?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
IB/rxe: check for allocation failure on elem
[linux/fpc-iii.git] / fs / ceph / mds_client.c
blob3d2639c30018a5956314fe00cdd5fe8f8c81afb9
1 #include <linux/ceph/ceph_debug.h>
2
3 #include <linux/fs.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
6 #include <linux/gfp.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11 #include <linux/ratelimit.h>
12
13 #include "super.h"
14 #include "mds_client.h"
15
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
22
23 /*
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
30 *
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
35 * requests.
36 *
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
40 *
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
46 */
47
48 struct ceph_reconnect_state {
49 int nr_caps;
50 struct ceph_pagelist *pagelist;
51 unsigned msg_version;
52 };
53
54 static void __wake_requests(struct ceph_mds_client *mdsc,
55 struct list_head *head);
56
57 static const struct ceph_connection_operations mds_con_ops;
58
59
60 /*
61 * mds reply parsing
62 */
63
64 /*
65 * parse individual inode info
66 */
67 static int parse_reply_info_in(void **p, void *end,
68 struct ceph_mds_reply_info_in *info,
69 u64 features)
70 {
71 int err = -EIO;
72
73 info->in = *p;
74 *p += sizeof(struct ceph_mds_reply_inode) +
75 sizeof(*info->in->fragtree.splits) *
76 le32_to_cpu(info->in->fragtree.nsplits);
77
78 ceph_decode_32_safe(p, end, info->symlink_len, bad);
79 ceph_decode_need(p, end, info->symlink_len, bad);
80 info->symlink = *p;
81 *p += info->symlink_len;
82
83 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
84 ceph_decode_copy_safe(p, end, &info->dir_layout,
85 sizeof(info->dir_layout), bad);
86 else
87 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
88
89 ceph_decode_32_safe(p, end, info->xattr_len, bad);
90 ceph_decode_need(p, end, info->xattr_len, bad);
91 info->xattr_data = *p;
92 *p += info->xattr_len;
93
94 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
95 ceph_decode_64_safe(p, end, info->inline_version, bad);
96 ceph_decode_32_safe(p, end, info->inline_len, bad);
97 ceph_decode_need(p, end, info->inline_len, bad);
98 info->inline_data = *p;
99 *p += info->inline_len;
100 } else
101 info->inline_version = CEPH_INLINE_NONE;
102
103 info->pool_ns_len = 0;
104 info->pool_ns_data = NULL;
105 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
106 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
107 if (info->pool_ns_len > 0) {
108 ceph_decode_need(p, end, info->pool_ns_len, bad);
109 info->pool_ns_data = *p;
110 *p += info->pool_ns_len;
111 }
112 }
113
114 return 0;
115 bad:
116 return err;
117 }
118
119 /*
120 * parse a normal reply, which may contain a (dir+)dentry and/or a
121 * target inode.
122 */
123 static int parse_reply_info_trace(void **p, void *end,
124 struct ceph_mds_reply_info_parsed *info,
125 u64 features)
126 {
127 int err;
128
129 if (info->head->is_dentry) {
130 err = parse_reply_info_in(p, end, &info->diri, features);
131 if (err < 0)
132 goto out_bad;
133
134 if (unlikely(*p + sizeof(*info->dirfrag) > end))
135 goto bad;
136 info->dirfrag = *p;
137 *p += sizeof(*info->dirfrag) +
138 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
139 if (unlikely(*p > end))
140 goto bad;
141
142 ceph_decode_32_safe(p, end, info->dname_len, bad);
143 ceph_decode_need(p, end, info->dname_len, bad);
144 info->dname = *p;
145 *p += info->dname_len;
146 info->dlease = *p;
147 *p += sizeof(*info->dlease);
148 }
149
150 if (info->head->is_target) {
151 err = parse_reply_info_in(p, end, &info->targeti, features);
152 if (err < 0)
153 goto out_bad;
154 }
155
156 if (unlikely(*p != end))
157 goto bad;
158 return 0;
159
160 bad:
161 err = -EIO;
162 out_bad:
163 pr_err("problem parsing mds trace %d\n", err);
164 return err;
165 }
166
167 /*
168 * parse readdir results
169 */
170 static int parse_reply_info_dir(void **p, void *end,
171 struct ceph_mds_reply_info_parsed *info,
172 u64 features)
173 {
174 u32 num, i = 0;
175 int err;
176
177 info->dir_dir = *p;
178 if (*p + sizeof(*info->dir_dir) > end)
179 goto bad;
180 *p += sizeof(*info->dir_dir) +
181 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
182 if (*p > end)
183 goto bad;
184
185 ceph_decode_need(p, end, sizeof(num) + 2, bad);
186 num = ceph_decode_32(p);
187 {
188 u16 flags = ceph_decode_16(p);
189 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
190 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
191 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
192 }
193 if (num == 0)
194 goto done;
195
196 BUG_ON(!info->dir_entries);
197 if ((unsigned long)(info->dir_entries + num) >
198 (unsigned long)info->dir_entries + info->dir_buf_size) {
199 pr_err("dir contents are larger than expected\n");
200 WARN_ON(1);
201 goto bad;
202 }
203
204 info->dir_nr = num;
205 while (num) {
206 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
207 /* dentry */
208 ceph_decode_need(p, end, sizeof(u32)*2, bad);
209 rde->name_len = ceph_decode_32(p);
210 ceph_decode_need(p, end, rde->name_len, bad);
211 rde->name = *p;
212 *p += rde->name_len;
213 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
214 rde->lease = *p;
215 *p += sizeof(struct ceph_mds_reply_lease);
216
217 /* inode */
218 err = parse_reply_info_in(p, end, &rde->inode, features);
219 if (err < 0)
220 goto out_bad;
221 /* ceph_readdir_prepopulate() will update it */
222 rde->offset = 0;
223 i++;
224 num--;
225 }
226
227 done:
228 if (*p != end)
229 goto bad;
230 return 0;
231
232 bad:
233 err = -EIO;
234 out_bad:
235 pr_err("problem parsing dir contents %d\n", err);
236 return err;
237 }
238
239 /*
240 * parse fcntl F_GETLK results
241 */
242 static int parse_reply_info_filelock(void **p, void *end,
243 struct ceph_mds_reply_info_parsed *info,
244 u64 features)
245 {
246 if (*p + sizeof(*info->filelock_reply) > end)
247 goto bad;
248
249 info->filelock_reply = *p;
250 *p += sizeof(*info->filelock_reply);
251
252 if (unlikely(*p != end))
253 goto bad;
254 return 0;
255
256 bad:
257 return -EIO;
258 }
259
260 /*
261 * parse create results
262 */
263 static int parse_reply_info_create(void **p, void *end,
264 struct ceph_mds_reply_info_parsed *info,
265 u64 features)
266 {
267 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
268 if (*p == end) {
269 info->has_create_ino = false;
270 } else {
271 info->has_create_ino = true;
272 info->ino = ceph_decode_64(p);
273 }
274 }
275
276 if (unlikely(*p != end))
277 goto bad;
278 return 0;
279
280 bad:
281 return -EIO;
282 }
283
284 /*
285 * parse extra results
286 */
287 static int parse_reply_info_extra(void **p, void *end,
288 struct ceph_mds_reply_info_parsed *info,
289 u64 features)
290 {
291 u32 op = le32_to_cpu(info->head->op);
292
293 if (op == CEPH_MDS_OP_GETFILELOCK)
294 return parse_reply_info_filelock(p, end, info, features);
295 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
296 return parse_reply_info_dir(p, end, info, features);
297 else if (op == CEPH_MDS_OP_CREATE)
298 return parse_reply_info_create(p, end, info, features);
299 else
300 return -EIO;
301 }
302
303 /*
304 * parse entire mds reply
305 */
306 static int parse_reply_info(struct ceph_msg *msg,
307 struct ceph_mds_reply_info_parsed *info,
308 u64 features)
309 {
310 void *p, *end;
311 u32 len;
312 int err;
313
314 info->head = msg->front.iov_base;
315 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
316 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
317
318 /* trace */
319 ceph_decode_32_safe(&p, end, len, bad);
320 if (len > 0) {
321 ceph_decode_need(&p, end, len, bad);
322 err = parse_reply_info_trace(&p, p+len, info, features);
323 if (err < 0)
324 goto out_bad;
325 }
326
327 /* extra */
328 ceph_decode_32_safe(&p, end, len, bad);
329 if (len > 0) {
330 ceph_decode_need(&p, end, len, bad);
331 err = parse_reply_info_extra(&p, p+len, info, features);
332 if (err < 0)
333 goto out_bad;
334 }
335
336 /* snap blob */
337 ceph_decode_32_safe(&p, end, len, bad);
338 info->snapblob_len = len;
339 info->snapblob = p;
340 p += len;
341
342 if (p != end)
343 goto bad;
344 return 0;
345
346 bad:
347 err = -EIO;
348 out_bad:
349 pr_err("mds parse_reply err %d\n", err);
350 return err;
351 }
352
353 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
354 {
355 if (!info->dir_entries)
356 return;
357 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
358 }
359
360
361 /*
362 * sessions
363 */
364 const char *ceph_session_state_name(int s)
365 {
366 switch (s) {
367 case CEPH_MDS_SESSION_NEW: return "new";
368 case CEPH_MDS_SESSION_OPENING: return "opening";
369 case CEPH_MDS_SESSION_OPEN: return "open";
370 case CEPH_MDS_SESSION_HUNG: return "hung";
371 case CEPH_MDS_SESSION_CLOSING: return "closing";
372 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
373 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
374 case CEPH_MDS_SESSION_REJECTED: return "rejected";
375 default: return "???";
376 }
377 }
378
379 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
380 {
381 if (atomic_inc_not_zero(&s->s_ref)) {
382 dout("mdsc get_session %p %d -> %d\n", s,
383 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
384 return s;
385 } else {
386 dout("mdsc get_session %p 0 -- FAIL", s);
387 return NULL;
388 }
389 }
390
391 void ceph_put_mds_session(struct ceph_mds_session *s)
392 {
393 dout("mdsc put_session %p %d -> %d\n", s,
394 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
395 if (atomic_dec_and_test(&s->s_ref)) {
396 if (s->s_auth.authorizer)
397 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
398 kfree(s);
399 }
400 }
401
402 /*
403 * called under mdsc->mutex
404 */
405 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
406 int mds)
407 {
408 struct ceph_mds_session *session;
409
410 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
411 return NULL;
412 session = mdsc->sessions[mds];
413 dout("lookup_mds_session %p %d\n", session,
414 atomic_read(&session->s_ref));
415 get_session(session);
416 return session;
417 }
418
419 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
420 {
421 if (mds >= mdsc->max_sessions)
422 return false;
423 return mdsc->sessions[mds];
424 }
425
426 static int __verify_registered_session(struct ceph_mds_client *mdsc,
427 struct ceph_mds_session *s)
428 {
429 if (s->s_mds >= mdsc->max_sessions ||
430 mdsc->sessions[s->s_mds] != s)
431 return -ENOENT;
432 return 0;
433 }
434
435 /*
436 * create+register a new session for given mds.
437 * called under mdsc->mutex.
438 */
439 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
440 int mds)
441 {
442 struct ceph_mds_session *s;
443
444 if (mds >= mdsc->mdsmap->m_max_mds)
445 return ERR_PTR(-EINVAL);
446
447 s = kzalloc(sizeof(*s), GFP_NOFS);
448 if (!s)
449 return ERR_PTR(-ENOMEM);
450 s->s_mdsc = mdsc;
451 s->s_mds = mds;
452 s->s_state = CEPH_MDS_SESSION_NEW;
453 s->s_ttl = 0;
454 s->s_seq = 0;
455 mutex_init(&s->s_mutex);
456
457 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
458
459 spin_lock_init(&s->s_gen_ttl_lock);
460 s->s_cap_gen = 0;
461 s->s_cap_ttl = jiffies - 1;
462
463 spin_lock_init(&s->s_cap_lock);
464 s->s_renew_requested = 0;
465 s->s_renew_seq = 0;
466 INIT_LIST_HEAD(&s->s_caps);
467 s->s_nr_caps = 0;
468 s->s_trim_caps = 0;
469 atomic_set(&s->s_ref, 1);
470 INIT_LIST_HEAD(&s->s_waiting);
471 INIT_LIST_HEAD(&s->s_unsafe);
472 s->s_num_cap_releases = 0;
473 s->s_cap_reconnect = 0;
474 s->s_cap_iterator = NULL;
475 INIT_LIST_HEAD(&s->s_cap_releases);
476 INIT_LIST_HEAD(&s->s_cap_flushing);
477
478 dout("register_session mds%d\n", mds);
479 if (mds >= mdsc->max_sessions) {
480 int newmax = 1 << get_count_order(mds+1);
481 struct ceph_mds_session **sa;
482
483 dout("register_session realloc to %d\n", newmax);
484 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
485 if (sa == NULL)
486 goto fail_realloc;
487 if (mdsc->sessions) {
488 memcpy(sa, mdsc->sessions,
489 mdsc->max_sessions * sizeof(void *));
490 kfree(mdsc->sessions);
491 }
492 mdsc->sessions = sa;
493 mdsc->max_sessions = newmax;
494 }
495 mdsc->sessions[mds] = s;
496 atomic_inc(&mdsc->num_sessions);
497 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
498
499 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
500 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
501
502 return s;
503
504 fail_realloc:
505 kfree(s);
506 return ERR_PTR(-ENOMEM);
507 }
508
509 /*
510 * called under mdsc->mutex
511 */
512 static void __unregister_session(struct ceph_mds_client *mdsc,
513 struct ceph_mds_session *s)
514 {
515 dout("__unregister_session mds%d %p\n", s->s_mds, s);
516 BUG_ON(mdsc->sessions[s->s_mds] != s);
517 mdsc->sessions[s->s_mds] = NULL;
518 ceph_con_close(&s->s_con);
519 ceph_put_mds_session(s);
520 atomic_dec(&mdsc->num_sessions);
521 }
522
523 /*
524 * drop session refs in request.
525 *
526 * should be last request ref, or hold mdsc->mutex
527 */
528 static void put_request_session(struct ceph_mds_request *req)
529 {
530 if (req->r_session) {
531 ceph_put_mds_session(req->r_session);
532 req->r_session = NULL;
533 }
534 }
535
536 void ceph_mdsc_release_request(struct kref *kref)
537 {
538 struct ceph_mds_request *req = container_of(kref,
539 struct ceph_mds_request,
540 r_kref);
541 destroy_reply_info(&req->r_reply_info);
542 if (req->r_request)
543 ceph_msg_put(req->r_request);
544 if (req->r_reply)
545 ceph_msg_put(req->r_reply);
546 if (req->r_inode) {
547 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
548 iput(req->r_inode);
549 }
550 if (req->r_locked_dir)
551 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
552 iput(req->r_target_inode);
553 if (req->r_dentry)
554 dput(req->r_dentry);
555 if (req->r_old_dentry)
556 dput(req->r_old_dentry);
557 if (req->r_old_dentry_dir) {
558 /*
559 * track (and drop pins for) r_old_dentry_dir
560 * separately, since r_old_dentry's d_parent may have
561 * changed between the dir mutex being dropped and
562 * this request being freed.
563 */
564 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
565 CEPH_CAP_PIN);
566 iput(req->r_old_dentry_dir);
567 }
568 kfree(req->r_path1);
569 kfree(req->r_path2);
570 if (req->r_pagelist)
571 ceph_pagelist_release(req->r_pagelist);
572 put_request_session(req);
573 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
574 kfree(req);
575 }
576
577 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
578
579 /*
580 * lookup session, bump ref if found.
581 *
582 * called under mdsc->mutex.
583 */
584 static struct ceph_mds_request *
585 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
586 {
587 struct ceph_mds_request *req;
588
589 req = lookup_request(&mdsc->request_tree, tid);
590 if (req)
591 ceph_mdsc_get_request(req);
592
593 return req;
594 }
595
596 /*
597 * Register an in-flight request, and assign a tid. Link to directory
598 * are modifying (if any).
599 *
600 * Called under mdsc->mutex.
601 */
602 static void __register_request(struct ceph_mds_client *mdsc,
603 struct ceph_mds_request *req,
604 struct inode *dir)
605 {
606 req->r_tid = ++mdsc->last_tid;
607 if (req->r_num_caps)
608 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
609 req->r_num_caps);
610 dout("__register_request %p tid %lld\n", req, req->r_tid);
611 ceph_mdsc_get_request(req);
612 insert_request(&mdsc->request_tree, req);
613
614 req->r_uid = current_fsuid();
615 req->r_gid = current_fsgid();
616
617 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
618 mdsc->oldest_tid = req->r_tid;
619
620 if (dir) {
621 ihold(dir);
622 req->r_unsafe_dir = dir;
623 }
624 }
625
626 static void __unregister_request(struct ceph_mds_client *mdsc,
627 struct ceph_mds_request *req)
628 {
629 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
630
631 /* Never leave an unregistered request on an unsafe list! */
632 list_del_init(&req->r_unsafe_item);
633
634 if (req->r_tid == mdsc->oldest_tid) {
635 struct rb_node *p = rb_next(&req->r_node);
636 mdsc->oldest_tid = 0;
637 while (p) {
638 struct ceph_mds_request *next_req =
639 rb_entry(p, struct ceph_mds_request, r_node);
640 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
641 mdsc->oldest_tid = next_req->r_tid;
642 break;
643 }
644 p = rb_next(p);
645 }
646 }
647
648 erase_request(&mdsc->request_tree, req);
649
650 if (req->r_unsafe_dir && req->r_got_unsafe) {
651 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
652 spin_lock(&ci->i_unsafe_lock);
653 list_del_init(&req->r_unsafe_dir_item);
654 spin_unlock(&ci->i_unsafe_lock);
655 }
656 if (req->r_target_inode && req->r_got_unsafe) {
657 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
658 spin_lock(&ci->i_unsafe_lock);
659 list_del_init(&req->r_unsafe_target_item);
660 spin_unlock(&ci->i_unsafe_lock);
661 }
662
663 if (req->r_unsafe_dir) {
664 iput(req->r_unsafe_dir);
665 req->r_unsafe_dir = NULL;
666 }
667
668 complete_all(&req->r_safe_completion);
669
670 ceph_mdsc_put_request(req);
671 }
672
673 /*
674 * Choose mds to send request to next. If there is a hint set in the
675 * request (e.g., due to a prior forward hint from the mds), use that.
676 * Otherwise, consult frag tree and/or caps to identify the
677 * appropriate mds. If all else fails, choose randomly.
678 *
679 * Called under mdsc->mutex.
680 */
681 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
682 {
683 /*
684 * we don't need to worry about protecting the d_parent access
685 * here because we never renaming inside the snapped namespace
686 * except to resplice to another snapdir, and either the old or new
687 * result is a valid result.
688 */
689 while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
690 dentry = dentry->d_parent;
691 return dentry;
692 }
693
694 static int __choose_mds(struct ceph_mds_client *mdsc,
695 struct ceph_mds_request *req)
696 {
697 struct inode *inode;
698 struct ceph_inode_info *ci;
699 struct ceph_cap *cap;
700 int mode = req->r_direct_mode;
701 int mds = -1;
702 u32 hash = req->r_direct_hash;
703 bool is_hash = req->r_direct_is_hash;
704
705 /*
706 * is there a specific mds we should try? ignore hint if we have
707 * no session and the mds is not up (active or recovering).
708 */
709 if (req->r_resend_mds >= 0 &&
710 (__have_session(mdsc, req->r_resend_mds) ||
711 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
712 dout("choose_mds using resend_mds mds%d\n",
713 req->r_resend_mds);
714 return req->r_resend_mds;
715 }
716
717 if (mode == USE_RANDOM_MDS)
718 goto random;
719
720 inode = NULL;
721 if (req->r_inode) {
722 inode = req->r_inode;
723 } else if (req->r_dentry) {
724 /* ignore race with rename; old or new d_parent is okay */
725 struct dentry *parent = req->r_dentry->d_parent;
726 struct inode *dir = d_inode(parent);
727
728 if (dir->i_sb != mdsc->fsc->sb) {
729 /* not this fs! */
730 inode = d_inode(req->r_dentry);
731 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
732 /* direct snapped/virtual snapdir requests
733 * based on parent dir inode */
734 struct dentry *dn = get_nonsnap_parent(parent);
735 inode = d_inode(dn);
736 dout("__choose_mds using nonsnap parent %p\n", inode);
737 } else {
738 /* dentry target */
739 inode = d_inode(req->r_dentry);
740 if (!inode || mode == USE_AUTH_MDS) {
741 /* dir + name */
742 inode = dir;
743 hash = ceph_dentry_hash(dir, req->r_dentry);
744 is_hash = true;
745 }
746 }
747 }
748
749 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
750 (int)hash, mode);
751 if (!inode)
752 goto random;
753 ci = ceph_inode(inode);
754
755 if (is_hash && S_ISDIR(inode->i_mode)) {
756 struct ceph_inode_frag frag;
757 int found;
758
759 ceph_choose_frag(ci, hash, &frag, &found);
760 if (found) {
761 if (mode == USE_ANY_MDS && frag.ndist > 0) {
762 u8 r;
763
764 /* choose a random replica */
765 get_random_bytes(&r, 1);
766 r %= frag.ndist;
767 mds = frag.dist[r];
768 dout("choose_mds %p %llx.%llx "
769 "frag %u mds%d (%d/%d)\n",
770 inode, ceph_vinop(inode),
771 frag.frag, mds,
772 (int)r, frag.ndist);
773 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
774 CEPH_MDS_STATE_ACTIVE)
775 return mds;
776 }
777
778 /* since this file/dir wasn't known to be
779 * replicated, then we want to look for the
780 * authoritative mds. */
781 mode = USE_AUTH_MDS;
782 if (frag.mds >= 0) {
783 /* choose auth mds */
784 mds = frag.mds;
785 dout("choose_mds %p %llx.%llx "
786 "frag %u mds%d (auth)\n",
787 inode, ceph_vinop(inode), frag.frag, mds);
788 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
789 CEPH_MDS_STATE_ACTIVE)
790 return mds;
791 }
792 }
793 }
794
795 spin_lock(&ci->i_ceph_lock);
796 cap = NULL;
797 if (mode == USE_AUTH_MDS)
798 cap = ci->i_auth_cap;
799 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
800 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
801 if (!cap) {
802 spin_unlock(&ci->i_ceph_lock);
803 goto random;
804 }
805 mds = cap->session->s_mds;
806 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
807 inode, ceph_vinop(inode), mds,
808 cap == ci->i_auth_cap ? "auth " : "", cap);
809 spin_unlock(&ci->i_ceph_lock);
810 return mds;
811
812 random:
813 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
814 dout("choose_mds chose random mds%d\n", mds);
815 return mds;
816 }
817
818
819 /*
820 * session messages
821 */
822 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
823 {
824 struct ceph_msg *msg;
825 struct ceph_mds_session_head *h;
826
827 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
828 false);
829 if (!msg) {
830 pr_err("create_session_msg ENOMEM creating msg\n");
831 return NULL;
832 }
833 h = msg->front.iov_base;
834 h->op = cpu_to_le32(op);
835 h->seq = cpu_to_le64(seq);
836
837 return msg;
838 }
839
840 /*
841 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
842 * to include additional client metadata fields.
843 */
844 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
845 {
846 struct ceph_msg *msg;
847 struct ceph_mds_session_head *h;
848 int i = -1;
849 int metadata_bytes = 0;
850 int metadata_key_count = 0;
851 struct ceph_options *opt = mdsc->fsc->client->options;
852 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
853 void *p;
854
855 const char* metadata[][2] = {
856 {"hostname", utsname()->nodename},
857 {"kernel_version", utsname()->release},
858 {"entity_id", opt->name ? : ""},
859 {"root", fsopt->server_path ? : "/"},
860 {NULL, NULL}
861 };
862
863 /* Calculate serialized length of metadata */
864 metadata_bytes = 4; /* map length */
865 for (i = 0; metadata[i][0] != NULL; ++i) {
866 metadata_bytes += 8 + strlen(metadata[i][0]) +
867 strlen(metadata[i][1]);
868 metadata_key_count++;
869 }
870
871 /* Allocate the message */
872 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
873 GFP_NOFS, false);
874 if (!msg) {
875 pr_err("create_session_msg ENOMEM creating msg\n");
876 return NULL;
877 }
878 h = msg->front.iov_base;
879 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
880 h->seq = cpu_to_le64(seq);
881
882 /*
883 * Serialize client metadata into waiting buffer space, using
884 * the format that userspace expects for map<string, string>
885 *
886 * ClientSession messages with metadata are v2
887 */
888 msg->hdr.version = cpu_to_le16(2);
889 msg->hdr.compat_version = cpu_to_le16(1);
890
891 /* The write pointer, following the session_head structure */
892 p = msg->front.iov_base + sizeof(*h);
893
894 /* Number of entries in the map */
895 ceph_encode_32(&p, metadata_key_count);
896
897 /* Two length-prefixed strings for each entry in the map */
898 for (i = 0; metadata[i][0] != NULL; ++i) {
899 size_t const key_len = strlen(metadata[i][0]);
900 size_t const val_len = strlen(metadata[i][1]);
901
902 ceph_encode_32(&p, key_len);
903 memcpy(p, metadata[i][0], key_len);
904 p += key_len;
905 ceph_encode_32(&p, val_len);
906 memcpy(p, metadata[i][1], val_len);
907 p += val_len;
908 }
909
910 return msg;
911 }
912
913 /*
914 * send session open request.
915 *
916 * called under mdsc->mutex
917 */
918 static int __open_session(struct ceph_mds_client *mdsc,
919 struct ceph_mds_session *session)
920 {
921 struct ceph_msg *msg;
922 int mstate;
923 int mds = session->s_mds;
924
925 /* wait for mds to go active? */
926 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
927 dout("open_session to mds%d (%s)\n", mds,
928 ceph_mds_state_name(mstate));
929 session->s_state = CEPH_MDS_SESSION_OPENING;
930 session->s_renew_requested = jiffies;
931
932 /* send connect message */
933 msg = create_session_open_msg(mdsc, session->s_seq);
934 if (!msg)
935 return -ENOMEM;
936 ceph_con_send(&session->s_con, msg);
937 return 0;
938 }
939
940 /*
941 * open sessions for any export targets for the given mds
942 *
943 * called under mdsc->mutex
944 */
945 static struct ceph_mds_session *
946 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
947 {
948 struct ceph_mds_session *session;
949
950 session = __ceph_lookup_mds_session(mdsc, target);
951 if (!session) {
952 session = register_session(mdsc, target);
953 if (IS_ERR(session))
954 return session;
955 }
956 if (session->s_state == CEPH_MDS_SESSION_NEW ||
957 session->s_state == CEPH_MDS_SESSION_CLOSING)
958 __open_session(mdsc, session);
959
960 return session;
961 }
962
963 struct ceph_mds_session *
964 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
965 {
966 struct ceph_mds_session *session;
967
968 dout("open_export_target_session to mds%d\n", target);
969
970 mutex_lock(&mdsc->mutex);
971 session = __open_export_target_session(mdsc, target);
972 mutex_unlock(&mdsc->mutex);
973
974 return session;
975 }
976
977 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
978 struct ceph_mds_session *session)
979 {
980 struct ceph_mds_info *mi;
981 struct ceph_mds_session *ts;
982 int i, mds = session->s_mds;
983
984 if (mds >= mdsc->mdsmap->m_max_mds)
985 return;
986
987 mi = &mdsc->mdsmap->m_info[mds];
988 dout("open_export_target_sessions for mds%d (%d targets)\n",
989 session->s_mds, mi->num_export_targets);
990
991 for (i = 0; i < mi->num_export_targets; i++) {
992 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
993 if (!IS_ERR(ts))
994 ceph_put_mds_session(ts);
995 }
996 }
997
998 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
999 struct ceph_mds_session *session)
1000 {
1001 mutex_lock(&mdsc->mutex);
1002 __open_export_target_sessions(mdsc, session);
1003 mutex_unlock(&mdsc->mutex);
1004 }
1005
1006 /*
1007 * session caps
1008 */
1009
1010 /* caller holds s_cap_lock, we drop it */
1011 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1012 struct ceph_mds_session *session)
1013 __releases(session->s_cap_lock)
1014 {
1015 LIST_HEAD(tmp_list);
1016 list_splice_init(&session->s_cap_releases, &tmp_list);
1017 session->s_num_cap_releases = 0;
1018 spin_unlock(&session->s_cap_lock);
1019
1020 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1021 while (!list_empty(&tmp_list)) {
1022 struct ceph_cap *cap;
1023 /* zero out the in-progress message */
1024 cap = list_first_entry(&tmp_list,
1025 struct ceph_cap, session_caps);
1026 list_del(&cap->session_caps);
1027 ceph_put_cap(mdsc, cap);
1028 }
1029 }
1030
1031 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1032 struct ceph_mds_session *session)
1033 {
1034 struct ceph_mds_request *req;
1035 struct rb_node *p;
1036
1037 dout("cleanup_session_requests mds%d\n", session->s_mds);
1038 mutex_lock(&mdsc->mutex);
1039 while (!list_empty(&session->s_unsafe)) {
1040 req = list_first_entry(&session->s_unsafe,
1041 struct ceph_mds_request, r_unsafe_item);
1042 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1043 req->r_tid);
1044 __unregister_request(mdsc, req);
1045 }
1046 /* zero r_attempts, so kick_requests() will re-send requests */
1047 p = rb_first(&mdsc->request_tree);
1048 while (p) {
1049 req = rb_entry(p, struct ceph_mds_request, r_node);
1050 p = rb_next(p);
1051 if (req->r_session &&
1052 req->r_session->s_mds == session->s_mds)
1053 req->r_attempts = 0;
1054 }
1055 mutex_unlock(&mdsc->mutex);
1056 }
1057
1058 /*
1059 * Helper to safely iterate over all caps associated with a session, with
1060 * special care taken to handle a racing __ceph_remove_cap().
1061 *
1062 * Caller must hold session s_mutex.
1063 */
1064 static int iterate_session_caps(struct ceph_mds_session *session,
1065 int (*cb)(struct inode *, struct ceph_cap *,
1066 void *), void *arg)
1067 {
1068 struct list_head *p;
1069 struct ceph_cap *cap;
1070 struct inode *inode, *last_inode = NULL;
1071 struct ceph_cap *old_cap = NULL;
1072 int ret;
1073
1074 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1075 spin_lock(&session->s_cap_lock);
1076 p = session->s_caps.next;
1077 while (p != &session->s_caps) {
1078 cap = list_entry(p, struct ceph_cap, session_caps);
1079 inode = igrab(&cap->ci->vfs_inode);
1080 if (!inode) {
1081 p = p->next;
1082 continue;
1083 }
1084 session->s_cap_iterator = cap;
1085 spin_unlock(&session->s_cap_lock);
1086
1087 if (last_inode) {
1088 iput(last_inode);
1089 last_inode = NULL;
1090 }
1091 if (old_cap) {
1092 ceph_put_cap(session->s_mdsc, old_cap);
1093 old_cap = NULL;
1094 }
1095
1096 ret = cb(inode, cap, arg);
1097 last_inode = inode;
1098
1099 spin_lock(&session->s_cap_lock);
1100 p = p->next;
1101 if (cap->ci == NULL) {
1102 dout("iterate_session_caps finishing cap %p removal\n",
1103 cap);
1104 BUG_ON(cap->session != session);
1105 cap->session = NULL;
1106 list_del_init(&cap->session_caps);
1107 session->s_nr_caps--;
1108 if (cap->queue_release) {
1109 list_add_tail(&cap->session_caps,
1110 &session->s_cap_releases);
1111 session->s_num_cap_releases++;
1112 } else {
1113 old_cap = cap; /* put_cap it w/o locks held */
1114 }
1115 }
1116 if (ret < 0)
1117 goto out;
1118 }
1119 ret = 0;
1120 out:
1121 session->s_cap_iterator = NULL;
1122 spin_unlock(&session->s_cap_lock);
1123
1124 iput(last_inode);
1125 if (old_cap)
1126 ceph_put_cap(session->s_mdsc, old_cap);
1127
1128 return ret;
1129 }
1130
1131 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1132 void *arg)
1133 {
1134 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1135 struct ceph_inode_info *ci = ceph_inode(inode);
1136 LIST_HEAD(to_remove);
1137 bool drop = false;
1138 bool invalidate = false;
1139
1140 dout("removing cap %p, ci is %p, inode is %p\n",
1141 cap, ci, &ci->vfs_inode);
1142 spin_lock(&ci->i_ceph_lock);
1143 __ceph_remove_cap(cap, false);
1144 if (!ci->i_auth_cap) {
1145 struct ceph_cap_flush *cf;
1146 struct ceph_mds_client *mdsc = fsc->mdsc;
1147
1148 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1149
1150 if (ci->i_wrbuffer_ref > 0 &&
1151 ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1152 invalidate = true;
1153
1154 while (!list_empty(&ci->i_cap_flush_list)) {
1155 cf = list_first_entry(&ci->i_cap_flush_list,
1156 struct ceph_cap_flush, i_list);
1157 list_move(&cf->i_list, &to_remove);
1158 }
1159
1160 spin_lock(&mdsc->cap_dirty_lock);
1161
1162 list_for_each_entry(cf, &to_remove, i_list)
1163 list_del(&cf->g_list);
1164
1165 if (!list_empty(&ci->i_dirty_item)) {
1166 pr_warn_ratelimited(
1167 " dropping dirty %s state for %p %lld\n",
1168 ceph_cap_string(ci->i_dirty_caps),
1169 inode, ceph_ino(inode));
1170 ci->i_dirty_caps = 0;
1171 list_del_init(&ci->i_dirty_item);
1172 drop = true;
1173 }
1174 if (!list_empty(&ci->i_flushing_item)) {
1175 pr_warn_ratelimited(
1176 " dropping dirty+flushing %s state for %p %lld\n",
1177 ceph_cap_string(ci->i_flushing_caps),
1178 inode, ceph_ino(inode));
1179 ci->i_flushing_caps = 0;
1180 list_del_init(&ci->i_flushing_item);
1181 mdsc->num_cap_flushing--;
1182 drop = true;
1183 }
1184 spin_unlock(&mdsc->cap_dirty_lock);
1185
1186 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1187 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1188 ci->i_prealloc_cap_flush = NULL;
1189 }
1190 }
1191 spin_unlock(&ci->i_ceph_lock);
1192 while (!list_empty(&to_remove)) {
1193 struct ceph_cap_flush *cf;
1194 cf = list_first_entry(&to_remove,
1195 struct ceph_cap_flush, i_list);
1196 list_del(&cf->i_list);
1197 ceph_free_cap_flush(cf);
1198 }
1199
1200 wake_up_all(&ci->i_cap_wq);
1201 if (invalidate)
1202 ceph_queue_invalidate(inode);
1203 if (drop)
1204 iput(inode);
1205 return 0;
1206 }
1207
1208 /*
1209 * caller must hold session s_mutex
1210 */
1211 static void remove_session_caps(struct ceph_mds_session *session)
1212 {
1213 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1214 struct super_block *sb = fsc->sb;
1215 dout("remove_session_caps on %p\n", session);
1216 iterate_session_caps(session, remove_session_caps_cb, fsc);
1217
1218 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1219
1220 spin_lock(&session->s_cap_lock);
1221 if (session->s_nr_caps > 0) {
1222 struct inode *inode;
1223 struct ceph_cap *cap, *prev = NULL;
1224 struct ceph_vino vino;
1225 /*
1226 * iterate_session_caps() skips inodes that are being
1227 * deleted, we need to wait until deletions are complete.
1228 * __wait_on_freeing_inode() is designed for the job,
1229 * but it is not exported, so use lookup inode function
1230 * to access it.
1231 */
1232 while (!list_empty(&session->s_caps)) {
1233 cap = list_entry(session->s_caps.next,
1234 struct ceph_cap, session_caps);
1235 if (cap == prev)
1236 break;
1237 prev = cap;
1238 vino = cap->ci->i_vino;
1239 spin_unlock(&session->s_cap_lock);
1240
1241 inode = ceph_find_inode(sb, vino);
1242 iput(inode);
1243
1244 spin_lock(&session->s_cap_lock);
1245 }
1246 }
1247
1248 // drop cap expires and unlock s_cap_lock
1249 cleanup_cap_releases(session->s_mdsc, session);
1250
1251 BUG_ON(session->s_nr_caps > 0);
1252 BUG_ON(!list_empty(&session->s_cap_flushing));
1253 }
1254
1255 /*
1256 * wake up any threads waiting on this session's caps. if the cap is
1257 * old (didn't get renewed on the client reconnect), remove it now.
1258 *
1259 * caller must hold s_mutex.
1260 */
1261 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1262 void *arg)
1263 {
1264 struct ceph_inode_info *ci = ceph_inode(inode);
1265
1266 if (arg) {
1267 spin_lock(&ci->i_ceph_lock);
1268 ci->i_wanted_max_size = 0;
1269 ci->i_requested_max_size = 0;
1270 spin_unlock(&ci->i_ceph_lock);
1271 }
1272 wake_up_all(&ci->i_cap_wq);
1273 return 0;
1274 }
1275
1276 static void wake_up_session_caps(struct ceph_mds_session *session,
1277 int reconnect)
1278 {
1279 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1280 iterate_session_caps(session, wake_up_session_cb,
1281 (void *)(unsigned long)reconnect);
1282 }
1283
1284 /*
1285 * Send periodic message to MDS renewing all currently held caps. The
1286 * ack will reset the expiration for all caps from this session.
1287 *
1288 * caller holds s_mutex
1289 */
1290 static int send_renew_caps(struct ceph_mds_client *mdsc,
1291 struct ceph_mds_session *session)
1292 {
1293 struct ceph_msg *msg;
1294 int state;
1295
1296 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1297 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1298 pr_info("mds%d caps stale\n", session->s_mds);
1299 session->s_renew_requested = jiffies;
1300
1301 /* do not try to renew caps until a recovering mds has reconnected
1302 * with its clients. */
1303 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1304 if (state < CEPH_MDS_STATE_RECONNECT) {
1305 dout("send_renew_caps ignoring mds%d (%s)\n",
1306 session->s_mds, ceph_mds_state_name(state));
1307 return 0;
1308 }
1309
1310 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1311 ceph_mds_state_name(state));
1312 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1313 ++session->s_renew_seq);
1314 if (!msg)
1315 return -ENOMEM;
1316 ceph_con_send(&session->s_con, msg);
1317 return 0;
1318 }
1319
1320 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1321 struct ceph_mds_session *session, u64 seq)
1322 {
1323 struct ceph_msg *msg;
1324
1325 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1326 session->s_mds, ceph_session_state_name(session->s_state), seq);
1327 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1328 if (!msg)
1329 return -ENOMEM;
1330 ceph_con_send(&session->s_con, msg);
1331 return 0;
1332 }
1333
1334
1335 /*
1336 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1337 *
1338 * Called under session->s_mutex
1339 */
1340 static void renewed_caps(struct ceph_mds_client *mdsc,
1341 struct ceph_mds_session *session, int is_renew)
1342 {
1343 int was_stale;
1344 int wake = 0;
1345
1346 spin_lock(&session->s_cap_lock);
1347 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1348
1349 session->s_cap_ttl = session->s_renew_requested +
1350 mdsc->mdsmap->m_session_timeout*HZ;
1351
1352 if (was_stale) {
1353 if (time_before(jiffies, session->s_cap_ttl)) {
1354 pr_info("mds%d caps renewed\n", session->s_mds);
1355 wake = 1;
1356 } else {
1357 pr_info("mds%d caps still stale\n", session->s_mds);
1358 }
1359 }
1360 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1361 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1362 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1363 spin_unlock(&session->s_cap_lock);
1364
1365 if (wake)
1366 wake_up_session_caps(session, 0);
1367 }
1368
1369 /*
1370 * send a session close request
1371 */
1372 static int request_close_session(struct ceph_mds_client *mdsc,
1373 struct ceph_mds_session *session)
1374 {
1375 struct ceph_msg *msg;
1376
1377 dout("request_close_session mds%d state %s seq %lld\n",
1378 session->s_mds, ceph_session_state_name(session->s_state),
1379 session->s_seq);
1380 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1381 if (!msg)
1382 return -ENOMEM;
1383 ceph_con_send(&session->s_con, msg);
1384 return 1;
1385 }
1386
1387 /*
1388 * Called with s_mutex held.
1389 */
1390 static int __close_session(struct ceph_mds_client *mdsc,
1391 struct ceph_mds_session *session)
1392 {
1393 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1394 return 0;
1395 session->s_state = CEPH_MDS_SESSION_CLOSING;
1396 return request_close_session(mdsc, session);
1397 }
1398
1399 static bool drop_negative_children(struct dentry *dentry)
1400 {
1401 struct dentry *child;
1402 bool all_negative = true;
1403
1404 if (!d_is_dir(dentry))
1405 goto out;
1406
1407 spin_lock(&dentry->d_lock);
1408 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1409 if (d_really_is_positive(child)) {
1410 all_negative = false;
1411 break;
1412 }
1413 }
1414 spin_unlock(&dentry->d_lock);
1415
1416 if (all_negative)
1417 shrink_dcache_parent(dentry);
1418 out:
1419 return all_negative;
1420 }
1421
1422 /*
1423 * Trim old(er) caps.
1424 *
1425 * Because we can't cache an inode without one or more caps, we do
1426 * this indirectly: if a cap is unused, we prune its aliases, at which
1427 * point the inode will hopefully get dropped to.
1428 *
1429 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1430 * memory pressure from the MDS, though, so it needn't be perfect.
1431 */
1432 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1433 {
1434 struct ceph_mds_session *session = arg;
1435 struct ceph_inode_info *ci = ceph_inode(inode);
1436 int used, wanted, oissued, mine;
1437
1438 if (session->s_trim_caps <= 0)
1439 return -1;
1440
1441 spin_lock(&ci->i_ceph_lock);
1442 mine = cap->issued | cap->implemented;
1443 used = __ceph_caps_used(ci);
1444 wanted = __ceph_caps_file_wanted(ci);
1445 oissued = __ceph_caps_issued_other(ci, cap);
1446
1447 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1448 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1449 ceph_cap_string(used), ceph_cap_string(wanted));
1450 if (cap == ci->i_auth_cap) {
1451 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1452 !list_empty(&ci->i_cap_snaps))
1453 goto out;
1454 if ((used | wanted) & CEPH_CAP_ANY_WR)
1455 goto out;
1456 }
1457 /* The inode has cached pages, but it's no longer used.
1458 * we can safely drop it */
1459 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1460 !(oissued & CEPH_CAP_FILE_CACHE)) {
1461 used = 0;
1462 oissued = 0;
1463 }
1464 if ((used | wanted) & ~oissued & mine)
1465 goto out; /* we need these caps */
1466
1467 if (oissued) {
1468 /* we aren't the only cap.. just remove us */
1469 __ceph_remove_cap(cap, true);
1470 session->s_trim_caps--;
1471 } else {
1472 struct dentry *dentry;
1473 /* try dropping referring dentries */
1474 spin_unlock(&ci->i_ceph_lock);
1475 dentry = d_find_any_alias(inode);
1476 if (dentry && drop_negative_children(dentry)) {
1477 int count;
1478 dput(dentry);
1479 d_prune_aliases(inode);
1480 count = atomic_read(&inode->i_count);
1481 if (count == 1)
1482 session->s_trim_caps--;
1483 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1484 inode, cap, count);
1485 } else {
1486 dput(dentry);
1487 }
1488 return 0;
1489 }
1490
1491 out:
1492 spin_unlock(&ci->i_ceph_lock);
1493 return 0;
1494 }
1495
1496 /*
1497 * Trim session cap count down to some max number.
1498 */
1499 static int trim_caps(struct ceph_mds_client *mdsc,
1500 struct ceph_mds_session *session,
1501 int max_caps)
1502 {
1503 int trim_caps = session->s_nr_caps - max_caps;
1504
1505 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1506 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1507 if (trim_caps > 0) {
1508 session->s_trim_caps = trim_caps;
1509 iterate_session_caps(session, trim_caps_cb, session);
1510 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1511 session->s_mds, session->s_nr_caps, max_caps,
1512 trim_caps - session->s_trim_caps);
1513 session->s_trim_caps = 0;
1514 }
1515
1516 ceph_send_cap_releases(mdsc, session);
1517 return 0;
1518 }
1519
1520 static int check_caps_flush(struct ceph_mds_client *mdsc,
1521 u64 want_flush_tid)
1522 {
1523 int ret = 1;
1524
1525 spin_lock(&mdsc->cap_dirty_lock);
1526 if (!list_empty(&mdsc->cap_flush_list)) {
1527 struct ceph_cap_flush *cf =
1528 list_first_entry(&mdsc->cap_flush_list,
1529 struct ceph_cap_flush, g_list);
1530 if (cf->tid <= want_flush_tid) {
1531 dout("check_caps_flush still flushing tid "
1532 "%llu <= %llu\n", cf->tid, want_flush_tid);
1533 ret = 0;
1534 }
1535 }
1536 spin_unlock(&mdsc->cap_dirty_lock);
1537 return ret;
1538 }
1539
1540 /*
1541 * flush all dirty inode data to disk.
1542 *
1543 * returns true if we've flushed through want_flush_tid
1544 */
1545 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1546 u64 want_flush_tid)
1547 {
1548 dout("check_caps_flush want %llu\n", want_flush_tid);
1549
1550 wait_event(mdsc->cap_flushing_wq,
1551 check_caps_flush(mdsc, want_flush_tid));
1552
1553 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1554 }
1555
1556 /*
1557 * called under s_mutex
1558 */
1559 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1560 struct ceph_mds_session *session)
1561 {
1562 struct ceph_msg *msg = NULL;
1563 struct ceph_mds_cap_release *head;
1564 struct ceph_mds_cap_item *item;
1565 struct ceph_cap *cap;
1566 LIST_HEAD(tmp_list);
1567 int num_cap_releases;
1568
1569 spin_lock(&session->s_cap_lock);
1570 again:
1571 list_splice_init(&session->s_cap_releases, &tmp_list);
1572 num_cap_releases = session->s_num_cap_releases;
1573 session->s_num_cap_releases = 0;
1574 spin_unlock(&session->s_cap_lock);
1575
1576 while (!list_empty(&tmp_list)) {
1577 if (!msg) {
1578 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1579 PAGE_SIZE, GFP_NOFS, false);
1580 if (!msg)
1581 goto out_err;
1582 head = msg->front.iov_base;
1583 head->num = cpu_to_le32(0);
1584 msg->front.iov_len = sizeof(*head);
1585 }
1586 cap = list_first_entry(&tmp_list, struct ceph_cap,
1587 session_caps);
1588 list_del(&cap->session_caps);
1589 num_cap_releases--;
1590
1591 head = msg->front.iov_base;
1592 le32_add_cpu(&head->num, 1);
1593 item = msg->front.iov_base + msg->front.iov_len;
1594 item->ino = cpu_to_le64(cap->cap_ino);
1595 item->cap_id = cpu_to_le64(cap->cap_id);
1596 item->migrate_seq = cpu_to_le32(cap->mseq);
1597 item->seq = cpu_to_le32(cap->issue_seq);
1598 msg->front.iov_len += sizeof(*item);
1599
1600 ceph_put_cap(mdsc, cap);
1601
1602 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1603 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1604 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1605 ceph_con_send(&session->s_con, msg);
1606 msg = NULL;
1607 }
1608 }
1609
1610 BUG_ON(num_cap_releases != 0);
1611
1612 spin_lock(&session->s_cap_lock);
1613 if (!list_empty(&session->s_cap_releases))
1614 goto again;
1615 spin_unlock(&session->s_cap_lock);
1616
1617 if (msg) {
1618 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1619 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1620 ceph_con_send(&session->s_con, msg);
1621 }
1622 return;
1623 out_err:
1624 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1625 session->s_mds);
1626 spin_lock(&session->s_cap_lock);
1627 list_splice(&tmp_list, &session->s_cap_releases);
1628 session->s_num_cap_releases += num_cap_releases;
1629 spin_unlock(&session->s_cap_lock);
1630 }
1631
1632 /*
1633 * requests
1634 */
1635
1636 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1637 struct inode *dir)
1638 {
1639 struct ceph_inode_info *ci = ceph_inode(dir);
1640 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1641 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1642 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1643 int order, num_entries;
1644
1645 spin_lock(&ci->i_ceph_lock);
1646 num_entries = ci->i_files + ci->i_subdirs;
1647 spin_unlock(&ci->i_ceph_lock);
1648 num_entries = max(num_entries, 1);
1649 num_entries = min(num_entries, opt->max_readdir);
1650
1651 order = get_order(size * num_entries);
1652 while (order >= 0) {
1653 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1654 __GFP_NOWARN,
1655 order);
1656 if (rinfo->dir_entries)
1657 break;
1658 order--;
1659 }
1660 if (!rinfo->dir_entries)
1661 return -ENOMEM;
1662
1663 num_entries = (PAGE_SIZE << order) / size;
1664 num_entries = min(num_entries, opt->max_readdir);
1665
1666 rinfo->dir_buf_size = PAGE_SIZE << order;
1667 req->r_num_caps = num_entries + 1;
1668 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1669 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1670 return 0;
1671 }
1672
1673 /*
1674 * Create an mds request.
1675 */
1676 struct ceph_mds_request *
1677 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1678 {
1679 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1680
1681 if (!req)
1682 return ERR_PTR(-ENOMEM);
1683
1684 mutex_init(&req->r_fill_mutex);
1685 req->r_mdsc = mdsc;
1686 req->r_started = jiffies;
1687 req->r_resend_mds = -1;
1688 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1689 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1690 req->r_fmode = -1;
1691 kref_init(&req->r_kref);
1692 RB_CLEAR_NODE(&req->r_node);
1693 INIT_LIST_HEAD(&req->r_wait);
1694 init_completion(&req->r_completion);
1695 init_completion(&req->r_safe_completion);
1696 INIT_LIST_HEAD(&req->r_unsafe_item);
1697
1698 req->r_stamp = current_fs_time(mdsc->fsc->sb);
1699
1700 req->r_op = op;
1701 req->r_direct_mode = mode;
1702 return req;
1703 }
1704
1705 /*
1706 * return oldest (lowest) request, tid in request tree, 0 if none.
1707 *
1708 * called under mdsc->mutex.
1709 */
1710 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1711 {
1712 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1713 return NULL;
1714 return rb_entry(rb_first(&mdsc->request_tree),
1715 struct ceph_mds_request, r_node);
1716 }
1717
1718 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1719 {
1720 return mdsc->oldest_tid;
1721 }
1722
1723 /*
1724 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1725 * on build_path_from_dentry in fs/cifs/dir.c.
1726 *
1727 * If @stop_on_nosnap, generate path relative to the first non-snapped
1728 * inode.
1729 *
1730 * Encode hidden .snap dirs as a double /, i.e.
1731 * foo/.snap/bar -> foo//bar
1732 */
1733 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1734 int stop_on_nosnap)
1735 {
1736 struct dentry *temp;
1737 char *path;
1738 int len, pos;
1739 unsigned seq;
1740
1741 if (dentry == NULL)
1742 return ERR_PTR(-EINVAL);
1743
1744 retry:
1745 len = 0;
1746 seq = read_seqbegin(&rename_lock);
1747 rcu_read_lock();
1748 for (temp = dentry; !IS_ROOT(temp);) {
1749 struct inode *inode = d_inode(temp);
1750 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1751 len++; /* slash only */
1752 else if (stop_on_nosnap && inode &&
1753 ceph_snap(inode) == CEPH_NOSNAP)
1754 break;
1755 else
1756 len += 1 + temp->d_name.len;
1757 temp = temp->d_parent;
1758 }
1759 rcu_read_unlock();
1760 if (len)
1761 len--; /* no leading '/' */
1762
1763 path = kmalloc(len+1, GFP_NOFS);
1764 if (path == NULL)
1765 return ERR_PTR(-ENOMEM);
1766 pos = len;
1767 path[pos] = 0; /* trailing null */
1768 rcu_read_lock();
1769 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1770 struct inode *inode;
1771
1772 spin_lock(&temp->d_lock);
1773 inode = d_inode(temp);
1774 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1775 dout("build_path path+%d: %p SNAPDIR\n",
1776 pos, temp);
1777 } else if (stop_on_nosnap && inode &&
1778 ceph_snap(inode) == CEPH_NOSNAP) {
1779 spin_unlock(&temp->d_lock);
1780 break;
1781 } else {
1782 pos -= temp->d_name.len;
1783 if (pos < 0) {
1784 spin_unlock(&temp->d_lock);
1785 break;
1786 }
1787 strncpy(path + pos, temp->d_name.name,
1788 temp->d_name.len);
1789 }
1790 spin_unlock(&temp->d_lock);
1791 if (pos)
1792 path[--pos] = '/';
1793 temp = temp->d_parent;
1794 }
1795 rcu_read_unlock();
1796 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1797 pr_err("build_path did not end path lookup where "
1798 "expected, namelen is %d, pos is %d\n", len, pos);
1799 /* presumably this is only possible if racing with a
1800 rename of one of the parent directories (we can not
1801 lock the dentries above us to prevent this, but
1802 retrying should be harmless) */
1803 kfree(path);
1804 goto retry;
1805 }
1806
1807 *base = ceph_ino(d_inode(temp));
1808 *plen = len;
1809 dout("build_path on %p %d built %llx '%.*s'\n",
1810 dentry, d_count(dentry), *base, len, path);
1811 return path;
1812 }
1813
1814 static int build_dentry_path(struct dentry *dentry,
1815 const char **ppath, int *ppathlen, u64 *pino,
1816 int *pfreepath)
1817 {
1818 char *path;
1819 struct inode *dir;
1820
1821 rcu_read_lock();
1822 dir = d_inode_rcu(dentry->d_parent);
1823 if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1824 *pino = ceph_ino(dir);
1825 rcu_read_unlock();
1826 *ppath = dentry->d_name.name;
1827 *ppathlen = dentry->d_name.len;
1828 return 0;
1829 }
1830 rcu_read_unlock();
1831 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1832 if (IS_ERR(path))
1833 return PTR_ERR(path);
1834 *ppath = path;
1835 *pfreepath = 1;
1836 return 0;
1837 }
1838
1839 static int build_inode_path(struct inode *inode,
1840 const char **ppath, int *ppathlen, u64 *pino,
1841 int *pfreepath)
1842 {
1843 struct dentry *dentry;
1844 char *path;
1845
1846 if (ceph_snap(inode) == CEPH_NOSNAP) {
1847 *pino = ceph_ino(inode);
1848 *ppathlen = 0;
1849 return 0;
1850 }
1851 dentry = d_find_alias(inode);
1852 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1853 dput(dentry);
1854 if (IS_ERR(path))
1855 return PTR_ERR(path);
1856 *ppath = path;
1857 *pfreepath = 1;
1858 return 0;
1859 }
1860
1861 /*
1862 * request arguments may be specified via an inode *, a dentry *, or
1863 * an explicit ino+path.
1864 */
1865 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1866 const char *rpath, u64 rino,
1867 const char **ppath, int *pathlen,
1868 u64 *ino, int *freepath)
1869 {
1870 int r = 0;
1871
1872 if (rinode) {
1873 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1874 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1875 ceph_snap(rinode));
1876 } else if (rdentry) {
1877 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1878 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1879 *ppath);
1880 } else if (rpath || rino) {
1881 *ino = rino;
1882 *ppath = rpath;
1883 *pathlen = rpath ? strlen(rpath) : 0;
1884 dout(" path %.*s\n", *pathlen, rpath);
1885 }
1886
1887 return r;
1888 }
1889
1890 /*
1891 * called under mdsc->mutex
1892 */
1893 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1894 struct ceph_mds_request *req,
1895 int mds, bool drop_cap_releases)
1896 {
1897 struct ceph_msg *msg;
1898 struct ceph_mds_request_head *head;
1899 const char *path1 = NULL;
1900 const char *path2 = NULL;
1901 u64 ino1 = 0, ino2 = 0;
1902 int pathlen1 = 0, pathlen2 = 0;
1903 int freepath1 = 0, freepath2 = 0;
1904 int len;
1905 u16 releases;
1906 void *p, *end;
1907 int ret;
1908
1909 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1910 req->r_path1, req->r_ino1.ino,
1911 &path1, &pathlen1, &ino1, &freepath1);
1912 if (ret < 0) {
1913 msg = ERR_PTR(ret);
1914 goto out;
1915 }
1916
1917 ret = set_request_path_attr(NULL, req->r_old_dentry,
1918 req->r_path2, req->r_ino2.ino,
1919 &path2, &pathlen2, &ino2, &freepath2);
1920 if (ret < 0) {
1921 msg = ERR_PTR(ret);
1922 goto out_free1;
1923 }
1924
1925 len = sizeof(*head) +
1926 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1927 sizeof(struct ceph_timespec);
1928
1929 /* calculate (max) length for cap releases */
1930 len += sizeof(struct ceph_mds_request_release) *
1931 (!!req->r_inode_drop + !!req->r_dentry_drop +
1932 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1933 if (req->r_dentry_drop)
1934 len += req->r_dentry->d_name.len;
1935 if (req->r_old_dentry_drop)
1936 len += req->r_old_dentry->d_name.len;
1937
1938 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1939 if (!msg) {
1940 msg = ERR_PTR(-ENOMEM);
1941 goto out_free2;
1942 }
1943
1944 msg->hdr.version = cpu_to_le16(2);
1945 msg->hdr.tid = cpu_to_le64(req->r_tid);
1946
1947 head = msg->front.iov_base;
1948 p = msg->front.iov_base + sizeof(*head);
1949 end = msg->front.iov_base + msg->front.iov_len;
1950
1951 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1952 head->op = cpu_to_le32(req->r_op);
1953 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1954 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1955 head->args = req->r_args;
1956
1957 ceph_encode_filepath(&p, end, ino1, path1);
1958 ceph_encode_filepath(&p, end, ino2, path2);
1959
1960 /* make note of release offset, in case we need to replay */
1961 req->r_request_release_offset = p - msg->front.iov_base;
1962
1963 /* cap releases */
1964 releases = 0;
1965 if (req->r_inode_drop)
1966 releases += ceph_encode_inode_release(&p,
1967 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1968 mds, req->r_inode_drop, req->r_inode_unless, 0);
1969 if (req->r_dentry_drop)
1970 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1971 mds, req->r_dentry_drop, req->r_dentry_unless);
1972 if (req->r_old_dentry_drop)
1973 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1974 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1975 if (req->r_old_inode_drop)
1976 releases += ceph_encode_inode_release(&p,
1977 d_inode(req->r_old_dentry),
1978 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1979
1980 if (drop_cap_releases) {
1981 releases = 0;
1982 p = msg->front.iov_base + req->r_request_release_offset;
1983 }
1984
1985 head->num_releases = cpu_to_le16(releases);
1986
1987 /* time stamp */
1988 {
1989 struct ceph_timespec ts;
1990 ceph_encode_timespec(&ts, &req->r_stamp);
1991 ceph_encode_copy(&p, &ts, sizeof(ts));
1992 }
1993
1994 BUG_ON(p > end);
1995 msg->front.iov_len = p - msg->front.iov_base;
1996 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1997
1998 if (req->r_pagelist) {
1999 struct ceph_pagelist *pagelist = req->r_pagelist;
2000 atomic_inc(&pagelist->refcnt);
2001 ceph_msg_data_add_pagelist(msg, pagelist);
2002 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2003 } else {
2004 msg->hdr.data_len = 0;
2005 }
2006
2007 msg->hdr.data_off = cpu_to_le16(0);
2008
2009 out_free2:
2010 if (freepath2)
2011 kfree((char *)path2);
2012 out_free1:
2013 if (freepath1)
2014 kfree((char *)path1);
2015 out:
2016 return msg;
2017 }
2018
2019 /*
2020 * called under mdsc->mutex if error, under no mutex if
2021 * success.
2022 */
2023 static void complete_request(struct ceph_mds_client *mdsc,
2024 struct ceph_mds_request *req)
2025 {
2026 if (req->r_callback)
2027 req->r_callback(mdsc, req);
2028 else
2029 complete_all(&req->r_completion);
2030 }
2031
2032 /*
2033 * called under mdsc->mutex
2034 */
2035 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2036 struct ceph_mds_request *req,
2037 int mds, bool drop_cap_releases)
2038 {
2039 struct ceph_mds_request_head *rhead;
2040 struct ceph_msg *msg;
2041 int flags = 0;
2042
2043 req->r_attempts++;
2044 if (req->r_inode) {
2045 struct ceph_cap *cap =
2046 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2047
2048 if (cap)
2049 req->r_sent_on_mseq = cap->mseq;
2050 else
2051 req->r_sent_on_mseq = -1;
2052 }
2053 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2054 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2055
2056 if (req->r_got_unsafe) {
2057 void *p;
2058 /*
2059 * Replay. Do not regenerate message (and rebuild
2060 * paths, etc.); just use the original message.
2061 * Rebuilding paths will break for renames because
2062 * d_move mangles the src name.
2063 */
2064 msg = req->r_request;
2065 rhead = msg->front.iov_base;
2066
2067 flags = le32_to_cpu(rhead->flags);
2068 flags |= CEPH_MDS_FLAG_REPLAY;
2069 rhead->flags = cpu_to_le32(flags);
2070
2071 if (req->r_target_inode)
2072 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2073
2074 rhead->num_retry = req->r_attempts - 1;
2075
2076 /* remove cap/dentry releases from message */
2077 rhead->num_releases = 0;
2078
2079 /* time stamp */
2080 p = msg->front.iov_base + req->r_request_release_offset;
2081 {
2082 struct ceph_timespec ts;
2083 ceph_encode_timespec(&ts, &req->r_stamp);
2084 ceph_encode_copy(&p, &ts, sizeof(ts));
2085 }
2086
2087 msg->front.iov_len = p - msg->front.iov_base;
2088 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2089 return 0;
2090 }
2091
2092 if (req->r_request) {
2093 ceph_msg_put(req->r_request);
2094 req->r_request = NULL;
2095 }
2096 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2097 if (IS_ERR(msg)) {
2098 req->r_err = PTR_ERR(msg);
2099 return PTR_ERR(msg);
2100 }
2101 req->r_request = msg;
2102
2103 rhead = msg->front.iov_base;
2104 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2105 if (req->r_got_unsafe)
2106 flags |= CEPH_MDS_FLAG_REPLAY;
2107 if (req->r_locked_dir)
2108 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2109 rhead->flags = cpu_to_le32(flags);
2110 rhead->num_fwd = req->r_num_fwd;
2111 rhead->num_retry = req->r_attempts - 1;
2112 rhead->ino = 0;
2113
2114 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2115 return 0;
2116 }
2117
2118 /*
2119 * send request, or put it on the appropriate wait list.
2120 */
2121 static int __do_request(struct ceph_mds_client *mdsc,
2122 struct ceph_mds_request *req)
2123 {
2124 struct ceph_mds_session *session = NULL;
2125 int mds = -1;
2126 int err = 0;
2127
2128 if (req->r_err || req->r_got_result) {
2129 if (req->r_aborted)
2130 __unregister_request(mdsc, req);
2131 goto out;
2132 }
2133
2134 if (req->r_timeout &&
2135 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2136 dout("do_request timed out\n");
2137 err = -EIO;
2138 goto finish;
2139 }
2140 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2141 dout("do_request forced umount\n");
2142 err = -EIO;
2143 goto finish;
2144 }
2145
2146 put_request_session(req);
2147
2148 mds = __choose_mds(mdsc, req);
2149 if (mds < 0 ||
2150 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2151 if (mdsc->mdsmap_err) {
2152 err = mdsc->mdsmap_err;
2153 dout("do_request mdsmap err %d\n", err);
2154 goto finish;
2155 }
2156 dout("do_request no mds or not active, waiting for map\n");
2157 list_add(&req->r_wait, &mdsc->waiting_for_map);
2158 goto out;
2159 }
2160
2161 /* get, open session */
2162 session = __ceph_lookup_mds_session(mdsc, mds);
2163 if (!session) {
2164 session = register_session(mdsc, mds);
2165 if (IS_ERR(session)) {
2166 err = PTR_ERR(session);
2167 goto finish;
2168 }
2169 }
2170 req->r_session = get_session(session);
2171
2172 dout("do_request mds%d session %p state %s\n", mds, session,
2173 ceph_session_state_name(session->s_state));
2174 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2175 session->s_state != CEPH_MDS_SESSION_HUNG) {
2176 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2177 err = -EACCES;
2178 goto out_session;
2179 }
2180 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2181 session->s_state == CEPH_MDS_SESSION_CLOSING)
2182 __open_session(mdsc, session);
2183 list_add(&req->r_wait, &session->s_waiting);
2184 goto out_session;
2185 }
2186
2187 /* send request */
2188 req->r_resend_mds = -1; /* forget any previous mds hint */
2189
2190 if (req->r_request_started == 0) /* note request start time */
2191 req->r_request_started = jiffies;
2192
2193 err = __prepare_send_request(mdsc, req, mds, false);
2194 if (!err) {
2195 ceph_msg_get(req->r_request);
2196 ceph_con_send(&session->s_con, req->r_request);
2197 }
2198
2199 out_session:
2200 ceph_put_mds_session(session);
2201 finish:
2202 if (err) {
2203 dout("__do_request early error %d\n", err);
2204 req->r_err = err;
2205 complete_request(mdsc, req);
2206 __unregister_request(mdsc, req);
2207 }
2208 out:
2209 return err;
2210 }
2211
2212 /*
2213 * called under mdsc->mutex
2214 */
2215 static void __wake_requests(struct ceph_mds_client *mdsc,
2216 struct list_head *head)
2217 {
2218 struct ceph_mds_request *req;
2219 LIST_HEAD(tmp_list);
2220
2221 list_splice_init(head, &tmp_list);
2222
2223 while (!list_empty(&tmp_list)) {
2224 req = list_entry(tmp_list.next,
2225 struct ceph_mds_request, r_wait);
2226 list_del_init(&req->r_wait);
2227 dout(" wake request %p tid %llu\n", req, req->r_tid);
2228 __do_request(mdsc, req);
2229 }
2230 }
2231
2232 /*
2233 * Wake up threads with requests pending for @mds, so that they can
2234 * resubmit their requests to a possibly different mds.
2235 */
2236 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2237 {
2238 struct ceph_mds_request *req;
2239 struct rb_node *p = rb_first(&mdsc->request_tree);
2240
2241 dout("kick_requests mds%d\n", mds);
2242 while (p) {
2243 req = rb_entry(p, struct ceph_mds_request, r_node);
2244 p = rb_next(p);
2245 if (req->r_got_unsafe)
2246 continue;
2247 if (req->r_attempts > 0)
2248 continue; /* only new requests */
2249 if (req->r_session &&
2250 req->r_session->s_mds == mds) {
2251 dout(" kicking tid %llu\n", req->r_tid);
2252 list_del_init(&req->r_wait);
2253 __do_request(mdsc, req);
2254 }
2255 }
2256 }
2257
2258 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2259 struct ceph_mds_request *req)
2260 {
2261 dout("submit_request on %p\n", req);
2262 mutex_lock(&mdsc->mutex);
2263 __register_request(mdsc, req, NULL);
2264 __do_request(mdsc, req);
2265 mutex_unlock(&mdsc->mutex);
2266 }
2267
2268 /*
2269 * Synchrously perform an mds request. Take care of all of the
2270 * session setup, forwarding, retry details.
2271 */
2272 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2273 struct inode *dir,
2274 struct ceph_mds_request *req)
2275 {
2276 int err;
2277
2278 dout("do_request on %p\n", req);
2279
2280 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2281 if (req->r_inode)
2282 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2283 if (req->r_locked_dir)
2284 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2285 if (req->r_old_dentry_dir)
2286 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2287 CEPH_CAP_PIN);
2288
2289 /* issue */
2290 mutex_lock(&mdsc->mutex);
2291 __register_request(mdsc, req, dir);
2292 __do_request(mdsc, req);
2293
2294 if (req->r_err) {
2295 err = req->r_err;
2296 goto out;
2297 }
2298
2299 /* wait */
2300 mutex_unlock(&mdsc->mutex);
2301 dout("do_request waiting\n");
2302 if (!req->r_timeout && req->r_wait_for_completion) {
2303 err = req->r_wait_for_completion(mdsc, req);
2304 } else {
2305 long timeleft = wait_for_completion_killable_timeout(
2306 &req->r_completion,
2307 ceph_timeout_jiffies(req->r_timeout));
2308 if (timeleft > 0)
2309 err = 0;
2310 else if (!timeleft)
2311 err = -EIO; /* timed out */
2312 else
2313 err = timeleft; /* killed */
2314 }
2315 dout("do_request waited, got %d\n", err);
2316 mutex_lock(&mdsc->mutex);
2317
2318 /* only abort if we didn't race with a real reply */
2319 if (req->r_got_result) {
2320 err = le32_to_cpu(req->r_reply_info.head->result);
2321 } else if (err < 0) {
2322 dout("aborted request %lld with %d\n", req->r_tid, err);
2323
2324 /*
2325 * ensure we aren't running concurrently with
2326 * ceph_fill_trace or ceph_readdir_prepopulate, which
2327 * rely on locks (dir mutex) held by our caller.
2328 */
2329 mutex_lock(&req->r_fill_mutex);
2330 req->r_err = err;
2331 req->r_aborted = true;
2332 mutex_unlock(&req->r_fill_mutex);
2333
2334 if (req->r_locked_dir &&
2335 (req->r_op & CEPH_MDS_OP_WRITE))
2336 ceph_invalidate_dir_request(req);
2337 } else {
2338 err = req->r_err;
2339 }
2340
2341 out:
2342 mutex_unlock(&mdsc->mutex);
2343 dout("do_request %p done, result %d\n", req, err);
2344 return err;
2345 }
2346
2347 /*
2348 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2349 * namespace request.
2350 */
2351 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2352 {
2353 struct inode *inode = req->r_locked_dir;
2354
2355 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2356
2357 ceph_dir_clear_complete(inode);
2358 if (req->r_dentry)
2359 ceph_invalidate_dentry_lease(req->r_dentry);
2360 if (req->r_old_dentry)
2361 ceph_invalidate_dentry_lease(req->r_old_dentry);
2362 }
2363
2364 /*
2365 * Handle mds reply.
2366 *
2367 * We take the session mutex and parse and process the reply immediately.
2368 * This preserves the logical ordering of replies, capabilities, etc., sent
2369 * by the MDS as they are applied to our local cache.
2370 */
2371 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2372 {
2373 struct ceph_mds_client *mdsc = session->s_mdsc;
2374 struct ceph_mds_request *req;
2375 struct ceph_mds_reply_head *head = msg->front.iov_base;
2376 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2377 struct ceph_snap_realm *realm;
2378 u64 tid;
2379 int err, result;
2380 int mds = session->s_mds;
2381
2382 if (msg->front.iov_len < sizeof(*head)) {
2383 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2384 ceph_msg_dump(msg);
2385 return;
2386 }
2387
2388 /* get request, session */
2389 tid = le64_to_cpu(msg->hdr.tid);
2390 mutex_lock(&mdsc->mutex);
2391 req = lookup_get_request(mdsc, tid);
2392 if (!req) {
2393 dout("handle_reply on unknown tid %llu\n", tid);
2394 mutex_unlock(&mdsc->mutex);
2395 return;
2396 }
2397 dout("handle_reply %p\n", req);
2398
2399 /* correct session? */
2400 if (req->r_session != session) {
2401 pr_err("mdsc_handle_reply got %llu on session mds%d"
2402 " not mds%d\n", tid, session->s_mds,
2403 req->r_session ? req->r_session->s_mds : -1);
2404 mutex_unlock(&mdsc->mutex);
2405 goto out;
2406 }
2407
2408 /* dup? */
2409 if ((req->r_got_unsafe && !head->safe) ||
2410 (req->r_got_safe && head->safe)) {
2411 pr_warn("got a dup %s reply on %llu from mds%d\n",
2412 head->safe ? "safe" : "unsafe", tid, mds);
2413 mutex_unlock(&mdsc->mutex);
2414 goto out;
2415 }
2416 if (req->r_got_safe) {
2417 pr_warn("got unsafe after safe on %llu from mds%d\n",
2418 tid, mds);
2419 mutex_unlock(&mdsc->mutex);
2420 goto out;
2421 }
2422
2423 result = le32_to_cpu(head->result);
2424
2425 /*
2426 * Handle an ESTALE
2427 * if we're not talking to the authority, send to them
2428 * if the authority has changed while we weren't looking,
2429 * send to new authority
2430 * Otherwise we just have to return an ESTALE
2431 */
2432 if (result == -ESTALE) {
2433 dout("got ESTALE on request %llu", req->r_tid);
2434 req->r_resend_mds = -1;
2435 if (req->r_direct_mode != USE_AUTH_MDS) {
2436 dout("not using auth, setting for that now");
2437 req->r_direct_mode = USE_AUTH_MDS;
2438 __do_request(mdsc, req);
2439 mutex_unlock(&mdsc->mutex);
2440 goto out;
2441 } else {
2442 int mds = __choose_mds(mdsc, req);
2443 if (mds >= 0 && mds != req->r_session->s_mds) {
2444 dout("but auth changed, so resending");
2445 __do_request(mdsc, req);
2446 mutex_unlock(&mdsc->mutex);
2447 goto out;
2448 }
2449 }
2450 dout("have to return ESTALE on request %llu", req->r_tid);
2451 }
2452
2453
2454 if (head->safe) {
2455 req->r_got_safe = true;
2456 __unregister_request(mdsc, req);
2457
2458 if (req->r_got_unsafe) {
2459 /*
2460 * We already handled the unsafe response, now do the
2461 * cleanup. No need to examine the response; the MDS
2462 * doesn't include any result info in the safe
2463 * response. And even if it did, there is nothing
2464 * useful we could do with a revised return value.
2465 */
2466 dout("got safe reply %llu, mds%d\n", tid, mds);
2467
2468 /* last unsafe request during umount? */
2469 if (mdsc->stopping && !__get_oldest_req(mdsc))
2470 complete_all(&mdsc->safe_umount_waiters);
2471 mutex_unlock(&mdsc->mutex);
2472 goto out;
2473 }
2474 } else {
2475 req->r_got_unsafe = true;
2476 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2477 if (req->r_unsafe_dir) {
2478 struct ceph_inode_info *ci =
2479 ceph_inode(req->r_unsafe_dir);
2480 spin_lock(&ci->i_unsafe_lock);
2481 list_add_tail(&req->r_unsafe_dir_item,
2482 &ci->i_unsafe_dirops);
2483 spin_unlock(&ci->i_unsafe_lock);
2484 }
2485 }
2486
2487 dout("handle_reply tid %lld result %d\n", tid, result);
2488 rinfo = &req->r_reply_info;
2489 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2490 mutex_unlock(&mdsc->mutex);
2491
2492 mutex_lock(&session->s_mutex);
2493 if (err < 0) {
2494 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2495 ceph_msg_dump(msg);
2496 goto out_err;
2497 }
2498
2499 /* snap trace */
2500 realm = NULL;
2501 if (rinfo->snapblob_len) {
2502 down_write(&mdsc->snap_rwsem);
2503 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2504 rinfo->snapblob + rinfo->snapblob_len,
2505 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2506 &realm);
2507 downgrade_write(&mdsc->snap_rwsem);
2508 } else {
2509 down_read(&mdsc->snap_rwsem);
2510 }
2511
2512 /* insert trace into our cache */
2513 mutex_lock(&req->r_fill_mutex);
2514 current->journal_info = req;
2515 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2516 if (err == 0) {
2517 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2518 req->r_op == CEPH_MDS_OP_LSSNAP))
2519 ceph_readdir_prepopulate(req, req->r_session);
2520 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2521 }
2522 current->journal_info = NULL;
2523 mutex_unlock(&req->r_fill_mutex);
2524
2525 up_read(&mdsc->snap_rwsem);
2526 if (realm)
2527 ceph_put_snap_realm(mdsc, realm);
2528
2529 if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2530 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2531 spin_lock(&ci->i_unsafe_lock);
2532 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2533 spin_unlock(&ci->i_unsafe_lock);
2534 }
2535 out_err:
2536 mutex_lock(&mdsc->mutex);
2537 if (!req->r_aborted) {
2538 if (err) {
2539 req->r_err = err;
2540 } else {
2541 req->r_reply = ceph_msg_get(msg);
2542 req->r_got_result = true;
2543 }
2544 } else {
2545 dout("reply arrived after request %lld was aborted\n", tid);
2546 }
2547 mutex_unlock(&mdsc->mutex);
2548
2549 mutex_unlock(&session->s_mutex);
2550
2551 /* kick calling process */
2552 complete_request(mdsc, req);
2553 out:
2554 ceph_mdsc_put_request(req);
2555 return;
2556 }
2557
2558
2559
2560 /*
2561 * handle mds notification that our request has been forwarded.
2562 */
2563 static void handle_forward(struct ceph_mds_client *mdsc,
2564 struct ceph_mds_session *session,
2565 struct ceph_msg *msg)
2566 {
2567 struct ceph_mds_request *req;
2568 u64 tid = le64_to_cpu(msg->hdr.tid);
2569 u32 next_mds;
2570 u32 fwd_seq;
2571 int err = -EINVAL;
2572 void *p = msg->front.iov_base;
2573 void *end = p + msg->front.iov_len;
2574
2575 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2576 next_mds = ceph_decode_32(&p);
2577 fwd_seq = ceph_decode_32(&p);
2578
2579 mutex_lock(&mdsc->mutex);
2580 req = lookup_get_request(mdsc, tid);
2581 if (!req) {
2582 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2583 goto out; /* dup reply? */
2584 }
2585
2586 if (req->r_aborted) {
2587 dout("forward tid %llu aborted, unregistering\n", tid);
2588 __unregister_request(mdsc, req);
2589 } else if (fwd_seq <= req->r_num_fwd) {
2590 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2591 tid, next_mds, req->r_num_fwd, fwd_seq);
2592 } else {
2593 /* resend. forward race not possible; mds would drop */
2594 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2595 BUG_ON(req->r_err);
2596 BUG_ON(req->r_got_result);
2597 req->r_attempts = 0;
2598 req->r_num_fwd = fwd_seq;
2599 req->r_resend_mds = next_mds;
2600 put_request_session(req);
2601 __do_request(mdsc, req);
2602 }
2603 ceph_mdsc_put_request(req);
2604 out:
2605 mutex_unlock(&mdsc->mutex);
2606 return;
2607
2608 bad:
2609 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2610 }
2611
2612 /*
2613 * handle a mds session control message
2614 */
2615 static void handle_session(struct ceph_mds_session *session,
2616 struct ceph_msg *msg)
2617 {
2618 struct ceph_mds_client *mdsc = session->s_mdsc;
2619 u32 op;
2620 u64 seq;
2621 int mds = session->s_mds;
2622 struct ceph_mds_session_head *h = msg->front.iov_base;
2623 int wake = 0;
2624
2625 /* decode */
2626 if (msg->front.iov_len != sizeof(*h))
2627 goto bad;
2628 op = le32_to_cpu(h->op);
2629 seq = le64_to_cpu(h->seq);
2630
2631 mutex_lock(&mdsc->mutex);
2632 if (op == CEPH_SESSION_CLOSE)
2633 __unregister_session(mdsc, session);
2634 /* FIXME: this ttl calculation is generous */
2635 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2636 mutex_unlock(&mdsc->mutex);
2637
2638 mutex_lock(&session->s_mutex);
2639
2640 dout("handle_session mds%d %s %p state %s seq %llu\n",
2641 mds, ceph_session_op_name(op), session,
2642 ceph_session_state_name(session->s_state), seq);
2643
2644 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2645 session->s_state = CEPH_MDS_SESSION_OPEN;
2646 pr_info("mds%d came back\n", session->s_mds);
2647 }
2648
2649 switch (op) {
2650 case CEPH_SESSION_OPEN:
2651 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2652 pr_info("mds%d reconnect success\n", session->s_mds);
2653 session->s_state = CEPH_MDS_SESSION_OPEN;
2654 renewed_caps(mdsc, session, 0);
2655 wake = 1;
2656 if (mdsc->stopping)
2657 __close_session(mdsc, session);
2658 break;
2659
2660 case CEPH_SESSION_RENEWCAPS:
2661 if (session->s_renew_seq == seq)
2662 renewed_caps(mdsc, session, 1);
2663 break;
2664
2665 case CEPH_SESSION_CLOSE:
2666 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2667 pr_info("mds%d reconnect denied\n", session->s_mds);
2668 cleanup_session_requests(mdsc, session);
2669 remove_session_caps(session);
2670 wake = 2; /* for good measure */
2671 wake_up_all(&mdsc->session_close_wq);
2672 break;
2673
2674 case CEPH_SESSION_STALE:
2675 pr_info("mds%d caps went stale, renewing\n",
2676 session->s_mds);
2677 spin_lock(&session->s_gen_ttl_lock);
2678 session->s_cap_gen++;
2679 session->s_cap_ttl = jiffies - 1;
2680 spin_unlock(&session->s_gen_ttl_lock);
2681 send_renew_caps(mdsc, session);
2682 break;
2683
2684 case CEPH_SESSION_RECALL_STATE:
2685 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2686 break;
2687
2688 case CEPH_SESSION_FLUSHMSG:
2689 send_flushmsg_ack(mdsc, session, seq);
2690 break;
2691
2692 case CEPH_SESSION_FORCE_RO:
2693 dout("force_session_readonly %p\n", session);
2694 spin_lock(&session->s_cap_lock);
2695 session->s_readonly = true;
2696 spin_unlock(&session->s_cap_lock);
2697 wake_up_session_caps(session, 0);
2698 break;
2699
2700 case CEPH_SESSION_REJECT:
2701 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2702 pr_info("mds%d rejected session\n", session->s_mds);
2703 session->s_state = CEPH_MDS_SESSION_REJECTED;
2704 cleanup_session_requests(mdsc, session);
2705 remove_session_caps(session);
2706 wake = 2; /* for good measure */
2707 break;
2708
2709 default:
2710 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2711 WARN_ON(1);
2712 }
2713
2714 mutex_unlock(&session->s_mutex);
2715 if (wake) {
2716 mutex_lock(&mdsc->mutex);
2717 __wake_requests(mdsc, &session->s_waiting);
2718 if (wake == 2)
2719 kick_requests(mdsc, mds);
2720 mutex_unlock(&mdsc->mutex);
2721 }
2722 return;
2723
2724 bad:
2725 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2726 (int)msg->front.iov_len);
2727 ceph_msg_dump(msg);
2728 return;
2729 }
2730
2731
2732 /*
2733 * called under session->mutex.
2734 */
2735 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2736 struct ceph_mds_session *session)
2737 {
2738 struct ceph_mds_request *req, *nreq;
2739 struct rb_node *p;
2740 int err;
2741
2742 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2743
2744 mutex_lock(&mdsc->mutex);
2745 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2746 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2747 if (!err) {
2748 ceph_msg_get(req->r_request);
2749 ceph_con_send(&session->s_con, req->r_request);
2750 }
2751 }
2752
2753 /*
2754 * also re-send old requests when MDS enters reconnect stage. So that MDS
2755 * can process completed request in clientreplay stage.
2756 */
2757 p = rb_first(&mdsc->request_tree);
2758 while (p) {
2759 req = rb_entry(p, struct ceph_mds_request, r_node);
2760 p = rb_next(p);
2761 if (req->r_got_unsafe)
2762 continue;
2763 if (req->r_attempts == 0)
2764 continue; /* only old requests */
2765 if (req->r_session &&
2766 req->r_session->s_mds == session->s_mds) {
2767 err = __prepare_send_request(mdsc, req,
2768 session->s_mds, true);
2769 if (!err) {
2770 ceph_msg_get(req->r_request);
2771 ceph_con_send(&session->s_con, req->r_request);
2772 }
2773 }
2774 }
2775 mutex_unlock(&mdsc->mutex);
2776 }
2777
2778 /*
2779 * Encode information about a cap for a reconnect with the MDS.
2780 */
2781 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2782 void *arg)
2783 {
2784 union {
2785 struct ceph_mds_cap_reconnect v2;
2786 struct ceph_mds_cap_reconnect_v1 v1;
2787 } rec;
2788 struct ceph_inode_info *ci;
2789 struct ceph_reconnect_state *recon_state = arg;
2790 struct ceph_pagelist *pagelist = recon_state->pagelist;
2791 char *path;
2792 int pathlen, err;
2793 u64 pathbase;
2794 u64 snap_follows;
2795 struct dentry *dentry;
2796
2797 ci = cap->ci;
2798
2799 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2800 inode, ceph_vinop(inode), cap, cap->cap_id,
2801 ceph_cap_string(cap->issued));
2802 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2803 if (err)
2804 return err;
2805
2806 dentry = d_find_alias(inode);
2807 if (dentry) {
2808 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2809 if (IS_ERR(path)) {
2810 err = PTR_ERR(path);
2811 goto out_dput;
2812 }
2813 } else {
2814 path = NULL;
2815 pathlen = 0;
2816 pathbase = 0;
2817 }
2818
2819 spin_lock(&ci->i_ceph_lock);
2820 cap->seq = 0; /* reset cap seq */
2821 cap->issue_seq = 0; /* and issue_seq */
2822 cap->mseq = 0; /* and migrate_seq */
2823 cap->cap_gen = cap->session->s_cap_gen;
2824
2825 if (recon_state->msg_version >= 2) {
2826 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2827 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2828 rec.v2.issued = cpu_to_le32(cap->issued);
2829 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2830 rec.v2.pathbase = cpu_to_le64(pathbase);
2831 rec.v2.flock_len = 0;
2832 } else {
2833 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2834 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2835 rec.v1.issued = cpu_to_le32(cap->issued);
2836 rec.v1.size = cpu_to_le64(inode->i_size);
2837 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2838 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2839 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2840 rec.v1.pathbase = cpu_to_le64(pathbase);
2841 }
2842
2843 if (list_empty(&ci->i_cap_snaps)) {
2844 snap_follows = 0;
2845 } else {
2846 struct ceph_cap_snap *capsnap =
2847 list_first_entry(&ci->i_cap_snaps,
2848 struct ceph_cap_snap, ci_item);
2849 snap_follows = capsnap->follows;
2850 }
2851 spin_unlock(&ci->i_ceph_lock);
2852
2853 if (recon_state->msg_version >= 2) {
2854 int num_fcntl_locks, num_flock_locks;
2855 struct ceph_filelock *flocks;
2856 size_t struct_len, total_len = 0;
2857 u8 struct_v = 0;
2858
2859 encode_again:
2860 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2861 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2862 sizeof(struct ceph_filelock), GFP_NOFS);
2863 if (!flocks) {
2864 err = -ENOMEM;
2865 goto out_free;
2866 }
2867 err = ceph_encode_locks_to_buffer(inode, flocks,
2868 num_fcntl_locks,
2869 num_flock_locks);
2870 if (err) {
2871 kfree(flocks);
2872 if (err == -ENOSPC)
2873 goto encode_again;
2874 goto out_free;
2875 }
2876
2877 if (recon_state->msg_version >= 3) {
2878 /* version, compat_version and struct_len */
2879 total_len = 2 * sizeof(u8) + sizeof(u32);
2880 struct_v = 2;
2881 }
2882 /*
2883 * number of encoded locks is stable, so copy to pagelist
2884 */
2885 struct_len = 2 * sizeof(u32) +
2886 (num_fcntl_locks + num_flock_locks) *
2887 sizeof(struct ceph_filelock);
2888 rec.v2.flock_len = cpu_to_le32(struct_len);
2889
2890 struct_len += sizeof(rec.v2);
2891 struct_len += sizeof(u32) + pathlen;
2892
2893 if (struct_v >= 2)
2894 struct_len += sizeof(u64); /* snap_follows */
2895
2896 total_len += struct_len;
2897 err = ceph_pagelist_reserve(pagelist, total_len);
2898
2899 if (!err) {
2900 if (recon_state->msg_version >= 3) {
2901 ceph_pagelist_encode_8(pagelist, struct_v);
2902 ceph_pagelist_encode_8(pagelist, 1);
2903 ceph_pagelist_encode_32(pagelist, struct_len);
2904 }
2905 ceph_pagelist_encode_string(pagelist, path, pathlen);
2906 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2907 ceph_locks_to_pagelist(flocks, pagelist,
2908 num_fcntl_locks,
2909 num_flock_locks);
2910 if (struct_v >= 2)
2911 ceph_pagelist_encode_64(pagelist, snap_follows);
2912 }
2913 kfree(flocks);
2914 } else {
2915 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2916 err = ceph_pagelist_reserve(pagelist, size);
2917 if (!err) {
2918 ceph_pagelist_encode_string(pagelist, path, pathlen);
2919 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2920 }
2921 }
2922
2923 recon_state->nr_caps++;
2924 out_free:
2925 kfree(path);
2926 out_dput:
2927 dput(dentry);
2928 return err;
2929 }
2930
2931
2932 /*
2933 * If an MDS fails and recovers, clients need to reconnect in order to
2934 * reestablish shared state. This includes all caps issued through
2935 * this session _and_ the snap_realm hierarchy. Because it's not
2936 * clear which snap realms the mds cares about, we send everything we
2937 * know about.. that ensures we'll then get any new info the
2938 * recovering MDS might have.
2939 *
2940 * This is a relatively heavyweight operation, but it's rare.
2941 *
2942 * called with mdsc->mutex held.
2943 */
2944 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2945 struct ceph_mds_session *session)
2946 {
2947 struct ceph_msg *reply;
2948 struct rb_node *p;
2949 int mds = session->s_mds;
2950 int err = -ENOMEM;
2951 int s_nr_caps;
2952 struct ceph_pagelist *pagelist;
2953 struct ceph_reconnect_state recon_state;
2954
2955 pr_info("mds%d reconnect start\n", mds);
2956
2957 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2958 if (!pagelist)
2959 goto fail_nopagelist;
2960 ceph_pagelist_init(pagelist);
2961
2962 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2963 if (!reply)
2964 goto fail_nomsg;
2965
2966 mutex_lock(&session->s_mutex);
2967 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2968 session->s_seq = 0;
2969
2970 dout("session %p state %s\n", session,
2971 ceph_session_state_name(session->s_state));
2972
2973 spin_lock(&session->s_gen_ttl_lock);
2974 session->s_cap_gen++;
2975 spin_unlock(&session->s_gen_ttl_lock);
2976
2977 spin_lock(&session->s_cap_lock);
2978 /* don't know if session is readonly */
2979 session->s_readonly = 0;
2980 /*
2981 * notify __ceph_remove_cap() that we are composing cap reconnect.
2982 * If a cap get released before being added to the cap reconnect,
2983 * __ceph_remove_cap() should skip queuing cap release.
2984 */
2985 session->s_cap_reconnect = 1;
2986 /* drop old cap expires; we're about to reestablish that state */
2987 cleanup_cap_releases(mdsc, session);
2988
2989 /* trim unused caps to reduce MDS's cache rejoin time */
2990 if (mdsc->fsc->sb->s_root)
2991 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2992
2993 ceph_con_close(&session->s_con);
2994 ceph_con_open(&session->s_con,
2995 CEPH_ENTITY_TYPE_MDS, mds,
2996 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2997
2998 /* replay unsafe requests */
2999 replay_unsafe_requests(mdsc, session);
3000
3001 down_read(&mdsc->snap_rwsem);
3002
3003 /* traverse this session's caps */
3004 s_nr_caps = session->s_nr_caps;
3005 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3006 if (err)
3007 goto fail;
3008
3009 recon_state.nr_caps = 0;
3010 recon_state.pagelist = pagelist;
3011 if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3012 recon_state.msg_version = 3;
3013 else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3014 recon_state.msg_version = 2;
3015 else
3016 recon_state.msg_version = 1;
3017 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3018 if (err < 0)
3019 goto fail;
3020
3021 spin_lock(&session->s_cap_lock);
3022 session->s_cap_reconnect = 0;
3023 spin_unlock(&session->s_cap_lock);
3024
3025 /*
3026 * snaprealms. we provide mds with the ino, seq (version), and
3027 * parent for all of our realms. If the mds has any newer info,
3028 * it will tell us.
3029 */
3030 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3031 struct ceph_snap_realm *realm =
3032 rb_entry(p, struct ceph_snap_realm, node);
3033 struct ceph_mds_snaprealm_reconnect sr_rec;
3034
3035 dout(" adding snap realm %llx seq %lld parent %llx\n",
3036 realm->ino, realm->seq, realm->parent_ino);
3037 sr_rec.ino = cpu_to_le64(realm->ino);
3038 sr_rec.seq = cpu_to_le64(realm->seq);
3039 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3040 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3041 if (err)
3042 goto fail;
3043 }
3044
3045 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3046
3047 /* raced with cap release? */
3048 if (s_nr_caps != recon_state.nr_caps) {
3049 struct page *page = list_first_entry(&pagelist->head,
3050 struct page, lru);
3051 __le32 *addr = kmap_atomic(page);
3052 *addr = cpu_to_le32(recon_state.nr_caps);
3053 kunmap_atomic(addr);
3054 }
3055
3056 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3057 ceph_msg_data_add_pagelist(reply, pagelist);
3058
3059 ceph_early_kick_flushing_caps(mdsc, session);
3060
3061 ceph_con_send(&session->s_con, reply);
3062
3063 mutex_unlock(&session->s_mutex);
3064
3065 mutex_lock(&mdsc->mutex);
3066 __wake_requests(mdsc, &session->s_waiting);
3067 mutex_unlock(&mdsc->mutex);
3068
3069 up_read(&mdsc->snap_rwsem);
3070 return;
3071
3072 fail:
3073 ceph_msg_put(reply);
3074 up_read(&mdsc->snap_rwsem);
3075 mutex_unlock(&session->s_mutex);
3076 fail_nomsg:
3077 ceph_pagelist_release(pagelist);
3078 fail_nopagelist:
3079 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3080 return;
3081 }
3082
3083
3084 /*
3085 * compare old and new mdsmaps, kicking requests
3086 * and closing out old connections as necessary
3087 *
3088 * called under mdsc->mutex.
3089 */
3090 static void check_new_map(struct ceph_mds_client *mdsc,
3091 struct ceph_mdsmap *newmap,
3092 struct ceph_mdsmap *oldmap)
3093 {
3094 int i;
3095 int oldstate, newstate;
3096 struct ceph_mds_session *s;
3097
3098 dout("check_new_map new %u old %u\n",
3099 newmap->m_epoch, oldmap->m_epoch);
3100
3101 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3102 if (mdsc->sessions[i] == NULL)
3103 continue;
3104 s = mdsc->sessions[i];
3105 oldstate = ceph_mdsmap_get_state(oldmap, i);
3106 newstate = ceph_mdsmap_get_state(newmap, i);
3107
3108 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3109 i, ceph_mds_state_name(oldstate),
3110 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3111 ceph_mds_state_name(newstate),
3112 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3113 ceph_session_state_name(s->s_state));
3114
3115 if (i >= newmap->m_max_mds ||
3116 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3117 ceph_mdsmap_get_addr(newmap, i),
3118 sizeof(struct ceph_entity_addr))) {
3119 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3120 /* the session never opened, just close it
3121 * out now */
3122 __wake_requests(mdsc, &s->s_waiting);
3123 __unregister_session(mdsc, s);
3124 } else {
3125 /* just close it */
3126 mutex_unlock(&mdsc->mutex);
3127 mutex_lock(&s->s_mutex);
3128 mutex_lock(&mdsc->mutex);
3129 ceph_con_close(&s->s_con);
3130 mutex_unlock(&s->s_mutex);
3131 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3132 }
3133 } else if (oldstate == newstate) {
3134 continue; /* nothing new with this mds */
3135 }
3136
3137 /*
3138 * send reconnect?
3139 */
3140 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3141 newstate >= CEPH_MDS_STATE_RECONNECT) {
3142 mutex_unlock(&mdsc->mutex);
3143 send_mds_reconnect(mdsc, s);
3144 mutex_lock(&mdsc->mutex);
3145 }
3146
3147 /*
3148 * kick request on any mds that has gone active.
3149 */
3150 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3151 newstate >= CEPH_MDS_STATE_ACTIVE) {
3152 if (oldstate != CEPH_MDS_STATE_CREATING &&
3153 oldstate != CEPH_MDS_STATE_STARTING)
3154 pr_info("mds%d recovery completed\n", s->s_mds);
3155 kick_requests(mdsc, i);
3156 ceph_kick_flushing_caps(mdsc, s);
3157 wake_up_session_caps(s, 1);
3158 }
3159 }
3160
3161 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3162 s = mdsc->sessions[i];
3163 if (!s)
3164 continue;
3165 if (!ceph_mdsmap_is_laggy(newmap, i))
3166 continue;
3167 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3168 s->s_state == CEPH_MDS_SESSION_HUNG ||
3169 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3170 dout(" connecting to export targets of laggy mds%d\n",
3171 i);
3172 __open_export_target_sessions(mdsc, s);
3173 }
3174 }
3175 }
3176
3177
3178
3179 /*
3180 * leases
3181 */
3182
3183 /*
3184 * caller must hold session s_mutex, dentry->d_lock
3185 */
3186 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3187 {
3188 struct ceph_dentry_info *di = ceph_dentry(dentry);
3189
3190 ceph_put_mds_session(di->lease_session);
3191 di->lease_session = NULL;
3192 }
3193
3194 static void handle_lease(struct ceph_mds_client *mdsc,
3195 struct ceph_mds_session *session,
3196 struct ceph_msg *msg)
3197 {
3198 struct super_block *sb = mdsc->fsc->sb;
3199 struct inode *inode;
3200 struct dentry *parent, *dentry;
3201 struct ceph_dentry_info *di;
3202 int mds = session->s_mds;
3203 struct ceph_mds_lease *h = msg->front.iov_base;
3204 u32 seq;
3205 struct ceph_vino vino;
3206 struct qstr dname;
3207 int release = 0;
3208
3209 dout("handle_lease from mds%d\n", mds);
3210
3211 /* decode */
3212 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3213 goto bad;
3214 vino.ino = le64_to_cpu(h->ino);
3215 vino.snap = CEPH_NOSNAP;
3216 seq = le32_to_cpu(h->seq);
3217 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3218 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3219 if (dname.len != get_unaligned_le32(h+1))
3220 goto bad;
3221
3222 /* lookup inode */
3223 inode = ceph_find_inode(sb, vino);
3224 dout("handle_lease %s, ino %llx %p %.*s\n",
3225 ceph_lease_op_name(h->action), vino.ino, inode,
3226 dname.len, dname.name);
3227
3228 mutex_lock(&session->s_mutex);
3229 session->s_seq++;
3230
3231 if (inode == NULL) {
3232 dout("handle_lease no inode %llx\n", vino.ino);
3233 goto release;
3234 }
3235
3236 /* dentry */
3237 parent = d_find_alias(inode);
3238 if (!parent) {
3239 dout("no parent dentry on inode %p\n", inode);
3240 WARN_ON(1);
3241 goto release; /* hrm... */
3242 }
3243 dname.hash = full_name_hash(parent, dname.name, dname.len);
3244 dentry = d_lookup(parent, &dname);
3245 dput(parent);
3246 if (!dentry)
3247 goto release;
3248
3249 spin_lock(&dentry->d_lock);
3250 di = ceph_dentry(dentry);
3251 switch (h->action) {
3252 case CEPH_MDS_LEASE_REVOKE:
3253 if (di->lease_session == session) {
3254 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3255 h->seq = cpu_to_le32(di->lease_seq);
3256 __ceph_mdsc_drop_dentry_lease(dentry);
3257 }
3258 release = 1;
3259 break;
3260
3261 case CEPH_MDS_LEASE_RENEW:
3262 if (di->lease_session == session &&
3263 di->lease_gen == session->s_cap_gen &&
3264 di->lease_renew_from &&
3265 di->lease_renew_after == 0) {
3266 unsigned long duration =
3267 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3268
3269 di->lease_seq = seq;
3270 di->time = di->lease_renew_from + duration;
3271 di->lease_renew_after = di->lease_renew_from +
3272 (duration >> 1);
3273 di->lease_renew_from = 0;
3274 }
3275 break;
3276 }
3277 spin_unlock(&dentry->d_lock);
3278 dput(dentry);
3279
3280 if (!release)
3281 goto out;
3282
3283 release:
3284 /* let's just reuse the same message */
3285 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3286 ceph_msg_get(msg);
3287 ceph_con_send(&session->s_con, msg);
3288
3289 out:
3290 iput(inode);
3291 mutex_unlock(&session->s_mutex);
3292 return;
3293
3294 bad:
3295 pr_err("corrupt lease message\n");
3296 ceph_msg_dump(msg);
3297 }
3298
3299 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3300 struct inode *inode,
3301 struct dentry *dentry, char action,
3302 u32 seq)
3303 {
3304 struct ceph_msg *msg;
3305 struct ceph_mds_lease *lease;
3306 int len = sizeof(*lease) + sizeof(u32);
3307 int dnamelen = 0;
3308
3309 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3310 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3311 dnamelen = dentry->d_name.len;
3312 len += dnamelen;
3313
3314 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3315 if (!msg)
3316 return;
3317 lease = msg->front.iov_base;
3318 lease->action = action;
3319 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3320 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3321 lease->seq = cpu_to_le32(seq);
3322 put_unaligned_le32(dnamelen, lease + 1);
3323 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3324
3325 /*
3326 * if this is a preemptive lease RELEASE, no need to
3327 * flush request stream, since the actual request will
3328 * soon follow.
3329 */
3330 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3331
3332 ceph_con_send(&session->s_con, msg);
3333 }
3334
3335 /*
3336 * drop all leases (and dentry refs) in preparation for umount
3337 */
3338 static void drop_leases(struct ceph_mds_client *mdsc)
3339 {
3340 int i;
3341
3342 dout("drop_leases\n");
3343 mutex_lock(&mdsc->mutex);
3344 for (i = 0; i < mdsc->max_sessions; i++) {
3345 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3346 if (!s)
3347 continue;
3348 mutex_unlock(&mdsc->mutex);
3349 mutex_lock(&s->s_mutex);
3350 mutex_unlock(&s->s_mutex);
3351 ceph_put_mds_session(s);
3352 mutex_lock(&mdsc->mutex);
3353 }
3354 mutex_unlock(&mdsc->mutex);
3355 }
3356
3357
3358
3359 /*
3360 * delayed work -- periodically trim expired leases, renew caps with mds
3361 */
3362 static void schedule_delayed(struct ceph_mds_client *mdsc)
3363 {
3364 int delay = 5;
3365 unsigned hz = round_jiffies_relative(HZ * delay);
3366 schedule_delayed_work(&mdsc->delayed_work, hz);
3367 }
3368
3369 static void delayed_work(struct work_struct *work)
3370 {
3371 int i;
3372 struct ceph_mds_client *mdsc =
3373 container_of(work, struct ceph_mds_client, delayed_work.work);
3374 int renew_interval;
3375 int renew_caps;
3376
3377 dout("mdsc delayed_work\n");
3378 ceph_check_delayed_caps(mdsc);
3379
3380 mutex_lock(&mdsc->mutex);
3381 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3382 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3383 mdsc->last_renew_caps);
3384 if (renew_caps)
3385 mdsc->last_renew_caps = jiffies;
3386
3387 for (i = 0; i < mdsc->max_sessions; i++) {
3388 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3389 if (s == NULL)
3390 continue;
3391 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3392 dout("resending session close request for mds%d\n",
3393 s->s_mds);
3394 request_close_session(mdsc, s);
3395 ceph_put_mds_session(s);
3396 continue;
3397 }
3398 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3399 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3400 s->s_state = CEPH_MDS_SESSION_HUNG;
3401 pr_info("mds%d hung\n", s->s_mds);
3402 }
3403 }
3404 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3405 /* this mds is failed or recovering, just wait */
3406 ceph_put_mds_session(s);
3407 continue;
3408 }
3409 mutex_unlock(&mdsc->mutex);
3410
3411 mutex_lock(&s->s_mutex);
3412 if (renew_caps)
3413 send_renew_caps(mdsc, s);
3414 else
3415 ceph_con_keepalive(&s->s_con);
3416 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3417 s->s_state == CEPH_MDS_SESSION_HUNG)
3418 ceph_send_cap_releases(mdsc, s);
3419 mutex_unlock(&s->s_mutex);
3420 ceph_put_mds_session(s);
3421
3422 mutex_lock(&mdsc->mutex);
3423 }
3424 mutex_unlock(&mdsc->mutex);
3425
3426 schedule_delayed(mdsc);
3427 }
3428
3429 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3430
3431 {
3432 struct ceph_mds_client *mdsc;
3433
3434 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3435 if (!mdsc)
3436 return -ENOMEM;
3437 mdsc->fsc = fsc;
3438 fsc->mdsc = mdsc;
3439 mutex_init(&mdsc->mutex);
3440 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3441 if (mdsc->mdsmap == NULL) {
3442 kfree(mdsc);
3443 return -ENOMEM;
3444 }
3445
3446 init_completion(&mdsc->safe_umount_waiters);
3447 init_waitqueue_head(&mdsc->session_close_wq);
3448 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3449 mdsc->sessions = NULL;
3450 atomic_set(&mdsc->num_sessions, 0);
3451 mdsc->max_sessions = 0;
3452 mdsc->stopping = 0;
3453 mdsc->last_snap_seq = 0;
3454 init_rwsem(&mdsc->snap_rwsem);
3455 mdsc->snap_realms = RB_ROOT;
3456 INIT_LIST_HEAD(&mdsc->snap_empty);
3457 spin_lock_init(&mdsc->snap_empty_lock);
3458 mdsc->last_tid = 0;
3459 mdsc->oldest_tid = 0;
3460 mdsc->request_tree = RB_ROOT;
3461 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3462 mdsc->last_renew_caps = jiffies;
3463 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3464 spin_lock_init(&mdsc->cap_delay_lock);
3465 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3466 spin_lock_init(&mdsc->snap_flush_lock);
3467 mdsc->last_cap_flush_tid = 1;
3468 INIT_LIST_HEAD(&mdsc->cap_flush_list);
3469 INIT_LIST_HEAD(&mdsc->cap_dirty);
3470 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3471 mdsc->num_cap_flushing = 0;
3472 spin_lock_init(&mdsc->cap_dirty_lock);
3473 init_waitqueue_head(&mdsc->cap_flushing_wq);
3474 spin_lock_init(&mdsc->dentry_lru_lock);
3475 INIT_LIST_HEAD(&mdsc->dentry_lru);
3476
3477 ceph_caps_init(mdsc);
3478 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3479
3480 init_rwsem(&mdsc->pool_perm_rwsem);
3481 mdsc->pool_perm_tree = RB_ROOT;
3482
3483 return 0;
3484 }
3485
3486 /*
3487 * Wait for safe replies on open mds requests. If we time out, drop
3488 * all requests from the tree to avoid dangling dentry refs.
3489 */
3490 static void wait_requests(struct ceph_mds_client *mdsc)
3491 {
3492 struct ceph_options *opts = mdsc->fsc->client->options;
3493 struct ceph_mds_request *req;
3494
3495 mutex_lock(&mdsc->mutex);
3496 if (__get_oldest_req(mdsc)) {
3497 mutex_unlock(&mdsc->mutex);
3498
3499 dout("wait_requests waiting for requests\n");
3500 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3501 ceph_timeout_jiffies(opts->mount_timeout));
3502
3503 /* tear down remaining requests */
3504 mutex_lock(&mdsc->mutex);
3505 while ((req = __get_oldest_req(mdsc))) {
3506 dout("wait_requests timed out on tid %llu\n",
3507 req->r_tid);
3508 __unregister_request(mdsc, req);
3509 }
3510 }
3511 mutex_unlock(&mdsc->mutex);
3512 dout("wait_requests done\n");
3513 }
3514
3515 /*
3516 * called before mount is ro, and before dentries are torn down.
3517 * (hmm, does this still race with new lookups?)
3518 */
3519 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3520 {
3521 dout("pre_umount\n");
3522 mdsc->stopping = 1;
3523
3524 drop_leases(mdsc);
3525 ceph_flush_dirty_caps(mdsc);
3526 wait_requests(mdsc);
3527
3528 /*
3529 * wait for reply handlers to drop their request refs and
3530 * their inode/dcache refs
3531 */
3532 ceph_msgr_flush();
3533 }
3534
3535 /*
3536 * wait for all write mds requests to flush.
3537 */
3538 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3539 {
3540 struct ceph_mds_request *req = NULL, *nextreq;
3541 struct rb_node *n;
3542
3543 mutex_lock(&mdsc->mutex);
3544 dout("wait_unsafe_requests want %lld\n", want_tid);
3545 restart:
3546 req = __get_oldest_req(mdsc);
3547 while (req && req->r_tid <= want_tid) {
3548 /* find next request */
3549 n = rb_next(&req->r_node);
3550 if (n)
3551 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3552 else
3553 nextreq = NULL;
3554 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3555 (req->r_op & CEPH_MDS_OP_WRITE)) {
3556 /* write op */
3557 ceph_mdsc_get_request(req);
3558 if (nextreq)
3559 ceph_mdsc_get_request(nextreq);
3560 mutex_unlock(&mdsc->mutex);
3561 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3562 req->r_tid, want_tid);
3563 wait_for_completion(&req->r_safe_completion);
3564 mutex_lock(&mdsc->mutex);
3565 ceph_mdsc_put_request(req);
3566 if (!nextreq)
3567 break; /* next dne before, so we're done! */
3568 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3569 /* next request was removed from tree */
3570 ceph_mdsc_put_request(nextreq);
3571 goto restart;
3572 }
3573 ceph_mdsc_put_request(nextreq); /* won't go away */
3574 }
3575 req = nextreq;
3576 }
3577 mutex_unlock(&mdsc->mutex);
3578 dout("wait_unsafe_requests done\n");
3579 }
3580
3581 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3582 {
3583 u64 want_tid, want_flush;
3584
3585 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3586 return;
3587
3588 dout("sync\n");
3589 mutex_lock(&mdsc->mutex);
3590 want_tid = mdsc->last_tid;
3591 mutex_unlock(&mdsc->mutex);
3592
3593 ceph_flush_dirty_caps(mdsc);
3594 spin_lock(&mdsc->cap_dirty_lock);
3595 want_flush = mdsc->last_cap_flush_tid;
3596 if (!list_empty(&mdsc->cap_flush_list)) {
3597 struct ceph_cap_flush *cf =
3598 list_last_entry(&mdsc->cap_flush_list,
3599 struct ceph_cap_flush, g_list);
3600 cf->wake = true;
3601 }
3602 spin_unlock(&mdsc->cap_dirty_lock);
3603
3604 dout("sync want tid %lld flush_seq %lld\n",
3605 want_tid, want_flush);
3606
3607 wait_unsafe_requests(mdsc, want_tid);
3608 wait_caps_flush(mdsc, want_flush);
3609 }
3610
3611 /*
3612 * true if all sessions are closed, or we force unmount
3613 */
3614 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3615 {
3616 if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3617 return true;
3618 return atomic_read(&mdsc->num_sessions) <= skipped;
3619 }
3620
3621 /*
3622 * called after sb is ro.
3623 */
3624 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3625 {
3626 struct ceph_options *opts = mdsc->fsc->client->options;
3627 struct ceph_mds_session *session;
3628 int i;
3629 int skipped = 0;
3630
3631 dout("close_sessions\n");
3632
3633 /* close sessions */
3634 mutex_lock(&mdsc->mutex);
3635 for (i = 0; i < mdsc->max_sessions; i++) {
3636 session = __ceph_lookup_mds_session(mdsc, i);
3637 if (!session)
3638 continue;
3639 mutex_unlock(&mdsc->mutex);
3640 mutex_lock(&session->s_mutex);
3641 if (__close_session(mdsc, session) <= 0)
3642 skipped++;
3643 mutex_unlock(&session->s_mutex);
3644 ceph_put_mds_session(session);
3645 mutex_lock(&mdsc->mutex);
3646 }
3647 mutex_unlock(&mdsc->mutex);
3648
3649 dout("waiting for sessions to close\n");
3650 wait_event_timeout(mdsc->session_close_wq,
3651 done_closing_sessions(mdsc, skipped),
3652 ceph_timeout_jiffies(opts->mount_timeout));
3653
3654 /* tear down remaining sessions */
3655 mutex_lock(&mdsc->mutex);
3656 for (i = 0; i < mdsc->max_sessions; i++) {
3657 if (mdsc->sessions[i]) {
3658 session = get_session(mdsc->sessions[i]);
3659 __unregister_session(mdsc, session);
3660 mutex_unlock(&mdsc->mutex);
3661 mutex_lock(&session->s_mutex);
3662 remove_session_caps(session);
3663 mutex_unlock(&session->s_mutex);
3664 ceph_put_mds_session(session);
3665 mutex_lock(&mdsc->mutex);
3666 }
3667 }
3668 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3669 mutex_unlock(&mdsc->mutex);
3670
3671 ceph_cleanup_empty_realms(mdsc);
3672
3673 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3674
3675 dout("stopped\n");
3676 }
3677
3678 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3679 {
3680 struct ceph_mds_session *session;
3681 int mds;
3682
3683 dout("force umount\n");
3684
3685 mutex_lock(&mdsc->mutex);
3686 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3687 session = __ceph_lookup_mds_session(mdsc, mds);
3688 if (!session)
3689 continue;
3690 mutex_unlock(&mdsc->mutex);
3691 mutex_lock(&session->s_mutex);
3692 __close_session(mdsc, session);
3693 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3694 cleanup_session_requests(mdsc, session);
3695 remove_session_caps(session);
3696 }
3697 mutex_unlock(&session->s_mutex);
3698 ceph_put_mds_session(session);
3699 mutex_lock(&mdsc->mutex);
3700 kick_requests(mdsc, mds);
3701 }
3702 __wake_requests(mdsc, &mdsc->waiting_for_map);
3703 mutex_unlock(&mdsc->mutex);
3704 }
3705
3706 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3707 {
3708 dout("stop\n");
3709 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3710 if (mdsc->mdsmap)
3711 ceph_mdsmap_destroy(mdsc->mdsmap);
3712 kfree(mdsc->sessions);
3713 ceph_caps_finalize(mdsc);
3714 ceph_pool_perm_destroy(mdsc);
3715 }
3716
3717 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3718 {
3719 struct ceph_mds_client *mdsc = fsc->mdsc;
3720
3721 dout("mdsc_destroy %p\n", mdsc);
3722 ceph_mdsc_stop(mdsc);
3723
3724 /* flush out any connection work with references to us */
3725 ceph_msgr_flush();
3726
3727 fsc->mdsc = NULL;
3728 kfree(mdsc);
3729 dout("mdsc_destroy %p done\n", mdsc);
3730 }
3731
3732 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3733 {
3734 struct ceph_fs_client *fsc = mdsc->fsc;
3735 const char *mds_namespace = fsc->mount_options->mds_namespace;
3736 void *p = msg->front.iov_base;
3737 void *end = p + msg->front.iov_len;
3738 u32 epoch;
3739 u32 map_len;
3740 u32 num_fs;
3741 u32 mount_fscid = (u32)-1;
3742 u8 struct_v, struct_cv;
3743 int err = -EINVAL;
3744
3745 ceph_decode_need(&p, end, sizeof(u32), bad);
3746 epoch = ceph_decode_32(&p);
3747
3748 dout("handle_fsmap epoch %u\n", epoch);
3749
3750 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3751 struct_v = ceph_decode_8(&p);
3752 struct_cv = ceph_decode_8(&p);
3753 map_len = ceph_decode_32(&p);
3754
3755 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3756 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3757
3758 num_fs = ceph_decode_32(&p);
3759 while (num_fs-- > 0) {
3760 void *info_p, *info_end;
3761 u32 info_len;
3762 u8 info_v, info_cv;
3763 u32 fscid, namelen;
3764
3765 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3766 info_v = ceph_decode_8(&p);
3767 info_cv = ceph_decode_8(&p);
3768 info_len = ceph_decode_32(&p);
3769 ceph_decode_need(&p, end, info_len, bad);
3770 info_p = p;
3771 info_end = p + info_len;
3772 p = info_end;
3773
3774 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3775 fscid = ceph_decode_32(&info_p);
3776 namelen = ceph_decode_32(&info_p);
3777 ceph_decode_need(&info_p, info_end, namelen, bad);
3778
3779 if (mds_namespace &&
3780 strlen(mds_namespace) == namelen &&
3781 !strncmp(mds_namespace, (char *)info_p, namelen)) {
3782 mount_fscid = fscid;
3783 break;
3784 }
3785 }
3786
3787 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3788 if (mount_fscid != (u32)-1) {
3789 fsc->client->monc.fs_cluster_id = mount_fscid;
3790 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3791 0, true);
3792 ceph_monc_renew_subs(&fsc->client->monc);
3793 } else {
3794 err = -ENOENT;
3795 goto err_out;
3796 }
3797 return;
3798 bad:
3799 pr_err("error decoding fsmap\n");
3800 err_out:
3801 mutex_lock(&mdsc->mutex);
3802 mdsc->mdsmap_err = -ENOENT;
3803 __wake_requests(mdsc, &mdsc->waiting_for_map);
3804 mutex_unlock(&mdsc->mutex);
3805 return;
3806 }
3807
3808 /*
3809 * handle mds map update.
3810 */
3811 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3812 {
3813 u32 epoch;
3814 u32 maplen;
3815 void *p = msg->front.iov_base;
3816 void *end = p + msg->front.iov_len;
3817 struct ceph_mdsmap *newmap, *oldmap;
3818 struct ceph_fsid fsid;
3819 int err = -EINVAL;
3820
3821 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3822 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3823 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3824 return;
3825 epoch = ceph_decode_32(&p);
3826 maplen = ceph_decode_32(&p);
3827 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3828
3829 /* do we need it? */
3830 mutex_lock(&mdsc->mutex);
3831 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3832 dout("handle_map epoch %u <= our %u\n",
3833 epoch, mdsc->mdsmap->m_epoch);
3834 mutex_unlock(&mdsc->mutex);
3835 return;
3836 }
3837
3838 newmap = ceph_mdsmap_decode(&p, end);
3839 if (IS_ERR(newmap)) {
3840 err = PTR_ERR(newmap);
3841 goto bad_unlock;
3842 }
3843
3844 /* swap into place */
3845 if (mdsc->mdsmap) {
3846 oldmap = mdsc->mdsmap;
3847 mdsc->mdsmap = newmap;
3848 check_new_map(mdsc, newmap, oldmap);
3849 ceph_mdsmap_destroy(oldmap);
3850 } else {
3851 mdsc->mdsmap = newmap; /* first mds map */
3852 }
3853 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3854
3855 __wake_requests(mdsc, &mdsc->waiting_for_map);
3856 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3857 mdsc->mdsmap->m_epoch);
3858
3859 mutex_unlock(&mdsc->mutex);
3860 schedule_delayed(mdsc);
3861 return;
3862
3863 bad_unlock:
3864 mutex_unlock(&mdsc->mutex);
3865 bad:
3866 pr_err("error decoding mdsmap %d\n", err);
3867 return;
3868 }
3869
3870 static struct ceph_connection *con_get(struct ceph_connection *con)
3871 {
3872 struct ceph_mds_session *s = con->private;
3873
3874 if (get_session(s)) {
3875 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3876 return con;
3877 }
3878 dout("mdsc con_get %p FAIL\n", s);
3879 return NULL;
3880 }
3881
3882 static void con_put(struct ceph_connection *con)
3883 {
3884 struct ceph_mds_session *s = con->private;
3885
3886 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3887 ceph_put_mds_session(s);
3888 }
3889
3890 /*
3891 * if the client is unresponsive for long enough, the mds will kill
3892 * the session entirely.
3893 */
3894 static void peer_reset(struct ceph_connection *con)
3895 {
3896 struct ceph_mds_session *s = con->private;
3897 struct ceph_mds_client *mdsc = s->s_mdsc;
3898
3899 pr_warn("mds%d closed our session\n", s->s_mds);
3900 send_mds_reconnect(mdsc, s);
3901 }
3902
3903 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3904 {
3905 struct ceph_mds_session *s = con->private;
3906 struct ceph_mds_client *mdsc = s->s_mdsc;
3907 int type = le16_to_cpu(msg->hdr.type);
3908
3909 mutex_lock(&mdsc->mutex);
3910 if (__verify_registered_session(mdsc, s) < 0) {
3911 mutex_unlock(&mdsc->mutex);
3912 goto out;
3913 }
3914 mutex_unlock(&mdsc->mutex);
3915
3916 switch (type) {
3917 case CEPH_MSG_MDS_MAP:
3918 ceph_mdsc_handle_mdsmap(mdsc, msg);
3919 break;
3920 case CEPH_MSG_FS_MAP_USER:
3921 ceph_mdsc_handle_fsmap(mdsc, msg);
3922 break;
3923 case CEPH_MSG_CLIENT_SESSION:
3924 handle_session(s, msg);
3925 break;
3926 case CEPH_MSG_CLIENT_REPLY:
3927 handle_reply(s, msg);
3928 break;
3929 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3930 handle_forward(mdsc, s, msg);
3931 break;
3932 case CEPH_MSG_CLIENT_CAPS:
3933 ceph_handle_caps(s, msg);
3934 break;
3935 case CEPH_MSG_CLIENT_SNAP:
3936 ceph_handle_snap(mdsc, s, msg);
3937 break;
3938 case CEPH_MSG_CLIENT_LEASE:
3939 handle_lease(mdsc, s, msg);
3940 break;
3941
3942 default:
3943 pr_err("received unknown message type %d %s\n", type,
3944 ceph_msg_type_name(type));
3945 }
3946 out:
3947 ceph_msg_put(msg);
3948 }
3949
3950 /*
3951 * authentication
3952 */
3953
3954 /*
3955 * Note: returned pointer is the address of a structure that's
3956 * managed separately. Caller must *not* attempt to free it.
3957 */
3958 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3959 int *proto, int force_new)
3960 {
3961 struct ceph_mds_session *s = con->private;
3962 struct ceph_mds_client *mdsc = s->s_mdsc;
3963 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3964 struct ceph_auth_handshake *auth = &s->s_auth;
3965
3966 if (force_new && auth->authorizer) {
3967 ceph_auth_destroy_authorizer(auth->authorizer);
3968 auth->authorizer = NULL;
3969 }
3970 if (!auth->authorizer) {
3971 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3972 auth);
3973 if (ret)
3974 return ERR_PTR(ret);
3975 } else {
3976 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3977 auth);
3978 if (ret)
3979 return ERR_PTR(ret);
3980 }
3981 *proto = ac->protocol;
3982
3983 return auth;
3984 }
3985
3986
3987 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3988 {
3989 struct ceph_mds_session *s = con->private;
3990 struct ceph_mds_client *mdsc = s->s_mdsc;
3991 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3992
3993 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3994 }
3995
3996 static int invalidate_authorizer(struct ceph_connection *con)
3997 {
3998 struct ceph_mds_session *s = con->private;
3999 struct ceph_mds_client *mdsc = s->s_mdsc;
4000 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4001
4002 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4003
4004 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4005 }
4006
4007 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4008 struct ceph_msg_header *hdr, int *skip)
4009 {
4010 struct ceph_msg *msg;
4011 int type = (int) le16_to_cpu(hdr->type);
4012 int front_len = (int) le32_to_cpu(hdr->front_len);
4013
4014 if (con->in_msg)
4015 return con->in_msg;
4016
4017 *skip = 0;
4018 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4019 if (!msg) {
4020 pr_err("unable to allocate msg type %d len %d\n",
4021 type, front_len);
4022 return NULL;
4023 }
4024
4025 return msg;
4026 }
4027
4028 static int mds_sign_message(struct ceph_msg *msg)
4029 {
4030 struct ceph_mds_session *s = msg->con->private;
4031 struct ceph_auth_handshake *auth = &s->s_auth;
4032
4033 return ceph_auth_sign_message(auth, msg);
4034 }
4035
4036 static int mds_check_message_signature(struct ceph_msg *msg)
4037 {
4038 struct ceph_mds_session *s = msg->con->private;
4039 struct ceph_auth_handshake *auth = &s->s_auth;
4040
4041 return ceph_auth_check_message_signature(auth, msg);
4042 }
4043
4044 static const struct ceph_connection_operations mds_con_ops = {
4045 .get = con_get,
4046 .put = con_put,
4047 .dispatch = dispatch,
4048 .get_authorizer = get_authorizer,
4049 .verify_authorizer_reply = verify_authorizer_reply,
4050 .invalidate_authorizer = invalidate_authorizer,
4051 .peer_reset = peer_reset,
4052 .alloc_msg = mds_alloc_msg,
4053 .sign_message = mds_sign_message,
4054 .check_message_signature = mds_check_message_signature,
4055 };
4056
4057 /* eof */
Linux with added FPC-III support