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52 * are no longer valid.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 *
1538 create_session_full_msg(struct ceph_mds_client *mdsc, int op, u64 seq)
1540 struct ceph_msg *msg;
1541 struct ceph_mds_session_head *h;
1543 int extra_bytes = 0;
1544 int metadata_key_count = 0;
1545 struct ceph_options *opt = mdsc->fsc->client->options;
1546 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1547 struct ceph_client *cl = mdsc->fsc->client;
1552 const char* metadata[][2] = {
1553 {"hostname", mdsc->nodename},
1554 {"kernel_version", init_utsname()->release},
1555 {"entity_id", opt->name ? : ""},
1556 {"root", fsopt->server_path ? : "/"},
1560 /* Calculate serialized length of metadata */
1561 extra_bytes = 4; /* map length */
1562 for (i = 0; metadata[i][0]; ++i) {
1563 extra_bytes += 8 + strlen(metadata[i][0]) +
1564 strlen(metadata[i][1]);
1565 metadata_key_count++;
1568 /* supported feature */
1570 count = ARRAY_SIZE(feature_bits);
1572 size = FEATURE_BYTES(count);
1573 extra_bytes += 4 + size;
1577 count = ARRAY_SIZE(metric_bits);
1579 size = METRIC_BYTES(count);
1580 extra_bytes += 2 + 4 + 4 + size;
1582 /* flags, mds auth caps and oldest_client_tid */
1583 extra_bytes += 4 + 4 + 8;
1585 /* Allocate the message */
1586 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1589 pr_err_client(cl, "ENOMEM creating session open msg\n");
1590 return ERR_PTR(-ENOMEM);
1592 p = msg->front.iov_base;
1593 end = p + msg->front.iov_len;
1596 h->op = cpu_to_le32(op);
1597 h->seq = cpu_to_le64(seq);
1600 * Serialize client metadata into waiting buffer space, using
1601 * the format that userspace expects for map<string, string>
1603 * ClientSession messages with metadata are v7
1605 msg->hdr.version = cpu_to_le16(7);
1606 msg->hdr.compat_version = cpu_to_le16(1);
1608 /* The write pointer, following the session_head structure */
1611 /* Number of entries in the map */
1612 ceph_encode_32(&p, metadata_key_count);
1614 /* Two length-prefixed strings for each entry in the map */
1615 for (i = 0; metadata[i][0]; ++i) {
1616 size_t const key_len = strlen(metadata[i][0]);
1617 size_t const val_len = strlen(metadata[i][1]);
1619 ceph_encode_32(&p, key_len);
1620 memcpy(p, metadata[i][0], key_len);
1622 ceph_encode_32(&p, val_len);
1623 memcpy(p, metadata[i][1], val_len);
1627 ret = encode_supported_features(&p, end);
1629 pr_err_client(cl, "encode_supported_features failed!\n");
1631 return ERR_PTR(ret);
1634 ret = encode_metric_spec(&p, end);
1636 pr_err_client(cl, "encode_metric_spec failed!\n");
1638 return ERR_PTR(ret);
1641 /* version == 5, flags */
1642 ceph_encode_32(&p, 0);
1644 /* version == 6, mds auth caps */
1645 ceph_encode_32(&p, 0);
1647 /* version == 7, oldest_client_tid */
1648 ceph_encode_64(&p, mdsc->oldest_tid);
1650 msg->front.iov_len = p - msg->front.iov_base;
1651 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1657 * send session open request.
1659 * called under mdsc->mutex
1661 static int __open_session(struct ceph_mds_client *mdsc,
1662 struct ceph_mds_session *session)
1664 struct ceph_msg *msg;
1666 int mds = session->s_mds;
1668 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
1671 /* wait for mds to go active? */
1672 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1673 doutc(mdsc->fsc->client, "open_session to mds%d (%s)\n", mds,
1674 ceph_mds_state_name(mstate));
1675 session->s_state = CEPH_MDS_SESSION_OPENING;
1676 session->s_renew_requested = jiffies;
1678 /* send connect message */
1679 msg = create_session_full_msg(mdsc, CEPH_SESSION_REQUEST_OPEN,
1682 return PTR_ERR(msg);
1683 ceph_con_send(&session->s_con, msg);
1688 * open sessions for any export targets for the given mds
1690 * called under mdsc->mutex
1692 static struct ceph_mds_session *
1693 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1695 struct ceph_mds_session *session;
1698 session = __ceph_lookup_mds_session(mdsc, target);
1700 session = register_session(mdsc, target);
1701 if (IS_ERR(session))
1704 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1705 session->s_state == CEPH_MDS_SESSION_CLOSING) {
1706 ret = __open_session(mdsc, session);
1708 return ERR_PTR(ret);
1714 struct ceph_mds_session *
1715 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1717 struct ceph_mds_session *session;
1718 struct ceph_client *cl = mdsc->fsc->client;
1720 doutc(cl, "to mds%d\n", target);
1722 mutex_lock(&mdsc->mutex);
1723 session = __open_export_target_session(mdsc, target);
1724 mutex_unlock(&mdsc->mutex);
1729 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1730 struct ceph_mds_session *session)
1732 struct ceph_mds_info *mi;
1733 struct ceph_mds_session *ts;
1734 int i, mds = session->s_mds;
1735 struct ceph_client *cl = mdsc->fsc->client;
1737 if (mds >= mdsc->mdsmap->possible_max_rank)
1740 mi = &mdsc->mdsmap->m_info[mds];
1741 doutc(cl, "for mds%d (%d targets)\n", session->s_mds,
1742 mi->num_export_targets);
1744 for (i = 0; i < mi->num_export_targets; i++) {
1745 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1746 ceph_put_mds_session(ts);
1750 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1751 struct ceph_mds_session *session)
1753 mutex_lock(&mdsc->mutex);
1754 __open_export_target_sessions(mdsc, session);
1755 mutex_unlock(&mdsc->mutex);
1762 static void detach_cap_releases(struct ceph_mds_session *session,
1763 struct list_head *target)
1765 struct ceph_client *cl = session->s_mdsc->fsc->client;
1767 lockdep_assert_held(&session->s_cap_lock);
1769 list_splice_init(&session->s_cap_releases, target);
1770 session->s_num_cap_releases = 0;
1771 doutc(cl, "mds%d\n", session->s_mds);
1774 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1775 struct list_head *dispose)
1777 while (!list_empty(dispose)) {
1778 struct ceph_cap *cap;
1779 /* zero out the in-progress message */
1780 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1781 list_del(&cap->session_caps);
1782 ceph_put_cap(mdsc, cap);
1786 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1787 struct ceph_mds_session *session)
1789 struct ceph_client *cl = mdsc->fsc->client;
1790 struct ceph_mds_request *req;
1793 doutc(cl, "mds%d\n", session->s_mds);
1794 mutex_lock(&mdsc->mutex);
1795 while (!list_empty(&session->s_unsafe)) {
1796 req = list_first_entry(&session->s_unsafe,
1797 struct ceph_mds_request, r_unsafe_item);
1798 pr_warn_ratelimited_client(cl, " dropping unsafe request %llu\n",
1800 if (req->r_target_inode)
1801 mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1802 if (req->r_unsafe_dir)
1803 mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1804 __unregister_request(mdsc, req);
1806 /* zero r_attempts, so kick_requests() will re-send requests */
1807 p = rb_first(&mdsc->request_tree);
1809 req = rb_entry(p, struct ceph_mds_request, r_node);
1811 if (req->r_session &&
1812 req->r_session->s_mds == session->s_mds)
1813 req->r_attempts = 0;
1815 mutex_unlock(&mdsc->mutex);
1819 * Helper to safely iterate over all caps associated with a session, with
1820 * special care taken to handle a racing __ceph_remove_cap().
1822 * Caller must hold session s_mutex.
1824 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1825 int (*cb)(struct inode *, int mds, void *),
1828 struct ceph_client *cl = session->s_mdsc->fsc->client;
1829 struct list_head *p;
1830 struct ceph_cap *cap;
1831 struct inode *inode, *last_inode = NULL;
1832 struct ceph_cap *old_cap = NULL;
1835 doutc(cl, "%p mds%d\n", session, session->s_mds);
1836 spin_lock(&session->s_cap_lock);
1837 p = session->s_caps.next;
1838 while (p != &session->s_caps) {
1841 cap = list_entry(p, struct ceph_cap, session_caps);
1842 inode = igrab(&cap->ci->netfs.inode);
1847 session->s_cap_iterator = cap;
1849 spin_unlock(&session->s_cap_lock);
1856 ceph_put_cap(session->s_mdsc, old_cap);
1860 ret = cb(inode, mds, arg);
1863 spin_lock(&session->s_cap_lock);
1866 doutc(cl, "finishing cap %p removal\n", cap);
1867 BUG_ON(cap->session != session);
1868 cap->session = NULL;
1869 list_del_init(&cap->session_caps);
1870 session->s_nr_caps--;
1871 atomic64_dec(&session->s_mdsc->metric.total_caps);
1872 if (cap->queue_release)
1873 __ceph_queue_cap_release(session, cap);
1875 old_cap = cap; /* put_cap it w/o locks held */
1882 session->s_cap_iterator = NULL;
1883 spin_unlock(&session->s_cap_lock);
1887 ceph_put_cap(session->s_mdsc, old_cap);
1892 static int remove_session_caps_cb(struct inode *inode, int mds, void *arg)
1894 struct ceph_inode_info *ci = ceph_inode(inode);
1895 struct ceph_client *cl = ceph_inode_to_client(inode);
1896 bool invalidate = false;
1897 struct ceph_cap *cap;
1900 spin_lock(&ci->i_ceph_lock);
1901 cap = __get_cap_for_mds(ci, mds);
1903 doutc(cl, " removing cap %p, ci is %p, inode is %p\n",
1904 cap, ci, &ci->netfs.inode);
1906 iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1908 spin_unlock(&ci->i_ceph_lock);
1911 wake_up_all(&ci->i_cap_wq);
1913 ceph_queue_invalidate(inode);
1920 * caller must hold session s_mutex
1922 static void remove_session_caps(struct ceph_mds_session *session)
1924 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1925 struct super_block *sb = fsc->sb;
1928 doutc(fsc->client, "on %p\n", session);
1929 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1931 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1933 spin_lock(&session->s_cap_lock);
1934 if (session->s_nr_caps > 0) {
1935 struct inode *inode;
1936 struct ceph_cap *cap, *prev = NULL;
1937 struct ceph_vino vino;
1939 * iterate_session_caps() skips inodes that are being
1940 * deleted, we need to wait until deletions are complete.
1941 * __wait_on_freeing_inode() is designed for the job,
1942 * but it is not exported, so use lookup inode function
1945 while (!list_empty(&session->s_caps)) {
1946 cap = list_entry(session->s_caps.next,
1947 struct ceph_cap, session_caps);
1951 vino = cap->ci->i_vino;
1952 spin_unlock(&session->s_cap_lock);
1954 inode = ceph_find_inode(sb, vino);
1957 spin_lock(&session->s_cap_lock);
1961 // drop cap expires and unlock s_cap_lock
1962 detach_cap_releases(session, &dispose);
1964 BUG_ON(session->s_nr_caps > 0);
1965 BUG_ON(!list_empty(&session->s_cap_flushing));
1966 spin_unlock(&session->s_cap_lock);
1967 dispose_cap_releases(session->s_mdsc, &dispose);
1977 * wake up any threads waiting on this session's caps. if the cap is
1978 * old (didn't get renewed on the client reconnect), remove it now.
1980 * caller must hold s_mutex.
1982 static int wake_up_session_cb(struct inode *inode, int mds, void *arg)
1984 struct ceph_inode_info *ci = ceph_inode(inode);
1985 unsigned long ev = (unsigned long)arg;
1987 if (ev == RECONNECT) {
1988 spin_lock(&ci->i_ceph_lock);
1989 ci->i_wanted_max_size = 0;
1990 ci->i_requested_max_size = 0;
1991 spin_unlock(&ci->i_ceph_lock);
1992 } else if (ev == RENEWCAPS) {
1993 struct ceph_cap *cap;
1995 spin_lock(&ci->i_ceph_lock);
1996 cap = __get_cap_for_mds(ci, mds);
1997 /* mds did not re-issue stale cap */
1998 if (cap && cap->cap_gen < atomic_read(&cap->session->s_cap_gen))
1999 cap->issued = cap->implemented = CEPH_CAP_PIN;
2000 spin_unlock(&ci->i_ceph_lock);
2001 } else if (ev == FORCE_RO) {
2003 wake_up_all(&ci->i_cap_wq);
2007 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
2009 struct ceph_client *cl = session->s_mdsc->fsc->client;
2011 doutc(cl, "session %p mds%d\n", session, session->s_mds);
2012 ceph_iterate_session_caps(session, wake_up_session_cb,
2013 (void *)(unsigned long)ev);
2017 * Send periodic message to MDS renewing all currently held caps. The
2018 * ack will reset the expiration for all caps from this session.
2020 * caller holds s_mutex
2022 static int send_renew_caps(struct ceph_mds_client *mdsc,
2023 struct ceph_mds_session *session)
2025 struct ceph_client *cl = mdsc->fsc->client;
2026 struct ceph_msg *msg;
2029 if (time_after_eq(jiffies, session->s_cap_ttl) &&
2030 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
2031 pr_info_client(cl, "mds%d caps stale\n", session->s_mds);
2032 session->s_renew_requested = jiffies;
2034 /* do not try to renew caps until a recovering mds has reconnected
2035 * with its clients. */
2036 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
2037 if (state < CEPH_MDS_STATE_RECONNECT) {
2038 doutc(cl, "ignoring mds%d (%s)\n", session->s_mds,
2039 ceph_mds_state_name(state));
2043 doutc(cl, "to mds%d (%s)\n", session->s_mds,
2044 ceph_mds_state_name(state));
2045 msg = create_session_full_msg(mdsc, CEPH_SESSION_REQUEST_RENEWCAPS,
2046 ++session->s_renew_seq);
2048 return PTR_ERR(msg);
2049 ceph_con_send(&session->s_con, msg);
2053 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
2054 struct ceph_mds_session *session, u64 seq)
2056 struct ceph_client *cl = mdsc->fsc->client;
2057 struct ceph_msg *msg;
2059 doutc(cl, "to mds%d (%s)s seq %lld\n", session->s_mds,
2060 ceph_session_state_name(session->s_state), seq);
2061 msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
2064 ceph_con_send(&session->s_con, msg);
2070 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
2072 * Called under session->s_mutex
2074 static void renewed_caps(struct ceph_mds_client *mdsc,
2075 struct ceph_mds_session *session, int is_renew)
2077 struct ceph_client *cl = mdsc->fsc->client;
2081 spin_lock(&session->s_cap_lock);
2082 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
2084 session->s_cap_ttl = session->s_renew_requested +
2085 mdsc->mdsmap->m_session_timeout*HZ;
2088 if (time_before(jiffies, session->s_cap_ttl)) {
2089 pr_info_client(cl, "mds%d caps renewed\n",
2093 pr_info_client(cl, "mds%d caps still stale\n",
2097 doutc(cl, "mds%d ttl now %lu, was %s, now %s\n", session->s_mds,
2098 session->s_cap_ttl, was_stale ? "stale" : "fresh",
2099 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
2100 spin_unlock(&session->s_cap_lock);
2103 wake_up_session_caps(session, RENEWCAPS);
2107 * send a session close request
2109 static int request_close_session(struct ceph_mds_session *session)
2111 struct ceph_client *cl = session->s_mdsc->fsc->client;
2112 struct ceph_msg *msg;
2114 doutc(cl, "mds%d state %s seq %lld\n", session->s_mds,
2115 ceph_session_state_name(session->s_state), session->s_seq);
2116 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
2120 ceph_con_send(&session->s_con, msg);
2125 * Called with s_mutex held.
2127 static int __close_session(struct ceph_mds_client *mdsc,
2128 struct ceph_mds_session *session)
2130 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
2132 session->s_state = CEPH_MDS_SESSION_CLOSING;
2133 return request_close_session(session);
2136 static bool drop_negative_children(struct dentry *dentry)
2138 struct dentry *child;
2139 bool all_negative = true;
2141 if (!d_is_dir(dentry))
2144 spin_lock(&dentry->d_lock);
2145 hlist_for_each_entry(child, &dentry->d_children, d_sib) {
2146 if (d_really_is_positive(child)) {
2147 all_negative = false;
2151 spin_unlock(&dentry->d_lock);
2154 shrink_dcache_parent(dentry);
2156 return all_negative;
2160 * Trim old(er) caps.
2162 * Because we can't cache an inode without one or more caps, we do
2163 * this indirectly: if a cap is unused, we prune its aliases, at which
2164 * point the inode will hopefully get dropped to.
2166 * Yes, this is a bit sloppy. Our only real goal here is to respond to
2167 * memory pressure from the MDS, though, so it needn't be perfect.
2169 static int trim_caps_cb(struct inode *inode, int mds, void *arg)
2171 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
2172 struct ceph_client *cl = mdsc->fsc->client;
2173 int *remaining = arg;
2174 struct ceph_inode_info *ci = ceph_inode(inode);
2175 int used, wanted, oissued, mine;
2176 struct ceph_cap *cap;
2178 if (*remaining <= 0)
2181 spin_lock(&ci->i_ceph_lock);
2182 cap = __get_cap_for_mds(ci, mds);
2184 spin_unlock(&ci->i_ceph_lock);
2187 mine = cap->issued | cap->implemented;
2188 used = __ceph_caps_used(ci);
2189 wanted = __ceph_caps_file_wanted(ci);
2190 oissued = __ceph_caps_issued_other(ci, cap);
2192 doutc(cl, "%p %llx.%llx cap %p mine %s oissued %s used %s wanted %s\n",
2193 inode, ceph_vinop(inode), cap, ceph_cap_string(mine),
2194 ceph_cap_string(oissued), ceph_cap_string(used),
2195 ceph_cap_string(wanted));
2196 if (cap == ci->i_auth_cap) {
2197 if (ci->i_dirty_caps || ci->i_flushing_caps ||
2198 !list_empty(&ci->i_cap_snaps))
2200 if ((used | wanted) & CEPH_CAP_ANY_WR)
2202 /* Note: it's possible that i_filelock_ref becomes non-zero
2203 * after dropping auth caps. It doesn't hurt because reply
2204 * of lock mds request will re-add auth caps. */
2205 if (atomic_read(&ci->i_filelock_ref) > 0)
2208 /* The inode has cached pages, but it's no longer used.
2209 * we can safely drop it */
2210 if (S_ISREG(inode->i_mode) &&
2211 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
2212 !(oissued & CEPH_CAP_FILE_CACHE)) {
2216 if ((used | wanted) & ~oissued & mine)
2217 goto out; /* we need these caps */
2220 /* we aren't the only cap.. just remove us */
2221 ceph_remove_cap(mdsc, cap, true);
2224 struct dentry *dentry;
2225 /* try dropping referring dentries */
2226 spin_unlock(&ci->i_ceph_lock);
2227 dentry = d_find_any_alias(inode);
2228 if (dentry && drop_negative_children(dentry)) {
2231 d_prune_aliases(inode);
2232 count = atomic_read(&inode->i_count);
2235 doutc(cl, "%p %llx.%llx cap %p pruned, count now %d\n",
2236 inode, ceph_vinop(inode), cap, count);
2244 spin_unlock(&ci->i_ceph_lock);
2249 * Trim session cap count down to some max number.
2251 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2252 struct ceph_mds_session *session,
2255 struct ceph_client *cl = mdsc->fsc->client;
2256 int trim_caps = session->s_nr_caps - max_caps;
2258 doutc(cl, "mds%d start: %d / %d, trim %d\n", session->s_mds,
2259 session->s_nr_caps, max_caps, trim_caps);
2260 if (trim_caps > 0) {
2261 int remaining = trim_caps;
2263 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2264 doutc(cl, "mds%d done: %d / %d, trimmed %d\n",
2265 session->s_mds, session->s_nr_caps, max_caps,
2266 trim_caps - remaining);
2269 ceph_flush_cap_releases(mdsc, session);
2273 static int check_caps_flush(struct ceph_mds_client *mdsc,
2276 struct ceph_client *cl = mdsc->fsc->client;
2279 spin_lock(&mdsc->cap_dirty_lock);
2280 if (!list_empty(&mdsc->cap_flush_list)) {
2281 struct ceph_cap_flush *cf =
2282 list_first_entry(&mdsc->cap_flush_list,
2283 struct ceph_cap_flush, g_list);
2284 if (cf->tid <= want_flush_tid) {
2285 doutc(cl, "still flushing tid %llu <= %llu\n",
2286 cf->tid, want_flush_tid);
2290 spin_unlock(&mdsc->cap_dirty_lock);
2295 * flush all dirty inode data to disk.
2297 * returns true if we've flushed through want_flush_tid
2299 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2302 struct ceph_client *cl = mdsc->fsc->client;
2304 doutc(cl, "want %llu\n", want_flush_tid);
2306 wait_event(mdsc->cap_flushing_wq,
2307 check_caps_flush(mdsc, want_flush_tid));
2309 doutc(cl, "ok, flushed thru %llu\n", want_flush_tid);
2313 * called under s_mutex
2315 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2316 struct ceph_mds_session *session)
2318 struct ceph_client *cl = mdsc->fsc->client;
2319 struct ceph_msg *msg = NULL;
2320 struct ceph_mds_cap_release *head;
2321 struct ceph_mds_cap_item *item;
2322 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2323 struct ceph_cap *cap;
2324 LIST_HEAD(tmp_list);
2325 int num_cap_releases;
2326 __le32 barrier, *cap_barrier;
2328 down_read(&osdc->lock);
2329 barrier = cpu_to_le32(osdc->epoch_barrier);
2330 up_read(&osdc->lock);
2332 spin_lock(&session->s_cap_lock);
2334 list_splice_init(&session->s_cap_releases, &tmp_list);
2335 num_cap_releases = session->s_num_cap_releases;
2336 session->s_num_cap_releases = 0;
2337 spin_unlock(&session->s_cap_lock);
2339 while (!list_empty(&tmp_list)) {
2341 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2342 PAGE_SIZE, GFP_NOFS, false);
2345 head = msg->front.iov_base;
2346 head->num = cpu_to_le32(0);
2347 msg->front.iov_len = sizeof(*head);
2349 msg->hdr.version = cpu_to_le16(2);
2350 msg->hdr.compat_version = cpu_to_le16(1);
2353 cap = list_first_entry(&tmp_list, struct ceph_cap,
2355 list_del(&cap->session_caps);
2358 head = msg->front.iov_base;
2359 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2361 item = msg->front.iov_base + msg->front.iov_len;
2362 item->ino = cpu_to_le64(cap->cap_ino);
2363 item->cap_id = cpu_to_le64(cap->cap_id);
2364 item->migrate_seq = cpu_to_le32(cap->mseq);
2365 item->seq = cpu_to_le32(cap->issue_seq);
2366 msg->front.iov_len += sizeof(*item);
2368 ceph_put_cap(mdsc, cap);
2370 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2371 // Append cap_barrier field
2372 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2373 *cap_barrier = barrier;
2374 msg->front.iov_len += sizeof(*cap_barrier);
2376 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2377 doutc(cl, "mds%d %p\n", session->s_mds, msg);
2378 ceph_con_send(&session->s_con, msg);
2383 BUG_ON(num_cap_releases != 0);
2385 spin_lock(&session->s_cap_lock);
2386 if (!list_empty(&session->s_cap_releases))
2388 spin_unlock(&session->s_cap_lock);
2391 // Append cap_barrier field
2392 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2393 *cap_barrier = barrier;
2394 msg->front.iov_len += sizeof(*cap_barrier);
2396 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2397 doutc(cl, "mds%d %p\n", session->s_mds, msg);
2398 ceph_con_send(&session->s_con, msg);
2402 pr_err_client(cl, "mds%d, failed to allocate message\n",
2404 spin_lock(&session->s_cap_lock);
2405 list_splice(&tmp_list, &session->s_cap_releases);
2406 session->s_num_cap_releases += num_cap_releases;
2407 spin_unlock(&session->s_cap_lock);
2410 static void ceph_cap_release_work(struct work_struct *work)
2412 struct ceph_mds_session *session =
2413 container_of(work, struct ceph_mds_session, s_cap_release_work);
2415 mutex_lock(&session->s_mutex);
2416 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2417 session->s_state == CEPH_MDS_SESSION_HUNG)
2418 ceph_send_cap_releases(session->s_mdsc, session);
2419 mutex_unlock(&session->s_mutex);
2420 ceph_put_mds_session(session);
2423 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2424 struct ceph_mds_session *session)
2426 struct ceph_client *cl = mdsc->fsc->client;
2430 ceph_get_mds_session(session);
2431 if (queue_work(mdsc->fsc->cap_wq,
2432 &session->s_cap_release_work)) {
2433 doutc(cl, "cap release work queued\n");
2435 ceph_put_mds_session(session);
2436 doutc(cl, "failed to queue cap release work\n");
2441 * caller holds session->s_cap_lock
2443 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2444 struct ceph_cap *cap)
2446 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2447 session->s_num_cap_releases++;
2449 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2450 ceph_flush_cap_releases(session->s_mdsc, session);
2453 static void ceph_cap_reclaim_work(struct work_struct *work)
2455 struct ceph_mds_client *mdsc =
2456 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2457 int ret = ceph_trim_dentries(mdsc);
2459 ceph_queue_cap_reclaim_work(mdsc);
2462 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2464 struct ceph_client *cl = mdsc->fsc->client;
2468 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2469 doutc(cl, "caps reclaim work queued\n");
2471 doutc(cl, "failed to queue caps release work\n");
2475 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2480 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2481 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2482 atomic_set(&mdsc->cap_reclaim_pending, 0);
2483 ceph_queue_cap_reclaim_work(mdsc);
2487 void ceph_queue_cap_unlink_work(struct ceph_mds_client *mdsc)
2489 struct ceph_client *cl = mdsc->fsc->client;
2493 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_unlink_work)) {
2494 doutc(cl, "caps unlink work queued\n");
2496 doutc(cl, "failed to queue caps unlink work\n");
2500 static void ceph_cap_unlink_work(struct work_struct *work)
2502 struct ceph_mds_client *mdsc =
2503 container_of(work, struct ceph_mds_client, cap_unlink_work);
2504 struct ceph_client *cl = mdsc->fsc->client;
2506 doutc(cl, "begin\n");
2507 spin_lock(&mdsc->cap_delay_lock);
2508 while (!list_empty(&mdsc->cap_unlink_delay_list)) {
2509 struct ceph_inode_info *ci;
2510 struct inode *inode;
2512 ci = list_first_entry(&mdsc->cap_unlink_delay_list,
2513 struct ceph_inode_info,
2515 list_del_init(&ci->i_cap_delay_list);
2517 inode = igrab(&ci->netfs.inode);
2519 spin_unlock(&mdsc->cap_delay_lock);
2520 doutc(cl, "on %p %llx.%llx\n", inode,
2522 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2524 spin_lock(&mdsc->cap_delay_lock);
2527 spin_unlock(&mdsc->cap_delay_lock);
2528 doutc(cl, "done\n");
2535 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2538 struct ceph_inode_info *ci = ceph_inode(dir);
2539 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2540 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2541 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2542 unsigned int num_entries;
2545 spin_lock(&ci->i_ceph_lock);
2546 num_entries = ci->i_files + ci->i_subdirs;
2547 spin_unlock(&ci->i_ceph_lock);
2548 num_entries = max(num_entries, 1U);
2549 num_entries = min(num_entries, opt->max_readdir);
2551 order = get_order(size * num_entries);
2552 while (order >= 0) {
2553 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2557 if (rinfo->dir_entries)
2561 if (!rinfo->dir_entries)
2564 num_entries = (PAGE_SIZE << order) / size;
2565 num_entries = min(num_entries, opt->max_readdir);
2567 rinfo->dir_buf_size = PAGE_SIZE << order;
2568 req->r_num_caps = num_entries + 1;
2569 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2570 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2575 * Create an mds request.
2577 struct ceph_mds_request *
2578 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2580 struct ceph_mds_request *req;
2582 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2584 return ERR_PTR(-ENOMEM);
2586 mutex_init(&req->r_fill_mutex);
2588 req->r_started = jiffies;
2589 req->r_start_latency = ktime_get();
2590 req->r_resend_mds = -1;
2591 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2592 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2594 req->r_feature_needed = -1;
2595 kref_init(&req->r_kref);
2596 RB_CLEAR_NODE(&req->r_node);
2597 INIT_LIST_HEAD(&req->r_wait);
2598 init_completion(&req->r_completion);
2599 init_completion(&req->r_safe_completion);
2600 INIT_LIST_HEAD(&req->r_unsafe_item);
2602 ktime_get_coarse_real_ts64(&req->r_stamp);
2605 req->r_direct_mode = mode;
2610 * return oldest (lowest) request, tid in request tree, 0 if none.
2612 * called under mdsc->mutex.
2614 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2616 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2618 return rb_entry(rb_first(&mdsc->request_tree),
2619 struct ceph_mds_request, r_node);
2622 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2624 return mdsc->oldest_tid;
2627 #if IS_ENABLED(CONFIG_FS_ENCRYPTION)
2628 static u8 *get_fscrypt_altname(const struct ceph_mds_request *req, u32 *plen)
2630 struct inode *dir = req->r_parent;
2631 struct dentry *dentry = req->r_dentry;
2632 u8 *cryptbuf = NULL;
2636 /* only encode if we have parent and dentry */
2637 if (!dir || !dentry)
2640 /* No-op unless this is encrypted */
2641 if (!IS_ENCRYPTED(dir))
2644 ret = ceph_fscrypt_prepare_readdir(dir);
2646 return ERR_PTR(ret);
2648 /* No key? Just ignore it. */
2649 if (!fscrypt_has_encryption_key(dir))
2652 if (!fscrypt_fname_encrypted_size(dir, dentry->d_name.len, NAME_MAX,
2655 return ERR_PTR(-ENAMETOOLONG);
2658 /* No need to append altname if name is short enough */
2659 if (len <= CEPH_NOHASH_NAME_MAX) {
2664 cryptbuf = kmalloc(len, GFP_KERNEL);
2666 return ERR_PTR(-ENOMEM);
2668 ret = fscrypt_fname_encrypt(dir, &dentry->d_name, cryptbuf, len);
2671 return ERR_PTR(ret);
2678 static u8 *get_fscrypt_altname(const struct ceph_mds_request *req, u32 *plen)
2686 * ceph_mdsc_build_path - build a path string to a given dentry
2688 * @dentry: dentry to which path should be built
2689 * @plen: returned length of string
2690 * @pbase: returned base inode number
2691 * @for_wire: is this path going to be sent to the MDS?
2693 * Build a string that represents the path to the dentry. This is mostly called
2694 * for two different purposes:
2696 * 1) we need to build a path string to send to the MDS (for_wire == true)
2697 * 2) we need a path string for local presentation (e.g. debugfs)
2698 * (for_wire == false)
2700 * The path is built in reverse, starting with the dentry. Walk back up toward
2701 * the root, building the path until the first non-snapped inode is reached
2702 * (for_wire) or the root inode is reached (!for_wire).
2704 * Encode hidden .snap dirs as a double /, i.e.
2705 * foo/.snap/bar -> foo//bar
2707 char *ceph_mdsc_build_path(struct ceph_mds_client *mdsc, struct dentry *dentry,
2708 int *plen, u64 *pbase, int for_wire)
2710 struct ceph_client *cl = mdsc->fsc->client;
2712 struct inode *inode;
2719 return ERR_PTR(-EINVAL);
2723 return ERR_PTR(-ENOMEM);
2728 seq = read_seqbegin(&rename_lock);
2731 struct dentry *parent;
2733 spin_lock(&cur->d_lock);
2734 inode = d_inode(cur);
2735 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2736 doutc(cl, "path+%d: %p SNAPDIR\n", pos, cur);
2737 spin_unlock(&cur->d_lock);
2738 parent = dget_parent(cur);
2739 } else if (for_wire && inode && dentry != cur &&
2740 ceph_snap(inode) == CEPH_NOSNAP) {
2741 spin_unlock(&cur->d_lock);
2742 pos++; /* get rid of any prepended '/' */
2744 } else if (!for_wire || !IS_ENCRYPTED(d_inode(cur->d_parent))) {
2745 pos -= cur->d_name.len;
2747 spin_unlock(&cur->d_lock);
2750 memcpy(path + pos, cur->d_name.name, cur->d_name.len);
2751 spin_unlock(&cur->d_lock);
2752 parent = dget_parent(cur);
2758 * Proactively copy name into buf, in case we need to
2761 memcpy(buf, cur->d_name.name, cur->d_name.len);
2762 len = cur->d_name.len;
2763 spin_unlock(&cur->d_lock);
2764 parent = dget_parent(cur);
2766 ret = ceph_fscrypt_prepare_readdir(d_inode(parent));
2770 return ERR_PTR(ret);
2773 if (fscrypt_has_encryption_key(d_inode(parent))) {
2774 len = ceph_encode_encrypted_fname(d_inode(parent),
2779 return ERR_PTR(len);
2787 memcpy(path + pos, buf, len);
2792 /* Are we at the root? */
2796 /* Are we out of buffer? */
2802 inode = d_inode(cur);
2803 base = inode ? ceph_ino(inode) : 0;
2806 if (read_seqretry(&rename_lock, seq))
2811 * A rename didn't occur, but somehow we didn't end up where
2812 * we thought we would. Throw a warning and try again.
2814 pr_warn_client(cl, "did not end path lookup where expected (pos = %d)\n",
2820 *plen = PATH_MAX - 1 - pos;
2821 doutc(cl, "on %p %d built %llx '%.*s'\n", dentry, d_count(dentry),
2822 base, *plen, path + pos);
2826 static int build_dentry_path(struct ceph_mds_client *mdsc, struct dentry *dentry,
2827 struct inode *dir, const char **ppath, int *ppathlen,
2828 u64 *pino, bool *pfreepath, bool parent_locked)
2834 dir = d_inode_rcu(dentry->d_parent);
2835 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP &&
2836 !IS_ENCRYPTED(dir)) {
2837 *pino = ceph_ino(dir);
2839 *ppath = dentry->d_name.name;
2840 *ppathlen = dentry->d_name.len;
2844 path = ceph_mdsc_build_path(mdsc, dentry, ppathlen, pino, 1);
2846 return PTR_ERR(path);
2852 static int build_inode_path(struct inode *inode,
2853 const char **ppath, int *ppathlen, u64 *pino,
2856 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
2857 struct dentry *dentry;
2860 if (ceph_snap(inode) == CEPH_NOSNAP) {
2861 *pino = ceph_ino(inode);
2865 dentry = d_find_alias(inode);
2866 path = ceph_mdsc_build_path(mdsc, dentry, ppathlen, pino, 1);
2869 return PTR_ERR(path);
2876 * request arguments may be specified via an inode *, a dentry *, or
2877 * an explicit ino+path.
2879 static int set_request_path_attr(struct ceph_mds_client *mdsc, struct inode *rinode,
2880 struct dentry *rdentry, struct inode *rdiri,
2881 const char *rpath, u64 rino, const char **ppath,
2882 int *pathlen, u64 *ino, bool *freepath,
2885 struct ceph_client *cl = mdsc->fsc->client;
2889 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2890 doutc(cl, " inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2892 } else if (rdentry) {
2893 r = build_dentry_path(mdsc, rdentry, rdiri, ppath, pathlen, ino,
2894 freepath, parent_locked);
2895 doutc(cl, " dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen, *ppath);
2896 } else if (rpath || rino) {
2899 *pathlen = rpath ? strlen(rpath) : 0;
2900 doutc(cl, " path %.*s\n", *pathlen, rpath);
2906 static void encode_mclientrequest_tail(void **p,
2907 const struct ceph_mds_request *req)
2909 struct ceph_timespec ts;
2912 ceph_encode_timespec64(&ts, &req->r_stamp);
2913 ceph_encode_copy(p, &ts, sizeof(ts));
2916 ceph_encode_32(p, req->r_cred->group_info->ngroups);
2917 for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2918 ceph_encode_64(p, from_kgid(&init_user_ns,
2919 req->r_cred->group_info->gid[i]));
2922 ceph_encode_32(p, req->r_altname_len);
2923 ceph_encode_copy(p, req->r_altname, req->r_altname_len);
2925 /* v6: fscrypt_auth and fscrypt_file */
2926 if (req->r_fscrypt_auth) {
2927 u32 authlen = ceph_fscrypt_auth_len(req->r_fscrypt_auth);
2929 ceph_encode_32(p, authlen);
2930 ceph_encode_copy(p, req->r_fscrypt_auth, authlen);
2932 ceph_encode_32(p, 0);
2934 if (test_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags)) {
2935 ceph_encode_32(p, sizeof(__le64));
2936 ceph_encode_64(p, req->r_fscrypt_file);
2938 ceph_encode_32(p, 0);
2942 static inline u16 mds_supported_head_version(struct ceph_mds_session *session)
2944 if (!test_bit(CEPHFS_FEATURE_32BITS_RETRY_FWD, &session->s_features))
2947 if (!test_bit(CEPHFS_FEATURE_HAS_OWNER_UIDGID, &session->s_features))
2950 return CEPH_MDS_REQUEST_HEAD_VERSION;
2953 static struct ceph_mds_request_head_legacy *
2954 find_legacy_request_head(void *p, u64 features)
2956 bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2957 struct ceph_mds_request_head_old *ohead;
2960 return (struct ceph_mds_request_head_legacy *)p;
2961 ohead = (struct ceph_mds_request_head_old *)p;
2962 return (struct ceph_mds_request_head_legacy *)&ohead->oldest_client_tid;
2966 * called under mdsc->mutex
2968 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2969 struct ceph_mds_request *req,
2970 bool drop_cap_releases)
2972 int mds = session->s_mds;
2973 struct ceph_mds_client *mdsc = session->s_mdsc;
2974 struct ceph_client *cl = mdsc->fsc->client;
2975 struct ceph_msg *msg;
2976 struct ceph_mds_request_head_legacy *lhead;
2977 const char *path1 = NULL;
2978 const char *path2 = NULL;
2979 u64 ino1 = 0, ino2 = 0;
2980 int pathlen1 = 0, pathlen2 = 0;
2981 bool freepath1 = false, freepath2 = false;
2982 struct dentry *old_dentry = NULL;
2987 bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2988 u16 request_head_version = mds_supported_head_version(session);
2989 kuid_t caller_fsuid = req->r_cred->fsuid;
2990 kgid_t caller_fsgid = req->r_cred->fsgid;
2992 ret = set_request_path_attr(mdsc, req->r_inode, req->r_dentry,
2993 req->r_parent, req->r_path1, req->r_ino1.ino,
2994 &path1, &pathlen1, &ino1, &freepath1,
2995 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2996 &req->r_req_flags));
3002 /* If r_old_dentry is set, then assume that its parent is locked */
3003 if (req->r_old_dentry &&
3004 !(req->r_old_dentry->d_flags & DCACHE_DISCONNECTED))
3005 old_dentry = req->r_old_dentry;
3006 ret = set_request_path_attr(mdsc, NULL, old_dentry,
3007 req->r_old_dentry_dir,
3008 req->r_path2, req->r_ino2.ino,
3009 &path2, &pathlen2, &ino2, &freepath2, true);
3015 req->r_altname = get_fscrypt_altname(req, &req->r_altname_len);
3016 if (IS_ERR(req->r_altname)) {
3017 msg = ERR_CAST(req->r_altname);
3018 req->r_altname = NULL;
3023 * For old cephs without supporting the 32bit retry/fwd feature
3024 * it will copy the raw memories directly when decoding the
3025 * requests. While new cephs will decode the head depending the
3026 * version member, so we need to make sure it will be compatible
3030 len = sizeof(struct ceph_mds_request_head_legacy);
3031 else if (request_head_version == 1)
3032 len = sizeof(struct ceph_mds_request_head_old);
3033 else if (request_head_version == 2)
3034 len = offsetofend(struct ceph_mds_request_head, ext_num_fwd);
3036 len = sizeof(struct ceph_mds_request_head);
3039 len += 2 * (1 + sizeof(u32) + sizeof(u64));
3040 len += pathlen1 + pathlen2;
3043 len += sizeof(struct ceph_mds_request_release) *
3044 (!!req->r_inode_drop + !!req->r_dentry_drop +
3045 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
3047 if (req->r_dentry_drop)
3049 if (req->r_old_dentry_drop)
3052 /* MClientRequest tail */
3055 len += sizeof(struct ceph_timespec);
3058 len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
3060 /* alternate name */
3061 len += sizeof(u32) + req->r_altname_len;
3064 len += sizeof(u32); // fscrypt_auth
3065 if (req->r_fscrypt_auth)
3066 len += ceph_fscrypt_auth_len(req->r_fscrypt_auth);
3070 if (test_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags))
3071 len += sizeof(__le64);
3073 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
3075 msg = ERR_PTR(-ENOMEM);
3079 msg->hdr.tid = cpu_to_le64(req->r_tid);
3081 lhead = find_legacy_request_head(msg->front.iov_base,
3082 session->s_con.peer_features);
3084 if ((req->r_mnt_idmap != &nop_mnt_idmap) &&
3085 !test_bit(CEPHFS_FEATURE_HAS_OWNER_UIDGID, &session->s_features)) {
3086 WARN_ON_ONCE(!IS_CEPH_MDS_OP_NEWINODE(req->r_op));
3088 if (enable_unsafe_idmap) {
3089 pr_warn_once_client(cl,
3090 "idmapped mount is used and CEPHFS_FEATURE_HAS_OWNER_UIDGID"
3091 " is not supported by MDS. UID/GID-based restrictions may"
3092 " not work properly.\n");
3094 caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
3095 VFSUIDT_INIT(req->r_cred->fsuid));
3096 caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
3097 VFSGIDT_INIT(req->r_cred->fsgid));
3099 pr_err_ratelimited_client(cl,
3100 "idmapped mount is used and CEPHFS_FEATURE_HAS_OWNER_UIDGID"
3101 " is not supported by MDS. Fail request with -EIO.\n");
3109 * The ceph_mds_request_head_legacy didn't contain a version field, and
3110 * one was added when we moved the message version from 3->4.
3113 msg->hdr.version = cpu_to_le16(3);
3114 p = msg->front.iov_base + sizeof(*lhead);
3115 } else if (request_head_version == 1) {
3116 struct ceph_mds_request_head_old *ohead = msg->front.iov_base;
3118 msg->hdr.version = cpu_to_le16(4);
3119 ohead->version = cpu_to_le16(1);
3120 p = msg->front.iov_base + sizeof(*ohead);
3121 } else if (request_head_version == 2) {
3122 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3124 msg->hdr.version = cpu_to_le16(6);
3125 nhead->version = cpu_to_le16(2);
3127 p = msg->front.iov_base + offsetofend(struct ceph_mds_request_head, ext_num_fwd);
3129 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3133 msg->hdr.version = cpu_to_le16(6);
3134 nhead->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
3135 nhead->struct_len = cpu_to_le32(sizeof(struct ceph_mds_request_head));
3137 if (IS_CEPH_MDS_OP_NEWINODE(req->r_op)) {
3138 owner_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
3139 VFSUIDT_INIT(req->r_cred->fsuid));
3140 owner_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
3141 VFSGIDT_INIT(req->r_cred->fsgid));
3142 nhead->owner_uid = cpu_to_le32(from_kuid(&init_user_ns, owner_fsuid));
3143 nhead->owner_gid = cpu_to_le32(from_kgid(&init_user_ns, owner_fsgid));
3145 nhead->owner_uid = cpu_to_le32(-1);
3146 nhead->owner_gid = cpu_to_le32(-1);
3149 p = msg->front.iov_base + sizeof(*nhead);
3152 end = msg->front.iov_base + msg->front.iov_len;
3154 lhead->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
3155 lhead->op = cpu_to_le32(req->r_op);
3156 lhead->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
3158 lhead->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
3160 lhead->ino = cpu_to_le64(req->r_deleg_ino);
3161 lhead->args = req->r_args;
3163 ceph_encode_filepath(&p, end, ino1, path1);
3164 ceph_encode_filepath(&p, end, ino2, path2);
3166 /* make note of release offset, in case we need to replay */
3167 req->r_request_release_offset = p - msg->front.iov_base;
3171 if (req->r_inode_drop)
3172 releases += ceph_encode_inode_release(&p,
3173 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
3174 mds, req->r_inode_drop, req->r_inode_unless,
3175 req->r_op == CEPH_MDS_OP_READDIR);
3176 if (req->r_dentry_drop) {
3177 ret = ceph_encode_dentry_release(&p, req->r_dentry,
3178 req->r_parent, mds, req->r_dentry_drop,
3179 req->r_dentry_unless);
3184 if (req->r_old_dentry_drop) {
3185 ret = ceph_encode_dentry_release(&p, req->r_old_dentry,
3186 req->r_old_dentry_dir, mds,
3187 req->r_old_dentry_drop,
3188 req->r_old_dentry_unless);
3193 if (req->r_old_inode_drop)
3194 releases += ceph_encode_inode_release(&p,
3195 d_inode(req->r_old_dentry),
3196 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
3198 if (drop_cap_releases) {
3200 p = msg->front.iov_base + req->r_request_release_offset;
3203 lhead->num_releases = cpu_to_le16(releases);
3205 encode_mclientrequest_tail(&p, req);
3207 if (WARN_ON_ONCE(p > end)) {
3209 msg = ERR_PTR(-ERANGE);
3213 msg->front.iov_len = p - msg->front.iov_base;
3214 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
3216 if (req->r_pagelist) {
3217 struct ceph_pagelist *pagelist = req->r_pagelist;
3218 ceph_msg_data_add_pagelist(msg, pagelist);
3219 msg->hdr.data_len = cpu_to_le32(pagelist->length);
3221 msg->hdr.data_len = 0;
3224 msg->hdr.data_off = cpu_to_le16(0);
3228 ceph_mdsc_free_path((char *)path2, pathlen2);
3231 ceph_mdsc_free_path((char *)path1, pathlen1);
3241 * called under mdsc->mutex if error, under no mutex if
3244 static void complete_request(struct ceph_mds_client *mdsc,
3245 struct ceph_mds_request *req)
3247 req->r_end_latency = ktime_get();
3249 if (req->r_callback)
3250 req->r_callback(mdsc, req);
3251 complete_all(&req->r_completion);
3255 * called under mdsc->mutex
3257 static int __prepare_send_request(struct ceph_mds_session *session,
3258 struct ceph_mds_request *req,
3259 bool drop_cap_releases)
3261 int mds = session->s_mds;
3262 struct ceph_mds_client *mdsc = session->s_mdsc;
3263 struct ceph_client *cl = mdsc->fsc->client;
3264 struct ceph_mds_request_head_legacy *lhead;
3265 struct ceph_mds_request_head *nhead;
3266 struct ceph_msg *msg;
3267 int flags = 0, old_max_retry;
3268 bool old_version = !test_bit(CEPHFS_FEATURE_32BITS_RETRY_FWD,
3269 &session->s_features);
3272 * Avoid inifinite retrying after overflow. The client will
3273 * increase the retry count and if the MDS is old version,
3274 * so we limit to retry at most 256 times.
3276 if (req->r_attempts) {
3277 old_max_retry = sizeof_field(struct ceph_mds_request_head_old,
3279 old_max_retry = 1 << (old_max_retry * BITS_PER_BYTE);
3280 if ((old_version && req->r_attempts >= old_max_retry) ||
3281 ((uint32_t)req->r_attempts >= U32_MAX)) {
3282 pr_warn_ratelimited_client(cl, "request tid %llu seq overflow\n",
3290 struct ceph_cap *cap =
3291 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
3294 req->r_sent_on_mseq = cap->mseq;
3296 req->r_sent_on_mseq = -1;
3298 doutc(cl, "%p tid %lld %s (attempt %d)\n", req, req->r_tid,
3299 ceph_mds_op_name(req->r_op), req->r_attempts);
3301 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3305 * Replay. Do not regenerate message (and rebuild
3306 * paths, etc.); just use the original message.
3307 * Rebuilding paths will break for renames because
3308 * d_move mangles the src name.
3310 msg = req->r_request;
3311 lhead = find_legacy_request_head(msg->front.iov_base,
3312 session->s_con.peer_features);
3314 flags = le32_to_cpu(lhead->flags);
3315 flags |= CEPH_MDS_FLAG_REPLAY;
3316 lhead->flags = cpu_to_le32(flags);
3318 if (req->r_target_inode)
3319 lhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
3321 lhead->num_retry = req->r_attempts - 1;
3323 nhead = (struct ceph_mds_request_head*)msg->front.iov_base;
3324 nhead->ext_num_retry = cpu_to_le32(req->r_attempts - 1);
3327 /* remove cap/dentry releases from message */
3328 lhead->num_releases = 0;
3330 p = msg->front.iov_base + req->r_request_release_offset;
3331 encode_mclientrequest_tail(&p, req);
3333 msg->front.iov_len = p - msg->front.iov_base;
3334 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
3338 if (req->r_request) {
3339 ceph_msg_put(req->r_request);
3340 req->r_request = NULL;
3342 msg = create_request_message(session, req, drop_cap_releases);
3344 req->r_err = PTR_ERR(msg);
3345 return PTR_ERR(msg);
3347 req->r_request = msg;
3349 lhead = find_legacy_request_head(msg->front.iov_base,
3350 session->s_con.peer_features);
3351 lhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
3352 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3353 flags |= CEPH_MDS_FLAG_REPLAY;
3354 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
3355 flags |= CEPH_MDS_FLAG_ASYNC;
3357 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
3358 lhead->flags = cpu_to_le32(flags);
3359 lhead->num_fwd = req->r_num_fwd;
3360 lhead->num_retry = req->r_attempts - 1;
3362 nhead = (struct ceph_mds_request_head*)msg->front.iov_base;
3363 nhead->ext_num_fwd = cpu_to_le32(req->r_num_fwd);
3364 nhead->ext_num_retry = cpu_to_le32(req->r_attempts - 1);
3367 doutc(cl, " r_parent = %p\n", req->r_parent);
3372 * called under mdsc->mutex
3374 static int __send_request(struct ceph_mds_session *session,
3375 struct ceph_mds_request *req,
3376 bool drop_cap_releases)
3380 err = __prepare_send_request(session, req, drop_cap_releases);
3382 ceph_msg_get(req->r_request);
3383 ceph_con_send(&session->s_con, req->r_request);
3390 * send request, or put it on the appropriate wait list.
3392 static void __do_request(struct ceph_mds_client *mdsc,
3393 struct ceph_mds_request *req)
3395 struct ceph_client *cl = mdsc->fsc->client;
3396 struct ceph_mds_session *session = NULL;
3401 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3402 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
3403 __unregister_request(mdsc, req);
3407 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO) {
3408 doutc(cl, "metadata corrupted\n");
3412 if (req->r_timeout &&
3413 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
3414 doutc(cl, "timed out\n");
3418 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
3419 doutc(cl, "forced umount\n");
3423 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
3424 if (mdsc->mdsmap_err) {
3425 err = mdsc->mdsmap_err;
3426 doutc(cl, "mdsmap err %d\n", err);
3429 if (mdsc->mdsmap->m_epoch == 0) {
3430 doutc(cl, "no mdsmap, waiting for map\n");
3431 list_add(&req->r_wait, &mdsc->waiting_for_map);
3434 if (!(mdsc->fsc->mount_options->flags &
3435 CEPH_MOUNT_OPT_MOUNTWAIT) &&
3436 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
3437 err = -EHOSTUNREACH;
3442 put_request_session(req);
3444 mds = __choose_mds(mdsc, req, &random);
3446 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
3447 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
3451 doutc(cl, "no mds or not active, waiting for map\n");
3452 list_add(&req->r_wait, &mdsc->waiting_for_map);
3456 /* get, open session */
3457 session = __ceph_lookup_mds_session(mdsc, mds);
3459 session = register_session(mdsc, mds);
3460 if (IS_ERR(session)) {
3461 err = PTR_ERR(session);
3465 req->r_session = ceph_get_mds_session(session);
3467 doutc(cl, "mds%d session %p state %s\n", mds, session,
3468 ceph_session_state_name(session->s_state));
3471 * The old ceph will crash the MDSs when see unknown OPs
3473 if (req->r_feature_needed > 0 &&
3474 !test_bit(req->r_feature_needed, &session->s_features)) {
3479 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
3480 session->s_state != CEPH_MDS_SESSION_HUNG) {
3482 * We cannot queue async requests since the caps and delegated
3483 * inodes are bound to the session. Just return -EJUKEBOX and
3484 * let the caller retry a sync request in that case.
3486 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
3492 * If the session has been REJECTED, then return a hard error,
3493 * unless it's a CLEANRECOVER mount, in which case we'll queue
3494 * it to the mdsc queue.
3496 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
3497 if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
3498 list_add(&req->r_wait, &mdsc->waiting_for_map);
3504 if (session->s_state == CEPH_MDS_SESSION_NEW ||
3505 session->s_state == CEPH_MDS_SESSION_CLOSING) {
3506 err = __open_session(mdsc, session);
3509 /* retry the same mds later */
3511 req->r_resend_mds = mds;
3513 list_add(&req->r_wait, &session->s_waiting);
3518 req->r_resend_mds = -1; /* forget any previous mds hint */
3520 if (req->r_request_started == 0) /* note request start time */
3521 req->r_request_started = jiffies;
3524 * For async create we will choose the auth MDS of frag in parent
3525 * directory to send the request and ususally this works fine, but
3526 * if the migrated the dirtory to another MDS before it could handle
3527 * it the request will be forwarded.
3529 * And then the auth cap will be changed.
3531 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags) && req->r_num_fwd) {
3532 struct ceph_dentry_info *di = ceph_dentry(req->r_dentry);
3533 struct ceph_inode_info *ci;
3534 struct ceph_cap *cap;
3537 * The request maybe handled very fast and the new inode
3538 * hasn't been linked to the dentry yet. We need to wait
3539 * for the ceph_finish_async_create(), which shouldn't be
3540 * stuck too long or fail in thoery, to finish when forwarding
3543 if (!d_inode(req->r_dentry)) {
3544 err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT,
3547 mutex_lock(&req->r_fill_mutex);
3548 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3549 mutex_unlock(&req->r_fill_mutex);
3554 ci = ceph_inode(d_inode(req->r_dentry));
3556 spin_lock(&ci->i_ceph_lock);
3557 cap = ci->i_auth_cap;
3558 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE && mds != cap->mds) {
3559 doutc(cl, "session changed for auth cap %d -> %d\n",
3560 cap->session->s_mds, session->s_mds);
3562 /* Remove the auth cap from old session */
3563 spin_lock(&cap->session->s_cap_lock);
3564 cap->session->s_nr_caps--;
3565 list_del_init(&cap->session_caps);
3566 spin_unlock(&cap->session->s_cap_lock);
3568 /* Add the auth cap to the new session */
3570 cap->session = session;
3571 spin_lock(&session->s_cap_lock);
3572 session->s_nr_caps++;
3573 list_add_tail(&cap->session_caps, &session->s_caps);
3574 spin_unlock(&session->s_cap_lock);
3576 change_auth_cap_ses(ci, session);
3578 spin_unlock(&ci->i_ceph_lock);
3581 err = __send_request(session, req, false);
3584 ceph_put_mds_session(session);
3587 doutc(cl, "early error %d\n", err);
3589 complete_request(mdsc, req);
3590 __unregister_request(mdsc, req);
3596 * called under mdsc->mutex
3598 static void __wake_requests(struct ceph_mds_client *mdsc,
3599 struct list_head *head)
3601 struct ceph_client *cl = mdsc->fsc->client;
3602 struct ceph_mds_request *req;
3603 LIST_HEAD(tmp_list);
3605 list_splice_init(head, &tmp_list);
3607 while (!list_empty(&tmp_list)) {
3608 req = list_entry(tmp_list.next,
3609 struct ceph_mds_request, r_wait);
3610 list_del_init(&req->r_wait);
3611 doutc(cl, " wake request %p tid %llu\n", req,
3613 __do_request(mdsc, req);
3618 * Wake up threads with requests pending for @mds, so that they can
3619 * resubmit their requests to a possibly different mds.
3621 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
3623 struct ceph_client *cl = mdsc->fsc->client;
3624 struct ceph_mds_request *req;
3625 struct rb_node *p = rb_first(&mdsc->request_tree);
3627 doutc(cl, "kick_requests mds%d\n", mds);
3629 req = rb_entry(p, struct ceph_mds_request, r_node);
3631 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3633 if (req->r_attempts > 0)
3634 continue; /* only new requests */
3635 if (req->r_session &&
3636 req->r_session->s_mds == mds) {
3637 doutc(cl, " kicking tid %llu\n", req->r_tid);
3638 list_del_init(&req->r_wait);
3639 __do_request(mdsc, req);
3644 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
3645 struct ceph_mds_request *req)
3647 struct ceph_client *cl = mdsc->fsc->client;
3650 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
3652 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3653 if (req->r_parent) {
3654 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3655 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3656 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3657 spin_lock(&ci->i_ceph_lock);
3658 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3659 __ceph_touch_fmode(ci, mdsc, fmode);
3660 spin_unlock(&ci->i_ceph_lock);
3662 if (req->r_old_dentry_dir)
3663 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3667 err = ceph_wait_on_async_create(req->r_inode);
3669 doutc(cl, "wait for async create returned: %d\n", err);
3674 if (!err && req->r_old_inode) {
3675 err = ceph_wait_on_async_create(req->r_old_inode);
3677 doutc(cl, "wait for async create returned: %d\n", err);
3682 doutc(cl, "submit_request on %p for inode %p\n", req, dir);
3683 mutex_lock(&mdsc->mutex);
3684 __register_request(mdsc, req, dir);
3685 __do_request(mdsc, req);
3687 mutex_unlock(&mdsc->mutex);
3691 int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3692 struct ceph_mds_request *req,
3693 ceph_mds_request_wait_callback_t wait_func)
3695 struct ceph_client *cl = mdsc->fsc->client;
3699 doutc(cl, "do_request waiting\n");
3701 err = wait_func(mdsc, req);
3703 long timeleft = wait_for_completion_killable_timeout(
3705 ceph_timeout_jiffies(req->r_timeout));
3709 err = -ETIMEDOUT; /* timed out */
3711 err = timeleft; /* killed */
3713 doutc(cl, "do_request waited, got %d\n", err);
3714 mutex_lock(&mdsc->mutex);
3716 /* only abort if we didn't race with a real reply */
3717 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3718 err = le32_to_cpu(req->r_reply_info.head->result);
3719 } else if (err < 0) {
3720 doutc(cl, "aborted request %lld with %d\n", req->r_tid, err);
3723 * ensure we aren't running concurrently with
3724 * ceph_fill_trace or ceph_readdir_prepopulate, which
3725 * rely on locks (dir mutex) held by our caller.
3727 mutex_lock(&req->r_fill_mutex);
3729 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3730 mutex_unlock(&req->r_fill_mutex);
3732 if (req->r_parent &&
3733 (req->r_op & CEPH_MDS_OP_WRITE))
3734 ceph_invalidate_dir_request(req);
3739 mutex_unlock(&mdsc->mutex);
3744 * Synchrously perform an mds request. Take care of all of the
3745 * session setup, forwarding, retry details.
3747 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3749 struct ceph_mds_request *req)
3751 struct ceph_client *cl = mdsc->fsc->client;
3754 doutc(cl, "do_request on %p\n", req);
3757 err = ceph_mdsc_submit_request(mdsc, dir, req);
3759 err = ceph_mdsc_wait_request(mdsc, req, NULL);
3760 doutc(cl, "do_request %p done, result %d\n", req, err);
3765 * Invalidate dir's completeness, dentry lease state on an aborted MDS
3766 * namespace request.
3768 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3770 struct inode *dir = req->r_parent;
3771 struct inode *old_dir = req->r_old_dentry_dir;
3772 struct ceph_client *cl = req->r_mdsc->fsc->client;
3774 doutc(cl, "invalidate_dir_request %p %p (complete, lease(s))\n",
3777 ceph_dir_clear_complete(dir);
3779 ceph_dir_clear_complete(old_dir);
3781 ceph_invalidate_dentry_lease(req->r_dentry);
3782 if (req->r_old_dentry)
3783 ceph_invalidate_dentry_lease(req->r_old_dentry);
3789 * We take the session mutex and parse and process the reply immediately.
3790 * This preserves the logical ordering of replies, capabilities, etc., sent
3791 * by the MDS as they are applied to our local cache.
3793 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3795 struct ceph_mds_client *mdsc = session->s_mdsc;
3796 struct ceph_client *cl = mdsc->fsc->client;
3797 struct ceph_mds_request *req;
3798 struct ceph_mds_reply_head *head = msg->front.iov_base;
3799 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
3800 struct ceph_snap_realm *realm;
3803 int mds = session->s_mds;
3804 bool close_sessions = false;
3806 if (msg->front.iov_len < sizeof(*head)) {
3807 pr_err_client(cl, "got corrupt (short) reply\n");
3812 /* get request, session */
3813 tid = le64_to_cpu(msg->hdr.tid);
3814 mutex_lock(&mdsc->mutex);
3815 req = lookup_get_request(mdsc, tid);
3817 doutc(cl, "on unknown tid %llu\n", tid);
3818 mutex_unlock(&mdsc->mutex);
3821 doutc(cl, "handle_reply %p\n", req);
3823 /* correct session? */
3824 if (req->r_session != session) {
3825 pr_err_client(cl, "got %llu on session mds%d not mds%d\n",
3826 tid, session->s_mds,
3827 req->r_session ? req->r_session->s_mds : -1);
3828 mutex_unlock(&mdsc->mutex);
3833 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3834 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3835 pr_warn_client(cl, "got a dup %s reply on %llu from mds%d\n",
3836 head->safe ? "safe" : "unsafe", tid, mds);
3837 mutex_unlock(&mdsc->mutex);
3840 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3841 pr_warn_client(cl, "got unsafe after safe on %llu from mds%d\n",
3843 mutex_unlock(&mdsc->mutex);
3847 result = le32_to_cpu(head->result);
3850 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3851 __unregister_request(mdsc, req);
3853 /* last request during umount? */
3854 if (mdsc->stopping && !__get_oldest_req(mdsc))
3855 complete_all(&mdsc->safe_umount_waiters);
3857 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3859 * We already handled the unsafe response, now do the
3860 * cleanup. No need to examine the response; the MDS
3861 * doesn't include any result info in the safe
3862 * response. And even if it did, there is nothing
3863 * useful we could do with a revised return value.
3865 doutc(cl, "got safe reply %llu, mds%d\n", tid, mds);
3867 mutex_unlock(&mdsc->mutex);
3871 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3872 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3875 doutc(cl, "tid %lld result %d\n", tid, result);
3876 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3877 err = parse_reply_info(session, msg, req, (u64)-1);
3879 err = parse_reply_info(session, msg, req,
3880 session->s_con.peer_features);
3881 mutex_unlock(&mdsc->mutex);
3883 /* Must find target inode outside of mutexes to avoid deadlocks */
3884 rinfo = &req->r_reply_info;
3885 if ((err >= 0) && rinfo->head->is_target) {
3886 struct inode *in = xchg(&req->r_new_inode, NULL);
3887 struct ceph_vino tvino = {
3888 .ino = le64_to_cpu(rinfo->targeti.in->ino),
3889 .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3893 * If we ended up opening an existing inode, discard
3896 if (req->r_op == CEPH_MDS_OP_CREATE &&
3897 !req->r_reply_info.has_create_ino) {
3898 /* This should never happen on an async create */
3899 WARN_ON_ONCE(req->r_deleg_ino);
3904 in = ceph_get_inode(mdsc->fsc->sb, tvino, in);
3907 mutex_lock(&session->s_mutex);
3910 req->r_target_inode = in;
3913 mutex_lock(&session->s_mutex);
3915 pr_err_client(cl, "got corrupt reply mds%d(tid:%lld)\n",
3923 if (rinfo->snapblob_len) {
3924 down_write(&mdsc->snap_rwsem);
3925 err = ceph_update_snap_trace(mdsc, rinfo->snapblob,
3926 rinfo->snapblob + rinfo->snapblob_len,
3927 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3930 up_write(&mdsc->snap_rwsem);
3931 close_sessions = true;
3936 downgrade_write(&mdsc->snap_rwsem);
3938 down_read(&mdsc->snap_rwsem);
3941 /* insert trace into our cache */
3942 mutex_lock(&req->r_fill_mutex);
3943 current->journal_info = req;
3944 err = ceph_fill_trace(mdsc->fsc->sb, req);
3946 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3947 req->r_op == CEPH_MDS_OP_LSSNAP))
3948 err = ceph_readdir_prepopulate(req, req->r_session);
3950 current->journal_info = NULL;
3951 mutex_unlock(&req->r_fill_mutex);
3953 up_read(&mdsc->snap_rwsem);
3955 ceph_put_snap_realm(mdsc, realm);
3958 if (req->r_target_inode &&
3959 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3960 struct ceph_inode_info *ci =
3961 ceph_inode(req->r_target_inode);
3962 spin_lock(&ci->i_unsafe_lock);
3963 list_add_tail(&req->r_unsafe_target_item,
3964 &ci->i_unsafe_iops);
3965 spin_unlock(&ci->i_unsafe_lock);
3968 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3971 mutex_lock(&mdsc->mutex);
3972 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3976 req->r_reply = ceph_msg_get(msg);
3977 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3980 doutc(cl, "reply arrived after request %lld was aborted\n", tid);
3982 mutex_unlock(&mdsc->mutex);
3984 mutex_unlock(&session->s_mutex);
3986 /* kick calling process */
3987 complete_request(mdsc, req);
3989 ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3990 req->r_end_latency, err);
3992 ceph_mdsc_put_request(req);
3994 /* Defer closing the sessions after s_mutex lock being released */
3996 ceph_mdsc_close_sessions(mdsc);
4003 * handle mds notification that our request has been forwarded.
4005 static void handle_forward(struct ceph_mds_client *mdsc,
4006 struct ceph_mds_session *session,
4007 struct ceph_msg *msg)
4009 struct ceph_client *cl = mdsc->fsc->client;
4010 struct ceph_mds_request *req;
4011 u64 tid = le64_to_cpu(msg->hdr.tid);
4015 void *p = msg->front.iov_base;
4016 void *end = p + msg->front.iov_len;
4017 bool aborted = false;
4019 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
4020 next_mds = ceph_decode_32(&p);
4021 fwd_seq = ceph_decode_32(&p);
4023 mutex_lock(&mdsc->mutex);
4024 req = lookup_get_request(mdsc, tid);
4026 mutex_unlock(&mdsc->mutex);
4027 doutc(cl, "forward tid %llu to mds%d - req dne\n", tid, next_mds);
4028 return; /* dup reply? */
4031 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
4032 doutc(cl, "forward tid %llu aborted, unregistering\n", tid);
4033 __unregister_request(mdsc, req);
4034 } else if (fwd_seq <= req->r_num_fwd || (uint32_t)fwd_seq >= U32_MAX) {
4036 * Avoid inifinite retrying after overflow.
4038 * The MDS will increase the fwd count and in client side
4039 * if the num_fwd is less than the one saved in request
4040 * that means the MDS is an old version and overflowed of
4043 mutex_lock(&req->r_fill_mutex);
4044 req->r_err = -EMULTIHOP;
4045 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
4046 mutex_unlock(&req->r_fill_mutex);
4048 pr_warn_ratelimited_client(cl, "forward tid %llu seq overflow\n",
4051 /* resend. forward race not possible; mds would drop */
4052 doutc(cl, "forward tid %llu to mds%d (we resend)\n", tid, next_mds);
4054 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
4055 req->r_attempts = 0;
4056 req->r_num_fwd = fwd_seq;
4057 req->r_resend_mds = next_mds;
4058 put_request_session(req);
4059 __do_request(mdsc, req);
4061 mutex_unlock(&mdsc->mutex);
4063 /* kick calling process */
4065 complete_request(mdsc, req);
4066 ceph_mdsc_put_request(req);
4070 pr_err_client(cl, "decode error err=%d\n", err);
4074 static int __decode_session_metadata(void **p, void *end,
4077 /* map<string,string> */
4080 ceph_decode_32_safe(p, end, n, bad);
4083 ceph_decode_32_safe(p, end, len, bad);
4084 ceph_decode_need(p, end, len, bad);
4085 err_str = !strncmp(*p, "error_string", len);
4087 ceph_decode_32_safe(p, end, len, bad);
4088 ceph_decode_need(p, end, len, bad);
4090 * Match "blocklisted (blacklisted)" from newer MDSes,
4091 * or "blacklisted" from older MDSes.
4093 if (err_str && strnstr(*p, "blacklisted", len))
4094 *blocklisted = true;
4103 * handle a mds session control message
4105 static void handle_session(struct ceph_mds_session *session,
4106 struct ceph_msg *msg)
4108 struct ceph_mds_client *mdsc = session->s_mdsc;
4109 struct ceph_client *cl = mdsc->fsc->client;
4110 int mds = session->s_mds;
4111 int msg_version = le16_to_cpu(msg->hdr.version);
4112 void *p = msg->front.iov_base;
4113 void *end = p + msg->front.iov_len;
4114 struct ceph_mds_session_head *h;
4116 u64 seq, features = 0;
4118 bool blocklisted = false;
4121 ceph_decode_need(&p, end, sizeof(*h), bad);
4125 op = le32_to_cpu(h->op);
4126 seq = le64_to_cpu(h->seq);
4128 if (msg_version >= 3) {
4130 /* version >= 2 and < 5, decode metadata, skip otherwise
4131 * as it's handled via flags.
4133 if (msg_version >= 5)
4134 ceph_decode_skip_map(&p, end, string, string, bad);
4135 else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
4138 /* version >= 3, feature bits */
4139 ceph_decode_32_safe(&p, end, len, bad);
4141 ceph_decode_64_safe(&p, end, features, bad);
4142 p += len - sizeof(features);
4146 if (msg_version >= 5) {
4150 ceph_decode_skip_16(&p, end, bad); /* struct_v, struct_cv */
4151 ceph_decode_32_safe(&p, end, len, bad); /* len */
4152 ceph_decode_skip_n(&p, end, len, bad); /* metric_spec */
4154 /* version >= 5, flags */
4155 ceph_decode_32_safe(&p, end, flags, bad);
4156 if (flags & CEPH_SESSION_BLOCKLISTED) {
4157 pr_warn_client(cl, "mds%d session blocklisted\n",
4163 mutex_lock(&mdsc->mutex);
4164 if (op == CEPH_SESSION_CLOSE) {
4165 ceph_get_mds_session(session);
4166 __unregister_session(mdsc, session);
4168 /* FIXME: this ttl calculation is generous */
4169 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
4170 mutex_unlock(&mdsc->mutex);
4172 mutex_lock(&session->s_mutex);
4174 doutc(cl, "mds%d %s %p state %s seq %llu\n", mds,
4175 ceph_session_op_name(op), session,
4176 ceph_session_state_name(session->s_state), seq);
4178 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
4179 session->s_state = CEPH_MDS_SESSION_OPEN;
4180 pr_info_client(cl, "mds%d came back\n", session->s_mds);
4184 case CEPH_SESSION_OPEN:
4185 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
4186 pr_info_client(cl, "mds%d reconnect success\n",
4189 session->s_features = features;
4190 if (session->s_state == CEPH_MDS_SESSION_OPEN) {
4191 pr_notice_client(cl, "mds%d is already opened\n",
4194 session->s_state = CEPH_MDS_SESSION_OPEN;
4195 renewed_caps(mdsc, session, 0);
4196 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT,
4197 &session->s_features))
4198 metric_schedule_delayed(&mdsc->metric);
4202 * The connection maybe broken and the session in client
4203 * side has been reinitialized, need to update the seq
4206 if (!session->s_seq && seq)
4207 session->s_seq = seq;
4211 __close_session(mdsc, session);
4214 case CEPH_SESSION_RENEWCAPS:
4215 if (session->s_renew_seq == seq)
4216 renewed_caps(mdsc, session, 1);
4219 case CEPH_SESSION_CLOSE:
4220 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
4221 pr_info_client(cl, "mds%d reconnect denied\n",
4223 session->s_state = CEPH_MDS_SESSION_CLOSED;
4224 cleanup_session_requests(mdsc, session);
4225 remove_session_caps(session);
4226 wake = 2; /* for good measure */
4227 wake_up_all(&mdsc->session_close_wq);
4230 case CEPH_SESSION_STALE:
4231 pr_info_client(cl, "mds%d caps went stale, renewing\n",
4233 atomic_inc(&session->s_cap_gen);
4234 session->s_cap_ttl = jiffies - 1;
4235 send_renew_caps(mdsc, session);
4238 case CEPH_SESSION_RECALL_STATE:
4239 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
4242 case CEPH_SESSION_FLUSHMSG:
4243 /* flush cap releases */
4244 spin_lock(&session->s_cap_lock);
4245 if (session->s_num_cap_releases)
4246 ceph_flush_cap_releases(mdsc, session);
4247 spin_unlock(&session->s_cap_lock);
4249 send_flushmsg_ack(mdsc, session, seq);
4252 case CEPH_SESSION_FORCE_RO:
4253 doutc(cl, "force_session_readonly %p\n", session);
4254 spin_lock(&session->s_cap_lock);
4255 session->s_readonly = true;
4256 spin_unlock(&session->s_cap_lock);
4257 wake_up_session_caps(session, FORCE_RO);
4260 case CEPH_SESSION_REJECT:
4261 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
4262 pr_info_client(cl, "mds%d rejected session\n",
4264 session->s_state = CEPH_MDS_SESSION_REJECTED;
4265 cleanup_session_requests(mdsc, session);
4266 remove_session_caps(session);
4268 mdsc->fsc->blocklisted = true;
4269 wake = 2; /* for good measure */
4273 pr_err_client(cl, "bad op %d mds%d\n", op, mds);
4277 mutex_unlock(&session->s_mutex);
4279 mutex_lock(&mdsc->mutex);
4280 __wake_requests(mdsc, &session->s_waiting);
4282 kick_requests(mdsc, mds);
4283 mutex_unlock(&mdsc->mutex);
4285 if (op == CEPH_SESSION_CLOSE)
4286 ceph_put_mds_session(session);
4290 pr_err_client(cl, "corrupt message mds%d len %d\n", mds,
4291 (int)msg->front.iov_len);
4296 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
4298 struct ceph_client *cl = req->r_mdsc->fsc->client;
4301 dcaps = xchg(&req->r_dir_caps, 0);
4303 doutc(cl, "releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
4304 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
4308 void ceph_mdsc_release_dir_caps_async(struct ceph_mds_request *req)
4310 struct ceph_client *cl = req->r_mdsc->fsc->client;
4313 dcaps = xchg(&req->r_dir_caps, 0);
4315 doutc(cl, "releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
4316 ceph_put_cap_refs_async(ceph_inode(req->r_parent), dcaps);
4321 * called under session->mutex.
4323 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
4324 struct ceph_mds_session *session)
4326 struct ceph_mds_request *req, *nreq;
4329 doutc(mdsc->fsc->client, "mds%d\n", session->s_mds);
4331 mutex_lock(&mdsc->mutex);
4332 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
4333 __send_request(session, req, true);
4336 * also re-send old requests when MDS enters reconnect stage. So that MDS
4337 * can process completed request in clientreplay stage.
4339 p = rb_first(&mdsc->request_tree);
4341 req = rb_entry(p, struct ceph_mds_request, r_node);
4343 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
4345 if (req->r_attempts == 0)
4346 continue; /* only old requests */
4347 if (!req->r_session)
4349 if (req->r_session->s_mds != session->s_mds)
4352 ceph_mdsc_release_dir_caps_async(req);
4354 __send_request(session, req, true);
4356 mutex_unlock(&mdsc->mutex);
4359 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
4361 struct ceph_msg *reply;
4362 struct ceph_pagelist *_pagelist;
4367 if (!recon_state->allow_multi)
4370 /* can't handle message that contains both caps and realm */
4371 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
4373 /* pre-allocate new pagelist */
4374 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
4378 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4382 /* placeholder for nr_caps */
4383 err = ceph_pagelist_encode_32(_pagelist, 0);
4387 if (recon_state->nr_caps) {
4388 /* currently encoding caps */
4389 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
4393 /* placeholder for nr_realms (currently encoding relams) */
4394 err = ceph_pagelist_encode_32(_pagelist, 0);
4399 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
4403 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
4404 addr = kmap_atomic(page);
4405 if (recon_state->nr_caps) {
4406 /* currently encoding caps */
4407 *addr = cpu_to_le32(recon_state->nr_caps);
4409 /* currently encoding relams */
4410 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
4412 kunmap_atomic(addr);
4414 reply->hdr.version = cpu_to_le16(5);
4415 reply->hdr.compat_version = cpu_to_le16(4);
4417 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
4418 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
4420 ceph_con_send(&recon_state->session->s_con, reply);
4421 ceph_pagelist_release(recon_state->pagelist);
4423 recon_state->pagelist = _pagelist;
4424 recon_state->nr_caps = 0;
4425 recon_state->nr_realms = 0;
4426 recon_state->msg_version = 5;
4429 ceph_msg_put(reply);
4431 ceph_pagelist_release(_pagelist);
4435 static struct dentry* d_find_primary(struct inode *inode)
4437 struct dentry *alias, *dn = NULL;
4439 if (hlist_empty(&inode->i_dentry))
4442 spin_lock(&inode->i_lock);
4443 if (hlist_empty(&inode->i_dentry))
4446 if (S_ISDIR(inode->i_mode)) {
4447 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
4448 if (!IS_ROOT(alias))
4453 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
4454 spin_lock(&alias->d_lock);
4455 if (!d_unhashed(alias) &&
4456 (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
4457 dn = dget_dlock(alias);
4459 spin_unlock(&alias->d_lock);
4464 spin_unlock(&inode->i_lock);
4469 * Encode information about a cap for a reconnect with the MDS.
4471 static int reconnect_caps_cb(struct inode *inode, int mds, void *arg)
4473 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
4474 struct ceph_client *cl = ceph_inode_to_client(inode);
4476 struct ceph_mds_cap_reconnect v2;
4477 struct ceph_mds_cap_reconnect_v1 v1;
4479 struct ceph_inode_info *ci = ceph_inode(inode);
4480 struct ceph_reconnect_state *recon_state = arg;
4481 struct ceph_pagelist *pagelist = recon_state->pagelist;
4482 struct dentry *dentry;
4483 struct ceph_cap *cap;
4485 int pathlen = 0, err;
4489 dentry = d_find_primary(inode);
4491 /* set pathbase to parent dir when msg_version >= 2 */
4492 path = ceph_mdsc_build_path(mdsc, dentry, &pathlen, &pathbase,
4493 recon_state->msg_version >= 2);
4496 err = PTR_ERR(path);
4504 spin_lock(&ci->i_ceph_lock);
4505 cap = __get_cap_for_mds(ci, mds);
4507 spin_unlock(&ci->i_ceph_lock);
4511 doutc(cl, " adding %p ino %llx.%llx cap %p %lld %s\n", inode,
4512 ceph_vinop(inode), cap, cap->cap_id,
4513 ceph_cap_string(cap->issued));
4515 cap->seq = 0; /* reset cap seq */
4516 cap->issue_seq = 0; /* and issue_seq */
4517 cap->mseq = 0; /* and migrate_seq */
4518 cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
4520 /* These are lost when the session goes away */
4521 if (S_ISDIR(inode->i_mode)) {
4522 if (cap->issued & CEPH_CAP_DIR_CREATE) {
4523 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
4524 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
4526 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
4529 if (recon_state->msg_version >= 2) {
4530 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
4531 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
4532 rec.v2.issued = cpu_to_le32(cap->issued);
4533 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4534 rec.v2.pathbase = cpu_to_le64(pathbase);
4535 rec.v2.flock_len = (__force __le32)
4536 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
4538 struct timespec64 ts;
4540 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
4541 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
4542 rec.v1.issued = cpu_to_le32(cap->issued);
4543 rec.v1.size = cpu_to_le64(i_size_read(inode));
4544 ts = inode_get_mtime(inode);
4545 ceph_encode_timespec64(&rec.v1.mtime, &ts);
4546 ts = inode_get_atime(inode);
4547 ceph_encode_timespec64(&rec.v1.atime, &ts);
4548 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4549 rec.v1.pathbase = cpu_to_le64(pathbase);
4552 if (list_empty(&ci->i_cap_snaps)) {
4553 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
4555 struct ceph_cap_snap *capsnap =
4556 list_first_entry(&ci->i_cap_snaps,
4557 struct ceph_cap_snap, ci_item);
4558 snap_follows = capsnap->follows;
4560 spin_unlock(&ci->i_ceph_lock);
4562 if (recon_state->msg_version >= 2) {
4563 int num_fcntl_locks, num_flock_locks;
4564 struct ceph_filelock *flocks = NULL;
4565 size_t struct_len, total_len = sizeof(u64);
4569 if (rec.v2.flock_len) {
4570 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
4572 num_fcntl_locks = 0;
4573 num_flock_locks = 0;
4575 if (num_fcntl_locks + num_flock_locks > 0) {
4576 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
4577 sizeof(struct ceph_filelock),
4583 err = ceph_encode_locks_to_buffer(inode, flocks,
4598 if (recon_state->msg_version >= 3) {
4599 /* version, compat_version and struct_len */
4600 total_len += 2 * sizeof(u8) + sizeof(u32);
4604 * number of encoded locks is stable, so copy to pagelist
4606 struct_len = 2 * sizeof(u32) +
4607 (num_fcntl_locks + num_flock_locks) *
4608 sizeof(struct ceph_filelock);
4609 rec.v2.flock_len = cpu_to_le32(struct_len);
4611 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
4614 struct_len += sizeof(u64); /* snap_follows */
4616 total_len += struct_len;
4618 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
4619 err = send_reconnect_partial(recon_state);
4621 goto out_freeflocks;
4622 pagelist = recon_state->pagelist;
4625 err = ceph_pagelist_reserve(pagelist, total_len);
4627 goto out_freeflocks;
4629 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4630 if (recon_state->msg_version >= 3) {
4631 ceph_pagelist_encode_8(pagelist, struct_v);
4632 ceph_pagelist_encode_8(pagelist, 1);
4633 ceph_pagelist_encode_32(pagelist, struct_len);
4635 ceph_pagelist_encode_string(pagelist, path, pathlen);
4636 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
4637 ceph_locks_to_pagelist(flocks, pagelist,
4638 num_fcntl_locks, num_flock_locks);
4640 ceph_pagelist_encode_64(pagelist, snap_follows);
4644 err = ceph_pagelist_reserve(pagelist,
4645 sizeof(u64) + sizeof(u32) +
4646 pathlen + sizeof(rec.v1));
4650 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4651 ceph_pagelist_encode_string(pagelist, path, pathlen);
4652 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
4656 ceph_mdsc_free_path(path, pathlen);
4658 recon_state->nr_caps++;
4662 static int encode_snap_realms(struct ceph_mds_client *mdsc,
4663 struct ceph_reconnect_state *recon_state)
4666 struct ceph_pagelist *pagelist = recon_state->pagelist;
4667 struct ceph_client *cl = mdsc->fsc->client;
4670 if (recon_state->msg_version >= 4) {
4671 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
4677 * snaprealms. we provide mds with the ino, seq (version), and
4678 * parent for all of our realms. If the mds has any newer info,
4681 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
4682 struct ceph_snap_realm *realm =
4683 rb_entry(p, struct ceph_snap_realm, node);
4684 struct ceph_mds_snaprealm_reconnect sr_rec;
4686 if (recon_state->msg_version >= 4) {
4687 size_t need = sizeof(u8) * 2 + sizeof(u32) +
4690 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
4691 err = send_reconnect_partial(recon_state);
4694 pagelist = recon_state->pagelist;
4697 err = ceph_pagelist_reserve(pagelist, need);
4701 ceph_pagelist_encode_8(pagelist, 1);
4702 ceph_pagelist_encode_8(pagelist, 1);
4703 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
4706 doutc(cl, " adding snap realm %llx seq %lld parent %llx\n",
4707 realm->ino, realm->seq, realm->parent_ino);
4708 sr_rec.ino = cpu_to_le64(realm->ino);
4709 sr_rec.seq = cpu_to_le64(realm->seq);
4710 sr_rec.parent = cpu_to_le64(realm->parent_ino);
4712 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
4716 recon_state->nr_realms++;
4724 * If an MDS fails and recovers, clients need to reconnect in order to
4725 * reestablish shared state. This includes all caps issued through
4726 * this session _and_ the snap_realm hierarchy. Because it's not
4727 * clear which snap realms the mds cares about, we send everything we
4728 * know about.. that ensures we'll then get any new info the
4729 * recovering MDS might have.
4731 * This is a relatively heavyweight operation, but it's rare.
4733 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
4734 struct ceph_mds_session *session)
4736 struct ceph_client *cl = mdsc->fsc->client;
4737 struct ceph_msg *reply;
4738 int mds = session->s_mds;
4740 struct ceph_reconnect_state recon_state = {
4745 pr_info_client(cl, "mds%d reconnect start\n", mds);
4747 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4748 if (!recon_state.pagelist)
4749 goto fail_nopagelist;
4751 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4755 xa_destroy(&session->s_delegated_inos);
4757 mutex_lock(&session->s_mutex);
4758 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4761 doutc(cl, "session %p state %s\n", session,
4762 ceph_session_state_name(session->s_state));
4764 atomic_inc(&session->s_cap_gen);
4766 spin_lock(&session->s_cap_lock);
4767 /* don't know if session is readonly */
4768 session->s_readonly = 0;
4770 * notify __ceph_remove_cap() that we are composing cap reconnect.
4771 * If a cap get released before being added to the cap reconnect,
4772 * __ceph_remove_cap() should skip queuing cap release.
4774 session->s_cap_reconnect = 1;
4775 /* drop old cap expires; we're about to reestablish that state */
4776 detach_cap_releases(session, &dispose);
4777 spin_unlock(&session->s_cap_lock);
4778 dispose_cap_releases(mdsc, &dispose);
4780 /* trim unused caps to reduce MDS's cache rejoin time */
4781 if (mdsc->fsc->sb->s_root)
4782 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4784 ceph_con_close(&session->s_con);
4785 ceph_con_open(&session->s_con,
4786 CEPH_ENTITY_TYPE_MDS, mds,
4787 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4789 /* replay unsafe requests */
4790 replay_unsafe_requests(mdsc, session);
4792 ceph_early_kick_flushing_caps(mdsc, session);
4794 down_read(&mdsc->snap_rwsem);
4796 /* placeholder for nr_caps */
4797 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4801 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4802 recon_state.msg_version = 3;
4803 recon_state.allow_multi = true;
4804 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4805 recon_state.msg_version = 3;
4807 recon_state.msg_version = 2;
4809 /* trsaverse this session's caps */
4810 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4812 spin_lock(&session->s_cap_lock);
4813 session->s_cap_reconnect = 0;
4814 spin_unlock(&session->s_cap_lock);
4819 /* check if all realms can be encoded into current message */
4820 if (mdsc->num_snap_realms) {
4822 recon_state.pagelist->length +
4823 mdsc->num_snap_realms *
4824 sizeof(struct ceph_mds_snaprealm_reconnect);
4825 if (recon_state.msg_version >= 4) {
4826 /* number of realms */
4827 total_len += sizeof(u32);
4828 /* version, compat_version and struct_len */
4829 total_len += mdsc->num_snap_realms *
4830 (2 * sizeof(u8) + sizeof(u32));
4832 if (total_len > RECONNECT_MAX_SIZE) {
4833 if (!recon_state.allow_multi) {
4837 if (recon_state.nr_caps) {
4838 err = send_reconnect_partial(&recon_state);
4842 recon_state.msg_version = 5;
4846 err = encode_snap_realms(mdsc, &recon_state);
4850 if (recon_state.msg_version >= 5) {
4851 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4856 if (recon_state.nr_caps || recon_state.nr_realms) {
4858 list_first_entry(&recon_state.pagelist->head,
4860 __le32 *addr = kmap_atomic(page);
4861 if (recon_state.nr_caps) {
4862 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4863 *addr = cpu_to_le32(recon_state.nr_caps);
4864 } else if (recon_state.msg_version >= 4) {
4865 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4867 kunmap_atomic(addr);
4870 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4871 if (recon_state.msg_version >= 4)
4872 reply->hdr.compat_version = cpu_to_le16(4);
4874 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4875 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4877 ceph_con_send(&session->s_con, reply);
4879 mutex_unlock(&session->s_mutex);
4881 mutex_lock(&mdsc->mutex);
4882 __wake_requests(mdsc, &session->s_waiting);
4883 mutex_unlock(&mdsc->mutex);
4885 up_read(&mdsc->snap_rwsem);
4886 ceph_pagelist_release(recon_state.pagelist);
4890 ceph_msg_put(reply);
4891 up_read(&mdsc->snap_rwsem);
4892 mutex_unlock(&session->s_mutex);
4894 ceph_pagelist_release(recon_state.pagelist);
4896 pr_err_client(cl, "error %d preparing reconnect for mds%d\n",
4903 * compare old and new mdsmaps, kicking requests
4904 * and closing out old connections as necessary
4906 * called under mdsc->mutex.
4908 static void check_new_map(struct ceph_mds_client *mdsc,
4909 struct ceph_mdsmap *newmap,
4910 struct ceph_mdsmap *oldmap)
4913 int oldstate, newstate;
4914 struct ceph_mds_session *s;
4915 unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
4916 struct ceph_client *cl = mdsc->fsc->client;
4918 doutc(cl, "new %u old %u\n", newmap->m_epoch, oldmap->m_epoch);
4920 if (newmap->m_info) {
4921 for (i = 0; i < newmap->possible_max_rank; i++) {
4922 for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
4923 set_bit(newmap->m_info[i].export_targets[j], targets);
4927 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4928 if (!mdsc->sessions[i])
4930 s = mdsc->sessions[i];
4931 oldstate = ceph_mdsmap_get_state(oldmap, i);
4932 newstate = ceph_mdsmap_get_state(newmap, i);
4934 doutc(cl, "mds%d state %s%s -> %s%s (session %s)\n",
4935 i, ceph_mds_state_name(oldstate),
4936 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4937 ceph_mds_state_name(newstate),
4938 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4939 ceph_session_state_name(s->s_state));
4941 if (i >= newmap->possible_max_rank) {
4942 /* force close session for stopped mds */
4943 ceph_get_mds_session(s);
4944 __unregister_session(mdsc, s);
4945 __wake_requests(mdsc, &s->s_waiting);
4946 mutex_unlock(&mdsc->mutex);
4948 mutex_lock(&s->s_mutex);
4949 cleanup_session_requests(mdsc, s);
4950 remove_session_caps(s);
4951 mutex_unlock(&s->s_mutex);
4953 ceph_put_mds_session(s);
4955 mutex_lock(&mdsc->mutex);
4956 kick_requests(mdsc, i);
4960 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4961 ceph_mdsmap_get_addr(newmap, i),
4962 sizeof(struct ceph_entity_addr))) {
4964 mutex_unlock(&mdsc->mutex);
4965 mutex_lock(&s->s_mutex);
4966 mutex_lock(&mdsc->mutex);
4967 ceph_con_close(&s->s_con);
4968 mutex_unlock(&s->s_mutex);
4969 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4970 } else if (oldstate == newstate) {
4971 continue; /* nothing new with this mds */
4977 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4978 newstate >= CEPH_MDS_STATE_RECONNECT) {
4979 mutex_unlock(&mdsc->mutex);
4980 clear_bit(i, targets);
4981 send_mds_reconnect(mdsc, s);
4982 mutex_lock(&mdsc->mutex);
4986 * kick request on any mds that has gone active.
4988 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4989 newstate >= CEPH_MDS_STATE_ACTIVE) {
4990 if (oldstate != CEPH_MDS_STATE_CREATING &&
4991 oldstate != CEPH_MDS_STATE_STARTING)
4992 pr_info_client(cl, "mds%d recovery completed\n",
4994 kick_requests(mdsc, i);
4995 mutex_unlock(&mdsc->mutex);
4996 mutex_lock(&s->s_mutex);
4997 mutex_lock(&mdsc->mutex);
4998 ceph_kick_flushing_caps(mdsc, s);
4999 mutex_unlock(&s->s_mutex);
5000 wake_up_session_caps(s, RECONNECT);
5005 * Only open and reconnect sessions that don't exist yet.
5007 for (i = 0; i < newmap->possible_max_rank; i++) {
5009 * In case the import MDS is crashed just after
5010 * the EImportStart journal is flushed, so when
5011 * a standby MDS takes over it and is replaying
5012 * the EImportStart journal the new MDS daemon
5013 * will wait the client to reconnect it, but the
5014 * client may never register/open the session yet.
5016 * Will try to reconnect that MDS daemon if the
5017 * rank number is in the export targets array and
5018 * is the up:reconnect state.
5020 newstate = ceph_mdsmap_get_state(newmap, i);
5021 if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
5025 * The session maybe registered and opened by some
5026 * requests which were choosing random MDSes during
5027 * the mdsc->mutex's unlock/lock gap below in rare
5028 * case. But the related MDS daemon will just queue
5029 * that requests and be still waiting for the client's
5030 * reconnection request in up:reconnect state.
5032 s = __ceph_lookup_mds_session(mdsc, i);
5034 s = __open_export_target_session(mdsc, i);
5038 "failed to open export target session, err %d\n",
5043 doutc(cl, "send reconnect to export target mds.%d\n", i);
5044 mutex_unlock(&mdsc->mutex);
5045 send_mds_reconnect(mdsc, s);
5046 ceph_put_mds_session(s);
5047 mutex_lock(&mdsc->mutex);
5050 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
5051 s = mdsc->sessions[i];
5054 if (!ceph_mdsmap_is_laggy(newmap, i))
5056 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
5057 s->s_state == CEPH_MDS_SESSION_HUNG ||
5058 s->s_state == CEPH_MDS_SESSION_CLOSING) {
5059 doutc(cl, " connecting to export targets of laggy mds%d\n", i);
5060 __open_export_target_sessions(mdsc, s);
5072 * caller must hold session s_mutex, dentry->d_lock
5074 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
5076 struct ceph_dentry_info *di = ceph_dentry(dentry);
5078 ceph_put_mds_session(di->lease_session);
5079 di->lease_session = NULL;
5082 static void handle_lease(struct ceph_mds_client *mdsc,
5083 struct ceph_mds_session *session,
5084 struct ceph_msg *msg)
5086 struct ceph_client *cl = mdsc->fsc->client;
5087 struct super_block *sb = mdsc->fsc->sb;
5088 struct inode *inode;
5089 struct dentry *parent, *dentry;
5090 struct ceph_dentry_info *di;
5091 int mds = session->s_mds;
5092 struct ceph_mds_lease *h = msg->front.iov_base;
5094 struct ceph_vino vino;
5098 doutc(cl, "from mds%d\n", mds);
5100 if (!ceph_inc_mds_stopping_blocker(mdsc, session))
5104 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
5106 vino.ino = le64_to_cpu(h->ino);
5107 vino.snap = CEPH_NOSNAP;
5108 seq = le32_to_cpu(h->seq);
5109 dname.len = get_unaligned_le32(h + 1);
5110 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
5112 dname.name = (void *)(h + 1) + sizeof(u32);
5115 inode = ceph_find_inode(sb, vino);
5116 doutc(cl, "%s, ino %llx %p %.*s\n", ceph_lease_op_name(h->action),
5117 vino.ino, inode, dname.len, dname.name);
5119 mutex_lock(&session->s_mutex);
5121 doutc(cl, "no inode %llx\n", vino.ino);
5126 parent = d_find_alias(inode);
5128 doutc(cl, "no parent dentry on inode %p\n", inode);
5130 goto release; /* hrm... */
5132 dname.hash = full_name_hash(parent, dname.name, dname.len);
5133 dentry = d_lookup(parent, &dname);
5138 spin_lock(&dentry->d_lock);
5139 di = ceph_dentry(dentry);
5140 switch (h->action) {
5141 case CEPH_MDS_LEASE_REVOKE:
5142 if (di->lease_session == session) {
5143 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
5144 h->seq = cpu_to_le32(di->lease_seq);
5145 __ceph_mdsc_drop_dentry_lease(dentry);
5150 case CEPH_MDS_LEASE_RENEW:
5151 if (di->lease_session == session &&
5152 di->lease_gen == atomic_read(&session->s_cap_gen) &&
5153 di->lease_renew_from &&
5154 di->lease_renew_after == 0) {
5155 unsigned long duration =
5156 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
5158 di->lease_seq = seq;
5159 di->time = di->lease_renew_from + duration;
5160 di->lease_renew_after = di->lease_renew_from +
5162 di->lease_renew_from = 0;
5166 spin_unlock(&dentry->d_lock);
5173 /* let's just reuse the same message */
5174 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
5176 ceph_con_send(&session->s_con, msg);
5179 mutex_unlock(&session->s_mutex);
5182 ceph_dec_mds_stopping_blocker(mdsc);
5186 ceph_dec_mds_stopping_blocker(mdsc);
5188 pr_err_client(cl, "corrupt lease message\n");
5192 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
5193 struct dentry *dentry, char action,
5196 struct ceph_client *cl = session->s_mdsc->fsc->client;
5197 struct ceph_msg *msg;
5198 struct ceph_mds_lease *lease;
5200 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
5202 doutc(cl, "identry %p %s to mds%d\n", dentry, ceph_lease_op_name(action),
5205 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
5208 lease = msg->front.iov_base;
5209 lease->action = action;
5210 lease->seq = cpu_to_le32(seq);
5212 spin_lock(&dentry->d_lock);
5213 dir = d_inode(dentry->d_parent);
5214 lease->ino = cpu_to_le64(ceph_ino(dir));
5215 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
5217 put_unaligned_le32(dentry->d_name.len, lease + 1);
5218 memcpy((void *)(lease + 1) + 4,
5219 dentry->d_name.name, dentry->d_name.len);
5220 spin_unlock(&dentry->d_lock);
5222 ceph_con_send(&session->s_con, msg);
5226 * lock unlock the session, to wait ongoing session activities
5228 static void lock_unlock_session(struct ceph_mds_session *s)
5230 mutex_lock(&s->s_mutex);
5231 mutex_unlock(&s->s_mutex);
5234 static void maybe_recover_session(struct ceph_mds_client *mdsc)
5236 struct ceph_client *cl = mdsc->fsc->client;
5237 struct ceph_fs_client *fsc = mdsc->fsc;
5239 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
5242 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
5245 if (!READ_ONCE(fsc->blocklisted))
5248 pr_info_client(cl, "auto reconnect after blocklisted\n");
5249 ceph_force_reconnect(fsc->sb);
5252 bool check_session_state(struct ceph_mds_session *s)
5254 struct ceph_client *cl = s->s_mdsc->fsc->client;
5256 switch (s->s_state) {
5257 case CEPH_MDS_SESSION_OPEN:
5258 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
5259 s->s_state = CEPH_MDS_SESSION_HUNG;
5260 pr_info_client(cl, "mds%d hung\n", s->s_mds);
5263 case CEPH_MDS_SESSION_CLOSING:
5264 case CEPH_MDS_SESSION_NEW:
5265 case CEPH_MDS_SESSION_RESTARTING:
5266 case CEPH_MDS_SESSION_CLOSED:
5267 case CEPH_MDS_SESSION_REJECTED:
5275 * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
5276 * then we need to retransmit that request.
5278 void inc_session_sequence(struct ceph_mds_session *s)
5280 struct ceph_client *cl = s->s_mdsc->fsc->client;
5282 lockdep_assert_held(&s->s_mutex);
5286 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
5289 doutc(cl, "resending session close request for mds%d\n", s->s_mds);
5290 ret = request_close_session(s);
5292 pr_err_client(cl, "unable to close session to mds%d: %d\n",
5298 * delayed work -- periodically trim expired leases, renew caps with mds. If
5299 * the @delay parameter is set to 0 or if it's more than 5 secs, the default
5300 * workqueue delay value of 5 secs will be used.
5302 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
5304 unsigned long max_delay = HZ * 5;
5306 /* 5 secs default delay */
5307 if (!delay || (delay > max_delay))
5309 schedule_delayed_work(&mdsc->delayed_work,
5310 round_jiffies_relative(delay));
5313 static void delayed_work(struct work_struct *work)
5315 struct ceph_mds_client *mdsc =
5316 container_of(work, struct ceph_mds_client, delayed_work.work);
5317 unsigned long delay;
5322 doutc(mdsc->fsc->client, "mdsc delayed_work\n");
5324 if (mdsc->stopping >= CEPH_MDSC_STOPPING_FLUSHED)
5327 mutex_lock(&mdsc->mutex);
5328 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
5329 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
5330 mdsc->last_renew_caps);
5332 mdsc->last_renew_caps = jiffies;
5334 for (i = 0; i < mdsc->max_sessions; i++) {
5335 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
5339 if (!check_session_state(s)) {
5340 ceph_put_mds_session(s);
5343 mutex_unlock(&mdsc->mutex);
5345 mutex_lock(&s->s_mutex);
5347 send_renew_caps(mdsc, s);
5349 ceph_con_keepalive(&s->s_con);
5350 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
5351 s->s_state == CEPH_MDS_SESSION_HUNG)
5352 ceph_send_cap_releases(mdsc, s);
5353 mutex_unlock(&s->s_mutex);
5354 ceph_put_mds_session(s);
5356 mutex_lock(&mdsc->mutex);
5358 mutex_unlock(&mdsc->mutex);
5360 delay = ceph_check_delayed_caps(mdsc);
5362 ceph_queue_cap_reclaim_work(mdsc);
5364 ceph_trim_snapid_map(mdsc);
5366 maybe_recover_session(mdsc);
5368 schedule_delayed(mdsc, delay);
5371 int ceph_mdsc_init(struct ceph_fs_client *fsc)
5374 struct ceph_mds_client *mdsc;
5377 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
5381 mutex_init(&mdsc->mutex);
5382 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
5383 if (!mdsc->mdsmap) {
5388 init_completion(&mdsc->safe_umount_waiters);
5389 spin_lock_init(&mdsc->stopping_lock);
5390 atomic_set(&mdsc->stopping_blockers, 0);
5391 init_completion(&mdsc->stopping_waiter);
5392 init_waitqueue_head(&mdsc->session_close_wq);
5393 INIT_LIST_HEAD(&mdsc->waiting_for_map);
5394 mdsc->quotarealms_inodes = RB_ROOT;
5395 mutex_init(&mdsc->quotarealms_inodes_mutex);
5396 init_rwsem(&mdsc->snap_rwsem);
5397 mdsc->snap_realms = RB_ROOT;
5398 INIT_LIST_HEAD(&mdsc->snap_empty);
5399 spin_lock_init(&mdsc->snap_empty_lock);
5400 mdsc->request_tree = RB_ROOT;
5401 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
5402 mdsc->last_renew_caps = jiffies;
5403 INIT_LIST_HEAD(&mdsc->cap_delay_list);
5404 INIT_LIST_HEAD(&mdsc->cap_wait_list);
5405 spin_lock_init(&mdsc->cap_delay_lock);
5406 INIT_LIST_HEAD(&mdsc->cap_unlink_delay_list);
5407 INIT_LIST_HEAD(&mdsc->snap_flush_list);
5408 spin_lock_init(&mdsc->snap_flush_lock);
5409 mdsc->last_cap_flush_tid = 1;
5410 INIT_LIST_HEAD(&mdsc->cap_flush_list);
5411 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
5412 spin_lock_init(&mdsc->cap_dirty_lock);
5413 init_waitqueue_head(&mdsc->cap_flushing_wq);
5414 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
5415 INIT_WORK(&mdsc->cap_unlink_work, ceph_cap_unlink_work);
5416 err = ceph_metric_init(&mdsc->metric);
5420 spin_lock_init(&mdsc->dentry_list_lock);
5421 INIT_LIST_HEAD(&mdsc->dentry_leases);
5422 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
5424 ceph_caps_init(mdsc);
5425 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
5427 spin_lock_init(&mdsc->snapid_map_lock);
5428 mdsc->snapid_map_tree = RB_ROOT;
5429 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
5431 init_rwsem(&mdsc->pool_perm_rwsem);
5432 mdsc->pool_perm_tree = RB_ROOT;
5434 strscpy(mdsc->nodename, utsname()->nodename,
5435 sizeof(mdsc->nodename));
5441 kfree(mdsc->mdsmap);
5448 * Wait for safe replies on open mds requests. If we time out, drop
5449 * all requests from the tree to avoid dangling dentry refs.
5451 static void wait_requests(struct ceph_mds_client *mdsc)
5453 struct ceph_client *cl = mdsc->fsc->client;
5454 struct ceph_options *opts = mdsc->fsc->client->options;
5455 struct ceph_mds_request *req;
5457 mutex_lock(&mdsc->mutex);
5458 if (__get_oldest_req(mdsc)) {
5459 mutex_unlock(&mdsc->mutex);
5461 doutc(cl, "waiting for requests\n");
5462 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
5463 ceph_timeout_jiffies(opts->mount_timeout));
5465 /* tear down remaining requests */
5466 mutex_lock(&mdsc->mutex);
5467 while ((req = __get_oldest_req(mdsc))) {
5468 doutc(cl, "timed out on tid %llu\n", req->r_tid);
5469 list_del_init(&req->r_wait);
5470 __unregister_request(mdsc, req);
5473 mutex_unlock(&mdsc->mutex);
5474 doutc(cl, "done\n");
5477 void send_flush_mdlog(struct ceph_mds_session *s)
5479 struct ceph_client *cl = s->s_mdsc->fsc->client;
5480 struct ceph_msg *msg;
5483 * Pre-luminous MDS crashes when it sees an unknown session request
5485 if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
5488 mutex_lock(&s->s_mutex);
5489 doutc(cl, "request mdlog flush to mds%d (%s)s seq %lld\n",
5490 s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
5491 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
5494 pr_err_client(cl, "failed to request mdlog flush to mds%d (%s) seq %lld\n",
5495 s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
5497 ceph_con_send(&s->s_con, msg);
5499 mutex_unlock(&s->s_mutex);
5503 * called before mount is ro, and before dentries are torn down.
5504 * (hmm, does this still race with new lookups?)
5506 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
5508 doutc(mdsc->fsc->client, "begin\n");
5509 mdsc->stopping = CEPH_MDSC_STOPPING_BEGIN;
5511 ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
5512 ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
5513 ceph_flush_dirty_caps(mdsc);
5514 wait_requests(mdsc);
5517 * wait for reply handlers to drop their request refs and
5518 * their inode/dcache refs
5522 ceph_cleanup_quotarealms_inodes(mdsc);
5523 doutc(mdsc->fsc->client, "done\n");
5527 * flush the mdlog and wait for all write mds requests to flush.
5529 static void flush_mdlog_and_wait_mdsc_unsafe_requests(struct ceph_mds_client *mdsc,
5532 struct ceph_client *cl = mdsc->fsc->client;
5533 struct ceph_mds_request *req = NULL, *nextreq;
5534 struct ceph_mds_session *last_session = NULL;
5537 mutex_lock(&mdsc->mutex);
5538 doutc(cl, "want %lld\n", want_tid);
5540 req = __get_oldest_req(mdsc);
5541 while (req && req->r_tid <= want_tid) {
5542 /* find next request */
5543 n = rb_next(&req->r_node);
5545 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
5548 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
5549 (req->r_op & CEPH_MDS_OP_WRITE)) {
5550 struct ceph_mds_session *s = req->r_session;
5558 ceph_mdsc_get_request(req);
5560 ceph_mdsc_get_request(nextreq);
5561 s = ceph_get_mds_session(s);
5562 mutex_unlock(&mdsc->mutex);
5564 /* send flush mdlog request to MDS */
5565 if (last_session != s) {
5566 send_flush_mdlog(s);
5567 ceph_put_mds_session(last_session);
5570 ceph_put_mds_session(s);
5572 doutc(cl, "wait on %llu (want %llu)\n",
5573 req->r_tid, want_tid);
5574 wait_for_completion(&req->r_safe_completion);
5576 mutex_lock(&mdsc->mutex);
5577 ceph_mdsc_put_request(req);
5579 break; /* next dne before, so we're done! */
5580 if (RB_EMPTY_NODE(&nextreq->r_node)) {
5581 /* next request was removed from tree */
5582 ceph_mdsc_put_request(nextreq);
5585 ceph_mdsc_put_request(nextreq); /* won't go away */
5589 mutex_unlock(&mdsc->mutex);
5590 ceph_put_mds_session(last_session);
5591 doutc(cl, "done\n");
5594 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
5596 struct ceph_client *cl = mdsc->fsc->client;
5597 u64 want_tid, want_flush;
5599 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
5602 doutc(cl, "sync\n");
5603 mutex_lock(&mdsc->mutex);
5604 want_tid = mdsc->last_tid;
5605 mutex_unlock(&mdsc->mutex);
5607 ceph_flush_dirty_caps(mdsc);
5608 spin_lock(&mdsc->cap_dirty_lock);
5609 want_flush = mdsc->last_cap_flush_tid;
5610 if (!list_empty(&mdsc->cap_flush_list)) {
5611 struct ceph_cap_flush *cf =
5612 list_last_entry(&mdsc->cap_flush_list,
5613 struct ceph_cap_flush, g_list);
5616 spin_unlock(&mdsc->cap_dirty_lock);
5618 doutc(cl, "sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
5620 flush_mdlog_and_wait_mdsc_unsafe_requests(mdsc, want_tid);
5621 wait_caps_flush(mdsc, want_flush);
5625 * true if all sessions are closed, or we force unmount
5627 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
5629 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
5631 return atomic_read(&mdsc->num_sessions) <= skipped;
5635 * called after sb is ro or when metadata corrupted.
5637 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
5639 struct ceph_options *opts = mdsc->fsc->client->options;
5640 struct ceph_client *cl = mdsc->fsc->client;
5641 struct ceph_mds_session *session;
5645 doutc(cl, "begin\n");
5647 /* close sessions */
5648 mutex_lock(&mdsc->mutex);
5649 for (i = 0; i < mdsc->max_sessions; i++) {
5650 session = __ceph_lookup_mds_session(mdsc, i);
5653 mutex_unlock(&mdsc->mutex);
5654 mutex_lock(&session->s_mutex);
5655 if (__close_session(mdsc, session) <= 0)
5657 mutex_unlock(&session->s_mutex);
5658 ceph_put_mds_session(session);
5659 mutex_lock(&mdsc->mutex);
5661 mutex_unlock(&mdsc->mutex);
5663 doutc(cl, "waiting for sessions to close\n");
5664 wait_event_timeout(mdsc->session_close_wq,
5665 done_closing_sessions(mdsc, skipped),
5666 ceph_timeout_jiffies(opts->mount_timeout));
5668 /* tear down remaining sessions */
5669 mutex_lock(&mdsc->mutex);
5670 for (i = 0; i < mdsc->max_sessions; i++) {
5671 if (mdsc->sessions[i]) {
5672 session = ceph_get_mds_session(mdsc->sessions[i]);
5673 __unregister_session(mdsc, session);
5674 mutex_unlock(&mdsc->mutex);
5675 mutex_lock(&session->s_mutex);
5676 remove_session_caps(session);
5677 mutex_unlock(&session->s_mutex);
5678 ceph_put_mds_session(session);
5679 mutex_lock(&mdsc->mutex);
5682 WARN_ON(!list_empty(&mdsc->cap_delay_list));
5683 mutex_unlock(&mdsc->mutex);
5685 ceph_cleanup_snapid_map(mdsc);
5686 ceph_cleanup_global_and_empty_realms(mdsc);
5688 cancel_work_sync(&mdsc->cap_reclaim_work);
5689 cancel_work_sync(&mdsc->cap_unlink_work);
5690 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
5692 doutc(cl, "done\n");
5695 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
5697 struct ceph_mds_session *session;
5700 doutc(mdsc->fsc->client, "force umount\n");
5702 mutex_lock(&mdsc->mutex);
5703 for (mds = 0; mds < mdsc->max_sessions; mds++) {
5704 session = __ceph_lookup_mds_session(mdsc, mds);
5708 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
5709 __unregister_session(mdsc, session);
5710 __wake_requests(mdsc, &session->s_waiting);
5711 mutex_unlock(&mdsc->mutex);
5713 mutex_lock(&session->s_mutex);
5714 __close_session(mdsc, session);
5715 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
5716 cleanup_session_requests(mdsc, session);
5717 remove_session_caps(session);
5719 mutex_unlock(&session->s_mutex);
5720 ceph_put_mds_session(session);
5722 mutex_lock(&mdsc->mutex);
5723 kick_requests(mdsc, mds);
5725 __wake_requests(mdsc, &mdsc->waiting_for_map);
5726 mutex_unlock(&mdsc->mutex);
5729 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
5731 doutc(mdsc->fsc->client, "stop\n");
5733 * Make sure the delayed work stopped before releasing
5736 * Because the cancel_delayed_work_sync() will only
5737 * guarantee that the work finishes executing. But the
5738 * delayed work will re-arm itself again after that.
5740 flush_delayed_work(&mdsc->delayed_work);
5743 ceph_mdsmap_destroy(mdsc->mdsmap);
5744 kfree(mdsc->sessions);
5745 ceph_caps_finalize(mdsc);
5746 ceph_pool_perm_destroy(mdsc);
5749 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
5751 struct ceph_mds_client *mdsc = fsc->mdsc;
5752 doutc(fsc->client, "%p\n", mdsc);
5757 /* flush out any connection work with references to us */
5760 ceph_mdsc_stop(mdsc);
5762 ceph_metric_destroy(&mdsc->metric);
5766 doutc(fsc->client, "%p done\n", mdsc);
5769 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5771 struct ceph_fs_client *fsc = mdsc->fsc;
5772 struct ceph_client *cl = fsc->client;
5773 const char *mds_namespace = fsc->mount_options->mds_namespace;
5774 void *p = msg->front.iov_base;
5775 void *end = p + msg->front.iov_len;
5778 u32 mount_fscid = (u32)-1;
5781 ceph_decode_need(&p, end, sizeof(u32), bad);
5782 epoch = ceph_decode_32(&p);
5784 doutc(cl, "epoch %u\n", epoch);
5786 /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
5787 ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
5789 ceph_decode_32_safe(&p, end, num_fs, bad);
5790 while (num_fs-- > 0) {
5791 void *info_p, *info_end;
5795 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
5796 p += 2; // info_v, info_cv
5797 info_len = ceph_decode_32(&p);
5798 ceph_decode_need(&p, end, info_len, bad);
5800 info_end = p + info_len;
5803 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
5804 fscid = ceph_decode_32(&info_p);
5805 namelen = ceph_decode_32(&info_p);
5806 ceph_decode_need(&info_p, info_end, namelen, bad);
5808 if (mds_namespace &&
5809 strlen(mds_namespace) == namelen &&
5810 !strncmp(mds_namespace, (char *)info_p, namelen)) {
5811 mount_fscid = fscid;
5816 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
5817 if (mount_fscid != (u32)-1) {
5818 fsc->client->monc.fs_cluster_id = mount_fscid;
5819 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
5821 ceph_monc_renew_subs(&fsc->client->monc);
5829 pr_err_client(cl, "error decoding fsmap %d. Shutting down mount.\n",
5831 ceph_umount_begin(mdsc->fsc->sb);
5834 mutex_lock(&mdsc->mutex);
5835 mdsc->mdsmap_err = err;
5836 __wake_requests(mdsc, &mdsc->waiting_for_map);
5837 mutex_unlock(&mdsc->mutex);
5841 * handle mds map update.
5843 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5845 struct ceph_client *cl = mdsc->fsc->client;
5848 void *p = msg->front.iov_base;
5849 void *end = p + msg->front.iov_len;
5850 struct ceph_mdsmap *newmap, *oldmap;
5851 struct ceph_fsid fsid;
5854 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5855 ceph_decode_copy(&p, &fsid, sizeof(fsid));
5856 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5858 epoch = ceph_decode_32(&p);
5859 maplen = ceph_decode_32(&p);
5860 doutc(cl, "epoch %u len %d\n", epoch, (int)maplen);
5862 /* do we need it? */
5863 mutex_lock(&mdsc->mutex);
5864 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5865 doutc(cl, "epoch %u <= our %u\n", epoch, mdsc->mdsmap->m_epoch);
5866 mutex_unlock(&mdsc->mutex);
5870 newmap = ceph_mdsmap_decode(mdsc, &p, end, ceph_msgr2(mdsc->fsc->client));
5871 if (IS_ERR(newmap)) {
5872 err = PTR_ERR(newmap);
5876 /* swap into place */
5878 oldmap = mdsc->mdsmap;
5879 mdsc->mdsmap = newmap;
5880 check_new_map(mdsc, newmap, oldmap);
5881 ceph_mdsmap_destroy(oldmap);
5883 mdsc->mdsmap = newmap; /* first mds map */
5885 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5888 __wake_requests(mdsc, &mdsc->waiting_for_map);
5889 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5890 mdsc->mdsmap->m_epoch);
5892 mutex_unlock(&mdsc->mutex);
5893 schedule_delayed(mdsc, 0);
5897 mutex_unlock(&mdsc->mutex);
5899 pr_err_client(cl, "error decoding mdsmap %d. Shutting down mount.\n",
5901 ceph_umount_begin(mdsc->fsc->sb);
5906 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5908 struct ceph_mds_session *s = con->private;
5910 if (ceph_get_mds_session(s))
5915 static void mds_put_con(struct ceph_connection *con)
5917 struct ceph_mds_session *s = con->private;
5919 ceph_put_mds_session(s);
5923 * if the client is unresponsive for long enough, the mds will kill
5924 * the session entirely.
5926 static void mds_peer_reset(struct ceph_connection *con)
5928 struct ceph_mds_session *s = con->private;
5929 struct ceph_mds_client *mdsc = s->s_mdsc;
5931 pr_warn_client(mdsc->fsc->client, "mds%d closed our session\n",
5933 if (READ_ONCE(mdsc->fsc->mount_state) != CEPH_MOUNT_FENCE_IO &&
5934 ceph_mdsmap_get_state(mdsc->mdsmap, s->s_mds) >= CEPH_MDS_STATE_RECONNECT)
5935 send_mds_reconnect(mdsc, s);
5938 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5940 struct ceph_mds_session *s = con->private;
5941 struct ceph_mds_client *mdsc = s->s_mdsc;
5942 struct ceph_client *cl = mdsc->fsc->client;
5943 int type = le16_to_cpu(msg->hdr.type);
5945 mutex_lock(&mdsc->mutex);
5946 if (__verify_registered_session(mdsc, s) < 0) {
5947 mutex_unlock(&mdsc->mutex);
5950 mutex_unlock(&mdsc->mutex);
5953 case CEPH_MSG_MDS_MAP:
5954 ceph_mdsc_handle_mdsmap(mdsc, msg);
5956 case CEPH_MSG_FS_MAP_USER:
5957 ceph_mdsc_handle_fsmap(mdsc, msg);
5959 case CEPH_MSG_CLIENT_SESSION:
5960 handle_session(s, msg);
5962 case CEPH_MSG_CLIENT_REPLY:
5963 handle_reply(s, msg);
5965 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5966 handle_forward(mdsc, s, msg);
5968 case CEPH_MSG_CLIENT_CAPS:
5969 ceph_handle_caps(s, msg);
5971 case CEPH_MSG_CLIENT_SNAP:
5972 ceph_handle_snap(mdsc, s, msg);
5974 case CEPH_MSG_CLIENT_LEASE:
5975 handle_lease(mdsc, s, msg);
5977 case CEPH_MSG_CLIENT_QUOTA:
5978 ceph_handle_quota(mdsc, s, msg);
5982 pr_err_client(cl, "received unknown message type %d %s\n",
5983 type, ceph_msg_type_name(type));
5994 * Note: returned pointer is the address of a structure that's
5995 * managed separately. Caller must *not* attempt to free it.
5997 static struct ceph_auth_handshake *
5998 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
6000 struct ceph_mds_session *s = con->private;
6001 struct ceph_mds_client *mdsc = s->s_mdsc;
6002 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6003 struct ceph_auth_handshake *auth = &s->s_auth;
6006 ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
6007 force_new, proto, NULL, NULL);
6009 return ERR_PTR(ret);
6014 static int mds_add_authorizer_challenge(struct ceph_connection *con,
6015 void *challenge_buf, int challenge_buf_len)
6017 struct ceph_mds_session *s = con->private;
6018 struct ceph_mds_client *mdsc = s->s_mdsc;
6019 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6021 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
6022 challenge_buf, challenge_buf_len);
6025 static int mds_verify_authorizer_reply(struct ceph_connection *con)
6027 struct ceph_mds_session *s = con->private;
6028 struct ceph_mds_client *mdsc = s->s_mdsc;
6029 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6030 struct ceph_auth_handshake *auth = &s->s_auth;
6032 return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
6033 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
6034 NULL, NULL, NULL, NULL);
6037 static int mds_invalidate_authorizer(struct ceph_connection *con)
6039 struct ceph_mds_session *s = con->private;
6040 struct ceph_mds_client *mdsc = s->s_mdsc;
6041 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6043 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
6045 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
6048 static int mds_get_auth_request(struct ceph_connection *con,
6049 void *buf, int *buf_len,
6050 void **authorizer, int *authorizer_len)
6052 struct ceph_mds_session *s = con->private;
6053 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6054 struct ceph_auth_handshake *auth = &s->s_auth;
6057 ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
6062 *authorizer = auth->authorizer_buf;
6063 *authorizer_len = auth->authorizer_buf_len;
6067 static int mds_handle_auth_reply_more(struct ceph_connection *con,
6068 void *reply, int reply_len,
6069 void *buf, int *buf_len,
6070 void **authorizer, int *authorizer_len)
6072 struct ceph_mds_session *s = con->private;
6073 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6074 struct ceph_auth_handshake *auth = &s->s_auth;
6077 ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
6082 *authorizer = auth->authorizer_buf;
6083 *authorizer_len = auth->authorizer_buf_len;
6087 static int mds_handle_auth_done(struct ceph_connection *con,
6088 u64 global_id, void *reply, int reply_len,
6089 u8 *session_key, int *session_key_len,
6090 u8 *con_secret, int *con_secret_len)
6092 struct ceph_mds_session *s = con->private;
6093 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6094 struct ceph_auth_handshake *auth = &s->s_auth;
6096 return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
6097 session_key, session_key_len,
6098 con_secret, con_secret_len);
6101 static int mds_handle_auth_bad_method(struct ceph_connection *con,
6102 int used_proto, int result,
6103 const int *allowed_protos, int proto_cnt,
6104 const int *allowed_modes, int mode_cnt)
6106 struct ceph_mds_session *s = con->private;
6107 struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
6110 if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
6112 allowed_protos, proto_cnt,
6113 allowed_modes, mode_cnt)) {
6114 ret = ceph_monc_validate_auth(monc);
6122 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
6123 struct ceph_msg_header *hdr, int *skip)
6125 struct ceph_msg *msg;
6126 int type = (int) le16_to_cpu(hdr->type);
6127 int front_len = (int) le32_to_cpu(hdr->front_len);
6133 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
6135 pr_err("unable to allocate msg type %d len %d\n",
6143 static int mds_sign_message(struct ceph_msg *msg)
6145 struct ceph_mds_session *s = msg->con->private;
6146 struct ceph_auth_handshake *auth = &s->s_auth;
6148 return ceph_auth_sign_message(auth, msg);
6151 static int mds_check_message_signature(struct ceph_msg *msg)
6153 struct ceph_mds_session *s = msg->con->private;
6154 struct ceph_auth_handshake *auth = &s->s_auth;
6156 return ceph_auth_check_message_signature(auth, msg);
6159 static const struct ceph_connection_operations mds_con_ops = {
6162 .alloc_msg = mds_alloc_msg,
6163 .dispatch = mds_dispatch,
6164 .peer_reset = mds_peer_reset,
6165 .get_authorizer = mds_get_authorizer,
6166 .add_authorizer_challenge = mds_add_authorizer_challenge,
6167 .verify_authorizer_reply = mds_verify_authorizer_reply,
6168 .invalidate_authorizer = mds_invalidate_authorizer,
6169 .sign_message = mds_sign_message,
6170 .check_message_signature = mds_check_message_signature,
6171 .get_auth_request = mds_get_auth_request,
6172 .handle_auth_reply_more = mds_handle_auth_reply_more,
6173 .handle_auth_done = mds_handle_auth_done,
6174 .handle_auth_bad_method = mds_handle_auth_bad_method,