1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/wait.h>
6 #include <linux/slab.h>
8 #include <linux/sched.h>
9 #include <linux/debugfs.h>
10 #include <linux/seq_file.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>
23 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
26 * A cluster of MDS (metadata server) daemons is responsible for
27 * managing the file system namespace (the directory hierarchy and
28 * inodes) and for coordinating shared access to storage. Metadata is
29 * partitioning hierarchically across a number of servers, and that
30 * partition varies over time as the cluster adjusts the distribution
31 * in order to balance load.
33 * The MDS client is primarily responsible to managing synchronous
34 * metadata requests for operations like open, unlink, and so forth.
35 * If there is a MDS failure, we find out about it when we (possibly
36 * request and) receive a new MDS map, and can resubmit affected
39 * For the most part, though, we take advantage of a lossless
40 * communications channel to the MDS, and do not need to worry about
41 * timing out or resubmitting requests.
43 * We maintain a stateful "session" with each MDS we interact with.
44 * Within each session, we sent periodic heartbeat messages to ensure
45 * any capabilities or leases we have been issues remain valid. If
46 * the session times out and goes stale, our leases and capabilities
47 * are no longer valid.
50 struct ceph_reconnect_state {
51 struct ceph_mds_session *session;
52 int nr_caps, nr_realms;
53 struct ceph_pagelist *pagelist;
58 static void __wake_requests(struct ceph_mds_client *mdsc,
59 struct list_head *head);
60 static void ceph_cap_release_work(struct work_struct *work);
61 static void ceph_cap_reclaim_work(struct work_struct *work);
63 static const struct ceph_connection_operations mds_con_ops;
70 static int parse_reply_info_quota(void **p, void *end,
71 struct ceph_mds_reply_info_in *info)
73 u8 struct_v, struct_compat;
76 ceph_decode_8_safe(p, end, struct_v, bad);
77 ceph_decode_8_safe(p, end, struct_compat, bad);
78 /* struct_v is expected to be >= 1. we only
79 * understand encoding with struct_compat == 1. */
80 if (!struct_v || struct_compat != 1)
82 ceph_decode_32_safe(p, end, struct_len, bad);
83 ceph_decode_need(p, end, struct_len, bad);
84 end = *p + struct_len;
85 ceph_decode_64_safe(p, end, info->max_bytes, bad);
86 ceph_decode_64_safe(p, end, info->max_files, bad);
94 * parse individual inode info
96 static int parse_reply_info_in(void **p, void *end,
97 struct ceph_mds_reply_info_in *info,
103 if (features == (u64)-1) {
106 ceph_decode_8_safe(p, end, struct_v, bad);
107 ceph_decode_8_safe(p, end, struct_compat, bad);
108 /* struct_v is expected to be >= 1. we only understand
109 * encoding with struct_compat == 1. */
110 if (!struct_v || struct_compat != 1)
112 ceph_decode_32_safe(p, end, struct_len, bad);
113 ceph_decode_need(p, end, struct_len, bad);
114 end = *p + struct_len;
117 ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
119 *p += sizeof(struct ceph_mds_reply_inode) +
120 sizeof(*info->in->fragtree.splits) *
121 le32_to_cpu(info->in->fragtree.nsplits);
123 ceph_decode_32_safe(p, end, info->symlink_len, bad);
124 ceph_decode_need(p, end, info->symlink_len, bad);
126 *p += info->symlink_len;
128 ceph_decode_copy_safe(p, end, &info->dir_layout,
129 sizeof(info->dir_layout), bad);
130 ceph_decode_32_safe(p, end, info->xattr_len, bad);
131 ceph_decode_need(p, end, info->xattr_len, bad);
132 info->xattr_data = *p;
133 *p += info->xattr_len;
135 if (features == (u64)-1) {
137 ceph_decode_64_safe(p, end, info->inline_version, bad);
138 ceph_decode_32_safe(p, end, info->inline_len, bad);
139 ceph_decode_need(p, end, info->inline_len, bad);
140 info->inline_data = *p;
141 *p += info->inline_len;
143 err = parse_reply_info_quota(p, end, info);
147 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
148 if (info->pool_ns_len > 0) {
149 ceph_decode_need(p, end, info->pool_ns_len, bad);
150 info->pool_ns_data = *p;
151 *p += info->pool_ns_len;
153 /* btime, change_attr */
155 struct ceph_timespec btime;
157 ceph_decode_need(p, end, sizeof(btime), bad);
158 ceph_decode_copy(p, &btime, sizeof(btime));
159 ceph_decode_64_safe(p, end, change_attr, bad);
164 ceph_decode_32_safe(p, end, info->dir_pin, bad);
166 info->dir_pin = -ENODATA;
171 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
172 ceph_decode_64_safe(p, end, info->inline_version, bad);
173 ceph_decode_32_safe(p, end, info->inline_len, bad);
174 ceph_decode_need(p, end, info->inline_len, bad);
175 info->inline_data = *p;
176 *p += info->inline_len;
178 info->inline_version = CEPH_INLINE_NONE;
180 if (features & CEPH_FEATURE_MDS_QUOTA) {
181 err = parse_reply_info_quota(p, end, info);
189 info->pool_ns_len = 0;
190 info->pool_ns_data = NULL;
191 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
192 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
193 if (info->pool_ns_len > 0) {
194 ceph_decode_need(p, end, info->pool_ns_len, bad);
195 info->pool_ns_data = *p;
196 *p += info->pool_ns_len;
200 info->dir_pin = -ENODATA;
209 static int parse_reply_info_dir(void **p, void *end,
210 struct ceph_mds_reply_dirfrag **dirfrag,
213 if (features == (u64)-1) {
214 u8 struct_v, struct_compat;
216 ceph_decode_8_safe(p, end, struct_v, bad);
217 ceph_decode_8_safe(p, end, struct_compat, bad);
218 /* struct_v is expected to be >= 1. we only understand
219 * encoding whose struct_compat == 1. */
220 if (!struct_v || struct_compat != 1)
222 ceph_decode_32_safe(p, end, struct_len, bad);
223 ceph_decode_need(p, end, struct_len, bad);
224 end = *p + struct_len;
227 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
229 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
230 if (unlikely(*p > end))
232 if (features == (u64)-1)
239 static int parse_reply_info_lease(void **p, void *end,
240 struct ceph_mds_reply_lease **lease,
243 if (features == (u64)-1) {
244 u8 struct_v, struct_compat;
246 ceph_decode_8_safe(p, end, struct_v, bad);
247 ceph_decode_8_safe(p, end, struct_compat, bad);
248 /* struct_v is expected to be >= 1. we only understand
249 * encoding whose struct_compat == 1. */
250 if (!struct_v || struct_compat != 1)
252 ceph_decode_32_safe(p, end, struct_len, bad);
253 ceph_decode_need(p, end, struct_len, bad);
254 end = *p + struct_len;
257 ceph_decode_need(p, end, sizeof(**lease), bad);
259 *p += sizeof(**lease);
260 if (features == (u64)-1)
268 * parse a normal reply, which may contain a (dir+)dentry and/or a
271 static int parse_reply_info_trace(void **p, void *end,
272 struct ceph_mds_reply_info_parsed *info,
277 if (info->head->is_dentry) {
278 err = parse_reply_info_in(p, end, &info->diri, features);
282 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
286 ceph_decode_32_safe(p, end, info->dname_len, bad);
287 ceph_decode_need(p, end, info->dname_len, bad);
289 *p += info->dname_len;
291 err = parse_reply_info_lease(p, end, &info->dlease, features);
296 if (info->head->is_target) {
297 err = parse_reply_info_in(p, end, &info->targeti, features);
302 if (unlikely(*p != end))
309 pr_err("problem parsing mds trace %d\n", err);
314 * parse readdir results
316 static int parse_reply_info_readdir(void **p, void *end,
317 struct ceph_mds_reply_info_parsed *info,
323 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
327 ceph_decode_need(p, end, sizeof(num) + 2, bad);
328 num = ceph_decode_32(p);
330 u16 flags = ceph_decode_16(p);
331 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
332 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
333 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
334 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
339 BUG_ON(!info->dir_entries);
340 if ((unsigned long)(info->dir_entries + num) >
341 (unsigned long)info->dir_entries + info->dir_buf_size) {
342 pr_err("dir contents are larger than expected\n");
349 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
351 ceph_decode_32_safe(p, end, rde->name_len, bad);
352 ceph_decode_need(p, end, rde->name_len, bad);
355 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
358 err = parse_reply_info_lease(p, end, &rde->lease, features);
362 err = parse_reply_info_in(p, end, &rde->inode, features);
365 /* ceph_readdir_prepopulate() will update it */
379 pr_err("problem parsing dir contents %d\n", err);
384 * parse fcntl F_GETLK results
386 static int parse_reply_info_filelock(void **p, void *end,
387 struct ceph_mds_reply_info_parsed *info,
390 if (*p + sizeof(*info->filelock_reply) > end)
393 info->filelock_reply = *p;
394 *p += sizeof(*info->filelock_reply);
396 if (unlikely(*p != end))
405 * parse create results
407 static int parse_reply_info_create(void **p, void *end,
408 struct ceph_mds_reply_info_parsed *info,
411 if (features == (u64)-1 ||
412 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
414 info->has_create_ino = false;
416 info->has_create_ino = true;
417 info->ino = ceph_decode_64(p);
421 if (unlikely(*p != end))
430 * parse extra results
432 static int parse_reply_info_extra(void **p, void *end,
433 struct ceph_mds_reply_info_parsed *info,
436 u32 op = le32_to_cpu(info->head->op);
438 if (op == CEPH_MDS_OP_GETFILELOCK)
439 return parse_reply_info_filelock(p, end, info, features);
440 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
441 return parse_reply_info_readdir(p, end, info, features);
442 else if (op == CEPH_MDS_OP_CREATE)
443 return parse_reply_info_create(p, end, info, features);
449 * parse entire mds reply
451 static int parse_reply_info(struct ceph_msg *msg,
452 struct ceph_mds_reply_info_parsed *info,
459 info->head = msg->front.iov_base;
460 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
461 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
464 ceph_decode_32_safe(&p, end, len, bad);
466 ceph_decode_need(&p, end, len, bad);
467 err = parse_reply_info_trace(&p, p+len, info, features);
473 ceph_decode_32_safe(&p, end, len, bad);
475 ceph_decode_need(&p, end, len, bad);
476 err = parse_reply_info_extra(&p, p+len, info, features);
482 ceph_decode_32_safe(&p, end, len, bad);
483 info->snapblob_len = len;
494 pr_err("mds parse_reply err %d\n", err);
498 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
500 if (!info->dir_entries)
502 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
509 const char *ceph_session_state_name(int s)
512 case CEPH_MDS_SESSION_NEW: return "new";
513 case CEPH_MDS_SESSION_OPENING: return "opening";
514 case CEPH_MDS_SESSION_OPEN: return "open";
515 case CEPH_MDS_SESSION_HUNG: return "hung";
516 case CEPH_MDS_SESSION_CLOSING: return "closing";
517 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
518 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
519 case CEPH_MDS_SESSION_REJECTED: return "rejected";
520 default: return "???";
524 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
526 if (refcount_inc_not_zero(&s->s_ref)) {
527 dout("mdsc get_session %p %d -> %d\n", s,
528 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
531 dout("mdsc get_session %p 0 -- FAIL\n", s);
536 void ceph_put_mds_session(struct ceph_mds_session *s)
538 dout("mdsc put_session %p %d -> %d\n", s,
539 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
540 if (refcount_dec_and_test(&s->s_ref)) {
541 if (s->s_auth.authorizer)
542 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
548 * called under mdsc->mutex
550 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
553 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
555 return get_session(mdsc->sessions[mds]);
558 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
560 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
566 static int __verify_registered_session(struct ceph_mds_client *mdsc,
567 struct ceph_mds_session *s)
569 if (s->s_mds >= mdsc->max_sessions ||
570 mdsc->sessions[s->s_mds] != s)
576 * create+register a new session for given mds.
577 * called under mdsc->mutex.
579 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
582 struct ceph_mds_session *s;
584 if (mds >= mdsc->mdsmap->m_num_mds)
585 return ERR_PTR(-EINVAL);
587 s = kzalloc(sizeof(*s), GFP_NOFS);
589 return ERR_PTR(-ENOMEM);
591 if (mds >= mdsc->max_sessions) {
592 int newmax = 1 << get_count_order(mds + 1);
593 struct ceph_mds_session **sa;
595 dout("%s: realloc to %d\n", __func__, newmax);
596 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
599 if (mdsc->sessions) {
600 memcpy(sa, mdsc->sessions,
601 mdsc->max_sessions * sizeof(void *));
602 kfree(mdsc->sessions);
605 mdsc->max_sessions = newmax;
608 dout("%s: mds%d\n", __func__, mds);
611 s->s_state = CEPH_MDS_SESSION_NEW;
614 mutex_init(&s->s_mutex);
616 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
618 spin_lock_init(&s->s_gen_ttl_lock);
620 s->s_cap_ttl = jiffies - 1;
622 spin_lock_init(&s->s_cap_lock);
623 s->s_renew_requested = 0;
625 INIT_LIST_HEAD(&s->s_caps);
628 refcount_set(&s->s_ref, 1);
629 INIT_LIST_HEAD(&s->s_waiting);
630 INIT_LIST_HEAD(&s->s_unsafe);
631 s->s_num_cap_releases = 0;
632 s->s_cap_reconnect = 0;
633 s->s_cap_iterator = NULL;
634 INIT_LIST_HEAD(&s->s_cap_releases);
635 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
637 INIT_LIST_HEAD(&s->s_cap_flushing);
639 mdsc->sessions[mds] = s;
640 atomic_inc(&mdsc->num_sessions);
641 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
643 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
644 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
650 return ERR_PTR(-ENOMEM);
654 * called under mdsc->mutex
656 static void __unregister_session(struct ceph_mds_client *mdsc,
657 struct ceph_mds_session *s)
659 dout("__unregister_session mds%d %p\n", s->s_mds, s);
660 BUG_ON(mdsc->sessions[s->s_mds] != s);
661 mdsc->sessions[s->s_mds] = NULL;
663 ceph_con_close(&s->s_con);
664 ceph_put_mds_session(s);
665 atomic_dec(&mdsc->num_sessions);
669 * drop session refs in request.
671 * should be last request ref, or hold mdsc->mutex
673 static void put_request_session(struct ceph_mds_request *req)
675 if (req->r_session) {
676 ceph_put_mds_session(req->r_session);
677 req->r_session = NULL;
681 void ceph_mdsc_release_request(struct kref *kref)
683 struct ceph_mds_request *req = container_of(kref,
684 struct ceph_mds_request,
686 destroy_reply_info(&req->r_reply_info);
688 ceph_msg_put(req->r_request);
690 ceph_msg_put(req->r_reply);
692 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
693 /* avoid calling iput_final() in mds dispatch threads */
694 ceph_async_iput(req->r_inode);
697 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
698 ceph_async_iput(req->r_target_inode);
701 if (req->r_old_dentry)
702 dput(req->r_old_dentry);
703 if (req->r_old_dentry_dir) {
705 * track (and drop pins for) r_old_dentry_dir
706 * separately, since r_old_dentry's d_parent may have
707 * changed between the dir mutex being dropped and
708 * this request being freed.
710 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
712 ceph_async_iput(req->r_old_dentry_dir);
717 ceph_pagelist_release(req->r_pagelist);
718 put_request_session(req);
719 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
723 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
726 * lookup session, bump ref if found.
728 * called under mdsc->mutex.
730 static struct ceph_mds_request *
731 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
733 struct ceph_mds_request *req;
735 req = lookup_request(&mdsc->request_tree, tid);
737 ceph_mdsc_get_request(req);
743 * Register an in-flight request, and assign a tid. Link to directory
744 * are modifying (if any).
746 * Called under mdsc->mutex.
748 static void __register_request(struct ceph_mds_client *mdsc,
749 struct ceph_mds_request *req,
754 req->r_tid = ++mdsc->last_tid;
755 if (req->r_num_caps) {
756 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
759 pr_err("__register_request %p "
760 "failed to reserve caps: %d\n", req, ret);
761 /* set req->r_err to fail early from __do_request */
766 dout("__register_request %p tid %lld\n", req, req->r_tid);
767 ceph_mdsc_get_request(req);
768 insert_request(&mdsc->request_tree, req);
770 req->r_uid = current_fsuid();
771 req->r_gid = current_fsgid();
773 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
774 mdsc->oldest_tid = req->r_tid;
778 req->r_unsafe_dir = dir;
782 static void __unregister_request(struct ceph_mds_client *mdsc,
783 struct ceph_mds_request *req)
785 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
787 /* Never leave an unregistered request on an unsafe list! */
788 list_del_init(&req->r_unsafe_item);
790 if (req->r_tid == mdsc->oldest_tid) {
791 struct rb_node *p = rb_next(&req->r_node);
792 mdsc->oldest_tid = 0;
794 struct ceph_mds_request *next_req =
795 rb_entry(p, struct ceph_mds_request, r_node);
796 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
797 mdsc->oldest_tid = next_req->r_tid;
804 erase_request(&mdsc->request_tree, req);
806 if (req->r_unsafe_dir &&
807 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
808 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
809 spin_lock(&ci->i_unsafe_lock);
810 list_del_init(&req->r_unsafe_dir_item);
811 spin_unlock(&ci->i_unsafe_lock);
813 if (req->r_target_inode &&
814 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
815 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
816 spin_lock(&ci->i_unsafe_lock);
817 list_del_init(&req->r_unsafe_target_item);
818 spin_unlock(&ci->i_unsafe_lock);
821 if (req->r_unsafe_dir) {
822 /* avoid calling iput_final() in mds dispatch threads */
823 ceph_async_iput(req->r_unsafe_dir);
824 req->r_unsafe_dir = NULL;
827 complete_all(&req->r_safe_completion);
829 ceph_mdsc_put_request(req);
833 * Walk back up the dentry tree until we hit a dentry representing a
834 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
835 * when calling this) to ensure that the objects won't disappear while we're
836 * working with them. Once we hit a candidate dentry, we attempt to take a
837 * reference to it, and return that as the result.
839 static struct inode *get_nonsnap_parent(struct dentry *dentry)
841 struct inode *inode = NULL;
843 while (dentry && !IS_ROOT(dentry)) {
844 inode = d_inode_rcu(dentry);
845 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
847 dentry = dentry->d_parent;
850 inode = igrab(inode);
855 * Choose mds to send request to next. If there is a hint set in the
856 * request (e.g., due to a prior forward hint from the mds), use that.
857 * Otherwise, consult frag tree and/or caps to identify the
858 * appropriate mds. If all else fails, choose randomly.
860 * Called under mdsc->mutex.
862 static int __choose_mds(struct ceph_mds_client *mdsc,
863 struct ceph_mds_request *req)
866 struct ceph_inode_info *ci;
867 struct ceph_cap *cap;
868 int mode = req->r_direct_mode;
870 u32 hash = req->r_direct_hash;
871 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
874 * is there a specific mds we should try? ignore hint if we have
875 * no session and the mds is not up (active or recovering).
877 if (req->r_resend_mds >= 0 &&
878 (__have_session(mdsc, req->r_resend_mds) ||
879 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
880 dout("choose_mds using resend_mds mds%d\n",
882 return req->r_resend_mds;
885 if (mode == USE_RANDOM_MDS)
890 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
891 inode = req->r_inode;
894 /* req->r_dentry is non-null for LSSNAP request */
896 inode = get_nonsnap_parent(req->r_dentry);
898 dout("__choose_mds using snapdir's parent %p\n", inode);
900 } else if (req->r_dentry) {
901 /* ignore race with rename; old or new d_parent is okay */
902 struct dentry *parent;
906 parent = req->r_dentry->d_parent;
907 dir = req->r_parent ? : d_inode_rcu(parent);
909 if (!dir || dir->i_sb != mdsc->fsc->sb) {
910 /* not this fs or parent went negative */
911 inode = d_inode(req->r_dentry);
914 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
915 /* direct snapped/virtual snapdir requests
916 * based on parent dir inode */
917 inode = get_nonsnap_parent(parent);
918 dout("__choose_mds using nonsnap parent %p\n", inode);
921 inode = d_inode(req->r_dentry);
922 if (!inode || mode == USE_AUTH_MDS) {
925 hash = ceph_dentry_hash(dir, req->r_dentry);
934 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
938 ci = ceph_inode(inode);
940 if (is_hash && S_ISDIR(inode->i_mode)) {
941 struct ceph_inode_frag frag;
944 ceph_choose_frag(ci, hash, &frag, &found);
946 if (mode == USE_ANY_MDS && frag.ndist > 0) {
949 /* choose a random replica */
950 get_random_bytes(&r, 1);
953 dout("choose_mds %p %llx.%llx "
954 "frag %u mds%d (%d/%d)\n",
955 inode, ceph_vinop(inode),
958 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
959 CEPH_MDS_STATE_ACTIVE)
963 /* since this file/dir wasn't known to be
964 * replicated, then we want to look for the
965 * authoritative mds. */
968 /* choose auth mds */
970 dout("choose_mds %p %llx.%llx "
971 "frag %u mds%d (auth)\n",
972 inode, ceph_vinop(inode), frag.frag, mds);
973 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
974 CEPH_MDS_STATE_ACTIVE)
980 spin_lock(&ci->i_ceph_lock);
982 if (mode == USE_AUTH_MDS)
983 cap = ci->i_auth_cap;
984 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
985 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
987 spin_unlock(&ci->i_ceph_lock);
988 ceph_async_iput(inode);
991 mds = cap->session->s_mds;
992 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
993 inode, ceph_vinop(inode), mds,
994 cap == ci->i_auth_cap ? "auth " : "", cap);
995 spin_unlock(&ci->i_ceph_lock);
997 /* avoid calling iput_final() while holding mdsc->mutex or
998 * in mds dispatch threads */
999 ceph_async_iput(inode);
1003 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1004 dout("choose_mds chose random mds%d\n", mds);
1012 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1014 struct ceph_msg *msg;
1015 struct ceph_mds_session_head *h;
1017 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1020 pr_err("create_session_msg ENOMEM creating msg\n");
1023 h = msg->front.iov_base;
1024 h->op = cpu_to_le32(op);
1025 h->seq = cpu_to_le64(seq);
1030 static void encode_supported_features(void **p, void *end)
1032 static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1033 static const size_t count = ARRAY_SIZE(bits);
1037 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
1039 BUG_ON(*p + 4 + size > end);
1040 ceph_encode_32(p, size);
1041 memset(*p, 0, size);
1042 for (i = 0; i < count; i++)
1043 ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
1046 BUG_ON(*p + 4 > end);
1047 ceph_encode_32(p, 0);
1052 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1053 * to include additional client metadata fields.
1055 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1057 struct ceph_msg *msg;
1058 struct ceph_mds_session_head *h;
1060 int extra_bytes = 0;
1061 int metadata_key_count = 0;
1062 struct ceph_options *opt = mdsc->fsc->client->options;
1063 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1066 const char* metadata[][2] = {
1067 {"hostname", mdsc->nodename},
1068 {"kernel_version", init_utsname()->release},
1069 {"entity_id", opt->name ? : ""},
1070 {"root", fsopt->server_path ? : "/"},
1074 /* Calculate serialized length of metadata */
1075 extra_bytes = 4; /* map length */
1076 for (i = 0; metadata[i][0]; ++i) {
1077 extra_bytes += 8 + strlen(metadata[i][0]) +
1078 strlen(metadata[i][1]);
1079 metadata_key_count++;
1081 /* supported feature */
1082 extra_bytes += 4 + 8;
1084 /* Allocate the message */
1085 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1088 pr_err("create_session_msg ENOMEM creating msg\n");
1091 p = msg->front.iov_base;
1092 end = p + msg->front.iov_len;
1095 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1096 h->seq = cpu_to_le64(seq);
1099 * Serialize client metadata into waiting buffer space, using
1100 * the format that userspace expects for map<string, string>
1102 * ClientSession messages with metadata are v2
1104 msg->hdr.version = cpu_to_le16(3);
1105 msg->hdr.compat_version = cpu_to_le16(1);
1107 /* The write pointer, following the session_head structure */
1110 /* Number of entries in the map */
1111 ceph_encode_32(&p, metadata_key_count);
1113 /* Two length-prefixed strings for each entry in the map */
1114 for (i = 0; metadata[i][0]; ++i) {
1115 size_t const key_len = strlen(metadata[i][0]);
1116 size_t const val_len = strlen(metadata[i][1]);
1118 ceph_encode_32(&p, key_len);
1119 memcpy(p, metadata[i][0], key_len);
1121 ceph_encode_32(&p, val_len);
1122 memcpy(p, metadata[i][1], val_len);
1126 encode_supported_features(&p, end);
1127 msg->front.iov_len = p - msg->front.iov_base;
1128 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1134 * send session open request.
1136 * called under mdsc->mutex
1138 static int __open_session(struct ceph_mds_client *mdsc,
1139 struct ceph_mds_session *session)
1141 struct ceph_msg *msg;
1143 int mds = session->s_mds;
1145 /* wait for mds to go active? */
1146 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1147 dout("open_session to mds%d (%s)\n", mds,
1148 ceph_mds_state_name(mstate));
1149 session->s_state = CEPH_MDS_SESSION_OPENING;
1150 session->s_renew_requested = jiffies;
1152 /* send connect message */
1153 msg = create_session_open_msg(mdsc, session->s_seq);
1156 ceph_con_send(&session->s_con, msg);
1161 * open sessions for any export targets for the given mds
1163 * called under mdsc->mutex
1165 static struct ceph_mds_session *
1166 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1168 struct ceph_mds_session *session;
1170 session = __ceph_lookup_mds_session(mdsc, target);
1172 session = register_session(mdsc, target);
1173 if (IS_ERR(session))
1176 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1177 session->s_state == CEPH_MDS_SESSION_CLOSING)
1178 __open_session(mdsc, session);
1183 struct ceph_mds_session *
1184 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1186 struct ceph_mds_session *session;
1188 dout("open_export_target_session to mds%d\n", target);
1190 mutex_lock(&mdsc->mutex);
1191 session = __open_export_target_session(mdsc, target);
1192 mutex_unlock(&mdsc->mutex);
1197 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1198 struct ceph_mds_session *session)
1200 struct ceph_mds_info *mi;
1201 struct ceph_mds_session *ts;
1202 int i, mds = session->s_mds;
1204 if (mds >= mdsc->mdsmap->m_num_mds)
1207 mi = &mdsc->mdsmap->m_info[mds];
1208 dout("open_export_target_sessions for mds%d (%d targets)\n",
1209 session->s_mds, mi->num_export_targets);
1211 for (i = 0; i < mi->num_export_targets; i++) {
1212 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1214 ceph_put_mds_session(ts);
1218 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1219 struct ceph_mds_session *session)
1221 mutex_lock(&mdsc->mutex);
1222 __open_export_target_sessions(mdsc, session);
1223 mutex_unlock(&mdsc->mutex);
1230 static void detach_cap_releases(struct ceph_mds_session *session,
1231 struct list_head *target)
1233 lockdep_assert_held(&session->s_cap_lock);
1235 list_splice_init(&session->s_cap_releases, target);
1236 session->s_num_cap_releases = 0;
1237 dout("dispose_cap_releases mds%d\n", session->s_mds);
1240 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1241 struct list_head *dispose)
1243 while (!list_empty(dispose)) {
1244 struct ceph_cap *cap;
1245 /* zero out the in-progress message */
1246 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1247 list_del(&cap->session_caps);
1248 ceph_put_cap(mdsc, cap);
1252 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1253 struct ceph_mds_session *session)
1255 struct ceph_mds_request *req;
1258 dout("cleanup_session_requests mds%d\n", session->s_mds);
1259 mutex_lock(&mdsc->mutex);
1260 while (!list_empty(&session->s_unsafe)) {
1261 req = list_first_entry(&session->s_unsafe,
1262 struct ceph_mds_request, r_unsafe_item);
1263 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1265 __unregister_request(mdsc, req);
1267 /* zero r_attempts, so kick_requests() will re-send requests */
1268 p = rb_first(&mdsc->request_tree);
1270 req = rb_entry(p, struct ceph_mds_request, r_node);
1272 if (req->r_session &&
1273 req->r_session->s_mds == session->s_mds)
1274 req->r_attempts = 0;
1276 mutex_unlock(&mdsc->mutex);
1280 * Helper to safely iterate over all caps associated with a session, with
1281 * special care taken to handle a racing __ceph_remove_cap().
1283 * Caller must hold session s_mutex.
1285 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1286 int (*cb)(struct inode *, struct ceph_cap *,
1289 struct list_head *p;
1290 struct ceph_cap *cap;
1291 struct inode *inode, *last_inode = NULL;
1292 struct ceph_cap *old_cap = NULL;
1295 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1296 spin_lock(&session->s_cap_lock);
1297 p = session->s_caps.next;
1298 while (p != &session->s_caps) {
1299 cap = list_entry(p, struct ceph_cap, session_caps);
1300 inode = igrab(&cap->ci->vfs_inode);
1305 session->s_cap_iterator = cap;
1306 spin_unlock(&session->s_cap_lock);
1309 /* avoid calling iput_final() while holding
1310 * s_mutex or in mds dispatch threads */
1311 ceph_async_iput(last_inode);
1315 ceph_put_cap(session->s_mdsc, old_cap);
1319 ret = cb(inode, cap, arg);
1322 spin_lock(&session->s_cap_lock);
1325 dout("iterate_session_caps finishing cap %p removal\n",
1327 BUG_ON(cap->session != session);
1328 cap->session = NULL;
1329 list_del_init(&cap->session_caps);
1330 session->s_nr_caps--;
1331 if (cap->queue_release)
1332 __ceph_queue_cap_release(session, cap);
1334 old_cap = cap; /* put_cap it w/o locks held */
1341 session->s_cap_iterator = NULL;
1342 spin_unlock(&session->s_cap_lock);
1344 ceph_async_iput(last_inode);
1346 ceph_put_cap(session->s_mdsc, old_cap);
1351 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1354 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1355 struct ceph_inode_info *ci = ceph_inode(inode);
1356 LIST_HEAD(to_remove);
1358 bool invalidate = false;
1360 dout("removing cap %p, ci is %p, inode is %p\n",
1361 cap, ci, &ci->vfs_inode);
1362 spin_lock(&ci->i_ceph_lock);
1363 if (cap->mds_wanted | cap->issued)
1364 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1365 __ceph_remove_cap(cap, false);
1366 if (!ci->i_auth_cap) {
1367 struct ceph_cap_flush *cf;
1368 struct ceph_mds_client *mdsc = fsc->mdsc;
1370 if (ci->i_wrbuffer_ref > 0 &&
1371 READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1374 while (!list_empty(&ci->i_cap_flush_list)) {
1375 cf = list_first_entry(&ci->i_cap_flush_list,
1376 struct ceph_cap_flush, i_list);
1377 list_move(&cf->i_list, &to_remove);
1380 spin_lock(&mdsc->cap_dirty_lock);
1382 list_for_each_entry(cf, &to_remove, i_list)
1383 list_del(&cf->g_list);
1385 if (!list_empty(&ci->i_dirty_item)) {
1386 pr_warn_ratelimited(
1387 " dropping dirty %s state for %p %lld\n",
1388 ceph_cap_string(ci->i_dirty_caps),
1389 inode, ceph_ino(inode));
1390 ci->i_dirty_caps = 0;
1391 list_del_init(&ci->i_dirty_item);
1394 if (!list_empty(&ci->i_flushing_item)) {
1395 pr_warn_ratelimited(
1396 " dropping dirty+flushing %s state for %p %lld\n",
1397 ceph_cap_string(ci->i_flushing_caps),
1398 inode, ceph_ino(inode));
1399 ci->i_flushing_caps = 0;
1400 list_del_init(&ci->i_flushing_item);
1401 mdsc->num_cap_flushing--;
1404 spin_unlock(&mdsc->cap_dirty_lock);
1406 if (atomic_read(&ci->i_filelock_ref) > 0) {
1407 /* make further file lock syscall return -EIO */
1408 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1409 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1410 inode, ceph_ino(inode));
1413 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1414 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1415 ci->i_prealloc_cap_flush = NULL;
1419 ci->i_wrbuffer_ref_head == 0 &&
1420 ci->i_wr_ref == 0 &&
1421 ci->i_dirty_caps == 0 &&
1422 ci->i_flushing_caps == 0) {
1423 ceph_put_snap_context(ci->i_head_snapc);
1424 ci->i_head_snapc = NULL;
1427 spin_unlock(&ci->i_ceph_lock);
1428 while (!list_empty(&to_remove)) {
1429 struct ceph_cap_flush *cf;
1430 cf = list_first_entry(&to_remove,
1431 struct ceph_cap_flush, i_list);
1432 list_del(&cf->i_list);
1433 ceph_free_cap_flush(cf);
1436 wake_up_all(&ci->i_cap_wq);
1438 ceph_queue_invalidate(inode);
1445 * caller must hold session s_mutex
1447 static void remove_session_caps(struct ceph_mds_session *session)
1449 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1450 struct super_block *sb = fsc->sb;
1453 dout("remove_session_caps on %p\n", session);
1454 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1456 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1458 spin_lock(&session->s_cap_lock);
1459 if (session->s_nr_caps > 0) {
1460 struct inode *inode;
1461 struct ceph_cap *cap, *prev = NULL;
1462 struct ceph_vino vino;
1464 * iterate_session_caps() skips inodes that are being
1465 * deleted, we need to wait until deletions are complete.
1466 * __wait_on_freeing_inode() is designed for the job,
1467 * but it is not exported, so use lookup inode function
1470 while (!list_empty(&session->s_caps)) {
1471 cap = list_entry(session->s_caps.next,
1472 struct ceph_cap, session_caps);
1476 vino = cap->ci->i_vino;
1477 spin_unlock(&session->s_cap_lock);
1479 inode = ceph_find_inode(sb, vino);
1480 /* avoid calling iput_final() while holding s_mutex */
1481 ceph_async_iput(inode);
1483 spin_lock(&session->s_cap_lock);
1487 // drop cap expires and unlock s_cap_lock
1488 detach_cap_releases(session, &dispose);
1490 BUG_ON(session->s_nr_caps > 0);
1491 BUG_ON(!list_empty(&session->s_cap_flushing));
1492 spin_unlock(&session->s_cap_lock);
1493 dispose_cap_releases(session->s_mdsc, &dispose);
1503 * wake up any threads waiting on this session's caps. if the cap is
1504 * old (didn't get renewed on the client reconnect), remove it now.
1506 * caller must hold s_mutex.
1508 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1511 struct ceph_inode_info *ci = ceph_inode(inode);
1512 unsigned long ev = (unsigned long)arg;
1514 if (ev == RECONNECT) {
1515 spin_lock(&ci->i_ceph_lock);
1516 ci->i_wanted_max_size = 0;
1517 ci->i_requested_max_size = 0;
1518 spin_unlock(&ci->i_ceph_lock);
1519 } else if (ev == RENEWCAPS) {
1520 if (cap->cap_gen < cap->session->s_cap_gen) {
1521 /* mds did not re-issue stale cap */
1522 spin_lock(&ci->i_ceph_lock);
1523 cap->issued = cap->implemented = CEPH_CAP_PIN;
1524 /* make sure mds knows what we want */
1525 if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1526 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1527 spin_unlock(&ci->i_ceph_lock);
1529 } else if (ev == FORCE_RO) {
1531 wake_up_all(&ci->i_cap_wq);
1535 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1537 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1538 ceph_iterate_session_caps(session, wake_up_session_cb,
1539 (void *)(unsigned long)ev);
1543 * Send periodic message to MDS renewing all currently held caps. The
1544 * ack will reset the expiration for all caps from this session.
1546 * caller holds s_mutex
1548 static int send_renew_caps(struct ceph_mds_client *mdsc,
1549 struct ceph_mds_session *session)
1551 struct ceph_msg *msg;
1554 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1555 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1556 pr_info("mds%d caps stale\n", session->s_mds);
1557 session->s_renew_requested = jiffies;
1559 /* do not try to renew caps until a recovering mds has reconnected
1560 * with its clients. */
1561 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1562 if (state < CEPH_MDS_STATE_RECONNECT) {
1563 dout("send_renew_caps ignoring mds%d (%s)\n",
1564 session->s_mds, ceph_mds_state_name(state));
1568 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1569 ceph_mds_state_name(state));
1570 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1571 ++session->s_renew_seq);
1574 ceph_con_send(&session->s_con, msg);
1578 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1579 struct ceph_mds_session *session, u64 seq)
1581 struct ceph_msg *msg;
1583 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1584 session->s_mds, ceph_session_state_name(session->s_state), seq);
1585 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1588 ceph_con_send(&session->s_con, msg);
1594 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1596 * Called under session->s_mutex
1598 static void renewed_caps(struct ceph_mds_client *mdsc,
1599 struct ceph_mds_session *session, int is_renew)
1604 spin_lock(&session->s_cap_lock);
1605 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1607 session->s_cap_ttl = session->s_renew_requested +
1608 mdsc->mdsmap->m_session_timeout*HZ;
1611 if (time_before(jiffies, session->s_cap_ttl)) {
1612 pr_info("mds%d caps renewed\n", session->s_mds);
1615 pr_info("mds%d caps still stale\n", session->s_mds);
1618 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1619 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1620 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1621 spin_unlock(&session->s_cap_lock);
1624 wake_up_session_caps(session, RENEWCAPS);
1628 * send a session close request
1630 static int request_close_session(struct ceph_mds_client *mdsc,
1631 struct ceph_mds_session *session)
1633 struct ceph_msg *msg;
1635 dout("request_close_session mds%d state %s seq %lld\n",
1636 session->s_mds, ceph_session_state_name(session->s_state),
1638 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1641 ceph_con_send(&session->s_con, msg);
1646 * Called with s_mutex held.
1648 static int __close_session(struct ceph_mds_client *mdsc,
1649 struct ceph_mds_session *session)
1651 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1653 session->s_state = CEPH_MDS_SESSION_CLOSING;
1654 return request_close_session(mdsc, session);
1657 static bool drop_negative_children(struct dentry *dentry)
1659 struct dentry *child;
1660 bool all_negative = true;
1662 if (!d_is_dir(dentry))
1665 spin_lock(&dentry->d_lock);
1666 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1667 if (d_really_is_positive(child)) {
1668 all_negative = false;
1672 spin_unlock(&dentry->d_lock);
1675 shrink_dcache_parent(dentry);
1677 return all_negative;
1681 * Trim old(er) caps.
1683 * Because we can't cache an inode without one or more caps, we do
1684 * this indirectly: if a cap is unused, we prune its aliases, at which
1685 * point the inode will hopefully get dropped to.
1687 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1688 * memory pressure from the MDS, though, so it needn't be perfect.
1690 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1692 struct ceph_mds_session *session = arg;
1693 struct ceph_inode_info *ci = ceph_inode(inode);
1694 int used, wanted, oissued, mine;
1696 if (session->s_trim_caps <= 0)
1699 spin_lock(&ci->i_ceph_lock);
1700 mine = cap->issued | cap->implemented;
1701 used = __ceph_caps_used(ci);
1702 wanted = __ceph_caps_file_wanted(ci);
1703 oissued = __ceph_caps_issued_other(ci, cap);
1705 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1706 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1707 ceph_cap_string(used), ceph_cap_string(wanted));
1708 if (cap == ci->i_auth_cap) {
1709 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1710 !list_empty(&ci->i_cap_snaps))
1712 if ((used | wanted) & CEPH_CAP_ANY_WR)
1714 /* Note: it's possible that i_filelock_ref becomes non-zero
1715 * after dropping auth caps. It doesn't hurt because reply
1716 * of lock mds request will re-add auth caps. */
1717 if (atomic_read(&ci->i_filelock_ref) > 0)
1720 /* The inode has cached pages, but it's no longer used.
1721 * we can safely drop it */
1722 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1723 !(oissued & CEPH_CAP_FILE_CACHE)) {
1727 if ((used | wanted) & ~oissued & mine)
1728 goto out; /* we need these caps */
1731 /* we aren't the only cap.. just remove us */
1732 __ceph_remove_cap(cap, true);
1733 session->s_trim_caps--;
1735 struct dentry *dentry;
1736 /* try dropping referring dentries */
1737 spin_unlock(&ci->i_ceph_lock);
1738 dentry = d_find_any_alias(inode);
1739 if (dentry && drop_negative_children(dentry)) {
1742 d_prune_aliases(inode);
1743 count = atomic_read(&inode->i_count);
1745 session->s_trim_caps--;
1746 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1755 spin_unlock(&ci->i_ceph_lock);
1760 * Trim session cap count down to some max number.
1762 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1763 struct ceph_mds_session *session,
1766 int trim_caps = session->s_nr_caps - max_caps;
1768 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1769 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1770 if (trim_caps > 0) {
1771 session->s_trim_caps = trim_caps;
1772 ceph_iterate_session_caps(session, trim_caps_cb, session);
1773 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1774 session->s_mds, session->s_nr_caps, max_caps,
1775 trim_caps - session->s_trim_caps);
1776 session->s_trim_caps = 0;
1779 ceph_flush_cap_releases(mdsc, session);
1783 static int check_caps_flush(struct ceph_mds_client *mdsc,
1788 spin_lock(&mdsc->cap_dirty_lock);
1789 if (!list_empty(&mdsc->cap_flush_list)) {
1790 struct ceph_cap_flush *cf =
1791 list_first_entry(&mdsc->cap_flush_list,
1792 struct ceph_cap_flush, g_list);
1793 if (cf->tid <= want_flush_tid) {
1794 dout("check_caps_flush still flushing tid "
1795 "%llu <= %llu\n", cf->tid, want_flush_tid);
1799 spin_unlock(&mdsc->cap_dirty_lock);
1804 * flush all dirty inode data to disk.
1806 * returns true if we've flushed through want_flush_tid
1808 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1811 dout("check_caps_flush want %llu\n", want_flush_tid);
1813 wait_event(mdsc->cap_flushing_wq,
1814 check_caps_flush(mdsc, want_flush_tid));
1816 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1820 * called under s_mutex
1822 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1823 struct ceph_mds_session *session)
1825 struct ceph_msg *msg = NULL;
1826 struct ceph_mds_cap_release *head;
1827 struct ceph_mds_cap_item *item;
1828 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1829 struct ceph_cap *cap;
1830 LIST_HEAD(tmp_list);
1831 int num_cap_releases;
1832 __le32 barrier, *cap_barrier;
1834 down_read(&osdc->lock);
1835 barrier = cpu_to_le32(osdc->epoch_barrier);
1836 up_read(&osdc->lock);
1838 spin_lock(&session->s_cap_lock);
1840 list_splice_init(&session->s_cap_releases, &tmp_list);
1841 num_cap_releases = session->s_num_cap_releases;
1842 session->s_num_cap_releases = 0;
1843 spin_unlock(&session->s_cap_lock);
1845 while (!list_empty(&tmp_list)) {
1847 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1848 PAGE_SIZE, GFP_NOFS, false);
1851 head = msg->front.iov_base;
1852 head->num = cpu_to_le32(0);
1853 msg->front.iov_len = sizeof(*head);
1855 msg->hdr.version = cpu_to_le16(2);
1856 msg->hdr.compat_version = cpu_to_le16(1);
1859 cap = list_first_entry(&tmp_list, struct ceph_cap,
1861 list_del(&cap->session_caps);
1864 head = msg->front.iov_base;
1865 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1867 item = msg->front.iov_base + msg->front.iov_len;
1868 item->ino = cpu_to_le64(cap->cap_ino);
1869 item->cap_id = cpu_to_le64(cap->cap_id);
1870 item->migrate_seq = cpu_to_le32(cap->mseq);
1871 item->seq = cpu_to_le32(cap->issue_seq);
1872 msg->front.iov_len += sizeof(*item);
1874 ceph_put_cap(mdsc, cap);
1876 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1877 // Append cap_barrier field
1878 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1879 *cap_barrier = barrier;
1880 msg->front.iov_len += sizeof(*cap_barrier);
1882 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1883 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1884 ceph_con_send(&session->s_con, msg);
1889 BUG_ON(num_cap_releases != 0);
1891 spin_lock(&session->s_cap_lock);
1892 if (!list_empty(&session->s_cap_releases))
1894 spin_unlock(&session->s_cap_lock);
1897 // Append cap_barrier field
1898 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1899 *cap_barrier = barrier;
1900 msg->front.iov_len += sizeof(*cap_barrier);
1902 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1903 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1904 ceph_con_send(&session->s_con, msg);
1908 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1910 spin_lock(&session->s_cap_lock);
1911 list_splice(&tmp_list, &session->s_cap_releases);
1912 session->s_num_cap_releases += num_cap_releases;
1913 spin_unlock(&session->s_cap_lock);
1916 static void ceph_cap_release_work(struct work_struct *work)
1918 struct ceph_mds_session *session =
1919 container_of(work, struct ceph_mds_session, s_cap_release_work);
1921 mutex_lock(&session->s_mutex);
1922 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1923 session->s_state == CEPH_MDS_SESSION_HUNG)
1924 ceph_send_cap_releases(session->s_mdsc, session);
1925 mutex_unlock(&session->s_mutex);
1926 ceph_put_mds_session(session);
1929 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1930 struct ceph_mds_session *session)
1935 get_session(session);
1936 if (queue_work(mdsc->fsc->cap_wq,
1937 &session->s_cap_release_work)) {
1938 dout("cap release work queued\n");
1940 ceph_put_mds_session(session);
1941 dout("failed to queue cap release work\n");
1946 * caller holds session->s_cap_lock
1948 void __ceph_queue_cap_release(struct ceph_mds_session *session,
1949 struct ceph_cap *cap)
1951 list_add_tail(&cap->session_caps, &session->s_cap_releases);
1952 session->s_num_cap_releases++;
1954 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
1955 ceph_flush_cap_releases(session->s_mdsc, session);
1958 static void ceph_cap_reclaim_work(struct work_struct *work)
1960 struct ceph_mds_client *mdsc =
1961 container_of(work, struct ceph_mds_client, cap_reclaim_work);
1962 int ret = ceph_trim_dentries(mdsc);
1964 ceph_queue_cap_reclaim_work(mdsc);
1967 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
1972 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
1973 dout("caps reclaim work queued\n");
1975 dout("failed to queue caps release work\n");
1979 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
1984 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
1985 if (!(val % CEPH_CAPS_PER_RELEASE)) {
1986 atomic_set(&mdsc->cap_reclaim_pending, 0);
1987 ceph_queue_cap_reclaim_work(mdsc);
1995 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1998 struct ceph_inode_info *ci = ceph_inode(dir);
1999 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2000 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2001 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2002 int order, num_entries;
2004 spin_lock(&ci->i_ceph_lock);
2005 num_entries = ci->i_files + ci->i_subdirs;
2006 spin_unlock(&ci->i_ceph_lock);
2007 num_entries = max(num_entries, 1);
2008 num_entries = min(num_entries, opt->max_readdir);
2010 order = get_order(size * num_entries);
2011 while (order >= 0) {
2012 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2015 if (rinfo->dir_entries)
2019 if (!rinfo->dir_entries)
2022 num_entries = (PAGE_SIZE << order) / size;
2023 num_entries = min(num_entries, opt->max_readdir);
2025 rinfo->dir_buf_size = PAGE_SIZE << order;
2026 req->r_num_caps = num_entries + 1;
2027 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2028 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2033 * Create an mds request.
2035 struct ceph_mds_request *
2036 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2038 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2039 struct timespec64 ts;
2042 return ERR_PTR(-ENOMEM);
2044 mutex_init(&req->r_fill_mutex);
2046 req->r_started = jiffies;
2047 req->r_resend_mds = -1;
2048 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2049 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2051 kref_init(&req->r_kref);
2052 RB_CLEAR_NODE(&req->r_node);
2053 INIT_LIST_HEAD(&req->r_wait);
2054 init_completion(&req->r_completion);
2055 init_completion(&req->r_safe_completion);
2056 INIT_LIST_HEAD(&req->r_unsafe_item);
2058 ktime_get_coarse_real_ts64(&ts);
2059 req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
2062 req->r_direct_mode = mode;
2067 * return oldest (lowest) request, tid in request tree, 0 if none.
2069 * called under mdsc->mutex.
2071 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2073 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2075 return rb_entry(rb_first(&mdsc->request_tree),
2076 struct ceph_mds_request, r_node);
2079 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2081 return mdsc->oldest_tid;
2085 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2086 * on build_path_from_dentry in fs/cifs/dir.c.
2088 * If @stop_on_nosnap, generate path relative to the first non-snapped
2091 * Encode hidden .snap dirs as a double /, i.e.
2092 * foo/.snap/bar -> foo//bar
2094 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2097 struct dentry *temp;
2104 return ERR_PTR(-EINVAL);
2108 return ERR_PTR(-ENOMEM);
2113 seq = read_seqbegin(&rename_lock);
2117 struct inode *inode;
2119 spin_lock(&temp->d_lock);
2120 inode = d_inode(temp);
2121 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2122 dout("build_path path+%d: %p SNAPDIR\n",
2124 } else if (stop_on_nosnap && inode && dentry != temp &&
2125 ceph_snap(inode) == CEPH_NOSNAP) {
2126 spin_unlock(&temp->d_lock);
2127 pos++; /* get rid of any prepended '/' */
2130 pos -= temp->d_name.len;
2132 spin_unlock(&temp->d_lock);
2135 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2137 spin_unlock(&temp->d_lock);
2138 temp = temp->d_parent;
2140 /* Are we at the root? */
2144 /* Are we out of buffer? */
2150 base = ceph_ino(d_inode(temp));
2152 if (pos < 0 || read_seqretry(&rename_lock, seq)) {
2153 pr_err("build_path did not end path lookup where "
2154 "expected, pos is %d\n", pos);
2155 /* presumably this is only possible if racing with a
2156 rename of one of the parent directories (we can not
2157 lock the dentries above us to prevent this, but
2158 retrying should be harmless) */
2163 *plen = PATH_MAX - 1 - pos;
2164 dout("build_path on %p %d built %llx '%.*s'\n",
2165 dentry, d_count(dentry), base, *plen, path + pos);
2169 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2170 const char **ppath, int *ppathlen, u64 *pino,
2171 bool *pfreepath, bool parent_locked)
2177 dir = d_inode_rcu(dentry->d_parent);
2178 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2179 *pino = ceph_ino(dir);
2181 *ppath = dentry->d_name.name;
2182 *ppathlen = dentry->d_name.len;
2186 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2188 return PTR_ERR(path);
2194 static int build_inode_path(struct inode *inode,
2195 const char **ppath, int *ppathlen, u64 *pino,
2198 struct dentry *dentry;
2201 if (ceph_snap(inode) == CEPH_NOSNAP) {
2202 *pino = ceph_ino(inode);
2206 dentry = d_find_alias(inode);
2207 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2210 return PTR_ERR(path);
2217 * request arguments may be specified via an inode *, a dentry *, or
2218 * an explicit ino+path.
2220 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2221 struct inode *rdiri, const char *rpath,
2222 u64 rino, const char **ppath, int *pathlen,
2223 u64 *ino, bool *freepath, bool parent_locked)
2228 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2229 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2231 } else if (rdentry) {
2232 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2233 freepath, parent_locked);
2234 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2236 } else if (rpath || rino) {
2239 *pathlen = rpath ? strlen(rpath) : 0;
2240 dout(" path %.*s\n", *pathlen, rpath);
2247 * called under mdsc->mutex
2249 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2250 struct ceph_mds_request *req,
2251 int mds, bool drop_cap_releases)
2253 struct ceph_msg *msg;
2254 struct ceph_mds_request_head *head;
2255 const char *path1 = NULL;
2256 const char *path2 = NULL;
2257 u64 ino1 = 0, ino2 = 0;
2258 int pathlen1 = 0, pathlen2 = 0;
2259 bool freepath1 = false, freepath2 = false;
2265 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2266 req->r_parent, req->r_path1, req->r_ino1.ino,
2267 &path1, &pathlen1, &ino1, &freepath1,
2268 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2269 &req->r_req_flags));
2275 /* If r_old_dentry is set, then assume that its parent is locked */
2276 ret = set_request_path_attr(NULL, req->r_old_dentry,
2277 req->r_old_dentry_dir,
2278 req->r_path2, req->r_ino2.ino,
2279 &path2, &pathlen2, &ino2, &freepath2, true);
2285 len = sizeof(*head) +
2286 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2287 sizeof(struct ceph_timespec);
2289 /* calculate (max) length for cap releases */
2290 len += sizeof(struct ceph_mds_request_release) *
2291 (!!req->r_inode_drop + !!req->r_dentry_drop +
2292 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2293 if (req->r_dentry_drop)
2295 if (req->r_old_dentry_drop)
2298 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2300 msg = ERR_PTR(-ENOMEM);
2304 msg->hdr.version = cpu_to_le16(2);
2305 msg->hdr.tid = cpu_to_le64(req->r_tid);
2307 head = msg->front.iov_base;
2308 p = msg->front.iov_base + sizeof(*head);
2309 end = msg->front.iov_base + msg->front.iov_len;
2311 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2312 head->op = cpu_to_le32(req->r_op);
2313 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2314 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2315 head->args = req->r_args;
2317 ceph_encode_filepath(&p, end, ino1, path1);
2318 ceph_encode_filepath(&p, end, ino2, path2);
2320 /* make note of release offset, in case we need to replay */
2321 req->r_request_release_offset = p - msg->front.iov_base;
2325 if (req->r_inode_drop)
2326 releases += ceph_encode_inode_release(&p,
2327 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2328 mds, req->r_inode_drop, req->r_inode_unless, 0);
2329 if (req->r_dentry_drop)
2330 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2331 req->r_parent, mds, req->r_dentry_drop,
2332 req->r_dentry_unless);
2333 if (req->r_old_dentry_drop)
2334 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2335 req->r_old_dentry_dir, mds,
2336 req->r_old_dentry_drop,
2337 req->r_old_dentry_unless);
2338 if (req->r_old_inode_drop)
2339 releases += ceph_encode_inode_release(&p,
2340 d_inode(req->r_old_dentry),
2341 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2343 if (drop_cap_releases) {
2345 p = msg->front.iov_base + req->r_request_release_offset;
2348 head->num_releases = cpu_to_le16(releases);
2352 struct ceph_timespec ts;
2353 ceph_encode_timespec64(&ts, &req->r_stamp);
2354 ceph_encode_copy(&p, &ts, sizeof(ts));
2358 msg->front.iov_len = p - msg->front.iov_base;
2359 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2361 if (req->r_pagelist) {
2362 struct ceph_pagelist *pagelist = req->r_pagelist;
2363 ceph_msg_data_add_pagelist(msg, pagelist);
2364 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2366 msg->hdr.data_len = 0;
2369 msg->hdr.data_off = cpu_to_le16(0);
2373 ceph_mdsc_free_path((char *)path2, pathlen2);
2376 ceph_mdsc_free_path((char *)path1, pathlen1);
2382 * called under mdsc->mutex if error, under no mutex if
2385 static void complete_request(struct ceph_mds_client *mdsc,
2386 struct ceph_mds_request *req)
2388 if (req->r_callback)
2389 req->r_callback(mdsc, req);
2390 complete_all(&req->r_completion);
2394 * called under mdsc->mutex
2396 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2397 struct ceph_mds_request *req,
2398 int mds, bool drop_cap_releases)
2400 struct ceph_mds_request_head *rhead;
2401 struct ceph_msg *msg;
2406 struct ceph_cap *cap =
2407 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2410 req->r_sent_on_mseq = cap->mseq;
2412 req->r_sent_on_mseq = -1;
2414 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2415 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2417 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2420 * Replay. Do not regenerate message (and rebuild
2421 * paths, etc.); just use the original message.
2422 * Rebuilding paths will break for renames because
2423 * d_move mangles the src name.
2425 msg = req->r_request;
2426 rhead = msg->front.iov_base;
2428 flags = le32_to_cpu(rhead->flags);
2429 flags |= CEPH_MDS_FLAG_REPLAY;
2430 rhead->flags = cpu_to_le32(flags);
2432 if (req->r_target_inode)
2433 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2435 rhead->num_retry = req->r_attempts - 1;
2437 /* remove cap/dentry releases from message */
2438 rhead->num_releases = 0;
2441 p = msg->front.iov_base + req->r_request_release_offset;
2443 struct ceph_timespec ts;
2444 ceph_encode_timespec64(&ts, &req->r_stamp);
2445 ceph_encode_copy(&p, &ts, sizeof(ts));
2448 msg->front.iov_len = p - msg->front.iov_base;
2449 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2453 if (req->r_request) {
2454 ceph_msg_put(req->r_request);
2455 req->r_request = NULL;
2457 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2459 req->r_err = PTR_ERR(msg);
2460 return PTR_ERR(msg);
2462 req->r_request = msg;
2464 rhead = msg->front.iov_base;
2465 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2466 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2467 flags |= CEPH_MDS_FLAG_REPLAY;
2469 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2470 rhead->flags = cpu_to_le32(flags);
2471 rhead->num_fwd = req->r_num_fwd;
2472 rhead->num_retry = req->r_attempts - 1;
2475 dout(" r_parent = %p\n", req->r_parent);
2480 * send request, or put it on the appropriate wait list.
2482 static void __do_request(struct ceph_mds_client *mdsc,
2483 struct ceph_mds_request *req)
2485 struct ceph_mds_session *session = NULL;
2489 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2490 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2491 __unregister_request(mdsc, req);
2495 if (req->r_timeout &&
2496 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2497 dout("do_request timed out\n");
2501 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2502 dout("do_request forced umount\n");
2506 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2507 if (mdsc->mdsmap_err) {
2508 err = mdsc->mdsmap_err;
2509 dout("do_request mdsmap err %d\n", err);
2512 if (mdsc->mdsmap->m_epoch == 0) {
2513 dout("do_request no mdsmap, waiting for map\n");
2514 list_add(&req->r_wait, &mdsc->waiting_for_map);
2517 if (!(mdsc->fsc->mount_options->flags &
2518 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2519 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2521 pr_info("probably no mds server is up\n");
2526 put_request_session(req);
2528 mds = __choose_mds(mdsc, req);
2530 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2531 dout("do_request no mds or not active, waiting for map\n");
2532 list_add(&req->r_wait, &mdsc->waiting_for_map);
2536 /* get, open session */
2537 session = __ceph_lookup_mds_session(mdsc, mds);
2539 session = register_session(mdsc, mds);
2540 if (IS_ERR(session)) {
2541 err = PTR_ERR(session);
2545 req->r_session = get_session(session);
2547 dout("do_request mds%d session %p state %s\n", mds, session,
2548 ceph_session_state_name(session->s_state));
2549 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2550 session->s_state != CEPH_MDS_SESSION_HUNG) {
2551 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2555 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2556 session->s_state == CEPH_MDS_SESSION_CLOSING)
2557 __open_session(mdsc, session);
2558 list_add(&req->r_wait, &session->s_waiting);
2563 req->r_resend_mds = -1; /* forget any previous mds hint */
2565 if (req->r_request_started == 0) /* note request start time */
2566 req->r_request_started = jiffies;
2568 err = __prepare_send_request(mdsc, req, mds, false);
2570 ceph_msg_get(req->r_request);
2571 ceph_con_send(&session->s_con, req->r_request);
2575 ceph_put_mds_session(session);
2578 dout("__do_request early error %d\n", err);
2580 complete_request(mdsc, req);
2581 __unregister_request(mdsc, req);
2587 * called under mdsc->mutex
2589 static void __wake_requests(struct ceph_mds_client *mdsc,
2590 struct list_head *head)
2592 struct ceph_mds_request *req;
2593 LIST_HEAD(tmp_list);
2595 list_splice_init(head, &tmp_list);
2597 while (!list_empty(&tmp_list)) {
2598 req = list_entry(tmp_list.next,
2599 struct ceph_mds_request, r_wait);
2600 list_del_init(&req->r_wait);
2601 dout(" wake request %p tid %llu\n", req, req->r_tid);
2602 __do_request(mdsc, req);
2607 * Wake up threads with requests pending for @mds, so that they can
2608 * resubmit their requests to a possibly different mds.
2610 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2612 struct ceph_mds_request *req;
2613 struct rb_node *p = rb_first(&mdsc->request_tree);
2615 dout("kick_requests mds%d\n", mds);
2617 req = rb_entry(p, struct ceph_mds_request, r_node);
2619 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2621 if (req->r_attempts > 0)
2622 continue; /* only new requests */
2623 if (req->r_session &&
2624 req->r_session->s_mds == mds) {
2625 dout(" kicking tid %llu\n", req->r_tid);
2626 list_del_init(&req->r_wait);
2627 __do_request(mdsc, req);
2632 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2633 struct ceph_mds_request *req)
2637 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2639 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2641 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2642 if (req->r_old_dentry_dir)
2643 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2646 dout("submit_request on %p for inode %p\n", req, dir);
2647 mutex_lock(&mdsc->mutex);
2648 __register_request(mdsc, req, dir);
2649 __do_request(mdsc, req);
2651 mutex_unlock(&mdsc->mutex);
2655 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2656 struct ceph_mds_request *req)
2661 dout("do_request waiting\n");
2662 if (!req->r_timeout && req->r_wait_for_completion) {
2663 err = req->r_wait_for_completion(mdsc, req);
2665 long timeleft = wait_for_completion_killable_timeout(
2667 ceph_timeout_jiffies(req->r_timeout));
2671 err = -EIO; /* timed out */
2673 err = timeleft; /* killed */
2675 dout("do_request waited, got %d\n", err);
2676 mutex_lock(&mdsc->mutex);
2678 /* only abort if we didn't race with a real reply */
2679 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2680 err = le32_to_cpu(req->r_reply_info.head->result);
2681 } else if (err < 0) {
2682 dout("aborted request %lld with %d\n", req->r_tid, err);
2685 * ensure we aren't running concurrently with
2686 * ceph_fill_trace or ceph_readdir_prepopulate, which
2687 * rely on locks (dir mutex) held by our caller.
2689 mutex_lock(&req->r_fill_mutex);
2691 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2692 mutex_unlock(&req->r_fill_mutex);
2694 if (req->r_parent &&
2695 (req->r_op & CEPH_MDS_OP_WRITE))
2696 ceph_invalidate_dir_request(req);
2701 mutex_unlock(&mdsc->mutex);
2706 * Synchrously perform an mds request. Take care of all of the
2707 * session setup, forwarding, retry details.
2709 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2711 struct ceph_mds_request *req)
2715 dout("do_request on %p\n", req);
2718 err = ceph_mdsc_submit_request(mdsc, dir, req);
2720 err = ceph_mdsc_wait_request(mdsc, req);
2721 dout("do_request %p done, result %d\n", req, err);
2726 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2727 * namespace request.
2729 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2731 struct inode *dir = req->r_parent;
2732 struct inode *old_dir = req->r_old_dentry_dir;
2734 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2736 ceph_dir_clear_complete(dir);
2738 ceph_dir_clear_complete(old_dir);
2740 ceph_invalidate_dentry_lease(req->r_dentry);
2741 if (req->r_old_dentry)
2742 ceph_invalidate_dentry_lease(req->r_old_dentry);
2748 * We take the session mutex and parse and process the reply immediately.
2749 * This preserves the logical ordering of replies, capabilities, etc., sent
2750 * by the MDS as they are applied to our local cache.
2752 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2754 struct ceph_mds_client *mdsc = session->s_mdsc;
2755 struct ceph_mds_request *req;
2756 struct ceph_mds_reply_head *head = msg->front.iov_base;
2757 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2758 struct ceph_snap_realm *realm;
2761 int mds = session->s_mds;
2763 if (msg->front.iov_len < sizeof(*head)) {
2764 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2769 /* get request, session */
2770 tid = le64_to_cpu(msg->hdr.tid);
2771 mutex_lock(&mdsc->mutex);
2772 req = lookup_get_request(mdsc, tid);
2774 dout("handle_reply on unknown tid %llu\n", tid);
2775 mutex_unlock(&mdsc->mutex);
2778 dout("handle_reply %p\n", req);
2780 /* correct session? */
2781 if (req->r_session != session) {
2782 pr_err("mdsc_handle_reply got %llu on session mds%d"
2783 " not mds%d\n", tid, session->s_mds,
2784 req->r_session ? req->r_session->s_mds : -1);
2785 mutex_unlock(&mdsc->mutex);
2790 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2791 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2792 pr_warn("got a dup %s reply on %llu from mds%d\n",
2793 head->safe ? "safe" : "unsafe", tid, mds);
2794 mutex_unlock(&mdsc->mutex);
2797 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2798 pr_warn("got unsafe after safe on %llu from mds%d\n",
2800 mutex_unlock(&mdsc->mutex);
2804 result = le32_to_cpu(head->result);
2808 * if we're not talking to the authority, send to them
2809 * if the authority has changed while we weren't looking,
2810 * send to new authority
2811 * Otherwise we just have to return an ESTALE
2813 if (result == -ESTALE) {
2814 dout("got ESTALE on request %llu\n", req->r_tid);
2815 req->r_resend_mds = -1;
2816 if (req->r_direct_mode != USE_AUTH_MDS) {
2817 dout("not using auth, setting for that now\n");
2818 req->r_direct_mode = USE_AUTH_MDS;
2819 __do_request(mdsc, req);
2820 mutex_unlock(&mdsc->mutex);
2823 int mds = __choose_mds(mdsc, req);
2824 if (mds >= 0 && mds != req->r_session->s_mds) {
2825 dout("but auth changed, so resending\n");
2826 __do_request(mdsc, req);
2827 mutex_unlock(&mdsc->mutex);
2831 dout("have to return ESTALE on request %llu\n", req->r_tid);
2836 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2837 __unregister_request(mdsc, req);
2839 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2841 * We already handled the unsafe response, now do the
2842 * cleanup. No need to examine the response; the MDS
2843 * doesn't include any result info in the safe
2844 * response. And even if it did, there is nothing
2845 * useful we could do with a revised return value.
2847 dout("got safe reply %llu, mds%d\n", tid, mds);
2849 /* last unsafe request during umount? */
2850 if (mdsc->stopping && !__get_oldest_req(mdsc))
2851 complete_all(&mdsc->safe_umount_waiters);
2852 mutex_unlock(&mdsc->mutex);
2856 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2857 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2858 if (req->r_unsafe_dir) {
2859 struct ceph_inode_info *ci =
2860 ceph_inode(req->r_unsafe_dir);
2861 spin_lock(&ci->i_unsafe_lock);
2862 list_add_tail(&req->r_unsafe_dir_item,
2863 &ci->i_unsafe_dirops);
2864 spin_unlock(&ci->i_unsafe_lock);
2868 dout("handle_reply tid %lld result %d\n", tid, result);
2869 rinfo = &req->r_reply_info;
2870 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2871 err = parse_reply_info(msg, rinfo, (u64)-1);
2873 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2874 mutex_unlock(&mdsc->mutex);
2876 mutex_lock(&session->s_mutex);
2878 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2885 if (rinfo->snapblob_len) {
2886 down_write(&mdsc->snap_rwsem);
2887 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2888 rinfo->snapblob + rinfo->snapblob_len,
2889 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2891 downgrade_write(&mdsc->snap_rwsem);
2893 down_read(&mdsc->snap_rwsem);
2896 /* insert trace into our cache */
2897 mutex_lock(&req->r_fill_mutex);
2898 current->journal_info = req;
2899 err = ceph_fill_trace(mdsc->fsc->sb, req);
2901 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2902 req->r_op == CEPH_MDS_OP_LSSNAP))
2903 ceph_readdir_prepopulate(req, req->r_session);
2905 current->journal_info = NULL;
2906 mutex_unlock(&req->r_fill_mutex);
2908 up_read(&mdsc->snap_rwsem);
2910 ceph_put_snap_realm(mdsc, realm);
2913 if (req->r_target_inode &&
2914 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2915 struct ceph_inode_info *ci =
2916 ceph_inode(req->r_target_inode);
2917 spin_lock(&ci->i_unsafe_lock);
2918 list_add_tail(&req->r_unsafe_target_item,
2919 &ci->i_unsafe_iops);
2920 spin_unlock(&ci->i_unsafe_lock);
2923 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2926 mutex_lock(&mdsc->mutex);
2927 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2931 req->r_reply = ceph_msg_get(msg);
2932 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2935 dout("reply arrived after request %lld was aborted\n", tid);
2937 mutex_unlock(&mdsc->mutex);
2939 mutex_unlock(&session->s_mutex);
2941 /* kick calling process */
2942 complete_request(mdsc, req);
2944 ceph_mdsc_put_request(req);
2951 * handle mds notification that our request has been forwarded.
2953 static void handle_forward(struct ceph_mds_client *mdsc,
2954 struct ceph_mds_session *session,
2955 struct ceph_msg *msg)
2957 struct ceph_mds_request *req;
2958 u64 tid = le64_to_cpu(msg->hdr.tid);
2962 void *p = msg->front.iov_base;
2963 void *end = p + msg->front.iov_len;
2965 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2966 next_mds = ceph_decode_32(&p);
2967 fwd_seq = ceph_decode_32(&p);
2969 mutex_lock(&mdsc->mutex);
2970 req = lookup_get_request(mdsc, tid);
2972 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2973 goto out; /* dup reply? */
2976 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2977 dout("forward tid %llu aborted, unregistering\n", tid);
2978 __unregister_request(mdsc, req);
2979 } else if (fwd_seq <= req->r_num_fwd) {
2980 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2981 tid, next_mds, req->r_num_fwd, fwd_seq);
2983 /* resend. forward race not possible; mds would drop */
2984 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2986 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2987 req->r_attempts = 0;
2988 req->r_num_fwd = fwd_seq;
2989 req->r_resend_mds = next_mds;
2990 put_request_session(req);
2991 __do_request(mdsc, req);
2993 ceph_mdsc_put_request(req);
2995 mutex_unlock(&mdsc->mutex);
2999 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3002 static int __decode_and_drop_session_metadata(void **p, void *end)
3004 /* map<string,string> */
3006 ceph_decode_32_safe(p, end, n, bad);
3009 ceph_decode_32_safe(p, end, len, bad);
3010 ceph_decode_need(p, end, len, bad);
3012 ceph_decode_32_safe(p, end, len, bad);
3013 ceph_decode_need(p, end, len, bad);
3022 * handle a mds session control message
3024 static void handle_session(struct ceph_mds_session *session,
3025 struct ceph_msg *msg)
3027 struct ceph_mds_client *mdsc = session->s_mdsc;
3028 int mds = session->s_mds;
3029 int msg_version = le16_to_cpu(msg->hdr.version);
3030 void *p = msg->front.iov_base;
3031 void *end = p + msg->front.iov_len;
3032 struct ceph_mds_session_head *h;
3035 unsigned long features = 0;
3039 ceph_decode_need(&p, end, sizeof(*h), bad);
3043 op = le32_to_cpu(h->op);
3044 seq = le64_to_cpu(h->seq);
3046 if (msg_version >= 3) {
3048 /* version >= 2, metadata */
3049 if (__decode_and_drop_session_metadata(&p, end) < 0)
3051 /* version >= 3, feature bits */
3052 ceph_decode_32_safe(&p, end, len, bad);
3053 ceph_decode_need(&p, end, len, bad);
3054 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3058 mutex_lock(&mdsc->mutex);
3059 if (op == CEPH_SESSION_CLOSE) {
3060 get_session(session);
3061 __unregister_session(mdsc, session);
3063 /* FIXME: this ttl calculation is generous */
3064 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3065 mutex_unlock(&mdsc->mutex);
3067 mutex_lock(&session->s_mutex);
3069 dout("handle_session mds%d %s %p state %s seq %llu\n",
3070 mds, ceph_session_op_name(op), session,
3071 ceph_session_state_name(session->s_state), seq);
3073 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3074 session->s_state = CEPH_MDS_SESSION_OPEN;
3075 pr_info("mds%d came back\n", session->s_mds);
3079 case CEPH_SESSION_OPEN:
3080 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3081 pr_info("mds%d reconnect success\n", session->s_mds);
3082 session->s_state = CEPH_MDS_SESSION_OPEN;
3083 session->s_features = features;
3084 renewed_caps(mdsc, session, 0);
3087 __close_session(mdsc, session);
3090 case CEPH_SESSION_RENEWCAPS:
3091 if (session->s_renew_seq == seq)
3092 renewed_caps(mdsc, session, 1);
3095 case CEPH_SESSION_CLOSE:
3096 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3097 pr_info("mds%d reconnect denied\n", session->s_mds);
3098 cleanup_session_requests(mdsc, session);
3099 remove_session_caps(session);
3100 wake = 2; /* for good measure */
3101 wake_up_all(&mdsc->session_close_wq);
3104 case CEPH_SESSION_STALE:
3105 pr_info("mds%d caps went stale, renewing\n",
3107 spin_lock(&session->s_gen_ttl_lock);
3108 session->s_cap_gen++;
3109 session->s_cap_ttl = jiffies - 1;
3110 spin_unlock(&session->s_gen_ttl_lock);
3111 send_renew_caps(mdsc, session);
3114 case CEPH_SESSION_RECALL_STATE:
3115 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3118 case CEPH_SESSION_FLUSHMSG:
3119 send_flushmsg_ack(mdsc, session, seq);
3122 case CEPH_SESSION_FORCE_RO:
3123 dout("force_session_readonly %p\n", session);
3124 spin_lock(&session->s_cap_lock);
3125 session->s_readonly = true;
3126 spin_unlock(&session->s_cap_lock);
3127 wake_up_session_caps(session, FORCE_RO);
3130 case CEPH_SESSION_REJECT:
3131 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3132 pr_info("mds%d rejected session\n", session->s_mds);
3133 session->s_state = CEPH_MDS_SESSION_REJECTED;
3134 cleanup_session_requests(mdsc, session);
3135 remove_session_caps(session);
3136 wake = 2; /* for good measure */
3140 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3144 mutex_unlock(&session->s_mutex);
3146 mutex_lock(&mdsc->mutex);
3147 __wake_requests(mdsc, &session->s_waiting);
3149 kick_requests(mdsc, mds);
3150 mutex_unlock(&mdsc->mutex);
3152 if (op == CEPH_SESSION_CLOSE)
3153 ceph_put_mds_session(session);
3157 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3158 (int)msg->front.iov_len);
3165 * called under session->mutex.
3167 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3168 struct ceph_mds_session *session)
3170 struct ceph_mds_request *req, *nreq;
3174 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3176 mutex_lock(&mdsc->mutex);
3177 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
3178 err = __prepare_send_request(mdsc, req, session->s_mds, true);
3180 ceph_msg_get(req->r_request);
3181 ceph_con_send(&session->s_con, req->r_request);
3186 * also re-send old requests when MDS enters reconnect stage. So that MDS
3187 * can process completed request in clientreplay stage.
3189 p = rb_first(&mdsc->request_tree);
3191 req = rb_entry(p, struct ceph_mds_request, r_node);
3193 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3195 if (req->r_attempts == 0)
3196 continue; /* only old requests */
3197 if (req->r_session &&
3198 req->r_session->s_mds == session->s_mds) {
3199 err = __prepare_send_request(mdsc, req,
3200 session->s_mds, true);
3202 ceph_msg_get(req->r_request);
3203 ceph_con_send(&session->s_con, req->r_request);
3207 mutex_unlock(&mdsc->mutex);
3210 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3212 struct ceph_msg *reply;
3213 struct ceph_pagelist *_pagelist;
3218 if (!recon_state->allow_multi)
3221 /* can't handle message that contains both caps and realm */
3222 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3224 /* pre-allocate new pagelist */
3225 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3229 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3233 /* placeholder for nr_caps */
3234 err = ceph_pagelist_encode_32(_pagelist, 0);
3238 if (recon_state->nr_caps) {
3239 /* currently encoding caps */
3240 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3244 /* placeholder for nr_realms (currently encoding relams) */
3245 err = ceph_pagelist_encode_32(_pagelist, 0);
3250 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3254 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3255 addr = kmap_atomic(page);
3256 if (recon_state->nr_caps) {
3257 /* currently encoding caps */
3258 *addr = cpu_to_le32(recon_state->nr_caps);
3260 /* currently encoding relams */
3261 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3263 kunmap_atomic(addr);
3265 reply->hdr.version = cpu_to_le16(5);
3266 reply->hdr.compat_version = cpu_to_le16(4);
3268 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3269 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3271 ceph_con_send(&recon_state->session->s_con, reply);
3272 ceph_pagelist_release(recon_state->pagelist);
3274 recon_state->pagelist = _pagelist;
3275 recon_state->nr_caps = 0;
3276 recon_state->nr_realms = 0;
3277 recon_state->msg_version = 5;
3280 ceph_msg_put(reply);
3282 ceph_pagelist_release(_pagelist);
3287 * Encode information about a cap for a reconnect with the MDS.
3289 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3293 struct ceph_mds_cap_reconnect v2;
3294 struct ceph_mds_cap_reconnect_v1 v1;
3296 struct ceph_inode_info *ci = cap->ci;
3297 struct ceph_reconnect_state *recon_state = arg;
3298 struct ceph_pagelist *pagelist = recon_state->pagelist;
3302 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3303 inode, ceph_vinop(inode), cap, cap->cap_id,
3304 ceph_cap_string(cap->issued));
3306 spin_lock(&ci->i_ceph_lock);
3307 cap->seq = 0; /* reset cap seq */
3308 cap->issue_seq = 0; /* and issue_seq */
3309 cap->mseq = 0; /* and migrate_seq */
3310 cap->cap_gen = cap->session->s_cap_gen;
3312 if (recon_state->msg_version >= 2) {
3313 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3314 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3315 rec.v2.issued = cpu_to_le32(cap->issued);
3316 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3317 rec.v2.pathbase = 0;
3318 rec.v2.flock_len = (__force __le32)
3319 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3321 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3322 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3323 rec.v1.issued = cpu_to_le32(cap->issued);
3324 rec.v1.size = cpu_to_le64(inode->i_size);
3325 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3326 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3327 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3328 rec.v1.pathbase = 0;
3331 if (list_empty(&ci->i_cap_snaps)) {
3332 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3334 struct ceph_cap_snap *capsnap =
3335 list_first_entry(&ci->i_cap_snaps,
3336 struct ceph_cap_snap, ci_item);
3337 snap_follows = capsnap->follows;
3339 spin_unlock(&ci->i_ceph_lock);
3341 if (recon_state->msg_version >= 2) {
3342 int num_fcntl_locks, num_flock_locks;
3343 struct ceph_filelock *flocks = NULL;
3344 size_t struct_len, total_len = sizeof(u64);
3348 if (rec.v2.flock_len) {
3349 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3351 num_fcntl_locks = 0;
3352 num_flock_locks = 0;
3354 if (num_fcntl_locks + num_flock_locks > 0) {
3355 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3356 sizeof(struct ceph_filelock),
3362 err = ceph_encode_locks_to_buffer(inode, flocks,
3377 if (recon_state->msg_version >= 3) {
3378 /* version, compat_version and struct_len */
3379 total_len += 2 * sizeof(u8) + sizeof(u32);
3383 * number of encoded locks is stable, so copy to pagelist
3385 struct_len = 2 * sizeof(u32) +
3386 (num_fcntl_locks + num_flock_locks) *
3387 sizeof(struct ceph_filelock);
3388 rec.v2.flock_len = cpu_to_le32(struct_len);
3390 struct_len += sizeof(u32) + sizeof(rec.v2);
3393 struct_len += sizeof(u64); /* snap_follows */
3395 total_len += struct_len;
3397 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3398 err = send_reconnect_partial(recon_state);
3400 goto out_freeflocks;
3401 pagelist = recon_state->pagelist;
3404 err = ceph_pagelist_reserve(pagelist, total_len);
3406 goto out_freeflocks;
3408 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3409 if (recon_state->msg_version >= 3) {
3410 ceph_pagelist_encode_8(pagelist, struct_v);
3411 ceph_pagelist_encode_8(pagelist, 1);
3412 ceph_pagelist_encode_32(pagelist, struct_len);
3414 ceph_pagelist_encode_string(pagelist, NULL, 0);
3415 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3416 ceph_locks_to_pagelist(flocks, pagelist,
3417 num_fcntl_locks, num_flock_locks);
3419 ceph_pagelist_encode_64(pagelist, snap_follows);
3426 struct dentry *dentry;
3428 dentry = d_find_alias(inode);
3430 path = ceph_mdsc_build_path(dentry,
3431 &pathlen, &pathbase, 0);
3434 err = PTR_ERR(path);
3437 rec.v1.pathbase = cpu_to_le64(pathbase);
3440 err = ceph_pagelist_reserve(pagelist,
3441 sizeof(u64) + sizeof(u32) +
3442 pathlen + sizeof(rec.v1));
3447 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3448 ceph_pagelist_encode_string(pagelist, path, pathlen);
3449 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3451 ceph_mdsc_free_path(path, pathlen);
3456 recon_state->nr_caps++;
3460 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3461 struct ceph_reconnect_state *recon_state)
3464 struct ceph_pagelist *pagelist = recon_state->pagelist;
3467 if (recon_state->msg_version >= 4) {
3468 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3474 * snaprealms. we provide mds with the ino, seq (version), and
3475 * parent for all of our realms. If the mds has any newer info,
3478 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3479 struct ceph_snap_realm *realm =
3480 rb_entry(p, struct ceph_snap_realm, node);
3481 struct ceph_mds_snaprealm_reconnect sr_rec;
3483 if (recon_state->msg_version >= 4) {
3484 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3487 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3488 err = send_reconnect_partial(recon_state);
3491 pagelist = recon_state->pagelist;
3494 err = ceph_pagelist_reserve(pagelist, need);
3498 ceph_pagelist_encode_8(pagelist, 1);
3499 ceph_pagelist_encode_8(pagelist, 1);
3500 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3503 dout(" adding snap realm %llx seq %lld parent %llx\n",
3504 realm->ino, realm->seq, realm->parent_ino);
3505 sr_rec.ino = cpu_to_le64(realm->ino);
3506 sr_rec.seq = cpu_to_le64(realm->seq);
3507 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3509 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3513 recon_state->nr_realms++;
3521 * If an MDS fails and recovers, clients need to reconnect in order to
3522 * reestablish shared state. This includes all caps issued through
3523 * this session _and_ the snap_realm hierarchy. Because it's not
3524 * clear which snap realms the mds cares about, we send everything we
3525 * know about.. that ensures we'll then get any new info the
3526 * recovering MDS might have.
3528 * This is a relatively heavyweight operation, but it's rare.
3530 * called with mdsc->mutex held.
3532 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3533 struct ceph_mds_session *session)
3535 struct ceph_msg *reply;
3536 int mds = session->s_mds;
3538 struct ceph_reconnect_state recon_state = {
3543 pr_info("mds%d reconnect start\n", mds);
3545 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3546 if (!recon_state.pagelist)
3547 goto fail_nopagelist;
3549 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3553 mutex_lock(&session->s_mutex);
3554 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3557 dout("session %p state %s\n", session,
3558 ceph_session_state_name(session->s_state));
3560 spin_lock(&session->s_gen_ttl_lock);
3561 session->s_cap_gen++;
3562 spin_unlock(&session->s_gen_ttl_lock);
3564 spin_lock(&session->s_cap_lock);
3565 /* don't know if session is readonly */
3566 session->s_readonly = 0;
3568 * notify __ceph_remove_cap() that we are composing cap reconnect.
3569 * If a cap get released before being added to the cap reconnect,
3570 * __ceph_remove_cap() should skip queuing cap release.
3572 session->s_cap_reconnect = 1;
3573 /* drop old cap expires; we're about to reestablish that state */
3574 detach_cap_releases(session, &dispose);
3575 spin_unlock(&session->s_cap_lock);
3576 dispose_cap_releases(mdsc, &dispose);
3578 /* trim unused caps to reduce MDS's cache rejoin time */
3579 if (mdsc->fsc->sb->s_root)
3580 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3582 ceph_con_close(&session->s_con);
3583 ceph_con_open(&session->s_con,
3584 CEPH_ENTITY_TYPE_MDS, mds,
3585 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3587 /* replay unsafe requests */
3588 replay_unsafe_requests(mdsc, session);
3590 ceph_early_kick_flushing_caps(mdsc, session);
3592 down_read(&mdsc->snap_rwsem);
3594 /* placeholder for nr_caps */
3595 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3599 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3600 recon_state.msg_version = 3;
3601 recon_state.allow_multi = true;
3602 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3603 recon_state.msg_version = 3;
3605 recon_state.msg_version = 2;
3607 /* trsaverse this session's caps */
3608 err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3610 spin_lock(&session->s_cap_lock);
3611 session->s_cap_reconnect = 0;
3612 spin_unlock(&session->s_cap_lock);
3617 /* check if all realms can be encoded into current message */
3618 if (mdsc->num_snap_realms) {
3620 recon_state.pagelist->length +
3621 mdsc->num_snap_realms *
3622 sizeof(struct ceph_mds_snaprealm_reconnect);
3623 if (recon_state.msg_version >= 4) {
3624 /* number of realms */
3625 total_len += sizeof(u32);
3626 /* version, compat_version and struct_len */
3627 total_len += mdsc->num_snap_realms *
3628 (2 * sizeof(u8) + sizeof(u32));
3630 if (total_len > RECONNECT_MAX_SIZE) {
3631 if (!recon_state.allow_multi) {
3635 if (recon_state.nr_caps) {
3636 err = send_reconnect_partial(&recon_state);
3640 recon_state.msg_version = 5;
3644 err = encode_snap_realms(mdsc, &recon_state);
3648 if (recon_state.msg_version >= 5) {
3649 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3654 if (recon_state.nr_caps || recon_state.nr_realms) {
3656 list_first_entry(&recon_state.pagelist->head,
3658 __le32 *addr = kmap_atomic(page);
3659 if (recon_state.nr_caps) {
3660 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3661 *addr = cpu_to_le32(recon_state.nr_caps);
3662 } else if (recon_state.msg_version >= 4) {
3663 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3665 kunmap_atomic(addr);
3668 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3669 if (recon_state.msg_version >= 4)
3670 reply->hdr.compat_version = cpu_to_le16(4);
3672 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3673 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3675 ceph_con_send(&session->s_con, reply);
3677 mutex_unlock(&session->s_mutex);
3679 mutex_lock(&mdsc->mutex);
3680 __wake_requests(mdsc, &session->s_waiting);
3681 mutex_unlock(&mdsc->mutex);
3683 up_read(&mdsc->snap_rwsem);
3684 ceph_pagelist_release(recon_state.pagelist);
3688 ceph_msg_put(reply);
3689 up_read(&mdsc->snap_rwsem);
3690 mutex_unlock(&session->s_mutex);
3692 ceph_pagelist_release(recon_state.pagelist);
3694 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3700 * compare old and new mdsmaps, kicking requests
3701 * and closing out old connections as necessary
3703 * called under mdsc->mutex.
3705 static void check_new_map(struct ceph_mds_client *mdsc,
3706 struct ceph_mdsmap *newmap,
3707 struct ceph_mdsmap *oldmap)
3710 int oldstate, newstate;
3711 struct ceph_mds_session *s;
3713 dout("check_new_map new %u old %u\n",
3714 newmap->m_epoch, oldmap->m_epoch);
3716 for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3717 if (!mdsc->sessions[i])
3719 s = mdsc->sessions[i];
3720 oldstate = ceph_mdsmap_get_state(oldmap, i);
3721 newstate = ceph_mdsmap_get_state(newmap, i);
3723 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3724 i, ceph_mds_state_name(oldstate),
3725 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3726 ceph_mds_state_name(newstate),
3727 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3728 ceph_session_state_name(s->s_state));
3730 if (i >= newmap->m_num_mds ||
3731 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3732 ceph_mdsmap_get_addr(newmap, i),
3733 sizeof(struct ceph_entity_addr))) {
3734 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3735 /* the session never opened, just close it
3738 __unregister_session(mdsc, s);
3739 __wake_requests(mdsc, &s->s_waiting);
3740 ceph_put_mds_session(s);
3741 } else if (i >= newmap->m_num_mds) {
3742 /* force close session for stopped mds */
3744 __unregister_session(mdsc, s);
3745 __wake_requests(mdsc, &s->s_waiting);
3746 kick_requests(mdsc, i);
3747 mutex_unlock(&mdsc->mutex);
3749 mutex_lock(&s->s_mutex);
3750 cleanup_session_requests(mdsc, s);
3751 remove_session_caps(s);
3752 mutex_unlock(&s->s_mutex);
3754 ceph_put_mds_session(s);
3756 mutex_lock(&mdsc->mutex);
3759 mutex_unlock(&mdsc->mutex);
3760 mutex_lock(&s->s_mutex);
3761 mutex_lock(&mdsc->mutex);
3762 ceph_con_close(&s->s_con);
3763 mutex_unlock(&s->s_mutex);
3764 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3766 } else if (oldstate == newstate) {
3767 continue; /* nothing new with this mds */
3773 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3774 newstate >= CEPH_MDS_STATE_RECONNECT) {
3775 mutex_unlock(&mdsc->mutex);
3776 send_mds_reconnect(mdsc, s);
3777 mutex_lock(&mdsc->mutex);
3781 * kick request on any mds that has gone active.
3783 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3784 newstate >= CEPH_MDS_STATE_ACTIVE) {
3785 if (oldstate != CEPH_MDS_STATE_CREATING &&
3786 oldstate != CEPH_MDS_STATE_STARTING)
3787 pr_info("mds%d recovery completed\n", s->s_mds);
3788 kick_requests(mdsc, i);
3789 ceph_kick_flushing_caps(mdsc, s);
3790 wake_up_session_caps(s, RECONNECT);
3794 for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3795 s = mdsc->sessions[i];
3798 if (!ceph_mdsmap_is_laggy(newmap, i))
3800 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3801 s->s_state == CEPH_MDS_SESSION_HUNG ||
3802 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3803 dout(" connecting to export targets of laggy mds%d\n",
3805 __open_export_target_sessions(mdsc, s);
3817 * caller must hold session s_mutex, dentry->d_lock
3819 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3821 struct ceph_dentry_info *di = ceph_dentry(dentry);
3823 ceph_put_mds_session(di->lease_session);
3824 di->lease_session = NULL;
3827 static void handle_lease(struct ceph_mds_client *mdsc,
3828 struct ceph_mds_session *session,
3829 struct ceph_msg *msg)
3831 struct super_block *sb = mdsc->fsc->sb;
3832 struct inode *inode;
3833 struct dentry *parent, *dentry;
3834 struct ceph_dentry_info *di;
3835 int mds = session->s_mds;
3836 struct ceph_mds_lease *h = msg->front.iov_base;
3838 struct ceph_vino vino;
3842 dout("handle_lease from mds%d\n", mds);
3845 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3847 vino.ino = le64_to_cpu(h->ino);
3848 vino.snap = CEPH_NOSNAP;
3849 seq = le32_to_cpu(h->seq);
3850 dname.len = get_unaligned_le32(h + 1);
3851 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3853 dname.name = (void *)(h + 1) + sizeof(u32);
3856 inode = ceph_find_inode(sb, vino);
3857 dout("handle_lease %s, ino %llx %p %.*s\n",
3858 ceph_lease_op_name(h->action), vino.ino, inode,
3859 dname.len, dname.name);
3861 mutex_lock(&session->s_mutex);
3865 dout("handle_lease no inode %llx\n", vino.ino);
3870 parent = d_find_alias(inode);
3872 dout("no parent dentry on inode %p\n", inode);
3874 goto release; /* hrm... */
3876 dname.hash = full_name_hash(parent, dname.name, dname.len);
3877 dentry = d_lookup(parent, &dname);
3882 spin_lock(&dentry->d_lock);
3883 di = ceph_dentry(dentry);
3884 switch (h->action) {
3885 case CEPH_MDS_LEASE_REVOKE:
3886 if (di->lease_session == session) {
3887 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3888 h->seq = cpu_to_le32(di->lease_seq);
3889 __ceph_mdsc_drop_dentry_lease(dentry);
3894 case CEPH_MDS_LEASE_RENEW:
3895 if (di->lease_session == session &&
3896 di->lease_gen == session->s_cap_gen &&
3897 di->lease_renew_from &&
3898 di->lease_renew_after == 0) {
3899 unsigned long duration =
3900 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3902 di->lease_seq = seq;
3903 di->time = di->lease_renew_from + duration;
3904 di->lease_renew_after = di->lease_renew_from +
3906 di->lease_renew_from = 0;
3910 spin_unlock(&dentry->d_lock);
3917 /* let's just reuse the same message */
3918 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3920 ceph_con_send(&session->s_con, msg);
3923 mutex_unlock(&session->s_mutex);
3924 /* avoid calling iput_final() in mds dispatch threads */
3925 ceph_async_iput(inode);
3929 pr_err("corrupt lease message\n");
3933 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3934 struct inode *inode,
3935 struct dentry *dentry, char action,
3938 struct ceph_msg *msg;
3939 struct ceph_mds_lease *lease;
3940 int len = sizeof(*lease) + sizeof(u32);
3943 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3944 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3945 dnamelen = dentry->d_name.len;
3948 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3951 lease = msg->front.iov_base;
3952 lease->action = action;
3953 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3954 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3955 lease->seq = cpu_to_le32(seq);
3956 put_unaligned_le32(dnamelen, lease + 1);
3957 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3960 * if this is a preemptive lease RELEASE, no need to
3961 * flush request stream, since the actual request will
3964 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3966 ceph_con_send(&session->s_con, msg);
3970 * lock unlock sessions, to wait ongoing session activities
3972 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
3976 mutex_lock(&mdsc->mutex);
3977 for (i = 0; i < mdsc->max_sessions; i++) {
3978 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3981 mutex_unlock(&mdsc->mutex);
3982 mutex_lock(&s->s_mutex);
3983 mutex_unlock(&s->s_mutex);
3984 ceph_put_mds_session(s);
3985 mutex_lock(&mdsc->mutex);
3987 mutex_unlock(&mdsc->mutex);
3993 * delayed work -- periodically trim expired leases, renew caps with mds
3995 static void schedule_delayed(struct ceph_mds_client *mdsc)
3998 unsigned hz = round_jiffies_relative(HZ * delay);
3999 schedule_delayed_work(&mdsc->delayed_work, hz);
4002 static void delayed_work(struct work_struct *work)
4005 struct ceph_mds_client *mdsc =
4006 container_of(work, struct ceph_mds_client, delayed_work.work);
4010 dout("mdsc delayed_work\n");
4012 mutex_lock(&mdsc->mutex);
4013 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4014 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4015 mdsc->last_renew_caps);
4017 mdsc->last_renew_caps = jiffies;
4019 for (i = 0; i < mdsc->max_sessions; i++) {
4020 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4023 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4024 dout("resending session close request for mds%d\n",
4026 request_close_session(mdsc, s);
4027 ceph_put_mds_session(s);
4030 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4031 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4032 s->s_state = CEPH_MDS_SESSION_HUNG;
4033 pr_info("mds%d hung\n", s->s_mds);
4036 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
4037 /* this mds is failed or recovering, just wait */
4038 ceph_put_mds_session(s);
4041 mutex_unlock(&mdsc->mutex);
4043 mutex_lock(&s->s_mutex);
4045 send_renew_caps(mdsc, s);
4047 ceph_con_keepalive(&s->s_con);
4048 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4049 s->s_state == CEPH_MDS_SESSION_HUNG)
4050 ceph_send_cap_releases(mdsc, s);
4051 mutex_unlock(&s->s_mutex);
4052 ceph_put_mds_session(s);
4054 mutex_lock(&mdsc->mutex);
4056 mutex_unlock(&mdsc->mutex);
4058 ceph_check_delayed_caps(mdsc);
4060 ceph_queue_cap_reclaim_work(mdsc);
4062 ceph_trim_snapid_map(mdsc);
4064 schedule_delayed(mdsc);
4067 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4070 struct ceph_mds_client *mdsc;
4072 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4076 mutex_init(&mdsc->mutex);
4077 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4078 if (!mdsc->mdsmap) {
4084 init_completion(&mdsc->safe_umount_waiters);
4085 init_waitqueue_head(&mdsc->session_close_wq);
4086 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4087 mdsc->sessions = NULL;
4088 atomic_set(&mdsc->num_sessions, 0);
4089 mdsc->max_sessions = 0;
4091 atomic64_set(&mdsc->quotarealms_count, 0);
4092 mdsc->quotarealms_inodes = RB_ROOT;
4093 mutex_init(&mdsc->quotarealms_inodes_mutex);
4094 mdsc->last_snap_seq = 0;
4095 init_rwsem(&mdsc->snap_rwsem);
4096 mdsc->snap_realms = RB_ROOT;
4097 INIT_LIST_HEAD(&mdsc->snap_empty);
4098 mdsc->num_snap_realms = 0;
4099 spin_lock_init(&mdsc->snap_empty_lock);
4101 mdsc->oldest_tid = 0;
4102 mdsc->request_tree = RB_ROOT;
4103 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4104 mdsc->last_renew_caps = jiffies;
4105 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4106 spin_lock_init(&mdsc->cap_delay_lock);
4107 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4108 spin_lock_init(&mdsc->snap_flush_lock);
4109 mdsc->last_cap_flush_tid = 1;
4110 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4111 INIT_LIST_HEAD(&mdsc->cap_dirty);
4112 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4113 mdsc->num_cap_flushing = 0;
4114 spin_lock_init(&mdsc->cap_dirty_lock);
4115 init_waitqueue_head(&mdsc->cap_flushing_wq);
4116 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4117 atomic_set(&mdsc->cap_reclaim_pending, 0);
4119 spin_lock_init(&mdsc->dentry_list_lock);
4120 INIT_LIST_HEAD(&mdsc->dentry_leases);
4121 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4123 ceph_caps_init(mdsc);
4124 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4126 spin_lock_init(&mdsc->snapid_map_lock);
4127 mdsc->snapid_map_tree = RB_ROOT;
4128 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4130 init_rwsem(&mdsc->pool_perm_rwsem);
4131 mdsc->pool_perm_tree = RB_ROOT;
4133 strscpy(mdsc->nodename, utsname()->nodename,
4134 sizeof(mdsc->nodename));
4139 * Wait for safe replies on open mds requests. If we time out, drop
4140 * all requests from the tree to avoid dangling dentry refs.
4142 static void wait_requests(struct ceph_mds_client *mdsc)
4144 struct ceph_options *opts = mdsc->fsc->client->options;
4145 struct ceph_mds_request *req;
4147 mutex_lock(&mdsc->mutex);
4148 if (__get_oldest_req(mdsc)) {
4149 mutex_unlock(&mdsc->mutex);
4151 dout("wait_requests waiting for requests\n");
4152 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4153 ceph_timeout_jiffies(opts->mount_timeout));
4155 /* tear down remaining requests */
4156 mutex_lock(&mdsc->mutex);
4157 while ((req = __get_oldest_req(mdsc))) {
4158 dout("wait_requests timed out on tid %llu\n",
4160 __unregister_request(mdsc, req);
4163 mutex_unlock(&mdsc->mutex);
4164 dout("wait_requests done\n");
4168 * called before mount is ro, and before dentries are torn down.
4169 * (hmm, does this still race with new lookups?)
4171 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4173 dout("pre_umount\n");
4176 lock_unlock_sessions(mdsc);
4177 ceph_flush_dirty_caps(mdsc);
4178 wait_requests(mdsc);
4181 * wait for reply handlers to drop their request refs and
4182 * their inode/dcache refs
4186 ceph_cleanup_quotarealms_inodes(mdsc);
4190 * wait for all write mds requests to flush.
4192 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4194 struct ceph_mds_request *req = NULL, *nextreq;
4197 mutex_lock(&mdsc->mutex);
4198 dout("wait_unsafe_requests want %lld\n", want_tid);
4200 req = __get_oldest_req(mdsc);
4201 while (req && req->r_tid <= want_tid) {
4202 /* find next request */
4203 n = rb_next(&req->r_node);
4205 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4208 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4209 (req->r_op & CEPH_MDS_OP_WRITE)) {
4211 ceph_mdsc_get_request(req);
4213 ceph_mdsc_get_request(nextreq);
4214 mutex_unlock(&mdsc->mutex);
4215 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
4216 req->r_tid, want_tid);
4217 wait_for_completion(&req->r_safe_completion);
4218 mutex_lock(&mdsc->mutex);
4219 ceph_mdsc_put_request(req);
4221 break; /* next dne before, so we're done! */
4222 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4223 /* next request was removed from tree */
4224 ceph_mdsc_put_request(nextreq);
4227 ceph_mdsc_put_request(nextreq); /* won't go away */
4231 mutex_unlock(&mdsc->mutex);
4232 dout("wait_unsafe_requests done\n");
4235 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4237 u64 want_tid, want_flush;
4239 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4243 mutex_lock(&mdsc->mutex);
4244 want_tid = mdsc->last_tid;
4245 mutex_unlock(&mdsc->mutex);
4247 ceph_flush_dirty_caps(mdsc);
4248 spin_lock(&mdsc->cap_dirty_lock);
4249 want_flush = mdsc->last_cap_flush_tid;
4250 if (!list_empty(&mdsc->cap_flush_list)) {
4251 struct ceph_cap_flush *cf =
4252 list_last_entry(&mdsc->cap_flush_list,
4253 struct ceph_cap_flush, g_list);
4256 spin_unlock(&mdsc->cap_dirty_lock);
4258 dout("sync want tid %lld flush_seq %lld\n",
4259 want_tid, want_flush);
4261 wait_unsafe_requests(mdsc, want_tid);
4262 wait_caps_flush(mdsc, want_flush);
4266 * true if all sessions are closed, or we force unmount
4268 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4270 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4272 return atomic_read(&mdsc->num_sessions) <= skipped;
4276 * called after sb is ro.
4278 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4280 struct ceph_options *opts = mdsc->fsc->client->options;
4281 struct ceph_mds_session *session;
4285 dout("close_sessions\n");
4287 /* close sessions */
4288 mutex_lock(&mdsc->mutex);
4289 for (i = 0; i < mdsc->max_sessions; i++) {
4290 session = __ceph_lookup_mds_session(mdsc, i);
4293 mutex_unlock(&mdsc->mutex);
4294 mutex_lock(&session->s_mutex);
4295 if (__close_session(mdsc, session) <= 0)
4297 mutex_unlock(&session->s_mutex);
4298 ceph_put_mds_session(session);
4299 mutex_lock(&mdsc->mutex);
4301 mutex_unlock(&mdsc->mutex);
4303 dout("waiting for sessions to close\n");
4304 wait_event_timeout(mdsc->session_close_wq,
4305 done_closing_sessions(mdsc, skipped),
4306 ceph_timeout_jiffies(opts->mount_timeout));
4308 /* tear down remaining sessions */
4309 mutex_lock(&mdsc->mutex);
4310 for (i = 0; i < mdsc->max_sessions; i++) {
4311 if (mdsc->sessions[i]) {
4312 session = get_session(mdsc->sessions[i]);
4313 __unregister_session(mdsc, session);
4314 mutex_unlock(&mdsc->mutex);
4315 mutex_lock(&session->s_mutex);
4316 remove_session_caps(session);
4317 mutex_unlock(&session->s_mutex);
4318 ceph_put_mds_session(session);
4319 mutex_lock(&mdsc->mutex);
4322 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4323 mutex_unlock(&mdsc->mutex);
4325 ceph_cleanup_snapid_map(mdsc);
4326 ceph_cleanup_empty_realms(mdsc);
4328 cancel_work_sync(&mdsc->cap_reclaim_work);
4329 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4334 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4336 struct ceph_mds_session *session;
4339 dout("force umount\n");
4341 mutex_lock(&mdsc->mutex);
4342 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4343 session = __ceph_lookup_mds_session(mdsc, mds);
4346 mutex_unlock(&mdsc->mutex);
4347 mutex_lock(&session->s_mutex);
4348 __close_session(mdsc, session);
4349 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4350 cleanup_session_requests(mdsc, session);
4351 remove_session_caps(session);
4353 mutex_unlock(&session->s_mutex);
4354 ceph_put_mds_session(session);
4355 mutex_lock(&mdsc->mutex);
4356 kick_requests(mdsc, mds);
4358 __wake_requests(mdsc, &mdsc->waiting_for_map);
4359 mutex_unlock(&mdsc->mutex);
4362 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4365 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4367 ceph_mdsmap_destroy(mdsc->mdsmap);
4368 kfree(mdsc->sessions);
4369 ceph_caps_finalize(mdsc);
4370 ceph_pool_perm_destroy(mdsc);
4373 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4375 struct ceph_mds_client *mdsc = fsc->mdsc;
4376 dout("mdsc_destroy %p\n", mdsc);
4381 /* flush out any connection work with references to us */
4384 ceph_mdsc_stop(mdsc);
4388 dout("mdsc_destroy %p done\n", mdsc);
4391 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4393 struct ceph_fs_client *fsc = mdsc->fsc;
4394 const char *mds_namespace = fsc->mount_options->mds_namespace;
4395 void *p = msg->front.iov_base;
4396 void *end = p + msg->front.iov_len;
4400 u32 mount_fscid = (u32)-1;
4401 u8 struct_v, struct_cv;
4404 ceph_decode_need(&p, end, sizeof(u32), bad);
4405 epoch = ceph_decode_32(&p);
4407 dout("handle_fsmap epoch %u\n", epoch);
4409 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4410 struct_v = ceph_decode_8(&p);
4411 struct_cv = ceph_decode_8(&p);
4412 map_len = ceph_decode_32(&p);
4414 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4415 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4417 num_fs = ceph_decode_32(&p);
4418 while (num_fs-- > 0) {
4419 void *info_p, *info_end;
4424 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4425 info_v = ceph_decode_8(&p);
4426 info_cv = ceph_decode_8(&p);
4427 info_len = ceph_decode_32(&p);
4428 ceph_decode_need(&p, end, info_len, bad);
4430 info_end = p + info_len;
4433 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4434 fscid = ceph_decode_32(&info_p);
4435 namelen = ceph_decode_32(&info_p);
4436 ceph_decode_need(&info_p, info_end, namelen, bad);
4438 if (mds_namespace &&
4439 strlen(mds_namespace) == namelen &&
4440 !strncmp(mds_namespace, (char *)info_p, namelen)) {
4441 mount_fscid = fscid;
4446 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4447 if (mount_fscid != (u32)-1) {
4448 fsc->client->monc.fs_cluster_id = mount_fscid;
4449 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4451 ceph_monc_renew_subs(&fsc->client->monc);
4459 pr_err("error decoding fsmap\n");
4461 mutex_lock(&mdsc->mutex);
4462 mdsc->mdsmap_err = err;
4463 __wake_requests(mdsc, &mdsc->waiting_for_map);
4464 mutex_unlock(&mdsc->mutex);
4468 * handle mds map update.
4470 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4474 void *p = msg->front.iov_base;
4475 void *end = p + msg->front.iov_len;
4476 struct ceph_mdsmap *newmap, *oldmap;
4477 struct ceph_fsid fsid;
4480 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4481 ceph_decode_copy(&p, &fsid, sizeof(fsid));
4482 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4484 epoch = ceph_decode_32(&p);
4485 maplen = ceph_decode_32(&p);
4486 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4488 /* do we need it? */
4489 mutex_lock(&mdsc->mutex);
4490 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4491 dout("handle_map epoch %u <= our %u\n",
4492 epoch, mdsc->mdsmap->m_epoch);
4493 mutex_unlock(&mdsc->mutex);
4497 newmap = ceph_mdsmap_decode(&p, end);
4498 if (IS_ERR(newmap)) {
4499 err = PTR_ERR(newmap);
4503 /* swap into place */
4505 oldmap = mdsc->mdsmap;
4506 mdsc->mdsmap = newmap;
4507 check_new_map(mdsc, newmap, oldmap);
4508 ceph_mdsmap_destroy(oldmap);
4510 mdsc->mdsmap = newmap; /* first mds map */
4512 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4515 __wake_requests(mdsc, &mdsc->waiting_for_map);
4516 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4517 mdsc->mdsmap->m_epoch);
4519 mutex_unlock(&mdsc->mutex);
4520 schedule_delayed(mdsc);
4524 mutex_unlock(&mdsc->mutex);
4526 pr_err("error decoding mdsmap %d\n", err);
4530 static struct ceph_connection *con_get(struct ceph_connection *con)
4532 struct ceph_mds_session *s = con->private;
4534 if (get_session(s)) {
4535 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4538 dout("mdsc con_get %p FAIL\n", s);
4542 static void con_put(struct ceph_connection *con)
4544 struct ceph_mds_session *s = con->private;
4546 dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4547 ceph_put_mds_session(s);
4551 * if the client is unresponsive for long enough, the mds will kill
4552 * the session entirely.
4554 static void peer_reset(struct ceph_connection *con)
4556 struct ceph_mds_session *s = con->private;
4557 struct ceph_mds_client *mdsc = s->s_mdsc;
4559 pr_warn("mds%d closed our session\n", s->s_mds);
4560 send_mds_reconnect(mdsc, s);
4563 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4565 struct ceph_mds_session *s = con->private;
4566 struct ceph_mds_client *mdsc = s->s_mdsc;
4567 int type = le16_to_cpu(msg->hdr.type);
4569 mutex_lock(&mdsc->mutex);
4570 if (__verify_registered_session(mdsc, s) < 0) {
4571 mutex_unlock(&mdsc->mutex);
4574 mutex_unlock(&mdsc->mutex);
4577 case CEPH_MSG_MDS_MAP:
4578 ceph_mdsc_handle_mdsmap(mdsc, msg);
4580 case CEPH_MSG_FS_MAP_USER:
4581 ceph_mdsc_handle_fsmap(mdsc, msg);
4583 case CEPH_MSG_CLIENT_SESSION:
4584 handle_session(s, msg);
4586 case CEPH_MSG_CLIENT_REPLY:
4587 handle_reply(s, msg);
4589 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4590 handle_forward(mdsc, s, msg);
4592 case CEPH_MSG_CLIENT_CAPS:
4593 ceph_handle_caps(s, msg);
4595 case CEPH_MSG_CLIENT_SNAP:
4596 ceph_handle_snap(mdsc, s, msg);
4598 case CEPH_MSG_CLIENT_LEASE:
4599 handle_lease(mdsc, s, msg);
4601 case CEPH_MSG_CLIENT_QUOTA:
4602 ceph_handle_quota(mdsc, s, msg);
4606 pr_err("received unknown message type %d %s\n", type,
4607 ceph_msg_type_name(type));
4618 * Note: returned pointer is the address of a structure that's
4619 * managed separately. Caller must *not* attempt to free it.
4621 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4622 int *proto, int force_new)
4624 struct ceph_mds_session *s = con->private;
4625 struct ceph_mds_client *mdsc = s->s_mdsc;
4626 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4627 struct ceph_auth_handshake *auth = &s->s_auth;
4629 if (force_new && auth->authorizer) {
4630 ceph_auth_destroy_authorizer(auth->authorizer);
4631 auth->authorizer = NULL;
4633 if (!auth->authorizer) {
4634 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4637 return ERR_PTR(ret);
4639 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4642 return ERR_PTR(ret);
4644 *proto = ac->protocol;
4649 static int add_authorizer_challenge(struct ceph_connection *con,
4650 void *challenge_buf, int challenge_buf_len)
4652 struct ceph_mds_session *s = con->private;
4653 struct ceph_mds_client *mdsc = s->s_mdsc;
4654 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4656 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4657 challenge_buf, challenge_buf_len);
4660 static int verify_authorizer_reply(struct ceph_connection *con)
4662 struct ceph_mds_session *s = con->private;
4663 struct ceph_mds_client *mdsc = s->s_mdsc;
4664 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4666 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4669 static int invalidate_authorizer(struct ceph_connection *con)
4671 struct ceph_mds_session *s = con->private;
4672 struct ceph_mds_client *mdsc = s->s_mdsc;
4673 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4675 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4677 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4680 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4681 struct ceph_msg_header *hdr, int *skip)
4683 struct ceph_msg *msg;
4684 int type = (int) le16_to_cpu(hdr->type);
4685 int front_len = (int) le32_to_cpu(hdr->front_len);
4691 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4693 pr_err("unable to allocate msg type %d len %d\n",
4701 static int mds_sign_message(struct ceph_msg *msg)
4703 struct ceph_mds_session *s = msg->con->private;
4704 struct ceph_auth_handshake *auth = &s->s_auth;
4706 return ceph_auth_sign_message(auth, msg);
4709 static int mds_check_message_signature(struct ceph_msg *msg)
4711 struct ceph_mds_session *s = msg->con->private;
4712 struct ceph_auth_handshake *auth = &s->s_auth;
4714 return ceph_auth_check_message_signature(auth, msg);
4717 static const struct ceph_connection_operations mds_con_ops = {
4720 .dispatch = dispatch,
4721 .get_authorizer = get_authorizer,
4722 .add_authorizer_challenge = add_authorizer_challenge,
4723 .verify_authorizer_reply = verify_authorizer_reply,
4724 .invalidate_authorizer = invalidate_authorizer,
4725 .peer_reset = peer_reset,
4726 .alloc_msg = mds_alloc_msg,
4727 .sign_message = mds_sign_message,
4728 .check_message_signature = mds_check_message_signature,