?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 3.12.28
[linux/fpc-iii.git] / fs / ceph / mds_client.c
blob788901552eb1c46b34eca985237df654481c7b56
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/sched.h>
7 #include <linux/debugfs.h>
8 #include <linux/seq_file.h>
9
10 #include "super.h"
11 #include "mds_client.h"
12
13 #include <linux/ceph/ceph_features.h>
14 #include <linux/ceph/messenger.h>
15 #include <linux/ceph/decode.h>
16 #include <linux/ceph/pagelist.h>
17 #include <linux/ceph/auth.h>
18 #include <linux/ceph/debugfs.h>
19
20 /*
21 * A cluster of MDS (metadata server) daemons is responsible for
22 * managing the file system namespace (the directory hierarchy and
23 * inodes) and for coordinating shared access to storage. Metadata is
24 * partitioning hierarchically across a number of servers, and that
25 * partition varies over time as the cluster adjusts the distribution
26 * in order to balance load.
27 *
28 * The MDS client is primarily responsible to managing synchronous
29 * metadata requests for operations like open, unlink, and so forth.
30 * If there is a MDS failure, we find out about it when we (possibly
31 * request and) receive a new MDS map, and can resubmit affected
32 * requests.
33 *
34 * For the most part, though, we take advantage of a lossless
35 * communications channel to the MDS, and do not need to worry about
36 * timing out or resubmitting requests.
37 *
38 * We maintain a stateful "session" with each MDS we interact with.
39 * Within each session, we sent periodic heartbeat messages to ensure
40 * any capabilities or leases we have been issues remain valid. If
41 * the session times out and goes stale, our leases and capabilities
42 * are no longer valid.
43 */
44
45 struct ceph_reconnect_state {
46 struct ceph_pagelist *pagelist;
47 bool flock;
48 };
49
50 static void __wake_requests(struct ceph_mds_client *mdsc,
51 struct list_head *head);
52
53 static const struct ceph_connection_operations mds_con_ops;
54
55
56 /*
57 * mds reply parsing
58 */
59
60 /*
61 * parse individual inode info
62 */
63 static int parse_reply_info_in(void **p, void *end,
64 struct ceph_mds_reply_info_in *info,
65 int features)
66 {
67 int err = -EIO;
68
69 info->in = *p;
70 *p += sizeof(struct ceph_mds_reply_inode) +
71 sizeof(*info->in->fragtree.splits) *
72 le32_to_cpu(info->in->fragtree.nsplits);
73
74 ceph_decode_32_safe(p, end, info->symlink_len, bad);
75 ceph_decode_need(p, end, info->symlink_len, bad);
76 info->symlink = *p;
77 *p += info->symlink_len;
78
79 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
80 ceph_decode_copy_safe(p, end, &info->dir_layout,
81 sizeof(info->dir_layout), bad);
82 else
83 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
84
85 ceph_decode_32_safe(p, end, info->xattr_len, bad);
86 ceph_decode_need(p, end, info->xattr_len, bad);
87 info->xattr_data = *p;
88 *p += info->xattr_len;
89 return 0;
90 bad:
91 return err;
92 }
93
94 /*
95 * parse a normal reply, which may contain a (dir+)dentry and/or a
96 * target inode.
97 */
98 static int parse_reply_info_trace(void **p, void *end,
99 struct ceph_mds_reply_info_parsed *info,
100 int features)
101 {
102 int err;
103
104 if (info->head->is_dentry) {
105 err = parse_reply_info_in(p, end, &info->diri, features);
106 if (err < 0)
107 goto out_bad;
108
109 if (unlikely(*p + sizeof(*info->dirfrag) > end))
110 goto bad;
111 info->dirfrag = *p;
112 *p += sizeof(*info->dirfrag) +
113 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
114 if (unlikely(*p > end))
115 goto bad;
116
117 ceph_decode_32_safe(p, end, info->dname_len, bad);
118 ceph_decode_need(p, end, info->dname_len, bad);
119 info->dname = *p;
120 *p += info->dname_len;
121 info->dlease = *p;
122 *p += sizeof(*info->dlease);
123 }
124
125 if (info->head->is_target) {
126 err = parse_reply_info_in(p, end, &info->targeti, features);
127 if (err < 0)
128 goto out_bad;
129 }
130
131 if (unlikely(*p != end))
132 goto bad;
133 return 0;
134
135 bad:
136 err = -EIO;
137 out_bad:
138 pr_err("problem parsing mds trace %d\n", err);
139 return err;
140 }
141
142 /*
143 * parse readdir results
144 */
145 static int parse_reply_info_dir(void **p, void *end,
146 struct ceph_mds_reply_info_parsed *info,
147 int features)
148 {
149 u32 num, i = 0;
150 int err;
151
152 info->dir_dir = *p;
153 if (*p + sizeof(*info->dir_dir) > end)
154 goto bad;
155 *p += sizeof(*info->dir_dir) +
156 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
157 if (*p > end)
158 goto bad;
159
160 ceph_decode_need(p, end, sizeof(num) + 2, bad);
161 num = ceph_decode_32(p);
162 info->dir_end = ceph_decode_8(p);
163 info->dir_complete = ceph_decode_8(p);
164 if (num == 0)
165 goto done;
166
167 /* alloc large array */
168 info->dir_nr = num;
169 info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
170 sizeof(*info->dir_dname) +
171 sizeof(*info->dir_dname_len) +
172 sizeof(*info->dir_dlease),
173 GFP_NOFS);
174 if (info->dir_in == NULL) {
175 err = -ENOMEM;
176 goto out_bad;
177 }
178 info->dir_dname = (void *)(info->dir_in + num);
179 info->dir_dname_len = (void *)(info->dir_dname + num);
180 info->dir_dlease = (void *)(info->dir_dname_len + num);
181
182 while (num) {
183 /* dentry */
184 ceph_decode_need(p, end, sizeof(u32)*2, bad);
185 info->dir_dname_len[i] = ceph_decode_32(p);
186 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
187 info->dir_dname[i] = *p;
188 *p += info->dir_dname_len[i];
189 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
190 info->dir_dname[i]);
191 info->dir_dlease[i] = *p;
192 *p += sizeof(struct ceph_mds_reply_lease);
193
194 /* inode */
195 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
196 if (err < 0)
197 goto out_bad;
198 i++;
199 num--;
200 }
201
202 done:
203 if (*p != end)
204 goto bad;
205 return 0;
206
207 bad:
208 err = -EIO;
209 out_bad:
210 pr_err("problem parsing dir contents %d\n", err);
211 return err;
212 }
213
214 /*
215 * parse fcntl F_GETLK results
216 */
217 static int parse_reply_info_filelock(void **p, void *end,
218 struct ceph_mds_reply_info_parsed *info,
219 int features)
220 {
221 if (*p + sizeof(*info->filelock_reply) > end)
222 goto bad;
223
224 info->filelock_reply = *p;
225 *p += sizeof(*info->filelock_reply);
226
227 if (unlikely(*p != end))
228 goto bad;
229 return 0;
230
231 bad:
232 return -EIO;
233 }
234
235 /*
236 * parse create results
237 */
238 static int parse_reply_info_create(void **p, void *end,
239 struct ceph_mds_reply_info_parsed *info,
240 int features)
241 {
242 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
243 if (*p == end) {
244 info->has_create_ino = false;
245 } else {
246 info->has_create_ino = true;
247 info->ino = ceph_decode_64(p);
248 }
249 }
250
251 if (unlikely(*p != end))
252 goto bad;
253 return 0;
254
255 bad:
256 return -EIO;
257 }
258
259 /*
260 * parse extra results
261 */
262 static int parse_reply_info_extra(void **p, void *end,
263 struct ceph_mds_reply_info_parsed *info,
264 int features)
265 {
266 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
267 return parse_reply_info_filelock(p, end, info, features);
268 else if (info->head->op == CEPH_MDS_OP_READDIR ||
269 info->head->op == CEPH_MDS_OP_LSSNAP)
270 return parse_reply_info_dir(p, end, info, features);
271 else if (info->head->op == CEPH_MDS_OP_CREATE)
272 return parse_reply_info_create(p, end, info, features);
273 else
274 return -EIO;
275 }
276
277 /*
278 * parse entire mds reply
279 */
280 static int parse_reply_info(struct ceph_msg *msg,
281 struct ceph_mds_reply_info_parsed *info,
282 int features)
283 {
284 void *p, *end;
285 u32 len;
286 int err;
287
288 info->head = msg->front.iov_base;
289 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
290 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
291
292 /* trace */
293 ceph_decode_32_safe(&p, end, len, bad);
294 if (len > 0) {
295 ceph_decode_need(&p, end, len, bad);
296 err = parse_reply_info_trace(&p, p+len, info, features);
297 if (err < 0)
298 goto out_bad;
299 }
300
301 /* extra */
302 ceph_decode_32_safe(&p, end, len, bad);
303 if (len > 0) {
304 ceph_decode_need(&p, end, len, bad);
305 err = parse_reply_info_extra(&p, p+len, info, features);
306 if (err < 0)
307 goto out_bad;
308 }
309
310 /* snap blob */
311 ceph_decode_32_safe(&p, end, len, bad);
312 info->snapblob_len = len;
313 info->snapblob = p;
314 p += len;
315
316 if (p != end)
317 goto bad;
318 return 0;
319
320 bad:
321 err = -EIO;
322 out_bad:
323 pr_err("mds parse_reply err %d\n", err);
324 return err;
325 }
326
327 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
328 {
329 kfree(info->dir_in);
330 }
331
332
333 /*
334 * sessions
335 */
336 static const char *session_state_name(int s)
337 {
338 switch (s) {
339 case CEPH_MDS_SESSION_NEW: return "new";
340 case CEPH_MDS_SESSION_OPENING: return "opening";
341 case CEPH_MDS_SESSION_OPEN: return "open";
342 case CEPH_MDS_SESSION_HUNG: return "hung";
343 case CEPH_MDS_SESSION_CLOSING: return "closing";
344 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
345 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
346 default: return "???";
347 }
348 }
349
350 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
351 {
352 if (atomic_inc_not_zero(&s->s_ref)) {
353 dout("mdsc get_session %p %d -> %d\n", s,
354 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
355 return s;
356 } else {
357 dout("mdsc get_session %p 0 -- FAIL", s);
358 return NULL;
359 }
360 }
361
362 void ceph_put_mds_session(struct ceph_mds_session *s)
363 {
364 dout("mdsc put_session %p %d -> %d\n", s,
365 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
366 if (atomic_dec_and_test(&s->s_ref)) {
367 if (s->s_auth.authorizer)
368 ceph_auth_destroy_authorizer(
369 s->s_mdsc->fsc->client->monc.auth,
370 s->s_auth.authorizer);
371 kfree(s);
372 }
373 }
374
375 /*
376 * called under mdsc->mutex
377 */
378 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
379 int mds)
380 {
381 struct ceph_mds_session *session;
382
383 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
384 return NULL;
385 session = mdsc->sessions[mds];
386 dout("lookup_mds_session %p %d\n", session,
387 atomic_read(&session->s_ref));
388 get_session(session);
389 return session;
390 }
391
392 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
393 {
394 if (mds >= mdsc->max_sessions)
395 return false;
396 return mdsc->sessions[mds];
397 }
398
399 static int __verify_registered_session(struct ceph_mds_client *mdsc,
400 struct ceph_mds_session *s)
401 {
402 if (s->s_mds >= mdsc->max_sessions ||
403 mdsc->sessions[s->s_mds] != s)
404 return -ENOENT;
405 return 0;
406 }
407
408 /*
409 * create+register a new session for given mds.
410 * called under mdsc->mutex.
411 */
412 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
413 int mds)
414 {
415 struct ceph_mds_session *s;
416
417 if (mds >= mdsc->mdsmap->m_max_mds)
418 return ERR_PTR(-EINVAL);
419
420 s = kzalloc(sizeof(*s), GFP_NOFS);
421 if (!s)
422 return ERR_PTR(-ENOMEM);
423 s->s_mdsc = mdsc;
424 s->s_mds = mds;
425 s->s_state = CEPH_MDS_SESSION_NEW;
426 s->s_ttl = 0;
427 s->s_seq = 0;
428 mutex_init(&s->s_mutex);
429
430 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
431
432 spin_lock_init(&s->s_gen_ttl_lock);
433 s->s_cap_gen = 0;
434 s->s_cap_ttl = jiffies - 1;
435
436 spin_lock_init(&s->s_cap_lock);
437 s->s_renew_requested = 0;
438 s->s_renew_seq = 0;
439 INIT_LIST_HEAD(&s->s_caps);
440 s->s_nr_caps = 0;
441 s->s_trim_caps = 0;
442 atomic_set(&s->s_ref, 1);
443 INIT_LIST_HEAD(&s->s_waiting);
444 INIT_LIST_HEAD(&s->s_unsafe);
445 s->s_num_cap_releases = 0;
446 s->s_cap_iterator = NULL;
447 INIT_LIST_HEAD(&s->s_cap_releases);
448 INIT_LIST_HEAD(&s->s_cap_releases_done);
449 INIT_LIST_HEAD(&s->s_cap_flushing);
450 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
451
452 dout("register_session mds%d\n", mds);
453 if (mds >= mdsc->max_sessions) {
454 int newmax = 1 << get_count_order(mds+1);
455 struct ceph_mds_session **sa;
456
457 dout("register_session realloc to %d\n", newmax);
458 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
459 if (sa == NULL)
460 goto fail_realloc;
461 if (mdsc->sessions) {
462 memcpy(sa, mdsc->sessions,
463 mdsc->max_sessions * sizeof(void *));
464 kfree(mdsc->sessions);
465 }
466 mdsc->sessions = sa;
467 mdsc->max_sessions = newmax;
468 }
469 mdsc->sessions[mds] = s;
470 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
471
472 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
473 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
474
475 return s;
476
477 fail_realloc:
478 kfree(s);
479 return ERR_PTR(-ENOMEM);
480 }
481
482 /*
483 * called under mdsc->mutex
484 */
485 static void __unregister_session(struct ceph_mds_client *mdsc,
486 struct ceph_mds_session *s)
487 {
488 dout("__unregister_session mds%d %p\n", s->s_mds, s);
489 BUG_ON(mdsc->sessions[s->s_mds] != s);
490 mdsc->sessions[s->s_mds] = NULL;
491 ceph_con_close(&s->s_con);
492 ceph_put_mds_session(s);
493 }
494
495 /*
496 * drop session refs in request.
497 *
498 * should be last request ref, or hold mdsc->mutex
499 */
500 static void put_request_session(struct ceph_mds_request *req)
501 {
502 if (req->r_session) {
503 ceph_put_mds_session(req->r_session);
504 req->r_session = NULL;
505 }
506 }
507
508 void ceph_mdsc_release_request(struct kref *kref)
509 {
510 struct ceph_mds_request *req = container_of(kref,
511 struct ceph_mds_request,
512 r_kref);
513 if (req->r_request)
514 ceph_msg_put(req->r_request);
515 if (req->r_reply) {
516 ceph_msg_put(req->r_reply);
517 destroy_reply_info(&req->r_reply_info);
518 }
519 if (req->r_inode) {
520 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
521 iput(req->r_inode);
522 }
523 if (req->r_locked_dir)
524 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
525 if (req->r_target_inode)
526 iput(req->r_target_inode);
527 if (req->r_dentry)
528 dput(req->r_dentry);
529 if (req->r_old_dentry) {
530 /*
531 * track (and drop pins for) r_old_dentry_dir
532 * separately, since r_old_dentry's d_parent may have
533 * changed between the dir mutex being dropped and
534 * this request being freed.
535 */
536 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
537 CEPH_CAP_PIN);
538 dput(req->r_old_dentry);
539 iput(req->r_old_dentry_dir);
540 }
541 kfree(req->r_path1);
542 kfree(req->r_path2);
543 put_request_session(req);
544 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
545 kfree(req);
546 }
547
548 /*
549 * lookup session, bump ref if found.
550 *
551 * called under mdsc->mutex.
552 */
553 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
554 u64 tid)
555 {
556 struct ceph_mds_request *req;
557 struct rb_node *n = mdsc->request_tree.rb_node;
558
559 while (n) {
560 req = rb_entry(n, struct ceph_mds_request, r_node);
561 if (tid < req->r_tid)
562 n = n->rb_left;
563 else if (tid > req->r_tid)
564 n = n->rb_right;
565 else {
566 ceph_mdsc_get_request(req);
567 return req;
568 }
569 }
570 return NULL;
571 }
572
573 static void __insert_request(struct ceph_mds_client *mdsc,
574 struct ceph_mds_request *new)
575 {
576 struct rb_node **p = &mdsc->request_tree.rb_node;
577 struct rb_node *parent = NULL;
578 struct ceph_mds_request *req = NULL;
579
580 while (*p) {
581 parent = *p;
582 req = rb_entry(parent, struct ceph_mds_request, r_node);
583 if (new->r_tid < req->r_tid)
584 p = &(*p)->rb_left;
585 else if (new->r_tid > req->r_tid)
586 p = &(*p)->rb_right;
587 else
588 BUG();
589 }
590
591 rb_link_node(&new->r_node, parent, p);
592 rb_insert_color(&new->r_node, &mdsc->request_tree);
593 }
594
595 /*
596 * Register an in-flight request, and assign a tid. Link to directory
597 * are modifying (if any).
598 *
599 * Called under mdsc->mutex.
600 */
601 static void __register_request(struct ceph_mds_client *mdsc,
602 struct ceph_mds_request *req,
603 struct inode *dir)
604 {
605 req->r_tid = ++mdsc->last_tid;
606 if (req->r_num_caps)
607 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
608 req->r_num_caps);
609 dout("__register_request %p tid %lld\n", req, req->r_tid);
610 ceph_mdsc_get_request(req);
611 __insert_request(mdsc, req);
612
613 req->r_uid = current_fsuid();
614 req->r_gid = current_fsgid();
615
616 if (dir) {
617 struct ceph_inode_info *ci = ceph_inode(dir);
618
619 ihold(dir);
620 spin_lock(&ci->i_unsafe_lock);
621 req->r_unsafe_dir = dir;
622 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
623 spin_unlock(&ci->i_unsafe_lock);
624 }
625 }
626
627 static void __unregister_request(struct ceph_mds_client *mdsc,
628 struct ceph_mds_request *req)
629 {
630 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
631 rb_erase(&req->r_node, &mdsc->request_tree);
632 RB_CLEAR_NODE(&req->r_node);
633
634 if (req->r_unsafe_dir) {
635 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
636
637 spin_lock(&ci->i_unsafe_lock);
638 list_del_init(&req->r_unsafe_dir_item);
639 spin_unlock(&ci->i_unsafe_lock);
640
641 iput(req->r_unsafe_dir);
642 req->r_unsafe_dir = NULL;
643 }
644
645 complete_all(&req->r_safe_completion);
646
647 ceph_mdsc_put_request(req);
648 }
649
650 /*
651 * Choose mds to send request to next. If there is a hint set in the
652 * request (e.g., due to a prior forward hint from the mds), use that.
653 * Otherwise, consult frag tree and/or caps to identify the
654 * appropriate mds. If all else fails, choose randomly.
655 *
656 * Called under mdsc->mutex.
657 */
658 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
659 {
660 /*
661 * we don't need to worry about protecting the d_parent access
662 * here because we never renaming inside the snapped namespace
663 * except to resplice to another snapdir, and either the old or new
664 * result is a valid result.
665 */
666 while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
667 dentry = dentry->d_parent;
668 return dentry;
669 }
670
671 static int __choose_mds(struct ceph_mds_client *mdsc,
672 struct ceph_mds_request *req)
673 {
674 struct inode *inode;
675 struct ceph_inode_info *ci;
676 struct ceph_cap *cap;
677 int mode = req->r_direct_mode;
678 int mds = -1;
679 u32 hash = req->r_direct_hash;
680 bool is_hash = req->r_direct_is_hash;
681
682 /*
683 * is there a specific mds we should try? ignore hint if we have
684 * no session and the mds is not up (active or recovering).
685 */
686 if (req->r_resend_mds >= 0 &&
687 (__have_session(mdsc, req->r_resend_mds) ||
688 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
689 dout("choose_mds using resend_mds mds%d\n",
690 req->r_resend_mds);
691 return req->r_resend_mds;
692 }
693
694 if (mode == USE_RANDOM_MDS)
695 goto random;
696
697 inode = NULL;
698 if (req->r_inode) {
699 inode = req->r_inode;
700 } else if (req->r_dentry) {
701 /* ignore race with rename; old or new d_parent is okay */
702 struct dentry *parent = req->r_dentry->d_parent;
703 struct inode *dir = parent->d_inode;
704
705 if (dir->i_sb != mdsc->fsc->sb) {
706 /* not this fs! */
707 inode = req->r_dentry->d_inode;
708 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
709 /* direct snapped/virtual snapdir requests
710 * based on parent dir inode */
711 struct dentry *dn = get_nonsnap_parent(parent);
712 inode = dn->d_inode;
713 dout("__choose_mds using nonsnap parent %p\n", inode);
714 } else if (req->r_dentry->d_inode) {
715 /* dentry target */
716 inode = req->r_dentry->d_inode;
717 } else {
718 /* dir + name */
719 inode = dir;
720 hash = ceph_dentry_hash(dir, req->r_dentry);
721 is_hash = true;
722 }
723 }
724
725 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
726 (int)hash, mode);
727 if (!inode)
728 goto random;
729 ci = ceph_inode(inode);
730
731 if (is_hash && S_ISDIR(inode->i_mode)) {
732 struct ceph_inode_frag frag;
733 int found;
734
735 ceph_choose_frag(ci, hash, &frag, &found);
736 if (found) {
737 if (mode == USE_ANY_MDS && frag.ndist > 0) {
738 u8 r;
739
740 /* choose a random replica */
741 get_random_bytes(&r, 1);
742 r %= frag.ndist;
743 mds = frag.dist[r];
744 dout("choose_mds %p %llx.%llx "
745 "frag %u mds%d (%d/%d)\n",
746 inode, ceph_vinop(inode),
747 frag.frag, mds,
748 (int)r, frag.ndist);
749 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
750 CEPH_MDS_STATE_ACTIVE)
751 return mds;
752 }
753
754 /* since this file/dir wasn't known to be
755 * replicated, then we want to look for the
756 * authoritative mds. */
757 mode = USE_AUTH_MDS;
758 if (frag.mds >= 0) {
759 /* choose auth mds */
760 mds = frag.mds;
761 dout("choose_mds %p %llx.%llx "
762 "frag %u mds%d (auth)\n",
763 inode, ceph_vinop(inode), frag.frag, mds);
764 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
765 CEPH_MDS_STATE_ACTIVE)
766 return mds;
767 }
768 }
769 }
770
771 spin_lock(&ci->i_ceph_lock);
772 cap = NULL;
773 if (mode == USE_AUTH_MDS)
774 cap = ci->i_auth_cap;
775 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
776 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
777 if (!cap) {
778 spin_unlock(&ci->i_ceph_lock);
779 goto random;
780 }
781 mds = cap->session->s_mds;
782 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
783 inode, ceph_vinop(inode), mds,
784 cap == ci->i_auth_cap ? "auth " : "", cap);
785 spin_unlock(&ci->i_ceph_lock);
786 return mds;
787
788 random:
789 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
790 dout("choose_mds chose random mds%d\n", mds);
791 return mds;
792 }
793
794
795 /*
796 * session messages
797 */
798 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
799 {
800 struct ceph_msg *msg;
801 struct ceph_mds_session_head *h;
802
803 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
804 false);
805 if (!msg) {
806 pr_err("create_session_msg ENOMEM creating msg\n");
807 return NULL;
808 }
809 h = msg->front.iov_base;
810 h->op = cpu_to_le32(op);
811 h->seq = cpu_to_le64(seq);
812 return msg;
813 }
814
815 /*
816 * send session open request.
817 *
818 * called under mdsc->mutex
819 */
820 static int __open_session(struct ceph_mds_client *mdsc,
821 struct ceph_mds_session *session)
822 {
823 struct ceph_msg *msg;
824 int mstate;
825 int mds = session->s_mds;
826
827 /* wait for mds to go active? */
828 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
829 dout("open_session to mds%d (%s)\n", mds,
830 ceph_mds_state_name(mstate));
831 session->s_state = CEPH_MDS_SESSION_OPENING;
832 session->s_renew_requested = jiffies;
833
834 /* send connect message */
835 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
836 if (!msg)
837 return -ENOMEM;
838 ceph_con_send(&session->s_con, msg);
839 return 0;
840 }
841
842 /*
843 * open sessions for any export targets for the given mds
844 *
845 * called under mdsc->mutex
846 */
847 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
848 struct ceph_mds_session *session)
849 {
850 struct ceph_mds_info *mi;
851 struct ceph_mds_session *ts;
852 int i, mds = session->s_mds;
853 int target;
854
855 if (mds >= mdsc->mdsmap->m_max_mds)
856 return;
857 mi = &mdsc->mdsmap->m_info[mds];
858 dout("open_export_target_sessions for mds%d (%d targets)\n",
859 session->s_mds, mi->num_export_targets);
860
861 for (i = 0; i < mi->num_export_targets; i++) {
862 target = mi->export_targets[i];
863 ts = __ceph_lookup_mds_session(mdsc, target);
864 if (!ts) {
865 ts = register_session(mdsc, target);
866 if (IS_ERR(ts))
867 return;
868 }
869 if (session->s_state == CEPH_MDS_SESSION_NEW ||
870 session->s_state == CEPH_MDS_SESSION_CLOSING)
871 __open_session(mdsc, session);
872 else
873 dout(" mds%d target mds%d %p is %s\n", session->s_mds,
874 i, ts, session_state_name(ts->s_state));
875 ceph_put_mds_session(ts);
876 }
877 }
878
879 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
880 struct ceph_mds_session *session)
881 {
882 mutex_lock(&mdsc->mutex);
883 __open_export_target_sessions(mdsc, session);
884 mutex_unlock(&mdsc->mutex);
885 }
886
887 /*
888 * session caps
889 */
890
891 /*
892 * Free preallocated cap messages assigned to this session
893 */
894 static void cleanup_cap_releases(struct ceph_mds_session *session)
895 {
896 struct ceph_msg *msg;
897
898 spin_lock(&session->s_cap_lock);
899 while (!list_empty(&session->s_cap_releases)) {
900 msg = list_first_entry(&session->s_cap_releases,
901 struct ceph_msg, list_head);
902 list_del_init(&msg->list_head);
903 ceph_msg_put(msg);
904 }
905 while (!list_empty(&session->s_cap_releases_done)) {
906 msg = list_first_entry(&session->s_cap_releases_done,
907 struct ceph_msg, list_head);
908 list_del_init(&msg->list_head);
909 ceph_msg_put(msg);
910 }
911 spin_unlock(&session->s_cap_lock);
912 }
913
914 /*
915 * Helper to safely iterate over all caps associated with a session, with
916 * special care taken to handle a racing __ceph_remove_cap().
917 *
918 * Caller must hold session s_mutex.
919 */
920 static int iterate_session_caps(struct ceph_mds_session *session,
921 int (*cb)(struct inode *, struct ceph_cap *,
922 void *), void *arg)
923 {
924 struct list_head *p;
925 struct ceph_cap *cap;
926 struct inode *inode, *last_inode = NULL;
927 struct ceph_cap *old_cap = NULL;
928 int ret;
929
930 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
931 spin_lock(&session->s_cap_lock);
932 p = session->s_caps.next;
933 while (p != &session->s_caps) {
934 cap = list_entry(p, struct ceph_cap, session_caps);
935 inode = igrab(&cap->ci->vfs_inode);
936 if (!inode) {
937 p = p->next;
938 continue;
939 }
940 session->s_cap_iterator = cap;
941 spin_unlock(&session->s_cap_lock);
942
943 if (last_inode) {
944 iput(last_inode);
945 last_inode = NULL;
946 }
947 if (old_cap) {
948 ceph_put_cap(session->s_mdsc, old_cap);
949 old_cap = NULL;
950 }
951
952 ret = cb(inode, cap, arg);
953 last_inode = inode;
954
955 spin_lock(&session->s_cap_lock);
956 p = p->next;
957 if (cap->ci == NULL) {
958 dout("iterate_session_caps finishing cap %p removal\n",
959 cap);
960 BUG_ON(cap->session != session);
961 list_del_init(&cap->session_caps);
962 session->s_nr_caps--;
963 cap->session = NULL;
964 old_cap = cap; /* put_cap it w/o locks held */
965 }
966 if (ret < 0)
967 goto out;
968 }
969 ret = 0;
970 out:
971 session->s_cap_iterator = NULL;
972 spin_unlock(&session->s_cap_lock);
973
974 if (last_inode)
975 iput(last_inode);
976 if (old_cap)
977 ceph_put_cap(session->s_mdsc, old_cap);
978
979 return ret;
980 }
981
982 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
983 void *arg)
984 {
985 struct ceph_inode_info *ci = ceph_inode(inode);
986 int drop = 0;
987
988 dout("removing cap %p, ci is %p, inode is %p\n",
989 cap, ci, &ci->vfs_inode);
990 spin_lock(&ci->i_ceph_lock);
991 __ceph_remove_cap(cap);
992 if (!__ceph_is_any_real_caps(ci)) {
993 struct ceph_mds_client *mdsc =
994 ceph_sb_to_client(inode->i_sb)->mdsc;
995
996 spin_lock(&mdsc->cap_dirty_lock);
997 if (!list_empty(&ci->i_dirty_item)) {
998 pr_info(" dropping dirty %s state for %p %lld\n",
999 ceph_cap_string(ci->i_dirty_caps),
1000 inode, ceph_ino(inode));
1001 ci->i_dirty_caps = 0;
1002 list_del_init(&ci->i_dirty_item);
1003 drop = 1;
1004 }
1005 if (!list_empty(&ci->i_flushing_item)) {
1006 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1007 ceph_cap_string(ci->i_flushing_caps),
1008 inode, ceph_ino(inode));
1009 ci->i_flushing_caps = 0;
1010 list_del_init(&ci->i_flushing_item);
1011 mdsc->num_cap_flushing--;
1012 drop = 1;
1013 }
1014 if (drop && ci->i_wrbuffer_ref) {
1015 pr_info(" dropping dirty data for %p %lld\n",
1016 inode, ceph_ino(inode));
1017 ci->i_wrbuffer_ref = 0;
1018 ci->i_wrbuffer_ref_head = 0;
1019 drop++;
1020 }
1021 spin_unlock(&mdsc->cap_dirty_lock);
1022 }
1023 spin_unlock(&ci->i_ceph_lock);
1024 while (drop--)
1025 iput(inode);
1026 return 0;
1027 }
1028
1029 /*
1030 * caller must hold session s_mutex
1031 */
1032 static void remove_session_caps(struct ceph_mds_session *session)
1033 {
1034 dout("remove_session_caps on %p\n", session);
1035 iterate_session_caps(session, remove_session_caps_cb, NULL);
1036
1037 spin_lock(&session->s_cap_lock);
1038 if (session->s_nr_caps > 0) {
1039 struct super_block *sb = session->s_mdsc->fsc->sb;
1040 struct inode *inode;
1041 struct ceph_cap *cap, *prev = NULL;
1042 struct ceph_vino vino;
1043 /*
1044 * iterate_session_caps() skips inodes that are being
1045 * deleted, we need to wait until deletions are complete.
1046 * __wait_on_freeing_inode() is designed for the job,
1047 * but it is not exported, so use lookup inode function
1048 * to access it.
1049 */
1050 while (!list_empty(&session->s_caps)) {
1051 cap = list_entry(session->s_caps.next,
1052 struct ceph_cap, session_caps);
1053 if (cap == prev)
1054 break;
1055 prev = cap;
1056 vino = cap->ci->i_vino;
1057 spin_unlock(&session->s_cap_lock);
1058
1059 inode = ceph_find_inode(sb, vino);
1060 iput(inode);
1061
1062 spin_lock(&session->s_cap_lock);
1063 }
1064 }
1065 spin_unlock(&session->s_cap_lock);
1066
1067 BUG_ON(session->s_nr_caps > 0);
1068 BUG_ON(!list_empty(&session->s_cap_flushing));
1069 cleanup_cap_releases(session);
1070 }
1071
1072 /*
1073 * wake up any threads waiting on this session's caps. if the cap is
1074 * old (didn't get renewed on the client reconnect), remove it now.
1075 *
1076 * caller must hold s_mutex.
1077 */
1078 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1079 void *arg)
1080 {
1081 struct ceph_inode_info *ci = ceph_inode(inode);
1082
1083 wake_up_all(&ci->i_cap_wq);
1084 if (arg) {
1085 spin_lock(&ci->i_ceph_lock);
1086 ci->i_wanted_max_size = 0;
1087 ci->i_requested_max_size = 0;
1088 spin_unlock(&ci->i_ceph_lock);
1089 }
1090 return 0;
1091 }
1092
1093 static void wake_up_session_caps(struct ceph_mds_session *session,
1094 int reconnect)
1095 {
1096 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1097 iterate_session_caps(session, wake_up_session_cb,
1098 (void *)(unsigned long)reconnect);
1099 }
1100
1101 /*
1102 * Send periodic message to MDS renewing all currently held caps. The
1103 * ack will reset the expiration for all caps from this session.
1104 *
1105 * caller holds s_mutex
1106 */
1107 static int send_renew_caps(struct ceph_mds_client *mdsc,
1108 struct ceph_mds_session *session)
1109 {
1110 struct ceph_msg *msg;
1111 int state;
1112
1113 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1114 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1115 pr_info("mds%d caps stale\n", session->s_mds);
1116 session->s_renew_requested = jiffies;
1117
1118 /* do not try to renew caps until a recovering mds has reconnected
1119 * with its clients. */
1120 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1121 if (state < CEPH_MDS_STATE_RECONNECT) {
1122 dout("send_renew_caps ignoring mds%d (%s)\n",
1123 session->s_mds, ceph_mds_state_name(state));
1124 return 0;
1125 }
1126
1127 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1128 ceph_mds_state_name(state));
1129 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1130 ++session->s_renew_seq);
1131 if (!msg)
1132 return -ENOMEM;
1133 ceph_con_send(&session->s_con, msg);
1134 return 0;
1135 }
1136
1137 /*
1138 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1139 *
1140 * Called under session->s_mutex
1141 */
1142 static void renewed_caps(struct ceph_mds_client *mdsc,
1143 struct ceph_mds_session *session, int is_renew)
1144 {
1145 int was_stale;
1146 int wake = 0;
1147
1148 spin_lock(&session->s_cap_lock);
1149 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1150
1151 session->s_cap_ttl = session->s_renew_requested +
1152 mdsc->mdsmap->m_session_timeout*HZ;
1153
1154 if (was_stale) {
1155 if (time_before(jiffies, session->s_cap_ttl)) {
1156 pr_info("mds%d caps renewed\n", session->s_mds);
1157 wake = 1;
1158 } else {
1159 pr_info("mds%d caps still stale\n", session->s_mds);
1160 }
1161 }
1162 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1163 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1164 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1165 spin_unlock(&session->s_cap_lock);
1166
1167 if (wake)
1168 wake_up_session_caps(session, 0);
1169 }
1170
1171 /*
1172 * send a session close request
1173 */
1174 static int request_close_session(struct ceph_mds_client *mdsc,
1175 struct ceph_mds_session *session)
1176 {
1177 struct ceph_msg *msg;
1178
1179 dout("request_close_session mds%d state %s seq %lld\n",
1180 session->s_mds, session_state_name(session->s_state),
1181 session->s_seq);
1182 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1183 if (!msg)
1184 return -ENOMEM;
1185 ceph_con_send(&session->s_con, msg);
1186 return 0;
1187 }
1188
1189 /*
1190 * Called with s_mutex held.
1191 */
1192 static int __close_session(struct ceph_mds_client *mdsc,
1193 struct ceph_mds_session *session)
1194 {
1195 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1196 return 0;
1197 session->s_state = CEPH_MDS_SESSION_CLOSING;
1198 return request_close_session(mdsc, session);
1199 }
1200
1201 /*
1202 * Trim old(er) caps.
1203 *
1204 * Because we can't cache an inode without one or more caps, we do
1205 * this indirectly: if a cap is unused, we prune its aliases, at which
1206 * point the inode will hopefully get dropped to.
1207 *
1208 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1209 * memory pressure from the MDS, though, so it needn't be perfect.
1210 */
1211 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1212 {
1213 struct ceph_mds_session *session = arg;
1214 struct ceph_inode_info *ci = ceph_inode(inode);
1215 int used, oissued, mine;
1216
1217 if (session->s_trim_caps <= 0)
1218 return -1;
1219
1220 spin_lock(&ci->i_ceph_lock);
1221 mine = cap->issued | cap->implemented;
1222 used = __ceph_caps_used(ci);
1223 oissued = __ceph_caps_issued_other(ci, cap);
1224
1225 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1226 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1227 ceph_cap_string(used));
1228 if (ci->i_dirty_caps)
1229 goto out; /* dirty caps */
1230 if ((used & ~oissued) & mine)
1231 goto out; /* we need these caps */
1232
1233 session->s_trim_caps--;
1234 if (oissued) {
1235 /* we aren't the only cap.. just remove us */
1236 __queue_cap_release(session, ceph_ino(inode), cap->cap_id,
1237 cap->mseq, cap->issue_seq);
1238 __ceph_remove_cap(cap);
1239 } else {
1240 /* try to drop referring dentries */
1241 spin_unlock(&ci->i_ceph_lock);
1242 d_prune_aliases(inode);
1243 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1244 inode, cap, atomic_read(&inode->i_count));
1245 return 0;
1246 }
1247
1248 out:
1249 spin_unlock(&ci->i_ceph_lock);
1250 return 0;
1251 }
1252
1253 /*
1254 * Trim session cap count down to some max number.
1255 */
1256 static int trim_caps(struct ceph_mds_client *mdsc,
1257 struct ceph_mds_session *session,
1258 int max_caps)
1259 {
1260 int trim_caps = session->s_nr_caps - max_caps;
1261
1262 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1263 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1264 if (trim_caps > 0) {
1265 session->s_trim_caps = trim_caps;
1266 iterate_session_caps(session, trim_caps_cb, session);
1267 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1268 session->s_mds, session->s_nr_caps, max_caps,
1269 trim_caps - session->s_trim_caps);
1270 session->s_trim_caps = 0;
1271 }
1272 return 0;
1273 }
1274
1275 /*
1276 * Allocate cap_release messages. If there is a partially full message
1277 * in the queue, try to allocate enough to cover it's remainder, so that
1278 * we can send it immediately.
1279 *
1280 * Called under s_mutex.
1281 */
1282 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1283 struct ceph_mds_session *session)
1284 {
1285 struct ceph_msg *msg, *partial = NULL;
1286 struct ceph_mds_cap_release *head;
1287 int err = -ENOMEM;
1288 int extra = mdsc->fsc->mount_options->cap_release_safety;
1289 int num;
1290
1291 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1292 extra);
1293
1294 spin_lock(&session->s_cap_lock);
1295
1296 if (!list_empty(&session->s_cap_releases)) {
1297 msg = list_first_entry(&session->s_cap_releases,
1298 struct ceph_msg,
1299 list_head);
1300 head = msg->front.iov_base;
1301 num = le32_to_cpu(head->num);
1302 if (num) {
1303 dout(" partial %p with (%d/%d)\n", msg, num,
1304 (int)CEPH_CAPS_PER_RELEASE);
1305 extra += CEPH_CAPS_PER_RELEASE - num;
1306 partial = msg;
1307 }
1308 }
1309 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1310 spin_unlock(&session->s_cap_lock);
1311 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1312 GFP_NOFS, false);
1313 if (!msg)
1314 goto out_unlocked;
1315 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1316 (int)msg->front.iov_len);
1317 head = msg->front.iov_base;
1318 head->num = cpu_to_le32(0);
1319 msg->front.iov_len = sizeof(*head);
1320 spin_lock(&session->s_cap_lock);
1321 list_add(&msg->list_head, &session->s_cap_releases);
1322 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1323 }
1324
1325 if (partial) {
1326 head = partial->front.iov_base;
1327 num = le32_to_cpu(head->num);
1328 dout(" queueing partial %p with %d/%d\n", partial, num,
1329 (int)CEPH_CAPS_PER_RELEASE);
1330 list_move_tail(&partial->list_head,
1331 &session->s_cap_releases_done);
1332 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1333 }
1334 err = 0;
1335 spin_unlock(&session->s_cap_lock);
1336 out_unlocked:
1337 return err;
1338 }
1339
1340 /*
1341 * flush all dirty inode data to disk.
1342 *
1343 * returns true if we've flushed through want_flush_seq
1344 */
1345 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1346 {
1347 int mds, ret = 1;
1348
1349 dout("check_cap_flush want %lld\n", want_flush_seq);
1350 mutex_lock(&mdsc->mutex);
1351 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1352 struct ceph_mds_session *session = mdsc->sessions[mds];
1353
1354 if (!session)
1355 continue;
1356 get_session(session);
1357 mutex_unlock(&mdsc->mutex);
1358
1359 mutex_lock(&session->s_mutex);
1360 if (!list_empty(&session->s_cap_flushing)) {
1361 struct ceph_inode_info *ci =
1362 list_entry(session->s_cap_flushing.next,
1363 struct ceph_inode_info,
1364 i_flushing_item);
1365 struct inode *inode = &ci->vfs_inode;
1366
1367 spin_lock(&ci->i_ceph_lock);
1368 if (ci->i_cap_flush_seq <= want_flush_seq) {
1369 dout("check_cap_flush still flushing %p "
1370 "seq %lld <= %lld to mds%d\n", inode,
1371 ci->i_cap_flush_seq, want_flush_seq,
1372 session->s_mds);
1373 ret = 0;
1374 }
1375 spin_unlock(&ci->i_ceph_lock);
1376 }
1377 mutex_unlock(&session->s_mutex);
1378 ceph_put_mds_session(session);
1379
1380 if (!ret)
1381 return ret;
1382 mutex_lock(&mdsc->mutex);
1383 }
1384
1385 mutex_unlock(&mdsc->mutex);
1386 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1387 return ret;
1388 }
1389
1390 /*
1391 * called under s_mutex
1392 */
1393 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1394 struct ceph_mds_session *session)
1395 {
1396 struct ceph_msg *msg;
1397
1398 dout("send_cap_releases mds%d\n", session->s_mds);
1399 spin_lock(&session->s_cap_lock);
1400 while (!list_empty(&session->s_cap_releases_done)) {
1401 msg = list_first_entry(&session->s_cap_releases_done,
1402 struct ceph_msg, list_head);
1403 list_del_init(&msg->list_head);
1404 spin_unlock(&session->s_cap_lock);
1405 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1406 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1407 ceph_con_send(&session->s_con, msg);
1408 spin_lock(&session->s_cap_lock);
1409 }
1410 spin_unlock(&session->s_cap_lock);
1411 }
1412
1413 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1414 struct ceph_mds_session *session)
1415 {
1416 struct ceph_msg *msg;
1417 struct ceph_mds_cap_release *head;
1418 unsigned num;
1419
1420 dout("discard_cap_releases mds%d\n", session->s_mds);
1421 spin_lock(&session->s_cap_lock);
1422
1423 /* zero out the in-progress message */
1424 msg = list_first_entry(&session->s_cap_releases,
1425 struct ceph_msg, list_head);
1426 head = msg->front.iov_base;
1427 num = le32_to_cpu(head->num);
1428 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1429 head->num = cpu_to_le32(0);
1430 msg->front.iov_len = sizeof(*head);
1431 session->s_num_cap_releases += num;
1432
1433 /* requeue completed messages */
1434 while (!list_empty(&session->s_cap_releases_done)) {
1435 msg = list_first_entry(&session->s_cap_releases_done,
1436 struct ceph_msg, list_head);
1437 list_del_init(&msg->list_head);
1438
1439 head = msg->front.iov_base;
1440 num = le32_to_cpu(head->num);
1441 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1442 num);
1443 session->s_num_cap_releases += num;
1444 head->num = cpu_to_le32(0);
1445 msg->front.iov_len = sizeof(*head);
1446 list_add(&msg->list_head, &session->s_cap_releases);
1447 }
1448
1449 spin_unlock(&session->s_cap_lock);
1450 }
1451
1452 /*
1453 * requests
1454 */
1455
1456 /*
1457 * Create an mds request.
1458 */
1459 struct ceph_mds_request *
1460 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1461 {
1462 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1463
1464 if (!req)
1465 return ERR_PTR(-ENOMEM);
1466
1467 mutex_init(&req->r_fill_mutex);
1468 req->r_mdsc = mdsc;
1469 req->r_started = jiffies;
1470 req->r_resend_mds = -1;
1471 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1472 req->r_fmode = -1;
1473 kref_init(&req->r_kref);
1474 INIT_LIST_HEAD(&req->r_wait);
1475 init_completion(&req->r_completion);
1476 init_completion(&req->r_safe_completion);
1477 INIT_LIST_HEAD(&req->r_unsafe_item);
1478
1479 req->r_op = op;
1480 req->r_direct_mode = mode;
1481 return req;
1482 }
1483
1484 /*
1485 * return oldest (lowest) request, tid in request tree, 0 if none.
1486 *
1487 * called under mdsc->mutex.
1488 */
1489 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1490 {
1491 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1492 return NULL;
1493 return rb_entry(rb_first(&mdsc->request_tree),
1494 struct ceph_mds_request, r_node);
1495 }
1496
1497 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1498 {
1499 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1500
1501 if (req)
1502 return req->r_tid;
1503 return 0;
1504 }
1505
1506 /*
1507 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1508 * on build_path_from_dentry in fs/cifs/dir.c.
1509 *
1510 * If @stop_on_nosnap, generate path relative to the first non-snapped
1511 * inode.
1512 *
1513 * Encode hidden .snap dirs as a double /, i.e.
1514 * foo/.snap/bar -> foo//bar
1515 */
1516 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1517 int stop_on_nosnap)
1518 {
1519 struct dentry *temp;
1520 char *path;
1521 int len, pos;
1522 unsigned seq;
1523
1524 if (dentry == NULL)
1525 return ERR_PTR(-EINVAL);
1526
1527 retry:
1528 len = 0;
1529 seq = read_seqbegin(&rename_lock);
1530 rcu_read_lock();
1531 for (temp = dentry; !IS_ROOT(temp);) {
1532 struct inode *inode = temp->d_inode;
1533 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1534 len++; /* slash only */
1535 else if (stop_on_nosnap && inode &&
1536 ceph_snap(inode) == CEPH_NOSNAP)
1537 break;
1538 else
1539 len += 1 + temp->d_name.len;
1540 temp = temp->d_parent;
1541 }
1542 rcu_read_unlock();
1543 if (len)
1544 len--; /* no leading '/' */
1545
1546 path = kmalloc(len+1, GFP_NOFS);
1547 if (path == NULL)
1548 return ERR_PTR(-ENOMEM);
1549 pos = len;
1550 path[pos] = 0; /* trailing null */
1551 rcu_read_lock();
1552 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1553 struct inode *inode;
1554
1555 spin_lock(&temp->d_lock);
1556 inode = temp->d_inode;
1557 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1558 dout("build_path path+%d: %p SNAPDIR\n",
1559 pos, temp);
1560 } else if (stop_on_nosnap && inode &&
1561 ceph_snap(inode) == CEPH_NOSNAP) {
1562 spin_unlock(&temp->d_lock);
1563 break;
1564 } else {
1565 pos -= temp->d_name.len;
1566 if (pos < 0) {
1567 spin_unlock(&temp->d_lock);
1568 break;
1569 }
1570 strncpy(path + pos, temp->d_name.name,
1571 temp->d_name.len);
1572 }
1573 spin_unlock(&temp->d_lock);
1574 if (pos)
1575 path[--pos] = '/';
1576 temp = temp->d_parent;
1577 }
1578 rcu_read_unlock();
1579 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1580 pr_err("build_path did not end path lookup where "
1581 "expected, namelen is %d, pos is %d\n", len, pos);
1582 /* presumably this is only possible if racing with a
1583 rename of one of the parent directories (we can not
1584 lock the dentries above us to prevent this, but
1585 retrying should be harmless) */
1586 kfree(path);
1587 goto retry;
1588 }
1589
1590 *base = ceph_ino(temp->d_inode);
1591 *plen = len;
1592 dout("build_path on %p %d built %llx '%.*s'\n",
1593 dentry, d_count(dentry), *base, len, path);
1594 return path;
1595 }
1596
1597 static int build_dentry_path(struct dentry *dentry,
1598 const char **ppath, int *ppathlen, u64 *pino,
1599 int *pfreepath)
1600 {
1601 char *path;
1602
1603 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1604 *pino = ceph_ino(dentry->d_parent->d_inode);
1605 *ppath = dentry->d_name.name;
1606 *ppathlen = dentry->d_name.len;
1607 return 0;
1608 }
1609 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1610 if (IS_ERR(path))
1611 return PTR_ERR(path);
1612 *ppath = path;
1613 *pfreepath = 1;
1614 return 0;
1615 }
1616
1617 static int build_inode_path(struct inode *inode,
1618 const char **ppath, int *ppathlen, u64 *pino,
1619 int *pfreepath)
1620 {
1621 struct dentry *dentry;
1622 char *path;
1623
1624 if (ceph_snap(inode) == CEPH_NOSNAP) {
1625 *pino = ceph_ino(inode);
1626 *ppathlen = 0;
1627 return 0;
1628 }
1629 dentry = d_find_alias(inode);
1630 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1631 dput(dentry);
1632 if (IS_ERR(path))
1633 return PTR_ERR(path);
1634 *ppath = path;
1635 *pfreepath = 1;
1636 return 0;
1637 }
1638
1639 /*
1640 * request arguments may be specified via an inode *, a dentry *, or
1641 * an explicit ino+path.
1642 */
1643 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1644 const char *rpath, u64 rino,
1645 const char **ppath, int *pathlen,
1646 u64 *ino, int *freepath)
1647 {
1648 int r = 0;
1649
1650 if (rinode) {
1651 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1652 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1653 ceph_snap(rinode));
1654 } else if (rdentry) {
1655 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1656 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1657 *ppath);
1658 } else if (rpath || rino) {
1659 *ino = rino;
1660 *ppath = rpath;
1661 *pathlen = rpath ? strlen(rpath) : 0;
1662 dout(" path %.*s\n", *pathlen, rpath);
1663 }
1664
1665 return r;
1666 }
1667
1668 /*
1669 * called under mdsc->mutex
1670 */
1671 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1672 struct ceph_mds_request *req,
1673 int mds)
1674 {
1675 struct ceph_msg *msg;
1676 struct ceph_mds_request_head *head;
1677 const char *path1 = NULL;
1678 const char *path2 = NULL;
1679 u64 ino1 = 0, ino2 = 0;
1680 int pathlen1 = 0, pathlen2 = 0;
1681 int freepath1 = 0, freepath2 = 0;
1682 int len;
1683 u16 releases;
1684 void *p, *end;
1685 int ret;
1686
1687 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1688 req->r_path1, req->r_ino1.ino,
1689 &path1, &pathlen1, &ino1, &freepath1);
1690 if (ret < 0) {
1691 msg = ERR_PTR(ret);
1692 goto out;
1693 }
1694
1695 ret = set_request_path_attr(NULL, req->r_old_dentry,
1696 req->r_path2, req->r_ino2.ino,
1697 &path2, &pathlen2, &ino2, &freepath2);
1698 if (ret < 0) {
1699 msg = ERR_PTR(ret);
1700 goto out_free1;
1701 }
1702
1703 len = sizeof(*head) +
1704 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1705
1706 /* calculate (max) length for cap releases */
1707 len += sizeof(struct ceph_mds_request_release) *
1708 (!!req->r_inode_drop + !!req->r_dentry_drop +
1709 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1710 if (req->r_dentry_drop)
1711 len += req->r_dentry->d_name.len;
1712 if (req->r_old_dentry_drop)
1713 len += req->r_old_dentry->d_name.len;
1714
1715 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1716 if (!msg) {
1717 msg = ERR_PTR(-ENOMEM);
1718 goto out_free2;
1719 }
1720
1721 msg->hdr.tid = cpu_to_le64(req->r_tid);
1722
1723 head = msg->front.iov_base;
1724 p = msg->front.iov_base + sizeof(*head);
1725 end = msg->front.iov_base + msg->front.iov_len;
1726
1727 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1728 head->op = cpu_to_le32(req->r_op);
1729 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1730 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1731 head->args = req->r_args;
1732
1733 ceph_encode_filepath(&p, end, ino1, path1);
1734 ceph_encode_filepath(&p, end, ino2, path2);
1735
1736 /* make note of release offset, in case we need to replay */
1737 req->r_request_release_offset = p - msg->front.iov_base;
1738
1739 /* cap releases */
1740 releases = 0;
1741 if (req->r_inode_drop)
1742 releases += ceph_encode_inode_release(&p,
1743 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1744 mds, req->r_inode_drop, req->r_inode_unless, 0);
1745 if (req->r_dentry_drop)
1746 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1747 mds, req->r_dentry_drop, req->r_dentry_unless);
1748 if (req->r_old_dentry_drop)
1749 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1750 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1751 if (req->r_old_inode_drop)
1752 releases += ceph_encode_inode_release(&p,
1753 req->r_old_dentry->d_inode,
1754 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1755 head->num_releases = cpu_to_le16(releases);
1756
1757 BUG_ON(p > end);
1758 msg->front.iov_len = p - msg->front.iov_base;
1759 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1760
1761 if (req->r_data_len) {
1762 /* outbound data set only by ceph_sync_setxattr() */
1763 BUG_ON(!req->r_pages);
1764 ceph_msg_data_add_pages(msg, req->r_pages, req->r_data_len, 0);
1765 }
1766
1767 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1768 msg->hdr.data_off = cpu_to_le16(0);
1769
1770 out_free2:
1771 if (freepath2)
1772 kfree((char *)path2);
1773 out_free1:
1774 if (freepath1)
1775 kfree((char *)path1);
1776 out:
1777 return msg;
1778 }
1779
1780 /*
1781 * called under mdsc->mutex if error, under no mutex if
1782 * success.
1783 */
1784 static void complete_request(struct ceph_mds_client *mdsc,
1785 struct ceph_mds_request *req)
1786 {
1787 if (req->r_callback)
1788 req->r_callback(mdsc, req);
1789 else
1790 complete_all(&req->r_completion);
1791 }
1792
1793 /*
1794 * called under mdsc->mutex
1795 */
1796 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1797 struct ceph_mds_request *req,
1798 int mds)
1799 {
1800 struct ceph_mds_request_head *rhead;
1801 struct ceph_msg *msg;
1802 int flags = 0;
1803
1804 req->r_attempts++;
1805 if (req->r_inode) {
1806 struct ceph_cap *cap =
1807 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1808
1809 if (cap)
1810 req->r_sent_on_mseq = cap->mseq;
1811 else
1812 req->r_sent_on_mseq = -1;
1813 }
1814 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1815 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1816
1817 if (req->r_got_unsafe) {
1818 /*
1819 * Replay. Do not regenerate message (and rebuild
1820 * paths, etc.); just use the original message.
1821 * Rebuilding paths will break for renames because
1822 * d_move mangles the src name.
1823 */
1824 msg = req->r_request;
1825 rhead = msg->front.iov_base;
1826
1827 flags = le32_to_cpu(rhead->flags);
1828 flags |= CEPH_MDS_FLAG_REPLAY;
1829 rhead->flags = cpu_to_le32(flags);
1830
1831 if (req->r_target_inode)
1832 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1833
1834 rhead->num_retry = req->r_attempts - 1;
1835
1836 /* remove cap/dentry releases from message */
1837 rhead->num_releases = 0;
1838 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1839 msg->front.iov_len = req->r_request_release_offset;
1840 return 0;
1841 }
1842
1843 if (req->r_request) {
1844 ceph_msg_put(req->r_request);
1845 req->r_request = NULL;
1846 }
1847 msg = create_request_message(mdsc, req, mds);
1848 if (IS_ERR(msg)) {
1849 req->r_err = PTR_ERR(msg);
1850 complete_request(mdsc, req);
1851 return PTR_ERR(msg);
1852 }
1853 req->r_request = msg;
1854
1855 rhead = msg->front.iov_base;
1856 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1857 if (req->r_got_unsafe)
1858 flags |= CEPH_MDS_FLAG_REPLAY;
1859 if (req->r_locked_dir)
1860 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1861 rhead->flags = cpu_to_le32(flags);
1862 rhead->num_fwd = req->r_num_fwd;
1863 rhead->num_retry = req->r_attempts - 1;
1864 rhead->ino = 0;
1865
1866 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1867 return 0;
1868 }
1869
1870 /*
1871 * send request, or put it on the appropriate wait list.
1872 */
1873 static int __do_request(struct ceph_mds_client *mdsc,
1874 struct ceph_mds_request *req)
1875 {
1876 struct ceph_mds_session *session = NULL;
1877 int mds = -1;
1878 int err = -EAGAIN;
1879
1880 if (req->r_err || req->r_got_result) {
1881 if (req->r_aborted)
1882 __unregister_request(mdsc, req);
1883 goto out;
1884 }
1885
1886 if (req->r_timeout &&
1887 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1888 dout("do_request timed out\n");
1889 err = -EIO;
1890 goto finish;
1891 }
1892
1893 put_request_session(req);
1894
1895 mds = __choose_mds(mdsc, req);
1896 if (mds < 0 ||
1897 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1898 dout("do_request no mds or not active, waiting for map\n");
1899 list_add(&req->r_wait, &mdsc->waiting_for_map);
1900 goto out;
1901 }
1902
1903 /* get, open session */
1904 session = __ceph_lookup_mds_session(mdsc, mds);
1905 if (!session) {
1906 session = register_session(mdsc, mds);
1907 if (IS_ERR(session)) {
1908 err = PTR_ERR(session);
1909 goto finish;
1910 }
1911 }
1912 req->r_session = get_session(session);
1913
1914 dout("do_request mds%d session %p state %s\n", mds, session,
1915 session_state_name(session->s_state));
1916 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1917 session->s_state != CEPH_MDS_SESSION_HUNG) {
1918 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1919 session->s_state == CEPH_MDS_SESSION_CLOSING)
1920 __open_session(mdsc, session);
1921 list_add(&req->r_wait, &session->s_waiting);
1922 goto out_session;
1923 }
1924
1925 /* send request */
1926 req->r_resend_mds = -1; /* forget any previous mds hint */
1927
1928 if (req->r_request_started == 0) /* note request start time */
1929 req->r_request_started = jiffies;
1930
1931 err = __prepare_send_request(mdsc, req, mds);
1932 if (!err) {
1933 ceph_msg_get(req->r_request);
1934 ceph_con_send(&session->s_con, req->r_request);
1935 }
1936
1937 out_session:
1938 ceph_put_mds_session(session);
1939 out:
1940 return err;
1941
1942 finish:
1943 req->r_err = err;
1944 complete_request(mdsc, req);
1945 goto out;
1946 }
1947
1948 /*
1949 * called under mdsc->mutex
1950 */
1951 static void __wake_requests(struct ceph_mds_client *mdsc,
1952 struct list_head *head)
1953 {
1954 struct ceph_mds_request *req;
1955 LIST_HEAD(tmp_list);
1956
1957 list_splice_init(head, &tmp_list);
1958
1959 while (!list_empty(&tmp_list)) {
1960 req = list_entry(tmp_list.next,
1961 struct ceph_mds_request, r_wait);
1962 list_del_init(&req->r_wait);
1963 dout(" wake request %p tid %llu\n", req, req->r_tid);
1964 __do_request(mdsc, req);
1965 }
1966 }
1967
1968 /*
1969 * Wake up threads with requests pending for @mds, so that they can
1970 * resubmit their requests to a possibly different mds.
1971 */
1972 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1973 {
1974 struct ceph_mds_request *req;
1975 struct rb_node *p;
1976
1977 dout("kick_requests mds%d\n", mds);
1978 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1979 req = rb_entry(p, struct ceph_mds_request, r_node);
1980 if (req->r_got_unsafe)
1981 continue;
1982 if (req->r_session &&
1983 req->r_session->s_mds == mds) {
1984 dout(" kicking tid %llu\n", req->r_tid);
1985 __do_request(mdsc, req);
1986 }
1987 }
1988 }
1989
1990 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1991 struct ceph_mds_request *req)
1992 {
1993 dout("submit_request on %p\n", req);
1994 mutex_lock(&mdsc->mutex);
1995 __register_request(mdsc, req, NULL);
1996 __do_request(mdsc, req);
1997 mutex_unlock(&mdsc->mutex);
1998 }
1999
2000 /*
2001 * Synchrously perform an mds request. Take care of all of the
2002 * session setup, forwarding, retry details.
2003 */
2004 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2005 struct inode *dir,
2006 struct ceph_mds_request *req)
2007 {
2008 int err;
2009
2010 dout("do_request on %p\n", req);
2011
2012 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2013 if (req->r_inode)
2014 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2015 if (req->r_locked_dir)
2016 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2017 if (req->r_old_dentry)
2018 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2019 CEPH_CAP_PIN);
2020
2021 /* issue */
2022 mutex_lock(&mdsc->mutex);
2023 __register_request(mdsc, req, dir);
2024 __do_request(mdsc, req);
2025
2026 if (req->r_err) {
2027 err = req->r_err;
2028 __unregister_request(mdsc, req);
2029 dout("do_request early error %d\n", err);
2030 goto out;
2031 }
2032
2033 /* wait */
2034 mutex_unlock(&mdsc->mutex);
2035 dout("do_request waiting\n");
2036 if (req->r_timeout) {
2037 err = (long)wait_for_completion_killable_timeout(
2038 &req->r_completion, req->r_timeout);
2039 if (err == 0)
2040 err = -EIO;
2041 } else {
2042 err = wait_for_completion_killable(&req->r_completion);
2043 }
2044 dout("do_request waited, got %d\n", err);
2045 mutex_lock(&mdsc->mutex);
2046
2047 /* only abort if we didn't race with a real reply */
2048 if (req->r_got_result) {
2049 err = le32_to_cpu(req->r_reply_info.head->result);
2050 } else if (err < 0) {
2051 dout("aborted request %lld with %d\n", req->r_tid, err);
2052
2053 /*
2054 * ensure we aren't running concurrently with
2055 * ceph_fill_trace or ceph_readdir_prepopulate, which
2056 * rely on locks (dir mutex) held by our caller.
2057 */
2058 mutex_lock(&req->r_fill_mutex);
2059 req->r_err = err;
2060 req->r_aborted = true;
2061 mutex_unlock(&req->r_fill_mutex);
2062
2063 if (req->r_locked_dir &&
2064 (req->r_op & CEPH_MDS_OP_WRITE))
2065 ceph_invalidate_dir_request(req);
2066 } else {
2067 err = req->r_err;
2068 }
2069
2070 out:
2071 mutex_unlock(&mdsc->mutex);
2072 dout("do_request %p done, result %d\n", req, err);
2073 return err;
2074 }
2075
2076 /*
2077 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2078 * namespace request.
2079 */
2080 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2081 {
2082 struct inode *inode = req->r_locked_dir;
2083
2084 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2085
2086 ceph_dir_clear_complete(inode);
2087 if (req->r_dentry)
2088 ceph_invalidate_dentry_lease(req->r_dentry);
2089 if (req->r_old_dentry)
2090 ceph_invalidate_dentry_lease(req->r_old_dentry);
2091 }
2092
2093 /*
2094 * Handle mds reply.
2095 *
2096 * We take the session mutex and parse and process the reply immediately.
2097 * This preserves the logical ordering of replies, capabilities, etc., sent
2098 * by the MDS as they are applied to our local cache.
2099 */
2100 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2101 {
2102 struct ceph_mds_client *mdsc = session->s_mdsc;
2103 struct ceph_mds_request *req;
2104 struct ceph_mds_reply_head *head = msg->front.iov_base;
2105 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2106 u64 tid;
2107 int err, result;
2108 int mds = session->s_mds;
2109
2110 if (msg->front.iov_len < sizeof(*head)) {
2111 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2112 ceph_msg_dump(msg);
2113 return;
2114 }
2115
2116 /* get request, session */
2117 tid = le64_to_cpu(msg->hdr.tid);
2118 mutex_lock(&mdsc->mutex);
2119 req = __lookup_request(mdsc, tid);
2120 if (!req) {
2121 dout("handle_reply on unknown tid %llu\n", tid);
2122 mutex_unlock(&mdsc->mutex);
2123 return;
2124 }
2125 dout("handle_reply %p\n", req);
2126
2127 /* correct session? */
2128 if (req->r_session != session) {
2129 pr_err("mdsc_handle_reply got %llu on session mds%d"
2130 " not mds%d\n", tid, session->s_mds,
2131 req->r_session ? req->r_session->s_mds : -1);
2132 mutex_unlock(&mdsc->mutex);
2133 goto out;
2134 }
2135
2136 /* dup? */
2137 if ((req->r_got_unsafe && !head->safe) ||
2138 (req->r_got_safe && head->safe)) {
2139 pr_warning("got a dup %s reply on %llu from mds%d\n",
2140 head->safe ? "safe" : "unsafe", tid, mds);
2141 mutex_unlock(&mdsc->mutex);
2142 goto out;
2143 }
2144 if (req->r_got_safe && !head->safe) {
2145 pr_warning("got unsafe after safe on %llu from mds%d\n",
2146 tid, mds);
2147 mutex_unlock(&mdsc->mutex);
2148 goto out;
2149 }
2150
2151 result = le32_to_cpu(head->result);
2152
2153 /*
2154 * Handle an ESTALE
2155 * if we're not talking to the authority, send to them
2156 * if the authority has changed while we weren't looking,
2157 * send to new authority
2158 * Otherwise we just have to return an ESTALE
2159 */
2160 if (result == -ESTALE) {
2161 dout("got ESTALE on request %llu", req->r_tid);
2162 if (!req->r_inode) {
2163 /* do nothing; not an authority problem */
2164 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2165 dout("not using auth, setting for that now");
2166 req->r_direct_mode = USE_AUTH_MDS;
2167 __do_request(mdsc, req);
2168 mutex_unlock(&mdsc->mutex);
2169 goto out;
2170 } else {
2171 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2172 struct ceph_cap *cap = NULL;
2173
2174 if (req->r_session)
2175 cap = ceph_get_cap_for_mds(ci,
2176 req->r_session->s_mds);
2177
2178 dout("already using auth");
2179 if ((!cap || cap != ci->i_auth_cap) ||
2180 (cap->mseq != req->r_sent_on_mseq)) {
2181 dout("but cap changed, so resending");
2182 __do_request(mdsc, req);
2183 mutex_unlock(&mdsc->mutex);
2184 goto out;
2185 }
2186 }
2187 dout("have to return ESTALE on request %llu", req->r_tid);
2188 }
2189
2190
2191 if (head->safe) {
2192 req->r_got_safe = true;
2193 __unregister_request(mdsc, req);
2194
2195 if (req->r_got_unsafe) {
2196 /*
2197 * We already handled the unsafe response, now do the
2198 * cleanup. No need to examine the response; the MDS
2199 * doesn't include any result info in the safe
2200 * response. And even if it did, there is nothing
2201 * useful we could do with a revised return value.
2202 */
2203 dout("got safe reply %llu, mds%d\n", tid, mds);
2204 list_del_init(&req->r_unsafe_item);
2205
2206 /* last unsafe request during umount? */
2207 if (mdsc->stopping && !__get_oldest_req(mdsc))
2208 complete_all(&mdsc->safe_umount_waiters);
2209 mutex_unlock(&mdsc->mutex);
2210 goto out;
2211 }
2212 } else {
2213 req->r_got_unsafe = true;
2214 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2215 }
2216
2217 dout("handle_reply tid %lld result %d\n", tid, result);
2218 rinfo = &req->r_reply_info;
2219 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2220 mutex_unlock(&mdsc->mutex);
2221
2222 mutex_lock(&session->s_mutex);
2223 if (err < 0) {
2224 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2225 ceph_msg_dump(msg);
2226 goto out_err;
2227 }
2228
2229 /* snap trace */
2230 if (rinfo->snapblob_len) {
2231 down_write(&mdsc->snap_rwsem);
2232 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2233 rinfo->snapblob + rinfo->snapblob_len,
2234 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2235 downgrade_write(&mdsc->snap_rwsem);
2236 } else {
2237 down_read(&mdsc->snap_rwsem);
2238 }
2239
2240 /* insert trace into our cache */
2241 mutex_lock(&req->r_fill_mutex);
2242 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2243 if (err == 0) {
2244 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2245 req->r_op == CEPH_MDS_OP_LSSNAP) &&
2246 rinfo->dir_nr)
2247 ceph_readdir_prepopulate(req, req->r_session);
2248 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2249 }
2250 mutex_unlock(&req->r_fill_mutex);
2251
2252 up_read(&mdsc->snap_rwsem);
2253 out_err:
2254 mutex_lock(&mdsc->mutex);
2255 if (!req->r_aborted) {
2256 if (err) {
2257 req->r_err = err;
2258 } else {
2259 req->r_reply = msg;
2260 ceph_msg_get(msg);
2261 req->r_got_result = true;
2262 }
2263 } else {
2264 dout("reply arrived after request %lld was aborted\n", tid);
2265 }
2266 mutex_unlock(&mdsc->mutex);
2267
2268 ceph_add_cap_releases(mdsc, req->r_session);
2269 mutex_unlock(&session->s_mutex);
2270
2271 /* kick calling process */
2272 complete_request(mdsc, req);
2273 out:
2274 ceph_mdsc_put_request(req);
2275 return;
2276 }
2277
2278
2279
2280 /*
2281 * handle mds notification that our request has been forwarded.
2282 */
2283 static void handle_forward(struct ceph_mds_client *mdsc,
2284 struct ceph_mds_session *session,
2285 struct ceph_msg *msg)
2286 {
2287 struct ceph_mds_request *req;
2288 u64 tid = le64_to_cpu(msg->hdr.tid);
2289 u32 next_mds;
2290 u32 fwd_seq;
2291 int err = -EINVAL;
2292 void *p = msg->front.iov_base;
2293 void *end = p + msg->front.iov_len;
2294
2295 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2296 next_mds = ceph_decode_32(&p);
2297 fwd_seq = ceph_decode_32(&p);
2298
2299 mutex_lock(&mdsc->mutex);
2300 req = __lookup_request(mdsc, tid);
2301 if (!req) {
2302 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2303 goto out; /* dup reply? */
2304 }
2305
2306 if (req->r_aborted) {
2307 dout("forward tid %llu aborted, unregistering\n", tid);
2308 __unregister_request(mdsc, req);
2309 } else if (fwd_seq <= req->r_num_fwd) {
2310 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2311 tid, next_mds, req->r_num_fwd, fwd_seq);
2312 } else {
2313 /* resend. forward race not possible; mds would drop */
2314 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2315 BUG_ON(req->r_err);
2316 BUG_ON(req->r_got_result);
2317 req->r_num_fwd = fwd_seq;
2318 req->r_resend_mds = next_mds;
2319 put_request_session(req);
2320 __do_request(mdsc, req);
2321 }
2322 ceph_mdsc_put_request(req);
2323 out:
2324 mutex_unlock(&mdsc->mutex);
2325 return;
2326
2327 bad:
2328 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2329 }
2330
2331 /*
2332 * handle a mds session control message
2333 */
2334 static void handle_session(struct ceph_mds_session *session,
2335 struct ceph_msg *msg)
2336 {
2337 struct ceph_mds_client *mdsc = session->s_mdsc;
2338 u32 op;
2339 u64 seq;
2340 int mds = session->s_mds;
2341 struct ceph_mds_session_head *h = msg->front.iov_base;
2342 int wake = 0;
2343
2344 /* decode */
2345 if (msg->front.iov_len != sizeof(*h))
2346 goto bad;
2347 op = le32_to_cpu(h->op);
2348 seq = le64_to_cpu(h->seq);
2349
2350 mutex_lock(&mdsc->mutex);
2351 if (op == CEPH_SESSION_CLOSE)
2352 __unregister_session(mdsc, session);
2353 /* FIXME: this ttl calculation is generous */
2354 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2355 mutex_unlock(&mdsc->mutex);
2356
2357 mutex_lock(&session->s_mutex);
2358
2359 dout("handle_session mds%d %s %p state %s seq %llu\n",
2360 mds, ceph_session_op_name(op), session,
2361 session_state_name(session->s_state), seq);
2362
2363 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2364 session->s_state = CEPH_MDS_SESSION_OPEN;
2365 pr_info("mds%d came back\n", session->s_mds);
2366 }
2367
2368 switch (op) {
2369 case CEPH_SESSION_OPEN:
2370 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2371 pr_info("mds%d reconnect success\n", session->s_mds);
2372 session->s_state = CEPH_MDS_SESSION_OPEN;
2373 renewed_caps(mdsc, session, 0);
2374 wake = 1;
2375 if (mdsc->stopping)
2376 __close_session(mdsc, session);
2377 break;
2378
2379 case CEPH_SESSION_RENEWCAPS:
2380 if (session->s_renew_seq == seq)
2381 renewed_caps(mdsc, session, 1);
2382 break;
2383
2384 case CEPH_SESSION_CLOSE:
2385 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2386 pr_info("mds%d reconnect denied\n", session->s_mds);
2387 remove_session_caps(session);
2388 wake = 1; /* for good measure */
2389 wake_up_all(&mdsc->session_close_wq);
2390 kick_requests(mdsc, mds);
2391 break;
2392
2393 case CEPH_SESSION_STALE:
2394 pr_info("mds%d caps went stale, renewing\n",
2395 session->s_mds);
2396 spin_lock(&session->s_gen_ttl_lock);
2397 session->s_cap_gen++;
2398 session->s_cap_ttl = jiffies - 1;
2399 spin_unlock(&session->s_gen_ttl_lock);
2400 send_renew_caps(mdsc, session);
2401 break;
2402
2403 case CEPH_SESSION_RECALL_STATE:
2404 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2405 break;
2406
2407 default:
2408 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2409 WARN_ON(1);
2410 }
2411
2412 mutex_unlock(&session->s_mutex);
2413 if (wake) {
2414 mutex_lock(&mdsc->mutex);
2415 __wake_requests(mdsc, &session->s_waiting);
2416 mutex_unlock(&mdsc->mutex);
2417 }
2418 return;
2419
2420 bad:
2421 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2422 (int)msg->front.iov_len);
2423 ceph_msg_dump(msg);
2424 return;
2425 }
2426
2427
2428 /*
2429 * called under session->mutex.
2430 */
2431 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2432 struct ceph_mds_session *session)
2433 {
2434 struct ceph_mds_request *req, *nreq;
2435 int err;
2436
2437 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2438
2439 mutex_lock(&mdsc->mutex);
2440 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2441 err = __prepare_send_request(mdsc, req, session->s_mds);
2442 if (!err) {
2443 ceph_msg_get(req->r_request);
2444 ceph_con_send(&session->s_con, req->r_request);
2445 }
2446 }
2447 mutex_unlock(&mdsc->mutex);
2448 }
2449
2450 /*
2451 * Encode information about a cap for a reconnect with the MDS.
2452 */
2453 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2454 void *arg)
2455 {
2456 union {
2457 struct ceph_mds_cap_reconnect v2;
2458 struct ceph_mds_cap_reconnect_v1 v1;
2459 } rec;
2460 size_t reclen;
2461 struct ceph_inode_info *ci;
2462 struct ceph_reconnect_state *recon_state = arg;
2463 struct ceph_pagelist *pagelist = recon_state->pagelist;
2464 char *path;
2465 int pathlen, err;
2466 u64 pathbase;
2467 struct dentry *dentry;
2468
2469 ci = cap->ci;
2470
2471 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2472 inode, ceph_vinop(inode), cap, cap->cap_id,
2473 ceph_cap_string(cap->issued));
2474 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2475 if (err)
2476 return err;
2477
2478 dentry = d_find_alias(inode);
2479 if (dentry) {
2480 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2481 if (IS_ERR(path)) {
2482 err = PTR_ERR(path);
2483 goto out_dput;
2484 }
2485 } else {
2486 path = NULL;
2487 pathlen = 0;
2488 }
2489 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2490 if (err)
2491 goto out_free;
2492
2493 spin_lock(&ci->i_ceph_lock);
2494 cap->seq = 0; /* reset cap seq */
2495 cap->issue_seq = 0; /* and issue_seq */
2496 cap->mseq = 0; /* and migrate_seq */
2497
2498 if (recon_state->flock) {
2499 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2500 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2501 rec.v2.issued = cpu_to_le32(cap->issued);
2502 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2503 rec.v2.pathbase = cpu_to_le64(pathbase);
2504 rec.v2.flock_len = 0;
2505 reclen = sizeof(rec.v2);
2506 } else {
2507 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2508 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2509 rec.v1.issued = cpu_to_le32(cap->issued);
2510 rec.v1.size = cpu_to_le64(inode->i_size);
2511 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2512 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2513 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2514 rec.v1.pathbase = cpu_to_le64(pathbase);
2515 reclen = sizeof(rec.v1);
2516 }
2517 spin_unlock(&ci->i_ceph_lock);
2518
2519 if (recon_state->flock) {
2520 int num_fcntl_locks, num_flock_locks;
2521 struct ceph_filelock *flocks;
2522
2523 encode_again:
2524 spin_lock(&inode->i_lock);
2525 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2526 spin_unlock(&inode->i_lock);
2527 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2528 sizeof(struct ceph_filelock), GFP_NOFS);
2529 if (!flocks) {
2530 err = -ENOMEM;
2531 goto out_free;
2532 }
2533 spin_lock(&inode->i_lock);
2534 err = ceph_encode_locks_to_buffer(inode, flocks,
2535 num_fcntl_locks,
2536 num_flock_locks);
2537 spin_unlock(&inode->i_lock);
2538 if (err) {
2539 kfree(flocks);
2540 if (err == -ENOSPC)
2541 goto encode_again;
2542 goto out_free;
2543 }
2544 /*
2545 * number of encoded locks is stable, so copy to pagelist
2546 */
2547 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2548 (num_fcntl_locks+num_flock_locks) *
2549 sizeof(struct ceph_filelock));
2550 err = ceph_pagelist_append(pagelist, &rec, reclen);
2551 if (!err)
2552 err = ceph_locks_to_pagelist(flocks, pagelist,
2553 num_fcntl_locks,
2554 num_flock_locks);
2555 kfree(flocks);
2556 } else {
2557 err = ceph_pagelist_append(pagelist, &rec, reclen);
2558 }
2559 out_free:
2560 kfree(path);
2561 out_dput:
2562 dput(dentry);
2563 return err;
2564 }
2565
2566
2567 /*
2568 * If an MDS fails and recovers, clients need to reconnect in order to
2569 * reestablish shared state. This includes all caps issued through
2570 * this session _and_ the snap_realm hierarchy. Because it's not
2571 * clear which snap realms the mds cares about, we send everything we
2572 * know about.. that ensures we'll then get any new info the
2573 * recovering MDS might have.
2574 *
2575 * This is a relatively heavyweight operation, but it's rare.
2576 *
2577 * called with mdsc->mutex held.
2578 */
2579 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2580 struct ceph_mds_session *session)
2581 {
2582 struct ceph_msg *reply;
2583 struct rb_node *p;
2584 int mds = session->s_mds;
2585 int err = -ENOMEM;
2586 struct ceph_pagelist *pagelist;
2587 struct ceph_reconnect_state recon_state;
2588
2589 pr_info("mds%d reconnect start\n", mds);
2590
2591 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2592 if (!pagelist)
2593 goto fail_nopagelist;
2594 ceph_pagelist_init(pagelist);
2595
2596 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2597 if (!reply)
2598 goto fail_nomsg;
2599
2600 mutex_lock(&session->s_mutex);
2601 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2602 session->s_seq = 0;
2603
2604 ceph_con_close(&session->s_con);
2605 ceph_con_open(&session->s_con,
2606 CEPH_ENTITY_TYPE_MDS, mds,
2607 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2608
2609 /* replay unsafe requests */
2610 replay_unsafe_requests(mdsc, session);
2611
2612 down_read(&mdsc->snap_rwsem);
2613
2614 dout("session %p state %s\n", session,
2615 session_state_name(session->s_state));
2616
2617 /* drop old cap expires; we're about to reestablish that state */
2618 discard_cap_releases(mdsc, session);
2619
2620 /* traverse this session's caps */
2621 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2622 if (err)
2623 goto fail;
2624
2625 recon_state.pagelist = pagelist;
2626 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2627 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2628 if (err < 0)
2629 goto fail;
2630
2631 /*
2632 * snaprealms. we provide mds with the ino, seq (version), and
2633 * parent for all of our realms. If the mds has any newer info,
2634 * it will tell us.
2635 */
2636 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2637 struct ceph_snap_realm *realm =
2638 rb_entry(p, struct ceph_snap_realm, node);
2639 struct ceph_mds_snaprealm_reconnect sr_rec;
2640
2641 dout(" adding snap realm %llx seq %lld parent %llx\n",
2642 realm->ino, realm->seq, realm->parent_ino);
2643 sr_rec.ino = cpu_to_le64(realm->ino);
2644 sr_rec.seq = cpu_to_le64(realm->seq);
2645 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2646 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2647 if (err)
2648 goto fail;
2649 }
2650
2651 if (recon_state.flock)
2652 reply->hdr.version = cpu_to_le16(2);
2653 if (pagelist->length) {
2654 /* set up outbound data if we have any */
2655 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2656 ceph_msg_data_add_pagelist(reply, pagelist);
2657 }
2658 ceph_con_send(&session->s_con, reply);
2659
2660 mutex_unlock(&session->s_mutex);
2661
2662 mutex_lock(&mdsc->mutex);
2663 __wake_requests(mdsc, &session->s_waiting);
2664 mutex_unlock(&mdsc->mutex);
2665
2666 up_read(&mdsc->snap_rwsem);
2667 return;
2668
2669 fail:
2670 ceph_msg_put(reply);
2671 up_read(&mdsc->snap_rwsem);
2672 mutex_unlock(&session->s_mutex);
2673 fail_nomsg:
2674 ceph_pagelist_release(pagelist);
2675 kfree(pagelist);
2676 fail_nopagelist:
2677 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2678 return;
2679 }
2680
2681
2682 /*
2683 * compare old and new mdsmaps, kicking requests
2684 * and closing out old connections as necessary
2685 *
2686 * called under mdsc->mutex.
2687 */
2688 static void check_new_map(struct ceph_mds_client *mdsc,
2689 struct ceph_mdsmap *newmap,
2690 struct ceph_mdsmap *oldmap)
2691 {
2692 int i;
2693 int oldstate, newstate;
2694 struct ceph_mds_session *s;
2695
2696 dout("check_new_map new %u old %u\n",
2697 newmap->m_epoch, oldmap->m_epoch);
2698
2699 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2700 if (mdsc->sessions[i] == NULL)
2701 continue;
2702 s = mdsc->sessions[i];
2703 oldstate = ceph_mdsmap_get_state(oldmap, i);
2704 newstate = ceph_mdsmap_get_state(newmap, i);
2705
2706 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2707 i, ceph_mds_state_name(oldstate),
2708 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2709 ceph_mds_state_name(newstate),
2710 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2711 session_state_name(s->s_state));
2712
2713 if (i >= newmap->m_max_mds ||
2714 memcmp(ceph_mdsmap_get_addr(oldmap, i),
2715 ceph_mdsmap_get_addr(newmap, i),
2716 sizeof(struct ceph_entity_addr))) {
2717 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2718 /* the session never opened, just close it
2719 * out now */
2720 __wake_requests(mdsc, &s->s_waiting);
2721 __unregister_session(mdsc, s);
2722 } else {
2723 /* just close it */
2724 mutex_unlock(&mdsc->mutex);
2725 mutex_lock(&s->s_mutex);
2726 mutex_lock(&mdsc->mutex);
2727 ceph_con_close(&s->s_con);
2728 mutex_unlock(&s->s_mutex);
2729 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2730 }
2731
2732 /* kick any requests waiting on the recovering mds */
2733 kick_requests(mdsc, i);
2734 } else if (oldstate == newstate) {
2735 continue; /* nothing new with this mds */
2736 }
2737
2738 /*
2739 * send reconnect?
2740 */
2741 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2742 newstate >= CEPH_MDS_STATE_RECONNECT) {
2743 mutex_unlock(&mdsc->mutex);
2744 send_mds_reconnect(mdsc, s);
2745 mutex_lock(&mdsc->mutex);
2746 }
2747
2748 /*
2749 * kick request on any mds that has gone active.
2750 */
2751 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2752 newstate >= CEPH_MDS_STATE_ACTIVE) {
2753 if (oldstate != CEPH_MDS_STATE_CREATING &&
2754 oldstate != CEPH_MDS_STATE_STARTING)
2755 pr_info("mds%d recovery completed\n", s->s_mds);
2756 kick_requests(mdsc, i);
2757 ceph_kick_flushing_caps(mdsc, s);
2758 wake_up_session_caps(s, 1);
2759 }
2760 }
2761
2762 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2763 s = mdsc->sessions[i];
2764 if (!s)
2765 continue;
2766 if (!ceph_mdsmap_is_laggy(newmap, i))
2767 continue;
2768 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2769 s->s_state == CEPH_MDS_SESSION_HUNG ||
2770 s->s_state == CEPH_MDS_SESSION_CLOSING) {
2771 dout(" connecting to export targets of laggy mds%d\n",
2772 i);
2773 __open_export_target_sessions(mdsc, s);
2774 }
2775 }
2776 }
2777
2778
2779
2780 /*
2781 * leases
2782 */
2783
2784 /*
2785 * caller must hold session s_mutex, dentry->d_lock
2786 */
2787 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2788 {
2789 struct ceph_dentry_info *di = ceph_dentry(dentry);
2790
2791 ceph_put_mds_session(di->lease_session);
2792 di->lease_session = NULL;
2793 }
2794
2795 static void handle_lease(struct ceph_mds_client *mdsc,
2796 struct ceph_mds_session *session,
2797 struct ceph_msg *msg)
2798 {
2799 struct super_block *sb = mdsc->fsc->sb;
2800 struct inode *inode;
2801 struct dentry *parent, *dentry;
2802 struct ceph_dentry_info *di;
2803 int mds = session->s_mds;
2804 struct ceph_mds_lease *h = msg->front.iov_base;
2805 u32 seq;
2806 struct ceph_vino vino;
2807 struct qstr dname;
2808 int release = 0;
2809
2810 dout("handle_lease from mds%d\n", mds);
2811
2812 /* decode */
2813 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2814 goto bad;
2815 vino.ino = le64_to_cpu(h->ino);
2816 vino.snap = CEPH_NOSNAP;
2817 seq = le32_to_cpu(h->seq);
2818 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2819 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2820 if (dname.len != get_unaligned_le32(h+1))
2821 goto bad;
2822
2823 mutex_lock(&session->s_mutex);
2824 session->s_seq++;
2825
2826 /* lookup inode */
2827 inode = ceph_find_inode(sb, vino);
2828 dout("handle_lease %s, ino %llx %p %.*s\n",
2829 ceph_lease_op_name(h->action), vino.ino, inode,
2830 dname.len, dname.name);
2831 if (inode == NULL) {
2832 dout("handle_lease no inode %llx\n", vino.ino);
2833 goto release;
2834 }
2835
2836 /* dentry */
2837 parent = d_find_alias(inode);
2838 if (!parent) {
2839 dout("no parent dentry on inode %p\n", inode);
2840 WARN_ON(1);
2841 goto release; /* hrm... */
2842 }
2843 dname.hash = full_name_hash(dname.name, dname.len);
2844 dentry = d_lookup(parent, &dname);
2845 dput(parent);
2846 if (!dentry)
2847 goto release;
2848
2849 spin_lock(&dentry->d_lock);
2850 di = ceph_dentry(dentry);
2851 switch (h->action) {
2852 case CEPH_MDS_LEASE_REVOKE:
2853 if (di->lease_session == session) {
2854 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2855 h->seq = cpu_to_le32(di->lease_seq);
2856 __ceph_mdsc_drop_dentry_lease(dentry);
2857 }
2858 release = 1;
2859 break;
2860
2861 case CEPH_MDS_LEASE_RENEW:
2862 if (di->lease_session == session &&
2863 di->lease_gen == session->s_cap_gen &&
2864 di->lease_renew_from &&
2865 di->lease_renew_after == 0) {
2866 unsigned long duration =
2867 le32_to_cpu(h->duration_ms) * HZ / 1000;
2868
2869 di->lease_seq = seq;
2870 dentry->d_time = di->lease_renew_from + duration;
2871 di->lease_renew_after = di->lease_renew_from +
2872 (duration >> 1);
2873 di->lease_renew_from = 0;
2874 }
2875 break;
2876 }
2877 spin_unlock(&dentry->d_lock);
2878 dput(dentry);
2879
2880 if (!release)
2881 goto out;
2882
2883 release:
2884 /* let's just reuse the same message */
2885 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2886 ceph_msg_get(msg);
2887 ceph_con_send(&session->s_con, msg);
2888
2889 out:
2890 iput(inode);
2891 mutex_unlock(&session->s_mutex);
2892 return;
2893
2894 bad:
2895 pr_err("corrupt lease message\n");
2896 ceph_msg_dump(msg);
2897 }
2898
2899 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2900 struct inode *inode,
2901 struct dentry *dentry, char action,
2902 u32 seq)
2903 {
2904 struct ceph_msg *msg;
2905 struct ceph_mds_lease *lease;
2906 int len = sizeof(*lease) + sizeof(u32);
2907 int dnamelen = 0;
2908
2909 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2910 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2911 dnamelen = dentry->d_name.len;
2912 len += dnamelen;
2913
2914 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2915 if (!msg)
2916 return;
2917 lease = msg->front.iov_base;
2918 lease->action = action;
2919 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2920 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2921 lease->seq = cpu_to_le32(seq);
2922 put_unaligned_le32(dnamelen, lease + 1);
2923 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2924
2925 /*
2926 * if this is a preemptive lease RELEASE, no need to
2927 * flush request stream, since the actual request will
2928 * soon follow.
2929 */
2930 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2931
2932 ceph_con_send(&session->s_con, msg);
2933 }
2934
2935 /*
2936 * Preemptively release a lease we expect to invalidate anyway.
2937 * Pass @inode always, @dentry is optional.
2938 */
2939 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2940 struct dentry *dentry)
2941 {
2942 struct ceph_dentry_info *di;
2943 struct ceph_mds_session *session;
2944 u32 seq;
2945
2946 BUG_ON(inode == NULL);
2947 BUG_ON(dentry == NULL);
2948
2949 /* is dentry lease valid? */
2950 spin_lock(&dentry->d_lock);
2951 di = ceph_dentry(dentry);
2952 if (!di || !di->lease_session ||
2953 di->lease_session->s_mds < 0 ||
2954 di->lease_gen != di->lease_session->s_cap_gen ||
2955 !time_before(jiffies, dentry->d_time)) {
2956 dout("lease_release inode %p dentry %p -- "
2957 "no lease\n",
2958 inode, dentry);
2959 spin_unlock(&dentry->d_lock);
2960 return;
2961 }
2962
2963 /* we do have a lease on this dentry; note mds and seq */
2964 session = ceph_get_mds_session(di->lease_session);
2965 seq = di->lease_seq;
2966 __ceph_mdsc_drop_dentry_lease(dentry);
2967 spin_unlock(&dentry->d_lock);
2968
2969 dout("lease_release inode %p dentry %p to mds%d\n",
2970 inode, dentry, session->s_mds);
2971 ceph_mdsc_lease_send_msg(session, inode, dentry,
2972 CEPH_MDS_LEASE_RELEASE, seq);
2973 ceph_put_mds_session(session);
2974 }
2975
2976 /*
2977 * drop all leases (and dentry refs) in preparation for umount
2978 */
2979 static void drop_leases(struct ceph_mds_client *mdsc)
2980 {
2981 int i;
2982
2983 dout("drop_leases\n");
2984 mutex_lock(&mdsc->mutex);
2985 for (i = 0; i < mdsc->max_sessions; i++) {
2986 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2987 if (!s)
2988 continue;
2989 mutex_unlock(&mdsc->mutex);
2990 mutex_lock(&s->s_mutex);
2991 mutex_unlock(&s->s_mutex);
2992 ceph_put_mds_session(s);
2993 mutex_lock(&mdsc->mutex);
2994 }
2995 mutex_unlock(&mdsc->mutex);
2996 }
2997
2998
2999
3000 /*
3001 * delayed work -- periodically trim expired leases, renew caps with mds
3002 */
3003 static void schedule_delayed(struct ceph_mds_client *mdsc)
3004 {
3005 int delay = 5;
3006 unsigned hz = round_jiffies_relative(HZ * delay);
3007 schedule_delayed_work(&mdsc->delayed_work, hz);
3008 }
3009
3010 static void delayed_work(struct work_struct *work)
3011 {
3012 int i;
3013 struct ceph_mds_client *mdsc =
3014 container_of(work, struct ceph_mds_client, delayed_work.work);
3015 int renew_interval;
3016 int renew_caps;
3017
3018 dout("mdsc delayed_work\n");
3019 ceph_check_delayed_caps(mdsc);
3020
3021 mutex_lock(&mdsc->mutex);
3022 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3023 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3024 mdsc->last_renew_caps);
3025 if (renew_caps)
3026 mdsc->last_renew_caps = jiffies;
3027
3028 for (i = 0; i < mdsc->max_sessions; i++) {
3029 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3030 if (s == NULL)
3031 continue;
3032 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3033 dout("resending session close request for mds%d\n",
3034 s->s_mds);
3035 request_close_session(mdsc, s);
3036 ceph_put_mds_session(s);
3037 continue;
3038 }
3039 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3040 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3041 s->s_state = CEPH_MDS_SESSION_HUNG;
3042 pr_info("mds%d hung\n", s->s_mds);
3043 }
3044 }
3045 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3046 /* this mds is failed or recovering, just wait */
3047 ceph_put_mds_session(s);
3048 continue;
3049 }
3050 mutex_unlock(&mdsc->mutex);
3051
3052 mutex_lock(&s->s_mutex);
3053 if (renew_caps)
3054 send_renew_caps(mdsc, s);
3055 else
3056 ceph_con_keepalive(&s->s_con);
3057 ceph_add_cap_releases(mdsc, s);
3058 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3059 s->s_state == CEPH_MDS_SESSION_HUNG)
3060 ceph_send_cap_releases(mdsc, s);
3061 mutex_unlock(&s->s_mutex);
3062 ceph_put_mds_session(s);
3063
3064 mutex_lock(&mdsc->mutex);
3065 }
3066 mutex_unlock(&mdsc->mutex);
3067
3068 schedule_delayed(mdsc);
3069 }
3070
3071 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3072
3073 {
3074 struct ceph_mds_client *mdsc;
3075
3076 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3077 if (!mdsc)
3078 return -ENOMEM;
3079 mdsc->fsc = fsc;
3080 fsc->mdsc = mdsc;
3081 mutex_init(&mdsc->mutex);
3082 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3083 if (mdsc->mdsmap == NULL) {
3084 kfree(mdsc);
3085 return -ENOMEM;
3086 }
3087
3088 init_completion(&mdsc->safe_umount_waiters);
3089 init_waitqueue_head(&mdsc->session_close_wq);
3090 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3091 mdsc->sessions = NULL;
3092 mdsc->max_sessions = 0;
3093 mdsc->stopping = 0;
3094 init_rwsem(&mdsc->snap_rwsem);
3095 mdsc->snap_realms = RB_ROOT;
3096 INIT_LIST_HEAD(&mdsc->snap_empty);
3097 spin_lock_init(&mdsc->snap_empty_lock);
3098 mdsc->last_tid = 0;
3099 mdsc->request_tree = RB_ROOT;
3100 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3101 mdsc->last_renew_caps = jiffies;
3102 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3103 spin_lock_init(&mdsc->cap_delay_lock);
3104 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3105 spin_lock_init(&mdsc->snap_flush_lock);
3106 mdsc->cap_flush_seq = 0;
3107 INIT_LIST_HEAD(&mdsc->cap_dirty);
3108 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3109 mdsc->num_cap_flushing = 0;
3110 spin_lock_init(&mdsc->cap_dirty_lock);
3111 init_waitqueue_head(&mdsc->cap_flushing_wq);
3112 spin_lock_init(&mdsc->dentry_lru_lock);
3113 INIT_LIST_HEAD(&mdsc->dentry_lru);
3114
3115 ceph_caps_init(mdsc);
3116 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3117
3118 return 0;
3119 }
3120
3121 /*
3122 * Wait for safe replies on open mds requests. If we time out, drop
3123 * all requests from the tree to avoid dangling dentry refs.
3124 */
3125 static void wait_requests(struct ceph_mds_client *mdsc)
3126 {
3127 struct ceph_mds_request *req;
3128 struct ceph_fs_client *fsc = mdsc->fsc;
3129
3130 mutex_lock(&mdsc->mutex);
3131 if (__get_oldest_req(mdsc)) {
3132 mutex_unlock(&mdsc->mutex);
3133
3134 dout("wait_requests waiting for requests\n");
3135 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3136 fsc->client->options->mount_timeout * HZ);
3137
3138 /* tear down remaining requests */
3139 mutex_lock(&mdsc->mutex);
3140 while ((req = __get_oldest_req(mdsc))) {
3141 dout("wait_requests timed out on tid %llu\n",
3142 req->r_tid);
3143 __unregister_request(mdsc, req);
3144 }
3145 }
3146 mutex_unlock(&mdsc->mutex);
3147 dout("wait_requests done\n");
3148 }
3149
3150 /*
3151 * called before mount is ro, and before dentries are torn down.
3152 * (hmm, does this still race with new lookups?)
3153 */
3154 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3155 {
3156 dout("pre_umount\n");
3157 mdsc->stopping = 1;
3158
3159 drop_leases(mdsc);
3160 ceph_flush_dirty_caps(mdsc);
3161 wait_requests(mdsc);
3162
3163 /*
3164 * wait for reply handlers to drop their request refs and
3165 * their inode/dcache refs
3166 */
3167 ceph_msgr_flush();
3168 }
3169
3170 /*
3171 * wait for all write mds requests to flush.
3172 */
3173 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3174 {
3175 struct ceph_mds_request *req = NULL, *nextreq;
3176 struct rb_node *n;
3177
3178 mutex_lock(&mdsc->mutex);
3179 dout("wait_unsafe_requests want %lld\n", want_tid);
3180 restart:
3181 req = __get_oldest_req(mdsc);
3182 while (req && req->r_tid <= want_tid) {
3183 /* find next request */
3184 n = rb_next(&req->r_node);
3185 if (n)
3186 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3187 else
3188 nextreq = NULL;
3189 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3190 /* write op */
3191 ceph_mdsc_get_request(req);
3192 if (nextreq)
3193 ceph_mdsc_get_request(nextreq);
3194 mutex_unlock(&mdsc->mutex);
3195 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3196 req->r_tid, want_tid);
3197 wait_for_completion(&req->r_safe_completion);
3198 mutex_lock(&mdsc->mutex);
3199 ceph_mdsc_put_request(req);
3200 if (!nextreq)
3201 break; /* next dne before, so we're done! */
3202 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3203 /* next request was removed from tree */
3204 ceph_mdsc_put_request(nextreq);
3205 goto restart;
3206 }
3207 ceph_mdsc_put_request(nextreq); /* won't go away */
3208 }
3209 req = nextreq;
3210 }
3211 mutex_unlock(&mdsc->mutex);
3212 dout("wait_unsafe_requests done\n");
3213 }
3214
3215 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3216 {
3217 u64 want_tid, want_flush;
3218
3219 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3220 return;
3221
3222 dout("sync\n");
3223 mutex_lock(&mdsc->mutex);
3224 want_tid = mdsc->last_tid;
3225 want_flush = mdsc->cap_flush_seq;
3226 mutex_unlock(&mdsc->mutex);
3227 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3228
3229 ceph_flush_dirty_caps(mdsc);
3230
3231 wait_unsafe_requests(mdsc, want_tid);
3232 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3233 }
3234
3235 /*
3236 * true if all sessions are closed, or we force unmount
3237 */
3238 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3239 {
3240 int i, n = 0;
3241
3242 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3243 return true;
3244
3245 mutex_lock(&mdsc->mutex);
3246 for (i = 0; i < mdsc->max_sessions; i++)
3247 if (mdsc->sessions[i])
3248 n++;
3249 mutex_unlock(&mdsc->mutex);
3250 return n == 0;
3251 }
3252
3253 /*
3254 * called after sb is ro.
3255 */
3256 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3257 {
3258 struct ceph_mds_session *session;
3259 int i;
3260 struct ceph_fs_client *fsc = mdsc->fsc;
3261 unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3262
3263 dout("close_sessions\n");
3264
3265 /* close sessions */
3266 mutex_lock(&mdsc->mutex);
3267 for (i = 0; i < mdsc->max_sessions; i++) {
3268 session = __ceph_lookup_mds_session(mdsc, i);
3269 if (!session)
3270 continue;
3271 mutex_unlock(&mdsc->mutex);
3272 mutex_lock(&session->s_mutex);
3273 __close_session(mdsc, session);
3274 mutex_unlock(&session->s_mutex);
3275 ceph_put_mds_session(session);
3276 mutex_lock(&mdsc->mutex);
3277 }
3278 mutex_unlock(&mdsc->mutex);
3279
3280 dout("waiting for sessions to close\n");
3281 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3282 timeout);
3283
3284 /* tear down remaining sessions */
3285 mutex_lock(&mdsc->mutex);
3286 for (i = 0; i < mdsc->max_sessions; i++) {
3287 if (mdsc->sessions[i]) {
3288 session = get_session(mdsc->sessions[i]);
3289 __unregister_session(mdsc, session);
3290 mutex_unlock(&mdsc->mutex);
3291 mutex_lock(&session->s_mutex);
3292 remove_session_caps(session);
3293 mutex_unlock(&session->s_mutex);
3294 ceph_put_mds_session(session);
3295 mutex_lock(&mdsc->mutex);
3296 }
3297 }
3298 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3299 mutex_unlock(&mdsc->mutex);
3300
3301 ceph_cleanup_empty_realms(mdsc);
3302
3303 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3304
3305 dout("stopped\n");
3306 }
3307
3308 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3309 {
3310 dout("stop\n");
3311 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3312 if (mdsc->mdsmap)
3313 ceph_mdsmap_destroy(mdsc->mdsmap);
3314 kfree(mdsc->sessions);
3315 ceph_caps_finalize(mdsc);
3316 }
3317
3318 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3319 {
3320 struct ceph_mds_client *mdsc = fsc->mdsc;
3321
3322 dout("mdsc_destroy %p\n", mdsc);
3323 ceph_mdsc_stop(mdsc);
3324
3325 /* flush out any connection work with references to us */
3326 ceph_msgr_flush();
3327
3328 fsc->mdsc = NULL;
3329 kfree(mdsc);
3330 dout("mdsc_destroy %p done\n", mdsc);
3331 }
3332
3333
3334 /*
3335 * handle mds map update.
3336 */
3337 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3338 {
3339 u32 epoch;
3340 u32 maplen;
3341 void *p = msg->front.iov_base;
3342 void *end = p + msg->front.iov_len;
3343 struct ceph_mdsmap *newmap, *oldmap;
3344 struct ceph_fsid fsid;
3345 int err = -EINVAL;
3346
3347 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3348 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3349 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3350 return;
3351 epoch = ceph_decode_32(&p);
3352 maplen = ceph_decode_32(&p);
3353 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3354
3355 /* do we need it? */
3356 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3357 mutex_lock(&mdsc->mutex);
3358 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3359 dout("handle_map epoch %u <= our %u\n",
3360 epoch, mdsc->mdsmap->m_epoch);
3361 mutex_unlock(&mdsc->mutex);
3362 return;
3363 }
3364
3365 newmap = ceph_mdsmap_decode(&p, end);
3366 if (IS_ERR(newmap)) {
3367 err = PTR_ERR(newmap);
3368 goto bad_unlock;
3369 }
3370
3371 /* swap into place */
3372 if (mdsc->mdsmap) {
3373 oldmap = mdsc->mdsmap;
3374 mdsc->mdsmap = newmap;
3375 check_new_map(mdsc, newmap, oldmap);
3376 ceph_mdsmap_destroy(oldmap);
3377 } else {
3378 mdsc->mdsmap = newmap; /* first mds map */
3379 }
3380 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3381
3382 __wake_requests(mdsc, &mdsc->waiting_for_map);
3383
3384 mutex_unlock(&mdsc->mutex);
3385 schedule_delayed(mdsc);
3386 return;
3387
3388 bad_unlock:
3389 mutex_unlock(&mdsc->mutex);
3390 bad:
3391 pr_err("error decoding mdsmap %d\n", err);
3392 return;
3393 }
3394
3395 static struct ceph_connection *con_get(struct ceph_connection *con)
3396 {
3397 struct ceph_mds_session *s = con->private;
3398
3399 if (get_session(s)) {
3400 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3401 return con;
3402 }
3403 dout("mdsc con_get %p FAIL\n", s);
3404 return NULL;
3405 }
3406
3407 static void con_put(struct ceph_connection *con)
3408 {
3409 struct ceph_mds_session *s = con->private;
3410
3411 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3412 ceph_put_mds_session(s);
3413 }
3414
3415 /*
3416 * if the client is unresponsive for long enough, the mds will kill
3417 * the session entirely.
3418 */
3419 static void peer_reset(struct ceph_connection *con)
3420 {
3421 struct ceph_mds_session *s = con->private;
3422 struct ceph_mds_client *mdsc = s->s_mdsc;
3423
3424 pr_warning("mds%d closed our session\n", s->s_mds);
3425 send_mds_reconnect(mdsc, s);
3426 }
3427
3428 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3429 {
3430 struct ceph_mds_session *s = con->private;
3431 struct ceph_mds_client *mdsc = s->s_mdsc;
3432 int type = le16_to_cpu(msg->hdr.type);
3433
3434 mutex_lock(&mdsc->mutex);
3435 if (__verify_registered_session(mdsc, s) < 0) {
3436 mutex_unlock(&mdsc->mutex);
3437 goto out;
3438 }
3439 mutex_unlock(&mdsc->mutex);
3440
3441 switch (type) {
3442 case CEPH_MSG_MDS_MAP:
3443 ceph_mdsc_handle_map(mdsc, msg);
3444 break;
3445 case CEPH_MSG_CLIENT_SESSION:
3446 handle_session(s, msg);
3447 break;
3448 case CEPH_MSG_CLIENT_REPLY:
3449 handle_reply(s, msg);
3450 break;
3451 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3452 handle_forward(mdsc, s, msg);
3453 break;
3454 case CEPH_MSG_CLIENT_CAPS:
3455 ceph_handle_caps(s, msg);
3456 break;
3457 case CEPH_MSG_CLIENT_SNAP:
3458 ceph_handle_snap(mdsc, s, msg);
3459 break;
3460 case CEPH_MSG_CLIENT_LEASE:
3461 handle_lease(mdsc, s, msg);
3462 break;
3463
3464 default:
3465 pr_err("received unknown message type %d %s\n", type,
3466 ceph_msg_type_name(type));
3467 }
3468 out:
3469 ceph_msg_put(msg);
3470 }
3471
3472 /*
3473 * authentication
3474 */
3475
3476 /*
3477 * Note: returned pointer is the address of a structure that's
3478 * managed separately. Caller must *not* attempt to free it.
3479 */
3480 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3481 int *proto, int force_new)
3482 {
3483 struct ceph_mds_session *s = con->private;
3484 struct ceph_mds_client *mdsc = s->s_mdsc;
3485 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3486 struct ceph_auth_handshake *auth = &s->s_auth;
3487
3488 if (force_new && auth->authorizer) {
3489 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3490 auth->authorizer = NULL;
3491 }
3492 if (!auth->authorizer) {
3493 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3494 auth);
3495 if (ret)
3496 return ERR_PTR(ret);
3497 } else {
3498 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3499 auth);
3500 if (ret)
3501 return ERR_PTR(ret);
3502 }
3503 *proto = ac->protocol;
3504
3505 return auth;
3506 }
3507
3508
3509 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3510 {
3511 struct ceph_mds_session *s = con->private;
3512 struct ceph_mds_client *mdsc = s->s_mdsc;
3513 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3514
3515 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3516 }
3517
3518 static int invalidate_authorizer(struct ceph_connection *con)
3519 {
3520 struct ceph_mds_session *s = con->private;
3521 struct ceph_mds_client *mdsc = s->s_mdsc;
3522 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3523
3524 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3525
3526 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3527 }
3528
3529 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3530 struct ceph_msg_header *hdr, int *skip)
3531 {
3532 struct ceph_msg *msg;
3533 int type = (int) le16_to_cpu(hdr->type);
3534 int front_len = (int) le32_to_cpu(hdr->front_len);
3535
3536 if (con->in_msg)
3537 return con->in_msg;
3538
3539 *skip = 0;
3540 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3541 if (!msg) {
3542 pr_err("unable to allocate msg type %d len %d\n",
3543 type, front_len);
3544 return NULL;
3545 }
3546
3547 return msg;
3548 }
3549
3550 static const struct ceph_connection_operations mds_con_ops = {
3551 .get = con_get,
3552 .put = con_put,
3553 .dispatch = dispatch,
3554 .get_authorizer = get_authorizer,
3555 .verify_authorizer_reply = verify_authorizer_reply,
3556 .invalidate_authorizer = invalidate_authorizer,
3557 .peer_reset = peer_reset,
3558 .alloc_msg = mds_alloc_msg,
3559 };
3560
3561 /* eof */
Linux with added FPC-III support