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>
12 #include <linux/bits.h>
13 #include <linux/ktime.h>
14 #include <linux/bitmap.h>
15 #include <linux/mnt_idmapping.h>
18 #include "mds_client.h"
21 #include <linux/ceph/ceph_features.h>
22 #include <linux/ceph/messenger.h>
23 #include <linux/ceph/decode.h>
24 #include <linux/ceph/pagelist.h>
25 #include <linux/ceph/auth.h>
26 #include <linux/ceph/debugfs.h>
28 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
31 * A cluster of MDS (metadata server) daemons is responsible for
32 * managing the file system namespace (the directory hierarchy and
33 * inodes) and for coordinating shared access to storage. Metadata is
34 * partitioning hierarchically across a number of servers, and that
35 * partition varies over time as the cluster adjusts the distribution
36 * in order to balance load.
38 * The MDS client is primarily responsible to managing synchronous
39 * metadata requests for operations like open, unlink, and so forth.
40 * If there is a MDS failure, we find out about it when we (possibly
41 * request and) receive a new MDS map, and can resubmit affected
44 * For the most part, though, we take advantage of a lossless
45 * communications channel to the MDS, and do not need to worry about
46 * timing out or resubmitting requests.
48 * We maintain a stateful "session" with each MDS we interact with.
49 * Within each session, we sent periodic heartbeat messages to ensure
50 * any capabilities or leases we have been issues remain valid. If
51 * the session times out and goes stale, our leases and capabilities
52 * are no longer valid.
55 struct ceph_reconnect_state {
56 struct ceph_mds_session *session;
57 int nr_caps, nr_realms;
58 struct ceph_pagelist *pagelist;
63 static void __wake_requests(struct ceph_mds_client *mdsc,
64 struct list_head *head);
65 static void ceph_cap_release_work(struct work_struct *work);
66 static void ceph_cap_reclaim_work(struct work_struct *work);
68 static const struct ceph_connection_operations mds_con_ops;
75 static int parse_reply_info_quota(void **p, void *end,
76 struct ceph_mds_reply_info_in *info)
78 u8 struct_v, struct_compat;
81 ceph_decode_8_safe(p, end, struct_v, bad);
82 ceph_decode_8_safe(p, end, struct_compat, bad);
83 /* struct_v is expected to be >= 1. we only
84 * understand encoding with struct_compat == 1. */
85 if (!struct_v || struct_compat != 1)
87 ceph_decode_32_safe(p, end, struct_len, bad);
88 ceph_decode_need(p, end, struct_len, bad);
89 end = *p + struct_len;
90 ceph_decode_64_safe(p, end, info->max_bytes, bad);
91 ceph_decode_64_safe(p, end, info->max_files, bad);
99 * parse individual inode info
101 static int parse_reply_info_in(void **p, void *end,
102 struct ceph_mds_reply_info_in *info,
108 if (features == (u64)-1) {
111 ceph_decode_8_safe(p, end, struct_v, bad);
112 ceph_decode_8_safe(p, end, struct_compat, bad);
113 /* struct_v is expected to be >= 1. we only understand
114 * encoding with struct_compat == 1. */
115 if (!struct_v || struct_compat != 1)
117 ceph_decode_32_safe(p, end, struct_len, bad);
118 ceph_decode_need(p, end, struct_len, bad);
119 end = *p + struct_len;
122 ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
124 *p += sizeof(struct ceph_mds_reply_inode) +
125 sizeof(*info->in->fragtree.splits) *
126 le32_to_cpu(info->in->fragtree.nsplits);
128 ceph_decode_32_safe(p, end, info->symlink_len, bad);
129 ceph_decode_need(p, end, info->symlink_len, bad);
131 *p += info->symlink_len;
133 ceph_decode_copy_safe(p, end, &info->dir_layout,
134 sizeof(info->dir_layout), bad);
135 ceph_decode_32_safe(p, end, info->xattr_len, bad);
136 ceph_decode_need(p, end, info->xattr_len, bad);
137 info->xattr_data = *p;
138 *p += info->xattr_len;
140 if (features == (u64)-1) {
142 ceph_decode_64_safe(p, end, info->inline_version, bad);
143 ceph_decode_32_safe(p, end, info->inline_len, bad);
144 ceph_decode_need(p, end, info->inline_len, bad);
145 info->inline_data = *p;
146 *p += info->inline_len;
148 err = parse_reply_info_quota(p, end, info);
152 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
153 if (info->pool_ns_len > 0) {
154 ceph_decode_need(p, end, info->pool_ns_len, bad);
155 info->pool_ns_data = *p;
156 *p += info->pool_ns_len;
160 ceph_decode_need(p, end, sizeof(info->btime), bad);
161 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
163 /* change attribute */
164 ceph_decode_64_safe(p, end, info->change_attr, bad);
168 ceph_decode_32_safe(p, end, info->dir_pin, bad);
170 info->dir_pin = -ENODATA;
173 /* snapshot birth time, remains zero for v<=2 */
175 ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
176 ceph_decode_copy(p, &info->snap_btime,
177 sizeof(info->snap_btime));
179 memset(&info->snap_btime, 0, sizeof(info->snap_btime));
182 /* snapshot count, remains zero for v<=3 */
184 ceph_decode_64_safe(p, end, info->rsnaps, bad);
192 ceph_decode_32_safe(p, end, alen, bad);
198 ceph_decode_32_safe(p, end, len, bad);
199 ceph_decode_skip_n(p, end, len, bad);
201 ceph_decode_32_safe(p, end, len, bad);
202 ceph_decode_skip_n(p, end, len, bad);
206 /* fscrypt flag -- ignore */
208 ceph_decode_skip_8(p, end, bad);
210 info->fscrypt_auth = NULL;
211 info->fscrypt_auth_len = 0;
212 info->fscrypt_file = NULL;
213 info->fscrypt_file_len = 0;
215 ceph_decode_32_safe(p, end, info->fscrypt_auth_len, bad);
216 if (info->fscrypt_auth_len) {
217 info->fscrypt_auth = kmalloc(info->fscrypt_auth_len,
219 if (!info->fscrypt_auth)
221 ceph_decode_copy_safe(p, end, info->fscrypt_auth,
222 info->fscrypt_auth_len, bad);
224 ceph_decode_32_safe(p, end, info->fscrypt_file_len, bad);
225 if (info->fscrypt_file_len) {
226 info->fscrypt_file = kmalloc(info->fscrypt_file_len,
228 if (!info->fscrypt_file)
230 ceph_decode_copy_safe(p, end, info->fscrypt_file,
231 info->fscrypt_file_len, bad);
236 /* legacy (unversioned) struct */
237 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
238 ceph_decode_64_safe(p, end, info->inline_version, bad);
239 ceph_decode_32_safe(p, end, info->inline_len, bad);
240 ceph_decode_need(p, end, info->inline_len, bad);
241 info->inline_data = *p;
242 *p += info->inline_len;
244 info->inline_version = CEPH_INLINE_NONE;
246 if (features & CEPH_FEATURE_MDS_QUOTA) {
247 err = parse_reply_info_quota(p, end, info);
255 info->pool_ns_len = 0;
256 info->pool_ns_data = NULL;
257 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
258 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
259 if (info->pool_ns_len > 0) {
260 ceph_decode_need(p, end, info->pool_ns_len, bad);
261 info->pool_ns_data = *p;
262 *p += info->pool_ns_len;
266 if (features & CEPH_FEATURE_FS_BTIME) {
267 ceph_decode_need(p, end, sizeof(info->btime), bad);
268 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
269 ceph_decode_64_safe(p, end, info->change_attr, bad);
272 info->dir_pin = -ENODATA;
273 /* info->snap_btime and info->rsnaps remain zero */
282 static int parse_reply_info_dir(void **p, void *end,
283 struct ceph_mds_reply_dirfrag **dirfrag,
286 if (features == (u64)-1) {
287 u8 struct_v, struct_compat;
289 ceph_decode_8_safe(p, end, struct_v, bad);
290 ceph_decode_8_safe(p, end, struct_compat, bad);
291 /* struct_v is expected to be >= 1. we only understand
292 * encoding whose struct_compat == 1. */
293 if (!struct_v || struct_compat != 1)
295 ceph_decode_32_safe(p, end, struct_len, bad);
296 ceph_decode_need(p, end, struct_len, bad);
297 end = *p + struct_len;
300 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
302 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
303 if (unlikely(*p > end))
305 if (features == (u64)-1)
312 static int parse_reply_info_lease(void **p, void *end,
313 struct ceph_mds_reply_lease **lease,
314 u64 features, u32 *altname_len, u8 **altname)
320 if (features == (u64)-1) {
323 ceph_decode_8_safe(p, end, struct_v, bad);
324 ceph_decode_8_safe(p, end, struct_compat, bad);
326 /* struct_v is expected to be >= 1. we only understand
327 * encoding whose struct_compat == 1. */
328 if (!struct_v || struct_compat != 1)
331 ceph_decode_32_safe(p, end, struct_len, bad);
333 struct_len = sizeof(**lease);
338 lend = *p + struct_len;
339 ceph_decode_need(p, end, struct_len, bad);
341 *p += sizeof(**lease);
343 if (features == (u64)-1) {
345 ceph_decode_32_safe(p, end, *altname_len, bad);
346 ceph_decode_need(p, end, *altname_len, bad);
361 * parse a normal reply, which may contain a (dir+)dentry and/or a
364 static int parse_reply_info_trace(void **p, void *end,
365 struct ceph_mds_reply_info_parsed *info,
370 if (info->head->is_dentry) {
371 err = parse_reply_info_in(p, end, &info->diri, features);
375 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
379 ceph_decode_32_safe(p, end, info->dname_len, bad);
380 ceph_decode_need(p, end, info->dname_len, bad);
382 *p += info->dname_len;
384 err = parse_reply_info_lease(p, end, &info->dlease, features,
385 &info->altname_len, &info->altname);
390 if (info->head->is_target) {
391 err = parse_reply_info_in(p, end, &info->targeti, features);
396 if (unlikely(*p != end))
403 pr_err("problem parsing mds trace %d\n", err);
408 * parse readdir results
410 static int parse_reply_info_readdir(void **p, void *end,
411 struct ceph_mds_request *req,
414 struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
415 struct ceph_client *cl = req->r_mdsc->fsc->client;
419 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
423 ceph_decode_need(p, end, sizeof(num) + 2, bad);
424 num = ceph_decode_32(p);
426 u16 flags = ceph_decode_16(p);
427 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
428 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
429 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
430 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
435 BUG_ON(!info->dir_entries);
436 if ((unsigned long)(info->dir_entries + num) >
437 (unsigned long)info->dir_entries + info->dir_buf_size) {
438 pr_err_client(cl, "dir contents are larger than expected\n");
445 struct inode *inode = d_inode(req->r_dentry);
446 struct ceph_inode_info *ci = ceph_inode(inode);
447 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
448 struct fscrypt_str tname = FSTR_INIT(NULL, 0);
449 struct fscrypt_str oname = FSTR_INIT(NULL, 0);
450 struct ceph_fname fname;
451 u32 altname_len, _name_len;
455 ceph_decode_32_safe(p, end, _name_len, bad);
456 ceph_decode_need(p, end, _name_len, bad);
459 doutc(cl, "parsed dir dname '%.*s'\n", _name_len, _name);
461 if (info->hash_order)
462 rde->raw_hash = ceph_str_hash(ci->i_dir_layout.dl_dir_hash,
466 err = parse_reply_info_lease(p, end, &rde->lease, features,
467 &altname_len, &altname);
472 * Try to dencrypt the dentry names and update them
473 * in the ceph_mds_reply_dir_entry struct.
477 fname.name_len = _name_len;
478 fname.ctext = altname;
479 fname.ctext_len = altname_len;
481 * The _name_len maybe larger than altname_len, such as
482 * when the human readable name length is in range of
483 * (CEPH_NOHASH_NAME_MAX, CEPH_NOHASH_NAME_MAX + SHA256_DIGEST_SIZE),
484 * then the copy in ceph_fname_to_usr will corrupt the
485 * data if there has no encryption key.
487 * Just set the no_copy flag and then if there has no
488 * encryption key the oname.name will be assigned to
491 fname.no_copy = true;
492 if (altname_len == 0) {
494 * Set tname to _name, and this will be used
495 * to do the base64_decode in-place. It's
496 * safe because the decoded string should
497 * always be shorter, which is 3/4 of origin
503 * Set oname to _name too, and this will be
504 * used to do the dencryption in-place.
507 oname.len = _name_len;
510 * This will do the decryption only in-place
511 * from altname cryptext directly.
513 oname.name = altname;
514 oname.len = altname_len;
516 rde->is_nokey = false;
517 err = ceph_fname_to_usr(&fname, &tname, &oname, &rde->is_nokey);
519 pr_err_client(cl, "unable to decode %.*s, got %d\n",
520 _name_len, _name, err);
523 rde->name = oname.name;
524 rde->name_len = oname.len;
527 err = parse_reply_info_in(p, end, &rde->inode, features);
530 /* ceph_readdir_prepopulate() will update it */
537 /* Skip over any unrecognized fields */
544 pr_err_client(cl, "problem parsing dir contents %d\n", err);
549 * parse fcntl F_GETLK results
551 static int parse_reply_info_filelock(void **p, void *end,
552 struct ceph_mds_reply_info_parsed *info,
555 if (*p + sizeof(*info->filelock_reply) > end)
558 info->filelock_reply = *p;
560 /* Skip over any unrecognized fields */
568 #if BITS_PER_LONG == 64
570 #define DELEGATED_INO_AVAILABLE xa_mk_value(1)
572 static int ceph_parse_deleg_inos(void **p, void *end,
573 struct ceph_mds_session *s)
575 struct ceph_client *cl = s->s_mdsc->fsc->client;
578 ceph_decode_32_safe(p, end, sets, bad);
579 doutc(cl, "got %u sets of delegated inodes\n", sets);
583 ceph_decode_64_safe(p, end, start, bad);
584 ceph_decode_64_safe(p, end, len, bad);
586 /* Don't accept a delegation of system inodes */
587 if (start < CEPH_INO_SYSTEM_BASE) {
588 pr_warn_ratelimited_client(cl,
589 "ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
594 int err = xa_insert(&s->s_delegated_inos, start++,
595 DELEGATED_INO_AVAILABLE,
598 doutc(cl, "added delegated inode 0x%llx\n", start - 1);
599 } else if (err == -EBUSY) {
601 "MDS delegated inode 0x%llx more than once.\n",
613 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
618 xa_for_each(&s->s_delegated_inos, ino, val) {
619 val = xa_erase(&s->s_delegated_inos, ino);
620 if (val == DELEGATED_INO_AVAILABLE)
626 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
628 return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
631 #else /* BITS_PER_LONG == 64 */
633 * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
634 * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
637 static int ceph_parse_deleg_inos(void **p, void *end,
638 struct ceph_mds_session *s)
642 ceph_decode_32_safe(p, end, sets, bad);
644 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
650 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
655 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
659 #endif /* BITS_PER_LONG == 64 */
662 * parse create results
664 static int parse_reply_info_create(void **p, void *end,
665 struct ceph_mds_reply_info_parsed *info,
666 u64 features, struct ceph_mds_session *s)
670 if (features == (u64)-1 ||
671 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
673 /* Malformed reply? */
674 info->has_create_ino = false;
675 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
676 info->has_create_ino = true;
677 /* struct_v, struct_compat, and len */
678 ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
679 ceph_decode_64_safe(p, end, info->ino, bad);
680 ret = ceph_parse_deleg_inos(p, end, s);
685 ceph_decode_64_safe(p, end, info->ino, bad);
686 info->has_create_ino = true;
693 /* Skip over any unrecognized fields */
700 static int parse_reply_info_getvxattr(void **p, void *end,
701 struct ceph_mds_reply_info_parsed *info,
706 ceph_decode_skip_8(p, end, bad); /* skip current version: 1 */
707 ceph_decode_skip_8(p, end, bad); /* skip first version: 1 */
708 ceph_decode_skip_32(p, end, bad); /* skip payload length */
710 ceph_decode_32_safe(p, end, value_len, bad);
712 if (value_len == end - *p) {
713 info->xattr_info.xattr_value = *p;
714 info->xattr_info.xattr_value_len = value_len;
723 * parse extra results
725 static int parse_reply_info_extra(void **p, void *end,
726 struct ceph_mds_request *req,
727 u64 features, struct ceph_mds_session *s)
729 struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
730 u32 op = le32_to_cpu(info->head->op);
732 if (op == CEPH_MDS_OP_GETFILELOCK)
733 return parse_reply_info_filelock(p, end, info, features);
734 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
735 return parse_reply_info_readdir(p, end, req, features);
736 else if (op == CEPH_MDS_OP_CREATE)
737 return parse_reply_info_create(p, end, info, features, s);
738 else if (op == CEPH_MDS_OP_GETVXATTR)
739 return parse_reply_info_getvxattr(p, end, info, features);
745 * parse entire mds reply
747 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
748 struct ceph_mds_request *req, u64 features)
750 struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
751 struct ceph_client *cl = s->s_mdsc->fsc->client;
756 info->head = msg->front.iov_base;
757 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
758 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
761 ceph_decode_32_safe(&p, end, len, bad);
763 ceph_decode_need(&p, end, len, bad);
764 err = parse_reply_info_trace(&p, p+len, info, features);
770 ceph_decode_32_safe(&p, end, len, bad);
772 ceph_decode_need(&p, end, len, bad);
773 err = parse_reply_info_extra(&p, p+len, req, features, s);
779 ceph_decode_32_safe(&p, end, len, bad);
780 info->snapblob_len = len;
791 pr_err_client(cl, "mds parse_reply err %d\n", err);
796 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
800 kfree(info->diri.fscrypt_auth);
801 kfree(info->diri.fscrypt_file);
802 kfree(info->targeti.fscrypt_auth);
803 kfree(info->targeti.fscrypt_file);
804 if (!info->dir_entries)
807 for (i = 0; i < info->dir_nr; i++) {
808 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
810 kfree(rde->inode.fscrypt_auth);
811 kfree(rde->inode.fscrypt_file);
813 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
817 * In async unlink case the kclient won't wait for the first reply
818 * from MDS and just drop all the links and unhash the dentry and then
819 * succeeds immediately.
821 * For any new create/link/rename,etc requests followed by using the
822 * same file names we must wait for the first reply of the inflight
823 * unlink request, or the MDS possibly will fail these following
824 * requests with -EEXIST if the inflight async unlink request was
825 * delayed for some reasons.
827 * And the worst case is that for the none async openc request it will
828 * successfully open the file if the CDentry hasn't been unlinked yet,
829 * but later the previous delayed async unlink request will remove the
830 * CDenty. That means the just created file is possiblly deleted later
833 * We need to wait for the inflight async unlink requests to finish
834 * when creating new files/directories by using the same file names.
836 int ceph_wait_on_conflict_unlink(struct dentry *dentry)
838 struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
839 struct ceph_client *cl = fsc->client;
840 struct dentry *pdentry = dentry->d_parent;
841 struct dentry *udentry, *found = NULL;
842 struct ceph_dentry_info *di;
844 u32 hash = dentry->d_name.hash;
847 dname.name = dentry->d_name.name;
848 dname.len = dentry->d_name.len;
851 hash_for_each_possible_rcu(fsc->async_unlink_conflict, di,
853 udentry = di->dentry;
855 spin_lock(&udentry->d_lock);
856 if (udentry->d_name.hash != hash)
858 if (unlikely(udentry->d_parent != pdentry))
860 if (!hash_hashed(&di->hnode))
863 if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
864 pr_warn_client(cl, "dentry %p:%pd async unlink bit is not set\n",
867 if (!d_same_name(udentry, pdentry, &dname))
870 found = dget_dlock(udentry);
871 spin_unlock(&udentry->d_lock);
874 spin_unlock(&udentry->d_lock);
881 doutc(cl, "dentry %p:%pd conflict with old %p:%pd\n", dentry, dentry,
884 err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT,
894 const char *ceph_session_state_name(int s)
897 case CEPH_MDS_SESSION_NEW: return "new";
898 case CEPH_MDS_SESSION_OPENING: return "opening";
899 case CEPH_MDS_SESSION_OPEN: return "open";
900 case CEPH_MDS_SESSION_HUNG: return "hung";
901 case CEPH_MDS_SESSION_CLOSING: return "closing";
902 case CEPH_MDS_SESSION_CLOSED: return "closed";
903 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
904 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
905 case CEPH_MDS_SESSION_REJECTED: return "rejected";
906 default: return "???";
910 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
912 if (refcount_inc_not_zero(&s->s_ref))
917 void ceph_put_mds_session(struct ceph_mds_session *s)
919 if (IS_ERR_OR_NULL(s))
922 if (refcount_dec_and_test(&s->s_ref)) {
923 if (s->s_auth.authorizer)
924 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
925 WARN_ON(mutex_is_locked(&s->s_mutex));
926 xa_destroy(&s->s_delegated_inos);
932 * called under mdsc->mutex
934 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
937 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
939 return ceph_get_mds_session(mdsc->sessions[mds]);
942 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
944 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
950 static int __verify_registered_session(struct ceph_mds_client *mdsc,
951 struct ceph_mds_session *s)
953 if (s->s_mds >= mdsc->max_sessions ||
954 mdsc->sessions[s->s_mds] != s)
960 * create+register a new session for given mds.
961 * called under mdsc->mutex.
963 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
966 struct ceph_client *cl = mdsc->fsc->client;
967 struct ceph_mds_session *s;
969 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
970 return ERR_PTR(-EIO);
972 if (mds >= mdsc->mdsmap->possible_max_rank)
973 return ERR_PTR(-EINVAL);
975 s = kzalloc(sizeof(*s), GFP_NOFS);
977 return ERR_PTR(-ENOMEM);
979 if (mds >= mdsc->max_sessions) {
980 int newmax = 1 << get_count_order(mds + 1);
981 struct ceph_mds_session **sa;
983 doutc(cl, "realloc to %d\n", newmax);
984 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
987 if (mdsc->sessions) {
988 memcpy(sa, mdsc->sessions,
989 mdsc->max_sessions * sizeof(void *));
990 kfree(mdsc->sessions);
993 mdsc->max_sessions = newmax;
996 doutc(cl, "mds%d\n", mds);
999 s->s_state = CEPH_MDS_SESSION_NEW;
1000 mutex_init(&s->s_mutex);
1002 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
1004 atomic_set(&s->s_cap_gen, 1);
1005 s->s_cap_ttl = jiffies - 1;
1007 spin_lock_init(&s->s_cap_lock);
1008 INIT_LIST_HEAD(&s->s_caps);
1009 refcount_set(&s->s_ref, 1);
1010 INIT_LIST_HEAD(&s->s_waiting);
1011 INIT_LIST_HEAD(&s->s_unsafe);
1012 xa_init(&s->s_delegated_inos);
1013 INIT_LIST_HEAD(&s->s_cap_releases);
1014 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
1016 INIT_LIST_HEAD(&s->s_cap_dirty);
1017 INIT_LIST_HEAD(&s->s_cap_flushing);
1019 mdsc->sessions[mds] = s;
1020 atomic_inc(&mdsc->num_sessions);
1021 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
1023 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
1024 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
1030 return ERR_PTR(-ENOMEM);
1034 * called under mdsc->mutex
1036 static void __unregister_session(struct ceph_mds_client *mdsc,
1037 struct ceph_mds_session *s)
1039 doutc(mdsc->fsc->client, "mds%d %p\n", s->s_mds, s);
1040 BUG_ON(mdsc->sessions[s->s_mds] != s);
1041 mdsc->sessions[s->s_mds] = NULL;
1042 ceph_con_close(&s->s_con);
1043 ceph_put_mds_session(s);
1044 atomic_dec(&mdsc->num_sessions);
1048 * drop session refs in request.
1050 * should be last request ref, or hold mdsc->mutex
1052 static void put_request_session(struct ceph_mds_request *req)
1054 if (req->r_session) {
1055 ceph_put_mds_session(req->r_session);
1056 req->r_session = NULL;
1060 void ceph_mdsc_iterate_sessions(struct ceph_mds_client *mdsc,
1061 void (*cb)(struct ceph_mds_session *),
1066 mutex_lock(&mdsc->mutex);
1067 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
1068 struct ceph_mds_session *s;
1070 s = __ceph_lookup_mds_session(mdsc, mds);
1074 if (check_state && !check_session_state(s)) {
1075 ceph_put_mds_session(s);
1079 mutex_unlock(&mdsc->mutex);
1081 ceph_put_mds_session(s);
1082 mutex_lock(&mdsc->mutex);
1084 mutex_unlock(&mdsc->mutex);
1087 void ceph_mdsc_release_request(struct kref *kref)
1089 struct ceph_mds_request *req = container_of(kref,
1090 struct ceph_mds_request,
1092 ceph_mdsc_release_dir_caps_async(req);
1093 destroy_reply_info(&req->r_reply_info);
1095 ceph_msg_put(req->r_request);
1097 ceph_msg_put(req->r_reply);
1099 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1102 if (req->r_parent) {
1103 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
1104 iput(req->r_parent);
1106 iput(req->r_target_inode);
1107 iput(req->r_new_inode);
1109 dput(req->r_dentry);
1110 if (req->r_old_dentry)
1111 dput(req->r_old_dentry);
1112 if (req->r_old_dentry_dir) {
1114 * track (and drop pins for) r_old_dentry_dir
1115 * separately, since r_old_dentry's d_parent may have
1116 * changed between the dir mutex being dropped and
1117 * this request being freed.
1119 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
1121 iput(req->r_old_dentry_dir);
1123 kfree(req->r_path1);
1124 kfree(req->r_path2);
1125 put_cred(req->r_cred);
1126 if (req->r_mnt_idmap)
1127 mnt_idmap_put(req->r_mnt_idmap);
1128 if (req->r_pagelist)
1129 ceph_pagelist_release(req->r_pagelist);
1130 kfree(req->r_fscrypt_auth);
1131 kfree(req->r_altname);
1132 put_request_session(req);
1133 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
1134 WARN_ON_ONCE(!list_empty(&req->r_wait));
1135 kmem_cache_free(ceph_mds_request_cachep, req);
1138 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
1141 * lookup session, bump ref if found.
1143 * called under mdsc->mutex.
1145 static struct ceph_mds_request *
1146 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
1148 struct ceph_mds_request *req;
1150 req = lookup_request(&mdsc->request_tree, tid);
1152 ceph_mdsc_get_request(req);
1158 * Register an in-flight request, and assign a tid. Link to directory
1159 * are modifying (if any).
1161 * Called under mdsc->mutex.
1163 static void __register_request(struct ceph_mds_client *mdsc,
1164 struct ceph_mds_request *req,
1167 struct ceph_client *cl = mdsc->fsc->client;
1170 req->r_tid = ++mdsc->last_tid;
1171 if (req->r_num_caps) {
1172 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
1175 pr_err_client(cl, "%p failed to reserve caps: %d\n",
1177 /* set req->r_err to fail early from __do_request */
1182 doutc(cl, "%p tid %lld\n", req, req->r_tid);
1183 ceph_mdsc_get_request(req);
1184 insert_request(&mdsc->request_tree, req);
1186 req->r_cred = get_current_cred();
1187 if (!req->r_mnt_idmap)
1188 req->r_mnt_idmap = &nop_mnt_idmap;
1190 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
1191 mdsc->oldest_tid = req->r_tid;
1194 struct ceph_inode_info *ci = ceph_inode(dir);
1197 req->r_unsafe_dir = dir;
1198 spin_lock(&ci->i_unsafe_lock);
1199 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
1200 spin_unlock(&ci->i_unsafe_lock);
1204 static void __unregister_request(struct ceph_mds_client *mdsc,
1205 struct ceph_mds_request *req)
1207 doutc(mdsc->fsc->client, "%p tid %lld\n", req, req->r_tid);
1209 /* Never leave an unregistered request on an unsafe list! */
1210 list_del_init(&req->r_unsafe_item);
1212 if (req->r_tid == mdsc->oldest_tid) {
1213 struct rb_node *p = rb_next(&req->r_node);
1214 mdsc->oldest_tid = 0;
1216 struct ceph_mds_request *next_req =
1217 rb_entry(p, struct ceph_mds_request, r_node);
1218 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
1219 mdsc->oldest_tid = next_req->r_tid;
1226 erase_request(&mdsc->request_tree, req);
1228 if (req->r_unsafe_dir) {
1229 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
1230 spin_lock(&ci->i_unsafe_lock);
1231 list_del_init(&req->r_unsafe_dir_item);
1232 spin_unlock(&ci->i_unsafe_lock);
1234 if (req->r_target_inode &&
1235 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
1236 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
1237 spin_lock(&ci->i_unsafe_lock);
1238 list_del_init(&req->r_unsafe_target_item);
1239 spin_unlock(&ci->i_unsafe_lock);
1242 if (req->r_unsafe_dir) {
1243 iput(req->r_unsafe_dir);
1244 req->r_unsafe_dir = NULL;
1247 complete_all(&req->r_safe_completion);
1249 ceph_mdsc_put_request(req);
1253 * Walk back up the dentry tree until we hit a dentry representing a
1254 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
1255 * when calling this) to ensure that the objects won't disappear while we're
1256 * working with them. Once we hit a candidate dentry, we attempt to take a
1257 * reference to it, and return that as the result.
1259 static struct inode *get_nonsnap_parent(struct dentry *dentry)
1261 struct inode *inode = NULL;
1263 while (dentry && !IS_ROOT(dentry)) {
1264 inode = d_inode_rcu(dentry);
1265 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
1267 dentry = dentry->d_parent;
1270 inode = igrab(inode);
1275 * Choose mds to send request to next. If there is a hint set in the
1276 * request (e.g., due to a prior forward hint from the mds), use that.
1277 * Otherwise, consult frag tree and/or caps to identify the
1278 * appropriate mds. If all else fails, choose randomly.
1280 * Called under mdsc->mutex.
1282 static int __choose_mds(struct ceph_mds_client *mdsc,
1283 struct ceph_mds_request *req,
1286 struct inode *inode;
1287 struct ceph_inode_info *ci;
1288 struct ceph_cap *cap;
1289 int mode = req->r_direct_mode;
1291 u32 hash = req->r_direct_hash;
1292 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1293 struct ceph_client *cl = mdsc->fsc->client;
1299 * is there a specific mds we should try? ignore hint if we have
1300 * no session and the mds is not up (active or recovering).
1302 if (req->r_resend_mds >= 0 &&
1303 (__have_session(mdsc, req->r_resend_mds) ||
1304 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1305 doutc(cl, "using resend_mds mds%d\n", req->r_resend_mds);
1306 return req->r_resend_mds;
1309 if (mode == USE_RANDOM_MDS)
1314 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1315 inode = req->r_inode;
1318 /* req->r_dentry is non-null for LSSNAP request */
1320 inode = get_nonsnap_parent(req->r_dentry);
1322 doutc(cl, "using snapdir's parent %p %llx.%llx\n",
1323 inode, ceph_vinop(inode));
1325 } else if (req->r_dentry) {
1326 /* ignore race with rename; old or new d_parent is okay */
1327 struct dentry *parent;
1331 parent = READ_ONCE(req->r_dentry->d_parent);
1332 dir = req->r_parent ? : d_inode_rcu(parent);
1334 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1335 /* not this fs or parent went negative */
1336 inode = d_inode(req->r_dentry);
1339 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1340 /* direct snapped/virtual snapdir requests
1341 * based on parent dir inode */
1342 inode = get_nonsnap_parent(parent);
1343 doutc(cl, "using nonsnap parent %p %llx.%llx\n",
1344 inode, ceph_vinop(inode));
1347 inode = d_inode(req->r_dentry);
1348 if (!inode || mode == USE_AUTH_MDS) {
1351 hash = ceph_dentry_hash(dir, req->r_dentry);
1363 doutc(cl, "%p %llx.%llx is_hash=%d (0x%x) mode %d\n", inode,
1364 ceph_vinop(inode), (int)is_hash, hash, mode);
1365 ci = ceph_inode(inode);
1367 if (is_hash && S_ISDIR(inode->i_mode)) {
1368 struct ceph_inode_frag frag;
1371 ceph_choose_frag(ci, hash, &frag, &found);
1373 if (mode == USE_ANY_MDS && frag.ndist > 0) {
1376 /* choose a random replica */
1377 get_random_bytes(&r, 1);
1380 doutc(cl, "%p %llx.%llx frag %u mds%d (%d/%d)\n",
1381 inode, ceph_vinop(inode), frag.frag,
1382 mds, (int)r, frag.ndist);
1383 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1384 CEPH_MDS_STATE_ACTIVE &&
1385 !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1389 /* since this file/dir wasn't known to be
1390 * replicated, then we want to look for the
1391 * authoritative mds. */
1392 if (frag.mds >= 0) {
1393 /* choose auth mds */
1395 doutc(cl, "%p %llx.%llx frag %u mds%d (auth)\n",
1396 inode, ceph_vinop(inode), frag.frag, mds);
1397 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1398 CEPH_MDS_STATE_ACTIVE) {
1399 if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1404 mode = USE_AUTH_MDS;
1408 spin_lock(&ci->i_ceph_lock);
1410 if (mode == USE_AUTH_MDS)
1411 cap = ci->i_auth_cap;
1412 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1413 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1415 spin_unlock(&ci->i_ceph_lock);
1419 mds = cap->session->s_mds;
1420 doutc(cl, "%p %llx.%llx mds%d (%scap %p)\n", inode,
1421 ceph_vinop(inode), mds,
1422 cap == ci->i_auth_cap ? "auth " : "", cap);
1423 spin_unlock(&ci->i_ceph_lock);
1432 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1433 doutc(cl, "chose random mds%d\n", mds);
1441 struct ceph_msg *ceph_create_session_msg(u32 op, u64 seq)
1443 struct ceph_msg *msg;
1444 struct ceph_mds_session_head *h;
1446 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1449 pr_err("ENOMEM creating session %s msg\n",
1450 ceph_session_op_name(op));
1453 h = msg->front.iov_base;
1454 h->op = cpu_to_le32(op);
1455 h->seq = cpu_to_le64(seq);
1460 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1461 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1462 static int encode_supported_features(void **p, void *end)
1464 static const size_t count = ARRAY_SIZE(feature_bits);
1468 size_t size = FEATURE_BYTES(count);
1471 if (WARN_ON_ONCE(*p + 4 + size > end))
1474 ceph_encode_32(p, size);
1475 memset(*p, 0, size);
1476 for (i = 0; i < count; i++) {
1477 bit = feature_bits[i];
1478 ((unsigned char *)(*p))[bit / 8] |= BIT(bit % 8);
1482 if (WARN_ON_ONCE(*p + 4 > end))
1485 ceph_encode_32(p, 0);
1491 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1492 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1493 static int encode_metric_spec(void **p, void *end)
1495 static const size_t count = ARRAY_SIZE(metric_bits);
1498 if (WARN_ON_ONCE(*p + 2 > end))
1501 ceph_encode_8(p, 1); /* version */
1502 ceph_encode_8(p, 1); /* compat */
1506 size_t size = METRIC_BYTES(count);
1508 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1511 /* metric spec info length */
1512 ceph_encode_32(p, 4 + size);
1515 ceph_encode_32(p, size);
1516 memset(*p, 0, size);
1517 for (i = 0; i < count; i++)
1518 ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1521 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1524 /* metric spec info length */
1525 ceph_encode_32(p, 4);
1527 ceph_encode_32(p, 0);
1534 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1535 * to include additional client metadata fields.
1537 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1539 struct ceph_msg *msg;
1540 struct ceph_mds_session_head *h;
1542 int extra_bytes = 0;
1543 int metadata_key_count = 0;
1544 struct ceph_options *opt = mdsc->fsc->client->options;
1545 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1546 struct ceph_client *cl = mdsc->fsc->client;
1551 const char* metadata[][2] = {
1552 {"hostname", mdsc->nodename},
1553 {"kernel_version", init_utsname()->release},
1554 {"entity_id", opt->name ? : ""},
1555 {"root", fsopt->server_path ? : "/"},
1559 /* Calculate serialized length of metadata */
1560 extra_bytes = 4; /* map length */
1561 for (i = 0; metadata[i][0]; ++i) {
1562 extra_bytes += 8 + strlen(metadata[i][0]) +
1563 strlen(metadata[i][1]);
1564 metadata_key_count++;
1567 /* supported feature */
1569 count = ARRAY_SIZE(feature_bits);
1571 size = FEATURE_BYTES(count);
1572 extra_bytes += 4 + size;
1576 count = ARRAY_SIZE(metric_bits);
1578 size = METRIC_BYTES(count);
1579 extra_bytes += 2 + 4 + 4 + size;
1581 /* Allocate the message */
1582 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1585 pr_err_client(cl, "ENOMEM creating session open msg\n");
1586 return ERR_PTR(-ENOMEM);
1588 p = msg->front.iov_base;
1589 end = p + msg->front.iov_len;
1592 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1593 h->seq = cpu_to_le64(seq);
1596 * Serialize client metadata into waiting buffer space, using
1597 * the format that userspace expects for map<string, string>
1599 * ClientSession messages with metadata are v4
1601 msg->hdr.version = cpu_to_le16(4);
1602 msg->hdr.compat_version = cpu_to_le16(1);
1604 /* The write pointer, following the session_head structure */
1607 /* Number of entries in the map */
1608 ceph_encode_32(&p, metadata_key_count);
1610 /* Two length-prefixed strings for each entry in the map */
1611 for (i = 0; metadata[i][0]; ++i) {
1612 size_t const key_len = strlen(metadata[i][0]);
1613 size_t const val_len = strlen(metadata[i][1]);
1615 ceph_encode_32(&p, key_len);
1616 memcpy(p, metadata[i][0], key_len);
1618 ceph_encode_32(&p, val_len);
1619 memcpy(p, metadata[i][1], val_len);
1623 ret = encode_supported_features(&p, end);
1625 pr_err_client(cl, "encode_supported_features failed!\n");
1627 return ERR_PTR(ret);
1630 ret = encode_metric_spec(&p, end);
1632 pr_err_client(cl, "encode_metric_spec failed!\n");
1634 return ERR_PTR(ret);
1637 msg->front.iov_len = p - msg->front.iov_base;
1638 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1644 * send session open request.
1646 * called under mdsc->mutex
1648 static int __open_session(struct ceph_mds_client *mdsc,
1649 struct ceph_mds_session *session)
1651 struct ceph_msg *msg;
1653 int mds = session->s_mds;
1655 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
1658 /* wait for mds to go active? */
1659 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1660 doutc(mdsc->fsc->client, "open_session to mds%d (%s)\n", mds,
1661 ceph_mds_state_name(mstate));
1662 session->s_state = CEPH_MDS_SESSION_OPENING;
1663 session->s_renew_requested = jiffies;
1665 /* send connect message */
1666 msg = create_session_open_msg(mdsc, session->s_seq);
1668 return PTR_ERR(msg);
1669 ceph_con_send(&session->s_con, msg);
1674 * open sessions for any export targets for the given mds
1676 * called under mdsc->mutex
1678 static struct ceph_mds_session *
1679 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1681 struct ceph_mds_session *session;
1684 session = __ceph_lookup_mds_session(mdsc, target);
1686 session = register_session(mdsc, target);
1687 if (IS_ERR(session))
1690 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1691 session->s_state == CEPH_MDS_SESSION_CLOSING) {
1692 ret = __open_session(mdsc, session);
1694 return ERR_PTR(ret);
1700 struct ceph_mds_session *
1701 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1703 struct ceph_mds_session *session;
1704 struct ceph_client *cl = mdsc->fsc->client;
1706 doutc(cl, "to mds%d\n", target);
1708 mutex_lock(&mdsc->mutex);
1709 session = __open_export_target_session(mdsc, target);
1710 mutex_unlock(&mdsc->mutex);
1715 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1716 struct ceph_mds_session *session)
1718 struct ceph_mds_info *mi;
1719 struct ceph_mds_session *ts;
1720 int i, mds = session->s_mds;
1721 struct ceph_client *cl = mdsc->fsc->client;
1723 if (mds >= mdsc->mdsmap->possible_max_rank)
1726 mi = &mdsc->mdsmap->m_info[mds];
1727 doutc(cl, "for mds%d (%d targets)\n", session->s_mds,
1728 mi->num_export_targets);
1730 for (i = 0; i < mi->num_export_targets; i++) {
1731 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1732 ceph_put_mds_session(ts);
1736 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1737 struct ceph_mds_session *session)
1739 mutex_lock(&mdsc->mutex);
1740 __open_export_target_sessions(mdsc, session);
1741 mutex_unlock(&mdsc->mutex);
1748 static void detach_cap_releases(struct ceph_mds_session *session,
1749 struct list_head *target)
1751 struct ceph_client *cl = session->s_mdsc->fsc->client;
1753 lockdep_assert_held(&session->s_cap_lock);
1755 list_splice_init(&session->s_cap_releases, target);
1756 session->s_num_cap_releases = 0;
1757 doutc(cl, "mds%d\n", session->s_mds);
1760 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1761 struct list_head *dispose)
1763 while (!list_empty(dispose)) {
1764 struct ceph_cap *cap;
1765 /* zero out the in-progress message */
1766 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1767 list_del(&cap->session_caps);
1768 ceph_put_cap(mdsc, cap);
1772 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1773 struct ceph_mds_session *session)
1775 struct ceph_client *cl = mdsc->fsc->client;
1776 struct ceph_mds_request *req;
1779 doutc(cl, "mds%d\n", session->s_mds);
1780 mutex_lock(&mdsc->mutex);
1781 while (!list_empty(&session->s_unsafe)) {
1782 req = list_first_entry(&session->s_unsafe,
1783 struct ceph_mds_request, r_unsafe_item);
1784 pr_warn_ratelimited_client(cl, " dropping unsafe request %llu\n",
1786 if (req->r_target_inode)
1787 mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1788 if (req->r_unsafe_dir)
1789 mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1790 __unregister_request(mdsc, req);
1792 /* zero r_attempts, so kick_requests() will re-send requests */
1793 p = rb_first(&mdsc->request_tree);
1795 req = rb_entry(p, struct ceph_mds_request, r_node);
1797 if (req->r_session &&
1798 req->r_session->s_mds == session->s_mds)
1799 req->r_attempts = 0;
1801 mutex_unlock(&mdsc->mutex);
1805 * Helper to safely iterate over all caps associated with a session, with
1806 * special care taken to handle a racing __ceph_remove_cap().
1808 * Caller must hold session s_mutex.
1810 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1811 int (*cb)(struct inode *, int mds, void *),
1814 struct ceph_client *cl = session->s_mdsc->fsc->client;
1815 struct list_head *p;
1816 struct ceph_cap *cap;
1817 struct inode *inode, *last_inode = NULL;
1818 struct ceph_cap *old_cap = NULL;
1821 doutc(cl, "%p mds%d\n", session, session->s_mds);
1822 spin_lock(&session->s_cap_lock);
1823 p = session->s_caps.next;
1824 while (p != &session->s_caps) {
1827 cap = list_entry(p, struct ceph_cap, session_caps);
1828 inode = igrab(&cap->ci->netfs.inode);
1833 session->s_cap_iterator = cap;
1835 spin_unlock(&session->s_cap_lock);
1842 ceph_put_cap(session->s_mdsc, old_cap);
1846 ret = cb(inode, mds, arg);
1849 spin_lock(&session->s_cap_lock);
1852 doutc(cl, "finishing cap %p removal\n", cap);
1853 BUG_ON(cap->session != session);
1854 cap->session = NULL;
1855 list_del_init(&cap->session_caps);
1856 session->s_nr_caps--;
1857 atomic64_dec(&session->s_mdsc->metric.total_caps);
1858 if (cap->queue_release)
1859 __ceph_queue_cap_release(session, cap);
1861 old_cap = cap; /* put_cap it w/o locks held */
1868 session->s_cap_iterator = NULL;
1869 spin_unlock(&session->s_cap_lock);
1873 ceph_put_cap(session->s_mdsc, old_cap);
1878 static int remove_session_caps_cb(struct inode *inode, int mds, void *arg)
1880 struct ceph_inode_info *ci = ceph_inode(inode);
1881 struct ceph_client *cl = ceph_inode_to_client(inode);
1882 bool invalidate = false;
1883 struct ceph_cap *cap;
1886 spin_lock(&ci->i_ceph_lock);
1887 cap = __get_cap_for_mds(ci, mds);
1889 doutc(cl, " removing cap %p, ci is %p, inode is %p\n",
1890 cap, ci, &ci->netfs.inode);
1892 iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1894 spin_unlock(&ci->i_ceph_lock);
1897 wake_up_all(&ci->i_cap_wq);
1899 ceph_queue_invalidate(inode);
1906 * caller must hold session s_mutex
1908 static void remove_session_caps(struct ceph_mds_session *session)
1910 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1911 struct super_block *sb = fsc->sb;
1914 doutc(fsc->client, "on %p\n", session);
1915 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1917 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1919 spin_lock(&session->s_cap_lock);
1920 if (session->s_nr_caps > 0) {
1921 struct inode *inode;
1922 struct ceph_cap *cap, *prev = NULL;
1923 struct ceph_vino vino;
1925 * iterate_session_caps() skips inodes that are being
1926 * deleted, we need to wait until deletions are complete.
1927 * __wait_on_freeing_inode() is designed for the job,
1928 * but it is not exported, so use lookup inode function
1931 while (!list_empty(&session->s_caps)) {
1932 cap = list_entry(session->s_caps.next,
1933 struct ceph_cap, session_caps);
1937 vino = cap->ci->i_vino;
1938 spin_unlock(&session->s_cap_lock);
1940 inode = ceph_find_inode(sb, vino);
1943 spin_lock(&session->s_cap_lock);
1947 // drop cap expires and unlock s_cap_lock
1948 detach_cap_releases(session, &dispose);
1950 BUG_ON(session->s_nr_caps > 0);
1951 BUG_ON(!list_empty(&session->s_cap_flushing));
1952 spin_unlock(&session->s_cap_lock);
1953 dispose_cap_releases(session->s_mdsc, &dispose);
1963 * wake up any threads waiting on this session's caps. if the cap is
1964 * old (didn't get renewed on the client reconnect), remove it now.
1966 * caller must hold s_mutex.
1968 static int wake_up_session_cb(struct inode *inode, int mds, void *arg)
1970 struct ceph_inode_info *ci = ceph_inode(inode);
1971 unsigned long ev = (unsigned long)arg;
1973 if (ev == RECONNECT) {
1974 spin_lock(&ci->i_ceph_lock);
1975 ci->i_wanted_max_size = 0;
1976 ci->i_requested_max_size = 0;
1977 spin_unlock(&ci->i_ceph_lock);
1978 } else if (ev == RENEWCAPS) {
1979 struct ceph_cap *cap;
1981 spin_lock(&ci->i_ceph_lock);
1982 cap = __get_cap_for_mds(ci, mds);
1983 /* mds did not re-issue stale cap */
1984 if (cap && cap->cap_gen < atomic_read(&cap->session->s_cap_gen))
1985 cap->issued = cap->implemented = CEPH_CAP_PIN;
1986 spin_unlock(&ci->i_ceph_lock);
1987 } else if (ev == FORCE_RO) {
1989 wake_up_all(&ci->i_cap_wq);
1993 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1995 struct ceph_client *cl = session->s_mdsc->fsc->client;
1997 doutc(cl, "session %p mds%d\n", session, session->s_mds);
1998 ceph_iterate_session_caps(session, wake_up_session_cb,
1999 (void *)(unsigned long)ev);
2003 * Send periodic message to MDS renewing all currently held caps. The
2004 * ack will reset the expiration for all caps from this session.
2006 * caller holds s_mutex
2008 static int send_renew_caps(struct ceph_mds_client *mdsc,
2009 struct ceph_mds_session *session)
2011 struct ceph_client *cl = mdsc->fsc->client;
2012 struct ceph_msg *msg;
2015 if (time_after_eq(jiffies, session->s_cap_ttl) &&
2016 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
2017 pr_info_client(cl, "mds%d caps stale\n", session->s_mds);
2018 session->s_renew_requested = jiffies;
2020 /* do not try to renew caps until a recovering mds has reconnected
2021 * with its clients. */
2022 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
2023 if (state < CEPH_MDS_STATE_RECONNECT) {
2024 doutc(cl, "ignoring mds%d (%s)\n", session->s_mds,
2025 ceph_mds_state_name(state));
2029 doutc(cl, "to mds%d (%s)\n", session->s_mds,
2030 ceph_mds_state_name(state));
2031 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
2032 ++session->s_renew_seq);
2035 ceph_con_send(&session->s_con, msg);
2039 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
2040 struct ceph_mds_session *session, u64 seq)
2042 struct ceph_client *cl = mdsc->fsc->client;
2043 struct ceph_msg *msg;
2045 doutc(cl, "to mds%d (%s)s seq %lld\n", session->s_mds,
2046 ceph_session_state_name(session->s_state), seq);
2047 msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
2050 ceph_con_send(&session->s_con, msg);
2056 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
2058 * Called under session->s_mutex
2060 static void renewed_caps(struct ceph_mds_client *mdsc,
2061 struct ceph_mds_session *session, int is_renew)
2063 struct ceph_client *cl = mdsc->fsc->client;
2067 spin_lock(&session->s_cap_lock);
2068 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
2070 session->s_cap_ttl = session->s_renew_requested +
2071 mdsc->mdsmap->m_session_timeout*HZ;
2074 if (time_before(jiffies, session->s_cap_ttl)) {
2075 pr_info_client(cl, "mds%d caps renewed\n",
2079 pr_info_client(cl, "mds%d caps still stale\n",
2083 doutc(cl, "mds%d ttl now %lu, was %s, now %s\n", session->s_mds,
2084 session->s_cap_ttl, was_stale ? "stale" : "fresh",
2085 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
2086 spin_unlock(&session->s_cap_lock);
2089 wake_up_session_caps(session, RENEWCAPS);
2093 * send a session close request
2095 static int request_close_session(struct ceph_mds_session *session)
2097 struct ceph_client *cl = session->s_mdsc->fsc->client;
2098 struct ceph_msg *msg;
2100 doutc(cl, "mds%d state %s seq %lld\n", session->s_mds,
2101 ceph_session_state_name(session->s_state), session->s_seq);
2102 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
2106 ceph_con_send(&session->s_con, msg);
2111 * Called with s_mutex held.
2113 static int __close_session(struct ceph_mds_client *mdsc,
2114 struct ceph_mds_session *session)
2116 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
2118 session->s_state = CEPH_MDS_SESSION_CLOSING;
2119 return request_close_session(session);
2122 static bool drop_negative_children(struct dentry *dentry)
2124 struct dentry *child;
2125 bool all_negative = true;
2127 if (!d_is_dir(dentry))
2130 spin_lock(&dentry->d_lock);
2131 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
2132 if (d_really_is_positive(child)) {
2133 all_negative = false;
2137 spin_unlock(&dentry->d_lock);
2140 shrink_dcache_parent(dentry);
2142 return all_negative;
2146 * Trim old(er) caps.
2148 * Because we can't cache an inode without one or more caps, we do
2149 * this indirectly: if a cap is unused, we prune its aliases, at which
2150 * point the inode will hopefully get dropped to.
2152 * Yes, this is a bit sloppy. Our only real goal here is to respond to
2153 * memory pressure from the MDS, though, so it needn't be perfect.
2155 static int trim_caps_cb(struct inode *inode, int mds, void *arg)
2157 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
2158 struct ceph_client *cl = mdsc->fsc->client;
2159 int *remaining = arg;
2160 struct ceph_inode_info *ci = ceph_inode(inode);
2161 int used, wanted, oissued, mine;
2162 struct ceph_cap *cap;
2164 if (*remaining <= 0)
2167 spin_lock(&ci->i_ceph_lock);
2168 cap = __get_cap_for_mds(ci, mds);
2170 spin_unlock(&ci->i_ceph_lock);
2173 mine = cap->issued | cap->implemented;
2174 used = __ceph_caps_used(ci);
2175 wanted = __ceph_caps_file_wanted(ci);
2176 oissued = __ceph_caps_issued_other(ci, cap);
2178 doutc(cl, "%p %llx.%llx cap %p mine %s oissued %s used %s wanted %s\n",
2179 inode, ceph_vinop(inode), cap, ceph_cap_string(mine),
2180 ceph_cap_string(oissued), ceph_cap_string(used),
2181 ceph_cap_string(wanted));
2182 if (cap == ci->i_auth_cap) {
2183 if (ci->i_dirty_caps || ci->i_flushing_caps ||
2184 !list_empty(&ci->i_cap_snaps))
2186 if ((used | wanted) & CEPH_CAP_ANY_WR)
2188 /* Note: it's possible that i_filelock_ref becomes non-zero
2189 * after dropping auth caps. It doesn't hurt because reply
2190 * of lock mds request will re-add auth caps. */
2191 if (atomic_read(&ci->i_filelock_ref) > 0)
2194 /* The inode has cached pages, but it's no longer used.
2195 * we can safely drop it */
2196 if (S_ISREG(inode->i_mode) &&
2197 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
2198 !(oissued & CEPH_CAP_FILE_CACHE)) {
2202 if ((used | wanted) & ~oissued & mine)
2203 goto out; /* we need these caps */
2206 /* we aren't the only cap.. just remove us */
2207 ceph_remove_cap(mdsc, cap, true);
2210 struct dentry *dentry;
2211 /* try dropping referring dentries */
2212 spin_unlock(&ci->i_ceph_lock);
2213 dentry = d_find_any_alias(inode);
2214 if (dentry && drop_negative_children(dentry)) {
2217 d_prune_aliases(inode);
2218 count = atomic_read(&inode->i_count);
2221 doutc(cl, "%p %llx.%llx cap %p pruned, count now %d\n",
2222 inode, ceph_vinop(inode), cap, count);
2230 spin_unlock(&ci->i_ceph_lock);
2235 * Trim session cap count down to some max number.
2237 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2238 struct ceph_mds_session *session,
2241 struct ceph_client *cl = mdsc->fsc->client;
2242 int trim_caps = session->s_nr_caps - max_caps;
2244 doutc(cl, "mds%d start: %d / %d, trim %d\n", session->s_mds,
2245 session->s_nr_caps, max_caps, trim_caps);
2246 if (trim_caps > 0) {
2247 int remaining = trim_caps;
2249 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2250 doutc(cl, "mds%d done: %d / %d, trimmed %d\n",
2251 session->s_mds, session->s_nr_caps, max_caps,
2252 trim_caps - remaining);
2255 ceph_flush_cap_releases(mdsc, session);
2259 static int check_caps_flush(struct ceph_mds_client *mdsc,
2262 struct ceph_client *cl = mdsc->fsc->client;
2265 spin_lock(&mdsc->cap_dirty_lock);
2266 if (!list_empty(&mdsc->cap_flush_list)) {
2267 struct ceph_cap_flush *cf =
2268 list_first_entry(&mdsc->cap_flush_list,
2269 struct ceph_cap_flush, g_list);
2270 if (cf->tid <= want_flush_tid) {
2271 doutc(cl, "still flushing tid %llu <= %llu\n",
2272 cf->tid, want_flush_tid);
2276 spin_unlock(&mdsc->cap_dirty_lock);
2281 * flush all dirty inode data to disk.
2283 * returns true if we've flushed through want_flush_tid
2285 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2288 struct ceph_client *cl = mdsc->fsc->client;
2290 doutc(cl, "want %llu\n", want_flush_tid);
2292 wait_event(mdsc->cap_flushing_wq,
2293 check_caps_flush(mdsc, want_flush_tid));
2295 doutc(cl, "ok, flushed thru %llu\n", want_flush_tid);
2299 * called under s_mutex
2301 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2302 struct ceph_mds_session *session)
2304 struct ceph_client *cl = mdsc->fsc->client;
2305 struct ceph_msg *msg = NULL;
2306 struct ceph_mds_cap_release *head;
2307 struct ceph_mds_cap_item *item;
2308 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2309 struct ceph_cap *cap;
2310 LIST_HEAD(tmp_list);
2311 int num_cap_releases;
2312 __le32 barrier, *cap_barrier;
2314 down_read(&osdc->lock);
2315 barrier = cpu_to_le32(osdc->epoch_barrier);
2316 up_read(&osdc->lock);
2318 spin_lock(&session->s_cap_lock);
2320 list_splice_init(&session->s_cap_releases, &tmp_list);
2321 num_cap_releases = session->s_num_cap_releases;
2322 session->s_num_cap_releases = 0;
2323 spin_unlock(&session->s_cap_lock);
2325 while (!list_empty(&tmp_list)) {
2327 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2328 PAGE_SIZE, GFP_NOFS, false);
2331 head = msg->front.iov_base;
2332 head->num = cpu_to_le32(0);
2333 msg->front.iov_len = sizeof(*head);
2335 msg->hdr.version = cpu_to_le16(2);
2336 msg->hdr.compat_version = cpu_to_le16(1);
2339 cap = list_first_entry(&tmp_list, struct ceph_cap,
2341 list_del(&cap->session_caps);
2344 head = msg->front.iov_base;
2345 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2347 item = msg->front.iov_base + msg->front.iov_len;
2348 item->ino = cpu_to_le64(cap->cap_ino);
2349 item->cap_id = cpu_to_le64(cap->cap_id);
2350 item->migrate_seq = cpu_to_le32(cap->mseq);
2351 item->seq = cpu_to_le32(cap->issue_seq);
2352 msg->front.iov_len += sizeof(*item);
2354 ceph_put_cap(mdsc, cap);
2356 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2357 // Append cap_barrier field
2358 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2359 *cap_barrier = barrier;
2360 msg->front.iov_len += sizeof(*cap_barrier);
2362 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2363 doutc(cl, "mds%d %p\n", session->s_mds, msg);
2364 ceph_con_send(&session->s_con, msg);
2369 BUG_ON(num_cap_releases != 0);
2371 spin_lock(&session->s_cap_lock);
2372 if (!list_empty(&session->s_cap_releases))
2374 spin_unlock(&session->s_cap_lock);
2377 // Append cap_barrier field
2378 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2379 *cap_barrier = barrier;
2380 msg->front.iov_len += sizeof(*cap_barrier);
2382 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2383 doutc(cl, "mds%d %p\n", session->s_mds, msg);
2384 ceph_con_send(&session->s_con, msg);
2388 pr_err_client(cl, "mds%d, failed to allocate message\n",
2390 spin_lock(&session->s_cap_lock);
2391 list_splice(&tmp_list, &session->s_cap_releases);
2392 session->s_num_cap_releases += num_cap_releases;
2393 spin_unlock(&session->s_cap_lock);
2396 static void ceph_cap_release_work(struct work_struct *work)
2398 struct ceph_mds_session *session =
2399 container_of(work, struct ceph_mds_session, s_cap_release_work);
2401 mutex_lock(&session->s_mutex);
2402 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2403 session->s_state == CEPH_MDS_SESSION_HUNG)
2404 ceph_send_cap_releases(session->s_mdsc, session);
2405 mutex_unlock(&session->s_mutex);
2406 ceph_put_mds_session(session);
2409 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2410 struct ceph_mds_session *session)
2412 struct ceph_client *cl = mdsc->fsc->client;
2416 ceph_get_mds_session(session);
2417 if (queue_work(mdsc->fsc->cap_wq,
2418 &session->s_cap_release_work)) {
2419 doutc(cl, "cap release work queued\n");
2421 ceph_put_mds_session(session);
2422 doutc(cl, "failed to queue cap release work\n");
2427 * caller holds session->s_cap_lock
2429 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2430 struct ceph_cap *cap)
2432 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2433 session->s_num_cap_releases++;
2435 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2436 ceph_flush_cap_releases(session->s_mdsc, session);
2439 static void ceph_cap_reclaim_work(struct work_struct *work)
2441 struct ceph_mds_client *mdsc =
2442 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2443 int ret = ceph_trim_dentries(mdsc);
2445 ceph_queue_cap_reclaim_work(mdsc);
2448 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2450 struct ceph_client *cl = mdsc->fsc->client;
2454 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2455 doutc(cl, "caps reclaim work queued\n");
2457 doutc(cl, "failed to queue caps release work\n");
2461 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2466 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2467 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2468 atomic_set(&mdsc->cap_reclaim_pending, 0);
2469 ceph_queue_cap_reclaim_work(mdsc);
2477 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2480 struct ceph_inode_info *ci = ceph_inode(dir);
2481 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2482 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2483 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2484 unsigned int num_entries;
2487 spin_lock(&ci->i_ceph_lock);
2488 num_entries = ci->i_files + ci->i_subdirs;
2489 spin_unlock(&ci->i_ceph_lock);
2490 num_entries = max(num_entries, 1U);
2491 num_entries = min(num_entries, opt->max_readdir);
2493 order = get_order(size * num_entries);
2494 while (order >= 0) {
2495 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2499 if (rinfo->dir_entries)
2503 if (!rinfo->dir_entries)
2506 num_entries = (PAGE_SIZE << order) / size;
2507 num_entries = min(num_entries, opt->max_readdir);
2509 rinfo->dir_buf_size = PAGE_SIZE << order;
2510 req->r_num_caps = num_entries + 1;
2511 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2512 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2517 * Create an mds request.
2519 struct ceph_mds_request *
2520 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2522 struct ceph_mds_request *req;
2524 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2526 return ERR_PTR(-ENOMEM);
2528 mutex_init(&req->r_fill_mutex);
2530 req->r_started = jiffies;
2531 req->r_start_latency = ktime_get();
2532 req->r_resend_mds = -1;
2533 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2534 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2536 req->r_feature_needed = -1;
2537 kref_init(&req->r_kref);
2538 RB_CLEAR_NODE(&req->r_node);
2539 INIT_LIST_HEAD(&req->r_wait);
2540 init_completion(&req->r_completion);
2541 init_completion(&req->r_safe_completion);
2542 INIT_LIST_HEAD(&req->r_unsafe_item);
2544 ktime_get_coarse_real_ts64(&req->r_stamp);
2547 req->r_direct_mode = mode;
2552 * return oldest (lowest) request, tid in request tree, 0 if none.
2554 * called under mdsc->mutex.
2556 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2558 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2560 return rb_entry(rb_first(&mdsc->request_tree),
2561 struct ceph_mds_request, r_node);
2564 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2566 return mdsc->oldest_tid;
2569 #if IS_ENABLED(CONFIG_FS_ENCRYPTION)
2570 static u8 *get_fscrypt_altname(const struct ceph_mds_request *req, u32 *plen)
2572 struct inode *dir = req->r_parent;
2573 struct dentry *dentry = req->r_dentry;
2574 u8 *cryptbuf = NULL;
2578 /* only encode if we have parent and dentry */
2579 if (!dir || !dentry)
2582 /* No-op unless this is encrypted */
2583 if (!IS_ENCRYPTED(dir))
2586 ret = ceph_fscrypt_prepare_readdir(dir);
2588 return ERR_PTR(ret);
2590 /* No key? Just ignore it. */
2591 if (!fscrypt_has_encryption_key(dir))
2594 if (!fscrypt_fname_encrypted_size(dir, dentry->d_name.len, NAME_MAX,
2597 return ERR_PTR(-ENAMETOOLONG);
2600 /* No need to append altname if name is short enough */
2601 if (len <= CEPH_NOHASH_NAME_MAX) {
2606 cryptbuf = kmalloc(len, GFP_KERNEL);
2608 return ERR_PTR(-ENOMEM);
2610 ret = fscrypt_fname_encrypt(dir, &dentry->d_name, cryptbuf, len);
2613 return ERR_PTR(ret);
2620 static u8 *get_fscrypt_altname(const struct ceph_mds_request *req, u32 *plen)
2628 * ceph_mdsc_build_path - build a path string to a given dentry
2630 * @dentry: dentry to which path should be built
2631 * @plen: returned length of string
2632 * @pbase: returned base inode number
2633 * @for_wire: is this path going to be sent to the MDS?
2635 * Build a string that represents the path to the dentry. This is mostly called
2636 * for two different purposes:
2638 * 1) we need to build a path string to send to the MDS (for_wire == true)
2639 * 2) we need a path string for local presentation (e.g. debugfs)
2640 * (for_wire == false)
2642 * The path is built in reverse, starting with the dentry. Walk back up toward
2643 * the root, building the path until the first non-snapped inode is reached
2644 * (for_wire) or the root inode is reached (!for_wire).
2646 * Encode hidden .snap dirs as a double /, i.e.
2647 * foo/.snap/bar -> foo//bar
2649 char *ceph_mdsc_build_path(struct ceph_mds_client *mdsc, struct dentry *dentry,
2650 int *plen, u64 *pbase, int for_wire)
2652 struct ceph_client *cl = mdsc->fsc->client;
2654 struct inode *inode;
2661 return ERR_PTR(-EINVAL);
2665 return ERR_PTR(-ENOMEM);
2670 seq = read_seqbegin(&rename_lock);
2673 struct dentry *parent;
2675 spin_lock(&cur->d_lock);
2676 inode = d_inode(cur);
2677 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2678 doutc(cl, "path+%d: %p SNAPDIR\n", pos, cur);
2679 spin_unlock(&cur->d_lock);
2680 parent = dget_parent(cur);
2681 } else if (for_wire && inode && dentry != cur &&
2682 ceph_snap(inode) == CEPH_NOSNAP) {
2683 spin_unlock(&cur->d_lock);
2684 pos++; /* get rid of any prepended '/' */
2686 } else if (!for_wire || !IS_ENCRYPTED(d_inode(cur->d_parent))) {
2687 pos -= cur->d_name.len;
2689 spin_unlock(&cur->d_lock);
2692 memcpy(path + pos, cur->d_name.name, cur->d_name.len);
2693 spin_unlock(&cur->d_lock);
2694 parent = dget_parent(cur);
2700 * Proactively copy name into buf, in case we need to
2703 memcpy(buf, cur->d_name.name, cur->d_name.len);
2704 len = cur->d_name.len;
2705 spin_unlock(&cur->d_lock);
2706 parent = dget_parent(cur);
2708 ret = ceph_fscrypt_prepare_readdir(d_inode(parent));
2712 return ERR_PTR(ret);
2715 if (fscrypt_has_encryption_key(d_inode(parent))) {
2716 len = ceph_encode_encrypted_fname(d_inode(parent),
2721 return ERR_PTR(len);
2729 memcpy(path + pos, buf, len);
2734 /* Are we at the root? */
2738 /* Are we out of buffer? */
2744 inode = d_inode(cur);
2745 base = inode ? ceph_ino(inode) : 0;
2748 if (read_seqretry(&rename_lock, seq))
2753 * A rename didn't occur, but somehow we didn't end up where
2754 * we thought we would. Throw a warning and try again.
2756 pr_warn_client(cl, "did not end path lookup where expected (pos = %d)\n",
2762 *plen = PATH_MAX - 1 - pos;
2763 doutc(cl, "on %p %d built %llx '%.*s'\n", dentry, d_count(dentry),
2764 base, *plen, path + pos);
2768 static int build_dentry_path(struct ceph_mds_client *mdsc, struct dentry *dentry,
2769 struct inode *dir, const char **ppath, int *ppathlen,
2770 u64 *pino, bool *pfreepath, bool parent_locked)
2776 dir = d_inode_rcu(dentry->d_parent);
2777 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP &&
2778 !IS_ENCRYPTED(dir)) {
2779 *pino = ceph_ino(dir);
2781 *ppath = dentry->d_name.name;
2782 *ppathlen = dentry->d_name.len;
2786 path = ceph_mdsc_build_path(mdsc, dentry, ppathlen, pino, 1);
2788 return PTR_ERR(path);
2794 static int build_inode_path(struct inode *inode,
2795 const char **ppath, int *ppathlen, u64 *pino,
2798 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
2799 struct dentry *dentry;
2802 if (ceph_snap(inode) == CEPH_NOSNAP) {
2803 *pino = ceph_ino(inode);
2807 dentry = d_find_alias(inode);
2808 path = ceph_mdsc_build_path(mdsc, dentry, ppathlen, pino, 1);
2811 return PTR_ERR(path);
2818 * request arguments may be specified via an inode *, a dentry *, or
2819 * an explicit ino+path.
2821 static int set_request_path_attr(struct ceph_mds_client *mdsc, struct inode *rinode,
2822 struct dentry *rdentry, struct inode *rdiri,
2823 const char *rpath, u64 rino, const char **ppath,
2824 int *pathlen, u64 *ino, bool *freepath,
2827 struct ceph_client *cl = mdsc->fsc->client;
2831 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2832 doutc(cl, " inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2834 } else if (rdentry) {
2835 r = build_dentry_path(mdsc, rdentry, rdiri, ppath, pathlen, ino,
2836 freepath, parent_locked);
2837 doutc(cl, " dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen, *ppath);
2838 } else if (rpath || rino) {
2841 *pathlen = rpath ? strlen(rpath) : 0;
2842 doutc(cl, " path %.*s\n", *pathlen, rpath);
2848 static void encode_mclientrequest_tail(void **p,
2849 const struct ceph_mds_request *req)
2851 struct ceph_timespec ts;
2854 ceph_encode_timespec64(&ts, &req->r_stamp);
2855 ceph_encode_copy(p, &ts, sizeof(ts));
2858 ceph_encode_32(p, req->r_cred->group_info->ngroups);
2859 for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2860 ceph_encode_64(p, from_kgid(&init_user_ns,
2861 req->r_cred->group_info->gid[i]));
2864 ceph_encode_32(p, req->r_altname_len);
2865 ceph_encode_copy(p, req->r_altname, req->r_altname_len);
2867 /* v6: fscrypt_auth and fscrypt_file */
2868 if (req->r_fscrypt_auth) {
2869 u32 authlen = ceph_fscrypt_auth_len(req->r_fscrypt_auth);
2871 ceph_encode_32(p, authlen);
2872 ceph_encode_copy(p, req->r_fscrypt_auth, authlen);
2874 ceph_encode_32(p, 0);
2876 if (test_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags)) {
2877 ceph_encode_32(p, sizeof(__le64));
2878 ceph_encode_64(p, req->r_fscrypt_file);
2880 ceph_encode_32(p, 0);
2884 static inline u16 mds_supported_head_version(struct ceph_mds_session *session)
2886 if (!test_bit(CEPHFS_FEATURE_32BITS_RETRY_FWD, &session->s_features))
2889 if (!test_bit(CEPHFS_FEATURE_HAS_OWNER_UIDGID, &session->s_features))
2892 return CEPH_MDS_REQUEST_HEAD_VERSION;
2895 static struct ceph_mds_request_head_legacy *
2896 find_legacy_request_head(void *p, u64 features)
2898 bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2899 struct ceph_mds_request_head_old *ohead;
2902 return (struct ceph_mds_request_head_legacy *)p;
2903 ohead = (struct ceph_mds_request_head_old *)p;
2904 return (struct ceph_mds_request_head_legacy *)&ohead->oldest_client_tid;
2908 * called under mdsc->mutex
2910 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2911 struct ceph_mds_request *req,
2912 bool drop_cap_releases)
2914 int mds = session->s_mds;
2915 struct ceph_mds_client *mdsc = session->s_mdsc;
2916 struct ceph_client *cl = mdsc->fsc->client;
2917 struct ceph_msg *msg;
2918 struct ceph_mds_request_head_legacy *lhead;
2919 const char *path1 = NULL;
2920 const char *path2 = NULL;
2921 u64 ino1 = 0, ino2 = 0;
2922 int pathlen1 = 0, pathlen2 = 0;
2923 bool freepath1 = false, freepath2 = false;
2924 struct dentry *old_dentry = NULL;
2929 bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2930 u16 request_head_version = mds_supported_head_version(session);
2931 kuid_t caller_fsuid = req->r_cred->fsuid;
2932 kgid_t caller_fsgid = req->r_cred->fsgid;
2934 ret = set_request_path_attr(mdsc, req->r_inode, req->r_dentry,
2935 req->r_parent, req->r_path1, req->r_ino1.ino,
2936 &path1, &pathlen1, &ino1, &freepath1,
2937 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2938 &req->r_req_flags));
2944 /* If r_old_dentry is set, then assume that its parent is locked */
2945 if (req->r_old_dentry &&
2946 !(req->r_old_dentry->d_flags & DCACHE_DISCONNECTED))
2947 old_dentry = req->r_old_dentry;
2948 ret = set_request_path_attr(mdsc, NULL, old_dentry,
2949 req->r_old_dentry_dir,
2950 req->r_path2, req->r_ino2.ino,
2951 &path2, &pathlen2, &ino2, &freepath2, true);
2957 req->r_altname = get_fscrypt_altname(req, &req->r_altname_len);
2958 if (IS_ERR(req->r_altname)) {
2959 msg = ERR_CAST(req->r_altname);
2960 req->r_altname = NULL;
2965 * For old cephs without supporting the 32bit retry/fwd feature
2966 * it will copy the raw memories directly when decoding the
2967 * requests. While new cephs will decode the head depending the
2968 * version member, so we need to make sure it will be compatible
2972 len = sizeof(struct ceph_mds_request_head_legacy);
2973 else if (request_head_version == 1)
2974 len = sizeof(struct ceph_mds_request_head_old);
2975 else if (request_head_version == 2)
2976 len = offsetofend(struct ceph_mds_request_head, ext_num_fwd);
2978 len = sizeof(struct ceph_mds_request_head);
2981 len += 2 * (1 + sizeof(u32) + sizeof(u64));
2982 len += pathlen1 + pathlen2;
2985 len += sizeof(struct ceph_mds_request_release) *
2986 (!!req->r_inode_drop + !!req->r_dentry_drop +
2987 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2989 if (req->r_dentry_drop)
2991 if (req->r_old_dentry_drop)
2994 /* MClientRequest tail */
2997 len += sizeof(struct ceph_timespec);
3000 len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
3002 /* alternate name */
3003 len += sizeof(u32) + req->r_altname_len;
3006 len += sizeof(u32); // fscrypt_auth
3007 if (req->r_fscrypt_auth)
3008 len += ceph_fscrypt_auth_len(req->r_fscrypt_auth);
3012 if (test_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags))
3013 len += sizeof(__le64);
3015 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
3017 msg = ERR_PTR(-ENOMEM);
3021 msg->hdr.tid = cpu_to_le64(req->r_tid);
3023 lhead = find_legacy_request_head(msg->front.iov_base,
3024 session->s_con.peer_features);
3026 if ((req->r_mnt_idmap != &nop_mnt_idmap) &&
3027 !test_bit(CEPHFS_FEATURE_HAS_OWNER_UIDGID, &session->s_features)) {
3028 WARN_ON_ONCE(!IS_CEPH_MDS_OP_NEWINODE(req->r_op));
3030 if (enable_unsafe_idmap) {
3031 pr_warn_once_client(cl,
3032 "idmapped mount is used and CEPHFS_FEATURE_HAS_OWNER_UIDGID"
3033 " is not supported by MDS. UID/GID-based restrictions may"
3034 " not work properly.\n");
3036 caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
3037 VFSUIDT_INIT(req->r_cred->fsuid));
3038 caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
3039 VFSGIDT_INIT(req->r_cred->fsgid));
3041 pr_err_ratelimited_client(cl,
3042 "idmapped mount is used and CEPHFS_FEATURE_HAS_OWNER_UIDGID"
3043 " is not supported by MDS. Fail request with -EIO.\n");
3051 * The ceph_mds_request_head_legacy didn't contain a version field, and
3052 * one was added when we moved the message version from 3->4.
3055 msg->hdr.version = cpu_to_le16(3);
3056 p = msg->front.iov_base + sizeof(*lhead);
3057 } else if (request_head_version == 1) {
3058 struct ceph_mds_request_head_old *ohead = msg->front.iov_base;
3060 msg->hdr.version = cpu_to_le16(4);
3061 ohead->version = cpu_to_le16(1);
3062 p = msg->front.iov_base + sizeof(*ohead);
3063 } else if (request_head_version == 2) {
3064 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3066 msg->hdr.version = cpu_to_le16(6);
3067 nhead->version = cpu_to_le16(2);
3069 p = msg->front.iov_base + offsetofend(struct ceph_mds_request_head, ext_num_fwd);
3071 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3075 msg->hdr.version = cpu_to_le16(6);
3076 nhead->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
3077 nhead->struct_len = cpu_to_le32(sizeof(struct ceph_mds_request_head));
3079 if (IS_CEPH_MDS_OP_NEWINODE(req->r_op)) {
3080 owner_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
3081 VFSUIDT_INIT(req->r_cred->fsuid));
3082 owner_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
3083 VFSGIDT_INIT(req->r_cred->fsgid));
3084 nhead->owner_uid = cpu_to_le32(from_kuid(&init_user_ns, owner_fsuid));
3085 nhead->owner_gid = cpu_to_le32(from_kgid(&init_user_ns, owner_fsgid));
3087 nhead->owner_uid = cpu_to_le32(-1);
3088 nhead->owner_gid = cpu_to_le32(-1);
3091 p = msg->front.iov_base + sizeof(*nhead);
3094 end = msg->front.iov_base + msg->front.iov_len;
3096 lhead->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
3097 lhead->op = cpu_to_le32(req->r_op);
3098 lhead->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
3100 lhead->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
3102 lhead->ino = cpu_to_le64(req->r_deleg_ino);
3103 lhead->args = req->r_args;
3105 ceph_encode_filepath(&p, end, ino1, path1);
3106 ceph_encode_filepath(&p, end, ino2, path2);
3108 /* make note of release offset, in case we need to replay */
3109 req->r_request_release_offset = p - msg->front.iov_base;
3113 if (req->r_inode_drop)
3114 releases += ceph_encode_inode_release(&p,
3115 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
3116 mds, req->r_inode_drop, req->r_inode_unless,
3117 req->r_op == CEPH_MDS_OP_READDIR);
3118 if (req->r_dentry_drop) {
3119 ret = ceph_encode_dentry_release(&p, req->r_dentry,
3120 req->r_parent, mds, req->r_dentry_drop,
3121 req->r_dentry_unless);
3126 if (req->r_old_dentry_drop) {
3127 ret = ceph_encode_dentry_release(&p, req->r_old_dentry,
3128 req->r_old_dentry_dir, mds,
3129 req->r_old_dentry_drop,
3130 req->r_old_dentry_unless);
3135 if (req->r_old_inode_drop)
3136 releases += ceph_encode_inode_release(&p,
3137 d_inode(req->r_old_dentry),
3138 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
3140 if (drop_cap_releases) {
3142 p = msg->front.iov_base + req->r_request_release_offset;
3145 lhead->num_releases = cpu_to_le16(releases);
3147 encode_mclientrequest_tail(&p, req);
3149 if (WARN_ON_ONCE(p > end)) {
3151 msg = ERR_PTR(-ERANGE);
3155 msg->front.iov_len = p - msg->front.iov_base;
3156 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
3158 if (req->r_pagelist) {
3159 struct ceph_pagelist *pagelist = req->r_pagelist;
3160 ceph_msg_data_add_pagelist(msg, pagelist);
3161 msg->hdr.data_len = cpu_to_le32(pagelist->length);
3163 msg->hdr.data_len = 0;
3166 msg->hdr.data_off = cpu_to_le16(0);
3170 ceph_mdsc_free_path((char *)path2, pathlen2);
3173 ceph_mdsc_free_path((char *)path1, pathlen1);
3183 * called under mdsc->mutex if error, under no mutex if
3186 static void complete_request(struct ceph_mds_client *mdsc,
3187 struct ceph_mds_request *req)
3189 req->r_end_latency = ktime_get();
3191 if (req->r_callback)
3192 req->r_callback(mdsc, req);
3193 complete_all(&req->r_completion);
3197 * called under mdsc->mutex
3199 static int __prepare_send_request(struct ceph_mds_session *session,
3200 struct ceph_mds_request *req,
3201 bool drop_cap_releases)
3203 int mds = session->s_mds;
3204 struct ceph_mds_client *mdsc = session->s_mdsc;
3205 struct ceph_client *cl = mdsc->fsc->client;
3206 struct ceph_mds_request_head_legacy *lhead;
3207 struct ceph_mds_request_head *nhead;
3208 struct ceph_msg *msg;
3209 int flags = 0, old_max_retry;
3210 bool old_version = !test_bit(CEPHFS_FEATURE_32BITS_RETRY_FWD,
3211 &session->s_features);
3214 * Avoid inifinite retrying after overflow. The client will
3215 * increase the retry count and if the MDS is old version,
3216 * so we limit to retry at most 256 times.
3218 if (req->r_attempts) {
3219 old_max_retry = sizeof_field(struct ceph_mds_request_head_old,
3221 old_max_retry = 1 << (old_max_retry * BITS_PER_BYTE);
3222 if ((old_version && req->r_attempts >= old_max_retry) ||
3223 ((uint32_t)req->r_attempts >= U32_MAX)) {
3224 pr_warn_ratelimited_client(cl, "request tid %llu seq overflow\n",
3232 struct ceph_cap *cap =
3233 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
3236 req->r_sent_on_mseq = cap->mseq;
3238 req->r_sent_on_mseq = -1;
3240 doutc(cl, "%p tid %lld %s (attempt %d)\n", req, req->r_tid,
3241 ceph_mds_op_name(req->r_op), req->r_attempts);
3243 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3247 * Replay. Do not regenerate message (and rebuild
3248 * paths, etc.); just use the original message.
3249 * Rebuilding paths will break for renames because
3250 * d_move mangles the src name.
3252 msg = req->r_request;
3253 lhead = find_legacy_request_head(msg->front.iov_base,
3254 session->s_con.peer_features);
3256 flags = le32_to_cpu(lhead->flags);
3257 flags |= CEPH_MDS_FLAG_REPLAY;
3258 lhead->flags = cpu_to_le32(flags);
3260 if (req->r_target_inode)
3261 lhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
3263 lhead->num_retry = req->r_attempts - 1;
3265 nhead = (struct ceph_mds_request_head*)msg->front.iov_base;
3266 nhead->ext_num_retry = cpu_to_le32(req->r_attempts - 1);
3269 /* remove cap/dentry releases from message */
3270 lhead->num_releases = 0;
3272 p = msg->front.iov_base + req->r_request_release_offset;
3273 encode_mclientrequest_tail(&p, req);
3275 msg->front.iov_len = p - msg->front.iov_base;
3276 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
3280 if (req->r_request) {
3281 ceph_msg_put(req->r_request);
3282 req->r_request = NULL;
3284 msg = create_request_message(session, req, drop_cap_releases);
3286 req->r_err = PTR_ERR(msg);
3287 return PTR_ERR(msg);
3289 req->r_request = msg;
3291 lhead = find_legacy_request_head(msg->front.iov_base,
3292 session->s_con.peer_features);
3293 lhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
3294 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3295 flags |= CEPH_MDS_FLAG_REPLAY;
3296 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
3297 flags |= CEPH_MDS_FLAG_ASYNC;
3299 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
3300 lhead->flags = cpu_to_le32(flags);
3301 lhead->num_fwd = req->r_num_fwd;
3302 lhead->num_retry = req->r_attempts - 1;
3304 nhead = (struct ceph_mds_request_head*)msg->front.iov_base;
3305 nhead->ext_num_fwd = cpu_to_le32(req->r_num_fwd);
3306 nhead->ext_num_retry = cpu_to_le32(req->r_attempts - 1);
3309 doutc(cl, " r_parent = %p\n", req->r_parent);
3314 * called under mdsc->mutex
3316 static int __send_request(struct ceph_mds_session *session,
3317 struct ceph_mds_request *req,
3318 bool drop_cap_releases)
3322 err = __prepare_send_request(session, req, drop_cap_releases);
3324 ceph_msg_get(req->r_request);
3325 ceph_con_send(&session->s_con, req->r_request);
3332 * send request, or put it on the appropriate wait list.
3334 static void __do_request(struct ceph_mds_client *mdsc,
3335 struct ceph_mds_request *req)
3337 struct ceph_client *cl = mdsc->fsc->client;
3338 struct ceph_mds_session *session = NULL;
3343 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3344 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
3345 __unregister_request(mdsc, req);
3349 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO) {
3350 doutc(cl, "metadata corrupted\n");
3354 if (req->r_timeout &&
3355 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
3356 doutc(cl, "timed out\n");
3360 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
3361 doutc(cl, "forced umount\n");
3365 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
3366 if (mdsc->mdsmap_err) {
3367 err = mdsc->mdsmap_err;
3368 doutc(cl, "mdsmap err %d\n", err);
3371 if (mdsc->mdsmap->m_epoch == 0) {
3372 doutc(cl, "no mdsmap, waiting for map\n");
3373 list_add(&req->r_wait, &mdsc->waiting_for_map);
3376 if (!(mdsc->fsc->mount_options->flags &
3377 CEPH_MOUNT_OPT_MOUNTWAIT) &&
3378 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
3379 err = -EHOSTUNREACH;
3384 put_request_session(req);
3386 mds = __choose_mds(mdsc, req, &random);
3388 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
3389 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
3393 doutc(cl, "no mds or not active, waiting for map\n");
3394 list_add(&req->r_wait, &mdsc->waiting_for_map);
3398 /* get, open session */
3399 session = __ceph_lookup_mds_session(mdsc, mds);
3401 session = register_session(mdsc, mds);
3402 if (IS_ERR(session)) {
3403 err = PTR_ERR(session);
3407 req->r_session = ceph_get_mds_session(session);
3409 doutc(cl, "mds%d session %p state %s\n", mds, session,
3410 ceph_session_state_name(session->s_state));
3413 * The old ceph will crash the MDSs when see unknown OPs
3415 if (req->r_feature_needed > 0 &&
3416 !test_bit(req->r_feature_needed, &session->s_features)) {
3421 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
3422 session->s_state != CEPH_MDS_SESSION_HUNG) {
3424 * We cannot queue async requests since the caps and delegated
3425 * inodes are bound to the session. Just return -EJUKEBOX and
3426 * let the caller retry a sync request in that case.
3428 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
3434 * If the session has been REJECTED, then return a hard error,
3435 * unless it's a CLEANRECOVER mount, in which case we'll queue
3436 * it to the mdsc queue.
3438 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
3439 if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
3440 list_add(&req->r_wait, &mdsc->waiting_for_map);
3446 if (session->s_state == CEPH_MDS_SESSION_NEW ||
3447 session->s_state == CEPH_MDS_SESSION_CLOSING) {
3448 err = __open_session(mdsc, session);
3451 /* retry the same mds later */
3453 req->r_resend_mds = mds;
3455 list_add(&req->r_wait, &session->s_waiting);
3460 req->r_resend_mds = -1; /* forget any previous mds hint */
3462 if (req->r_request_started == 0) /* note request start time */
3463 req->r_request_started = jiffies;
3466 * For async create we will choose the auth MDS of frag in parent
3467 * directory to send the request and ususally this works fine, but
3468 * if the migrated the dirtory to another MDS before it could handle
3469 * it the request will be forwarded.
3471 * And then the auth cap will be changed.
3473 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags) && req->r_num_fwd) {
3474 struct ceph_dentry_info *di = ceph_dentry(req->r_dentry);
3475 struct ceph_inode_info *ci;
3476 struct ceph_cap *cap;
3479 * The request maybe handled very fast and the new inode
3480 * hasn't been linked to the dentry yet. We need to wait
3481 * for the ceph_finish_async_create(), which shouldn't be
3482 * stuck too long or fail in thoery, to finish when forwarding
3485 if (!d_inode(req->r_dentry)) {
3486 err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT,
3489 mutex_lock(&req->r_fill_mutex);
3490 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3491 mutex_unlock(&req->r_fill_mutex);
3496 ci = ceph_inode(d_inode(req->r_dentry));
3498 spin_lock(&ci->i_ceph_lock);
3499 cap = ci->i_auth_cap;
3500 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE && mds != cap->mds) {
3501 doutc(cl, "session changed for auth cap %d -> %d\n",
3502 cap->session->s_mds, session->s_mds);
3504 /* Remove the auth cap from old session */
3505 spin_lock(&cap->session->s_cap_lock);
3506 cap->session->s_nr_caps--;
3507 list_del_init(&cap->session_caps);
3508 spin_unlock(&cap->session->s_cap_lock);
3510 /* Add the auth cap to the new session */
3512 cap->session = session;
3513 spin_lock(&session->s_cap_lock);
3514 session->s_nr_caps++;
3515 list_add_tail(&cap->session_caps, &session->s_caps);
3516 spin_unlock(&session->s_cap_lock);
3518 change_auth_cap_ses(ci, session);
3520 spin_unlock(&ci->i_ceph_lock);
3523 err = __send_request(session, req, false);
3526 ceph_put_mds_session(session);
3529 doutc(cl, "early error %d\n", err);
3531 complete_request(mdsc, req);
3532 __unregister_request(mdsc, req);
3538 * called under mdsc->mutex
3540 static void __wake_requests(struct ceph_mds_client *mdsc,
3541 struct list_head *head)
3543 struct ceph_client *cl = mdsc->fsc->client;
3544 struct ceph_mds_request *req;
3545 LIST_HEAD(tmp_list);
3547 list_splice_init(head, &tmp_list);
3549 while (!list_empty(&tmp_list)) {
3550 req = list_entry(tmp_list.next,
3551 struct ceph_mds_request, r_wait);
3552 list_del_init(&req->r_wait);
3553 doutc(cl, " wake request %p tid %llu\n", req,
3555 __do_request(mdsc, req);
3560 * Wake up threads with requests pending for @mds, so that they can
3561 * resubmit their requests to a possibly different mds.
3563 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
3565 struct ceph_client *cl = mdsc->fsc->client;
3566 struct ceph_mds_request *req;
3567 struct rb_node *p = rb_first(&mdsc->request_tree);
3569 doutc(cl, "kick_requests mds%d\n", mds);
3571 req = rb_entry(p, struct ceph_mds_request, r_node);
3573 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3575 if (req->r_attempts > 0)
3576 continue; /* only new requests */
3577 if (req->r_session &&
3578 req->r_session->s_mds == mds) {
3579 doutc(cl, " kicking tid %llu\n", req->r_tid);
3580 list_del_init(&req->r_wait);
3581 __do_request(mdsc, req);
3586 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
3587 struct ceph_mds_request *req)
3589 struct ceph_client *cl = mdsc->fsc->client;
3592 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
3594 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3595 if (req->r_parent) {
3596 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3597 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3598 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3599 spin_lock(&ci->i_ceph_lock);
3600 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3601 __ceph_touch_fmode(ci, mdsc, fmode);
3602 spin_unlock(&ci->i_ceph_lock);
3604 if (req->r_old_dentry_dir)
3605 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3609 err = ceph_wait_on_async_create(req->r_inode);
3611 doutc(cl, "wait for async create returned: %d\n", err);
3616 if (!err && req->r_old_inode) {
3617 err = ceph_wait_on_async_create(req->r_old_inode);
3619 doutc(cl, "wait for async create returned: %d\n", err);
3624 doutc(cl, "submit_request on %p for inode %p\n", req, dir);
3625 mutex_lock(&mdsc->mutex);
3626 __register_request(mdsc, req, dir);
3627 __do_request(mdsc, req);
3629 mutex_unlock(&mdsc->mutex);
3633 int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3634 struct ceph_mds_request *req,
3635 ceph_mds_request_wait_callback_t wait_func)
3637 struct ceph_client *cl = mdsc->fsc->client;
3641 doutc(cl, "do_request waiting\n");
3643 err = wait_func(mdsc, req);
3645 long timeleft = wait_for_completion_killable_timeout(
3647 ceph_timeout_jiffies(req->r_timeout));
3651 err = -ETIMEDOUT; /* timed out */
3653 err = timeleft; /* killed */
3655 doutc(cl, "do_request waited, got %d\n", err);
3656 mutex_lock(&mdsc->mutex);
3658 /* only abort if we didn't race with a real reply */
3659 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3660 err = le32_to_cpu(req->r_reply_info.head->result);
3661 } else if (err < 0) {
3662 doutc(cl, "aborted request %lld with %d\n", req->r_tid, err);
3665 * ensure we aren't running concurrently with
3666 * ceph_fill_trace or ceph_readdir_prepopulate, which
3667 * rely on locks (dir mutex) held by our caller.
3669 mutex_lock(&req->r_fill_mutex);
3671 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3672 mutex_unlock(&req->r_fill_mutex);
3674 if (req->r_parent &&
3675 (req->r_op & CEPH_MDS_OP_WRITE))
3676 ceph_invalidate_dir_request(req);
3681 mutex_unlock(&mdsc->mutex);
3686 * Synchrously perform an mds request. Take care of all of the
3687 * session setup, forwarding, retry details.
3689 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3691 struct ceph_mds_request *req)
3693 struct ceph_client *cl = mdsc->fsc->client;
3696 doutc(cl, "do_request on %p\n", req);
3699 err = ceph_mdsc_submit_request(mdsc, dir, req);
3701 err = ceph_mdsc_wait_request(mdsc, req, NULL);
3702 doutc(cl, "do_request %p done, result %d\n", req, err);
3707 * Invalidate dir's completeness, dentry lease state on an aborted MDS
3708 * namespace request.
3710 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3712 struct inode *dir = req->r_parent;
3713 struct inode *old_dir = req->r_old_dentry_dir;
3714 struct ceph_client *cl = req->r_mdsc->fsc->client;
3716 doutc(cl, "invalidate_dir_request %p %p (complete, lease(s))\n",
3719 ceph_dir_clear_complete(dir);
3721 ceph_dir_clear_complete(old_dir);
3723 ceph_invalidate_dentry_lease(req->r_dentry);
3724 if (req->r_old_dentry)
3725 ceph_invalidate_dentry_lease(req->r_old_dentry);
3731 * We take the session mutex and parse and process the reply immediately.
3732 * This preserves the logical ordering of replies, capabilities, etc., sent
3733 * by the MDS as they are applied to our local cache.
3735 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3737 struct ceph_mds_client *mdsc = session->s_mdsc;
3738 struct ceph_client *cl = mdsc->fsc->client;
3739 struct ceph_mds_request *req;
3740 struct ceph_mds_reply_head *head = msg->front.iov_base;
3741 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
3742 struct ceph_snap_realm *realm;
3745 int mds = session->s_mds;
3746 bool close_sessions = false;
3748 if (msg->front.iov_len < sizeof(*head)) {
3749 pr_err_client(cl, "got corrupt (short) reply\n");
3754 /* get request, session */
3755 tid = le64_to_cpu(msg->hdr.tid);
3756 mutex_lock(&mdsc->mutex);
3757 req = lookup_get_request(mdsc, tid);
3759 doutc(cl, "on unknown tid %llu\n", tid);
3760 mutex_unlock(&mdsc->mutex);
3763 doutc(cl, "handle_reply %p\n", req);
3765 /* correct session? */
3766 if (req->r_session != session) {
3767 pr_err_client(cl, "got %llu on session mds%d not mds%d\n",
3768 tid, session->s_mds,
3769 req->r_session ? req->r_session->s_mds : -1);
3770 mutex_unlock(&mdsc->mutex);
3775 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3776 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3777 pr_warn_client(cl, "got a dup %s reply on %llu from mds%d\n",
3778 head->safe ? "safe" : "unsafe", tid, mds);
3779 mutex_unlock(&mdsc->mutex);
3782 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3783 pr_warn_client(cl, "got unsafe after safe on %llu from mds%d\n",
3785 mutex_unlock(&mdsc->mutex);
3789 result = le32_to_cpu(head->result);
3792 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3793 __unregister_request(mdsc, req);
3795 /* last request during umount? */
3796 if (mdsc->stopping && !__get_oldest_req(mdsc))
3797 complete_all(&mdsc->safe_umount_waiters);
3799 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3801 * We already handled the unsafe response, now do the
3802 * cleanup. No need to examine the response; the MDS
3803 * doesn't include any result info in the safe
3804 * response. And even if it did, there is nothing
3805 * useful we could do with a revised return value.
3807 doutc(cl, "got safe reply %llu, mds%d\n", tid, mds);
3809 mutex_unlock(&mdsc->mutex);
3813 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3814 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3817 doutc(cl, "tid %lld result %d\n", tid, result);
3818 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3819 err = parse_reply_info(session, msg, req, (u64)-1);
3821 err = parse_reply_info(session, msg, req,
3822 session->s_con.peer_features);
3823 mutex_unlock(&mdsc->mutex);
3825 /* Must find target inode outside of mutexes to avoid deadlocks */
3826 rinfo = &req->r_reply_info;
3827 if ((err >= 0) && rinfo->head->is_target) {
3828 struct inode *in = xchg(&req->r_new_inode, NULL);
3829 struct ceph_vino tvino = {
3830 .ino = le64_to_cpu(rinfo->targeti.in->ino),
3831 .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3835 * If we ended up opening an existing inode, discard
3838 if (req->r_op == CEPH_MDS_OP_CREATE &&
3839 !req->r_reply_info.has_create_ino) {
3840 /* This should never happen on an async create */
3841 WARN_ON_ONCE(req->r_deleg_ino);
3846 in = ceph_get_inode(mdsc->fsc->sb, tvino, in);
3849 mutex_lock(&session->s_mutex);
3852 req->r_target_inode = in;
3855 mutex_lock(&session->s_mutex);
3857 pr_err_client(cl, "got corrupt reply mds%d(tid:%lld)\n",
3865 if (rinfo->snapblob_len) {
3866 down_write(&mdsc->snap_rwsem);
3867 err = ceph_update_snap_trace(mdsc, rinfo->snapblob,
3868 rinfo->snapblob + rinfo->snapblob_len,
3869 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3872 up_write(&mdsc->snap_rwsem);
3873 close_sessions = true;
3878 downgrade_write(&mdsc->snap_rwsem);
3880 down_read(&mdsc->snap_rwsem);
3883 /* insert trace into our cache */
3884 mutex_lock(&req->r_fill_mutex);
3885 current->journal_info = req;
3886 err = ceph_fill_trace(mdsc->fsc->sb, req);
3888 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3889 req->r_op == CEPH_MDS_OP_LSSNAP))
3890 err = ceph_readdir_prepopulate(req, req->r_session);
3892 current->journal_info = NULL;
3893 mutex_unlock(&req->r_fill_mutex);
3895 up_read(&mdsc->snap_rwsem);
3897 ceph_put_snap_realm(mdsc, realm);
3900 if (req->r_target_inode &&
3901 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3902 struct ceph_inode_info *ci =
3903 ceph_inode(req->r_target_inode);
3904 spin_lock(&ci->i_unsafe_lock);
3905 list_add_tail(&req->r_unsafe_target_item,
3906 &ci->i_unsafe_iops);
3907 spin_unlock(&ci->i_unsafe_lock);
3910 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3913 mutex_lock(&mdsc->mutex);
3914 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3918 req->r_reply = ceph_msg_get(msg);
3919 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3922 doutc(cl, "reply arrived after request %lld was aborted\n", tid);
3924 mutex_unlock(&mdsc->mutex);
3926 mutex_unlock(&session->s_mutex);
3928 /* kick calling process */
3929 complete_request(mdsc, req);
3931 ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3932 req->r_end_latency, err);
3934 ceph_mdsc_put_request(req);
3936 /* Defer closing the sessions after s_mutex lock being released */
3938 ceph_mdsc_close_sessions(mdsc);
3945 * handle mds notification that our request has been forwarded.
3947 static void handle_forward(struct ceph_mds_client *mdsc,
3948 struct ceph_mds_session *session,
3949 struct ceph_msg *msg)
3951 struct ceph_client *cl = mdsc->fsc->client;
3952 struct ceph_mds_request *req;
3953 u64 tid = le64_to_cpu(msg->hdr.tid);
3957 void *p = msg->front.iov_base;
3958 void *end = p + msg->front.iov_len;
3959 bool aborted = false;
3961 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3962 next_mds = ceph_decode_32(&p);
3963 fwd_seq = ceph_decode_32(&p);
3965 mutex_lock(&mdsc->mutex);
3966 req = lookup_get_request(mdsc, tid);
3968 mutex_unlock(&mdsc->mutex);
3969 doutc(cl, "forward tid %llu to mds%d - req dne\n", tid, next_mds);
3970 return; /* dup reply? */
3973 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3974 doutc(cl, "forward tid %llu aborted, unregistering\n", tid);
3975 __unregister_request(mdsc, req);
3976 } else if (fwd_seq <= req->r_num_fwd || (uint32_t)fwd_seq >= U32_MAX) {
3978 * Avoid inifinite retrying after overflow.
3980 * The MDS will increase the fwd count and in client side
3981 * if the num_fwd is less than the one saved in request
3982 * that means the MDS is an old version and overflowed of
3985 mutex_lock(&req->r_fill_mutex);
3986 req->r_err = -EMULTIHOP;
3987 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3988 mutex_unlock(&req->r_fill_mutex);
3990 pr_warn_ratelimited_client(cl, "forward tid %llu seq overflow\n",
3993 /* resend. forward race not possible; mds would drop */
3994 doutc(cl, "forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3996 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3997 req->r_attempts = 0;
3998 req->r_num_fwd = fwd_seq;
3999 req->r_resend_mds = next_mds;
4000 put_request_session(req);
4001 __do_request(mdsc, req);
4003 mutex_unlock(&mdsc->mutex);
4005 /* kick calling process */
4007 complete_request(mdsc, req);
4008 ceph_mdsc_put_request(req);
4012 pr_err_client(cl, "decode error err=%d\n", err);
4016 static int __decode_session_metadata(void **p, void *end,
4019 /* map<string,string> */
4022 ceph_decode_32_safe(p, end, n, bad);
4025 ceph_decode_32_safe(p, end, len, bad);
4026 ceph_decode_need(p, end, len, bad);
4027 err_str = !strncmp(*p, "error_string", len);
4029 ceph_decode_32_safe(p, end, len, bad);
4030 ceph_decode_need(p, end, len, bad);
4032 * Match "blocklisted (blacklisted)" from newer MDSes,
4033 * or "blacklisted" from older MDSes.
4035 if (err_str && strnstr(*p, "blacklisted", len))
4036 *blocklisted = true;
4045 * handle a mds session control message
4047 static void handle_session(struct ceph_mds_session *session,
4048 struct ceph_msg *msg)
4050 struct ceph_mds_client *mdsc = session->s_mdsc;
4051 struct ceph_client *cl = mdsc->fsc->client;
4052 int mds = session->s_mds;
4053 int msg_version = le16_to_cpu(msg->hdr.version);
4054 void *p = msg->front.iov_base;
4055 void *end = p + msg->front.iov_len;
4056 struct ceph_mds_session_head *h;
4058 u64 seq, features = 0;
4060 bool blocklisted = false;
4063 ceph_decode_need(&p, end, sizeof(*h), bad);
4067 op = le32_to_cpu(h->op);
4068 seq = le64_to_cpu(h->seq);
4070 if (msg_version >= 3) {
4072 /* version >= 2 and < 5, decode metadata, skip otherwise
4073 * as it's handled via flags.
4075 if (msg_version >= 5)
4076 ceph_decode_skip_map(&p, end, string, string, bad);
4077 else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
4080 /* version >= 3, feature bits */
4081 ceph_decode_32_safe(&p, end, len, bad);
4083 ceph_decode_64_safe(&p, end, features, bad);
4084 p += len - sizeof(features);
4088 if (msg_version >= 5) {
4092 ceph_decode_skip_16(&p, end, bad); /* struct_v, struct_cv */
4093 ceph_decode_32_safe(&p, end, len, bad); /* len */
4094 ceph_decode_skip_n(&p, end, len, bad); /* metric_spec */
4096 /* version >= 5, flags */
4097 ceph_decode_32_safe(&p, end, flags, bad);
4098 if (flags & CEPH_SESSION_BLOCKLISTED) {
4099 pr_warn_client(cl, "mds%d session blocklisted\n",
4105 mutex_lock(&mdsc->mutex);
4106 if (op == CEPH_SESSION_CLOSE) {
4107 ceph_get_mds_session(session);
4108 __unregister_session(mdsc, session);
4110 /* FIXME: this ttl calculation is generous */
4111 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
4112 mutex_unlock(&mdsc->mutex);
4114 mutex_lock(&session->s_mutex);
4116 doutc(cl, "mds%d %s %p state %s seq %llu\n", mds,
4117 ceph_session_op_name(op), session,
4118 ceph_session_state_name(session->s_state), seq);
4120 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
4121 session->s_state = CEPH_MDS_SESSION_OPEN;
4122 pr_info_client(cl, "mds%d came back\n", session->s_mds);
4126 case CEPH_SESSION_OPEN:
4127 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
4128 pr_info_client(cl, "mds%d reconnect success\n",
4131 session->s_features = features;
4132 if (session->s_state == CEPH_MDS_SESSION_OPEN) {
4133 pr_notice_client(cl, "mds%d is already opened\n",
4136 session->s_state = CEPH_MDS_SESSION_OPEN;
4137 renewed_caps(mdsc, session, 0);
4138 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT,
4139 &session->s_features))
4140 metric_schedule_delayed(&mdsc->metric);
4144 * The connection maybe broken and the session in client
4145 * side has been reinitialized, need to update the seq
4148 if (!session->s_seq && seq)
4149 session->s_seq = seq;
4153 __close_session(mdsc, session);
4156 case CEPH_SESSION_RENEWCAPS:
4157 if (session->s_renew_seq == seq)
4158 renewed_caps(mdsc, session, 1);
4161 case CEPH_SESSION_CLOSE:
4162 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
4163 pr_info_client(cl, "mds%d reconnect denied\n",
4165 session->s_state = CEPH_MDS_SESSION_CLOSED;
4166 cleanup_session_requests(mdsc, session);
4167 remove_session_caps(session);
4168 wake = 2; /* for good measure */
4169 wake_up_all(&mdsc->session_close_wq);
4172 case CEPH_SESSION_STALE:
4173 pr_info_client(cl, "mds%d caps went stale, renewing\n",
4175 atomic_inc(&session->s_cap_gen);
4176 session->s_cap_ttl = jiffies - 1;
4177 send_renew_caps(mdsc, session);
4180 case CEPH_SESSION_RECALL_STATE:
4181 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
4184 case CEPH_SESSION_FLUSHMSG:
4185 /* flush cap releases */
4186 spin_lock(&session->s_cap_lock);
4187 if (session->s_num_cap_releases)
4188 ceph_flush_cap_releases(mdsc, session);
4189 spin_unlock(&session->s_cap_lock);
4191 send_flushmsg_ack(mdsc, session, seq);
4194 case CEPH_SESSION_FORCE_RO:
4195 doutc(cl, "force_session_readonly %p\n", session);
4196 spin_lock(&session->s_cap_lock);
4197 session->s_readonly = true;
4198 spin_unlock(&session->s_cap_lock);
4199 wake_up_session_caps(session, FORCE_RO);
4202 case CEPH_SESSION_REJECT:
4203 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
4204 pr_info_client(cl, "mds%d rejected session\n",
4206 session->s_state = CEPH_MDS_SESSION_REJECTED;
4207 cleanup_session_requests(mdsc, session);
4208 remove_session_caps(session);
4210 mdsc->fsc->blocklisted = true;
4211 wake = 2; /* for good measure */
4215 pr_err_client(cl, "bad op %d mds%d\n", op, mds);
4219 mutex_unlock(&session->s_mutex);
4221 mutex_lock(&mdsc->mutex);
4222 __wake_requests(mdsc, &session->s_waiting);
4224 kick_requests(mdsc, mds);
4225 mutex_unlock(&mdsc->mutex);
4227 if (op == CEPH_SESSION_CLOSE)
4228 ceph_put_mds_session(session);
4232 pr_err_client(cl, "corrupt message mds%d len %d\n", mds,
4233 (int)msg->front.iov_len);
4238 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
4240 struct ceph_client *cl = req->r_mdsc->fsc->client;
4243 dcaps = xchg(&req->r_dir_caps, 0);
4245 doutc(cl, "releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
4246 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
4250 void ceph_mdsc_release_dir_caps_async(struct ceph_mds_request *req)
4252 struct ceph_client *cl = req->r_mdsc->fsc->client;
4255 dcaps = xchg(&req->r_dir_caps, 0);
4257 doutc(cl, "releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
4258 ceph_put_cap_refs_async(ceph_inode(req->r_parent), dcaps);
4263 * called under session->mutex.
4265 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
4266 struct ceph_mds_session *session)
4268 struct ceph_mds_request *req, *nreq;
4271 doutc(mdsc->fsc->client, "mds%d\n", session->s_mds);
4273 mutex_lock(&mdsc->mutex);
4274 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
4275 __send_request(session, req, true);
4278 * also re-send old requests when MDS enters reconnect stage. So that MDS
4279 * can process completed request in clientreplay stage.
4281 p = rb_first(&mdsc->request_tree);
4283 req = rb_entry(p, struct ceph_mds_request, r_node);
4285 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
4287 if (req->r_attempts == 0)
4288 continue; /* only old requests */
4289 if (!req->r_session)
4291 if (req->r_session->s_mds != session->s_mds)
4294 ceph_mdsc_release_dir_caps_async(req);
4296 __send_request(session, req, true);
4298 mutex_unlock(&mdsc->mutex);
4301 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
4303 struct ceph_msg *reply;
4304 struct ceph_pagelist *_pagelist;
4309 if (!recon_state->allow_multi)
4312 /* can't handle message that contains both caps and realm */
4313 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
4315 /* pre-allocate new pagelist */
4316 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
4320 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4324 /* placeholder for nr_caps */
4325 err = ceph_pagelist_encode_32(_pagelist, 0);
4329 if (recon_state->nr_caps) {
4330 /* currently encoding caps */
4331 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
4335 /* placeholder for nr_realms (currently encoding relams) */
4336 err = ceph_pagelist_encode_32(_pagelist, 0);
4341 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
4345 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
4346 addr = kmap_atomic(page);
4347 if (recon_state->nr_caps) {
4348 /* currently encoding caps */
4349 *addr = cpu_to_le32(recon_state->nr_caps);
4351 /* currently encoding relams */
4352 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
4354 kunmap_atomic(addr);
4356 reply->hdr.version = cpu_to_le16(5);
4357 reply->hdr.compat_version = cpu_to_le16(4);
4359 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
4360 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
4362 ceph_con_send(&recon_state->session->s_con, reply);
4363 ceph_pagelist_release(recon_state->pagelist);
4365 recon_state->pagelist = _pagelist;
4366 recon_state->nr_caps = 0;
4367 recon_state->nr_realms = 0;
4368 recon_state->msg_version = 5;
4371 ceph_msg_put(reply);
4373 ceph_pagelist_release(_pagelist);
4377 static struct dentry* d_find_primary(struct inode *inode)
4379 struct dentry *alias, *dn = NULL;
4381 if (hlist_empty(&inode->i_dentry))
4384 spin_lock(&inode->i_lock);
4385 if (hlist_empty(&inode->i_dentry))
4388 if (S_ISDIR(inode->i_mode)) {
4389 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
4390 if (!IS_ROOT(alias))
4395 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
4396 spin_lock(&alias->d_lock);
4397 if (!d_unhashed(alias) &&
4398 (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
4399 dn = dget_dlock(alias);
4401 spin_unlock(&alias->d_lock);
4406 spin_unlock(&inode->i_lock);
4411 * Encode information about a cap for a reconnect with the MDS.
4413 static int reconnect_caps_cb(struct inode *inode, int mds, void *arg)
4415 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
4416 struct ceph_client *cl = ceph_inode_to_client(inode);
4418 struct ceph_mds_cap_reconnect v2;
4419 struct ceph_mds_cap_reconnect_v1 v1;
4421 struct ceph_inode_info *ci = ceph_inode(inode);
4422 struct ceph_reconnect_state *recon_state = arg;
4423 struct ceph_pagelist *pagelist = recon_state->pagelist;
4424 struct dentry *dentry;
4425 struct ceph_cap *cap;
4427 int pathlen = 0, err;
4431 dentry = d_find_primary(inode);
4433 /* set pathbase to parent dir when msg_version >= 2 */
4434 path = ceph_mdsc_build_path(mdsc, dentry, &pathlen, &pathbase,
4435 recon_state->msg_version >= 2);
4438 err = PTR_ERR(path);
4446 spin_lock(&ci->i_ceph_lock);
4447 cap = __get_cap_for_mds(ci, mds);
4449 spin_unlock(&ci->i_ceph_lock);
4453 doutc(cl, " adding %p ino %llx.%llx cap %p %lld %s\n", inode,
4454 ceph_vinop(inode), cap, cap->cap_id,
4455 ceph_cap_string(cap->issued));
4457 cap->seq = 0; /* reset cap seq */
4458 cap->issue_seq = 0; /* and issue_seq */
4459 cap->mseq = 0; /* and migrate_seq */
4460 cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
4462 /* These are lost when the session goes away */
4463 if (S_ISDIR(inode->i_mode)) {
4464 if (cap->issued & CEPH_CAP_DIR_CREATE) {
4465 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
4466 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
4468 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
4471 if (recon_state->msg_version >= 2) {
4472 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
4473 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
4474 rec.v2.issued = cpu_to_le32(cap->issued);
4475 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4476 rec.v2.pathbase = cpu_to_le64(pathbase);
4477 rec.v2.flock_len = (__force __le32)
4478 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
4480 struct timespec64 ts;
4482 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
4483 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
4484 rec.v1.issued = cpu_to_le32(cap->issued);
4485 rec.v1.size = cpu_to_le64(i_size_read(inode));
4486 ts = inode_get_mtime(inode);
4487 ceph_encode_timespec64(&rec.v1.mtime, &ts);
4488 ts = inode_get_atime(inode);
4489 ceph_encode_timespec64(&rec.v1.atime, &ts);
4490 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4491 rec.v1.pathbase = cpu_to_le64(pathbase);
4494 if (list_empty(&ci->i_cap_snaps)) {
4495 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
4497 struct ceph_cap_snap *capsnap =
4498 list_first_entry(&ci->i_cap_snaps,
4499 struct ceph_cap_snap, ci_item);
4500 snap_follows = capsnap->follows;
4502 spin_unlock(&ci->i_ceph_lock);
4504 if (recon_state->msg_version >= 2) {
4505 int num_fcntl_locks, num_flock_locks;
4506 struct ceph_filelock *flocks = NULL;
4507 size_t struct_len, total_len = sizeof(u64);
4511 if (rec.v2.flock_len) {
4512 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
4514 num_fcntl_locks = 0;
4515 num_flock_locks = 0;
4517 if (num_fcntl_locks + num_flock_locks > 0) {
4518 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
4519 sizeof(struct ceph_filelock),
4525 err = ceph_encode_locks_to_buffer(inode, flocks,
4540 if (recon_state->msg_version >= 3) {
4541 /* version, compat_version and struct_len */
4542 total_len += 2 * sizeof(u8) + sizeof(u32);
4546 * number of encoded locks is stable, so copy to pagelist
4548 struct_len = 2 * sizeof(u32) +
4549 (num_fcntl_locks + num_flock_locks) *
4550 sizeof(struct ceph_filelock);
4551 rec.v2.flock_len = cpu_to_le32(struct_len);
4553 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
4556 struct_len += sizeof(u64); /* snap_follows */
4558 total_len += struct_len;
4560 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
4561 err = send_reconnect_partial(recon_state);
4563 goto out_freeflocks;
4564 pagelist = recon_state->pagelist;
4567 err = ceph_pagelist_reserve(pagelist, total_len);
4569 goto out_freeflocks;
4571 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4572 if (recon_state->msg_version >= 3) {
4573 ceph_pagelist_encode_8(pagelist, struct_v);
4574 ceph_pagelist_encode_8(pagelist, 1);
4575 ceph_pagelist_encode_32(pagelist, struct_len);
4577 ceph_pagelist_encode_string(pagelist, path, pathlen);
4578 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
4579 ceph_locks_to_pagelist(flocks, pagelist,
4580 num_fcntl_locks, num_flock_locks);
4582 ceph_pagelist_encode_64(pagelist, snap_follows);
4586 err = ceph_pagelist_reserve(pagelist,
4587 sizeof(u64) + sizeof(u32) +
4588 pathlen + sizeof(rec.v1));
4592 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4593 ceph_pagelist_encode_string(pagelist, path, pathlen);
4594 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
4598 ceph_mdsc_free_path(path, pathlen);
4600 recon_state->nr_caps++;
4604 static int encode_snap_realms(struct ceph_mds_client *mdsc,
4605 struct ceph_reconnect_state *recon_state)
4608 struct ceph_pagelist *pagelist = recon_state->pagelist;
4609 struct ceph_client *cl = mdsc->fsc->client;
4612 if (recon_state->msg_version >= 4) {
4613 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
4619 * snaprealms. we provide mds with the ino, seq (version), and
4620 * parent for all of our realms. If the mds has any newer info,
4623 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
4624 struct ceph_snap_realm *realm =
4625 rb_entry(p, struct ceph_snap_realm, node);
4626 struct ceph_mds_snaprealm_reconnect sr_rec;
4628 if (recon_state->msg_version >= 4) {
4629 size_t need = sizeof(u8) * 2 + sizeof(u32) +
4632 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
4633 err = send_reconnect_partial(recon_state);
4636 pagelist = recon_state->pagelist;
4639 err = ceph_pagelist_reserve(pagelist, need);
4643 ceph_pagelist_encode_8(pagelist, 1);
4644 ceph_pagelist_encode_8(pagelist, 1);
4645 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
4648 doutc(cl, " adding snap realm %llx seq %lld parent %llx\n",
4649 realm->ino, realm->seq, realm->parent_ino);
4650 sr_rec.ino = cpu_to_le64(realm->ino);
4651 sr_rec.seq = cpu_to_le64(realm->seq);
4652 sr_rec.parent = cpu_to_le64(realm->parent_ino);
4654 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
4658 recon_state->nr_realms++;
4666 * If an MDS fails and recovers, clients need to reconnect in order to
4667 * reestablish shared state. This includes all caps issued through
4668 * this session _and_ the snap_realm hierarchy. Because it's not
4669 * clear which snap realms the mds cares about, we send everything we
4670 * know about.. that ensures we'll then get any new info the
4671 * recovering MDS might have.
4673 * This is a relatively heavyweight operation, but it's rare.
4675 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
4676 struct ceph_mds_session *session)
4678 struct ceph_client *cl = mdsc->fsc->client;
4679 struct ceph_msg *reply;
4680 int mds = session->s_mds;
4682 struct ceph_reconnect_state recon_state = {
4687 pr_info_client(cl, "mds%d reconnect start\n", mds);
4689 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4690 if (!recon_state.pagelist)
4691 goto fail_nopagelist;
4693 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4697 xa_destroy(&session->s_delegated_inos);
4699 mutex_lock(&session->s_mutex);
4700 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4703 doutc(cl, "session %p state %s\n", session,
4704 ceph_session_state_name(session->s_state));
4706 atomic_inc(&session->s_cap_gen);
4708 spin_lock(&session->s_cap_lock);
4709 /* don't know if session is readonly */
4710 session->s_readonly = 0;
4712 * notify __ceph_remove_cap() that we are composing cap reconnect.
4713 * If a cap get released before being added to the cap reconnect,
4714 * __ceph_remove_cap() should skip queuing cap release.
4716 session->s_cap_reconnect = 1;
4717 /* drop old cap expires; we're about to reestablish that state */
4718 detach_cap_releases(session, &dispose);
4719 spin_unlock(&session->s_cap_lock);
4720 dispose_cap_releases(mdsc, &dispose);
4722 /* trim unused caps to reduce MDS's cache rejoin time */
4723 if (mdsc->fsc->sb->s_root)
4724 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4726 ceph_con_close(&session->s_con);
4727 ceph_con_open(&session->s_con,
4728 CEPH_ENTITY_TYPE_MDS, mds,
4729 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4731 /* replay unsafe requests */
4732 replay_unsafe_requests(mdsc, session);
4734 ceph_early_kick_flushing_caps(mdsc, session);
4736 down_read(&mdsc->snap_rwsem);
4738 /* placeholder for nr_caps */
4739 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4743 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4744 recon_state.msg_version = 3;
4745 recon_state.allow_multi = true;
4746 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4747 recon_state.msg_version = 3;
4749 recon_state.msg_version = 2;
4751 /* trsaverse this session's caps */
4752 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4754 spin_lock(&session->s_cap_lock);
4755 session->s_cap_reconnect = 0;
4756 spin_unlock(&session->s_cap_lock);
4761 /* check if all realms can be encoded into current message */
4762 if (mdsc->num_snap_realms) {
4764 recon_state.pagelist->length +
4765 mdsc->num_snap_realms *
4766 sizeof(struct ceph_mds_snaprealm_reconnect);
4767 if (recon_state.msg_version >= 4) {
4768 /* number of realms */
4769 total_len += sizeof(u32);
4770 /* version, compat_version and struct_len */
4771 total_len += mdsc->num_snap_realms *
4772 (2 * sizeof(u8) + sizeof(u32));
4774 if (total_len > RECONNECT_MAX_SIZE) {
4775 if (!recon_state.allow_multi) {
4779 if (recon_state.nr_caps) {
4780 err = send_reconnect_partial(&recon_state);
4784 recon_state.msg_version = 5;
4788 err = encode_snap_realms(mdsc, &recon_state);
4792 if (recon_state.msg_version >= 5) {
4793 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4798 if (recon_state.nr_caps || recon_state.nr_realms) {
4800 list_first_entry(&recon_state.pagelist->head,
4802 __le32 *addr = kmap_atomic(page);
4803 if (recon_state.nr_caps) {
4804 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4805 *addr = cpu_to_le32(recon_state.nr_caps);
4806 } else if (recon_state.msg_version >= 4) {
4807 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4809 kunmap_atomic(addr);
4812 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4813 if (recon_state.msg_version >= 4)
4814 reply->hdr.compat_version = cpu_to_le16(4);
4816 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4817 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4819 ceph_con_send(&session->s_con, reply);
4821 mutex_unlock(&session->s_mutex);
4823 mutex_lock(&mdsc->mutex);
4824 __wake_requests(mdsc, &session->s_waiting);
4825 mutex_unlock(&mdsc->mutex);
4827 up_read(&mdsc->snap_rwsem);
4828 ceph_pagelist_release(recon_state.pagelist);
4832 ceph_msg_put(reply);
4833 up_read(&mdsc->snap_rwsem);
4834 mutex_unlock(&session->s_mutex);
4836 ceph_pagelist_release(recon_state.pagelist);
4838 pr_err_client(cl, "error %d preparing reconnect for mds%d\n",
4845 * compare old and new mdsmaps, kicking requests
4846 * and closing out old connections as necessary
4848 * called under mdsc->mutex.
4850 static void check_new_map(struct ceph_mds_client *mdsc,
4851 struct ceph_mdsmap *newmap,
4852 struct ceph_mdsmap *oldmap)
4855 int oldstate, newstate;
4856 struct ceph_mds_session *s;
4857 unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
4858 struct ceph_client *cl = mdsc->fsc->client;
4860 doutc(cl, "new %u old %u\n", newmap->m_epoch, oldmap->m_epoch);
4862 if (newmap->m_info) {
4863 for (i = 0; i < newmap->possible_max_rank; i++) {
4864 for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
4865 set_bit(newmap->m_info[i].export_targets[j], targets);
4869 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4870 if (!mdsc->sessions[i])
4872 s = mdsc->sessions[i];
4873 oldstate = ceph_mdsmap_get_state(oldmap, i);
4874 newstate = ceph_mdsmap_get_state(newmap, i);
4876 doutc(cl, "mds%d state %s%s -> %s%s (session %s)\n",
4877 i, ceph_mds_state_name(oldstate),
4878 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4879 ceph_mds_state_name(newstate),
4880 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4881 ceph_session_state_name(s->s_state));
4883 if (i >= newmap->possible_max_rank) {
4884 /* force close session for stopped mds */
4885 ceph_get_mds_session(s);
4886 __unregister_session(mdsc, s);
4887 __wake_requests(mdsc, &s->s_waiting);
4888 mutex_unlock(&mdsc->mutex);
4890 mutex_lock(&s->s_mutex);
4891 cleanup_session_requests(mdsc, s);
4892 remove_session_caps(s);
4893 mutex_unlock(&s->s_mutex);
4895 ceph_put_mds_session(s);
4897 mutex_lock(&mdsc->mutex);
4898 kick_requests(mdsc, i);
4902 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4903 ceph_mdsmap_get_addr(newmap, i),
4904 sizeof(struct ceph_entity_addr))) {
4906 mutex_unlock(&mdsc->mutex);
4907 mutex_lock(&s->s_mutex);
4908 mutex_lock(&mdsc->mutex);
4909 ceph_con_close(&s->s_con);
4910 mutex_unlock(&s->s_mutex);
4911 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4912 } else if (oldstate == newstate) {
4913 continue; /* nothing new with this mds */
4919 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4920 newstate >= CEPH_MDS_STATE_RECONNECT) {
4921 mutex_unlock(&mdsc->mutex);
4922 clear_bit(i, targets);
4923 send_mds_reconnect(mdsc, s);
4924 mutex_lock(&mdsc->mutex);
4928 * kick request on any mds that has gone active.
4930 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4931 newstate >= CEPH_MDS_STATE_ACTIVE) {
4932 if (oldstate != CEPH_MDS_STATE_CREATING &&
4933 oldstate != CEPH_MDS_STATE_STARTING)
4934 pr_info_client(cl, "mds%d recovery completed\n",
4936 kick_requests(mdsc, i);
4937 mutex_unlock(&mdsc->mutex);
4938 mutex_lock(&s->s_mutex);
4939 mutex_lock(&mdsc->mutex);
4940 ceph_kick_flushing_caps(mdsc, s);
4941 mutex_unlock(&s->s_mutex);
4942 wake_up_session_caps(s, RECONNECT);
4947 * Only open and reconnect sessions that don't exist yet.
4949 for (i = 0; i < newmap->possible_max_rank; i++) {
4951 * In case the import MDS is crashed just after
4952 * the EImportStart journal is flushed, so when
4953 * a standby MDS takes over it and is replaying
4954 * the EImportStart journal the new MDS daemon
4955 * will wait the client to reconnect it, but the
4956 * client may never register/open the session yet.
4958 * Will try to reconnect that MDS daemon if the
4959 * rank number is in the export targets array and
4960 * is the up:reconnect state.
4962 newstate = ceph_mdsmap_get_state(newmap, i);
4963 if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
4967 * The session maybe registered and opened by some
4968 * requests which were choosing random MDSes during
4969 * the mdsc->mutex's unlock/lock gap below in rare
4970 * case. But the related MDS daemon will just queue
4971 * that requests and be still waiting for the client's
4972 * reconnection request in up:reconnect state.
4974 s = __ceph_lookup_mds_session(mdsc, i);
4976 s = __open_export_target_session(mdsc, i);
4980 "failed to open export target session, err %d\n",
4985 doutc(cl, "send reconnect to export target mds.%d\n", i);
4986 mutex_unlock(&mdsc->mutex);
4987 send_mds_reconnect(mdsc, s);
4988 ceph_put_mds_session(s);
4989 mutex_lock(&mdsc->mutex);
4992 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4993 s = mdsc->sessions[i];
4996 if (!ceph_mdsmap_is_laggy(newmap, i))
4998 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4999 s->s_state == CEPH_MDS_SESSION_HUNG ||
5000 s->s_state == CEPH_MDS_SESSION_CLOSING) {
5001 doutc(cl, " connecting to export targets of laggy mds%d\n", i);
5002 __open_export_target_sessions(mdsc, s);
5014 * caller must hold session s_mutex, dentry->d_lock
5016 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
5018 struct ceph_dentry_info *di = ceph_dentry(dentry);
5020 ceph_put_mds_session(di->lease_session);
5021 di->lease_session = NULL;
5024 static void handle_lease(struct ceph_mds_client *mdsc,
5025 struct ceph_mds_session *session,
5026 struct ceph_msg *msg)
5028 struct ceph_client *cl = mdsc->fsc->client;
5029 struct super_block *sb = mdsc->fsc->sb;
5030 struct inode *inode;
5031 struct dentry *parent, *dentry;
5032 struct ceph_dentry_info *di;
5033 int mds = session->s_mds;
5034 struct ceph_mds_lease *h = msg->front.iov_base;
5036 struct ceph_vino vino;
5040 doutc(cl, "from mds%d\n", mds);
5042 if (!ceph_inc_mds_stopping_blocker(mdsc, session))
5046 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
5048 vino.ino = le64_to_cpu(h->ino);
5049 vino.snap = CEPH_NOSNAP;
5050 seq = le32_to_cpu(h->seq);
5051 dname.len = get_unaligned_le32(h + 1);
5052 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
5054 dname.name = (void *)(h + 1) + sizeof(u32);
5057 inode = ceph_find_inode(sb, vino);
5058 doutc(cl, "%s, ino %llx %p %.*s\n", ceph_lease_op_name(h->action),
5059 vino.ino, inode, dname.len, dname.name);
5061 mutex_lock(&session->s_mutex);
5063 doutc(cl, "no inode %llx\n", vino.ino);
5068 parent = d_find_alias(inode);
5070 doutc(cl, "no parent dentry on inode %p\n", inode);
5072 goto release; /* hrm... */
5074 dname.hash = full_name_hash(parent, dname.name, dname.len);
5075 dentry = d_lookup(parent, &dname);
5080 spin_lock(&dentry->d_lock);
5081 di = ceph_dentry(dentry);
5082 switch (h->action) {
5083 case CEPH_MDS_LEASE_REVOKE:
5084 if (di->lease_session == session) {
5085 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
5086 h->seq = cpu_to_le32(di->lease_seq);
5087 __ceph_mdsc_drop_dentry_lease(dentry);
5092 case CEPH_MDS_LEASE_RENEW:
5093 if (di->lease_session == session &&
5094 di->lease_gen == atomic_read(&session->s_cap_gen) &&
5095 di->lease_renew_from &&
5096 di->lease_renew_after == 0) {
5097 unsigned long duration =
5098 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
5100 di->lease_seq = seq;
5101 di->time = di->lease_renew_from + duration;
5102 di->lease_renew_after = di->lease_renew_from +
5104 di->lease_renew_from = 0;
5108 spin_unlock(&dentry->d_lock);
5115 /* let's just reuse the same message */
5116 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
5118 ceph_con_send(&session->s_con, msg);
5121 mutex_unlock(&session->s_mutex);
5124 ceph_dec_mds_stopping_blocker(mdsc);
5128 ceph_dec_mds_stopping_blocker(mdsc);
5130 pr_err_client(cl, "corrupt lease message\n");
5134 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
5135 struct dentry *dentry, char action,
5138 struct ceph_client *cl = session->s_mdsc->fsc->client;
5139 struct ceph_msg *msg;
5140 struct ceph_mds_lease *lease;
5142 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
5144 doutc(cl, "identry %p %s to mds%d\n", dentry, ceph_lease_op_name(action),
5147 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
5150 lease = msg->front.iov_base;
5151 lease->action = action;
5152 lease->seq = cpu_to_le32(seq);
5154 spin_lock(&dentry->d_lock);
5155 dir = d_inode(dentry->d_parent);
5156 lease->ino = cpu_to_le64(ceph_ino(dir));
5157 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
5159 put_unaligned_le32(dentry->d_name.len, lease + 1);
5160 memcpy((void *)(lease + 1) + 4,
5161 dentry->d_name.name, dentry->d_name.len);
5162 spin_unlock(&dentry->d_lock);
5164 ceph_con_send(&session->s_con, msg);
5168 * lock unlock the session, to wait ongoing session activities
5170 static void lock_unlock_session(struct ceph_mds_session *s)
5172 mutex_lock(&s->s_mutex);
5173 mutex_unlock(&s->s_mutex);
5176 static void maybe_recover_session(struct ceph_mds_client *mdsc)
5178 struct ceph_client *cl = mdsc->fsc->client;
5179 struct ceph_fs_client *fsc = mdsc->fsc;
5181 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
5184 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
5187 if (!READ_ONCE(fsc->blocklisted))
5190 pr_info_client(cl, "auto reconnect after blocklisted\n");
5191 ceph_force_reconnect(fsc->sb);
5194 bool check_session_state(struct ceph_mds_session *s)
5196 struct ceph_client *cl = s->s_mdsc->fsc->client;
5198 switch (s->s_state) {
5199 case CEPH_MDS_SESSION_OPEN:
5200 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
5201 s->s_state = CEPH_MDS_SESSION_HUNG;
5202 pr_info_client(cl, "mds%d hung\n", s->s_mds);
5205 case CEPH_MDS_SESSION_CLOSING:
5206 case CEPH_MDS_SESSION_NEW:
5207 case CEPH_MDS_SESSION_RESTARTING:
5208 case CEPH_MDS_SESSION_CLOSED:
5209 case CEPH_MDS_SESSION_REJECTED:
5217 * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
5218 * then we need to retransmit that request.
5220 void inc_session_sequence(struct ceph_mds_session *s)
5222 struct ceph_client *cl = s->s_mdsc->fsc->client;
5224 lockdep_assert_held(&s->s_mutex);
5228 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
5231 doutc(cl, "resending session close request for mds%d\n", s->s_mds);
5232 ret = request_close_session(s);
5234 pr_err_client(cl, "unable to close session to mds%d: %d\n",
5240 * delayed work -- periodically trim expired leases, renew caps with mds. If
5241 * the @delay parameter is set to 0 or if it's more than 5 secs, the default
5242 * workqueue delay value of 5 secs will be used.
5244 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
5246 unsigned long max_delay = HZ * 5;
5248 /* 5 secs default delay */
5249 if (!delay || (delay > max_delay))
5251 schedule_delayed_work(&mdsc->delayed_work,
5252 round_jiffies_relative(delay));
5255 static void delayed_work(struct work_struct *work)
5257 struct ceph_mds_client *mdsc =
5258 container_of(work, struct ceph_mds_client, delayed_work.work);
5259 unsigned long delay;
5264 doutc(mdsc->fsc->client, "mdsc delayed_work\n");
5266 if (mdsc->stopping >= CEPH_MDSC_STOPPING_FLUSHED)
5269 mutex_lock(&mdsc->mutex);
5270 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
5271 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
5272 mdsc->last_renew_caps);
5274 mdsc->last_renew_caps = jiffies;
5276 for (i = 0; i < mdsc->max_sessions; i++) {
5277 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
5281 if (!check_session_state(s)) {
5282 ceph_put_mds_session(s);
5285 mutex_unlock(&mdsc->mutex);
5287 mutex_lock(&s->s_mutex);
5289 send_renew_caps(mdsc, s);
5291 ceph_con_keepalive(&s->s_con);
5292 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
5293 s->s_state == CEPH_MDS_SESSION_HUNG)
5294 ceph_send_cap_releases(mdsc, s);
5295 mutex_unlock(&s->s_mutex);
5296 ceph_put_mds_session(s);
5298 mutex_lock(&mdsc->mutex);
5300 mutex_unlock(&mdsc->mutex);
5302 delay = ceph_check_delayed_caps(mdsc);
5304 ceph_queue_cap_reclaim_work(mdsc);
5306 ceph_trim_snapid_map(mdsc);
5308 maybe_recover_session(mdsc);
5310 schedule_delayed(mdsc, delay);
5313 int ceph_mdsc_init(struct ceph_fs_client *fsc)
5316 struct ceph_mds_client *mdsc;
5319 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
5323 mutex_init(&mdsc->mutex);
5324 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
5325 if (!mdsc->mdsmap) {
5330 init_completion(&mdsc->safe_umount_waiters);
5331 spin_lock_init(&mdsc->stopping_lock);
5332 atomic_set(&mdsc->stopping_blockers, 0);
5333 init_completion(&mdsc->stopping_waiter);
5334 init_waitqueue_head(&mdsc->session_close_wq);
5335 INIT_LIST_HEAD(&mdsc->waiting_for_map);
5336 mdsc->quotarealms_inodes = RB_ROOT;
5337 mutex_init(&mdsc->quotarealms_inodes_mutex);
5338 init_rwsem(&mdsc->snap_rwsem);
5339 mdsc->snap_realms = RB_ROOT;
5340 INIT_LIST_HEAD(&mdsc->snap_empty);
5341 spin_lock_init(&mdsc->snap_empty_lock);
5342 mdsc->request_tree = RB_ROOT;
5343 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
5344 mdsc->last_renew_caps = jiffies;
5345 INIT_LIST_HEAD(&mdsc->cap_delay_list);
5346 INIT_LIST_HEAD(&mdsc->cap_wait_list);
5347 spin_lock_init(&mdsc->cap_delay_lock);
5348 INIT_LIST_HEAD(&mdsc->snap_flush_list);
5349 spin_lock_init(&mdsc->snap_flush_lock);
5350 mdsc->last_cap_flush_tid = 1;
5351 INIT_LIST_HEAD(&mdsc->cap_flush_list);
5352 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
5353 spin_lock_init(&mdsc->cap_dirty_lock);
5354 init_waitqueue_head(&mdsc->cap_flushing_wq);
5355 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
5356 err = ceph_metric_init(&mdsc->metric);
5360 spin_lock_init(&mdsc->dentry_list_lock);
5361 INIT_LIST_HEAD(&mdsc->dentry_leases);
5362 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
5364 ceph_caps_init(mdsc);
5365 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
5367 spin_lock_init(&mdsc->snapid_map_lock);
5368 mdsc->snapid_map_tree = RB_ROOT;
5369 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
5371 init_rwsem(&mdsc->pool_perm_rwsem);
5372 mdsc->pool_perm_tree = RB_ROOT;
5374 strscpy(mdsc->nodename, utsname()->nodename,
5375 sizeof(mdsc->nodename));
5381 kfree(mdsc->mdsmap);
5388 * Wait for safe replies on open mds requests. If we time out, drop
5389 * all requests from the tree to avoid dangling dentry refs.
5391 static void wait_requests(struct ceph_mds_client *mdsc)
5393 struct ceph_client *cl = mdsc->fsc->client;
5394 struct ceph_options *opts = mdsc->fsc->client->options;
5395 struct ceph_mds_request *req;
5397 mutex_lock(&mdsc->mutex);
5398 if (__get_oldest_req(mdsc)) {
5399 mutex_unlock(&mdsc->mutex);
5401 doutc(cl, "waiting for requests\n");
5402 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
5403 ceph_timeout_jiffies(opts->mount_timeout));
5405 /* tear down remaining requests */
5406 mutex_lock(&mdsc->mutex);
5407 while ((req = __get_oldest_req(mdsc))) {
5408 doutc(cl, "timed out on tid %llu\n", req->r_tid);
5409 list_del_init(&req->r_wait);
5410 __unregister_request(mdsc, req);
5413 mutex_unlock(&mdsc->mutex);
5414 doutc(cl, "done\n");
5417 void send_flush_mdlog(struct ceph_mds_session *s)
5419 struct ceph_client *cl = s->s_mdsc->fsc->client;
5420 struct ceph_msg *msg;
5423 * Pre-luminous MDS crashes when it sees an unknown session request
5425 if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
5428 mutex_lock(&s->s_mutex);
5429 doutc(cl, "request mdlog flush to mds%d (%s)s seq %lld\n",
5430 s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
5431 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
5434 pr_err_client(cl, "failed to request mdlog flush to mds%d (%s) seq %lld\n",
5435 s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
5437 ceph_con_send(&s->s_con, msg);
5439 mutex_unlock(&s->s_mutex);
5443 * called before mount is ro, and before dentries are torn down.
5444 * (hmm, does this still race with new lookups?)
5446 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
5448 doutc(mdsc->fsc->client, "begin\n");
5449 mdsc->stopping = CEPH_MDSC_STOPPING_BEGIN;
5451 ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
5452 ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
5453 ceph_flush_dirty_caps(mdsc);
5454 wait_requests(mdsc);
5457 * wait for reply handlers to drop their request refs and
5458 * their inode/dcache refs
5462 ceph_cleanup_quotarealms_inodes(mdsc);
5463 doutc(mdsc->fsc->client, "done\n");
5467 * flush the mdlog and wait for all write mds requests to flush.
5469 static void flush_mdlog_and_wait_mdsc_unsafe_requests(struct ceph_mds_client *mdsc,
5472 struct ceph_client *cl = mdsc->fsc->client;
5473 struct ceph_mds_request *req = NULL, *nextreq;
5474 struct ceph_mds_session *last_session = NULL;
5477 mutex_lock(&mdsc->mutex);
5478 doutc(cl, "want %lld\n", want_tid);
5480 req = __get_oldest_req(mdsc);
5481 while (req && req->r_tid <= want_tid) {
5482 /* find next request */
5483 n = rb_next(&req->r_node);
5485 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
5488 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
5489 (req->r_op & CEPH_MDS_OP_WRITE)) {
5490 struct ceph_mds_session *s = req->r_session;
5498 ceph_mdsc_get_request(req);
5500 ceph_mdsc_get_request(nextreq);
5501 s = ceph_get_mds_session(s);
5502 mutex_unlock(&mdsc->mutex);
5504 /* send flush mdlog request to MDS */
5505 if (last_session != s) {
5506 send_flush_mdlog(s);
5507 ceph_put_mds_session(last_session);
5510 ceph_put_mds_session(s);
5512 doutc(cl, "wait on %llu (want %llu)\n",
5513 req->r_tid, want_tid);
5514 wait_for_completion(&req->r_safe_completion);
5516 mutex_lock(&mdsc->mutex);
5517 ceph_mdsc_put_request(req);
5519 break; /* next dne before, so we're done! */
5520 if (RB_EMPTY_NODE(&nextreq->r_node)) {
5521 /* next request was removed from tree */
5522 ceph_mdsc_put_request(nextreq);
5525 ceph_mdsc_put_request(nextreq); /* won't go away */
5529 mutex_unlock(&mdsc->mutex);
5530 ceph_put_mds_session(last_session);
5531 doutc(cl, "done\n");
5534 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
5536 struct ceph_client *cl = mdsc->fsc->client;
5537 u64 want_tid, want_flush;
5539 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
5542 doutc(cl, "sync\n");
5543 mutex_lock(&mdsc->mutex);
5544 want_tid = mdsc->last_tid;
5545 mutex_unlock(&mdsc->mutex);
5547 ceph_flush_dirty_caps(mdsc);
5548 spin_lock(&mdsc->cap_dirty_lock);
5549 want_flush = mdsc->last_cap_flush_tid;
5550 if (!list_empty(&mdsc->cap_flush_list)) {
5551 struct ceph_cap_flush *cf =
5552 list_last_entry(&mdsc->cap_flush_list,
5553 struct ceph_cap_flush, g_list);
5556 spin_unlock(&mdsc->cap_dirty_lock);
5558 doutc(cl, "sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
5560 flush_mdlog_and_wait_mdsc_unsafe_requests(mdsc, want_tid);
5561 wait_caps_flush(mdsc, want_flush);
5565 * true if all sessions are closed, or we force unmount
5567 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
5569 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
5571 return atomic_read(&mdsc->num_sessions) <= skipped;
5575 * called after sb is ro or when metadata corrupted.
5577 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
5579 struct ceph_options *opts = mdsc->fsc->client->options;
5580 struct ceph_client *cl = mdsc->fsc->client;
5581 struct ceph_mds_session *session;
5585 doutc(cl, "begin\n");
5587 /* close sessions */
5588 mutex_lock(&mdsc->mutex);
5589 for (i = 0; i < mdsc->max_sessions; i++) {
5590 session = __ceph_lookup_mds_session(mdsc, i);
5593 mutex_unlock(&mdsc->mutex);
5594 mutex_lock(&session->s_mutex);
5595 if (__close_session(mdsc, session) <= 0)
5597 mutex_unlock(&session->s_mutex);
5598 ceph_put_mds_session(session);
5599 mutex_lock(&mdsc->mutex);
5601 mutex_unlock(&mdsc->mutex);
5603 doutc(cl, "waiting for sessions to close\n");
5604 wait_event_timeout(mdsc->session_close_wq,
5605 done_closing_sessions(mdsc, skipped),
5606 ceph_timeout_jiffies(opts->mount_timeout));
5608 /* tear down remaining sessions */
5609 mutex_lock(&mdsc->mutex);
5610 for (i = 0; i < mdsc->max_sessions; i++) {
5611 if (mdsc->sessions[i]) {
5612 session = ceph_get_mds_session(mdsc->sessions[i]);
5613 __unregister_session(mdsc, session);
5614 mutex_unlock(&mdsc->mutex);
5615 mutex_lock(&session->s_mutex);
5616 remove_session_caps(session);
5617 mutex_unlock(&session->s_mutex);
5618 ceph_put_mds_session(session);
5619 mutex_lock(&mdsc->mutex);
5622 WARN_ON(!list_empty(&mdsc->cap_delay_list));
5623 mutex_unlock(&mdsc->mutex);
5625 ceph_cleanup_snapid_map(mdsc);
5626 ceph_cleanup_global_and_empty_realms(mdsc);
5628 cancel_work_sync(&mdsc->cap_reclaim_work);
5629 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
5631 doutc(cl, "done\n");
5634 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
5636 struct ceph_mds_session *session;
5639 doutc(mdsc->fsc->client, "force umount\n");
5641 mutex_lock(&mdsc->mutex);
5642 for (mds = 0; mds < mdsc->max_sessions; mds++) {
5643 session = __ceph_lookup_mds_session(mdsc, mds);
5647 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
5648 __unregister_session(mdsc, session);
5649 __wake_requests(mdsc, &session->s_waiting);
5650 mutex_unlock(&mdsc->mutex);
5652 mutex_lock(&session->s_mutex);
5653 __close_session(mdsc, session);
5654 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
5655 cleanup_session_requests(mdsc, session);
5656 remove_session_caps(session);
5658 mutex_unlock(&session->s_mutex);
5659 ceph_put_mds_session(session);
5661 mutex_lock(&mdsc->mutex);
5662 kick_requests(mdsc, mds);
5664 __wake_requests(mdsc, &mdsc->waiting_for_map);
5665 mutex_unlock(&mdsc->mutex);
5668 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
5670 doutc(mdsc->fsc->client, "stop\n");
5672 * Make sure the delayed work stopped before releasing
5675 * Because the cancel_delayed_work_sync() will only
5676 * guarantee that the work finishes executing. But the
5677 * delayed work will re-arm itself again after that.
5679 flush_delayed_work(&mdsc->delayed_work);
5682 ceph_mdsmap_destroy(mdsc->mdsmap);
5683 kfree(mdsc->sessions);
5684 ceph_caps_finalize(mdsc);
5685 ceph_pool_perm_destroy(mdsc);
5688 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
5690 struct ceph_mds_client *mdsc = fsc->mdsc;
5691 doutc(fsc->client, "%p\n", mdsc);
5696 /* flush out any connection work with references to us */
5699 ceph_mdsc_stop(mdsc);
5701 ceph_metric_destroy(&mdsc->metric);
5705 doutc(fsc->client, "%p done\n", mdsc);
5708 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5710 struct ceph_fs_client *fsc = mdsc->fsc;
5711 struct ceph_client *cl = fsc->client;
5712 const char *mds_namespace = fsc->mount_options->mds_namespace;
5713 void *p = msg->front.iov_base;
5714 void *end = p + msg->front.iov_len;
5717 u32 mount_fscid = (u32)-1;
5720 ceph_decode_need(&p, end, sizeof(u32), bad);
5721 epoch = ceph_decode_32(&p);
5723 doutc(cl, "epoch %u\n", epoch);
5725 /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
5726 ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
5728 ceph_decode_32_safe(&p, end, num_fs, bad);
5729 while (num_fs-- > 0) {
5730 void *info_p, *info_end;
5734 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
5735 p += 2; // info_v, info_cv
5736 info_len = ceph_decode_32(&p);
5737 ceph_decode_need(&p, end, info_len, bad);
5739 info_end = p + info_len;
5742 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
5743 fscid = ceph_decode_32(&info_p);
5744 namelen = ceph_decode_32(&info_p);
5745 ceph_decode_need(&info_p, info_end, namelen, bad);
5747 if (mds_namespace &&
5748 strlen(mds_namespace) == namelen &&
5749 !strncmp(mds_namespace, (char *)info_p, namelen)) {
5750 mount_fscid = fscid;
5755 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
5756 if (mount_fscid != (u32)-1) {
5757 fsc->client->monc.fs_cluster_id = mount_fscid;
5758 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
5760 ceph_monc_renew_subs(&fsc->client->monc);
5768 pr_err_client(cl, "error decoding fsmap %d. Shutting down mount.\n",
5770 ceph_umount_begin(mdsc->fsc->sb);
5773 mutex_lock(&mdsc->mutex);
5774 mdsc->mdsmap_err = err;
5775 __wake_requests(mdsc, &mdsc->waiting_for_map);
5776 mutex_unlock(&mdsc->mutex);
5780 * handle mds map update.
5782 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5784 struct ceph_client *cl = mdsc->fsc->client;
5787 void *p = msg->front.iov_base;
5788 void *end = p + msg->front.iov_len;
5789 struct ceph_mdsmap *newmap, *oldmap;
5790 struct ceph_fsid fsid;
5793 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5794 ceph_decode_copy(&p, &fsid, sizeof(fsid));
5795 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5797 epoch = ceph_decode_32(&p);
5798 maplen = ceph_decode_32(&p);
5799 doutc(cl, "epoch %u len %d\n", epoch, (int)maplen);
5801 /* do we need it? */
5802 mutex_lock(&mdsc->mutex);
5803 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5804 doutc(cl, "epoch %u <= our %u\n", epoch, mdsc->mdsmap->m_epoch);
5805 mutex_unlock(&mdsc->mutex);
5809 newmap = ceph_mdsmap_decode(mdsc, &p, end, ceph_msgr2(mdsc->fsc->client));
5810 if (IS_ERR(newmap)) {
5811 err = PTR_ERR(newmap);
5815 /* swap into place */
5817 oldmap = mdsc->mdsmap;
5818 mdsc->mdsmap = newmap;
5819 check_new_map(mdsc, newmap, oldmap);
5820 ceph_mdsmap_destroy(oldmap);
5822 mdsc->mdsmap = newmap; /* first mds map */
5824 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5827 __wake_requests(mdsc, &mdsc->waiting_for_map);
5828 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5829 mdsc->mdsmap->m_epoch);
5831 mutex_unlock(&mdsc->mutex);
5832 schedule_delayed(mdsc, 0);
5836 mutex_unlock(&mdsc->mutex);
5838 pr_err_client(cl, "error decoding mdsmap %d. Shutting down mount.\n",
5840 ceph_umount_begin(mdsc->fsc->sb);
5845 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5847 struct ceph_mds_session *s = con->private;
5849 if (ceph_get_mds_session(s))
5854 static void mds_put_con(struct ceph_connection *con)
5856 struct ceph_mds_session *s = con->private;
5858 ceph_put_mds_session(s);
5862 * if the client is unresponsive for long enough, the mds will kill
5863 * the session entirely.
5865 static void mds_peer_reset(struct ceph_connection *con)
5867 struct ceph_mds_session *s = con->private;
5868 struct ceph_mds_client *mdsc = s->s_mdsc;
5870 pr_warn_client(mdsc->fsc->client, "mds%d closed our session\n",
5872 if (READ_ONCE(mdsc->fsc->mount_state) != CEPH_MOUNT_FENCE_IO)
5873 send_mds_reconnect(mdsc, s);
5876 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5878 struct ceph_mds_session *s = con->private;
5879 struct ceph_mds_client *mdsc = s->s_mdsc;
5880 struct ceph_client *cl = mdsc->fsc->client;
5881 int type = le16_to_cpu(msg->hdr.type);
5883 mutex_lock(&mdsc->mutex);
5884 if (__verify_registered_session(mdsc, s) < 0) {
5885 mutex_unlock(&mdsc->mutex);
5888 mutex_unlock(&mdsc->mutex);
5891 case CEPH_MSG_MDS_MAP:
5892 ceph_mdsc_handle_mdsmap(mdsc, msg);
5894 case CEPH_MSG_FS_MAP_USER:
5895 ceph_mdsc_handle_fsmap(mdsc, msg);
5897 case CEPH_MSG_CLIENT_SESSION:
5898 handle_session(s, msg);
5900 case CEPH_MSG_CLIENT_REPLY:
5901 handle_reply(s, msg);
5903 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5904 handle_forward(mdsc, s, msg);
5906 case CEPH_MSG_CLIENT_CAPS:
5907 ceph_handle_caps(s, msg);
5909 case CEPH_MSG_CLIENT_SNAP:
5910 ceph_handle_snap(mdsc, s, msg);
5912 case CEPH_MSG_CLIENT_LEASE:
5913 handle_lease(mdsc, s, msg);
5915 case CEPH_MSG_CLIENT_QUOTA:
5916 ceph_handle_quota(mdsc, s, msg);
5920 pr_err_client(cl, "received unknown message type %d %s\n",
5921 type, ceph_msg_type_name(type));
5932 * Note: returned pointer is the address of a structure that's
5933 * managed separately. Caller must *not* attempt to free it.
5935 static struct ceph_auth_handshake *
5936 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5938 struct ceph_mds_session *s = con->private;
5939 struct ceph_mds_client *mdsc = s->s_mdsc;
5940 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5941 struct ceph_auth_handshake *auth = &s->s_auth;
5944 ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5945 force_new, proto, NULL, NULL);
5947 return ERR_PTR(ret);
5952 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5953 void *challenge_buf, int challenge_buf_len)
5955 struct ceph_mds_session *s = con->private;
5956 struct ceph_mds_client *mdsc = s->s_mdsc;
5957 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5959 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5960 challenge_buf, challenge_buf_len);
5963 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5965 struct ceph_mds_session *s = con->private;
5966 struct ceph_mds_client *mdsc = s->s_mdsc;
5967 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5968 struct ceph_auth_handshake *auth = &s->s_auth;
5970 return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5971 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5972 NULL, NULL, NULL, NULL);
5975 static int mds_invalidate_authorizer(struct ceph_connection *con)
5977 struct ceph_mds_session *s = con->private;
5978 struct ceph_mds_client *mdsc = s->s_mdsc;
5979 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5981 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5983 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5986 static int mds_get_auth_request(struct ceph_connection *con,
5987 void *buf, int *buf_len,
5988 void **authorizer, int *authorizer_len)
5990 struct ceph_mds_session *s = con->private;
5991 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5992 struct ceph_auth_handshake *auth = &s->s_auth;
5995 ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
6000 *authorizer = auth->authorizer_buf;
6001 *authorizer_len = auth->authorizer_buf_len;
6005 static int mds_handle_auth_reply_more(struct ceph_connection *con,
6006 void *reply, int reply_len,
6007 void *buf, int *buf_len,
6008 void **authorizer, int *authorizer_len)
6010 struct ceph_mds_session *s = con->private;
6011 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6012 struct ceph_auth_handshake *auth = &s->s_auth;
6015 ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
6020 *authorizer = auth->authorizer_buf;
6021 *authorizer_len = auth->authorizer_buf_len;
6025 static int mds_handle_auth_done(struct ceph_connection *con,
6026 u64 global_id, void *reply, int reply_len,
6027 u8 *session_key, int *session_key_len,
6028 u8 *con_secret, int *con_secret_len)
6030 struct ceph_mds_session *s = con->private;
6031 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6032 struct ceph_auth_handshake *auth = &s->s_auth;
6034 return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
6035 session_key, session_key_len,
6036 con_secret, con_secret_len);
6039 static int mds_handle_auth_bad_method(struct ceph_connection *con,
6040 int used_proto, int result,
6041 const int *allowed_protos, int proto_cnt,
6042 const int *allowed_modes, int mode_cnt)
6044 struct ceph_mds_session *s = con->private;
6045 struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
6048 if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
6050 allowed_protos, proto_cnt,
6051 allowed_modes, mode_cnt)) {
6052 ret = ceph_monc_validate_auth(monc);
6060 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
6061 struct ceph_msg_header *hdr, int *skip)
6063 struct ceph_msg *msg;
6064 int type = (int) le16_to_cpu(hdr->type);
6065 int front_len = (int) le32_to_cpu(hdr->front_len);
6071 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
6073 pr_err("unable to allocate msg type %d len %d\n",
6081 static int mds_sign_message(struct ceph_msg *msg)
6083 struct ceph_mds_session *s = msg->con->private;
6084 struct ceph_auth_handshake *auth = &s->s_auth;
6086 return ceph_auth_sign_message(auth, msg);
6089 static int mds_check_message_signature(struct ceph_msg *msg)
6091 struct ceph_mds_session *s = msg->con->private;
6092 struct ceph_auth_handshake *auth = &s->s_auth;
6094 return ceph_auth_check_message_signature(auth, msg);
6097 static const struct ceph_connection_operations mds_con_ops = {
6100 .alloc_msg = mds_alloc_msg,
6101 .dispatch = mds_dispatch,
6102 .peer_reset = mds_peer_reset,
6103 .get_authorizer = mds_get_authorizer,
6104 .add_authorizer_challenge = mds_add_authorizer_challenge,
6105 .verify_authorizer_reply = mds_verify_authorizer_reply,
6106 .invalidate_authorizer = mds_invalidate_authorizer,
6107 .sign_message = mds_sign_message,
6108 .check_message_signature = mds_check_message_signature,
6109 .get_auth_request = mds_get_auth_request,
6110 .handle_auth_reply_more = mds_handle_auth_reply_more,
6111 .handle_auth_done = mds_handle_auth_done,
6112 .handle_auth_bad_method = mds_handle_auth_bad_method,