find_alias(inode);
3160 dout("no parent dentry on inode %p\n", inode);
3162 goto release; /* hrm... */
3164 dname.hash = full_name_hash(dname.name, dname.len);
3165 dentry = d_lookup(parent, &dname);
3170 spin_lock(&dentry->d_lock);
3171 di = ceph_dentry(dentry);
3172 switch (h->action) {
3173 case CEPH_MDS_LEASE_REVOKE:
3174 if (di->lease_session == session) {
3175 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3176 h->seq = cpu_to_le32(di->lease_seq);
3177 __ceph_mdsc_drop_dentry_lease(dentry);
3182 case CEPH_MDS_LEASE_RENEW:
3183 if (di->lease_session == session &&
3184 di->lease_gen == session->s_cap_gen &&
3185 di->lease_renew_from &&
3186 di->lease_renew_after == 0) {
3187 unsigned long duration =
3188 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3190 di->lease_seq = seq;
3191 dentry->d_time = di->lease_renew_from + duration;
3192 di->lease_renew_after = di->lease_renew_from +
3194 di->lease_renew_from = 0;
3198 spin_unlock(&dentry->d_lock);
3205 /* let's just reuse the same message */
3206 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3208 ceph_con_send(&session->s_con, msg);
3212 mutex_unlock(&session->s_mutex);
3216 pr_err("corrupt lease message\n");
3220 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3221 struct inode *inode,
3222 struct dentry *dentry, char action,
3225 struct ceph_msg *msg;
3226 struct ceph_mds_lease *lease;
3227 int len = sizeof(*lease) + sizeof(u32);
3230 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3231 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3232 dnamelen = dentry->d_name.len;
3235 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3238 lease = msg->front.iov_base;
3239 lease->action = action;
3240 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3241 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3242 lease->seq = cpu_to_le32(seq);
3243 put_unaligned_le32(dnamelen, lease + 1);
3244 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3247 * if this is a preemptive lease RELEASE, no need to
3248 * flush request stream, since the actual request will
3251 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3253 ceph_con_send(&session->s_con, msg);
3257 * Preemptively release a lease we expect to invalidate anyway.
3258 * Pass @inode always, @dentry is optional.
3260 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3261 struct dentry *dentry)
3263 struct ceph_dentry_info *di;
3264 struct ceph_mds_session *session;
3267 BUG_ON(inode == NULL);
3268 BUG_ON(dentry == NULL);
3270 /* is dentry lease valid? */
3271 spin_lock(&dentry->d_lock);
3272 di = ceph_dentry(dentry);
3273 if (!di || !di->lease_session ||
3274 di->lease_session->s_mds < 0 ||
3275 di->lease_gen != di->lease_session->s_cap_gen ||
3276 !time_before(jiffies, dentry->d_time)) {
3277 dout("lease_release inode %p dentry %p -- "
3280 spin_unlock(&dentry->d_lock);
3284 /* we do have a lease on this dentry; note mds and seq */
3285 session = ceph_get_mds_session(di->lease_session);
3286 seq = di->lease_seq;
3287 __ceph_mdsc_drop_dentry_lease(dentry);
3288 spin_unlock(&dentry->d_lock);
3290 dout("lease_release inode %p dentry %p to mds%d\n",
3291 inode, dentry, session->s_mds);
3292 ceph_mdsc_lease_send_msg(session, inode, dentry,
3293 CEPH_MDS_LEASE_RELEASE, seq);
3294 ceph_put_mds_session(session);
3298 * drop all leases (and dentry refs) in preparation for umount
3300 static void drop_leases(struct ceph_mds_client *mdsc)
3304 dout("drop_leases\n");
3305 mutex_lock(&mdsc->mutex);
3306 for (i = 0; i < mdsc->max_sessions; i++) {
3307 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3310 mutex_unlock(&mdsc->mutex);
3311 mutex_lock(&s->s_mutex);
3312 mutex_unlock(&s->s_mutex);
3313 ceph_put_mds_session(s);
3314 mutex_lock(&mdsc->mutex);
3316 mutex_unlock(&mdsc->mutex);
3322 * delayed work -- periodically trim expired leases, renew caps with mds
3324 static void schedule_delayed(struct ceph_mds_client *mdsc)
3327 unsigned hz = round_jiffies_relative(HZ * delay);
3328 schedule_delayed_work(&mdsc->delayed_work, hz);
3331 static void delayed_work(struct work_struct *work)
3334 struct ceph_mds_client *mdsc =
3335 container_of(work, struct ceph_mds_client, delayed_work.work);
3339 dout("mdsc delayed_work\n");
3340 ceph_check_delayed_caps(mdsc);
3342 mutex_lock(&mdsc->mutex);
3343 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3344 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3345 mdsc->last_renew_caps);
3347 mdsc->last_renew_caps = jiffies;
3349 for (i = 0; i < mdsc->max_sessions; i++) {
3350 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3353 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3354 dout("resending session close request for mds%d\n",
3356 request_close_session(mdsc, s);
3357 ceph_put_mds_session(s);
3360 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3361 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3362 s->s_state = CEPH_MDS_SESSION_HUNG;
3363 pr_info("mds%d hung\n", s->s_mds);
3366 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3367 /* this mds is failed or recovering, just wait */
3368 ceph_put_mds_session(s);
3371 mutex_unlock(&mdsc->mutex);
3373 mutex_lock(&s->s_mutex);
3375 send_renew_caps(mdsc, s);
3377 ceph_con_keepalive(&s->s_con);
3378 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3379 s->s_state == CEPH_MDS_SESSION_HUNG)
3380 ceph_send_cap_releases(mdsc, s);
3381 mutex_unlock(&s->s_mutex);
3382 ceph_put_mds_session(s);
3384 mutex_lock(&mdsc->mutex);
3386 mutex_unlock(&mdsc->mutex);
3388 schedule_delayed(mdsc);
3391 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3394 struct ceph_mds_client *mdsc;
3396 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3401 mutex_init(&mdsc->mutex);
3402 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3403 if (mdsc->mdsmap == NULL) {
3408 init_completion(&mdsc->safe_umount_waiters);
3409 init_waitqueue_head(&mdsc->session_close_wq);
3410 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3411 mdsc->sessions = NULL;
3412 atomic_set(&mdsc->num_sessions, 0);
3413 mdsc->max_sessions = 0;
3415 mdsc->last_snap_seq = 0;
3416 init_rwsem(&mdsc->snap_rwsem);
3417 mdsc->snap_realms = RB_ROOT;
3418 INIT_LIST_HEAD(&mdsc->snap_empty);
3419 spin_lock_init(&mdsc->snap_empty_lock);
3421 mdsc->oldest_tid = 0;
3422 mdsc->request_tree = RB_ROOT;
3423 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3424 mdsc->last_renew_caps = jiffies;
3425 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3426 spin_lock_init(&mdsc->cap_delay_lock);
3427 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3428 spin_lock_init(&mdsc->snap_flush_lock);
3429 mdsc->cap_flush_seq = 0;
3430 mdsc->last_cap_flush_tid = 1;
3431 INIT_LIST_HEAD(&mdsc->cap_dirty);
3432 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3433 mdsc->num_cap_flushing = 0;
3434 spin_lock_init(&mdsc->cap_dirty_lock);
3435 init_waitqueue_head(&mdsc->cap_flushing_wq);
3436 spin_lock_init(&mdsc->dentry_lru_lock);
3437 INIT_LIST_HEAD(&mdsc->dentry_lru);
3439 ceph_caps_init(mdsc);
3440 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3442 init_rwsem(&mdsc->pool_perm_rwsem);
3443 mdsc->pool_perm_tree = RB_ROOT;
3449 * Wait for safe replies on open mds requests. If we time out, drop
3450 * all requests from the tree to avoid dangling dentry refs.
3452 static void wait_requests(struct ceph_mds_client *mdsc)
3454 struct ceph_options *opts = mdsc->fsc->client->options;
3455 struct ceph_mds_request *req;
3457 mutex_lock(&mdsc->mutex);
3458 if (__get_oldest_req(mdsc)) {
3459 mutex_unlock(&mdsc->mutex);
3461 dout("wait_requests waiting for requests\n");
3462 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3463 ceph_timeout_jiffies(opts->mount_timeout));
3465 /* tear down remaining requests */
3466 mutex_lock(&mdsc->mutex);
3467 while ((req = __get_oldest_req(mdsc))) {
3468 dout("wait_requests timed out on tid %llu\n",
3470 __unregister_request(mdsc, req);
3473 mutex_unlock(&mdsc->mutex);
3474 dout("wait_requests done\n");
3478 * called before mount is ro, and before dentries are torn down.
3479 * (hmm, does this still race with new lookups?)
3481 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3483 dout("pre_umount\n");
3487 ceph_flush_dirty_caps(mdsc);
3488 wait_requests(mdsc);
3491 * wait for reply handlers to drop their request refs and
3492 * their inode/dcache refs
3498 * wait for all write mds requests to flush.
3500 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3502 struct ceph_mds_request *req = NULL, *nextreq;
3505 mutex_lock(&mdsc->mutex);
3506 dout("wait_unsafe_requests want %lld\n", want_tid);
3508 req = __get_oldest_req(mdsc);
3509 while (req && req->r_tid <= want_tid) {
3510 /* find next request */
3511 n = rb_next(&req->r_node);
3513 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3516 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3517 (req->r_op & CEPH_MDS_OP_WRITE)) {
3519 ceph_mdsc_get_request(req);
3521 ceph_mdsc_get_request(nextreq);
3522 mutex_unlock(&mdsc->mutex);
3523 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3524 req->r_tid, want_tid);
3525 wait_for_completion(&req->r_safe_completion);
3526 mutex_lock(&mdsc->mutex);
3527 ceph_mdsc_put_request(req);
3529 break; /* next dne before, so we're done! */
3530 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3531 /* next request was removed from tree */
3532 ceph_mdsc_put_request(nextreq);
3535 ceph_mdsc_put_request(nextreq); /* won't go away */
3539 mutex_unlock(&mdsc->mutex);
3540 dout("wait_unsafe_requests done\n");
3543 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3545 u64 want_tid, want_flush, want_snap;
3547 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3551 mutex_lock(&mdsc->mutex);
3552 want_tid = mdsc->last_tid;
3553 mutex_unlock(&mdsc->mutex);
3555 ceph_flush_dirty_caps(mdsc);
3556 spin_lock(&mdsc->cap_dirty_lock);
3557 want_flush = mdsc->cap_flush_seq;
3558 spin_unlock(&mdsc->cap_dirty_lock);
3560 down_read(&mdsc->snap_rwsem);
3561 want_snap = mdsc->last_snap_seq;
3562 up_read(&mdsc->snap_rwsem);
3564 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3565 want_tid, want_flush, want_snap);
3567 wait_unsafe_requests(mdsc, want_tid);
3568 wait_caps_flush(mdsc, want_flush, want_snap);
3572 * true if all sessions are closed, or we force unmount
3574 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3576 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3578 return atomic_read(&mdsc->num_sessions) == 0;
3582 * called after sb is ro.
3584 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3586 struct ceph_options *opts = mdsc->fsc->client->options;
3587 struct ceph_mds_session *session;
3590 dout("close_sessions\n");
3592 /* close sessions */
3593 mutex_lock(&mdsc->mutex);
3594 for (i = 0; i < mdsc->max_sessions; i++) {
3595 session = __ceph_lookup_mds_session(mdsc, i);
3598 mutex_unlock(&mdsc->mutex);
3599 mutex_lock(&session->s_mutex);
3600 __close_session(mdsc, session);
3601 mutex_unlock(&session->s_mutex);
3602 ceph_put_mds_session(session);
3603 mutex_lock(&mdsc->mutex);
3605 mutex_unlock(&mdsc->mutex);
3607 dout("waiting for sessions to close\n");
3608 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3609 ceph_timeout_jiffies(opts->mount_timeout));
3611 /* tear down remaining sessions */
3612 mutex_lock(&mdsc->mutex);
3613 for (i = 0; i < mdsc->max_sessions; i++) {
3614 if (mdsc->sessions[i]) {
3615 session = get_session(mdsc->sessions[i]);
3616 __unregister_session(mdsc, session);
3617 mutex_unlock(&mdsc->mutex);
3618 mutex_lock(&session->s_mutex);
3619 remove_session_caps(session);
3620 mutex_unlock(&session->s_mutex);
3621 ceph_put_mds_session(session);
3622 mutex_lock(&mdsc->mutex);
3625 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3626 mutex_unlock(&mdsc->mutex);
3628 ceph_cleanup_empty_realms(mdsc);
3630 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3635 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3638 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3640 ceph_mdsmap_destroy(mdsc->mdsmap);
3641 kfree(mdsc->sessions);
3642 ceph_caps_finalize(mdsc);
3643 ceph_pool_perm_destroy(mdsc);
3646 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3648 struct ceph_mds_client *mdsc = fsc->mdsc;
3650 dout("mdsc_destroy %p\n", mdsc);
3651 ceph_mdsc_stop(mdsc);
3653 /* flush out any connection work with references to us */
3658 dout("mdsc_destroy %p done\n", mdsc);
3663 * handle mds map update.
3665 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3669 void *p = msg->front.iov_base;
3670 void *end = p + msg->front.iov_len;
3671 struct ceph_mdsmap *newmap, *oldmap;
3672 struct ceph_fsid fsid;
3675 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3676 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3677 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3679 epoch = ceph_decode_32(&p);
3680 maplen = ceph_decode_32(&p);
3681 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3683 /* do we need it? */
3684 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3685 mutex_lock(&mdsc->mutex);
3686 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3687 dout("handle_map epoch %u <= our %u\n",
3688 epoch, mdsc->mdsmap->m_epoch);
3689 mutex_unlock(&mdsc->mutex);
3693 newmap = ceph_mdsmap_decode(&p, end);
3694 if (IS_ERR(newmap)) {
3695 err = PTR_ERR(newmap);
3699 /* swap into place */
3701 oldmap = mdsc->mdsmap;
3702 mdsc->mdsmap = newmap;
3703 check_new_map(mdsc, newmap, oldmap);
3704 ceph_mdsmap_destroy(oldmap);
3706 mdsc->mdsmap = newmap; /* first mds map */
3708 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3710 __wake_requests(mdsc, &mdsc->waiting_for_map);
3712 mutex_unlock(&mdsc->mutex);
3713 schedule_delayed(mdsc);
3717 mutex_unlock(&mdsc->mutex);
3719 pr_err("error decoding mdsmap %d\n", err);
3723 static struct ceph_connection *con_get(struct ceph_connection *con)
3725 struct ceph_mds_session *s = con->private;
3727 if (get_session(s)) {
3728 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3731 dout("mdsc con_get %p FAIL\n", s);
3735 static void con_put(struct ceph_connection *con)
3737 struct ceph_mds_session *s = con->private;
3739 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3740 ceph_put_mds_session(s);
3744 * if the client is unresponsive for long enough, the mds will kill
3745 * the session entirely.
3747 static void peer_reset(struct ceph_connection *con)
3749 struct ceph_mds_session *s = con->private;
3750 struct ceph_mds_client *mdsc = s->s_mdsc;
3752 pr_warn("mds%d closed our session\n", s->s_mds);
3753 send_mds_reconnect(mdsc, s);
3756 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3758 struct ceph_mds_session *s = con->private;
3759 struct ceph_mds_client *mdsc = s->s_mdsc;
3760 int type = le16_to_cpu(msg->hdr.type);
3762 mutex_lock(&mdsc->mutex);
3763 if (__verify_registered_session(mdsc, s) < 0) {
3764 mutex_unlock(&mdsc->mutex);
3767 mutex_unlock(&mdsc->mutex);
3770 case CEPH_MSG_MDS_MAP:
3771 ceph_mdsc_handle_map(mdsc, msg);
3773 case CEPH_MSG_CLIENT_SESSION:
3774 handle_session(s, msg);
3776 case CEPH_MSG_CLIENT_REPLY:
3777 handle_reply(s, msg);
3779 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3780 handle_forward(mdsc, s, msg);
3782 case CEPH_MSG_CLIENT_CAPS:
3783 ceph_handle_caps(s, msg);
3785 case CEPH_MSG_CLIENT_SNAP:
3786 ceph_handle_snap(mdsc, s, msg);
3788 case CEPH_MSG_CLIENT_LEASE:
3789 handle_lease(mdsc, s, msg);
3793 pr_err("received unknown message type %d %s\n", type,
3794 ceph_msg_type_name(type));
3805 * Note: returned pointer is the address of a structure that's
3806 * managed separately. Caller must *not* attempt to free it.
3808 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3809 int *proto, int force_new)
3811 struct ceph_mds_session *s = con->private;
3812 struct ceph_mds_client *mdsc = s->s_mdsc;
3813 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3814 struct ceph_auth_handshake *auth = &s->s_auth;
3816 if (force_new && auth->authorizer) {
3817 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3818 auth->authorizer = NULL;
3820 if (!auth->authorizer) {
3821 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3824 return ERR_PTR(ret);
3826 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3829 return ERR_PTR(ret);
3831 *proto = ac->protocol;
3837 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3839 struct ceph_mds_session *s = con->private;
3840 struct ceph_mds_client *mdsc = s->s_mdsc;
3841 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3843 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3846 static int invalidate_authorizer(struct ceph_connection *con)
3848 struct ceph_mds_session *s = con->private;
3849 struct ceph_mds_client *mdsc = s->s_mdsc;
3850 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3852 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3854 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3857 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3858 struct ceph_msg_header *hdr, int *skip)
3860 struct ceph_msg *msg;
3861 int type = (int) le16_to_cpu(hdr->type);
3862 int front_len = (int) le32_to_cpu(hdr->front_len);
3868 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3870 pr_err("unable to allocate msg type %d len %d\n",
3878 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3880 struct ceph_mds_session *s = con->private;
3881 struct ceph_auth_handshake *auth = &s->s_auth;
3882 return ceph_auth_sign_message(auth, msg);
3885 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3887 struct ceph_mds_session *s = con->private;
3888 struct ceph_auth_handshake *auth = &s->s_auth;
3889 return ceph_auth_check_message_signature(auth, msg);
3892 static const struct ceph_connection_operations mds_con_ops = {
3895 .dispatch = dispatch,
3896 .get_authorizer = get_authorizer,
3897 .verify_authorizer_reply = verify_authorizer_reply,
3898 .invalidate_authorizer = invalidate_authorizer,
3899 .peer_reset = peer_reset,
3900 .alloc_msg = mds_alloc_msg,
3901 .sign_message = sign_message,
3902 .check_message_signature = check_message_signature,