xref: /openbmc/linux/fs/ceph/mds_client.c (revision 3e26a691)
1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/fs.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
6 #include <linux/gfp.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11 #include <linux/ratelimit.h>
12 
13 #include "super.h"
14 #include "mds_client.h"
15 
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
22 
23 /*
24  * A cluster of MDS (metadata server) daemons is responsible for
25  * managing the file system namespace (the directory hierarchy and
26  * inodes) and for coordinating shared access to storage.  Metadata is
27  * partitioning hierarchically across a number of servers, and that
28  * partition varies over time as the cluster adjusts the distribution
29  * in order to balance load.
30  *
31  * The MDS client is primarily responsible to managing synchronous
32  * metadata requests for operations like open, unlink, and so forth.
33  * If there is a MDS failure, we find out about it when we (possibly
34  * request and) receive a new MDS map, and can resubmit affected
35  * requests.
36  *
37  * For the most part, though, we take advantage of a lossless
38  * communications channel to the MDS, and do not need to worry about
39  * timing out or resubmitting requests.
40  *
41  * We maintain a stateful "session" with each MDS we interact with.
42  * Within each session, we sent periodic heartbeat messages to ensure
43  * any capabilities or leases we have been issues remain valid.  If
44  * the session times out and goes stale, our leases and capabilities
45  * are no longer valid.
46  */
47 
48 struct ceph_reconnect_state {
49 	int nr_caps;
50 	struct ceph_pagelist *pagelist;
51 	bool flock;
52 };
53 
54 static void __wake_requests(struct ceph_mds_client *mdsc,
55 			    struct list_head *head);
56 
57 static const struct ceph_connection_operations mds_con_ops;
58 
59 
60 /*
61  * mds reply parsing
62  */
63 
64 /*
65  * parse individual inode info
66  */
67 static int parse_reply_info_in(void **p, void *end,
68 			       struct ceph_mds_reply_info_in *info,
69 			       u64 features)
70 {
71 	int err = -EIO;
72 
73 	info->in = *p;
74 	*p += sizeof(struct ceph_mds_reply_inode) +
75 		sizeof(*info->in->fragtree.splits) *
76 		le32_to_cpu(info->in->fragtree.nsplits);
77 
78 	ceph_decode_32_safe(p, end, info->symlink_len, bad);
79 	ceph_decode_need(p, end, info->symlink_len, bad);
80 	info->symlink = *p;
81 	*p += info->symlink_len;
82 
83 	if (features & CEPH_FEATURE_DIRLAYOUTHASH)
84 		ceph_decode_copy_safe(p, end, &info->dir_layout,
85 				      sizeof(info->dir_layout), bad);
86 	else
87 		memset(&info->dir_layout, 0, sizeof(info->dir_layout));
88 
89 	ceph_decode_32_safe(p, end, info->xattr_len, bad);
90 	ceph_decode_need(p, end, info->xattr_len, bad);
91 	info->xattr_data = *p;
92 	*p += info->xattr_len;
93 
94 	if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
95 		ceph_decode_64_safe(p, end, info->inline_version, bad);
96 		ceph_decode_32_safe(p, end, info->inline_len, bad);
97 		ceph_decode_need(p, end, info->inline_len, bad);
98 		info->inline_data = *p;
99 		*p += info->inline_len;
100 	} else
101 		info->inline_version = CEPH_INLINE_NONE;
102 
103 	if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
104 		ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
105 		ceph_decode_need(p, end, info->pool_ns_len, bad);
106 		*p += info->pool_ns_len;
107 	} else {
108 		info->pool_ns_len = 0;
109 	}
110 
111 	return 0;
112 bad:
113 	return err;
114 }
115 
116 /*
117  * parse a normal reply, which may contain a (dir+)dentry and/or a
118  * target inode.
119  */
120 static int parse_reply_info_trace(void **p, void *end,
121 				  struct ceph_mds_reply_info_parsed *info,
122 				  u64 features)
123 {
124 	int err;
125 
126 	if (info->head->is_dentry) {
127 		err = parse_reply_info_in(p, end, &info->diri, features);
128 		if (err < 0)
129 			goto out_bad;
130 
131 		if (unlikely(*p + sizeof(*info->dirfrag) > end))
132 			goto bad;
133 		info->dirfrag = *p;
134 		*p += sizeof(*info->dirfrag) +
135 			sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
136 		if (unlikely(*p > end))
137 			goto bad;
138 
139 		ceph_decode_32_safe(p, end, info->dname_len, bad);
140 		ceph_decode_need(p, end, info->dname_len, bad);
141 		info->dname = *p;
142 		*p += info->dname_len;
143 		info->dlease = *p;
144 		*p += sizeof(*info->dlease);
145 	}
146 
147 	if (info->head->is_target) {
148 		err = parse_reply_info_in(p, end, &info->targeti, features);
149 		if (err < 0)
150 			goto out_bad;
151 	}
152 
153 	if (unlikely(*p != end))
154 		goto bad;
155 	return 0;
156 
157 bad:
158 	err = -EIO;
159 out_bad:
160 	pr_err("problem parsing mds trace %d\n", err);
161 	return err;
162 }
163 
164 /*
165  * parse readdir results
166  */
167 static int parse_reply_info_dir(void **p, void *end,
168 				struct ceph_mds_reply_info_parsed *info,
169 				u64 features)
170 {
171 	u32 num, i = 0;
172 	int err;
173 
174 	info->dir_dir = *p;
175 	if (*p + sizeof(*info->dir_dir) > end)
176 		goto bad;
177 	*p += sizeof(*info->dir_dir) +
178 		sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
179 	if (*p > end)
180 		goto bad;
181 
182 	ceph_decode_need(p, end, sizeof(num) + 2, bad);
183 	num = ceph_decode_32(p);
184 	info->dir_end = ceph_decode_8(p);
185 	info->dir_complete = ceph_decode_8(p);
186 	if (num == 0)
187 		goto done;
188 
189 	BUG_ON(!info->dir_in);
190 	info->dir_dname = (void *)(info->dir_in + num);
191 	info->dir_dname_len = (void *)(info->dir_dname + num);
192 	info->dir_dlease = (void *)(info->dir_dname_len + num);
193 	if ((unsigned long)(info->dir_dlease + num) >
194 	    (unsigned long)info->dir_in + info->dir_buf_size) {
195 		pr_err("dir contents are larger than expected\n");
196 		WARN_ON(1);
197 		goto bad;
198 	}
199 
200 	info->dir_nr = num;
201 	while (num) {
202 		/* dentry */
203 		ceph_decode_need(p, end, sizeof(u32)*2, bad);
204 		info->dir_dname_len[i] = ceph_decode_32(p);
205 		ceph_decode_need(p, end, info->dir_dname_len[i], bad);
206 		info->dir_dname[i] = *p;
207 		*p += info->dir_dname_len[i];
208 		dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
209 		     info->dir_dname[i]);
210 		info->dir_dlease[i] = *p;
211 		*p += sizeof(struct ceph_mds_reply_lease);
212 
213 		/* inode */
214 		err = parse_reply_info_in(p, end, &info->dir_in[i], features);
215 		if (err < 0)
216 			goto out_bad;
217 		i++;
218 		num--;
219 	}
220 
221 done:
222 	if (*p != end)
223 		goto bad;
224 	return 0;
225 
226 bad:
227 	err = -EIO;
228 out_bad:
229 	pr_err("problem parsing dir contents %d\n", err);
230 	return err;
231 }
232 
233 /*
234  * parse fcntl F_GETLK results
235  */
236 static int parse_reply_info_filelock(void **p, void *end,
237 				     struct ceph_mds_reply_info_parsed *info,
238 				     u64 features)
239 {
240 	if (*p + sizeof(*info->filelock_reply) > end)
241 		goto bad;
242 
243 	info->filelock_reply = *p;
244 	*p += sizeof(*info->filelock_reply);
245 
246 	if (unlikely(*p != end))
247 		goto bad;
248 	return 0;
249 
250 bad:
251 	return -EIO;
252 }
253 
254 /*
255  * parse create results
256  */
257 static int parse_reply_info_create(void **p, void *end,
258 				  struct ceph_mds_reply_info_parsed *info,
259 				  u64 features)
260 {
261 	if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
262 		if (*p == end) {
263 			info->has_create_ino = false;
264 		} else {
265 			info->has_create_ino = true;
266 			info->ino = ceph_decode_64(p);
267 		}
268 	}
269 
270 	if (unlikely(*p != end))
271 		goto bad;
272 	return 0;
273 
274 bad:
275 	return -EIO;
276 }
277 
278 /*
279  * parse extra results
280  */
281 static int parse_reply_info_extra(void **p, void *end,
282 				  struct ceph_mds_reply_info_parsed *info,
283 				  u64 features)
284 {
285 	if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
286 		return parse_reply_info_filelock(p, end, info, features);
287 	else if (info->head->op == CEPH_MDS_OP_READDIR ||
288 		 info->head->op == CEPH_MDS_OP_LSSNAP)
289 		return parse_reply_info_dir(p, end, info, features);
290 	else if (info->head->op == CEPH_MDS_OP_CREATE)
291 		return parse_reply_info_create(p, end, info, features);
292 	else
293 		return -EIO;
294 }
295 
296 /*
297  * parse entire mds reply
298  */
299 static int parse_reply_info(struct ceph_msg *msg,
300 			    struct ceph_mds_reply_info_parsed *info,
301 			    u64 features)
302 {
303 	void *p, *end;
304 	u32 len;
305 	int err;
306 
307 	info->head = msg->front.iov_base;
308 	p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
309 	end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
310 
311 	/* trace */
312 	ceph_decode_32_safe(&p, end, len, bad);
313 	if (len > 0) {
314 		ceph_decode_need(&p, end, len, bad);
315 		err = parse_reply_info_trace(&p, p+len, info, features);
316 		if (err < 0)
317 			goto out_bad;
318 	}
319 
320 	/* extra */
321 	ceph_decode_32_safe(&p, end, len, bad);
322 	if (len > 0) {
323 		ceph_decode_need(&p, end, len, bad);
324 		err = parse_reply_info_extra(&p, p+len, info, features);
325 		if (err < 0)
326 			goto out_bad;
327 	}
328 
329 	/* snap blob */
330 	ceph_decode_32_safe(&p, end, len, bad);
331 	info->snapblob_len = len;
332 	info->snapblob = p;
333 	p += len;
334 
335 	if (p != end)
336 		goto bad;
337 	return 0;
338 
339 bad:
340 	err = -EIO;
341 out_bad:
342 	pr_err("mds parse_reply err %d\n", err);
343 	return err;
344 }
345 
346 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
347 {
348 	if (!info->dir_in)
349 		return;
350 	free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
351 }
352 
353 
354 /*
355  * sessions
356  */
357 const char *ceph_session_state_name(int s)
358 {
359 	switch (s) {
360 	case CEPH_MDS_SESSION_NEW: return "new";
361 	case CEPH_MDS_SESSION_OPENING: return "opening";
362 	case CEPH_MDS_SESSION_OPEN: return "open";
363 	case CEPH_MDS_SESSION_HUNG: return "hung";
364 	case CEPH_MDS_SESSION_CLOSING: return "closing";
365 	case CEPH_MDS_SESSION_RESTARTING: return "restarting";
366 	case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
367 	default: return "???";
368 	}
369 }
370 
371 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
372 {
373 	if (atomic_inc_not_zero(&s->s_ref)) {
374 		dout("mdsc get_session %p %d -> %d\n", s,
375 		     atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
376 		return s;
377 	} else {
378 		dout("mdsc get_session %p 0 -- FAIL", s);
379 		return NULL;
380 	}
381 }
382 
383 void ceph_put_mds_session(struct ceph_mds_session *s)
384 {
385 	dout("mdsc put_session %p %d -> %d\n", s,
386 	     atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
387 	if (atomic_dec_and_test(&s->s_ref)) {
388 		if (s->s_auth.authorizer)
389 			ceph_auth_destroy_authorizer(
390 				s->s_mdsc->fsc->client->monc.auth,
391 				s->s_auth.authorizer);
392 		kfree(s);
393 	}
394 }
395 
396 /*
397  * called under mdsc->mutex
398  */
399 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
400 						   int mds)
401 {
402 	struct ceph_mds_session *session;
403 
404 	if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
405 		return NULL;
406 	session = mdsc->sessions[mds];
407 	dout("lookup_mds_session %p %d\n", session,
408 	     atomic_read(&session->s_ref));
409 	get_session(session);
410 	return session;
411 }
412 
413 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
414 {
415 	if (mds >= mdsc->max_sessions)
416 		return false;
417 	return mdsc->sessions[mds];
418 }
419 
420 static int __verify_registered_session(struct ceph_mds_client *mdsc,
421 				       struct ceph_mds_session *s)
422 {
423 	if (s->s_mds >= mdsc->max_sessions ||
424 	    mdsc->sessions[s->s_mds] != s)
425 		return -ENOENT;
426 	return 0;
427 }
428 
429 /*
430  * create+register a new session for given mds.
431  * called under mdsc->mutex.
432  */
433 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
434 						 int mds)
435 {
436 	struct ceph_mds_session *s;
437 
438 	if (mds >= mdsc->mdsmap->m_max_mds)
439 		return ERR_PTR(-EINVAL);
440 
441 	s = kzalloc(sizeof(*s), GFP_NOFS);
442 	if (!s)
443 		return ERR_PTR(-ENOMEM);
444 	s->s_mdsc = mdsc;
445 	s->s_mds = mds;
446 	s->s_state = CEPH_MDS_SESSION_NEW;
447 	s->s_ttl = 0;
448 	s->s_seq = 0;
449 	mutex_init(&s->s_mutex);
450 
451 	ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
452 
453 	spin_lock_init(&s->s_gen_ttl_lock);
454 	s->s_cap_gen = 0;
455 	s->s_cap_ttl = jiffies - 1;
456 
457 	spin_lock_init(&s->s_cap_lock);
458 	s->s_renew_requested = 0;
459 	s->s_renew_seq = 0;
460 	INIT_LIST_HEAD(&s->s_caps);
461 	s->s_nr_caps = 0;
462 	s->s_trim_caps = 0;
463 	atomic_set(&s->s_ref, 1);
464 	INIT_LIST_HEAD(&s->s_waiting);
465 	INIT_LIST_HEAD(&s->s_unsafe);
466 	s->s_num_cap_releases = 0;
467 	s->s_cap_reconnect = 0;
468 	s->s_cap_iterator = NULL;
469 	INIT_LIST_HEAD(&s->s_cap_releases);
470 	INIT_LIST_HEAD(&s->s_cap_flushing);
471 	INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
472 
473 	dout("register_session mds%d\n", mds);
474 	if (mds >= mdsc->max_sessions) {
475 		int newmax = 1 << get_count_order(mds+1);
476 		struct ceph_mds_session **sa;
477 
478 		dout("register_session realloc to %d\n", newmax);
479 		sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
480 		if (sa == NULL)
481 			goto fail_realloc;
482 		if (mdsc->sessions) {
483 			memcpy(sa, mdsc->sessions,
484 			       mdsc->max_sessions * sizeof(void *));
485 			kfree(mdsc->sessions);
486 		}
487 		mdsc->sessions = sa;
488 		mdsc->max_sessions = newmax;
489 	}
490 	mdsc->sessions[mds] = s;
491 	atomic_inc(&mdsc->num_sessions);
492 	atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
493 
494 	ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
495 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
496 
497 	return s;
498 
499 fail_realloc:
500 	kfree(s);
501 	return ERR_PTR(-ENOMEM);
502 }
503 
504 /*
505  * called under mdsc->mutex
506  */
507 static void __unregister_session(struct ceph_mds_client *mdsc,
508 			       struct ceph_mds_session *s)
509 {
510 	dout("__unregister_session mds%d %p\n", s->s_mds, s);
511 	BUG_ON(mdsc->sessions[s->s_mds] != s);
512 	mdsc->sessions[s->s_mds] = NULL;
513 	ceph_con_close(&s->s_con);
514 	ceph_put_mds_session(s);
515 	atomic_dec(&mdsc->num_sessions);
516 }
517 
518 /*
519  * drop session refs in request.
520  *
521  * should be last request ref, or hold mdsc->mutex
522  */
523 static void put_request_session(struct ceph_mds_request *req)
524 {
525 	if (req->r_session) {
526 		ceph_put_mds_session(req->r_session);
527 		req->r_session = NULL;
528 	}
529 }
530 
531 void ceph_mdsc_release_request(struct kref *kref)
532 {
533 	struct ceph_mds_request *req = container_of(kref,
534 						    struct ceph_mds_request,
535 						    r_kref);
536 	destroy_reply_info(&req->r_reply_info);
537 	if (req->r_request)
538 		ceph_msg_put(req->r_request);
539 	if (req->r_reply)
540 		ceph_msg_put(req->r_reply);
541 	if (req->r_inode) {
542 		ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
543 		iput(req->r_inode);
544 	}
545 	if (req->r_locked_dir)
546 		ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
547 	iput(req->r_target_inode);
548 	if (req->r_dentry)
549 		dput(req->r_dentry);
550 	if (req->r_old_dentry)
551 		dput(req->r_old_dentry);
552 	if (req->r_old_dentry_dir) {
553 		/*
554 		 * track (and drop pins for) r_old_dentry_dir
555 		 * separately, since r_old_dentry's d_parent may have
556 		 * changed between the dir mutex being dropped and
557 		 * this request being freed.
558 		 */
559 		ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
560 				  CEPH_CAP_PIN);
561 		iput(req->r_old_dentry_dir);
562 	}
563 	kfree(req->r_path1);
564 	kfree(req->r_path2);
565 	if (req->r_pagelist)
566 		ceph_pagelist_release(req->r_pagelist);
567 	put_request_session(req);
568 	ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
569 	kfree(req);
570 }
571 
572 /*
573  * lookup session, bump ref if found.
574  *
575  * called under mdsc->mutex.
576  */
577 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
578 					     u64 tid)
579 {
580 	struct ceph_mds_request *req;
581 	struct rb_node *n = mdsc->request_tree.rb_node;
582 
583 	while (n) {
584 		req = rb_entry(n, struct ceph_mds_request, r_node);
585 		if (tid < req->r_tid)
586 			n = n->rb_left;
587 		else if (tid > req->r_tid)
588 			n = n->rb_right;
589 		else {
590 			ceph_mdsc_get_request(req);
591 			return req;
592 		}
593 	}
594 	return NULL;
595 }
596 
597 static void __insert_request(struct ceph_mds_client *mdsc,
598 			     struct ceph_mds_request *new)
599 {
600 	struct rb_node **p = &mdsc->request_tree.rb_node;
601 	struct rb_node *parent = NULL;
602 	struct ceph_mds_request *req = NULL;
603 
604 	while (*p) {
605 		parent = *p;
606 		req = rb_entry(parent, struct ceph_mds_request, r_node);
607 		if (new->r_tid < req->r_tid)
608 			p = &(*p)->rb_left;
609 		else if (new->r_tid > req->r_tid)
610 			p = &(*p)->rb_right;
611 		else
612 			BUG();
613 	}
614 
615 	rb_link_node(&new->r_node, parent, p);
616 	rb_insert_color(&new->r_node, &mdsc->request_tree);
617 }
618 
619 /*
620  * Register an in-flight request, and assign a tid.  Link to directory
621  * are modifying (if any).
622  *
623  * Called under mdsc->mutex.
624  */
625 static void __register_request(struct ceph_mds_client *mdsc,
626 			       struct ceph_mds_request *req,
627 			       struct inode *dir)
628 {
629 	req->r_tid = ++mdsc->last_tid;
630 	if (req->r_num_caps)
631 		ceph_reserve_caps(mdsc, &req->r_caps_reservation,
632 				  req->r_num_caps);
633 	dout("__register_request %p tid %lld\n", req, req->r_tid);
634 	ceph_mdsc_get_request(req);
635 	__insert_request(mdsc, req);
636 
637 	req->r_uid = current_fsuid();
638 	req->r_gid = current_fsgid();
639 
640 	if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
641 		mdsc->oldest_tid = req->r_tid;
642 
643 	if (dir) {
644 		ihold(dir);
645 		req->r_unsafe_dir = dir;
646 	}
647 }
648 
649 static void __unregister_request(struct ceph_mds_client *mdsc,
650 				 struct ceph_mds_request *req)
651 {
652 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
653 
654 	if (req->r_tid == mdsc->oldest_tid) {
655 		struct rb_node *p = rb_next(&req->r_node);
656 		mdsc->oldest_tid = 0;
657 		while (p) {
658 			struct ceph_mds_request *next_req =
659 				rb_entry(p, struct ceph_mds_request, r_node);
660 			if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
661 				mdsc->oldest_tid = next_req->r_tid;
662 				break;
663 			}
664 			p = rb_next(p);
665 		}
666 	}
667 
668 	rb_erase(&req->r_node, &mdsc->request_tree);
669 	RB_CLEAR_NODE(&req->r_node);
670 
671 	if (req->r_unsafe_dir && req->r_got_unsafe) {
672 		struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
673 		spin_lock(&ci->i_unsafe_lock);
674 		list_del_init(&req->r_unsafe_dir_item);
675 		spin_unlock(&ci->i_unsafe_lock);
676 	}
677 	if (req->r_target_inode && req->r_got_unsafe) {
678 		struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
679 		spin_lock(&ci->i_unsafe_lock);
680 		list_del_init(&req->r_unsafe_target_item);
681 		spin_unlock(&ci->i_unsafe_lock);
682 	}
683 
684 	if (req->r_unsafe_dir) {
685 		iput(req->r_unsafe_dir);
686 		req->r_unsafe_dir = NULL;
687 	}
688 
689 	complete_all(&req->r_safe_completion);
690 
691 	ceph_mdsc_put_request(req);
692 }
693 
694 /*
695  * Choose mds to send request to next.  If there is a hint set in the
696  * request (e.g., due to a prior forward hint from the mds), use that.
697  * Otherwise, consult frag tree and/or caps to identify the
698  * appropriate mds.  If all else fails, choose randomly.
699  *
700  * Called under mdsc->mutex.
701  */
702 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
703 {
704 	/*
705 	 * we don't need to worry about protecting the d_parent access
706 	 * here because we never renaming inside the snapped namespace
707 	 * except to resplice to another snapdir, and either the old or new
708 	 * result is a valid result.
709 	 */
710 	while (!IS_ROOT(dentry) && ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
711 		dentry = dentry->d_parent;
712 	return dentry;
713 }
714 
715 static int __choose_mds(struct ceph_mds_client *mdsc,
716 			struct ceph_mds_request *req)
717 {
718 	struct inode *inode;
719 	struct ceph_inode_info *ci;
720 	struct ceph_cap *cap;
721 	int mode = req->r_direct_mode;
722 	int mds = -1;
723 	u32 hash = req->r_direct_hash;
724 	bool is_hash = req->r_direct_is_hash;
725 
726 	/*
727 	 * is there a specific mds we should try?  ignore hint if we have
728 	 * no session and the mds is not up (active or recovering).
729 	 */
730 	if (req->r_resend_mds >= 0 &&
731 	    (__have_session(mdsc, req->r_resend_mds) ||
732 	     ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
733 		dout("choose_mds using resend_mds mds%d\n",
734 		     req->r_resend_mds);
735 		return req->r_resend_mds;
736 	}
737 
738 	if (mode == USE_RANDOM_MDS)
739 		goto random;
740 
741 	inode = NULL;
742 	if (req->r_inode) {
743 		inode = req->r_inode;
744 	} else if (req->r_dentry) {
745 		/* ignore race with rename; old or new d_parent is okay */
746 		struct dentry *parent = req->r_dentry->d_parent;
747 		struct inode *dir = d_inode(parent);
748 
749 		if (dir->i_sb != mdsc->fsc->sb) {
750 			/* not this fs! */
751 			inode = d_inode(req->r_dentry);
752 		} else if (ceph_snap(dir) != CEPH_NOSNAP) {
753 			/* direct snapped/virtual snapdir requests
754 			 * based on parent dir inode */
755 			struct dentry *dn = get_nonsnap_parent(parent);
756 			inode = d_inode(dn);
757 			dout("__choose_mds using nonsnap parent %p\n", inode);
758 		} else {
759 			/* dentry target */
760 			inode = d_inode(req->r_dentry);
761 			if (!inode || mode == USE_AUTH_MDS) {
762 				/* dir + name */
763 				inode = dir;
764 				hash = ceph_dentry_hash(dir, req->r_dentry);
765 				is_hash = true;
766 			}
767 		}
768 	}
769 
770 	dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
771 	     (int)hash, mode);
772 	if (!inode)
773 		goto random;
774 	ci = ceph_inode(inode);
775 
776 	if (is_hash && S_ISDIR(inode->i_mode)) {
777 		struct ceph_inode_frag frag;
778 		int found;
779 
780 		ceph_choose_frag(ci, hash, &frag, &found);
781 		if (found) {
782 			if (mode == USE_ANY_MDS && frag.ndist > 0) {
783 				u8 r;
784 
785 				/* choose a random replica */
786 				get_random_bytes(&r, 1);
787 				r %= frag.ndist;
788 				mds = frag.dist[r];
789 				dout("choose_mds %p %llx.%llx "
790 				     "frag %u mds%d (%d/%d)\n",
791 				     inode, ceph_vinop(inode),
792 				     frag.frag, mds,
793 				     (int)r, frag.ndist);
794 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
795 				    CEPH_MDS_STATE_ACTIVE)
796 					return mds;
797 			}
798 
799 			/* since this file/dir wasn't known to be
800 			 * replicated, then we want to look for the
801 			 * authoritative mds. */
802 			mode = USE_AUTH_MDS;
803 			if (frag.mds >= 0) {
804 				/* choose auth mds */
805 				mds = frag.mds;
806 				dout("choose_mds %p %llx.%llx "
807 				     "frag %u mds%d (auth)\n",
808 				     inode, ceph_vinop(inode), frag.frag, mds);
809 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
810 				    CEPH_MDS_STATE_ACTIVE)
811 					return mds;
812 			}
813 		}
814 	}
815 
816 	spin_lock(&ci->i_ceph_lock);
817 	cap = NULL;
818 	if (mode == USE_AUTH_MDS)
819 		cap = ci->i_auth_cap;
820 	if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
821 		cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
822 	if (!cap) {
823 		spin_unlock(&ci->i_ceph_lock);
824 		goto random;
825 	}
826 	mds = cap->session->s_mds;
827 	dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
828 	     inode, ceph_vinop(inode), mds,
829 	     cap == ci->i_auth_cap ? "auth " : "", cap);
830 	spin_unlock(&ci->i_ceph_lock);
831 	return mds;
832 
833 random:
834 	mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
835 	dout("choose_mds chose random mds%d\n", mds);
836 	return mds;
837 }
838 
839 
840 /*
841  * session messages
842  */
843 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
844 {
845 	struct ceph_msg *msg;
846 	struct ceph_mds_session_head *h;
847 
848 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
849 			   false);
850 	if (!msg) {
851 		pr_err("create_session_msg ENOMEM creating msg\n");
852 		return NULL;
853 	}
854 	h = msg->front.iov_base;
855 	h->op = cpu_to_le32(op);
856 	h->seq = cpu_to_le64(seq);
857 
858 	return msg;
859 }
860 
861 /*
862  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
863  * to include additional client metadata fields.
864  */
865 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
866 {
867 	struct ceph_msg *msg;
868 	struct ceph_mds_session_head *h;
869 	int i = -1;
870 	int metadata_bytes = 0;
871 	int metadata_key_count = 0;
872 	struct ceph_options *opt = mdsc->fsc->client->options;
873 	void *p;
874 
875 	const char* metadata[][2] = {
876 		{"hostname", utsname()->nodename},
877 		{"kernel_version", utsname()->release},
878 		{"entity_id", opt->name ? opt->name : ""},
879 		{NULL, NULL}
880 	};
881 
882 	/* Calculate serialized length of metadata */
883 	metadata_bytes = 4;  /* map length */
884 	for (i = 0; metadata[i][0] != NULL; ++i) {
885 		metadata_bytes += 8 + strlen(metadata[i][0]) +
886 			strlen(metadata[i][1]);
887 		metadata_key_count++;
888 	}
889 
890 	/* Allocate the message */
891 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
892 			   GFP_NOFS, false);
893 	if (!msg) {
894 		pr_err("create_session_msg ENOMEM creating msg\n");
895 		return NULL;
896 	}
897 	h = msg->front.iov_base;
898 	h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
899 	h->seq = cpu_to_le64(seq);
900 
901 	/*
902 	 * Serialize client metadata into waiting buffer space, using
903 	 * the format that userspace expects for map<string, string>
904 	 *
905 	 * ClientSession messages with metadata are v2
906 	 */
907 	msg->hdr.version = cpu_to_le16(2);
908 	msg->hdr.compat_version = cpu_to_le16(1);
909 
910 	/* The write pointer, following the session_head structure */
911 	p = msg->front.iov_base + sizeof(*h);
912 
913 	/* Number of entries in the map */
914 	ceph_encode_32(&p, metadata_key_count);
915 
916 	/* Two length-prefixed strings for each entry in the map */
917 	for (i = 0; metadata[i][0] != NULL; ++i) {
918 		size_t const key_len = strlen(metadata[i][0]);
919 		size_t const val_len = strlen(metadata[i][1]);
920 
921 		ceph_encode_32(&p, key_len);
922 		memcpy(p, metadata[i][0], key_len);
923 		p += key_len;
924 		ceph_encode_32(&p, val_len);
925 		memcpy(p, metadata[i][1], val_len);
926 		p += val_len;
927 	}
928 
929 	return msg;
930 }
931 
932 /*
933  * send session open request.
934  *
935  * called under mdsc->mutex
936  */
937 static int __open_session(struct ceph_mds_client *mdsc,
938 			  struct ceph_mds_session *session)
939 {
940 	struct ceph_msg *msg;
941 	int mstate;
942 	int mds = session->s_mds;
943 
944 	/* wait for mds to go active? */
945 	mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
946 	dout("open_session to mds%d (%s)\n", mds,
947 	     ceph_mds_state_name(mstate));
948 	session->s_state = CEPH_MDS_SESSION_OPENING;
949 	session->s_renew_requested = jiffies;
950 
951 	/* send connect message */
952 	msg = create_session_open_msg(mdsc, session->s_seq);
953 	if (!msg)
954 		return -ENOMEM;
955 	ceph_con_send(&session->s_con, msg);
956 	return 0;
957 }
958 
959 /*
960  * open sessions for any export targets for the given mds
961  *
962  * called under mdsc->mutex
963  */
964 static struct ceph_mds_session *
965 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
966 {
967 	struct ceph_mds_session *session;
968 
969 	session = __ceph_lookup_mds_session(mdsc, target);
970 	if (!session) {
971 		session = register_session(mdsc, target);
972 		if (IS_ERR(session))
973 			return session;
974 	}
975 	if (session->s_state == CEPH_MDS_SESSION_NEW ||
976 	    session->s_state == CEPH_MDS_SESSION_CLOSING)
977 		__open_session(mdsc, session);
978 
979 	return session;
980 }
981 
982 struct ceph_mds_session *
983 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
984 {
985 	struct ceph_mds_session *session;
986 
987 	dout("open_export_target_session to mds%d\n", target);
988 
989 	mutex_lock(&mdsc->mutex);
990 	session = __open_export_target_session(mdsc, target);
991 	mutex_unlock(&mdsc->mutex);
992 
993 	return session;
994 }
995 
996 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
997 					  struct ceph_mds_session *session)
998 {
999 	struct ceph_mds_info *mi;
1000 	struct ceph_mds_session *ts;
1001 	int i, mds = session->s_mds;
1002 
1003 	if (mds >= mdsc->mdsmap->m_max_mds)
1004 		return;
1005 
1006 	mi = &mdsc->mdsmap->m_info[mds];
1007 	dout("open_export_target_sessions for mds%d (%d targets)\n",
1008 	     session->s_mds, mi->num_export_targets);
1009 
1010 	for (i = 0; i < mi->num_export_targets; i++) {
1011 		ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1012 		if (!IS_ERR(ts))
1013 			ceph_put_mds_session(ts);
1014 	}
1015 }
1016 
1017 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1018 					   struct ceph_mds_session *session)
1019 {
1020 	mutex_lock(&mdsc->mutex);
1021 	__open_export_target_sessions(mdsc, session);
1022 	mutex_unlock(&mdsc->mutex);
1023 }
1024 
1025 /*
1026  * session caps
1027  */
1028 
1029 /* caller holds s_cap_lock, we drop it */
1030 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1031 				 struct ceph_mds_session *session)
1032 	__releases(session->s_cap_lock)
1033 {
1034 	LIST_HEAD(tmp_list);
1035 	list_splice_init(&session->s_cap_releases, &tmp_list);
1036 	session->s_num_cap_releases = 0;
1037 	spin_unlock(&session->s_cap_lock);
1038 
1039 	dout("cleanup_cap_releases mds%d\n", session->s_mds);
1040 	while (!list_empty(&tmp_list)) {
1041 		struct ceph_cap *cap;
1042 		/* zero out the in-progress message */
1043 		cap = list_first_entry(&tmp_list,
1044 					struct ceph_cap, session_caps);
1045 		list_del(&cap->session_caps);
1046 		ceph_put_cap(mdsc, cap);
1047 	}
1048 }
1049 
1050 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1051 				     struct ceph_mds_session *session)
1052 {
1053 	struct ceph_mds_request *req;
1054 	struct rb_node *p;
1055 
1056 	dout("cleanup_session_requests mds%d\n", session->s_mds);
1057 	mutex_lock(&mdsc->mutex);
1058 	while (!list_empty(&session->s_unsafe)) {
1059 		req = list_first_entry(&session->s_unsafe,
1060 				       struct ceph_mds_request, r_unsafe_item);
1061 		list_del_init(&req->r_unsafe_item);
1062 		pr_warn_ratelimited(" dropping unsafe request %llu\n",
1063 				    req->r_tid);
1064 		__unregister_request(mdsc, req);
1065 	}
1066 	/* zero r_attempts, so kick_requests() will re-send requests */
1067 	p = rb_first(&mdsc->request_tree);
1068 	while (p) {
1069 		req = rb_entry(p, struct ceph_mds_request, r_node);
1070 		p = rb_next(p);
1071 		if (req->r_session &&
1072 		    req->r_session->s_mds == session->s_mds)
1073 			req->r_attempts = 0;
1074 	}
1075 	mutex_unlock(&mdsc->mutex);
1076 }
1077 
1078 /*
1079  * Helper to safely iterate over all caps associated with a session, with
1080  * special care taken to handle a racing __ceph_remove_cap().
1081  *
1082  * Caller must hold session s_mutex.
1083  */
1084 static int iterate_session_caps(struct ceph_mds_session *session,
1085 				 int (*cb)(struct inode *, struct ceph_cap *,
1086 					    void *), void *arg)
1087 {
1088 	struct list_head *p;
1089 	struct ceph_cap *cap;
1090 	struct inode *inode, *last_inode = NULL;
1091 	struct ceph_cap *old_cap = NULL;
1092 	int ret;
1093 
1094 	dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1095 	spin_lock(&session->s_cap_lock);
1096 	p = session->s_caps.next;
1097 	while (p != &session->s_caps) {
1098 		cap = list_entry(p, struct ceph_cap, session_caps);
1099 		inode = igrab(&cap->ci->vfs_inode);
1100 		if (!inode) {
1101 			p = p->next;
1102 			continue;
1103 		}
1104 		session->s_cap_iterator = cap;
1105 		spin_unlock(&session->s_cap_lock);
1106 
1107 		if (last_inode) {
1108 			iput(last_inode);
1109 			last_inode = NULL;
1110 		}
1111 		if (old_cap) {
1112 			ceph_put_cap(session->s_mdsc, old_cap);
1113 			old_cap = NULL;
1114 		}
1115 
1116 		ret = cb(inode, cap, arg);
1117 		last_inode = inode;
1118 
1119 		spin_lock(&session->s_cap_lock);
1120 		p = p->next;
1121 		if (cap->ci == NULL) {
1122 			dout("iterate_session_caps  finishing cap %p removal\n",
1123 			     cap);
1124 			BUG_ON(cap->session != session);
1125 			cap->session = NULL;
1126 			list_del_init(&cap->session_caps);
1127 			session->s_nr_caps--;
1128 			if (cap->queue_release) {
1129 				list_add_tail(&cap->session_caps,
1130 					      &session->s_cap_releases);
1131 				session->s_num_cap_releases++;
1132 			} else {
1133 				old_cap = cap;  /* put_cap it w/o locks held */
1134 			}
1135 		}
1136 		if (ret < 0)
1137 			goto out;
1138 	}
1139 	ret = 0;
1140 out:
1141 	session->s_cap_iterator = NULL;
1142 	spin_unlock(&session->s_cap_lock);
1143 
1144 	iput(last_inode);
1145 	if (old_cap)
1146 		ceph_put_cap(session->s_mdsc, old_cap);
1147 
1148 	return ret;
1149 }
1150 
1151 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1152 				  void *arg)
1153 {
1154 	struct ceph_inode_info *ci = ceph_inode(inode);
1155 	LIST_HEAD(to_remove);
1156 	int drop = 0;
1157 
1158 	dout("removing cap %p, ci is %p, inode is %p\n",
1159 	     cap, ci, &ci->vfs_inode);
1160 	spin_lock(&ci->i_ceph_lock);
1161 	__ceph_remove_cap(cap, false);
1162 	if (!ci->i_auth_cap) {
1163 		struct ceph_cap_flush *cf;
1164 		struct ceph_mds_client *mdsc =
1165 			ceph_sb_to_client(inode->i_sb)->mdsc;
1166 
1167 		while (true) {
1168 			struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1169 			if (!n)
1170 				break;
1171 			cf = rb_entry(n, struct ceph_cap_flush, i_node);
1172 			rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1173 			list_add(&cf->list, &to_remove);
1174 		}
1175 
1176 		spin_lock(&mdsc->cap_dirty_lock);
1177 
1178 		list_for_each_entry(cf, &to_remove, list)
1179 			rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1180 
1181 		if (!list_empty(&ci->i_dirty_item)) {
1182 			pr_warn_ratelimited(
1183 				" dropping dirty %s state for %p %lld\n",
1184 				ceph_cap_string(ci->i_dirty_caps),
1185 				inode, ceph_ino(inode));
1186 			ci->i_dirty_caps = 0;
1187 			list_del_init(&ci->i_dirty_item);
1188 			drop = 1;
1189 		}
1190 		if (!list_empty(&ci->i_flushing_item)) {
1191 			pr_warn_ratelimited(
1192 				" dropping dirty+flushing %s state for %p %lld\n",
1193 				ceph_cap_string(ci->i_flushing_caps),
1194 				inode, ceph_ino(inode));
1195 			ci->i_flushing_caps = 0;
1196 			list_del_init(&ci->i_flushing_item);
1197 			mdsc->num_cap_flushing--;
1198 			drop = 1;
1199 		}
1200 		spin_unlock(&mdsc->cap_dirty_lock);
1201 
1202 		if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1203 			list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1204 			ci->i_prealloc_cap_flush = NULL;
1205 		}
1206 	}
1207 	spin_unlock(&ci->i_ceph_lock);
1208 	while (!list_empty(&to_remove)) {
1209 		struct ceph_cap_flush *cf;
1210 		cf = list_first_entry(&to_remove,
1211 				      struct ceph_cap_flush, list);
1212 		list_del(&cf->list);
1213 		ceph_free_cap_flush(cf);
1214 	}
1215 	while (drop--)
1216 		iput(inode);
1217 	return 0;
1218 }
1219 
1220 /*
1221  * caller must hold session s_mutex
1222  */
1223 static void remove_session_caps(struct ceph_mds_session *session)
1224 {
1225 	dout("remove_session_caps on %p\n", session);
1226 	iterate_session_caps(session, remove_session_caps_cb, NULL);
1227 
1228 	spin_lock(&session->s_cap_lock);
1229 	if (session->s_nr_caps > 0) {
1230 		struct super_block *sb = session->s_mdsc->fsc->sb;
1231 		struct inode *inode;
1232 		struct ceph_cap *cap, *prev = NULL;
1233 		struct ceph_vino vino;
1234 		/*
1235 		 * iterate_session_caps() skips inodes that are being
1236 		 * deleted, we need to wait until deletions are complete.
1237 		 * __wait_on_freeing_inode() is designed for the job,
1238 		 * but it is not exported, so use lookup inode function
1239 		 * to access it.
1240 		 */
1241 		while (!list_empty(&session->s_caps)) {
1242 			cap = list_entry(session->s_caps.next,
1243 					 struct ceph_cap, session_caps);
1244 			if (cap == prev)
1245 				break;
1246 			prev = cap;
1247 			vino = cap->ci->i_vino;
1248 			spin_unlock(&session->s_cap_lock);
1249 
1250 			inode = ceph_find_inode(sb, vino);
1251 			iput(inode);
1252 
1253 			spin_lock(&session->s_cap_lock);
1254 		}
1255 	}
1256 
1257 	// drop cap expires and unlock s_cap_lock
1258 	cleanup_cap_releases(session->s_mdsc, session);
1259 
1260 	BUG_ON(session->s_nr_caps > 0);
1261 	BUG_ON(!list_empty(&session->s_cap_flushing));
1262 }
1263 
1264 /*
1265  * wake up any threads waiting on this session's caps.  if the cap is
1266  * old (didn't get renewed on the client reconnect), remove it now.
1267  *
1268  * caller must hold s_mutex.
1269  */
1270 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1271 			      void *arg)
1272 {
1273 	struct ceph_inode_info *ci = ceph_inode(inode);
1274 
1275 	wake_up_all(&ci->i_cap_wq);
1276 	if (arg) {
1277 		spin_lock(&ci->i_ceph_lock);
1278 		ci->i_wanted_max_size = 0;
1279 		ci->i_requested_max_size = 0;
1280 		spin_unlock(&ci->i_ceph_lock);
1281 	}
1282 	return 0;
1283 }
1284 
1285 static void wake_up_session_caps(struct ceph_mds_session *session,
1286 				 int reconnect)
1287 {
1288 	dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1289 	iterate_session_caps(session, wake_up_session_cb,
1290 			     (void *)(unsigned long)reconnect);
1291 }
1292 
1293 /*
1294  * Send periodic message to MDS renewing all currently held caps.  The
1295  * ack will reset the expiration for all caps from this session.
1296  *
1297  * caller holds s_mutex
1298  */
1299 static int send_renew_caps(struct ceph_mds_client *mdsc,
1300 			   struct ceph_mds_session *session)
1301 {
1302 	struct ceph_msg *msg;
1303 	int state;
1304 
1305 	if (time_after_eq(jiffies, session->s_cap_ttl) &&
1306 	    time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1307 		pr_info("mds%d caps stale\n", session->s_mds);
1308 	session->s_renew_requested = jiffies;
1309 
1310 	/* do not try to renew caps until a recovering mds has reconnected
1311 	 * with its clients. */
1312 	state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1313 	if (state < CEPH_MDS_STATE_RECONNECT) {
1314 		dout("send_renew_caps ignoring mds%d (%s)\n",
1315 		     session->s_mds, ceph_mds_state_name(state));
1316 		return 0;
1317 	}
1318 
1319 	dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1320 		ceph_mds_state_name(state));
1321 	msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1322 				 ++session->s_renew_seq);
1323 	if (!msg)
1324 		return -ENOMEM;
1325 	ceph_con_send(&session->s_con, msg);
1326 	return 0;
1327 }
1328 
1329 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1330 			     struct ceph_mds_session *session, u64 seq)
1331 {
1332 	struct ceph_msg *msg;
1333 
1334 	dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1335 	     session->s_mds, ceph_session_state_name(session->s_state), seq);
1336 	msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1337 	if (!msg)
1338 		return -ENOMEM;
1339 	ceph_con_send(&session->s_con, msg);
1340 	return 0;
1341 }
1342 
1343 
1344 /*
1345  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1346  *
1347  * Called under session->s_mutex
1348  */
1349 static void renewed_caps(struct ceph_mds_client *mdsc,
1350 			 struct ceph_mds_session *session, int is_renew)
1351 {
1352 	int was_stale;
1353 	int wake = 0;
1354 
1355 	spin_lock(&session->s_cap_lock);
1356 	was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1357 
1358 	session->s_cap_ttl = session->s_renew_requested +
1359 		mdsc->mdsmap->m_session_timeout*HZ;
1360 
1361 	if (was_stale) {
1362 		if (time_before(jiffies, session->s_cap_ttl)) {
1363 			pr_info("mds%d caps renewed\n", session->s_mds);
1364 			wake = 1;
1365 		} else {
1366 			pr_info("mds%d caps still stale\n", session->s_mds);
1367 		}
1368 	}
1369 	dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1370 	     session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1371 	     time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1372 	spin_unlock(&session->s_cap_lock);
1373 
1374 	if (wake)
1375 		wake_up_session_caps(session, 0);
1376 }
1377 
1378 /*
1379  * send a session close request
1380  */
1381 static int request_close_session(struct ceph_mds_client *mdsc,
1382 				 struct ceph_mds_session *session)
1383 {
1384 	struct ceph_msg *msg;
1385 
1386 	dout("request_close_session mds%d state %s seq %lld\n",
1387 	     session->s_mds, ceph_session_state_name(session->s_state),
1388 	     session->s_seq);
1389 	msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1390 	if (!msg)
1391 		return -ENOMEM;
1392 	ceph_con_send(&session->s_con, msg);
1393 	return 0;
1394 }
1395 
1396 /*
1397  * Called with s_mutex held.
1398  */
1399 static int __close_session(struct ceph_mds_client *mdsc,
1400 			 struct ceph_mds_session *session)
1401 {
1402 	if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1403 		return 0;
1404 	session->s_state = CEPH_MDS_SESSION_CLOSING;
1405 	return request_close_session(mdsc, session);
1406 }
1407 
1408 /*
1409  * Trim old(er) caps.
1410  *
1411  * Because we can't cache an inode without one or more caps, we do
1412  * this indirectly: if a cap is unused, we prune its aliases, at which
1413  * point the inode will hopefully get dropped to.
1414  *
1415  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1416  * memory pressure from the MDS, though, so it needn't be perfect.
1417  */
1418 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1419 {
1420 	struct ceph_mds_session *session = arg;
1421 	struct ceph_inode_info *ci = ceph_inode(inode);
1422 	int used, wanted, oissued, mine;
1423 
1424 	if (session->s_trim_caps <= 0)
1425 		return -1;
1426 
1427 	spin_lock(&ci->i_ceph_lock);
1428 	mine = cap->issued | cap->implemented;
1429 	used = __ceph_caps_used(ci);
1430 	wanted = __ceph_caps_file_wanted(ci);
1431 	oissued = __ceph_caps_issued_other(ci, cap);
1432 
1433 	dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1434 	     inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1435 	     ceph_cap_string(used), ceph_cap_string(wanted));
1436 	if (cap == ci->i_auth_cap) {
1437 		if (ci->i_dirty_caps || ci->i_flushing_caps ||
1438 		    !list_empty(&ci->i_cap_snaps))
1439 			goto out;
1440 		if ((used | wanted) & CEPH_CAP_ANY_WR)
1441 			goto out;
1442 	}
1443 	/* The inode has cached pages, but it's no longer used.
1444 	 * we can safely drop it */
1445 	if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1446 	    !(oissued & CEPH_CAP_FILE_CACHE)) {
1447 	  used = 0;
1448 	  oissued = 0;
1449 	}
1450 	if ((used | wanted) & ~oissued & mine)
1451 		goto out;   /* we need these caps */
1452 
1453 	session->s_trim_caps--;
1454 	if (oissued) {
1455 		/* we aren't the only cap.. just remove us */
1456 		__ceph_remove_cap(cap, true);
1457 	} else {
1458 		/* try dropping referring dentries */
1459 		spin_unlock(&ci->i_ceph_lock);
1460 		d_prune_aliases(inode);
1461 		dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1462 		     inode, cap, atomic_read(&inode->i_count));
1463 		return 0;
1464 	}
1465 
1466 out:
1467 	spin_unlock(&ci->i_ceph_lock);
1468 	return 0;
1469 }
1470 
1471 /*
1472  * Trim session cap count down to some max number.
1473  */
1474 static int trim_caps(struct ceph_mds_client *mdsc,
1475 		     struct ceph_mds_session *session,
1476 		     int max_caps)
1477 {
1478 	int trim_caps = session->s_nr_caps - max_caps;
1479 
1480 	dout("trim_caps mds%d start: %d / %d, trim %d\n",
1481 	     session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1482 	if (trim_caps > 0) {
1483 		session->s_trim_caps = trim_caps;
1484 		iterate_session_caps(session, trim_caps_cb, session);
1485 		dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1486 		     session->s_mds, session->s_nr_caps, max_caps,
1487 			trim_caps - session->s_trim_caps);
1488 		session->s_trim_caps = 0;
1489 	}
1490 
1491 	ceph_send_cap_releases(mdsc, session);
1492 	return 0;
1493 }
1494 
1495 static int check_capsnap_flush(struct ceph_inode_info *ci,
1496 			       u64 want_snap_seq)
1497 {
1498 	int ret = 1;
1499 	spin_lock(&ci->i_ceph_lock);
1500 	if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1501 		struct ceph_cap_snap *capsnap =
1502 			list_first_entry(&ci->i_cap_snaps,
1503 					 struct ceph_cap_snap, ci_item);
1504 		ret = capsnap->follows >= want_snap_seq;
1505 	}
1506 	spin_unlock(&ci->i_ceph_lock);
1507 	return ret;
1508 }
1509 
1510 static int check_caps_flush(struct ceph_mds_client *mdsc,
1511 			    u64 want_flush_tid)
1512 {
1513 	struct rb_node *n;
1514 	struct ceph_cap_flush *cf;
1515 	int ret = 1;
1516 
1517 	spin_lock(&mdsc->cap_dirty_lock);
1518 	n = rb_first(&mdsc->cap_flush_tree);
1519 	cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1520 	if (cf && cf->tid <= want_flush_tid) {
1521 		dout("check_caps_flush still flushing tid %llu <= %llu\n",
1522 		     cf->tid, want_flush_tid);
1523 		ret = 0;
1524 	}
1525 	spin_unlock(&mdsc->cap_dirty_lock);
1526 	return ret;
1527 }
1528 
1529 /*
1530  * flush all dirty inode data to disk.
1531  *
1532  * returns true if we've flushed through want_flush_tid
1533  */
1534 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1535 			    u64 want_flush_tid, u64 want_snap_seq)
1536 {
1537 	int mds;
1538 
1539 	dout("check_caps_flush want %llu snap want %llu\n",
1540 	     want_flush_tid, want_snap_seq);
1541 	mutex_lock(&mdsc->mutex);
1542 	for (mds = 0; mds < mdsc->max_sessions; ) {
1543 		struct ceph_mds_session *session = mdsc->sessions[mds];
1544 		struct inode *inode = NULL;
1545 
1546 		if (!session) {
1547 			mds++;
1548 			continue;
1549 		}
1550 		get_session(session);
1551 		mutex_unlock(&mdsc->mutex);
1552 
1553 		mutex_lock(&session->s_mutex);
1554 		if (!list_empty(&session->s_cap_snaps_flushing)) {
1555 			struct ceph_cap_snap *capsnap =
1556 				list_first_entry(&session->s_cap_snaps_flushing,
1557 						 struct ceph_cap_snap,
1558 						 flushing_item);
1559 			struct ceph_inode_info *ci = capsnap->ci;
1560 			if (!check_capsnap_flush(ci, want_snap_seq)) {
1561 				dout("check_cap_flush still flushing snap %p "
1562 				     "follows %lld <= %lld to mds%d\n",
1563 				     &ci->vfs_inode, capsnap->follows,
1564 				     want_snap_seq, mds);
1565 				inode = igrab(&ci->vfs_inode);
1566 			}
1567 		}
1568 		mutex_unlock(&session->s_mutex);
1569 		ceph_put_mds_session(session);
1570 
1571 		if (inode) {
1572 			wait_event(mdsc->cap_flushing_wq,
1573 				   check_capsnap_flush(ceph_inode(inode),
1574 						       want_snap_seq));
1575 			iput(inode);
1576 		} else {
1577 			mds++;
1578 		}
1579 
1580 		mutex_lock(&mdsc->mutex);
1581 	}
1582 	mutex_unlock(&mdsc->mutex);
1583 
1584 	wait_event(mdsc->cap_flushing_wq,
1585 		   check_caps_flush(mdsc, want_flush_tid));
1586 
1587 	dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1588 }
1589 
1590 /*
1591  * called under s_mutex
1592  */
1593 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1594 			    struct ceph_mds_session *session)
1595 {
1596 	struct ceph_msg *msg = NULL;
1597 	struct ceph_mds_cap_release *head;
1598 	struct ceph_mds_cap_item *item;
1599 	struct ceph_cap *cap;
1600 	LIST_HEAD(tmp_list);
1601 	int num_cap_releases;
1602 
1603 	spin_lock(&session->s_cap_lock);
1604 again:
1605 	list_splice_init(&session->s_cap_releases, &tmp_list);
1606 	num_cap_releases = session->s_num_cap_releases;
1607 	session->s_num_cap_releases = 0;
1608 	spin_unlock(&session->s_cap_lock);
1609 
1610 	while (!list_empty(&tmp_list)) {
1611 		if (!msg) {
1612 			msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1613 					PAGE_CACHE_SIZE, GFP_NOFS, false);
1614 			if (!msg)
1615 				goto out_err;
1616 			head = msg->front.iov_base;
1617 			head->num = cpu_to_le32(0);
1618 			msg->front.iov_len = sizeof(*head);
1619 		}
1620 		cap = list_first_entry(&tmp_list, struct ceph_cap,
1621 					session_caps);
1622 		list_del(&cap->session_caps);
1623 		num_cap_releases--;
1624 
1625 		head = msg->front.iov_base;
1626 		le32_add_cpu(&head->num, 1);
1627 		item = msg->front.iov_base + msg->front.iov_len;
1628 		item->ino = cpu_to_le64(cap->cap_ino);
1629 		item->cap_id = cpu_to_le64(cap->cap_id);
1630 		item->migrate_seq = cpu_to_le32(cap->mseq);
1631 		item->seq = cpu_to_le32(cap->issue_seq);
1632 		msg->front.iov_len += sizeof(*item);
1633 
1634 		ceph_put_cap(mdsc, cap);
1635 
1636 		if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1637 			msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1638 			dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1639 			ceph_con_send(&session->s_con, msg);
1640 			msg = NULL;
1641 		}
1642 	}
1643 
1644 	BUG_ON(num_cap_releases != 0);
1645 
1646 	spin_lock(&session->s_cap_lock);
1647 	if (!list_empty(&session->s_cap_releases))
1648 		goto again;
1649 	spin_unlock(&session->s_cap_lock);
1650 
1651 	if (msg) {
1652 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1653 		dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1654 		ceph_con_send(&session->s_con, msg);
1655 	}
1656 	return;
1657 out_err:
1658 	pr_err("send_cap_releases mds%d, failed to allocate message\n",
1659 		session->s_mds);
1660 	spin_lock(&session->s_cap_lock);
1661 	list_splice(&tmp_list, &session->s_cap_releases);
1662 	session->s_num_cap_releases += num_cap_releases;
1663 	spin_unlock(&session->s_cap_lock);
1664 }
1665 
1666 /*
1667  * requests
1668  */
1669 
1670 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1671 				    struct inode *dir)
1672 {
1673 	struct ceph_inode_info *ci = ceph_inode(dir);
1674 	struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1675 	struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1676 	size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1677 		      sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1678 	int order, num_entries;
1679 
1680 	spin_lock(&ci->i_ceph_lock);
1681 	num_entries = ci->i_files + ci->i_subdirs;
1682 	spin_unlock(&ci->i_ceph_lock);
1683 	num_entries = max(num_entries, 1);
1684 	num_entries = min(num_entries, opt->max_readdir);
1685 
1686 	order = get_order(size * num_entries);
1687 	while (order >= 0) {
1688 		rinfo->dir_in = (void*)__get_free_pages(GFP_KERNEL |
1689 							__GFP_NOWARN,
1690 							order);
1691 		if (rinfo->dir_in)
1692 			break;
1693 		order--;
1694 	}
1695 	if (!rinfo->dir_in)
1696 		return -ENOMEM;
1697 
1698 	num_entries = (PAGE_SIZE << order) / size;
1699 	num_entries = min(num_entries, opt->max_readdir);
1700 
1701 	rinfo->dir_buf_size = PAGE_SIZE << order;
1702 	req->r_num_caps = num_entries + 1;
1703 	req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1704 	req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1705 	return 0;
1706 }
1707 
1708 /*
1709  * Create an mds request.
1710  */
1711 struct ceph_mds_request *
1712 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1713 {
1714 	struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1715 
1716 	if (!req)
1717 		return ERR_PTR(-ENOMEM);
1718 
1719 	mutex_init(&req->r_fill_mutex);
1720 	req->r_mdsc = mdsc;
1721 	req->r_started = jiffies;
1722 	req->r_resend_mds = -1;
1723 	INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1724 	INIT_LIST_HEAD(&req->r_unsafe_target_item);
1725 	req->r_fmode = -1;
1726 	kref_init(&req->r_kref);
1727 	INIT_LIST_HEAD(&req->r_wait);
1728 	init_completion(&req->r_completion);
1729 	init_completion(&req->r_safe_completion);
1730 	INIT_LIST_HEAD(&req->r_unsafe_item);
1731 
1732 	req->r_stamp = current_fs_time(mdsc->fsc->sb);
1733 
1734 	req->r_op = op;
1735 	req->r_direct_mode = mode;
1736 	return req;
1737 }
1738 
1739 /*
1740  * return oldest (lowest) request, tid in request tree, 0 if none.
1741  *
1742  * called under mdsc->mutex.
1743  */
1744 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1745 {
1746 	if (RB_EMPTY_ROOT(&mdsc->request_tree))
1747 		return NULL;
1748 	return rb_entry(rb_first(&mdsc->request_tree),
1749 			struct ceph_mds_request, r_node);
1750 }
1751 
1752 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1753 {
1754 	return mdsc->oldest_tid;
1755 }
1756 
1757 /*
1758  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1759  * on build_path_from_dentry in fs/cifs/dir.c.
1760  *
1761  * If @stop_on_nosnap, generate path relative to the first non-snapped
1762  * inode.
1763  *
1764  * Encode hidden .snap dirs as a double /, i.e.
1765  *   foo/.snap/bar -> foo//bar
1766  */
1767 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1768 			   int stop_on_nosnap)
1769 {
1770 	struct dentry *temp;
1771 	char *path;
1772 	int len, pos;
1773 	unsigned seq;
1774 
1775 	if (dentry == NULL)
1776 		return ERR_PTR(-EINVAL);
1777 
1778 retry:
1779 	len = 0;
1780 	seq = read_seqbegin(&rename_lock);
1781 	rcu_read_lock();
1782 	for (temp = dentry; !IS_ROOT(temp);) {
1783 		struct inode *inode = d_inode(temp);
1784 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1785 			len++;  /* slash only */
1786 		else if (stop_on_nosnap && inode &&
1787 			 ceph_snap(inode) == CEPH_NOSNAP)
1788 			break;
1789 		else
1790 			len += 1 + temp->d_name.len;
1791 		temp = temp->d_parent;
1792 	}
1793 	rcu_read_unlock();
1794 	if (len)
1795 		len--;  /* no leading '/' */
1796 
1797 	path = kmalloc(len+1, GFP_NOFS);
1798 	if (path == NULL)
1799 		return ERR_PTR(-ENOMEM);
1800 	pos = len;
1801 	path[pos] = 0;	/* trailing null */
1802 	rcu_read_lock();
1803 	for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1804 		struct inode *inode;
1805 
1806 		spin_lock(&temp->d_lock);
1807 		inode = d_inode(temp);
1808 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1809 			dout("build_path path+%d: %p SNAPDIR\n",
1810 			     pos, temp);
1811 		} else if (stop_on_nosnap && inode &&
1812 			   ceph_snap(inode) == CEPH_NOSNAP) {
1813 			spin_unlock(&temp->d_lock);
1814 			break;
1815 		} else {
1816 			pos -= temp->d_name.len;
1817 			if (pos < 0) {
1818 				spin_unlock(&temp->d_lock);
1819 				break;
1820 			}
1821 			strncpy(path + pos, temp->d_name.name,
1822 				temp->d_name.len);
1823 		}
1824 		spin_unlock(&temp->d_lock);
1825 		if (pos)
1826 			path[--pos] = '/';
1827 		temp = temp->d_parent;
1828 	}
1829 	rcu_read_unlock();
1830 	if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1831 		pr_err("build_path did not end path lookup where "
1832 		       "expected, namelen is %d, pos is %d\n", len, pos);
1833 		/* presumably this is only possible if racing with a
1834 		   rename of one of the parent directories (we can not
1835 		   lock the dentries above us to prevent this, but
1836 		   retrying should be harmless) */
1837 		kfree(path);
1838 		goto retry;
1839 	}
1840 
1841 	*base = ceph_ino(d_inode(temp));
1842 	*plen = len;
1843 	dout("build_path on %p %d built %llx '%.*s'\n",
1844 	     dentry, d_count(dentry), *base, len, path);
1845 	return path;
1846 }
1847 
1848 static int build_dentry_path(struct dentry *dentry,
1849 			     const char **ppath, int *ppathlen, u64 *pino,
1850 			     int *pfreepath)
1851 {
1852 	char *path;
1853 
1854 	if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP) {
1855 		*pino = ceph_ino(d_inode(dentry->d_parent));
1856 		*ppath = dentry->d_name.name;
1857 		*ppathlen = dentry->d_name.len;
1858 		return 0;
1859 	}
1860 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1861 	if (IS_ERR(path))
1862 		return PTR_ERR(path);
1863 	*ppath = path;
1864 	*pfreepath = 1;
1865 	return 0;
1866 }
1867 
1868 static int build_inode_path(struct inode *inode,
1869 			    const char **ppath, int *ppathlen, u64 *pino,
1870 			    int *pfreepath)
1871 {
1872 	struct dentry *dentry;
1873 	char *path;
1874 
1875 	if (ceph_snap(inode) == CEPH_NOSNAP) {
1876 		*pino = ceph_ino(inode);
1877 		*ppathlen = 0;
1878 		return 0;
1879 	}
1880 	dentry = d_find_alias(inode);
1881 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1882 	dput(dentry);
1883 	if (IS_ERR(path))
1884 		return PTR_ERR(path);
1885 	*ppath = path;
1886 	*pfreepath = 1;
1887 	return 0;
1888 }
1889 
1890 /*
1891  * request arguments may be specified via an inode *, a dentry *, or
1892  * an explicit ino+path.
1893  */
1894 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1895 				  const char *rpath, u64 rino,
1896 				  const char **ppath, int *pathlen,
1897 				  u64 *ino, int *freepath)
1898 {
1899 	int r = 0;
1900 
1901 	if (rinode) {
1902 		r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1903 		dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1904 		     ceph_snap(rinode));
1905 	} else if (rdentry) {
1906 		r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1907 		dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1908 		     *ppath);
1909 	} else if (rpath || rino) {
1910 		*ino = rino;
1911 		*ppath = rpath;
1912 		*pathlen = rpath ? strlen(rpath) : 0;
1913 		dout(" path %.*s\n", *pathlen, rpath);
1914 	}
1915 
1916 	return r;
1917 }
1918 
1919 /*
1920  * called under mdsc->mutex
1921  */
1922 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1923 					       struct ceph_mds_request *req,
1924 					       int mds, bool drop_cap_releases)
1925 {
1926 	struct ceph_msg *msg;
1927 	struct ceph_mds_request_head *head;
1928 	const char *path1 = NULL;
1929 	const char *path2 = NULL;
1930 	u64 ino1 = 0, ino2 = 0;
1931 	int pathlen1 = 0, pathlen2 = 0;
1932 	int freepath1 = 0, freepath2 = 0;
1933 	int len;
1934 	u16 releases;
1935 	void *p, *end;
1936 	int ret;
1937 
1938 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
1939 			      req->r_path1, req->r_ino1.ino,
1940 			      &path1, &pathlen1, &ino1, &freepath1);
1941 	if (ret < 0) {
1942 		msg = ERR_PTR(ret);
1943 		goto out;
1944 	}
1945 
1946 	ret = set_request_path_attr(NULL, req->r_old_dentry,
1947 			      req->r_path2, req->r_ino2.ino,
1948 			      &path2, &pathlen2, &ino2, &freepath2);
1949 	if (ret < 0) {
1950 		msg = ERR_PTR(ret);
1951 		goto out_free1;
1952 	}
1953 
1954 	len = sizeof(*head) +
1955 		pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1956 		sizeof(struct ceph_timespec);
1957 
1958 	/* calculate (max) length for cap releases */
1959 	len += sizeof(struct ceph_mds_request_release) *
1960 		(!!req->r_inode_drop + !!req->r_dentry_drop +
1961 		 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1962 	if (req->r_dentry_drop)
1963 		len += req->r_dentry->d_name.len;
1964 	if (req->r_old_dentry_drop)
1965 		len += req->r_old_dentry->d_name.len;
1966 
1967 	msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1968 	if (!msg) {
1969 		msg = ERR_PTR(-ENOMEM);
1970 		goto out_free2;
1971 	}
1972 
1973 	msg->hdr.version = cpu_to_le16(2);
1974 	msg->hdr.tid = cpu_to_le64(req->r_tid);
1975 
1976 	head = msg->front.iov_base;
1977 	p = msg->front.iov_base + sizeof(*head);
1978 	end = msg->front.iov_base + msg->front.iov_len;
1979 
1980 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1981 	head->op = cpu_to_le32(req->r_op);
1982 	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1983 	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1984 	head->args = req->r_args;
1985 
1986 	ceph_encode_filepath(&p, end, ino1, path1);
1987 	ceph_encode_filepath(&p, end, ino2, path2);
1988 
1989 	/* make note of release offset, in case we need to replay */
1990 	req->r_request_release_offset = p - msg->front.iov_base;
1991 
1992 	/* cap releases */
1993 	releases = 0;
1994 	if (req->r_inode_drop)
1995 		releases += ceph_encode_inode_release(&p,
1996 		      req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1997 		      mds, req->r_inode_drop, req->r_inode_unless, 0);
1998 	if (req->r_dentry_drop)
1999 		releases += ceph_encode_dentry_release(&p, req->r_dentry,
2000 		       mds, req->r_dentry_drop, req->r_dentry_unless);
2001 	if (req->r_old_dentry_drop)
2002 		releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2003 		       mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
2004 	if (req->r_old_inode_drop)
2005 		releases += ceph_encode_inode_release(&p,
2006 		      d_inode(req->r_old_dentry),
2007 		      mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2008 
2009 	if (drop_cap_releases) {
2010 		releases = 0;
2011 		p = msg->front.iov_base + req->r_request_release_offset;
2012 	}
2013 
2014 	head->num_releases = cpu_to_le16(releases);
2015 
2016 	/* time stamp */
2017 	{
2018 		struct ceph_timespec ts;
2019 		ceph_encode_timespec(&ts, &req->r_stamp);
2020 		ceph_encode_copy(&p, &ts, sizeof(ts));
2021 	}
2022 
2023 	BUG_ON(p > end);
2024 	msg->front.iov_len = p - msg->front.iov_base;
2025 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2026 
2027 	if (req->r_pagelist) {
2028 		struct ceph_pagelist *pagelist = req->r_pagelist;
2029 		atomic_inc(&pagelist->refcnt);
2030 		ceph_msg_data_add_pagelist(msg, pagelist);
2031 		msg->hdr.data_len = cpu_to_le32(pagelist->length);
2032 	} else {
2033 		msg->hdr.data_len = 0;
2034 	}
2035 
2036 	msg->hdr.data_off = cpu_to_le16(0);
2037 
2038 out_free2:
2039 	if (freepath2)
2040 		kfree((char *)path2);
2041 out_free1:
2042 	if (freepath1)
2043 		kfree((char *)path1);
2044 out:
2045 	return msg;
2046 }
2047 
2048 /*
2049  * called under mdsc->mutex if error, under no mutex if
2050  * success.
2051  */
2052 static void complete_request(struct ceph_mds_client *mdsc,
2053 			     struct ceph_mds_request *req)
2054 {
2055 	if (req->r_callback)
2056 		req->r_callback(mdsc, req);
2057 	else
2058 		complete_all(&req->r_completion);
2059 }
2060 
2061 /*
2062  * called under mdsc->mutex
2063  */
2064 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2065 				  struct ceph_mds_request *req,
2066 				  int mds, bool drop_cap_releases)
2067 {
2068 	struct ceph_mds_request_head *rhead;
2069 	struct ceph_msg *msg;
2070 	int flags = 0;
2071 
2072 	req->r_attempts++;
2073 	if (req->r_inode) {
2074 		struct ceph_cap *cap =
2075 			ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2076 
2077 		if (cap)
2078 			req->r_sent_on_mseq = cap->mseq;
2079 		else
2080 			req->r_sent_on_mseq = -1;
2081 	}
2082 	dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2083 	     req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2084 
2085 	if (req->r_got_unsafe) {
2086 		void *p;
2087 		/*
2088 		 * Replay.  Do not regenerate message (and rebuild
2089 		 * paths, etc.); just use the original message.
2090 		 * Rebuilding paths will break for renames because
2091 		 * d_move mangles the src name.
2092 		 */
2093 		msg = req->r_request;
2094 		rhead = msg->front.iov_base;
2095 
2096 		flags = le32_to_cpu(rhead->flags);
2097 		flags |= CEPH_MDS_FLAG_REPLAY;
2098 		rhead->flags = cpu_to_le32(flags);
2099 
2100 		if (req->r_target_inode)
2101 			rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2102 
2103 		rhead->num_retry = req->r_attempts - 1;
2104 
2105 		/* remove cap/dentry releases from message */
2106 		rhead->num_releases = 0;
2107 
2108 		/* time stamp */
2109 		p = msg->front.iov_base + req->r_request_release_offset;
2110 		{
2111 			struct ceph_timespec ts;
2112 			ceph_encode_timespec(&ts, &req->r_stamp);
2113 			ceph_encode_copy(&p, &ts, sizeof(ts));
2114 		}
2115 
2116 		msg->front.iov_len = p - msg->front.iov_base;
2117 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2118 		return 0;
2119 	}
2120 
2121 	if (req->r_request) {
2122 		ceph_msg_put(req->r_request);
2123 		req->r_request = NULL;
2124 	}
2125 	msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2126 	if (IS_ERR(msg)) {
2127 		req->r_err = PTR_ERR(msg);
2128 		return PTR_ERR(msg);
2129 	}
2130 	req->r_request = msg;
2131 
2132 	rhead = msg->front.iov_base;
2133 	rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2134 	if (req->r_got_unsafe)
2135 		flags |= CEPH_MDS_FLAG_REPLAY;
2136 	if (req->r_locked_dir)
2137 		flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2138 	rhead->flags = cpu_to_le32(flags);
2139 	rhead->num_fwd = req->r_num_fwd;
2140 	rhead->num_retry = req->r_attempts - 1;
2141 	rhead->ino = 0;
2142 
2143 	dout(" r_locked_dir = %p\n", req->r_locked_dir);
2144 	return 0;
2145 }
2146 
2147 /*
2148  * send request, or put it on the appropriate wait list.
2149  */
2150 static int __do_request(struct ceph_mds_client *mdsc,
2151 			struct ceph_mds_request *req)
2152 {
2153 	struct ceph_mds_session *session = NULL;
2154 	int mds = -1;
2155 	int err = 0;
2156 
2157 	if (req->r_err || req->r_got_result) {
2158 		if (req->r_aborted)
2159 			__unregister_request(mdsc, req);
2160 		goto out;
2161 	}
2162 
2163 	if (req->r_timeout &&
2164 	    time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2165 		dout("do_request timed out\n");
2166 		err = -EIO;
2167 		goto finish;
2168 	}
2169 	if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2170 		dout("do_request forced umount\n");
2171 		err = -EIO;
2172 		goto finish;
2173 	}
2174 
2175 	put_request_session(req);
2176 
2177 	mds = __choose_mds(mdsc, req);
2178 	if (mds < 0 ||
2179 	    ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2180 		dout("do_request no mds or not active, waiting for map\n");
2181 		list_add(&req->r_wait, &mdsc->waiting_for_map);
2182 		goto out;
2183 	}
2184 
2185 	/* get, open session */
2186 	session = __ceph_lookup_mds_session(mdsc, mds);
2187 	if (!session) {
2188 		session = register_session(mdsc, mds);
2189 		if (IS_ERR(session)) {
2190 			err = PTR_ERR(session);
2191 			goto finish;
2192 		}
2193 	}
2194 	req->r_session = get_session(session);
2195 
2196 	dout("do_request mds%d session %p state %s\n", mds, session,
2197 	     ceph_session_state_name(session->s_state));
2198 	if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2199 	    session->s_state != CEPH_MDS_SESSION_HUNG) {
2200 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
2201 		    session->s_state == CEPH_MDS_SESSION_CLOSING)
2202 			__open_session(mdsc, session);
2203 		list_add(&req->r_wait, &session->s_waiting);
2204 		goto out_session;
2205 	}
2206 
2207 	/* send request */
2208 	req->r_resend_mds = -1;   /* forget any previous mds hint */
2209 
2210 	if (req->r_request_started == 0)   /* note request start time */
2211 		req->r_request_started = jiffies;
2212 
2213 	err = __prepare_send_request(mdsc, req, mds, false);
2214 	if (!err) {
2215 		ceph_msg_get(req->r_request);
2216 		ceph_con_send(&session->s_con, req->r_request);
2217 	}
2218 
2219 out_session:
2220 	ceph_put_mds_session(session);
2221 finish:
2222 	if (err) {
2223 		dout("__do_request early error %d\n", err);
2224 		req->r_err = err;
2225 		complete_request(mdsc, req);
2226 		__unregister_request(mdsc, req);
2227 	}
2228 out:
2229 	return err;
2230 }
2231 
2232 /*
2233  * called under mdsc->mutex
2234  */
2235 static void __wake_requests(struct ceph_mds_client *mdsc,
2236 			    struct list_head *head)
2237 {
2238 	struct ceph_mds_request *req;
2239 	LIST_HEAD(tmp_list);
2240 
2241 	list_splice_init(head, &tmp_list);
2242 
2243 	while (!list_empty(&tmp_list)) {
2244 		req = list_entry(tmp_list.next,
2245 				 struct ceph_mds_request, r_wait);
2246 		list_del_init(&req->r_wait);
2247 		dout(" wake request %p tid %llu\n", req, req->r_tid);
2248 		__do_request(mdsc, req);
2249 	}
2250 }
2251 
2252 /*
2253  * Wake up threads with requests pending for @mds, so that they can
2254  * resubmit their requests to a possibly different mds.
2255  */
2256 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2257 {
2258 	struct ceph_mds_request *req;
2259 	struct rb_node *p = rb_first(&mdsc->request_tree);
2260 
2261 	dout("kick_requests mds%d\n", mds);
2262 	while (p) {
2263 		req = rb_entry(p, struct ceph_mds_request, r_node);
2264 		p = rb_next(p);
2265 		if (req->r_got_unsafe)
2266 			continue;
2267 		if (req->r_attempts > 0)
2268 			continue; /* only new requests */
2269 		if (req->r_session &&
2270 		    req->r_session->s_mds == mds) {
2271 			dout(" kicking tid %llu\n", req->r_tid);
2272 			list_del_init(&req->r_wait);
2273 			__do_request(mdsc, req);
2274 		}
2275 	}
2276 }
2277 
2278 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2279 			      struct ceph_mds_request *req)
2280 {
2281 	dout("submit_request on %p\n", req);
2282 	mutex_lock(&mdsc->mutex);
2283 	__register_request(mdsc, req, NULL);
2284 	__do_request(mdsc, req);
2285 	mutex_unlock(&mdsc->mutex);
2286 }
2287 
2288 /*
2289  * Synchrously perform an mds request.  Take care of all of the
2290  * session setup, forwarding, retry details.
2291  */
2292 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2293 			 struct inode *dir,
2294 			 struct ceph_mds_request *req)
2295 {
2296 	int err;
2297 
2298 	dout("do_request on %p\n", req);
2299 
2300 	/* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2301 	if (req->r_inode)
2302 		ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2303 	if (req->r_locked_dir)
2304 		ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2305 	if (req->r_old_dentry_dir)
2306 		ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2307 				  CEPH_CAP_PIN);
2308 
2309 	/* deny access to directories with pool_ns layouts */
2310 	if (req->r_inode && S_ISDIR(req->r_inode->i_mode) &&
2311 	    ceph_inode(req->r_inode)->i_pool_ns_len)
2312 		return -EIO;
2313 	if (req->r_locked_dir &&
2314 	    ceph_inode(req->r_locked_dir)->i_pool_ns_len)
2315 		return -EIO;
2316 
2317 	/* issue */
2318 	mutex_lock(&mdsc->mutex);
2319 	__register_request(mdsc, req, dir);
2320 	__do_request(mdsc, req);
2321 
2322 	if (req->r_err) {
2323 		err = req->r_err;
2324 		goto out;
2325 	}
2326 
2327 	/* wait */
2328 	mutex_unlock(&mdsc->mutex);
2329 	dout("do_request waiting\n");
2330 	if (!req->r_timeout && req->r_wait_for_completion) {
2331 		err = req->r_wait_for_completion(mdsc, req);
2332 	} else {
2333 		long timeleft = wait_for_completion_killable_timeout(
2334 					&req->r_completion,
2335 					ceph_timeout_jiffies(req->r_timeout));
2336 		if (timeleft > 0)
2337 			err = 0;
2338 		else if (!timeleft)
2339 			err = -EIO;  /* timed out */
2340 		else
2341 			err = timeleft;  /* killed */
2342 	}
2343 	dout("do_request waited, got %d\n", err);
2344 	mutex_lock(&mdsc->mutex);
2345 
2346 	/* only abort if we didn't race with a real reply */
2347 	if (req->r_got_result) {
2348 		err = le32_to_cpu(req->r_reply_info.head->result);
2349 	} else if (err < 0) {
2350 		dout("aborted request %lld with %d\n", req->r_tid, err);
2351 
2352 		/*
2353 		 * ensure we aren't running concurrently with
2354 		 * ceph_fill_trace or ceph_readdir_prepopulate, which
2355 		 * rely on locks (dir mutex) held by our caller.
2356 		 */
2357 		mutex_lock(&req->r_fill_mutex);
2358 		req->r_err = err;
2359 		req->r_aborted = true;
2360 		mutex_unlock(&req->r_fill_mutex);
2361 
2362 		if (req->r_locked_dir &&
2363 		    (req->r_op & CEPH_MDS_OP_WRITE))
2364 			ceph_invalidate_dir_request(req);
2365 	} else {
2366 		err = req->r_err;
2367 	}
2368 
2369 out:
2370 	mutex_unlock(&mdsc->mutex);
2371 	dout("do_request %p done, result %d\n", req, err);
2372 	return err;
2373 }
2374 
2375 /*
2376  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2377  * namespace request.
2378  */
2379 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2380 {
2381 	struct inode *inode = req->r_locked_dir;
2382 
2383 	dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2384 
2385 	ceph_dir_clear_complete(inode);
2386 	if (req->r_dentry)
2387 		ceph_invalidate_dentry_lease(req->r_dentry);
2388 	if (req->r_old_dentry)
2389 		ceph_invalidate_dentry_lease(req->r_old_dentry);
2390 }
2391 
2392 /*
2393  * Handle mds reply.
2394  *
2395  * We take the session mutex and parse and process the reply immediately.
2396  * This preserves the logical ordering of replies, capabilities, etc., sent
2397  * by the MDS as they are applied to our local cache.
2398  */
2399 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2400 {
2401 	struct ceph_mds_client *mdsc = session->s_mdsc;
2402 	struct ceph_mds_request *req;
2403 	struct ceph_mds_reply_head *head = msg->front.iov_base;
2404 	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2405 	struct ceph_snap_realm *realm;
2406 	u64 tid;
2407 	int err, result;
2408 	int mds = session->s_mds;
2409 
2410 	if (msg->front.iov_len < sizeof(*head)) {
2411 		pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2412 		ceph_msg_dump(msg);
2413 		return;
2414 	}
2415 
2416 	/* get request, session */
2417 	tid = le64_to_cpu(msg->hdr.tid);
2418 	mutex_lock(&mdsc->mutex);
2419 	req = __lookup_request(mdsc, tid);
2420 	if (!req) {
2421 		dout("handle_reply on unknown tid %llu\n", tid);
2422 		mutex_unlock(&mdsc->mutex);
2423 		return;
2424 	}
2425 	dout("handle_reply %p\n", req);
2426 
2427 	/* correct session? */
2428 	if (req->r_session != session) {
2429 		pr_err("mdsc_handle_reply got %llu on session mds%d"
2430 		       " not mds%d\n", tid, session->s_mds,
2431 		       req->r_session ? req->r_session->s_mds : -1);
2432 		mutex_unlock(&mdsc->mutex);
2433 		goto out;
2434 	}
2435 
2436 	/* dup? */
2437 	if ((req->r_got_unsafe && !head->safe) ||
2438 	    (req->r_got_safe && head->safe)) {
2439 		pr_warn("got a dup %s reply on %llu from mds%d\n",
2440 			   head->safe ? "safe" : "unsafe", tid, mds);
2441 		mutex_unlock(&mdsc->mutex);
2442 		goto out;
2443 	}
2444 	if (req->r_got_safe) {
2445 		pr_warn("got unsafe after safe on %llu from mds%d\n",
2446 			   tid, mds);
2447 		mutex_unlock(&mdsc->mutex);
2448 		goto out;
2449 	}
2450 
2451 	result = le32_to_cpu(head->result);
2452 
2453 	/*
2454 	 * Handle an ESTALE
2455 	 * if we're not talking to the authority, send to them
2456 	 * if the authority has changed while we weren't looking,
2457 	 * send to new authority
2458 	 * Otherwise we just have to return an ESTALE
2459 	 */
2460 	if (result == -ESTALE) {
2461 		dout("got ESTALE on request %llu", req->r_tid);
2462 		req->r_resend_mds = -1;
2463 		if (req->r_direct_mode != USE_AUTH_MDS) {
2464 			dout("not using auth, setting for that now");
2465 			req->r_direct_mode = USE_AUTH_MDS;
2466 			__do_request(mdsc, req);
2467 			mutex_unlock(&mdsc->mutex);
2468 			goto out;
2469 		} else  {
2470 			int mds = __choose_mds(mdsc, req);
2471 			if (mds >= 0 && mds != req->r_session->s_mds) {
2472 				dout("but auth changed, so resending");
2473 				__do_request(mdsc, req);
2474 				mutex_unlock(&mdsc->mutex);
2475 				goto out;
2476 			}
2477 		}
2478 		dout("have to return ESTALE on request %llu", req->r_tid);
2479 	}
2480 
2481 
2482 	if (head->safe) {
2483 		req->r_got_safe = true;
2484 		__unregister_request(mdsc, req);
2485 
2486 		if (req->r_got_unsafe) {
2487 			/*
2488 			 * We already handled the unsafe response, now do the
2489 			 * cleanup.  No need to examine the response; the MDS
2490 			 * doesn't include any result info in the safe
2491 			 * response.  And even if it did, there is nothing
2492 			 * useful we could do with a revised return value.
2493 			 */
2494 			dout("got safe reply %llu, mds%d\n", tid, mds);
2495 			list_del_init(&req->r_unsafe_item);
2496 
2497 			/* last unsafe request during umount? */
2498 			if (mdsc->stopping && !__get_oldest_req(mdsc))
2499 				complete_all(&mdsc->safe_umount_waiters);
2500 			mutex_unlock(&mdsc->mutex);
2501 			goto out;
2502 		}
2503 	} else {
2504 		req->r_got_unsafe = true;
2505 		list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2506 		if (req->r_unsafe_dir) {
2507 			struct ceph_inode_info *ci =
2508 					ceph_inode(req->r_unsafe_dir);
2509 			spin_lock(&ci->i_unsafe_lock);
2510 			list_add_tail(&req->r_unsafe_dir_item,
2511 				      &ci->i_unsafe_dirops);
2512 			spin_unlock(&ci->i_unsafe_lock);
2513 		}
2514 	}
2515 
2516 	dout("handle_reply tid %lld result %d\n", tid, result);
2517 	rinfo = &req->r_reply_info;
2518 	err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2519 	mutex_unlock(&mdsc->mutex);
2520 
2521 	mutex_lock(&session->s_mutex);
2522 	if (err < 0) {
2523 		pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2524 		ceph_msg_dump(msg);
2525 		goto out_err;
2526 	}
2527 
2528 	/* snap trace */
2529 	realm = NULL;
2530 	if (rinfo->snapblob_len) {
2531 		down_write(&mdsc->snap_rwsem);
2532 		ceph_update_snap_trace(mdsc, rinfo->snapblob,
2533 				rinfo->snapblob + rinfo->snapblob_len,
2534 				le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2535 				&realm);
2536 		downgrade_write(&mdsc->snap_rwsem);
2537 	} else {
2538 		down_read(&mdsc->snap_rwsem);
2539 	}
2540 
2541 	/* insert trace into our cache */
2542 	mutex_lock(&req->r_fill_mutex);
2543 	current->journal_info = req;
2544 	err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2545 	if (err == 0) {
2546 		if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2547 				    req->r_op == CEPH_MDS_OP_LSSNAP))
2548 			ceph_readdir_prepopulate(req, req->r_session);
2549 		ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2550 	}
2551 	current->journal_info = NULL;
2552 	mutex_unlock(&req->r_fill_mutex);
2553 
2554 	up_read(&mdsc->snap_rwsem);
2555 	if (realm)
2556 		ceph_put_snap_realm(mdsc, realm);
2557 
2558 	if (err == 0 && req->r_got_unsafe && req->r_target_inode) {
2559 		struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2560 		spin_lock(&ci->i_unsafe_lock);
2561 		list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2562 		spin_unlock(&ci->i_unsafe_lock);
2563 	}
2564 out_err:
2565 	mutex_lock(&mdsc->mutex);
2566 	if (!req->r_aborted) {
2567 		if (err) {
2568 			req->r_err = err;
2569 		} else {
2570 			req->r_reply =  ceph_msg_get(msg);
2571 			req->r_got_result = true;
2572 		}
2573 	} else {
2574 		dout("reply arrived after request %lld was aborted\n", tid);
2575 	}
2576 	mutex_unlock(&mdsc->mutex);
2577 
2578 	mutex_unlock(&session->s_mutex);
2579 
2580 	/* kick calling process */
2581 	complete_request(mdsc, req);
2582 out:
2583 	ceph_mdsc_put_request(req);
2584 	return;
2585 }
2586 
2587 
2588 
2589 /*
2590  * handle mds notification that our request has been forwarded.
2591  */
2592 static void handle_forward(struct ceph_mds_client *mdsc,
2593 			   struct ceph_mds_session *session,
2594 			   struct ceph_msg *msg)
2595 {
2596 	struct ceph_mds_request *req;
2597 	u64 tid = le64_to_cpu(msg->hdr.tid);
2598 	u32 next_mds;
2599 	u32 fwd_seq;
2600 	int err = -EINVAL;
2601 	void *p = msg->front.iov_base;
2602 	void *end = p + msg->front.iov_len;
2603 
2604 	ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2605 	next_mds = ceph_decode_32(&p);
2606 	fwd_seq = ceph_decode_32(&p);
2607 
2608 	mutex_lock(&mdsc->mutex);
2609 	req = __lookup_request(mdsc, tid);
2610 	if (!req) {
2611 		dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2612 		goto out;  /* dup reply? */
2613 	}
2614 
2615 	if (req->r_aborted) {
2616 		dout("forward tid %llu aborted, unregistering\n", tid);
2617 		__unregister_request(mdsc, req);
2618 	} else if (fwd_seq <= req->r_num_fwd) {
2619 		dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2620 		     tid, next_mds, req->r_num_fwd, fwd_seq);
2621 	} else {
2622 		/* resend. forward race not possible; mds would drop */
2623 		dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2624 		BUG_ON(req->r_err);
2625 		BUG_ON(req->r_got_result);
2626 		req->r_attempts = 0;
2627 		req->r_num_fwd = fwd_seq;
2628 		req->r_resend_mds = next_mds;
2629 		put_request_session(req);
2630 		__do_request(mdsc, req);
2631 	}
2632 	ceph_mdsc_put_request(req);
2633 out:
2634 	mutex_unlock(&mdsc->mutex);
2635 	return;
2636 
2637 bad:
2638 	pr_err("mdsc_handle_forward decode error err=%d\n", err);
2639 }
2640 
2641 /*
2642  * handle a mds session control message
2643  */
2644 static void handle_session(struct ceph_mds_session *session,
2645 			   struct ceph_msg *msg)
2646 {
2647 	struct ceph_mds_client *mdsc = session->s_mdsc;
2648 	u32 op;
2649 	u64 seq;
2650 	int mds = session->s_mds;
2651 	struct ceph_mds_session_head *h = msg->front.iov_base;
2652 	int wake = 0;
2653 
2654 	/* decode */
2655 	if (msg->front.iov_len != sizeof(*h))
2656 		goto bad;
2657 	op = le32_to_cpu(h->op);
2658 	seq = le64_to_cpu(h->seq);
2659 
2660 	mutex_lock(&mdsc->mutex);
2661 	if (op == CEPH_SESSION_CLOSE)
2662 		__unregister_session(mdsc, session);
2663 	/* FIXME: this ttl calculation is generous */
2664 	session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2665 	mutex_unlock(&mdsc->mutex);
2666 
2667 	mutex_lock(&session->s_mutex);
2668 
2669 	dout("handle_session mds%d %s %p state %s seq %llu\n",
2670 	     mds, ceph_session_op_name(op), session,
2671 	     ceph_session_state_name(session->s_state), seq);
2672 
2673 	if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2674 		session->s_state = CEPH_MDS_SESSION_OPEN;
2675 		pr_info("mds%d came back\n", session->s_mds);
2676 	}
2677 
2678 	switch (op) {
2679 	case CEPH_SESSION_OPEN:
2680 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2681 			pr_info("mds%d reconnect success\n", session->s_mds);
2682 		session->s_state = CEPH_MDS_SESSION_OPEN;
2683 		renewed_caps(mdsc, session, 0);
2684 		wake = 1;
2685 		if (mdsc->stopping)
2686 			__close_session(mdsc, session);
2687 		break;
2688 
2689 	case CEPH_SESSION_RENEWCAPS:
2690 		if (session->s_renew_seq == seq)
2691 			renewed_caps(mdsc, session, 1);
2692 		break;
2693 
2694 	case CEPH_SESSION_CLOSE:
2695 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2696 			pr_info("mds%d reconnect denied\n", session->s_mds);
2697 		cleanup_session_requests(mdsc, session);
2698 		remove_session_caps(session);
2699 		wake = 2; /* for good measure */
2700 		wake_up_all(&mdsc->session_close_wq);
2701 		break;
2702 
2703 	case CEPH_SESSION_STALE:
2704 		pr_info("mds%d caps went stale, renewing\n",
2705 			session->s_mds);
2706 		spin_lock(&session->s_gen_ttl_lock);
2707 		session->s_cap_gen++;
2708 		session->s_cap_ttl = jiffies - 1;
2709 		spin_unlock(&session->s_gen_ttl_lock);
2710 		send_renew_caps(mdsc, session);
2711 		break;
2712 
2713 	case CEPH_SESSION_RECALL_STATE:
2714 		trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2715 		break;
2716 
2717 	case CEPH_SESSION_FLUSHMSG:
2718 		send_flushmsg_ack(mdsc, session, seq);
2719 		break;
2720 
2721 	case CEPH_SESSION_FORCE_RO:
2722 		dout("force_session_readonly %p\n", session);
2723 		spin_lock(&session->s_cap_lock);
2724 		session->s_readonly = true;
2725 		spin_unlock(&session->s_cap_lock);
2726 		wake_up_session_caps(session, 0);
2727 		break;
2728 
2729 	default:
2730 		pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2731 		WARN_ON(1);
2732 	}
2733 
2734 	mutex_unlock(&session->s_mutex);
2735 	if (wake) {
2736 		mutex_lock(&mdsc->mutex);
2737 		__wake_requests(mdsc, &session->s_waiting);
2738 		if (wake == 2)
2739 			kick_requests(mdsc, mds);
2740 		mutex_unlock(&mdsc->mutex);
2741 	}
2742 	return;
2743 
2744 bad:
2745 	pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2746 	       (int)msg->front.iov_len);
2747 	ceph_msg_dump(msg);
2748 	return;
2749 }
2750 
2751 
2752 /*
2753  * called under session->mutex.
2754  */
2755 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2756 				   struct ceph_mds_session *session)
2757 {
2758 	struct ceph_mds_request *req, *nreq;
2759 	struct rb_node *p;
2760 	int err;
2761 
2762 	dout("replay_unsafe_requests mds%d\n", session->s_mds);
2763 
2764 	mutex_lock(&mdsc->mutex);
2765 	list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2766 		err = __prepare_send_request(mdsc, req, session->s_mds, true);
2767 		if (!err) {
2768 			ceph_msg_get(req->r_request);
2769 			ceph_con_send(&session->s_con, req->r_request);
2770 		}
2771 	}
2772 
2773 	/*
2774 	 * also re-send old requests when MDS enters reconnect stage. So that MDS
2775 	 * can process completed request in clientreplay stage.
2776 	 */
2777 	p = rb_first(&mdsc->request_tree);
2778 	while (p) {
2779 		req = rb_entry(p, struct ceph_mds_request, r_node);
2780 		p = rb_next(p);
2781 		if (req->r_got_unsafe)
2782 			continue;
2783 		if (req->r_attempts == 0)
2784 			continue; /* only old requests */
2785 		if (req->r_session &&
2786 		    req->r_session->s_mds == session->s_mds) {
2787 			err = __prepare_send_request(mdsc, req,
2788 						     session->s_mds, true);
2789 			if (!err) {
2790 				ceph_msg_get(req->r_request);
2791 				ceph_con_send(&session->s_con, req->r_request);
2792 			}
2793 		}
2794 	}
2795 	mutex_unlock(&mdsc->mutex);
2796 }
2797 
2798 /*
2799  * Encode information about a cap for a reconnect with the MDS.
2800  */
2801 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2802 			  void *arg)
2803 {
2804 	union {
2805 		struct ceph_mds_cap_reconnect v2;
2806 		struct ceph_mds_cap_reconnect_v1 v1;
2807 	} rec;
2808 	size_t reclen;
2809 	struct ceph_inode_info *ci;
2810 	struct ceph_reconnect_state *recon_state = arg;
2811 	struct ceph_pagelist *pagelist = recon_state->pagelist;
2812 	char *path;
2813 	int pathlen, err;
2814 	u64 pathbase;
2815 	struct dentry *dentry;
2816 
2817 	ci = cap->ci;
2818 
2819 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2820 	     inode, ceph_vinop(inode), cap, cap->cap_id,
2821 	     ceph_cap_string(cap->issued));
2822 	err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2823 	if (err)
2824 		return err;
2825 
2826 	dentry = d_find_alias(inode);
2827 	if (dentry) {
2828 		path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2829 		if (IS_ERR(path)) {
2830 			err = PTR_ERR(path);
2831 			goto out_dput;
2832 		}
2833 	} else {
2834 		path = NULL;
2835 		pathlen = 0;
2836 	}
2837 	err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2838 	if (err)
2839 		goto out_free;
2840 
2841 	spin_lock(&ci->i_ceph_lock);
2842 	cap->seq = 0;        /* reset cap seq */
2843 	cap->issue_seq = 0;  /* and issue_seq */
2844 	cap->mseq = 0;       /* and migrate_seq */
2845 	cap->cap_gen = cap->session->s_cap_gen;
2846 
2847 	if (recon_state->flock) {
2848 		rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2849 		rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2850 		rec.v2.issued = cpu_to_le32(cap->issued);
2851 		rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2852 		rec.v2.pathbase = cpu_to_le64(pathbase);
2853 		rec.v2.flock_len = 0;
2854 		reclen = sizeof(rec.v2);
2855 	} else {
2856 		rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2857 		rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2858 		rec.v1.issued = cpu_to_le32(cap->issued);
2859 		rec.v1.size = cpu_to_le64(inode->i_size);
2860 		ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2861 		ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2862 		rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2863 		rec.v1.pathbase = cpu_to_le64(pathbase);
2864 		reclen = sizeof(rec.v1);
2865 	}
2866 	spin_unlock(&ci->i_ceph_lock);
2867 
2868 	if (recon_state->flock) {
2869 		int num_fcntl_locks, num_flock_locks;
2870 		struct ceph_filelock *flocks;
2871 
2872 encode_again:
2873 		ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2874 		flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2875 				 sizeof(struct ceph_filelock), GFP_NOFS);
2876 		if (!flocks) {
2877 			err = -ENOMEM;
2878 			goto out_free;
2879 		}
2880 		err = ceph_encode_locks_to_buffer(inode, flocks,
2881 						  num_fcntl_locks,
2882 						  num_flock_locks);
2883 		if (err) {
2884 			kfree(flocks);
2885 			if (err == -ENOSPC)
2886 				goto encode_again;
2887 			goto out_free;
2888 		}
2889 		/*
2890 		 * number of encoded locks is stable, so copy to pagelist
2891 		 */
2892 		rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2893 				    (num_fcntl_locks+num_flock_locks) *
2894 				    sizeof(struct ceph_filelock));
2895 		err = ceph_pagelist_append(pagelist, &rec, reclen);
2896 		if (!err)
2897 			err = ceph_locks_to_pagelist(flocks, pagelist,
2898 						     num_fcntl_locks,
2899 						     num_flock_locks);
2900 		kfree(flocks);
2901 	} else {
2902 		err = ceph_pagelist_append(pagelist, &rec, reclen);
2903 	}
2904 
2905 	recon_state->nr_caps++;
2906 out_free:
2907 	kfree(path);
2908 out_dput:
2909 	dput(dentry);
2910 	return err;
2911 }
2912 
2913 
2914 /*
2915  * If an MDS fails and recovers, clients need to reconnect in order to
2916  * reestablish shared state.  This includes all caps issued through
2917  * this session _and_ the snap_realm hierarchy.  Because it's not
2918  * clear which snap realms the mds cares about, we send everything we
2919  * know about.. that ensures we'll then get any new info the
2920  * recovering MDS might have.
2921  *
2922  * This is a relatively heavyweight operation, but it's rare.
2923  *
2924  * called with mdsc->mutex held.
2925  */
2926 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2927 			       struct ceph_mds_session *session)
2928 {
2929 	struct ceph_msg *reply;
2930 	struct rb_node *p;
2931 	int mds = session->s_mds;
2932 	int err = -ENOMEM;
2933 	int s_nr_caps;
2934 	struct ceph_pagelist *pagelist;
2935 	struct ceph_reconnect_state recon_state;
2936 
2937 	pr_info("mds%d reconnect start\n", mds);
2938 
2939 	pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2940 	if (!pagelist)
2941 		goto fail_nopagelist;
2942 	ceph_pagelist_init(pagelist);
2943 
2944 	reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2945 	if (!reply)
2946 		goto fail_nomsg;
2947 
2948 	mutex_lock(&session->s_mutex);
2949 	session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2950 	session->s_seq = 0;
2951 
2952 	dout("session %p state %s\n", session,
2953 	     ceph_session_state_name(session->s_state));
2954 
2955 	spin_lock(&session->s_gen_ttl_lock);
2956 	session->s_cap_gen++;
2957 	spin_unlock(&session->s_gen_ttl_lock);
2958 
2959 	spin_lock(&session->s_cap_lock);
2960 	/* don't know if session is readonly */
2961 	session->s_readonly = 0;
2962 	/*
2963 	 * notify __ceph_remove_cap() that we are composing cap reconnect.
2964 	 * If a cap get released before being added to the cap reconnect,
2965 	 * __ceph_remove_cap() should skip queuing cap release.
2966 	 */
2967 	session->s_cap_reconnect = 1;
2968 	/* drop old cap expires; we're about to reestablish that state */
2969 	cleanup_cap_releases(mdsc, session);
2970 
2971 	/* trim unused caps to reduce MDS's cache rejoin time */
2972 	if (mdsc->fsc->sb->s_root)
2973 		shrink_dcache_parent(mdsc->fsc->sb->s_root);
2974 
2975 	ceph_con_close(&session->s_con);
2976 	ceph_con_open(&session->s_con,
2977 		      CEPH_ENTITY_TYPE_MDS, mds,
2978 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2979 
2980 	/* replay unsafe requests */
2981 	replay_unsafe_requests(mdsc, session);
2982 
2983 	down_read(&mdsc->snap_rwsem);
2984 
2985 	/* traverse this session's caps */
2986 	s_nr_caps = session->s_nr_caps;
2987 	err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2988 	if (err)
2989 		goto fail;
2990 
2991 	recon_state.nr_caps = 0;
2992 	recon_state.pagelist = pagelist;
2993 	recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2994 	err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2995 	if (err < 0)
2996 		goto fail;
2997 
2998 	spin_lock(&session->s_cap_lock);
2999 	session->s_cap_reconnect = 0;
3000 	spin_unlock(&session->s_cap_lock);
3001 
3002 	/*
3003 	 * snaprealms.  we provide mds with the ino, seq (version), and
3004 	 * parent for all of our realms.  If the mds has any newer info,
3005 	 * it will tell us.
3006 	 */
3007 	for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3008 		struct ceph_snap_realm *realm =
3009 			rb_entry(p, struct ceph_snap_realm, node);
3010 		struct ceph_mds_snaprealm_reconnect sr_rec;
3011 
3012 		dout(" adding snap realm %llx seq %lld parent %llx\n",
3013 		     realm->ino, realm->seq, realm->parent_ino);
3014 		sr_rec.ino = cpu_to_le64(realm->ino);
3015 		sr_rec.seq = cpu_to_le64(realm->seq);
3016 		sr_rec.parent = cpu_to_le64(realm->parent_ino);
3017 		err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3018 		if (err)
3019 			goto fail;
3020 	}
3021 
3022 	if (recon_state.flock)
3023 		reply->hdr.version = cpu_to_le16(2);
3024 
3025 	/* raced with cap release? */
3026 	if (s_nr_caps != recon_state.nr_caps) {
3027 		struct page *page = list_first_entry(&pagelist->head,
3028 						     struct page, lru);
3029 		__le32 *addr = kmap_atomic(page);
3030 		*addr = cpu_to_le32(recon_state.nr_caps);
3031 		kunmap_atomic(addr);
3032 	}
3033 
3034 	reply->hdr.data_len = cpu_to_le32(pagelist->length);
3035 	ceph_msg_data_add_pagelist(reply, pagelist);
3036 
3037 	ceph_early_kick_flushing_caps(mdsc, session);
3038 
3039 	ceph_con_send(&session->s_con, reply);
3040 
3041 	mutex_unlock(&session->s_mutex);
3042 
3043 	mutex_lock(&mdsc->mutex);
3044 	__wake_requests(mdsc, &session->s_waiting);
3045 	mutex_unlock(&mdsc->mutex);
3046 
3047 	up_read(&mdsc->snap_rwsem);
3048 	return;
3049 
3050 fail:
3051 	ceph_msg_put(reply);
3052 	up_read(&mdsc->snap_rwsem);
3053 	mutex_unlock(&session->s_mutex);
3054 fail_nomsg:
3055 	ceph_pagelist_release(pagelist);
3056 fail_nopagelist:
3057 	pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3058 	return;
3059 }
3060 
3061 
3062 /*
3063  * compare old and new mdsmaps, kicking requests
3064  * and closing out old connections as necessary
3065  *
3066  * called under mdsc->mutex.
3067  */
3068 static void check_new_map(struct ceph_mds_client *mdsc,
3069 			  struct ceph_mdsmap *newmap,
3070 			  struct ceph_mdsmap *oldmap)
3071 {
3072 	int i;
3073 	int oldstate, newstate;
3074 	struct ceph_mds_session *s;
3075 
3076 	dout("check_new_map new %u old %u\n",
3077 	     newmap->m_epoch, oldmap->m_epoch);
3078 
3079 	for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
3080 		if (mdsc->sessions[i] == NULL)
3081 			continue;
3082 		s = mdsc->sessions[i];
3083 		oldstate = ceph_mdsmap_get_state(oldmap, i);
3084 		newstate = ceph_mdsmap_get_state(newmap, i);
3085 
3086 		dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3087 		     i, ceph_mds_state_name(oldstate),
3088 		     ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3089 		     ceph_mds_state_name(newstate),
3090 		     ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3091 		     ceph_session_state_name(s->s_state));
3092 
3093 		if (i >= newmap->m_max_mds ||
3094 		    memcmp(ceph_mdsmap_get_addr(oldmap, i),
3095 			   ceph_mdsmap_get_addr(newmap, i),
3096 			   sizeof(struct ceph_entity_addr))) {
3097 			if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3098 				/* the session never opened, just close it
3099 				 * out now */
3100 				__wake_requests(mdsc, &s->s_waiting);
3101 				__unregister_session(mdsc, s);
3102 			} else {
3103 				/* just close it */
3104 				mutex_unlock(&mdsc->mutex);
3105 				mutex_lock(&s->s_mutex);
3106 				mutex_lock(&mdsc->mutex);
3107 				ceph_con_close(&s->s_con);
3108 				mutex_unlock(&s->s_mutex);
3109 				s->s_state = CEPH_MDS_SESSION_RESTARTING;
3110 			}
3111 		} else if (oldstate == newstate) {
3112 			continue;  /* nothing new with this mds */
3113 		}
3114 
3115 		/*
3116 		 * send reconnect?
3117 		 */
3118 		if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3119 		    newstate >= CEPH_MDS_STATE_RECONNECT) {
3120 			mutex_unlock(&mdsc->mutex);
3121 			send_mds_reconnect(mdsc, s);
3122 			mutex_lock(&mdsc->mutex);
3123 		}
3124 
3125 		/*
3126 		 * kick request on any mds that has gone active.
3127 		 */
3128 		if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3129 		    newstate >= CEPH_MDS_STATE_ACTIVE) {
3130 			if (oldstate != CEPH_MDS_STATE_CREATING &&
3131 			    oldstate != CEPH_MDS_STATE_STARTING)
3132 				pr_info("mds%d recovery completed\n", s->s_mds);
3133 			kick_requests(mdsc, i);
3134 			ceph_kick_flushing_caps(mdsc, s);
3135 			wake_up_session_caps(s, 1);
3136 		}
3137 	}
3138 
3139 	for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3140 		s = mdsc->sessions[i];
3141 		if (!s)
3142 			continue;
3143 		if (!ceph_mdsmap_is_laggy(newmap, i))
3144 			continue;
3145 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3146 		    s->s_state == CEPH_MDS_SESSION_HUNG ||
3147 		    s->s_state == CEPH_MDS_SESSION_CLOSING) {
3148 			dout(" connecting to export targets of laggy mds%d\n",
3149 			     i);
3150 			__open_export_target_sessions(mdsc, s);
3151 		}
3152 	}
3153 }
3154 
3155 
3156 
3157 /*
3158  * leases
3159  */
3160 
3161 /*
3162  * caller must hold session s_mutex, dentry->d_lock
3163  */
3164 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3165 {
3166 	struct ceph_dentry_info *di = ceph_dentry(dentry);
3167 
3168 	ceph_put_mds_session(di->lease_session);
3169 	di->lease_session = NULL;
3170 }
3171 
3172 static void handle_lease(struct ceph_mds_client *mdsc,
3173 			 struct ceph_mds_session *session,
3174 			 struct ceph_msg *msg)
3175 {
3176 	struct super_block *sb = mdsc->fsc->sb;
3177 	struct inode *inode;
3178 	struct dentry *parent, *dentry;
3179 	struct ceph_dentry_info *di;
3180 	int mds = session->s_mds;
3181 	struct ceph_mds_lease *h = msg->front.iov_base;
3182 	u32 seq;
3183 	struct ceph_vino vino;
3184 	struct qstr dname;
3185 	int release = 0;
3186 
3187 	dout("handle_lease from mds%d\n", mds);
3188 
3189 	/* decode */
3190 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3191 		goto bad;
3192 	vino.ino = le64_to_cpu(h->ino);
3193 	vino.snap = CEPH_NOSNAP;
3194 	seq = le32_to_cpu(h->seq);
3195 	dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3196 	dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3197 	if (dname.len != get_unaligned_le32(h+1))
3198 		goto bad;
3199 
3200 	/* lookup inode */
3201 	inode = ceph_find_inode(sb, vino);
3202 	dout("handle_lease %s, ino %llx %p %.*s\n",
3203 	     ceph_lease_op_name(h->action), vino.ino, inode,
3204 	     dname.len, dname.name);
3205 
3206 	mutex_lock(&session->s_mutex);
3207 	session->s_seq++;
3208 
3209 	if (inode == NULL) {
3210 		dout("handle_lease no inode %llx\n", vino.ino);
3211 		goto release;
3212 	}
3213 
3214 	/* dentry */
3215 	parent = d_find_alias(inode);
3216 	if (!parent) {
3217 		dout("no parent dentry on inode %p\n", inode);
3218 		WARN_ON(1);
3219 		goto release;  /* hrm... */
3220 	}
3221 	dname.hash = full_name_hash(dname.name, dname.len);
3222 	dentry = d_lookup(parent, &dname);
3223 	dput(parent);
3224 	if (!dentry)
3225 		goto release;
3226 
3227 	spin_lock(&dentry->d_lock);
3228 	di = ceph_dentry(dentry);
3229 	switch (h->action) {
3230 	case CEPH_MDS_LEASE_REVOKE:
3231 		if (di->lease_session == session) {
3232 			if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3233 				h->seq = cpu_to_le32(di->lease_seq);
3234 			__ceph_mdsc_drop_dentry_lease(dentry);
3235 		}
3236 		release = 1;
3237 		break;
3238 
3239 	case CEPH_MDS_LEASE_RENEW:
3240 		if (di->lease_session == session &&
3241 		    di->lease_gen == session->s_cap_gen &&
3242 		    di->lease_renew_from &&
3243 		    di->lease_renew_after == 0) {
3244 			unsigned long duration =
3245 				msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3246 
3247 			di->lease_seq = seq;
3248 			dentry->d_time = di->lease_renew_from + duration;
3249 			di->lease_renew_after = di->lease_renew_from +
3250 				(duration >> 1);
3251 			di->lease_renew_from = 0;
3252 		}
3253 		break;
3254 	}
3255 	spin_unlock(&dentry->d_lock);
3256 	dput(dentry);
3257 
3258 	if (!release)
3259 		goto out;
3260 
3261 release:
3262 	/* let's just reuse the same message */
3263 	h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3264 	ceph_msg_get(msg);
3265 	ceph_con_send(&session->s_con, msg);
3266 
3267 out:
3268 	iput(inode);
3269 	mutex_unlock(&session->s_mutex);
3270 	return;
3271 
3272 bad:
3273 	pr_err("corrupt lease message\n");
3274 	ceph_msg_dump(msg);
3275 }
3276 
3277 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3278 			      struct inode *inode,
3279 			      struct dentry *dentry, char action,
3280 			      u32 seq)
3281 {
3282 	struct ceph_msg *msg;
3283 	struct ceph_mds_lease *lease;
3284 	int len = sizeof(*lease) + sizeof(u32);
3285 	int dnamelen = 0;
3286 
3287 	dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3288 	     inode, dentry, ceph_lease_op_name(action), session->s_mds);
3289 	dnamelen = dentry->d_name.len;
3290 	len += dnamelen;
3291 
3292 	msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3293 	if (!msg)
3294 		return;
3295 	lease = msg->front.iov_base;
3296 	lease->action = action;
3297 	lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3298 	lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3299 	lease->seq = cpu_to_le32(seq);
3300 	put_unaligned_le32(dnamelen, lease + 1);
3301 	memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3302 
3303 	/*
3304 	 * if this is a preemptive lease RELEASE, no need to
3305 	 * flush request stream, since the actual request will
3306 	 * soon follow.
3307 	 */
3308 	msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3309 
3310 	ceph_con_send(&session->s_con, msg);
3311 }
3312 
3313 /*
3314  * Preemptively release a lease we expect to invalidate anyway.
3315  * Pass @inode always, @dentry is optional.
3316  */
3317 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3318 			     struct dentry *dentry)
3319 {
3320 	struct ceph_dentry_info *di;
3321 	struct ceph_mds_session *session;
3322 	u32 seq;
3323 
3324 	BUG_ON(inode == NULL);
3325 	BUG_ON(dentry == NULL);
3326 
3327 	/* is dentry lease valid? */
3328 	spin_lock(&dentry->d_lock);
3329 	di = ceph_dentry(dentry);
3330 	if (!di || !di->lease_session ||
3331 	    di->lease_session->s_mds < 0 ||
3332 	    di->lease_gen != di->lease_session->s_cap_gen ||
3333 	    !time_before(jiffies, dentry->d_time)) {
3334 		dout("lease_release inode %p dentry %p -- "
3335 		     "no lease\n",
3336 		     inode, dentry);
3337 		spin_unlock(&dentry->d_lock);
3338 		return;
3339 	}
3340 
3341 	/* we do have a lease on this dentry; note mds and seq */
3342 	session = ceph_get_mds_session(di->lease_session);
3343 	seq = di->lease_seq;
3344 	__ceph_mdsc_drop_dentry_lease(dentry);
3345 	spin_unlock(&dentry->d_lock);
3346 
3347 	dout("lease_release inode %p dentry %p to mds%d\n",
3348 	     inode, dentry, session->s_mds);
3349 	ceph_mdsc_lease_send_msg(session, inode, dentry,
3350 				 CEPH_MDS_LEASE_RELEASE, seq);
3351 	ceph_put_mds_session(session);
3352 }
3353 
3354 /*
3355  * drop all leases (and dentry refs) in preparation for umount
3356  */
3357 static void drop_leases(struct ceph_mds_client *mdsc)
3358 {
3359 	int i;
3360 
3361 	dout("drop_leases\n");
3362 	mutex_lock(&mdsc->mutex);
3363 	for (i = 0; i < mdsc->max_sessions; i++) {
3364 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3365 		if (!s)
3366 			continue;
3367 		mutex_unlock(&mdsc->mutex);
3368 		mutex_lock(&s->s_mutex);
3369 		mutex_unlock(&s->s_mutex);
3370 		ceph_put_mds_session(s);
3371 		mutex_lock(&mdsc->mutex);
3372 	}
3373 	mutex_unlock(&mdsc->mutex);
3374 }
3375 
3376 
3377 
3378 /*
3379  * delayed work -- periodically trim expired leases, renew caps with mds
3380  */
3381 static void schedule_delayed(struct ceph_mds_client *mdsc)
3382 {
3383 	int delay = 5;
3384 	unsigned hz = round_jiffies_relative(HZ * delay);
3385 	schedule_delayed_work(&mdsc->delayed_work, hz);
3386 }
3387 
3388 static void delayed_work(struct work_struct *work)
3389 {
3390 	int i;
3391 	struct ceph_mds_client *mdsc =
3392 		container_of(work, struct ceph_mds_client, delayed_work.work);
3393 	int renew_interval;
3394 	int renew_caps;
3395 
3396 	dout("mdsc delayed_work\n");
3397 	ceph_check_delayed_caps(mdsc);
3398 
3399 	mutex_lock(&mdsc->mutex);
3400 	renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3401 	renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3402 				   mdsc->last_renew_caps);
3403 	if (renew_caps)
3404 		mdsc->last_renew_caps = jiffies;
3405 
3406 	for (i = 0; i < mdsc->max_sessions; i++) {
3407 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3408 		if (s == NULL)
3409 			continue;
3410 		if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3411 			dout("resending session close request for mds%d\n",
3412 			     s->s_mds);
3413 			request_close_session(mdsc, s);
3414 			ceph_put_mds_session(s);
3415 			continue;
3416 		}
3417 		if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3418 			if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3419 				s->s_state = CEPH_MDS_SESSION_HUNG;
3420 				pr_info("mds%d hung\n", s->s_mds);
3421 			}
3422 		}
3423 		if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3424 			/* this mds is failed or recovering, just wait */
3425 			ceph_put_mds_session(s);
3426 			continue;
3427 		}
3428 		mutex_unlock(&mdsc->mutex);
3429 
3430 		mutex_lock(&s->s_mutex);
3431 		if (renew_caps)
3432 			send_renew_caps(mdsc, s);
3433 		else
3434 			ceph_con_keepalive(&s->s_con);
3435 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3436 		    s->s_state == CEPH_MDS_SESSION_HUNG)
3437 			ceph_send_cap_releases(mdsc, s);
3438 		mutex_unlock(&s->s_mutex);
3439 		ceph_put_mds_session(s);
3440 
3441 		mutex_lock(&mdsc->mutex);
3442 	}
3443 	mutex_unlock(&mdsc->mutex);
3444 
3445 	schedule_delayed(mdsc);
3446 }
3447 
3448 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3449 
3450 {
3451 	struct ceph_mds_client *mdsc;
3452 
3453 	mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3454 	if (!mdsc)
3455 		return -ENOMEM;
3456 	mdsc->fsc = fsc;
3457 	fsc->mdsc = mdsc;
3458 	mutex_init(&mdsc->mutex);
3459 	mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3460 	if (mdsc->mdsmap == NULL) {
3461 		kfree(mdsc);
3462 		return -ENOMEM;
3463 	}
3464 
3465 	init_completion(&mdsc->safe_umount_waiters);
3466 	init_waitqueue_head(&mdsc->session_close_wq);
3467 	INIT_LIST_HEAD(&mdsc->waiting_for_map);
3468 	mdsc->sessions = NULL;
3469 	atomic_set(&mdsc->num_sessions, 0);
3470 	mdsc->max_sessions = 0;
3471 	mdsc->stopping = 0;
3472 	mdsc->last_snap_seq = 0;
3473 	init_rwsem(&mdsc->snap_rwsem);
3474 	mdsc->snap_realms = RB_ROOT;
3475 	INIT_LIST_HEAD(&mdsc->snap_empty);
3476 	spin_lock_init(&mdsc->snap_empty_lock);
3477 	mdsc->last_tid = 0;
3478 	mdsc->oldest_tid = 0;
3479 	mdsc->request_tree = RB_ROOT;
3480 	INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3481 	mdsc->last_renew_caps = jiffies;
3482 	INIT_LIST_HEAD(&mdsc->cap_delay_list);
3483 	spin_lock_init(&mdsc->cap_delay_lock);
3484 	INIT_LIST_HEAD(&mdsc->snap_flush_list);
3485 	spin_lock_init(&mdsc->snap_flush_lock);
3486 	mdsc->last_cap_flush_tid = 1;
3487 	mdsc->cap_flush_tree = RB_ROOT;
3488 	INIT_LIST_HEAD(&mdsc->cap_dirty);
3489 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3490 	mdsc->num_cap_flushing = 0;
3491 	spin_lock_init(&mdsc->cap_dirty_lock);
3492 	init_waitqueue_head(&mdsc->cap_flushing_wq);
3493 	spin_lock_init(&mdsc->dentry_lru_lock);
3494 	INIT_LIST_HEAD(&mdsc->dentry_lru);
3495 
3496 	ceph_caps_init(mdsc);
3497 	ceph_adjust_min_caps(mdsc, fsc->min_caps);
3498 
3499 	init_rwsem(&mdsc->pool_perm_rwsem);
3500 	mdsc->pool_perm_tree = RB_ROOT;
3501 
3502 	return 0;
3503 }
3504 
3505 /*
3506  * Wait for safe replies on open mds requests.  If we time out, drop
3507  * all requests from the tree to avoid dangling dentry refs.
3508  */
3509 static void wait_requests(struct ceph_mds_client *mdsc)
3510 {
3511 	struct ceph_options *opts = mdsc->fsc->client->options;
3512 	struct ceph_mds_request *req;
3513 
3514 	mutex_lock(&mdsc->mutex);
3515 	if (__get_oldest_req(mdsc)) {
3516 		mutex_unlock(&mdsc->mutex);
3517 
3518 		dout("wait_requests waiting for requests\n");
3519 		wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3520 				    ceph_timeout_jiffies(opts->mount_timeout));
3521 
3522 		/* tear down remaining requests */
3523 		mutex_lock(&mdsc->mutex);
3524 		while ((req = __get_oldest_req(mdsc))) {
3525 			dout("wait_requests timed out on tid %llu\n",
3526 			     req->r_tid);
3527 			__unregister_request(mdsc, req);
3528 		}
3529 	}
3530 	mutex_unlock(&mdsc->mutex);
3531 	dout("wait_requests done\n");
3532 }
3533 
3534 /*
3535  * called before mount is ro, and before dentries are torn down.
3536  * (hmm, does this still race with new lookups?)
3537  */
3538 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3539 {
3540 	dout("pre_umount\n");
3541 	mdsc->stopping = 1;
3542 
3543 	drop_leases(mdsc);
3544 	ceph_flush_dirty_caps(mdsc);
3545 	wait_requests(mdsc);
3546 
3547 	/*
3548 	 * wait for reply handlers to drop their request refs and
3549 	 * their inode/dcache refs
3550 	 */
3551 	ceph_msgr_flush();
3552 }
3553 
3554 /*
3555  * wait for all write mds requests to flush.
3556  */
3557 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3558 {
3559 	struct ceph_mds_request *req = NULL, *nextreq;
3560 	struct rb_node *n;
3561 
3562 	mutex_lock(&mdsc->mutex);
3563 	dout("wait_unsafe_requests want %lld\n", want_tid);
3564 restart:
3565 	req = __get_oldest_req(mdsc);
3566 	while (req && req->r_tid <= want_tid) {
3567 		/* find next request */
3568 		n = rb_next(&req->r_node);
3569 		if (n)
3570 			nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3571 		else
3572 			nextreq = NULL;
3573 		if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3574 		    (req->r_op & CEPH_MDS_OP_WRITE)) {
3575 			/* write op */
3576 			ceph_mdsc_get_request(req);
3577 			if (nextreq)
3578 				ceph_mdsc_get_request(nextreq);
3579 			mutex_unlock(&mdsc->mutex);
3580 			dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3581 			     req->r_tid, want_tid);
3582 			wait_for_completion(&req->r_safe_completion);
3583 			mutex_lock(&mdsc->mutex);
3584 			ceph_mdsc_put_request(req);
3585 			if (!nextreq)
3586 				break;  /* next dne before, so we're done! */
3587 			if (RB_EMPTY_NODE(&nextreq->r_node)) {
3588 				/* next request was removed from tree */
3589 				ceph_mdsc_put_request(nextreq);
3590 				goto restart;
3591 			}
3592 			ceph_mdsc_put_request(nextreq);  /* won't go away */
3593 		}
3594 		req = nextreq;
3595 	}
3596 	mutex_unlock(&mdsc->mutex);
3597 	dout("wait_unsafe_requests done\n");
3598 }
3599 
3600 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3601 {
3602 	u64 want_tid, want_flush, want_snap;
3603 
3604 	if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3605 		return;
3606 
3607 	dout("sync\n");
3608 	mutex_lock(&mdsc->mutex);
3609 	want_tid = mdsc->last_tid;
3610 	mutex_unlock(&mdsc->mutex);
3611 
3612 	ceph_flush_dirty_caps(mdsc);
3613 	spin_lock(&mdsc->cap_dirty_lock);
3614 	want_flush = mdsc->last_cap_flush_tid;
3615 	spin_unlock(&mdsc->cap_dirty_lock);
3616 
3617 	down_read(&mdsc->snap_rwsem);
3618 	want_snap = mdsc->last_snap_seq;
3619 	up_read(&mdsc->snap_rwsem);
3620 
3621 	dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3622 	     want_tid, want_flush, want_snap);
3623 
3624 	wait_unsafe_requests(mdsc, want_tid);
3625 	wait_caps_flush(mdsc, want_flush, want_snap);
3626 }
3627 
3628 /*
3629  * true if all sessions are closed, or we force unmount
3630  */
3631 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3632 {
3633 	if (ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3634 		return true;
3635 	return atomic_read(&mdsc->num_sessions) == 0;
3636 }
3637 
3638 /*
3639  * called after sb is ro.
3640  */
3641 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3642 {
3643 	struct ceph_options *opts = mdsc->fsc->client->options;
3644 	struct ceph_mds_session *session;
3645 	int i;
3646 
3647 	dout("close_sessions\n");
3648 
3649 	/* close sessions */
3650 	mutex_lock(&mdsc->mutex);
3651 	for (i = 0; i < mdsc->max_sessions; i++) {
3652 		session = __ceph_lookup_mds_session(mdsc, i);
3653 		if (!session)
3654 			continue;
3655 		mutex_unlock(&mdsc->mutex);
3656 		mutex_lock(&session->s_mutex);
3657 		__close_session(mdsc, session);
3658 		mutex_unlock(&session->s_mutex);
3659 		ceph_put_mds_session(session);
3660 		mutex_lock(&mdsc->mutex);
3661 	}
3662 	mutex_unlock(&mdsc->mutex);
3663 
3664 	dout("waiting for sessions to close\n");
3665 	wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3666 			   ceph_timeout_jiffies(opts->mount_timeout));
3667 
3668 	/* tear down remaining sessions */
3669 	mutex_lock(&mdsc->mutex);
3670 	for (i = 0; i < mdsc->max_sessions; i++) {
3671 		if (mdsc->sessions[i]) {
3672 			session = get_session(mdsc->sessions[i]);
3673 			__unregister_session(mdsc, session);
3674 			mutex_unlock(&mdsc->mutex);
3675 			mutex_lock(&session->s_mutex);
3676 			remove_session_caps(session);
3677 			mutex_unlock(&session->s_mutex);
3678 			ceph_put_mds_session(session);
3679 			mutex_lock(&mdsc->mutex);
3680 		}
3681 	}
3682 	WARN_ON(!list_empty(&mdsc->cap_delay_list));
3683 	mutex_unlock(&mdsc->mutex);
3684 
3685 	ceph_cleanup_empty_realms(mdsc);
3686 
3687 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3688 
3689 	dout("stopped\n");
3690 }
3691 
3692 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3693 {
3694 	struct ceph_mds_session *session;
3695 	int mds;
3696 
3697 	dout("force umount\n");
3698 
3699 	mutex_lock(&mdsc->mutex);
3700 	for (mds = 0; mds < mdsc->max_sessions; mds++) {
3701 		session = __ceph_lookup_mds_session(mdsc, mds);
3702 		if (!session)
3703 			continue;
3704 		mutex_unlock(&mdsc->mutex);
3705 		mutex_lock(&session->s_mutex);
3706 		__close_session(mdsc, session);
3707 		if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3708 			cleanup_session_requests(mdsc, session);
3709 			remove_session_caps(session);
3710 		}
3711 		mutex_unlock(&session->s_mutex);
3712 		ceph_put_mds_session(session);
3713 		mutex_lock(&mdsc->mutex);
3714 		kick_requests(mdsc, mds);
3715 	}
3716 	__wake_requests(mdsc, &mdsc->waiting_for_map);
3717 	mutex_unlock(&mdsc->mutex);
3718 }
3719 
3720 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3721 {
3722 	dout("stop\n");
3723 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3724 	if (mdsc->mdsmap)
3725 		ceph_mdsmap_destroy(mdsc->mdsmap);
3726 	kfree(mdsc->sessions);
3727 	ceph_caps_finalize(mdsc);
3728 	ceph_pool_perm_destroy(mdsc);
3729 }
3730 
3731 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3732 {
3733 	struct ceph_mds_client *mdsc = fsc->mdsc;
3734 
3735 	dout("mdsc_destroy %p\n", mdsc);
3736 	ceph_mdsc_stop(mdsc);
3737 
3738 	/* flush out any connection work with references to us */
3739 	ceph_msgr_flush();
3740 
3741 	fsc->mdsc = NULL;
3742 	kfree(mdsc);
3743 	dout("mdsc_destroy %p done\n", mdsc);
3744 }
3745 
3746 
3747 /*
3748  * handle mds map update.
3749  */
3750 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3751 {
3752 	u32 epoch;
3753 	u32 maplen;
3754 	void *p = msg->front.iov_base;
3755 	void *end = p + msg->front.iov_len;
3756 	struct ceph_mdsmap *newmap, *oldmap;
3757 	struct ceph_fsid fsid;
3758 	int err = -EINVAL;
3759 
3760 	ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3761 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
3762 	if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3763 		return;
3764 	epoch = ceph_decode_32(&p);
3765 	maplen = ceph_decode_32(&p);
3766 	dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3767 
3768 	/* do we need it? */
3769 	mutex_lock(&mdsc->mutex);
3770 	if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3771 		dout("handle_map epoch %u <= our %u\n",
3772 		     epoch, mdsc->mdsmap->m_epoch);
3773 		mutex_unlock(&mdsc->mutex);
3774 		return;
3775 	}
3776 
3777 	newmap = ceph_mdsmap_decode(&p, end);
3778 	if (IS_ERR(newmap)) {
3779 		err = PTR_ERR(newmap);
3780 		goto bad_unlock;
3781 	}
3782 
3783 	/* swap into place */
3784 	if (mdsc->mdsmap) {
3785 		oldmap = mdsc->mdsmap;
3786 		mdsc->mdsmap = newmap;
3787 		check_new_map(mdsc, newmap, oldmap);
3788 		ceph_mdsmap_destroy(oldmap);
3789 	} else {
3790 		mdsc->mdsmap = newmap;  /* first mds map */
3791 	}
3792 	mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3793 
3794 	__wake_requests(mdsc, &mdsc->waiting_for_map);
3795 	ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3796 			  mdsc->mdsmap->m_epoch);
3797 
3798 	mutex_unlock(&mdsc->mutex);
3799 	schedule_delayed(mdsc);
3800 	return;
3801 
3802 bad_unlock:
3803 	mutex_unlock(&mdsc->mutex);
3804 bad:
3805 	pr_err("error decoding mdsmap %d\n", err);
3806 	return;
3807 }
3808 
3809 static struct ceph_connection *con_get(struct ceph_connection *con)
3810 {
3811 	struct ceph_mds_session *s = con->private;
3812 
3813 	if (get_session(s)) {
3814 		dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3815 		return con;
3816 	}
3817 	dout("mdsc con_get %p FAIL\n", s);
3818 	return NULL;
3819 }
3820 
3821 static void con_put(struct ceph_connection *con)
3822 {
3823 	struct ceph_mds_session *s = con->private;
3824 
3825 	dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3826 	ceph_put_mds_session(s);
3827 }
3828 
3829 /*
3830  * if the client is unresponsive for long enough, the mds will kill
3831  * the session entirely.
3832  */
3833 static void peer_reset(struct ceph_connection *con)
3834 {
3835 	struct ceph_mds_session *s = con->private;
3836 	struct ceph_mds_client *mdsc = s->s_mdsc;
3837 
3838 	pr_warn("mds%d closed our session\n", s->s_mds);
3839 	send_mds_reconnect(mdsc, s);
3840 }
3841 
3842 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3843 {
3844 	struct ceph_mds_session *s = con->private;
3845 	struct ceph_mds_client *mdsc = s->s_mdsc;
3846 	int type = le16_to_cpu(msg->hdr.type);
3847 
3848 	mutex_lock(&mdsc->mutex);
3849 	if (__verify_registered_session(mdsc, s) < 0) {
3850 		mutex_unlock(&mdsc->mutex);
3851 		goto out;
3852 	}
3853 	mutex_unlock(&mdsc->mutex);
3854 
3855 	switch (type) {
3856 	case CEPH_MSG_MDS_MAP:
3857 		ceph_mdsc_handle_map(mdsc, msg);
3858 		break;
3859 	case CEPH_MSG_CLIENT_SESSION:
3860 		handle_session(s, msg);
3861 		break;
3862 	case CEPH_MSG_CLIENT_REPLY:
3863 		handle_reply(s, msg);
3864 		break;
3865 	case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3866 		handle_forward(mdsc, s, msg);
3867 		break;
3868 	case CEPH_MSG_CLIENT_CAPS:
3869 		ceph_handle_caps(s, msg);
3870 		break;
3871 	case CEPH_MSG_CLIENT_SNAP:
3872 		ceph_handle_snap(mdsc, s, msg);
3873 		break;
3874 	case CEPH_MSG_CLIENT_LEASE:
3875 		handle_lease(mdsc, s, msg);
3876 		break;
3877 
3878 	default:
3879 		pr_err("received unknown message type %d %s\n", type,
3880 		       ceph_msg_type_name(type));
3881 	}
3882 out:
3883 	ceph_msg_put(msg);
3884 }
3885 
3886 /*
3887  * authentication
3888  */
3889 
3890 /*
3891  * Note: returned pointer is the address of a structure that's
3892  * managed separately.  Caller must *not* attempt to free it.
3893  */
3894 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3895 					int *proto, int force_new)
3896 {
3897 	struct ceph_mds_session *s = con->private;
3898 	struct ceph_mds_client *mdsc = s->s_mdsc;
3899 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3900 	struct ceph_auth_handshake *auth = &s->s_auth;
3901 
3902 	if (force_new && auth->authorizer) {
3903 		ceph_auth_destroy_authorizer(ac, auth->authorizer);
3904 		auth->authorizer = NULL;
3905 	}
3906 	if (!auth->authorizer) {
3907 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3908 						      auth);
3909 		if (ret)
3910 			return ERR_PTR(ret);
3911 	} else {
3912 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3913 						      auth);
3914 		if (ret)
3915 			return ERR_PTR(ret);
3916 	}
3917 	*proto = ac->protocol;
3918 
3919 	return auth;
3920 }
3921 
3922 
3923 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3924 {
3925 	struct ceph_mds_session *s = con->private;
3926 	struct ceph_mds_client *mdsc = s->s_mdsc;
3927 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3928 
3929 	return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3930 }
3931 
3932 static int invalidate_authorizer(struct ceph_connection *con)
3933 {
3934 	struct ceph_mds_session *s = con->private;
3935 	struct ceph_mds_client *mdsc = s->s_mdsc;
3936 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3937 
3938 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3939 
3940 	return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3941 }
3942 
3943 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3944 				struct ceph_msg_header *hdr, int *skip)
3945 {
3946 	struct ceph_msg *msg;
3947 	int type = (int) le16_to_cpu(hdr->type);
3948 	int front_len = (int) le32_to_cpu(hdr->front_len);
3949 
3950 	if (con->in_msg)
3951 		return con->in_msg;
3952 
3953 	*skip = 0;
3954 	msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3955 	if (!msg) {
3956 		pr_err("unable to allocate msg type %d len %d\n",
3957 		       type, front_len);
3958 		return NULL;
3959 	}
3960 
3961 	return msg;
3962 }
3963 
3964 static int mds_sign_message(struct ceph_msg *msg)
3965 {
3966        struct ceph_mds_session *s = msg->con->private;
3967        struct ceph_auth_handshake *auth = &s->s_auth;
3968 
3969        return ceph_auth_sign_message(auth, msg);
3970 }
3971 
3972 static int mds_check_message_signature(struct ceph_msg *msg)
3973 {
3974        struct ceph_mds_session *s = msg->con->private;
3975        struct ceph_auth_handshake *auth = &s->s_auth;
3976 
3977        return ceph_auth_check_message_signature(auth, msg);
3978 }
3979 
3980 static const struct ceph_connection_operations mds_con_ops = {
3981 	.get = con_get,
3982 	.put = con_put,
3983 	.dispatch = dispatch,
3984 	.get_authorizer = get_authorizer,
3985 	.verify_authorizer_reply = verify_authorizer_reply,
3986 	.invalidate_authorizer = invalidate_authorizer,
3987 	.peer_reset = peer_reset,
3988 	.alloc_msg = mds_alloc_msg,
3989 	.sign_message = mds_sign_message,
3990 	.check_message_signature = mds_check_message_signature,
3991 };
3992 
3993 /* eof */
3994