1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
6 #include <linux/sched.h>
7 #include <linux/debugfs.h>
8 #include <linux/seq_file.h>
11 #include "mds_client.h"
13 #include <linux/ceph/ceph_features.h>
14 #include <linux/ceph/messenger.h>
15 #include <linux/ceph/decode.h>
16 #include <linux/ceph/pagelist.h>
17 #include <linux/ceph/auth.h>
18 #include <linux/ceph/debugfs.h>
21 * A cluster of MDS (metadata server) daemons is responsible for
22 * managing the file system namespace (the directory hierarchy and
23 * inodes) and for coordinating shared access to storage. Metadata is
24 * partitioning hierarchically across a number of servers, and that
25 * partition varies over time as the cluster adjusts the distribution
26 * in order to balance load.
28 * The MDS client is primarily responsible to managing synchronous
29 * metadata requests for operations like open, unlink, and so forth.
30 * If there is a MDS failure, we find out about it when we (possibly
31 * request and) receive a new MDS map, and can resubmit affected
34 * For the most part, though, we take advantage of a lossless
35 * communications channel to the MDS, and do not need to worry about
36 * timing out or resubmitting requests.
38 * We maintain a stateful "session" with each MDS we interact with.
39 * Within each session, we sent periodic heartbeat messages to ensure
40 * any capabilities or leases we have been issues remain valid. If
41 * the session times out and goes stale, our leases and capabilities
42 * are no longer valid.
45 struct ceph_reconnect_state {
46 struct ceph_pagelist *pagelist;
50 static void __wake_requests(struct ceph_mds_client *mdsc,
51 struct list_head *head);
53 static const struct ceph_connection_operations mds_con_ops;
61 * parse individual inode info
63 static int parse_reply_info_in(void **p, void *end,
64 struct ceph_mds_reply_info_in *info,
70 *p += sizeof(struct ceph_mds_reply_inode) +
71 sizeof(*info->in->fragtree.splits) *
72 le32_to_cpu(info->in->fragtree.nsplits);
74 ceph_decode_32_safe(p, end, info->symlink_len, bad);
75 ceph_decode_need(p, end, info->symlink_len, bad);
77 *p += info->symlink_len;
79 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
80 ceph_decode_copy_safe(p, end, &info->dir_layout,
81 sizeof(info->dir_layout), bad);
83 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
85 ceph_decode_32_safe(p, end, info->xattr_len, bad);
86 ceph_decode_need(p, end, info->xattr_len, bad);
87 info->xattr_data = *p;
88 *p += info->xattr_len;
95 * parse a normal reply, which may contain a (dir+)dentry and/or a
98 static int parse_reply_info_trace(void **p, void *end,
99 struct ceph_mds_reply_info_parsed *info,
104 if (info->head->is_dentry) {
105 err = parse_reply_info_in(p, end, &info->diri, features);
109 if (unlikely(*p + sizeof(*info->dirfrag) > end))
112 *p += sizeof(*info->dirfrag) +
113 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
114 if (unlikely(*p > end))
117 ceph_decode_32_safe(p, end, info->dname_len, bad);
118 ceph_decode_need(p, end, info->dname_len, bad);
120 *p += info->dname_len;
122 *p += sizeof(*info->dlease);
125 if (info->head->is_target) {
126 err = parse_reply_info_in(p, end, &info->targeti, features);
131 if (unlikely(*p != end))
138 pr_err("problem parsing mds trace %d\n", err);
143 * parse readdir results
145 static int parse_reply_info_dir(void **p, void *end,
146 struct ceph_mds_reply_info_parsed *info,
153 if (*p + sizeof(*info->dir_dir) > end)
155 *p += sizeof(*info->dir_dir) +
156 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
160 ceph_decode_need(p, end, sizeof(num) + 2, bad);
161 num = ceph_decode_32(p);
162 info->dir_end = ceph_decode_8(p);
163 info->dir_complete = ceph_decode_8(p);
167 /* alloc large array */
169 info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
170 sizeof(*info->dir_dname) +
171 sizeof(*info->dir_dname_len) +
172 sizeof(*info->dir_dlease),
174 if (info->dir_in == NULL) {
178 info->dir_dname = (void *)(info->dir_in + num);
179 info->dir_dname_len = (void *)(info->dir_dname + num);
180 info->dir_dlease = (void *)(info->dir_dname_len + num);
184 ceph_decode_need(p, end, sizeof(u32)*2, bad);
185 info->dir_dname_len[i] = ceph_decode_32(p);
186 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
187 info->dir_dname[i] = *p;
188 *p += info->dir_dname_len[i];
189 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
191 info->dir_dlease[i] = *p;
192 *p += sizeof(struct ceph_mds_reply_lease);
195 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
210 pr_err("problem parsing dir contents %d\n", err);
215 * parse fcntl F_GETLK results
217 static int parse_reply_info_filelock(void **p, void *end,
218 struct ceph_mds_reply_info_parsed *info,
221 if (*p + sizeof(*info->filelock_reply) > end)
224 info->filelock_reply = *p;
225 *p += sizeof(*info->filelock_reply);
227 if (unlikely(*p != end))
236 * parse create results
238 static int parse_reply_info_create(void **p, void *end,
239 struct ceph_mds_reply_info_parsed *info,
242 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
244 info->has_create_ino = false;
246 info->has_create_ino = true;
247 info->ino = ceph_decode_64(p);
251 if (unlikely(*p != end))
260 * parse extra results
262 static int parse_reply_info_extra(void **p, void *end,
263 struct ceph_mds_reply_info_parsed *info,
266 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
267 return parse_reply_info_filelock(p, end, info, features);
268 else if (info->head->op == CEPH_MDS_OP_READDIR ||
269 info->head->op == CEPH_MDS_OP_LSSNAP)
270 return parse_reply_info_dir(p, end, info, features);
271 else if (info->head->op == CEPH_MDS_OP_CREATE)
272 return parse_reply_info_create(p, end, info, features);
278 * parse entire mds reply
280 static int parse_reply_info(struct ceph_msg *msg,
281 struct ceph_mds_reply_info_parsed *info,
288 info->head = msg->front.iov_base;
289 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
290 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
293 ceph_decode_32_safe(&p, end, len, bad);
295 ceph_decode_need(&p, end, len, bad);
296 err = parse_reply_info_trace(&p, p+len, info, features);
302 ceph_decode_32_safe(&p, end, len, bad);
304 ceph_decode_need(&p, end, len, bad);
305 err = parse_reply_info_extra(&p, p+len, info, features);
311 ceph_decode_32_safe(&p, end, len, bad);
312 info->snapblob_len = len;
323 pr_err("mds parse_reply err %d\n", err);
327 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
336 static const char *session_state_name(int s)
339 case CEPH_MDS_SESSION_NEW: return "new";
340 case CEPH_MDS_SESSION_OPENING: return "opening";
341 case CEPH_MDS_SESSION_OPEN: return "open";
342 case CEPH_MDS_SESSION_HUNG: return "hung";
343 case CEPH_MDS_SESSION_CLOSING: return "closing";
344 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
345 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
346 default: return "???";
350 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
352 if (atomic_inc_not_zero(&s->s_ref)) {
353 dout("mdsc get_session %p %d -> %d\n", s,
354 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
357 dout("mdsc get_session %p 0 -- FAIL", s);
362 void ceph_put_mds_session(struct ceph_mds_session *s)
364 dout("mdsc put_session %p %d -> %d\n", s,
365 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
366 if (atomic_dec_and_test(&s->s_ref)) {
367 if (s->s_auth.authorizer)
368 ceph_auth_destroy_authorizer(
369 s->s_mdsc->fsc->client->monc.auth,
370 s->s_auth.authorizer);
376 * called under mdsc->mutex
378 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
381 struct ceph_mds_session *session;
383 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
385 session = mdsc->sessions[mds];
386 dout("lookup_mds_session %p %d\n", session,
387 atomic_read(&session->s_ref));
388 get_session(session);
392 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
394 if (mds >= mdsc->max_sessions)
396 return mdsc->sessions[mds];
399 static int __verify_registered_session(struct ceph_mds_client *mdsc,
400 struct ceph_mds_session *s)
402 if (s->s_mds >= mdsc->max_sessions ||
403 mdsc->sessions[s->s_mds] != s)
409 * create+register a new session for given mds.
410 * called under mdsc->mutex.
412 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
415 struct ceph_mds_session *s;
417 s = kzalloc(sizeof(*s), GFP_NOFS);
419 return ERR_PTR(-ENOMEM);
422 s->s_state = CEPH_MDS_SESSION_NEW;
425 mutex_init(&s->s_mutex);
427 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
429 spin_lock_init(&s->s_gen_ttl_lock);
431 s->s_cap_ttl = jiffies - 1;
433 spin_lock_init(&s->s_cap_lock);
434 s->s_renew_requested = 0;
436 INIT_LIST_HEAD(&s->s_caps);
439 atomic_set(&s->s_ref, 1);
440 INIT_LIST_HEAD(&s->s_waiting);
441 INIT_LIST_HEAD(&s->s_unsafe);
442 s->s_num_cap_releases = 0;
443 s->s_cap_iterator = NULL;
444 INIT_LIST_HEAD(&s->s_cap_releases);
445 INIT_LIST_HEAD(&s->s_cap_releases_done);
446 INIT_LIST_HEAD(&s->s_cap_flushing);
447 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
449 dout("register_session mds%d\n", mds);
450 if (mds >= mdsc->max_sessions) {
451 int newmax = 1 << get_count_order(mds+1);
452 struct ceph_mds_session **sa;
454 dout("register_session realloc to %d\n", newmax);
455 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
458 if (mdsc->sessions) {
459 memcpy(sa, mdsc->sessions,
460 mdsc->max_sessions * sizeof(void *));
461 kfree(mdsc->sessions);
464 mdsc->max_sessions = newmax;
466 mdsc->sessions[mds] = s;
467 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
469 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
470 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
476 return ERR_PTR(-ENOMEM);
480 * called under mdsc->mutex
482 static void __unregister_session(struct ceph_mds_client *mdsc,
483 struct ceph_mds_session *s)
485 dout("__unregister_session mds%d %p\n", s->s_mds, s);
486 BUG_ON(mdsc->sessions[s->s_mds] != s);
487 mdsc->sessions[s->s_mds] = NULL;
488 ceph_con_close(&s->s_con);
489 ceph_put_mds_session(s);
493 * drop session refs in request.
495 * should be last request ref, or hold mdsc->mutex
497 static void put_request_session(struct ceph_mds_request *req)
499 if (req->r_session) {
500 ceph_put_mds_session(req->r_session);
501 req->r_session = NULL;
505 void ceph_mdsc_release_request(struct kref *kref)
507 struct ceph_mds_request *req = container_of(kref,
508 struct ceph_mds_request,
511 ceph_msg_put(req->r_request);
513 ceph_msg_put(req->r_reply);
514 destroy_reply_info(&req->r_reply_info);
517 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
520 if (req->r_locked_dir)
521 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
522 if (req->r_target_inode)
523 iput(req->r_target_inode);
526 if (req->r_old_dentry) {
528 * track (and drop pins for) r_old_dentry_dir
529 * separately, since r_old_dentry's d_parent may have
530 * changed between the dir mutex being dropped and
531 * this request being freed.
533 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
535 dput(req->r_old_dentry);
536 iput(req->r_old_dentry_dir);
540 put_request_session(req);
541 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
546 * lookup session, bump ref if found.
548 * called under mdsc->mutex.
550 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
553 struct ceph_mds_request *req;
554 struct rb_node *n = mdsc->request_tree.rb_node;
557 req = rb_entry(n, struct ceph_mds_request, r_node);
558 if (tid < req->r_tid)
560 else if (tid > req->r_tid)
563 ceph_mdsc_get_request(req);
570 static void __insert_request(struct ceph_mds_client *mdsc,
571 struct ceph_mds_request *new)
573 struct rb_node **p = &mdsc->request_tree.rb_node;
574 struct rb_node *parent = NULL;
575 struct ceph_mds_request *req = NULL;
579 req = rb_entry(parent, struct ceph_mds_request, r_node);
580 if (new->r_tid < req->r_tid)
582 else if (new->r_tid > req->r_tid)
588 rb_link_node(&new->r_node, parent, p);
589 rb_insert_color(&new->r_node, &mdsc->request_tree);
593 * Register an in-flight request, and assign a tid. Link to directory
594 * are modifying (if any).
596 * Called under mdsc->mutex.
598 static void __register_request(struct ceph_mds_client *mdsc,
599 struct ceph_mds_request *req,
602 req->r_tid = ++mdsc->last_tid;
604 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
606 dout("__register_request %p tid %lld\n", req, req->r_tid);
607 ceph_mdsc_get_request(req);
608 __insert_request(mdsc, req);
610 req->r_uid = current_fsuid();
611 req->r_gid = current_fsgid();
614 struct ceph_inode_info *ci = ceph_inode(dir);
617 spin_lock(&ci->i_unsafe_lock);
618 req->r_unsafe_dir = dir;
619 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
620 spin_unlock(&ci->i_unsafe_lock);
624 static void __unregister_request(struct ceph_mds_client *mdsc,
625 struct ceph_mds_request *req)
627 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
628 rb_erase(&req->r_node, &mdsc->request_tree);
629 RB_CLEAR_NODE(&req->r_node);
631 if (req->r_unsafe_dir) {
632 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
634 spin_lock(&ci->i_unsafe_lock);
635 list_del_init(&req->r_unsafe_dir_item);
636 spin_unlock(&ci->i_unsafe_lock);
638 iput(req->r_unsafe_dir);
639 req->r_unsafe_dir = NULL;
642 complete_all(&req->r_safe_completion);
644 ceph_mdsc_put_request(req);
648 * Choose mds to send request to next. If there is a hint set in the
649 * request (e.g., due to a prior forward hint from the mds), use that.
650 * Otherwise, consult frag tree and/or caps to identify the
651 * appropriate mds. If all else fails, choose randomly.
653 * Called under mdsc->mutex.
655 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
658 * we don't need to worry about protecting the d_parent access
659 * here because we never renaming inside the snapped namespace
660 * except to resplice to another snapdir, and either the old or new
661 * result is a valid result.
663 while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
664 dentry = dentry->d_parent;
668 static int __choose_mds(struct ceph_mds_client *mdsc,
669 struct ceph_mds_request *req)
672 struct ceph_inode_info *ci;
673 struct ceph_cap *cap;
674 int mode = req->r_direct_mode;
676 u32 hash = req->r_direct_hash;
677 bool is_hash = req->r_direct_is_hash;
680 * is there a specific mds we should try? ignore hint if we have
681 * no session and the mds is not up (active or recovering).
683 if (req->r_resend_mds >= 0 &&
684 (__have_session(mdsc, req->r_resend_mds) ||
685 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
686 dout("choose_mds using resend_mds mds%d\n",
688 return req->r_resend_mds;
691 if (mode == USE_RANDOM_MDS)
696 inode = req->r_inode;
697 } else if (req->r_dentry) {
698 /* ignore race with rename; old or new d_parent is okay */
699 struct dentry *parent = req->r_dentry->d_parent;
700 struct inode *dir = parent->d_inode;
702 if (dir->i_sb != mdsc->fsc->sb) {
704 inode = req->r_dentry->d_inode;
705 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
706 /* direct snapped/virtual snapdir requests
707 * based on parent dir inode */
708 struct dentry *dn = get_nonsnap_parent(parent);
710 dout("__choose_mds using nonsnap parent %p\n", inode);
711 } else if (req->r_dentry->d_inode) {
713 inode = req->r_dentry->d_inode;
717 hash = ceph_dentry_hash(dir, req->r_dentry);
722 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
726 ci = ceph_inode(inode);
728 if (is_hash && S_ISDIR(inode->i_mode)) {
729 struct ceph_inode_frag frag;
732 ceph_choose_frag(ci, hash, &frag, &found);
734 if (mode == USE_ANY_MDS && frag.ndist > 0) {
737 /* choose a random replica */
738 get_random_bytes(&r, 1);
741 dout("choose_mds %p %llx.%llx "
742 "frag %u mds%d (%d/%d)\n",
743 inode, ceph_vinop(inode),
746 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
747 CEPH_MDS_STATE_ACTIVE)
751 /* since this file/dir wasn't known to be
752 * replicated, then we want to look for the
753 * authoritative mds. */
756 /* choose auth mds */
758 dout("choose_mds %p %llx.%llx "
759 "frag %u mds%d (auth)\n",
760 inode, ceph_vinop(inode), frag.frag, mds);
761 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
762 CEPH_MDS_STATE_ACTIVE)
768 spin_lock(&ci->i_ceph_lock);
770 if (mode == USE_AUTH_MDS)
771 cap = ci->i_auth_cap;
772 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
773 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
775 spin_unlock(&ci->i_ceph_lock);
778 mds = cap->session->s_mds;
779 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
780 inode, ceph_vinop(inode), mds,
781 cap == ci->i_auth_cap ? "auth " : "", cap);
782 spin_unlock(&ci->i_ceph_lock);
786 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
787 dout("choose_mds chose random mds%d\n", mds);
795 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
797 struct ceph_msg *msg;
798 struct ceph_mds_session_head *h;
800 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
803 pr_err("create_session_msg ENOMEM creating msg\n");
806 h = msg->front.iov_base;
807 h->op = cpu_to_le32(op);
808 h->seq = cpu_to_le64(seq);
813 * send session open request.
815 * called under mdsc->mutex
817 static int __open_session(struct ceph_mds_client *mdsc,
818 struct ceph_mds_session *session)
820 struct ceph_msg *msg;
822 int mds = session->s_mds;
824 /* wait for mds to go active? */
825 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
826 dout("open_session to mds%d (%s)\n", mds,
827 ceph_mds_state_name(mstate));
828 session->s_state = CEPH_MDS_SESSION_OPENING;
829 session->s_renew_requested = jiffies;
831 /* send connect message */
832 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
835 ceph_con_send(&session->s_con, msg);
840 * open sessions for any export targets for the given mds
842 * called under mdsc->mutex
844 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
845 struct ceph_mds_session *session)
847 struct ceph_mds_info *mi;
848 struct ceph_mds_session *ts;
849 int i, mds = session->s_mds;
852 if (mds >= mdsc->mdsmap->m_max_mds)
854 mi = &mdsc->mdsmap->m_info[mds];
855 dout("open_export_target_sessions for mds%d (%d targets)\n",
856 session->s_mds, mi->num_export_targets);
858 for (i = 0; i < mi->num_export_targets; i++) {
859 target = mi->export_targets[i];
860 ts = __ceph_lookup_mds_session(mdsc, target);
862 ts = register_session(mdsc, target);
866 if (session->s_state == CEPH_MDS_SESSION_NEW ||
867 session->s_state == CEPH_MDS_SESSION_CLOSING)
868 __open_session(mdsc, session);
870 dout(" mds%d target mds%d %p is %s\n", session->s_mds,
871 i, ts, session_state_name(ts->s_state));
872 ceph_put_mds_session(ts);
876 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
877 struct ceph_mds_session *session)
879 mutex_lock(&mdsc->mutex);
880 __open_export_target_sessions(mdsc, session);
881 mutex_unlock(&mdsc->mutex);
889 * Free preallocated cap messages assigned to this session
891 static void cleanup_cap_releases(struct ceph_mds_session *session)
893 struct ceph_msg *msg;
895 spin_lock(&session->s_cap_lock);
896 while (!list_empty(&session->s_cap_releases)) {
897 msg = list_first_entry(&session->s_cap_releases,
898 struct ceph_msg, list_head);
899 list_del_init(&msg->list_head);
902 while (!list_empty(&session->s_cap_releases_done)) {
903 msg = list_first_entry(&session->s_cap_releases_done,
904 struct ceph_msg, list_head);
905 list_del_init(&msg->list_head);
908 spin_unlock(&session->s_cap_lock);
912 * Helper to safely iterate over all caps associated with a session, with
913 * special care taken to handle a racing __ceph_remove_cap().
915 * Caller must hold session s_mutex.
917 static int iterate_session_caps(struct ceph_mds_session *session,
918 int (*cb)(struct inode *, struct ceph_cap *,
922 struct ceph_cap *cap;
923 struct inode *inode, *last_inode = NULL;
924 struct ceph_cap *old_cap = NULL;
927 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
928 spin_lock(&session->s_cap_lock);
929 p = session->s_caps.next;
930 while (p != &session->s_caps) {
931 cap = list_entry(p, struct ceph_cap, session_caps);
932 inode = igrab(&cap->ci->vfs_inode);
937 session->s_cap_iterator = cap;
938 spin_unlock(&session->s_cap_lock);
945 ceph_put_cap(session->s_mdsc, old_cap);
949 ret = cb(inode, cap, arg);
952 spin_lock(&session->s_cap_lock);
954 if (cap->ci == NULL) {
955 dout("iterate_session_caps finishing cap %p removal\n",
957 BUG_ON(cap->session != session);
958 list_del_init(&cap->session_caps);
959 session->s_nr_caps--;
961 old_cap = cap; /* put_cap it w/o locks held */
968 session->s_cap_iterator = NULL;
969 spin_unlock(&session->s_cap_lock);
974 ceph_put_cap(session->s_mdsc, old_cap);
979 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
982 struct ceph_inode_info *ci = ceph_inode(inode);
985 dout("removing cap %p, ci is %p, inode is %p\n",
986 cap, ci, &ci->vfs_inode);
987 spin_lock(&ci->i_ceph_lock);
988 __ceph_remove_cap(cap);
989 if (!__ceph_is_any_real_caps(ci)) {
990 struct ceph_mds_client *mdsc =
991 ceph_sb_to_client(inode->i_sb)->mdsc;
993 spin_lock(&mdsc->cap_dirty_lock);
994 if (!list_empty(&ci->i_dirty_item)) {
995 pr_info(" dropping dirty %s state for %p %lld\n",
996 ceph_cap_string(ci->i_dirty_caps),
997 inode, ceph_ino(inode));
998 ci->i_dirty_caps = 0;
999 list_del_init(&ci->i_dirty_item);
1002 if (!list_empty(&ci->i_flushing_item)) {
1003 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1004 ceph_cap_string(ci->i_flushing_caps),
1005 inode, ceph_ino(inode));
1006 ci->i_flushing_caps = 0;
1007 list_del_init(&ci->i_flushing_item);
1008 mdsc->num_cap_flushing--;
1011 if (drop && ci->i_wrbuffer_ref) {
1012 pr_info(" dropping dirty data for %p %lld\n",
1013 inode, ceph_ino(inode));
1014 ci->i_wrbuffer_ref = 0;
1015 ci->i_wrbuffer_ref_head = 0;
1018 spin_unlock(&mdsc->cap_dirty_lock);
1020 spin_unlock(&ci->i_ceph_lock);
1027 * caller must hold session s_mutex
1029 static void remove_session_caps(struct ceph_mds_session *session)
1031 dout("remove_session_caps on %p\n", session);
1032 iterate_session_caps(session, remove_session_caps_cb, NULL);
1033 BUG_ON(session->s_nr_caps > 0);
1034 BUG_ON(!list_empty(&session->s_cap_flushing));
1035 cleanup_cap_releases(session);
1039 * wake up any threads waiting on this session's caps. if the cap is
1040 * old (didn't get renewed on the client reconnect), remove it now.
1042 * caller must hold s_mutex.
1044 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1047 struct ceph_inode_info *ci = ceph_inode(inode);
1049 wake_up_all(&ci->i_cap_wq);
1051 spin_lock(&ci->i_ceph_lock);
1052 ci->i_wanted_max_size = 0;
1053 ci->i_requested_max_size = 0;
1054 spin_unlock(&ci->i_ceph_lock);
1059 static void wake_up_session_caps(struct ceph_mds_session *session,
1062 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1063 iterate_session_caps(session, wake_up_session_cb,
1064 (void *)(unsigned long)reconnect);
1068 * Send periodic message to MDS renewing all currently held caps. The
1069 * ack will reset the expiration for all caps from this session.
1071 * caller holds s_mutex
1073 static int send_renew_caps(struct ceph_mds_client *mdsc,
1074 struct ceph_mds_session *session)
1076 struct ceph_msg *msg;
1079 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1080 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1081 pr_info("mds%d caps stale\n", session->s_mds);
1082 session->s_renew_requested = jiffies;
1084 /* do not try to renew caps until a recovering mds has reconnected
1085 * with its clients. */
1086 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1087 if (state < CEPH_MDS_STATE_RECONNECT) {
1088 dout("send_renew_caps ignoring mds%d (%s)\n",
1089 session->s_mds, ceph_mds_state_name(state));
1093 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1094 ceph_mds_state_name(state));
1095 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1096 ++session->s_renew_seq);
1099 ceph_con_send(&session->s_con, msg);
1104 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1106 * Called under session->s_mutex
1108 static void renewed_caps(struct ceph_mds_client *mdsc,
1109 struct ceph_mds_session *session, int is_renew)
1114 spin_lock(&session->s_cap_lock);
1115 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1117 session->s_cap_ttl = session->s_renew_requested +
1118 mdsc->mdsmap->m_session_timeout*HZ;
1121 if (time_before(jiffies, session->s_cap_ttl)) {
1122 pr_info("mds%d caps renewed\n", session->s_mds);
1125 pr_info("mds%d caps still stale\n", session->s_mds);
1128 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1129 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1130 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1131 spin_unlock(&session->s_cap_lock);
1134 wake_up_session_caps(session, 0);
1138 * send a session close request
1140 static int request_close_session(struct ceph_mds_client *mdsc,
1141 struct ceph_mds_session *session)
1143 struct ceph_msg *msg;
1145 dout("request_close_session mds%d state %s seq %lld\n",
1146 session->s_mds, session_state_name(session->s_state),
1148 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1151 ceph_con_send(&session->s_con, msg);
1156 * Called with s_mutex held.
1158 static int __close_session(struct ceph_mds_client *mdsc,
1159 struct ceph_mds_session *session)
1161 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1163 session->s_state = CEPH_MDS_SESSION_CLOSING;
1164 return request_close_session(mdsc, session);
1168 * Trim old(er) caps.
1170 * Because we can't cache an inode without one or more caps, we do
1171 * this indirectly: if a cap is unused, we prune its aliases, at which
1172 * point the inode will hopefully get dropped to.
1174 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1175 * memory pressure from the MDS, though, so it needn't be perfect.
1177 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1179 struct ceph_mds_session *session = arg;
1180 struct ceph_inode_info *ci = ceph_inode(inode);
1181 int used, oissued, mine;
1183 if (session->s_trim_caps <= 0)
1186 spin_lock(&ci->i_ceph_lock);
1187 mine = cap->issued | cap->implemented;
1188 used = __ceph_caps_used(ci);
1189 oissued = __ceph_caps_issued_other(ci, cap);
1191 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1192 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1193 ceph_cap_string(used));
1194 if (ci->i_dirty_caps)
1195 goto out; /* dirty caps */
1196 if ((used & ~oissued) & mine)
1197 goto out; /* we need these caps */
1199 session->s_trim_caps--;
1201 /* we aren't the only cap.. just remove us */
1202 __queue_cap_release(session, ceph_ino(inode), cap->cap_id,
1203 cap->mseq, cap->issue_seq);
1204 __ceph_remove_cap(cap);
1206 /* try to drop referring dentries */
1207 spin_unlock(&ci->i_ceph_lock);
1208 d_prune_aliases(inode);
1209 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1210 inode, cap, atomic_read(&inode->i_count));
1215 spin_unlock(&ci->i_ceph_lock);
1220 * Trim session cap count down to some max number.
1222 static int trim_caps(struct ceph_mds_client *mdsc,
1223 struct ceph_mds_session *session,
1226 int trim_caps = session->s_nr_caps - max_caps;
1228 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1229 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1230 if (trim_caps > 0) {
1231 session->s_trim_caps = trim_caps;
1232 iterate_session_caps(session, trim_caps_cb, session);
1233 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1234 session->s_mds, session->s_nr_caps, max_caps,
1235 trim_caps - session->s_trim_caps);
1236 session->s_trim_caps = 0;
1242 * Allocate cap_release messages. If there is a partially full message
1243 * in the queue, try to allocate enough to cover it's remainder, so that
1244 * we can send it immediately.
1246 * Called under s_mutex.
1248 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1249 struct ceph_mds_session *session)
1251 struct ceph_msg *msg, *partial = NULL;
1252 struct ceph_mds_cap_release *head;
1254 int extra = mdsc->fsc->mount_options->cap_release_safety;
1257 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1260 spin_lock(&session->s_cap_lock);
1262 if (!list_empty(&session->s_cap_releases)) {
1263 msg = list_first_entry(&session->s_cap_releases,
1266 head = msg->front.iov_base;
1267 num = le32_to_cpu(head->num);
1269 dout(" partial %p with (%d/%d)\n", msg, num,
1270 (int)CEPH_CAPS_PER_RELEASE);
1271 extra += CEPH_CAPS_PER_RELEASE - num;
1275 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1276 spin_unlock(&session->s_cap_lock);
1277 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1281 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1282 (int)msg->front.iov_len);
1283 head = msg->front.iov_base;
1284 head->num = cpu_to_le32(0);
1285 msg->front.iov_len = sizeof(*head);
1286 spin_lock(&session->s_cap_lock);
1287 list_add(&msg->list_head, &session->s_cap_releases);
1288 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1292 head = partial->front.iov_base;
1293 num = le32_to_cpu(head->num);
1294 dout(" queueing partial %p with %d/%d\n", partial, num,
1295 (int)CEPH_CAPS_PER_RELEASE);
1296 list_move_tail(&partial->list_head,
1297 &session->s_cap_releases_done);
1298 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1301 spin_unlock(&session->s_cap_lock);
1307 * flush all dirty inode data to disk.
1309 * returns true if we've flushed through want_flush_seq
1311 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1315 dout("check_cap_flush want %lld\n", want_flush_seq);
1316 mutex_lock(&mdsc->mutex);
1317 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1318 struct ceph_mds_session *session = mdsc->sessions[mds];
1322 get_session(session);
1323 mutex_unlock(&mdsc->mutex);
1325 mutex_lock(&session->s_mutex);
1326 if (!list_empty(&session->s_cap_flushing)) {
1327 struct ceph_inode_info *ci =
1328 list_entry(session->s_cap_flushing.next,
1329 struct ceph_inode_info,
1331 struct inode *inode = &ci->vfs_inode;
1333 spin_lock(&ci->i_ceph_lock);
1334 if (ci->i_cap_flush_seq <= want_flush_seq) {
1335 dout("check_cap_flush still flushing %p "
1336 "seq %lld <= %lld to mds%d\n", inode,
1337 ci->i_cap_flush_seq, want_flush_seq,
1341 spin_unlock(&ci->i_ceph_lock);
1343 mutex_unlock(&session->s_mutex);
1344 ceph_put_mds_session(session);
1348 mutex_lock(&mdsc->mutex);
1351 mutex_unlock(&mdsc->mutex);
1352 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1357 * called under s_mutex
1359 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1360 struct ceph_mds_session *session)
1362 struct ceph_msg *msg;
1364 dout("send_cap_releases mds%d\n", session->s_mds);
1365 spin_lock(&session->s_cap_lock);
1366 while (!list_empty(&session->s_cap_releases_done)) {
1367 msg = list_first_entry(&session->s_cap_releases_done,
1368 struct ceph_msg, list_head);
1369 list_del_init(&msg->list_head);
1370 spin_unlock(&session->s_cap_lock);
1371 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1372 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1373 ceph_con_send(&session->s_con, msg);
1374 spin_lock(&session->s_cap_lock);
1376 spin_unlock(&session->s_cap_lock);
1379 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1380 struct ceph_mds_session *session)
1382 struct ceph_msg *msg;
1383 struct ceph_mds_cap_release *head;
1386 dout("discard_cap_releases mds%d\n", session->s_mds);
1387 spin_lock(&session->s_cap_lock);
1389 /* zero out the in-progress message */
1390 msg = list_first_entry(&session->s_cap_releases,
1391 struct ceph_msg, list_head);
1392 head = msg->front.iov_base;
1393 num = le32_to_cpu(head->num);
1394 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1395 head->num = cpu_to_le32(0);
1396 session->s_num_cap_releases += num;
1398 /* requeue completed messages */
1399 while (!list_empty(&session->s_cap_releases_done)) {
1400 msg = list_first_entry(&session->s_cap_releases_done,
1401 struct ceph_msg, list_head);
1402 list_del_init(&msg->list_head);
1404 head = msg->front.iov_base;
1405 num = le32_to_cpu(head->num);
1406 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1408 session->s_num_cap_releases += num;
1409 head->num = cpu_to_le32(0);
1410 msg->front.iov_len = sizeof(*head);
1411 list_add(&msg->list_head, &session->s_cap_releases);
1414 spin_unlock(&session->s_cap_lock);
1422 * Create an mds request.
1424 struct ceph_mds_request *
1425 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1427 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1430 return ERR_PTR(-ENOMEM);
1432 mutex_init(&req->r_fill_mutex);
1434 req->r_started = jiffies;
1435 req->r_resend_mds = -1;
1436 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1438 kref_init(&req->r_kref);
1439 INIT_LIST_HEAD(&req->r_wait);
1440 init_completion(&req->r_completion);
1441 init_completion(&req->r_safe_completion);
1442 INIT_LIST_HEAD(&req->r_unsafe_item);
1445 req->r_direct_mode = mode;
1450 * return oldest (lowest) request, tid in request tree, 0 if none.
1452 * called under mdsc->mutex.
1454 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1456 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1458 return rb_entry(rb_first(&mdsc->request_tree),
1459 struct ceph_mds_request, r_node);
1462 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1464 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1472 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1473 * on build_path_from_dentry in fs/cifs/dir.c.
1475 * If @stop_on_nosnap, generate path relative to the first non-snapped
1478 * Encode hidden .snap dirs as a double /, i.e.
1479 * foo/.snap/bar -> foo//bar
1481 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1484 struct dentry *temp;
1490 return ERR_PTR(-EINVAL);
1494 seq = read_seqbegin(&rename_lock);
1496 for (temp = dentry; !IS_ROOT(temp);) {
1497 struct inode *inode = temp->d_inode;
1498 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1499 len++; /* slash only */
1500 else if (stop_on_nosnap && inode &&
1501 ceph_snap(inode) == CEPH_NOSNAP)
1504 len += 1 + temp->d_name.len;
1505 temp = temp->d_parent;
1509 len--; /* no leading '/' */
1511 path = kmalloc(len+1, GFP_NOFS);
1513 return ERR_PTR(-ENOMEM);
1515 path[pos] = 0; /* trailing null */
1517 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1518 struct inode *inode;
1520 spin_lock(&temp->d_lock);
1521 inode = temp->d_inode;
1522 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1523 dout("build_path path+%d: %p SNAPDIR\n",
1525 } else if (stop_on_nosnap && inode &&
1526 ceph_snap(inode) == CEPH_NOSNAP) {
1527 spin_unlock(&temp->d_lock);
1530 pos -= temp->d_name.len;
1532 spin_unlock(&temp->d_lock);
1535 strncpy(path + pos, temp->d_name.name,
1538 spin_unlock(&temp->d_lock);
1541 temp = temp->d_parent;
1544 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1545 pr_err("build_path did not end path lookup where "
1546 "expected, namelen is %d, pos is %d\n", len, pos);
1547 /* presumably this is only possible if racing with a
1548 rename of one of the parent directories (we can not
1549 lock the dentries above us to prevent this, but
1550 retrying should be harmless) */
1555 *base = ceph_ino(temp->d_inode);
1557 dout("build_path on %p %d built %llx '%.*s'\n",
1558 dentry, dentry->d_count, *base, len, path);
1562 static int build_dentry_path(struct dentry *dentry,
1563 const char **ppath, int *ppathlen, u64 *pino,
1568 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1569 *pino = ceph_ino(dentry->d_parent->d_inode);
1570 *ppath = dentry->d_name.name;
1571 *ppathlen = dentry->d_name.len;
1574 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1576 return PTR_ERR(path);
1582 static int build_inode_path(struct inode *inode,
1583 const char **ppath, int *ppathlen, u64 *pino,
1586 struct dentry *dentry;
1589 if (ceph_snap(inode) == CEPH_NOSNAP) {
1590 *pino = ceph_ino(inode);
1594 dentry = d_find_alias(inode);
1595 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1598 return PTR_ERR(path);
1605 * request arguments may be specified via an inode *, a dentry *, or
1606 * an explicit ino+path.
1608 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1609 const char *rpath, u64 rino,
1610 const char **ppath, int *pathlen,
1611 u64 *ino, int *freepath)
1616 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1617 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1619 } else if (rdentry) {
1620 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1621 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1623 } else if (rpath || rino) {
1626 *pathlen = rpath ? strlen(rpath) : 0;
1627 dout(" path %.*s\n", *pathlen, rpath);
1634 * called under mdsc->mutex
1636 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1637 struct ceph_mds_request *req,
1640 struct ceph_msg *msg;
1641 struct ceph_mds_request_head *head;
1642 const char *path1 = NULL;
1643 const char *path2 = NULL;
1644 u64 ino1 = 0, ino2 = 0;
1645 int pathlen1 = 0, pathlen2 = 0;
1646 int freepath1 = 0, freepath2 = 0;
1652 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1653 req->r_path1, req->r_ino1.ino,
1654 &path1, &pathlen1, &ino1, &freepath1);
1660 ret = set_request_path_attr(NULL, req->r_old_dentry,
1661 req->r_path2, req->r_ino2.ino,
1662 &path2, &pathlen2, &ino2, &freepath2);
1668 len = sizeof(*head) +
1669 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1671 /* calculate (max) length for cap releases */
1672 len += sizeof(struct ceph_mds_request_release) *
1673 (!!req->r_inode_drop + !!req->r_dentry_drop +
1674 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1675 if (req->r_dentry_drop)
1676 len += req->r_dentry->d_name.len;
1677 if (req->r_old_dentry_drop)
1678 len += req->r_old_dentry->d_name.len;
1680 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1682 msg = ERR_PTR(-ENOMEM);
1686 msg->hdr.tid = cpu_to_le64(req->r_tid);
1688 head = msg->front.iov_base;
1689 p = msg->front.iov_base + sizeof(*head);
1690 end = msg->front.iov_base + msg->front.iov_len;
1692 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1693 head->op = cpu_to_le32(req->r_op);
1694 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1695 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1696 head->args = req->r_args;
1698 ceph_encode_filepath(&p, end, ino1, path1);
1699 ceph_encode_filepath(&p, end, ino2, path2);
1701 /* make note of release offset, in case we need to replay */
1702 req->r_request_release_offset = p - msg->front.iov_base;
1706 if (req->r_inode_drop)
1707 releases += ceph_encode_inode_release(&p,
1708 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1709 mds, req->r_inode_drop, req->r_inode_unless, 0);
1710 if (req->r_dentry_drop)
1711 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1712 mds, req->r_dentry_drop, req->r_dentry_unless);
1713 if (req->r_old_dentry_drop)
1714 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1715 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1716 if (req->r_old_inode_drop)
1717 releases += ceph_encode_inode_release(&p,
1718 req->r_old_dentry->d_inode,
1719 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1720 head->num_releases = cpu_to_le16(releases);
1723 msg->front.iov_len = p - msg->front.iov_base;
1724 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1726 if (req->r_data_len) {
1727 /* outbound data set only by ceph_sync_setxattr() */
1728 BUG_ON(!req->r_pages);
1729 ceph_msg_data_add_pages(msg, req->r_pages, req->r_data_len, 0);
1732 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1733 msg->hdr.data_off = cpu_to_le16(0);
1737 kfree((char *)path2);
1740 kfree((char *)path1);
1746 * called under mdsc->mutex if error, under no mutex if
1749 static void complete_request(struct ceph_mds_client *mdsc,
1750 struct ceph_mds_request *req)
1752 if (req->r_callback)
1753 req->r_callback(mdsc, req);
1755 complete_all(&req->r_completion);
1759 * called under mdsc->mutex
1761 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1762 struct ceph_mds_request *req,
1765 struct ceph_mds_request_head *rhead;
1766 struct ceph_msg *msg;
1771 struct ceph_cap *cap =
1772 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1775 req->r_sent_on_mseq = cap->mseq;
1777 req->r_sent_on_mseq = -1;
1779 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1780 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1782 if (req->r_got_unsafe) {
1784 * Replay. Do not regenerate message (and rebuild
1785 * paths, etc.); just use the original message.
1786 * Rebuilding paths will break for renames because
1787 * d_move mangles the src name.
1789 msg = req->r_request;
1790 rhead = msg->front.iov_base;
1792 flags = le32_to_cpu(rhead->flags);
1793 flags |= CEPH_MDS_FLAG_REPLAY;
1794 rhead->flags = cpu_to_le32(flags);
1796 if (req->r_target_inode)
1797 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1799 rhead->num_retry = req->r_attempts - 1;
1801 /* remove cap/dentry releases from message */
1802 rhead->num_releases = 0;
1803 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1804 msg->front.iov_len = req->r_request_release_offset;
1808 if (req->r_request) {
1809 ceph_msg_put(req->r_request);
1810 req->r_request = NULL;
1812 msg = create_request_message(mdsc, req, mds);
1814 req->r_err = PTR_ERR(msg);
1815 complete_request(mdsc, req);
1816 return PTR_ERR(msg);
1818 req->r_request = msg;
1820 rhead = msg->front.iov_base;
1821 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1822 if (req->r_got_unsafe)
1823 flags |= CEPH_MDS_FLAG_REPLAY;
1824 if (req->r_locked_dir)
1825 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1826 rhead->flags = cpu_to_le32(flags);
1827 rhead->num_fwd = req->r_num_fwd;
1828 rhead->num_retry = req->r_attempts - 1;
1831 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1836 * send request, or put it on the appropriate wait list.
1838 static int __do_request(struct ceph_mds_client *mdsc,
1839 struct ceph_mds_request *req)
1841 struct ceph_mds_session *session = NULL;
1845 if (req->r_err || req->r_got_result) {
1847 __unregister_request(mdsc, req);
1851 if (req->r_timeout &&
1852 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1853 dout("do_request timed out\n");
1858 put_request_session(req);
1860 mds = __choose_mds(mdsc, req);
1862 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1863 dout("do_request no mds or not active, waiting for map\n");
1864 list_add(&req->r_wait, &mdsc->waiting_for_map);
1868 /* get, open session */
1869 session = __ceph_lookup_mds_session(mdsc, mds);
1871 session = register_session(mdsc, mds);
1872 if (IS_ERR(session)) {
1873 err = PTR_ERR(session);
1877 req->r_session = get_session(session);
1879 dout("do_request mds%d session %p state %s\n", mds, session,
1880 session_state_name(session->s_state));
1881 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1882 session->s_state != CEPH_MDS_SESSION_HUNG) {
1883 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1884 session->s_state == CEPH_MDS_SESSION_CLOSING)
1885 __open_session(mdsc, session);
1886 list_add(&req->r_wait, &session->s_waiting);
1891 req->r_resend_mds = -1; /* forget any previous mds hint */
1893 if (req->r_request_started == 0) /* note request start time */
1894 req->r_request_started = jiffies;
1896 err = __prepare_send_request(mdsc, req, mds);
1898 ceph_msg_get(req->r_request);
1899 ceph_con_send(&session->s_con, req->r_request);
1903 ceph_put_mds_session(session);
1909 complete_request(mdsc, req);
1914 * called under mdsc->mutex
1916 static void __wake_requests(struct ceph_mds_client *mdsc,
1917 struct list_head *head)
1919 struct ceph_mds_request *req;
1920 LIST_HEAD(tmp_list);
1922 list_splice_init(head, &tmp_list);
1924 while (!list_empty(&tmp_list)) {
1925 req = list_entry(tmp_list.next,
1926 struct ceph_mds_request, r_wait);
1927 list_del_init(&req->r_wait);
1928 dout(" wake request %p tid %llu\n", req, req->r_tid);
1929 __do_request(mdsc, req);
1934 * Wake up threads with requests pending for @mds, so that they can
1935 * resubmit their requests to a possibly different mds.
1937 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1939 struct ceph_mds_request *req;
1942 dout("kick_requests mds%d\n", mds);
1943 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1944 req = rb_entry(p, struct ceph_mds_request, r_node);
1945 if (req->r_got_unsafe)
1947 if (req->r_session &&
1948 req->r_session->s_mds == mds) {
1949 dout(" kicking tid %llu\n", req->r_tid);
1950 __do_request(mdsc, req);
1955 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1956 struct ceph_mds_request *req)
1958 dout("submit_request on %p\n", req);
1959 mutex_lock(&mdsc->mutex);
1960 __register_request(mdsc, req, NULL);
1961 __do_request(mdsc, req);
1962 mutex_unlock(&mdsc->mutex);
1966 * Synchrously perform an mds request. Take care of all of the
1967 * session setup, forwarding, retry details.
1969 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1971 struct ceph_mds_request *req)
1975 dout("do_request on %p\n", req);
1977 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1979 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1980 if (req->r_locked_dir)
1981 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1982 if (req->r_old_dentry)
1983 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
1987 mutex_lock(&mdsc->mutex);
1988 __register_request(mdsc, req, dir);
1989 __do_request(mdsc, req);
1993 __unregister_request(mdsc, req);
1994 dout("do_request early error %d\n", err);
1999 mutex_unlock(&mdsc->mutex);
2000 dout("do_request waiting\n");
2001 if (req->r_timeout) {
2002 err = (long)wait_for_completion_killable_timeout(
2003 &req->r_completion, req->r_timeout);
2007 err = wait_for_completion_killable(&req->r_completion);
2009 dout("do_request waited, got %d\n", err);
2010 mutex_lock(&mdsc->mutex);
2012 /* only abort if we didn't race with a real reply */
2013 if (req->r_got_result) {
2014 err = le32_to_cpu(req->r_reply_info.head->result);
2015 } else if (err < 0) {
2016 dout("aborted request %lld with %d\n", req->r_tid, err);
2019 * ensure we aren't running concurrently with
2020 * ceph_fill_trace or ceph_readdir_prepopulate, which
2021 * rely on locks (dir mutex) held by our caller.
2023 mutex_lock(&req->r_fill_mutex);
2025 req->r_aborted = true;
2026 mutex_unlock(&req->r_fill_mutex);
2028 if (req->r_locked_dir &&
2029 (req->r_op & CEPH_MDS_OP_WRITE))
2030 ceph_invalidate_dir_request(req);
2036 mutex_unlock(&mdsc->mutex);
2037 dout("do_request %p done, result %d\n", req, err);
2042 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2043 * namespace request.
2045 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2047 struct inode *inode = req->r_locked_dir;
2049 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2051 ceph_dir_clear_complete(inode);
2053 ceph_invalidate_dentry_lease(req->r_dentry);
2054 if (req->r_old_dentry)
2055 ceph_invalidate_dentry_lease(req->r_old_dentry);
2061 * We take the session mutex and parse and process the reply immediately.
2062 * This preserves the logical ordering of replies, capabilities, etc., sent
2063 * by the MDS as they are applied to our local cache.
2065 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2067 struct ceph_mds_client *mdsc = session->s_mdsc;
2068 struct ceph_mds_request *req;
2069 struct ceph_mds_reply_head *head = msg->front.iov_base;
2070 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2073 int mds = session->s_mds;
2075 if (msg->front.iov_len < sizeof(*head)) {
2076 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2081 /* get request, session */
2082 tid = le64_to_cpu(msg->hdr.tid);
2083 mutex_lock(&mdsc->mutex);
2084 req = __lookup_request(mdsc, tid);
2086 dout("handle_reply on unknown tid %llu\n", tid);
2087 mutex_unlock(&mdsc->mutex);
2090 dout("handle_reply %p\n", req);
2092 /* correct session? */
2093 if (req->r_session != session) {
2094 pr_err("mdsc_handle_reply got %llu on session mds%d"
2095 " not mds%d\n", tid, session->s_mds,
2096 req->r_session ? req->r_session->s_mds : -1);
2097 mutex_unlock(&mdsc->mutex);
2102 if ((req->r_got_unsafe && !head->safe) ||
2103 (req->r_got_safe && head->safe)) {
2104 pr_warning("got a dup %s reply on %llu from mds%d\n",
2105 head->safe ? "safe" : "unsafe", tid, mds);
2106 mutex_unlock(&mdsc->mutex);
2109 if (req->r_got_safe && !head->safe) {
2110 pr_warning("got unsafe after safe on %llu from mds%d\n",
2112 mutex_unlock(&mdsc->mutex);
2116 result = le32_to_cpu(head->result);
2120 * if we're not talking to the authority, send to them
2121 * if the authority has changed while we weren't looking,
2122 * send to new authority
2123 * Otherwise we just have to return an ESTALE
2125 if (result == -ESTALE) {
2126 dout("got ESTALE on request %llu", req->r_tid);
2127 if (!req->r_inode) {
2128 /* do nothing; not an authority problem */
2129 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2130 dout("not using auth, setting for that now");
2131 req->r_direct_mode = USE_AUTH_MDS;
2132 __do_request(mdsc, req);
2133 mutex_unlock(&mdsc->mutex);
2136 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2137 struct ceph_cap *cap = NULL;
2140 cap = ceph_get_cap_for_mds(ci,
2141 req->r_session->s_mds);
2143 dout("already using auth");
2144 if ((!cap || cap != ci->i_auth_cap) ||
2145 (cap->mseq != req->r_sent_on_mseq)) {
2146 dout("but cap changed, so resending");
2147 __do_request(mdsc, req);
2148 mutex_unlock(&mdsc->mutex);
2152 dout("have to return ESTALE on request %llu", req->r_tid);
2157 req->r_got_safe = true;
2158 __unregister_request(mdsc, req);
2160 if (req->r_got_unsafe) {
2162 * We already handled the unsafe response, now do the
2163 * cleanup. No need to examine the response; the MDS
2164 * doesn't include any result info in the safe
2165 * response. And even if it did, there is nothing
2166 * useful we could do with a revised return value.
2168 dout("got safe reply %llu, mds%d\n", tid, mds);
2169 list_del_init(&req->r_unsafe_item);
2171 /* last unsafe request during umount? */
2172 if (mdsc->stopping && !__get_oldest_req(mdsc))
2173 complete_all(&mdsc->safe_umount_waiters);
2174 mutex_unlock(&mdsc->mutex);
2178 req->r_got_unsafe = true;
2179 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2182 dout("handle_reply tid %lld result %d\n", tid, result);
2183 rinfo = &req->r_reply_info;
2184 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2185 mutex_unlock(&mdsc->mutex);
2187 mutex_lock(&session->s_mutex);
2189 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2195 if (rinfo->snapblob_len) {
2196 down_write(&mdsc->snap_rwsem);
2197 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2198 rinfo->snapblob + rinfo->snapblob_len,
2199 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2200 downgrade_write(&mdsc->snap_rwsem);
2202 down_read(&mdsc->snap_rwsem);
2205 /* insert trace into our cache */
2206 mutex_lock(&req->r_fill_mutex);
2207 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2209 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2210 req->r_op == CEPH_MDS_OP_LSSNAP) &&
2212 ceph_readdir_prepopulate(req, req->r_session);
2213 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2215 mutex_unlock(&req->r_fill_mutex);
2217 up_read(&mdsc->snap_rwsem);
2219 mutex_lock(&mdsc->mutex);
2220 if (!req->r_aborted) {
2226 req->r_got_result = true;
2229 dout("reply arrived after request %lld was aborted\n", tid);
2231 mutex_unlock(&mdsc->mutex);
2233 ceph_add_cap_releases(mdsc, req->r_session);
2234 mutex_unlock(&session->s_mutex);
2236 /* kick calling process */
2237 complete_request(mdsc, req);
2239 ceph_mdsc_put_request(req);
2246 * handle mds notification that our request has been forwarded.
2248 static void handle_forward(struct ceph_mds_client *mdsc,
2249 struct ceph_mds_session *session,
2250 struct ceph_msg *msg)
2252 struct ceph_mds_request *req;
2253 u64 tid = le64_to_cpu(msg->hdr.tid);
2257 void *p = msg->front.iov_base;
2258 void *end = p + msg->front.iov_len;
2260 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2261 next_mds = ceph_decode_32(&p);
2262 fwd_seq = ceph_decode_32(&p);
2264 mutex_lock(&mdsc->mutex);
2265 req = __lookup_request(mdsc, tid);
2267 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2268 goto out; /* dup reply? */
2271 if (req->r_aborted) {
2272 dout("forward tid %llu aborted, unregistering\n", tid);
2273 __unregister_request(mdsc, req);
2274 } else if (fwd_seq <= req->r_num_fwd) {
2275 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2276 tid, next_mds, req->r_num_fwd, fwd_seq);
2278 /* resend. forward race not possible; mds would drop */
2279 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2281 BUG_ON(req->r_got_result);
2282 req->r_num_fwd = fwd_seq;
2283 req->r_resend_mds = next_mds;
2284 put_request_session(req);
2285 __do_request(mdsc, req);
2287 ceph_mdsc_put_request(req);
2289 mutex_unlock(&mdsc->mutex);
2293 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2297 * handle a mds session control message
2299 static void handle_session(struct ceph_mds_session *session,
2300 struct ceph_msg *msg)
2302 struct ceph_mds_client *mdsc = session->s_mdsc;
2305 int mds = session->s_mds;
2306 struct ceph_mds_session_head *h = msg->front.iov_base;
2310 if (msg->front.iov_len != sizeof(*h))
2312 op = le32_to_cpu(h->op);
2313 seq = le64_to_cpu(h->seq);
2315 mutex_lock(&mdsc->mutex);
2316 if (op == CEPH_SESSION_CLOSE)
2317 __unregister_session(mdsc, session);
2318 /* FIXME: this ttl calculation is generous */
2319 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2320 mutex_unlock(&mdsc->mutex);
2322 mutex_lock(&session->s_mutex);
2324 dout("handle_session mds%d %s %p state %s seq %llu\n",
2325 mds, ceph_session_op_name(op), session,
2326 session_state_name(session->s_state), seq);
2328 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2329 session->s_state = CEPH_MDS_SESSION_OPEN;
2330 pr_info("mds%d came back\n", session->s_mds);
2334 case CEPH_SESSION_OPEN:
2335 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2336 pr_info("mds%d reconnect success\n", session->s_mds);
2337 session->s_state = CEPH_MDS_SESSION_OPEN;
2338 renewed_caps(mdsc, session, 0);
2341 __close_session(mdsc, session);
2344 case CEPH_SESSION_RENEWCAPS:
2345 if (session->s_renew_seq == seq)
2346 renewed_caps(mdsc, session, 1);
2349 case CEPH_SESSION_CLOSE:
2350 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2351 pr_info("mds%d reconnect denied\n", session->s_mds);
2352 remove_session_caps(session);
2353 wake = 1; /* for good measure */
2354 wake_up_all(&mdsc->session_close_wq);
2355 kick_requests(mdsc, mds);
2358 case CEPH_SESSION_STALE:
2359 pr_info("mds%d caps went stale, renewing\n",
2361 spin_lock(&session->s_gen_ttl_lock);
2362 session->s_cap_gen++;
2363 session->s_cap_ttl = jiffies - 1;
2364 spin_unlock(&session->s_gen_ttl_lock);
2365 send_renew_caps(mdsc, session);
2368 case CEPH_SESSION_RECALL_STATE:
2369 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2373 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2377 mutex_unlock(&session->s_mutex);
2379 mutex_lock(&mdsc->mutex);
2380 __wake_requests(mdsc, &session->s_waiting);
2381 mutex_unlock(&mdsc->mutex);
2386 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2387 (int)msg->front.iov_len);
2394 * called under session->mutex.
2396 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2397 struct ceph_mds_session *session)
2399 struct ceph_mds_request *req, *nreq;
2402 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2404 mutex_lock(&mdsc->mutex);
2405 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2406 err = __prepare_send_request(mdsc, req, session->s_mds);
2408 ceph_msg_get(req->r_request);
2409 ceph_con_send(&session->s_con, req->r_request);
2412 mutex_unlock(&mdsc->mutex);
2416 * Encode information about a cap for a reconnect with the MDS.
2418 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2422 struct ceph_mds_cap_reconnect v2;
2423 struct ceph_mds_cap_reconnect_v1 v1;
2426 struct ceph_inode_info *ci;
2427 struct ceph_reconnect_state *recon_state = arg;
2428 struct ceph_pagelist *pagelist = recon_state->pagelist;
2432 struct dentry *dentry;
2436 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2437 inode, ceph_vinop(inode), cap, cap->cap_id,
2438 ceph_cap_string(cap->issued));
2439 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2443 dentry = d_find_alias(inode);
2445 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2447 err = PTR_ERR(path);
2454 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2458 spin_lock(&ci->i_ceph_lock);
2459 cap->seq = 0; /* reset cap seq */
2460 cap->issue_seq = 0; /* and issue_seq */
2462 if (recon_state->flock) {
2463 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2464 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2465 rec.v2.issued = cpu_to_le32(cap->issued);
2466 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2467 rec.v2.pathbase = cpu_to_le64(pathbase);
2468 rec.v2.flock_len = 0;
2469 reclen = sizeof(rec.v2);
2471 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2472 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2473 rec.v1.issued = cpu_to_le32(cap->issued);
2474 rec.v1.size = cpu_to_le64(inode->i_size);
2475 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2476 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2477 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2478 rec.v1.pathbase = cpu_to_le64(pathbase);
2479 reclen = sizeof(rec.v1);
2481 spin_unlock(&ci->i_ceph_lock);
2483 if (recon_state->flock) {
2484 int num_fcntl_locks, num_flock_locks;
2485 struct ceph_filelock *flocks;
2489 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2491 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2492 sizeof(struct ceph_filelock), GFP_NOFS);
2498 err = ceph_encode_locks_to_buffer(inode, flocks,
2509 * number of encoded locks is stable, so copy to pagelist
2511 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2512 (num_fcntl_locks+num_flock_locks) *
2513 sizeof(struct ceph_filelock));
2514 err = ceph_pagelist_append(pagelist, &rec, reclen);
2516 err = ceph_locks_to_pagelist(flocks, pagelist,
2521 err = ceph_pagelist_append(pagelist, &rec, reclen);
2532 * If an MDS fails and recovers, clients need to reconnect in order to
2533 * reestablish shared state. This includes all caps issued through
2534 * this session _and_ the snap_realm hierarchy. Because it's not
2535 * clear which snap realms the mds cares about, we send everything we
2536 * know about.. that ensures we'll then get any new info the
2537 * recovering MDS might have.
2539 * This is a relatively heavyweight operation, but it's rare.
2541 * called with mdsc->mutex held.
2543 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2544 struct ceph_mds_session *session)
2546 struct ceph_msg *reply;
2548 int mds = session->s_mds;
2550 struct ceph_pagelist *pagelist;
2551 struct ceph_reconnect_state recon_state;
2553 pr_info("mds%d reconnect start\n", mds);
2555 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2557 goto fail_nopagelist;
2558 ceph_pagelist_init(pagelist);
2560 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2564 mutex_lock(&session->s_mutex);
2565 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2568 ceph_con_close(&session->s_con);
2569 ceph_con_open(&session->s_con,
2570 CEPH_ENTITY_TYPE_MDS, mds,
2571 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2573 /* replay unsafe requests */
2574 replay_unsafe_requests(mdsc, session);
2576 down_read(&mdsc->snap_rwsem);
2578 dout("session %p state %s\n", session,
2579 session_state_name(session->s_state));
2581 /* drop old cap expires; we're about to reestablish that state */
2582 discard_cap_releases(mdsc, session);
2584 /* traverse this session's caps */
2585 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2589 recon_state.pagelist = pagelist;
2590 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2591 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2596 * snaprealms. we provide mds with the ino, seq (version), and
2597 * parent for all of our realms. If the mds has any newer info,
2600 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2601 struct ceph_snap_realm *realm =
2602 rb_entry(p, struct ceph_snap_realm, node);
2603 struct ceph_mds_snaprealm_reconnect sr_rec;
2605 dout(" adding snap realm %llx seq %lld parent %llx\n",
2606 realm->ino, realm->seq, realm->parent_ino);
2607 sr_rec.ino = cpu_to_le64(realm->ino);
2608 sr_rec.seq = cpu_to_le64(realm->seq);
2609 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2610 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2615 if (recon_state.flock)
2616 reply->hdr.version = cpu_to_le16(2);
2617 if (pagelist->length) {
2618 /* set up outbound data if we have any */
2619 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2620 ceph_msg_data_add_pagelist(reply, pagelist);
2622 ceph_con_send(&session->s_con, reply);
2624 mutex_unlock(&session->s_mutex);
2626 mutex_lock(&mdsc->mutex);
2627 __wake_requests(mdsc, &session->s_waiting);
2628 mutex_unlock(&mdsc->mutex);
2630 up_read(&mdsc->snap_rwsem);
2634 ceph_msg_put(reply);
2635 up_read(&mdsc->snap_rwsem);
2636 mutex_unlock(&session->s_mutex);
2638 ceph_pagelist_release(pagelist);
2641 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2647 * compare old and new mdsmaps, kicking requests
2648 * and closing out old connections as necessary
2650 * called under mdsc->mutex.
2652 static void check_new_map(struct ceph_mds_client *mdsc,
2653 struct ceph_mdsmap *newmap,
2654 struct ceph_mdsmap *oldmap)
2657 int oldstate, newstate;
2658 struct ceph_mds_session *s;
2660 dout("check_new_map new %u old %u\n",
2661 newmap->m_epoch, oldmap->m_epoch);
2663 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2664 if (mdsc->sessions[i] == NULL)
2666 s = mdsc->sessions[i];
2667 oldstate = ceph_mdsmap_get_state(oldmap, i);
2668 newstate = ceph_mdsmap_get_state(newmap, i);
2670 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2671 i, ceph_mds_state_name(oldstate),
2672 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2673 ceph_mds_state_name(newstate),
2674 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2675 session_state_name(s->s_state));
2677 if (i >= newmap->m_max_mds ||
2678 memcmp(ceph_mdsmap_get_addr(oldmap, i),
2679 ceph_mdsmap_get_addr(newmap, i),
2680 sizeof(struct ceph_entity_addr))) {
2681 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2682 /* the session never opened, just close it
2684 __wake_requests(mdsc, &s->s_waiting);
2685 __unregister_session(mdsc, s);
2688 mutex_unlock(&mdsc->mutex);
2689 mutex_lock(&s->s_mutex);
2690 mutex_lock(&mdsc->mutex);
2691 ceph_con_close(&s->s_con);
2692 mutex_unlock(&s->s_mutex);
2693 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2696 /* kick any requests waiting on the recovering mds */
2697 kick_requests(mdsc, i);
2698 } else if (oldstate == newstate) {
2699 continue; /* nothing new with this mds */
2705 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2706 newstate >= CEPH_MDS_STATE_RECONNECT) {
2707 mutex_unlock(&mdsc->mutex);
2708 send_mds_reconnect(mdsc, s);
2709 mutex_lock(&mdsc->mutex);
2713 * kick request on any mds that has gone active.
2715 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2716 newstate >= CEPH_MDS_STATE_ACTIVE) {
2717 if (oldstate != CEPH_MDS_STATE_CREATING &&
2718 oldstate != CEPH_MDS_STATE_STARTING)
2719 pr_info("mds%d recovery completed\n", s->s_mds);
2720 kick_requests(mdsc, i);
2721 ceph_kick_flushing_caps(mdsc, s);
2722 wake_up_session_caps(s, 1);
2726 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2727 s = mdsc->sessions[i];
2730 if (!ceph_mdsmap_is_laggy(newmap, i))
2732 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2733 s->s_state == CEPH_MDS_SESSION_HUNG ||
2734 s->s_state == CEPH_MDS_SESSION_CLOSING) {
2735 dout(" connecting to export targets of laggy mds%d\n",
2737 __open_export_target_sessions(mdsc, s);
2749 * caller must hold session s_mutex, dentry->d_lock
2751 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2753 struct ceph_dentry_info *di = ceph_dentry(dentry);
2755 ceph_put_mds_session(di->lease_session);
2756 di->lease_session = NULL;
2759 static void handle_lease(struct ceph_mds_client *mdsc,
2760 struct ceph_mds_session *session,
2761 struct ceph_msg *msg)
2763 struct super_block *sb = mdsc->fsc->sb;
2764 struct inode *inode;
2765 struct dentry *parent, *dentry;
2766 struct ceph_dentry_info *di;
2767 int mds = session->s_mds;
2768 struct ceph_mds_lease *h = msg->front.iov_base;
2770 struct ceph_vino vino;
2774 dout("handle_lease from mds%d\n", mds);
2777 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2779 vino.ino = le64_to_cpu(h->ino);
2780 vino.snap = CEPH_NOSNAP;
2781 seq = le32_to_cpu(h->seq);
2782 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2783 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2784 if (dname.len != get_unaligned_le32(h+1))
2787 mutex_lock(&session->s_mutex);
2791 inode = ceph_find_inode(sb, vino);
2792 dout("handle_lease %s, ino %llx %p %.*s\n",
2793 ceph_lease_op_name(h->action), vino.ino, inode,
2794 dname.len, dname.name);
2795 if (inode == NULL) {
2796 dout("handle_lease no inode %llx\n", vino.ino);
2801 parent = d_find_alias(inode);
2803 dout("no parent dentry on inode %p\n", inode);
2805 goto release; /* hrm... */
2807 dname.hash = full_name_hash(dname.name, dname.len);
2808 dentry = d_lookup(parent, &dname);
2813 spin_lock(&dentry->d_lock);
2814 di = ceph_dentry(dentry);
2815 switch (h->action) {
2816 case CEPH_MDS_LEASE_REVOKE:
2817 if (di->lease_session == session) {
2818 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2819 h->seq = cpu_to_le32(di->lease_seq);
2820 __ceph_mdsc_drop_dentry_lease(dentry);
2825 case CEPH_MDS_LEASE_RENEW:
2826 if (di->lease_session == session &&
2827 di->lease_gen == session->s_cap_gen &&
2828 di->lease_renew_from &&
2829 di->lease_renew_after == 0) {
2830 unsigned long duration =
2831 le32_to_cpu(h->duration_ms) * HZ / 1000;
2833 di->lease_seq = seq;
2834 dentry->d_time = di->lease_renew_from + duration;
2835 di->lease_renew_after = di->lease_renew_from +
2837 di->lease_renew_from = 0;
2841 spin_unlock(&dentry->d_lock);
2848 /* let's just reuse the same message */
2849 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2851 ceph_con_send(&session->s_con, msg);
2855 mutex_unlock(&session->s_mutex);
2859 pr_err("corrupt lease message\n");
2863 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2864 struct inode *inode,
2865 struct dentry *dentry, char action,
2868 struct ceph_msg *msg;
2869 struct ceph_mds_lease *lease;
2870 int len = sizeof(*lease) + sizeof(u32);
2873 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2874 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2875 dnamelen = dentry->d_name.len;
2878 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2881 lease = msg->front.iov_base;
2882 lease->action = action;
2883 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2884 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2885 lease->seq = cpu_to_le32(seq);
2886 put_unaligned_le32(dnamelen, lease + 1);
2887 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2890 * if this is a preemptive lease RELEASE, no need to
2891 * flush request stream, since the actual request will
2894 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2896 ceph_con_send(&session->s_con, msg);
2900 * Preemptively release a lease we expect to invalidate anyway.
2901 * Pass @inode always, @dentry is optional.
2903 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2904 struct dentry *dentry)
2906 struct ceph_dentry_info *di;
2907 struct ceph_mds_session *session;
2910 BUG_ON(inode == NULL);
2911 BUG_ON(dentry == NULL);
2913 /* is dentry lease valid? */
2914 spin_lock(&dentry->d_lock);
2915 di = ceph_dentry(dentry);
2916 if (!di || !di->lease_session ||
2917 di->lease_session->s_mds < 0 ||
2918 di->lease_gen != di->lease_session->s_cap_gen ||
2919 !time_before(jiffies, dentry->d_time)) {
2920 dout("lease_release inode %p dentry %p -- "
2923 spin_unlock(&dentry->d_lock);
2927 /* we do have a lease on this dentry; note mds and seq */
2928 session = ceph_get_mds_session(di->lease_session);
2929 seq = di->lease_seq;
2930 __ceph_mdsc_drop_dentry_lease(dentry);
2931 spin_unlock(&dentry->d_lock);
2933 dout("lease_release inode %p dentry %p to mds%d\n",
2934 inode, dentry, session->s_mds);
2935 ceph_mdsc_lease_send_msg(session, inode, dentry,
2936 CEPH_MDS_LEASE_RELEASE, seq);
2937 ceph_put_mds_session(session);
2941 * drop all leases (and dentry refs) in preparation for umount
2943 static void drop_leases(struct ceph_mds_client *mdsc)
2947 dout("drop_leases\n");
2948 mutex_lock(&mdsc->mutex);
2949 for (i = 0; i < mdsc->max_sessions; i++) {
2950 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2953 mutex_unlock(&mdsc->mutex);
2954 mutex_lock(&s->s_mutex);
2955 mutex_unlock(&s->s_mutex);
2956 ceph_put_mds_session(s);
2957 mutex_lock(&mdsc->mutex);
2959 mutex_unlock(&mdsc->mutex);
2965 * delayed work -- periodically trim expired leases, renew caps with mds
2967 static void schedule_delayed(struct ceph_mds_client *mdsc)
2970 unsigned hz = round_jiffies_relative(HZ * delay);
2971 schedule_delayed_work(&mdsc->delayed_work, hz);
2974 static void delayed_work(struct work_struct *work)
2977 struct ceph_mds_client *mdsc =
2978 container_of(work, struct ceph_mds_client, delayed_work.work);
2982 dout("mdsc delayed_work\n");
2983 ceph_check_delayed_caps(mdsc);
2985 mutex_lock(&mdsc->mutex);
2986 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2987 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2988 mdsc->last_renew_caps);
2990 mdsc->last_renew_caps = jiffies;
2992 for (i = 0; i < mdsc->max_sessions; i++) {
2993 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2996 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2997 dout("resending session close request for mds%d\n",
2999 request_close_session(mdsc, s);
3000 ceph_put_mds_session(s);
3003 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3004 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3005 s->s_state = CEPH_MDS_SESSION_HUNG;
3006 pr_info("mds%d hung\n", s->s_mds);
3009 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3010 /* this mds is failed or recovering, just wait */
3011 ceph_put_mds_session(s);
3014 mutex_unlock(&mdsc->mutex);
3016 mutex_lock(&s->s_mutex);
3018 send_renew_caps(mdsc, s);
3020 ceph_con_keepalive(&s->s_con);
3021 ceph_add_cap_releases(mdsc, s);
3022 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3023 s->s_state == CEPH_MDS_SESSION_HUNG)
3024 ceph_send_cap_releases(mdsc, s);
3025 mutex_unlock(&s->s_mutex);
3026 ceph_put_mds_session(s);
3028 mutex_lock(&mdsc->mutex);
3030 mutex_unlock(&mdsc->mutex);
3032 schedule_delayed(mdsc);
3035 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3038 struct ceph_mds_client *mdsc;
3040 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3045 mutex_init(&mdsc->mutex);
3046 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3047 if (mdsc->mdsmap == NULL)
3050 init_completion(&mdsc->safe_umount_waiters);
3051 init_waitqueue_head(&mdsc->session_close_wq);
3052 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3053 mdsc->sessions = NULL;
3054 mdsc->max_sessions = 0;
3056 init_rwsem(&mdsc->snap_rwsem);
3057 mdsc->snap_realms = RB_ROOT;
3058 INIT_LIST_HEAD(&mdsc->snap_empty);
3059 spin_lock_init(&mdsc->snap_empty_lock);
3061 mdsc->request_tree = RB_ROOT;
3062 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3063 mdsc->last_renew_caps = jiffies;
3064 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3065 spin_lock_init(&mdsc->cap_delay_lock);
3066 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3067 spin_lock_init(&mdsc->snap_flush_lock);
3068 mdsc->cap_flush_seq = 0;
3069 INIT_LIST_HEAD(&mdsc->cap_dirty);
3070 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3071 mdsc->num_cap_flushing = 0;
3072 spin_lock_init(&mdsc->cap_dirty_lock);
3073 init_waitqueue_head(&mdsc->cap_flushing_wq);
3074 spin_lock_init(&mdsc->dentry_lru_lock);
3075 INIT_LIST_HEAD(&mdsc->dentry_lru);
3077 ceph_caps_init(mdsc);
3078 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3084 * Wait for safe replies on open mds requests. If we time out, drop
3085 * all requests from the tree to avoid dangling dentry refs.
3087 static void wait_requests(struct ceph_mds_client *mdsc)
3089 struct ceph_mds_request *req;
3090 struct ceph_fs_client *fsc = mdsc->fsc;
3092 mutex_lock(&mdsc->mutex);
3093 if (__get_oldest_req(mdsc)) {
3094 mutex_unlock(&mdsc->mutex);
3096 dout("wait_requests waiting for requests\n");
3097 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3098 fsc->client->options->mount_timeout * HZ);
3100 /* tear down remaining requests */
3101 mutex_lock(&mdsc->mutex);
3102 while ((req = __get_oldest_req(mdsc))) {
3103 dout("wait_requests timed out on tid %llu\n",
3105 __unregister_request(mdsc, req);
3108 mutex_unlock(&mdsc->mutex);
3109 dout("wait_requests done\n");
3113 * called before mount is ro, and before dentries are torn down.
3114 * (hmm, does this still race with new lookups?)
3116 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3118 dout("pre_umount\n");
3122 ceph_flush_dirty_caps(mdsc);
3123 wait_requests(mdsc);
3126 * wait for reply handlers to drop their request refs and
3127 * their inode/dcache refs
3133 * wait for all write mds requests to flush.
3135 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3137 struct ceph_mds_request *req = NULL, *nextreq;
3140 mutex_lock(&mdsc->mutex);
3141 dout("wait_unsafe_requests want %lld\n", want_tid);
3143 req = __get_oldest_req(mdsc);
3144 while (req && req->r_tid <= want_tid) {
3145 /* find next request */
3146 n = rb_next(&req->r_node);
3148 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3151 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3153 ceph_mdsc_get_request(req);
3155 ceph_mdsc_get_request(nextreq);
3156 mutex_unlock(&mdsc->mutex);
3157 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3158 req->r_tid, want_tid);
3159 wait_for_completion(&req->r_safe_completion);
3160 mutex_lock(&mdsc->mutex);
3161 ceph_mdsc_put_request(req);
3163 break; /* next dne before, so we're done! */
3164 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3165 /* next request was removed from tree */
3166 ceph_mdsc_put_request(nextreq);
3169 ceph_mdsc_put_request(nextreq); /* won't go away */
3173 mutex_unlock(&mdsc->mutex);
3174 dout("wait_unsafe_requests done\n");
3177 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3179 u64 want_tid, want_flush;
3181 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3185 mutex_lock(&mdsc->mutex);
3186 want_tid = mdsc->last_tid;
3187 want_flush = mdsc->cap_flush_seq;
3188 mutex_unlock(&mdsc->mutex);
3189 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3191 ceph_flush_dirty_caps(mdsc);
3193 wait_unsafe_requests(mdsc, want_tid);
3194 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3198 * true if all sessions are closed, or we force unmount
3200 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3204 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3207 mutex_lock(&mdsc->mutex);
3208 for (i = 0; i < mdsc->max_sessions; i++)
3209 if (mdsc->sessions[i])
3211 mutex_unlock(&mdsc->mutex);
3216 * called after sb is ro.
3218 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3220 struct ceph_mds_session *session;
3222 struct ceph_fs_client *fsc = mdsc->fsc;
3223 unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3225 dout("close_sessions\n");
3227 /* close sessions */
3228 mutex_lock(&mdsc->mutex);
3229 for (i = 0; i < mdsc->max_sessions; i++) {
3230 session = __ceph_lookup_mds_session(mdsc, i);
3233 mutex_unlock(&mdsc->mutex);
3234 mutex_lock(&session->s_mutex);
3235 __close_session(mdsc, session);
3236 mutex_unlock(&session->s_mutex);
3237 ceph_put_mds_session(session);
3238 mutex_lock(&mdsc->mutex);
3240 mutex_unlock(&mdsc->mutex);
3242 dout("waiting for sessions to close\n");
3243 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3246 /* tear down remaining sessions */
3247 mutex_lock(&mdsc->mutex);
3248 for (i = 0; i < mdsc->max_sessions; i++) {
3249 if (mdsc->sessions[i]) {
3250 session = get_session(mdsc->sessions[i]);
3251 __unregister_session(mdsc, session);
3252 mutex_unlock(&mdsc->mutex);
3253 mutex_lock(&session->s_mutex);
3254 remove_session_caps(session);
3255 mutex_unlock(&session->s_mutex);
3256 ceph_put_mds_session(session);
3257 mutex_lock(&mdsc->mutex);
3260 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3261 mutex_unlock(&mdsc->mutex);
3263 ceph_cleanup_empty_realms(mdsc);
3265 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3270 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3273 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3275 ceph_mdsmap_destroy(mdsc->mdsmap);
3276 kfree(mdsc->sessions);
3277 ceph_caps_finalize(mdsc);
3280 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3282 struct ceph_mds_client *mdsc = fsc->mdsc;
3284 dout("mdsc_destroy %p\n", mdsc);
3285 ceph_mdsc_stop(mdsc);
3287 /* flush out any connection work with references to us */
3292 dout("mdsc_destroy %p done\n", mdsc);
3297 * handle mds map update.
3299 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3303 void *p = msg->front.iov_base;
3304 void *end = p + msg->front.iov_len;
3305 struct ceph_mdsmap *newmap, *oldmap;
3306 struct ceph_fsid fsid;
3309 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3310 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3311 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3313 epoch = ceph_decode_32(&p);
3314 maplen = ceph_decode_32(&p);
3315 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3317 /* do we need it? */
3318 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3319 mutex_lock(&mdsc->mutex);
3320 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3321 dout("handle_map epoch %u <= our %u\n",
3322 epoch, mdsc->mdsmap->m_epoch);
3323 mutex_unlock(&mdsc->mutex);
3327 newmap = ceph_mdsmap_decode(&p, end);
3328 if (IS_ERR(newmap)) {
3329 err = PTR_ERR(newmap);
3333 /* swap into place */
3335 oldmap = mdsc->mdsmap;
3336 mdsc->mdsmap = newmap;
3337 check_new_map(mdsc, newmap, oldmap);
3338 ceph_mdsmap_destroy(oldmap);
3340 mdsc->mdsmap = newmap; /* first mds map */
3342 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3344 __wake_requests(mdsc, &mdsc->waiting_for_map);
3346 mutex_unlock(&mdsc->mutex);
3347 schedule_delayed(mdsc);
3351 mutex_unlock(&mdsc->mutex);
3353 pr_err("error decoding mdsmap %d\n", err);
3357 static struct ceph_connection *con_get(struct ceph_connection *con)
3359 struct ceph_mds_session *s = con->private;
3361 if (get_session(s)) {
3362 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3365 dout("mdsc con_get %p FAIL\n", s);
3369 static void con_put(struct ceph_connection *con)
3371 struct ceph_mds_session *s = con->private;
3373 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3374 ceph_put_mds_session(s);
3378 * if the client is unresponsive for long enough, the mds will kill
3379 * the session entirely.
3381 static void peer_reset(struct ceph_connection *con)
3383 struct ceph_mds_session *s = con->private;
3384 struct ceph_mds_client *mdsc = s->s_mdsc;
3386 pr_warning("mds%d closed our session\n", s->s_mds);
3387 send_mds_reconnect(mdsc, s);
3390 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3392 struct ceph_mds_session *s = con->private;
3393 struct ceph_mds_client *mdsc = s->s_mdsc;
3394 int type = le16_to_cpu(msg->hdr.type);
3396 mutex_lock(&mdsc->mutex);
3397 if (__verify_registered_session(mdsc, s) < 0) {
3398 mutex_unlock(&mdsc->mutex);
3401 mutex_unlock(&mdsc->mutex);
3404 case CEPH_MSG_MDS_MAP:
3405 ceph_mdsc_handle_map(mdsc, msg);
3407 case CEPH_MSG_CLIENT_SESSION:
3408 handle_session(s, msg);
3410 case CEPH_MSG_CLIENT_REPLY:
3411 handle_reply(s, msg);
3413 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3414 handle_forward(mdsc, s, msg);
3416 case CEPH_MSG_CLIENT_CAPS:
3417 ceph_handle_caps(s, msg);
3419 case CEPH_MSG_CLIENT_SNAP:
3420 ceph_handle_snap(mdsc, s, msg);
3422 case CEPH_MSG_CLIENT_LEASE:
3423 handle_lease(mdsc, s, msg);
3427 pr_err("received unknown message type %d %s\n", type,
3428 ceph_msg_type_name(type));
3439 * Note: returned pointer is the address of a structure that's
3440 * managed separately. Caller must *not* attempt to free it.
3442 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3443 int *proto, int force_new)
3445 struct ceph_mds_session *s = con->private;
3446 struct ceph_mds_client *mdsc = s->s_mdsc;
3447 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3448 struct ceph_auth_handshake *auth = &s->s_auth;
3450 if (force_new && auth->authorizer) {
3451 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3452 auth->authorizer = NULL;
3454 if (!auth->authorizer) {
3455 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3458 return ERR_PTR(ret);
3460 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3463 return ERR_PTR(ret);
3465 *proto = ac->protocol;
3471 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3473 struct ceph_mds_session *s = con->private;
3474 struct ceph_mds_client *mdsc = s->s_mdsc;
3475 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3477 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3480 static int invalidate_authorizer(struct ceph_connection *con)
3482 struct ceph_mds_session *s = con->private;
3483 struct ceph_mds_client *mdsc = s->s_mdsc;
3484 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3486 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3488 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3491 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3492 struct ceph_msg_header *hdr, int *skip)
3494 struct ceph_msg *msg;
3495 int type = (int) le16_to_cpu(hdr->type);
3496 int front_len = (int) le32_to_cpu(hdr->front_len);
3502 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3504 pr_err("unable to allocate msg type %d len %d\n",
3512 static const struct ceph_connection_operations mds_con_ops = {
3515 .dispatch = dispatch,
3516 .get_authorizer = get_authorizer,
3517 .verify_authorizer_reply = verify_authorizer_reply,
3518 .invalidate_authorizer = invalidate_authorizer,
3519 .peer_reset = peer_reset,
3520 .alloc_msg = mds_alloc_msg,