Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/ceph/ceph_debug.h>
3
4#include <linux/fs.h>
5#include <linux/wait.h>
6#include <linux/slab.h>
7#include <linux/gfp.h>
8#include <linux/sched.h>
9#include <linux/debugfs.h>
10#include <linux/seq_file.h>
11#include <linux/ratelimit.h>
12
13#include "super.h"
14#include "mds_client.h"
15
16#include <linux/ceph/ceph_features.h>
17#include <linux/ceph/messenger.h>
18#include <linux/ceph/decode.h>
19#include <linux/ceph/pagelist.h>
20#include <linux/ceph/auth.h>
21#include <linux/ceph/debugfs.h>
22
23#define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
24
25/*
26 * A cluster of MDS (metadata server) daemons is responsible for
27 * managing the file system namespace (the directory hierarchy and
28 * inodes) and for coordinating shared access to storage. Metadata is
29 * partitioning hierarchically across a number of servers, and that
30 * partition varies over time as the cluster adjusts the distribution
31 * in order to balance load.
32 *
33 * The MDS client is primarily responsible to managing synchronous
34 * metadata requests for operations like open, unlink, and so forth.
35 * If there is a MDS failure, we find out about it when we (possibly
36 * request and) receive a new MDS map, and can resubmit affected
37 * requests.
38 *
39 * For the most part, though, we take advantage of a lossless
40 * communications channel to the MDS, and do not need to worry about
41 * timing out or resubmitting requests.
42 *
43 * We maintain a stateful "session" with each MDS we interact with.
44 * Within each session, we sent periodic heartbeat messages to ensure
45 * any capabilities or leases we have been issues remain valid. If
46 * the session times out and goes stale, our leases and capabilities
47 * are no longer valid.
48 */
49
50struct ceph_reconnect_state {
51 struct ceph_mds_session *session;
52 int nr_caps, nr_realms;
53 struct ceph_pagelist *pagelist;
54 unsigned msg_version;
55 bool allow_multi;
56};
57
58static void __wake_requests(struct ceph_mds_client *mdsc,
59 struct list_head *head);
60static void ceph_cap_release_work(struct work_struct *work);
61static void ceph_cap_reclaim_work(struct work_struct *work);
62
63static const struct ceph_connection_operations mds_con_ops;
64
65
66/*
67 * mds reply parsing
68 */
69
70static int parse_reply_info_quota(void **p, void *end,
71 struct ceph_mds_reply_info_in *info)
72{
73 u8 struct_v, struct_compat;
74 u32 struct_len;
75
76 ceph_decode_8_safe(p, end, struct_v, bad);
77 ceph_decode_8_safe(p, end, struct_compat, bad);
78 /* struct_v is expected to be >= 1. we only
79 * understand encoding with struct_compat == 1. */
80 if (!struct_v || struct_compat != 1)
81 goto bad;
82 ceph_decode_32_safe(p, end, struct_len, bad);
83 ceph_decode_need(p, end, struct_len, bad);
84 end = *p + struct_len;
85 ceph_decode_64_safe(p, end, info->max_bytes, bad);
86 ceph_decode_64_safe(p, end, info->max_files, bad);
87 *p = end;
88 return 0;
89bad:
90 return -EIO;
91}
92
93/*
94 * parse individual inode info
95 */
96static int parse_reply_info_in(void **p, void *end,
97 struct ceph_mds_reply_info_in *info,
98 u64 features)
99{
100 int err = 0;
101 u8 struct_v = 0;
102
103 if (features == (u64)-1) {
104 u32 struct_len;
105 u8 struct_compat;
106 ceph_decode_8_safe(p, end, struct_v, bad);
107 ceph_decode_8_safe(p, end, struct_compat, bad);
108 /* struct_v is expected to be >= 1. we only understand
109 * encoding with struct_compat == 1. */
110 if (!struct_v || struct_compat != 1)
111 goto bad;
112 ceph_decode_32_safe(p, end, struct_len, bad);
113 ceph_decode_need(p, end, struct_len, bad);
114 end = *p + struct_len;
115 }
116
117 ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
118 info->in = *p;
119 *p += sizeof(struct ceph_mds_reply_inode) +
120 sizeof(*info->in->fragtree.splits) *
121 le32_to_cpu(info->in->fragtree.nsplits);
122
123 ceph_decode_32_safe(p, end, info->symlink_len, bad);
124 ceph_decode_need(p, end, info->symlink_len, bad);
125 info->symlink = *p;
126 *p += info->symlink_len;
127
128 ceph_decode_copy_safe(p, end, &info->dir_layout,
129 sizeof(info->dir_layout), bad);
130 ceph_decode_32_safe(p, end, info->xattr_len, bad);
131 ceph_decode_need(p, end, info->xattr_len, bad);
132 info->xattr_data = *p;
133 *p += info->xattr_len;
134
135 if (features == (u64)-1) {
136 /* inline data */
137 ceph_decode_64_safe(p, end, info->inline_version, bad);
138 ceph_decode_32_safe(p, end, info->inline_len, bad);
139 ceph_decode_need(p, end, info->inline_len, bad);
140 info->inline_data = *p;
141 *p += info->inline_len;
142 /* quota */
143 err = parse_reply_info_quota(p, end, info);
144 if (err < 0)
145 goto out_bad;
146 /* pool namespace */
147 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
148 if (info->pool_ns_len > 0) {
149 ceph_decode_need(p, end, info->pool_ns_len, bad);
150 info->pool_ns_data = *p;
151 *p += info->pool_ns_len;
152 }
153
154 /* btime */
155 ceph_decode_need(p, end, sizeof(info->btime), bad);
156 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
157
158 /* change attribute */
159 ceph_decode_64_safe(p, end, info->change_attr, bad);
160
161 /* dir pin */
162 if (struct_v >= 2) {
163 ceph_decode_32_safe(p, end, info->dir_pin, bad);
164 } else {
165 info->dir_pin = -ENODATA;
166 }
167
168 /* snapshot birth time, remains zero for v<=2 */
169 if (struct_v >= 3) {
170 ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
171 ceph_decode_copy(p, &info->snap_btime,
172 sizeof(info->snap_btime));
173 } else {
174 memset(&info->snap_btime, 0, sizeof(info->snap_btime));
175 }
176
177 *p = end;
178 } else {
179 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
180 ceph_decode_64_safe(p, end, info->inline_version, bad);
181 ceph_decode_32_safe(p, end, info->inline_len, bad);
182 ceph_decode_need(p, end, info->inline_len, bad);
183 info->inline_data = *p;
184 *p += info->inline_len;
185 } else
186 info->inline_version = CEPH_INLINE_NONE;
187
188 if (features & CEPH_FEATURE_MDS_QUOTA) {
189 err = parse_reply_info_quota(p, end, info);
190 if (err < 0)
191 goto out_bad;
192 } else {
193 info->max_bytes = 0;
194 info->max_files = 0;
195 }
196
197 info->pool_ns_len = 0;
198 info->pool_ns_data = NULL;
199 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
200 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
201 if (info->pool_ns_len > 0) {
202 ceph_decode_need(p, end, info->pool_ns_len, bad);
203 info->pool_ns_data = *p;
204 *p += info->pool_ns_len;
205 }
206 }
207
208 if (features & CEPH_FEATURE_FS_BTIME) {
209 ceph_decode_need(p, end, sizeof(info->btime), bad);
210 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
211 ceph_decode_64_safe(p, end, info->change_attr, bad);
212 }
213
214 info->dir_pin = -ENODATA;
215 /* info->snap_btime remains zero */
216 }
217 return 0;
218bad:
219 err = -EIO;
220out_bad:
221 return err;
222}
223
224static int parse_reply_info_dir(void **p, void *end,
225 struct ceph_mds_reply_dirfrag **dirfrag,
226 u64 features)
227{
228 if (features == (u64)-1) {
229 u8 struct_v, struct_compat;
230 u32 struct_len;
231 ceph_decode_8_safe(p, end, struct_v, bad);
232 ceph_decode_8_safe(p, end, struct_compat, bad);
233 /* struct_v is expected to be >= 1. we only understand
234 * encoding whose struct_compat == 1. */
235 if (!struct_v || struct_compat != 1)
236 goto bad;
237 ceph_decode_32_safe(p, end, struct_len, bad);
238 ceph_decode_need(p, end, struct_len, bad);
239 end = *p + struct_len;
240 }
241
242 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
243 *dirfrag = *p;
244 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
245 if (unlikely(*p > end))
246 goto bad;
247 if (features == (u64)-1)
248 *p = end;
249 return 0;
250bad:
251 return -EIO;
252}
253
254static int parse_reply_info_lease(void **p, void *end,
255 struct ceph_mds_reply_lease **lease,
256 u64 features)
257{
258 if (features == (u64)-1) {
259 u8 struct_v, struct_compat;
260 u32 struct_len;
261 ceph_decode_8_safe(p, end, struct_v, bad);
262 ceph_decode_8_safe(p, end, struct_compat, bad);
263 /* struct_v is expected to be >= 1. we only understand
264 * encoding whose struct_compat == 1. */
265 if (!struct_v || struct_compat != 1)
266 goto bad;
267 ceph_decode_32_safe(p, end, struct_len, bad);
268 ceph_decode_need(p, end, struct_len, bad);
269 end = *p + struct_len;
270 }
271
272 ceph_decode_need(p, end, sizeof(**lease), bad);
273 *lease = *p;
274 *p += sizeof(**lease);
275 if (features == (u64)-1)
276 *p = end;
277 return 0;
278bad:
279 return -EIO;
280}
281
282/*
283 * parse a normal reply, which may contain a (dir+)dentry and/or a
284 * target inode.
285 */
286static int parse_reply_info_trace(void **p, void *end,
287 struct ceph_mds_reply_info_parsed *info,
288 u64 features)
289{
290 int err;
291
292 if (info->head->is_dentry) {
293 err = parse_reply_info_in(p, end, &info->diri, features);
294 if (err < 0)
295 goto out_bad;
296
297 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
298 if (err < 0)
299 goto out_bad;
300
301 ceph_decode_32_safe(p, end, info->dname_len, bad);
302 ceph_decode_need(p, end, info->dname_len, bad);
303 info->dname = *p;
304 *p += info->dname_len;
305
306 err = parse_reply_info_lease(p, end, &info->dlease, features);
307 if (err < 0)
308 goto out_bad;
309 }
310
311 if (info->head->is_target) {
312 err = parse_reply_info_in(p, end, &info->targeti, features);
313 if (err < 0)
314 goto out_bad;
315 }
316
317 if (unlikely(*p != end))
318 goto bad;
319 return 0;
320
321bad:
322 err = -EIO;
323out_bad:
324 pr_err("problem parsing mds trace %d\n", err);
325 return err;
326}
327
328/*
329 * parse readdir results
330 */
331static int parse_reply_info_readdir(void **p, void *end,
332 struct ceph_mds_reply_info_parsed *info,
333 u64 features)
334{
335 u32 num, i = 0;
336 int err;
337
338 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
339 if (err < 0)
340 goto out_bad;
341
342 ceph_decode_need(p, end, sizeof(num) + 2, bad);
343 num = ceph_decode_32(p);
344 {
345 u16 flags = ceph_decode_16(p);
346 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
347 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
348 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
349 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
350 }
351 if (num == 0)
352 goto done;
353
354 BUG_ON(!info->dir_entries);
355 if ((unsigned long)(info->dir_entries + num) >
356 (unsigned long)info->dir_entries + info->dir_buf_size) {
357 pr_err("dir contents are larger than expected\n");
358 WARN_ON(1);
359 goto bad;
360 }
361
362 info->dir_nr = num;
363 while (num) {
364 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
365 /* dentry */
366 ceph_decode_32_safe(p, end, rde->name_len, bad);
367 ceph_decode_need(p, end, rde->name_len, bad);
368 rde->name = *p;
369 *p += rde->name_len;
370 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
371
372 /* dentry lease */
373 err = parse_reply_info_lease(p, end, &rde->lease, features);
374 if (err)
375 goto out_bad;
376 /* inode */
377 err = parse_reply_info_in(p, end, &rde->inode, features);
378 if (err < 0)
379 goto out_bad;
380 /* ceph_readdir_prepopulate() will update it */
381 rde->offset = 0;
382 i++;
383 num--;
384 }
385
386done:
387 /* Skip over any unrecognized fields */
388 *p = end;
389 return 0;
390
391bad:
392 err = -EIO;
393out_bad:
394 pr_err("problem parsing dir contents %d\n", err);
395 return err;
396}
397
398/*
399 * parse fcntl F_GETLK results
400 */
401static int parse_reply_info_filelock(void **p, void *end,
402 struct ceph_mds_reply_info_parsed *info,
403 u64 features)
404{
405 if (*p + sizeof(*info->filelock_reply) > end)
406 goto bad;
407
408 info->filelock_reply = *p;
409
410 /* Skip over any unrecognized fields */
411 *p = end;
412 return 0;
413bad:
414 return -EIO;
415}
416
417/*
418 * parse create results
419 */
420static int parse_reply_info_create(void **p, void *end,
421 struct ceph_mds_reply_info_parsed *info,
422 u64 features)
423{
424 if (features == (u64)-1 ||
425 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
426 /* Malformed reply? */
427 if (*p == end) {
428 info->has_create_ino = false;
429 } else {
430 info->has_create_ino = true;
431 ceph_decode_64_safe(p, end, info->ino, bad);
432 }
433 } else {
434 if (*p != end)
435 goto bad;
436 }
437
438 /* Skip over any unrecognized fields */
439 *p = end;
440 return 0;
441bad:
442 return -EIO;
443}
444
445/*
446 * parse extra results
447 */
448static int parse_reply_info_extra(void **p, void *end,
449 struct ceph_mds_reply_info_parsed *info,
450 u64 features)
451{
452 u32 op = le32_to_cpu(info->head->op);
453
454 if (op == CEPH_MDS_OP_GETFILELOCK)
455 return parse_reply_info_filelock(p, end, info, features);
456 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
457 return parse_reply_info_readdir(p, end, info, features);
458 else if (op == CEPH_MDS_OP_CREATE)
459 return parse_reply_info_create(p, end, info, features);
460 else
461 return -EIO;
462}
463
464/*
465 * parse entire mds reply
466 */
467static int parse_reply_info(struct ceph_msg *msg,
468 struct ceph_mds_reply_info_parsed *info,
469 u64 features)
470{
471 void *p, *end;
472 u32 len;
473 int err;
474
475 info->head = msg->front.iov_base;
476 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
477 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
478
479 /* trace */
480 ceph_decode_32_safe(&p, end, len, bad);
481 if (len > 0) {
482 ceph_decode_need(&p, end, len, bad);
483 err = parse_reply_info_trace(&p, p+len, info, features);
484 if (err < 0)
485 goto out_bad;
486 }
487
488 /* extra */
489 ceph_decode_32_safe(&p, end, len, bad);
490 if (len > 0) {
491 ceph_decode_need(&p, end, len, bad);
492 err = parse_reply_info_extra(&p, p+len, info, features);
493 if (err < 0)
494 goto out_bad;
495 }
496
497 /* snap blob */
498 ceph_decode_32_safe(&p, end, len, bad);
499 info->snapblob_len = len;
500 info->snapblob = p;
501 p += len;
502
503 if (p != end)
504 goto bad;
505 return 0;
506
507bad:
508 err = -EIO;
509out_bad:
510 pr_err("mds parse_reply err %d\n", err);
511 return err;
512}
513
514static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
515{
516 if (!info->dir_entries)
517 return;
518 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
519}
520
521
522/*
523 * sessions
524 */
525const char *ceph_session_state_name(int s)
526{
527 switch (s) {
528 case CEPH_MDS_SESSION_NEW: return "new";
529 case CEPH_MDS_SESSION_OPENING: return "opening";
530 case CEPH_MDS_SESSION_OPEN: return "open";
531 case CEPH_MDS_SESSION_HUNG: return "hung";
532 case CEPH_MDS_SESSION_CLOSING: return "closing";
533 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
534 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
535 case CEPH_MDS_SESSION_REJECTED: return "rejected";
536 default: return "???";
537 }
538}
539
540static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
541{
542 if (refcount_inc_not_zero(&s->s_ref)) {
543 dout("mdsc get_session %p %d -> %d\n", s,
544 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
545 return s;
546 } else {
547 dout("mdsc get_session %p 0 -- FAIL\n", s);
548 return NULL;
549 }
550}
551
552void ceph_put_mds_session(struct ceph_mds_session *s)
553{
554 dout("mdsc put_session %p %d -> %d\n", s,
555 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
556 if (refcount_dec_and_test(&s->s_ref)) {
557 if (s->s_auth.authorizer)
558 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
559 kfree(s);
560 }
561}
562
563/*
564 * called under mdsc->mutex
565 */
566struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
567 int mds)
568{
569 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
570 return NULL;
571 return get_session(mdsc->sessions[mds]);
572}
573
574static bool __have_session(struct ceph_mds_client *mdsc, int mds)
575{
576 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
577 return false;
578 else
579 return true;
580}
581
582static int __verify_registered_session(struct ceph_mds_client *mdsc,
583 struct ceph_mds_session *s)
584{
585 if (s->s_mds >= mdsc->max_sessions ||
586 mdsc->sessions[s->s_mds] != s)
587 return -ENOENT;
588 return 0;
589}
590
591/*
592 * create+register a new session for given mds.
593 * called under mdsc->mutex.
594 */
595static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
596 int mds)
597{
598 struct ceph_mds_session *s;
599
600 if (mds >= mdsc->mdsmap->m_num_mds)
601 return ERR_PTR(-EINVAL);
602
603 s = kzalloc(sizeof(*s), GFP_NOFS);
604 if (!s)
605 return ERR_PTR(-ENOMEM);
606
607 if (mds >= mdsc->max_sessions) {
608 int newmax = 1 << get_count_order(mds + 1);
609 struct ceph_mds_session **sa;
610
611 dout("%s: realloc to %d\n", __func__, newmax);
612 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
613 if (!sa)
614 goto fail_realloc;
615 if (mdsc->sessions) {
616 memcpy(sa, mdsc->sessions,
617 mdsc->max_sessions * sizeof(void *));
618 kfree(mdsc->sessions);
619 }
620 mdsc->sessions = sa;
621 mdsc->max_sessions = newmax;
622 }
623
624 dout("%s: mds%d\n", __func__, mds);
625 s->s_mdsc = mdsc;
626 s->s_mds = mds;
627 s->s_state = CEPH_MDS_SESSION_NEW;
628 s->s_ttl = 0;
629 s->s_seq = 0;
630 mutex_init(&s->s_mutex);
631
632 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
633
634 spin_lock_init(&s->s_gen_ttl_lock);
635 s->s_cap_gen = 1;
636 s->s_cap_ttl = jiffies - 1;
637
638 spin_lock_init(&s->s_cap_lock);
639 s->s_renew_requested = 0;
640 s->s_renew_seq = 0;
641 INIT_LIST_HEAD(&s->s_caps);
642 s->s_nr_caps = 0;
643 refcount_set(&s->s_ref, 1);
644 INIT_LIST_HEAD(&s->s_waiting);
645 INIT_LIST_HEAD(&s->s_unsafe);
646 s->s_num_cap_releases = 0;
647 s->s_cap_reconnect = 0;
648 s->s_cap_iterator = NULL;
649 INIT_LIST_HEAD(&s->s_cap_releases);
650 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
651
652 INIT_LIST_HEAD(&s->s_cap_flushing);
653
654 mdsc->sessions[mds] = s;
655 atomic_inc(&mdsc->num_sessions);
656 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
657
658 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
659 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
660
661 return s;
662
663fail_realloc:
664 kfree(s);
665 return ERR_PTR(-ENOMEM);
666}
667
668/*
669 * called under mdsc->mutex
670 */
671static void __unregister_session(struct ceph_mds_client *mdsc,
672 struct ceph_mds_session *s)
673{
674 dout("__unregister_session mds%d %p\n", s->s_mds, s);
675 BUG_ON(mdsc->sessions[s->s_mds] != s);
676 mdsc->sessions[s->s_mds] = NULL;
677 s->s_state = 0;
678 ceph_con_close(&s->s_con);
679 ceph_put_mds_session(s);
680 atomic_dec(&mdsc->num_sessions);
681}
682
683/*
684 * drop session refs in request.
685 *
686 * should be last request ref, or hold mdsc->mutex
687 */
688static void put_request_session(struct ceph_mds_request *req)
689{
690 if (req->r_session) {
691 ceph_put_mds_session(req->r_session);
692 req->r_session = NULL;
693 }
694}
695
696void ceph_mdsc_release_request(struct kref *kref)
697{
698 struct ceph_mds_request *req = container_of(kref,
699 struct ceph_mds_request,
700 r_kref);
701 destroy_reply_info(&req->r_reply_info);
702 if (req->r_request)
703 ceph_msg_put(req->r_request);
704 if (req->r_reply)
705 ceph_msg_put(req->r_reply);
706 if (req->r_inode) {
707 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
708 /* avoid calling iput_final() in mds dispatch threads */
709 ceph_async_iput(req->r_inode);
710 }
711 if (req->r_parent)
712 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
713 ceph_async_iput(req->r_target_inode);
714 if (req->r_dentry)
715 dput(req->r_dentry);
716 if (req->r_old_dentry)
717 dput(req->r_old_dentry);
718 if (req->r_old_dentry_dir) {
719 /*
720 * track (and drop pins for) r_old_dentry_dir
721 * separately, since r_old_dentry's d_parent may have
722 * changed between the dir mutex being dropped and
723 * this request being freed.
724 */
725 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
726 CEPH_CAP_PIN);
727 ceph_async_iput(req->r_old_dentry_dir);
728 }
729 kfree(req->r_path1);
730 kfree(req->r_path2);
731 if (req->r_pagelist)
732 ceph_pagelist_release(req->r_pagelist);
733 put_request_session(req);
734 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
735 WARN_ON_ONCE(!list_empty(&req->r_wait));
736 kfree(req);
737}
738
739DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
740
741/*
742 * lookup session, bump ref if found.
743 *
744 * called under mdsc->mutex.
745 */
746static struct ceph_mds_request *
747lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
748{
749 struct ceph_mds_request *req;
750
751 req = lookup_request(&mdsc->request_tree, tid);
752 if (req)
753 ceph_mdsc_get_request(req);
754
755 return req;
756}
757
758/*
759 * Register an in-flight request, and assign a tid. Link to directory
760 * are modifying (if any).
761 *
762 * Called under mdsc->mutex.
763 */
764static void __register_request(struct ceph_mds_client *mdsc,
765 struct ceph_mds_request *req,
766 struct inode *dir)
767{
768 int ret = 0;
769
770 req->r_tid = ++mdsc->last_tid;
771 if (req->r_num_caps) {
772 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
773 req->r_num_caps);
774 if (ret < 0) {
775 pr_err("__register_request %p "
776 "failed to reserve caps: %d\n", req, ret);
777 /* set req->r_err to fail early from __do_request */
778 req->r_err = ret;
779 return;
780 }
781 }
782 dout("__register_request %p tid %lld\n", req, req->r_tid);
783 ceph_mdsc_get_request(req);
784 insert_request(&mdsc->request_tree, req);
785
786 req->r_uid = current_fsuid();
787 req->r_gid = current_fsgid();
788
789 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
790 mdsc->oldest_tid = req->r_tid;
791
792 if (dir) {
793 ihold(dir);
794 req->r_unsafe_dir = dir;
795 }
796}
797
798static void __unregister_request(struct ceph_mds_client *mdsc,
799 struct ceph_mds_request *req)
800{
801 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
802
803 /* Never leave an unregistered request on an unsafe list! */
804 list_del_init(&req->r_unsafe_item);
805
806 if (req->r_tid == mdsc->oldest_tid) {
807 struct rb_node *p = rb_next(&req->r_node);
808 mdsc->oldest_tid = 0;
809 while (p) {
810 struct ceph_mds_request *next_req =
811 rb_entry(p, struct ceph_mds_request, r_node);
812 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
813 mdsc->oldest_tid = next_req->r_tid;
814 break;
815 }
816 p = rb_next(p);
817 }
818 }
819
820 erase_request(&mdsc->request_tree, req);
821
822 if (req->r_unsafe_dir &&
823 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
824 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
825 spin_lock(&ci->i_unsafe_lock);
826 list_del_init(&req->r_unsafe_dir_item);
827 spin_unlock(&ci->i_unsafe_lock);
828 }
829 if (req->r_target_inode &&
830 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
831 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
832 spin_lock(&ci->i_unsafe_lock);
833 list_del_init(&req->r_unsafe_target_item);
834 spin_unlock(&ci->i_unsafe_lock);
835 }
836
837 if (req->r_unsafe_dir) {
838 /* avoid calling iput_final() in mds dispatch threads */
839 ceph_async_iput(req->r_unsafe_dir);
840 req->r_unsafe_dir = NULL;
841 }
842
843 complete_all(&req->r_safe_completion);
844
845 ceph_mdsc_put_request(req);
846}
847
848/*
849 * Walk back up the dentry tree until we hit a dentry representing a
850 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
851 * when calling this) to ensure that the objects won't disappear while we're
852 * working with them. Once we hit a candidate dentry, we attempt to take a
853 * reference to it, and return that as the result.
854 */
855static struct inode *get_nonsnap_parent(struct dentry *dentry)
856{
857 struct inode *inode = NULL;
858
859 while (dentry && !IS_ROOT(dentry)) {
860 inode = d_inode_rcu(dentry);
861 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
862 break;
863 dentry = dentry->d_parent;
864 }
865 if (inode)
866 inode = igrab(inode);
867 return inode;
868}
869
870/*
871 * Choose mds to send request to next. If there is a hint set in the
872 * request (e.g., due to a prior forward hint from the mds), use that.
873 * Otherwise, consult frag tree and/or caps to identify the
874 * appropriate mds. If all else fails, choose randomly.
875 *
876 * Called under mdsc->mutex.
877 */
878static int __choose_mds(struct ceph_mds_client *mdsc,
879 struct ceph_mds_request *req)
880{
881 struct inode *inode;
882 struct ceph_inode_info *ci;
883 struct ceph_cap *cap;
884 int mode = req->r_direct_mode;
885 int mds = -1;
886 u32 hash = req->r_direct_hash;
887 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
888
889 /*
890 * is there a specific mds we should try? ignore hint if we have
891 * no session and the mds is not up (active or recovering).
892 */
893 if (req->r_resend_mds >= 0 &&
894 (__have_session(mdsc, req->r_resend_mds) ||
895 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
896 dout("choose_mds using resend_mds mds%d\n",
897 req->r_resend_mds);
898 return req->r_resend_mds;
899 }
900
901 if (mode == USE_RANDOM_MDS)
902 goto random;
903
904 inode = NULL;
905 if (req->r_inode) {
906 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
907 inode = req->r_inode;
908 ihold(inode);
909 } else {
910 /* req->r_dentry is non-null for LSSNAP request */
911 rcu_read_lock();
912 inode = get_nonsnap_parent(req->r_dentry);
913 rcu_read_unlock();
914 dout("__choose_mds using snapdir's parent %p\n", inode);
915 }
916 } else if (req->r_dentry) {
917 /* ignore race with rename; old or new d_parent is okay */
918 struct dentry *parent;
919 struct inode *dir;
920
921 rcu_read_lock();
922 parent = READ_ONCE(req->r_dentry->d_parent);
923 dir = req->r_parent ? : d_inode_rcu(parent);
924
925 if (!dir || dir->i_sb != mdsc->fsc->sb) {
926 /* not this fs or parent went negative */
927 inode = d_inode(req->r_dentry);
928 if (inode)
929 ihold(inode);
930 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
931 /* direct snapped/virtual snapdir requests
932 * based on parent dir inode */
933 inode = get_nonsnap_parent(parent);
934 dout("__choose_mds using nonsnap parent %p\n", inode);
935 } else {
936 /* dentry target */
937 inode = d_inode(req->r_dentry);
938 if (!inode || mode == USE_AUTH_MDS) {
939 /* dir + name */
940 inode = igrab(dir);
941 hash = ceph_dentry_hash(dir, req->r_dentry);
942 is_hash = true;
943 } else {
944 ihold(inode);
945 }
946 }
947 rcu_read_unlock();
948 }
949
950 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
951 (int)hash, mode);
952 if (!inode)
953 goto random;
954 ci = ceph_inode(inode);
955
956 if (is_hash && S_ISDIR(inode->i_mode)) {
957 struct ceph_inode_frag frag;
958 int found;
959
960 ceph_choose_frag(ci, hash, &frag, &found);
961 if (found) {
962 if (mode == USE_ANY_MDS && frag.ndist > 0) {
963 u8 r;
964
965 /* choose a random replica */
966 get_random_bytes(&r, 1);
967 r %= frag.ndist;
968 mds = frag.dist[r];
969 dout("choose_mds %p %llx.%llx "
970 "frag %u mds%d (%d/%d)\n",
971 inode, ceph_vinop(inode),
972 frag.frag, mds,
973 (int)r, frag.ndist);
974 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
975 CEPH_MDS_STATE_ACTIVE)
976 goto out;
977 }
978
979 /* since this file/dir wasn't known to be
980 * replicated, then we want to look for the
981 * authoritative mds. */
982 mode = USE_AUTH_MDS;
983 if (frag.mds >= 0) {
984 /* choose auth mds */
985 mds = frag.mds;
986 dout("choose_mds %p %llx.%llx "
987 "frag %u mds%d (auth)\n",
988 inode, ceph_vinop(inode), frag.frag, mds);
989 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
990 CEPH_MDS_STATE_ACTIVE)
991 goto out;
992 }
993 }
994 }
995
996 spin_lock(&ci->i_ceph_lock);
997 cap = NULL;
998 if (mode == USE_AUTH_MDS)
999 cap = ci->i_auth_cap;
1000 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1001 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1002 if (!cap) {
1003 spin_unlock(&ci->i_ceph_lock);
1004 ceph_async_iput(inode);
1005 goto random;
1006 }
1007 mds = cap->session->s_mds;
1008 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
1009 inode, ceph_vinop(inode), mds,
1010 cap == ci->i_auth_cap ? "auth " : "", cap);
1011 spin_unlock(&ci->i_ceph_lock);
1012out:
1013 /* avoid calling iput_final() while holding mdsc->mutex or
1014 * in mds dispatch threads */
1015 ceph_async_iput(inode);
1016 return mds;
1017
1018random:
1019 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1020 dout("choose_mds chose random mds%d\n", mds);
1021 return mds;
1022}
1023
1024
1025/*
1026 * session messages
1027 */
1028static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1029{
1030 struct ceph_msg *msg;
1031 struct ceph_mds_session_head *h;
1032
1033 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1034 false);
1035 if (!msg) {
1036 pr_err("create_session_msg ENOMEM creating msg\n");
1037 return NULL;
1038 }
1039 h = msg->front.iov_base;
1040 h->op = cpu_to_le32(op);
1041 h->seq = cpu_to_le64(seq);
1042
1043 return msg;
1044}
1045
1046static void encode_supported_features(void **p, void *end)
1047{
1048 static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1049 static const size_t count = ARRAY_SIZE(bits);
1050
1051 if (count > 0) {
1052 size_t i;
1053 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
1054
1055 BUG_ON(*p + 4 + size > end);
1056 ceph_encode_32(p, size);
1057 memset(*p, 0, size);
1058 for (i = 0; i < count; i++)
1059 ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
1060 *p += size;
1061 } else {
1062 BUG_ON(*p + 4 > end);
1063 ceph_encode_32(p, 0);
1064 }
1065}
1066
1067/*
1068 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1069 * to include additional client metadata fields.
1070 */
1071static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1072{
1073 struct ceph_msg *msg;
1074 struct ceph_mds_session_head *h;
1075 int i = -1;
1076 int extra_bytes = 0;
1077 int metadata_key_count = 0;
1078 struct ceph_options *opt = mdsc->fsc->client->options;
1079 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1080 void *p, *end;
1081
1082 const char* metadata[][2] = {
1083 {"hostname", mdsc->nodename},
1084 {"kernel_version", init_utsname()->release},
1085 {"entity_id", opt->name ? : ""},
1086 {"root", fsopt->server_path ? : "/"},
1087 {NULL, NULL}
1088 };
1089
1090 /* Calculate serialized length of metadata */
1091 extra_bytes = 4; /* map length */
1092 for (i = 0; metadata[i][0]; ++i) {
1093 extra_bytes += 8 + strlen(metadata[i][0]) +
1094 strlen(metadata[i][1]);
1095 metadata_key_count++;
1096 }
1097 /* supported feature */
1098 extra_bytes += 4 + 8;
1099
1100 /* Allocate the message */
1101 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1102 GFP_NOFS, false);
1103 if (!msg) {
1104 pr_err("create_session_msg ENOMEM creating msg\n");
1105 return NULL;
1106 }
1107 p = msg->front.iov_base;
1108 end = p + msg->front.iov_len;
1109
1110 h = p;
1111 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1112 h->seq = cpu_to_le64(seq);
1113
1114 /*
1115 * Serialize client metadata into waiting buffer space, using
1116 * the format that userspace expects for map<string, string>
1117 *
1118 * ClientSession messages with metadata are v2
1119 */
1120 msg->hdr.version = cpu_to_le16(3);
1121 msg->hdr.compat_version = cpu_to_le16(1);
1122
1123 /* The write pointer, following the session_head structure */
1124 p += sizeof(*h);
1125
1126 /* Number of entries in the map */
1127 ceph_encode_32(&p, metadata_key_count);
1128
1129 /* Two length-prefixed strings for each entry in the map */
1130 for (i = 0; metadata[i][0]; ++i) {
1131 size_t const key_len = strlen(metadata[i][0]);
1132 size_t const val_len = strlen(metadata[i][1]);
1133
1134 ceph_encode_32(&p, key_len);
1135 memcpy(p, metadata[i][0], key_len);
1136 p += key_len;
1137 ceph_encode_32(&p, val_len);
1138 memcpy(p, metadata[i][1], val_len);
1139 p += val_len;
1140 }
1141
1142 encode_supported_features(&p, end);
1143 msg->front.iov_len = p - msg->front.iov_base;
1144 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1145
1146 return msg;
1147}
1148
1149/*
1150 * send session open request.
1151 *
1152 * called under mdsc->mutex
1153 */
1154static int __open_session(struct ceph_mds_client *mdsc,
1155 struct ceph_mds_session *session)
1156{
1157 struct ceph_msg *msg;
1158 int mstate;
1159 int mds = session->s_mds;
1160
1161 /* wait for mds to go active? */
1162 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1163 dout("open_session to mds%d (%s)\n", mds,
1164 ceph_mds_state_name(mstate));
1165 session->s_state = CEPH_MDS_SESSION_OPENING;
1166 session->s_renew_requested = jiffies;
1167
1168 /* send connect message */
1169 msg = create_session_open_msg(mdsc, session->s_seq);
1170 if (!msg)
1171 return -ENOMEM;
1172 ceph_con_send(&session->s_con, msg);
1173 return 0;
1174}
1175
1176/*
1177 * open sessions for any export targets for the given mds
1178 *
1179 * called under mdsc->mutex
1180 */
1181static struct ceph_mds_session *
1182__open_export_target_session(struct ceph_mds_client *mdsc, int target)
1183{
1184 struct ceph_mds_session *session;
1185
1186 session = __ceph_lookup_mds_session(mdsc, target);
1187 if (!session) {
1188 session = register_session(mdsc, target);
1189 if (IS_ERR(session))
1190 return session;
1191 }
1192 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1193 session->s_state == CEPH_MDS_SESSION_CLOSING)
1194 __open_session(mdsc, session);
1195
1196 return session;
1197}
1198
1199struct ceph_mds_session *
1200ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1201{
1202 struct ceph_mds_session *session;
1203
1204 dout("open_export_target_session to mds%d\n", target);
1205
1206 mutex_lock(&mdsc->mutex);
1207 session = __open_export_target_session(mdsc, target);
1208 mutex_unlock(&mdsc->mutex);
1209
1210 return session;
1211}
1212
1213static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1214 struct ceph_mds_session *session)
1215{
1216 struct ceph_mds_info *mi;
1217 struct ceph_mds_session *ts;
1218 int i, mds = session->s_mds;
1219
1220 if (mds >= mdsc->mdsmap->m_num_mds)
1221 return;
1222
1223 mi = &mdsc->mdsmap->m_info[mds];
1224 dout("open_export_target_sessions for mds%d (%d targets)\n",
1225 session->s_mds, mi->num_export_targets);
1226
1227 for (i = 0; i < mi->num_export_targets; i++) {
1228 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1229 if (!IS_ERR(ts))
1230 ceph_put_mds_session(ts);
1231 }
1232}
1233
1234void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1235 struct ceph_mds_session *session)
1236{
1237 mutex_lock(&mdsc->mutex);
1238 __open_export_target_sessions(mdsc, session);
1239 mutex_unlock(&mdsc->mutex);
1240}
1241
1242/*
1243 * session caps
1244 */
1245
1246static void detach_cap_releases(struct ceph_mds_session *session,
1247 struct list_head *target)
1248{
1249 lockdep_assert_held(&session->s_cap_lock);
1250
1251 list_splice_init(&session->s_cap_releases, target);
1252 session->s_num_cap_releases = 0;
1253 dout("dispose_cap_releases mds%d\n", session->s_mds);
1254}
1255
1256static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1257 struct list_head *dispose)
1258{
1259 while (!list_empty(dispose)) {
1260 struct ceph_cap *cap;
1261 /* zero out the in-progress message */
1262 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1263 list_del(&cap->session_caps);
1264 ceph_put_cap(mdsc, cap);
1265 }
1266}
1267
1268static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1269 struct ceph_mds_session *session)
1270{
1271 struct ceph_mds_request *req;
1272 struct rb_node *p;
1273 struct ceph_inode_info *ci;
1274
1275 dout("cleanup_session_requests mds%d\n", session->s_mds);
1276 mutex_lock(&mdsc->mutex);
1277 while (!list_empty(&session->s_unsafe)) {
1278 req = list_first_entry(&session->s_unsafe,
1279 struct ceph_mds_request, r_unsafe_item);
1280 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1281 req->r_tid);
1282 if (req->r_target_inode) {
1283 /* dropping unsafe change of inode's attributes */
1284 ci = ceph_inode(req->r_target_inode);
1285 errseq_set(&ci->i_meta_err, -EIO);
1286 }
1287 if (req->r_unsafe_dir) {
1288 /* dropping unsafe directory operation */
1289 ci = ceph_inode(req->r_unsafe_dir);
1290 errseq_set(&ci->i_meta_err, -EIO);
1291 }
1292 __unregister_request(mdsc, req);
1293 }
1294 /* zero r_attempts, so kick_requests() will re-send requests */
1295 p = rb_first(&mdsc->request_tree);
1296 while (p) {
1297 req = rb_entry(p, struct ceph_mds_request, r_node);
1298 p = rb_next(p);
1299 if (req->r_session &&
1300 req->r_session->s_mds == session->s_mds)
1301 req->r_attempts = 0;
1302 }
1303 mutex_unlock(&mdsc->mutex);
1304}
1305
1306/*
1307 * Helper to safely iterate over all caps associated with a session, with
1308 * special care taken to handle a racing __ceph_remove_cap().
1309 *
1310 * Caller must hold session s_mutex.
1311 */
1312int ceph_iterate_session_caps(struct ceph_mds_session *session,
1313 int (*cb)(struct inode *, struct ceph_cap *,
1314 void *), void *arg)
1315{
1316 struct list_head *p;
1317 struct ceph_cap *cap;
1318 struct inode *inode, *last_inode = NULL;
1319 struct ceph_cap *old_cap = NULL;
1320 int ret;
1321
1322 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1323 spin_lock(&session->s_cap_lock);
1324 p = session->s_caps.next;
1325 while (p != &session->s_caps) {
1326 cap = list_entry(p, struct ceph_cap, session_caps);
1327 inode = igrab(&cap->ci->vfs_inode);
1328 if (!inode) {
1329 p = p->next;
1330 continue;
1331 }
1332 session->s_cap_iterator = cap;
1333 spin_unlock(&session->s_cap_lock);
1334
1335 if (last_inode) {
1336 /* avoid calling iput_final() while holding
1337 * s_mutex or in mds dispatch threads */
1338 ceph_async_iput(last_inode);
1339 last_inode = NULL;
1340 }
1341 if (old_cap) {
1342 ceph_put_cap(session->s_mdsc, old_cap);
1343 old_cap = NULL;
1344 }
1345
1346 ret = cb(inode, cap, arg);
1347 last_inode = inode;
1348
1349 spin_lock(&session->s_cap_lock);
1350 p = p->next;
1351 if (!cap->ci) {
1352 dout("iterate_session_caps finishing cap %p removal\n",
1353 cap);
1354 BUG_ON(cap->session != session);
1355 cap->session = NULL;
1356 list_del_init(&cap->session_caps);
1357 session->s_nr_caps--;
1358 if (cap->queue_release)
1359 __ceph_queue_cap_release(session, cap);
1360 else
1361 old_cap = cap; /* put_cap it w/o locks held */
1362 }
1363 if (ret < 0)
1364 goto out;
1365 }
1366 ret = 0;
1367out:
1368 session->s_cap_iterator = NULL;
1369 spin_unlock(&session->s_cap_lock);
1370
1371 ceph_async_iput(last_inode);
1372 if (old_cap)
1373 ceph_put_cap(session->s_mdsc, old_cap);
1374
1375 return ret;
1376}
1377
1378static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1379 void *arg)
1380{
1381 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1382 struct ceph_inode_info *ci = ceph_inode(inode);
1383 LIST_HEAD(to_remove);
1384 bool dirty_dropped = false;
1385 bool invalidate = false;
1386
1387 dout("removing cap %p, ci is %p, inode is %p\n",
1388 cap, ci, &ci->vfs_inode);
1389 spin_lock(&ci->i_ceph_lock);
1390 if (cap->mds_wanted | cap->issued)
1391 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1392 __ceph_remove_cap(cap, false);
1393 if (!ci->i_auth_cap) {
1394 struct ceph_cap_flush *cf;
1395 struct ceph_mds_client *mdsc = fsc->mdsc;
1396
1397 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1398 if (inode->i_data.nrpages > 0)
1399 invalidate = true;
1400 if (ci->i_wrbuffer_ref > 0)
1401 mapping_set_error(&inode->i_data, -EIO);
1402 }
1403
1404 while (!list_empty(&ci->i_cap_flush_list)) {
1405 cf = list_first_entry(&ci->i_cap_flush_list,
1406 struct ceph_cap_flush, i_list);
1407 list_move(&cf->i_list, &to_remove);
1408 }
1409
1410 spin_lock(&mdsc->cap_dirty_lock);
1411
1412 list_for_each_entry(cf, &to_remove, i_list)
1413 list_del(&cf->g_list);
1414
1415 if (!list_empty(&ci->i_dirty_item)) {
1416 pr_warn_ratelimited(
1417 " dropping dirty %s state for %p %lld\n",
1418 ceph_cap_string(ci->i_dirty_caps),
1419 inode, ceph_ino(inode));
1420 ci->i_dirty_caps = 0;
1421 list_del_init(&ci->i_dirty_item);
1422 dirty_dropped = true;
1423 }
1424 if (!list_empty(&ci->i_flushing_item)) {
1425 pr_warn_ratelimited(
1426 " dropping dirty+flushing %s state for %p %lld\n",
1427 ceph_cap_string(ci->i_flushing_caps),
1428 inode, ceph_ino(inode));
1429 ci->i_flushing_caps = 0;
1430 list_del_init(&ci->i_flushing_item);
1431 mdsc->num_cap_flushing--;
1432 dirty_dropped = true;
1433 }
1434 spin_unlock(&mdsc->cap_dirty_lock);
1435
1436 if (dirty_dropped) {
1437 errseq_set(&ci->i_meta_err, -EIO);
1438
1439 if (ci->i_wrbuffer_ref_head == 0 &&
1440 ci->i_wr_ref == 0 &&
1441 ci->i_dirty_caps == 0 &&
1442 ci->i_flushing_caps == 0) {
1443 ceph_put_snap_context(ci->i_head_snapc);
1444 ci->i_head_snapc = NULL;
1445 }
1446 }
1447
1448 if (atomic_read(&ci->i_filelock_ref) > 0) {
1449 /* make further file lock syscall return -EIO */
1450 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1451 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1452 inode, ceph_ino(inode));
1453 }
1454
1455 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1456 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1457 ci->i_prealloc_cap_flush = NULL;
1458 }
1459 }
1460 spin_unlock(&ci->i_ceph_lock);
1461 while (!list_empty(&to_remove)) {
1462 struct ceph_cap_flush *cf;
1463 cf = list_first_entry(&to_remove,
1464 struct ceph_cap_flush, i_list);
1465 list_del(&cf->i_list);
1466 ceph_free_cap_flush(cf);
1467 }
1468
1469 wake_up_all(&ci->i_cap_wq);
1470 if (invalidate)
1471 ceph_queue_invalidate(inode);
1472 if (dirty_dropped)
1473 iput(inode);
1474 return 0;
1475}
1476
1477/*
1478 * caller must hold session s_mutex
1479 */
1480static void remove_session_caps(struct ceph_mds_session *session)
1481{
1482 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1483 struct super_block *sb = fsc->sb;
1484 LIST_HEAD(dispose);
1485
1486 dout("remove_session_caps on %p\n", session);
1487 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1488
1489 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1490
1491 spin_lock(&session->s_cap_lock);
1492 if (session->s_nr_caps > 0) {
1493 struct inode *inode;
1494 struct ceph_cap *cap, *prev = NULL;
1495 struct ceph_vino vino;
1496 /*
1497 * iterate_session_caps() skips inodes that are being
1498 * deleted, we need to wait until deletions are complete.
1499 * __wait_on_freeing_inode() is designed for the job,
1500 * but it is not exported, so use lookup inode function
1501 * to access it.
1502 */
1503 while (!list_empty(&session->s_caps)) {
1504 cap = list_entry(session->s_caps.next,
1505 struct ceph_cap, session_caps);
1506 if (cap == prev)
1507 break;
1508 prev = cap;
1509 vino = cap->ci->i_vino;
1510 spin_unlock(&session->s_cap_lock);
1511
1512 inode = ceph_find_inode(sb, vino);
1513 /* avoid calling iput_final() while holding s_mutex */
1514 ceph_async_iput(inode);
1515
1516 spin_lock(&session->s_cap_lock);
1517 }
1518 }
1519
1520 // drop cap expires and unlock s_cap_lock
1521 detach_cap_releases(session, &dispose);
1522
1523 BUG_ON(session->s_nr_caps > 0);
1524 BUG_ON(!list_empty(&session->s_cap_flushing));
1525 spin_unlock(&session->s_cap_lock);
1526 dispose_cap_releases(session->s_mdsc, &dispose);
1527}
1528
1529enum {
1530 RECONNECT,
1531 RENEWCAPS,
1532 FORCE_RO,
1533};
1534
1535/*
1536 * wake up any threads waiting on this session's caps. if the cap is
1537 * old (didn't get renewed on the client reconnect), remove it now.
1538 *
1539 * caller must hold s_mutex.
1540 */
1541static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1542 void *arg)
1543{
1544 struct ceph_inode_info *ci = ceph_inode(inode);
1545 unsigned long ev = (unsigned long)arg;
1546
1547 if (ev == RECONNECT) {
1548 spin_lock(&ci->i_ceph_lock);
1549 ci->i_wanted_max_size = 0;
1550 ci->i_requested_max_size = 0;
1551 spin_unlock(&ci->i_ceph_lock);
1552 } else if (ev == RENEWCAPS) {
1553 if (cap->cap_gen < cap->session->s_cap_gen) {
1554 /* mds did not re-issue stale cap */
1555 spin_lock(&ci->i_ceph_lock);
1556 cap->issued = cap->implemented = CEPH_CAP_PIN;
1557 /* make sure mds knows what we want */
1558 if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1559 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1560 spin_unlock(&ci->i_ceph_lock);
1561 }
1562 } else if (ev == FORCE_RO) {
1563 }
1564 wake_up_all(&ci->i_cap_wq);
1565 return 0;
1566}
1567
1568static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1569{
1570 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1571 ceph_iterate_session_caps(session, wake_up_session_cb,
1572 (void *)(unsigned long)ev);
1573}
1574
1575/*
1576 * Send periodic message to MDS renewing all currently held caps. The
1577 * ack will reset the expiration for all caps from this session.
1578 *
1579 * caller holds s_mutex
1580 */
1581static int send_renew_caps(struct ceph_mds_client *mdsc,
1582 struct ceph_mds_session *session)
1583{
1584 struct ceph_msg *msg;
1585 int state;
1586
1587 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1588 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1589 pr_info("mds%d caps stale\n", session->s_mds);
1590 session->s_renew_requested = jiffies;
1591
1592 /* do not try to renew caps until a recovering mds has reconnected
1593 * with its clients. */
1594 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1595 if (state < CEPH_MDS_STATE_RECONNECT) {
1596 dout("send_renew_caps ignoring mds%d (%s)\n",
1597 session->s_mds, ceph_mds_state_name(state));
1598 return 0;
1599 }
1600
1601 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1602 ceph_mds_state_name(state));
1603 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1604 ++session->s_renew_seq);
1605 if (!msg)
1606 return -ENOMEM;
1607 ceph_con_send(&session->s_con, msg);
1608 return 0;
1609}
1610
1611static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1612 struct ceph_mds_session *session, u64 seq)
1613{
1614 struct ceph_msg *msg;
1615
1616 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1617 session->s_mds, ceph_session_state_name(session->s_state), seq);
1618 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1619 if (!msg)
1620 return -ENOMEM;
1621 ceph_con_send(&session->s_con, msg);
1622 return 0;
1623}
1624
1625
1626/*
1627 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1628 *
1629 * Called under session->s_mutex
1630 */
1631static void renewed_caps(struct ceph_mds_client *mdsc,
1632 struct ceph_mds_session *session, int is_renew)
1633{
1634 int was_stale;
1635 int wake = 0;
1636
1637 spin_lock(&session->s_cap_lock);
1638 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1639
1640 session->s_cap_ttl = session->s_renew_requested +
1641 mdsc->mdsmap->m_session_timeout*HZ;
1642
1643 if (was_stale) {
1644 if (time_before(jiffies, session->s_cap_ttl)) {
1645 pr_info("mds%d caps renewed\n", session->s_mds);
1646 wake = 1;
1647 } else {
1648 pr_info("mds%d caps still stale\n", session->s_mds);
1649 }
1650 }
1651 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1652 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1653 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1654 spin_unlock(&session->s_cap_lock);
1655
1656 if (wake)
1657 wake_up_session_caps(session, RENEWCAPS);
1658}
1659
1660/*
1661 * send a session close request
1662 */
1663static int request_close_session(struct ceph_mds_client *mdsc,
1664 struct ceph_mds_session *session)
1665{
1666 struct ceph_msg *msg;
1667
1668 dout("request_close_session mds%d state %s seq %lld\n",
1669 session->s_mds, ceph_session_state_name(session->s_state),
1670 session->s_seq);
1671 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1672 if (!msg)
1673 return -ENOMEM;
1674 ceph_con_send(&session->s_con, msg);
1675 return 1;
1676}
1677
1678/*
1679 * Called with s_mutex held.
1680 */
1681static int __close_session(struct ceph_mds_client *mdsc,
1682 struct ceph_mds_session *session)
1683{
1684 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1685 return 0;
1686 session->s_state = CEPH_MDS_SESSION_CLOSING;
1687 return request_close_session(mdsc, session);
1688}
1689
1690static bool drop_negative_children(struct dentry *dentry)
1691{
1692 struct dentry *child;
1693 bool all_negative = true;
1694
1695 if (!d_is_dir(dentry))
1696 goto out;
1697
1698 spin_lock(&dentry->d_lock);
1699 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1700 if (d_really_is_positive(child)) {
1701 all_negative = false;
1702 break;
1703 }
1704 }
1705 spin_unlock(&dentry->d_lock);
1706
1707 if (all_negative)
1708 shrink_dcache_parent(dentry);
1709out:
1710 return all_negative;
1711}
1712
1713/*
1714 * Trim old(er) caps.
1715 *
1716 * Because we can't cache an inode without one or more caps, we do
1717 * this indirectly: if a cap is unused, we prune its aliases, at which
1718 * point the inode will hopefully get dropped to.
1719 *
1720 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1721 * memory pressure from the MDS, though, so it needn't be perfect.
1722 */
1723static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1724{
1725 int *remaining = arg;
1726 struct ceph_inode_info *ci = ceph_inode(inode);
1727 int used, wanted, oissued, mine;
1728
1729 if (*remaining <= 0)
1730 return -1;
1731
1732 spin_lock(&ci->i_ceph_lock);
1733 mine = cap->issued | cap->implemented;
1734 used = __ceph_caps_used(ci);
1735 wanted = __ceph_caps_file_wanted(ci);
1736 oissued = __ceph_caps_issued_other(ci, cap);
1737
1738 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1739 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1740 ceph_cap_string(used), ceph_cap_string(wanted));
1741 if (cap == ci->i_auth_cap) {
1742 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1743 !list_empty(&ci->i_cap_snaps))
1744 goto out;
1745 if ((used | wanted) & CEPH_CAP_ANY_WR)
1746 goto out;
1747 /* Note: it's possible that i_filelock_ref becomes non-zero
1748 * after dropping auth caps. It doesn't hurt because reply
1749 * of lock mds request will re-add auth caps. */
1750 if (atomic_read(&ci->i_filelock_ref) > 0)
1751 goto out;
1752 }
1753 /* The inode has cached pages, but it's no longer used.
1754 * we can safely drop it */
1755 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1756 !(oissued & CEPH_CAP_FILE_CACHE)) {
1757 used = 0;
1758 oissued = 0;
1759 }
1760 if ((used | wanted) & ~oissued & mine)
1761 goto out; /* we need these caps */
1762
1763 if (oissued) {
1764 /* we aren't the only cap.. just remove us */
1765 __ceph_remove_cap(cap, true);
1766 (*remaining)--;
1767 } else {
1768 struct dentry *dentry;
1769 /* try dropping referring dentries */
1770 spin_unlock(&ci->i_ceph_lock);
1771 dentry = d_find_any_alias(inode);
1772 if (dentry && drop_negative_children(dentry)) {
1773 int count;
1774 dput(dentry);
1775 d_prune_aliases(inode);
1776 count = atomic_read(&inode->i_count);
1777 if (count == 1)
1778 (*remaining)--;
1779 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1780 inode, cap, count);
1781 } else {
1782 dput(dentry);
1783 }
1784 return 0;
1785 }
1786
1787out:
1788 spin_unlock(&ci->i_ceph_lock);
1789 return 0;
1790}
1791
1792/*
1793 * Trim session cap count down to some max number.
1794 */
1795int ceph_trim_caps(struct ceph_mds_client *mdsc,
1796 struct ceph_mds_session *session,
1797 int max_caps)
1798{
1799 int trim_caps = session->s_nr_caps - max_caps;
1800
1801 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1802 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1803 if (trim_caps > 0) {
1804 int remaining = trim_caps;
1805
1806 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1807 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1808 session->s_mds, session->s_nr_caps, max_caps,
1809 trim_caps - remaining);
1810 }
1811
1812 ceph_flush_cap_releases(mdsc, session);
1813 return 0;
1814}
1815
1816static int check_caps_flush(struct ceph_mds_client *mdsc,
1817 u64 want_flush_tid)
1818{
1819 int ret = 1;
1820
1821 spin_lock(&mdsc->cap_dirty_lock);
1822 if (!list_empty(&mdsc->cap_flush_list)) {
1823 struct ceph_cap_flush *cf =
1824 list_first_entry(&mdsc->cap_flush_list,
1825 struct ceph_cap_flush, g_list);
1826 if (cf->tid <= want_flush_tid) {
1827 dout("check_caps_flush still flushing tid "
1828 "%llu <= %llu\n", cf->tid, want_flush_tid);
1829 ret = 0;
1830 }
1831 }
1832 spin_unlock(&mdsc->cap_dirty_lock);
1833 return ret;
1834}
1835
1836/*
1837 * flush all dirty inode data to disk.
1838 *
1839 * returns true if we've flushed through want_flush_tid
1840 */
1841static void wait_caps_flush(struct ceph_mds_client *mdsc,
1842 u64 want_flush_tid)
1843{
1844 dout("check_caps_flush want %llu\n", want_flush_tid);
1845
1846 wait_event(mdsc->cap_flushing_wq,
1847 check_caps_flush(mdsc, want_flush_tid));
1848
1849 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1850}
1851
1852/*
1853 * called under s_mutex
1854 */
1855static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1856 struct ceph_mds_session *session)
1857{
1858 struct ceph_msg *msg = NULL;
1859 struct ceph_mds_cap_release *head;
1860 struct ceph_mds_cap_item *item;
1861 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1862 struct ceph_cap *cap;
1863 LIST_HEAD(tmp_list);
1864 int num_cap_releases;
1865 __le32 barrier, *cap_barrier;
1866
1867 down_read(&osdc->lock);
1868 barrier = cpu_to_le32(osdc->epoch_barrier);
1869 up_read(&osdc->lock);
1870
1871 spin_lock(&session->s_cap_lock);
1872again:
1873 list_splice_init(&session->s_cap_releases, &tmp_list);
1874 num_cap_releases = session->s_num_cap_releases;
1875 session->s_num_cap_releases = 0;
1876 spin_unlock(&session->s_cap_lock);
1877
1878 while (!list_empty(&tmp_list)) {
1879 if (!msg) {
1880 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1881 PAGE_SIZE, GFP_NOFS, false);
1882 if (!msg)
1883 goto out_err;
1884 head = msg->front.iov_base;
1885 head->num = cpu_to_le32(0);
1886 msg->front.iov_len = sizeof(*head);
1887
1888 msg->hdr.version = cpu_to_le16(2);
1889 msg->hdr.compat_version = cpu_to_le16(1);
1890 }
1891
1892 cap = list_first_entry(&tmp_list, struct ceph_cap,
1893 session_caps);
1894 list_del(&cap->session_caps);
1895 num_cap_releases--;
1896
1897 head = msg->front.iov_base;
1898 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1899 &head->num);
1900 item = msg->front.iov_base + msg->front.iov_len;
1901 item->ino = cpu_to_le64(cap->cap_ino);
1902 item->cap_id = cpu_to_le64(cap->cap_id);
1903 item->migrate_seq = cpu_to_le32(cap->mseq);
1904 item->seq = cpu_to_le32(cap->issue_seq);
1905 msg->front.iov_len += sizeof(*item);
1906
1907 ceph_put_cap(mdsc, cap);
1908
1909 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1910 // Append cap_barrier field
1911 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1912 *cap_barrier = barrier;
1913 msg->front.iov_len += sizeof(*cap_barrier);
1914
1915 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1916 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1917 ceph_con_send(&session->s_con, msg);
1918 msg = NULL;
1919 }
1920 }
1921
1922 BUG_ON(num_cap_releases != 0);
1923
1924 spin_lock(&session->s_cap_lock);
1925 if (!list_empty(&session->s_cap_releases))
1926 goto again;
1927 spin_unlock(&session->s_cap_lock);
1928
1929 if (msg) {
1930 // Append cap_barrier field
1931 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1932 *cap_barrier = barrier;
1933 msg->front.iov_len += sizeof(*cap_barrier);
1934
1935 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1936 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1937 ceph_con_send(&session->s_con, msg);
1938 }
1939 return;
1940out_err:
1941 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1942 session->s_mds);
1943 spin_lock(&session->s_cap_lock);
1944 list_splice(&tmp_list, &session->s_cap_releases);
1945 session->s_num_cap_releases += num_cap_releases;
1946 spin_unlock(&session->s_cap_lock);
1947}
1948
1949static void ceph_cap_release_work(struct work_struct *work)
1950{
1951 struct ceph_mds_session *session =
1952 container_of(work, struct ceph_mds_session, s_cap_release_work);
1953
1954 mutex_lock(&session->s_mutex);
1955 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1956 session->s_state == CEPH_MDS_SESSION_HUNG)
1957 ceph_send_cap_releases(session->s_mdsc, session);
1958 mutex_unlock(&session->s_mutex);
1959 ceph_put_mds_session(session);
1960}
1961
1962void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1963 struct ceph_mds_session *session)
1964{
1965 if (mdsc->stopping)
1966 return;
1967
1968 get_session(session);
1969 if (queue_work(mdsc->fsc->cap_wq,
1970 &session->s_cap_release_work)) {
1971 dout("cap release work queued\n");
1972 } else {
1973 ceph_put_mds_session(session);
1974 dout("failed to queue cap release work\n");
1975 }
1976}
1977
1978/*
1979 * caller holds session->s_cap_lock
1980 */
1981void __ceph_queue_cap_release(struct ceph_mds_session *session,
1982 struct ceph_cap *cap)
1983{
1984 list_add_tail(&cap->session_caps, &session->s_cap_releases);
1985 session->s_num_cap_releases++;
1986
1987 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
1988 ceph_flush_cap_releases(session->s_mdsc, session);
1989}
1990
1991static void ceph_cap_reclaim_work(struct work_struct *work)
1992{
1993 struct ceph_mds_client *mdsc =
1994 container_of(work, struct ceph_mds_client, cap_reclaim_work);
1995 int ret = ceph_trim_dentries(mdsc);
1996 if (ret == -EAGAIN)
1997 ceph_queue_cap_reclaim_work(mdsc);
1998}
1999
2000void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2001{
2002 if (mdsc->stopping)
2003 return;
2004
2005 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2006 dout("caps reclaim work queued\n");
2007 } else {
2008 dout("failed to queue caps release work\n");
2009 }
2010}
2011
2012void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2013{
2014 int val;
2015 if (!nr)
2016 return;
2017 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2018 if (!(val % CEPH_CAPS_PER_RELEASE)) {
2019 atomic_set(&mdsc->cap_reclaim_pending, 0);
2020 ceph_queue_cap_reclaim_work(mdsc);
2021 }
2022}
2023
2024/*
2025 * requests
2026 */
2027
2028int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2029 struct inode *dir)
2030{
2031 struct ceph_inode_info *ci = ceph_inode(dir);
2032 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2033 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2034 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2035 int order, num_entries;
2036
2037 spin_lock(&ci->i_ceph_lock);
2038 num_entries = ci->i_files + ci->i_subdirs;
2039 spin_unlock(&ci->i_ceph_lock);
2040 num_entries = max(num_entries, 1);
2041 num_entries = min(num_entries, opt->max_readdir);
2042
2043 order = get_order(size * num_entries);
2044 while (order >= 0) {
2045 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2046 __GFP_NOWARN,
2047 order);
2048 if (rinfo->dir_entries)
2049 break;
2050 order--;
2051 }
2052 if (!rinfo->dir_entries)
2053 return -ENOMEM;
2054
2055 num_entries = (PAGE_SIZE << order) / size;
2056 num_entries = min(num_entries, opt->max_readdir);
2057
2058 rinfo->dir_buf_size = PAGE_SIZE << order;
2059 req->r_num_caps = num_entries + 1;
2060 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2061 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2062 return 0;
2063}
2064
2065/*
2066 * Create an mds request.
2067 */
2068struct ceph_mds_request *
2069ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2070{
2071 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2072 struct timespec64 ts;
2073
2074 if (!req)
2075 return ERR_PTR(-ENOMEM);
2076
2077 mutex_init(&req->r_fill_mutex);
2078 req->r_mdsc = mdsc;
2079 req->r_started = jiffies;
2080 req->r_resend_mds = -1;
2081 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2082 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2083 req->r_fmode = -1;
2084 kref_init(&req->r_kref);
2085 RB_CLEAR_NODE(&req->r_node);
2086 INIT_LIST_HEAD(&req->r_wait);
2087 init_completion(&req->r_completion);
2088 init_completion(&req->r_safe_completion);
2089 INIT_LIST_HEAD(&req->r_unsafe_item);
2090
2091 ktime_get_coarse_real_ts64(&ts);
2092 req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
2093
2094 req->r_op = op;
2095 req->r_direct_mode = mode;
2096 return req;
2097}
2098
2099/*
2100 * return oldest (lowest) request, tid in request tree, 0 if none.
2101 *
2102 * called under mdsc->mutex.
2103 */
2104static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2105{
2106 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2107 return NULL;
2108 return rb_entry(rb_first(&mdsc->request_tree),
2109 struct ceph_mds_request, r_node);
2110}
2111
2112static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2113{
2114 return mdsc->oldest_tid;
2115}
2116
2117/*
2118 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2119 * on build_path_from_dentry in fs/cifs/dir.c.
2120 *
2121 * If @stop_on_nosnap, generate path relative to the first non-snapped
2122 * inode.
2123 *
2124 * Encode hidden .snap dirs as a double /, i.e.
2125 * foo/.snap/bar -> foo//bar
2126 */
2127char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2128 int stop_on_nosnap)
2129{
2130 struct dentry *temp;
2131 char *path;
2132 int pos;
2133 unsigned seq;
2134 u64 base;
2135
2136 if (!dentry)
2137 return ERR_PTR(-EINVAL);
2138
2139 path = __getname();
2140 if (!path)
2141 return ERR_PTR(-ENOMEM);
2142retry:
2143 pos = PATH_MAX - 1;
2144 path[pos] = '\0';
2145
2146 seq = read_seqbegin(&rename_lock);
2147 rcu_read_lock();
2148 temp = dentry;
2149 for (;;) {
2150 struct inode *inode;
2151
2152 spin_lock(&temp->d_lock);
2153 inode = d_inode(temp);
2154 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2155 dout("build_path path+%d: %p SNAPDIR\n",
2156 pos, temp);
2157 } else if (stop_on_nosnap && inode && dentry != temp &&
2158 ceph_snap(inode) == CEPH_NOSNAP) {
2159 spin_unlock(&temp->d_lock);
2160 pos++; /* get rid of any prepended '/' */
2161 break;
2162 } else {
2163 pos -= temp->d_name.len;
2164 if (pos < 0) {
2165 spin_unlock(&temp->d_lock);
2166 break;
2167 }
2168 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2169 }
2170 spin_unlock(&temp->d_lock);
2171 temp = READ_ONCE(temp->d_parent);
2172
2173 /* Are we at the root? */
2174 if (IS_ROOT(temp))
2175 break;
2176
2177 /* Are we out of buffer? */
2178 if (--pos < 0)
2179 break;
2180
2181 path[pos] = '/';
2182 }
2183 base = ceph_ino(d_inode(temp));
2184 rcu_read_unlock();
2185 if (pos < 0 || read_seqretry(&rename_lock, seq)) {
2186 pr_err("build_path did not end path lookup where "
2187 "expected, pos is %d\n", pos);
2188 /* presumably this is only possible if racing with a
2189 rename of one of the parent directories (we can not
2190 lock the dentries above us to prevent this, but
2191 retrying should be harmless) */
2192 goto retry;
2193 }
2194
2195 *pbase = base;
2196 *plen = PATH_MAX - 1 - pos;
2197 dout("build_path on %p %d built %llx '%.*s'\n",
2198 dentry, d_count(dentry), base, *plen, path + pos);
2199 return path + pos;
2200}
2201
2202static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2203 const char **ppath, int *ppathlen, u64 *pino,
2204 bool *pfreepath, bool parent_locked)
2205{
2206 char *path;
2207
2208 rcu_read_lock();
2209 if (!dir)
2210 dir = d_inode_rcu(dentry->d_parent);
2211 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2212 *pino = ceph_ino(dir);
2213 rcu_read_unlock();
2214 *ppath = dentry->d_name.name;
2215 *ppathlen = dentry->d_name.len;
2216 return 0;
2217 }
2218 rcu_read_unlock();
2219 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2220 if (IS_ERR(path))
2221 return PTR_ERR(path);
2222 *ppath = path;
2223 *pfreepath = true;
2224 return 0;
2225}
2226
2227static int build_inode_path(struct inode *inode,
2228 const char **ppath, int *ppathlen, u64 *pino,
2229 bool *pfreepath)
2230{
2231 struct dentry *dentry;
2232 char *path;
2233
2234 if (ceph_snap(inode) == CEPH_NOSNAP) {
2235 *pino = ceph_ino(inode);
2236 *ppathlen = 0;
2237 return 0;
2238 }
2239 dentry = d_find_alias(inode);
2240 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2241 dput(dentry);
2242 if (IS_ERR(path))
2243 return PTR_ERR(path);
2244 *ppath = path;
2245 *pfreepath = true;
2246 return 0;
2247}
2248
2249/*
2250 * request arguments may be specified via an inode *, a dentry *, or
2251 * an explicit ino+path.
2252 */
2253static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2254 struct inode *rdiri, const char *rpath,
2255 u64 rino, const char **ppath, int *pathlen,
2256 u64 *ino, bool *freepath, bool parent_locked)
2257{
2258 int r = 0;
2259
2260 if (rinode) {
2261 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2262 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2263 ceph_snap(rinode));
2264 } else if (rdentry) {
2265 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2266 freepath, parent_locked);
2267 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2268 *ppath);
2269 } else if (rpath || rino) {
2270 *ino = rino;
2271 *ppath = rpath;
2272 *pathlen = rpath ? strlen(rpath) : 0;
2273 dout(" path %.*s\n", *pathlen, rpath);
2274 }
2275
2276 return r;
2277}
2278
2279/*
2280 * called under mdsc->mutex
2281 */
2282static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2283 struct ceph_mds_request *req,
2284 int mds, bool drop_cap_releases)
2285{
2286 struct ceph_msg *msg;
2287 struct ceph_mds_request_head *head;
2288 const char *path1 = NULL;
2289 const char *path2 = NULL;
2290 u64 ino1 = 0, ino2 = 0;
2291 int pathlen1 = 0, pathlen2 = 0;
2292 bool freepath1 = false, freepath2 = false;
2293 int len;
2294 u16 releases;
2295 void *p, *end;
2296 int ret;
2297
2298 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2299 req->r_parent, req->r_path1, req->r_ino1.ino,
2300 &path1, &pathlen1, &ino1, &freepath1,
2301 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2302 &req->r_req_flags));
2303 if (ret < 0) {
2304 msg = ERR_PTR(ret);
2305 goto out;
2306 }
2307
2308 /* If r_old_dentry is set, then assume that its parent is locked */
2309 ret = set_request_path_attr(NULL, req->r_old_dentry,
2310 req->r_old_dentry_dir,
2311 req->r_path2, req->r_ino2.ino,
2312 &path2, &pathlen2, &ino2, &freepath2, true);
2313 if (ret < 0) {
2314 msg = ERR_PTR(ret);
2315 goto out_free1;
2316 }
2317
2318 len = sizeof(*head) +
2319 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2320 sizeof(struct ceph_timespec);
2321
2322 /* calculate (max) length for cap releases */
2323 len += sizeof(struct ceph_mds_request_release) *
2324 (!!req->r_inode_drop + !!req->r_dentry_drop +
2325 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2326 if (req->r_dentry_drop)
2327 len += pathlen1;
2328 if (req->r_old_dentry_drop)
2329 len += pathlen2;
2330
2331 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2332 if (!msg) {
2333 msg = ERR_PTR(-ENOMEM);
2334 goto out_free2;
2335 }
2336
2337 msg->hdr.version = cpu_to_le16(2);
2338 msg->hdr.tid = cpu_to_le64(req->r_tid);
2339
2340 head = msg->front.iov_base;
2341 p = msg->front.iov_base + sizeof(*head);
2342 end = msg->front.iov_base + msg->front.iov_len;
2343
2344 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2345 head->op = cpu_to_le32(req->r_op);
2346 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2347 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2348 head->args = req->r_args;
2349
2350 ceph_encode_filepath(&p, end, ino1, path1);
2351 ceph_encode_filepath(&p, end, ino2, path2);
2352
2353 /* make note of release offset, in case we need to replay */
2354 req->r_request_release_offset = p - msg->front.iov_base;
2355
2356 /* cap releases */
2357 releases = 0;
2358 if (req->r_inode_drop)
2359 releases += ceph_encode_inode_release(&p,
2360 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2361 mds, req->r_inode_drop, req->r_inode_unless, 0);
2362 if (req->r_dentry_drop)
2363 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2364 req->r_parent, mds, req->r_dentry_drop,
2365 req->r_dentry_unless);
2366 if (req->r_old_dentry_drop)
2367 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2368 req->r_old_dentry_dir, mds,
2369 req->r_old_dentry_drop,
2370 req->r_old_dentry_unless);
2371 if (req->r_old_inode_drop)
2372 releases += ceph_encode_inode_release(&p,
2373 d_inode(req->r_old_dentry),
2374 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2375
2376 if (drop_cap_releases) {
2377 releases = 0;
2378 p = msg->front.iov_base + req->r_request_release_offset;
2379 }
2380
2381 head->num_releases = cpu_to_le16(releases);
2382
2383 /* time stamp */
2384 {
2385 struct ceph_timespec ts;
2386 ceph_encode_timespec64(&ts, &req->r_stamp);
2387 ceph_encode_copy(&p, &ts, sizeof(ts));
2388 }
2389
2390 BUG_ON(p > end);
2391 msg->front.iov_len = p - msg->front.iov_base;
2392 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2393
2394 if (req->r_pagelist) {
2395 struct ceph_pagelist *pagelist = req->r_pagelist;
2396 ceph_msg_data_add_pagelist(msg, pagelist);
2397 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2398 } else {
2399 msg->hdr.data_len = 0;
2400 }
2401
2402 msg->hdr.data_off = cpu_to_le16(0);
2403
2404out_free2:
2405 if (freepath2)
2406 ceph_mdsc_free_path((char *)path2, pathlen2);
2407out_free1:
2408 if (freepath1)
2409 ceph_mdsc_free_path((char *)path1, pathlen1);
2410out:
2411 return msg;
2412}
2413
2414/*
2415 * called under mdsc->mutex if error, under no mutex if
2416 * success.
2417 */
2418static void complete_request(struct ceph_mds_client *mdsc,
2419 struct ceph_mds_request *req)
2420{
2421 if (req->r_callback)
2422 req->r_callback(mdsc, req);
2423 complete_all(&req->r_completion);
2424}
2425
2426/*
2427 * called under mdsc->mutex
2428 */
2429static int __prepare_send_request(struct ceph_mds_client *mdsc,
2430 struct ceph_mds_request *req,
2431 int mds, bool drop_cap_releases)
2432{
2433 struct ceph_mds_request_head *rhead;
2434 struct ceph_msg *msg;
2435 int flags = 0;
2436
2437 req->r_attempts++;
2438 if (req->r_inode) {
2439 struct ceph_cap *cap =
2440 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2441
2442 if (cap)
2443 req->r_sent_on_mseq = cap->mseq;
2444 else
2445 req->r_sent_on_mseq = -1;
2446 }
2447 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2448 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2449
2450 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2451 void *p;
2452 /*
2453 * Replay. Do not regenerate message (and rebuild
2454 * paths, etc.); just use the original message.
2455 * Rebuilding paths will break for renames because
2456 * d_move mangles the src name.
2457 */
2458 msg = req->r_request;
2459 rhead = msg->front.iov_base;
2460
2461 flags = le32_to_cpu(rhead->flags);
2462 flags |= CEPH_MDS_FLAG_REPLAY;
2463 rhead->flags = cpu_to_le32(flags);
2464
2465 if (req->r_target_inode)
2466 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2467
2468 rhead->num_retry = req->r_attempts - 1;
2469
2470 /* remove cap/dentry releases from message */
2471 rhead->num_releases = 0;
2472
2473 /* time stamp */
2474 p = msg->front.iov_base + req->r_request_release_offset;
2475 {
2476 struct ceph_timespec ts;
2477 ceph_encode_timespec64(&ts, &req->r_stamp);
2478 ceph_encode_copy(&p, &ts, sizeof(ts));
2479 }
2480
2481 msg->front.iov_len = p - msg->front.iov_base;
2482 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2483 return 0;
2484 }
2485
2486 if (req->r_request) {
2487 ceph_msg_put(req->r_request);
2488 req->r_request = NULL;
2489 }
2490 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2491 if (IS_ERR(msg)) {
2492 req->r_err = PTR_ERR(msg);
2493 return PTR_ERR(msg);
2494 }
2495 req->r_request = msg;
2496
2497 rhead = msg->front.iov_base;
2498 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2499 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2500 flags |= CEPH_MDS_FLAG_REPLAY;
2501 if (req->r_parent)
2502 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2503 rhead->flags = cpu_to_le32(flags);
2504 rhead->num_fwd = req->r_num_fwd;
2505 rhead->num_retry = req->r_attempts - 1;
2506 rhead->ino = 0;
2507
2508 dout(" r_parent = %p\n", req->r_parent);
2509 return 0;
2510}
2511
2512/*
2513 * send request, or put it on the appropriate wait list.
2514 */
2515static void __do_request(struct ceph_mds_client *mdsc,
2516 struct ceph_mds_request *req)
2517{
2518 struct ceph_mds_session *session = NULL;
2519 int mds = -1;
2520 int err = 0;
2521
2522 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2523 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2524 __unregister_request(mdsc, req);
2525 return;
2526 }
2527
2528 if (req->r_timeout &&
2529 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2530 dout("do_request timed out\n");
2531 err = -EIO;
2532 goto finish;
2533 }
2534 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2535 dout("do_request forced umount\n");
2536 err = -EIO;
2537 goto finish;
2538 }
2539 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2540 if (mdsc->mdsmap_err) {
2541 err = mdsc->mdsmap_err;
2542 dout("do_request mdsmap err %d\n", err);
2543 goto finish;
2544 }
2545 if (mdsc->mdsmap->m_epoch == 0) {
2546 dout("do_request no mdsmap, waiting for map\n");
2547 list_add(&req->r_wait, &mdsc->waiting_for_map);
2548 return;
2549 }
2550 if (!(mdsc->fsc->mount_options->flags &
2551 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2552 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2553 err = -ENOENT;
2554 pr_info("probably no mds server is up\n");
2555 goto finish;
2556 }
2557 }
2558
2559 put_request_session(req);
2560
2561 mds = __choose_mds(mdsc, req);
2562 if (mds < 0 ||
2563 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2564 dout("do_request no mds or not active, waiting for map\n");
2565 list_add(&req->r_wait, &mdsc->waiting_for_map);
2566 return;
2567 }
2568
2569 /* get, open session */
2570 session = __ceph_lookup_mds_session(mdsc, mds);
2571 if (!session) {
2572 session = register_session(mdsc, mds);
2573 if (IS_ERR(session)) {
2574 err = PTR_ERR(session);
2575 goto finish;
2576 }
2577 }
2578 req->r_session = get_session(session);
2579
2580 dout("do_request mds%d session %p state %s\n", mds, session,
2581 ceph_session_state_name(session->s_state));
2582 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2583 session->s_state != CEPH_MDS_SESSION_HUNG) {
2584 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2585 err = -EACCES;
2586 goto out_session;
2587 }
2588 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2589 session->s_state == CEPH_MDS_SESSION_CLOSING)
2590 __open_session(mdsc, session);
2591 list_add(&req->r_wait, &session->s_waiting);
2592 goto out_session;
2593 }
2594
2595 /* send request */
2596 req->r_resend_mds = -1; /* forget any previous mds hint */
2597
2598 if (req->r_request_started == 0) /* note request start time */
2599 req->r_request_started = jiffies;
2600
2601 err = __prepare_send_request(mdsc, req, mds, false);
2602 if (!err) {
2603 ceph_msg_get(req->r_request);
2604 ceph_con_send(&session->s_con, req->r_request);
2605 }
2606
2607out_session:
2608 ceph_put_mds_session(session);
2609finish:
2610 if (err) {
2611 dout("__do_request early error %d\n", err);
2612 req->r_err = err;
2613 complete_request(mdsc, req);
2614 __unregister_request(mdsc, req);
2615 }
2616 return;
2617}
2618
2619/*
2620 * called under mdsc->mutex
2621 */
2622static void __wake_requests(struct ceph_mds_client *mdsc,
2623 struct list_head *head)
2624{
2625 struct ceph_mds_request *req;
2626 LIST_HEAD(tmp_list);
2627
2628 list_splice_init(head, &tmp_list);
2629
2630 while (!list_empty(&tmp_list)) {
2631 req = list_entry(tmp_list.next,
2632 struct ceph_mds_request, r_wait);
2633 list_del_init(&req->r_wait);
2634 dout(" wake request %p tid %llu\n", req, req->r_tid);
2635 __do_request(mdsc, req);
2636 }
2637}
2638
2639/*
2640 * Wake up threads with requests pending for @mds, so that they can
2641 * resubmit their requests to a possibly different mds.
2642 */
2643static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2644{
2645 struct ceph_mds_request *req;
2646 struct rb_node *p = rb_first(&mdsc->request_tree);
2647
2648 dout("kick_requests mds%d\n", mds);
2649 while (p) {
2650 req = rb_entry(p, struct ceph_mds_request, r_node);
2651 p = rb_next(p);
2652 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2653 continue;
2654 if (req->r_attempts > 0)
2655 continue; /* only new requests */
2656 if (req->r_session &&
2657 req->r_session->s_mds == mds) {
2658 dout(" kicking tid %llu\n", req->r_tid);
2659 list_del_init(&req->r_wait);
2660 __do_request(mdsc, req);
2661 }
2662 }
2663}
2664
2665int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2666 struct ceph_mds_request *req)
2667{
2668 int err;
2669
2670 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2671 if (req->r_inode)
2672 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2673 if (req->r_parent)
2674 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2675 if (req->r_old_dentry_dir)
2676 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2677 CEPH_CAP_PIN);
2678
2679 dout("submit_request on %p for inode %p\n", req, dir);
2680 mutex_lock(&mdsc->mutex);
2681 __register_request(mdsc, req, dir);
2682 __do_request(mdsc, req);
2683 err = req->r_err;
2684 mutex_unlock(&mdsc->mutex);
2685 return err;
2686}
2687
2688static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2689 struct ceph_mds_request *req)
2690{
2691 int err;
2692
2693 /* wait */
2694 dout("do_request waiting\n");
2695 if (!req->r_timeout && req->r_wait_for_completion) {
2696 err = req->r_wait_for_completion(mdsc, req);
2697 } else {
2698 long timeleft = wait_for_completion_killable_timeout(
2699 &req->r_completion,
2700 ceph_timeout_jiffies(req->r_timeout));
2701 if (timeleft > 0)
2702 err = 0;
2703 else if (!timeleft)
2704 err = -EIO; /* timed out */
2705 else
2706 err = timeleft; /* killed */
2707 }
2708 dout("do_request waited, got %d\n", err);
2709 mutex_lock(&mdsc->mutex);
2710
2711 /* only abort if we didn't race with a real reply */
2712 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2713 err = le32_to_cpu(req->r_reply_info.head->result);
2714 } else if (err < 0) {
2715 dout("aborted request %lld with %d\n", req->r_tid, err);
2716
2717 /*
2718 * ensure we aren't running concurrently with
2719 * ceph_fill_trace or ceph_readdir_prepopulate, which
2720 * rely on locks (dir mutex) held by our caller.
2721 */
2722 mutex_lock(&req->r_fill_mutex);
2723 req->r_err = err;
2724 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2725 mutex_unlock(&req->r_fill_mutex);
2726
2727 if (req->r_parent &&
2728 (req->r_op & CEPH_MDS_OP_WRITE))
2729 ceph_invalidate_dir_request(req);
2730 } else {
2731 err = req->r_err;
2732 }
2733
2734 mutex_unlock(&mdsc->mutex);
2735 return err;
2736}
2737
2738/*
2739 * Synchrously perform an mds request. Take care of all of the
2740 * session setup, forwarding, retry details.
2741 */
2742int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2743 struct inode *dir,
2744 struct ceph_mds_request *req)
2745{
2746 int err;
2747
2748 dout("do_request on %p\n", req);
2749
2750 /* issue */
2751 err = ceph_mdsc_submit_request(mdsc, dir, req);
2752 if (!err)
2753 err = ceph_mdsc_wait_request(mdsc, req);
2754 dout("do_request %p done, result %d\n", req, err);
2755 return err;
2756}
2757
2758/*
2759 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2760 * namespace request.
2761 */
2762void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2763{
2764 struct inode *dir = req->r_parent;
2765 struct inode *old_dir = req->r_old_dentry_dir;
2766
2767 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2768
2769 ceph_dir_clear_complete(dir);
2770 if (old_dir)
2771 ceph_dir_clear_complete(old_dir);
2772 if (req->r_dentry)
2773 ceph_invalidate_dentry_lease(req->r_dentry);
2774 if (req->r_old_dentry)
2775 ceph_invalidate_dentry_lease(req->r_old_dentry);
2776}
2777
2778/*
2779 * Handle mds reply.
2780 *
2781 * We take the session mutex and parse and process the reply immediately.
2782 * This preserves the logical ordering of replies, capabilities, etc., sent
2783 * by the MDS as they are applied to our local cache.
2784 */
2785static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2786{
2787 struct ceph_mds_client *mdsc = session->s_mdsc;
2788 struct ceph_mds_request *req;
2789 struct ceph_mds_reply_head *head = msg->front.iov_base;
2790 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2791 struct ceph_snap_realm *realm;
2792 u64 tid;
2793 int err, result;
2794 int mds = session->s_mds;
2795
2796 if (msg->front.iov_len < sizeof(*head)) {
2797 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2798 ceph_msg_dump(msg);
2799 return;
2800 }
2801
2802 /* get request, session */
2803 tid = le64_to_cpu(msg->hdr.tid);
2804 mutex_lock(&mdsc->mutex);
2805 req = lookup_get_request(mdsc, tid);
2806 if (!req) {
2807 dout("handle_reply on unknown tid %llu\n", tid);
2808 mutex_unlock(&mdsc->mutex);
2809 return;
2810 }
2811 dout("handle_reply %p\n", req);
2812
2813 /* correct session? */
2814 if (req->r_session != session) {
2815 pr_err("mdsc_handle_reply got %llu on session mds%d"
2816 " not mds%d\n", tid, session->s_mds,
2817 req->r_session ? req->r_session->s_mds : -1);
2818 mutex_unlock(&mdsc->mutex);
2819 goto out;
2820 }
2821
2822 /* dup? */
2823 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2824 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2825 pr_warn("got a dup %s reply on %llu from mds%d\n",
2826 head->safe ? "safe" : "unsafe", tid, mds);
2827 mutex_unlock(&mdsc->mutex);
2828 goto out;
2829 }
2830 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2831 pr_warn("got unsafe after safe on %llu from mds%d\n",
2832 tid, mds);
2833 mutex_unlock(&mdsc->mutex);
2834 goto out;
2835 }
2836
2837 result = le32_to_cpu(head->result);
2838
2839 /*
2840 * Handle an ESTALE
2841 * if we're not talking to the authority, send to them
2842 * if the authority has changed while we weren't looking,
2843 * send to new authority
2844 * Otherwise we just have to return an ESTALE
2845 */
2846 if (result == -ESTALE) {
2847 dout("got ESTALE on request %llu\n", req->r_tid);
2848 req->r_resend_mds = -1;
2849 if (req->r_direct_mode != USE_AUTH_MDS) {
2850 dout("not using auth, setting for that now\n");
2851 req->r_direct_mode = USE_AUTH_MDS;
2852 __do_request(mdsc, req);
2853 mutex_unlock(&mdsc->mutex);
2854 goto out;
2855 } else {
2856 int mds = __choose_mds(mdsc, req);
2857 if (mds >= 0 && mds != req->r_session->s_mds) {
2858 dout("but auth changed, so resending\n");
2859 __do_request(mdsc, req);
2860 mutex_unlock(&mdsc->mutex);
2861 goto out;
2862 }
2863 }
2864 dout("have to return ESTALE on request %llu\n", req->r_tid);
2865 }
2866
2867
2868 if (head->safe) {
2869 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2870 __unregister_request(mdsc, req);
2871
2872 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2873 /*
2874 * We already handled the unsafe response, now do the
2875 * cleanup. No need to examine the response; the MDS
2876 * doesn't include any result info in the safe
2877 * response. And even if it did, there is nothing
2878 * useful we could do with a revised return value.
2879 */
2880 dout("got safe reply %llu, mds%d\n", tid, mds);
2881
2882 /* last unsafe request during umount? */
2883 if (mdsc->stopping && !__get_oldest_req(mdsc))
2884 complete_all(&mdsc->safe_umount_waiters);
2885 mutex_unlock(&mdsc->mutex);
2886 goto out;
2887 }
2888 } else {
2889 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2890 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2891 if (req->r_unsafe_dir) {
2892 struct ceph_inode_info *ci =
2893 ceph_inode(req->r_unsafe_dir);
2894 spin_lock(&ci->i_unsafe_lock);
2895 list_add_tail(&req->r_unsafe_dir_item,
2896 &ci->i_unsafe_dirops);
2897 spin_unlock(&ci->i_unsafe_lock);
2898 }
2899 }
2900
2901 dout("handle_reply tid %lld result %d\n", tid, result);
2902 rinfo = &req->r_reply_info;
2903 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2904 err = parse_reply_info(msg, rinfo, (u64)-1);
2905 else
2906 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2907 mutex_unlock(&mdsc->mutex);
2908
2909 mutex_lock(&session->s_mutex);
2910 if (err < 0) {
2911 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2912 ceph_msg_dump(msg);
2913 goto out_err;
2914 }
2915
2916 /* snap trace */
2917 realm = NULL;
2918 if (rinfo->snapblob_len) {
2919 down_write(&mdsc->snap_rwsem);
2920 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2921 rinfo->snapblob + rinfo->snapblob_len,
2922 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2923 &realm);
2924 downgrade_write(&mdsc->snap_rwsem);
2925 } else {
2926 down_read(&mdsc->snap_rwsem);
2927 }
2928
2929 /* insert trace into our cache */
2930 mutex_lock(&req->r_fill_mutex);
2931 current->journal_info = req;
2932 err = ceph_fill_trace(mdsc->fsc->sb, req);
2933 if (err == 0) {
2934 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2935 req->r_op == CEPH_MDS_OP_LSSNAP))
2936 ceph_readdir_prepopulate(req, req->r_session);
2937 }
2938 current->journal_info = NULL;
2939 mutex_unlock(&req->r_fill_mutex);
2940
2941 up_read(&mdsc->snap_rwsem);
2942 if (realm)
2943 ceph_put_snap_realm(mdsc, realm);
2944
2945 if (err == 0) {
2946 if (req->r_target_inode &&
2947 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2948 struct ceph_inode_info *ci =
2949 ceph_inode(req->r_target_inode);
2950 spin_lock(&ci->i_unsafe_lock);
2951 list_add_tail(&req->r_unsafe_target_item,
2952 &ci->i_unsafe_iops);
2953 spin_unlock(&ci->i_unsafe_lock);
2954 }
2955
2956 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2957 }
2958out_err:
2959 mutex_lock(&mdsc->mutex);
2960 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2961 if (err) {
2962 req->r_err = err;
2963 } else {
2964 req->r_reply = ceph_msg_get(msg);
2965 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2966 }
2967 } else {
2968 dout("reply arrived after request %lld was aborted\n", tid);
2969 }
2970 mutex_unlock(&mdsc->mutex);
2971
2972 mutex_unlock(&session->s_mutex);
2973
2974 /* kick calling process */
2975 complete_request(mdsc, req);
2976out:
2977 ceph_mdsc_put_request(req);
2978 return;
2979}
2980
2981
2982
2983/*
2984 * handle mds notification that our request has been forwarded.
2985 */
2986static void handle_forward(struct ceph_mds_client *mdsc,
2987 struct ceph_mds_session *session,
2988 struct ceph_msg *msg)
2989{
2990 struct ceph_mds_request *req;
2991 u64 tid = le64_to_cpu(msg->hdr.tid);
2992 u32 next_mds;
2993 u32 fwd_seq;
2994 int err = -EINVAL;
2995 void *p = msg->front.iov_base;
2996 void *end = p + msg->front.iov_len;
2997
2998 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2999 next_mds = ceph_decode_32(&p);
3000 fwd_seq = ceph_decode_32(&p);
3001
3002 mutex_lock(&mdsc->mutex);
3003 req = lookup_get_request(mdsc, tid);
3004 if (!req) {
3005 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3006 goto out; /* dup reply? */
3007 }
3008
3009 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3010 dout("forward tid %llu aborted, unregistering\n", tid);
3011 __unregister_request(mdsc, req);
3012 } else if (fwd_seq <= req->r_num_fwd) {
3013 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3014 tid, next_mds, req->r_num_fwd, fwd_seq);
3015 } else {
3016 /* resend. forward race not possible; mds would drop */
3017 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3018 BUG_ON(req->r_err);
3019 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3020 req->r_attempts = 0;
3021 req->r_num_fwd = fwd_seq;
3022 req->r_resend_mds = next_mds;
3023 put_request_session(req);
3024 __do_request(mdsc, req);
3025 }
3026 ceph_mdsc_put_request(req);
3027out:
3028 mutex_unlock(&mdsc->mutex);
3029 return;
3030
3031bad:
3032 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3033}
3034
3035static int __decode_session_metadata(void **p, void *end,
3036 bool *blacklisted)
3037{
3038 /* map<string,string> */
3039 u32 n;
3040 bool err_str;
3041 ceph_decode_32_safe(p, end, n, bad);
3042 while (n-- > 0) {
3043 u32 len;
3044 ceph_decode_32_safe(p, end, len, bad);
3045 ceph_decode_need(p, end, len, bad);
3046 err_str = !strncmp(*p, "error_string", len);
3047 *p += len;
3048 ceph_decode_32_safe(p, end, len, bad);
3049 ceph_decode_need(p, end, len, bad);
3050 if (err_str && strnstr(*p, "blacklisted", len))
3051 *blacklisted = true;
3052 *p += len;
3053 }
3054 return 0;
3055bad:
3056 return -1;
3057}
3058
3059/*
3060 * handle a mds session control message
3061 */
3062static void handle_session(struct ceph_mds_session *session,
3063 struct ceph_msg *msg)
3064{
3065 struct ceph_mds_client *mdsc = session->s_mdsc;
3066 int mds = session->s_mds;
3067 int msg_version = le16_to_cpu(msg->hdr.version);
3068 void *p = msg->front.iov_base;
3069 void *end = p + msg->front.iov_len;
3070 struct ceph_mds_session_head *h;
3071 u32 op;
3072 u64 seq;
3073 unsigned long features = 0;
3074 int wake = 0;
3075 bool blacklisted = false;
3076
3077 /* decode */
3078 ceph_decode_need(&p, end, sizeof(*h), bad);
3079 h = p;
3080 p += sizeof(*h);
3081
3082 op = le32_to_cpu(h->op);
3083 seq = le64_to_cpu(h->seq);
3084
3085 if (msg_version >= 3) {
3086 u32 len;
3087 /* version >= 2, metadata */
3088 if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3089 goto bad;
3090 /* version >= 3, feature bits */
3091 ceph_decode_32_safe(&p, end, len, bad);
3092 ceph_decode_need(&p, end, len, bad);
3093 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3094 p += len;
3095 }
3096
3097 mutex_lock(&mdsc->mutex);
3098 if (op == CEPH_SESSION_CLOSE) {
3099 get_session(session);
3100 __unregister_session(mdsc, session);
3101 }
3102 /* FIXME: this ttl calculation is generous */
3103 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3104 mutex_unlock(&mdsc->mutex);
3105
3106 mutex_lock(&session->s_mutex);
3107
3108 dout("handle_session mds%d %s %p state %s seq %llu\n",
3109 mds, ceph_session_op_name(op), session,
3110 ceph_session_state_name(session->s_state), seq);
3111
3112 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3113 session->s_state = CEPH_MDS_SESSION_OPEN;
3114 pr_info("mds%d came back\n", session->s_mds);
3115 }
3116
3117 switch (op) {
3118 case CEPH_SESSION_OPEN:
3119 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3120 pr_info("mds%d reconnect success\n", session->s_mds);
3121 session->s_state = CEPH_MDS_SESSION_OPEN;
3122 session->s_features = features;
3123 renewed_caps(mdsc, session, 0);
3124 wake = 1;
3125 if (mdsc->stopping)
3126 __close_session(mdsc, session);
3127 break;
3128
3129 case CEPH_SESSION_RENEWCAPS:
3130 if (session->s_renew_seq == seq)
3131 renewed_caps(mdsc, session, 1);
3132 break;
3133
3134 case CEPH_SESSION_CLOSE:
3135 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3136 pr_info("mds%d reconnect denied\n", session->s_mds);
3137 cleanup_session_requests(mdsc, session);
3138 remove_session_caps(session);
3139 wake = 2; /* for good measure */
3140 wake_up_all(&mdsc->session_close_wq);
3141 break;
3142
3143 case CEPH_SESSION_STALE:
3144 pr_info("mds%d caps went stale, renewing\n",
3145 session->s_mds);
3146 spin_lock(&session->s_gen_ttl_lock);
3147 session->s_cap_gen++;
3148 session->s_cap_ttl = jiffies - 1;
3149 spin_unlock(&session->s_gen_ttl_lock);
3150 send_renew_caps(mdsc, session);
3151 break;
3152
3153 case CEPH_SESSION_RECALL_STATE:
3154 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3155 break;
3156
3157 case CEPH_SESSION_FLUSHMSG:
3158 send_flushmsg_ack(mdsc, session, seq);
3159 break;
3160
3161 case CEPH_SESSION_FORCE_RO:
3162 dout("force_session_readonly %p\n", session);
3163 spin_lock(&session->s_cap_lock);
3164 session->s_readonly = true;
3165 spin_unlock(&session->s_cap_lock);
3166 wake_up_session_caps(session, FORCE_RO);
3167 break;
3168
3169 case CEPH_SESSION_REJECT:
3170 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3171 pr_info("mds%d rejected session\n", session->s_mds);
3172 session->s_state = CEPH_MDS_SESSION_REJECTED;
3173 cleanup_session_requests(mdsc, session);
3174 remove_session_caps(session);
3175 if (blacklisted)
3176 mdsc->fsc->blacklisted = true;
3177 wake = 2; /* for good measure */
3178 break;
3179
3180 default:
3181 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3182 WARN_ON(1);
3183 }
3184
3185 mutex_unlock(&session->s_mutex);
3186 if (wake) {
3187 mutex_lock(&mdsc->mutex);
3188 __wake_requests(mdsc, &session->s_waiting);
3189 if (wake == 2)
3190 kick_requests(mdsc, mds);
3191 mutex_unlock(&mdsc->mutex);
3192 }
3193 if (op == CEPH_SESSION_CLOSE)
3194 ceph_put_mds_session(session);
3195 return;
3196
3197bad:
3198 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3199 (int)msg->front.iov_len);
3200 ceph_msg_dump(msg);
3201 return;
3202}
3203
3204
3205/*
3206 * called under session->mutex.
3207 */
3208static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3209 struct ceph_mds_session *session)
3210{
3211 struct ceph_mds_request *req, *nreq;
3212 struct rb_node *p;
3213 int err;
3214
3215 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3216
3217 mutex_lock(&mdsc->mutex);
3218 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
3219 err = __prepare_send_request(mdsc, req, session->s_mds, true);
3220 if (!err) {
3221 ceph_msg_get(req->r_request);
3222 ceph_con_send(&session->s_con, req->r_request);
3223 }
3224 }
3225
3226 /*
3227 * also re-send old requests when MDS enters reconnect stage. So that MDS
3228 * can process completed request in clientreplay stage.
3229 */
3230 p = rb_first(&mdsc->request_tree);
3231 while (p) {
3232 req = rb_entry(p, struct ceph_mds_request, r_node);
3233 p = rb_next(p);
3234 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3235 continue;
3236 if (req->r_attempts == 0)
3237 continue; /* only old requests */
3238 if (req->r_session &&
3239 req->r_session->s_mds == session->s_mds) {
3240 err = __prepare_send_request(mdsc, req,
3241 session->s_mds, true);
3242 if (!err) {
3243 ceph_msg_get(req->r_request);
3244 ceph_con_send(&session->s_con, req->r_request);
3245 }
3246 }
3247 }
3248 mutex_unlock(&mdsc->mutex);
3249}
3250
3251static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3252{
3253 struct ceph_msg *reply;
3254 struct ceph_pagelist *_pagelist;
3255 struct page *page;
3256 __le32 *addr;
3257 int err = -ENOMEM;
3258
3259 if (!recon_state->allow_multi)
3260 return -ENOSPC;
3261
3262 /* can't handle message that contains both caps and realm */
3263 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3264
3265 /* pre-allocate new pagelist */
3266 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3267 if (!_pagelist)
3268 return -ENOMEM;
3269
3270 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3271 if (!reply)
3272 goto fail_msg;
3273
3274 /* placeholder for nr_caps */
3275 err = ceph_pagelist_encode_32(_pagelist, 0);
3276 if (err < 0)
3277 goto fail;
3278
3279 if (recon_state->nr_caps) {
3280 /* currently encoding caps */
3281 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3282 if (err)
3283 goto fail;
3284 } else {
3285 /* placeholder for nr_realms (currently encoding relams) */
3286 err = ceph_pagelist_encode_32(_pagelist, 0);
3287 if (err < 0)
3288 goto fail;
3289 }
3290
3291 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3292 if (err)
3293 goto fail;
3294
3295 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3296 addr = kmap_atomic(page);
3297 if (recon_state->nr_caps) {
3298 /* currently encoding caps */
3299 *addr = cpu_to_le32(recon_state->nr_caps);
3300 } else {
3301 /* currently encoding relams */
3302 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3303 }
3304 kunmap_atomic(addr);
3305
3306 reply->hdr.version = cpu_to_le16(5);
3307 reply->hdr.compat_version = cpu_to_le16(4);
3308
3309 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3310 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3311
3312 ceph_con_send(&recon_state->session->s_con, reply);
3313 ceph_pagelist_release(recon_state->pagelist);
3314
3315 recon_state->pagelist = _pagelist;
3316 recon_state->nr_caps = 0;
3317 recon_state->nr_realms = 0;
3318 recon_state->msg_version = 5;
3319 return 0;
3320fail:
3321 ceph_msg_put(reply);
3322fail_msg:
3323 ceph_pagelist_release(_pagelist);
3324 return err;
3325}
3326
3327/*
3328 * Encode information about a cap for a reconnect with the MDS.
3329 */
3330static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3331 void *arg)
3332{
3333 union {
3334 struct ceph_mds_cap_reconnect v2;
3335 struct ceph_mds_cap_reconnect_v1 v1;
3336 } rec;
3337 struct ceph_inode_info *ci = cap->ci;
3338 struct ceph_reconnect_state *recon_state = arg;
3339 struct ceph_pagelist *pagelist = recon_state->pagelist;
3340 int err;
3341 u64 snap_follows;
3342
3343 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3344 inode, ceph_vinop(inode), cap, cap->cap_id,
3345 ceph_cap_string(cap->issued));
3346
3347 spin_lock(&ci->i_ceph_lock);
3348 cap->seq = 0; /* reset cap seq */
3349 cap->issue_seq = 0; /* and issue_seq */
3350 cap->mseq = 0; /* and migrate_seq */
3351 cap->cap_gen = cap->session->s_cap_gen;
3352
3353 if (recon_state->msg_version >= 2) {
3354 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3355 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3356 rec.v2.issued = cpu_to_le32(cap->issued);
3357 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3358 rec.v2.pathbase = 0;
3359 rec.v2.flock_len = (__force __le32)
3360 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3361 } else {
3362 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3363 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3364 rec.v1.issued = cpu_to_le32(cap->issued);
3365 rec.v1.size = cpu_to_le64(inode->i_size);
3366 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3367 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3368 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3369 rec.v1.pathbase = 0;
3370 }
3371
3372 if (list_empty(&ci->i_cap_snaps)) {
3373 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3374 } else {
3375 struct ceph_cap_snap *capsnap =
3376 list_first_entry(&ci->i_cap_snaps,
3377 struct ceph_cap_snap, ci_item);
3378 snap_follows = capsnap->follows;
3379 }
3380 spin_unlock(&ci->i_ceph_lock);
3381
3382 if (recon_state->msg_version >= 2) {
3383 int num_fcntl_locks, num_flock_locks;
3384 struct ceph_filelock *flocks = NULL;
3385 size_t struct_len, total_len = sizeof(u64);
3386 u8 struct_v = 0;
3387
3388encode_again:
3389 if (rec.v2.flock_len) {
3390 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3391 } else {
3392 num_fcntl_locks = 0;
3393 num_flock_locks = 0;
3394 }
3395 if (num_fcntl_locks + num_flock_locks > 0) {
3396 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3397 sizeof(struct ceph_filelock),
3398 GFP_NOFS);
3399 if (!flocks) {
3400 err = -ENOMEM;
3401 goto out_err;
3402 }
3403 err = ceph_encode_locks_to_buffer(inode, flocks,
3404 num_fcntl_locks,
3405 num_flock_locks);
3406 if (err) {
3407 kfree(flocks);
3408 flocks = NULL;
3409 if (err == -ENOSPC)
3410 goto encode_again;
3411 goto out_err;
3412 }
3413 } else {
3414 kfree(flocks);
3415 flocks = NULL;
3416 }
3417
3418 if (recon_state->msg_version >= 3) {
3419 /* version, compat_version and struct_len */
3420 total_len += 2 * sizeof(u8) + sizeof(u32);
3421 struct_v = 2;
3422 }
3423 /*
3424 * number of encoded locks is stable, so copy to pagelist
3425 */
3426 struct_len = 2 * sizeof(u32) +
3427 (num_fcntl_locks + num_flock_locks) *
3428 sizeof(struct ceph_filelock);
3429 rec.v2.flock_len = cpu_to_le32(struct_len);
3430
3431 struct_len += sizeof(u32) + sizeof(rec.v2);
3432
3433 if (struct_v >= 2)
3434 struct_len += sizeof(u64); /* snap_follows */
3435
3436 total_len += struct_len;
3437
3438 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3439 err = send_reconnect_partial(recon_state);
3440 if (err)
3441 goto out_freeflocks;
3442 pagelist = recon_state->pagelist;
3443 }
3444
3445 err = ceph_pagelist_reserve(pagelist, total_len);
3446 if (err)
3447 goto out_freeflocks;
3448
3449 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3450 if (recon_state->msg_version >= 3) {
3451 ceph_pagelist_encode_8(pagelist, struct_v);
3452 ceph_pagelist_encode_8(pagelist, 1);
3453 ceph_pagelist_encode_32(pagelist, struct_len);
3454 }
3455 ceph_pagelist_encode_string(pagelist, NULL, 0);
3456 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3457 ceph_locks_to_pagelist(flocks, pagelist,
3458 num_fcntl_locks, num_flock_locks);
3459 if (struct_v >= 2)
3460 ceph_pagelist_encode_64(pagelist, snap_follows);
3461out_freeflocks:
3462 kfree(flocks);
3463 } else {
3464 u64 pathbase = 0;
3465 int pathlen = 0;
3466 char *path = NULL;
3467 struct dentry *dentry;
3468
3469 dentry = d_find_alias(inode);
3470 if (dentry) {
3471 path = ceph_mdsc_build_path(dentry,
3472 &pathlen, &pathbase, 0);
3473 dput(dentry);
3474 if (IS_ERR(path)) {
3475 err = PTR_ERR(path);
3476 goto out_err;
3477 }
3478 rec.v1.pathbase = cpu_to_le64(pathbase);
3479 }
3480
3481 err = ceph_pagelist_reserve(pagelist,
3482 sizeof(u64) + sizeof(u32) +
3483 pathlen + sizeof(rec.v1));
3484 if (err) {
3485 goto out_freepath;
3486 }
3487
3488 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3489 ceph_pagelist_encode_string(pagelist, path, pathlen);
3490 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3491out_freepath:
3492 ceph_mdsc_free_path(path, pathlen);
3493 }
3494
3495out_err:
3496 if (err >= 0)
3497 recon_state->nr_caps++;
3498 return err;
3499}
3500
3501static int encode_snap_realms(struct ceph_mds_client *mdsc,
3502 struct ceph_reconnect_state *recon_state)
3503{
3504 struct rb_node *p;
3505 struct ceph_pagelist *pagelist = recon_state->pagelist;
3506 int err = 0;
3507
3508 if (recon_state->msg_version >= 4) {
3509 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3510 if (err < 0)
3511 goto fail;
3512 }
3513
3514 /*
3515 * snaprealms. we provide mds with the ino, seq (version), and
3516 * parent for all of our realms. If the mds has any newer info,
3517 * it will tell us.
3518 */
3519 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3520 struct ceph_snap_realm *realm =
3521 rb_entry(p, struct ceph_snap_realm, node);
3522 struct ceph_mds_snaprealm_reconnect sr_rec;
3523
3524 if (recon_state->msg_version >= 4) {
3525 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3526 sizeof(sr_rec);
3527
3528 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3529 err = send_reconnect_partial(recon_state);
3530 if (err)
3531 goto fail;
3532 pagelist = recon_state->pagelist;
3533 }
3534
3535 err = ceph_pagelist_reserve(pagelist, need);
3536 if (err)
3537 goto fail;
3538
3539 ceph_pagelist_encode_8(pagelist, 1);
3540 ceph_pagelist_encode_8(pagelist, 1);
3541 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3542 }
3543
3544 dout(" adding snap realm %llx seq %lld parent %llx\n",
3545 realm->ino, realm->seq, realm->parent_ino);
3546 sr_rec.ino = cpu_to_le64(realm->ino);
3547 sr_rec.seq = cpu_to_le64(realm->seq);
3548 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3549
3550 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3551 if (err)
3552 goto fail;
3553
3554 recon_state->nr_realms++;
3555 }
3556fail:
3557 return err;
3558}
3559
3560
3561/*
3562 * If an MDS fails and recovers, clients need to reconnect in order to
3563 * reestablish shared state. This includes all caps issued through
3564 * this session _and_ the snap_realm hierarchy. Because it's not
3565 * clear which snap realms the mds cares about, we send everything we
3566 * know about.. that ensures we'll then get any new info the
3567 * recovering MDS might have.
3568 *
3569 * This is a relatively heavyweight operation, but it's rare.
3570 *
3571 * called with mdsc->mutex held.
3572 */
3573static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3574 struct ceph_mds_session *session)
3575{
3576 struct ceph_msg *reply;
3577 int mds = session->s_mds;
3578 int err = -ENOMEM;
3579 struct ceph_reconnect_state recon_state = {
3580 .session = session,
3581 };
3582 LIST_HEAD(dispose);
3583
3584 pr_info("mds%d reconnect start\n", mds);
3585
3586 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3587 if (!recon_state.pagelist)
3588 goto fail_nopagelist;
3589
3590 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3591 if (!reply)
3592 goto fail_nomsg;
3593
3594 mutex_lock(&session->s_mutex);
3595 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3596 session->s_seq = 0;
3597
3598 dout("session %p state %s\n", session,
3599 ceph_session_state_name(session->s_state));
3600
3601 spin_lock(&session->s_gen_ttl_lock);
3602 session->s_cap_gen++;
3603 spin_unlock(&session->s_gen_ttl_lock);
3604
3605 spin_lock(&session->s_cap_lock);
3606 /* don't know if session is readonly */
3607 session->s_readonly = 0;
3608 /*
3609 * notify __ceph_remove_cap() that we are composing cap reconnect.
3610 * If a cap get released before being added to the cap reconnect,
3611 * __ceph_remove_cap() should skip queuing cap release.
3612 */
3613 session->s_cap_reconnect = 1;
3614 /* drop old cap expires; we're about to reestablish that state */
3615 detach_cap_releases(session, &dispose);
3616 spin_unlock(&session->s_cap_lock);
3617 dispose_cap_releases(mdsc, &dispose);
3618
3619 /* trim unused caps to reduce MDS's cache rejoin time */
3620 if (mdsc->fsc->sb->s_root)
3621 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3622
3623 ceph_con_close(&session->s_con);
3624 ceph_con_open(&session->s_con,
3625 CEPH_ENTITY_TYPE_MDS, mds,
3626 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3627
3628 /* replay unsafe requests */
3629 replay_unsafe_requests(mdsc, session);
3630
3631 ceph_early_kick_flushing_caps(mdsc, session);
3632
3633 down_read(&mdsc->snap_rwsem);
3634
3635 /* placeholder for nr_caps */
3636 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3637 if (err)
3638 goto fail;
3639
3640 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3641 recon_state.msg_version = 3;
3642 recon_state.allow_multi = true;
3643 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3644 recon_state.msg_version = 3;
3645 } else {
3646 recon_state.msg_version = 2;
3647 }
3648 /* trsaverse this session's caps */
3649 err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3650
3651 spin_lock(&session->s_cap_lock);
3652 session->s_cap_reconnect = 0;
3653 spin_unlock(&session->s_cap_lock);
3654
3655 if (err < 0)
3656 goto fail;
3657
3658 /* check if all realms can be encoded into current message */
3659 if (mdsc->num_snap_realms) {
3660 size_t total_len =
3661 recon_state.pagelist->length +
3662 mdsc->num_snap_realms *
3663 sizeof(struct ceph_mds_snaprealm_reconnect);
3664 if (recon_state.msg_version >= 4) {
3665 /* number of realms */
3666 total_len += sizeof(u32);
3667 /* version, compat_version and struct_len */
3668 total_len += mdsc->num_snap_realms *
3669 (2 * sizeof(u8) + sizeof(u32));
3670 }
3671 if (total_len > RECONNECT_MAX_SIZE) {
3672 if (!recon_state.allow_multi) {
3673 err = -ENOSPC;
3674 goto fail;
3675 }
3676 if (recon_state.nr_caps) {
3677 err = send_reconnect_partial(&recon_state);
3678 if (err)
3679 goto fail;
3680 }
3681 recon_state.msg_version = 5;
3682 }
3683 }
3684
3685 err = encode_snap_realms(mdsc, &recon_state);
3686 if (err < 0)
3687 goto fail;
3688
3689 if (recon_state.msg_version >= 5) {
3690 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3691 if (err < 0)
3692 goto fail;
3693 }
3694
3695 if (recon_state.nr_caps || recon_state.nr_realms) {
3696 struct page *page =
3697 list_first_entry(&recon_state.pagelist->head,
3698 struct page, lru);
3699 __le32 *addr = kmap_atomic(page);
3700 if (recon_state.nr_caps) {
3701 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3702 *addr = cpu_to_le32(recon_state.nr_caps);
3703 } else if (recon_state.msg_version >= 4) {
3704 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3705 }
3706 kunmap_atomic(addr);
3707 }
3708
3709 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3710 if (recon_state.msg_version >= 4)
3711 reply->hdr.compat_version = cpu_to_le16(4);
3712
3713 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3714 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3715
3716 ceph_con_send(&session->s_con, reply);
3717
3718 mutex_unlock(&session->s_mutex);
3719
3720 mutex_lock(&mdsc->mutex);
3721 __wake_requests(mdsc, &session->s_waiting);
3722 mutex_unlock(&mdsc->mutex);
3723
3724 up_read(&mdsc->snap_rwsem);
3725 ceph_pagelist_release(recon_state.pagelist);
3726 return;
3727
3728fail:
3729 ceph_msg_put(reply);
3730 up_read(&mdsc->snap_rwsem);
3731 mutex_unlock(&session->s_mutex);
3732fail_nomsg:
3733 ceph_pagelist_release(recon_state.pagelist);
3734fail_nopagelist:
3735 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3736 return;
3737}
3738
3739
3740/*
3741 * compare old and new mdsmaps, kicking requests
3742 * and closing out old connections as necessary
3743 *
3744 * called under mdsc->mutex.
3745 */
3746static void check_new_map(struct ceph_mds_client *mdsc,
3747 struct ceph_mdsmap *newmap,
3748 struct ceph_mdsmap *oldmap)
3749{
3750 int i;
3751 int oldstate, newstate;
3752 struct ceph_mds_session *s;
3753
3754 dout("check_new_map new %u old %u\n",
3755 newmap->m_epoch, oldmap->m_epoch);
3756
3757 for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3758 if (!mdsc->sessions[i])
3759 continue;
3760 s = mdsc->sessions[i];
3761 oldstate = ceph_mdsmap_get_state(oldmap, i);
3762 newstate = ceph_mdsmap_get_state(newmap, i);
3763
3764 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3765 i, ceph_mds_state_name(oldstate),
3766 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3767 ceph_mds_state_name(newstate),
3768 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3769 ceph_session_state_name(s->s_state));
3770
3771 if (i >= newmap->m_num_mds) {
3772 /* force close session for stopped mds */
3773 get_session(s);
3774 __unregister_session(mdsc, s);
3775 __wake_requests(mdsc, &s->s_waiting);
3776 mutex_unlock(&mdsc->mutex);
3777
3778 mutex_lock(&s->s_mutex);
3779 cleanup_session_requests(mdsc, s);
3780 remove_session_caps(s);
3781 mutex_unlock(&s->s_mutex);
3782
3783 ceph_put_mds_session(s);
3784
3785 mutex_lock(&mdsc->mutex);
3786 kick_requests(mdsc, i);
3787 continue;
3788 }
3789
3790 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
3791 ceph_mdsmap_get_addr(newmap, i),
3792 sizeof(struct ceph_entity_addr))) {
3793 /* just close it */
3794 mutex_unlock(&mdsc->mutex);
3795 mutex_lock(&s->s_mutex);
3796 mutex_lock(&mdsc->mutex);
3797 ceph_con_close(&s->s_con);
3798 mutex_unlock(&s->s_mutex);
3799 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3800 } else if (oldstate == newstate) {
3801 continue; /* nothing new with this mds */
3802 }
3803
3804 /*
3805 * send reconnect?
3806 */
3807 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3808 newstate >= CEPH_MDS_STATE_RECONNECT) {
3809 mutex_unlock(&mdsc->mutex);
3810 send_mds_reconnect(mdsc, s);
3811 mutex_lock(&mdsc->mutex);
3812 }
3813
3814 /*
3815 * kick request on any mds that has gone active.
3816 */
3817 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3818 newstate >= CEPH_MDS_STATE_ACTIVE) {
3819 if (oldstate != CEPH_MDS_STATE_CREATING &&
3820 oldstate != CEPH_MDS_STATE_STARTING)
3821 pr_info("mds%d recovery completed\n", s->s_mds);
3822 kick_requests(mdsc, i);
3823 ceph_kick_flushing_caps(mdsc, s);
3824 wake_up_session_caps(s, RECONNECT);
3825 }
3826 }
3827
3828 for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3829 s = mdsc->sessions[i];
3830 if (!s)
3831 continue;
3832 if (!ceph_mdsmap_is_laggy(newmap, i))
3833 continue;
3834 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3835 s->s_state == CEPH_MDS_SESSION_HUNG ||
3836 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3837 dout(" connecting to export targets of laggy mds%d\n",
3838 i);
3839 __open_export_target_sessions(mdsc, s);
3840 }
3841 }
3842}
3843
3844
3845
3846/*
3847 * leases
3848 */
3849
3850/*
3851 * caller must hold session s_mutex, dentry->d_lock
3852 */
3853void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3854{
3855 struct ceph_dentry_info *di = ceph_dentry(dentry);
3856
3857 ceph_put_mds_session(di->lease_session);
3858 di->lease_session = NULL;
3859}
3860
3861static void handle_lease(struct ceph_mds_client *mdsc,
3862 struct ceph_mds_session *session,
3863 struct ceph_msg *msg)
3864{
3865 struct super_block *sb = mdsc->fsc->sb;
3866 struct inode *inode;
3867 struct dentry *parent, *dentry;
3868 struct ceph_dentry_info *di;
3869 int mds = session->s_mds;
3870 struct ceph_mds_lease *h = msg->front.iov_base;
3871 u32 seq;
3872 struct ceph_vino vino;
3873 struct qstr dname;
3874 int release = 0;
3875
3876 dout("handle_lease from mds%d\n", mds);
3877
3878 /* decode */
3879 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3880 goto bad;
3881 vino.ino = le64_to_cpu(h->ino);
3882 vino.snap = CEPH_NOSNAP;
3883 seq = le32_to_cpu(h->seq);
3884 dname.len = get_unaligned_le32(h + 1);
3885 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3886 goto bad;
3887 dname.name = (void *)(h + 1) + sizeof(u32);
3888
3889 /* lookup inode */
3890 inode = ceph_find_inode(sb, vino);
3891 dout("handle_lease %s, ino %llx %p %.*s\n",
3892 ceph_lease_op_name(h->action), vino.ino, inode,
3893 dname.len, dname.name);
3894
3895 mutex_lock(&session->s_mutex);
3896 session->s_seq++;
3897
3898 if (!inode) {
3899 dout("handle_lease no inode %llx\n", vino.ino);
3900 goto release;
3901 }
3902
3903 /* dentry */
3904 parent = d_find_alias(inode);
3905 if (!parent) {
3906 dout("no parent dentry on inode %p\n", inode);
3907 WARN_ON(1);
3908 goto release; /* hrm... */
3909 }
3910 dname.hash = full_name_hash(parent, dname.name, dname.len);
3911 dentry = d_lookup(parent, &dname);
3912 dput(parent);
3913 if (!dentry)
3914 goto release;
3915
3916 spin_lock(&dentry->d_lock);
3917 di = ceph_dentry(dentry);
3918 switch (h->action) {
3919 case CEPH_MDS_LEASE_REVOKE:
3920 if (di->lease_session == session) {
3921 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3922 h->seq = cpu_to_le32(di->lease_seq);
3923 __ceph_mdsc_drop_dentry_lease(dentry);
3924 }
3925 release = 1;
3926 break;
3927
3928 case CEPH_MDS_LEASE_RENEW:
3929 if (di->lease_session == session &&
3930 di->lease_gen == session->s_cap_gen &&
3931 di->lease_renew_from &&
3932 di->lease_renew_after == 0) {
3933 unsigned long duration =
3934 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3935
3936 di->lease_seq = seq;
3937 di->time = di->lease_renew_from + duration;
3938 di->lease_renew_after = di->lease_renew_from +
3939 (duration >> 1);
3940 di->lease_renew_from = 0;
3941 }
3942 break;
3943 }
3944 spin_unlock(&dentry->d_lock);
3945 dput(dentry);
3946
3947 if (!release)
3948 goto out;
3949
3950release:
3951 /* let's just reuse the same message */
3952 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3953 ceph_msg_get(msg);
3954 ceph_con_send(&session->s_con, msg);
3955
3956out:
3957 mutex_unlock(&session->s_mutex);
3958 /* avoid calling iput_final() in mds dispatch threads */
3959 ceph_async_iput(inode);
3960 return;
3961
3962bad:
3963 pr_err("corrupt lease message\n");
3964 ceph_msg_dump(msg);
3965}
3966
3967void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3968 struct dentry *dentry, char action,
3969 u32 seq)
3970{
3971 struct ceph_msg *msg;
3972 struct ceph_mds_lease *lease;
3973 struct inode *dir;
3974 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
3975
3976 dout("lease_send_msg identry %p %s to mds%d\n",
3977 dentry, ceph_lease_op_name(action), session->s_mds);
3978
3979 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3980 if (!msg)
3981 return;
3982 lease = msg->front.iov_base;
3983 lease->action = action;
3984 lease->seq = cpu_to_le32(seq);
3985
3986 spin_lock(&dentry->d_lock);
3987 dir = d_inode(dentry->d_parent);
3988 lease->ino = cpu_to_le64(ceph_ino(dir));
3989 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
3990
3991 put_unaligned_le32(dentry->d_name.len, lease + 1);
3992 memcpy((void *)(lease + 1) + 4,
3993 dentry->d_name.name, dentry->d_name.len);
3994 spin_unlock(&dentry->d_lock);
3995 /*
3996 * if this is a preemptive lease RELEASE, no need to
3997 * flush request stream, since the actual request will
3998 * soon follow.
3999 */
4000 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4001
4002 ceph_con_send(&session->s_con, msg);
4003}
4004
4005/*
4006 * lock unlock sessions, to wait ongoing session activities
4007 */
4008static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4009{
4010 int i;
4011
4012 mutex_lock(&mdsc->mutex);
4013 for (i = 0; i < mdsc->max_sessions; i++) {
4014 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4015 if (!s)
4016 continue;
4017 mutex_unlock(&mdsc->mutex);
4018 mutex_lock(&s->s_mutex);
4019 mutex_unlock(&s->s_mutex);
4020 ceph_put_mds_session(s);
4021 mutex_lock(&mdsc->mutex);
4022 }
4023 mutex_unlock(&mdsc->mutex);
4024}
4025
4026static void maybe_recover_session(struct ceph_mds_client *mdsc)
4027{
4028 struct ceph_fs_client *fsc = mdsc->fsc;
4029
4030 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4031 return;
4032
4033 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4034 return;
4035
4036 if (!READ_ONCE(fsc->blacklisted))
4037 return;
4038
4039 if (fsc->last_auto_reconnect &&
4040 time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4041 return;
4042
4043 pr_info("auto reconnect after blacklisted\n");
4044 fsc->last_auto_reconnect = jiffies;
4045 ceph_force_reconnect(fsc->sb);
4046}
4047
4048/*
4049 * delayed work -- periodically trim expired leases, renew caps with mds
4050 */
4051static void schedule_delayed(struct ceph_mds_client *mdsc)
4052{
4053 int delay = 5;
4054 unsigned hz = round_jiffies_relative(HZ * delay);
4055 schedule_delayed_work(&mdsc->delayed_work, hz);
4056}
4057
4058static void delayed_work(struct work_struct *work)
4059{
4060 int i;
4061 struct ceph_mds_client *mdsc =
4062 container_of(work, struct ceph_mds_client, delayed_work.work);
4063 int renew_interval;
4064 int renew_caps;
4065
4066 dout("mdsc delayed_work\n");
4067
4068 mutex_lock(&mdsc->mutex);
4069 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4070 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4071 mdsc->last_renew_caps);
4072 if (renew_caps)
4073 mdsc->last_renew_caps = jiffies;
4074
4075 for (i = 0; i < mdsc->max_sessions; i++) {
4076 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4077 if (!s)
4078 continue;
4079 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4080 dout("resending session close request for mds%d\n",
4081 s->s_mds);
4082 request_close_session(mdsc, s);
4083 ceph_put_mds_session(s);
4084 continue;
4085 }
4086 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4087 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4088 s->s_state = CEPH_MDS_SESSION_HUNG;
4089 pr_info("mds%d hung\n", s->s_mds);
4090 }
4091 }
4092 if (s->s_state == CEPH_MDS_SESSION_NEW ||
4093 s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4094 s->s_state == CEPH_MDS_SESSION_REJECTED) {
4095 /* this mds is failed or recovering, just wait */
4096 ceph_put_mds_session(s);
4097 continue;
4098 }
4099 mutex_unlock(&mdsc->mutex);
4100
4101 mutex_lock(&s->s_mutex);
4102 if (renew_caps)
4103 send_renew_caps(mdsc, s);
4104 else
4105 ceph_con_keepalive(&s->s_con);
4106 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4107 s->s_state == CEPH_MDS_SESSION_HUNG)
4108 ceph_send_cap_releases(mdsc, s);
4109 mutex_unlock(&s->s_mutex);
4110 ceph_put_mds_session(s);
4111
4112 mutex_lock(&mdsc->mutex);
4113 }
4114 mutex_unlock(&mdsc->mutex);
4115
4116 ceph_check_delayed_caps(mdsc);
4117
4118 ceph_queue_cap_reclaim_work(mdsc);
4119
4120 ceph_trim_snapid_map(mdsc);
4121
4122 maybe_recover_session(mdsc);
4123
4124 schedule_delayed(mdsc);
4125}
4126
4127int ceph_mdsc_init(struct ceph_fs_client *fsc)
4128
4129{
4130 struct ceph_mds_client *mdsc;
4131
4132 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4133 if (!mdsc)
4134 return -ENOMEM;
4135 mdsc->fsc = fsc;
4136 mutex_init(&mdsc->mutex);
4137 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4138 if (!mdsc->mdsmap) {
4139 kfree(mdsc);
4140 return -ENOMEM;
4141 }
4142
4143 fsc->mdsc = mdsc;
4144 init_completion(&mdsc->safe_umount_waiters);
4145 init_waitqueue_head(&mdsc->session_close_wq);
4146 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4147 mdsc->sessions = NULL;
4148 atomic_set(&mdsc->num_sessions, 0);
4149 mdsc->max_sessions = 0;
4150 mdsc->stopping = 0;
4151 atomic64_set(&mdsc->quotarealms_count, 0);
4152 mdsc->quotarealms_inodes = RB_ROOT;
4153 mutex_init(&mdsc->quotarealms_inodes_mutex);
4154 mdsc->last_snap_seq = 0;
4155 init_rwsem(&mdsc->snap_rwsem);
4156 mdsc->snap_realms = RB_ROOT;
4157 INIT_LIST_HEAD(&mdsc->snap_empty);
4158 mdsc->num_snap_realms = 0;
4159 spin_lock_init(&mdsc->snap_empty_lock);
4160 mdsc->last_tid = 0;
4161 mdsc->oldest_tid = 0;
4162 mdsc->request_tree = RB_ROOT;
4163 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4164 mdsc->last_renew_caps = jiffies;
4165 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4166 spin_lock_init(&mdsc->cap_delay_lock);
4167 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4168 spin_lock_init(&mdsc->snap_flush_lock);
4169 mdsc->last_cap_flush_tid = 1;
4170 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4171 INIT_LIST_HEAD(&mdsc->cap_dirty);
4172 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4173 mdsc->num_cap_flushing = 0;
4174 spin_lock_init(&mdsc->cap_dirty_lock);
4175 init_waitqueue_head(&mdsc->cap_flushing_wq);
4176 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4177 atomic_set(&mdsc->cap_reclaim_pending, 0);
4178
4179 spin_lock_init(&mdsc->dentry_list_lock);
4180 INIT_LIST_HEAD(&mdsc->dentry_leases);
4181 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4182
4183 ceph_caps_init(mdsc);
4184 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4185
4186 spin_lock_init(&mdsc->snapid_map_lock);
4187 mdsc->snapid_map_tree = RB_ROOT;
4188 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4189
4190 init_rwsem(&mdsc->pool_perm_rwsem);
4191 mdsc->pool_perm_tree = RB_ROOT;
4192
4193 strscpy(mdsc->nodename, utsname()->nodename,
4194 sizeof(mdsc->nodename));
4195 return 0;
4196}
4197
4198/*
4199 * Wait for safe replies on open mds requests. If we time out, drop
4200 * all requests from the tree to avoid dangling dentry refs.
4201 */
4202static void wait_requests(struct ceph_mds_client *mdsc)
4203{
4204 struct ceph_options *opts = mdsc->fsc->client->options;
4205 struct ceph_mds_request *req;
4206
4207 mutex_lock(&mdsc->mutex);
4208 if (__get_oldest_req(mdsc)) {
4209 mutex_unlock(&mdsc->mutex);
4210
4211 dout("wait_requests waiting for requests\n");
4212 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4213 ceph_timeout_jiffies(opts->mount_timeout));
4214
4215 /* tear down remaining requests */
4216 mutex_lock(&mdsc->mutex);
4217 while ((req = __get_oldest_req(mdsc))) {
4218 dout("wait_requests timed out on tid %llu\n",
4219 req->r_tid);
4220 list_del_init(&req->r_wait);
4221 __unregister_request(mdsc, req);
4222 }
4223 }
4224 mutex_unlock(&mdsc->mutex);
4225 dout("wait_requests done\n");
4226}
4227
4228/*
4229 * called before mount is ro, and before dentries are torn down.
4230 * (hmm, does this still race with new lookups?)
4231 */
4232void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4233{
4234 dout("pre_umount\n");
4235 mdsc->stopping = 1;
4236
4237 lock_unlock_sessions(mdsc);
4238 ceph_flush_dirty_caps(mdsc);
4239 wait_requests(mdsc);
4240
4241 /*
4242 * wait for reply handlers to drop their request refs and
4243 * their inode/dcache refs
4244 */
4245 ceph_msgr_flush();
4246
4247 ceph_cleanup_quotarealms_inodes(mdsc);
4248}
4249
4250/*
4251 * wait for all write mds requests to flush.
4252 */
4253static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4254{
4255 struct ceph_mds_request *req = NULL, *nextreq;
4256 struct rb_node *n;
4257
4258 mutex_lock(&mdsc->mutex);
4259 dout("wait_unsafe_requests want %lld\n", want_tid);
4260restart:
4261 req = __get_oldest_req(mdsc);
4262 while (req && req->r_tid <= want_tid) {
4263 /* find next request */
4264 n = rb_next(&req->r_node);
4265 if (n)
4266 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4267 else
4268 nextreq = NULL;
4269 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4270 (req->r_op & CEPH_MDS_OP_WRITE)) {
4271 /* write op */
4272 ceph_mdsc_get_request(req);
4273 if (nextreq)
4274 ceph_mdsc_get_request(nextreq);
4275 mutex_unlock(&mdsc->mutex);
4276 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
4277 req->r_tid, want_tid);
4278 wait_for_completion(&req->r_safe_completion);
4279 mutex_lock(&mdsc->mutex);
4280 ceph_mdsc_put_request(req);
4281 if (!nextreq)
4282 break; /* next dne before, so we're done! */
4283 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4284 /* next request was removed from tree */
4285 ceph_mdsc_put_request(nextreq);
4286 goto restart;
4287 }
4288 ceph_mdsc_put_request(nextreq); /* won't go away */
4289 }
4290 req = nextreq;
4291 }
4292 mutex_unlock(&mdsc->mutex);
4293 dout("wait_unsafe_requests done\n");
4294}
4295
4296void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4297{
4298 u64 want_tid, want_flush;
4299
4300 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4301 return;
4302
4303 dout("sync\n");
4304 mutex_lock(&mdsc->mutex);
4305 want_tid = mdsc->last_tid;
4306 mutex_unlock(&mdsc->mutex);
4307
4308 ceph_flush_dirty_caps(mdsc);
4309 spin_lock(&mdsc->cap_dirty_lock);
4310 want_flush = mdsc->last_cap_flush_tid;
4311 if (!list_empty(&mdsc->cap_flush_list)) {
4312 struct ceph_cap_flush *cf =
4313 list_last_entry(&mdsc->cap_flush_list,
4314 struct ceph_cap_flush, g_list);
4315 cf->wake = true;
4316 }
4317 spin_unlock(&mdsc->cap_dirty_lock);
4318
4319 dout("sync want tid %lld flush_seq %lld\n",
4320 want_tid, want_flush);
4321
4322 wait_unsafe_requests(mdsc, want_tid);
4323 wait_caps_flush(mdsc, want_flush);
4324}
4325
4326/*
4327 * true if all sessions are closed, or we force unmount
4328 */
4329static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4330{
4331 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4332 return true;
4333 return atomic_read(&mdsc->num_sessions) <= skipped;
4334}
4335
4336/*
4337 * called after sb is ro.
4338 */
4339void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4340{
4341 struct ceph_options *opts = mdsc->fsc->client->options;
4342 struct ceph_mds_session *session;
4343 int i;
4344 int skipped = 0;
4345
4346 dout("close_sessions\n");
4347
4348 /* close sessions */
4349 mutex_lock(&mdsc->mutex);
4350 for (i = 0; i < mdsc->max_sessions; i++) {
4351 session = __ceph_lookup_mds_session(mdsc, i);
4352 if (!session)
4353 continue;
4354 mutex_unlock(&mdsc->mutex);
4355 mutex_lock(&session->s_mutex);
4356 if (__close_session(mdsc, session) <= 0)
4357 skipped++;
4358 mutex_unlock(&session->s_mutex);
4359 ceph_put_mds_session(session);
4360 mutex_lock(&mdsc->mutex);
4361 }
4362 mutex_unlock(&mdsc->mutex);
4363
4364 dout("waiting for sessions to close\n");
4365 wait_event_timeout(mdsc->session_close_wq,
4366 done_closing_sessions(mdsc, skipped),
4367 ceph_timeout_jiffies(opts->mount_timeout));
4368
4369 /* tear down remaining sessions */
4370 mutex_lock(&mdsc->mutex);
4371 for (i = 0; i < mdsc->max_sessions; i++) {
4372 if (mdsc->sessions[i]) {
4373 session = get_session(mdsc->sessions[i]);
4374 __unregister_session(mdsc, session);
4375 mutex_unlock(&mdsc->mutex);
4376 mutex_lock(&session->s_mutex);
4377 remove_session_caps(session);
4378 mutex_unlock(&session->s_mutex);
4379 ceph_put_mds_session(session);
4380 mutex_lock(&mdsc->mutex);
4381 }
4382 }
4383 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4384 mutex_unlock(&mdsc->mutex);
4385
4386 ceph_cleanup_snapid_map(mdsc);
4387 ceph_cleanup_empty_realms(mdsc);
4388
4389 cancel_work_sync(&mdsc->cap_reclaim_work);
4390 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4391
4392 dout("stopped\n");
4393}
4394
4395void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4396{
4397 struct ceph_mds_session *session;
4398 int mds;
4399
4400 dout("force umount\n");
4401
4402 mutex_lock(&mdsc->mutex);
4403 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4404 session = __ceph_lookup_mds_session(mdsc, mds);
4405 if (!session)
4406 continue;
4407
4408 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4409 __unregister_session(mdsc, session);
4410 __wake_requests(mdsc, &session->s_waiting);
4411 mutex_unlock(&mdsc->mutex);
4412
4413 mutex_lock(&session->s_mutex);
4414 __close_session(mdsc, session);
4415 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4416 cleanup_session_requests(mdsc, session);
4417 remove_session_caps(session);
4418 }
4419 mutex_unlock(&session->s_mutex);
4420 ceph_put_mds_session(session);
4421
4422 mutex_lock(&mdsc->mutex);
4423 kick_requests(mdsc, mds);
4424 }
4425 __wake_requests(mdsc, &mdsc->waiting_for_map);
4426 mutex_unlock(&mdsc->mutex);
4427}
4428
4429static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4430{
4431 dout("stop\n");
4432 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4433 if (mdsc->mdsmap)
4434 ceph_mdsmap_destroy(mdsc->mdsmap);
4435 kfree(mdsc->sessions);
4436 ceph_caps_finalize(mdsc);
4437 ceph_pool_perm_destroy(mdsc);
4438}
4439
4440void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4441{
4442 struct ceph_mds_client *mdsc = fsc->mdsc;
4443 dout("mdsc_destroy %p\n", mdsc);
4444
4445 if (!mdsc)
4446 return;
4447
4448 /* flush out any connection work with references to us */
4449 ceph_msgr_flush();
4450
4451 ceph_mdsc_stop(mdsc);
4452
4453 fsc->mdsc = NULL;
4454 kfree(mdsc);
4455 dout("mdsc_destroy %p done\n", mdsc);
4456}
4457
4458void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4459{
4460 struct ceph_fs_client *fsc = mdsc->fsc;
4461 const char *mds_namespace = fsc->mount_options->mds_namespace;
4462 void *p = msg->front.iov_base;
4463 void *end = p + msg->front.iov_len;
4464 u32 epoch;
4465 u32 map_len;
4466 u32 num_fs;
4467 u32 mount_fscid = (u32)-1;
4468 u8 struct_v, struct_cv;
4469 int err = -EINVAL;
4470
4471 ceph_decode_need(&p, end, sizeof(u32), bad);
4472 epoch = ceph_decode_32(&p);
4473
4474 dout("handle_fsmap epoch %u\n", epoch);
4475
4476 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4477 struct_v = ceph_decode_8(&p);
4478 struct_cv = ceph_decode_8(&p);
4479 map_len = ceph_decode_32(&p);
4480
4481 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4482 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4483
4484 num_fs = ceph_decode_32(&p);
4485 while (num_fs-- > 0) {
4486 void *info_p, *info_end;
4487 u32 info_len;
4488 u8 info_v, info_cv;
4489 u32 fscid, namelen;
4490
4491 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4492 info_v = ceph_decode_8(&p);
4493 info_cv = ceph_decode_8(&p);
4494 info_len = ceph_decode_32(&p);
4495 ceph_decode_need(&p, end, info_len, bad);
4496 info_p = p;
4497 info_end = p + info_len;
4498 p = info_end;
4499
4500 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4501 fscid = ceph_decode_32(&info_p);
4502 namelen = ceph_decode_32(&info_p);
4503 ceph_decode_need(&info_p, info_end, namelen, bad);
4504
4505 if (mds_namespace &&
4506 strlen(mds_namespace) == namelen &&
4507 !strncmp(mds_namespace, (char *)info_p, namelen)) {
4508 mount_fscid = fscid;
4509 break;
4510 }
4511 }
4512
4513 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4514 if (mount_fscid != (u32)-1) {
4515 fsc->client->monc.fs_cluster_id = mount_fscid;
4516 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4517 0, true);
4518 ceph_monc_renew_subs(&fsc->client->monc);
4519 } else {
4520 err = -ENOENT;
4521 goto err_out;
4522 }
4523 return;
4524
4525bad:
4526 pr_err("error decoding fsmap\n");
4527err_out:
4528 mutex_lock(&mdsc->mutex);
4529 mdsc->mdsmap_err = err;
4530 __wake_requests(mdsc, &mdsc->waiting_for_map);
4531 mutex_unlock(&mdsc->mutex);
4532}
4533
4534/*
4535 * handle mds map update.
4536 */
4537void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4538{
4539 u32 epoch;
4540 u32 maplen;
4541 void *p = msg->front.iov_base;
4542 void *end = p + msg->front.iov_len;
4543 struct ceph_mdsmap *newmap, *oldmap;
4544 struct ceph_fsid fsid;
4545 int err = -EINVAL;
4546
4547 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4548 ceph_decode_copy(&p, &fsid, sizeof(fsid));
4549 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4550 return;
4551 epoch = ceph_decode_32(&p);
4552 maplen = ceph_decode_32(&p);
4553 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4554
4555 /* do we need it? */
4556 mutex_lock(&mdsc->mutex);
4557 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4558 dout("handle_map epoch %u <= our %u\n",
4559 epoch, mdsc->mdsmap->m_epoch);
4560 mutex_unlock(&mdsc->mutex);
4561 return;
4562 }
4563
4564 newmap = ceph_mdsmap_decode(&p, end);
4565 if (IS_ERR(newmap)) {
4566 err = PTR_ERR(newmap);
4567 goto bad_unlock;
4568 }
4569
4570 /* swap into place */
4571 if (mdsc->mdsmap) {
4572 oldmap = mdsc->mdsmap;
4573 mdsc->mdsmap = newmap;
4574 check_new_map(mdsc, newmap, oldmap);
4575 ceph_mdsmap_destroy(oldmap);
4576 } else {
4577 mdsc->mdsmap = newmap; /* first mds map */
4578 }
4579 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4580 MAX_LFS_FILESIZE);
4581
4582 __wake_requests(mdsc, &mdsc->waiting_for_map);
4583 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4584 mdsc->mdsmap->m_epoch);
4585
4586 mutex_unlock(&mdsc->mutex);
4587 schedule_delayed(mdsc);
4588 return;
4589
4590bad_unlock:
4591 mutex_unlock(&mdsc->mutex);
4592bad:
4593 pr_err("error decoding mdsmap %d\n", err);
4594 return;
4595}
4596
4597static struct ceph_connection *con_get(struct ceph_connection *con)
4598{
4599 struct ceph_mds_session *s = con->private;
4600
4601 if (get_session(s)) {
4602 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4603 return con;
4604 }
4605 dout("mdsc con_get %p FAIL\n", s);
4606 return NULL;
4607}
4608
4609static void con_put(struct ceph_connection *con)
4610{
4611 struct ceph_mds_session *s = con->private;
4612
4613 dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4614 ceph_put_mds_session(s);
4615}
4616
4617/*
4618 * if the client is unresponsive for long enough, the mds will kill
4619 * the session entirely.
4620 */
4621static void peer_reset(struct ceph_connection *con)
4622{
4623 struct ceph_mds_session *s = con->private;
4624 struct ceph_mds_client *mdsc = s->s_mdsc;
4625
4626 pr_warn("mds%d closed our session\n", s->s_mds);
4627 send_mds_reconnect(mdsc, s);
4628}
4629
4630static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4631{
4632 struct ceph_mds_session *s = con->private;
4633 struct ceph_mds_client *mdsc = s->s_mdsc;
4634 int type = le16_to_cpu(msg->hdr.type);
4635
4636 mutex_lock(&mdsc->mutex);
4637 if (__verify_registered_session(mdsc, s) < 0) {
4638 mutex_unlock(&mdsc->mutex);
4639 goto out;
4640 }
4641 mutex_unlock(&mdsc->mutex);
4642
4643 switch (type) {
4644 case CEPH_MSG_MDS_MAP:
4645 ceph_mdsc_handle_mdsmap(mdsc, msg);
4646 break;
4647 case CEPH_MSG_FS_MAP_USER:
4648 ceph_mdsc_handle_fsmap(mdsc, msg);
4649 break;
4650 case CEPH_MSG_CLIENT_SESSION:
4651 handle_session(s, msg);
4652 break;
4653 case CEPH_MSG_CLIENT_REPLY:
4654 handle_reply(s, msg);
4655 break;
4656 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4657 handle_forward(mdsc, s, msg);
4658 break;
4659 case CEPH_MSG_CLIENT_CAPS:
4660 ceph_handle_caps(s, msg);
4661 break;
4662 case CEPH_MSG_CLIENT_SNAP:
4663 ceph_handle_snap(mdsc, s, msg);
4664 break;
4665 case CEPH_MSG_CLIENT_LEASE:
4666 handle_lease(mdsc, s, msg);
4667 break;
4668 case CEPH_MSG_CLIENT_QUOTA:
4669 ceph_handle_quota(mdsc, s, msg);
4670 break;
4671
4672 default:
4673 pr_err("received unknown message type %d %s\n", type,
4674 ceph_msg_type_name(type));
4675 }
4676out:
4677 ceph_msg_put(msg);
4678}
4679
4680/*
4681 * authentication
4682 */
4683
4684/*
4685 * Note: returned pointer is the address of a structure that's
4686 * managed separately. Caller must *not* attempt to free it.
4687 */
4688static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4689 int *proto, int force_new)
4690{
4691 struct ceph_mds_session *s = con->private;
4692 struct ceph_mds_client *mdsc = s->s_mdsc;
4693 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4694 struct ceph_auth_handshake *auth = &s->s_auth;
4695
4696 if (force_new && auth->authorizer) {
4697 ceph_auth_destroy_authorizer(auth->authorizer);
4698 auth->authorizer = NULL;
4699 }
4700 if (!auth->authorizer) {
4701 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4702 auth);
4703 if (ret)
4704 return ERR_PTR(ret);
4705 } else {
4706 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4707 auth);
4708 if (ret)
4709 return ERR_PTR(ret);
4710 }
4711 *proto = ac->protocol;
4712
4713 return auth;
4714}
4715
4716static int add_authorizer_challenge(struct ceph_connection *con,
4717 void *challenge_buf, int challenge_buf_len)
4718{
4719 struct ceph_mds_session *s = con->private;
4720 struct ceph_mds_client *mdsc = s->s_mdsc;
4721 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4722
4723 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4724 challenge_buf, challenge_buf_len);
4725}
4726
4727static int verify_authorizer_reply(struct ceph_connection *con)
4728{
4729 struct ceph_mds_session *s = con->private;
4730 struct ceph_mds_client *mdsc = s->s_mdsc;
4731 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4732
4733 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4734}
4735
4736static int invalidate_authorizer(struct ceph_connection *con)
4737{
4738 struct ceph_mds_session *s = con->private;
4739 struct ceph_mds_client *mdsc = s->s_mdsc;
4740 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4741
4742 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4743
4744 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4745}
4746
4747static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4748 struct ceph_msg_header *hdr, int *skip)
4749{
4750 struct ceph_msg *msg;
4751 int type = (int) le16_to_cpu(hdr->type);
4752 int front_len = (int) le32_to_cpu(hdr->front_len);
4753
4754 if (con->in_msg)
4755 return con->in_msg;
4756
4757 *skip = 0;
4758 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4759 if (!msg) {
4760 pr_err("unable to allocate msg type %d len %d\n",
4761 type, front_len);
4762 return NULL;
4763 }
4764
4765 return msg;
4766}
4767
4768static int mds_sign_message(struct ceph_msg *msg)
4769{
4770 struct ceph_mds_session *s = msg->con->private;
4771 struct ceph_auth_handshake *auth = &s->s_auth;
4772
4773 return ceph_auth_sign_message(auth, msg);
4774}
4775
4776static int mds_check_message_signature(struct ceph_msg *msg)
4777{
4778 struct ceph_mds_session *s = msg->con->private;
4779 struct ceph_auth_handshake *auth = &s->s_auth;
4780
4781 return ceph_auth_check_message_signature(auth, msg);
4782}
4783
4784static const struct ceph_connection_operations mds_con_ops = {
4785 .get = con_get,
4786 .put = con_put,
4787 .dispatch = dispatch,
4788 .get_authorizer = get_authorizer,
4789 .add_authorizer_challenge = add_authorizer_challenge,
4790 .verify_authorizer_reply = verify_authorizer_reply,
4791 .invalidate_authorizer = invalidate_authorizer,
4792 .peer_reset = peer_reset,
4793 .alloc_msg = mds_alloc_msg,
4794 .sign_message = mds_sign_message,
4795 .check_message_signature = mds_check_message_signature,
4796};
4797
4798/* eof */
1#include <linux/ceph/ceph_debug.h>
2
3#include <linux/fs.h>
4#include <linux/wait.h>
5#include <linux/slab.h>
6#include <linux/gfp.h>
7#include <linux/sched.h>
8#include <linux/debugfs.h>
9#include <linux/seq_file.h>
10#include <linux/utsname.h>
11#include <linux/ratelimit.h>
12
13#include "super.h"
14#include "mds_client.h"
15
16#include <linux/ceph/ceph_features.h>
17#include <linux/ceph/messenger.h>
18#include <linux/ceph/decode.h>
19#include <linux/ceph/pagelist.h>
20#include <linux/ceph/auth.h>
21#include <linux/ceph/debugfs.h>
22
23/*
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
30 *
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
35 * requests.
36 *
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
40 *
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
46 */
47
48struct ceph_reconnect_state {
49 int nr_caps;
50 struct ceph_pagelist *pagelist;
51 unsigned msg_version;
52};
53
54static void __wake_requests(struct ceph_mds_client *mdsc,
55 struct list_head *head);
56
57static const struct ceph_connection_operations mds_con_ops;
58
59
60/*
61 * mds reply parsing
62 */
63
64/*
65 * parse individual inode info
66 */
67static int parse_reply_info_in(void **p, void *end,
68 struct ceph_mds_reply_info_in *info,
69 u64 features)
70{
71 int err = -EIO;
72
73 info->in = *p;
74 *p += sizeof(struct ceph_mds_reply_inode) +
75 sizeof(*info->in->fragtree.splits) *
76 le32_to_cpu(info->in->fragtree.nsplits);
77
78 ceph_decode_32_safe(p, end, info->symlink_len, bad);
79 ceph_decode_need(p, end, info->symlink_len, bad);
80 info->symlink = *p;
81 *p += info->symlink_len;
82
83 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
84 ceph_decode_copy_safe(p, end, &info->dir_layout,
85 sizeof(info->dir_layout), bad);
86 else
87 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
88
89 ceph_decode_32_safe(p, end, info->xattr_len, bad);
90 ceph_decode_need(p, end, info->xattr_len, bad);
91 info->xattr_data = *p;
92 *p += info->xattr_len;
93
94 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
95 ceph_decode_64_safe(p, end, info->inline_version, bad);
96 ceph_decode_32_safe(p, end, info->inline_len, bad);
97 ceph_decode_need(p, end, info->inline_len, bad);
98 info->inline_data = *p;
99 *p += info->inline_len;
100 } else
101 info->inline_version = CEPH_INLINE_NONE;
102
103 info->pool_ns_len = 0;
104 info->pool_ns_data = NULL;
105 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
106 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
107 if (info->pool_ns_len > 0) {
108 ceph_decode_need(p, end, info->pool_ns_len, bad);
109 info->pool_ns_data = *p;
110 *p += info->pool_ns_len;
111 }
112 }
113
114 return 0;
115bad:
116 return err;
117}
118
119/*
120 * parse a normal reply, which may contain a (dir+)dentry and/or a
121 * target inode.
122 */
123static int parse_reply_info_trace(void **p, void *end,
124 struct ceph_mds_reply_info_parsed *info,
125 u64 features)
126{
127 int err;
128
129 if (info->head->is_dentry) {
130 err = parse_reply_info_in(p, end, &info->diri, features);
131 if (err < 0)
132 goto out_bad;
133
134 if (unlikely(*p + sizeof(*info->dirfrag) > end))
135 goto bad;
136 info->dirfrag = *p;
137 *p += sizeof(*info->dirfrag) +
138 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
139 if (unlikely(*p > end))
140 goto bad;
141
142 ceph_decode_32_safe(p, end, info->dname_len, bad);
143 ceph_decode_need(p, end, info->dname_len, bad);
144 info->dname = *p;
145 *p += info->dname_len;
146 info->dlease = *p;
147 *p += sizeof(*info->dlease);
148 }
149
150 if (info->head->is_target) {
151 err = parse_reply_info_in(p, end, &info->targeti, features);
152 if (err < 0)
153 goto out_bad;
154 }
155
156 if (unlikely(*p != end))
157 goto bad;
158 return 0;
159
160bad:
161 err = -EIO;
162out_bad:
163 pr_err("problem parsing mds trace %d\n", err);
164 return err;
165}
166
167/*
168 * parse readdir results
169 */
170static int parse_reply_info_dir(void **p, void *end,
171 struct ceph_mds_reply_info_parsed *info,
172 u64 features)
173{
174 u32 num, i = 0;
175 int err;
176
177 info->dir_dir = *p;
178 if (*p + sizeof(*info->dir_dir) > end)
179 goto bad;
180 *p += sizeof(*info->dir_dir) +
181 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
182 if (*p > end)
183 goto bad;
184
185 ceph_decode_need(p, end, sizeof(num) + 2, bad);
186 num = ceph_decode_32(p);
187 {
188 u16 flags = ceph_decode_16(p);
189 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
190 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
191 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
192 }
193 if (num == 0)
194 goto done;
195
196 BUG_ON(!info->dir_entries);
197 if ((unsigned long)(info->dir_entries + num) >
198 (unsigned long)info->dir_entries + info->dir_buf_size) {
199 pr_err("dir contents are larger than expected\n");
200 WARN_ON(1);
201 goto bad;
202 }
203
204 info->dir_nr = num;
205 while (num) {
206 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
207 /* dentry */
208 ceph_decode_need(p, end, sizeof(u32)*2, bad);
209 rde->name_len = ceph_decode_32(p);
210 ceph_decode_need(p, end, rde->name_len, bad);
211 rde->name = *p;
212 *p += rde->name_len;
213 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
214 rde->lease = *p;
215 *p += sizeof(struct ceph_mds_reply_lease);
216
217 /* inode */
218 err = parse_reply_info_in(p, end, &rde->inode, features);
219 if (err < 0)
220 goto out_bad;
221 /* ceph_readdir_prepopulate() will update it */
222 rde->offset = 0;
223 i++;
224 num--;
225 }
226
227done:
228 if (*p != end)
229 goto bad;
230 return 0;
231
232bad:
233 err = -EIO;
234out_bad:
235 pr_err("problem parsing dir contents %d\n", err);
236 return err;
237}
238
239/*
240 * parse fcntl F_GETLK results
241 */
242static int parse_reply_info_filelock(void **p, void *end,
243 struct ceph_mds_reply_info_parsed *info,
244 u64 features)
245{
246 if (*p + sizeof(*info->filelock_reply) > end)
247 goto bad;
248
249 info->filelock_reply = *p;
250 *p += sizeof(*info->filelock_reply);
251
252 if (unlikely(*p != end))
253 goto bad;
254 return 0;
255
256bad:
257 return -EIO;
258}
259
260/*
261 * parse create results
262 */
263static int parse_reply_info_create(void **p, void *end,
264 struct ceph_mds_reply_info_parsed *info,
265 u64 features)
266{
267 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
268 if (*p == end) {
269 info->has_create_ino = false;
270 } else {
271 info->has_create_ino = true;
272 info->ino = ceph_decode_64(p);
273 }
274 }
275
276 if (unlikely(*p != end))
277 goto bad;
278 return 0;
279
280bad:
281 return -EIO;
282}
283
284/*
285 * parse extra results
286 */
287static int parse_reply_info_extra(void **p, void *end,
288 struct ceph_mds_reply_info_parsed *info,
289 u64 features)
290{
291 u32 op = le32_to_cpu(info->head->op);
292
293 if (op == CEPH_MDS_OP_GETFILELOCK)
294 return parse_reply_info_filelock(p, end, info, features);
295 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
296 return parse_reply_info_dir(p, end, info, features);
297 else if (op == CEPH_MDS_OP_CREATE)
298 return parse_reply_info_create(p, end, info, features);
299 else
300 return -EIO;
301}
302
303/*
304 * parse entire mds reply
305 */
306static int parse_reply_info(struct ceph_msg *msg,
307 struct ceph_mds_reply_info_parsed *info,
308 u64 features)
309{
310 void *p, *end;
311 u32 len;
312 int err;
313
314 info->head = msg->front.iov_base;
315 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
316 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
317
318 /* trace */
319 ceph_decode_32_safe(&p, end, len, bad);
320 if (len > 0) {
321 ceph_decode_need(&p, end, len, bad);
322 err = parse_reply_info_trace(&p, p+len, info, features);
323 if (err < 0)
324 goto out_bad;
325 }
326
327 /* extra */
328 ceph_decode_32_safe(&p, end, len, bad);
329 if (len > 0) {
330 ceph_decode_need(&p, end, len, bad);
331 err = parse_reply_info_extra(&p, p+len, info, features);
332 if (err < 0)
333 goto out_bad;
334 }
335
336 /* snap blob */
337 ceph_decode_32_safe(&p, end, len, bad);
338 info->snapblob_len = len;
339 info->snapblob = p;
340 p += len;
341
342 if (p != end)
343 goto bad;
344 return 0;
345
346bad:
347 err = -EIO;
348out_bad:
349 pr_err("mds parse_reply err %d\n", err);
350 return err;
351}
352
353static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
354{
355 if (!info->dir_entries)
356 return;
357 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
358}
359
360
361/*
362 * sessions
363 */
364const char *ceph_session_state_name(int s)
365{
366 switch (s) {
367 case CEPH_MDS_SESSION_NEW: return "new";
368 case CEPH_MDS_SESSION_OPENING: return "opening";
369 case CEPH_MDS_SESSION_OPEN: return "open";
370 case CEPH_MDS_SESSION_HUNG: return "hung";
371 case CEPH_MDS_SESSION_CLOSING: return "closing";
372 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
373 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
374 case CEPH_MDS_SESSION_REJECTED: return "rejected";
375 default: return "???";
376 }
377}
378
379static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
380{
381 if (atomic_inc_not_zero(&s->s_ref)) {
382 dout("mdsc get_session %p %d -> %d\n", s,
383 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
384 return s;
385 } else {
386 dout("mdsc get_session %p 0 -- FAIL", s);
387 return NULL;
388 }
389}
390
391void ceph_put_mds_session(struct ceph_mds_session *s)
392{
393 dout("mdsc put_session %p %d -> %d\n", s,
394 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
395 if (atomic_dec_and_test(&s->s_ref)) {
396 if (s->s_auth.authorizer)
397 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
398 kfree(s);
399 }
400}
401
402/*
403 * called under mdsc->mutex
404 */
405struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
406 int mds)
407{
408 struct ceph_mds_session *session;
409
410 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
411 return NULL;
412 session = mdsc->sessions[mds];
413 dout("lookup_mds_session %p %d\n", session,
414 atomic_read(&session->s_ref));
415 get_session(session);
416 return session;
417}
418
419static bool __have_session(struct ceph_mds_client *mdsc, int mds)
420{
421 if (mds >= mdsc->max_sessions)
422 return false;
423 return mdsc->sessions[mds];
424}
425
426static int __verify_registered_session(struct ceph_mds_client *mdsc,
427 struct ceph_mds_session *s)
428{
429 if (s->s_mds >= mdsc->max_sessions ||
430 mdsc->sessions[s->s_mds] != s)
431 return -ENOENT;
432 return 0;
433}
434
435/*
436 * create+register a new session for given mds.
437 * called under mdsc->mutex.
438 */
439static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
440 int mds)
441{
442 struct ceph_mds_session *s;
443
444 if (mds >= mdsc->mdsmap->m_max_mds)
445 return ERR_PTR(-EINVAL);
446
447 s = kzalloc(sizeof(*s), GFP_NOFS);
448 if (!s)
449 return ERR_PTR(-ENOMEM);
450 s->s_mdsc = mdsc;
451 s->s_mds = mds;
452 s->s_state = CEPH_MDS_SESSION_NEW;
453 s->s_ttl = 0;
454 s->s_seq = 0;
455 mutex_init(&s->s_mutex);
456
457 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
458
459 spin_lock_init(&s->s_gen_ttl_lock);
460 s->s_cap_gen = 0;
461 s->s_cap_ttl = jiffies - 1;
462
463 spin_lock_init(&s->s_cap_lock);
464 s->s_renew_requested = 0;
465 s->s_renew_seq = 0;
466 INIT_LIST_HEAD(&s->s_caps);
467 s->s_nr_caps = 0;
468 s->s_trim_caps = 0;
469 atomic_set(&s->s_ref, 1);
470 INIT_LIST_HEAD(&s->s_waiting);
471 INIT_LIST_HEAD(&s->s_unsafe);
472 s->s_num_cap_releases = 0;
473 s->s_cap_reconnect = 0;
474 s->s_cap_iterator = NULL;
475 INIT_LIST_HEAD(&s->s_cap_releases);
476 INIT_LIST_HEAD(&s->s_cap_flushing);
477
478 dout("register_session mds%d\n", mds);
479 if (mds >= mdsc->max_sessions) {
480 int newmax = 1 << get_count_order(mds+1);
481 struct ceph_mds_session **sa;
482
483 dout("register_session realloc to %d\n", newmax);
484 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
485 if (sa == NULL)
486 goto fail_realloc;
487 if (mdsc->sessions) {
488 memcpy(sa, mdsc->sessions,
489 mdsc->max_sessions * sizeof(void *));
490 kfree(mdsc->sessions);
491 }
492 mdsc->sessions = sa;
493 mdsc->max_sessions = newmax;
494 }
495 mdsc->sessions[mds] = s;
496 atomic_inc(&mdsc->num_sessions);
497 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
498
499 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
500 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
501
502 return s;
503
504fail_realloc:
505 kfree(s);
506 return ERR_PTR(-ENOMEM);
507}
508
509/*
510 * called under mdsc->mutex
511 */
512static void __unregister_session(struct ceph_mds_client *mdsc,
513 struct ceph_mds_session *s)
514{
515 dout("__unregister_session mds%d %p\n", s->s_mds, s);
516 BUG_ON(mdsc->sessions[s->s_mds] != s);
517 mdsc->sessions[s->s_mds] = NULL;
518 ceph_con_close(&s->s_con);
519 ceph_put_mds_session(s);
520 atomic_dec(&mdsc->num_sessions);
521}
522
523/*
524 * drop session refs in request.
525 *
526 * should be last request ref, or hold mdsc->mutex
527 */
528static void put_request_session(struct ceph_mds_request *req)
529{
530 if (req->r_session) {
531 ceph_put_mds_session(req->r_session);
532 req->r_session = NULL;
533 }
534}
535
536void ceph_mdsc_release_request(struct kref *kref)
537{
538 struct ceph_mds_request *req = container_of(kref,
539 struct ceph_mds_request,
540 r_kref);
541 destroy_reply_info(&req->r_reply_info);
542 if (req->r_request)
543 ceph_msg_put(req->r_request);
544 if (req->r_reply)
545 ceph_msg_put(req->r_reply);
546 if (req->r_inode) {
547 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
548 iput(req->r_inode);
549 }
550 if (req->r_locked_dir)
551 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
552 iput(req->r_target_inode);
553 if (req->r_dentry)
554 dput(req->r_dentry);
555 if (req->r_old_dentry)
556 dput(req->r_old_dentry);
557 if (req->r_old_dentry_dir) {
558 /*
559 * track (and drop pins for) r_old_dentry_dir
560 * separately, since r_old_dentry's d_parent may have
561 * changed between the dir mutex being dropped and
562 * this request being freed.
563 */
564 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
565 CEPH_CAP_PIN);
566 iput(req->r_old_dentry_dir);
567 }
568 kfree(req->r_path1);
569 kfree(req->r_path2);
570 if (req->r_pagelist)
571 ceph_pagelist_release(req->r_pagelist);
572 put_request_session(req);
573 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
574 kfree(req);
575}
576
577DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
578
579/*
580 * lookup session, bump ref if found.
581 *
582 * called under mdsc->mutex.
583 */
584static struct ceph_mds_request *
585lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
586{
587 struct ceph_mds_request *req;
588
589 req = lookup_request(&mdsc->request_tree, tid);
590 if (req)
591 ceph_mdsc_get_request(req);
592
593 return req;
594}
595
596/*
597 * Register an in-flight request, and assign a tid. Link to directory
598 * are modifying (if any).
599 *
600 * Called under mdsc->mutex.
601 */
602static void __register_request(struct ceph_mds_client *mdsc,
603 struct ceph_mds_request *req,
604 struct inode *dir)
605{
606 req->r_tid = ++mdsc->last_tid;
607 if (req->r_num_caps)
608 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
609 req->r_num_caps);
610 dout("__register_request %p tid %lld\n", req, req->r_tid);
611 ceph_mdsc_get_request(req);
612 insert_request(&mdsc->request_tree, req);
613
614 req->r_uid = current_fsuid();
615 req->r_gid = current_fsgid();
616
617 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
618 mdsc->oldest_tid = req->r_tid;
619
620 if (dir) {
621 ihold(dir);
622 req->r_unsafe_dir = dir;
623 }
624}
625
626static void __unregister_request(struct ceph_mds_client *mdsc,
627 struct ceph_mds_request *req)
628{
629 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
630
631 /* Never leave an unregistered request on an unsafe list! */
632 list_del_init(&req->r_unsafe_item);
633
634 if (req->r_tid == mdsc->oldest_tid) {
635 struct rb_node *p = rb_next(&req->r_node);
636 mdsc->oldest_tid = 0;
637 while (p) {
638 struct ceph_mds_request *next_req =
639 rb_entry(p, struct ceph_mds_request, r_node);
640 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
641 mdsc->oldest_tid = next_req->r_tid;
642 break;
643 }
644 p = rb_next(p);
645 }
646 }
647
648 erase_request(&mdsc->request_tree, req);
649
650 if (req->r_unsafe_dir && req->r_got_unsafe) {
651 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
652 spin_lock(&ci->i_unsafe_lock);
653 list_del_init(&req->r_unsafe_dir_item);
654 spin_unlock(&ci->i_unsafe_lock);
655 }
656 if (req->r_target_inode && req->r_got_unsafe) {
657 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
658 spin_lock(&ci->i_unsafe_lock);
659 list_del_init(&req->r_unsafe_target_item);
660 spin_unlock(&ci->i_unsafe_lock);
661 }
662
663 if (req->r_unsafe_dir) {
664 iput(req->r_unsafe_dir);
665 req->r_unsafe_dir = NULL;
666 }
667
668 complete_all(&req->r_safe_completion);
669
670 ceph_mdsc_put_request(req);
671}
672
673/*
674 * Choose mds to send request to next. If there is a hint set in the
675 * request (e.g., due to a prior forward hint from the mds), use that.
676 * Otherwise, consult frag tree and/or caps to identify the
677 * appropriate mds. If all else fails, choose randomly.
678 *
679 * Called under mdsc->mutex.
680 */
681static struct dentry *get_nonsnap_parent(struct dentry *dentry)
682{
683 /*
684 * we don't need to worry about protecting the d_parent access
685 * here because we never renaming inside the snapped namespace
686 * except to resplice to another snapdir, and either the old or new
687 * result is a valid result.
688 */
689 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
690 dentry = dentry->d_parent;
691 return dentry;
692}
693
694static int __choose_mds(struct ceph_mds_client *mdsc,
695 struct ceph_mds_request *req)
696{
697 struct inode *inode;
698 struct ceph_inode_info *ci;
699 struct ceph_cap *cap;
700 int mode = req->r_direct_mode;
701 int mds = -1;
702 u32 hash = req->r_direct_hash;
703 bool is_hash = req->r_direct_is_hash;
704
705 /*
706 * is there a specific mds we should try? ignore hint if we have
707 * no session and the mds is not up (active or recovering).
708 */
709 if (req->r_resend_mds >= 0 &&
710 (__have_session(mdsc, req->r_resend_mds) ||
711 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
712 dout("choose_mds using resend_mds mds%d\n",
713 req->r_resend_mds);
714 return req->r_resend_mds;
715 }
716
717 if (mode == USE_RANDOM_MDS)
718 goto random;
719
720 inode = NULL;
721 if (req->r_inode) {
722 inode = req->r_inode;
723 } else if (req->r_dentry) {
724 /* ignore race with rename; old or new d_parent is okay */
725 struct dentry *parent = req->r_dentry->d_parent;
726 struct inode *dir = d_inode(parent);
727
728 if (dir->i_sb != mdsc->fsc->sb) {
729 /* not this fs! */
730 inode = d_inode(req->r_dentry);
731 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
732 /* direct snapped/virtual snapdir requests
733 * based on parent dir inode */
734 struct dentry *dn = get_nonsnap_parent(parent);
735 inode = d_inode(dn);
736 dout("__choose_mds using nonsnap parent %p\n", inode);
737 } else {
738 /* dentry target */
739 inode = d_inode(req->r_dentry);
740 if (!inode || mode == USE_AUTH_MDS) {
741 /* dir + name */
742 inode = dir;
743 hash = ceph_dentry_hash(dir, req->r_dentry);
744 is_hash = true;
745 }
746 }
747 }
748
749 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
750 (int)hash, mode);
751 if (!inode)
752 goto random;
753 ci = ceph_inode(inode);
754
755 if (is_hash && S_ISDIR(inode->i_mode)) {
756 struct ceph_inode_frag frag;
757 int found;
758
759 ceph_choose_frag(ci, hash, &frag, &found);
760 if (found) {
761 if (mode == USE_ANY_MDS && frag.ndist > 0) {
762 u8 r;
763
764 /* choose a random replica */
765 get_random_bytes(&r, 1);
766 r %= frag.ndist;
767 mds = frag.dist[r];
768 dout("choose_mds %p %llx.%llx "
769 "frag %u mds%d (%d/%d)\n",
770 inode, ceph_vinop(inode),
771 frag.frag, mds,
772 (int)r, frag.ndist);
773 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
774 CEPH_MDS_STATE_ACTIVE)
775 return mds;
776 }
777
778 /* since this file/dir wasn't known to be
779 * replicated, then we want to look for the
780 * authoritative mds. */
781 mode = USE_AUTH_MDS;
782 if (frag.mds >= 0) {
783 /* choose auth mds */
784 mds = frag.mds;
785 dout("choose_mds %p %llx.%llx "
786 "frag %u mds%d (auth)\n",
787 inode, ceph_vinop(inode), frag.frag, mds);
788 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
789 CEPH_MDS_STATE_ACTIVE)
790 return mds;
791 }
792 }
793 }
794
795 spin_lock(&ci->i_ceph_lock);
796 cap = NULL;
797 if (mode == USE_AUTH_MDS)
798 cap = ci->i_auth_cap;
799 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
800 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
801 if (!cap) {
802 spin_unlock(&ci->i_ceph_lock);
803 goto random;
804 }
805 mds = cap->session->s_mds;
806 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
807 inode, ceph_vinop(inode), mds,
808 cap == ci->i_auth_cap ? "auth " : "", cap);
809 spin_unlock(&ci->i_ceph_lock);
810 return mds;
811
812random:
813 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
814 dout("choose_mds chose random mds%d\n", mds);
815 return mds;
816}
817
818
819/*
820 * session messages
821 */
822static struct ceph_msg *create_session_msg(u32 op, u64 seq)
823{
824 struct ceph_msg *msg;
825 struct ceph_mds_session_head *h;
826
827 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
828 false);
829 if (!msg) {
830 pr_err("create_session_msg ENOMEM creating msg\n");
831 return NULL;
832 }
833 h = msg->front.iov_base;
834 h->op = cpu_to_le32(op);
835 h->seq = cpu_to_le64(seq);
836
837 return msg;
838}
839
840/*
841 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
842 * to include additional client metadata fields.
843 */
844static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
845{
846 struct ceph_msg *msg;
847 struct ceph_mds_session_head *h;
848 int i = -1;
849 int metadata_bytes = 0;
850 int metadata_key_count = 0;
851 struct ceph_options *opt = mdsc->fsc->client->options;
852 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
853 void *p;
854
855 const char* metadata[][2] = {
856 {"hostname", utsname()->nodename},
857 {"kernel_version", utsname()->release},
858 {"entity_id", opt->name ? : ""},
859 {"root", fsopt->server_path ? : "/"},
860 {NULL, NULL}
861 };
862
863 /* Calculate serialized length of metadata */
864 metadata_bytes = 4; /* map length */
865 for (i = 0; metadata[i][0] != NULL; ++i) {
866 metadata_bytes += 8 + strlen(metadata[i][0]) +
867 strlen(metadata[i][1]);
868 metadata_key_count++;
869 }
870
871 /* Allocate the message */
872 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
873 GFP_NOFS, false);
874 if (!msg) {
875 pr_err("create_session_msg ENOMEM creating msg\n");
876 return NULL;
877 }
878 h = msg->front.iov_base;
879 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
880 h->seq = cpu_to_le64(seq);
881
882 /*
883 * Serialize client metadata into waiting buffer space, using
884 * the format that userspace expects for map<string, string>
885 *
886 * ClientSession messages with metadata are v2
887 */
888 msg->hdr.version = cpu_to_le16(2);
889 msg->hdr.compat_version = cpu_to_le16(1);
890
891 /* The write pointer, following the session_head structure */
892 p = msg->front.iov_base + sizeof(*h);
893
894 /* Number of entries in the map */
895 ceph_encode_32(&p, metadata_key_count);
896
897 /* Two length-prefixed strings for each entry in the map */
898 for (i = 0; metadata[i][0] != NULL; ++i) {
899 size_t const key_len = strlen(metadata[i][0]);
900 size_t const val_len = strlen(metadata[i][1]);
901
902 ceph_encode_32(&p, key_len);
903 memcpy(p, metadata[i][0], key_len);
904 p += key_len;
905 ceph_encode_32(&p, val_len);
906 memcpy(p, metadata[i][1], val_len);
907 p += val_len;
908 }
909
910 return msg;
911}
912
913/*
914 * send session open request.
915 *
916 * called under mdsc->mutex
917 */
918static int __open_session(struct ceph_mds_client *mdsc,
919 struct ceph_mds_session *session)
920{
921 struct ceph_msg *msg;
922 int mstate;
923 int mds = session->s_mds;
924
925 /* wait for mds to go active? */
926 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
927 dout("open_session to mds%d (%s)\n", mds,
928 ceph_mds_state_name(mstate));
929 session->s_state = CEPH_MDS_SESSION_OPENING;
930 session->s_renew_requested = jiffies;
931
932 /* send connect message */
933 msg = create_session_open_msg(mdsc, session->s_seq);
934 if (!msg)
935 return -ENOMEM;
936 ceph_con_send(&session->s_con, msg);
937 return 0;
938}
939
940/*
941 * open sessions for any export targets for the given mds
942 *
943 * called under mdsc->mutex
944 */
945static struct ceph_mds_session *
946__open_export_target_session(struct ceph_mds_client *mdsc, int target)
947{
948 struct ceph_mds_session *session;
949
950 session = __ceph_lookup_mds_session(mdsc, target);
951 if (!session) {
952 session = register_session(mdsc, target);
953 if (IS_ERR(session))
954 return session;
955 }
956 if (session->s_state == CEPH_MDS_SESSION_NEW ||
957 session->s_state == CEPH_MDS_SESSION_CLOSING)
958 __open_session(mdsc, session);
959
960 return session;
961}
962
963struct ceph_mds_session *
964ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
965{
966 struct ceph_mds_session *session;
967
968 dout("open_export_target_session to mds%d\n", target);
969
970 mutex_lock(&mdsc->mutex);
971 session = __open_export_target_session(mdsc, target);
972 mutex_unlock(&mdsc->mutex);
973
974 return session;
975}
976
977static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
978 struct ceph_mds_session *session)
979{
980 struct ceph_mds_info *mi;
981 struct ceph_mds_session *ts;
982 int i, mds = session->s_mds;
983
984 if (mds >= mdsc->mdsmap->m_max_mds)
985 return;
986
987 mi = &mdsc->mdsmap->m_info[mds];
988 dout("open_export_target_sessions for mds%d (%d targets)\n",
989 session->s_mds, mi->num_export_targets);
990
991 for (i = 0; i < mi->num_export_targets; i++) {
992 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
993 if (!IS_ERR(ts))
994 ceph_put_mds_session(ts);
995 }
996}
997
998void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
999 struct ceph_mds_session *session)
1000{
1001 mutex_lock(&mdsc->mutex);
1002 __open_export_target_sessions(mdsc, session);
1003 mutex_unlock(&mdsc->mutex);
1004}
1005
1006/*
1007 * session caps
1008 */
1009
1010/* caller holds s_cap_lock, we drop it */
1011static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1012 struct ceph_mds_session *session)
1013 __releases(session->s_cap_lock)
1014{
1015 LIST_HEAD(tmp_list);
1016 list_splice_init(&session->s_cap_releases, &tmp_list);
1017 session->s_num_cap_releases = 0;
1018 spin_unlock(&session->s_cap_lock);
1019
1020 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1021 while (!list_empty(&tmp_list)) {
1022 struct ceph_cap *cap;
1023 /* zero out the in-progress message */
1024 cap = list_first_entry(&tmp_list,
1025 struct ceph_cap, session_caps);
1026 list_del(&cap->session_caps);
1027 ceph_put_cap(mdsc, cap);
1028 }
1029}
1030
1031static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1032 struct ceph_mds_session *session)
1033{
1034 struct ceph_mds_request *req;
1035 struct rb_node *p;
1036
1037 dout("cleanup_session_requests mds%d\n", session->s_mds);
1038 mutex_lock(&mdsc->mutex);
1039 while (!list_empty(&session->s_unsafe)) {
1040 req = list_first_entry(&session->s_unsafe,
1041 struct ceph_mds_request, r_unsafe_item);
1042 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1043 req->r_tid);
1044 __unregister_request(mdsc, req);
1045 }
1046 /* zero r_attempts, so kick_requests() will re-send requests */
1047 p = rb_first(&mdsc->request_tree);
1048 while (p) {
1049 req = rb_entry(p, struct ceph_mds_request, r_node);
1050 p = rb_next(p);
1051 if (req->r_session &&
1052 req->r_session->s_mds == session->s_mds)
1053 req->r_attempts = 0;
1054 }
1055 mutex_unlock(&mdsc->mutex);
1056}
1057
1058/*
1059 * Helper to safely iterate over all caps associated with a session, with
1060 * special care taken to handle a racing __ceph_remove_cap().
1061 *
1062 * Caller must hold session s_mutex.
1063 */
1064static int iterate_session_caps(struct ceph_mds_session *session,
1065 int (*cb)(struct inode *, struct ceph_cap *,
1066 void *), void *arg)
1067{
1068 struct list_head *p;
1069 struct ceph_cap *cap;
1070 struct inode *inode, *last_inode = NULL;
1071 struct ceph_cap *old_cap = NULL;
1072 int ret;
1073
1074 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1075 spin_lock(&session->s_cap_lock);
1076 p = session->s_caps.next;
1077 while (p != &session->s_caps) {
1078 cap = list_entry(p, struct ceph_cap, session_caps);
1079 inode = igrab(&cap->ci->vfs_inode);
1080 if (!inode) {
1081 p = p->next;
1082 continue;
1083 }
1084 session->s_cap_iterator = cap;
1085 spin_unlock(&session->s_cap_lock);
1086
1087 if (last_inode) {
1088 iput(last_inode);
1089 last_inode = NULL;
1090 }
1091 if (old_cap) {
1092 ceph_put_cap(session->s_mdsc, old_cap);
1093 old_cap = NULL;
1094 }
1095
1096 ret = cb(inode, cap, arg);
1097 last_inode = inode;
1098
1099 spin_lock(&session->s_cap_lock);
1100 p = p->next;
1101 if (cap->ci == NULL) {
1102 dout("iterate_session_caps finishing cap %p removal\n",
1103 cap);
1104 BUG_ON(cap->session != session);
1105 cap->session = NULL;
1106 list_del_init(&cap->session_caps);
1107 session->s_nr_caps--;
1108 if (cap->queue_release) {
1109 list_add_tail(&cap->session_caps,
1110 &session->s_cap_releases);
1111 session->s_num_cap_releases++;
1112 } else {
1113 old_cap = cap; /* put_cap it w/o locks held */
1114 }
1115 }
1116 if (ret < 0)
1117 goto out;
1118 }
1119 ret = 0;
1120out:
1121 session->s_cap_iterator = NULL;
1122 spin_unlock(&session->s_cap_lock);
1123
1124 iput(last_inode);
1125 if (old_cap)
1126 ceph_put_cap(session->s_mdsc, old_cap);
1127
1128 return ret;
1129}
1130
1131static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1132 void *arg)
1133{
1134 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1135 struct ceph_inode_info *ci = ceph_inode(inode);
1136 LIST_HEAD(to_remove);
1137 bool drop = false;
1138 bool invalidate = false;
1139
1140 dout("removing cap %p, ci is %p, inode is %p\n",
1141 cap, ci, &ci->vfs_inode);
1142 spin_lock(&ci->i_ceph_lock);
1143 __ceph_remove_cap(cap, false);
1144 if (!ci->i_auth_cap) {
1145 struct ceph_cap_flush *cf;
1146 struct ceph_mds_client *mdsc = fsc->mdsc;
1147
1148 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1149
1150 if (ci->i_wrbuffer_ref > 0 &&
1151 ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1152 invalidate = true;
1153
1154 while (!list_empty(&ci->i_cap_flush_list)) {
1155 cf = list_first_entry(&ci->i_cap_flush_list,
1156 struct ceph_cap_flush, i_list);
1157 list_move(&cf->i_list, &to_remove);
1158 }
1159
1160 spin_lock(&mdsc->cap_dirty_lock);
1161
1162 list_for_each_entry(cf, &to_remove, i_list)
1163 list_del(&cf->g_list);
1164
1165 if (!list_empty(&ci->i_dirty_item)) {
1166 pr_warn_ratelimited(
1167 " dropping dirty %s state for %p %lld\n",
1168 ceph_cap_string(ci->i_dirty_caps),
1169 inode, ceph_ino(inode));
1170 ci->i_dirty_caps = 0;
1171 list_del_init(&ci->i_dirty_item);
1172 drop = true;
1173 }
1174 if (!list_empty(&ci->i_flushing_item)) {
1175 pr_warn_ratelimited(
1176 " dropping dirty+flushing %s state for %p %lld\n",
1177 ceph_cap_string(ci->i_flushing_caps),
1178 inode, ceph_ino(inode));
1179 ci->i_flushing_caps = 0;
1180 list_del_init(&ci->i_flushing_item);
1181 mdsc->num_cap_flushing--;
1182 drop = true;
1183 }
1184 spin_unlock(&mdsc->cap_dirty_lock);
1185
1186 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1187 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1188 ci->i_prealloc_cap_flush = NULL;
1189 }
1190 }
1191 spin_unlock(&ci->i_ceph_lock);
1192 while (!list_empty(&to_remove)) {
1193 struct ceph_cap_flush *cf;
1194 cf = list_first_entry(&to_remove,
1195 struct ceph_cap_flush, i_list);
1196 list_del(&cf->i_list);
1197 ceph_free_cap_flush(cf);
1198 }
1199
1200 wake_up_all(&ci->i_cap_wq);
1201 if (invalidate)
1202 ceph_queue_invalidate(inode);
1203 if (drop)
1204 iput(inode);
1205 return 0;
1206}
1207
1208/*
1209 * caller must hold session s_mutex
1210 */
1211static void remove_session_caps(struct ceph_mds_session *session)
1212{
1213 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1214 struct super_block *sb = fsc->sb;
1215 dout("remove_session_caps on %p\n", session);
1216 iterate_session_caps(session, remove_session_caps_cb, fsc);
1217
1218 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1219
1220 spin_lock(&session->s_cap_lock);
1221 if (session->s_nr_caps > 0) {
1222 struct inode *inode;
1223 struct ceph_cap *cap, *prev = NULL;
1224 struct ceph_vino vino;
1225 /*
1226 * iterate_session_caps() skips inodes that are being
1227 * deleted, we need to wait until deletions are complete.
1228 * __wait_on_freeing_inode() is designed for the job,
1229 * but it is not exported, so use lookup inode function
1230 * to access it.
1231 */
1232 while (!list_empty(&session->s_caps)) {
1233 cap = list_entry(session->s_caps.next,
1234 struct ceph_cap, session_caps);
1235 if (cap == prev)
1236 break;
1237 prev = cap;
1238 vino = cap->ci->i_vino;
1239 spin_unlock(&session->s_cap_lock);
1240
1241 inode = ceph_find_inode(sb, vino);
1242 iput(inode);
1243
1244 spin_lock(&session->s_cap_lock);
1245 }
1246 }
1247
1248 // drop cap expires and unlock s_cap_lock
1249 cleanup_cap_releases(session->s_mdsc, session);
1250
1251 BUG_ON(session->s_nr_caps > 0);
1252 BUG_ON(!list_empty(&session->s_cap_flushing));
1253}
1254
1255/*
1256 * wake up any threads waiting on this session's caps. if the cap is
1257 * old (didn't get renewed on the client reconnect), remove it now.
1258 *
1259 * caller must hold s_mutex.
1260 */
1261static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1262 void *arg)
1263{
1264 struct ceph_inode_info *ci = ceph_inode(inode);
1265
1266 if (arg) {
1267 spin_lock(&ci->i_ceph_lock);
1268 ci->i_wanted_max_size = 0;
1269 ci->i_requested_max_size = 0;
1270 spin_unlock(&ci->i_ceph_lock);
1271 }
1272 wake_up_all(&ci->i_cap_wq);
1273 return 0;
1274}
1275
1276static void wake_up_session_caps(struct ceph_mds_session *session,
1277 int reconnect)
1278{
1279 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1280 iterate_session_caps(session, wake_up_session_cb,
1281 (void *)(unsigned long)reconnect);
1282}
1283
1284/*
1285 * Send periodic message to MDS renewing all currently held caps. The
1286 * ack will reset the expiration for all caps from this session.
1287 *
1288 * caller holds s_mutex
1289 */
1290static int send_renew_caps(struct ceph_mds_client *mdsc,
1291 struct ceph_mds_session *session)
1292{
1293 struct ceph_msg *msg;
1294 int state;
1295
1296 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1297 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1298 pr_info("mds%d caps stale\n", session->s_mds);
1299 session->s_renew_requested = jiffies;
1300
1301 /* do not try to renew caps until a recovering mds has reconnected
1302 * with its clients. */
1303 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1304 if (state < CEPH_MDS_STATE_RECONNECT) {
1305 dout("send_renew_caps ignoring mds%d (%s)\n",
1306 session->s_mds, ceph_mds_state_name(state));
1307 return 0;
1308 }
1309
1310 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1311 ceph_mds_state_name(state));
1312 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1313 ++session->s_renew_seq);
1314 if (!msg)
1315 return -ENOMEM;
1316 ceph_con_send(&session->s_con, msg);
1317 return 0;
1318}
1319
1320static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1321 struct ceph_mds_session *session, u64 seq)
1322{
1323 struct ceph_msg *msg;
1324
1325 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1326 session->s_mds, ceph_session_state_name(session->s_state), seq);
1327 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1328 if (!msg)
1329 return -ENOMEM;
1330 ceph_con_send(&session->s_con, msg);
1331 return 0;
1332}
1333
1334
1335/*
1336 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1337 *
1338 * Called under session->s_mutex
1339 */
1340static void renewed_caps(struct ceph_mds_client *mdsc,
1341 struct ceph_mds_session *session, int is_renew)
1342{
1343 int was_stale;
1344 int wake = 0;
1345
1346 spin_lock(&session->s_cap_lock);
1347 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1348
1349 session->s_cap_ttl = session->s_renew_requested +
1350 mdsc->mdsmap->m_session_timeout*HZ;
1351
1352 if (was_stale) {
1353 if (time_before(jiffies, session->s_cap_ttl)) {
1354 pr_info("mds%d caps renewed\n", session->s_mds);
1355 wake = 1;
1356 } else {
1357 pr_info("mds%d caps still stale\n", session->s_mds);
1358 }
1359 }
1360 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1361 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1362 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1363 spin_unlock(&session->s_cap_lock);
1364
1365 if (wake)
1366 wake_up_session_caps(session, 0);
1367}
1368
1369/*
1370 * send a session close request
1371 */
1372static int request_close_session(struct ceph_mds_client *mdsc,
1373 struct ceph_mds_session *session)
1374{
1375 struct ceph_msg *msg;
1376
1377 dout("request_close_session mds%d state %s seq %lld\n",
1378 session->s_mds, ceph_session_state_name(session->s_state),
1379 session->s_seq);
1380 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1381 if (!msg)
1382 return -ENOMEM;
1383 ceph_con_send(&session->s_con, msg);
1384 return 1;
1385}
1386
1387/*
1388 * Called with s_mutex held.
1389 */
1390static int __close_session(struct ceph_mds_client *mdsc,
1391 struct ceph_mds_session *session)
1392{
1393 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1394 return 0;
1395 session->s_state = CEPH_MDS_SESSION_CLOSING;
1396 return request_close_session(mdsc, session);
1397}
1398
1399/*
1400 * Trim old(er) caps.
1401 *
1402 * Because we can't cache an inode without one or more caps, we do
1403 * this indirectly: if a cap is unused, we prune its aliases, at which
1404 * point the inode will hopefully get dropped to.
1405 *
1406 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1407 * memory pressure from the MDS, though, so it needn't be perfect.
1408 */
1409static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1410{
1411 struct ceph_mds_session *session = arg;
1412 struct ceph_inode_info *ci = ceph_inode(inode);
1413 int used, wanted, oissued, mine;
1414
1415 if (session->s_trim_caps <= 0)
1416 return -1;
1417
1418 spin_lock(&ci->i_ceph_lock);
1419 mine = cap->issued | cap->implemented;
1420 used = __ceph_caps_used(ci);
1421 wanted = __ceph_caps_file_wanted(ci);
1422 oissued = __ceph_caps_issued_other(ci, cap);
1423
1424 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1425 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1426 ceph_cap_string(used), ceph_cap_string(wanted));
1427 if (cap == ci->i_auth_cap) {
1428 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1429 !list_empty(&ci->i_cap_snaps))
1430 goto out;
1431 if ((used | wanted) & CEPH_CAP_ANY_WR)
1432 goto out;
1433 }
1434 /* The inode has cached pages, but it's no longer used.
1435 * we can safely drop it */
1436 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1437 !(oissued & CEPH_CAP_FILE_CACHE)) {
1438 used = 0;
1439 oissued = 0;
1440 }
1441 if ((used | wanted) & ~oissued & mine)
1442 goto out; /* we need these caps */
1443
1444 session->s_trim_caps--;
1445 if (oissued) {
1446 /* we aren't the only cap.. just remove us */
1447 __ceph_remove_cap(cap, true);
1448 } else {
1449 /* try dropping referring dentries */
1450 spin_unlock(&ci->i_ceph_lock);
1451 d_prune_aliases(inode);
1452 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1453 inode, cap, atomic_read(&inode->i_count));
1454 return 0;
1455 }
1456
1457out:
1458 spin_unlock(&ci->i_ceph_lock);
1459 return 0;
1460}
1461
1462/*
1463 * Trim session cap count down to some max number.
1464 */
1465static int trim_caps(struct ceph_mds_client *mdsc,
1466 struct ceph_mds_session *session,
1467 int max_caps)
1468{
1469 int trim_caps = session->s_nr_caps - max_caps;
1470
1471 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1472 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1473 if (trim_caps > 0) {
1474 session->s_trim_caps = trim_caps;
1475 iterate_session_caps(session, trim_caps_cb, session);
1476 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1477 session->s_mds, session->s_nr_caps, max_caps,
1478 trim_caps - session->s_trim_caps);
1479 session->s_trim_caps = 0;
1480 }
1481
1482 ceph_send_cap_releases(mdsc, session);
1483 return 0;
1484}
1485
1486static int check_caps_flush(struct ceph_mds_client *mdsc,
1487 u64 want_flush_tid)
1488{
1489 int ret = 1;
1490
1491 spin_lock(&mdsc->cap_dirty_lock);
1492 if (!list_empty(&mdsc->cap_flush_list)) {
1493 struct ceph_cap_flush *cf =
1494 list_first_entry(&mdsc->cap_flush_list,
1495 struct ceph_cap_flush, g_list);
1496 if (cf->tid <= want_flush_tid) {
1497 dout("check_caps_flush still flushing tid "
1498 "%llu <= %llu\n", cf->tid, want_flush_tid);
1499 ret = 0;
1500 }
1501 }
1502 spin_unlock(&mdsc->cap_dirty_lock);
1503 return ret;
1504}
1505
1506/*
1507 * flush all dirty inode data to disk.
1508 *
1509 * returns true if we've flushed through want_flush_tid
1510 */
1511static void wait_caps_flush(struct ceph_mds_client *mdsc,
1512 u64 want_flush_tid)
1513{
1514 dout("check_caps_flush want %llu\n", want_flush_tid);
1515
1516 wait_event(mdsc->cap_flushing_wq,
1517 check_caps_flush(mdsc, want_flush_tid));
1518
1519 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1520}
1521
1522/*
1523 * called under s_mutex
1524 */
1525void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1526 struct ceph_mds_session *session)
1527{
1528 struct ceph_msg *msg = NULL;
1529 struct ceph_mds_cap_release *head;
1530 struct ceph_mds_cap_item *item;
1531 struct ceph_cap *cap;
1532 LIST_HEAD(tmp_list);
1533 int num_cap_releases;
1534
1535 spin_lock(&session->s_cap_lock);
1536again:
1537 list_splice_init(&session->s_cap_releases, &tmp_list);
1538 num_cap_releases = session->s_num_cap_releases;
1539 session->s_num_cap_releases = 0;
1540 spin_unlock(&session->s_cap_lock);
1541
1542 while (!list_empty(&tmp_list)) {
1543 if (!msg) {
1544 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1545 PAGE_SIZE, GFP_NOFS, false);
1546 if (!msg)
1547 goto out_err;
1548 head = msg->front.iov_base;
1549 head->num = cpu_to_le32(0);
1550 msg->front.iov_len = sizeof(*head);
1551 }
1552 cap = list_first_entry(&tmp_list, struct ceph_cap,
1553 session_caps);
1554 list_del(&cap->session_caps);
1555 num_cap_releases--;
1556
1557 head = msg->front.iov_base;
1558 le32_add_cpu(&head->num, 1);
1559 item = msg->front.iov_base + msg->front.iov_len;
1560 item->ino = cpu_to_le64(cap->cap_ino);
1561 item->cap_id = cpu_to_le64(cap->cap_id);
1562 item->migrate_seq = cpu_to_le32(cap->mseq);
1563 item->seq = cpu_to_le32(cap->issue_seq);
1564 msg->front.iov_len += sizeof(*item);
1565
1566 ceph_put_cap(mdsc, cap);
1567
1568 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1569 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1570 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1571 ceph_con_send(&session->s_con, msg);
1572 msg = NULL;
1573 }
1574 }
1575
1576 BUG_ON(num_cap_releases != 0);
1577
1578 spin_lock(&session->s_cap_lock);
1579 if (!list_empty(&session->s_cap_releases))
1580 goto again;
1581 spin_unlock(&session->s_cap_lock);
1582
1583 if (msg) {
1584 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1585 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1586 ceph_con_send(&session->s_con, msg);
1587 }
1588 return;
1589out_err:
1590 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1591 session->s_mds);
1592 spin_lock(&session->s_cap_lock);
1593 list_splice(&tmp_list, &session->s_cap_releases);
1594 session->s_num_cap_releases += num_cap_releases;
1595 spin_unlock(&session->s_cap_lock);
1596}
1597
1598/*
1599 * requests
1600 */
1601
1602int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1603 struct inode *dir)
1604{
1605 struct ceph_inode_info *ci = ceph_inode(dir);
1606 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1607 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1608 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1609 int order, num_entries;
1610
1611 spin_lock(&ci->i_ceph_lock);
1612 num_entries = ci->i_files + ci->i_subdirs;
1613 spin_unlock(&ci->i_ceph_lock);
1614 num_entries = max(num_entries, 1);
1615 num_entries = min(num_entries, opt->max_readdir);
1616
1617 order = get_order(size * num_entries);
1618 while (order >= 0) {
1619 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1620 __GFP_NOWARN,
1621 order);
1622 if (rinfo->dir_entries)
1623 break;
1624 order--;
1625 }
1626 if (!rinfo->dir_entries)
1627 return -ENOMEM;
1628
1629 num_entries = (PAGE_SIZE << order) / size;
1630 num_entries = min(num_entries, opt->max_readdir);
1631
1632 rinfo->dir_buf_size = PAGE_SIZE << order;
1633 req->r_num_caps = num_entries + 1;
1634 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1635 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1636 return 0;
1637}
1638
1639/*
1640 * Create an mds request.
1641 */
1642struct ceph_mds_request *
1643ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1644{
1645 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1646
1647 if (!req)
1648 return ERR_PTR(-ENOMEM);
1649
1650 mutex_init(&req->r_fill_mutex);
1651 req->r_mdsc = mdsc;
1652 req->r_started = jiffies;
1653 req->r_resend_mds = -1;
1654 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1655 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1656 req->r_fmode = -1;
1657 kref_init(&req->r_kref);
1658 RB_CLEAR_NODE(&req->r_node);
1659 INIT_LIST_HEAD(&req->r_wait);
1660 init_completion(&req->r_completion);
1661 init_completion(&req->r_safe_completion);
1662 INIT_LIST_HEAD(&req->r_unsafe_item);
1663
1664 req->r_stamp = current_fs_time(mdsc->fsc->sb);
1665
1666 req->r_op = op;
1667 req->r_direct_mode = mode;
1668 return req;
1669}
1670
1671/*
1672 * return oldest (lowest) request, tid in request tree, 0 if none.
1673 *
1674 * called under mdsc->mutex.
1675 */
1676static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1677{
1678 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1679 return NULL;
1680 return rb_entry(rb_first(&mdsc->request_tree),
1681 struct ceph_mds_request, r_node);
1682}
1683
1684static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1685{
1686 return mdsc->oldest_tid;
1687}
1688
1689/*
1690 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1691 * on build_path_from_dentry in fs/cifs/dir.c.
1692 *
1693 * If @stop_on_nosnap, generate path relative to the first non-snapped
1694 * inode.
1695 *
1696 * Encode hidden .snap dirs as a double /, i.e.
1697 * foo/.snap/bar -> foo//bar
1698 */
1699char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1700 int stop_on_nosnap)
1701{
1702 struct dentry *temp;
1703 char *path;
1704 int len, pos;
1705 unsigned seq;
1706
1707 if (dentry == NULL)
1708 return ERR_PTR(-EINVAL);
1709
1710retry:
1711 len = 0;
1712 seq = read_seqbegin(&rename_lock);
1713 rcu_read_lock();
1714 for (temp = dentry; !IS_ROOT(temp);) {
1715 struct inode *inode = d_inode(temp);
1716 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1717 len++; /* slash only */
1718 else if (stop_on_nosnap && inode &&
1719 ceph_snap(inode) == CEPH_NOSNAP)
1720 break;
1721 else
1722 len += 1 + temp->d_name.len;
1723 temp = temp->d_parent;
1724 }
1725 rcu_read_unlock();
1726 if (len)
1727 len--; /* no leading '/' */
1728
1729 path = kmalloc(len+1, GFP_NOFS);
1730 if (path == NULL)
1731 return ERR_PTR(-ENOMEM);
1732 pos = len;
1733 path[pos] = 0; /* trailing null */
1734 rcu_read_lock();
1735 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1736 struct inode *inode;
1737
1738 spin_lock(&temp->d_lock);
1739 inode = d_inode(temp);
1740 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1741 dout("build_path path+%d: %p SNAPDIR\n",
1742 pos, temp);
1743 } else if (stop_on_nosnap && inode &&
1744 ceph_snap(inode) == CEPH_NOSNAP) {
1745 spin_unlock(&temp->d_lock);
1746 break;
1747 } else {
1748 pos -= temp->d_name.len;
1749 if (pos < 0) {
1750 spin_unlock(&temp->d_lock);
1751 break;
1752 }
1753 strncpy(path + pos, temp->d_name.name,
1754 temp->d_name.len);
1755 }
1756 spin_unlock(&temp->d_lock);
1757 if (pos)
1758 path[--pos] = '/';
1759 temp = temp->d_parent;
1760 }
1761 rcu_read_unlock();
1762 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1763 pr_err("build_path did not end path lookup where "
1764 "expected, namelen is %d, pos is %d\n", len, pos);
1765 /* presumably this is only possible if racing with a
1766 rename of one of the parent directories (we can not
1767 lock the dentries above us to prevent this, but
1768 retrying should be harmless) */
1769 kfree(path);
1770 goto retry;
1771 }
1772
1773 *base = ceph_ino(d_inode(temp));
1774 *plen = len;
1775 dout("build_path on %p %d built %llx '%.*s'\n",
1776 dentry, d_count(dentry), *base, len, path);
1777 return path;
1778}
1779
1780static int build_dentry_path(struct dentry *dentry,
1781 const char **ppath, int *ppathlen, u64 *pino,
1782 int *pfreepath)
1783{
1784 char *path;
1785
1786 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1787 *pino = ceph_ino(d_inode(dentry->d_parent));
1788 *ppath = dentry->d_name.name;
1789 *ppathlen = dentry->d_name.len;
1790 return 0;
1791 }
1792 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1793 if (IS_ERR(path))
1794 return PTR_ERR(path);
1795 *ppath = path;
1796 *pfreepath = 1;
1797 return 0;
1798}
1799
1800static int build_inode_path(struct inode *inode,
1801 const char **ppath, int *ppathlen, u64 *pino,
1802 int *pfreepath)
1803{
1804 struct dentry *dentry;
1805 char *path;
1806
1807 if (ceph_snap(inode) == CEPH_NOSNAP) {
1808 *pino = ceph_ino(inode);
1809 *ppathlen = 0;
1810 return 0;
1811 }
1812 dentry = d_find_alias(inode);
1813 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1814 dput(dentry);
1815 if (IS_ERR(path))
1816 return PTR_ERR(path);
1817 *ppath = path;
1818 *pfreepath = 1;
1819 return 0;
1820}
1821
1822/*
1823 * request arguments may be specified via an inode *, a dentry *, or
1824 * an explicit ino+path.
1825 */
1826static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1827 const char *rpath, u64 rino,
1828 const char **ppath, int *pathlen,
1829 u64 *ino, int *freepath)
1830{
1831 int r = 0;
1832
1833 if (rinode) {
1834 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1835 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1836 ceph_snap(rinode));
1837 } else if (rdentry) {
1838 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1839 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1840 *ppath);
1841 } else if (rpath || rino) {
1842 *ino = rino;
1843 *ppath = rpath;
1844 *pathlen = rpath ? strlen(rpath) : 0;
1845 dout(" path %.*s\n", *pathlen, rpath);
1846 }
1847
1848 return r;
1849}
1850
1851/*
1852 * called under mdsc->mutex
1853 */
1854static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1855 struct ceph_mds_request *req,
1856 int mds, bool drop_cap_releases)
1857{
1858 struct ceph_msg *msg;
1859 struct ceph_mds_request_head *head;
1860 const char *path1 = NULL;
1861 const char *path2 = NULL;
1862 u64 ino1 = 0, ino2 = 0;
1863 int pathlen1 = 0, pathlen2 = 0;
1864 int freepath1 = 0, freepath2 = 0;
1865 int len;
1866 u16 releases;
1867 void *p, *end;
1868 int ret;
1869
1870 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1871 req->r_path1, req->r_ino1.ino,
1872 &path1, &pathlen1, &ino1, &freepath1);
1873 if (ret < 0) {
1874 msg = ERR_PTR(ret);
1875 goto out;
1876 }
1877
1878 ret = set_request_path_attr(NULL, req->r_old_dentry,
1879 req->r_path2, req->r_ino2.ino,
1880 &path2, &pathlen2, &ino2, &freepath2);
1881 if (ret < 0) {
1882 msg = ERR_PTR(ret);
1883 goto out_free1;
1884 }
1885
1886 len = sizeof(*head) +
1887 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1888 sizeof(struct ceph_timespec);
1889
1890 /* calculate (max) length for cap releases */
1891 len += sizeof(struct ceph_mds_request_release) *
1892 (!!req->r_inode_drop + !!req->r_dentry_drop +
1893 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1894 if (req->r_dentry_drop)
1895 len += req->r_dentry->d_name.len;
1896 if (req->r_old_dentry_drop)
1897 len += req->r_old_dentry->d_name.len;
1898
1899 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1900 if (!msg) {
1901 msg = ERR_PTR(-ENOMEM);
1902 goto out_free2;
1903 }
1904
1905 msg->hdr.version = cpu_to_le16(2);
1906 msg->hdr.tid = cpu_to_le64(req->r_tid);
1907
1908 head = msg->front.iov_base;
1909 p = msg->front.iov_base + sizeof(*head);
1910 end = msg->front.iov_base + msg->front.iov_len;
1911
1912 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1913 head->op = cpu_to_le32(req->r_op);
1914 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1915 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1916 head->args = req->r_args;
1917
1918 ceph_encode_filepath(&p, end, ino1, path1);
1919 ceph_encode_filepath(&p, end, ino2, path2);
1920
1921 /* make note of release offset, in case we need to replay */
1922 req->r_request_release_offset = p - msg->front.iov_base;
1923
1924 /* cap releases */
1925 releases = 0;
1926 if (req->r_inode_drop)
1927 releases += ceph_encode_inode_release(&p,
1928 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1929 mds, req->r_inode_drop, req->r_inode_unless, 0);
1930 if (req->r_dentry_drop)
1931 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1932 mds, req->r_dentry_drop, req->r_dentry_unless);
1933 if (req->r_old_dentry_drop)
1934 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1935 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1936 if (req->r_old_inode_drop)
1937 releases += ceph_encode_inode_release(&p,
1938 d_inode(req->r_old_dentry),
1939 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1940
1941 if (drop_cap_releases) {
1942 releases = 0;
1943 p = msg->front.iov_base + req->r_request_release_offset;
1944 }
1945
1946 head->num_releases = cpu_to_le16(releases);
1947
1948 /* time stamp */
1949 {
1950 struct ceph_timespec ts;
1951 ceph_encode_timespec(&ts, &req->r_stamp);
1952 ceph_encode_copy(&p, &ts, sizeof(ts));
1953 }
1954
1955 BUG_ON(p > end);
1956 msg->front.iov_len = p - msg->front.iov_base;
1957 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1958
1959 if (req->r_pagelist) {
1960 struct ceph_pagelist *pagelist = req->r_pagelist;
1961 atomic_inc(&pagelist->refcnt);
1962 ceph_msg_data_add_pagelist(msg, pagelist);
1963 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1964 } else {
1965 msg->hdr.data_len = 0;
1966 }
1967
1968 msg->hdr.data_off = cpu_to_le16(0);
1969
1970out_free2:
1971 if (freepath2)
1972 kfree((char *)path2);
1973out_free1:
1974 if (freepath1)
1975 kfree((char *)path1);
1976out:
1977 return msg;
1978}
1979
1980/*
1981 * called under mdsc->mutex if error, under no mutex if
1982 * success.
1983 */
1984static void complete_request(struct ceph_mds_client *mdsc,
1985 struct ceph_mds_request *req)
1986{
1987 if (req->r_callback)
1988 req->r_callback(mdsc, req);
1989 else
1990 complete_all(&req->r_completion);
1991}
1992
1993/*
1994 * called under mdsc->mutex
1995 */
1996static int __prepare_send_request(struct ceph_mds_client *mdsc,
1997 struct ceph_mds_request *req,
1998 int mds, bool drop_cap_releases)
1999{
2000 struct ceph_mds_request_head *rhead;
2001 struct ceph_msg *msg;
2002 int flags = 0;
2003
2004 req->r_attempts++;
2005 if (req->r_inode) {
2006 struct ceph_cap *cap =
2007 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2008
2009 if (cap)
2010 req->r_sent_on_mseq = cap->mseq;
2011 else
2012 req->r_sent_on_mseq = -1;
2013 }
2014 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2015 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2016
2017 if (req->r_got_unsafe) {
2018 void *p;
2019 /*
2020 * Replay. Do not regenerate message (and rebuild
2021 * paths, etc.); just use the original message.
2022 * Rebuilding paths will break for renames because
2023 * d_move mangles the src name.
2024 */
2025 msg = req->r_request;
2026 rhead = msg->front.iov_base;
2027
2028 flags = le32_to_cpu(rhead->flags);
2029 flags |= CEPH_MDS_FLAG_REPLAY;
2030 rhead->flags = cpu_to_le32(flags);
2031
2032 if (req->r_target_inode)
2033 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2034
2035 rhead->num_retry = req->r_attempts - 1;
2036
2037 /* remove cap/dentry releases from message */
2038 rhead->num_releases = 0;
2039
2040 /* time stamp */
2041 p = msg->front.iov_base + req->r_request_release_offset;
2042 {
2043 struct ceph_timespec ts;
2044 ceph_encode_timespec(&ts, &req->r_stamp);
2045 ceph_encode_copy(&p, &ts, sizeof(ts));
2046 }
2047
2048 msg->front.iov_len = p - msg->front.iov_base;
2049 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2050 return 0;
2051 }
2052
2053 if (req->r_request) {
2054 ceph_msg_put(req->r_request);
2055 req->r_request = NULL;
2056 }
2057 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2058 if (IS_ERR(msg)) {
2059 req->r_err = PTR_ERR(msg);
2060 return PTR_ERR(msg);
2061 }
2062 req->r_request = msg;
2063
2064 rhead = msg->front.iov_base;
2065 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2066 if (req->r_got_unsafe)
2067 flags |= CEPH_MDS_FLAG_REPLAY;
2068 if (req->r_locked_dir)
2069 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2070 rhead->flags = cpu_to_le32(flags);
2071 rhead->num_fwd = req->r_num_fwd;
2072 rhead->num_retry = req->r_attempts - 1;
2073 rhead->ino = 0;
2074
2075 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2076 return 0;
2077}
2078
2079/*
2080 * send request, or put it on the appropriate wait list.
2081 */
2082static int __do_request(struct ceph_mds_client *mdsc,
2083 struct ceph_mds_request *req)
2084{
2085 struct ceph_mds_session *session = NULL;
2086 int mds = -1;
2087 int err = 0;
2088
2089 if (req->r_err || req->r_got_result) {
2090 if (req->r_aborted)
2091 __unregister_request(mdsc, req);
2092 goto out;
2093 }
2094
2095 if (req->r_timeout &&
2096 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2097 dout("do_request timed out\n");
2098 err = -EIO;
2099 goto finish;
2100 }
2101 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2102 dout("do_request forced umount\n");
2103 err = -EIO;
2104 goto finish;
2105 }
2106 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2107 if (mdsc->mdsmap_err) {
2108 err = mdsc->mdsmap_err;
2109 dout("do_request mdsmap err %d\n", err);
2110 goto finish;
2111 }
2112 if (mdsc->mdsmap->m_epoch == 0) {
2113 dout("do_request no mdsmap, waiting for map\n");
2114 list_add(&req->r_wait, &mdsc->waiting_for_map);
2115 goto finish;
2116 }
2117 if (!(mdsc->fsc->mount_options->flags &
2118 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2119 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2120 err = -ENOENT;
2121 pr_info("probably no mds server is up\n");
2122 goto finish;
2123 }
2124 }
2125
2126 put_request_session(req);
2127
2128 mds = __choose_mds(mdsc, req);
2129 if (mds < 0 ||
2130 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2131 dout("do_request no mds or not active, waiting for map\n");
2132 list_add(&req->r_wait, &mdsc->waiting_for_map);
2133 goto out;
2134 }
2135
2136 /* get, open session */
2137 session = __ceph_lookup_mds_session(mdsc, mds);
2138 if (!session) {
2139 session = register_session(mdsc, mds);
2140 if (IS_ERR(session)) {
2141 err = PTR_ERR(session);
2142 goto finish;
2143 }
2144 }
2145 req->r_session = get_session(session);
2146
2147 dout("do_request mds%d session %p state %s\n", mds, session,
2148 ceph_session_state_name(session->s_state));
2149 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2150 session->s_state != CEPH_MDS_SESSION_HUNG) {
2151 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2152 err = -EACCES;
2153 goto out_session;
2154 }
2155 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2156 session->s_state == CEPH_MDS_SESSION_CLOSING)
2157 __open_session(mdsc, session);
2158 list_add(&req->r_wait, &session->s_waiting);
2159 goto out_session;
2160 }
2161
2162 /* send request */
2163 req->r_resend_mds = -1; /* forget any previous mds hint */
2164
2165 if (req->r_request_started == 0) /* note request start time */
2166 req->r_request_started = jiffies;
2167
2168 err = __prepare_send_request(mdsc, req, mds, false);
2169 if (!err) {
2170 ceph_msg_get(req->r_request);
2171 ceph_con_send(&session->s_con, req->r_request);
2172 }
2173
2174out_session:
2175 ceph_put_mds_session(session);
2176finish:
2177 if (err) {
2178 dout("__do_request early error %d\n", err);
2179 req->r_err = err;
2180 complete_request(mdsc, req);
2181 __unregister_request(mdsc, req);
2182 }
2183out:
2184 return err;
2185}
2186
2187/*
2188 * called under mdsc->mutex
2189 */
2190static void __wake_requests(struct ceph_mds_client *mdsc,
2191 struct list_head *head)
2192{
2193 struct ceph_mds_request *req;
2194 LIST_HEAD(tmp_list);
2195
2196 list_splice_init(head, &tmp_list);
2197
2198 while (!list_empty(&tmp_list)) {
2199 req = list_entry(tmp_list.next,
2200 struct ceph_mds_request, r_wait);
2201 list_del_init(&req->r_wait);
2202 dout(" wake request %p tid %llu\n", req, req->r_tid);
2203 __do_request(mdsc, req);
2204 }
2205}
2206
2207/*
2208 * Wake up threads with requests pending for @mds, so that they can
2209 * resubmit their requests to a possibly different mds.
2210 */
2211static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2212{
2213 struct ceph_mds_request *req;
2214 struct rb_node *p = rb_first(&mdsc->request_tree);
2215
2216 dout("kick_requests mds%d\n", mds);
2217 while (p) {
2218 req = rb_entry(p, struct ceph_mds_request, r_node);
2219 p = rb_next(p);
2220 if (req->r_got_unsafe)
2221 continue;
2222 if (req->r_attempts > 0)
2223 continue; /* only new requests */
2224 if (req->r_session &&
2225 req->r_session->s_mds == mds) {
2226 dout(" kicking tid %llu\n", req->r_tid);
2227 list_del_init(&req->r_wait);
2228 __do_request(mdsc, req);
2229 }
2230 }
2231}
2232
2233void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2234 struct ceph_mds_request *req)
2235{
2236 dout("submit_request on %p\n", req);
2237 mutex_lock(&mdsc->mutex);
2238 __register_request(mdsc, req, NULL);
2239 __do_request(mdsc, req);
2240 mutex_unlock(&mdsc->mutex);
2241}
2242
2243/*
2244 * Synchrously perform an mds request. Take care of all of the
2245 * session setup, forwarding, retry details.
2246 */
2247int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2248 struct inode *dir,
2249 struct ceph_mds_request *req)
2250{
2251 int err;
2252
2253 dout("do_request on %p\n", req);
2254
2255 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2256 if (req->r_inode)
2257 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2258 if (req->r_locked_dir)
2259 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2260 if (req->r_old_dentry_dir)
2261 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2262 CEPH_CAP_PIN);
2263
2264 /* issue */
2265 mutex_lock(&mdsc->mutex);
2266 __register_request(mdsc, req, dir);
2267 __do_request(mdsc, req);
2268
2269 if (req->r_err) {
2270 err = req->r_err;
2271 goto out;
2272 }
2273
2274 /* wait */
2275 mutex_unlock(&mdsc->mutex);
2276 dout("do_request waiting\n");
2277 if (!req->r_timeout && req->r_wait_for_completion) {
2278 err = req->r_wait_for_completion(mdsc, req);
2279 } else {
2280 long timeleft = wait_for_completion_killable_timeout(
2281 &req->r_completion,
2282 ceph_timeout_jiffies(req->r_timeout));
2283 if (timeleft > 0)
2284 err = 0;
2285 else if (!timeleft)
2286 err = -EIO; /* timed out */
2287 else
2288 err = timeleft; /* killed */
2289 }
2290 dout("do_request waited, got %d\n", err);
2291 mutex_lock(&mdsc->mutex);
2292
2293 /* only abort if we didn't race with a real reply */
2294 if (req->r_got_result) {
2295 err = le32_to_cpu(req->r_reply_info.head->result);
2296 } else if (err < 0) {
2297 dout("aborted request %lld with %d\n", req->r_tid, err);
2298
2299 /*
2300 * ensure we aren't running concurrently with
2301 * ceph_fill_trace or ceph_readdir_prepopulate, which
2302 * rely on locks (dir mutex) held by our caller.
2303 */
2304 mutex_lock(&req->r_fill_mutex);
2305 req->r_err = err;
2306 req->r_aborted = true;
2307 mutex_unlock(&req->r_fill_mutex);
2308
2309 if (req->r_locked_dir &&
2310 (req->r_op & CEPH_MDS_OP_WRITE))
2311 ceph_invalidate_dir_request(req);
2312 } else {
2313 err = req->r_err;
2314 }
2315
2316out:
2317 mutex_unlock(&mdsc->mutex);
2318 dout("do_request %p done, result %d\n", req, err);
2319 return err;
2320}
2321
2322/*
2323 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2324 * namespace request.
2325 */
2326void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2327{
2328 struct inode *inode = req->r_locked_dir;
2329
2330 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2331
2332 ceph_dir_clear_complete(inode);
2333 if (req->r_dentry)
2334 ceph_invalidate_dentry_lease(req->r_dentry);
2335 if (req->r_old_dentry)
2336 ceph_invalidate_dentry_lease(req->r_old_dentry);
2337}
2338
2339/*
2340 * Handle mds reply.
2341 *
2342 * We take the session mutex and parse and process the reply immediately.
2343 * This preserves the logical ordering of replies, capabilities, etc., sent
2344 * by the MDS as they are applied to our local cache.
2345 */
2346static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2347{
2348 struct ceph_mds_client *mdsc = session->s_mdsc;
2349 struct ceph_mds_request *req;
2350 struct ceph_mds_reply_head *head = msg->front.iov_base;
2351 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2352 struct ceph_snap_realm *realm;
2353 u64 tid;
2354 int err, result;
2355 int mds = session->s_mds;
2356
2357 if (msg->front.iov_len < sizeof(*head)) {
2358 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2359 ceph_msg_dump(msg);
2360 return;
2361 }
2362
2363 /* get request, session */
2364 tid = le64_to_cpu(msg->hdr.tid);
2365 mutex_lock(&mdsc->mutex);
2366 req = lookup_get_request(mdsc, tid);
2367 if (!req) {
2368 dout("handle_reply on unknown tid %llu\n", tid);
2369 mutex_unlock(&mdsc->mutex);
2370 return;
2371 }
2372 dout("handle_reply %p\n", req);
2373
2374 /* correct session? */
2375 if (req->r_session != session) {
2376 pr_err("mdsc_handle_reply got %llu on session mds%d"
2377 " not mds%d\n", tid, session->s_mds,
2378 req->r_session ? req->r_session->s_mds : -1);
2379 mutex_unlock(&mdsc->mutex);
2380 goto out;
2381 }
2382
2383 /* dup? */
2384 if ((req->r_got_unsafe && !head->safe) ||
2385 (req->r_got_safe && head->safe)) {
2386 pr_warn("got a dup %s reply on %llu from mds%d\n",
2387 head->safe ? "safe" : "unsafe", tid, mds);
2388 mutex_unlock(&mdsc->mutex);
2389 goto out;
2390 }
2391 if (req->r_got_safe) {
2392 pr_warn("got unsafe after safe on %llu from mds%d\n",
2393 tid, mds);
2394 mutex_unlock(&mdsc->mutex);
2395 goto out;
2396 }
2397
2398 result = le32_to_cpu(head->result);
2399
2400 /*
2401 * Handle an ESTALE
2402 * if we're not talking to the authority, send to them
2403 * if the authority has changed while we weren't looking,
2404 * send to new authority
2405 * Otherwise we just have to return an ESTALE
2406 */
2407 if (result == -ESTALE) {
2408 dout("got ESTALE on request %llu", req->r_tid);
2409 req->r_resend_mds = -1;
2410 if (req->r_direct_mode != USE_AUTH_MDS) {
2411 dout("not using auth, setting for that now");
2412 req->r_direct_mode = USE_AUTH_MDS;
2413 __do_request(mdsc, req);
2414 mutex_unlock(&mdsc->mutex);
2415 goto out;
2416 } else {
2417 int mds = __choose_mds(mdsc, req);
2418 if (mds >= 0 && mds != req->r_session->s_mds) {
2419 dout("but auth changed, so resending");
2420 __do_request(mdsc, req);
2421 mutex_unlock(&mdsc->mutex);
2422 goto out;
2423 }
2424 }
2425 dout("have to return ESTALE on request %llu", req->r_tid);
2426 }
2427
2428
2429 if (head->safe) {
2430 req->r_got_safe = true;
2431 __unregister_request(mdsc, req);
2432
2433 if (req->r_got_unsafe) {
2434 /*
2435 * We already handled the unsafe response, now do the
2436 * cleanup. No need to examine the response; the MDS
2437 * doesn't include any result info in the safe
2438 * response. And even if it did, there is nothing
2439 * useful we could do with a revised return value.
2440 */
2441 dout("got safe reply %llu, mds%d\n", tid, mds);
2442
2443 /* last unsafe request during umount? */
2444 if (mdsc->stopping && !__get_oldest_req(mdsc))
2445 complete_all(&mdsc->safe_umount_waiters);
2446 mutex_unlock(&mdsc->mutex);
2447 goto out;
2448 }
2449 } else {
2450 req->r_got_unsafe = true;
2451 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2452 if (req->r_unsafe_dir) {
2453 struct ceph_inode_info *ci =
2454 ceph_inode(req->r_unsafe_dir);
2455 spin_lock(&ci->i_unsafe_lock);
2456 list_add_tail(&req->r_unsafe_dir_item,
2457 &ci->i_unsafe_dirops);
2458 spin_unlock(&ci->i_unsafe_lock);
2459 }
2460 }
2461
2462 dout("handle_reply tid %lld result %d\n", tid, result);
2463 rinfo = &req->r_reply_info;
2464 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2465 mutex_unlock(&mdsc->mutex);
2466
2467 mutex_lock(&session->s_mutex);
2468 if (err < 0) {
2469 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2470 ceph_msg_dump(msg);
2471 goto out_err;
2472 }
2473
2474 /* snap trace */
2475 realm = NULL;
2476 if (rinfo->snapblob_len) {
2477 down_write(&mdsc->snap_rwsem);
2478 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2479 rinfo->snapblob + rinfo->snapblob_len,
2480 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2481 &realm);
2482 downgrade_write(&mdsc->snap_rwsem);
2483 } else {
2484 down_read(&mdsc->snap_rwsem);
2485 }
2486
2487 /* insert trace into our cache */
2488 mutex_lock(&req->r_fill_mutex);
2489 current->journal_info = req;
2490 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2491 if (err == 0) {
2492 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2493 req->r_op == CEPH_MDS_OP_LSSNAP))
2494 ceph_readdir_prepopulate(req, req->r_session);
2495 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2496 }
2497 current->journal_info = NULL;
2498 mutex_unlock(&req->r_fill_mutex);
2499
2500 up_read(&mdsc->snap_rwsem);
2501 if (realm)
2502 ceph_put_snap_realm(mdsc, realm);
2503
2504 if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2505 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2506 spin_lock(&ci->i_unsafe_lock);
2507 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2508 spin_unlock(&ci->i_unsafe_lock);
2509 }
2510out_err:
2511 mutex_lock(&mdsc->mutex);
2512 if (!req->r_aborted) {
2513 if (err) {
2514 req->r_err = err;
2515 } else {
2516 req->r_reply = ceph_msg_get(msg);
2517 req->r_got_result = true;
2518 }
2519 } else {
2520 dout("reply arrived after request %lld was aborted\n", tid);
2521 }
2522 mutex_unlock(&mdsc->mutex);
2523
2524 mutex_unlock(&session->s_mutex);
2525
2526 /* kick calling process */
2527 complete_request(mdsc, req);
2528out:
2529 ceph_mdsc_put_request(req);
2530 return;
2531}
2532
2533
2534
2535/*
2536 * handle mds notification that our request has been forwarded.
2537 */
2538static void handle_forward(struct ceph_mds_client *mdsc,
2539 struct ceph_mds_session *session,
2540 struct ceph_msg *msg)
2541{
2542 struct ceph_mds_request *req;
2543 u64 tid = le64_to_cpu(msg->hdr.tid);
2544 u32 next_mds;
2545 u32 fwd_seq;
2546 int err = -EINVAL;
2547 void *p = msg->front.iov_base;
2548 void *end = p + msg->front.iov_len;
2549
2550 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2551 next_mds = ceph_decode_32(&p);
2552 fwd_seq = ceph_decode_32(&p);
2553
2554 mutex_lock(&mdsc->mutex);
2555 req = lookup_get_request(mdsc, tid);
2556 if (!req) {
2557 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2558 goto out; /* dup reply? */
2559 }
2560
2561 if (req->r_aborted) {
2562 dout("forward tid %llu aborted, unregistering\n", tid);
2563 __unregister_request(mdsc, req);
2564 } else if (fwd_seq <= req->r_num_fwd) {
2565 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2566 tid, next_mds, req->r_num_fwd, fwd_seq);
2567 } else {
2568 /* resend. forward race not possible; mds would drop */
2569 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2570 BUG_ON(req->r_err);
2571 BUG_ON(req->r_got_result);
2572 req->r_attempts = 0;
2573 req->r_num_fwd = fwd_seq;
2574 req->r_resend_mds = next_mds;
2575 put_request_session(req);
2576 __do_request(mdsc, req);
2577 }
2578 ceph_mdsc_put_request(req);
2579out:
2580 mutex_unlock(&mdsc->mutex);
2581 return;
2582
2583bad:
2584 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2585}
2586
2587/*
2588 * handle a mds session control message
2589 */
2590static void handle_session(struct ceph_mds_session *session,
2591 struct ceph_msg *msg)
2592{
2593 struct ceph_mds_client *mdsc = session->s_mdsc;
2594 u32 op;
2595 u64 seq;
2596 int mds = session->s_mds;
2597 struct ceph_mds_session_head *h = msg->front.iov_base;
2598 int wake = 0;
2599
2600 /* decode */
2601 if (msg->front.iov_len != sizeof(*h))
2602 goto bad;
2603 op = le32_to_cpu(h->op);
2604 seq = le64_to_cpu(h->seq);
2605
2606 mutex_lock(&mdsc->mutex);
2607 if (op == CEPH_SESSION_CLOSE)
2608 __unregister_session(mdsc, session);
2609 /* FIXME: this ttl calculation is generous */
2610 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2611 mutex_unlock(&mdsc->mutex);
2612
2613 mutex_lock(&session->s_mutex);
2614
2615 dout("handle_session mds%d %s %p state %s seq %llu\n",
2616 mds, ceph_session_op_name(op), session,
2617 ceph_session_state_name(session->s_state), seq);
2618
2619 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2620 session->s_state = CEPH_MDS_SESSION_OPEN;
2621 pr_info("mds%d came back\n", session->s_mds);
2622 }
2623
2624 switch (op) {
2625 case CEPH_SESSION_OPEN:
2626 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2627 pr_info("mds%d reconnect success\n", session->s_mds);
2628 session->s_state = CEPH_MDS_SESSION_OPEN;
2629 renewed_caps(mdsc, session, 0);
2630 wake = 1;
2631 if (mdsc->stopping)
2632 __close_session(mdsc, session);
2633 break;
2634
2635 case CEPH_SESSION_RENEWCAPS:
2636 if (session->s_renew_seq == seq)
2637 renewed_caps(mdsc, session, 1);
2638 break;
2639
2640 case CEPH_SESSION_CLOSE:
2641 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2642 pr_info("mds%d reconnect denied\n", session->s_mds);
2643 cleanup_session_requests(mdsc, session);
2644 remove_session_caps(session);
2645 wake = 2; /* for good measure */
2646 wake_up_all(&mdsc->session_close_wq);
2647 break;
2648
2649 case CEPH_SESSION_STALE:
2650 pr_info("mds%d caps went stale, renewing\n",
2651 session->s_mds);
2652 spin_lock(&session->s_gen_ttl_lock);
2653 session->s_cap_gen++;
2654 session->s_cap_ttl = jiffies - 1;
2655 spin_unlock(&session->s_gen_ttl_lock);
2656 send_renew_caps(mdsc, session);
2657 break;
2658
2659 case CEPH_SESSION_RECALL_STATE:
2660 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2661 break;
2662
2663 case CEPH_SESSION_FLUSHMSG:
2664 send_flushmsg_ack(mdsc, session, seq);
2665 break;
2666
2667 case CEPH_SESSION_FORCE_RO:
2668 dout("force_session_readonly %p\n", session);
2669 spin_lock(&session->s_cap_lock);
2670 session->s_readonly = true;
2671 spin_unlock(&session->s_cap_lock);
2672 wake_up_session_caps(session, 0);
2673 break;
2674
2675 case CEPH_SESSION_REJECT:
2676 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2677 pr_info("mds%d rejected session\n", session->s_mds);
2678 session->s_state = CEPH_MDS_SESSION_REJECTED;
2679 cleanup_session_requests(mdsc, session);
2680 remove_session_caps(session);
2681 wake = 2; /* for good measure */
2682 break;
2683
2684 default:
2685 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2686 WARN_ON(1);
2687 }
2688
2689 mutex_unlock(&session->s_mutex);
2690 if (wake) {
2691 mutex_lock(&mdsc->mutex);
2692 __wake_requests(mdsc, &session->s_waiting);
2693 if (wake == 2)
2694 kick_requests(mdsc, mds);
2695 mutex_unlock(&mdsc->mutex);
2696 }
2697 return;
2698
2699bad:
2700 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2701 (int)msg->front.iov_len);
2702 ceph_msg_dump(msg);
2703 return;
2704}
2705
2706
2707/*
2708 * called under session->mutex.
2709 */
2710static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2711 struct ceph_mds_session *session)
2712{
2713 struct ceph_mds_request *req, *nreq;
2714 struct rb_node *p;
2715 int err;
2716
2717 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2718
2719 mutex_lock(&mdsc->mutex);
2720 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2721 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2722 if (!err) {
2723 ceph_msg_get(req->r_request);
2724 ceph_con_send(&session->s_con, req->r_request);
2725 }
2726 }
2727
2728 /*
2729 * also re-send old requests when MDS enters reconnect stage. So that MDS
2730 * can process completed request in clientreplay stage.
2731 */
2732 p = rb_first(&mdsc->request_tree);
2733 while (p) {
2734 req = rb_entry(p, struct ceph_mds_request, r_node);
2735 p = rb_next(p);
2736 if (req->r_got_unsafe)
2737 continue;
2738 if (req->r_attempts == 0)
2739 continue; /* only old requests */
2740 if (req->r_session &&
2741 req->r_session->s_mds == session->s_mds) {
2742 err = __prepare_send_request(mdsc, req,
2743 session->s_mds, true);
2744 if (!err) {
2745 ceph_msg_get(req->r_request);
2746 ceph_con_send(&session->s_con, req->r_request);
2747 }
2748 }
2749 }
2750 mutex_unlock(&mdsc->mutex);
2751}
2752
2753/*
2754 * Encode information about a cap for a reconnect with the MDS.
2755 */
2756static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2757 void *arg)
2758{
2759 union {
2760 struct ceph_mds_cap_reconnect v2;
2761 struct ceph_mds_cap_reconnect_v1 v1;
2762 } rec;
2763 struct ceph_inode_info *ci;
2764 struct ceph_reconnect_state *recon_state = arg;
2765 struct ceph_pagelist *pagelist = recon_state->pagelist;
2766 char *path;
2767 int pathlen, err;
2768 u64 pathbase;
2769 u64 snap_follows;
2770 struct dentry *dentry;
2771
2772 ci = cap->ci;
2773
2774 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2775 inode, ceph_vinop(inode), cap, cap->cap_id,
2776 ceph_cap_string(cap->issued));
2777 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2778 if (err)
2779 return err;
2780
2781 dentry = d_find_alias(inode);
2782 if (dentry) {
2783 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2784 if (IS_ERR(path)) {
2785 err = PTR_ERR(path);
2786 goto out_dput;
2787 }
2788 } else {
2789 path = NULL;
2790 pathlen = 0;
2791 pathbase = 0;
2792 }
2793
2794 spin_lock(&ci->i_ceph_lock);
2795 cap->seq = 0; /* reset cap seq */
2796 cap->issue_seq = 0; /* and issue_seq */
2797 cap->mseq = 0; /* and migrate_seq */
2798 cap->cap_gen = cap->session->s_cap_gen;
2799
2800 if (recon_state->msg_version >= 2) {
2801 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2802 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2803 rec.v2.issued = cpu_to_le32(cap->issued);
2804 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2805 rec.v2.pathbase = cpu_to_le64(pathbase);
2806 rec.v2.flock_len = 0;
2807 } else {
2808 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2809 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2810 rec.v1.issued = cpu_to_le32(cap->issued);
2811 rec.v1.size = cpu_to_le64(inode->i_size);
2812 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2813 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2814 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2815 rec.v1.pathbase = cpu_to_le64(pathbase);
2816 }
2817
2818 if (list_empty(&ci->i_cap_snaps)) {
2819 snap_follows = 0;
2820 } else {
2821 struct ceph_cap_snap *capsnap =
2822 list_first_entry(&ci->i_cap_snaps,
2823 struct ceph_cap_snap, ci_item);
2824 snap_follows = capsnap->follows;
2825 }
2826 spin_unlock(&ci->i_ceph_lock);
2827
2828 if (recon_state->msg_version >= 2) {
2829 int num_fcntl_locks, num_flock_locks;
2830 struct ceph_filelock *flocks;
2831 size_t struct_len, total_len = 0;
2832 u8 struct_v = 0;
2833
2834encode_again:
2835 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2836 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2837 sizeof(struct ceph_filelock), GFP_NOFS);
2838 if (!flocks) {
2839 err = -ENOMEM;
2840 goto out_free;
2841 }
2842 err = ceph_encode_locks_to_buffer(inode, flocks,
2843 num_fcntl_locks,
2844 num_flock_locks);
2845 if (err) {
2846 kfree(flocks);
2847 if (err == -ENOSPC)
2848 goto encode_again;
2849 goto out_free;
2850 }
2851
2852 if (recon_state->msg_version >= 3) {
2853 /* version, compat_version and struct_len */
2854 total_len = 2 * sizeof(u8) + sizeof(u32);
2855 struct_v = 2;
2856 }
2857 /*
2858 * number of encoded locks is stable, so copy to pagelist
2859 */
2860 struct_len = 2 * sizeof(u32) +
2861 (num_fcntl_locks + num_flock_locks) *
2862 sizeof(struct ceph_filelock);
2863 rec.v2.flock_len = cpu_to_le32(struct_len);
2864
2865 struct_len += sizeof(rec.v2);
2866 struct_len += sizeof(u32) + pathlen;
2867
2868 if (struct_v >= 2)
2869 struct_len += sizeof(u64); /* snap_follows */
2870
2871 total_len += struct_len;
2872 err = ceph_pagelist_reserve(pagelist, total_len);
2873
2874 if (!err) {
2875 if (recon_state->msg_version >= 3) {
2876 ceph_pagelist_encode_8(pagelist, struct_v);
2877 ceph_pagelist_encode_8(pagelist, 1);
2878 ceph_pagelist_encode_32(pagelist, struct_len);
2879 }
2880 ceph_pagelist_encode_string(pagelist, path, pathlen);
2881 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2882 ceph_locks_to_pagelist(flocks, pagelist,
2883 num_fcntl_locks,
2884 num_flock_locks);
2885 if (struct_v >= 2)
2886 ceph_pagelist_encode_64(pagelist, snap_follows);
2887 }
2888 kfree(flocks);
2889 } else {
2890 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2891 err = ceph_pagelist_reserve(pagelist, size);
2892 if (!err) {
2893 ceph_pagelist_encode_string(pagelist, path, pathlen);
2894 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2895 }
2896 }
2897
2898 recon_state->nr_caps++;
2899out_free:
2900 kfree(path);
2901out_dput:
2902 dput(dentry);
2903 return err;
2904}
2905
2906
2907/*
2908 * If an MDS fails and recovers, clients need to reconnect in order to
2909 * reestablish shared state. This includes all caps issued through
2910 * this session _and_ the snap_realm hierarchy. Because it's not
2911 * clear which snap realms the mds cares about, we send everything we
2912 * know about.. that ensures we'll then get any new info the
2913 * recovering MDS might have.
2914 *
2915 * This is a relatively heavyweight operation, but it's rare.
2916 *
2917 * called with mdsc->mutex held.
2918 */
2919static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2920 struct ceph_mds_session *session)
2921{
2922 struct ceph_msg *reply;
2923 struct rb_node *p;
2924 int mds = session->s_mds;
2925 int err = -ENOMEM;
2926 int s_nr_caps;
2927 struct ceph_pagelist *pagelist;
2928 struct ceph_reconnect_state recon_state;
2929
2930 pr_info("mds%d reconnect start\n", mds);
2931
2932 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2933 if (!pagelist)
2934 goto fail_nopagelist;
2935 ceph_pagelist_init(pagelist);
2936
2937 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2938 if (!reply)
2939 goto fail_nomsg;
2940
2941 mutex_lock(&session->s_mutex);
2942 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2943 session->s_seq = 0;
2944
2945 dout("session %p state %s\n", session,
2946 ceph_session_state_name(session->s_state));
2947
2948 spin_lock(&session->s_gen_ttl_lock);
2949 session->s_cap_gen++;
2950 spin_unlock(&session->s_gen_ttl_lock);
2951
2952 spin_lock(&session->s_cap_lock);
2953 /* don't know if session is readonly */
2954 session->s_readonly = 0;
2955 /*
2956 * notify __ceph_remove_cap() that we are composing cap reconnect.
2957 * If a cap get released before being added to the cap reconnect,
2958 * __ceph_remove_cap() should skip queuing cap release.
2959 */
2960 session->s_cap_reconnect = 1;
2961 /* drop old cap expires; we're about to reestablish that state */
2962 cleanup_cap_releases(mdsc, session);
2963
2964 /* trim unused caps to reduce MDS's cache rejoin time */
2965 if (mdsc->fsc->sb->s_root)
2966 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2967
2968 ceph_con_close(&session->s_con);
2969 ceph_con_open(&session->s_con,
2970 CEPH_ENTITY_TYPE_MDS, mds,
2971 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2972
2973 /* replay unsafe requests */
2974 replay_unsafe_requests(mdsc, session);
2975
2976 down_read(&mdsc->snap_rwsem);
2977
2978 /* traverse this session's caps */
2979 s_nr_caps = session->s_nr_caps;
2980 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2981 if (err)
2982 goto fail;
2983
2984 recon_state.nr_caps = 0;
2985 recon_state.pagelist = pagelist;
2986 if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
2987 recon_state.msg_version = 3;
2988 else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
2989 recon_state.msg_version = 2;
2990 else
2991 recon_state.msg_version = 1;
2992 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2993 if (err < 0)
2994 goto fail;
2995
2996 spin_lock(&session->s_cap_lock);
2997 session->s_cap_reconnect = 0;
2998 spin_unlock(&session->s_cap_lock);
2999
3000 /*
3001 * snaprealms. we provide mds with the ino, seq (version), and
3002 * parent for all of our realms. If the mds has any newer info,
3003 * it will tell us.
3004 */
3005 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3006 struct ceph_snap_realm *realm =
3007 rb_entry(p, struct ceph_snap_realm, node);
3008 struct ceph_mds_snaprealm_reconnect sr_rec;
3009
3010 dout(" adding snap realm %llx seq %lld parent %llx\n",
3011 realm->ino, realm->seq, realm->parent_ino);
3012 sr_rec.ino = cpu_to_le64(realm->ino);
3013 sr_rec.seq = cpu_to_le64(realm->seq);
3014 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3015 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3016 if (err)
3017 goto fail;
3018 }
3019
3020 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3021
3022 /* raced with cap release? */
3023 if (s_nr_caps != recon_state.nr_caps) {
3024 struct page *page = list_first_entry(&pagelist->head,
3025 struct page, lru);
3026 __le32 *addr = kmap_atomic(page);
3027 *addr = cpu_to_le32(recon_state.nr_caps);
3028 kunmap_atomic(addr);
3029 }
3030
3031 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3032 ceph_msg_data_add_pagelist(reply, pagelist);
3033
3034 ceph_early_kick_flushing_caps(mdsc, session);
3035
3036 ceph_con_send(&session->s_con, reply);
3037
3038 mutex_unlock(&session->s_mutex);
3039
3040 mutex_lock(&mdsc->mutex);
3041 __wake_requests(mdsc, &session->s_waiting);
3042 mutex_unlock(&mdsc->mutex);
3043
3044 up_read(&mdsc->snap_rwsem);
3045 return;
3046
3047fail:
3048 ceph_msg_put(reply);
3049 up_read(&mdsc->snap_rwsem);
3050 mutex_unlock(&session->s_mutex);
3051fail_nomsg:
3052 ceph_pagelist_release(pagelist);
3053fail_nopagelist:
3054 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3055 return;
3056}
3057
3058
3059/*
3060 * compare old and new mdsmaps, kicking requests
3061 * and closing out old connections as necessary
3062 *
3063 * called under mdsc->mutex.
3064 */
3065static void check_new_map(struct ceph_mds_client *mdsc,
3066 struct ceph_mdsmap *newmap,
3067 struct ceph_mdsmap *oldmap)
3068{
3069 int i;
3070 int oldstate, newstate;
3071 struct ceph_mds_session *s;
3072
3073 dout("check_new_map new %u old %u\n",
3074 newmap->m_epoch, oldmap->m_epoch);
3075
3076 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3077 if (mdsc->sessions[i] == NULL)
3078 continue;
3079 s = mdsc->sessions[i];
3080 oldstate = ceph_mdsmap_get_state(oldmap, i);
3081 newstate = ceph_mdsmap_get_state(newmap, i);
3082
3083 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3084 i, ceph_mds_state_name(oldstate),
3085 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3086 ceph_mds_state_name(newstate),
3087 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3088 ceph_session_state_name(s->s_state));
3089
3090 if (i >= newmap->m_max_mds ||
3091 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3092 ceph_mdsmap_get_addr(newmap, i),
3093 sizeof(struct ceph_entity_addr))) {
3094 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3095 /* the session never opened, just close it
3096 * out now */
3097 __wake_requests(mdsc, &s->s_waiting);
3098 __unregister_session(mdsc, s);
3099 } else {
3100 /* just close it */
3101 mutex_unlock(&mdsc->mutex);
3102 mutex_lock(&s->s_mutex);
3103 mutex_lock(&mdsc->mutex);
3104 ceph_con_close(&s->s_con);
3105 mutex_unlock(&s->s_mutex);
3106 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3107 }
3108 } else if (oldstate == newstate) {
3109 continue; /* nothing new with this mds */
3110 }
3111
3112 /*
3113 * send reconnect?
3114 */
3115 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3116 newstate >= CEPH_MDS_STATE_RECONNECT) {
3117 mutex_unlock(&mdsc->mutex);
3118 send_mds_reconnect(mdsc, s);
3119 mutex_lock(&mdsc->mutex);
3120 }
3121
3122 /*
3123 * kick request on any mds that has gone active.
3124 */
3125 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3126 newstate >= CEPH_MDS_STATE_ACTIVE) {
3127 if (oldstate != CEPH_MDS_STATE_CREATING &&
3128 oldstate != CEPH_MDS_STATE_STARTING)
3129 pr_info("mds%d recovery completed\n", s->s_mds);
3130 kick_requests(mdsc, i);
3131 ceph_kick_flushing_caps(mdsc, s);
3132 wake_up_session_caps(s, 1);
3133 }
3134 }
3135
3136 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3137 s = mdsc->sessions[i];
3138 if (!s)
3139 continue;
3140 if (!ceph_mdsmap_is_laggy(newmap, i))
3141 continue;
3142 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3143 s->s_state == CEPH_MDS_SESSION_HUNG ||
3144 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3145 dout(" connecting to export targets of laggy mds%d\n",
3146 i);
3147 __open_export_target_sessions(mdsc, s);
3148 }
3149 }
3150}
3151
3152
3153
3154/*
3155 * leases
3156 */
3157
3158/*
3159 * caller must hold session s_mutex, dentry->d_lock
3160 */
3161void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3162{
3163 struct ceph_dentry_info *di = ceph_dentry(dentry);
3164
3165 ceph_put_mds_session(di->lease_session);
3166 di->lease_session = NULL;
3167}
3168
3169static void handle_lease(struct ceph_mds_client *mdsc,
3170 struct ceph_mds_session *session,
3171 struct ceph_msg *msg)
3172{
3173 struct super_block *sb = mdsc->fsc->sb;
3174 struct inode *inode;
3175 struct dentry *parent, *dentry;
3176 struct ceph_dentry_info *di;
3177 int mds = session->s_mds;
3178 struct ceph_mds_lease *h = msg->front.iov_base;
3179 u32 seq;
3180 struct ceph_vino vino;
3181 struct qstr dname;
3182 int release = 0;
3183
3184 dout("handle_lease from mds%d\n", mds);
3185
3186 /* decode */
3187 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3188 goto bad;
3189 vino.ino = le64_to_cpu(h->ino);
3190 vino.snap = CEPH_NOSNAP;
3191 seq = le32_to_cpu(h->seq);
3192 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3193 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3194 if (dname.len != get_unaligned_le32(h+1))
3195 goto bad;
3196
3197 /* lookup inode */
3198 inode = ceph_find_inode(sb, vino);
3199 dout("handle_lease %s, ino %llx %p %.*s\n",
3200 ceph_lease_op_name(h->action), vino.ino, inode,
3201 dname.len, dname.name);
3202
3203 mutex_lock(&session->s_mutex);
3204 session->s_seq++;
3205
3206 if (inode == NULL) {
3207 dout("handle_lease no inode %llx\n", vino.ino);
3208 goto release;
3209 }
3210
3211 /* dentry */
3212 parent = d_find_alias(inode);
3213 if (!parent) {
3214 dout("no parent dentry on inode %p\n", inode);
3215 WARN_ON(1);
3216 goto release; /* hrm... */
3217 }
3218 dname.hash = full_name_hash(parent, dname.name, dname.len);
3219 dentry = d_lookup(parent, &dname);
3220 dput(parent);
3221 if (!dentry)
3222 goto release;
3223
3224 spin_lock(&dentry->d_lock);
3225 di = ceph_dentry(dentry);
3226 switch (h->action) {
3227 case CEPH_MDS_LEASE_REVOKE:
3228 if (di->lease_session == session) {
3229 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3230 h->seq = cpu_to_le32(di->lease_seq);
3231 __ceph_mdsc_drop_dentry_lease(dentry);
3232 }
3233 release = 1;
3234 break;
3235
3236 case CEPH_MDS_LEASE_RENEW:
3237 if (di->lease_session == session &&
3238 di->lease_gen == session->s_cap_gen &&
3239 di->lease_renew_from &&
3240 di->lease_renew_after == 0) {
3241 unsigned long duration =
3242 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3243
3244 di->lease_seq = seq;
3245 di->time = di->lease_renew_from + duration;
3246 di->lease_renew_after = di->lease_renew_from +
3247 (duration >> 1);
3248 di->lease_renew_from = 0;
3249 }
3250 break;
3251 }
3252 spin_unlock(&dentry->d_lock);
3253 dput(dentry);
3254
3255 if (!release)
3256 goto out;
3257
3258release:
3259 /* let's just reuse the same message */
3260 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3261 ceph_msg_get(msg);
3262 ceph_con_send(&session->s_con, msg);
3263
3264out:
3265 iput(inode);
3266 mutex_unlock(&session->s_mutex);
3267 return;
3268
3269bad:
3270 pr_err("corrupt lease message\n");
3271 ceph_msg_dump(msg);
3272}
3273
3274void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3275 struct inode *inode,
3276 struct dentry *dentry, char action,
3277 u32 seq)
3278{
3279 struct ceph_msg *msg;
3280 struct ceph_mds_lease *lease;
3281 int len = sizeof(*lease) + sizeof(u32);
3282 int dnamelen = 0;
3283
3284 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3285 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3286 dnamelen = dentry->d_name.len;
3287 len += dnamelen;
3288
3289 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3290 if (!msg)
3291 return;
3292 lease = msg->front.iov_base;
3293 lease->action = action;
3294 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3295 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3296 lease->seq = cpu_to_le32(seq);
3297 put_unaligned_le32(dnamelen, lease + 1);
3298 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3299
3300 /*
3301 * if this is a preemptive lease RELEASE, no need to
3302 * flush request stream, since the actual request will
3303 * soon follow.
3304 */
3305 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3306
3307 ceph_con_send(&session->s_con, msg);
3308}
3309
3310/*
3311 * drop all leases (and dentry refs) in preparation for umount
3312 */
3313static void drop_leases(struct ceph_mds_client *mdsc)
3314{
3315 int i;
3316
3317 dout("drop_leases\n");
3318 mutex_lock(&mdsc->mutex);
3319 for (i = 0; i < mdsc->max_sessions; i++) {
3320 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3321 if (!s)
3322 continue;
3323 mutex_unlock(&mdsc->mutex);
3324 mutex_lock(&s->s_mutex);
3325 mutex_unlock(&s->s_mutex);
3326 ceph_put_mds_session(s);
3327 mutex_lock(&mdsc->mutex);
3328 }
3329 mutex_unlock(&mdsc->mutex);
3330}
3331
3332
3333
3334/*
3335 * delayed work -- periodically trim expired leases, renew caps with mds
3336 */
3337static void schedule_delayed(struct ceph_mds_client *mdsc)
3338{
3339 int delay = 5;
3340 unsigned hz = round_jiffies_relative(HZ * delay);
3341 schedule_delayed_work(&mdsc->delayed_work, hz);
3342}
3343
3344static void delayed_work(struct work_struct *work)
3345{
3346 int i;
3347 struct ceph_mds_client *mdsc =
3348 container_of(work, struct ceph_mds_client, delayed_work.work);
3349 int renew_interval;
3350 int renew_caps;
3351
3352 dout("mdsc delayed_work\n");
3353 ceph_check_delayed_caps(mdsc);
3354
3355 mutex_lock(&mdsc->mutex);
3356 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3357 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3358 mdsc->last_renew_caps);
3359 if (renew_caps)
3360 mdsc->last_renew_caps = jiffies;
3361
3362 for (i = 0; i < mdsc->max_sessions; i++) {
3363 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3364 if (s == NULL)
3365 continue;
3366 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3367 dout("resending session close request for mds%d\n",
3368 s->s_mds);
3369 request_close_session(mdsc, s);
3370 ceph_put_mds_session(s);
3371 continue;
3372 }
3373 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3374 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3375 s->s_state = CEPH_MDS_SESSION_HUNG;
3376 pr_info("mds%d hung\n", s->s_mds);
3377 }
3378 }
3379 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3380 /* this mds is failed or recovering, just wait */
3381 ceph_put_mds_session(s);
3382 continue;
3383 }
3384 mutex_unlock(&mdsc->mutex);
3385
3386 mutex_lock(&s->s_mutex);
3387 if (renew_caps)
3388 send_renew_caps(mdsc, s);
3389 else
3390 ceph_con_keepalive(&s->s_con);
3391 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3392 s->s_state == CEPH_MDS_SESSION_HUNG)
3393 ceph_send_cap_releases(mdsc, s);
3394 mutex_unlock(&s->s_mutex);
3395 ceph_put_mds_session(s);
3396
3397 mutex_lock(&mdsc->mutex);
3398 }
3399 mutex_unlock(&mdsc->mutex);
3400
3401 schedule_delayed(mdsc);
3402}
3403
3404int ceph_mdsc_init(struct ceph_fs_client *fsc)
3405
3406{
3407 struct ceph_mds_client *mdsc;
3408
3409 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3410 if (!mdsc)
3411 return -ENOMEM;
3412 mdsc->fsc = fsc;
3413 fsc->mdsc = mdsc;
3414 mutex_init(&mdsc->mutex);
3415 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3416 if (mdsc->mdsmap == NULL) {
3417 kfree(mdsc);
3418 return -ENOMEM;
3419 }
3420
3421 init_completion(&mdsc->safe_umount_waiters);
3422 init_waitqueue_head(&mdsc->session_close_wq);
3423 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3424 mdsc->sessions = NULL;
3425 atomic_set(&mdsc->num_sessions, 0);
3426 mdsc->max_sessions = 0;
3427 mdsc->stopping = 0;
3428 mdsc->last_snap_seq = 0;
3429 init_rwsem(&mdsc->snap_rwsem);
3430 mdsc->snap_realms = RB_ROOT;
3431 INIT_LIST_HEAD(&mdsc->snap_empty);
3432 spin_lock_init(&mdsc->snap_empty_lock);
3433 mdsc->last_tid = 0;
3434 mdsc->oldest_tid = 0;
3435 mdsc->request_tree = RB_ROOT;
3436 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3437 mdsc->last_renew_caps = jiffies;
3438 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3439 spin_lock_init(&mdsc->cap_delay_lock);
3440 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3441 spin_lock_init(&mdsc->snap_flush_lock);
3442 mdsc->last_cap_flush_tid = 1;
3443 INIT_LIST_HEAD(&mdsc->cap_flush_list);
3444 INIT_LIST_HEAD(&mdsc->cap_dirty);
3445 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3446 mdsc->num_cap_flushing = 0;
3447 spin_lock_init(&mdsc->cap_dirty_lock);
3448 init_waitqueue_head(&mdsc->cap_flushing_wq);
3449 spin_lock_init(&mdsc->dentry_lru_lock);
3450 INIT_LIST_HEAD(&mdsc->dentry_lru);
3451
3452 ceph_caps_init(mdsc);
3453 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3454
3455 init_rwsem(&mdsc->pool_perm_rwsem);
3456 mdsc->pool_perm_tree = RB_ROOT;
3457
3458 return 0;
3459}
3460
3461/*
3462 * Wait for safe replies on open mds requests. If we time out, drop
3463 * all requests from the tree to avoid dangling dentry refs.
3464 */
3465static void wait_requests(struct ceph_mds_client *mdsc)
3466{
3467 struct ceph_options *opts = mdsc->fsc->client->options;
3468 struct ceph_mds_request *req;
3469
3470 mutex_lock(&mdsc->mutex);
3471 if (__get_oldest_req(mdsc)) {
3472 mutex_unlock(&mdsc->mutex);
3473
3474 dout("wait_requests waiting for requests\n");
3475 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3476 ceph_timeout_jiffies(opts->mount_timeout));
3477
3478 /* tear down remaining requests */
3479 mutex_lock(&mdsc->mutex);
3480 while ((req = __get_oldest_req(mdsc))) {
3481 dout("wait_requests timed out on tid %llu\n",
3482 req->r_tid);
3483 __unregister_request(mdsc, req);
3484 }
3485 }
3486 mutex_unlock(&mdsc->mutex);
3487 dout("wait_requests done\n");
3488}
3489
3490/*
3491 * called before mount is ro, and before dentries are torn down.
3492 * (hmm, does this still race with new lookups?)
3493 */
3494void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3495{
3496 dout("pre_umount\n");
3497 mdsc->stopping = 1;
3498
3499 drop_leases(mdsc);
3500 ceph_flush_dirty_caps(mdsc);
3501 wait_requests(mdsc);
3502
3503 /*
3504 * wait for reply handlers to drop their request refs and
3505 * their inode/dcache refs
3506 */
3507 ceph_msgr_flush();
3508}
3509
3510/*
3511 * wait for all write mds requests to flush.
3512 */
3513static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3514{
3515 struct ceph_mds_request *req = NULL, *nextreq;
3516 struct rb_node *n;
3517
3518 mutex_lock(&mdsc->mutex);
3519 dout("wait_unsafe_requests want %lld\n", want_tid);
3520restart:
3521 req = __get_oldest_req(mdsc);
3522 while (req && req->r_tid <= want_tid) {
3523 /* find next request */
3524 n = rb_next(&req->r_node);
3525 if (n)
3526 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3527 else
3528 nextreq = NULL;
3529 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3530 (req->r_op & CEPH_MDS_OP_WRITE)) {
3531 /* write op */
3532 ceph_mdsc_get_request(req);
3533 if (nextreq)
3534 ceph_mdsc_get_request(nextreq);
3535 mutex_unlock(&mdsc->mutex);
3536 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3537 req->r_tid, want_tid);
3538 wait_for_completion(&req->r_safe_completion);
3539 mutex_lock(&mdsc->mutex);
3540 ceph_mdsc_put_request(req);
3541 if (!nextreq)
3542 break; /* next dne before, so we're done! */
3543 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3544 /* next request was removed from tree */
3545 ceph_mdsc_put_request(nextreq);
3546 goto restart;
3547 }
3548 ceph_mdsc_put_request(nextreq); /* won't go away */
3549 }
3550 req = nextreq;
3551 }
3552 mutex_unlock(&mdsc->mutex);
3553 dout("wait_unsafe_requests done\n");
3554}
3555
3556void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3557{
3558 u64 want_tid, want_flush;
3559
3560 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3561 return;
3562
3563 dout("sync\n");
3564 mutex_lock(&mdsc->mutex);
3565 want_tid = mdsc->last_tid;
3566 mutex_unlock(&mdsc->mutex);
3567
3568 ceph_flush_dirty_caps(mdsc);
3569 spin_lock(&mdsc->cap_dirty_lock);
3570 want_flush = mdsc->last_cap_flush_tid;
3571 if (!list_empty(&mdsc->cap_flush_list)) {
3572 struct ceph_cap_flush *cf =
3573 list_last_entry(&mdsc->cap_flush_list,
3574 struct ceph_cap_flush, g_list);
3575 cf->wake = true;
3576 }
3577 spin_unlock(&mdsc->cap_dirty_lock);
3578
3579 dout("sync want tid %lld flush_seq %lld\n",
3580 want_tid, want_flush);
3581
3582 wait_unsafe_requests(mdsc, want_tid);
3583 wait_caps_flush(mdsc, want_flush);
3584}
3585
3586/*
3587 * true if all sessions are closed, or we force unmount
3588 */
3589static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3590{
3591 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3592 return true;
3593 return atomic_read(&mdsc->num_sessions) <= skipped;
3594}
3595
3596/*
3597 * called after sb is ro.
3598 */
3599void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3600{
3601 struct ceph_options *opts = mdsc->fsc->client->options;
3602 struct ceph_mds_session *session;
3603 int i;
3604 int skipped = 0;
3605
3606 dout("close_sessions\n");
3607
3608 /* close sessions */
3609 mutex_lock(&mdsc->mutex);
3610 for (i = 0; i < mdsc->max_sessions; i++) {
3611 session = __ceph_lookup_mds_session(mdsc, i);
3612 if (!session)
3613 continue;
3614 mutex_unlock(&mdsc->mutex);
3615 mutex_lock(&session->s_mutex);
3616 if (__close_session(mdsc, session) <= 0)
3617 skipped++;
3618 mutex_unlock(&session->s_mutex);
3619 ceph_put_mds_session(session);
3620 mutex_lock(&mdsc->mutex);
3621 }
3622 mutex_unlock(&mdsc->mutex);
3623
3624 dout("waiting for sessions to close\n");
3625 wait_event_timeout(mdsc->session_close_wq,
3626 done_closing_sessions(mdsc, skipped),
3627 ceph_timeout_jiffies(opts->mount_timeout));
3628
3629 /* tear down remaining sessions */
3630 mutex_lock(&mdsc->mutex);
3631 for (i = 0; i < mdsc->max_sessions; i++) {
3632 if (mdsc->sessions[i]) {
3633 session = get_session(mdsc->sessions[i]);
3634 __unregister_session(mdsc, session);
3635 mutex_unlock(&mdsc->mutex);
3636 mutex_lock(&session->s_mutex);
3637 remove_session_caps(session);
3638 mutex_unlock(&session->s_mutex);
3639 ceph_put_mds_session(session);
3640 mutex_lock(&mdsc->mutex);
3641 }
3642 }
3643 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3644 mutex_unlock(&mdsc->mutex);
3645
3646 ceph_cleanup_empty_realms(mdsc);
3647
3648 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3649
3650 dout("stopped\n");
3651}
3652
3653void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3654{
3655 struct ceph_mds_session *session;
3656 int mds;
3657
3658 dout("force umount\n");
3659
3660 mutex_lock(&mdsc->mutex);
3661 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3662 session = __ceph_lookup_mds_session(mdsc, mds);
3663 if (!session)
3664 continue;
3665 mutex_unlock(&mdsc->mutex);
3666 mutex_lock(&session->s_mutex);
3667 __close_session(mdsc, session);
3668 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3669 cleanup_session_requests(mdsc, session);
3670 remove_session_caps(session);
3671 }
3672 mutex_unlock(&session->s_mutex);
3673 ceph_put_mds_session(session);
3674 mutex_lock(&mdsc->mutex);
3675 kick_requests(mdsc, mds);
3676 }
3677 __wake_requests(mdsc, &mdsc->waiting_for_map);
3678 mutex_unlock(&mdsc->mutex);
3679}
3680
3681static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3682{
3683 dout("stop\n");
3684 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3685 if (mdsc->mdsmap)
3686 ceph_mdsmap_destroy(mdsc->mdsmap);
3687 kfree(mdsc->sessions);
3688 ceph_caps_finalize(mdsc);
3689 ceph_pool_perm_destroy(mdsc);
3690}
3691
3692void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3693{
3694 struct ceph_mds_client *mdsc = fsc->mdsc;
3695
3696 dout("mdsc_destroy %p\n", mdsc);
3697 ceph_mdsc_stop(mdsc);
3698
3699 /* flush out any connection work with references to us */
3700 ceph_msgr_flush();
3701
3702 fsc->mdsc = NULL;
3703 kfree(mdsc);
3704 dout("mdsc_destroy %p done\n", mdsc);
3705}
3706
3707void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3708{
3709 struct ceph_fs_client *fsc = mdsc->fsc;
3710 const char *mds_namespace = fsc->mount_options->mds_namespace;
3711 void *p = msg->front.iov_base;
3712 void *end = p + msg->front.iov_len;
3713 u32 epoch;
3714 u32 map_len;
3715 u32 num_fs;
3716 u32 mount_fscid = (u32)-1;
3717 u8 struct_v, struct_cv;
3718 int err = -EINVAL;
3719
3720 ceph_decode_need(&p, end, sizeof(u32), bad);
3721 epoch = ceph_decode_32(&p);
3722
3723 dout("handle_fsmap epoch %u\n", epoch);
3724
3725 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3726 struct_v = ceph_decode_8(&p);
3727 struct_cv = ceph_decode_8(&p);
3728 map_len = ceph_decode_32(&p);
3729
3730 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3731 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3732
3733 num_fs = ceph_decode_32(&p);
3734 while (num_fs-- > 0) {
3735 void *info_p, *info_end;
3736 u32 info_len;
3737 u8 info_v, info_cv;
3738 u32 fscid, namelen;
3739
3740 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3741 info_v = ceph_decode_8(&p);
3742 info_cv = ceph_decode_8(&p);
3743 info_len = ceph_decode_32(&p);
3744 ceph_decode_need(&p, end, info_len, bad);
3745 info_p = p;
3746 info_end = p + info_len;
3747 p = info_end;
3748
3749 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3750 fscid = ceph_decode_32(&info_p);
3751 namelen = ceph_decode_32(&info_p);
3752 ceph_decode_need(&info_p, info_end, namelen, bad);
3753
3754 if (mds_namespace &&
3755 strlen(mds_namespace) == namelen &&
3756 !strncmp(mds_namespace, (char *)info_p, namelen)) {
3757 mount_fscid = fscid;
3758 break;
3759 }
3760 }
3761
3762 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3763 if (mount_fscid != (u32)-1) {
3764 fsc->client->monc.fs_cluster_id = mount_fscid;
3765 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3766 0, true);
3767 ceph_monc_renew_subs(&fsc->client->monc);
3768 } else {
3769 err = -ENOENT;
3770 goto err_out;
3771 }
3772 return;
3773bad:
3774 pr_err("error decoding fsmap\n");
3775err_out:
3776 mutex_lock(&mdsc->mutex);
3777 mdsc->mdsmap_err = -ENOENT;
3778 __wake_requests(mdsc, &mdsc->waiting_for_map);
3779 mutex_unlock(&mdsc->mutex);
3780 return;
3781}
3782
3783/*
3784 * handle mds map update.
3785 */
3786void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3787{
3788 u32 epoch;
3789 u32 maplen;
3790 void *p = msg->front.iov_base;
3791 void *end = p + msg->front.iov_len;
3792 struct ceph_mdsmap *newmap, *oldmap;
3793 struct ceph_fsid fsid;
3794 int err = -EINVAL;
3795
3796 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3797 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3798 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3799 return;
3800 epoch = ceph_decode_32(&p);
3801 maplen = ceph_decode_32(&p);
3802 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3803
3804 /* do we need it? */
3805 mutex_lock(&mdsc->mutex);
3806 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3807 dout("handle_map epoch %u <= our %u\n",
3808 epoch, mdsc->mdsmap->m_epoch);
3809 mutex_unlock(&mdsc->mutex);
3810 return;
3811 }
3812
3813 newmap = ceph_mdsmap_decode(&p, end);
3814 if (IS_ERR(newmap)) {
3815 err = PTR_ERR(newmap);
3816 goto bad_unlock;
3817 }
3818
3819 /* swap into place */
3820 if (mdsc->mdsmap) {
3821 oldmap = mdsc->mdsmap;
3822 mdsc->mdsmap = newmap;
3823 check_new_map(mdsc, newmap, oldmap);
3824 ceph_mdsmap_destroy(oldmap);
3825 } else {
3826 mdsc->mdsmap = newmap; /* first mds map */
3827 }
3828 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3829
3830 __wake_requests(mdsc, &mdsc->waiting_for_map);
3831 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3832 mdsc->mdsmap->m_epoch);
3833
3834 mutex_unlock(&mdsc->mutex);
3835 schedule_delayed(mdsc);
3836 return;
3837
3838bad_unlock:
3839 mutex_unlock(&mdsc->mutex);
3840bad:
3841 pr_err("error decoding mdsmap %d\n", err);
3842 return;
3843}
3844
3845static struct ceph_connection *con_get(struct ceph_connection *con)
3846{
3847 struct ceph_mds_session *s = con->private;
3848
3849 if (get_session(s)) {
3850 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3851 return con;
3852 }
3853 dout("mdsc con_get %p FAIL\n", s);
3854 return NULL;
3855}
3856
3857static void con_put(struct ceph_connection *con)
3858{
3859 struct ceph_mds_session *s = con->private;
3860
3861 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3862 ceph_put_mds_session(s);
3863}
3864
3865/*
3866 * if the client is unresponsive for long enough, the mds will kill
3867 * the session entirely.
3868 */
3869static void peer_reset(struct ceph_connection *con)
3870{
3871 struct ceph_mds_session *s = con->private;
3872 struct ceph_mds_client *mdsc = s->s_mdsc;
3873
3874 pr_warn("mds%d closed our session\n", s->s_mds);
3875 send_mds_reconnect(mdsc, s);
3876}
3877
3878static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3879{
3880 struct ceph_mds_session *s = con->private;
3881 struct ceph_mds_client *mdsc = s->s_mdsc;
3882 int type = le16_to_cpu(msg->hdr.type);
3883
3884 mutex_lock(&mdsc->mutex);
3885 if (__verify_registered_session(mdsc, s) < 0) {
3886 mutex_unlock(&mdsc->mutex);
3887 goto out;
3888 }
3889 mutex_unlock(&mdsc->mutex);
3890
3891 switch (type) {
3892 case CEPH_MSG_MDS_MAP:
3893 ceph_mdsc_handle_mdsmap(mdsc, msg);
3894 break;
3895 case CEPH_MSG_FS_MAP_USER:
3896 ceph_mdsc_handle_fsmap(mdsc, msg);
3897 break;
3898 case CEPH_MSG_CLIENT_SESSION:
3899 handle_session(s, msg);
3900 break;
3901 case CEPH_MSG_CLIENT_REPLY:
3902 handle_reply(s, msg);
3903 break;
3904 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3905 handle_forward(mdsc, s, msg);
3906 break;
3907 case CEPH_MSG_CLIENT_CAPS:
3908 ceph_handle_caps(s, msg);
3909 break;
3910 case CEPH_MSG_CLIENT_SNAP:
3911 ceph_handle_snap(mdsc, s, msg);
3912 break;
3913 case CEPH_MSG_CLIENT_LEASE:
3914 handle_lease(mdsc, s, msg);
3915 break;
3916
3917 default:
3918 pr_err("received unknown message type %d %s\n", type,
3919 ceph_msg_type_name(type));
3920 }
3921out:
3922 ceph_msg_put(msg);
3923}
3924
3925/*
3926 * authentication
3927 */
3928
3929/*
3930 * Note: returned pointer is the address of a structure that's
3931 * managed separately. Caller must *not* attempt to free it.
3932 */
3933static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3934 int *proto, int force_new)
3935{
3936 struct ceph_mds_session *s = con->private;
3937 struct ceph_mds_client *mdsc = s->s_mdsc;
3938 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3939 struct ceph_auth_handshake *auth = &s->s_auth;
3940
3941 if (force_new && auth->authorizer) {
3942 ceph_auth_destroy_authorizer(auth->authorizer);
3943 auth->authorizer = NULL;
3944 }
3945 if (!auth->authorizer) {
3946 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3947 auth);
3948 if (ret)
3949 return ERR_PTR(ret);
3950 } else {
3951 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3952 auth);
3953 if (ret)
3954 return ERR_PTR(ret);
3955 }
3956 *proto = ac->protocol;
3957
3958 return auth;
3959}
3960
3961
3962static int verify_authorizer_reply(struct ceph_connection *con)
3963{
3964 struct ceph_mds_session *s = con->private;
3965 struct ceph_mds_client *mdsc = s->s_mdsc;
3966 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3967
3968 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
3969}
3970
3971static int invalidate_authorizer(struct ceph_connection *con)
3972{
3973 struct ceph_mds_session *s = con->private;
3974 struct ceph_mds_client *mdsc = s->s_mdsc;
3975 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3976
3977 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3978
3979 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3980}
3981
3982static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3983 struct ceph_msg_header *hdr, int *skip)
3984{
3985 struct ceph_msg *msg;
3986 int type = (int) le16_to_cpu(hdr->type);
3987 int front_len = (int) le32_to_cpu(hdr->front_len);
3988
3989 if (con->in_msg)
3990 return con->in_msg;
3991
3992 *skip = 0;
3993 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3994 if (!msg) {
3995 pr_err("unable to allocate msg type %d len %d\n",
3996 type, front_len);
3997 return NULL;
3998 }
3999
4000 return msg;
4001}
4002
4003static int mds_sign_message(struct ceph_msg *msg)
4004{
4005 struct ceph_mds_session *s = msg->con->private;
4006 struct ceph_auth_handshake *auth = &s->s_auth;
4007
4008 return ceph_auth_sign_message(auth, msg);
4009}
4010
4011static int mds_check_message_signature(struct ceph_msg *msg)
4012{
4013 struct ceph_mds_session *s = msg->con->private;
4014 struct ceph_auth_handshake *auth = &s->s_auth;
4015
4016 return ceph_auth_check_message_signature(auth, msg);
4017}
4018
4019static const struct ceph_connection_operations mds_con_ops = {
4020 .get = con_get,
4021 .put = con_put,
4022 .dispatch = dispatch,
4023 .get_authorizer = get_authorizer,
4024 .verify_authorizer_reply = verify_authorizer_reply,
4025 .invalidate_authorizer = invalidate_authorizer,
4026 .peer_reset = peer_reset,
4027 .alloc_msg = mds_alloc_msg,
4028 .sign_message = mds_sign_message,
4029 .check_message_signature = mds_check_message_signature,
4030};
4031
4032/* eof */