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41 * are no longer valid.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