Loading...
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45 * are no longer valid.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 bool flock;
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 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
104 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
105 ceph_decode_need(p, end, info->pool_ns_len, bad);
106 *p += info->pool_ns_len;
107 } else {
108 info->pool_ns_len = 0;
109 }
110
111 return 0;
112bad:
113 return err;
114}
115
116/*
117 * parse a normal reply, which may contain a (dir+)dentry and/or a
118 * target inode.
119 */
120static int parse_reply_info_trace(void **p, void *end,
121 struct ceph_mds_reply_info_parsed *info,
122 u64 features)
123{
124 int err;
125
126 if (info->head->is_dentry) {
127 err = parse_reply_info_in(p, end, &info->diri, features);
128 if (err < 0)
129 goto out_bad;
130
131 if (unlikely(*p + sizeof(*info->dirfrag) > end))
132 goto bad;
133 info->dirfrag = *p;
134 *p += sizeof(*info->dirfrag) +
135 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
136 if (unlikely(*p > end))
137 goto bad;
138
139 ceph_decode_32_safe(p, end, info->dname_len, bad);
140 ceph_decode_need(p, end, info->dname_len, bad);
141 info->dname = *p;
142 *p += info->dname_len;
143 info->dlease = *p;
144 *p += sizeof(*info->dlease);
145 }
146
147 if (info->head->is_target) {
148 err = parse_reply_info_in(p, end, &info->targeti, features);
149 if (err < 0)
150 goto out_bad;
151 }
152
153 if (unlikely(*p != end))
154 goto bad;
155 return 0;
156
157bad:
158 err = -EIO;
159out_bad:
160 pr_err("problem parsing mds trace %d\n", err);
161 return err;
162}
163
164/*
165 * parse readdir results
166 */
167static int parse_reply_info_dir(void **p, void *end,
168 struct ceph_mds_reply_info_parsed *info,
169 u64 features)
170{
171 u32 num, i = 0;
172 int err;
173
174 info->dir_dir = *p;
175 if (*p + sizeof(*info->dir_dir) > end)
176 goto bad;
177 *p += sizeof(*info->dir_dir) +
178 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
179 if (*p > end)
180 goto bad;
181
182 ceph_decode_need(p, end, sizeof(num) + 2, bad);
183 num = ceph_decode_32(p);
184 info->dir_end = ceph_decode_8(p);
185 info->dir_complete = ceph_decode_8(p);
186 if (num == 0)
187 goto done;
188
189 BUG_ON(!info->dir_in);
190 info->dir_dname = (void *)(info->dir_in + num);
191 info->dir_dname_len = (void *)(info->dir_dname + num);
192 info->dir_dlease = (void *)(info->dir_dname_len + num);
193 if ((unsigned long)(info->dir_dlease + num) >
194 (unsigned long)info->dir_in + info->dir_buf_size) {
195 pr_err("dir contents are larger than expected\n");
196 WARN_ON(1);
197 goto bad;
198 }
199
200 info->dir_nr = num;
201 while (num) {
202 /* dentry */
203 ceph_decode_need(p, end, sizeof(u32)*2, bad);
204 info->dir_dname_len[i] = ceph_decode_32(p);
205 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
206 info->dir_dname[i] = *p;
207 *p += info->dir_dname_len[i];
208 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
209 info->dir_dname[i]);
210 info->dir_dlease[i] = *p;
211 *p += sizeof(struct ceph_mds_reply_lease);
212
213 /* inode */
214 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
215 if (err < 0)
216 goto out_bad;
217 i++;
218 num--;
219 }
220
221done:
222 if (*p != end)
223 goto bad;
224 return 0;
225
226bad:
227 err = -EIO;
228out_bad:
229 pr_err("problem parsing dir contents %d\n", err);
230 return err;
231}
232
233/*
234 * parse fcntl F_GETLK results
235 */
236static int parse_reply_info_filelock(void **p, void *end,
237 struct ceph_mds_reply_info_parsed *info,
238 u64 features)
239{
240 if (*p + sizeof(*info->filelock_reply) > end)
241 goto bad;
242
243 info->filelock_reply = *p;
244 *p += sizeof(*info->filelock_reply);
245
246 if (unlikely(*p != end))
247 goto bad;
248 return 0;
249
250bad:
251 return -EIO;
252}
253
254/*
255 * parse create results
256 */
257static int parse_reply_info_create(void **p, void *end,
258 struct ceph_mds_reply_info_parsed *info,
259 u64 features)
260{
261 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
262 if (*p == end) {
263 info->has_create_ino = false;
264 } else {
265 info->has_create_ino = true;
266 info->ino = ceph_decode_64(p);
267 }
268 }
269
270 if (unlikely(*p != end))
271 goto bad;
272 return 0;
273
274bad:
275 return -EIO;
276}
277
278/*
279 * parse extra results
280 */
281static int parse_reply_info_extra(void **p, void *end,
282 struct ceph_mds_reply_info_parsed *info,
283 u64 features)
284{
285 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
286 return parse_reply_info_filelock(p, end, info, features);
287 else if (info->head->op == CEPH_MDS_OP_READDIR ||
288 info->head->op == CEPH_MDS_OP_LSSNAP)
289 return parse_reply_info_dir(p, end, info, features);
290 else if (info->head->op == CEPH_MDS_OP_CREATE)
291 return parse_reply_info_create(p, end, info, features);
292 else
293 return -EIO;
294}
295
296/*
297 * parse entire mds reply
298 */
299static int parse_reply_info(struct ceph_msg *msg,
300 struct ceph_mds_reply_info_parsed *info,
301 u64 features)
302{
303 void *p, *end;
304 u32 len;
305 int err;
306
307 info->head = msg->front.iov_base;
308 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
309 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
310
311 /* trace */
312 ceph_decode_32_safe(&p, end, len, bad);
313 if (len > 0) {
314 ceph_decode_need(&p, end, len, bad);
315 err = parse_reply_info_trace(&p, p+len, info, features);
316 if (err < 0)
317 goto out_bad;
318 }
319
320 /* extra */
321 ceph_decode_32_safe(&p, end, len, bad);
322 if (len > 0) {
323 ceph_decode_need(&p, end, len, bad);
324 err = parse_reply_info_extra(&p, p+len, info, features);
325 if (err < 0)
326 goto out_bad;
327 }
328
329 /* snap blob */
330 ceph_decode_32_safe(&p, end, len, bad);
331 info->snapblob_len = len;
332 info->snapblob = p;
333 p += len;
334
335 if (p != end)
336 goto bad;
337 return 0;
338
339bad:
340 err = -EIO;
341out_bad:
342 pr_err("mds parse_reply err %d\n", err);
343 return err;
344}
345
346static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
347{
348 if (!info->dir_in)
349 return;
350 free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
351}
352
353
354/*
355 * sessions
356 */
357const char *ceph_session_state_name(int s)
358{
359 switch (s) {
360 case CEPH_MDS_SESSION_NEW: return "new";
361 case CEPH_MDS_SESSION_OPENING: return "opening";
362 case CEPH_MDS_SESSION_OPEN: return "open";
363 case CEPH_MDS_SESSION_HUNG: return "hung";
364 case CEPH_MDS_SESSION_CLOSING: return "closing";
365 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
366 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
367 default: return "???";
368 }
369}
370
371static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
372{
373 if (atomic_inc_not_zero(&s->s_ref)) {
374 dout("mdsc get_session %p %d -> %d\n", s,
375 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
376 return s;
377 } else {
378 dout("mdsc get_session %p 0 -- FAIL", s);
379 return NULL;
380 }
381}
382
383void ceph_put_mds_session(struct ceph_mds_session *s)
384{
385 dout("mdsc put_session %p %d -> %d\n", s,
386 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
387 if (atomic_dec_and_test(&s->s_ref)) {
388 if (s->s_auth.authorizer)
389 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
390 kfree(s);
391 }
392}
393
394/*
395 * called under mdsc->mutex
396 */
397struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
398 int mds)
399{
400 struct ceph_mds_session *session;
401
402 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
403 return NULL;
404 session = mdsc->sessions[mds];
405 dout("lookup_mds_session %p %d\n", session,
406 atomic_read(&session->s_ref));
407 get_session(session);
408 return session;
409}
410
411static bool __have_session(struct ceph_mds_client *mdsc, int mds)
412{
413 if (mds >= mdsc->max_sessions)
414 return false;
415 return mdsc->sessions[mds];
416}
417
418static int __verify_registered_session(struct ceph_mds_client *mdsc,
419 struct ceph_mds_session *s)
420{
421 if (s->s_mds >= mdsc->max_sessions ||
422 mdsc->sessions[s->s_mds] != s)
423 return -ENOENT;
424 return 0;
425}
426
427/*
428 * create+register a new session for given mds.
429 * called under mdsc->mutex.
430 */
431static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
432 int mds)
433{
434 struct ceph_mds_session *s;
435
436 if (mds >= mdsc->mdsmap->m_max_mds)
437 return ERR_PTR(-EINVAL);
438
439 s = kzalloc(sizeof(*s), GFP_NOFS);
440 if (!s)
441 return ERR_PTR(-ENOMEM);
442 s->s_mdsc = mdsc;
443 s->s_mds = mds;
444 s->s_state = CEPH_MDS_SESSION_NEW;
445 s->s_ttl = 0;
446 s->s_seq = 0;
447 mutex_init(&s->s_mutex);
448
449 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
450
451 spin_lock_init(&s->s_gen_ttl_lock);
452 s->s_cap_gen = 0;
453 s->s_cap_ttl = jiffies - 1;
454
455 spin_lock_init(&s->s_cap_lock);
456 s->s_renew_requested = 0;
457 s->s_renew_seq = 0;
458 INIT_LIST_HEAD(&s->s_caps);
459 s->s_nr_caps = 0;
460 s->s_trim_caps = 0;
461 atomic_set(&s->s_ref, 1);
462 INIT_LIST_HEAD(&s->s_waiting);
463 INIT_LIST_HEAD(&s->s_unsafe);
464 s->s_num_cap_releases = 0;
465 s->s_cap_reconnect = 0;
466 s->s_cap_iterator = NULL;
467 INIT_LIST_HEAD(&s->s_cap_releases);
468 INIT_LIST_HEAD(&s->s_cap_flushing);
469 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
470
471 dout("register_session mds%d\n", mds);
472 if (mds >= mdsc->max_sessions) {
473 int newmax = 1 << get_count_order(mds+1);
474 struct ceph_mds_session **sa;
475
476 dout("register_session realloc to %d\n", newmax);
477 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
478 if (sa == NULL)
479 goto fail_realloc;
480 if (mdsc->sessions) {
481 memcpy(sa, mdsc->sessions,
482 mdsc->max_sessions * sizeof(void *));
483 kfree(mdsc->sessions);
484 }
485 mdsc->sessions = sa;
486 mdsc->max_sessions = newmax;
487 }
488 mdsc->sessions[mds] = s;
489 atomic_inc(&mdsc->num_sessions);
490 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
491
492 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
493 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
494
495 return s;
496
497fail_realloc:
498 kfree(s);
499 return ERR_PTR(-ENOMEM);
500}
501
502/*
503 * called under mdsc->mutex
504 */
505static void __unregister_session(struct ceph_mds_client *mdsc,
506 struct ceph_mds_session *s)
507{
508 dout("__unregister_session mds%d %p\n", s->s_mds, s);
509 BUG_ON(mdsc->sessions[s->s_mds] != s);
510 mdsc->sessions[s->s_mds] = NULL;
511 ceph_con_close(&s->s_con);
512 ceph_put_mds_session(s);
513 atomic_dec(&mdsc->num_sessions);
514}
515
516/*
517 * drop session refs in request.
518 *
519 * should be last request ref, or hold mdsc->mutex
520 */
521static void put_request_session(struct ceph_mds_request *req)
522{
523 if (req->r_session) {
524 ceph_put_mds_session(req->r_session);
525 req->r_session = NULL;
526 }
527}
528
529void ceph_mdsc_release_request(struct kref *kref)
530{
531 struct ceph_mds_request *req = container_of(kref,
532 struct ceph_mds_request,
533 r_kref);
534 destroy_reply_info(&req->r_reply_info);
535 if (req->r_request)
536 ceph_msg_put(req->r_request);
537 if (req->r_reply)
538 ceph_msg_put(req->r_reply);
539 if (req->r_inode) {
540 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
541 iput(req->r_inode);
542 }
543 if (req->r_locked_dir)
544 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
545 iput(req->r_target_inode);
546 if (req->r_dentry)
547 dput(req->r_dentry);
548 if (req->r_old_dentry)
549 dput(req->r_old_dentry);
550 if (req->r_old_dentry_dir) {
551 /*
552 * track (and drop pins for) r_old_dentry_dir
553 * separately, since r_old_dentry's d_parent may have
554 * changed between the dir mutex being dropped and
555 * this request being freed.
556 */
557 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
558 CEPH_CAP_PIN);
559 iput(req->r_old_dentry_dir);
560 }
561 kfree(req->r_path1);
562 kfree(req->r_path2);
563 if (req->r_pagelist)
564 ceph_pagelist_release(req->r_pagelist);
565 put_request_session(req);
566 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
567 kfree(req);
568}
569
570/*
571 * lookup session, bump ref if found.
572 *
573 * called under mdsc->mutex.
574 */
575static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
576 u64 tid)
577{
578 struct ceph_mds_request *req;
579 struct rb_node *n = mdsc->request_tree.rb_node;
580
581 while (n) {
582 req = rb_entry(n, struct ceph_mds_request, r_node);
583 if (tid < req->r_tid)
584 n = n->rb_left;
585 else if (tid > req->r_tid)
586 n = n->rb_right;
587 else {
588 ceph_mdsc_get_request(req);
589 return req;
590 }
591 }
592 return NULL;
593}
594
595static void __insert_request(struct ceph_mds_client *mdsc,
596 struct ceph_mds_request *new)
597{
598 struct rb_node **p = &mdsc->request_tree.rb_node;
599 struct rb_node *parent = NULL;
600 struct ceph_mds_request *req = NULL;
601
602 while (*p) {
603 parent = *p;
604 req = rb_entry(parent, struct ceph_mds_request, r_node);
605 if (new->r_tid < req->r_tid)
606 p = &(*p)->rb_left;
607 else if (new->r_tid > req->r_tid)
608 p = &(*p)->rb_right;
609 else
610 BUG();
611 }
612
613 rb_link_node(&new->r_node, parent, p);
614 rb_insert_color(&new->r_node, &mdsc->request_tree);
615}
616
617/*
618 * Register an in-flight request, and assign a tid. Link to directory
619 * are modifying (if any).
620 *
621 * Called under mdsc->mutex.
622 */
623static void __register_request(struct ceph_mds_client *mdsc,
624 struct ceph_mds_request *req,
625 struct inode *dir)
626{
627 req->r_tid = ++mdsc->last_tid;
628 if (req->r_num_caps)
629 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
630 req->r_num_caps);
631 dout("__register_request %p tid %lld\n", req, req->r_tid);
632 ceph_mdsc_get_request(req);
633 __insert_request(mdsc, req);
634
635 req->r_uid = current_fsuid();
636 req->r_gid = current_fsgid();
637
638 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
639 mdsc->oldest_tid = req->r_tid;
640
641 if (dir) {
642 ihold(dir);
643 req->r_unsafe_dir = dir;
644 }
645}
646
647static void __unregister_request(struct ceph_mds_client *mdsc,
648 struct ceph_mds_request *req)
649{
650 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
651
652 if (req->r_tid == mdsc->oldest_tid) {
653 struct rb_node *p = rb_next(&req->r_node);
654 mdsc->oldest_tid = 0;
655 while (p) {
656 struct ceph_mds_request *next_req =
657 rb_entry(p, struct ceph_mds_request, r_node);
658 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
659 mdsc->oldest_tid = next_req->r_tid;
660 break;
661 }
662 p = rb_next(p);
663 }
664 }
665
666 rb_erase(&req->r_node, &mdsc->request_tree);
667 RB_CLEAR_NODE(&req->r_node);
668
669 if (req->r_unsafe_dir && req->r_got_unsafe) {
670 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
671 spin_lock(&ci->i_unsafe_lock);
672 list_del_init(&req->r_unsafe_dir_item);
673 spin_unlock(&ci->i_unsafe_lock);
674 }
675 if (req->r_target_inode && req->r_got_unsafe) {
676 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
677 spin_lock(&ci->i_unsafe_lock);
678 list_del_init(&req->r_unsafe_target_item);
679 spin_unlock(&ci->i_unsafe_lock);
680 }
681
682 if (req->r_unsafe_dir) {
683 iput(req->r_unsafe_dir);
684 req->r_unsafe_dir = NULL;
685 }
686
687 complete_all(&req->r_safe_completion);
688
689 ceph_mdsc_put_request(req);
690}
691
692/*
693 * Choose mds to send request to next. If there is a hint set in the
694 * request (e.g., due to a prior forward hint from the mds), use that.
695 * Otherwise, consult frag tree and/or caps to identify the
696 * appropriate mds. If all else fails, choose randomly.
697 *
698 * Called under mdsc->mutex.
699 */
700static struct dentry *get_nonsnap_parent(struct dentry *dentry)
701{
702 /*
703 * we don't need to worry about protecting the d_parent access
704 * here because we never renaming inside the snapped namespace
705 * except to resplice to another snapdir, and either the old or new
706 * result is a valid result.
707 */
708 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
709 dentry = dentry->d_parent;
710 return dentry;
711}
712
713static int __choose_mds(struct ceph_mds_client *mdsc,
714 struct ceph_mds_request *req)
715{
716 struct inode *inode;
717 struct ceph_inode_info *ci;
718 struct ceph_cap *cap;
719 int mode = req->r_direct_mode;
720 int mds = -1;
721 u32 hash = req->r_direct_hash;
722 bool is_hash = req->r_direct_is_hash;
723
724 /*
725 * is there a specific mds we should try? ignore hint if we have
726 * no session and the mds is not up (active or recovering).
727 */
728 if (req->r_resend_mds >= 0 &&
729 (__have_session(mdsc, req->r_resend_mds) ||
730 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
731 dout("choose_mds using resend_mds mds%d\n",
732 req->r_resend_mds);
733 return req->r_resend_mds;
734 }
735
736 if (mode == USE_RANDOM_MDS)
737 goto random;
738
739 inode = NULL;
740 if (req->r_inode) {
741 inode = req->r_inode;
742 } else if (req->r_dentry) {
743 /* ignore race with rename; old or new d_parent is okay */
744 struct dentry *parent = req->r_dentry->d_parent;
745 struct inode *dir = d_inode(parent);
746
747 if (dir->i_sb != mdsc->fsc->sb) {
748 /* not this fs! */
749 inode = d_inode(req->r_dentry);
750 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
751 /* direct snapped/virtual snapdir requests
752 * based on parent dir inode */
753 struct dentry *dn = get_nonsnap_parent(parent);
754 inode = d_inode(dn);
755 dout("__choose_mds using nonsnap parent %p\n", inode);
756 } else {
757 /* dentry target */
758 inode = d_inode(req->r_dentry);
759 if (!inode || mode == USE_AUTH_MDS) {
760 /* dir + name */
761 inode = dir;
762 hash = ceph_dentry_hash(dir, req->r_dentry);
763 is_hash = true;
764 }
765 }
766 }
767
768 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
769 (int)hash, mode);
770 if (!inode)
771 goto random;
772 ci = ceph_inode(inode);
773
774 if (is_hash && S_ISDIR(inode->i_mode)) {
775 struct ceph_inode_frag frag;
776 int found;
777
778 ceph_choose_frag(ci, hash, &frag, &found);
779 if (found) {
780 if (mode == USE_ANY_MDS && frag.ndist > 0) {
781 u8 r;
782
783 /* choose a random replica */
784 get_random_bytes(&r, 1);
785 r %= frag.ndist;
786 mds = frag.dist[r];
787 dout("choose_mds %p %llx.%llx "
788 "frag %u mds%d (%d/%d)\n",
789 inode, ceph_vinop(inode),
790 frag.frag, mds,
791 (int)r, frag.ndist);
792 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
793 CEPH_MDS_STATE_ACTIVE)
794 return mds;
795 }
796
797 /* since this file/dir wasn't known to be
798 * replicated, then we want to look for the
799 * authoritative mds. */
800 mode = USE_AUTH_MDS;
801 if (frag.mds >= 0) {
802 /* choose auth mds */
803 mds = frag.mds;
804 dout("choose_mds %p %llx.%llx "
805 "frag %u mds%d (auth)\n",
806 inode, ceph_vinop(inode), frag.frag, mds);
807 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
808 CEPH_MDS_STATE_ACTIVE)
809 return mds;
810 }
811 }
812 }
813
814 spin_lock(&ci->i_ceph_lock);
815 cap = NULL;
816 if (mode == USE_AUTH_MDS)
817 cap = ci->i_auth_cap;
818 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
819 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
820 if (!cap) {
821 spin_unlock(&ci->i_ceph_lock);
822 goto random;
823 }
824 mds = cap->session->s_mds;
825 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
826 inode, ceph_vinop(inode), mds,
827 cap == ci->i_auth_cap ? "auth " : "", cap);
828 spin_unlock(&ci->i_ceph_lock);
829 return mds;
830
831random:
832 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
833 dout("choose_mds chose random mds%d\n", mds);
834 return mds;
835}
836
837
838/*
839 * session messages
840 */
841static struct ceph_msg *create_session_msg(u32 op, u64 seq)
842{
843 struct ceph_msg *msg;
844 struct ceph_mds_session_head *h;
845
846 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
847 false);
848 if (!msg) {
849 pr_err("create_session_msg ENOMEM creating msg\n");
850 return NULL;
851 }
852 h = msg->front.iov_base;
853 h->op = cpu_to_le32(op);
854 h->seq = cpu_to_le64(seq);
855
856 return msg;
857}
858
859/*
860 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
861 * to include additional client metadata fields.
862 */
863static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
864{
865 struct ceph_msg *msg;
866 struct ceph_mds_session_head *h;
867 int i = -1;
868 int metadata_bytes = 0;
869 int metadata_key_count = 0;
870 struct ceph_options *opt = mdsc->fsc->client->options;
871 void *p;
872
873 const char* metadata[][2] = {
874 {"hostname", utsname()->nodename},
875 {"kernel_version", utsname()->release},
876 {"entity_id", opt->name ? opt->name : ""},
877 {NULL, NULL}
878 };
879
880 /* Calculate serialized length of metadata */
881 metadata_bytes = 4; /* map length */
882 for (i = 0; metadata[i][0] != NULL; ++i) {
883 metadata_bytes += 8 + strlen(metadata[i][0]) +
884 strlen(metadata[i][1]);
885 metadata_key_count++;
886 }
887
888 /* Allocate the message */
889 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
890 GFP_NOFS, false);
891 if (!msg) {
892 pr_err("create_session_msg ENOMEM creating msg\n");
893 return NULL;
894 }
895 h = msg->front.iov_base;
896 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
897 h->seq = cpu_to_le64(seq);
898
899 /*
900 * Serialize client metadata into waiting buffer space, using
901 * the format that userspace expects for map<string, string>
902 *
903 * ClientSession messages with metadata are v2
904 */
905 msg->hdr.version = cpu_to_le16(2);
906 msg->hdr.compat_version = cpu_to_le16(1);
907
908 /* The write pointer, following the session_head structure */
909 p = msg->front.iov_base + sizeof(*h);
910
911 /* Number of entries in the map */
912 ceph_encode_32(&p, metadata_key_count);
913
914 /* Two length-prefixed strings for each entry in the map */
915 for (i = 0; metadata[i][0] != NULL; ++i) {
916 size_t const key_len = strlen(metadata[i][0]);
917 size_t const val_len = strlen(metadata[i][1]);
918
919 ceph_encode_32(&p, key_len);
920 memcpy(p, metadata[i][0], key_len);
921 p += key_len;
922 ceph_encode_32(&p, val_len);
923 memcpy(p, metadata[i][1], val_len);
924 p += val_len;
925 }
926
927 return msg;
928}
929
930/*
931 * send session open request.
932 *
933 * called under mdsc->mutex
934 */
935static int __open_session(struct ceph_mds_client *mdsc,
936 struct ceph_mds_session *session)
937{
938 struct ceph_msg *msg;
939 int mstate;
940 int mds = session->s_mds;
941
942 /* wait for mds to go active? */
943 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
944 dout("open_session to mds%d (%s)\n", mds,
945 ceph_mds_state_name(mstate));
946 session->s_state = CEPH_MDS_SESSION_OPENING;
947 session->s_renew_requested = jiffies;
948
949 /* send connect message */
950 msg = create_session_open_msg(mdsc, session->s_seq);
951 if (!msg)
952 return -ENOMEM;
953 ceph_con_send(&session->s_con, msg);
954 return 0;
955}
956
957/*
958 * open sessions for any export targets for the given mds
959 *
960 * called under mdsc->mutex
961 */
962static struct ceph_mds_session *
963__open_export_target_session(struct ceph_mds_client *mdsc, int target)
964{
965 struct ceph_mds_session *session;
966
967 session = __ceph_lookup_mds_session(mdsc, target);
968 if (!session) {
969 session = register_session(mdsc, target);
970 if (IS_ERR(session))
971 return session;
972 }
973 if (session->s_state == CEPH_MDS_SESSION_NEW ||
974 session->s_state == CEPH_MDS_SESSION_CLOSING)
975 __open_session(mdsc, session);
976
977 return session;
978}
979
980struct ceph_mds_session *
981ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
982{
983 struct ceph_mds_session *session;
984
985 dout("open_export_target_session to mds%d\n", target);
986
987 mutex_lock(&mdsc->mutex);
988 session = __open_export_target_session(mdsc, target);
989 mutex_unlock(&mdsc->mutex);
990
991 return session;
992}
993
994static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
995 struct ceph_mds_session *session)
996{
997 struct ceph_mds_info *mi;
998 struct ceph_mds_session *ts;
999 int i, mds = session->s_mds;
1000
1001 if (mds >= mdsc->mdsmap->m_max_mds)
1002 return;
1003
1004 mi = &mdsc->mdsmap->m_info[mds];
1005 dout("open_export_target_sessions for mds%d (%d targets)\n",
1006 session->s_mds, mi->num_export_targets);
1007
1008 for (i = 0; i < mi->num_export_targets; i++) {
1009 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1010 if (!IS_ERR(ts))
1011 ceph_put_mds_session(ts);
1012 }
1013}
1014
1015void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1016 struct ceph_mds_session *session)
1017{
1018 mutex_lock(&mdsc->mutex);
1019 __open_export_target_sessions(mdsc, session);
1020 mutex_unlock(&mdsc->mutex);
1021}
1022
1023/*
1024 * session caps
1025 */
1026
1027/* caller holds s_cap_lock, we drop it */
1028static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1029 struct ceph_mds_session *session)
1030 __releases(session->s_cap_lock)
1031{
1032 LIST_HEAD(tmp_list);
1033 list_splice_init(&session->s_cap_releases, &tmp_list);
1034 session->s_num_cap_releases = 0;
1035 spin_unlock(&session->s_cap_lock);
1036
1037 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1038 while (!list_empty(&tmp_list)) {
1039 struct ceph_cap *cap;
1040 /* zero out the in-progress message */
1041 cap = list_first_entry(&tmp_list,
1042 struct ceph_cap, session_caps);
1043 list_del(&cap->session_caps);
1044 ceph_put_cap(mdsc, cap);
1045 }
1046}
1047
1048static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1049 struct ceph_mds_session *session)
1050{
1051 struct ceph_mds_request *req;
1052 struct rb_node *p;
1053
1054 dout("cleanup_session_requests mds%d\n", session->s_mds);
1055 mutex_lock(&mdsc->mutex);
1056 while (!list_empty(&session->s_unsafe)) {
1057 req = list_first_entry(&session->s_unsafe,
1058 struct ceph_mds_request, r_unsafe_item);
1059 list_del_init(&req->r_unsafe_item);
1060 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1061 req->r_tid);
1062 __unregister_request(mdsc, req);
1063 }
1064 /* zero r_attempts, so kick_requests() will re-send requests */
1065 p = rb_first(&mdsc->request_tree);
1066 while (p) {
1067 req = rb_entry(p, struct ceph_mds_request, r_node);
1068 p = rb_next(p);
1069 if (req->r_session &&
1070 req->r_session->s_mds == session->s_mds)
1071 req->r_attempts = 0;
1072 }
1073 mutex_unlock(&mdsc->mutex);
1074}
1075
1076/*
1077 * Helper to safely iterate over all caps associated with a session, with
1078 * special care taken to handle a racing __ceph_remove_cap().
1079 *
1080 * Caller must hold session s_mutex.
1081 */
1082static int iterate_session_caps(struct ceph_mds_session *session,
1083 int (*cb)(struct inode *, struct ceph_cap *,
1084 void *), void *arg)
1085{
1086 struct list_head *p;
1087 struct ceph_cap *cap;
1088 struct inode *inode, *last_inode = NULL;
1089 struct ceph_cap *old_cap = NULL;
1090 int ret;
1091
1092 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1093 spin_lock(&session->s_cap_lock);
1094 p = session->s_caps.next;
1095 while (p != &session->s_caps) {
1096 cap = list_entry(p, struct ceph_cap, session_caps);
1097 inode = igrab(&cap->ci->vfs_inode);
1098 if (!inode) {
1099 p = p->next;
1100 continue;
1101 }
1102 session->s_cap_iterator = cap;
1103 spin_unlock(&session->s_cap_lock);
1104
1105 if (last_inode) {
1106 iput(last_inode);
1107 last_inode = NULL;
1108 }
1109 if (old_cap) {
1110 ceph_put_cap(session->s_mdsc, old_cap);
1111 old_cap = NULL;
1112 }
1113
1114 ret = cb(inode, cap, arg);
1115 last_inode = inode;
1116
1117 spin_lock(&session->s_cap_lock);
1118 p = p->next;
1119 if (cap->ci == NULL) {
1120 dout("iterate_session_caps finishing cap %p removal\n",
1121 cap);
1122 BUG_ON(cap->session != session);
1123 cap->session = NULL;
1124 list_del_init(&cap->session_caps);
1125 session->s_nr_caps--;
1126 if (cap->queue_release) {
1127 list_add_tail(&cap->session_caps,
1128 &session->s_cap_releases);
1129 session->s_num_cap_releases++;
1130 } else {
1131 old_cap = cap; /* put_cap it w/o locks held */
1132 }
1133 }
1134 if (ret < 0)
1135 goto out;
1136 }
1137 ret = 0;
1138out:
1139 session->s_cap_iterator = NULL;
1140 spin_unlock(&session->s_cap_lock);
1141
1142 iput(last_inode);
1143 if (old_cap)
1144 ceph_put_cap(session->s_mdsc, old_cap);
1145
1146 return ret;
1147}
1148
1149static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1150 void *arg)
1151{
1152 struct ceph_inode_info *ci = ceph_inode(inode);
1153 LIST_HEAD(to_remove);
1154 int drop = 0;
1155
1156 dout("removing cap %p, ci is %p, inode is %p\n",
1157 cap, ci, &ci->vfs_inode);
1158 spin_lock(&ci->i_ceph_lock);
1159 __ceph_remove_cap(cap, false);
1160 if (!ci->i_auth_cap) {
1161 struct ceph_cap_flush *cf;
1162 struct ceph_mds_client *mdsc =
1163 ceph_sb_to_client(inode->i_sb)->mdsc;
1164
1165 while (true) {
1166 struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1167 if (!n)
1168 break;
1169 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1170 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1171 list_add(&cf->list, &to_remove);
1172 }
1173
1174 spin_lock(&mdsc->cap_dirty_lock);
1175
1176 list_for_each_entry(cf, &to_remove, list)
1177 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1178
1179 if (!list_empty(&ci->i_dirty_item)) {
1180 pr_warn_ratelimited(
1181 " dropping dirty %s state for %p %lld\n",
1182 ceph_cap_string(ci->i_dirty_caps),
1183 inode, ceph_ino(inode));
1184 ci->i_dirty_caps = 0;
1185 list_del_init(&ci->i_dirty_item);
1186 drop = 1;
1187 }
1188 if (!list_empty(&ci->i_flushing_item)) {
1189 pr_warn_ratelimited(
1190 " dropping dirty+flushing %s state for %p %lld\n",
1191 ceph_cap_string(ci->i_flushing_caps),
1192 inode, ceph_ino(inode));
1193 ci->i_flushing_caps = 0;
1194 list_del_init(&ci->i_flushing_item);
1195 mdsc->num_cap_flushing--;
1196 drop = 1;
1197 }
1198 spin_unlock(&mdsc->cap_dirty_lock);
1199
1200 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1201 list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1202 ci->i_prealloc_cap_flush = NULL;
1203 }
1204 }
1205 spin_unlock(&ci->i_ceph_lock);
1206 while (!list_empty(&to_remove)) {
1207 struct ceph_cap_flush *cf;
1208 cf = list_first_entry(&to_remove,
1209 struct ceph_cap_flush, list);
1210 list_del(&cf->list);
1211 ceph_free_cap_flush(cf);
1212 }
1213 while (drop--)
1214 iput(inode);
1215 return 0;
1216}
1217
1218/*
1219 * caller must hold session s_mutex
1220 */
1221static void remove_session_caps(struct ceph_mds_session *session)
1222{
1223 dout("remove_session_caps on %p\n", session);
1224 iterate_session_caps(session, remove_session_caps_cb, NULL);
1225
1226 spin_lock(&session->s_cap_lock);
1227 if (session->s_nr_caps > 0) {
1228 struct super_block *sb = session->s_mdsc->fsc->sb;
1229 struct inode *inode;
1230 struct ceph_cap *cap, *prev = NULL;
1231 struct ceph_vino vino;
1232 /*
1233 * iterate_session_caps() skips inodes that are being
1234 * deleted, we need to wait until deletions are complete.
1235 * __wait_on_freeing_inode() is designed for the job,
1236 * but it is not exported, so use lookup inode function
1237 * to access it.
1238 */
1239 while (!list_empty(&session->s_caps)) {
1240 cap = list_entry(session->s_caps.next,
1241 struct ceph_cap, session_caps);
1242 if (cap == prev)
1243 break;
1244 prev = cap;
1245 vino = cap->ci->i_vino;
1246 spin_unlock(&session->s_cap_lock);
1247
1248 inode = ceph_find_inode(sb, vino);
1249 iput(inode);
1250
1251 spin_lock(&session->s_cap_lock);
1252 }
1253 }
1254
1255 // drop cap expires and unlock s_cap_lock
1256 cleanup_cap_releases(session->s_mdsc, session);
1257
1258 BUG_ON(session->s_nr_caps > 0);
1259 BUG_ON(!list_empty(&session->s_cap_flushing));
1260}
1261
1262/*
1263 * wake up any threads waiting on this session's caps. if the cap is
1264 * old (didn't get renewed on the client reconnect), remove it now.
1265 *
1266 * caller must hold s_mutex.
1267 */
1268static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1269 void *arg)
1270{
1271 struct ceph_inode_info *ci = ceph_inode(inode);
1272
1273 wake_up_all(&ci->i_cap_wq);
1274 if (arg) {
1275 spin_lock(&ci->i_ceph_lock);
1276 ci->i_wanted_max_size = 0;
1277 ci->i_requested_max_size = 0;
1278 spin_unlock(&ci->i_ceph_lock);
1279 }
1280 return 0;
1281}
1282
1283static void wake_up_session_caps(struct ceph_mds_session *session,
1284 int reconnect)
1285{
1286 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1287 iterate_session_caps(session, wake_up_session_cb,
1288 (void *)(unsigned long)reconnect);
1289}
1290
1291/*
1292 * Send periodic message to MDS renewing all currently held caps. The
1293 * ack will reset the expiration for all caps from this session.
1294 *
1295 * caller holds s_mutex
1296 */
1297static int send_renew_caps(struct ceph_mds_client *mdsc,
1298 struct ceph_mds_session *session)
1299{
1300 struct ceph_msg *msg;
1301 int state;
1302
1303 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1304 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1305 pr_info("mds%d caps stale\n", session->s_mds);
1306 session->s_renew_requested = jiffies;
1307
1308 /* do not try to renew caps until a recovering mds has reconnected
1309 * with its clients. */
1310 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1311 if (state < CEPH_MDS_STATE_RECONNECT) {
1312 dout("send_renew_caps ignoring mds%d (%s)\n",
1313 session->s_mds, ceph_mds_state_name(state));
1314 return 0;
1315 }
1316
1317 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1318 ceph_mds_state_name(state));
1319 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1320 ++session->s_renew_seq);
1321 if (!msg)
1322 return -ENOMEM;
1323 ceph_con_send(&session->s_con, msg);
1324 return 0;
1325}
1326
1327static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1328 struct ceph_mds_session *session, u64 seq)
1329{
1330 struct ceph_msg *msg;
1331
1332 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1333 session->s_mds, ceph_session_state_name(session->s_state), seq);
1334 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1335 if (!msg)
1336 return -ENOMEM;
1337 ceph_con_send(&session->s_con, msg);
1338 return 0;
1339}
1340
1341
1342/*
1343 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1344 *
1345 * Called under session->s_mutex
1346 */
1347static void renewed_caps(struct ceph_mds_client *mdsc,
1348 struct ceph_mds_session *session, int is_renew)
1349{
1350 int was_stale;
1351 int wake = 0;
1352
1353 spin_lock(&session->s_cap_lock);
1354 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1355
1356 session->s_cap_ttl = session->s_renew_requested +
1357 mdsc->mdsmap->m_session_timeout*HZ;
1358
1359 if (was_stale) {
1360 if (time_before(jiffies, session->s_cap_ttl)) {
1361 pr_info("mds%d caps renewed\n", session->s_mds);
1362 wake = 1;
1363 } else {
1364 pr_info("mds%d caps still stale\n", session->s_mds);
1365 }
1366 }
1367 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1368 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1369 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1370 spin_unlock(&session->s_cap_lock);
1371
1372 if (wake)
1373 wake_up_session_caps(session, 0);
1374}
1375
1376/*
1377 * send a session close request
1378 */
1379static int request_close_session(struct ceph_mds_client *mdsc,
1380 struct ceph_mds_session *session)
1381{
1382 struct ceph_msg *msg;
1383
1384 dout("request_close_session mds%d state %s seq %lld\n",
1385 session->s_mds, ceph_session_state_name(session->s_state),
1386 session->s_seq);
1387 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1388 if (!msg)
1389 return -ENOMEM;
1390 ceph_con_send(&session->s_con, msg);
1391 return 0;
1392}
1393
1394/*
1395 * Called with s_mutex held.
1396 */
1397static int __close_session(struct ceph_mds_client *mdsc,
1398 struct ceph_mds_session *session)
1399{
1400 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1401 return 0;
1402 session->s_state = CEPH_MDS_SESSION_CLOSING;
1403 return request_close_session(mdsc, session);
1404}
1405
1406/*
1407 * Trim old(er) caps.
1408 *
1409 * Because we can't cache an inode without one or more caps, we do
1410 * this indirectly: if a cap is unused, we prune its aliases, at which
1411 * point the inode will hopefully get dropped to.
1412 *
1413 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1414 * memory pressure from the MDS, though, so it needn't be perfect.
1415 */
1416static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1417{
1418 struct ceph_mds_session *session = arg;
1419 struct ceph_inode_info *ci = ceph_inode(inode);
1420 int used, wanted, oissued, mine;
1421
1422 if (session->s_trim_caps <= 0)
1423 return -1;
1424
1425 spin_lock(&ci->i_ceph_lock);
1426 mine = cap->issued | cap->implemented;
1427 used = __ceph_caps_used(ci);
1428 wanted = __ceph_caps_file_wanted(ci);
1429 oissued = __ceph_caps_issued_other(ci, cap);
1430
1431 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1432 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1433 ceph_cap_string(used), ceph_cap_string(wanted));
1434 if (cap == ci->i_auth_cap) {
1435 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1436 !list_empty(&ci->i_cap_snaps))
1437 goto out;
1438 if ((used | wanted) & CEPH_CAP_ANY_WR)
1439 goto out;
1440 }
1441 /* The inode has cached pages, but it's no longer used.
1442 * we can safely drop it */
1443 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1444 !(oissued & CEPH_CAP_FILE_CACHE)) {
1445 used = 0;
1446 oissued = 0;
1447 }
1448 if ((used | wanted) & ~oissued & mine)
1449 goto out; /* we need these caps */
1450
1451 session->s_trim_caps--;
1452 if (oissued) {
1453 /* we aren't the only cap.. just remove us */
1454 __ceph_remove_cap(cap, true);
1455 } else {
1456 /* try dropping referring dentries */
1457 spin_unlock(&ci->i_ceph_lock);
1458 d_prune_aliases(inode);
1459 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1460 inode, cap, atomic_read(&inode->i_count));
1461 return 0;
1462 }
1463
1464out:
1465 spin_unlock(&ci->i_ceph_lock);
1466 return 0;
1467}
1468
1469/*
1470 * Trim session cap count down to some max number.
1471 */
1472static int trim_caps(struct ceph_mds_client *mdsc,
1473 struct ceph_mds_session *session,
1474 int max_caps)
1475{
1476 int trim_caps = session->s_nr_caps - max_caps;
1477
1478 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1479 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1480 if (trim_caps > 0) {
1481 session->s_trim_caps = trim_caps;
1482 iterate_session_caps(session, trim_caps_cb, session);
1483 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1484 session->s_mds, session->s_nr_caps, max_caps,
1485 trim_caps - session->s_trim_caps);
1486 session->s_trim_caps = 0;
1487 }
1488
1489 ceph_send_cap_releases(mdsc, session);
1490 return 0;
1491}
1492
1493static int check_capsnap_flush(struct ceph_inode_info *ci,
1494 u64 want_snap_seq)
1495{
1496 int ret = 1;
1497 spin_lock(&ci->i_ceph_lock);
1498 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1499 struct ceph_cap_snap *capsnap =
1500 list_first_entry(&ci->i_cap_snaps,
1501 struct ceph_cap_snap, ci_item);
1502 ret = capsnap->follows >= want_snap_seq;
1503 }
1504 spin_unlock(&ci->i_ceph_lock);
1505 return ret;
1506}
1507
1508static int check_caps_flush(struct ceph_mds_client *mdsc,
1509 u64 want_flush_tid)
1510{
1511 struct rb_node *n;
1512 struct ceph_cap_flush *cf;
1513 int ret = 1;
1514
1515 spin_lock(&mdsc->cap_dirty_lock);
1516 n = rb_first(&mdsc->cap_flush_tree);
1517 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1518 if (cf && cf->tid <= want_flush_tid) {
1519 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1520 cf->tid, want_flush_tid);
1521 ret = 0;
1522 }
1523 spin_unlock(&mdsc->cap_dirty_lock);
1524 return ret;
1525}
1526
1527/*
1528 * flush all dirty inode data to disk.
1529 *
1530 * returns true if we've flushed through want_flush_tid
1531 */
1532static void wait_caps_flush(struct ceph_mds_client *mdsc,
1533 u64 want_flush_tid, u64 want_snap_seq)
1534{
1535 int mds;
1536
1537 dout("check_caps_flush want %llu snap want %llu\n",
1538 want_flush_tid, want_snap_seq);
1539 mutex_lock(&mdsc->mutex);
1540 for (mds = 0; mds < mdsc->max_sessions; ) {
1541 struct ceph_mds_session *session = mdsc->sessions[mds];
1542 struct inode *inode = NULL;
1543
1544 if (!session) {
1545 mds++;
1546 continue;
1547 }
1548 get_session(session);
1549 mutex_unlock(&mdsc->mutex);
1550
1551 mutex_lock(&session->s_mutex);
1552 if (!list_empty(&session->s_cap_snaps_flushing)) {
1553 struct ceph_cap_snap *capsnap =
1554 list_first_entry(&session->s_cap_snaps_flushing,
1555 struct ceph_cap_snap,
1556 flushing_item);
1557 struct ceph_inode_info *ci = capsnap->ci;
1558 if (!check_capsnap_flush(ci, want_snap_seq)) {
1559 dout("check_cap_flush still flushing snap %p "
1560 "follows %lld <= %lld to mds%d\n",
1561 &ci->vfs_inode, capsnap->follows,
1562 want_snap_seq, mds);
1563 inode = igrab(&ci->vfs_inode);
1564 }
1565 }
1566 mutex_unlock(&session->s_mutex);
1567 ceph_put_mds_session(session);
1568
1569 if (inode) {
1570 wait_event(mdsc->cap_flushing_wq,
1571 check_capsnap_flush(ceph_inode(inode),
1572 want_snap_seq));
1573 iput(inode);
1574 } else {
1575 mds++;
1576 }
1577
1578 mutex_lock(&mdsc->mutex);
1579 }
1580 mutex_unlock(&mdsc->mutex);
1581
1582 wait_event(mdsc->cap_flushing_wq,
1583 check_caps_flush(mdsc, want_flush_tid));
1584
1585 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1586}
1587
1588/*
1589 * called under s_mutex
1590 */
1591void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1592 struct ceph_mds_session *session)
1593{
1594 struct ceph_msg *msg = NULL;
1595 struct ceph_mds_cap_release *head;
1596 struct ceph_mds_cap_item *item;
1597 struct ceph_cap *cap;
1598 LIST_HEAD(tmp_list);
1599 int num_cap_releases;
1600
1601 spin_lock(&session->s_cap_lock);
1602again:
1603 list_splice_init(&session->s_cap_releases, &tmp_list);
1604 num_cap_releases = session->s_num_cap_releases;
1605 session->s_num_cap_releases = 0;
1606 spin_unlock(&session->s_cap_lock);
1607
1608 while (!list_empty(&tmp_list)) {
1609 if (!msg) {
1610 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1611 PAGE_SIZE, GFP_NOFS, false);
1612 if (!msg)
1613 goto out_err;
1614 head = msg->front.iov_base;
1615 head->num = cpu_to_le32(0);
1616 msg->front.iov_len = sizeof(*head);
1617 }
1618 cap = list_first_entry(&tmp_list, struct ceph_cap,
1619 session_caps);
1620 list_del(&cap->session_caps);
1621 num_cap_releases--;
1622
1623 head = msg->front.iov_base;
1624 le32_add_cpu(&head->num, 1);
1625 item = msg->front.iov_base + msg->front.iov_len;
1626 item->ino = cpu_to_le64(cap->cap_ino);
1627 item->cap_id = cpu_to_le64(cap->cap_id);
1628 item->migrate_seq = cpu_to_le32(cap->mseq);
1629 item->seq = cpu_to_le32(cap->issue_seq);
1630 msg->front.iov_len += sizeof(*item);
1631
1632 ceph_put_cap(mdsc, cap);
1633
1634 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1635 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1636 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1637 ceph_con_send(&session->s_con, msg);
1638 msg = NULL;
1639 }
1640 }
1641
1642 BUG_ON(num_cap_releases != 0);
1643
1644 spin_lock(&session->s_cap_lock);
1645 if (!list_empty(&session->s_cap_releases))
1646 goto again;
1647 spin_unlock(&session->s_cap_lock);
1648
1649 if (msg) {
1650 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1651 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1652 ceph_con_send(&session->s_con, msg);
1653 }
1654 return;
1655out_err:
1656 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1657 session->s_mds);
1658 spin_lock(&session->s_cap_lock);
1659 list_splice(&tmp_list, &session->s_cap_releases);
1660 session->s_num_cap_releases += num_cap_releases;
1661 spin_unlock(&session->s_cap_lock);
1662}
1663
1664/*
1665 * requests
1666 */
1667
1668int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1669 struct inode *dir)
1670{
1671 struct ceph_inode_info *ci = ceph_inode(dir);
1672 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1673 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1674 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1675 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1676 int order, num_entries;
1677
1678 spin_lock(&ci->i_ceph_lock);
1679 num_entries = ci->i_files + ci->i_subdirs;
1680 spin_unlock(&ci->i_ceph_lock);
1681 num_entries = max(num_entries, 1);
1682 num_entries = min(num_entries, opt->max_readdir);
1683
1684 order = get_order(size * num_entries);
1685 while (order >= 0) {
1686 rinfo->dir_in = (void*)__get_free_pages(GFP_KERNEL |
1687 __GFP_NOWARN,
1688 order);
1689 if (rinfo->dir_in)
1690 break;
1691 order--;
1692 }
1693 if (!rinfo->dir_in)
1694 return -ENOMEM;
1695
1696 num_entries = (PAGE_SIZE << order) / size;
1697 num_entries = min(num_entries, opt->max_readdir);
1698
1699 rinfo->dir_buf_size = PAGE_SIZE << order;
1700 req->r_num_caps = num_entries + 1;
1701 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1702 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1703 return 0;
1704}
1705
1706/*
1707 * Create an mds request.
1708 */
1709struct ceph_mds_request *
1710ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1711{
1712 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1713
1714 if (!req)
1715 return ERR_PTR(-ENOMEM);
1716
1717 mutex_init(&req->r_fill_mutex);
1718 req->r_mdsc = mdsc;
1719 req->r_started = jiffies;
1720 req->r_resend_mds = -1;
1721 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1722 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1723 req->r_fmode = -1;
1724 kref_init(&req->r_kref);
1725 INIT_LIST_HEAD(&req->r_wait);
1726 init_completion(&req->r_completion);
1727 init_completion(&req->r_safe_completion);
1728 INIT_LIST_HEAD(&req->r_unsafe_item);
1729
1730 req->r_stamp = current_fs_time(mdsc->fsc->sb);
1731
1732 req->r_op = op;
1733 req->r_direct_mode = mode;
1734 return req;
1735}
1736
1737/*
1738 * return oldest (lowest) request, tid in request tree, 0 if none.
1739 *
1740 * called under mdsc->mutex.
1741 */
1742static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1743{
1744 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1745 return NULL;
1746 return rb_entry(rb_first(&mdsc->request_tree),
1747 struct ceph_mds_request, r_node);
1748}
1749
1750static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1751{
1752 return mdsc->oldest_tid;
1753}
1754
1755/*
1756 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1757 * on build_path_from_dentry in fs/cifs/dir.c.
1758 *
1759 * If @stop_on_nosnap, generate path relative to the first non-snapped
1760 * inode.
1761 *
1762 * Encode hidden .snap dirs as a double /, i.e.
1763 * foo/.snap/bar -> foo//bar
1764 */
1765char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1766 int stop_on_nosnap)
1767{
1768 struct dentry *temp;
1769 char *path;
1770 int len, pos;
1771 unsigned seq;
1772
1773 if (dentry == NULL)
1774 return ERR_PTR(-EINVAL);
1775
1776retry:
1777 len = 0;
1778 seq = read_seqbegin(&rename_lock);
1779 rcu_read_lock();
1780 for (temp = dentry; !IS_ROOT(temp);) {
1781 struct inode *inode = d_inode(temp);
1782 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1783 len++; /* slash only */
1784 else if (stop_on_nosnap && inode &&
1785 ceph_snap(inode) == CEPH_NOSNAP)
1786 break;
1787 else
1788 len += 1 + temp->d_name.len;
1789 temp = temp->d_parent;
1790 }
1791 rcu_read_unlock();
1792 if (len)
1793 len--; /* no leading '/' */
1794
1795 path = kmalloc(len+1, GFP_NOFS);
1796 if (path == NULL)
1797 return ERR_PTR(-ENOMEM);
1798 pos = len;
1799 path[pos] = 0; /* trailing null */
1800 rcu_read_lock();
1801 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1802 struct inode *inode;
1803
1804 spin_lock(&temp->d_lock);
1805 inode = d_inode(temp);
1806 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1807 dout("build_path path+%d: %p SNAPDIR\n",
1808 pos, temp);
1809 } else if (stop_on_nosnap && inode &&
1810 ceph_snap(inode) == CEPH_NOSNAP) {
1811 spin_unlock(&temp->d_lock);
1812 break;
1813 } else {
1814 pos -= temp->d_name.len;
1815 if (pos < 0) {
1816 spin_unlock(&temp->d_lock);
1817 break;
1818 }
1819 strncpy(path + pos, temp->d_name.name,
1820 temp->d_name.len);
1821 }
1822 spin_unlock(&temp->d_lock);
1823 if (pos)
1824 path[--pos] = '/';
1825 temp = temp->d_parent;
1826 }
1827 rcu_read_unlock();
1828 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1829 pr_err("build_path did not end path lookup where "
1830 "expected, namelen is %d, pos is %d\n", len, pos);
1831 /* presumably this is only possible if racing with a
1832 rename of one of the parent directories (we can not
1833 lock the dentries above us to prevent this, but
1834 retrying should be harmless) */
1835 kfree(path);
1836 goto retry;
1837 }
1838
1839 *base = ceph_ino(d_inode(temp));
1840 *plen = len;
1841 dout("build_path on %p %d built %llx '%.*s'\n",
1842 dentry, d_count(dentry), *base, len, path);
1843 return path;
1844}
1845
1846static int build_dentry_path(struct dentry *dentry,
1847 const char **ppath, int *ppathlen, u64 *pino,
1848 int *pfreepath)
1849{
1850 char *path;
1851
1852 if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1853 *pino = ceph_ino(d_inode(dentry->d_parent));
1854 *ppath = dentry->d_name.name;
1855 *ppathlen = dentry->d_name.len;
1856 return 0;
1857 }
1858 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1859 if (IS_ERR(path))
1860 return PTR_ERR(path);
1861 *ppath = path;
1862 *pfreepath = 1;
1863 return 0;
1864}
1865
1866static int build_inode_path(struct inode *inode,
1867 const char **ppath, int *ppathlen, u64 *pino,
1868 int *pfreepath)
1869{
1870 struct dentry *dentry;
1871 char *path;
1872
1873 if (ceph_snap(inode) == CEPH_NOSNAP) {
1874 *pino = ceph_ino(inode);
1875 *ppathlen = 0;
1876 return 0;
1877 }
1878 dentry = d_find_alias(inode);
1879 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1880 dput(dentry);
1881 if (IS_ERR(path))
1882 return PTR_ERR(path);
1883 *ppath = path;
1884 *pfreepath = 1;
1885 return 0;
1886}
1887
1888/*
1889 * request arguments may be specified via an inode *, a dentry *, or
1890 * an explicit ino+path.
1891 */
1892static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1893 const char *rpath, u64 rino,
1894 const char **ppath, int *pathlen,
1895 u64 *ino, int *freepath)
1896{
1897 int r = 0;
1898
1899 if (rinode) {
1900 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1901 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1902 ceph_snap(rinode));
1903 } else if (rdentry) {
1904 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1905 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1906 *ppath);
1907 } else if (rpath || rino) {
1908 *ino = rino;
1909 *ppath = rpath;
1910 *pathlen = rpath ? strlen(rpath) : 0;
1911 dout(" path %.*s\n", *pathlen, rpath);
1912 }
1913
1914 return r;
1915}
1916
1917/*
1918 * called under mdsc->mutex
1919 */
1920static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1921 struct ceph_mds_request *req,
1922 int mds, bool drop_cap_releases)
1923{
1924 struct ceph_msg *msg;
1925 struct ceph_mds_request_head *head;
1926 const char *path1 = NULL;
1927 const char *path2 = NULL;
1928 u64 ino1 = 0, ino2 = 0;
1929 int pathlen1 = 0, pathlen2 = 0;
1930 int freepath1 = 0, freepath2 = 0;
1931 int len;
1932 u16 releases;
1933 void *p, *end;
1934 int ret;
1935
1936 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1937 req->r_path1, req->r_ino1.ino,
1938 &path1, &pathlen1, &ino1, &freepath1);
1939 if (ret < 0) {
1940 msg = ERR_PTR(ret);
1941 goto out;
1942 }
1943
1944 ret = set_request_path_attr(NULL, req->r_old_dentry,
1945 req->r_path2, req->r_ino2.ino,
1946 &path2, &pathlen2, &ino2, &freepath2);
1947 if (ret < 0) {
1948 msg = ERR_PTR(ret);
1949 goto out_free1;
1950 }
1951
1952 len = sizeof(*head) +
1953 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1954 sizeof(struct ceph_timespec);
1955
1956 /* calculate (max) length for cap releases */
1957 len += sizeof(struct ceph_mds_request_release) *
1958 (!!req->r_inode_drop + !!req->r_dentry_drop +
1959 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1960 if (req->r_dentry_drop)
1961 len += req->r_dentry->d_name.len;
1962 if (req->r_old_dentry_drop)
1963 len += req->r_old_dentry->d_name.len;
1964
1965 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1966 if (!msg) {
1967 msg = ERR_PTR(-ENOMEM);
1968 goto out_free2;
1969 }
1970
1971 msg->hdr.version = cpu_to_le16(2);
1972 msg->hdr.tid = cpu_to_le64(req->r_tid);
1973
1974 head = msg->front.iov_base;
1975 p = msg->front.iov_base + sizeof(*head);
1976 end = msg->front.iov_base + msg->front.iov_len;
1977
1978 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1979 head->op = cpu_to_le32(req->r_op);
1980 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1981 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1982 head->args = req->r_args;
1983
1984 ceph_encode_filepath(&p, end, ino1, path1);
1985 ceph_encode_filepath(&p, end, ino2, path2);
1986
1987 /* make note of release offset, in case we need to replay */
1988 req->r_request_release_offset = p - msg->front.iov_base;
1989
1990 /* cap releases */
1991 releases = 0;
1992 if (req->r_inode_drop)
1993 releases += ceph_encode_inode_release(&p,
1994 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1995 mds, req->r_inode_drop, req->r_inode_unless, 0);
1996 if (req->r_dentry_drop)
1997 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1998 mds, req->r_dentry_drop, req->r_dentry_unless);
1999 if (req->r_old_dentry_drop)
2000 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2001 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
2002 if (req->r_old_inode_drop)
2003 releases += ceph_encode_inode_release(&p,
2004 d_inode(req->r_old_dentry),
2005 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2006
2007 if (drop_cap_releases) {
2008 releases = 0;
2009 p = msg->front.iov_base + req->r_request_release_offset;
2010 }
2011
2012 head->num_releases = cpu_to_le16(releases);
2013
2014 /* time stamp */
2015 {
2016 struct ceph_timespec ts;
2017 ceph_encode_timespec(&ts, &req->r_stamp);
2018 ceph_encode_copy(&p, &ts, sizeof(ts));
2019 }
2020
2021 BUG_ON(p > end);
2022 msg->front.iov_len = p - msg->front.iov_base;
2023 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2024
2025 if (req->r_pagelist) {
2026 struct ceph_pagelist *pagelist = req->r_pagelist;
2027 atomic_inc(&pagelist->refcnt);
2028 ceph_msg_data_add_pagelist(msg, pagelist);
2029 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2030 } else {
2031 msg->hdr.data_len = 0;
2032 }
2033
2034 msg->hdr.data_off = cpu_to_le16(0);
2035
2036out_free2:
2037 if (freepath2)
2038 kfree((char *)path2);
2039out_free1:
2040 if (freepath1)
2041 kfree((char *)path1);
2042out:
2043 return msg;
2044}
2045
2046/*
2047 * called under mdsc->mutex if error, under no mutex if
2048 * success.
2049 */
2050static void complete_request(struct ceph_mds_client *mdsc,
2051 struct ceph_mds_request *req)
2052{
2053 if (req->r_callback)
2054 req->r_callback(mdsc, req);
2055 else
2056 complete_all(&req->r_completion);
2057}
2058
2059/*
2060 * called under mdsc->mutex
2061 */
2062static int __prepare_send_request(struct ceph_mds_client *mdsc,
2063 struct ceph_mds_request *req,
2064 int mds, bool drop_cap_releases)
2065{
2066 struct ceph_mds_request_head *rhead;
2067 struct ceph_msg *msg;
2068 int flags = 0;
2069
2070 req->r_attempts++;
2071 if (req->r_inode) {
2072 struct ceph_cap *cap =
2073 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2074
2075 if (cap)
2076 req->r_sent_on_mseq = cap->mseq;
2077 else
2078 req->r_sent_on_mseq = -1;
2079 }
2080 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2081 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2082
2083 if (req->r_got_unsafe) {
2084 void *p;
2085 /*
2086 * Replay. Do not regenerate message (and rebuild
2087 * paths, etc.); just use the original message.
2088 * Rebuilding paths will break for renames because
2089 * d_move mangles the src name.
2090 */
2091 msg = req->r_request;
2092 rhead = msg->front.iov_base;
2093
2094 flags = le32_to_cpu(rhead->flags);
2095 flags |= CEPH_MDS_FLAG_REPLAY;
2096 rhead->flags = cpu_to_le32(flags);
2097
2098 if (req->r_target_inode)
2099 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2100
2101 rhead->num_retry = req->r_attempts - 1;
2102
2103 /* remove cap/dentry releases from message */
2104 rhead->num_releases = 0;
2105
2106 /* time stamp */
2107 p = msg->front.iov_base + req->r_request_release_offset;
2108 {
2109 struct ceph_timespec ts;
2110 ceph_encode_timespec(&ts, &req->r_stamp);
2111 ceph_encode_copy(&p, &ts, sizeof(ts));
2112 }
2113
2114 msg->front.iov_len = p - msg->front.iov_base;
2115 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2116 return 0;
2117 }
2118
2119 if (req->r_request) {
2120 ceph_msg_put(req->r_request);
2121 req->r_request = NULL;
2122 }
2123 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2124 if (IS_ERR(msg)) {
2125 req->r_err = PTR_ERR(msg);
2126 return PTR_ERR(msg);
2127 }
2128 req->r_request = msg;
2129
2130 rhead = msg->front.iov_base;
2131 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2132 if (req->r_got_unsafe)
2133 flags |= CEPH_MDS_FLAG_REPLAY;
2134 if (req->r_locked_dir)
2135 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2136 rhead->flags = cpu_to_le32(flags);
2137 rhead->num_fwd = req->r_num_fwd;
2138 rhead->num_retry = req->r_attempts - 1;
2139 rhead->ino = 0;
2140
2141 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2142 return 0;
2143}
2144
2145/*
2146 * send request, or put it on the appropriate wait list.
2147 */
2148static int __do_request(struct ceph_mds_client *mdsc,
2149 struct ceph_mds_request *req)
2150{
2151 struct ceph_mds_session *session = NULL;
2152 int mds = -1;
2153 int err = 0;
2154
2155 if (req->r_err || req->r_got_result) {
2156 if (req->r_aborted)
2157 __unregister_request(mdsc, req);
2158 goto out;
2159 }
2160
2161 if (req->r_timeout &&
2162 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2163 dout("do_request timed out\n");
2164 err = -EIO;
2165 goto finish;
2166 }
2167 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2168 dout("do_request forced umount\n");
2169 err = -EIO;
2170 goto finish;
2171 }
2172
2173 put_request_session(req);
2174
2175 mds = __choose_mds(mdsc, req);
2176 if (mds < 0 ||
2177 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2178 dout("do_request no mds or not active, waiting for map\n");
2179 list_add(&req->r_wait, &mdsc->waiting_for_map);
2180 goto out;
2181 }
2182
2183 /* get, open session */
2184 session = __ceph_lookup_mds_session(mdsc, mds);
2185 if (!session) {
2186 session = register_session(mdsc, mds);
2187 if (IS_ERR(session)) {
2188 err = PTR_ERR(session);
2189 goto finish;
2190 }
2191 }
2192 req->r_session = get_session(session);
2193
2194 dout("do_request mds%d session %p state %s\n", mds, session,
2195 ceph_session_state_name(session->s_state));
2196 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2197 session->s_state != CEPH_MDS_SESSION_HUNG) {
2198 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2199 session->s_state == CEPH_MDS_SESSION_CLOSING)
2200 __open_session(mdsc, session);
2201 list_add(&req->r_wait, &session->s_waiting);
2202 goto out_session;
2203 }
2204
2205 /* send request */
2206 req->r_resend_mds = -1; /* forget any previous mds hint */
2207
2208 if (req->r_request_started == 0) /* note request start time */
2209 req->r_request_started = jiffies;
2210
2211 err = __prepare_send_request(mdsc, req, mds, false);
2212 if (!err) {
2213 ceph_msg_get(req->r_request);
2214 ceph_con_send(&session->s_con, req->r_request);
2215 }
2216
2217out_session:
2218 ceph_put_mds_session(session);
2219finish:
2220 if (err) {
2221 dout("__do_request early error %d\n", err);
2222 req->r_err = err;
2223 complete_request(mdsc, req);
2224 __unregister_request(mdsc, req);
2225 }
2226out:
2227 return err;
2228}
2229
2230/*
2231 * called under mdsc->mutex
2232 */
2233static void __wake_requests(struct ceph_mds_client *mdsc,
2234 struct list_head *head)
2235{
2236 struct ceph_mds_request *req;
2237 LIST_HEAD(tmp_list);
2238
2239 list_splice_init(head, &tmp_list);
2240
2241 while (!list_empty(&tmp_list)) {
2242 req = list_entry(tmp_list.next,
2243 struct ceph_mds_request, r_wait);
2244 list_del_init(&req->r_wait);
2245 dout(" wake request %p tid %llu\n", req, req->r_tid);
2246 __do_request(mdsc, req);
2247 }
2248}
2249
2250/*
2251 * Wake up threads with requests pending for @mds, so that they can
2252 * resubmit their requests to a possibly different mds.
2253 */
2254static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2255{
2256 struct ceph_mds_request *req;
2257 struct rb_node *p = rb_first(&mdsc->request_tree);
2258
2259 dout("kick_requests mds%d\n", mds);
2260 while (p) {
2261 req = rb_entry(p, struct ceph_mds_request, r_node);
2262 p = rb_next(p);
2263 if (req->r_got_unsafe)
2264 continue;
2265 if (req->r_attempts > 0)
2266 continue; /* only new requests */
2267 if (req->r_session &&
2268 req->r_session->s_mds == mds) {
2269 dout(" kicking tid %llu\n", req->r_tid);
2270 list_del_init(&req->r_wait);
2271 __do_request(mdsc, req);
2272 }
2273 }
2274}
2275
2276void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2277 struct ceph_mds_request *req)
2278{
2279 dout("submit_request on %p\n", req);
2280 mutex_lock(&mdsc->mutex);
2281 __register_request(mdsc, req, NULL);
2282 __do_request(mdsc, req);
2283 mutex_unlock(&mdsc->mutex);
2284}
2285
2286/*
2287 * Synchrously perform an mds request. Take care of all of the
2288 * session setup, forwarding, retry details.
2289 */
2290int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2291 struct inode *dir,
2292 struct ceph_mds_request *req)
2293{
2294 int err;
2295
2296 dout("do_request on %p\n", req);
2297
2298 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2299 if (req->r_inode)
2300 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2301 if (req->r_locked_dir)
2302 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2303 if (req->r_old_dentry_dir)
2304 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2305 CEPH_CAP_PIN);
2306
2307 /* deny access to directories with pool_ns layouts */
2308 if (req->r_inode && S_ISDIR(req->r_inode->i_mode) &&
2309 ceph_inode(req->r_inode)->i_pool_ns_len)
2310 return -EIO;
2311 if (req->r_locked_dir &&
2312 ceph_inode(req->r_locked_dir)->i_pool_ns_len)
2313 return -EIO;
2314
2315 /* issue */
2316 mutex_lock(&mdsc->mutex);
2317 __register_request(mdsc, req, dir);
2318 __do_request(mdsc, req);
2319
2320 if (req->r_err) {
2321 err = req->r_err;
2322 goto out;
2323 }
2324
2325 /* wait */
2326 mutex_unlock(&mdsc->mutex);
2327 dout("do_request waiting\n");
2328 if (!req->r_timeout && req->r_wait_for_completion) {
2329 err = req->r_wait_for_completion(mdsc, req);
2330 } else {
2331 long timeleft = wait_for_completion_killable_timeout(
2332 &req->r_completion,
2333 ceph_timeout_jiffies(req->r_timeout));
2334 if (timeleft > 0)
2335 err = 0;
2336 else if (!timeleft)
2337 err = -EIO; /* timed out */
2338 else
2339 err = timeleft; /* killed */
2340 }
2341 dout("do_request waited, got %d\n", err);
2342 mutex_lock(&mdsc->mutex);
2343
2344 /* only abort if we didn't race with a real reply */
2345 if (req->r_got_result) {
2346 err = le32_to_cpu(req->r_reply_info.head->result);
2347 } else if (err < 0) {
2348 dout("aborted request %lld with %d\n", req->r_tid, err);
2349
2350 /*
2351 * ensure we aren't running concurrently with
2352 * ceph_fill_trace or ceph_readdir_prepopulate, which
2353 * rely on locks (dir mutex) held by our caller.
2354 */
2355 mutex_lock(&req->r_fill_mutex);
2356 req->r_err = err;
2357 req->r_aborted = true;
2358 mutex_unlock(&req->r_fill_mutex);
2359
2360 if (req->r_locked_dir &&
2361 (req->r_op & CEPH_MDS_OP_WRITE))
2362 ceph_invalidate_dir_request(req);
2363 } else {
2364 err = req->r_err;
2365 }
2366
2367out:
2368 mutex_unlock(&mdsc->mutex);
2369 dout("do_request %p done, result %d\n", req, err);
2370 return err;
2371}
2372
2373/*
2374 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2375 * namespace request.
2376 */
2377void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2378{
2379 struct inode *inode = req->r_locked_dir;
2380
2381 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2382
2383 ceph_dir_clear_complete(inode);
2384 if (req->r_dentry)
2385 ceph_invalidate_dentry_lease(req->r_dentry);
2386 if (req->r_old_dentry)
2387 ceph_invalidate_dentry_lease(req->r_old_dentry);
2388}
2389
2390/*
2391 * Handle mds reply.
2392 *
2393 * We take the session mutex and parse and process the reply immediately.
2394 * This preserves the logical ordering of replies, capabilities, etc., sent
2395 * by the MDS as they are applied to our local cache.
2396 */
2397static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2398{
2399 struct ceph_mds_client *mdsc = session->s_mdsc;
2400 struct ceph_mds_request *req;
2401 struct ceph_mds_reply_head *head = msg->front.iov_base;
2402 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2403 struct ceph_snap_realm *realm;
2404 u64 tid;
2405 int err, result;
2406 int mds = session->s_mds;
2407
2408 if (msg->front.iov_len < sizeof(*head)) {
2409 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2410 ceph_msg_dump(msg);
2411 return;
2412 }
2413
2414 /* get request, session */
2415 tid = le64_to_cpu(msg->hdr.tid);
2416 mutex_lock(&mdsc->mutex);
2417 req = __lookup_request(mdsc, tid);
2418 if (!req) {
2419 dout("handle_reply on unknown tid %llu\n", tid);
2420 mutex_unlock(&mdsc->mutex);
2421 return;
2422 }
2423 dout("handle_reply %p\n", req);
2424
2425 /* correct session? */
2426 if (req->r_session != session) {
2427 pr_err("mdsc_handle_reply got %llu on session mds%d"
2428 " not mds%d\n", tid, session->s_mds,
2429 req->r_session ? req->r_session->s_mds : -1);
2430 mutex_unlock(&mdsc->mutex);
2431 goto out;
2432 }
2433
2434 /* dup? */
2435 if ((req->r_got_unsafe && !head->safe) ||
2436 (req->r_got_safe && head->safe)) {
2437 pr_warn("got a dup %s reply on %llu from mds%d\n",
2438 head->safe ? "safe" : "unsafe", tid, mds);
2439 mutex_unlock(&mdsc->mutex);
2440 goto out;
2441 }
2442 if (req->r_got_safe) {
2443 pr_warn("got unsafe after safe on %llu from mds%d\n",
2444 tid, mds);
2445 mutex_unlock(&mdsc->mutex);
2446 goto out;
2447 }
2448
2449 result = le32_to_cpu(head->result);
2450
2451 /*
2452 * Handle an ESTALE
2453 * if we're not talking to the authority, send to them
2454 * if the authority has changed while we weren't looking,
2455 * send to new authority
2456 * Otherwise we just have to return an ESTALE
2457 */
2458 if (result == -ESTALE) {
2459 dout("got ESTALE on request %llu", req->r_tid);
2460 req->r_resend_mds = -1;
2461 if (req->r_direct_mode != USE_AUTH_MDS) {
2462 dout("not using auth, setting for that now");
2463 req->r_direct_mode = USE_AUTH_MDS;
2464 __do_request(mdsc, req);
2465 mutex_unlock(&mdsc->mutex);
2466 goto out;
2467 } else {
2468 int mds = __choose_mds(mdsc, req);
2469 if (mds >= 0 && mds != req->r_session->s_mds) {
2470 dout("but auth changed, so resending");
2471 __do_request(mdsc, req);
2472 mutex_unlock(&mdsc->mutex);
2473 goto out;
2474 }
2475 }
2476 dout("have to return ESTALE on request %llu", req->r_tid);
2477 }
2478
2479
2480 if (head->safe) {
2481 req->r_got_safe = true;
2482 __unregister_request(mdsc, req);
2483
2484 if (req->r_got_unsafe) {
2485 /*
2486 * We already handled the unsafe response, now do the
2487 * cleanup. No need to examine the response; the MDS
2488 * doesn't include any result info in the safe
2489 * response. And even if it did, there is nothing
2490 * useful we could do with a revised return value.
2491 */
2492 dout("got safe reply %llu, mds%d\n", tid, mds);
2493 list_del_init(&req->r_unsafe_item);
2494
2495 /* last unsafe request during umount? */
2496 if (mdsc->stopping && !__get_oldest_req(mdsc))
2497 complete_all(&mdsc->safe_umount_waiters);
2498 mutex_unlock(&mdsc->mutex);
2499 goto out;
2500 }
2501 } else {
2502 req->r_got_unsafe = true;
2503 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2504 if (req->r_unsafe_dir) {
2505 struct ceph_inode_info *ci =
2506 ceph_inode(req->r_unsafe_dir);
2507 spin_lock(&ci->i_unsafe_lock);
2508 list_add_tail(&req->r_unsafe_dir_item,
2509 &ci->i_unsafe_dirops);
2510 spin_unlock(&ci->i_unsafe_lock);
2511 }
2512 }
2513
2514 dout("handle_reply tid %lld result %d\n", tid, result);
2515 rinfo = &req->r_reply_info;
2516 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2517 mutex_unlock(&mdsc->mutex);
2518
2519 mutex_lock(&session->s_mutex);
2520 if (err < 0) {
2521 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2522 ceph_msg_dump(msg);
2523 goto out_err;
2524 }
2525
2526 /* snap trace */
2527 realm = NULL;
2528 if (rinfo->snapblob_len) {
2529 down_write(&mdsc->snap_rwsem);
2530 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2531 rinfo->snapblob + rinfo->snapblob_len,
2532 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2533 &realm);
2534 downgrade_write(&mdsc->snap_rwsem);
2535 } else {
2536 down_read(&mdsc->snap_rwsem);
2537 }
2538
2539 /* insert trace into our cache */
2540 mutex_lock(&req->r_fill_mutex);
2541 current->journal_info = req;
2542 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2543 if (err == 0) {
2544 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2545 req->r_op == CEPH_MDS_OP_LSSNAP))
2546 ceph_readdir_prepopulate(req, req->r_session);
2547 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2548 }
2549 current->journal_info = NULL;
2550 mutex_unlock(&req->r_fill_mutex);
2551
2552 up_read(&mdsc->snap_rwsem);
2553 if (realm)
2554 ceph_put_snap_realm(mdsc, realm);
2555
2556 if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2557 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2558 spin_lock(&ci->i_unsafe_lock);
2559 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2560 spin_unlock(&ci->i_unsafe_lock);
2561 }
2562out_err:
2563 mutex_lock(&mdsc->mutex);
2564 if (!req->r_aborted) {
2565 if (err) {
2566 req->r_err = err;
2567 } else {
2568 req->r_reply = ceph_msg_get(msg);
2569 req->r_got_result = true;
2570 }
2571 } else {
2572 dout("reply arrived after request %lld was aborted\n", tid);
2573 }
2574 mutex_unlock(&mdsc->mutex);
2575
2576 mutex_unlock(&session->s_mutex);
2577
2578 /* kick calling process */
2579 complete_request(mdsc, req);
2580out:
2581 ceph_mdsc_put_request(req);
2582 return;
2583}
2584
2585
2586
2587/*
2588 * handle mds notification that our request has been forwarded.
2589 */
2590static void handle_forward(struct ceph_mds_client *mdsc,
2591 struct ceph_mds_session *session,
2592 struct ceph_msg *msg)
2593{
2594 struct ceph_mds_request *req;
2595 u64 tid = le64_to_cpu(msg->hdr.tid);
2596 u32 next_mds;
2597 u32 fwd_seq;
2598 int err = -EINVAL;
2599 void *p = msg->front.iov_base;
2600 void *end = p + msg->front.iov_len;
2601
2602 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2603 next_mds = ceph_decode_32(&p);
2604 fwd_seq = ceph_decode_32(&p);
2605
2606 mutex_lock(&mdsc->mutex);
2607 req = __lookup_request(mdsc, tid);
2608 if (!req) {
2609 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2610 goto out; /* dup reply? */
2611 }
2612
2613 if (req->r_aborted) {
2614 dout("forward tid %llu aborted, unregistering\n", tid);
2615 __unregister_request(mdsc, req);
2616 } else if (fwd_seq <= req->r_num_fwd) {
2617 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2618 tid, next_mds, req->r_num_fwd, fwd_seq);
2619 } else {
2620 /* resend. forward race not possible; mds would drop */
2621 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2622 BUG_ON(req->r_err);
2623 BUG_ON(req->r_got_result);
2624 req->r_attempts = 0;
2625 req->r_num_fwd = fwd_seq;
2626 req->r_resend_mds = next_mds;
2627 put_request_session(req);
2628 __do_request(mdsc, req);
2629 }
2630 ceph_mdsc_put_request(req);
2631out:
2632 mutex_unlock(&mdsc->mutex);
2633 return;
2634
2635bad:
2636 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2637}
2638
2639/*
2640 * handle a mds session control message
2641 */
2642static void handle_session(struct ceph_mds_session *session,
2643 struct ceph_msg *msg)
2644{
2645 struct ceph_mds_client *mdsc = session->s_mdsc;
2646 u32 op;
2647 u64 seq;
2648 int mds = session->s_mds;
2649 struct ceph_mds_session_head *h = msg->front.iov_base;
2650 int wake = 0;
2651
2652 /* decode */
2653 if (msg->front.iov_len != sizeof(*h))
2654 goto bad;
2655 op = le32_to_cpu(h->op);
2656 seq = le64_to_cpu(h->seq);
2657
2658 mutex_lock(&mdsc->mutex);
2659 if (op == CEPH_SESSION_CLOSE)
2660 __unregister_session(mdsc, session);
2661 /* FIXME: this ttl calculation is generous */
2662 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2663 mutex_unlock(&mdsc->mutex);
2664
2665 mutex_lock(&session->s_mutex);
2666
2667 dout("handle_session mds%d %s %p state %s seq %llu\n",
2668 mds, ceph_session_op_name(op), session,
2669 ceph_session_state_name(session->s_state), seq);
2670
2671 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2672 session->s_state = CEPH_MDS_SESSION_OPEN;
2673 pr_info("mds%d came back\n", session->s_mds);
2674 }
2675
2676 switch (op) {
2677 case CEPH_SESSION_OPEN:
2678 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2679 pr_info("mds%d reconnect success\n", session->s_mds);
2680 session->s_state = CEPH_MDS_SESSION_OPEN;
2681 renewed_caps(mdsc, session, 0);
2682 wake = 1;
2683 if (mdsc->stopping)
2684 __close_session(mdsc, session);
2685 break;
2686
2687 case CEPH_SESSION_RENEWCAPS:
2688 if (session->s_renew_seq == seq)
2689 renewed_caps(mdsc, session, 1);
2690 break;
2691
2692 case CEPH_SESSION_CLOSE:
2693 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2694 pr_info("mds%d reconnect denied\n", session->s_mds);
2695 cleanup_session_requests(mdsc, session);
2696 remove_session_caps(session);
2697 wake = 2; /* for good measure */
2698 wake_up_all(&mdsc->session_close_wq);
2699 break;
2700
2701 case CEPH_SESSION_STALE:
2702 pr_info("mds%d caps went stale, renewing\n",
2703 session->s_mds);
2704 spin_lock(&session->s_gen_ttl_lock);
2705 session->s_cap_gen++;
2706 session->s_cap_ttl = jiffies - 1;
2707 spin_unlock(&session->s_gen_ttl_lock);
2708 send_renew_caps(mdsc, session);
2709 break;
2710
2711 case CEPH_SESSION_RECALL_STATE:
2712 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2713 break;
2714
2715 case CEPH_SESSION_FLUSHMSG:
2716 send_flushmsg_ack(mdsc, session, seq);
2717 break;
2718
2719 case CEPH_SESSION_FORCE_RO:
2720 dout("force_session_readonly %p\n", session);
2721 spin_lock(&session->s_cap_lock);
2722 session->s_readonly = true;
2723 spin_unlock(&session->s_cap_lock);
2724 wake_up_session_caps(session, 0);
2725 break;
2726
2727 default:
2728 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2729 WARN_ON(1);
2730 }
2731
2732 mutex_unlock(&session->s_mutex);
2733 if (wake) {
2734 mutex_lock(&mdsc->mutex);
2735 __wake_requests(mdsc, &session->s_waiting);
2736 if (wake == 2)
2737 kick_requests(mdsc, mds);
2738 mutex_unlock(&mdsc->mutex);
2739 }
2740 return;
2741
2742bad:
2743 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2744 (int)msg->front.iov_len);
2745 ceph_msg_dump(msg);
2746 return;
2747}
2748
2749
2750/*
2751 * called under session->mutex.
2752 */
2753static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2754 struct ceph_mds_session *session)
2755{
2756 struct ceph_mds_request *req, *nreq;
2757 struct rb_node *p;
2758 int err;
2759
2760 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2761
2762 mutex_lock(&mdsc->mutex);
2763 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2764 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2765 if (!err) {
2766 ceph_msg_get(req->r_request);
2767 ceph_con_send(&session->s_con, req->r_request);
2768 }
2769 }
2770
2771 /*
2772 * also re-send old requests when MDS enters reconnect stage. So that MDS
2773 * can process completed request in clientreplay stage.
2774 */
2775 p = rb_first(&mdsc->request_tree);
2776 while (p) {
2777 req = rb_entry(p, struct ceph_mds_request, r_node);
2778 p = rb_next(p);
2779 if (req->r_got_unsafe)
2780 continue;
2781 if (req->r_attempts == 0)
2782 continue; /* only old requests */
2783 if (req->r_session &&
2784 req->r_session->s_mds == session->s_mds) {
2785 err = __prepare_send_request(mdsc, req,
2786 session->s_mds, true);
2787 if (!err) {
2788 ceph_msg_get(req->r_request);
2789 ceph_con_send(&session->s_con, req->r_request);
2790 }
2791 }
2792 }
2793 mutex_unlock(&mdsc->mutex);
2794}
2795
2796/*
2797 * Encode information about a cap for a reconnect with the MDS.
2798 */
2799static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2800 void *arg)
2801{
2802 union {
2803 struct ceph_mds_cap_reconnect v2;
2804 struct ceph_mds_cap_reconnect_v1 v1;
2805 } rec;
2806 size_t reclen;
2807 struct ceph_inode_info *ci;
2808 struct ceph_reconnect_state *recon_state = arg;
2809 struct ceph_pagelist *pagelist = recon_state->pagelist;
2810 char *path;
2811 int pathlen, err;
2812 u64 pathbase;
2813 struct dentry *dentry;
2814
2815 ci = cap->ci;
2816
2817 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2818 inode, ceph_vinop(inode), cap, cap->cap_id,
2819 ceph_cap_string(cap->issued));
2820 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2821 if (err)
2822 return err;
2823
2824 dentry = d_find_alias(inode);
2825 if (dentry) {
2826 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2827 if (IS_ERR(path)) {
2828 err = PTR_ERR(path);
2829 goto out_dput;
2830 }
2831 } else {
2832 path = NULL;
2833 pathlen = 0;
2834 }
2835 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2836 if (err)
2837 goto out_free;
2838
2839 spin_lock(&ci->i_ceph_lock);
2840 cap->seq = 0; /* reset cap seq */
2841 cap->issue_seq = 0; /* and issue_seq */
2842 cap->mseq = 0; /* and migrate_seq */
2843 cap->cap_gen = cap->session->s_cap_gen;
2844
2845 if (recon_state->flock) {
2846 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2847 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2848 rec.v2.issued = cpu_to_le32(cap->issued);
2849 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2850 rec.v2.pathbase = cpu_to_le64(pathbase);
2851 rec.v2.flock_len = 0;
2852 reclen = sizeof(rec.v2);
2853 } else {
2854 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2855 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2856 rec.v1.issued = cpu_to_le32(cap->issued);
2857 rec.v1.size = cpu_to_le64(inode->i_size);
2858 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2859 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2860 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2861 rec.v1.pathbase = cpu_to_le64(pathbase);
2862 reclen = sizeof(rec.v1);
2863 }
2864 spin_unlock(&ci->i_ceph_lock);
2865
2866 if (recon_state->flock) {
2867 int num_fcntl_locks, num_flock_locks;
2868 struct ceph_filelock *flocks;
2869
2870encode_again:
2871 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2872 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2873 sizeof(struct ceph_filelock), GFP_NOFS);
2874 if (!flocks) {
2875 err = -ENOMEM;
2876 goto out_free;
2877 }
2878 err = ceph_encode_locks_to_buffer(inode, flocks,
2879 num_fcntl_locks,
2880 num_flock_locks);
2881 if (err) {
2882 kfree(flocks);
2883 if (err == -ENOSPC)
2884 goto encode_again;
2885 goto out_free;
2886 }
2887 /*
2888 * number of encoded locks is stable, so copy to pagelist
2889 */
2890 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2891 (num_fcntl_locks+num_flock_locks) *
2892 sizeof(struct ceph_filelock));
2893 err = ceph_pagelist_append(pagelist, &rec, reclen);
2894 if (!err)
2895 err = ceph_locks_to_pagelist(flocks, pagelist,
2896 num_fcntl_locks,
2897 num_flock_locks);
2898 kfree(flocks);
2899 } else {
2900 err = ceph_pagelist_append(pagelist, &rec, reclen);
2901 }
2902
2903 recon_state->nr_caps++;
2904out_free:
2905 kfree(path);
2906out_dput:
2907 dput(dentry);
2908 return err;
2909}
2910
2911
2912/*
2913 * If an MDS fails and recovers, clients need to reconnect in order to
2914 * reestablish shared state. This includes all caps issued through
2915 * this session _and_ the snap_realm hierarchy. Because it's not
2916 * clear which snap realms the mds cares about, we send everything we
2917 * know about.. that ensures we'll then get any new info the
2918 * recovering MDS might have.
2919 *
2920 * This is a relatively heavyweight operation, but it's rare.
2921 *
2922 * called with mdsc->mutex held.
2923 */
2924static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2925 struct ceph_mds_session *session)
2926{
2927 struct ceph_msg *reply;
2928 struct rb_node *p;
2929 int mds = session->s_mds;
2930 int err = -ENOMEM;
2931 int s_nr_caps;
2932 struct ceph_pagelist *pagelist;
2933 struct ceph_reconnect_state recon_state;
2934
2935 pr_info("mds%d reconnect start\n", mds);
2936
2937 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2938 if (!pagelist)
2939 goto fail_nopagelist;
2940 ceph_pagelist_init(pagelist);
2941
2942 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2943 if (!reply)
2944 goto fail_nomsg;
2945
2946 mutex_lock(&session->s_mutex);
2947 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2948 session->s_seq = 0;
2949
2950 dout("session %p state %s\n", session,
2951 ceph_session_state_name(session->s_state));
2952
2953 spin_lock(&session->s_gen_ttl_lock);
2954 session->s_cap_gen++;
2955 spin_unlock(&session->s_gen_ttl_lock);
2956
2957 spin_lock(&session->s_cap_lock);
2958 /* don't know if session is readonly */
2959 session->s_readonly = 0;
2960 /*
2961 * notify __ceph_remove_cap() that we are composing cap reconnect.
2962 * If a cap get released before being added to the cap reconnect,
2963 * __ceph_remove_cap() should skip queuing cap release.
2964 */
2965 session->s_cap_reconnect = 1;
2966 /* drop old cap expires; we're about to reestablish that state */
2967 cleanup_cap_releases(mdsc, session);
2968
2969 /* trim unused caps to reduce MDS's cache rejoin time */
2970 if (mdsc->fsc->sb->s_root)
2971 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2972
2973 ceph_con_close(&session->s_con);
2974 ceph_con_open(&session->s_con,
2975 CEPH_ENTITY_TYPE_MDS, mds,
2976 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2977
2978 /* replay unsafe requests */
2979 replay_unsafe_requests(mdsc, session);
2980
2981 down_read(&mdsc->snap_rwsem);
2982
2983 /* traverse this session's caps */
2984 s_nr_caps = session->s_nr_caps;
2985 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2986 if (err)
2987 goto fail;
2988
2989 recon_state.nr_caps = 0;
2990 recon_state.pagelist = pagelist;
2991 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
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 if (recon_state.flock)
3021 reply->hdr.version = cpu_to_le16(2);
3022
3023 /* raced with cap release? */
3024 if (s_nr_caps != recon_state.nr_caps) {
3025 struct page *page = list_first_entry(&pagelist->head,
3026 struct page, lru);
3027 __le32 *addr = kmap_atomic(page);
3028 *addr = cpu_to_le32(recon_state.nr_caps);
3029 kunmap_atomic(addr);
3030 }
3031
3032 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3033 ceph_msg_data_add_pagelist(reply, pagelist);
3034
3035 ceph_early_kick_flushing_caps(mdsc, session);
3036
3037 ceph_con_send(&session->s_con, reply);
3038
3039 mutex_unlock(&session->s_mutex);
3040
3041 mutex_lock(&mdsc->mutex);
3042 __wake_requests(mdsc, &session->s_waiting);
3043 mutex_unlock(&mdsc->mutex);
3044
3045 up_read(&mdsc->snap_rwsem);
3046 return;
3047
3048fail:
3049 ceph_msg_put(reply);
3050 up_read(&mdsc->snap_rwsem);
3051 mutex_unlock(&session->s_mutex);
3052fail_nomsg:
3053 ceph_pagelist_release(pagelist);
3054fail_nopagelist:
3055 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3056 return;
3057}
3058
3059
3060/*
3061 * compare old and new mdsmaps, kicking requests
3062 * and closing out old connections as necessary
3063 *
3064 * called under mdsc->mutex.
3065 */
3066static void check_new_map(struct ceph_mds_client *mdsc,
3067 struct ceph_mdsmap *newmap,
3068 struct ceph_mdsmap *oldmap)
3069{
3070 int i;
3071 int oldstate, newstate;
3072 struct ceph_mds_session *s;
3073
3074 dout("check_new_map new %u old %u\n",
3075 newmap->m_epoch, oldmap->m_epoch);
3076
3077 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3078 if (mdsc->sessions[i] == NULL)
3079 continue;
3080 s = mdsc->sessions[i];
3081 oldstate = ceph_mdsmap_get_state(oldmap, i);
3082 newstate = ceph_mdsmap_get_state(newmap, i);
3083
3084 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3085 i, ceph_mds_state_name(oldstate),
3086 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3087 ceph_mds_state_name(newstate),
3088 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3089 ceph_session_state_name(s->s_state));
3090
3091 if (i >= newmap->m_max_mds ||
3092 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3093 ceph_mdsmap_get_addr(newmap, i),
3094 sizeof(struct ceph_entity_addr))) {
3095 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3096 /* the session never opened, just close it
3097 * out now */
3098 __wake_requests(mdsc, &s->s_waiting);
3099 __unregister_session(mdsc, s);
3100 } else {
3101 /* just close it */
3102 mutex_unlock(&mdsc->mutex);
3103 mutex_lock(&s->s_mutex);
3104 mutex_lock(&mdsc->mutex);
3105 ceph_con_close(&s->s_con);
3106 mutex_unlock(&s->s_mutex);
3107 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3108 }
3109 } else if (oldstate == newstate) {
3110 continue; /* nothing new with this mds */
3111 }
3112
3113 /*
3114 * send reconnect?
3115 */
3116 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3117 newstate >= CEPH_MDS_STATE_RECONNECT) {
3118 mutex_unlock(&mdsc->mutex);
3119 send_mds_reconnect(mdsc, s);
3120 mutex_lock(&mdsc->mutex);
3121 }
3122
3123 /*
3124 * kick request on any mds that has gone active.
3125 */
3126 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3127 newstate >= CEPH_MDS_STATE_ACTIVE) {
3128 if (oldstate != CEPH_MDS_STATE_CREATING &&
3129 oldstate != CEPH_MDS_STATE_STARTING)
3130 pr_info("mds%d recovery completed\n", s->s_mds);
3131 kick_requests(mdsc, i);
3132 ceph_kick_flushing_caps(mdsc, s);
3133 wake_up_session_caps(s, 1);
3134 }
3135 }
3136
3137 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3138 s = mdsc->sessions[i];
3139 if (!s)
3140 continue;
3141 if (!ceph_mdsmap_is_laggy(newmap, i))
3142 continue;
3143 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3144 s->s_state == CEPH_MDS_SESSION_HUNG ||
3145 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3146 dout(" connecting to export targets of laggy mds%d\n",
3147 i);
3148 __open_export_target_sessions(mdsc, s);
3149 }
3150 }
3151}
3152
3153
3154
3155/*
3156 * leases
3157 */
3158
3159/*
3160 * caller must hold session s_mutex, dentry->d_lock
3161 */
3162void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3163{
3164 struct ceph_dentry_info *di = ceph_dentry(dentry);
3165
3166 ceph_put_mds_session(di->lease_session);
3167 di->lease_session = NULL;
3168}
3169
3170static void handle_lease(struct ceph_mds_client *mdsc,
3171 struct ceph_mds_session *session,
3172 struct ceph_msg *msg)
3173{
3174 struct super_block *sb = mdsc->fsc->sb;
3175 struct inode *inode;
3176 struct dentry *parent, *dentry;
3177 struct ceph_dentry_info *di;
3178 int mds = session->s_mds;
3179 struct ceph_mds_lease *h = msg->front.iov_base;
3180 u32 seq;
3181 struct ceph_vino vino;
3182 struct qstr dname;
3183 int release = 0;
3184
3185 dout("handle_lease from mds%d\n", mds);
3186
3187 /* decode */
3188 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3189 goto bad;
3190 vino.ino = le64_to_cpu(h->ino);
3191 vino.snap = CEPH_NOSNAP;
3192 seq = le32_to_cpu(h->seq);
3193 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3194 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3195 if (dname.len != get_unaligned_le32(h+1))
3196 goto bad;
3197
3198 /* lookup inode */
3199 inode = ceph_find_inode(sb, vino);
3200 dout("handle_lease %s, ino %llx %p %.*s\n",
3201 ceph_lease_op_name(h->action), vino.ino, inode,
3202 dname.len, dname.name);
3203
3204 mutex_lock(&session->s_mutex);
3205 session->s_seq++;
3206
3207 if (inode == NULL) {
3208 dout("handle_lease no inode %llx\n", vino.ino);
3209 goto release;
3210 }
3211
3212 /* dentry */
3213 parent = d_find_alias(inode);
3214 if (!parent) {
3215 dout("no parent dentry on inode %p\n", inode);
3216 WARN_ON(1);
3217 goto release; /* hrm... */
3218 }
3219 dname.hash = full_name_hash(dname.name, dname.len);
3220 dentry = d_lookup(parent, &dname);
3221 dput(parent);
3222 if (!dentry)
3223 goto release;
3224
3225 spin_lock(&dentry->d_lock);
3226 di = ceph_dentry(dentry);
3227 switch (h->action) {
3228 case CEPH_MDS_LEASE_REVOKE:
3229 if (di->lease_session == session) {
3230 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3231 h->seq = cpu_to_le32(di->lease_seq);
3232 __ceph_mdsc_drop_dentry_lease(dentry);
3233 }
3234 release = 1;
3235 break;
3236
3237 case CEPH_MDS_LEASE_RENEW:
3238 if (di->lease_session == session &&
3239 di->lease_gen == session->s_cap_gen &&
3240 di->lease_renew_from &&
3241 di->lease_renew_after == 0) {
3242 unsigned long duration =
3243 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3244
3245 di->lease_seq = seq;
3246 dentry->d_time = di->lease_renew_from + duration;
3247 di->lease_renew_after = di->lease_renew_from +
3248 (duration >> 1);
3249 di->lease_renew_from = 0;
3250 }
3251 break;
3252 }
3253 spin_unlock(&dentry->d_lock);
3254 dput(dentry);
3255
3256 if (!release)
3257 goto out;
3258
3259release:
3260 /* let's just reuse the same message */
3261 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3262 ceph_msg_get(msg);
3263 ceph_con_send(&session->s_con, msg);
3264
3265out:
3266 iput(inode);
3267 mutex_unlock(&session->s_mutex);
3268 return;
3269
3270bad:
3271 pr_err("corrupt lease message\n");
3272 ceph_msg_dump(msg);
3273}
3274
3275void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3276 struct inode *inode,
3277 struct dentry *dentry, char action,
3278 u32 seq)
3279{
3280 struct ceph_msg *msg;
3281 struct ceph_mds_lease *lease;
3282 int len = sizeof(*lease) + sizeof(u32);
3283 int dnamelen = 0;
3284
3285 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3286 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3287 dnamelen = dentry->d_name.len;
3288 len += dnamelen;
3289
3290 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3291 if (!msg)
3292 return;
3293 lease = msg->front.iov_base;
3294 lease->action = action;
3295 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3296 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3297 lease->seq = cpu_to_le32(seq);
3298 put_unaligned_le32(dnamelen, lease + 1);
3299 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3300
3301 /*
3302 * if this is a preemptive lease RELEASE, no need to
3303 * flush request stream, since the actual request will
3304 * soon follow.
3305 */
3306 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3307
3308 ceph_con_send(&session->s_con, msg);
3309}
3310
3311/*
3312 * Preemptively release a lease we expect to invalidate anyway.
3313 * Pass @inode always, @dentry is optional.
3314 */
3315void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3316 struct dentry *dentry)
3317{
3318 struct ceph_dentry_info *di;
3319 struct ceph_mds_session *session;
3320 u32 seq;
3321
3322 BUG_ON(inode == NULL);
3323 BUG_ON(dentry == NULL);
3324
3325 /* is dentry lease valid? */
3326 spin_lock(&dentry->d_lock);
3327 di = ceph_dentry(dentry);
3328 if (!di || !di->lease_session ||
3329 di->lease_session->s_mds < 0 ||
3330 di->lease_gen != di->lease_session->s_cap_gen ||
3331 !time_before(jiffies, dentry->d_time)) {
3332 dout("lease_release inode %p dentry %p -- "
3333 "no lease\n",
3334 inode, dentry);
3335 spin_unlock(&dentry->d_lock);
3336 return;
3337 }
3338
3339 /* we do have a lease on this dentry; note mds and seq */
3340 session = ceph_get_mds_session(di->lease_session);
3341 seq = di->lease_seq;
3342 __ceph_mdsc_drop_dentry_lease(dentry);
3343 spin_unlock(&dentry->d_lock);
3344
3345 dout("lease_release inode %p dentry %p to mds%d\n",
3346 inode, dentry, session->s_mds);
3347 ceph_mdsc_lease_send_msg(session, inode, dentry,
3348 CEPH_MDS_LEASE_RELEASE, seq);
3349 ceph_put_mds_session(session);
3350}
3351
3352/*
3353 * drop all leases (and dentry refs) in preparation for umount
3354 */
3355static void drop_leases(struct ceph_mds_client *mdsc)
3356{
3357 int i;
3358
3359 dout("drop_leases\n");
3360 mutex_lock(&mdsc->mutex);
3361 for (i = 0; i < mdsc->max_sessions; i++) {
3362 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3363 if (!s)
3364 continue;
3365 mutex_unlock(&mdsc->mutex);
3366 mutex_lock(&s->s_mutex);
3367 mutex_unlock(&s->s_mutex);
3368 ceph_put_mds_session(s);
3369 mutex_lock(&mdsc->mutex);
3370 }
3371 mutex_unlock(&mdsc->mutex);
3372}
3373
3374
3375
3376/*
3377 * delayed work -- periodically trim expired leases, renew caps with mds
3378 */
3379static void schedule_delayed(struct ceph_mds_client *mdsc)
3380{
3381 int delay = 5;
3382 unsigned hz = round_jiffies_relative(HZ * delay);
3383 schedule_delayed_work(&mdsc->delayed_work, hz);
3384}
3385
3386static void delayed_work(struct work_struct *work)
3387{
3388 int i;
3389 struct ceph_mds_client *mdsc =
3390 container_of(work, struct ceph_mds_client, delayed_work.work);
3391 int renew_interval;
3392 int renew_caps;
3393
3394 dout("mdsc delayed_work\n");
3395 ceph_check_delayed_caps(mdsc);
3396
3397 mutex_lock(&mdsc->mutex);
3398 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3399 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3400 mdsc->last_renew_caps);
3401 if (renew_caps)
3402 mdsc->last_renew_caps = jiffies;
3403
3404 for (i = 0; i < mdsc->max_sessions; i++) {
3405 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3406 if (s == NULL)
3407 continue;
3408 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3409 dout("resending session close request for mds%d\n",
3410 s->s_mds);
3411 request_close_session(mdsc, s);
3412 ceph_put_mds_session(s);
3413 continue;
3414 }
3415 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3416 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3417 s->s_state = CEPH_MDS_SESSION_HUNG;
3418 pr_info("mds%d hung\n", s->s_mds);
3419 }
3420 }
3421 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3422 /* this mds is failed or recovering, just wait */
3423 ceph_put_mds_session(s);
3424 continue;
3425 }
3426 mutex_unlock(&mdsc->mutex);
3427
3428 mutex_lock(&s->s_mutex);
3429 if (renew_caps)
3430 send_renew_caps(mdsc, s);
3431 else
3432 ceph_con_keepalive(&s->s_con);
3433 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3434 s->s_state == CEPH_MDS_SESSION_HUNG)
3435 ceph_send_cap_releases(mdsc, s);
3436 mutex_unlock(&s->s_mutex);
3437 ceph_put_mds_session(s);
3438
3439 mutex_lock(&mdsc->mutex);
3440 }
3441 mutex_unlock(&mdsc->mutex);
3442
3443 schedule_delayed(mdsc);
3444}
3445
3446int ceph_mdsc_init(struct ceph_fs_client *fsc)
3447
3448{
3449 struct ceph_mds_client *mdsc;
3450
3451 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3452 if (!mdsc)
3453 return -ENOMEM;
3454 mdsc->fsc = fsc;
3455 fsc->mdsc = mdsc;
3456 mutex_init(&mdsc->mutex);
3457 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3458 if (mdsc->mdsmap == NULL) {
3459 kfree(mdsc);
3460 return -ENOMEM;
3461 }
3462
3463 init_completion(&mdsc->safe_umount_waiters);
3464 init_waitqueue_head(&mdsc->session_close_wq);
3465 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3466 mdsc->sessions = NULL;
3467 atomic_set(&mdsc->num_sessions, 0);
3468 mdsc->max_sessions = 0;
3469 mdsc->stopping = 0;
3470 mdsc->last_snap_seq = 0;
3471 init_rwsem(&mdsc->snap_rwsem);
3472 mdsc->snap_realms = RB_ROOT;
3473 INIT_LIST_HEAD(&mdsc->snap_empty);
3474 spin_lock_init(&mdsc->snap_empty_lock);
3475 mdsc->last_tid = 0;
3476 mdsc->oldest_tid = 0;
3477 mdsc->request_tree = RB_ROOT;
3478 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3479 mdsc->last_renew_caps = jiffies;
3480 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3481 spin_lock_init(&mdsc->cap_delay_lock);
3482 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3483 spin_lock_init(&mdsc->snap_flush_lock);
3484 mdsc->last_cap_flush_tid = 1;
3485 mdsc->cap_flush_tree = RB_ROOT;
3486 INIT_LIST_HEAD(&mdsc->cap_dirty);
3487 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3488 mdsc->num_cap_flushing = 0;
3489 spin_lock_init(&mdsc->cap_dirty_lock);
3490 init_waitqueue_head(&mdsc->cap_flushing_wq);
3491 spin_lock_init(&mdsc->dentry_lru_lock);
3492 INIT_LIST_HEAD(&mdsc->dentry_lru);
3493
3494 ceph_caps_init(mdsc);
3495 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3496
3497 init_rwsem(&mdsc->pool_perm_rwsem);
3498 mdsc->pool_perm_tree = RB_ROOT;
3499
3500 return 0;
3501}
3502
3503/*
3504 * Wait for safe replies on open mds requests. If we time out, drop
3505 * all requests from the tree to avoid dangling dentry refs.
3506 */
3507static void wait_requests(struct ceph_mds_client *mdsc)
3508{
3509 struct ceph_options *opts = mdsc->fsc->client->options;
3510 struct ceph_mds_request *req;
3511
3512 mutex_lock(&mdsc->mutex);
3513 if (__get_oldest_req(mdsc)) {
3514 mutex_unlock(&mdsc->mutex);
3515
3516 dout("wait_requests waiting for requests\n");
3517 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3518 ceph_timeout_jiffies(opts->mount_timeout));
3519
3520 /* tear down remaining requests */
3521 mutex_lock(&mdsc->mutex);
3522 while ((req = __get_oldest_req(mdsc))) {
3523 dout("wait_requests timed out on tid %llu\n",
3524 req->r_tid);
3525 __unregister_request(mdsc, req);
3526 }
3527 }
3528 mutex_unlock(&mdsc->mutex);
3529 dout("wait_requests done\n");
3530}
3531
3532/*
3533 * called before mount is ro, and before dentries are torn down.
3534 * (hmm, does this still race with new lookups?)
3535 */
3536void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3537{
3538 dout("pre_umount\n");
3539 mdsc->stopping = 1;
3540
3541 drop_leases(mdsc);
3542 ceph_flush_dirty_caps(mdsc);
3543 wait_requests(mdsc);
3544
3545 /*
3546 * wait for reply handlers to drop their request refs and
3547 * their inode/dcache refs
3548 */
3549 ceph_msgr_flush();
3550}
3551
3552/*
3553 * wait for all write mds requests to flush.
3554 */
3555static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3556{
3557 struct ceph_mds_request *req = NULL, *nextreq;
3558 struct rb_node *n;
3559
3560 mutex_lock(&mdsc->mutex);
3561 dout("wait_unsafe_requests want %lld\n", want_tid);
3562restart:
3563 req = __get_oldest_req(mdsc);
3564 while (req && req->r_tid <= want_tid) {
3565 /* find next request */
3566 n = rb_next(&req->r_node);
3567 if (n)
3568 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3569 else
3570 nextreq = NULL;
3571 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3572 (req->r_op & CEPH_MDS_OP_WRITE)) {
3573 /* write op */
3574 ceph_mdsc_get_request(req);
3575 if (nextreq)
3576 ceph_mdsc_get_request(nextreq);
3577 mutex_unlock(&mdsc->mutex);
3578 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3579 req->r_tid, want_tid);
3580 wait_for_completion(&req->r_safe_completion);
3581 mutex_lock(&mdsc->mutex);
3582 ceph_mdsc_put_request(req);
3583 if (!nextreq)
3584 break; /* next dne before, so we're done! */
3585 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3586 /* next request was removed from tree */
3587 ceph_mdsc_put_request(nextreq);
3588 goto restart;
3589 }
3590 ceph_mdsc_put_request(nextreq); /* won't go away */
3591 }
3592 req = nextreq;
3593 }
3594 mutex_unlock(&mdsc->mutex);
3595 dout("wait_unsafe_requests done\n");
3596}
3597
3598void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3599{
3600 u64 want_tid, want_flush, want_snap;
3601
3602 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3603 return;
3604
3605 dout("sync\n");
3606 mutex_lock(&mdsc->mutex);
3607 want_tid = mdsc->last_tid;
3608 mutex_unlock(&mdsc->mutex);
3609
3610 ceph_flush_dirty_caps(mdsc);
3611 spin_lock(&mdsc->cap_dirty_lock);
3612 want_flush = mdsc->last_cap_flush_tid;
3613 spin_unlock(&mdsc->cap_dirty_lock);
3614
3615 down_read(&mdsc->snap_rwsem);
3616 want_snap = mdsc->last_snap_seq;
3617 up_read(&mdsc->snap_rwsem);
3618
3619 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3620 want_tid, want_flush, want_snap);
3621
3622 wait_unsafe_requests(mdsc, want_tid);
3623 wait_caps_flush(mdsc, want_flush, want_snap);
3624}
3625
3626/*
3627 * true if all sessions are closed, or we force unmount
3628 */
3629static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3630{
3631 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3632 return true;
3633 return atomic_read(&mdsc->num_sessions) == 0;
3634}
3635
3636/*
3637 * called after sb is ro.
3638 */
3639void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3640{
3641 struct ceph_options *opts = mdsc->fsc->client->options;
3642 struct ceph_mds_session *session;
3643 int i;
3644
3645 dout("close_sessions\n");
3646
3647 /* close sessions */
3648 mutex_lock(&mdsc->mutex);
3649 for (i = 0; i < mdsc->max_sessions; i++) {
3650 session = __ceph_lookup_mds_session(mdsc, i);
3651 if (!session)
3652 continue;
3653 mutex_unlock(&mdsc->mutex);
3654 mutex_lock(&session->s_mutex);
3655 __close_session(mdsc, session);
3656 mutex_unlock(&session->s_mutex);
3657 ceph_put_mds_session(session);
3658 mutex_lock(&mdsc->mutex);
3659 }
3660 mutex_unlock(&mdsc->mutex);
3661
3662 dout("waiting for sessions to close\n");
3663 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3664 ceph_timeout_jiffies(opts->mount_timeout));
3665
3666 /* tear down remaining sessions */
3667 mutex_lock(&mdsc->mutex);
3668 for (i = 0; i < mdsc->max_sessions; i++) {
3669 if (mdsc->sessions[i]) {
3670 session = get_session(mdsc->sessions[i]);
3671 __unregister_session(mdsc, session);
3672 mutex_unlock(&mdsc->mutex);
3673 mutex_lock(&session->s_mutex);
3674 remove_session_caps(session);
3675 mutex_unlock(&session->s_mutex);
3676 ceph_put_mds_session(session);
3677 mutex_lock(&mdsc->mutex);
3678 }
3679 }
3680 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3681 mutex_unlock(&mdsc->mutex);
3682
3683 ceph_cleanup_empty_realms(mdsc);
3684
3685 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3686
3687 dout("stopped\n");
3688}
3689
3690void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3691{
3692 struct ceph_mds_session *session;
3693 int mds;
3694
3695 dout("force umount\n");
3696
3697 mutex_lock(&mdsc->mutex);
3698 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3699 session = __ceph_lookup_mds_session(mdsc, mds);
3700 if (!session)
3701 continue;
3702 mutex_unlock(&mdsc->mutex);
3703 mutex_lock(&session->s_mutex);
3704 __close_session(mdsc, session);
3705 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3706 cleanup_session_requests(mdsc, session);
3707 remove_session_caps(session);
3708 }
3709 mutex_unlock(&session->s_mutex);
3710 ceph_put_mds_session(session);
3711 mutex_lock(&mdsc->mutex);
3712 kick_requests(mdsc, mds);
3713 }
3714 __wake_requests(mdsc, &mdsc->waiting_for_map);
3715 mutex_unlock(&mdsc->mutex);
3716}
3717
3718static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3719{
3720 dout("stop\n");
3721 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3722 if (mdsc->mdsmap)
3723 ceph_mdsmap_destroy(mdsc->mdsmap);
3724 kfree(mdsc->sessions);
3725 ceph_caps_finalize(mdsc);
3726 ceph_pool_perm_destroy(mdsc);
3727}
3728
3729void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3730{
3731 struct ceph_mds_client *mdsc = fsc->mdsc;
3732
3733 dout("mdsc_destroy %p\n", mdsc);
3734 ceph_mdsc_stop(mdsc);
3735
3736 /* flush out any connection work with references to us */
3737 ceph_msgr_flush();
3738
3739 fsc->mdsc = NULL;
3740 kfree(mdsc);
3741 dout("mdsc_destroy %p done\n", mdsc);
3742}
3743
3744
3745/*
3746 * handle mds map update.
3747 */
3748void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3749{
3750 u32 epoch;
3751 u32 maplen;
3752 void *p = msg->front.iov_base;
3753 void *end = p + msg->front.iov_len;
3754 struct ceph_mdsmap *newmap, *oldmap;
3755 struct ceph_fsid fsid;
3756 int err = -EINVAL;
3757
3758 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3759 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3760 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3761 return;
3762 epoch = ceph_decode_32(&p);
3763 maplen = ceph_decode_32(&p);
3764 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3765
3766 /* do we need it? */
3767 mutex_lock(&mdsc->mutex);
3768 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3769 dout("handle_map epoch %u <= our %u\n",
3770 epoch, mdsc->mdsmap->m_epoch);
3771 mutex_unlock(&mdsc->mutex);
3772 return;
3773 }
3774
3775 newmap = ceph_mdsmap_decode(&p, end);
3776 if (IS_ERR(newmap)) {
3777 err = PTR_ERR(newmap);
3778 goto bad_unlock;
3779 }
3780
3781 /* swap into place */
3782 if (mdsc->mdsmap) {
3783 oldmap = mdsc->mdsmap;
3784 mdsc->mdsmap = newmap;
3785 check_new_map(mdsc, newmap, oldmap);
3786 ceph_mdsmap_destroy(oldmap);
3787 } else {
3788 mdsc->mdsmap = newmap; /* first mds map */
3789 }
3790 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3791
3792 __wake_requests(mdsc, &mdsc->waiting_for_map);
3793 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3794 mdsc->mdsmap->m_epoch);
3795
3796 mutex_unlock(&mdsc->mutex);
3797 schedule_delayed(mdsc);
3798 return;
3799
3800bad_unlock:
3801 mutex_unlock(&mdsc->mutex);
3802bad:
3803 pr_err("error decoding mdsmap %d\n", err);
3804 return;
3805}
3806
3807static struct ceph_connection *con_get(struct ceph_connection *con)
3808{
3809 struct ceph_mds_session *s = con->private;
3810
3811 if (get_session(s)) {
3812 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3813 return con;
3814 }
3815 dout("mdsc con_get %p FAIL\n", s);
3816 return NULL;
3817}
3818
3819static void con_put(struct ceph_connection *con)
3820{
3821 struct ceph_mds_session *s = con->private;
3822
3823 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3824 ceph_put_mds_session(s);
3825}
3826
3827/*
3828 * if the client is unresponsive for long enough, the mds will kill
3829 * the session entirely.
3830 */
3831static void peer_reset(struct ceph_connection *con)
3832{
3833 struct ceph_mds_session *s = con->private;
3834 struct ceph_mds_client *mdsc = s->s_mdsc;
3835
3836 pr_warn("mds%d closed our session\n", s->s_mds);
3837 send_mds_reconnect(mdsc, s);
3838}
3839
3840static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3841{
3842 struct ceph_mds_session *s = con->private;
3843 struct ceph_mds_client *mdsc = s->s_mdsc;
3844 int type = le16_to_cpu(msg->hdr.type);
3845
3846 mutex_lock(&mdsc->mutex);
3847 if (__verify_registered_session(mdsc, s) < 0) {
3848 mutex_unlock(&mdsc->mutex);
3849 goto out;
3850 }
3851 mutex_unlock(&mdsc->mutex);
3852
3853 switch (type) {
3854 case CEPH_MSG_MDS_MAP:
3855 ceph_mdsc_handle_map(mdsc, msg);
3856 break;
3857 case CEPH_MSG_CLIENT_SESSION:
3858 handle_session(s, msg);
3859 break;
3860 case CEPH_MSG_CLIENT_REPLY:
3861 handle_reply(s, msg);
3862 break;
3863 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3864 handle_forward(mdsc, s, msg);
3865 break;
3866 case CEPH_MSG_CLIENT_CAPS:
3867 ceph_handle_caps(s, msg);
3868 break;
3869 case CEPH_MSG_CLIENT_SNAP:
3870 ceph_handle_snap(mdsc, s, msg);
3871 break;
3872 case CEPH_MSG_CLIENT_LEASE:
3873 handle_lease(mdsc, s, msg);
3874 break;
3875
3876 default:
3877 pr_err("received unknown message type %d %s\n", type,
3878 ceph_msg_type_name(type));
3879 }
3880out:
3881 ceph_msg_put(msg);
3882}
3883
3884/*
3885 * authentication
3886 */
3887
3888/*
3889 * Note: returned pointer is the address of a structure that's
3890 * managed separately. Caller must *not* attempt to free it.
3891 */
3892static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3893 int *proto, int force_new)
3894{
3895 struct ceph_mds_session *s = con->private;
3896 struct ceph_mds_client *mdsc = s->s_mdsc;
3897 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3898 struct ceph_auth_handshake *auth = &s->s_auth;
3899
3900 if (force_new && auth->authorizer) {
3901 ceph_auth_destroy_authorizer(auth->authorizer);
3902 auth->authorizer = NULL;
3903 }
3904 if (!auth->authorizer) {
3905 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3906 auth);
3907 if (ret)
3908 return ERR_PTR(ret);
3909 } else {
3910 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3911 auth);
3912 if (ret)
3913 return ERR_PTR(ret);
3914 }
3915 *proto = ac->protocol;
3916
3917 return auth;
3918}
3919
3920
3921static int verify_authorizer_reply(struct ceph_connection *con, int len)
3922{
3923 struct ceph_mds_session *s = con->private;
3924 struct ceph_mds_client *mdsc = s->s_mdsc;
3925 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3926
3927 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3928}
3929
3930static int invalidate_authorizer(struct ceph_connection *con)
3931{
3932 struct ceph_mds_session *s = con->private;
3933 struct ceph_mds_client *mdsc = s->s_mdsc;
3934 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3935
3936 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3937
3938 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3939}
3940
3941static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3942 struct ceph_msg_header *hdr, int *skip)
3943{
3944 struct ceph_msg *msg;
3945 int type = (int) le16_to_cpu(hdr->type);
3946 int front_len = (int) le32_to_cpu(hdr->front_len);
3947
3948 if (con->in_msg)
3949 return con->in_msg;
3950
3951 *skip = 0;
3952 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3953 if (!msg) {
3954 pr_err("unable to allocate msg type %d len %d\n",
3955 type, front_len);
3956 return NULL;
3957 }
3958
3959 return msg;
3960}
3961
3962static int mds_sign_message(struct ceph_msg *msg)
3963{
3964 struct ceph_mds_session *s = msg->con->private;
3965 struct ceph_auth_handshake *auth = &s->s_auth;
3966
3967 return ceph_auth_sign_message(auth, msg);
3968}
3969
3970static int mds_check_message_signature(struct ceph_msg *msg)
3971{
3972 struct ceph_mds_session *s = msg->con->private;
3973 struct ceph_auth_handshake *auth = &s->s_auth;
3974
3975 return ceph_auth_check_message_signature(auth, msg);
3976}
3977
3978static const struct ceph_connection_operations mds_con_ops = {
3979 .get = con_get,
3980 .put = con_put,
3981 .dispatch = dispatch,
3982 .get_authorizer = get_authorizer,
3983 .verify_authorizer_reply = verify_authorizer_reply,
3984 .invalidate_authorizer = invalidate_authorizer,
3985 .peer_reset = peer_reset,
3986 .alloc_msg = mds_alloc_msg,
3987 .sign_message = mds_sign_message,
3988 .check_message_signature = mds_check_message_signature,
3989};
3990
3991/* eof */
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#include <linux/bits.h>
13#include <linux/ktime.h>
14
15#include "super.h"
16#include "mds_client.h"
17
18#include <linux/ceph/ceph_features.h>
19#include <linux/ceph/messenger.h>
20#include <linux/ceph/decode.h>
21#include <linux/ceph/pagelist.h>
22#include <linux/ceph/auth.h>
23#include <linux/ceph/debugfs.h>
24
25#define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
26
27/*
28 * A cluster of MDS (metadata server) daemons is responsible for
29 * managing the file system namespace (the directory hierarchy and
30 * inodes) and for coordinating shared access to storage. Metadata is
31 * partitioning hierarchically across a number of servers, and that
32 * partition varies over time as the cluster adjusts the distribution
33 * in order to balance load.
34 *
35 * The MDS client is primarily responsible to managing synchronous
36 * metadata requests for operations like open, unlink, and so forth.
37 * If there is a MDS failure, we find out about it when we (possibly
38 * request and) receive a new MDS map, and can resubmit affected
39 * requests.
40 *
41 * For the most part, though, we take advantage of a lossless
42 * communications channel to the MDS, and do not need to worry about
43 * timing out or resubmitting requests.
44 *
45 * We maintain a stateful "session" with each MDS we interact with.
46 * Within each session, we sent periodic heartbeat messages to ensure
47 * any capabilities or leases we have been issues remain valid. If
48 * the session times out and goes stale, our leases and capabilities
49 * are no longer valid.
50 */
51
52struct ceph_reconnect_state {
53 struct ceph_mds_session *session;
54 int nr_caps, nr_realms;
55 struct ceph_pagelist *pagelist;
56 unsigned msg_version;
57 bool allow_multi;
58};
59
60static void __wake_requests(struct ceph_mds_client *mdsc,
61 struct list_head *head);
62static void ceph_cap_release_work(struct work_struct *work);
63static void ceph_cap_reclaim_work(struct work_struct *work);
64
65static const struct ceph_connection_operations mds_con_ops;
66
67
68/*
69 * mds reply parsing
70 */
71
72static int parse_reply_info_quota(void **p, void *end,
73 struct ceph_mds_reply_info_in *info)
74{
75 u8 struct_v, struct_compat;
76 u32 struct_len;
77
78 ceph_decode_8_safe(p, end, struct_v, bad);
79 ceph_decode_8_safe(p, end, struct_compat, bad);
80 /* struct_v is expected to be >= 1. we only
81 * understand encoding with struct_compat == 1. */
82 if (!struct_v || struct_compat != 1)
83 goto bad;
84 ceph_decode_32_safe(p, end, struct_len, bad);
85 ceph_decode_need(p, end, struct_len, bad);
86 end = *p + struct_len;
87 ceph_decode_64_safe(p, end, info->max_bytes, bad);
88 ceph_decode_64_safe(p, end, info->max_files, bad);
89 *p = end;
90 return 0;
91bad:
92 return -EIO;
93}
94
95/*
96 * parse individual inode info
97 */
98static int parse_reply_info_in(void **p, void *end,
99 struct ceph_mds_reply_info_in *info,
100 u64 features)
101{
102 int err = 0;
103 u8 struct_v = 0;
104
105 if (features == (u64)-1) {
106 u32 struct_len;
107 u8 struct_compat;
108 ceph_decode_8_safe(p, end, struct_v, bad);
109 ceph_decode_8_safe(p, end, struct_compat, bad);
110 /* struct_v is expected to be >= 1. we only understand
111 * encoding with struct_compat == 1. */
112 if (!struct_v || struct_compat != 1)
113 goto bad;
114 ceph_decode_32_safe(p, end, struct_len, bad);
115 ceph_decode_need(p, end, struct_len, bad);
116 end = *p + struct_len;
117 }
118
119 ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
120 info->in = *p;
121 *p += sizeof(struct ceph_mds_reply_inode) +
122 sizeof(*info->in->fragtree.splits) *
123 le32_to_cpu(info->in->fragtree.nsplits);
124
125 ceph_decode_32_safe(p, end, info->symlink_len, bad);
126 ceph_decode_need(p, end, info->symlink_len, bad);
127 info->symlink = *p;
128 *p += info->symlink_len;
129
130 ceph_decode_copy_safe(p, end, &info->dir_layout,
131 sizeof(info->dir_layout), bad);
132 ceph_decode_32_safe(p, end, info->xattr_len, bad);
133 ceph_decode_need(p, end, info->xattr_len, bad);
134 info->xattr_data = *p;
135 *p += info->xattr_len;
136
137 if (features == (u64)-1) {
138 /* inline data */
139 ceph_decode_64_safe(p, end, info->inline_version, bad);
140 ceph_decode_32_safe(p, end, info->inline_len, bad);
141 ceph_decode_need(p, end, info->inline_len, bad);
142 info->inline_data = *p;
143 *p += info->inline_len;
144 /* quota */
145 err = parse_reply_info_quota(p, end, info);
146 if (err < 0)
147 goto out_bad;
148 /* pool namespace */
149 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
150 if (info->pool_ns_len > 0) {
151 ceph_decode_need(p, end, info->pool_ns_len, bad);
152 info->pool_ns_data = *p;
153 *p += info->pool_ns_len;
154 }
155
156 /* btime */
157 ceph_decode_need(p, end, sizeof(info->btime), bad);
158 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
159
160 /* change attribute */
161 ceph_decode_64_safe(p, end, info->change_attr, bad);
162
163 /* dir pin */
164 if (struct_v >= 2) {
165 ceph_decode_32_safe(p, end, info->dir_pin, bad);
166 } else {
167 info->dir_pin = -ENODATA;
168 }
169
170 /* snapshot birth time, remains zero for v<=2 */
171 if (struct_v >= 3) {
172 ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
173 ceph_decode_copy(p, &info->snap_btime,
174 sizeof(info->snap_btime));
175 } else {
176 memset(&info->snap_btime, 0, sizeof(info->snap_btime));
177 }
178
179 /* snapshot count, remains zero for v<=3 */
180 if (struct_v >= 4) {
181 ceph_decode_64_safe(p, end, info->rsnaps, bad);
182 } else {
183 info->rsnaps = 0;
184 }
185
186 *p = end;
187 } else {
188 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
189 ceph_decode_64_safe(p, end, info->inline_version, bad);
190 ceph_decode_32_safe(p, end, info->inline_len, bad);
191 ceph_decode_need(p, end, info->inline_len, bad);
192 info->inline_data = *p;
193 *p += info->inline_len;
194 } else
195 info->inline_version = CEPH_INLINE_NONE;
196
197 if (features & CEPH_FEATURE_MDS_QUOTA) {
198 err = parse_reply_info_quota(p, end, info);
199 if (err < 0)
200 goto out_bad;
201 } else {
202 info->max_bytes = 0;
203 info->max_files = 0;
204 }
205
206 info->pool_ns_len = 0;
207 info->pool_ns_data = NULL;
208 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
209 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
210 if (info->pool_ns_len > 0) {
211 ceph_decode_need(p, end, info->pool_ns_len, bad);
212 info->pool_ns_data = *p;
213 *p += info->pool_ns_len;
214 }
215 }
216
217 if (features & CEPH_FEATURE_FS_BTIME) {
218 ceph_decode_need(p, end, sizeof(info->btime), bad);
219 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
220 ceph_decode_64_safe(p, end, info->change_attr, bad);
221 }
222
223 info->dir_pin = -ENODATA;
224 /* info->snap_btime and info->rsnaps remain zero */
225 }
226 return 0;
227bad:
228 err = -EIO;
229out_bad:
230 return err;
231}
232
233static int parse_reply_info_dir(void **p, void *end,
234 struct ceph_mds_reply_dirfrag **dirfrag,
235 u64 features)
236{
237 if (features == (u64)-1) {
238 u8 struct_v, struct_compat;
239 u32 struct_len;
240 ceph_decode_8_safe(p, end, struct_v, bad);
241 ceph_decode_8_safe(p, end, struct_compat, bad);
242 /* struct_v is expected to be >= 1. we only understand
243 * encoding whose struct_compat == 1. */
244 if (!struct_v || struct_compat != 1)
245 goto bad;
246 ceph_decode_32_safe(p, end, struct_len, bad);
247 ceph_decode_need(p, end, struct_len, bad);
248 end = *p + struct_len;
249 }
250
251 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
252 *dirfrag = *p;
253 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
254 if (unlikely(*p > end))
255 goto bad;
256 if (features == (u64)-1)
257 *p = end;
258 return 0;
259bad:
260 return -EIO;
261}
262
263static int parse_reply_info_lease(void **p, void *end,
264 struct ceph_mds_reply_lease **lease,
265 u64 features)
266{
267 if (features == (u64)-1) {
268 u8 struct_v, struct_compat;
269 u32 struct_len;
270 ceph_decode_8_safe(p, end, struct_v, bad);
271 ceph_decode_8_safe(p, end, struct_compat, bad);
272 /* struct_v is expected to be >= 1. we only understand
273 * encoding whose struct_compat == 1. */
274 if (!struct_v || struct_compat != 1)
275 goto bad;
276 ceph_decode_32_safe(p, end, struct_len, bad);
277 ceph_decode_need(p, end, struct_len, bad);
278 end = *p + struct_len;
279 }
280
281 ceph_decode_need(p, end, sizeof(**lease), bad);
282 *lease = *p;
283 *p += sizeof(**lease);
284 if (features == (u64)-1)
285 *p = end;
286 return 0;
287bad:
288 return -EIO;
289}
290
291/*
292 * parse a normal reply, which may contain a (dir+)dentry and/or a
293 * target inode.
294 */
295static int parse_reply_info_trace(void **p, void *end,
296 struct ceph_mds_reply_info_parsed *info,
297 u64 features)
298{
299 int err;
300
301 if (info->head->is_dentry) {
302 err = parse_reply_info_in(p, end, &info->diri, features);
303 if (err < 0)
304 goto out_bad;
305
306 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
307 if (err < 0)
308 goto out_bad;
309
310 ceph_decode_32_safe(p, end, info->dname_len, bad);
311 ceph_decode_need(p, end, info->dname_len, bad);
312 info->dname = *p;
313 *p += info->dname_len;
314
315 err = parse_reply_info_lease(p, end, &info->dlease, features);
316 if (err < 0)
317 goto out_bad;
318 }
319
320 if (info->head->is_target) {
321 err = parse_reply_info_in(p, end, &info->targeti, features);
322 if (err < 0)
323 goto out_bad;
324 }
325
326 if (unlikely(*p != end))
327 goto bad;
328 return 0;
329
330bad:
331 err = -EIO;
332out_bad:
333 pr_err("problem parsing mds trace %d\n", err);
334 return err;
335}
336
337/*
338 * parse readdir results
339 */
340static int parse_reply_info_readdir(void **p, void *end,
341 struct ceph_mds_reply_info_parsed *info,
342 u64 features)
343{
344 u32 num, i = 0;
345 int err;
346
347 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
348 if (err < 0)
349 goto out_bad;
350
351 ceph_decode_need(p, end, sizeof(num) + 2, bad);
352 num = ceph_decode_32(p);
353 {
354 u16 flags = ceph_decode_16(p);
355 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
356 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
357 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
358 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
359 }
360 if (num == 0)
361 goto done;
362
363 BUG_ON(!info->dir_entries);
364 if ((unsigned long)(info->dir_entries + num) >
365 (unsigned long)info->dir_entries + info->dir_buf_size) {
366 pr_err("dir contents are larger than expected\n");
367 WARN_ON(1);
368 goto bad;
369 }
370
371 info->dir_nr = num;
372 while (num) {
373 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
374 /* dentry */
375 ceph_decode_32_safe(p, end, rde->name_len, bad);
376 ceph_decode_need(p, end, rde->name_len, bad);
377 rde->name = *p;
378 *p += rde->name_len;
379 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
380
381 /* dentry lease */
382 err = parse_reply_info_lease(p, end, &rde->lease, features);
383 if (err)
384 goto out_bad;
385 /* inode */
386 err = parse_reply_info_in(p, end, &rde->inode, features);
387 if (err < 0)
388 goto out_bad;
389 /* ceph_readdir_prepopulate() will update it */
390 rde->offset = 0;
391 i++;
392 num--;
393 }
394
395done:
396 /* Skip over any unrecognized fields */
397 *p = end;
398 return 0;
399
400bad:
401 err = -EIO;
402out_bad:
403 pr_err("problem parsing dir contents %d\n", err);
404 return err;
405}
406
407/*
408 * parse fcntl F_GETLK results
409 */
410static int parse_reply_info_filelock(void **p, void *end,
411 struct ceph_mds_reply_info_parsed *info,
412 u64 features)
413{
414 if (*p + sizeof(*info->filelock_reply) > end)
415 goto bad;
416
417 info->filelock_reply = *p;
418
419 /* Skip over any unrecognized fields */
420 *p = end;
421 return 0;
422bad:
423 return -EIO;
424}
425
426
427#if BITS_PER_LONG == 64
428
429#define DELEGATED_INO_AVAILABLE xa_mk_value(1)
430
431static int ceph_parse_deleg_inos(void **p, void *end,
432 struct ceph_mds_session *s)
433{
434 u32 sets;
435
436 ceph_decode_32_safe(p, end, sets, bad);
437 dout("got %u sets of delegated inodes\n", sets);
438 while (sets--) {
439 u64 start, len, ino;
440
441 ceph_decode_64_safe(p, end, start, bad);
442 ceph_decode_64_safe(p, end, len, bad);
443
444 /* Don't accept a delegation of system inodes */
445 if (start < CEPH_INO_SYSTEM_BASE) {
446 pr_warn_ratelimited("ceph: ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
447 start, len);
448 continue;
449 }
450 while (len--) {
451 int err = xa_insert(&s->s_delegated_inos, ino = start++,
452 DELEGATED_INO_AVAILABLE,
453 GFP_KERNEL);
454 if (!err) {
455 dout("added delegated inode 0x%llx\n",
456 start - 1);
457 } else if (err == -EBUSY) {
458 pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
459 start - 1);
460 } else {
461 return err;
462 }
463 }
464 }
465 return 0;
466bad:
467 return -EIO;
468}
469
470u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
471{
472 unsigned long ino;
473 void *val;
474
475 xa_for_each(&s->s_delegated_inos, ino, val) {
476 val = xa_erase(&s->s_delegated_inos, ino);
477 if (val == DELEGATED_INO_AVAILABLE)
478 return ino;
479 }
480 return 0;
481}
482
483int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
484{
485 return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
486 GFP_KERNEL);
487}
488#else /* BITS_PER_LONG == 64 */
489/*
490 * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
491 * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
492 * and bottom words?
493 */
494static int ceph_parse_deleg_inos(void **p, void *end,
495 struct ceph_mds_session *s)
496{
497 u32 sets;
498
499 ceph_decode_32_safe(p, end, sets, bad);
500 if (sets)
501 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
502 return 0;
503bad:
504 return -EIO;
505}
506
507u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
508{
509 return 0;
510}
511
512int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
513{
514 return 0;
515}
516#endif /* BITS_PER_LONG == 64 */
517
518/*
519 * parse create results
520 */
521static int parse_reply_info_create(void **p, void *end,
522 struct ceph_mds_reply_info_parsed *info,
523 u64 features, struct ceph_mds_session *s)
524{
525 int ret;
526
527 if (features == (u64)-1 ||
528 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
529 if (*p == end) {
530 /* Malformed reply? */
531 info->has_create_ino = false;
532 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
533 info->has_create_ino = true;
534 /* struct_v, struct_compat, and len */
535 ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
536 ceph_decode_64_safe(p, end, info->ino, bad);
537 ret = ceph_parse_deleg_inos(p, end, s);
538 if (ret)
539 return ret;
540 } else {
541 /* legacy */
542 ceph_decode_64_safe(p, end, info->ino, bad);
543 info->has_create_ino = true;
544 }
545 } else {
546 if (*p != end)
547 goto bad;
548 }
549
550 /* Skip over any unrecognized fields */
551 *p = end;
552 return 0;
553bad:
554 return -EIO;
555}
556
557/*
558 * parse extra results
559 */
560static int parse_reply_info_extra(void **p, void *end,
561 struct ceph_mds_reply_info_parsed *info,
562 u64 features, struct ceph_mds_session *s)
563{
564 u32 op = le32_to_cpu(info->head->op);
565
566 if (op == CEPH_MDS_OP_GETFILELOCK)
567 return parse_reply_info_filelock(p, end, info, features);
568 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
569 return parse_reply_info_readdir(p, end, info, features);
570 else if (op == CEPH_MDS_OP_CREATE)
571 return parse_reply_info_create(p, end, info, features, s);
572 else
573 return -EIO;
574}
575
576/*
577 * parse entire mds reply
578 */
579static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
580 struct ceph_mds_reply_info_parsed *info,
581 u64 features)
582{
583 void *p, *end;
584 u32 len;
585 int err;
586
587 info->head = msg->front.iov_base;
588 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
589 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
590
591 /* trace */
592 ceph_decode_32_safe(&p, end, len, bad);
593 if (len > 0) {
594 ceph_decode_need(&p, end, len, bad);
595 err = parse_reply_info_trace(&p, p+len, info, features);
596 if (err < 0)
597 goto out_bad;
598 }
599
600 /* extra */
601 ceph_decode_32_safe(&p, end, len, bad);
602 if (len > 0) {
603 ceph_decode_need(&p, end, len, bad);
604 err = parse_reply_info_extra(&p, p+len, info, features, s);
605 if (err < 0)
606 goto out_bad;
607 }
608
609 /* snap blob */
610 ceph_decode_32_safe(&p, end, len, bad);
611 info->snapblob_len = len;
612 info->snapblob = p;
613 p += len;
614
615 if (p != end)
616 goto bad;
617 return 0;
618
619bad:
620 err = -EIO;
621out_bad:
622 pr_err("mds parse_reply err %d\n", err);
623 return err;
624}
625
626static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
627{
628 if (!info->dir_entries)
629 return;
630 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
631}
632
633
634/*
635 * sessions
636 */
637const char *ceph_session_state_name(int s)
638{
639 switch (s) {
640 case CEPH_MDS_SESSION_NEW: return "new";
641 case CEPH_MDS_SESSION_OPENING: return "opening";
642 case CEPH_MDS_SESSION_OPEN: return "open";
643 case CEPH_MDS_SESSION_HUNG: return "hung";
644 case CEPH_MDS_SESSION_CLOSING: return "closing";
645 case CEPH_MDS_SESSION_CLOSED: return "closed";
646 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
647 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
648 case CEPH_MDS_SESSION_REJECTED: return "rejected";
649 default: return "???";
650 }
651}
652
653struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
654{
655 if (refcount_inc_not_zero(&s->s_ref)) {
656 dout("mdsc get_session %p %d -> %d\n", s,
657 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
658 return s;
659 } else {
660 dout("mdsc get_session %p 0 -- FAIL\n", s);
661 return NULL;
662 }
663}
664
665void ceph_put_mds_session(struct ceph_mds_session *s)
666{
667 if (IS_ERR_OR_NULL(s))
668 return;
669
670 dout("mdsc put_session %p %d -> %d\n", s,
671 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
672 if (refcount_dec_and_test(&s->s_ref)) {
673 if (s->s_auth.authorizer)
674 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
675 WARN_ON(mutex_is_locked(&s->s_mutex));
676 xa_destroy(&s->s_delegated_inos);
677 kfree(s);
678 }
679}
680
681/*
682 * called under mdsc->mutex
683 */
684struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
685 int mds)
686{
687 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
688 return NULL;
689 return ceph_get_mds_session(mdsc->sessions[mds]);
690}
691
692static bool __have_session(struct ceph_mds_client *mdsc, int mds)
693{
694 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
695 return false;
696 else
697 return true;
698}
699
700static int __verify_registered_session(struct ceph_mds_client *mdsc,
701 struct ceph_mds_session *s)
702{
703 if (s->s_mds >= mdsc->max_sessions ||
704 mdsc->sessions[s->s_mds] != s)
705 return -ENOENT;
706 return 0;
707}
708
709/*
710 * create+register a new session for given mds.
711 * called under mdsc->mutex.
712 */
713static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
714 int mds)
715{
716 struct ceph_mds_session *s;
717
718 if (mds >= mdsc->mdsmap->possible_max_rank)
719 return ERR_PTR(-EINVAL);
720
721 s = kzalloc(sizeof(*s), GFP_NOFS);
722 if (!s)
723 return ERR_PTR(-ENOMEM);
724
725 if (mds >= mdsc->max_sessions) {
726 int newmax = 1 << get_count_order(mds + 1);
727 struct ceph_mds_session **sa;
728
729 dout("%s: realloc to %d\n", __func__, newmax);
730 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
731 if (!sa)
732 goto fail_realloc;
733 if (mdsc->sessions) {
734 memcpy(sa, mdsc->sessions,
735 mdsc->max_sessions * sizeof(void *));
736 kfree(mdsc->sessions);
737 }
738 mdsc->sessions = sa;
739 mdsc->max_sessions = newmax;
740 }
741
742 dout("%s: mds%d\n", __func__, mds);
743 s->s_mdsc = mdsc;
744 s->s_mds = mds;
745 s->s_state = CEPH_MDS_SESSION_NEW;
746 s->s_ttl = 0;
747 s->s_seq = 0;
748 mutex_init(&s->s_mutex);
749
750 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
751
752 atomic_set(&s->s_cap_gen, 1);
753 s->s_cap_ttl = jiffies - 1;
754
755 spin_lock_init(&s->s_cap_lock);
756 s->s_renew_requested = 0;
757 s->s_renew_seq = 0;
758 INIT_LIST_HEAD(&s->s_caps);
759 s->s_nr_caps = 0;
760 refcount_set(&s->s_ref, 1);
761 INIT_LIST_HEAD(&s->s_waiting);
762 INIT_LIST_HEAD(&s->s_unsafe);
763 xa_init(&s->s_delegated_inos);
764 s->s_num_cap_releases = 0;
765 s->s_cap_reconnect = 0;
766 s->s_cap_iterator = NULL;
767 INIT_LIST_HEAD(&s->s_cap_releases);
768 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
769
770 INIT_LIST_HEAD(&s->s_cap_dirty);
771 INIT_LIST_HEAD(&s->s_cap_flushing);
772
773 mdsc->sessions[mds] = s;
774 atomic_inc(&mdsc->num_sessions);
775 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
776
777 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
778 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
779
780 return s;
781
782fail_realloc:
783 kfree(s);
784 return ERR_PTR(-ENOMEM);
785}
786
787/*
788 * called under mdsc->mutex
789 */
790static void __unregister_session(struct ceph_mds_client *mdsc,
791 struct ceph_mds_session *s)
792{
793 dout("__unregister_session mds%d %p\n", s->s_mds, s);
794 BUG_ON(mdsc->sessions[s->s_mds] != s);
795 mdsc->sessions[s->s_mds] = NULL;
796 ceph_con_close(&s->s_con);
797 ceph_put_mds_session(s);
798 atomic_dec(&mdsc->num_sessions);
799}
800
801/*
802 * drop session refs in request.
803 *
804 * should be last request ref, or hold mdsc->mutex
805 */
806static void put_request_session(struct ceph_mds_request *req)
807{
808 if (req->r_session) {
809 ceph_put_mds_session(req->r_session);
810 req->r_session = NULL;
811 }
812}
813
814void ceph_mdsc_release_request(struct kref *kref)
815{
816 struct ceph_mds_request *req = container_of(kref,
817 struct ceph_mds_request,
818 r_kref);
819 ceph_mdsc_release_dir_caps_no_check(req);
820 destroy_reply_info(&req->r_reply_info);
821 if (req->r_request)
822 ceph_msg_put(req->r_request);
823 if (req->r_reply)
824 ceph_msg_put(req->r_reply);
825 if (req->r_inode) {
826 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
827 iput(req->r_inode);
828 }
829 if (req->r_parent) {
830 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
831 iput(req->r_parent);
832 }
833 iput(req->r_target_inode);
834 if (req->r_dentry)
835 dput(req->r_dentry);
836 if (req->r_old_dentry)
837 dput(req->r_old_dentry);
838 if (req->r_old_dentry_dir) {
839 /*
840 * track (and drop pins for) r_old_dentry_dir
841 * separately, since r_old_dentry's d_parent may have
842 * changed between the dir mutex being dropped and
843 * this request being freed.
844 */
845 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
846 CEPH_CAP_PIN);
847 iput(req->r_old_dentry_dir);
848 }
849 kfree(req->r_path1);
850 kfree(req->r_path2);
851 put_cred(req->r_cred);
852 if (req->r_pagelist)
853 ceph_pagelist_release(req->r_pagelist);
854 put_request_session(req);
855 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
856 WARN_ON_ONCE(!list_empty(&req->r_wait));
857 kmem_cache_free(ceph_mds_request_cachep, req);
858}
859
860DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
861
862/*
863 * lookup session, bump ref if found.
864 *
865 * called under mdsc->mutex.
866 */
867static struct ceph_mds_request *
868lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
869{
870 struct ceph_mds_request *req;
871
872 req = lookup_request(&mdsc->request_tree, tid);
873 if (req)
874 ceph_mdsc_get_request(req);
875
876 return req;
877}
878
879/*
880 * Register an in-flight request, and assign a tid. Link to directory
881 * are modifying (if any).
882 *
883 * Called under mdsc->mutex.
884 */
885static void __register_request(struct ceph_mds_client *mdsc,
886 struct ceph_mds_request *req,
887 struct inode *dir)
888{
889 int ret = 0;
890
891 req->r_tid = ++mdsc->last_tid;
892 if (req->r_num_caps) {
893 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
894 req->r_num_caps);
895 if (ret < 0) {
896 pr_err("__register_request %p "
897 "failed to reserve caps: %d\n", req, ret);
898 /* set req->r_err to fail early from __do_request */
899 req->r_err = ret;
900 return;
901 }
902 }
903 dout("__register_request %p tid %lld\n", req, req->r_tid);
904 ceph_mdsc_get_request(req);
905 insert_request(&mdsc->request_tree, req);
906
907 req->r_cred = get_current_cred();
908
909 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
910 mdsc->oldest_tid = req->r_tid;
911
912 if (dir) {
913 struct ceph_inode_info *ci = ceph_inode(dir);
914
915 ihold(dir);
916 req->r_unsafe_dir = dir;
917 spin_lock(&ci->i_unsafe_lock);
918 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
919 spin_unlock(&ci->i_unsafe_lock);
920 }
921}
922
923static void __unregister_request(struct ceph_mds_client *mdsc,
924 struct ceph_mds_request *req)
925{
926 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
927
928 /* Never leave an unregistered request on an unsafe list! */
929 list_del_init(&req->r_unsafe_item);
930
931 if (req->r_tid == mdsc->oldest_tid) {
932 struct rb_node *p = rb_next(&req->r_node);
933 mdsc->oldest_tid = 0;
934 while (p) {
935 struct ceph_mds_request *next_req =
936 rb_entry(p, struct ceph_mds_request, r_node);
937 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
938 mdsc->oldest_tid = next_req->r_tid;
939 break;
940 }
941 p = rb_next(p);
942 }
943 }
944
945 erase_request(&mdsc->request_tree, req);
946
947 if (req->r_unsafe_dir) {
948 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
949 spin_lock(&ci->i_unsafe_lock);
950 list_del_init(&req->r_unsafe_dir_item);
951 spin_unlock(&ci->i_unsafe_lock);
952 }
953 if (req->r_target_inode &&
954 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
955 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
956 spin_lock(&ci->i_unsafe_lock);
957 list_del_init(&req->r_unsafe_target_item);
958 spin_unlock(&ci->i_unsafe_lock);
959 }
960
961 if (req->r_unsafe_dir) {
962 iput(req->r_unsafe_dir);
963 req->r_unsafe_dir = NULL;
964 }
965
966 complete_all(&req->r_safe_completion);
967
968 ceph_mdsc_put_request(req);
969}
970
971/*
972 * Walk back up the dentry tree until we hit a dentry representing a
973 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
974 * when calling this) to ensure that the objects won't disappear while we're
975 * working with them. Once we hit a candidate dentry, we attempt to take a
976 * reference to it, and return that as the result.
977 */
978static struct inode *get_nonsnap_parent(struct dentry *dentry)
979{
980 struct inode *inode = NULL;
981
982 while (dentry && !IS_ROOT(dentry)) {
983 inode = d_inode_rcu(dentry);
984 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
985 break;
986 dentry = dentry->d_parent;
987 }
988 if (inode)
989 inode = igrab(inode);
990 return inode;
991}
992
993/*
994 * Choose mds to send request to next. If there is a hint set in the
995 * request (e.g., due to a prior forward hint from the mds), use that.
996 * Otherwise, consult frag tree and/or caps to identify the
997 * appropriate mds. If all else fails, choose randomly.
998 *
999 * Called under mdsc->mutex.
1000 */
1001static int __choose_mds(struct ceph_mds_client *mdsc,
1002 struct ceph_mds_request *req,
1003 bool *random)
1004{
1005 struct inode *inode;
1006 struct ceph_inode_info *ci;
1007 struct ceph_cap *cap;
1008 int mode = req->r_direct_mode;
1009 int mds = -1;
1010 u32 hash = req->r_direct_hash;
1011 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1012
1013 if (random)
1014 *random = false;
1015
1016 /*
1017 * is there a specific mds we should try? ignore hint if we have
1018 * no session and the mds is not up (active or recovering).
1019 */
1020 if (req->r_resend_mds >= 0 &&
1021 (__have_session(mdsc, req->r_resend_mds) ||
1022 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1023 dout("%s using resend_mds mds%d\n", __func__,
1024 req->r_resend_mds);
1025 return req->r_resend_mds;
1026 }
1027
1028 if (mode == USE_RANDOM_MDS)
1029 goto random;
1030
1031 inode = NULL;
1032 if (req->r_inode) {
1033 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1034 inode = req->r_inode;
1035 ihold(inode);
1036 } else {
1037 /* req->r_dentry is non-null for LSSNAP request */
1038 rcu_read_lock();
1039 inode = get_nonsnap_parent(req->r_dentry);
1040 rcu_read_unlock();
1041 dout("%s using snapdir's parent %p\n", __func__, inode);
1042 }
1043 } else if (req->r_dentry) {
1044 /* ignore race with rename; old or new d_parent is okay */
1045 struct dentry *parent;
1046 struct inode *dir;
1047
1048 rcu_read_lock();
1049 parent = READ_ONCE(req->r_dentry->d_parent);
1050 dir = req->r_parent ? : d_inode_rcu(parent);
1051
1052 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1053 /* not this fs or parent went negative */
1054 inode = d_inode(req->r_dentry);
1055 if (inode)
1056 ihold(inode);
1057 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1058 /* direct snapped/virtual snapdir requests
1059 * based on parent dir inode */
1060 inode = get_nonsnap_parent(parent);
1061 dout("%s using nonsnap parent %p\n", __func__, inode);
1062 } else {
1063 /* dentry target */
1064 inode = d_inode(req->r_dentry);
1065 if (!inode || mode == USE_AUTH_MDS) {
1066 /* dir + name */
1067 inode = igrab(dir);
1068 hash = ceph_dentry_hash(dir, req->r_dentry);
1069 is_hash = true;
1070 } else {
1071 ihold(inode);
1072 }
1073 }
1074 rcu_read_unlock();
1075 }
1076
1077 dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1078 hash, mode);
1079 if (!inode)
1080 goto random;
1081 ci = ceph_inode(inode);
1082
1083 if (is_hash && S_ISDIR(inode->i_mode)) {
1084 struct ceph_inode_frag frag;
1085 int found;
1086
1087 ceph_choose_frag(ci, hash, &frag, &found);
1088 if (found) {
1089 if (mode == USE_ANY_MDS && frag.ndist > 0) {
1090 u8 r;
1091
1092 /* choose a random replica */
1093 get_random_bytes(&r, 1);
1094 r %= frag.ndist;
1095 mds = frag.dist[r];
1096 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1097 __func__, inode, ceph_vinop(inode),
1098 frag.frag, mds, (int)r, frag.ndist);
1099 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1100 CEPH_MDS_STATE_ACTIVE &&
1101 !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1102 goto out;
1103 }
1104
1105 /* since this file/dir wasn't known to be
1106 * replicated, then we want to look for the
1107 * authoritative mds. */
1108 if (frag.mds >= 0) {
1109 /* choose auth mds */
1110 mds = frag.mds;
1111 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1112 __func__, inode, ceph_vinop(inode),
1113 frag.frag, mds);
1114 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1115 CEPH_MDS_STATE_ACTIVE) {
1116 if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1117 mds))
1118 goto out;
1119 }
1120 }
1121 mode = USE_AUTH_MDS;
1122 }
1123 }
1124
1125 spin_lock(&ci->i_ceph_lock);
1126 cap = NULL;
1127 if (mode == USE_AUTH_MDS)
1128 cap = ci->i_auth_cap;
1129 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1130 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1131 if (!cap) {
1132 spin_unlock(&ci->i_ceph_lock);
1133 iput(inode);
1134 goto random;
1135 }
1136 mds = cap->session->s_mds;
1137 dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1138 inode, ceph_vinop(inode), mds,
1139 cap == ci->i_auth_cap ? "auth " : "", cap);
1140 spin_unlock(&ci->i_ceph_lock);
1141out:
1142 iput(inode);
1143 return mds;
1144
1145random:
1146 if (random)
1147 *random = true;
1148
1149 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1150 dout("%s chose random mds%d\n", __func__, mds);
1151 return mds;
1152}
1153
1154
1155/*
1156 * session messages
1157 */
1158static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1159{
1160 struct ceph_msg *msg;
1161 struct ceph_mds_session_head *h;
1162
1163 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1164 false);
1165 if (!msg) {
1166 pr_err("create_session_msg ENOMEM creating msg\n");
1167 return NULL;
1168 }
1169 h = msg->front.iov_base;
1170 h->op = cpu_to_le32(op);
1171 h->seq = cpu_to_le64(seq);
1172
1173 return msg;
1174}
1175
1176static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1177#define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1178static int encode_supported_features(void **p, void *end)
1179{
1180 static const size_t count = ARRAY_SIZE(feature_bits);
1181
1182 if (count > 0) {
1183 size_t i;
1184 size_t size = FEATURE_BYTES(count);
1185
1186 if (WARN_ON_ONCE(*p + 4 + size > end))
1187 return -ERANGE;
1188
1189 ceph_encode_32(p, size);
1190 memset(*p, 0, size);
1191 for (i = 0; i < count; i++)
1192 ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1193 *p += size;
1194 } else {
1195 if (WARN_ON_ONCE(*p + 4 > end))
1196 return -ERANGE;
1197
1198 ceph_encode_32(p, 0);
1199 }
1200
1201 return 0;
1202}
1203
1204static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1205#define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1206static int encode_metric_spec(void **p, void *end)
1207{
1208 static const size_t count = ARRAY_SIZE(metric_bits);
1209
1210 /* header */
1211 if (WARN_ON_ONCE(*p + 2 > end))
1212 return -ERANGE;
1213
1214 ceph_encode_8(p, 1); /* version */
1215 ceph_encode_8(p, 1); /* compat */
1216
1217 if (count > 0) {
1218 size_t i;
1219 size_t size = METRIC_BYTES(count);
1220
1221 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1222 return -ERANGE;
1223
1224 /* metric spec info length */
1225 ceph_encode_32(p, 4 + size);
1226
1227 /* metric spec */
1228 ceph_encode_32(p, size);
1229 memset(*p, 0, size);
1230 for (i = 0; i < count; i++)
1231 ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1232 *p += size;
1233 } else {
1234 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1235 return -ERANGE;
1236
1237 /* metric spec info length */
1238 ceph_encode_32(p, 4);
1239 /* metric spec */
1240 ceph_encode_32(p, 0);
1241 }
1242
1243 return 0;
1244}
1245
1246/*
1247 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1248 * to include additional client metadata fields.
1249 */
1250static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1251{
1252 struct ceph_msg *msg;
1253 struct ceph_mds_session_head *h;
1254 int i;
1255 int extra_bytes = 0;
1256 int metadata_key_count = 0;
1257 struct ceph_options *opt = mdsc->fsc->client->options;
1258 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1259 size_t size, count;
1260 void *p, *end;
1261 int ret;
1262
1263 const char* metadata[][2] = {
1264 {"hostname", mdsc->nodename},
1265 {"kernel_version", init_utsname()->release},
1266 {"entity_id", opt->name ? : ""},
1267 {"root", fsopt->server_path ? : "/"},
1268 {NULL, NULL}
1269 };
1270
1271 /* Calculate serialized length of metadata */
1272 extra_bytes = 4; /* map length */
1273 for (i = 0; metadata[i][0]; ++i) {
1274 extra_bytes += 8 + strlen(metadata[i][0]) +
1275 strlen(metadata[i][1]);
1276 metadata_key_count++;
1277 }
1278
1279 /* supported feature */
1280 size = 0;
1281 count = ARRAY_SIZE(feature_bits);
1282 if (count > 0)
1283 size = FEATURE_BYTES(count);
1284 extra_bytes += 4 + size;
1285
1286 /* metric spec */
1287 size = 0;
1288 count = ARRAY_SIZE(metric_bits);
1289 if (count > 0)
1290 size = METRIC_BYTES(count);
1291 extra_bytes += 2 + 4 + 4 + size;
1292
1293 /* Allocate the message */
1294 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1295 GFP_NOFS, false);
1296 if (!msg) {
1297 pr_err("create_session_msg ENOMEM creating msg\n");
1298 return ERR_PTR(-ENOMEM);
1299 }
1300 p = msg->front.iov_base;
1301 end = p + msg->front.iov_len;
1302
1303 h = p;
1304 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1305 h->seq = cpu_to_le64(seq);
1306
1307 /*
1308 * Serialize client metadata into waiting buffer space, using
1309 * the format that userspace expects for map<string, string>
1310 *
1311 * ClientSession messages with metadata are v4
1312 */
1313 msg->hdr.version = cpu_to_le16(4);
1314 msg->hdr.compat_version = cpu_to_le16(1);
1315
1316 /* The write pointer, following the session_head structure */
1317 p += sizeof(*h);
1318
1319 /* Number of entries in the map */
1320 ceph_encode_32(&p, metadata_key_count);
1321
1322 /* Two length-prefixed strings for each entry in the map */
1323 for (i = 0; metadata[i][0]; ++i) {
1324 size_t const key_len = strlen(metadata[i][0]);
1325 size_t const val_len = strlen(metadata[i][1]);
1326
1327 ceph_encode_32(&p, key_len);
1328 memcpy(p, metadata[i][0], key_len);
1329 p += key_len;
1330 ceph_encode_32(&p, val_len);
1331 memcpy(p, metadata[i][1], val_len);
1332 p += val_len;
1333 }
1334
1335 ret = encode_supported_features(&p, end);
1336 if (ret) {
1337 pr_err("encode_supported_features failed!\n");
1338 ceph_msg_put(msg);
1339 return ERR_PTR(ret);
1340 }
1341
1342 ret = encode_metric_spec(&p, end);
1343 if (ret) {
1344 pr_err("encode_metric_spec failed!\n");
1345 ceph_msg_put(msg);
1346 return ERR_PTR(ret);
1347 }
1348
1349 msg->front.iov_len = p - msg->front.iov_base;
1350 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1351
1352 return msg;
1353}
1354
1355/*
1356 * send session open request.
1357 *
1358 * called under mdsc->mutex
1359 */
1360static int __open_session(struct ceph_mds_client *mdsc,
1361 struct ceph_mds_session *session)
1362{
1363 struct ceph_msg *msg;
1364 int mstate;
1365 int mds = session->s_mds;
1366
1367 /* wait for mds to go active? */
1368 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1369 dout("open_session to mds%d (%s)\n", mds,
1370 ceph_mds_state_name(mstate));
1371 session->s_state = CEPH_MDS_SESSION_OPENING;
1372 session->s_renew_requested = jiffies;
1373
1374 /* send connect message */
1375 msg = create_session_open_msg(mdsc, session->s_seq);
1376 if (IS_ERR(msg))
1377 return PTR_ERR(msg);
1378 ceph_con_send(&session->s_con, msg);
1379 return 0;
1380}
1381
1382/*
1383 * open sessions for any export targets for the given mds
1384 *
1385 * called under mdsc->mutex
1386 */
1387static struct ceph_mds_session *
1388__open_export_target_session(struct ceph_mds_client *mdsc, int target)
1389{
1390 struct ceph_mds_session *session;
1391 int ret;
1392
1393 session = __ceph_lookup_mds_session(mdsc, target);
1394 if (!session) {
1395 session = register_session(mdsc, target);
1396 if (IS_ERR(session))
1397 return session;
1398 }
1399 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1400 session->s_state == CEPH_MDS_SESSION_CLOSING) {
1401 ret = __open_session(mdsc, session);
1402 if (ret)
1403 return ERR_PTR(ret);
1404 }
1405
1406 return session;
1407}
1408
1409struct ceph_mds_session *
1410ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1411{
1412 struct ceph_mds_session *session;
1413
1414 dout("open_export_target_session to mds%d\n", target);
1415
1416 mutex_lock(&mdsc->mutex);
1417 session = __open_export_target_session(mdsc, target);
1418 mutex_unlock(&mdsc->mutex);
1419
1420 return session;
1421}
1422
1423static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1424 struct ceph_mds_session *session)
1425{
1426 struct ceph_mds_info *mi;
1427 struct ceph_mds_session *ts;
1428 int i, mds = session->s_mds;
1429
1430 if (mds >= mdsc->mdsmap->possible_max_rank)
1431 return;
1432
1433 mi = &mdsc->mdsmap->m_info[mds];
1434 dout("open_export_target_sessions for mds%d (%d targets)\n",
1435 session->s_mds, mi->num_export_targets);
1436
1437 for (i = 0; i < mi->num_export_targets; i++) {
1438 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1439 ceph_put_mds_session(ts);
1440 }
1441}
1442
1443void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1444 struct ceph_mds_session *session)
1445{
1446 mutex_lock(&mdsc->mutex);
1447 __open_export_target_sessions(mdsc, session);
1448 mutex_unlock(&mdsc->mutex);
1449}
1450
1451/*
1452 * session caps
1453 */
1454
1455static void detach_cap_releases(struct ceph_mds_session *session,
1456 struct list_head *target)
1457{
1458 lockdep_assert_held(&session->s_cap_lock);
1459
1460 list_splice_init(&session->s_cap_releases, target);
1461 session->s_num_cap_releases = 0;
1462 dout("dispose_cap_releases mds%d\n", session->s_mds);
1463}
1464
1465static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1466 struct list_head *dispose)
1467{
1468 while (!list_empty(dispose)) {
1469 struct ceph_cap *cap;
1470 /* zero out the in-progress message */
1471 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1472 list_del(&cap->session_caps);
1473 ceph_put_cap(mdsc, cap);
1474 }
1475}
1476
1477static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1478 struct ceph_mds_session *session)
1479{
1480 struct ceph_mds_request *req;
1481 struct rb_node *p;
1482
1483 dout("cleanup_session_requests mds%d\n", session->s_mds);
1484 mutex_lock(&mdsc->mutex);
1485 while (!list_empty(&session->s_unsafe)) {
1486 req = list_first_entry(&session->s_unsafe,
1487 struct ceph_mds_request, r_unsafe_item);
1488 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1489 req->r_tid);
1490 if (req->r_target_inode)
1491 mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1492 if (req->r_unsafe_dir)
1493 mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1494 __unregister_request(mdsc, req);
1495 }
1496 /* zero r_attempts, so kick_requests() will re-send requests */
1497 p = rb_first(&mdsc->request_tree);
1498 while (p) {
1499 req = rb_entry(p, struct ceph_mds_request, r_node);
1500 p = rb_next(p);
1501 if (req->r_session &&
1502 req->r_session->s_mds == session->s_mds)
1503 req->r_attempts = 0;
1504 }
1505 mutex_unlock(&mdsc->mutex);
1506}
1507
1508/*
1509 * Helper to safely iterate over all caps associated with a session, with
1510 * special care taken to handle a racing __ceph_remove_cap().
1511 *
1512 * Caller must hold session s_mutex.
1513 */
1514int ceph_iterate_session_caps(struct ceph_mds_session *session,
1515 int (*cb)(struct inode *, struct ceph_cap *,
1516 void *), void *arg)
1517{
1518 struct list_head *p;
1519 struct ceph_cap *cap;
1520 struct inode *inode, *last_inode = NULL;
1521 struct ceph_cap *old_cap = NULL;
1522 int ret;
1523
1524 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1525 spin_lock(&session->s_cap_lock);
1526 p = session->s_caps.next;
1527 while (p != &session->s_caps) {
1528 cap = list_entry(p, struct ceph_cap, session_caps);
1529 inode = igrab(&cap->ci->vfs_inode);
1530 if (!inode) {
1531 p = p->next;
1532 continue;
1533 }
1534 session->s_cap_iterator = cap;
1535 spin_unlock(&session->s_cap_lock);
1536
1537 if (last_inode) {
1538 iput(last_inode);
1539 last_inode = NULL;
1540 }
1541 if (old_cap) {
1542 ceph_put_cap(session->s_mdsc, old_cap);
1543 old_cap = NULL;
1544 }
1545
1546 ret = cb(inode, cap, arg);
1547 last_inode = inode;
1548
1549 spin_lock(&session->s_cap_lock);
1550 p = p->next;
1551 if (!cap->ci) {
1552 dout("iterate_session_caps finishing cap %p removal\n",
1553 cap);
1554 BUG_ON(cap->session != session);
1555 cap->session = NULL;
1556 list_del_init(&cap->session_caps);
1557 session->s_nr_caps--;
1558 atomic64_dec(&session->s_mdsc->metric.total_caps);
1559 if (cap->queue_release)
1560 __ceph_queue_cap_release(session, cap);
1561 else
1562 old_cap = cap; /* put_cap it w/o locks held */
1563 }
1564 if (ret < 0)
1565 goto out;
1566 }
1567 ret = 0;
1568out:
1569 session->s_cap_iterator = NULL;
1570 spin_unlock(&session->s_cap_lock);
1571
1572 iput(last_inode);
1573 if (old_cap)
1574 ceph_put_cap(session->s_mdsc, old_cap);
1575
1576 return ret;
1577}
1578
1579static int remove_capsnaps(struct ceph_mds_client *mdsc, struct inode *inode)
1580{
1581 struct ceph_inode_info *ci = ceph_inode(inode);
1582 struct ceph_cap_snap *capsnap;
1583 int capsnap_release = 0;
1584
1585 lockdep_assert_held(&ci->i_ceph_lock);
1586
1587 dout("removing capsnaps, ci is %p, inode is %p\n", ci, inode);
1588
1589 while (!list_empty(&ci->i_cap_snaps)) {
1590 capsnap = list_first_entry(&ci->i_cap_snaps,
1591 struct ceph_cap_snap, ci_item);
1592 __ceph_remove_capsnap(inode, capsnap, NULL, NULL);
1593 ceph_put_snap_context(capsnap->context);
1594 ceph_put_cap_snap(capsnap);
1595 capsnap_release++;
1596 }
1597 wake_up_all(&ci->i_cap_wq);
1598 wake_up_all(&mdsc->cap_flushing_wq);
1599 return capsnap_release;
1600}
1601
1602static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1603 void *arg)
1604{
1605 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1606 struct ceph_mds_client *mdsc = fsc->mdsc;
1607 struct ceph_inode_info *ci = ceph_inode(inode);
1608 LIST_HEAD(to_remove);
1609 bool dirty_dropped = false;
1610 bool invalidate = false;
1611 int capsnap_release = 0;
1612
1613 dout("removing cap %p, ci is %p, inode is %p\n",
1614 cap, ci, &ci->vfs_inode);
1615 spin_lock(&ci->i_ceph_lock);
1616 __ceph_remove_cap(cap, false);
1617 if (!ci->i_auth_cap) {
1618 struct ceph_cap_flush *cf;
1619
1620 if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1621 if (inode->i_data.nrpages > 0)
1622 invalidate = true;
1623 if (ci->i_wrbuffer_ref > 0)
1624 mapping_set_error(&inode->i_data, -EIO);
1625 }
1626
1627 while (!list_empty(&ci->i_cap_flush_list)) {
1628 cf = list_first_entry(&ci->i_cap_flush_list,
1629 struct ceph_cap_flush, i_list);
1630 list_move(&cf->i_list, &to_remove);
1631 }
1632
1633 spin_lock(&mdsc->cap_dirty_lock);
1634
1635 list_for_each_entry(cf, &to_remove, i_list)
1636 list_del_init(&cf->g_list);
1637
1638 if (!list_empty(&ci->i_dirty_item)) {
1639 pr_warn_ratelimited(
1640 " dropping dirty %s state for %p %lld\n",
1641 ceph_cap_string(ci->i_dirty_caps),
1642 inode, ceph_ino(inode));
1643 ci->i_dirty_caps = 0;
1644 list_del_init(&ci->i_dirty_item);
1645 dirty_dropped = true;
1646 }
1647 if (!list_empty(&ci->i_flushing_item)) {
1648 pr_warn_ratelimited(
1649 " dropping dirty+flushing %s state for %p %lld\n",
1650 ceph_cap_string(ci->i_flushing_caps),
1651 inode, ceph_ino(inode));
1652 ci->i_flushing_caps = 0;
1653 list_del_init(&ci->i_flushing_item);
1654 mdsc->num_cap_flushing--;
1655 dirty_dropped = true;
1656 }
1657 spin_unlock(&mdsc->cap_dirty_lock);
1658
1659 if (dirty_dropped) {
1660 mapping_set_error(inode->i_mapping, -EIO);
1661
1662 if (ci->i_wrbuffer_ref_head == 0 &&
1663 ci->i_wr_ref == 0 &&
1664 ci->i_dirty_caps == 0 &&
1665 ci->i_flushing_caps == 0) {
1666 ceph_put_snap_context(ci->i_head_snapc);
1667 ci->i_head_snapc = NULL;
1668 }
1669 }
1670
1671 if (atomic_read(&ci->i_filelock_ref) > 0) {
1672 /* make further file lock syscall return -EIO */
1673 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1674 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1675 inode, ceph_ino(inode));
1676 }
1677
1678 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1679 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1680 ci->i_prealloc_cap_flush = NULL;
1681 }
1682
1683 if (!list_empty(&ci->i_cap_snaps))
1684 capsnap_release = remove_capsnaps(mdsc, inode);
1685 }
1686 spin_unlock(&ci->i_ceph_lock);
1687 while (!list_empty(&to_remove)) {
1688 struct ceph_cap_flush *cf;
1689 cf = list_first_entry(&to_remove,
1690 struct ceph_cap_flush, i_list);
1691 list_del_init(&cf->i_list);
1692 if (!cf->is_capsnap)
1693 ceph_free_cap_flush(cf);
1694 }
1695
1696 wake_up_all(&ci->i_cap_wq);
1697 if (invalidate)
1698 ceph_queue_invalidate(inode);
1699 if (dirty_dropped)
1700 iput(inode);
1701 while (capsnap_release--)
1702 iput(inode);
1703 return 0;
1704}
1705
1706/*
1707 * caller must hold session s_mutex
1708 */
1709static void remove_session_caps(struct ceph_mds_session *session)
1710{
1711 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1712 struct super_block *sb = fsc->sb;
1713 LIST_HEAD(dispose);
1714
1715 dout("remove_session_caps on %p\n", session);
1716 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1717
1718 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1719
1720 spin_lock(&session->s_cap_lock);
1721 if (session->s_nr_caps > 0) {
1722 struct inode *inode;
1723 struct ceph_cap *cap, *prev = NULL;
1724 struct ceph_vino vino;
1725 /*
1726 * iterate_session_caps() skips inodes that are being
1727 * deleted, we need to wait until deletions are complete.
1728 * __wait_on_freeing_inode() is designed for the job,
1729 * but it is not exported, so use lookup inode function
1730 * to access it.
1731 */
1732 while (!list_empty(&session->s_caps)) {
1733 cap = list_entry(session->s_caps.next,
1734 struct ceph_cap, session_caps);
1735 if (cap == prev)
1736 break;
1737 prev = cap;
1738 vino = cap->ci->i_vino;
1739 spin_unlock(&session->s_cap_lock);
1740
1741 inode = ceph_find_inode(sb, vino);
1742 iput(inode);
1743
1744 spin_lock(&session->s_cap_lock);
1745 }
1746 }
1747
1748 // drop cap expires and unlock s_cap_lock
1749 detach_cap_releases(session, &dispose);
1750
1751 BUG_ON(session->s_nr_caps > 0);
1752 BUG_ON(!list_empty(&session->s_cap_flushing));
1753 spin_unlock(&session->s_cap_lock);
1754 dispose_cap_releases(session->s_mdsc, &dispose);
1755}
1756
1757enum {
1758 RECONNECT,
1759 RENEWCAPS,
1760 FORCE_RO,
1761};
1762
1763/*
1764 * wake up any threads waiting on this session's caps. if the cap is
1765 * old (didn't get renewed on the client reconnect), remove it now.
1766 *
1767 * caller must hold s_mutex.
1768 */
1769static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1770 void *arg)
1771{
1772 struct ceph_inode_info *ci = ceph_inode(inode);
1773 unsigned long ev = (unsigned long)arg;
1774
1775 if (ev == RECONNECT) {
1776 spin_lock(&ci->i_ceph_lock);
1777 ci->i_wanted_max_size = 0;
1778 ci->i_requested_max_size = 0;
1779 spin_unlock(&ci->i_ceph_lock);
1780 } else if (ev == RENEWCAPS) {
1781 if (cap->cap_gen < atomic_read(&cap->session->s_cap_gen)) {
1782 /* mds did not re-issue stale cap */
1783 spin_lock(&ci->i_ceph_lock);
1784 cap->issued = cap->implemented = CEPH_CAP_PIN;
1785 spin_unlock(&ci->i_ceph_lock);
1786 }
1787 } else if (ev == FORCE_RO) {
1788 }
1789 wake_up_all(&ci->i_cap_wq);
1790 return 0;
1791}
1792
1793static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1794{
1795 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1796 ceph_iterate_session_caps(session, wake_up_session_cb,
1797 (void *)(unsigned long)ev);
1798}
1799
1800/*
1801 * Send periodic message to MDS renewing all currently held caps. The
1802 * ack will reset the expiration for all caps from this session.
1803 *
1804 * caller holds s_mutex
1805 */
1806static int send_renew_caps(struct ceph_mds_client *mdsc,
1807 struct ceph_mds_session *session)
1808{
1809 struct ceph_msg *msg;
1810 int state;
1811
1812 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1813 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1814 pr_info("mds%d caps stale\n", session->s_mds);
1815 session->s_renew_requested = jiffies;
1816
1817 /* do not try to renew caps until a recovering mds has reconnected
1818 * with its clients. */
1819 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1820 if (state < CEPH_MDS_STATE_RECONNECT) {
1821 dout("send_renew_caps ignoring mds%d (%s)\n",
1822 session->s_mds, ceph_mds_state_name(state));
1823 return 0;
1824 }
1825
1826 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1827 ceph_mds_state_name(state));
1828 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1829 ++session->s_renew_seq);
1830 if (!msg)
1831 return -ENOMEM;
1832 ceph_con_send(&session->s_con, msg);
1833 return 0;
1834}
1835
1836static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1837 struct ceph_mds_session *session, u64 seq)
1838{
1839 struct ceph_msg *msg;
1840
1841 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1842 session->s_mds, ceph_session_state_name(session->s_state), seq);
1843 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1844 if (!msg)
1845 return -ENOMEM;
1846 ceph_con_send(&session->s_con, msg);
1847 return 0;
1848}
1849
1850
1851/*
1852 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1853 *
1854 * Called under session->s_mutex
1855 */
1856static void renewed_caps(struct ceph_mds_client *mdsc,
1857 struct ceph_mds_session *session, int is_renew)
1858{
1859 int was_stale;
1860 int wake = 0;
1861
1862 spin_lock(&session->s_cap_lock);
1863 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1864
1865 session->s_cap_ttl = session->s_renew_requested +
1866 mdsc->mdsmap->m_session_timeout*HZ;
1867
1868 if (was_stale) {
1869 if (time_before(jiffies, session->s_cap_ttl)) {
1870 pr_info("mds%d caps renewed\n", session->s_mds);
1871 wake = 1;
1872 } else {
1873 pr_info("mds%d caps still stale\n", session->s_mds);
1874 }
1875 }
1876 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1877 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1878 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1879 spin_unlock(&session->s_cap_lock);
1880
1881 if (wake)
1882 wake_up_session_caps(session, RENEWCAPS);
1883}
1884
1885/*
1886 * send a session close request
1887 */
1888static int request_close_session(struct ceph_mds_session *session)
1889{
1890 struct ceph_msg *msg;
1891
1892 dout("request_close_session mds%d state %s seq %lld\n",
1893 session->s_mds, ceph_session_state_name(session->s_state),
1894 session->s_seq);
1895 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1896 if (!msg)
1897 return -ENOMEM;
1898 ceph_con_send(&session->s_con, msg);
1899 return 1;
1900}
1901
1902/*
1903 * Called with s_mutex held.
1904 */
1905static int __close_session(struct ceph_mds_client *mdsc,
1906 struct ceph_mds_session *session)
1907{
1908 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1909 return 0;
1910 session->s_state = CEPH_MDS_SESSION_CLOSING;
1911 return request_close_session(session);
1912}
1913
1914static bool drop_negative_children(struct dentry *dentry)
1915{
1916 struct dentry *child;
1917 bool all_negative = true;
1918
1919 if (!d_is_dir(dentry))
1920 goto out;
1921
1922 spin_lock(&dentry->d_lock);
1923 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1924 if (d_really_is_positive(child)) {
1925 all_negative = false;
1926 break;
1927 }
1928 }
1929 spin_unlock(&dentry->d_lock);
1930
1931 if (all_negative)
1932 shrink_dcache_parent(dentry);
1933out:
1934 return all_negative;
1935}
1936
1937/*
1938 * Trim old(er) caps.
1939 *
1940 * Because we can't cache an inode without one or more caps, we do
1941 * this indirectly: if a cap is unused, we prune its aliases, at which
1942 * point the inode will hopefully get dropped to.
1943 *
1944 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1945 * memory pressure from the MDS, though, so it needn't be perfect.
1946 */
1947static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1948{
1949 int *remaining = arg;
1950 struct ceph_inode_info *ci = ceph_inode(inode);
1951 int used, wanted, oissued, mine;
1952
1953 if (*remaining <= 0)
1954 return -1;
1955
1956 spin_lock(&ci->i_ceph_lock);
1957 mine = cap->issued | cap->implemented;
1958 used = __ceph_caps_used(ci);
1959 wanted = __ceph_caps_file_wanted(ci);
1960 oissued = __ceph_caps_issued_other(ci, cap);
1961
1962 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1963 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1964 ceph_cap_string(used), ceph_cap_string(wanted));
1965 if (cap == ci->i_auth_cap) {
1966 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1967 !list_empty(&ci->i_cap_snaps))
1968 goto out;
1969 if ((used | wanted) & CEPH_CAP_ANY_WR)
1970 goto out;
1971 /* Note: it's possible that i_filelock_ref becomes non-zero
1972 * after dropping auth caps. It doesn't hurt because reply
1973 * of lock mds request will re-add auth caps. */
1974 if (atomic_read(&ci->i_filelock_ref) > 0)
1975 goto out;
1976 }
1977 /* The inode has cached pages, but it's no longer used.
1978 * we can safely drop it */
1979 if (S_ISREG(inode->i_mode) &&
1980 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1981 !(oissued & CEPH_CAP_FILE_CACHE)) {
1982 used = 0;
1983 oissued = 0;
1984 }
1985 if ((used | wanted) & ~oissued & mine)
1986 goto out; /* we need these caps */
1987
1988 if (oissued) {
1989 /* we aren't the only cap.. just remove us */
1990 __ceph_remove_cap(cap, true);
1991 (*remaining)--;
1992 } else {
1993 struct dentry *dentry;
1994 /* try dropping referring dentries */
1995 spin_unlock(&ci->i_ceph_lock);
1996 dentry = d_find_any_alias(inode);
1997 if (dentry && drop_negative_children(dentry)) {
1998 int count;
1999 dput(dentry);
2000 d_prune_aliases(inode);
2001 count = atomic_read(&inode->i_count);
2002 if (count == 1)
2003 (*remaining)--;
2004 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
2005 inode, cap, count);
2006 } else {
2007 dput(dentry);
2008 }
2009 return 0;
2010 }
2011
2012out:
2013 spin_unlock(&ci->i_ceph_lock);
2014 return 0;
2015}
2016
2017/*
2018 * Trim session cap count down to some max number.
2019 */
2020int ceph_trim_caps(struct ceph_mds_client *mdsc,
2021 struct ceph_mds_session *session,
2022 int max_caps)
2023{
2024 int trim_caps = session->s_nr_caps - max_caps;
2025
2026 dout("trim_caps mds%d start: %d / %d, trim %d\n",
2027 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
2028 if (trim_caps > 0) {
2029 int remaining = trim_caps;
2030
2031 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2032 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2033 session->s_mds, session->s_nr_caps, max_caps,
2034 trim_caps - remaining);
2035 }
2036
2037 ceph_flush_cap_releases(mdsc, session);
2038 return 0;
2039}
2040
2041static int check_caps_flush(struct ceph_mds_client *mdsc,
2042 u64 want_flush_tid)
2043{
2044 int ret = 1;
2045
2046 spin_lock(&mdsc->cap_dirty_lock);
2047 if (!list_empty(&mdsc->cap_flush_list)) {
2048 struct ceph_cap_flush *cf =
2049 list_first_entry(&mdsc->cap_flush_list,
2050 struct ceph_cap_flush, g_list);
2051 if (cf->tid <= want_flush_tid) {
2052 dout("check_caps_flush still flushing tid "
2053 "%llu <= %llu\n", cf->tid, want_flush_tid);
2054 ret = 0;
2055 }
2056 }
2057 spin_unlock(&mdsc->cap_dirty_lock);
2058 return ret;
2059}
2060
2061/*
2062 * flush all dirty inode data to disk.
2063 *
2064 * returns true if we've flushed through want_flush_tid
2065 */
2066static void wait_caps_flush(struct ceph_mds_client *mdsc,
2067 u64 want_flush_tid)
2068{
2069 dout("check_caps_flush want %llu\n", want_flush_tid);
2070
2071 wait_event(mdsc->cap_flushing_wq,
2072 check_caps_flush(mdsc, want_flush_tid));
2073
2074 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2075}
2076
2077/*
2078 * called under s_mutex
2079 */
2080static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2081 struct ceph_mds_session *session)
2082{
2083 struct ceph_msg *msg = NULL;
2084 struct ceph_mds_cap_release *head;
2085 struct ceph_mds_cap_item *item;
2086 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2087 struct ceph_cap *cap;
2088 LIST_HEAD(tmp_list);
2089 int num_cap_releases;
2090 __le32 barrier, *cap_barrier;
2091
2092 down_read(&osdc->lock);
2093 barrier = cpu_to_le32(osdc->epoch_barrier);
2094 up_read(&osdc->lock);
2095
2096 spin_lock(&session->s_cap_lock);
2097again:
2098 list_splice_init(&session->s_cap_releases, &tmp_list);
2099 num_cap_releases = session->s_num_cap_releases;
2100 session->s_num_cap_releases = 0;
2101 spin_unlock(&session->s_cap_lock);
2102
2103 while (!list_empty(&tmp_list)) {
2104 if (!msg) {
2105 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2106 PAGE_SIZE, GFP_NOFS, false);
2107 if (!msg)
2108 goto out_err;
2109 head = msg->front.iov_base;
2110 head->num = cpu_to_le32(0);
2111 msg->front.iov_len = sizeof(*head);
2112
2113 msg->hdr.version = cpu_to_le16(2);
2114 msg->hdr.compat_version = cpu_to_le16(1);
2115 }
2116
2117 cap = list_first_entry(&tmp_list, struct ceph_cap,
2118 session_caps);
2119 list_del(&cap->session_caps);
2120 num_cap_releases--;
2121
2122 head = msg->front.iov_base;
2123 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2124 &head->num);
2125 item = msg->front.iov_base + msg->front.iov_len;
2126 item->ino = cpu_to_le64(cap->cap_ino);
2127 item->cap_id = cpu_to_le64(cap->cap_id);
2128 item->migrate_seq = cpu_to_le32(cap->mseq);
2129 item->seq = cpu_to_le32(cap->issue_seq);
2130 msg->front.iov_len += sizeof(*item);
2131
2132 ceph_put_cap(mdsc, cap);
2133
2134 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2135 // Append cap_barrier field
2136 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2137 *cap_barrier = barrier;
2138 msg->front.iov_len += sizeof(*cap_barrier);
2139
2140 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2141 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2142 ceph_con_send(&session->s_con, msg);
2143 msg = NULL;
2144 }
2145 }
2146
2147 BUG_ON(num_cap_releases != 0);
2148
2149 spin_lock(&session->s_cap_lock);
2150 if (!list_empty(&session->s_cap_releases))
2151 goto again;
2152 spin_unlock(&session->s_cap_lock);
2153
2154 if (msg) {
2155 // Append cap_barrier field
2156 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2157 *cap_barrier = barrier;
2158 msg->front.iov_len += sizeof(*cap_barrier);
2159
2160 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2161 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2162 ceph_con_send(&session->s_con, msg);
2163 }
2164 return;
2165out_err:
2166 pr_err("send_cap_releases mds%d, failed to allocate message\n",
2167 session->s_mds);
2168 spin_lock(&session->s_cap_lock);
2169 list_splice(&tmp_list, &session->s_cap_releases);
2170 session->s_num_cap_releases += num_cap_releases;
2171 spin_unlock(&session->s_cap_lock);
2172}
2173
2174static void ceph_cap_release_work(struct work_struct *work)
2175{
2176 struct ceph_mds_session *session =
2177 container_of(work, struct ceph_mds_session, s_cap_release_work);
2178
2179 mutex_lock(&session->s_mutex);
2180 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2181 session->s_state == CEPH_MDS_SESSION_HUNG)
2182 ceph_send_cap_releases(session->s_mdsc, session);
2183 mutex_unlock(&session->s_mutex);
2184 ceph_put_mds_session(session);
2185}
2186
2187void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2188 struct ceph_mds_session *session)
2189{
2190 if (mdsc->stopping)
2191 return;
2192
2193 ceph_get_mds_session(session);
2194 if (queue_work(mdsc->fsc->cap_wq,
2195 &session->s_cap_release_work)) {
2196 dout("cap release work queued\n");
2197 } else {
2198 ceph_put_mds_session(session);
2199 dout("failed to queue cap release work\n");
2200 }
2201}
2202
2203/*
2204 * caller holds session->s_cap_lock
2205 */
2206void __ceph_queue_cap_release(struct ceph_mds_session *session,
2207 struct ceph_cap *cap)
2208{
2209 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2210 session->s_num_cap_releases++;
2211
2212 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2213 ceph_flush_cap_releases(session->s_mdsc, session);
2214}
2215
2216static void ceph_cap_reclaim_work(struct work_struct *work)
2217{
2218 struct ceph_mds_client *mdsc =
2219 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2220 int ret = ceph_trim_dentries(mdsc);
2221 if (ret == -EAGAIN)
2222 ceph_queue_cap_reclaim_work(mdsc);
2223}
2224
2225void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2226{
2227 if (mdsc->stopping)
2228 return;
2229
2230 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2231 dout("caps reclaim work queued\n");
2232 } else {
2233 dout("failed to queue caps release work\n");
2234 }
2235}
2236
2237void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2238{
2239 int val;
2240 if (!nr)
2241 return;
2242 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2243 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2244 atomic_set(&mdsc->cap_reclaim_pending, 0);
2245 ceph_queue_cap_reclaim_work(mdsc);
2246 }
2247}
2248
2249/*
2250 * requests
2251 */
2252
2253int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2254 struct inode *dir)
2255{
2256 struct ceph_inode_info *ci = ceph_inode(dir);
2257 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2258 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2259 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2260 unsigned int num_entries;
2261 int order;
2262
2263 spin_lock(&ci->i_ceph_lock);
2264 num_entries = ci->i_files + ci->i_subdirs;
2265 spin_unlock(&ci->i_ceph_lock);
2266 num_entries = max(num_entries, 1U);
2267 num_entries = min(num_entries, opt->max_readdir);
2268
2269 order = get_order(size * num_entries);
2270 while (order >= 0) {
2271 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2272 __GFP_NOWARN,
2273 order);
2274 if (rinfo->dir_entries)
2275 break;
2276 order--;
2277 }
2278 if (!rinfo->dir_entries)
2279 return -ENOMEM;
2280
2281 num_entries = (PAGE_SIZE << order) / size;
2282 num_entries = min(num_entries, opt->max_readdir);
2283
2284 rinfo->dir_buf_size = PAGE_SIZE << order;
2285 req->r_num_caps = num_entries + 1;
2286 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2287 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2288 return 0;
2289}
2290
2291/*
2292 * Create an mds request.
2293 */
2294struct ceph_mds_request *
2295ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2296{
2297 struct ceph_mds_request *req;
2298
2299 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2300 if (!req)
2301 return ERR_PTR(-ENOMEM);
2302
2303 mutex_init(&req->r_fill_mutex);
2304 req->r_mdsc = mdsc;
2305 req->r_started = jiffies;
2306 req->r_start_latency = ktime_get();
2307 req->r_resend_mds = -1;
2308 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2309 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2310 req->r_fmode = -1;
2311 kref_init(&req->r_kref);
2312 RB_CLEAR_NODE(&req->r_node);
2313 INIT_LIST_HEAD(&req->r_wait);
2314 init_completion(&req->r_completion);
2315 init_completion(&req->r_safe_completion);
2316 INIT_LIST_HEAD(&req->r_unsafe_item);
2317
2318 ktime_get_coarse_real_ts64(&req->r_stamp);
2319
2320 req->r_op = op;
2321 req->r_direct_mode = mode;
2322 return req;
2323}
2324
2325/*
2326 * return oldest (lowest) request, tid in request tree, 0 if none.
2327 *
2328 * called under mdsc->mutex.
2329 */
2330static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2331{
2332 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2333 return NULL;
2334 return rb_entry(rb_first(&mdsc->request_tree),
2335 struct ceph_mds_request, r_node);
2336}
2337
2338static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2339{
2340 return mdsc->oldest_tid;
2341}
2342
2343/*
2344 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2345 * on build_path_from_dentry in fs/cifs/dir.c.
2346 *
2347 * If @stop_on_nosnap, generate path relative to the first non-snapped
2348 * inode.
2349 *
2350 * Encode hidden .snap dirs as a double /, i.e.
2351 * foo/.snap/bar -> foo//bar
2352 */
2353char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2354 int stop_on_nosnap)
2355{
2356 struct dentry *temp;
2357 char *path;
2358 int pos;
2359 unsigned seq;
2360 u64 base;
2361
2362 if (!dentry)
2363 return ERR_PTR(-EINVAL);
2364
2365 path = __getname();
2366 if (!path)
2367 return ERR_PTR(-ENOMEM);
2368retry:
2369 pos = PATH_MAX - 1;
2370 path[pos] = '\0';
2371
2372 seq = read_seqbegin(&rename_lock);
2373 rcu_read_lock();
2374 temp = dentry;
2375 for (;;) {
2376 struct inode *inode;
2377
2378 spin_lock(&temp->d_lock);
2379 inode = d_inode(temp);
2380 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2381 dout("build_path path+%d: %p SNAPDIR\n",
2382 pos, temp);
2383 } else if (stop_on_nosnap && inode && dentry != temp &&
2384 ceph_snap(inode) == CEPH_NOSNAP) {
2385 spin_unlock(&temp->d_lock);
2386 pos++; /* get rid of any prepended '/' */
2387 break;
2388 } else {
2389 pos -= temp->d_name.len;
2390 if (pos < 0) {
2391 spin_unlock(&temp->d_lock);
2392 break;
2393 }
2394 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2395 }
2396 spin_unlock(&temp->d_lock);
2397 temp = READ_ONCE(temp->d_parent);
2398
2399 /* Are we at the root? */
2400 if (IS_ROOT(temp))
2401 break;
2402
2403 /* Are we out of buffer? */
2404 if (--pos < 0)
2405 break;
2406
2407 path[pos] = '/';
2408 }
2409 base = ceph_ino(d_inode(temp));
2410 rcu_read_unlock();
2411
2412 if (read_seqretry(&rename_lock, seq))
2413 goto retry;
2414
2415 if (pos < 0) {
2416 /*
2417 * A rename didn't occur, but somehow we didn't end up where
2418 * we thought we would. Throw a warning and try again.
2419 */
2420 pr_warn("build_path did not end path lookup where "
2421 "expected, pos is %d\n", pos);
2422 goto retry;
2423 }
2424
2425 *pbase = base;
2426 *plen = PATH_MAX - 1 - pos;
2427 dout("build_path on %p %d built %llx '%.*s'\n",
2428 dentry, d_count(dentry), base, *plen, path + pos);
2429 return path + pos;
2430}
2431
2432static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2433 const char **ppath, int *ppathlen, u64 *pino,
2434 bool *pfreepath, bool parent_locked)
2435{
2436 char *path;
2437
2438 rcu_read_lock();
2439 if (!dir)
2440 dir = d_inode_rcu(dentry->d_parent);
2441 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2442 *pino = ceph_ino(dir);
2443 rcu_read_unlock();
2444 *ppath = dentry->d_name.name;
2445 *ppathlen = dentry->d_name.len;
2446 return 0;
2447 }
2448 rcu_read_unlock();
2449 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2450 if (IS_ERR(path))
2451 return PTR_ERR(path);
2452 *ppath = path;
2453 *pfreepath = true;
2454 return 0;
2455}
2456
2457static int build_inode_path(struct inode *inode,
2458 const char **ppath, int *ppathlen, u64 *pino,
2459 bool *pfreepath)
2460{
2461 struct dentry *dentry;
2462 char *path;
2463
2464 if (ceph_snap(inode) == CEPH_NOSNAP) {
2465 *pino = ceph_ino(inode);
2466 *ppathlen = 0;
2467 return 0;
2468 }
2469 dentry = d_find_alias(inode);
2470 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2471 dput(dentry);
2472 if (IS_ERR(path))
2473 return PTR_ERR(path);
2474 *ppath = path;
2475 *pfreepath = true;
2476 return 0;
2477}
2478
2479/*
2480 * request arguments may be specified via an inode *, a dentry *, or
2481 * an explicit ino+path.
2482 */
2483static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2484 struct inode *rdiri, const char *rpath,
2485 u64 rino, const char **ppath, int *pathlen,
2486 u64 *ino, bool *freepath, bool parent_locked)
2487{
2488 int r = 0;
2489
2490 if (rinode) {
2491 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2492 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2493 ceph_snap(rinode));
2494 } else if (rdentry) {
2495 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2496 freepath, parent_locked);
2497 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2498 *ppath);
2499 } else if (rpath || rino) {
2500 *ino = rino;
2501 *ppath = rpath;
2502 *pathlen = rpath ? strlen(rpath) : 0;
2503 dout(" path %.*s\n", *pathlen, rpath);
2504 }
2505
2506 return r;
2507}
2508
2509static void encode_timestamp_and_gids(void **p,
2510 const struct ceph_mds_request *req)
2511{
2512 struct ceph_timespec ts;
2513 int i;
2514
2515 ceph_encode_timespec64(&ts, &req->r_stamp);
2516 ceph_encode_copy(p, &ts, sizeof(ts));
2517
2518 /* gid_list */
2519 ceph_encode_32(p, req->r_cred->group_info->ngroups);
2520 for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2521 ceph_encode_64(p, from_kgid(&init_user_ns,
2522 req->r_cred->group_info->gid[i]));
2523}
2524
2525/*
2526 * called under mdsc->mutex
2527 */
2528static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2529 struct ceph_mds_request *req,
2530 bool drop_cap_releases)
2531{
2532 int mds = session->s_mds;
2533 struct ceph_mds_client *mdsc = session->s_mdsc;
2534 struct ceph_msg *msg;
2535 struct ceph_mds_request_head_old *head;
2536 const char *path1 = NULL;
2537 const char *path2 = NULL;
2538 u64 ino1 = 0, ino2 = 0;
2539 int pathlen1 = 0, pathlen2 = 0;
2540 bool freepath1 = false, freepath2 = false;
2541 int len;
2542 u16 releases;
2543 void *p, *end;
2544 int ret;
2545 bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2546
2547 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2548 req->r_parent, req->r_path1, req->r_ino1.ino,
2549 &path1, &pathlen1, &ino1, &freepath1,
2550 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2551 &req->r_req_flags));
2552 if (ret < 0) {
2553 msg = ERR_PTR(ret);
2554 goto out;
2555 }
2556
2557 /* If r_old_dentry is set, then assume that its parent is locked */
2558 ret = set_request_path_attr(NULL, req->r_old_dentry,
2559 req->r_old_dentry_dir,
2560 req->r_path2, req->r_ino2.ino,
2561 &path2, &pathlen2, &ino2, &freepath2, true);
2562 if (ret < 0) {
2563 msg = ERR_PTR(ret);
2564 goto out_free1;
2565 }
2566
2567 len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2568 len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2569 sizeof(struct ceph_timespec);
2570 len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2571
2572 /* calculate (max) length for cap releases */
2573 len += sizeof(struct ceph_mds_request_release) *
2574 (!!req->r_inode_drop + !!req->r_dentry_drop +
2575 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2576
2577 if (req->r_dentry_drop)
2578 len += pathlen1;
2579 if (req->r_old_dentry_drop)
2580 len += pathlen2;
2581
2582 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2583 if (!msg) {
2584 msg = ERR_PTR(-ENOMEM);
2585 goto out_free2;
2586 }
2587
2588 msg->hdr.tid = cpu_to_le64(req->r_tid);
2589
2590 /*
2591 * The old ceph_mds_request_head didn't contain a version field, and
2592 * one was added when we moved the message version from 3->4.
2593 */
2594 if (legacy) {
2595 msg->hdr.version = cpu_to_le16(3);
2596 head = msg->front.iov_base;
2597 p = msg->front.iov_base + sizeof(*head);
2598 } else {
2599 struct ceph_mds_request_head *new_head = msg->front.iov_base;
2600
2601 msg->hdr.version = cpu_to_le16(4);
2602 new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2603 head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2604 p = msg->front.iov_base + sizeof(*new_head);
2605 }
2606
2607 end = msg->front.iov_base + msg->front.iov_len;
2608
2609 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2610 head->op = cpu_to_le32(req->r_op);
2611 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2612 req->r_cred->fsuid));
2613 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2614 req->r_cred->fsgid));
2615 head->ino = cpu_to_le64(req->r_deleg_ino);
2616 head->args = req->r_args;
2617
2618 ceph_encode_filepath(&p, end, ino1, path1);
2619 ceph_encode_filepath(&p, end, ino2, path2);
2620
2621 /* make note of release offset, in case we need to replay */
2622 req->r_request_release_offset = p - msg->front.iov_base;
2623
2624 /* cap releases */
2625 releases = 0;
2626 if (req->r_inode_drop)
2627 releases += ceph_encode_inode_release(&p,
2628 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2629 mds, req->r_inode_drop, req->r_inode_unless,
2630 req->r_op == CEPH_MDS_OP_READDIR);
2631 if (req->r_dentry_drop)
2632 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2633 req->r_parent, mds, req->r_dentry_drop,
2634 req->r_dentry_unless);
2635 if (req->r_old_dentry_drop)
2636 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2637 req->r_old_dentry_dir, mds,
2638 req->r_old_dentry_drop,
2639 req->r_old_dentry_unless);
2640 if (req->r_old_inode_drop)
2641 releases += ceph_encode_inode_release(&p,
2642 d_inode(req->r_old_dentry),
2643 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2644
2645 if (drop_cap_releases) {
2646 releases = 0;
2647 p = msg->front.iov_base + req->r_request_release_offset;
2648 }
2649
2650 head->num_releases = cpu_to_le16(releases);
2651
2652 encode_timestamp_and_gids(&p, req);
2653
2654 if (WARN_ON_ONCE(p > end)) {
2655 ceph_msg_put(msg);
2656 msg = ERR_PTR(-ERANGE);
2657 goto out_free2;
2658 }
2659
2660 msg->front.iov_len = p - msg->front.iov_base;
2661 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2662
2663 if (req->r_pagelist) {
2664 struct ceph_pagelist *pagelist = req->r_pagelist;
2665 ceph_msg_data_add_pagelist(msg, pagelist);
2666 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2667 } else {
2668 msg->hdr.data_len = 0;
2669 }
2670
2671 msg->hdr.data_off = cpu_to_le16(0);
2672
2673out_free2:
2674 if (freepath2)
2675 ceph_mdsc_free_path((char *)path2, pathlen2);
2676out_free1:
2677 if (freepath1)
2678 ceph_mdsc_free_path((char *)path1, pathlen1);
2679out:
2680 return msg;
2681}
2682
2683/*
2684 * called under mdsc->mutex if error, under no mutex if
2685 * success.
2686 */
2687static void complete_request(struct ceph_mds_client *mdsc,
2688 struct ceph_mds_request *req)
2689{
2690 req->r_end_latency = ktime_get();
2691
2692 if (req->r_callback)
2693 req->r_callback(mdsc, req);
2694 complete_all(&req->r_completion);
2695}
2696
2697static struct ceph_mds_request_head_old *
2698find_old_request_head(void *p, u64 features)
2699{
2700 bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2701 struct ceph_mds_request_head *new_head;
2702
2703 if (legacy)
2704 return (struct ceph_mds_request_head_old *)p;
2705 new_head = (struct ceph_mds_request_head *)p;
2706 return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2707}
2708
2709/*
2710 * called under mdsc->mutex
2711 */
2712static int __prepare_send_request(struct ceph_mds_session *session,
2713 struct ceph_mds_request *req,
2714 bool drop_cap_releases)
2715{
2716 int mds = session->s_mds;
2717 struct ceph_mds_client *mdsc = session->s_mdsc;
2718 struct ceph_mds_request_head_old *rhead;
2719 struct ceph_msg *msg;
2720 int flags = 0;
2721
2722 req->r_attempts++;
2723 if (req->r_inode) {
2724 struct ceph_cap *cap =
2725 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2726
2727 if (cap)
2728 req->r_sent_on_mseq = cap->mseq;
2729 else
2730 req->r_sent_on_mseq = -1;
2731 }
2732 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2733 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2734
2735 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2736 void *p;
2737
2738 /*
2739 * Replay. Do not regenerate message (and rebuild
2740 * paths, etc.); just use the original message.
2741 * Rebuilding paths will break for renames because
2742 * d_move mangles the src name.
2743 */
2744 msg = req->r_request;
2745 rhead = find_old_request_head(msg->front.iov_base,
2746 session->s_con.peer_features);
2747
2748 flags = le32_to_cpu(rhead->flags);
2749 flags |= CEPH_MDS_FLAG_REPLAY;
2750 rhead->flags = cpu_to_le32(flags);
2751
2752 if (req->r_target_inode)
2753 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2754
2755 rhead->num_retry = req->r_attempts - 1;
2756
2757 /* remove cap/dentry releases from message */
2758 rhead->num_releases = 0;
2759
2760 p = msg->front.iov_base + req->r_request_release_offset;
2761 encode_timestamp_and_gids(&p, req);
2762
2763 msg->front.iov_len = p - msg->front.iov_base;
2764 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2765 return 0;
2766 }
2767
2768 if (req->r_request) {
2769 ceph_msg_put(req->r_request);
2770 req->r_request = NULL;
2771 }
2772 msg = create_request_message(session, req, drop_cap_releases);
2773 if (IS_ERR(msg)) {
2774 req->r_err = PTR_ERR(msg);
2775 return PTR_ERR(msg);
2776 }
2777 req->r_request = msg;
2778
2779 rhead = find_old_request_head(msg->front.iov_base,
2780 session->s_con.peer_features);
2781 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2782 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2783 flags |= CEPH_MDS_FLAG_REPLAY;
2784 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2785 flags |= CEPH_MDS_FLAG_ASYNC;
2786 if (req->r_parent)
2787 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2788 rhead->flags = cpu_to_le32(flags);
2789 rhead->num_fwd = req->r_num_fwd;
2790 rhead->num_retry = req->r_attempts - 1;
2791
2792 dout(" r_parent = %p\n", req->r_parent);
2793 return 0;
2794}
2795
2796/*
2797 * called under mdsc->mutex
2798 */
2799static int __send_request(struct ceph_mds_session *session,
2800 struct ceph_mds_request *req,
2801 bool drop_cap_releases)
2802{
2803 int err;
2804
2805 err = __prepare_send_request(session, req, drop_cap_releases);
2806 if (!err) {
2807 ceph_msg_get(req->r_request);
2808 ceph_con_send(&session->s_con, req->r_request);
2809 }
2810
2811 return err;
2812}
2813
2814/*
2815 * send request, or put it on the appropriate wait list.
2816 */
2817static void __do_request(struct ceph_mds_client *mdsc,
2818 struct ceph_mds_request *req)
2819{
2820 struct ceph_mds_session *session = NULL;
2821 int mds = -1;
2822 int err = 0;
2823 bool random;
2824
2825 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2826 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2827 __unregister_request(mdsc, req);
2828 return;
2829 }
2830
2831 if (req->r_timeout &&
2832 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2833 dout("do_request timed out\n");
2834 err = -ETIMEDOUT;
2835 goto finish;
2836 }
2837 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2838 dout("do_request forced umount\n");
2839 err = -EIO;
2840 goto finish;
2841 }
2842 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2843 if (mdsc->mdsmap_err) {
2844 err = mdsc->mdsmap_err;
2845 dout("do_request mdsmap err %d\n", err);
2846 goto finish;
2847 }
2848 if (mdsc->mdsmap->m_epoch == 0) {
2849 dout("do_request no mdsmap, waiting for map\n");
2850 list_add(&req->r_wait, &mdsc->waiting_for_map);
2851 return;
2852 }
2853 if (!(mdsc->fsc->mount_options->flags &
2854 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2855 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2856 err = -EHOSTUNREACH;
2857 goto finish;
2858 }
2859 }
2860
2861 put_request_session(req);
2862
2863 mds = __choose_mds(mdsc, req, &random);
2864 if (mds < 0 ||
2865 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2866 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2867 err = -EJUKEBOX;
2868 goto finish;
2869 }
2870 dout("do_request no mds or not active, waiting for map\n");
2871 list_add(&req->r_wait, &mdsc->waiting_for_map);
2872 return;
2873 }
2874
2875 /* get, open session */
2876 session = __ceph_lookup_mds_session(mdsc, mds);
2877 if (!session) {
2878 session = register_session(mdsc, mds);
2879 if (IS_ERR(session)) {
2880 err = PTR_ERR(session);
2881 goto finish;
2882 }
2883 }
2884 req->r_session = ceph_get_mds_session(session);
2885
2886 dout("do_request mds%d session %p state %s\n", mds, session,
2887 ceph_session_state_name(session->s_state));
2888 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2889 session->s_state != CEPH_MDS_SESSION_HUNG) {
2890 /*
2891 * We cannot queue async requests since the caps and delegated
2892 * inodes are bound to the session. Just return -EJUKEBOX and
2893 * let the caller retry a sync request in that case.
2894 */
2895 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2896 err = -EJUKEBOX;
2897 goto out_session;
2898 }
2899
2900 /*
2901 * If the session has been REJECTED, then return a hard error,
2902 * unless it's a CLEANRECOVER mount, in which case we'll queue
2903 * it to the mdsc queue.
2904 */
2905 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2906 if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2907 list_add(&req->r_wait, &mdsc->waiting_for_map);
2908 else
2909 err = -EACCES;
2910 goto out_session;
2911 }
2912
2913 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2914 session->s_state == CEPH_MDS_SESSION_CLOSING) {
2915 err = __open_session(mdsc, session);
2916 if (err)
2917 goto out_session;
2918 /* retry the same mds later */
2919 if (random)
2920 req->r_resend_mds = mds;
2921 }
2922 list_add(&req->r_wait, &session->s_waiting);
2923 goto out_session;
2924 }
2925
2926 /* send request */
2927 req->r_resend_mds = -1; /* forget any previous mds hint */
2928
2929 if (req->r_request_started == 0) /* note request start time */
2930 req->r_request_started = jiffies;
2931
2932 err = __send_request(session, req, false);
2933
2934out_session:
2935 ceph_put_mds_session(session);
2936finish:
2937 if (err) {
2938 dout("__do_request early error %d\n", err);
2939 req->r_err = err;
2940 complete_request(mdsc, req);
2941 __unregister_request(mdsc, req);
2942 }
2943 return;
2944}
2945
2946/*
2947 * called under mdsc->mutex
2948 */
2949static void __wake_requests(struct ceph_mds_client *mdsc,
2950 struct list_head *head)
2951{
2952 struct ceph_mds_request *req;
2953 LIST_HEAD(tmp_list);
2954
2955 list_splice_init(head, &tmp_list);
2956
2957 while (!list_empty(&tmp_list)) {
2958 req = list_entry(tmp_list.next,
2959 struct ceph_mds_request, r_wait);
2960 list_del_init(&req->r_wait);
2961 dout(" wake request %p tid %llu\n", req, req->r_tid);
2962 __do_request(mdsc, req);
2963 }
2964}
2965
2966/*
2967 * Wake up threads with requests pending for @mds, so that they can
2968 * resubmit their requests to a possibly different mds.
2969 */
2970static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2971{
2972 struct ceph_mds_request *req;
2973 struct rb_node *p = rb_first(&mdsc->request_tree);
2974
2975 dout("kick_requests mds%d\n", mds);
2976 while (p) {
2977 req = rb_entry(p, struct ceph_mds_request, r_node);
2978 p = rb_next(p);
2979 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2980 continue;
2981 if (req->r_attempts > 0)
2982 continue; /* only new requests */
2983 if (req->r_session &&
2984 req->r_session->s_mds == mds) {
2985 dout(" kicking tid %llu\n", req->r_tid);
2986 list_del_init(&req->r_wait);
2987 __do_request(mdsc, req);
2988 }
2989 }
2990}
2991
2992int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2993 struct ceph_mds_request *req)
2994{
2995 int err = 0;
2996
2997 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2998 if (req->r_inode)
2999 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3000 if (req->r_parent) {
3001 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3002 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3003 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3004 spin_lock(&ci->i_ceph_lock);
3005 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3006 __ceph_touch_fmode(ci, mdsc, fmode);
3007 spin_unlock(&ci->i_ceph_lock);
3008 }
3009 if (req->r_old_dentry_dir)
3010 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3011 CEPH_CAP_PIN);
3012
3013 if (req->r_inode) {
3014 err = ceph_wait_on_async_create(req->r_inode);
3015 if (err) {
3016 dout("%s: wait for async create returned: %d\n",
3017 __func__, err);
3018 return err;
3019 }
3020 }
3021
3022 if (!err && req->r_old_inode) {
3023 err = ceph_wait_on_async_create(req->r_old_inode);
3024 if (err) {
3025 dout("%s: wait for async create returned: %d\n",
3026 __func__, err);
3027 return err;
3028 }
3029 }
3030
3031 dout("submit_request on %p for inode %p\n", req, dir);
3032 mutex_lock(&mdsc->mutex);
3033 __register_request(mdsc, req, dir);
3034 __do_request(mdsc, req);
3035 err = req->r_err;
3036 mutex_unlock(&mdsc->mutex);
3037 return err;
3038}
3039
3040static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3041 struct ceph_mds_request *req)
3042{
3043 int err;
3044
3045 /* wait */
3046 dout("do_request waiting\n");
3047 if (!req->r_timeout && req->r_wait_for_completion) {
3048 err = req->r_wait_for_completion(mdsc, req);
3049 } else {
3050 long timeleft = wait_for_completion_killable_timeout(
3051 &req->r_completion,
3052 ceph_timeout_jiffies(req->r_timeout));
3053 if (timeleft > 0)
3054 err = 0;
3055 else if (!timeleft)
3056 err = -ETIMEDOUT; /* timed out */
3057 else
3058 err = timeleft; /* killed */
3059 }
3060 dout("do_request waited, got %d\n", err);
3061 mutex_lock(&mdsc->mutex);
3062
3063 /* only abort if we didn't race with a real reply */
3064 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3065 err = le32_to_cpu(req->r_reply_info.head->result);
3066 } else if (err < 0) {
3067 dout("aborted request %lld with %d\n", req->r_tid, err);
3068
3069 /*
3070 * ensure we aren't running concurrently with
3071 * ceph_fill_trace or ceph_readdir_prepopulate, which
3072 * rely on locks (dir mutex) held by our caller.
3073 */
3074 mutex_lock(&req->r_fill_mutex);
3075 req->r_err = err;
3076 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3077 mutex_unlock(&req->r_fill_mutex);
3078
3079 if (req->r_parent &&
3080 (req->r_op & CEPH_MDS_OP_WRITE))
3081 ceph_invalidate_dir_request(req);
3082 } else {
3083 err = req->r_err;
3084 }
3085
3086 mutex_unlock(&mdsc->mutex);
3087 return err;
3088}
3089
3090/*
3091 * Synchrously perform an mds request. Take care of all of the
3092 * session setup, forwarding, retry details.
3093 */
3094int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3095 struct inode *dir,
3096 struct ceph_mds_request *req)
3097{
3098 int err;
3099
3100 dout("do_request on %p\n", req);
3101
3102 /* issue */
3103 err = ceph_mdsc_submit_request(mdsc, dir, req);
3104 if (!err)
3105 err = ceph_mdsc_wait_request(mdsc, req);
3106 dout("do_request %p done, result %d\n", req, err);
3107 return err;
3108}
3109
3110/*
3111 * Invalidate dir's completeness, dentry lease state on an aborted MDS
3112 * namespace request.
3113 */
3114void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3115{
3116 struct inode *dir = req->r_parent;
3117 struct inode *old_dir = req->r_old_dentry_dir;
3118
3119 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3120
3121 ceph_dir_clear_complete(dir);
3122 if (old_dir)
3123 ceph_dir_clear_complete(old_dir);
3124 if (req->r_dentry)
3125 ceph_invalidate_dentry_lease(req->r_dentry);
3126 if (req->r_old_dentry)
3127 ceph_invalidate_dentry_lease(req->r_old_dentry);
3128}
3129
3130/*
3131 * Handle mds reply.
3132 *
3133 * We take the session mutex and parse and process the reply immediately.
3134 * This preserves the logical ordering of replies, capabilities, etc., sent
3135 * by the MDS as they are applied to our local cache.
3136 */
3137static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3138{
3139 struct ceph_mds_client *mdsc = session->s_mdsc;
3140 struct ceph_mds_request *req;
3141 struct ceph_mds_reply_head *head = msg->front.iov_base;
3142 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
3143 struct ceph_snap_realm *realm;
3144 u64 tid;
3145 int err, result;
3146 int mds = session->s_mds;
3147
3148 if (msg->front.iov_len < sizeof(*head)) {
3149 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3150 ceph_msg_dump(msg);
3151 return;
3152 }
3153
3154 /* get request, session */
3155 tid = le64_to_cpu(msg->hdr.tid);
3156 mutex_lock(&mdsc->mutex);
3157 req = lookup_get_request(mdsc, tid);
3158 if (!req) {
3159 dout("handle_reply on unknown tid %llu\n", tid);
3160 mutex_unlock(&mdsc->mutex);
3161 return;
3162 }
3163 dout("handle_reply %p\n", req);
3164
3165 /* correct session? */
3166 if (req->r_session != session) {
3167 pr_err("mdsc_handle_reply got %llu on session mds%d"
3168 " not mds%d\n", tid, session->s_mds,
3169 req->r_session ? req->r_session->s_mds : -1);
3170 mutex_unlock(&mdsc->mutex);
3171 goto out;
3172 }
3173
3174 /* dup? */
3175 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3176 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3177 pr_warn("got a dup %s reply on %llu from mds%d\n",
3178 head->safe ? "safe" : "unsafe", tid, mds);
3179 mutex_unlock(&mdsc->mutex);
3180 goto out;
3181 }
3182 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3183 pr_warn("got unsafe after safe on %llu from mds%d\n",
3184 tid, mds);
3185 mutex_unlock(&mdsc->mutex);
3186 goto out;
3187 }
3188
3189 result = le32_to_cpu(head->result);
3190
3191 /*
3192 * Handle an ESTALE
3193 * if we're not talking to the authority, send to them
3194 * if the authority has changed while we weren't looking,
3195 * send to new authority
3196 * Otherwise we just have to return an ESTALE
3197 */
3198 if (result == -ESTALE) {
3199 dout("got ESTALE on request %llu\n", req->r_tid);
3200 req->r_resend_mds = -1;
3201 if (req->r_direct_mode != USE_AUTH_MDS) {
3202 dout("not using auth, setting for that now\n");
3203 req->r_direct_mode = USE_AUTH_MDS;
3204 __do_request(mdsc, req);
3205 mutex_unlock(&mdsc->mutex);
3206 goto out;
3207 } else {
3208 int mds = __choose_mds(mdsc, req, NULL);
3209 if (mds >= 0 && mds != req->r_session->s_mds) {
3210 dout("but auth changed, so resending\n");
3211 __do_request(mdsc, req);
3212 mutex_unlock(&mdsc->mutex);
3213 goto out;
3214 }
3215 }
3216 dout("have to return ESTALE on request %llu\n", req->r_tid);
3217 }
3218
3219
3220 if (head->safe) {
3221 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3222 __unregister_request(mdsc, req);
3223
3224 /* last request during umount? */
3225 if (mdsc->stopping && !__get_oldest_req(mdsc))
3226 complete_all(&mdsc->safe_umount_waiters);
3227
3228 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3229 /*
3230 * We already handled the unsafe response, now do the
3231 * cleanup. No need to examine the response; the MDS
3232 * doesn't include any result info in the safe
3233 * response. And even if it did, there is nothing
3234 * useful we could do with a revised return value.
3235 */
3236 dout("got safe reply %llu, mds%d\n", tid, mds);
3237
3238 mutex_unlock(&mdsc->mutex);
3239 goto out;
3240 }
3241 } else {
3242 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3243 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3244 }
3245
3246 dout("handle_reply tid %lld result %d\n", tid, result);
3247 rinfo = &req->r_reply_info;
3248 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3249 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3250 else
3251 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3252 mutex_unlock(&mdsc->mutex);
3253
3254 /* Must find target inode outside of mutexes to avoid deadlocks */
3255 if ((err >= 0) && rinfo->head->is_target) {
3256 struct inode *in;
3257 struct ceph_vino tvino = {
3258 .ino = le64_to_cpu(rinfo->targeti.in->ino),
3259 .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3260 };
3261
3262 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3263 if (IS_ERR(in)) {
3264 err = PTR_ERR(in);
3265 mutex_lock(&session->s_mutex);
3266 goto out_err;
3267 }
3268 req->r_target_inode = in;
3269 }
3270
3271 mutex_lock(&session->s_mutex);
3272 if (err < 0) {
3273 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3274 ceph_msg_dump(msg);
3275 goto out_err;
3276 }
3277
3278 /* snap trace */
3279 realm = NULL;
3280 if (rinfo->snapblob_len) {
3281 down_write(&mdsc->snap_rwsem);
3282 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3283 rinfo->snapblob + rinfo->snapblob_len,
3284 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3285 &realm);
3286 downgrade_write(&mdsc->snap_rwsem);
3287 } else {
3288 down_read(&mdsc->snap_rwsem);
3289 }
3290
3291 /* insert trace into our cache */
3292 mutex_lock(&req->r_fill_mutex);
3293 current->journal_info = req;
3294 err = ceph_fill_trace(mdsc->fsc->sb, req);
3295 if (err == 0) {
3296 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3297 req->r_op == CEPH_MDS_OP_LSSNAP))
3298 ceph_readdir_prepopulate(req, req->r_session);
3299 }
3300 current->journal_info = NULL;
3301 mutex_unlock(&req->r_fill_mutex);
3302
3303 up_read(&mdsc->snap_rwsem);
3304 if (realm)
3305 ceph_put_snap_realm(mdsc, realm);
3306
3307 if (err == 0) {
3308 if (req->r_target_inode &&
3309 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3310 struct ceph_inode_info *ci =
3311 ceph_inode(req->r_target_inode);
3312 spin_lock(&ci->i_unsafe_lock);
3313 list_add_tail(&req->r_unsafe_target_item,
3314 &ci->i_unsafe_iops);
3315 spin_unlock(&ci->i_unsafe_lock);
3316 }
3317
3318 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3319 }
3320out_err:
3321 mutex_lock(&mdsc->mutex);
3322 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3323 if (err) {
3324 req->r_err = err;
3325 } else {
3326 req->r_reply = ceph_msg_get(msg);
3327 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3328 }
3329 } else {
3330 dout("reply arrived after request %lld was aborted\n", tid);
3331 }
3332 mutex_unlock(&mdsc->mutex);
3333
3334 mutex_unlock(&session->s_mutex);
3335
3336 /* kick calling process */
3337 complete_request(mdsc, req);
3338
3339 ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3340 req->r_end_latency, err);
3341out:
3342 ceph_mdsc_put_request(req);
3343 return;
3344}
3345
3346
3347
3348/*
3349 * handle mds notification that our request has been forwarded.
3350 */
3351static void handle_forward(struct ceph_mds_client *mdsc,
3352 struct ceph_mds_session *session,
3353 struct ceph_msg *msg)
3354{
3355 struct ceph_mds_request *req;
3356 u64 tid = le64_to_cpu(msg->hdr.tid);
3357 u32 next_mds;
3358 u32 fwd_seq;
3359 int err = -EINVAL;
3360 void *p = msg->front.iov_base;
3361 void *end = p + msg->front.iov_len;
3362
3363 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3364 next_mds = ceph_decode_32(&p);
3365 fwd_seq = ceph_decode_32(&p);
3366
3367 mutex_lock(&mdsc->mutex);
3368 req = lookup_get_request(mdsc, tid);
3369 if (!req) {
3370 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3371 goto out; /* dup reply? */
3372 }
3373
3374 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3375 dout("forward tid %llu aborted, unregistering\n", tid);
3376 __unregister_request(mdsc, req);
3377 } else if (fwd_seq <= req->r_num_fwd) {
3378 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3379 tid, next_mds, req->r_num_fwd, fwd_seq);
3380 } else {
3381 /* resend. forward race not possible; mds would drop */
3382 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3383 BUG_ON(req->r_err);
3384 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3385 req->r_attempts = 0;
3386 req->r_num_fwd = fwd_seq;
3387 req->r_resend_mds = next_mds;
3388 put_request_session(req);
3389 __do_request(mdsc, req);
3390 }
3391 ceph_mdsc_put_request(req);
3392out:
3393 mutex_unlock(&mdsc->mutex);
3394 return;
3395
3396bad:
3397 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3398}
3399
3400static int __decode_session_metadata(void **p, void *end,
3401 bool *blocklisted)
3402{
3403 /* map<string,string> */
3404 u32 n;
3405 bool err_str;
3406 ceph_decode_32_safe(p, end, n, bad);
3407 while (n-- > 0) {
3408 u32 len;
3409 ceph_decode_32_safe(p, end, len, bad);
3410 ceph_decode_need(p, end, len, bad);
3411 err_str = !strncmp(*p, "error_string", len);
3412 *p += len;
3413 ceph_decode_32_safe(p, end, len, bad);
3414 ceph_decode_need(p, end, len, bad);
3415 /*
3416 * Match "blocklisted (blacklisted)" from newer MDSes,
3417 * or "blacklisted" from older MDSes.
3418 */
3419 if (err_str && strnstr(*p, "blacklisted", len))
3420 *blocklisted = true;
3421 *p += len;
3422 }
3423 return 0;
3424bad:
3425 return -1;
3426}
3427
3428/*
3429 * handle a mds session control message
3430 */
3431static void handle_session(struct ceph_mds_session *session,
3432 struct ceph_msg *msg)
3433{
3434 struct ceph_mds_client *mdsc = session->s_mdsc;
3435 int mds = session->s_mds;
3436 int msg_version = le16_to_cpu(msg->hdr.version);
3437 void *p = msg->front.iov_base;
3438 void *end = p + msg->front.iov_len;
3439 struct ceph_mds_session_head *h;
3440 u32 op;
3441 u64 seq, features = 0;
3442 int wake = 0;
3443 bool blocklisted = false;
3444
3445 /* decode */
3446 ceph_decode_need(&p, end, sizeof(*h), bad);
3447 h = p;
3448 p += sizeof(*h);
3449
3450 op = le32_to_cpu(h->op);
3451 seq = le64_to_cpu(h->seq);
3452
3453 if (msg_version >= 3) {
3454 u32 len;
3455 /* version >= 2, metadata */
3456 if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3457 goto bad;
3458 /* version >= 3, feature bits */
3459 ceph_decode_32_safe(&p, end, len, bad);
3460 if (len) {
3461 ceph_decode_64_safe(&p, end, features, bad);
3462 p += len - sizeof(features);
3463 }
3464 }
3465
3466 mutex_lock(&mdsc->mutex);
3467 if (op == CEPH_SESSION_CLOSE) {
3468 ceph_get_mds_session(session);
3469 __unregister_session(mdsc, session);
3470 }
3471 /* FIXME: this ttl calculation is generous */
3472 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3473 mutex_unlock(&mdsc->mutex);
3474
3475 mutex_lock(&session->s_mutex);
3476
3477 dout("handle_session mds%d %s %p state %s seq %llu\n",
3478 mds, ceph_session_op_name(op), session,
3479 ceph_session_state_name(session->s_state), seq);
3480
3481 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3482 session->s_state = CEPH_MDS_SESSION_OPEN;
3483 pr_info("mds%d came back\n", session->s_mds);
3484 }
3485
3486 switch (op) {
3487 case CEPH_SESSION_OPEN:
3488 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3489 pr_info("mds%d reconnect success\n", session->s_mds);
3490 session->s_state = CEPH_MDS_SESSION_OPEN;
3491 session->s_features = features;
3492 renewed_caps(mdsc, session, 0);
3493 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3494 metric_schedule_delayed(&mdsc->metric);
3495 wake = 1;
3496 if (mdsc->stopping)
3497 __close_session(mdsc, session);
3498 break;
3499
3500 case CEPH_SESSION_RENEWCAPS:
3501 if (session->s_renew_seq == seq)
3502 renewed_caps(mdsc, session, 1);
3503 break;
3504
3505 case CEPH_SESSION_CLOSE:
3506 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3507 pr_info("mds%d reconnect denied\n", session->s_mds);
3508 session->s_state = CEPH_MDS_SESSION_CLOSED;
3509 cleanup_session_requests(mdsc, session);
3510 remove_session_caps(session);
3511 wake = 2; /* for good measure */
3512 wake_up_all(&mdsc->session_close_wq);
3513 break;
3514
3515 case CEPH_SESSION_STALE:
3516 pr_info("mds%d caps went stale, renewing\n",
3517 session->s_mds);
3518 atomic_inc(&session->s_cap_gen);
3519 session->s_cap_ttl = jiffies - 1;
3520 send_renew_caps(mdsc, session);
3521 break;
3522
3523 case CEPH_SESSION_RECALL_STATE:
3524 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3525 break;
3526
3527 case CEPH_SESSION_FLUSHMSG:
3528 send_flushmsg_ack(mdsc, session, seq);
3529 break;
3530
3531 case CEPH_SESSION_FORCE_RO:
3532 dout("force_session_readonly %p\n", session);
3533 spin_lock(&session->s_cap_lock);
3534 session->s_readonly = true;
3535 spin_unlock(&session->s_cap_lock);
3536 wake_up_session_caps(session, FORCE_RO);
3537 break;
3538
3539 case CEPH_SESSION_REJECT:
3540 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3541 pr_info("mds%d rejected session\n", session->s_mds);
3542 session->s_state = CEPH_MDS_SESSION_REJECTED;
3543 cleanup_session_requests(mdsc, session);
3544 remove_session_caps(session);
3545 if (blocklisted)
3546 mdsc->fsc->blocklisted = true;
3547 wake = 2; /* for good measure */
3548 break;
3549
3550 default:
3551 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3552 WARN_ON(1);
3553 }
3554
3555 mutex_unlock(&session->s_mutex);
3556 if (wake) {
3557 mutex_lock(&mdsc->mutex);
3558 __wake_requests(mdsc, &session->s_waiting);
3559 if (wake == 2)
3560 kick_requests(mdsc, mds);
3561 mutex_unlock(&mdsc->mutex);
3562 }
3563 if (op == CEPH_SESSION_CLOSE)
3564 ceph_put_mds_session(session);
3565 return;
3566
3567bad:
3568 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3569 (int)msg->front.iov_len);
3570 ceph_msg_dump(msg);
3571 return;
3572}
3573
3574void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3575{
3576 int dcaps;
3577
3578 dcaps = xchg(&req->r_dir_caps, 0);
3579 if (dcaps) {
3580 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3581 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3582 }
3583}
3584
3585void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3586{
3587 int dcaps;
3588
3589 dcaps = xchg(&req->r_dir_caps, 0);
3590 if (dcaps) {
3591 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3592 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3593 dcaps);
3594 }
3595}
3596
3597/*
3598 * called under session->mutex.
3599 */
3600static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3601 struct ceph_mds_session *session)
3602{
3603 struct ceph_mds_request *req, *nreq;
3604 struct rb_node *p;
3605
3606 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3607
3608 mutex_lock(&mdsc->mutex);
3609 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3610 __send_request(session, req, true);
3611
3612 /*
3613 * also re-send old requests when MDS enters reconnect stage. So that MDS
3614 * can process completed request in clientreplay stage.
3615 */
3616 p = rb_first(&mdsc->request_tree);
3617 while (p) {
3618 req = rb_entry(p, struct ceph_mds_request, r_node);
3619 p = rb_next(p);
3620 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3621 continue;
3622 if (req->r_attempts == 0)
3623 continue; /* only old requests */
3624 if (!req->r_session)
3625 continue;
3626 if (req->r_session->s_mds != session->s_mds)
3627 continue;
3628
3629 ceph_mdsc_release_dir_caps_no_check(req);
3630
3631 __send_request(session, req, true);
3632 }
3633 mutex_unlock(&mdsc->mutex);
3634}
3635
3636static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3637{
3638 struct ceph_msg *reply;
3639 struct ceph_pagelist *_pagelist;
3640 struct page *page;
3641 __le32 *addr;
3642 int err = -ENOMEM;
3643
3644 if (!recon_state->allow_multi)
3645 return -ENOSPC;
3646
3647 /* can't handle message that contains both caps and realm */
3648 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3649
3650 /* pre-allocate new pagelist */
3651 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3652 if (!_pagelist)
3653 return -ENOMEM;
3654
3655 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3656 if (!reply)
3657 goto fail_msg;
3658
3659 /* placeholder for nr_caps */
3660 err = ceph_pagelist_encode_32(_pagelist, 0);
3661 if (err < 0)
3662 goto fail;
3663
3664 if (recon_state->nr_caps) {
3665 /* currently encoding caps */
3666 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3667 if (err)
3668 goto fail;
3669 } else {
3670 /* placeholder for nr_realms (currently encoding relams) */
3671 err = ceph_pagelist_encode_32(_pagelist, 0);
3672 if (err < 0)
3673 goto fail;
3674 }
3675
3676 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3677 if (err)
3678 goto fail;
3679
3680 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3681 addr = kmap_atomic(page);
3682 if (recon_state->nr_caps) {
3683 /* currently encoding caps */
3684 *addr = cpu_to_le32(recon_state->nr_caps);
3685 } else {
3686 /* currently encoding relams */
3687 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3688 }
3689 kunmap_atomic(addr);
3690
3691 reply->hdr.version = cpu_to_le16(5);
3692 reply->hdr.compat_version = cpu_to_le16(4);
3693
3694 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3695 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3696
3697 ceph_con_send(&recon_state->session->s_con, reply);
3698 ceph_pagelist_release(recon_state->pagelist);
3699
3700 recon_state->pagelist = _pagelist;
3701 recon_state->nr_caps = 0;
3702 recon_state->nr_realms = 0;
3703 recon_state->msg_version = 5;
3704 return 0;
3705fail:
3706 ceph_msg_put(reply);
3707fail_msg:
3708 ceph_pagelist_release(_pagelist);
3709 return err;
3710}
3711
3712static struct dentry* d_find_primary(struct inode *inode)
3713{
3714 struct dentry *alias, *dn = NULL;
3715
3716 if (hlist_empty(&inode->i_dentry))
3717 return NULL;
3718
3719 spin_lock(&inode->i_lock);
3720 if (hlist_empty(&inode->i_dentry))
3721 goto out_unlock;
3722
3723 if (S_ISDIR(inode->i_mode)) {
3724 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3725 if (!IS_ROOT(alias))
3726 dn = dget(alias);
3727 goto out_unlock;
3728 }
3729
3730 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3731 spin_lock(&alias->d_lock);
3732 if (!d_unhashed(alias) &&
3733 (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3734 dn = dget_dlock(alias);
3735 }
3736 spin_unlock(&alias->d_lock);
3737 if (dn)
3738 break;
3739 }
3740out_unlock:
3741 spin_unlock(&inode->i_lock);
3742 return dn;
3743}
3744
3745/*
3746 * Encode information about a cap for a reconnect with the MDS.
3747 */
3748static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3749 void *arg)
3750{
3751 union {
3752 struct ceph_mds_cap_reconnect v2;
3753 struct ceph_mds_cap_reconnect_v1 v1;
3754 } rec;
3755 struct ceph_inode_info *ci = cap->ci;
3756 struct ceph_reconnect_state *recon_state = arg;
3757 struct ceph_pagelist *pagelist = recon_state->pagelist;
3758 struct dentry *dentry;
3759 char *path;
3760 int pathlen, err;
3761 u64 pathbase;
3762 u64 snap_follows;
3763
3764 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3765 inode, ceph_vinop(inode), cap, cap->cap_id,
3766 ceph_cap_string(cap->issued));
3767
3768 dentry = d_find_primary(inode);
3769 if (dentry) {
3770 /* set pathbase to parent dir when msg_version >= 2 */
3771 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3772 recon_state->msg_version >= 2);
3773 dput(dentry);
3774 if (IS_ERR(path)) {
3775 err = PTR_ERR(path);
3776 goto out_err;
3777 }
3778 } else {
3779 path = NULL;
3780 pathlen = 0;
3781 pathbase = 0;
3782 }
3783
3784 spin_lock(&ci->i_ceph_lock);
3785 cap->seq = 0; /* reset cap seq */
3786 cap->issue_seq = 0; /* and issue_seq */
3787 cap->mseq = 0; /* and migrate_seq */
3788 cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
3789
3790 /* These are lost when the session goes away */
3791 if (S_ISDIR(inode->i_mode)) {
3792 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3793 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3794 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3795 }
3796 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3797 }
3798
3799 if (recon_state->msg_version >= 2) {
3800 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3801 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3802 rec.v2.issued = cpu_to_le32(cap->issued);
3803 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3804 rec.v2.pathbase = cpu_to_le64(pathbase);
3805 rec.v2.flock_len = (__force __le32)
3806 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3807 } else {
3808 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3809 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3810 rec.v1.issued = cpu_to_le32(cap->issued);
3811 rec.v1.size = cpu_to_le64(i_size_read(inode));
3812 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3813 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3814 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3815 rec.v1.pathbase = cpu_to_le64(pathbase);
3816 }
3817
3818 if (list_empty(&ci->i_cap_snaps)) {
3819 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3820 } else {
3821 struct ceph_cap_snap *capsnap =
3822 list_first_entry(&ci->i_cap_snaps,
3823 struct ceph_cap_snap, ci_item);
3824 snap_follows = capsnap->follows;
3825 }
3826 spin_unlock(&ci->i_ceph_lock);
3827
3828 if (recon_state->msg_version >= 2) {
3829 int num_fcntl_locks, num_flock_locks;
3830 struct ceph_filelock *flocks = NULL;
3831 size_t struct_len, total_len = sizeof(u64);
3832 u8 struct_v = 0;
3833
3834encode_again:
3835 if (rec.v2.flock_len) {
3836 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3837 } else {
3838 num_fcntl_locks = 0;
3839 num_flock_locks = 0;
3840 }
3841 if (num_fcntl_locks + num_flock_locks > 0) {
3842 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3843 sizeof(struct ceph_filelock),
3844 GFP_NOFS);
3845 if (!flocks) {
3846 err = -ENOMEM;
3847 goto out_err;
3848 }
3849 err = ceph_encode_locks_to_buffer(inode, flocks,
3850 num_fcntl_locks,
3851 num_flock_locks);
3852 if (err) {
3853 kfree(flocks);
3854 flocks = NULL;
3855 if (err == -ENOSPC)
3856 goto encode_again;
3857 goto out_err;
3858 }
3859 } else {
3860 kfree(flocks);
3861 flocks = NULL;
3862 }
3863
3864 if (recon_state->msg_version >= 3) {
3865 /* version, compat_version and struct_len */
3866 total_len += 2 * sizeof(u8) + sizeof(u32);
3867 struct_v = 2;
3868 }
3869 /*
3870 * number of encoded locks is stable, so copy to pagelist
3871 */
3872 struct_len = 2 * sizeof(u32) +
3873 (num_fcntl_locks + num_flock_locks) *
3874 sizeof(struct ceph_filelock);
3875 rec.v2.flock_len = cpu_to_le32(struct_len);
3876
3877 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3878
3879 if (struct_v >= 2)
3880 struct_len += sizeof(u64); /* snap_follows */
3881
3882 total_len += struct_len;
3883
3884 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3885 err = send_reconnect_partial(recon_state);
3886 if (err)
3887 goto out_freeflocks;
3888 pagelist = recon_state->pagelist;
3889 }
3890
3891 err = ceph_pagelist_reserve(pagelist, total_len);
3892 if (err)
3893 goto out_freeflocks;
3894
3895 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3896 if (recon_state->msg_version >= 3) {
3897 ceph_pagelist_encode_8(pagelist, struct_v);
3898 ceph_pagelist_encode_8(pagelist, 1);
3899 ceph_pagelist_encode_32(pagelist, struct_len);
3900 }
3901 ceph_pagelist_encode_string(pagelist, path, pathlen);
3902 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3903 ceph_locks_to_pagelist(flocks, pagelist,
3904 num_fcntl_locks, num_flock_locks);
3905 if (struct_v >= 2)
3906 ceph_pagelist_encode_64(pagelist, snap_follows);
3907out_freeflocks:
3908 kfree(flocks);
3909 } else {
3910 err = ceph_pagelist_reserve(pagelist,
3911 sizeof(u64) + sizeof(u32) +
3912 pathlen + sizeof(rec.v1));
3913 if (err)
3914 goto out_err;
3915
3916 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3917 ceph_pagelist_encode_string(pagelist, path, pathlen);
3918 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3919 }
3920
3921out_err:
3922 ceph_mdsc_free_path(path, pathlen);
3923 if (!err)
3924 recon_state->nr_caps++;
3925 return err;
3926}
3927
3928static int encode_snap_realms(struct ceph_mds_client *mdsc,
3929 struct ceph_reconnect_state *recon_state)
3930{
3931 struct rb_node *p;
3932 struct ceph_pagelist *pagelist = recon_state->pagelist;
3933 int err = 0;
3934
3935 if (recon_state->msg_version >= 4) {
3936 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3937 if (err < 0)
3938 goto fail;
3939 }
3940
3941 /*
3942 * snaprealms. we provide mds with the ino, seq (version), and
3943 * parent for all of our realms. If the mds has any newer info,
3944 * it will tell us.
3945 */
3946 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3947 struct ceph_snap_realm *realm =
3948 rb_entry(p, struct ceph_snap_realm, node);
3949 struct ceph_mds_snaprealm_reconnect sr_rec;
3950
3951 if (recon_state->msg_version >= 4) {
3952 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3953 sizeof(sr_rec);
3954
3955 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3956 err = send_reconnect_partial(recon_state);
3957 if (err)
3958 goto fail;
3959 pagelist = recon_state->pagelist;
3960 }
3961
3962 err = ceph_pagelist_reserve(pagelist, need);
3963 if (err)
3964 goto fail;
3965
3966 ceph_pagelist_encode_8(pagelist, 1);
3967 ceph_pagelist_encode_8(pagelist, 1);
3968 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3969 }
3970
3971 dout(" adding snap realm %llx seq %lld parent %llx\n",
3972 realm->ino, realm->seq, realm->parent_ino);
3973 sr_rec.ino = cpu_to_le64(realm->ino);
3974 sr_rec.seq = cpu_to_le64(realm->seq);
3975 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3976
3977 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3978 if (err)
3979 goto fail;
3980
3981 recon_state->nr_realms++;
3982 }
3983fail:
3984 return err;
3985}
3986
3987
3988/*
3989 * If an MDS fails and recovers, clients need to reconnect in order to
3990 * reestablish shared state. This includes all caps issued through
3991 * this session _and_ the snap_realm hierarchy. Because it's not
3992 * clear which snap realms the mds cares about, we send everything we
3993 * know about.. that ensures we'll then get any new info the
3994 * recovering MDS might have.
3995 *
3996 * This is a relatively heavyweight operation, but it's rare.
3997 */
3998static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3999 struct ceph_mds_session *session)
4000{
4001 struct ceph_msg *reply;
4002 int mds = session->s_mds;
4003 int err = -ENOMEM;
4004 struct ceph_reconnect_state recon_state = {
4005 .session = session,
4006 };
4007 LIST_HEAD(dispose);
4008
4009 pr_info("mds%d reconnect start\n", mds);
4010
4011 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4012 if (!recon_state.pagelist)
4013 goto fail_nopagelist;
4014
4015 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4016 if (!reply)
4017 goto fail_nomsg;
4018
4019 xa_destroy(&session->s_delegated_inos);
4020
4021 mutex_lock(&session->s_mutex);
4022 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4023 session->s_seq = 0;
4024
4025 dout("session %p state %s\n", session,
4026 ceph_session_state_name(session->s_state));
4027
4028 atomic_inc(&session->s_cap_gen);
4029
4030 spin_lock(&session->s_cap_lock);
4031 /* don't know if session is readonly */
4032 session->s_readonly = 0;
4033 /*
4034 * notify __ceph_remove_cap() that we are composing cap reconnect.
4035 * If a cap get released before being added to the cap reconnect,
4036 * __ceph_remove_cap() should skip queuing cap release.
4037 */
4038 session->s_cap_reconnect = 1;
4039 /* drop old cap expires; we're about to reestablish that state */
4040 detach_cap_releases(session, &dispose);
4041 spin_unlock(&session->s_cap_lock);
4042 dispose_cap_releases(mdsc, &dispose);
4043
4044 /* trim unused caps to reduce MDS's cache rejoin time */
4045 if (mdsc->fsc->sb->s_root)
4046 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4047
4048 ceph_con_close(&session->s_con);
4049 ceph_con_open(&session->s_con,
4050 CEPH_ENTITY_TYPE_MDS, mds,
4051 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4052
4053 /* replay unsafe requests */
4054 replay_unsafe_requests(mdsc, session);
4055
4056 ceph_early_kick_flushing_caps(mdsc, session);
4057
4058 down_read(&mdsc->snap_rwsem);
4059
4060 /* placeholder for nr_caps */
4061 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4062 if (err)
4063 goto fail;
4064
4065 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4066 recon_state.msg_version = 3;
4067 recon_state.allow_multi = true;
4068 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4069 recon_state.msg_version = 3;
4070 } else {
4071 recon_state.msg_version = 2;
4072 }
4073 /* trsaverse this session's caps */
4074 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4075
4076 spin_lock(&session->s_cap_lock);
4077 session->s_cap_reconnect = 0;
4078 spin_unlock(&session->s_cap_lock);
4079
4080 if (err < 0)
4081 goto fail;
4082
4083 /* check if all realms can be encoded into current message */
4084 if (mdsc->num_snap_realms) {
4085 size_t total_len =
4086 recon_state.pagelist->length +
4087 mdsc->num_snap_realms *
4088 sizeof(struct ceph_mds_snaprealm_reconnect);
4089 if (recon_state.msg_version >= 4) {
4090 /* number of realms */
4091 total_len += sizeof(u32);
4092 /* version, compat_version and struct_len */
4093 total_len += mdsc->num_snap_realms *
4094 (2 * sizeof(u8) + sizeof(u32));
4095 }
4096 if (total_len > RECONNECT_MAX_SIZE) {
4097 if (!recon_state.allow_multi) {
4098 err = -ENOSPC;
4099 goto fail;
4100 }
4101 if (recon_state.nr_caps) {
4102 err = send_reconnect_partial(&recon_state);
4103 if (err)
4104 goto fail;
4105 }
4106 recon_state.msg_version = 5;
4107 }
4108 }
4109
4110 err = encode_snap_realms(mdsc, &recon_state);
4111 if (err < 0)
4112 goto fail;
4113
4114 if (recon_state.msg_version >= 5) {
4115 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4116 if (err < 0)
4117 goto fail;
4118 }
4119
4120 if (recon_state.nr_caps || recon_state.nr_realms) {
4121 struct page *page =
4122 list_first_entry(&recon_state.pagelist->head,
4123 struct page, lru);
4124 __le32 *addr = kmap_atomic(page);
4125 if (recon_state.nr_caps) {
4126 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4127 *addr = cpu_to_le32(recon_state.nr_caps);
4128 } else if (recon_state.msg_version >= 4) {
4129 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4130 }
4131 kunmap_atomic(addr);
4132 }
4133
4134 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4135 if (recon_state.msg_version >= 4)
4136 reply->hdr.compat_version = cpu_to_le16(4);
4137
4138 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4139 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4140
4141 ceph_con_send(&session->s_con, reply);
4142
4143 mutex_unlock(&session->s_mutex);
4144
4145 mutex_lock(&mdsc->mutex);
4146 __wake_requests(mdsc, &session->s_waiting);
4147 mutex_unlock(&mdsc->mutex);
4148
4149 up_read(&mdsc->snap_rwsem);
4150 ceph_pagelist_release(recon_state.pagelist);
4151 return;
4152
4153fail:
4154 ceph_msg_put(reply);
4155 up_read(&mdsc->snap_rwsem);
4156 mutex_unlock(&session->s_mutex);
4157fail_nomsg:
4158 ceph_pagelist_release(recon_state.pagelist);
4159fail_nopagelist:
4160 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4161 return;
4162}
4163
4164
4165/*
4166 * compare old and new mdsmaps, kicking requests
4167 * and closing out old connections as necessary
4168 *
4169 * called under mdsc->mutex.
4170 */
4171static void check_new_map(struct ceph_mds_client *mdsc,
4172 struct ceph_mdsmap *newmap,
4173 struct ceph_mdsmap *oldmap)
4174{
4175 int i;
4176 int oldstate, newstate;
4177 struct ceph_mds_session *s;
4178
4179 dout("check_new_map new %u old %u\n",
4180 newmap->m_epoch, oldmap->m_epoch);
4181
4182 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4183 if (!mdsc->sessions[i])
4184 continue;
4185 s = mdsc->sessions[i];
4186 oldstate = ceph_mdsmap_get_state(oldmap, i);
4187 newstate = ceph_mdsmap_get_state(newmap, i);
4188
4189 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4190 i, ceph_mds_state_name(oldstate),
4191 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4192 ceph_mds_state_name(newstate),
4193 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4194 ceph_session_state_name(s->s_state));
4195
4196 if (i >= newmap->possible_max_rank) {
4197 /* force close session for stopped mds */
4198 ceph_get_mds_session(s);
4199 __unregister_session(mdsc, s);
4200 __wake_requests(mdsc, &s->s_waiting);
4201 mutex_unlock(&mdsc->mutex);
4202
4203 mutex_lock(&s->s_mutex);
4204 cleanup_session_requests(mdsc, s);
4205 remove_session_caps(s);
4206 mutex_unlock(&s->s_mutex);
4207
4208 ceph_put_mds_session(s);
4209
4210 mutex_lock(&mdsc->mutex);
4211 kick_requests(mdsc, i);
4212 continue;
4213 }
4214
4215 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4216 ceph_mdsmap_get_addr(newmap, i),
4217 sizeof(struct ceph_entity_addr))) {
4218 /* just close it */
4219 mutex_unlock(&mdsc->mutex);
4220 mutex_lock(&s->s_mutex);
4221 mutex_lock(&mdsc->mutex);
4222 ceph_con_close(&s->s_con);
4223 mutex_unlock(&s->s_mutex);
4224 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4225 } else if (oldstate == newstate) {
4226 continue; /* nothing new with this mds */
4227 }
4228
4229 /*
4230 * send reconnect?
4231 */
4232 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4233 newstate >= CEPH_MDS_STATE_RECONNECT) {
4234 mutex_unlock(&mdsc->mutex);
4235 send_mds_reconnect(mdsc, s);
4236 mutex_lock(&mdsc->mutex);
4237 }
4238
4239 /*
4240 * kick request on any mds that has gone active.
4241 */
4242 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4243 newstate >= CEPH_MDS_STATE_ACTIVE) {
4244 if (oldstate != CEPH_MDS_STATE_CREATING &&
4245 oldstate != CEPH_MDS_STATE_STARTING)
4246 pr_info("mds%d recovery completed\n", s->s_mds);
4247 kick_requests(mdsc, i);
4248 mutex_unlock(&mdsc->mutex);
4249 mutex_lock(&s->s_mutex);
4250 mutex_lock(&mdsc->mutex);
4251 ceph_kick_flushing_caps(mdsc, s);
4252 mutex_unlock(&s->s_mutex);
4253 wake_up_session_caps(s, RECONNECT);
4254 }
4255 }
4256
4257 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4258 s = mdsc->sessions[i];
4259 if (!s)
4260 continue;
4261 if (!ceph_mdsmap_is_laggy(newmap, i))
4262 continue;
4263 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4264 s->s_state == CEPH_MDS_SESSION_HUNG ||
4265 s->s_state == CEPH_MDS_SESSION_CLOSING) {
4266 dout(" connecting to export targets of laggy mds%d\n",
4267 i);
4268 __open_export_target_sessions(mdsc, s);
4269 }
4270 }
4271}
4272
4273
4274
4275/*
4276 * leases
4277 */
4278
4279/*
4280 * caller must hold session s_mutex, dentry->d_lock
4281 */
4282void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4283{
4284 struct ceph_dentry_info *di = ceph_dentry(dentry);
4285
4286 ceph_put_mds_session(di->lease_session);
4287 di->lease_session = NULL;
4288}
4289
4290static void handle_lease(struct ceph_mds_client *mdsc,
4291 struct ceph_mds_session *session,
4292 struct ceph_msg *msg)
4293{
4294 struct super_block *sb = mdsc->fsc->sb;
4295 struct inode *inode;
4296 struct dentry *parent, *dentry;
4297 struct ceph_dentry_info *di;
4298 int mds = session->s_mds;
4299 struct ceph_mds_lease *h = msg->front.iov_base;
4300 u32 seq;
4301 struct ceph_vino vino;
4302 struct qstr dname;
4303 int release = 0;
4304
4305 dout("handle_lease from mds%d\n", mds);
4306
4307 /* decode */
4308 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4309 goto bad;
4310 vino.ino = le64_to_cpu(h->ino);
4311 vino.snap = CEPH_NOSNAP;
4312 seq = le32_to_cpu(h->seq);
4313 dname.len = get_unaligned_le32(h + 1);
4314 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4315 goto bad;
4316 dname.name = (void *)(h + 1) + sizeof(u32);
4317
4318 /* lookup inode */
4319 inode = ceph_find_inode(sb, vino);
4320 dout("handle_lease %s, ino %llx %p %.*s\n",
4321 ceph_lease_op_name(h->action), vino.ino, inode,
4322 dname.len, dname.name);
4323
4324 mutex_lock(&session->s_mutex);
4325 inc_session_sequence(session);
4326
4327 if (!inode) {
4328 dout("handle_lease no inode %llx\n", vino.ino);
4329 goto release;
4330 }
4331
4332 /* dentry */
4333 parent = d_find_alias(inode);
4334 if (!parent) {
4335 dout("no parent dentry on inode %p\n", inode);
4336 WARN_ON(1);
4337 goto release; /* hrm... */
4338 }
4339 dname.hash = full_name_hash(parent, dname.name, dname.len);
4340 dentry = d_lookup(parent, &dname);
4341 dput(parent);
4342 if (!dentry)
4343 goto release;
4344
4345 spin_lock(&dentry->d_lock);
4346 di = ceph_dentry(dentry);
4347 switch (h->action) {
4348 case CEPH_MDS_LEASE_REVOKE:
4349 if (di->lease_session == session) {
4350 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4351 h->seq = cpu_to_le32(di->lease_seq);
4352 __ceph_mdsc_drop_dentry_lease(dentry);
4353 }
4354 release = 1;
4355 break;
4356
4357 case CEPH_MDS_LEASE_RENEW:
4358 if (di->lease_session == session &&
4359 di->lease_gen == atomic_read(&session->s_cap_gen) &&
4360 di->lease_renew_from &&
4361 di->lease_renew_after == 0) {
4362 unsigned long duration =
4363 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4364
4365 di->lease_seq = seq;
4366 di->time = di->lease_renew_from + duration;
4367 di->lease_renew_after = di->lease_renew_from +
4368 (duration >> 1);
4369 di->lease_renew_from = 0;
4370 }
4371 break;
4372 }
4373 spin_unlock(&dentry->d_lock);
4374 dput(dentry);
4375
4376 if (!release)
4377 goto out;
4378
4379release:
4380 /* let's just reuse the same message */
4381 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4382 ceph_msg_get(msg);
4383 ceph_con_send(&session->s_con, msg);
4384
4385out:
4386 mutex_unlock(&session->s_mutex);
4387 iput(inode);
4388 return;
4389
4390bad:
4391 pr_err("corrupt lease message\n");
4392 ceph_msg_dump(msg);
4393}
4394
4395void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4396 struct dentry *dentry, char action,
4397 u32 seq)
4398{
4399 struct ceph_msg *msg;
4400 struct ceph_mds_lease *lease;
4401 struct inode *dir;
4402 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4403
4404 dout("lease_send_msg identry %p %s to mds%d\n",
4405 dentry, ceph_lease_op_name(action), session->s_mds);
4406
4407 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4408 if (!msg)
4409 return;
4410 lease = msg->front.iov_base;
4411 lease->action = action;
4412 lease->seq = cpu_to_le32(seq);
4413
4414 spin_lock(&dentry->d_lock);
4415 dir = d_inode(dentry->d_parent);
4416 lease->ino = cpu_to_le64(ceph_ino(dir));
4417 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4418
4419 put_unaligned_le32(dentry->d_name.len, lease + 1);
4420 memcpy((void *)(lease + 1) + 4,
4421 dentry->d_name.name, dentry->d_name.len);
4422 spin_unlock(&dentry->d_lock);
4423 /*
4424 * if this is a preemptive lease RELEASE, no need to
4425 * flush request stream, since the actual request will
4426 * soon follow.
4427 */
4428 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4429
4430 ceph_con_send(&session->s_con, msg);
4431}
4432
4433/*
4434 * lock unlock sessions, to wait ongoing session activities
4435 */
4436static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4437{
4438 int i;
4439
4440 mutex_lock(&mdsc->mutex);
4441 for (i = 0; i < mdsc->max_sessions; i++) {
4442 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4443 if (!s)
4444 continue;
4445 mutex_unlock(&mdsc->mutex);
4446 mutex_lock(&s->s_mutex);
4447 mutex_unlock(&s->s_mutex);
4448 ceph_put_mds_session(s);
4449 mutex_lock(&mdsc->mutex);
4450 }
4451 mutex_unlock(&mdsc->mutex);
4452}
4453
4454static void maybe_recover_session(struct ceph_mds_client *mdsc)
4455{
4456 struct ceph_fs_client *fsc = mdsc->fsc;
4457
4458 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4459 return;
4460
4461 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4462 return;
4463
4464 if (!READ_ONCE(fsc->blocklisted))
4465 return;
4466
4467 pr_info("auto reconnect after blocklisted\n");
4468 ceph_force_reconnect(fsc->sb);
4469}
4470
4471bool check_session_state(struct ceph_mds_session *s)
4472{
4473 switch (s->s_state) {
4474 case CEPH_MDS_SESSION_OPEN:
4475 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4476 s->s_state = CEPH_MDS_SESSION_HUNG;
4477 pr_info("mds%d hung\n", s->s_mds);
4478 }
4479 break;
4480 case CEPH_MDS_SESSION_CLOSING:
4481 /* Should never reach this when we're unmounting */
4482 WARN_ON_ONCE(s->s_ttl);
4483 fallthrough;
4484 case CEPH_MDS_SESSION_NEW:
4485 case CEPH_MDS_SESSION_RESTARTING:
4486 case CEPH_MDS_SESSION_CLOSED:
4487 case CEPH_MDS_SESSION_REJECTED:
4488 return false;
4489 }
4490
4491 return true;
4492}
4493
4494/*
4495 * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4496 * then we need to retransmit that request.
4497 */
4498void inc_session_sequence(struct ceph_mds_session *s)
4499{
4500 lockdep_assert_held(&s->s_mutex);
4501
4502 s->s_seq++;
4503
4504 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4505 int ret;
4506
4507 dout("resending session close request for mds%d\n", s->s_mds);
4508 ret = request_close_session(s);
4509 if (ret < 0)
4510 pr_err("unable to close session to mds%d: %d\n",
4511 s->s_mds, ret);
4512 }
4513}
4514
4515/*
4516 * delayed work -- periodically trim expired leases, renew caps with mds. If
4517 * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4518 * workqueue delay value of 5 secs will be used.
4519 */
4520static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4521{
4522 unsigned long max_delay = HZ * 5;
4523
4524 /* 5 secs default delay */
4525 if (!delay || (delay > max_delay))
4526 delay = max_delay;
4527 schedule_delayed_work(&mdsc->delayed_work,
4528 round_jiffies_relative(delay));
4529}
4530
4531static void delayed_work(struct work_struct *work)
4532{
4533 struct ceph_mds_client *mdsc =
4534 container_of(work, struct ceph_mds_client, delayed_work.work);
4535 unsigned long delay;
4536 int renew_interval;
4537 int renew_caps;
4538 int i;
4539
4540 dout("mdsc delayed_work\n");
4541
4542 if (mdsc->stopping)
4543 return;
4544
4545 mutex_lock(&mdsc->mutex);
4546 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4547 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4548 mdsc->last_renew_caps);
4549 if (renew_caps)
4550 mdsc->last_renew_caps = jiffies;
4551
4552 for (i = 0; i < mdsc->max_sessions; i++) {
4553 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4554 if (!s)
4555 continue;
4556
4557 if (!check_session_state(s)) {
4558 ceph_put_mds_session(s);
4559 continue;
4560 }
4561 mutex_unlock(&mdsc->mutex);
4562
4563 mutex_lock(&s->s_mutex);
4564 if (renew_caps)
4565 send_renew_caps(mdsc, s);
4566 else
4567 ceph_con_keepalive(&s->s_con);
4568 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4569 s->s_state == CEPH_MDS_SESSION_HUNG)
4570 ceph_send_cap_releases(mdsc, s);
4571 mutex_unlock(&s->s_mutex);
4572 ceph_put_mds_session(s);
4573
4574 mutex_lock(&mdsc->mutex);
4575 }
4576 mutex_unlock(&mdsc->mutex);
4577
4578 delay = ceph_check_delayed_caps(mdsc);
4579
4580 ceph_queue_cap_reclaim_work(mdsc);
4581
4582 ceph_trim_snapid_map(mdsc);
4583
4584 maybe_recover_session(mdsc);
4585
4586 schedule_delayed(mdsc, delay);
4587}
4588
4589int ceph_mdsc_init(struct ceph_fs_client *fsc)
4590
4591{
4592 struct ceph_mds_client *mdsc;
4593 int err;
4594
4595 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4596 if (!mdsc)
4597 return -ENOMEM;
4598 mdsc->fsc = fsc;
4599 mutex_init(&mdsc->mutex);
4600 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4601 if (!mdsc->mdsmap) {
4602 err = -ENOMEM;
4603 goto err_mdsc;
4604 }
4605
4606 init_completion(&mdsc->safe_umount_waiters);
4607 init_waitqueue_head(&mdsc->session_close_wq);
4608 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4609 mdsc->sessions = NULL;
4610 atomic_set(&mdsc->num_sessions, 0);
4611 mdsc->max_sessions = 0;
4612 mdsc->stopping = 0;
4613 atomic64_set(&mdsc->quotarealms_count, 0);
4614 mdsc->quotarealms_inodes = RB_ROOT;
4615 mutex_init(&mdsc->quotarealms_inodes_mutex);
4616 mdsc->last_snap_seq = 0;
4617 init_rwsem(&mdsc->snap_rwsem);
4618 mdsc->snap_realms = RB_ROOT;
4619 INIT_LIST_HEAD(&mdsc->snap_empty);
4620 mdsc->num_snap_realms = 0;
4621 spin_lock_init(&mdsc->snap_empty_lock);
4622 mdsc->last_tid = 0;
4623 mdsc->oldest_tid = 0;
4624 mdsc->request_tree = RB_ROOT;
4625 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4626 mdsc->last_renew_caps = jiffies;
4627 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4628 INIT_LIST_HEAD(&mdsc->cap_wait_list);
4629 spin_lock_init(&mdsc->cap_delay_lock);
4630 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4631 spin_lock_init(&mdsc->snap_flush_lock);
4632 mdsc->last_cap_flush_tid = 1;
4633 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4634 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4635 mdsc->num_cap_flushing = 0;
4636 spin_lock_init(&mdsc->cap_dirty_lock);
4637 init_waitqueue_head(&mdsc->cap_flushing_wq);
4638 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4639 atomic_set(&mdsc->cap_reclaim_pending, 0);
4640 err = ceph_metric_init(&mdsc->metric);
4641 if (err)
4642 goto err_mdsmap;
4643
4644 spin_lock_init(&mdsc->dentry_list_lock);
4645 INIT_LIST_HEAD(&mdsc->dentry_leases);
4646 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4647
4648 ceph_caps_init(mdsc);
4649 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4650
4651 spin_lock_init(&mdsc->snapid_map_lock);
4652 mdsc->snapid_map_tree = RB_ROOT;
4653 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4654
4655 init_rwsem(&mdsc->pool_perm_rwsem);
4656 mdsc->pool_perm_tree = RB_ROOT;
4657
4658 strscpy(mdsc->nodename, utsname()->nodename,
4659 sizeof(mdsc->nodename));
4660
4661 fsc->mdsc = mdsc;
4662 return 0;
4663
4664err_mdsmap:
4665 kfree(mdsc->mdsmap);
4666err_mdsc:
4667 kfree(mdsc);
4668 return err;
4669}
4670
4671/*
4672 * Wait for safe replies on open mds requests. If we time out, drop
4673 * all requests from the tree to avoid dangling dentry refs.
4674 */
4675static void wait_requests(struct ceph_mds_client *mdsc)
4676{
4677 struct ceph_options *opts = mdsc->fsc->client->options;
4678 struct ceph_mds_request *req;
4679
4680 mutex_lock(&mdsc->mutex);
4681 if (__get_oldest_req(mdsc)) {
4682 mutex_unlock(&mdsc->mutex);
4683
4684 dout("wait_requests waiting for requests\n");
4685 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4686 ceph_timeout_jiffies(opts->mount_timeout));
4687
4688 /* tear down remaining requests */
4689 mutex_lock(&mdsc->mutex);
4690 while ((req = __get_oldest_req(mdsc))) {
4691 dout("wait_requests timed out on tid %llu\n",
4692 req->r_tid);
4693 list_del_init(&req->r_wait);
4694 __unregister_request(mdsc, req);
4695 }
4696 }
4697 mutex_unlock(&mdsc->mutex);
4698 dout("wait_requests done\n");
4699}
4700
4701/*
4702 * called before mount is ro, and before dentries are torn down.
4703 * (hmm, does this still race with new lookups?)
4704 */
4705void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4706{
4707 dout("pre_umount\n");
4708 mdsc->stopping = 1;
4709
4710 lock_unlock_sessions(mdsc);
4711 ceph_flush_dirty_caps(mdsc);
4712 wait_requests(mdsc);
4713
4714 /*
4715 * wait for reply handlers to drop their request refs and
4716 * their inode/dcache refs
4717 */
4718 ceph_msgr_flush();
4719
4720 ceph_cleanup_quotarealms_inodes(mdsc);
4721}
4722
4723/*
4724 * wait for all write mds requests to flush.
4725 */
4726static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4727{
4728 struct ceph_mds_request *req = NULL, *nextreq;
4729 struct rb_node *n;
4730
4731 mutex_lock(&mdsc->mutex);
4732 dout("wait_unsafe_requests want %lld\n", want_tid);
4733restart:
4734 req = __get_oldest_req(mdsc);
4735 while (req && req->r_tid <= want_tid) {
4736 /* find next request */
4737 n = rb_next(&req->r_node);
4738 if (n)
4739 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4740 else
4741 nextreq = NULL;
4742 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4743 (req->r_op & CEPH_MDS_OP_WRITE)) {
4744 /* write op */
4745 ceph_mdsc_get_request(req);
4746 if (nextreq)
4747 ceph_mdsc_get_request(nextreq);
4748 mutex_unlock(&mdsc->mutex);
4749 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
4750 req->r_tid, want_tid);
4751 wait_for_completion(&req->r_safe_completion);
4752 mutex_lock(&mdsc->mutex);
4753 ceph_mdsc_put_request(req);
4754 if (!nextreq)
4755 break; /* next dne before, so we're done! */
4756 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4757 /* next request was removed from tree */
4758 ceph_mdsc_put_request(nextreq);
4759 goto restart;
4760 }
4761 ceph_mdsc_put_request(nextreq); /* won't go away */
4762 }
4763 req = nextreq;
4764 }
4765 mutex_unlock(&mdsc->mutex);
4766 dout("wait_unsafe_requests done\n");
4767}
4768
4769void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4770{
4771 u64 want_tid, want_flush;
4772
4773 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4774 return;
4775
4776 dout("sync\n");
4777 mutex_lock(&mdsc->mutex);
4778 want_tid = mdsc->last_tid;
4779 mutex_unlock(&mdsc->mutex);
4780
4781 ceph_flush_dirty_caps(mdsc);
4782 spin_lock(&mdsc->cap_dirty_lock);
4783 want_flush = mdsc->last_cap_flush_tid;
4784 if (!list_empty(&mdsc->cap_flush_list)) {
4785 struct ceph_cap_flush *cf =
4786 list_last_entry(&mdsc->cap_flush_list,
4787 struct ceph_cap_flush, g_list);
4788 cf->wake = true;
4789 }
4790 spin_unlock(&mdsc->cap_dirty_lock);
4791
4792 dout("sync want tid %lld flush_seq %lld\n",
4793 want_tid, want_flush);
4794
4795 wait_unsafe_requests(mdsc, want_tid);
4796 wait_caps_flush(mdsc, want_flush);
4797}
4798
4799/*
4800 * true if all sessions are closed, or we force unmount
4801 */
4802static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4803{
4804 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4805 return true;
4806 return atomic_read(&mdsc->num_sessions) <= skipped;
4807}
4808
4809/*
4810 * called after sb is ro.
4811 */
4812void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4813{
4814 struct ceph_options *opts = mdsc->fsc->client->options;
4815 struct ceph_mds_session *session;
4816 int i;
4817 int skipped = 0;
4818
4819 dout("close_sessions\n");
4820
4821 /* close sessions */
4822 mutex_lock(&mdsc->mutex);
4823 for (i = 0; i < mdsc->max_sessions; i++) {
4824 session = __ceph_lookup_mds_session(mdsc, i);
4825 if (!session)
4826 continue;
4827 mutex_unlock(&mdsc->mutex);
4828 mutex_lock(&session->s_mutex);
4829 if (__close_session(mdsc, session) <= 0)
4830 skipped++;
4831 mutex_unlock(&session->s_mutex);
4832 ceph_put_mds_session(session);
4833 mutex_lock(&mdsc->mutex);
4834 }
4835 mutex_unlock(&mdsc->mutex);
4836
4837 dout("waiting for sessions to close\n");
4838 wait_event_timeout(mdsc->session_close_wq,
4839 done_closing_sessions(mdsc, skipped),
4840 ceph_timeout_jiffies(opts->mount_timeout));
4841
4842 /* tear down remaining sessions */
4843 mutex_lock(&mdsc->mutex);
4844 for (i = 0; i < mdsc->max_sessions; i++) {
4845 if (mdsc->sessions[i]) {
4846 session = ceph_get_mds_session(mdsc->sessions[i]);
4847 __unregister_session(mdsc, session);
4848 mutex_unlock(&mdsc->mutex);
4849 mutex_lock(&session->s_mutex);
4850 remove_session_caps(session);
4851 mutex_unlock(&session->s_mutex);
4852 ceph_put_mds_session(session);
4853 mutex_lock(&mdsc->mutex);
4854 }
4855 }
4856 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4857 mutex_unlock(&mdsc->mutex);
4858
4859 ceph_cleanup_snapid_map(mdsc);
4860 ceph_cleanup_empty_realms(mdsc);
4861
4862 cancel_work_sync(&mdsc->cap_reclaim_work);
4863 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4864
4865 dout("stopped\n");
4866}
4867
4868void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4869{
4870 struct ceph_mds_session *session;
4871 int mds;
4872
4873 dout("force umount\n");
4874
4875 mutex_lock(&mdsc->mutex);
4876 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4877 session = __ceph_lookup_mds_session(mdsc, mds);
4878 if (!session)
4879 continue;
4880
4881 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4882 __unregister_session(mdsc, session);
4883 __wake_requests(mdsc, &session->s_waiting);
4884 mutex_unlock(&mdsc->mutex);
4885
4886 mutex_lock(&session->s_mutex);
4887 __close_session(mdsc, session);
4888 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4889 cleanup_session_requests(mdsc, session);
4890 remove_session_caps(session);
4891 }
4892 mutex_unlock(&session->s_mutex);
4893 ceph_put_mds_session(session);
4894
4895 mutex_lock(&mdsc->mutex);
4896 kick_requests(mdsc, mds);
4897 }
4898 __wake_requests(mdsc, &mdsc->waiting_for_map);
4899 mutex_unlock(&mdsc->mutex);
4900}
4901
4902static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4903{
4904 dout("stop\n");
4905 /*
4906 * Make sure the delayed work stopped before releasing
4907 * the resources.
4908 *
4909 * Because the cancel_delayed_work_sync() will only
4910 * guarantee that the work finishes executing. But the
4911 * delayed work will re-arm itself again after that.
4912 */
4913 flush_delayed_work(&mdsc->delayed_work);
4914
4915 if (mdsc->mdsmap)
4916 ceph_mdsmap_destroy(mdsc->mdsmap);
4917 kfree(mdsc->sessions);
4918 ceph_caps_finalize(mdsc);
4919 ceph_pool_perm_destroy(mdsc);
4920}
4921
4922void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4923{
4924 struct ceph_mds_client *mdsc = fsc->mdsc;
4925 dout("mdsc_destroy %p\n", mdsc);
4926
4927 if (!mdsc)
4928 return;
4929
4930 /* flush out any connection work with references to us */
4931 ceph_msgr_flush();
4932
4933 ceph_mdsc_stop(mdsc);
4934
4935 ceph_metric_destroy(&mdsc->metric);
4936
4937 fsc->mdsc = NULL;
4938 kfree(mdsc);
4939 dout("mdsc_destroy %p done\n", mdsc);
4940}
4941
4942void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4943{
4944 struct ceph_fs_client *fsc = mdsc->fsc;
4945 const char *mds_namespace = fsc->mount_options->mds_namespace;
4946 void *p = msg->front.iov_base;
4947 void *end = p + msg->front.iov_len;
4948 u32 epoch;
4949 u32 num_fs;
4950 u32 mount_fscid = (u32)-1;
4951 int err = -EINVAL;
4952
4953 ceph_decode_need(&p, end, sizeof(u32), bad);
4954 epoch = ceph_decode_32(&p);
4955
4956 dout("handle_fsmap epoch %u\n", epoch);
4957
4958 /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
4959 ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
4960
4961 ceph_decode_32_safe(&p, end, num_fs, bad);
4962 while (num_fs-- > 0) {
4963 void *info_p, *info_end;
4964 u32 info_len;
4965 u32 fscid, namelen;
4966
4967 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4968 p += 2; // info_v, info_cv
4969 info_len = ceph_decode_32(&p);
4970 ceph_decode_need(&p, end, info_len, bad);
4971 info_p = p;
4972 info_end = p + info_len;
4973 p = info_end;
4974
4975 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4976 fscid = ceph_decode_32(&info_p);
4977 namelen = ceph_decode_32(&info_p);
4978 ceph_decode_need(&info_p, info_end, namelen, bad);
4979
4980 if (mds_namespace &&
4981 strlen(mds_namespace) == namelen &&
4982 !strncmp(mds_namespace, (char *)info_p, namelen)) {
4983 mount_fscid = fscid;
4984 break;
4985 }
4986 }
4987
4988 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4989 if (mount_fscid != (u32)-1) {
4990 fsc->client->monc.fs_cluster_id = mount_fscid;
4991 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4992 0, true);
4993 ceph_monc_renew_subs(&fsc->client->monc);
4994 } else {
4995 err = -ENOENT;
4996 goto err_out;
4997 }
4998 return;
4999
5000bad:
5001 pr_err("error decoding fsmap\n");
5002err_out:
5003 mutex_lock(&mdsc->mutex);
5004 mdsc->mdsmap_err = err;
5005 __wake_requests(mdsc, &mdsc->waiting_for_map);
5006 mutex_unlock(&mdsc->mutex);
5007}
5008
5009/*
5010 * handle mds map update.
5011 */
5012void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5013{
5014 u32 epoch;
5015 u32 maplen;
5016 void *p = msg->front.iov_base;
5017 void *end = p + msg->front.iov_len;
5018 struct ceph_mdsmap *newmap, *oldmap;
5019 struct ceph_fsid fsid;
5020 int err = -EINVAL;
5021
5022 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5023 ceph_decode_copy(&p, &fsid, sizeof(fsid));
5024 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5025 return;
5026 epoch = ceph_decode_32(&p);
5027 maplen = ceph_decode_32(&p);
5028 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5029
5030 /* do we need it? */
5031 mutex_lock(&mdsc->mutex);
5032 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5033 dout("handle_map epoch %u <= our %u\n",
5034 epoch, mdsc->mdsmap->m_epoch);
5035 mutex_unlock(&mdsc->mutex);
5036 return;
5037 }
5038
5039 newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5040 if (IS_ERR(newmap)) {
5041 err = PTR_ERR(newmap);
5042 goto bad_unlock;
5043 }
5044
5045 /* swap into place */
5046 if (mdsc->mdsmap) {
5047 oldmap = mdsc->mdsmap;
5048 mdsc->mdsmap = newmap;
5049 check_new_map(mdsc, newmap, oldmap);
5050 ceph_mdsmap_destroy(oldmap);
5051 } else {
5052 mdsc->mdsmap = newmap; /* first mds map */
5053 }
5054 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5055 MAX_LFS_FILESIZE);
5056
5057 __wake_requests(mdsc, &mdsc->waiting_for_map);
5058 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5059 mdsc->mdsmap->m_epoch);
5060
5061 mutex_unlock(&mdsc->mutex);
5062 schedule_delayed(mdsc, 0);
5063 return;
5064
5065bad_unlock:
5066 mutex_unlock(&mdsc->mutex);
5067bad:
5068 pr_err("error decoding mdsmap %d\n", err);
5069 return;
5070}
5071
5072static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5073{
5074 struct ceph_mds_session *s = con->private;
5075
5076 if (ceph_get_mds_session(s))
5077 return con;
5078 return NULL;
5079}
5080
5081static void mds_put_con(struct ceph_connection *con)
5082{
5083 struct ceph_mds_session *s = con->private;
5084
5085 ceph_put_mds_session(s);
5086}
5087
5088/*
5089 * if the client is unresponsive for long enough, the mds will kill
5090 * the session entirely.
5091 */
5092static void mds_peer_reset(struct ceph_connection *con)
5093{
5094 struct ceph_mds_session *s = con->private;
5095 struct ceph_mds_client *mdsc = s->s_mdsc;
5096
5097 pr_warn("mds%d closed our session\n", s->s_mds);
5098 send_mds_reconnect(mdsc, s);
5099}
5100
5101static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5102{
5103 struct ceph_mds_session *s = con->private;
5104 struct ceph_mds_client *mdsc = s->s_mdsc;
5105 int type = le16_to_cpu(msg->hdr.type);
5106
5107 mutex_lock(&mdsc->mutex);
5108 if (__verify_registered_session(mdsc, s) < 0) {
5109 mutex_unlock(&mdsc->mutex);
5110 goto out;
5111 }
5112 mutex_unlock(&mdsc->mutex);
5113
5114 switch (type) {
5115 case CEPH_MSG_MDS_MAP:
5116 ceph_mdsc_handle_mdsmap(mdsc, msg);
5117 break;
5118 case CEPH_MSG_FS_MAP_USER:
5119 ceph_mdsc_handle_fsmap(mdsc, msg);
5120 break;
5121 case CEPH_MSG_CLIENT_SESSION:
5122 handle_session(s, msg);
5123 break;
5124 case CEPH_MSG_CLIENT_REPLY:
5125 handle_reply(s, msg);
5126 break;
5127 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5128 handle_forward(mdsc, s, msg);
5129 break;
5130 case CEPH_MSG_CLIENT_CAPS:
5131 ceph_handle_caps(s, msg);
5132 break;
5133 case CEPH_MSG_CLIENT_SNAP:
5134 ceph_handle_snap(mdsc, s, msg);
5135 break;
5136 case CEPH_MSG_CLIENT_LEASE:
5137 handle_lease(mdsc, s, msg);
5138 break;
5139 case CEPH_MSG_CLIENT_QUOTA:
5140 ceph_handle_quota(mdsc, s, msg);
5141 break;
5142
5143 default:
5144 pr_err("received unknown message type %d %s\n", type,
5145 ceph_msg_type_name(type));
5146 }
5147out:
5148 ceph_msg_put(msg);
5149}
5150
5151/*
5152 * authentication
5153 */
5154
5155/*
5156 * Note: returned pointer is the address of a structure that's
5157 * managed separately. Caller must *not* attempt to free it.
5158 */
5159static struct ceph_auth_handshake *
5160mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5161{
5162 struct ceph_mds_session *s = con->private;
5163 struct ceph_mds_client *mdsc = s->s_mdsc;
5164 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5165 struct ceph_auth_handshake *auth = &s->s_auth;
5166 int ret;
5167
5168 ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5169 force_new, proto, NULL, NULL);
5170 if (ret)
5171 return ERR_PTR(ret);
5172
5173 return auth;
5174}
5175
5176static int mds_add_authorizer_challenge(struct ceph_connection *con,
5177 void *challenge_buf, int challenge_buf_len)
5178{
5179 struct ceph_mds_session *s = con->private;
5180 struct ceph_mds_client *mdsc = s->s_mdsc;
5181 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5182
5183 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5184 challenge_buf, challenge_buf_len);
5185}
5186
5187static int mds_verify_authorizer_reply(struct ceph_connection *con)
5188{
5189 struct ceph_mds_session *s = con->private;
5190 struct ceph_mds_client *mdsc = s->s_mdsc;
5191 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5192 struct ceph_auth_handshake *auth = &s->s_auth;
5193
5194 return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5195 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5196 NULL, NULL, NULL, NULL);
5197}
5198
5199static int mds_invalidate_authorizer(struct ceph_connection *con)
5200{
5201 struct ceph_mds_session *s = con->private;
5202 struct ceph_mds_client *mdsc = s->s_mdsc;
5203 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5204
5205 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5206
5207 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5208}
5209
5210static int mds_get_auth_request(struct ceph_connection *con,
5211 void *buf, int *buf_len,
5212 void **authorizer, int *authorizer_len)
5213{
5214 struct ceph_mds_session *s = con->private;
5215 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5216 struct ceph_auth_handshake *auth = &s->s_auth;
5217 int ret;
5218
5219 ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5220 buf, buf_len);
5221 if (ret)
5222 return ret;
5223
5224 *authorizer = auth->authorizer_buf;
5225 *authorizer_len = auth->authorizer_buf_len;
5226 return 0;
5227}
5228
5229static int mds_handle_auth_reply_more(struct ceph_connection *con,
5230 void *reply, int reply_len,
5231 void *buf, int *buf_len,
5232 void **authorizer, int *authorizer_len)
5233{
5234 struct ceph_mds_session *s = con->private;
5235 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5236 struct ceph_auth_handshake *auth = &s->s_auth;
5237 int ret;
5238
5239 ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5240 buf, buf_len);
5241 if (ret)
5242 return ret;
5243
5244 *authorizer = auth->authorizer_buf;
5245 *authorizer_len = auth->authorizer_buf_len;
5246 return 0;
5247}
5248
5249static int mds_handle_auth_done(struct ceph_connection *con,
5250 u64 global_id, void *reply, int reply_len,
5251 u8 *session_key, int *session_key_len,
5252 u8 *con_secret, int *con_secret_len)
5253{
5254 struct ceph_mds_session *s = con->private;
5255 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5256 struct ceph_auth_handshake *auth = &s->s_auth;
5257
5258 return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5259 session_key, session_key_len,
5260 con_secret, con_secret_len);
5261}
5262
5263static int mds_handle_auth_bad_method(struct ceph_connection *con,
5264 int used_proto, int result,
5265 const int *allowed_protos, int proto_cnt,
5266 const int *allowed_modes, int mode_cnt)
5267{
5268 struct ceph_mds_session *s = con->private;
5269 struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5270 int ret;
5271
5272 if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5273 used_proto, result,
5274 allowed_protos, proto_cnt,
5275 allowed_modes, mode_cnt)) {
5276 ret = ceph_monc_validate_auth(monc);
5277 if (ret)
5278 return ret;
5279 }
5280
5281 return -EACCES;
5282}
5283
5284static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5285 struct ceph_msg_header *hdr, int *skip)
5286{
5287 struct ceph_msg *msg;
5288 int type = (int) le16_to_cpu(hdr->type);
5289 int front_len = (int) le32_to_cpu(hdr->front_len);
5290
5291 if (con->in_msg)
5292 return con->in_msg;
5293
5294 *skip = 0;
5295 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5296 if (!msg) {
5297 pr_err("unable to allocate msg type %d len %d\n",
5298 type, front_len);
5299 return NULL;
5300 }
5301
5302 return msg;
5303}
5304
5305static int mds_sign_message(struct ceph_msg *msg)
5306{
5307 struct ceph_mds_session *s = msg->con->private;
5308 struct ceph_auth_handshake *auth = &s->s_auth;
5309
5310 return ceph_auth_sign_message(auth, msg);
5311}
5312
5313static int mds_check_message_signature(struct ceph_msg *msg)
5314{
5315 struct ceph_mds_session *s = msg->con->private;
5316 struct ceph_auth_handshake *auth = &s->s_auth;
5317
5318 return ceph_auth_check_message_signature(auth, msg);
5319}
5320
5321static const struct ceph_connection_operations mds_con_ops = {
5322 .get = mds_get_con,
5323 .put = mds_put_con,
5324 .alloc_msg = mds_alloc_msg,
5325 .dispatch = mds_dispatch,
5326 .peer_reset = mds_peer_reset,
5327 .get_authorizer = mds_get_authorizer,
5328 .add_authorizer_challenge = mds_add_authorizer_challenge,
5329 .verify_authorizer_reply = mds_verify_authorizer_reply,
5330 .invalidate_authorizer = mds_invalidate_authorizer,
5331 .sign_message = mds_sign_message,
5332 .check_message_signature = mds_check_message_signature,
5333 .get_auth_request = mds_get_auth_request,
5334 .handle_auth_reply_more = mds_handle_auth_reply_more,
5335 .handle_auth_done = mds_handle_auth_done,
5336 .handle_auth_bad_method = mds_handle_auth_bad_method,
5337};
5338
5339/* eof */