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45 * are no longer valid.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;
637 req->r_unsafe_dir = dir;
641 static void __unregister_request(struct ceph_mds_client *mdsc,
642 struct ceph_mds_request *req)
644 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
646 if (req->r_tid == mdsc->oldest_tid) {
647 struct rb_node *p = rb_next(&req->r_node);
648 mdsc->oldest_tid = 0;
650 struct ceph_mds_request *next_req =
651 rb_entry(p, struct ceph_mds_request, r_node);
652 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
653 mdsc->oldest_tid = next_req->r_tid;
660 rb_erase(&req->r_node, &mdsc->request_tree);
661 RB_CLEAR_NODE(&req->r_node);
663 if (req->r_unsafe_dir && req->r_got_unsafe) {
664 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
665 spin_lock(&ci->i_unsafe_lock);
666 list_del_init(&req->r_unsafe_dir_item);
667 spin_unlock(&ci->i_unsafe_lock);
670 if (req->r_unsafe_dir) {
671 iput(req->r_unsafe_dir);
672 req->r_unsafe_dir = NULL;
675 complete_all(&req->r_safe_completion);
677 ceph_mdsc_put_request(req);
681 * Choose mds to send request to next. If there is a hint set in the
682 * request (e.g., due to a prior forward hint from the mds), use that.
683 * Otherwise, consult frag tree and/or caps to identify the
684 * appropriate mds. If all else fails, choose randomly.
686 * Called under mdsc->mutex.
688 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
691 * we don't need to worry about protecting the d_parent access
692 * here because we never renaming inside the snapped namespace
693 * except to resplice to another snapdir, and either the old or new
694 * result is a valid result.
696 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
697 dentry = dentry->d_parent;
701 static int __choose_mds(struct ceph_mds_client *mdsc,
702 struct ceph_mds_request *req)
705 struct ceph_inode_info *ci;
706 struct ceph_cap *cap;
707 int mode = req->r_direct_mode;
709 u32 hash = req->r_direct_hash;
710 bool is_hash = req->r_direct_is_hash;
713 * is there a specific mds we should try? ignore hint if we have
714 * no session and the mds is not up (active or recovering).
716 if (req->r_resend_mds >= 0 &&
717 (__have_session(mdsc, req->r_resend_mds) ||
718 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
719 dout("choose_mds using resend_mds mds%d\n",
721 return req->r_resend_mds;
724 if (mode == USE_RANDOM_MDS)
729 inode = req->r_inode;
730 } else if (req->r_dentry) {
731 /* ignore race with rename; old or new d_parent is okay */
732 struct dentry *parent = req->r_dentry->d_parent;
733 struct inode *dir = d_inode(parent);
735 if (dir->i_sb != mdsc->fsc->sb) {
737 inode = d_inode(req->r_dentry);
738 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
739 /* direct snapped/virtual snapdir requests
740 * based on parent dir inode */
741 struct dentry *dn = get_nonsnap_parent(parent);
743 dout("__choose_mds using nonsnap parent %p\n", inode);
746 inode = d_inode(req->r_dentry);
747 if (!inode || mode == USE_AUTH_MDS) {
750 hash = ceph_dentry_hash(dir, req->r_dentry);
756 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
760 ci = ceph_inode(inode);
762 if (is_hash && S_ISDIR(inode->i_mode)) {
763 struct ceph_inode_frag frag;
766 ceph_choose_frag(ci, hash, &frag, &found);
768 if (mode == USE_ANY_MDS && frag.ndist > 0) {
771 /* choose a random replica */
772 get_random_bytes(&r, 1);
775 dout("choose_mds %p %llx.%llx "
776 "frag %u mds%d (%d/%d)\n",
777 inode, ceph_vinop(inode),
780 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
781 CEPH_MDS_STATE_ACTIVE)
785 /* since this file/dir wasn't known to be
786 * replicated, then we want to look for the
787 * authoritative mds. */
790 /* choose auth mds */
792 dout("choose_mds %p %llx.%llx "
793 "frag %u mds%d (auth)\n",
794 inode, ceph_vinop(inode), frag.frag, mds);
795 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
796 CEPH_MDS_STATE_ACTIVE)
802 spin_lock(&ci->i_ceph_lock);
804 if (mode == USE_AUTH_MDS)
805 cap = ci->i_auth_cap;
806 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
807 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
809 spin_unlock(&ci->i_ceph_lock);
812 mds = cap->session->s_mds;
813 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
814 inode, ceph_vinop(inode), mds,
815 cap == ci->i_auth_cap ? "auth " : "", cap);
816 spin_unlock(&ci->i_ceph_lock);
820 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
821 dout("choose_mds chose random mds%d\n", mds);
829 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
831 struct ceph_msg *msg;
832 struct ceph_mds_session_head *h;
834 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
837 pr_err("create_session_msg ENOMEM creating msg\n");
840 h = msg->front.iov_base;
841 h->op = cpu_to_le32(op);
842 h->seq = cpu_to_le64(seq);
848 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
849 * to include additional client metadata fields.
851 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
853 struct ceph_msg *msg;
854 struct ceph_mds_session_head *h;
856 int metadata_bytes = 0;
857 int metadata_key_count = 0;
858 struct ceph_options *opt = mdsc->fsc->client->options;
861 const char* metadata[][2] = {
862 {"hostname", utsname()->nodename},
863 {"kernel_version", utsname()->release},
864 {"entity_id", opt->name ? opt->name : ""},
868 /* Calculate serialized length of metadata */
869 metadata_bytes = 4; /* map length */
870 for (i = 0; metadata[i][0] != NULL; ++i) {
871 metadata_bytes += 8 + strlen(metadata[i][0]) +
872 strlen(metadata[i][1]);
873 metadata_key_count++;
876 /* Allocate the message */
877 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
880 pr_err("create_session_msg ENOMEM creating msg\n");
883 h = msg->front.iov_base;
884 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
885 h->seq = cpu_to_le64(seq);
888 * Serialize client metadata into waiting buffer space, using
889 * the format that userspace expects for map<string, string>
891 * ClientSession messages with metadata are v2
893 msg->hdr.version = cpu_to_le16(2);
894 msg->hdr.compat_version = cpu_to_le16(1);
896 /* The write pointer, following the session_head structure */
897 p = msg->front.iov_base + sizeof(*h);
899 /* Number of entries in the map */
900 ceph_encode_32(&p, metadata_key_count);
902 /* Two length-prefixed strings for each entry in the map */
903 for (i = 0; metadata[i][0] != NULL; ++i) {
904 size_t const key_len = strlen(metadata[i][0]);
905 size_t const val_len = strlen(metadata[i][1]);
907 ceph_encode_32(&p, key_len);
908 memcpy(p, metadata[i][0], key_len);
910 ceph_encode_32(&p, val_len);
911 memcpy(p, metadata[i][1], val_len);
919 * send session open request.
921 * called under mdsc->mutex
923 static int __open_session(struct ceph_mds_client *mdsc,
924 struct ceph_mds_session *session)
926 struct ceph_msg *msg;
928 int mds = session->s_mds;
930 /* wait for mds to go active? */
931 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
932 dout("open_session to mds%d (%s)\n", mds,
933 ceph_mds_state_name(mstate));
934 session->s_state = CEPH_MDS_SESSION_OPENING;
935 session->s_renew_requested = jiffies;
937 /* send connect message */
938 msg = create_session_open_msg(mdsc, session->s_seq);
941 ceph_con_send(&session->s_con, msg);
946 * open sessions for any export targets for the given mds
948 * called under mdsc->mutex
950 static struct ceph_mds_session *
951 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
953 struct ceph_mds_session *session;
955 session = __ceph_lookup_mds_session(mdsc, target);
957 session = register_session(mdsc, target);
961 if (session->s_state == CEPH_MDS_SESSION_NEW ||
962 session->s_state == CEPH_MDS_SESSION_CLOSING)
963 __open_session(mdsc, session);
968 struct ceph_mds_session *
969 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
971 struct ceph_mds_session *session;
973 dout("open_export_target_session to mds%d\n", target);
975 mutex_lock(&mdsc->mutex);
976 session = __open_export_target_session(mdsc, target);
977 mutex_unlock(&mdsc->mutex);
982 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
983 struct ceph_mds_session *session)
985 struct ceph_mds_info *mi;
986 struct ceph_mds_session *ts;
987 int i, mds = session->s_mds;
989 if (mds >= mdsc->mdsmap->m_max_mds)
992 mi = &mdsc->mdsmap->m_info[mds];
993 dout("open_export_target_sessions for mds%d (%d targets)\n",
994 session->s_mds, mi->num_export_targets);
996 for (i = 0; i < mi->num_export_targets; i++) {
997 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
999 ceph_put_mds_session(ts);
1003 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1004 struct ceph_mds_session *session)
1006 mutex_lock(&mdsc->mutex);
1007 __open_export_target_sessions(mdsc, session);
1008 mutex_unlock(&mdsc->mutex);
1015 /* caller holds s_cap_lock, we drop it */
1016 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1017 struct ceph_mds_session *session)
1018 __releases(session->s_cap_lock)
1020 LIST_HEAD(tmp_list);
1021 list_splice_init(&session->s_cap_releases, &tmp_list);
1022 session->s_num_cap_releases = 0;
1023 spin_unlock(&session->s_cap_lock);
1025 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1026 while (!list_empty(&tmp_list)) {
1027 struct ceph_cap *cap;
1028 /* zero out the in-progress message */
1029 cap = list_first_entry(&tmp_list,
1030 struct ceph_cap, session_caps);
1031 list_del(&cap->session_caps);
1032 ceph_put_cap(mdsc, cap);
1036 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1037 struct ceph_mds_session *session)
1039 struct ceph_mds_request *req;
1042 dout("cleanup_session_requests mds%d\n", session->s_mds);
1043 mutex_lock(&mdsc->mutex);
1044 while (!list_empty(&session->s_unsafe)) {
1045 req = list_first_entry(&session->s_unsafe,
1046 struct ceph_mds_request, r_unsafe_item);
1047 list_del_init(&req->r_unsafe_item);
1048 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1050 __unregister_request(mdsc, req);
1052 /* zero r_attempts, so kick_requests() will re-send requests */
1053 p = rb_first(&mdsc->request_tree);
1055 req = rb_entry(p, struct ceph_mds_request, r_node);
1057 if (req->r_session &&
1058 req->r_session->s_mds == session->s_mds)
1059 req->r_attempts = 0;
1061 mutex_unlock(&mdsc->mutex);
1065 * Helper to safely iterate over all caps associated with a session, with
1066 * special care taken to handle a racing __ceph_remove_cap().
1068 * Caller must hold session s_mutex.
1070 static int iterate_session_caps(struct ceph_mds_session *session,
1071 int (*cb)(struct inode *, struct ceph_cap *,
1074 struct list_head *p;
1075 struct ceph_cap *cap;
1076 struct inode *inode, *last_inode = NULL;
1077 struct ceph_cap *old_cap = NULL;
1080 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1081 spin_lock(&session->s_cap_lock);
1082 p = session->s_caps.next;
1083 while (p != &session->s_caps) {
1084 cap = list_entry(p, struct ceph_cap, session_caps);
1085 inode = igrab(&cap->ci->vfs_inode);
1090 session->s_cap_iterator = cap;
1091 spin_unlock(&session->s_cap_lock);
1098 ceph_put_cap(session->s_mdsc, old_cap);
1102 ret = cb(inode, cap, arg);
1105 spin_lock(&session->s_cap_lock);
1107 if (cap->ci == NULL) {
1108 dout("iterate_session_caps finishing cap %p removal\n",
1110 BUG_ON(cap->session != session);
1111 cap->session = NULL;
1112 list_del_init(&cap->session_caps);
1113 session->s_nr_caps--;
1114 if (cap->queue_release) {
1115 list_add_tail(&cap->session_caps,
1116 &session->s_cap_releases);
1117 session->s_num_cap_releases++;
1119 old_cap = cap; /* put_cap it w/o locks held */
1127 session->s_cap_iterator = NULL;
1128 spin_unlock(&session->s_cap_lock);
1132 ceph_put_cap(session->s_mdsc, old_cap);
1137 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1140 struct ceph_inode_info *ci = ceph_inode(inode);
1141 LIST_HEAD(to_remove);
1144 dout("removing cap %p, ci is %p, inode is %p\n",
1145 cap, ci, &ci->vfs_inode);
1146 spin_lock(&ci->i_ceph_lock);
1147 __ceph_remove_cap(cap, false);
1148 if (!ci->i_auth_cap) {
1149 struct ceph_cap_flush *cf;
1150 struct ceph_mds_client *mdsc =
1151 ceph_sb_to_client(inode->i_sb)->mdsc;
1154 struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1157 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1158 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1159 list_add(&cf->list, &to_remove);
1162 spin_lock(&mdsc->cap_dirty_lock);
1164 list_for_each_entry(cf, &to_remove, list)
1165 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
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);
1188 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1189 list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1190 ci->i_prealloc_cap_flush = NULL;
1193 spin_unlock(&ci->i_ceph_lock);
1194 while (!list_empty(&to_remove)) {
1195 struct ceph_cap_flush *cf;
1196 cf = list_first_entry(&to_remove,
1197 struct ceph_cap_flush, list);
1198 list_del(&cf->list);
1199 ceph_free_cap_flush(cf);
1207 * caller must hold session s_mutex
1209 static void remove_session_caps(struct ceph_mds_session *session)
1211 dout("remove_session_caps on %p\n", session);
1212 iterate_session_caps(session, remove_session_caps_cb, NULL);
1214 spin_lock(&session->s_cap_lock);
1215 if (session->s_nr_caps > 0) {
1216 struct super_block *sb = session->s_mdsc->fsc->sb;
1217 struct inode *inode;
1218 struct ceph_cap *cap, *prev = NULL;
1219 struct ceph_vino vino;
1221 * iterate_session_caps() skips inodes that are being
1222 * deleted, we need to wait until deletions are complete.
1223 * __wait_on_freeing_inode() is designed for the job,
1224 * but it is not exported, so use lookup inode function
1227 while (!list_empty(&session->s_caps)) {
1228 cap = list_entry(session->s_caps.next,
1229 struct ceph_cap, session_caps);
1233 vino = cap->ci->i_vino;
1234 spin_unlock(&session->s_cap_lock);
1236 inode = ceph_find_inode(sb, vino);
1239 spin_lock(&session->s_cap_lock);
1243 // drop cap expires and unlock s_cap_lock
1244 cleanup_cap_releases(session->s_mdsc, session);
1246 BUG_ON(session->s_nr_caps > 0);
1247 BUG_ON(!list_empty(&session->s_cap_flushing));
1251 * wake up any threads waiting on this session's caps. if the cap is
1252 * old (didn't get renewed on the client reconnect), remove it now.
1254 * caller must hold s_mutex.
1256 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1259 struct ceph_inode_info *ci = ceph_inode(inode);
1261 wake_up_all(&ci->i_cap_wq);
1263 spin_lock(&ci->i_ceph_lock);
1264 ci->i_wanted_max_size = 0;
1265 ci->i_requested_max_size = 0;
1266 spin_unlock(&ci->i_ceph_lock);
1271 static void wake_up_session_caps(struct ceph_mds_session *session,
1274 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1275 iterate_session_caps(session, wake_up_session_cb,
1276 (void *)(unsigned long)reconnect);
1280 * Send periodic message to MDS renewing all currently held caps. The
1281 * ack will reset the expiration for all caps from this session.
1283 * caller holds s_mutex
1285 static int send_renew_caps(struct ceph_mds_client *mdsc,
1286 struct ceph_mds_session *session)
1288 struct ceph_msg *msg;
1291 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1292 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1293 pr_info("mds%d caps stale\n", session->s_mds);
1294 session->s_renew_requested = jiffies;
1296 /* do not try to renew caps until a recovering mds has reconnected
1297 * with its clients. */
1298 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1299 if (state < CEPH_MDS_STATE_RECONNECT) {
1300 dout("send_renew_caps ignoring mds%d (%s)\n",
1301 session->s_mds, ceph_mds_state_name(state));
1305 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1306 ceph_mds_state_name(state));
1307 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1308 ++session->s_renew_seq);
1311 ceph_con_send(&session->s_con, msg);
1315 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1316 struct ceph_mds_session *session, u64 seq)
1318 struct ceph_msg *msg;
1320 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1321 session->s_mds, ceph_session_state_name(session->s_state), seq);
1322 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1325 ceph_con_send(&session->s_con, msg);
1331 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1333 * Called under session->s_mutex
1335 static void renewed_caps(struct ceph_mds_client *mdsc,
1336 struct ceph_mds_session *session, int is_renew)
1341 spin_lock(&session->s_cap_lock);
1342 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1344 session->s_cap_ttl = session->s_renew_requested +
1345 mdsc->mdsmap->m_session_timeout*HZ;
1348 if (time_before(jiffies, session->s_cap_ttl)) {
1349 pr_info("mds%d caps renewed\n", session->s_mds);
1352 pr_info("mds%d caps still stale\n", session->s_mds);
1355 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1356 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1357 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1358 spin_unlock(&session->s_cap_lock);
1361 wake_up_session_caps(session, 0);
1365 * send a session close request
1367 static int request_close_session(struct ceph_mds_client *mdsc,
1368 struct ceph_mds_session *session)
1370 struct ceph_msg *msg;
1372 dout("request_close_session mds%d state %s seq %lld\n",
1373 session->s_mds, ceph_session_state_name(session->s_state),
1375 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1378 ceph_con_send(&session->s_con, msg);
1383 * Called with s_mutex held.
1385 static int __close_session(struct ceph_mds_client *mdsc,
1386 struct ceph_mds_session *session)
1388 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1390 session->s_state = CEPH_MDS_SESSION_CLOSING;
1391 return request_close_session(mdsc, session);
1395 * Trim old(er) caps.
1397 * Because we can't cache an inode without one or more caps, we do
1398 * this indirectly: if a cap is unused, we prune its aliases, at which
1399 * point the inode will hopefully get dropped to.
1401 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1402 * memory pressure from the MDS, though, so it needn't be perfect.
1404 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1406 struct ceph_mds_session *session = arg;
1407 struct ceph_inode_info *ci = ceph_inode(inode);
1408 int used, wanted, oissued, mine;
1410 if (session->s_trim_caps <= 0)
1413 spin_lock(&ci->i_ceph_lock);
1414 mine = cap->issued | cap->implemented;
1415 used = __ceph_caps_used(ci);
1416 wanted = __ceph_caps_file_wanted(ci);
1417 oissued = __ceph_caps_issued_other(ci, cap);
1419 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1420 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1421 ceph_cap_string(used), ceph_cap_string(wanted));
1422 if (cap == ci->i_auth_cap) {
1423 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1424 !list_empty(&ci->i_cap_snaps))
1426 if ((used | wanted) & CEPH_CAP_ANY_WR)
1429 /* The inode has cached pages, but it's no longer used.
1430 * we can safely drop it */
1431 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1432 !(oissued & CEPH_CAP_FILE_CACHE)) {
1436 if ((used | wanted) & ~oissued & mine)
1437 goto out; /* we need these caps */
1439 session->s_trim_caps--;
1441 /* we aren't the only cap.. just remove us */
1442 __ceph_remove_cap(cap, true);
1444 /* try dropping referring dentries */
1445 spin_unlock(&ci->i_ceph_lock);
1446 d_prune_aliases(inode);
1447 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1448 inode, cap, atomic_read(&inode->i_count));
1453 spin_unlock(&ci->i_ceph_lock);
1458 * Trim session cap count down to some max number.
1460 static int trim_caps(struct ceph_mds_client *mdsc,
1461 struct ceph_mds_session *session,
1464 int trim_caps = session->s_nr_caps - max_caps;
1466 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1467 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1468 if (trim_caps > 0) {
1469 session->s_trim_caps = trim_caps;
1470 iterate_session_caps(session, trim_caps_cb, session);
1471 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1472 session->s_mds, session->s_nr_caps, max_caps,
1473 trim_caps - session->s_trim_caps);
1474 session->s_trim_caps = 0;
1477 ceph_send_cap_releases(mdsc, session);
1481 static int check_capsnap_flush(struct ceph_inode_info *ci,
1485 spin_lock(&ci->i_ceph_lock);
1486 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1487 struct ceph_cap_snap *capsnap =
1488 list_first_entry(&ci->i_cap_snaps,
1489 struct ceph_cap_snap, ci_item);
1490 ret = capsnap->follows >= want_snap_seq;
1492 spin_unlock(&ci->i_ceph_lock);
1496 static int check_caps_flush(struct ceph_mds_client *mdsc,
1500 struct ceph_cap_flush *cf;
1503 spin_lock(&mdsc->cap_dirty_lock);
1504 n = rb_first(&mdsc->cap_flush_tree);
1505 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1506 if (cf && cf->tid <= want_flush_tid) {
1507 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1508 cf->tid, want_flush_tid);
1511 spin_unlock(&mdsc->cap_dirty_lock);
1516 * flush all dirty inode data to disk.
1518 * returns true if we've flushed through want_flush_tid
1520 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1521 u64 want_flush_tid, u64 want_snap_seq)
1525 dout("check_caps_flush want %llu snap want %llu\n",
1526 want_flush_tid, want_snap_seq);
1527 mutex_lock(&mdsc->mutex);
1528 for (mds = 0; mds < mdsc->max_sessions; ) {
1529 struct ceph_mds_session *session = mdsc->sessions[mds];
1530 struct inode *inode = NULL;
1536 get_session(session);
1537 mutex_unlock(&mdsc->mutex);
1539 mutex_lock(&session->s_mutex);
1540 if (!list_empty(&session->s_cap_snaps_flushing)) {
1541 struct ceph_cap_snap *capsnap =
1542 list_first_entry(&session->s_cap_snaps_flushing,
1543 struct ceph_cap_snap,
1545 struct ceph_inode_info *ci = capsnap->ci;
1546 if (!check_capsnap_flush(ci, want_snap_seq)) {
1547 dout("check_cap_flush still flushing snap %p "
1548 "follows %lld <= %lld to mds%d\n",
1549 &ci->vfs_inode, capsnap->follows,
1550 want_snap_seq, mds);
1551 inode = igrab(&ci->vfs_inode);
1554 mutex_unlock(&session->s_mutex);
1555 ceph_put_mds_session(session);
1558 wait_event(mdsc->cap_flushing_wq,
1559 check_capsnap_flush(ceph_inode(inode),
1566 mutex_lock(&mdsc->mutex);
1568 mutex_unlock(&mdsc->mutex);
1570 wait_event(mdsc->cap_flushing_wq,
1571 check_caps_flush(mdsc, want_flush_tid));
1573 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1577 * called under s_mutex
1579 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1580 struct ceph_mds_session *session)
1582 struct ceph_msg *msg = NULL;
1583 struct ceph_mds_cap_release *head;
1584 struct ceph_mds_cap_item *item;
1585 struct ceph_cap *cap;
1586 LIST_HEAD(tmp_list);
1587 int num_cap_releases;
1589 spin_lock(&session->s_cap_lock);
1591 list_splice_init(&session->s_cap_releases, &tmp_list);
1592 num_cap_releases = session->s_num_cap_releases;
1593 session->s_num_cap_releases = 0;
1594 spin_unlock(&session->s_cap_lock);
1596 while (!list_empty(&tmp_list)) {
1598 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1599 PAGE_CACHE_SIZE, GFP_NOFS, false);
1602 head = msg->front.iov_base;
1603 head->num = cpu_to_le32(0);
1604 msg->front.iov_len = sizeof(*head);
1606 cap = list_first_entry(&tmp_list, struct ceph_cap,
1608 list_del(&cap->session_caps);
1611 head = msg->front.iov_base;
1612 le32_add_cpu(&head->num, 1);
1613 item = msg->front.iov_base + msg->front.iov_len;
1614 item->ino = cpu_to_le64(cap->cap_ino);
1615 item->cap_id = cpu_to_le64(cap->cap_id);
1616 item->migrate_seq = cpu_to_le32(cap->mseq);
1617 item->seq = cpu_to_le32(cap->issue_seq);
1618 msg->front.iov_len += sizeof(*item);
1620 ceph_put_cap(mdsc, cap);
1622 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1623 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1624 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1625 ceph_con_send(&session->s_con, msg);
1630 BUG_ON(num_cap_releases != 0);
1632 spin_lock(&session->s_cap_lock);
1633 if (!list_empty(&session->s_cap_releases))
1635 spin_unlock(&session->s_cap_lock);
1638 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1639 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1640 ceph_con_send(&session->s_con, msg);
1644 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1646 spin_lock(&session->s_cap_lock);
1647 list_splice(&tmp_list, &session->s_cap_releases);
1648 session->s_num_cap_releases += num_cap_releases;
1649 spin_unlock(&session->s_cap_lock);
1656 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1659 struct ceph_inode_info *ci = ceph_inode(dir);
1660 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1661 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1662 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1663 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1664 int order, num_entries;
1666 spin_lock(&ci->i_ceph_lock);
1667 num_entries = ci->i_files + ci->i_subdirs;
1668 spin_unlock(&ci->i_ceph_lock);
1669 num_entries = max(num_entries, 1);
1670 num_entries = min(num_entries, opt->max_readdir);
1672 order = get_order(size * num_entries);
1673 while (order >= 0) {
1674 rinfo->dir_in = (void*)__get_free_pages(GFP_KERNEL |
1684 num_entries = (PAGE_SIZE << order) / size;
1685 num_entries = min(num_entries, opt->max_readdir);
1687 rinfo->dir_buf_size = PAGE_SIZE << order;
1688 req->r_num_caps = num_entries + 1;
1689 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1690 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1695 * Create an mds request.
1697 struct ceph_mds_request *
1698 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1700 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1703 return ERR_PTR(-ENOMEM);
1705 mutex_init(&req->r_fill_mutex);
1707 req->r_started = jiffies;
1708 req->r_resend_mds = -1;
1709 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1711 kref_init(&req->r_kref);
1712 INIT_LIST_HEAD(&req->r_wait);
1713 init_completion(&req->r_completion);
1714 init_completion(&req->r_safe_completion);
1715 INIT_LIST_HEAD(&req->r_unsafe_item);
1717 req->r_stamp = CURRENT_TIME;
1720 req->r_direct_mode = mode;
1725 * return oldest (lowest) request, tid in request tree, 0 if none.
1727 * called under mdsc->mutex.
1729 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1731 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1733 return rb_entry(rb_first(&mdsc->request_tree),
1734 struct ceph_mds_request, r_node);
1737 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1739 return mdsc->oldest_tid;
1743 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1744 * on build_path_from_dentry in fs/cifs/dir.c.
1746 * If @stop_on_nosnap, generate path relative to the first non-snapped
1749 * Encode hidden .snap dirs as a double /, i.e.
1750 * foo/.snap/bar -> foo//bar
1752 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1755 struct dentry *temp;
1761 return ERR_PTR(-EINVAL);
1765 seq = read_seqbegin(&rename_lock);
1767 for (temp = dentry; !IS_ROOT(temp);) {
1768 struct inode *inode = d_inode(temp);
1769 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1770 len++; /* slash only */
1771 else if (stop_on_nosnap && inode &&
1772 ceph_snap(inode) == CEPH_NOSNAP)
1775 len += 1 + temp->d_name.len;
1776 temp = temp->d_parent;
1780 len--; /* no leading '/' */
1782 path = kmalloc(len+1, GFP_NOFS);
1784 return ERR_PTR(-ENOMEM);
1786 path[pos] = 0; /* trailing null */
1788 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1789 struct inode *inode;
1791 spin_lock(&temp->d_lock);
1792 inode = d_inode(temp);
1793 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1794 dout("build_path path+%d: %p SNAPDIR\n",
1796 } else if (stop_on_nosnap && inode &&
1797 ceph_snap(inode) == CEPH_NOSNAP) {
1798 spin_unlock(&temp->d_lock);
1801 pos -= temp->d_name.len;
1803 spin_unlock(&temp->d_lock);
1806 strncpy(path + pos, temp->d_name.name,
1809 spin_unlock(&temp->d_lock);
1812 temp = temp->d_parent;
1815 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1816 pr_err("build_path did not end path lookup where "
1817 "expected, namelen is %d, pos is %d\n", len, pos);
1818 /* presumably this is only possible if racing with a
1819 rename of one of the parent directories (we can not
1820 lock the dentries above us to prevent this, but
1821 retrying should be harmless) */
1826 *base = ceph_ino(d_inode(temp));
1828 dout("build_path on %p %d built %llx '%.*s'\n",
1829 dentry, d_count(dentry), *base, len, path);
1833 static int build_dentry_path(struct dentry *dentry,
1834 const char **ppath, int *ppathlen, u64 *pino,
1839 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1840 *pino = ceph_ino(d_inode(dentry->d_parent));
1841 *ppath = dentry->d_name.name;
1842 *ppathlen = dentry->d_name.len;
1845 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1847 return PTR_ERR(path);
1853 static int build_inode_path(struct inode *inode,
1854 const char **ppath, int *ppathlen, u64 *pino,
1857 struct dentry *dentry;
1860 if (ceph_snap(inode) == CEPH_NOSNAP) {
1861 *pino = ceph_ino(inode);
1865 dentry = d_find_alias(inode);
1866 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1869 return PTR_ERR(path);
1876 * request arguments may be specified via an inode *, a dentry *, or
1877 * an explicit ino+path.
1879 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1880 const char *rpath, u64 rino,
1881 const char **ppath, int *pathlen,
1882 u64 *ino, int *freepath)
1887 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1888 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1890 } else if (rdentry) {
1891 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1892 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1894 } else if (rpath || rino) {
1897 *pathlen = rpath ? strlen(rpath) : 0;
1898 dout(" path %.*s\n", *pathlen, rpath);
1905 * called under mdsc->mutex
1907 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1908 struct ceph_mds_request *req,
1909 int mds, bool drop_cap_releases)
1911 struct ceph_msg *msg;
1912 struct ceph_mds_request_head *head;
1913 const char *path1 = NULL;
1914 const char *path2 = NULL;
1915 u64 ino1 = 0, ino2 = 0;
1916 int pathlen1 = 0, pathlen2 = 0;
1917 int freepath1 = 0, freepath2 = 0;
1923 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1924 req->r_path1, req->r_ino1.ino,
1925 &path1, &pathlen1, &ino1, &freepath1);
1931 ret = set_request_path_attr(NULL, req->r_old_dentry,
1932 req->r_path2, req->r_ino2.ino,
1933 &path2, &pathlen2, &ino2, &freepath2);
1939 len = sizeof(*head) +
1940 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1941 sizeof(struct ceph_timespec);
1943 /* calculate (max) length for cap releases */
1944 len += sizeof(struct ceph_mds_request_release) *
1945 (!!req->r_inode_drop + !!req->r_dentry_drop +
1946 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1947 if (req->r_dentry_drop)
1948 len += req->r_dentry->d_name.len;
1949 if (req->r_old_dentry_drop)
1950 len += req->r_old_dentry->d_name.len;
1952 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1954 msg = ERR_PTR(-ENOMEM);
1958 msg->hdr.version = cpu_to_le16(2);
1959 msg->hdr.tid = cpu_to_le64(req->r_tid);
1961 head = msg->front.iov_base;
1962 p = msg->front.iov_base + sizeof(*head);
1963 end = msg->front.iov_base + msg->front.iov_len;
1965 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1966 head->op = cpu_to_le32(req->r_op);
1967 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1968 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1969 head->args = req->r_args;
1971 ceph_encode_filepath(&p, end, ino1, path1);
1972 ceph_encode_filepath(&p, end, ino2, path2);
1974 /* make note of release offset, in case we need to replay */
1975 req->r_request_release_offset = p - msg->front.iov_base;
1979 if (req->r_inode_drop)
1980 releases += ceph_encode_inode_release(&p,
1981 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1982 mds, req->r_inode_drop, req->r_inode_unless, 0);
1983 if (req->r_dentry_drop)
1984 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1985 mds, req->r_dentry_drop, req->r_dentry_unless);
1986 if (req->r_old_dentry_drop)
1987 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1988 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1989 if (req->r_old_inode_drop)
1990 releases += ceph_encode_inode_release(&p,
1991 d_inode(req->r_old_dentry),
1992 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1994 if (drop_cap_releases) {
1996 p = msg->front.iov_base + req->r_request_release_offset;
1999 head->num_releases = cpu_to_le16(releases);
2003 struct ceph_timespec ts;
2004 ceph_encode_timespec(&ts, &req->r_stamp);
2005 ceph_encode_copy(&p, &ts, sizeof(ts));
2009 msg->front.iov_len = p - msg->front.iov_base;
2010 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2012 if (req->r_pagelist) {
2013 struct ceph_pagelist *pagelist = req->r_pagelist;
2014 atomic_inc(&pagelist->refcnt);
2015 ceph_msg_data_add_pagelist(msg, pagelist);
2016 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2018 msg->hdr.data_len = 0;
2021 msg->hdr.data_off = cpu_to_le16(0);
2025 kfree((char *)path2);
2028 kfree((char *)path1);
2034 * called under mdsc->mutex if error, under no mutex if
2037 static void complete_request(struct ceph_mds_client *mdsc,
2038 struct ceph_mds_request *req)
2040 if (req->r_callback)
2041 req->r_callback(mdsc, req);
2043 complete_all(&req->r_completion);
2047 * called under mdsc->mutex
2049 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2050 struct ceph_mds_request *req,
2051 int mds, bool drop_cap_releases)
2053 struct ceph_mds_request_head *rhead;
2054 struct ceph_msg *msg;
2059 struct ceph_cap *cap =
2060 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2063 req->r_sent_on_mseq = cap->mseq;
2065 req->r_sent_on_mseq = -1;
2067 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2068 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2070 if (req->r_got_unsafe) {
2073 * Replay. Do not regenerate message (and rebuild
2074 * paths, etc.); just use the original message.
2075 * Rebuilding paths will break for renames because
2076 * d_move mangles the src name.
2078 msg = req->r_request;
2079 rhead = msg->front.iov_base;
2081 flags = le32_to_cpu(rhead->flags);
2082 flags |= CEPH_MDS_FLAG_REPLAY;
2083 rhead->flags = cpu_to_le32(flags);
2085 if (req->r_target_inode)
2086 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2088 rhead->num_retry = req->r_attempts - 1;
2090 /* remove cap/dentry releases from message */
2091 rhead->num_releases = 0;
2094 p = msg->front.iov_base + req->r_request_release_offset;
2096 struct ceph_timespec ts;
2097 ceph_encode_timespec(&ts, &req->r_stamp);
2098 ceph_encode_copy(&p, &ts, sizeof(ts));
2101 msg->front.iov_len = p - msg->front.iov_base;
2102 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2106 if (req->r_request) {
2107 ceph_msg_put(req->r_request);
2108 req->r_request = NULL;
2110 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2112 req->r_err = PTR_ERR(msg);
2113 return PTR_ERR(msg);
2115 req->r_request = msg;
2117 rhead = msg->front.iov_base;
2118 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2119 if (req->r_got_unsafe)
2120 flags |= CEPH_MDS_FLAG_REPLAY;
2121 if (req->r_locked_dir)
2122 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2123 rhead->flags = cpu_to_le32(flags);
2124 rhead->num_fwd = req->r_num_fwd;
2125 rhead->num_retry = req->r_attempts - 1;
2128 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2133 * send request, or put it on the appropriate wait list.
2135 static int __do_request(struct ceph_mds_client *mdsc,
2136 struct ceph_mds_request *req)
2138 struct ceph_mds_session *session = NULL;
2142 if (req->r_err || req->r_got_result) {
2144 __unregister_request(mdsc, req);
2148 if (req->r_timeout &&
2149 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2150 dout("do_request timed out\n");
2154 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2155 dout("do_request forced umount\n");
2160 put_request_session(req);
2162 mds = __choose_mds(mdsc, req);
2164 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2165 dout("do_request no mds or not active, waiting for map\n");
2166 list_add(&req->r_wait, &mdsc->waiting_for_map);
2170 /* get, open session */
2171 session = __ceph_lookup_mds_session(mdsc, mds);
2173 session = register_session(mdsc, mds);
2174 if (IS_ERR(session)) {
2175 err = PTR_ERR(session);
2179 req->r_session = get_session(session);
2181 dout("do_request mds%d session %p state %s\n", mds, session,
2182 ceph_session_state_name(session->s_state));
2183 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2184 session->s_state != CEPH_MDS_SESSION_HUNG) {
2185 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2186 session->s_state == CEPH_MDS_SESSION_CLOSING)
2187 __open_session(mdsc, session);
2188 list_add(&req->r_wait, &session->s_waiting);
2193 req->r_resend_mds = -1; /* forget any previous mds hint */
2195 if (req->r_request_started == 0) /* note request start time */
2196 req->r_request_started = jiffies;
2198 err = __prepare_send_request(mdsc, req, mds, false);
2200 ceph_msg_get(req->r_request);
2201 ceph_con_send(&session->s_con, req->r_request);
2205 ceph_put_mds_session(session);
2208 dout("__do_request early error %d\n", err);
2210 complete_request(mdsc, req);
2211 __unregister_request(mdsc, req);
2218 * called under mdsc->mutex
2220 static void __wake_requests(struct ceph_mds_client *mdsc,
2221 struct list_head *head)
2223 struct ceph_mds_request *req;
2224 LIST_HEAD(tmp_list);
2226 list_splice_init(head, &tmp_list);
2228 while (!list_empty(&tmp_list)) {
2229 req = list_entry(tmp_list.next,
2230 struct ceph_mds_request, r_wait);
2231 list_del_init(&req->r_wait);
2232 dout(" wake request %p tid %llu\n", req, req->r_tid);
2233 __do_request(mdsc, req);
2238 * Wake up threads with requests pending for @mds, so that they can
2239 * resubmit their requests to a possibly different mds.
2241 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2243 struct ceph_mds_request *req;
2244 struct rb_node *p = rb_first(&mdsc->request_tree);
2246 dout("kick_requests mds%d\n", mds);
2248 req = rb_entry(p, struct ceph_mds_request, r_node);
2250 if (req->r_got_unsafe)
2252 if (req->r_attempts > 0)
2253 continue; /* only new requests */
2254 if (req->r_session &&
2255 req->r_session->s_mds == mds) {
2256 dout(" kicking tid %llu\n", req->r_tid);
2257 list_del_init(&req->r_wait);
2258 __do_request(mdsc, req);
2263 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2264 struct ceph_mds_request *req)
2266 dout("submit_request on %p\n", req);
2267 mutex_lock(&mdsc->mutex);
2268 __register_request(mdsc, req, NULL);
2269 __do_request(mdsc, req);
2270 mutex_unlock(&mdsc->mutex);
2274 * Synchrously perform an mds request. Take care of all of the
2275 * session setup, forwarding, retry details.
2277 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2279 struct ceph_mds_request *req)
2283 dout("do_request on %p\n", req);
2285 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2287 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2288 if (req->r_locked_dir)
2289 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2290 if (req->r_old_dentry_dir)
2291 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2295 mutex_lock(&mdsc->mutex);
2296 __register_request(mdsc, req, dir);
2297 __do_request(mdsc, req);
2305 mutex_unlock(&mdsc->mutex);
2306 dout("do_request waiting\n");
2307 if (!req->r_timeout && req->r_wait_for_completion) {
2308 err = req->r_wait_for_completion(mdsc, req);
2310 long timeleft = wait_for_completion_killable_timeout(
2312 ceph_timeout_jiffies(req->r_timeout));
2316 err = -EIO; /* timed out */
2318 err = timeleft; /* killed */
2320 dout("do_request waited, got %d\n", err);
2321 mutex_lock(&mdsc->mutex);
2323 /* only abort if we didn't race with a real reply */
2324 if (req->r_got_result) {
2325 err = le32_to_cpu(req->r_reply_info.head->result);
2326 } else if (err < 0) {
2327 dout("aborted request %lld with %d\n", req->r_tid, err);
2330 * ensure we aren't running concurrently with
2331 * ceph_fill_trace or ceph_readdir_prepopulate, which
2332 * rely on locks (dir mutex) held by our caller.
2334 mutex_lock(&req->r_fill_mutex);
2336 req->r_aborted = true;
2337 mutex_unlock(&req->r_fill_mutex);
2339 if (req->r_locked_dir &&
2340 (req->r_op & CEPH_MDS_OP_WRITE))
2341 ceph_invalidate_dir_request(req);
2347 mutex_unlock(&mdsc->mutex);
2348 dout("do_request %p done, result %d\n", req, err);
2353 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2354 * namespace request.
2356 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2358 struct inode *inode = req->r_locked_dir;
2360 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2362 ceph_dir_clear_complete(inode);
2364 ceph_invalidate_dentry_lease(req->r_dentry);
2365 if (req->r_old_dentry)
2366 ceph_invalidate_dentry_lease(req->r_old_dentry);
2372 * We take the session mutex and parse and process the reply immediately.
2373 * This preserves the logical ordering of replies, capabilities, etc., sent
2374 * by the MDS as they are applied to our local cache.
2376 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2378 struct ceph_mds_client *mdsc = session->s_mdsc;
2379 struct ceph_mds_request *req;
2380 struct ceph_mds_reply_head *head = msg->front.iov_base;
2381 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2382 struct ceph_snap_realm *realm;
2385 int mds = session->s_mds;
2387 if (msg->front.iov_len < sizeof(*head)) {
2388 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2393 /* get request, session */
2394 tid = le64_to_cpu(msg->hdr.tid);
2395 mutex_lock(&mdsc->mutex);
2396 req = __lookup_request(mdsc, tid);
2398 dout("handle_reply on unknown tid %llu\n", tid);
2399 mutex_unlock(&mdsc->mutex);
2402 dout("handle_reply %p\n", req);
2404 /* correct session? */
2405 if (req->r_session != session) {
2406 pr_err("mdsc_handle_reply got %llu on session mds%d"
2407 " not mds%d\n", tid, session->s_mds,
2408 req->r_session ? req->r_session->s_mds : -1);
2409 mutex_unlock(&mdsc->mutex);
2414 if ((req->r_got_unsafe && !head->safe) ||
2415 (req->r_got_safe && head->safe)) {
2416 pr_warn("got a dup %s reply on %llu from mds%d\n",
2417 head->safe ? "safe" : "unsafe", tid, mds);
2418 mutex_unlock(&mdsc->mutex);
2421 if (req->r_got_safe) {
2422 pr_warn("got unsafe after safe on %llu from mds%d\n",
2424 mutex_unlock(&mdsc->mutex);
2428 result = le32_to_cpu(head->result);
2432 * if we're not talking to the authority, send to them
2433 * if the authority has changed while we weren't looking,
2434 * send to new authority
2435 * Otherwise we just have to return an ESTALE
2437 if (result == -ESTALE) {
2438 dout("got ESTALE on request %llu", req->r_tid);
2439 req->r_resend_mds = -1;
2440 if (req->r_direct_mode != USE_AUTH_MDS) {
2441 dout("not using auth, setting for that now");
2442 req->r_direct_mode = USE_AUTH_MDS;
2443 __do_request(mdsc, req);
2444 mutex_unlock(&mdsc->mutex);
2447 int mds = __choose_mds(mdsc, req);
2448 if (mds >= 0 && mds != req->r_session->s_mds) {
2449 dout("but auth changed, so resending");
2450 __do_request(mdsc, req);
2451 mutex_unlock(&mdsc->mutex);
2455 dout("have to return ESTALE on request %llu", req->r_tid);
2460 req->r_got_safe = true;
2461 __unregister_request(mdsc, req);
2463 if (req->r_got_unsafe) {
2465 * We already handled the unsafe response, now do the
2466 * cleanup. No need to examine the response; the MDS
2467 * doesn't include any result info in the safe
2468 * response. And even if it did, there is nothing
2469 * useful we could do with a revised return value.
2471 dout("got safe reply %llu, mds%d\n", tid, mds);
2472 list_del_init(&req->r_unsafe_item);
2474 /* last unsafe request during umount? */
2475 if (mdsc->stopping && !__get_oldest_req(mdsc))
2476 complete_all(&mdsc->safe_umount_waiters);
2477 mutex_unlock(&mdsc->mutex);
2481 req->r_got_unsafe = true;
2482 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2483 if (req->r_unsafe_dir) {
2484 struct ceph_inode_info *ci =
2485 ceph_inode(req->r_unsafe_dir);
2486 spin_lock(&ci->i_unsafe_lock);
2487 list_add_tail(&req->r_unsafe_dir_item,
2488 &ci->i_unsafe_dirops);
2489 spin_unlock(&ci->i_unsafe_lock);
2493 dout("handle_reply tid %lld result %d\n", tid, result);
2494 rinfo = &req->r_reply_info;
2495 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2496 mutex_unlock(&mdsc->mutex);
2498 mutex_lock(&session->s_mutex);
2500 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2507 if (rinfo->snapblob_len) {
2508 down_write(&mdsc->snap_rwsem);
2509 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2510 rinfo->snapblob + rinfo->snapblob_len,
2511 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2513 downgrade_write(&mdsc->snap_rwsem);
2515 down_read(&mdsc->snap_rwsem);
2518 /* insert trace into our cache */
2519 mutex_lock(&req->r_fill_mutex);
2520 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2522 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2523 req->r_op == CEPH_MDS_OP_LSSNAP))
2524 ceph_readdir_prepopulate(req, req->r_session);
2525 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2527 mutex_unlock(&req->r_fill_mutex);
2529 up_read(&mdsc->snap_rwsem);
2531 ceph_put_snap_realm(mdsc, realm);
2533 mutex_lock(&mdsc->mutex);
2534 if (!req->r_aborted) {
2538 req->r_reply = ceph_msg_get(msg);
2539 req->r_got_result = true;
2542 dout("reply arrived after request %lld was aborted\n", tid);
2544 mutex_unlock(&mdsc->mutex);
2546 mutex_unlock(&session->s_mutex);
2548 /* kick calling process */
2549 complete_request(mdsc, req);
2551 ceph_mdsc_put_request(req);
2558 * handle mds notification that our request has been forwarded.
2560 static void handle_forward(struct ceph_mds_client *mdsc,
2561 struct ceph_mds_session *session,
2562 struct ceph_msg *msg)
2564 struct ceph_mds_request *req;
2565 u64 tid = le64_to_cpu(msg->hdr.tid);
2569 void *p = msg->front.iov_base;
2570 void *end = p + msg->front.iov_len;
2572 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2573 next_mds = ceph_decode_32(&p);
2574 fwd_seq = ceph_decode_32(&p);
2576 mutex_lock(&mdsc->mutex);
2577 req = __lookup_request(mdsc, tid);
2579 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2580 goto out; /* dup reply? */
2583 if (req->r_aborted) {
2584 dout("forward tid %llu aborted, unregistering\n", tid);
2585 __unregister_request(mdsc, req);
2586 } else if (fwd_seq <= req->r_num_fwd) {
2587 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2588 tid, next_mds, req->r_num_fwd, fwd_seq);
2590 /* resend. forward race not possible; mds would drop */
2591 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2593 BUG_ON(req->r_got_result);
2594 req->r_attempts = 0;
2595 req->r_num_fwd = fwd_seq;
2596 req->r_resend_mds = next_mds;
2597 put_request_session(req);
2598 __do_request(mdsc, req);
2600 ceph_mdsc_put_request(req);
2602 mutex_unlock(&mdsc->mutex);
2606 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2610 * handle a mds session control message
2612 static void handle_session(struct ceph_mds_session *session,
2613 struct ceph_msg *msg)
2615 struct ceph_mds_client *mdsc = session->s_mdsc;
2618 int mds = session->s_mds;
2619 struct ceph_mds_session_head *h = msg->front.iov_base;
2623 if (msg->front.iov_len != sizeof(*h))
2625 op = le32_to_cpu(h->op);
2626 seq = le64_to_cpu(h->seq);
2628 mutex_lock(&mdsc->mutex);
2629 if (op == CEPH_SESSION_CLOSE)
2630 __unregister_session(mdsc, session);
2631 /* FIXME: this ttl calculation is generous */
2632 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2633 mutex_unlock(&mdsc->mutex);
2635 mutex_lock(&session->s_mutex);
2637 dout("handle_session mds%d %s %p state %s seq %llu\n",
2638 mds, ceph_session_op_name(op), session,
2639 ceph_session_state_name(session->s_state), seq);
2641 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2642 session->s_state = CEPH_MDS_SESSION_OPEN;
2643 pr_info("mds%d came back\n", session->s_mds);
2647 case CEPH_SESSION_OPEN:
2648 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2649 pr_info("mds%d reconnect success\n", session->s_mds);
2650 session->s_state = CEPH_MDS_SESSION_OPEN;
2651 renewed_caps(mdsc, session, 0);
2654 __close_session(mdsc, session);
2657 case CEPH_SESSION_RENEWCAPS:
2658 if (session->s_renew_seq == seq)
2659 renewed_caps(mdsc, session, 1);
2662 case CEPH_SESSION_CLOSE:
2663 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2664 pr_info("mds%d reconnect denied\n", session->s_mds);
2665 cleanup_session_requests(mdsc, session);
2666 remove_session_caps(session);
2667 wake = 2; /* for good measure */
2668 wake_up_all(&mdsc->session_close_wq);
2671 case CEPH_SESSION_STALE:
2672 pr_info("mds%d caps went stale, renewing\n",
2674 spin_lock(&session->s_gen_ttl_lock);
2675 session->s_cap_gen++;
2676 session->s_cap_ttl = jiffies - 1;
2677 spin_unlock(&session->s_gen_ttl_lock);
2678 send_renew_caps(mdsc, session);
2681 case CEPH_SESSION_RECALL_STATE:
2682 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2685 case CEPH_SESSION_FLUSHMSG:
2686 send_flushmsg_ack(mdsc, session, seq);
2689 case CEPH_SESSION_FORCE_RO:
2690 dout("force_session_readonly %p\n", session);
2691 spin_lock(&session->s_cap_lock);
2692 session->s_readonly = true;
2693 spin_unlock(&session->s_cap_lock);
2694 wake_up_session_caps(session, 0);
2698 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2702 mutex_unlock(&session->s_mutex);
2704 mutex_lock(&mdsc->mutex);
2705 __wake_requests(mdsc, &session->s_waiting);
2707 kick_requests(mdsc, mds);
2708 mutex_unlock(&mdsc->mutex);
2713 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2714 (int)msg->front.iov_len);
2721 * called under session->mutex.
2723 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2724 struct ceph_mds_session *session)
2726 struct ceph_mds_request *req, *nreq;
2730 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2732 mutex_lock(&mdsc->mutex);
2733 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2734 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2736 ceph_msg_get(req->r_request);
2737 ceph_con_send(&session->s_con, req->r_request);
2742 * also re-send old requests when MDS enters reconnect stage. So that MDS
2743 * can process completed request in clientreplay stage.
2745 p = rb_first(&mdsc->request_tree);
2747 req = rb_entry(p, struct ceph_mds_request, r_node);
2749 if (req->r_got_unsafe)
2751 if (req->r_attempts == 0)
2752 continue; /* only old requests */
2753 if (req->r_session &&
2754 req->r_session->s_mds == session->s_mds) {
2755 err = __prepare_send_request(mdsc, req,
2756 session->s_mds, true);
2758 ceph_msg_get(req->r_request);
2759 ceph_con_send(&session->s_con, req->r_request);
2763 mutex_unlock(&mdsc->mutex);
2767 * Encode information about a cap for a reconnect with the MDS.
2769 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2773 struct ceph_mds_cap_reconnect v2;
2774 struct ceph_mds_cap_reconnect_v1 v1;
2777 struct ceph_inode_info *ci;
2778 struct ceph_reconnect_state *recon_state = arg;
2779 struct ceph_pagelist *pagelist = recon_state->pagelist;
2783 struct dentry *dentry;
2787 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2788 inode, ceph_vinop(inode), cap, cap->cap_id,
2789 ceph_cap_string(cap->issued));
2790 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2794 dentry = d_find_alias(inode);
2796 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2798 err = PTR_ERR(path);
2805 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2809 spin_lock(&ci->i_ceph_lock);
2810 cap->seq = 0; /* reset cap seq */
2811 cap->issue_seq = 0; /* and issue_seq */
2812 cap->mseq = 0; /* and migrate_seq */
2813 cap->cap_gen = cap->session->s_cap_gen;
2815 if (recon_state->flock) {
2816 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2817 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2818 rec.v2.issued = cpu_to_le32(cap->issued);
2819 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2820 rec.v2.pathbase = cpu_to_le64(pathbase);
2821 rec.v2.flock_len = 0;
2822 reclen = sizeof(rec.v2);
2824 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2825 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2826 rec.v1.issued = cpu_to_le32(cap->issued);
2827 rec.v1.size = cpu_to_le64(inode->i_size);
2828 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2829 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2830 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2831 rec.v1.pathbase = cpu_to_le64(pathbase);
2832 reclen = sizeof(rec.v1);
2834 spin_unlock(&ci->i_ceph_lock);
2836 if (recon_state->flock) {
2837 int num_fcntl_locks, num_flock_locks;
2838 struct ceph_filelock *flocks;
2841 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2842 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2843 sizeof(struct ceph_filelock), GFP_NOFS);
2848 err = ceph_encode_locks_to_buffer(inode, flocks,
2858 * number of encoded locks is stable, so copy to pagelist
2860 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2861 (num_fcntl_locks+num_flock_locks) *
2862 sizeof(struct ceph_filelock));
2863 err = ceph_pagelist_append(pagelist, &rec, reclen);
2865 err = ceph_locks_to_pagelist(flocks, pagelist,
2870 err = ceph_pagelist_append(pagelist, &rec, reclen);
2873 recon_state->nr_caps++;
2883 * If an MDS fails and recovers, clients need to reconnect in order to
2884 * reestablish shared state. This includes all caps issued through
2885 * this session _and_ the snap_realm hierarchy. Because it's not
2886 * clear which snap realms the mds cares about, we send everything we
2887 * know about.. that ensures we'll then get any new info the
2888 * recovering MDS might have.
2890 * This is a relatively heavyweight operation, but it's rare.
2892 * called with mdsc->mutex held.
2894 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2895 struct ceph_mds_session *session)
2897 struct ceph_msg *reply;
2899 int mds = session->s_mds;
2902 struct ceph_pagelist *pagelist;
2903 struct ceph_reconnect_state recon_state;
2905 pr_info("mds%d reconnect start\n", mds);
2907 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2909 goto fail_nopagelist;
2910 ceph_pagelist_init(pagelist);
2912 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2916 mutex_lock(&session->s_mutex);
2917 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2920 dout("session %p state %s\n", session,
2921 ceph_session_state_name(session->s_state));
2923 spin_lock(&session->s_gen_ttl_lock);
2924 session->s_cap_gen++;
2925 spin_unlock(&session->s_gen_ttl_lock);
2927 spin_lock(&session->s_cap_lock);
2928 /* don't know if session is readonly */
2929 session->s_readonly = 0;
2931 * notify __ceph_remove_cap() that we are composing cap reconnect.
2932 * If a cap get released before being added to the cap reconnect,
2933 * __ceph_remove_cap() should skip queuing cap release.
2935 session->s_cap_reconnect = 1;
2936 /* drop old cap expires; we're about to reestablish that state */
2937 cleanup_cap_releases(mdsc, session);
2939 /* trim unused caps to reduce MDS's cache rejoin time */
2940 if (mdsc->fsc->sb->s_root)
2941 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2943 ceph_con_close(&session->s_con);
2944 ceph_con_open(&session->s_con,
2945 CEPH_ENTITY_TYPE_MDS, mds,
2946 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2948 /* replay unsafe requests */
2949 replay_unsafe_requests(mdsc, session);
2951 down_read(&mdsc->snap_rwsem);
2953 /* traverse this session's caps */
2954 s_nr_caps = session->s_nr_caps;
2955 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2959 recon_state.nr_caps = 0;
2960 recon_state.pagelist = pagelist;
2961 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2962 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2966 spin_lock(&session->s_cap_lock);
2967 session->s_cap_reconnect = 0;
2968 spin_unlock(&session->s_cap_lock);
2971 * snaprealms. we provide mds with the ino, seq (version), and
2972 * parent for all of our realms. If the mds has any newer info,
2975 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2976 struct ceph_snap_realm *realm =
2977 rb_entry(p, struct ceph_snap_realm, node);
2978 struct ceph_mds_snaprealm_reconnect sr_rec;
2980 dout(" adding snap realm %llx seq %lld parent %llx\n",
2981 realm->ino, realm->seq, realm->parent_ino);
2982 sr_rec.ino = cpu_to_le64(realm->ino);
2983 sr_rec.seq = cpu_to_le64(realm->seq);
2984 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2985 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2990 if (recon_state.flock)
2991 reply->hdr.version = cpu_to_le16(2);
2993 /* raced with cap release? */
2994 if (s_nr_caps != recon_state.nr_caps) {
2995 struct page *page = list_first_entry(&pagelist->head,
2997 __le32 *addr = kmap_atomic(page);
2998 *addr = cpu_to_le32(recon_state.nr_caps);
2999 kunmap_atomic(addr);
3002 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3003 ceph_msg_data_add_pagelist(reply, pagelist);
3005 ceph_early_kick_flushing_caps(mdsc, session);
3007 ceph_con_send(&session->s_con, reply);
3009 mutex_unlock(&session->s_mutex);
3011 mutex_lock(&mdsc->mutex);
3012 __wake_requests(mdsc, &session->s_waiting);
3013 mutex_unlock(&mdsc->mutex);
3015 up_read(&mdsc->snap_rwsem);
3019 ceph_msg_put(reply);
3020 up_read(&mdsc->snap_rwsem);
3021 mutex_unlock(&session->s_mutex);
3023 ceph_pagelist_release(pagelist);
3025 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3031 * compare old and new mdsmaps, kicking requests
3032 * and closing out old connections as necessary
3034 * called under mdsc->mutex.
3036 static void check_new_map(struct ceph_mds_client *mdsc,
3037 struct ceph_mdsmap *newmap,
3038 struct ceph_mdsmap *oldmap)
3041 int oldstate, newstate;
3042 struct ceph_mds_session *s;
3044 dout("check_new_map new %u old %u\n",
3045 newmap->m_epoch, oldmap->m_epoch);
3047 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3048 if (mdsc->sessions[i] == NULL)
3050 s = mdsc->sessions[i];
3051 oldstate = ceph_mdsmap_get_state(oldmap, i);
3052 newstate = ceph_mdsmap_get_state(newmap, i);
3054 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3055 i, ceph_mds_state_name(oldstate),
3056 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3057 ceph_mds_state_name(newstate),
3058 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3059 ceph_session_state_name(s->s_state));
3061 if (i >= newmap->m_max_mds ||
3062 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3063 ceph_mdsmap_get_addr(newmap, i),
3064 sizeof(struct ceph_entity_addr))) {
3065 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3066 /* the session never opened, just close it
3068 __wake_requests(mdsc, &s->s_waiting);
3069 __unregister_session(mdsc, s);
3072 mutex_unlock(&mdsc->mutex);
3073 mutex_lock(&s->s_mutex);
3074 mutex_lock(&mdsc->mutex);
3075 ceph_con_close(&s->s_con);
3076 mutex_unlock(&s->s_mutex);
3077 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3079 } else if (oldstate == newstate) {
3080 continue; /* nothing new with this mds */
3086 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3087 newstate >= CEPH_MDS_STATE_RECONNECT) {
3088 mutex_unlock(&mdsc->mutex);
3089 send_mds_reconnect(mdsc, s);
3090 mutex_lock(&mdsc->mutex);
3094 * kick request on any mds that has gone active.
3096 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3097 newstate >= CEPH_MDS_STATE_ACTIVE) {
3098 if (oldstate != CEPH_MDS_STATE_CREATING &&
3099 oldstate != CEPH_MDS_STATE_STARTING)
3100 pr_info("mds%d recovery completed\n", s->s_mds);
3101 kick_requests(mdsc, i);
3102 ceph_kick_flushing_caps(mdsc, s);
3103 wake_up_session_caps(s, 1);
3107 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3108 s = mdsc->sessions[i];
3111 if (!ceph_mdsmap_is_laggy(newmap, i))
3113 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3114 s->s_state == CEPH_MDS_SESSION_HUNG ||
3115 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3116 dout(" connecting to export targets of laggy mds%d\n",
3118 __open_export_target_sessions(mdsc, s);
3130 * caller must hold session s_mutex, dentry->d_lock
3132 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3134 struct ceph_dentry_info *di = ceph_dentry(dentry);
3136 ceph_put_mds_session(di->lease_session);
3137 di->lease_session = NULL;
3140 static void handle_lease(struct ceph_mds_client *mdsc,
3141 struct ceph_mds_session *session,
3142 struct ceph_msg *msg)
3144 struct super_block *sb = mdsc->fsc->sb;
3145 struct inode *inode;
3146 struct dentry *parent, *dentry;
3147 struct ceph_dentry_info *di;
3148 int mds = session->s_mds;
3149 struct ceph_mds_lease *h = msg->front.iov_base;
3151 struct ceph_vino vino;
3155 dout("handle_lease from mds%d\n", mds);
3158 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3160 vino.ino = le64_to_cpu(h->ino);
3161 vino.snap = CEPH_NOSNAP;
3162 seq = le32_to_cpu(h->seq);
3163 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3164 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3165 if (dname.len != get_unaligned_le32(h+1))
3169 inode = ceph_find_inode(sb, vino);
3170 dout("handle_lease %s, ino %llx %p %.*s\n",
3171 ceph_lease_op_name(h->action), vino.ino, inode,
3172 dname.len, dname.name);
3174 mutex_lock(&session->s_mutex);
3177 if (inode == NULL) {
3178 dout("handle_lease no inode %llx\n", vino.ino);
3183 parent = d_find_alias(inode);
3185 dout("no parent dentry on inode %p\n", inode);
3187 goto release; /* hrm... */
3189 dname.hash = full_name_hash(dname.name, dname.len);
3190 dentry = d_lookup(parent, &dname);
3195 spin_lock(&dentry->d_lock);
3196 di = ceph_dentry(dentry);
3197 switch (h->action) {
3198 case CEPH_MDS_LEASE_REVOKE:
3199 if (di->lease_session == session) {
3200 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3201 h->seq = cpu_to_le32(di->lease_seq);
3202 __ceph_mdsc_drop_dentry_lease(dentry);
3207 case CEPH_MDS_LEASE_RENEW:
3208 if (di->lease_session == session &&
3209 di->lease_gen == session->s_cap_gen &&
3210 di->lease_renew_from &&
3211 di->lease_renew_after == 0) {
3212 unsigned long duration =
3213 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3215 di->lease_seq = seq;
3216 dentry->d_time = di->lease_renew_from + duration;
3217 di->lease_renew_after = di->lease_renew_from +
3219 di->lease_renew_from = 0;
3223 spin_unlock(&dentry->d_lock);
3230 /* let's just reuse the same message */
3231 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3233 ceph_con_send(&session->s_con, msg);
3237 mutex_unlock(&session->s_mutex);
3241 pr_err("corrupt lease message\n");
3245 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3246 struct inode *inode,
3247 struct dentry *dentry, char action,
3250 struct ceph_msg *msg;
3251 struct ceph_mds_lease *lease;
3252 int len = sizeof(*lease) + sizeof(u32);
3255 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3256 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3257 dnamelen = dentry->d_name.len;
3260 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3263 lease = msg->front.iov_base;
3264 lease->action = action;
3265 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3266 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3267 lease->seq = cpu_to_le32(seq);
3268 put_unaligned_le32(dnamelen, lease + 1);
3269 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3272 * if this is a preemptive lease RELEASE, no need to
3273 * flush request stream, since the actual request will
3276 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3278 ceph_con_send(&session->s_con, msg);
3282 * Preemptively release a lease we expect to invalidate anyway.
3283 * Pass @inode always, @dentry is optional.
3285 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3286 struct dentry *dentry)
3288 struct ceph_dentry_info *di;
3289 struct ceph_mds_session *session;
3292 BUG_ON(inode == NULL);
3293 BUG_ON(dentry == NULL);
3295 /* is dentry lease valid? */
3296 spin_lock(&dentry->d_lock);
3297 di = ceph_dentry(dentry);
3298 if (!di || !di->lease_session ||
3299 di->lease_session->s_mds < 0 ||
3300 di->lease_gen != di->lease_session->s_cap_gen ||
3301 !time_before(jiffies, dentry->d_time)) {
3302 dout("lease_release inode %p dentry %p -- "
3305 spin_unlock(&dentry->d_lock);
3309 /* we do have a lease on this dentry; note mds and seq */
3310 session = ceph_get_mds_session(di->lease_session);
3311 seq = di->lease_seq;
3312 __ceph_mdsc_drop_dentry_lease(dentry);
3313 spin_unlock(&dentry->d_lock);
3315 dout("lease_release inode %p dentry %p to mds%d\n",
3316 inode, dentry, session->s_mds);
3317 ceph_mdsc_lease_send_msg(session, inode, dentry,
3318 CEPH_MDS_LEASE_RELEASE, seq);
3319 ceph_put_mds_session(session);
3323 * drop all leases (and dentry refs) in preparation for umount
3325 static void drop_leases(struct ceph_mds_client *mdsc)
3329 dout("drop_leases\n");
3330 mutex_lock(&mdsc->mutex);
3331 for (i = 0; i < mdsc->max_sessions; i++) {
3332 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3335 mutex_unlock(&mdsc->mutex);
3336 mutex_lock(&s->s_mutex);
3337 mutex_unlock(&s->s_mutex);
3338 ceph_put_mds_session(s);
3339 mutex_lock(&mdsc->mutex);
3341 mutex_unlock(&mdsc->mutex);
3347 * delayed work -- periodically trim expired leases, renew caps with mds
3349 static void schedule_delayed(struct ceph_mds_client *mdsc)
3352 unsigned hz = round_jiffies_relative(HZ * delay);
3353 schedule_delayed_work(&mdsc->delayed_work, hz);
3356 static void delayed_work(struct work_struct *work)
3359 struct ceph_mds_client *mdsc =
3360 container_of(work, struct ceph_mds_client, delayed_work.work);
3364 dout("mdsc delayed_work\n");
3365 ceph_check_delayed_caps(mdsc);
3367 mutex_lock(&mdsc->mutex);
3368 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3369 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3370 mdsc->last_renew_caps);
3372 mdsc->last_renew_caps = jiffies;
3374 for (i = 0; i < mdsc->max_sessions; i++) {
3375 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3378 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3379 dout("resending session close request for mds%d\n",
3381 request_close_session(mdsc, s);
3382 ceph_put_mds_session(s);
3385 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3386 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3387 s->s_state = CEPH_MDS_SESSION_HUNG;
3388 pr_info("mds%d hung\n", s->s_mds);
3391 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3392 /* this mds is failed or recovering, just wait */
3393 ceph_put_mds_session(s);
3396 mutex_unlock(&mdsc->mutex);
3398 mutex_lock(&s->s_mutex);
3400 send_renew_caps(mdsc, s);
3402 ceph_con_keepalive(&s->s_con);
3403 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3404 s->s_state == CEPH_MDS_SESSION_HUNG)
3405 ceph_send_cap_releases(mdsc, s);
3406 mutex_unlock(&s->s_mutex);
3407 ceph_put_mds_session(s);
3409 mutex_lock(&mdsc->mutex);
3411 mutex_unlock(&mdsc->mutex);
3413 schedule_delayed(mdsc);
3416 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3419 struct ceph_mds_client *mdsc;
3421 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3426 mutex_init(&mdsc->mutex);
3427 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3428 if (mdsc->mdsmap == NULL) {
3433 init_completion(&mdsc->safe_umount_waiters);
3434 init_waitqueue_head(&mdsc->session_close_wq);
3435 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3436 mdsc->sessions = NULL;
3437 atomic_set(&mdsc->num_sessions, 0);
3438 mdsc->max_sessions = 0;
3440 mdsc->last_snap_seq = 0;
3441 init_rwsem(&mdsc->snap_rwsem);
3442 mdsc->snap_realms = RB_ROOT;
3443 INIT_LIST_HEAD(&mdsc->snap_empty);
3444 spin_lock_init(&mdsc->snap_empty_lock);
3446 mdsc->oldest_tid = 0;
3447 mdsc->request_tree = RB_ROOT;
3448 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3449 mdsc->last_renew_caps = jiffies;
3450 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3451 spin_lock_init(&mdsc->cap_delay_lock);
3452 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3453 spin_lock_init(&mdsc->snap_flush_lock);
3454 mdsc->last_cap_flush_tid = 1;
3455 mdsc->cap_flush_tree = RB_ROOT;
3456 INIT_LIST_HEAD(&mdsc->cap_dirty);
3457 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3458 mdsc->num_cap_flushing = 0;
3459 spin_lock_init(&mdsc->cap_dirty_lock);
3460 init_waitqueue_head(&mdsc->cap_flushing_wq);
3461 spin_lock_init(&mdsc->dentry_lru_lock);
3462 INIT_LIST_HEAD(&mdsc->dentry_lru);
3464 ceph_caps_init(mdsc);
3465 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3467 init_rwsem(&mdsc->pool_perm_rwsem);
3468 mdsc->pool_perm_tree = RB_ROOT;
3474 * Wait for safe replies on open mds requests. If we time out, drop
3475 * all requests from the tree to avoid dangling dentry refs.
3477 static void wait_requests(struct ceph_mds_client *mdsc)
3479 struct ceph_options *opts = mdsc->fsc->client->options;
3480 struct ceph_mds_request *req;
3482 mutex_lock(&mdsc->mutex);
3483 if (__get_oldest_req(mdsc)) {
3484 mutex_unlock(&mdsc->mutex);
3486 dout("wait_requests waiting for requests\n");
3487 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3488 ceph_timeout_jiffies(opts->mount_timeout));
3490 /* tear down remaining requests */
3491 mutex_lock(&mdsc->mutex);
3492 while ((req = __get_oldest_req(mdsc))) {
3493 dout("wait_requests timed out on tid %llu\n",
3495 __unregister_request(mdsc, req);
3498 mutex_unlock(&mdsc->mutex);
3499 dout("wait_requests done\n");
3503 * called before mount is ro, and before dentries are torn down.
3504 * (hmm, does this still race with new lookups?)
3506 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3508 dout("pre_umount\n");
3512 ceph_flush_dirty_caps(mdsc);
3513 wait_requests(mdsc);
3516 * wait for reply handlers to drop their request refs and
3517 * their inode/dcache refs
3523 * wait for all write mds requests to flush.
3525 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3527 struct ceph_mds_request *req = NULL, *nextreq;
3530 mutex_lock(&mdsc->mutex);
3531 dout("wait_unsafe_requests want %lld\n", want_tid);
3533 req = __get_oldest_req(mdsc);
3534 while (req && req->r_tid <= want_tid) {
3535 /* find next request */
3536 n = rb_next(&req->r_node);
3538 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3541 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3542 (req->r_op & CEPH_MDS_OP_WRITE)) {
3544 ceph_mdsc_get_request(req);
3546 ceph_mdsc_get_request(nextreq);
3547 mutex_unlock(&mdsc->mutex);
3548 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3549 req->r_tid, want_tid);
3550 wait_for_completion(&req->r_safe_completion);
3551 mutex_lock(&mdsc->mutex);
3552 ceph_mdsc_put_request(req);
3554 break; /* next dne before, so we're done! */
3555 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3556 /* next request was removed from tree */
3557 ceph_mdsc_put_request(nextreq);
3560 ceph_mdsc_put_request(nextreq); /* won't go away */
3564 mutex_unlock(&mdsc->mutex);
3565 dout("wait_unsafe_requests done\n");
3568 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3570 u64 want_tid, want_flush, want_snap;
3572 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3576 mutex_lock(&mdsc->mutex);
3577 want_tid = mdsc->last_tid;
3578 mutex_unlock(&mdsc->mutex);
3580 ceph_flush_dirty_caps(mdsc);
3581 spin_lock(&mdsc->cap_dirty_lock);
3582 want_flush = mdsc->last_cap_flush_tid;
3583 spin_unlock(&mdsc->cap_dirty_lock);
3585 down_read(&mdsc->snap_rwsem);
3586 want_snap = mdsc->last_snap_seq;
3587 up_read(&mdsc->snap_rwsem);
3589 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3590 want_tid, want_flush, want_snap);
3592 wait_unsafe_requests(mdsc, want_tid);
3593 wait_caps_flush(mdsc, want_flush, want_snap);
3597 * true if all sessions are closed, or we force unmount
3599 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3601 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3603 return atomic_read(&mdsc->num_sessions) == 0;
3607 * called after sb is ro.
3609 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3611 struct ceph_options *opts = mdsc->fsc->client->options;
3612 struct ceph_mds_session *session;
3615 dout("close_sessions\n");
3617 /* close sessions */
3618 mutex_lock(&mdsc->mutex);
3619 for (i = 0; i < mdsc->max_sessions; i++) {
3620 session = __ceph_lookup_mds_session(mdsc, i);
3623 mutex_unlock(&mdsc->mutex);
3624 mutex_lock(&session->s_mutex);
3625 __close_session(mdsc, session);
3626 mutex_unlock(&session->s_mutex);
3627 ceph_put_mds_session(session);
3628 mutex_lock(&mdsc->mutex);
3630 mutex_unlock(&mdsc->mutex);
3632 dout("waiting for sessions to close\n");
3633 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3634 ceph_timeout_jiffies(opts->mount_timeout));
3636 /* tear down remaining sessions */
3637 mutex_lock(&mdsc->mutex);
3638 for (i = 0; i < mdsc->max_sessions; i++) {
3639 if (mdsc->sessions[i]) {
3640 session = get_session(mdsc->sessions[i]);
3641 __unregister_session(mdsc, session);
3642 mutex_unlock(&mdsc->mutex);
3643 mutex_lock(&session->s_mutex);
3644 remove_session_caps(session);
3645 mutex_unlock(&session->s_mutex);
3646 ceph_put_mds_session(session);
3647 mutex_lock(&mdsc->mutex);
3650 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3651 mutex_unlock(&mdsc->mutex);
3653 ceph_cleanup_empty_realms(mdsc);
3655 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3660 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3662 struct ceph_mds_session *session;
3665 dout("force umount\n");
3667 mutex_lock(&mdsc->mutex);
3668 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3669 session = __ceph_lookup_mds_session(mdsc, mds);
3672 mutex_unlock(&mdsc->mutex);
3673 mutex_lock(&session->s_mutex);
3674 __close_session(mdsc, session);
3675 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3676 cleanup_session_requests(mdsc, session);
3677 remove_session_caps(session);
3679 mutex_unlock(&session->s_mutex);
3680 ceph_put_mds_session(session);
3681 mutex_lock(&mdsc->mutex);
3682 kick_requests(mdsc, mds);
3684 __wake_requests(mdsc, &mdsc->waiting_for_map);
3685 mutex_unlock(&mdsc->mutex);
3688 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3691 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3693 ceph_mdsmap_destroy(mdsc->mdsmap);
3694 kfree(mdsc->sessions);
3695 ceph_caps_finalize(mdsc);
3696 ceph_pool_perm_destroy(mdsc);
3699 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3701 struct ceph_mds_client *mdsc = fsc->mdsc;
3703 dout("mdsc_destroy %p\n", mdsc);
3704 ceph_mdsc_stop(mdsc);
3706 /* flush out any connection work with references to us */
3711 dout("mdsc_destroy %p done\n", mdsc);
3716 * handle mds map update.
3718 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3722 void *p = msg->front.iov_base;
3723 void *end = p + msg->front.iov_len;
3724 struct ceph_mdsmap *newmap, *oldmap;
3725 struct ceph_fsid fsid;
3728 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3729 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3730 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3732 epoch = ceph_decode_32(&p);
3733 maplen = ceph_decode_32(&p);
3734 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3736 /* do we need it? */
3737 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3738 mutex_lock(&mdsc->mutex);
3739 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3740 dout("handle_map epoch %u <= our %u\n",
3741 epoch, mdsc->mdsmap->m_epoch);
3742 mutex_unlock(&mdsc->mutex);
3746 newmap = ceph_mdsmap_decode(&p, end);
3747 if (IS_ERR(newmap)) {
3748 err = PTR_ERR(newmap);
3752 /* swap into place */
3754 oldmap = mdsc->mdsmap;
3755 mdsc->mdsmap = newmap;
3756 check_new_map(mdsc, newmap, oldmap);
3757 ceph_mdsmap_destroy(oldmap);
3759 mdsc->mdsmap = newmap; /* first mds map */
3761 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3763 __wake_requests(mdsc, &mdsc->waiting_for_map);
3765 mutex_unlock(&mdsc->mutex);
3766 schedule_delayed(mdsc);
3770 mutex_unlock(&mdsc->mutex);
3772 pr_err("error decoding mdsmap %d\n", err);
3776 static struct ceph_connection *con_get(struct ceph_connection *con)
3778 struct ceph_mds_session *s = con->private;
3780 if (get_session(s)) {
3781 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3784 dout("mdsc con_get %p FAIL\n", s);
3788 static void con_put(struct ceph_connection *con)
3790 struct ceph_mds_session *s = con->private;
3792 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3793 ceph_put_mds_session(s);
3797 * if the client is unresponsive for long enough, the mds will kill
3798 * the session entirely.
3800 static void peer_reset(struct ceph_connection *con)
3802 struct ceph_mds_session *s = con->private;
3803 struct ceph_mds_client *mdsc = s->s_mdsc;
3805 pr_warn("mds%d closed our session\n", s->s_mds);
3806 send_mds_reconnect(mdsc, s);
3809 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3811 struct ceph_mds_session *s = con->private;
3812 struct ceph_mds_client *mdsc = s->s_mdsc;
3813 int type = le16_to_cpu(msg->hdr.type);
3815 mutex_lock(&mdsc->mutex);
3816 if (__verify_registered_session(mdsc, s) < 0) {
3817 mutex_unlock(&mdsc->mutex);
3820 mutex_unlock(&mdsc->mutex);
3823 case CEPH_MSG_MDS_MAP:
3824 ceph_mdsc_handle_map(mdsc, msg);
3826 case CEPH_MSG_CLIENT_SESSION:
3827 handle_session(s, msg);
3829 case CEPH_MSG_CLIENT_REPLY:
3830 handle_reply(s, msg);
3832 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3833 handle_forward(mdsc, s, msg);
3835 case CEPH_MSG_CLIENT_CAPS:
3836 ceph_handle_caps(s, msg);
3838 case CEPH_MSG_CLIENT_SNAP:
3839 ceph_handle_snap(mdsc, s, msg);
3841 case CEPH_MSG_CLIENT_LEASE:
3842 handle_lease(mdsc, s, msg);
3846 pr_err("received unknown message type %d %s\n", type,
3847 ceph_msg_type_name(type));
3858 * Note: returned pointer is the address of a structure that's
3859 * managed separately. Caller must *not* attempt to free it.
3861 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3862 int *proto, int force_new)
3864 struct ceph_mds_session *s = con->private;
3865 struct ceph_mds_client *mdsc = s->s_mdsc;
3866 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3867 struct ceph_auth_handshake *auth = &s->s_auth;
3869 if (force_new && auth->authorizer) {
3870 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3871 auth->authorizer = NULL;
3873 if (!auth->authorizer) {
3874 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3877 return ERR_PTR(ret);
3879 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3882 return ERR_PTR(ret);
3884 *proto = ac->protocol;
3890 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3892 struct ceph_mds_session *s = con->private;
3893 struct ceph_mds_client *mdsc = s->s_mdsc;
3894 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3896 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3899 static int invalidate_authorizer(struct ceph_connection *con)
3901 struct ceph_mds_session *s = con->private;
3902 struct ceph_mds_client *mdsc = s->s_mdsc;
3903 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3905 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3907 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3910 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3911 struct ceph_msg_header *hdr, int *skip)
3913 struct ceph_msg *msg;
3914 int type = (int) le16_to_cpu(hdr->type);
3915 int front_len = (int) le32_to_cpu(hdr->front_len);
3921 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3923 pr_err("unable to allocate msg type %d len %d\n",
3931 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3933 struct ceph_mds_session *s = con->private;
3934 struct ceph_auth_handshake *auth = &s->s_auth;
3935 return ceph_auth_sign_message(auth, msg);
3938 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3940 struct ceph_mds_session *s = con->private;
3941 struct ceph_auth_handshake *auth = &s->s_auth;
3942 return ceph_auth_check_message_signature(auth, msg);
3945 static const struct ceph_connection_operations mds_con_ops = {
3948 .dispatch = dispatch,
3949 .get_authorizer = get_authorizer,
3950 .verify_authorizer_reply = verify_authorizer_reply,
3951 .invalidate_authorizer = invalidate_authorizer,
3952 .peer_reset = peer_reset,
3953 .alloc_msg = mds_alloc_msg,
3954 .sign_message = sign_message,
3955 .check_message_signature = check_message_signature,