xref: /openbmc/linux/fs/ceph/mds_client.c (revision 5f2fb52fac15a8a8e10ce020dd532504a8abfc4e)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3 
4 #include <linux/fs.h>
5 #include <linux/wait.h>
6 #include <linux/slab.h>
7 #include <linux/gfp.h>
8 #include <linux/sched.h>
9 #include <linux/debugfs.h>
10 #include <linux/seq_file.h>
11 #include <linux/ratelimit.h>
12 
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 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
24 
25 /*
26  * A cluster of MDS (metadata server) daemons is responsible for
27  * managing the file system namespace (the directory hierarchy and
28  * inodes) and for coordinating shared access to storage.  Metadata is
29  * partitioning hierarchically across a number of servers, and that
30  * partition varies over time as the cluster adjusts the distribution
31  * in order to balance load.
32  *
33  * The MDS client is primarily responsible to managing synchronous
34  * metadata requests for operations like open, unlink, and so forth.
35  * If there is a MDS failure, we find out about it when we (possibly
36  * request and) receive a new MDS map, and can resubmit affected
37  * requests.
38  *
39  * For the most part, though, we take advantage of a lossless
40  * communications channel to the MDS, and do not need to worry about
41  * timing out or resubmitting requests.
42  *
43  * We maintain a stateful "session" with each MDS we interact with.
44  * Within each session, we sent periodic heartbeat messages to ensure
45  * any capabilities or leases we have been issues remain valid.  If
46  * the session times out and goes stale, our leases and capabilities
47  * are no longer valid.
48  */
49 
50 struct ceph_reconnect_state {
51 	struct ceph_mds_session *session;
52 	int nr_caps, nr_realms;
53 	struct ceph_pagelist *pagelist;
54 	unsigned msg_version;
55 	bool allow_multi;
56 };
57 
58 static void __wake_requests(struct ceph_mds_client *mdsc,
59 			    struct list_head *head);
60 static void ceph_cap_release_work(struct work_struct *work);
61 static void ceph_cap_reclaim_work(struct work_struct *work);
62 
63 static const struct ceph_connection_operations mds_con_ops;
64 
65 
66 /*
67  * mds reply parsing
68  */
69 
70 static int parse_reply_info_quota(void **p, void *end,
71 				  struct ceph_mds_reply_info_in *info)
72 {
73 	u8 struct_v, struct_compat;
74 	u32 struct_len;
75 
76 	ceph_decode_8_safe(p, end, struct_v, bad);
77 	ceph_decode_8_safe(p, end, struct_compat, bad);
78 	/* struct_v is expected to be >= 1. we only
79 	 * understand encoding with struct_compat == 1. */
80 	if (!struct_v || struct_compat != 1)
81 		goto bad;
82 	ceph_decode_32_safe(p, end, struct_len, bad);
83 	ceph_decode_need(p, end, struct_len, bad);
84 	end = *p + struct_len;
85 	ceph_decode_64_safe(p, end, info->max_bytes, bad);
86 	ceph_decode_64_safe(p, end, info->max_files, bad);
87 	*p = end;
88 	return 0;
89 bad:
90 	return -EIO;
91 }
92 
93 /*
94  * parse individual inode info
95  */
96 static int parse_reply_info_in(void **p, void *end,
97 			       struct ceph_mds_reply_info_in *info,
98 			       u64 features)
99 {
100 	int err = 0;
101 	u8 struct_v = 0;
102 
103 	if (features == (u64)-1) {
104 		u32 struct_len;
105 		u8 struct_compat;
106 		ceph_decode_8_safe(p, end, struct_v, bad);
107 		ceph_decode_8_safe(p, end, struct_compat, bad);
108 		/* struct_v is expected to be >= 1. we only understand
109 		 * encoding with struct_compat == 1. */
110 		if (!struct_v || struct_compat != 1)
111 			goto bad;
112 		ceph_decode_32_safe(p, end, struct_len, bad);
113 		ceph_decode_need(p, end, struct_len, bad);
114 		end = *p + struct_len;
115 	}
116 
117 	ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
118 	info->in = *p;
119 	*p += sizeof(struct ceph_mds_reply_inode) +
120 		sizeof(*info->in->fragtree.splits) *
121 		le32_to_cpu(info->in->fragtree.nsplits);
122 
123 	ceph_decode_32_safe(p, end, info->symlink_len, bad);
124 	ceph_decode_need(p, end, info->symlink_len, bad);
125 	info->symlink = *p;
126 	*p += info->symlink_len;
127 
128 	ceph_decode_copy_safe(p, end, &info->dir_layout,
129 			      sizeof(info->dir_layout), bad);
130 	ceph_decode_32_safe(p, end, info->xattr_len, bad);
131 	ceph_decode_need(p, end, info->xattr_len, bad);
132 	info->xattr_data = *p;
133 	*p += info->xattr_len;
134 
135 	if (features == (u64)-1) {
136 		/* inline data */
137 		ceph_decode_64_safe(p, end, info->inline_version, bad);
138 		ceph_decode_32_safe(p, end, info->inline_len, bad);
139 		ceph_decode_need(p, end, info->inline_len, bad);
140 		info->inline_data = *p;
141 		*p += info->inline_len;
142 		/* quota */
143 		err = parse_reply_info_quota(p, end, info);
144 		if (err < 0)
145 			goto out_bad;
146 		/* pool namespace */
147 		ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
148 		if (info->pool_ns_len > 0) {
149 			ceph_decode_need(p, end, info->pool_ns_len, bad);
150 			info->pool_ns_data = *p;
151 			*p += info->pool_ns_len;
152 		}
153 
154 		/* btime */
155 		ceph_decode_need(p, end, sizeof(info->btime), bad);
156 		ceph_decode_copy(p, &info->btime, sizeof(info->btime));
157 
158 		/* change attribute */
159 		ceph_decode_64_safe(p, end, info->change_attr, bad);
160 
161 		/* dir pin */
162 		if (struct_v >= 2) {
163 			ceph_decode_32_safe(p, end, info->dir_pin, bad);
164 		} else {
165 			info->dir_pin = -ENODATA;
166 		}
167 
168 		/* snapshot birth time, remains zero for v<=2 */
169 		if (struct_v >= 3) {
170 			ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
171 			ceph_decode_copy(p, &info->snap_btime,
172 					 sizeof(info->snap_btime));
173 		} else {
174 			memset(&info->snap_btime, 0, sizeof(info->snap_btime));
175 		}
176 
177 		*p = end;
178 	} else {
179 		if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
180 			ceph_decode_64_safe(p, end, info->inline_version, bad);
181 			ceph_decode_32_safe(p, end, info->inline_len, bad);
182 			ceph_decode_need(p, end, info->inline_len, bad);
183 			info->inline_data = *p;
184 			*p += info->inline_len;
185 		} else
186 			info->inline_version = CEPH_INLINE_NONE;
187 
188 		if (features & CEPH_FEATURE_MDS_QUOTA) {
189 			err = parse_reply_info_quota(p, end, info);
190 			if (err < 0)
191 				goto out_bad;
192 		} else {
193 			info->max_bytes = 0;
194 			info->max_files = 0;
195 		}
196 
197 		info->pool_ns_len = 0;
198 		info->pool_ns_data = NULL;
199 		if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
200 			ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
201 			if (info->pool_ns_len > 0) {
202 				ceph_decode_need(p, end, info->pool_ns_len, bad);
203 				info->pool_ns_data = *p;
204 				*p += info->pool_ns_len;
205 			}
206 		}
207 
208 		if (features & CEPH_FEATURE_FS_BTIME) {
209 			ceph_decode_need(p, end, sizeof(info->btime), bad);
210 			ceph_decode_copy(p, &info->btime, sizeof(info->btime));
211 			ceph_decode_64_safe(p, end, info->change_attr, bad);
212 		}
213 
214 		info->dir_pin = -ENODATA;
215 		/* info->snap_btime remains zero */
216 	}
217 	return 0;
218 bad:
219 	err = -EIO;
220 out_bad:
221 	return err;
222 }
223 
224 static int parse_reply_info_dir(void **p, void *end,
225 				struct ceph_mds_reply_dirfrag **dirfrag,
226 				u64 features)
227 {
228 	if (features == (u64)-1) {
229 		u8 struct_v, struct_compat;
230 		u32 struct_len;
231 		ceph_decode_8_safe(p, end, struct_v, bad);
232 		ceph_decode_8_safe(p, end, struct_compat, bad);
233 		/* struct_v is expected to be >= 1. we only understand
234 		 * encoding whose struct_compat == 1. */
235 		if (!struct_v || struct_compat != 1)
236 			goto bad;
237 		ceph_decode_32_safe(p, end, struct_len, bad);
238 		ceph_decode_need(p, end, struct_len, bad);
239 		end = *p + struct_len;
240 	}
241 
242 	ceph_decode_need(p, end, sizeof(**dirfrag), bad);
243 	*dirfrag = *p;
244 	*p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
245 	if (unlikely(*p > end))
246 		goto bad;
247 	if (features == (u64)-1)
248 		*p = end;
249 	return 0;
250 bad:
251 	return -EIO;
252 }
253 
254 static int parse_reply_info_lease(void **p, void *end,
255 				  struct ceph_mds_reply_lease **lease,
256 				  u64 features)
257 {
258 	if (features == (u64)-1) {
259 		u8 struct_v, struct_compat;
260 		u32 struct_len;
261 		ceph_decode_8_safe(p, end, struct_v, bad);
262 		ceph_decode_8_safe(p, end, struct_compat, bad);
263 		/* struct_v is expected to be >= 1. we only understand
264 		 * encoding whose struct_compat == 1. */
265 		if (!struct_v || struct_compat != 1)
266 			goto bad;
267 		ceph_decode_32_safe(p, end, struct_len, bad);
268 		ceph_decode_need(p, end, struct_len, bad);
269 		end = *p + struct_len;
270 	}
271 
272 	ceph_decode_need(p, end, sizeof(**lease), bad);
273 	*lease = *p;
274 	*p += sizeof(**lease);
275 	if (features == (u64)-1)
276 		*p = end;
277 	return 0;
278 bad:
279 	return -EIO;
280 }
281 
282 /*
283  * parse a normal reply, which may contain a (dir+)dentry and/or a
284  * target inode.
285  */
286 static int parse_reply_info_trace(void **p, void *end,
287 				  struct ceph_mds_reply_info_parsed *info,
288 				  u64 features)
289 {
290 	int err;
291 
292 	if (info->head->is_dentry) {
293 		err = parse_reply_info_in(p, end, &info->diri, features);
294 		if (err < 0)
295 			goto out_bad;
296 
297 		err = parse_reply_info_dir(p, end, &info->dirfrag, features);
298 		if (err < 0)
299 			goto out_bad;
300 
301 		ceph_decode_32_safe(p, end, info->dname_len, bad);
302 		ceph_decode_need(p, end, info->dname_len, bad);
303 		info->dname = *p;
304 		*p += info->dname_len;
305 
306 		err = parse_reply_info_lease(p, end, &info->dlease, features);
307 		if (err < 0)
308 			goto out_bad;
309 	}
310 
311 	if (info->head->is_target) {
312 		err = parse_reply_info_in(p, end, &info->targeti, features);
313 		if (err < 0)
314 			goto out_bad;
315 	}
316 
317 	if (unlikely(*p != end))
318 		goto bad;
319 	return 0;
320 
321 bad:
322 	err = -EIO;
323 out_bad:
324 	pr_err("problem parsing mds trace %d\n", err);
325 	return err;
326 }
327 
328 /*
329  * parse readdir results
330  */
331 static int parse_reply_info_readdir(void **p, void *end,
332 				struct ceph_mds_reply_info_parsed *info,
333 				u64 features)
334 {
335 	u32 num, i = 0;
336 	int err;
337 
338 	err = parse_reply_info_dir(p, end, &info->dir_dir, features);
339 	if (err < 0)
340 		goto out_bad;
341 
342 	ceph_decode_need(p, end, sizeof(num) + 2, bad);
343 	num = ceph_decode_32(p);
344 	{
345 		u16 flags = ceph_decode_16(p);
346 		info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
347 		info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
348 		info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
349 		info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
350 	}
351 	if (num == 0)
352 		goto done;
353 
354 	BUG_ON(!info->dir_entries);
355 	if ((unsigned long)(info->dir_entries + num) >
356 	    (unsigned long)info->dir_entries + info->dir_buf_size) {
357 		pr_err("dir contents are larger than expected\n");
358 		WARN_ON(1);
359 		goto bad;
360 	}
361 
362 	info->dir_nr = num;
363 	while (num) {
364 		struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
365 		/* dentry */
366 		ceph_decode_32_safe(p, end, rde->name_len, bad);
367 		ceph_decode_need(p, end, rde->name_len, bad);
368 		rde->name = *p;
369 		*p += rde->name_len;
370 		dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
371 
372 		/* dentry lease */
373 		err = parse_reply_info_lease(p, end, &rde->lease, features);
374 		if (err)
375 			goto out_bad;
376 		/* inode */
377 		err = parse_reply_info_in(p, end, &rde->inode, features);
378 		if (err < 0)
379 			goto out_bad;
380 		/* ceph_readdir_prepopulate() will update it */
381 		rde->offset = 0;
382 		i++;
383 		num--;
384 	}
385 
386 done:
387 	/* Skip over any unrecognized fields */
388 	*p = end;
389 	return 0;
390 
391 bad:
392 	err = -EIO;
393 out_bad:
394 	pr_err("problem parsing dir contents %d\n", err);
395 	return err;
396 }
397 
398 /*
399  * parse fcntl F_GETLK results
400  */
401 static int parse_reply_info_filelock(void **p, void *end,
402 				     struct ceph_mds_reply_info_parsed *info,
403 				     u64 features)
404 {
405 	if (*p + sizeof(*info->filelock_reply) > end)
406 		goto bad;
407 
408 	info->filelock_reply = *p;
409 
410 	/* Skip over any unrecognized fields */
411 	*p = end;
412 	return 0;
413 bad:
414 	return -EIO;
415 }
416 
417 /*
418  * parse create results
419  */
420 static int parse_reply_info_create(void **p, void *end,
421 				  struct ceph_mds_reply_info_parsed *info,
422 				  u64 features)
423 {
424 	if (features == (u64)-1 ||
425 	    (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
426 		/* Malformed reply? */
427 		if (*p == end) {
428 			info->has_create_ino = false;
429 		} else {
430 			info->has_create_ino = true;
431 			ceph_decode_64_safe(p, end, info->ino, bad);
432 		}
433 	} else {
434 		if (*p != end)
435 			goto bad;
436 	}
437 
438 	/* Skip over any unrecognized fields */
439 	*p = end;
440 	return 0;
441 bad:
442 	return -EIO;
443 }
444 
445 /*
446  * parse extra results
447  */
448 static int parse_reply_info_extra(void **p, void *end,
449 				  struct ceph_mds_reply_info_parsed *info,
450 				  u64 features)
451 {
452 	u32 op = le32_to_cpu(info->head->op);
453 
454 	if (op == CEPH_MDS_OP_GETFILELOCK)
455 		return parse_reply_info_filelock(p, end, info, features);
456 	else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
457 		return parse_reply_info_readdir(p, end, info, features);
458 	else if (op == CEPH_MDS_OP_CREATE)
459 		return parse_reply_info_create(p, end, info, features);
460 	else
461 		return -EIO;
462 }
463 
464 /*
465  * parse entire mds reply
466  */
467 static int parse_reply_info(struct ceph_msg *msg,
468 			    struct ceph_mds_reply_info_parsed *info,
469 			    u64 features)
470 {
471 	void *p, *end;
472 	u32 len;
473 	int err;
474 
475 	info->head = msg->front.iov_base;
476 	p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
477 	end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
478 
479 	/* trace */
480 	ceph_decode_32_safe(&p, end, len, bad);
481 	if (len > 0) {
482 		ceph_decode_need(&p, end, len, bad);
483 		err = parse_reply_info_trace(&p, p+len, info, features);
484 		if (err < 0)
485 			goto out_bad;
486 	}
487 
488 	/* extra */
489 	ceph_decode_32_safe(&p, end, len, bad);
490 	if (len > 0) {
491 		ceph_decode_need(&p, end, len, bad);
492 		err = parse_reply_info_extra(&p, p+len, info, features);
493 		if (err < 0)
494 			goto out_bad;
495 	}
496 
497 	/* snap blob */
498 	ceph_decode_32_safe(&p, end, len, bad);
499 	info->snapblob_len = len;
500 	info->snapblob = p;
501 	p += len;
502 
503 	if (p != end)
504 		goto bad;
505 	return 0;
506 
507 bad:
508 	err = -EIO;
509 out_bad:
510 	pr_err("mds parse_reply err %d\n", err);
511 	return err;
512 }
513 
514 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
515 {
516 	if (!info->dir_entries)
517 		return;
518 	free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
519 }
520 
521 
522 /*
523  * sessions
524  */
525 const char *ceph_session_state_name(int s)
526 {
527 	switch (s) {
528 	case CEPH_MDS_SESSION_NEW: return "new";
529 	case CEPH_MDS_SESSION_OPENING: return "opening";
530 	case CEPH_MDS_SESSION_OPEN: return "open";
531 	case CEPH_MDS_SESSION_HUNG: return "hung";
532 	case CEPH_MDS_SESSION_CLOSING: return "closing";
533 	case CEPH_MDS_SESSION_RESTARTING: return "restarting";
534 	case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
535 	case CEPH_MDS_SESSION_REJECTED: return "rejected";
536 	default: return "???";
537 	}
538 }
539 
540 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
541 {
542 	if (refcount_inc_not_zero(&s->s_ref)) {
543 		dout("mdsc get_session %p %d -> %d\n", s,
544 		     refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
545 		return s;
546 	} else {
547 		dout("mdsc get_session %p 0 -- FAIL\n", s);
548 		return NULL;
549 	}
550 }
551 
552 void ceph_put_mds_session(struct ceph_mds_session *s)
553 {
554 	dout("mdsc put_session %p %d -> %d\n", s,
555 	     refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
556 	if (refcount_dec_and_test(&s->s_ref)) {
557 		if (s->s_auth.authorizer)
558 			ceph_auth_destroy_authorizer(s->s_auth.authorizer);
559 		kfree(s);
560 	}
561 }
562 
563 /*
564  * called under mdsc->mutex
565  */
566 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
567 						   int mds)
568 {
569 	if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
570 		return NULL;
571 	return get_session(mdsc->sessions[mds]);
572 }
573 
574 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
575 {
576 	if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
577 		return false;
578 	else
579 		return true;
580 }
581 
582 static int __verify_registered_session(struct ceph_mds_client *mdsc,
583 				       struct ceph_mds_session *s)
584 {
585 	if (s->s_mds >= mdsc->max_sessions ||
586 	    mdsc->sessions[s->s_mds] != s)
587 		return -ENOENT;
588 	return 0;
589 }
590 
591 /*
592  * create+register a new session for given mds.
593  * called under mdsc->mutex.
594  */
595 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
596 						 int mds)
597 {
598 	struct ceph_mds_session *s;
599 
600 	if (mds >= mdsc->mdsmap->m_num_mds)
601 		return ERR_PTR(-EINVAL);
602 
603 	s = kzalloc(sizeof(*s), GFP_NOFS);
604 	if (!s)
605 		return ERR_PTR(-ENOMEM);
606 
607 	if (mds >= mdsc->max_sessions) {
608 		int newmax = 1 << get_count_order(mds + 1);
609 		struct ceph_mds_session **sa;
610 
611 		dout("%s: realloc to %d\n", __func__, newmax);
612 		sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
613 		if (!sa)
614 			goto fail_realloc;
615 		if (mdsc->sessions) {
616 			memcpy(sa, mdsc->sessions,
617 			       mdsc->max_sessions * sizeof(void *));
618 			kfree(mdsc->sessions);
619 		}
620 		mdsc->sessions = sa;
621 		mdsc->max_sessions = newmax;
622 	}
623 
624 	dout("%s: mds%d\n", __func__, mds);
625 	s->s_mdsc = mdsc;
626 	s->s_mds = mds;
627 	s->s_state = CEPH_MDS_SESSION_NEW;
628 	s->s_ttl = 0;
629 	s->s_seq = 0;
630 	mutex_init(&s->s_mutex);
631 
632 	ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
633 
634 	spin_lock_init(&s->s_gen_ttl_lock);
635 	s->s_cap_gen = 1;
636 	s->s_cap_ttl = jiffies - 1;
637 
638 	spin_lock_init(&s->s_cap_lock);
639 	s->s_renew_requested = 0;
640 	s->s_renew_seq = 0;
641 	INIT_LIST_HEAD(&s->s_caps);
642 	s->s_nr_caps = 0;
643 	refcount_set(&s->s_ref, 1);
644 	INIT_LIST_HEAD(&s->s_waiting);
645 	INIT_LIST_HEAD(&s->s_unsafe);
646 	s->s_num_cap_releases = 0;
647 	s->s_cap_reconnect = 0;
648 	s->s_cap_iterator = NULL;
649 	INIT_LIST_HEAD(&s->s_cap_releases);
650 	INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
651 
652 	INIT_LIST_HEAD(&s->s_cap_flushing);
653 
654 	mdsc->sessions[mds] = s;
655 	atomic_inc(&mdsc->num_sessions);
656 	refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
657 
658 	ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
659 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
660 
661 	return s;
662 
663 fail_realloc:
664 	kfree(s);
665 	return ERR_PTR(-ENOMEM);
666 }
667 
668 /*
669  * called under mdsc->mutex
670  */
671 static void __unregister_session(struct ceph_mds_client *mdsc,
672 			       struct ceph_mds_session *s)
673 {
674 	dout("__unregister_session mds%d %p\n", s->s_mds, s);
675 	BUG_ON(mdsc->sessions[s->s_mds] != s);
676 	mdsc->sessions[s->s_mds] = NULL;
677 	s->s_state = 0;
678 	ceph_con_close(&s->s_con);
679 	ceph_put_mds_session(s);
680 	atomic_dec(&mdsc->num_sessions);
681 }
682 
683 /*
684  * drop session refs in request.
685  *
686  * should be last request ref, or hold mdsc->mutex
687  */
688 static void put_request_session(struct ceph_mds_request *req)
689 {
690 	if (req->r_session) {
691 		ceph_put_mds_session(req->r_session);
692 		req->r_session = NULL;
693 	}
694 }
695 
696 void ceph_mdsc_release_request(struct kref *kref)
697 {
698 	struct ceph_mds_request *req = container_of(kref,
699 						    struct ceph_mds_request,
700 						    r_kref);
701 	destroy_reply_info(&req->r_reply_info);
702 	if (req->r_request)
703 		ceph_msg_put(req->r_request);
704 	if (req->r_reply)
705 		ceph_msg_put(req->r_reply);
706 	if (req->r_inode) {
707 		ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
708 		/* avoid calling iput_final() in mds dispatch threads */
709 		ceph_async_iput(req->r_inode);
710 	}
711 	if (req->r_parent) {
712 		ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
713 		ceph_async_iput(req->r_parent);
714 	}
715 	ceph_async_iput(req->r_target_inode);
716 	if (req->r_dentry)
717 		dput(req->r_dentry);
718 	if (req->r_old_dentry)
719 		dput(req->r_old_dentry);
720 	if (req->r_old_dentry_dir) {
721 		/*
722 		 * track (and drop pins for) r_old_dentry_dir
723 		 * separately, since r_old_dentry's d_parent may have
724 		 * changed between the dir mutex being dropped and
725 		 * this request being freed.
726 		 */
727 		ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
728 				  CEPH_CAP_PIN);
729 		ceph_async_iput(req->r_old_dentry_dir);
730 	}
731 	kfree(req->r_path1);
732 	kfree(req->r_path2);
733 	if (req->r_pagelist)
734 		ceph_pagelist_release(req->r_pagelist);
735 	put_request_session(req);
736 	ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
737 	WARN_ON_ONCE(!list_empty(&req->r_wait));
738 	kfree(req);
739 }
740 
741 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
742 
743 /*
744  * lookup session, bump ref if found.
745  *
746  * called under mdsc->mutex.
747  */
748 static struct ceph_mds_request *
749 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
750 {
751 	struct ceph_mds_request *req;
752 
753 	req = lookup_request(&mdsc->request_tree, tid);
754 	if (req)
755 		ceph_mdsc_get_request(req);
756 
757 	return req;
758 }
759 
760 /*
761  * Register an in-flight request, and assign a tid.  Link to directory
762  * are modifying (if any).
763  *
764  * Called under mdsc->mutex.
765  */
766 static void __register_request(struct ceph_mds_client *mdsc,
767 			       struct ceph_mds_request *req,
768 			       struct inode *dir)
769 {
770 	int ret = 0;
771 
772 	req->r_tid = ++mdsc->last_tid;
773 	if (req->r_num_caps) {
774 		ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
775 					req->r_num_caps);
776 		if (ret < 0) {
777 			pr_err("__register_request %p "
778 			       "failed to reserve caps: %d\n", req, ret);
779 			/* set req->r_err to fail early from __do_request */
780 			req->r_err = ret;
781 			return;
782 		}
783 	}
784 	dout("__register_request %p tid %lld\n", req, req->r_tid);
785 	ceph_mdsc_get_request(req);
786 	insert_request(&mdsc->request_tree, req);
787 
788 	req->r_uid = current_fsuid();
789 	req->r_gid = current_fsgid();
790 
791 	if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
792 		mdsc->oldest_tid = req->r_tid;
793 
794 	if (dir) {
795 		ihold(dir);
796 		req->r_unsafe_dir = dir;
797 	}
798 }
799 
800 static void __unregister_request(struct ceph_mds_client *mdsc,
801 				 struct ceph_mds_request *req)
802 {
803 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
804 
805 	/* Never leave an unregistered request on an unsafe list! */
806 	list_del_init(&req->r_unsafe_item);
807 
808 	if (req->r_tid == mdsc->oldest_tid) {
809 		struct rb_node *p = rb_next(&req->r_node);
810 		mdsc->oldest_tid = 0;
811 		while (p) {
812 			struct ceph_mds_request *next_req =
813 				rb_entry(p, struct ceph_mds_request, r_node);
814 			if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
815 				mdsc->oldest_tid = next_req->r_tid;
816 				break;
817 			}
818 			p = rb_next(p);
819 		}
820 	}
821 
822 	erase_request(&mdsc->request_tree, req);
823 
824 	if (req->r_unsafe_dir  &&
825 	    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
826 		struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
827 		spin_lock(&ci->i_unsafe_lock);
828 		list_del_init(&req->r_unsafe_dir_item);
829 		spin_unlock(&ci->i_unsafe_lock);
830 	}
831 	if (req->r_target_inode &&
832 	    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
833 		struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
834 		spin_lock(&ci->i_unsafe_lock);
835 		list_del_init(&req->r_unsafe_target_item);
836 		spin_unlock(&ci->i_unsafe_lock);
837 	}
838 
839 	if (req->r_unsafe_dir) {
840 		/* avoid calling iput_final() in mds dispatch threads */
841 		ceph_async_iput(req->r_unsafe_dir);
842 		req->r_unsafe_dir = NULL;
843 	}
844 
845 	complete_all(&req->r_safe_completion);
846 
847 	ceph_mdsc_put_request(req);
848 }
849 
850 /*
851  * Walk back up the dentry tree until we hit a dentry representing a
852  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
853  * when calling this) to ensure that the objects won't disappear while we're
854  * working with them. Once we hit a candidate dentry, we attempt to take a
855  * reference to it, and return that as the result.
856  */
857 static struct inode *get_nonsnap_parent(struct dentry *dentry)
858 {
859 	struct inode *inode = NULL;
860 
861 	while (dentry && !IS_ROOT(dentry)) {
862 		inode = d_inode_rcu(dentry);
863 		if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
864 			break;
865 		dentry = dentry->d_parent;
866 	}
867 	if (inode)
868 		inode = igrab(inode);
869 	return inode;
870 }
871 
872 /*
873  * Choose mds to send request to next.  If there is a hint set in the
874  * request (e.g., due to a prior forward hint from the mds), use that.
875  * Otherwise, consult frag tree and/or caps to identify the
876  * appropriate mds.  If all else fails, choose randomly.
877  *
878  * Called under mdsc->mutex.
879  */
880 static int __choose_mds(struct ceph_mds_client *mdsc,
881 			struct ceph_mds_request *req)
882 {
883 	struct inode *inode;
884 	struct ceph_inode_info *ci;
885 	struct ceph_cap *cap;
886 	int mode = req->r_direct_mode;
887 	int mds = -1;
888 	u32 hash = req->r_direct_hash;
889 	bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
890 
891 	/*
892 	 * is there a specific mds we should try?  ignore hint if we have
893 	 * no session and the mds is not up (active or recovering).
894 	 */
895 	if (req->r_resend_mds >= 0 &&
896 	    (__have_session(mdsc, req->r_resend_mds) ||
897 	     ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
898 		dout("choose_mds using resend_mds mds%d\n",
899 		     req->r_resend_mds);
900 		return req->r_resend_mds;
901 	}
902 
903 	if (mode == USE_RANDOM_MDS)
904 		goto random;
905 
906 	inode = NULL;
907 	if (req->r_inode) {
908 		if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
909 			inode = req->r_inode;
910 			ihold(inode);
911 		} else {
912 			/* req->r_dentry is non-null for LSSNAP request */
913 			rcu_read_lock();
914 			inode = get_nonsnap_parent(req->r_dentry);
915 			rcu_read_unlock();
916 			dout("__choose_mds using snapdir's parent %p\n", inode);
917 		}
918 	} else if (req->r_dentry) {
919 		/* ignore race with rename; old or new d_parent is okay */
920 		struct dentry *parent;
921 		struct inode *dir;
922 
923 		rcu_read_lock();
924 		parent = READ_ONCE(req->r_dentry->d_parent);
925 		dir = req->r_parent ? : d_inode_rcu(parent);
926 
927 		if (!dir || dir->i_sb != mdsc->fsc->sb) {
928 			/*  not this fs or parent went negative */
929 			inode = d_inode(req->r_dentry);
930 			if (inode)
931 				ihold(inode);
932 		} else if (ceph_snap(dir) != CEPH_NOSNAP) {
933 			/* direct snapped/virtual snapdir requests
934 			 * based on parent dir inode */
935 			inode = get_nonsnap_parent(parent);
936 			dout("__choose_mds using nonsnap parent %p\n", inode);
937 		} else {
938 			/* dentry target */
939 			inode = d_inode(req->r_dentry);
940 			if (!inode || mode == USE_AUTH_MDS) {
941 				/* dir + name */
942 				inode = igrab(dir);
943 				hash = ceph_dentry_hash(dir, req->r_dentry);
944 				is_hash = true;
945 			} else {
946 				ihold(inode);
947 			}
948 		}
949 		rcu_read_unlock();
950 	}
951 
952 	dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
953 	     (int)hash, mode);
954 	if (!inode)
955 		goto random;
956 	ci = ceph_inode(inode);
957 
958 	if (is_hash && S_ISDIR(inode->i_mode)) {
959 		struct ceph_inode_frag frag;
960 		int found;
961 
962 		ceph_choose_frag(ci, hash, &frag, &found);
963 		if (found) {
964 			if (mode == USE_ANY_MDS && frag.ndist > 0) {
965 				u8 r;
966 
967 				/* choose a random replica */
968 				get_random_bytes(&r, 1);
969 				r %= frag.ndist;
970 				mds = frag.dist[r];
971 				dout("choose_mds %p %llx.%llx "
972 				     "frag %u mds%d (%d/%d)\n",
973 				     inode, ceph_vinop(inode),
974 				     frag.frag, mds,
975 				     (int)r, frag.ndist);
976 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
977 				    CEPH_MDS_STATE_ACTIVE)
978 					goto out;
979 			}
980 
981 			/* since this file/dir wasn't known to be
982 			 * replicated, then we want to look for the
983 			 * authoritative mds. */
984 			mode = USE_AUTH_MDS;
985 			if (frag.mds >= 0) {
986 				/* choose auth mds */
987 				mds = frag.mds;
988 				dout("choose_mds %p %llx.%llx "
989 				     "frag %u mds%d (auth)\n",
990 				     inode, ceph_vinop(inode), frag.frag, mds);
991 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
992 				    CEPH_MDS_STATE_ACTIVE)
993 					goto out;
994 			}
995 		}
996 	}
997 
998 	spin_lock(&ci->i_ceph_lock);
999 	cap = NULL;
1000 	if (mode == USE_AUTH_MDS)
1001 		cap = ci->i_auth_cap;
1002 	if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1003 		cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1004 	if (!cap) {
1005 		spin_unlock(&ci->i_ceph_lock);
1006 		ceph_async_iput(inode);
1007 		goto random;
1008 	}
1009 	mds = cap->session->s_mds;
1010 	dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
1011 	     inode, ceph_vinop(inode), mds,
1012 	     cap == ci->i_auth_cap ? "auth " : "", cap);
1013 	spin_unlock(&ci->i_ceph_lock);
1014 out:
1015 	/* avoid calling iput_final() while holding mdsc->mutex or
1016 	 * in mds dispatch threads */
1017 	ceph_async_iput(inode);
1018 	return mds;
1019 
1020 random:
1021 	mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1022 	dout("choose_mds chose random mds%d\n", mds);
1023 	return mds;
1024 }
1025 
1026 
1027 /*
1028  * session messages
1029  */
1030 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1031 {
1032 	struct ceph_msg *msg;
1033 	struct ceph_mds_session_head *h;
1034 
1035 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1036 			   false);
1037 	if (!msg) {
1038 		pr_err("create_session_msg ENOMEM creating msg\n");
1039 		return NULL;
1040 	}
1041 	h = msg->front.iov_base;
1042 	h->op = cpu_to_le32(op);
1043 	h->seq = cpu_to_le64(seq);
1044 
1045 	return msg;
1046 }
1047 
1048 static void encode_supported_features(void **p, void *end)
1049 {
1050 	static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1051 	static const size_t count = ARRAY_SIZE(bits);
1052 
1053 	if (count > 0) {
1054 		size_t i;
1055 		size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
1056 
1057 		BUG_ON(*p + 4 + size > end);
1058 		ceph_encode_32(p, size);
1059 		memset(*p, 0, size);
1060 		for (i = 0; i < count; i++)
1061 			((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
1062 		*p += size;
1063 	} else {
1064 		BUG_ON(*p + 4 > end);
1065 		ceph_encode_32(p, 0);
1066 	}
1067 }
1068 
1069 /*
1070  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1071  * to include additional client metadata fields.
1072  */
1073 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1074 {
1075 	struct ceph_msg *msg;
1076 	struct ceph_mds_session_head *h;
1077 	int i = -1;
1078 	int extra_bytes = 0;
1079 	int metadata_key_count = 0;
1080 	struct ceph_options *opt = mdsc->fsc->client->options;
1081 	struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1082 	void *p, *end;
1083 
1084 	const char* metadata[][2] = {
1085 		{"hostname", mdsc->nodename},
1086 		{"kernel_version", init_utsname()->release},
1087 		{"entity_id", opt->name ? : ""},
1088 		{"root", fsopt->server_path ? : "/"},
1089 		{NULL, NULL}
1090 	};
1091 
1092 	/* Calculate serialized length of metadata */
1093 	extra_bytes = 4;  /* map length */
1094 	for (i = 0; metadata[i][0]; ++i) {
1095 		extra_bytes += 8 + strlen(metadata[i][0]) +
1096 			strlen(metadata[i][1]);
1097 		metadata_key_count++;
1098 	}
1099 	/* supported feature */
1100 	extra_bytes += 4 + 8;
1101 
1102 	/* Allocate the message */
1103 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1104 			   GFP_NOFS, false);
1105 	if (!msg) {
1106 		pr_err("create_session_msg ENOMEM creating msg\n");
1107 		return NULL;
1108 	}
1109 	p = msg->front.iov_base;
1110 	end = p + msg->front.iov_len;
1111 
1112 	h = p;
1113 	h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1114 	h->seq = cpu_to_le64(seq);
1115 
1116 	/*
1117 	 * Serialize client metadata into waiting buffer space, using
1118 	 * the format that userspace expects for map<string, string>
1119 	 *
1120 	 * ClientSession messages with metadata are v2
1121 	 */
1122 	msg->hdr.version = cpu_to_le16(3);
1123 	msg->hdr.compat_version = cpu_to_le16(1);
1124 
1125 	/* The write pointer, following the session_head structure */
1126 	p += sizeof(*h);
1127 
1128 	/* Number of entries in the map */
1129 	ceph_encode_32(&p, metadata_key_count);
1130 
1131 	/* Two length-prefixed strings for each entry in the map */
1132 	for (i = 0; metadata[i][0]; ++i) {
1133 		size_t const key_len = strlen(metadata[i][0]);
1134 		size_t const val_len = strlen(metadata[i][1]);
1135 
1136 		ceph_encode_32(&p, key_len);
1137 		memcpy(p, metadata[i][0], key_len);
1138 		p += key_len;
1139 		ceph_encode_32(&p, val_len);
1140 		memcpy(p, metadata[i][1], val_len);
1141 		p += val_len;
1142 	}
1143 
1144 	encode_supported_features(&p, end);
1145 	msg->front.iov_len = p - msg->front.iov_base;
1146 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1147 
1148 	return msg;
1149 }
1150 
1151 /*
1152  * send session open request.
1153  *
1154  * called under mdsc->mutex
1155  */
1156 static int __open_session(struct ceph_mds_client *mdsc,
1157 			  struct ceph_mds_session *session)
1158 {
1159 	struct ceph_msg *msg;
1160 	int mstate;
1161 	int mds = session->s_mds;
1162 
1163 	/* wait for mds to go active? */
1164 	mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1165 	dout("open_session to mds%d (%s)\n", mds,
1166 	     ceph_mds_state_name(mstate));
1167 	session->s_state = CEPH_MDS_SESSION_OPENING;
1168 	session->s_renew_requested = jiffies;
1169 
1170 	/* send connect message */
1171 	msg = create_session_open_msg(mdsc, session->s_seq);
1172 	if (!msg)
1173 		return -ENOMEM;
1174 	ceph_con_send(&session->s_con, msg);
1175 	return 0;
1176 }
1177 
1178 /*
1179  * open sessions for any export targets for the given mds
1180  *
1181  * called under mdsc->mutex
1182  */
1183 static struct ceph_mds_session *
1184 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1185 {
1186 	struct ceph_mds_session *session;
1187 
1188 	session = __ceph_lookup_mds_session(mdsc, target);
1189 	if (!session) {
1190 		session = register_session(mdsc, target);
1191 		if (IS_ERR(session))
1192 			return session;
1193 	}
1194 	if (session->s_state == CEPH_MDS_SESSION_NEW ||
1195 	    session->s_state == CEPH_MDS_SESSION_CLOSING)
1196 		__open_session(mdsc, session);
1197 
1198 	return session;
1199 }
1200 
1201 struct ceph_mds_session *
1202 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1203 {
1204 	struct ceph_mds_session *session;
1205 
1206 	dout("open_export_target_session to mds%d\n", target);
1207 
1208 	mutex_lock(&mdsc->mutex);
1209 	session = __open_export_target_session(mdsc, target);
1210 	mutex_unlock(&mdsc->mutex);
1211 
1212 	return session;
1213 }
1214 
1215 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1216 					  struct ceph_mds_session *session)
1217 {
1218 	struct ceph_mds_info *mi;
1219 	struct ceph_mds_session *ts;
1220 	int i, mds = session->s_mds;
1221 
1222 	if (mds >= mdsc->mdsmap->m_num_mds)
1223 		return;
1224 
1225 	mi = &mdsc->mdsmap->m_info[mds];
1226 	dout("open_export_target_sessions for mds%d (%d targets)\n",
1227 	     session->s_mds, mi->num_export_targets);
1228 
1229 	for (i = 0; i < mi->num_export_targets; i++) {
1230 		ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1231 		if (!IS_ERR(ts))
1232 			ceph_put_mds_session(ts);
1233 	}
1234 }
1235 
1236 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1237 					   struct ceph_mds_session *session)
1238 {
1239 	mutex_lock(&mdsc->mutex);
1240 	__open_export_target_sessions(mdsc, session);
1241 	mutex_unlock(&mdsc->mutex);
1242 }
1243 
1244 /*
1245  * session caps
1246  */
1247 
1248 static void detach_cap_releases(struct ceph_mds_session *session,
1249 				struct list_head *target)
1250 {
1251 	lockdep_assert_held(&session->s_cap_lock);
1252 
1253 	list_splice_init(&session->s_cap_releases, target);
1254 	session->s_num_cap_releases = 0;
1255 	dout("dispose_cap_releases mds%d\n", session->s_mds);
1256 }
1257 
1258 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1259 				 struct list_head *dispose)
1260 {
1261 	while (!list_empty(dispose)) {
1262 		struct ceph_cap *cap;
1263 		/* zero out the in-progress message */
1264 		cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1265 		list_del(&cap->session_caps);
1266 		ceph_put_cap(mdsc, cap);
1267 	}
1268 }
1269 
1270 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1271 				     struct ceph_mds_session *session)
1272 {
1273 	struct ceph_mds_request *req;
1274 	struct rb_node *p;
1275 	struct ceph_inode_info *ci;
1276 
1277 	dout("cleanup_session_requests mds%d\n", session->s_mds);
1278 	mutex_lock(&mdsc->mutex);
1279 	while (!list_empty(&session->s_unsafe)) {
1280 		req = list_first_entry(&session->s_unsafe,
1281 				       struct ceph_mds_request, r_unsafe_item);
1282 		pr_warn_ratelimited(" dropping unsafe request %llu\n",
1283 				    req->r_tid);
1284 		if (req->r_target_inode) {
1285 			/* dropping unsafe change of inode's attributes */
1286 			ci = ceph_inode(req->r_target_inode);
1287 			errseq_set(&ci->i_meta_err, -EIO);
1288 		}
1289 		if (req->r_unsafe_dir) {
1290 			/* dropping unsafe directory operation */
1291 			ci = ceph_inode(req->r_unsafe_dir);
1292 			errseq_set(&ci->i_meta_err, -EIO);
1293 		}
1294 		__unregister_request(mdsc, req);
1295 	}
1296 	/* zero r_attempts, so kick_requests() will re-send requests */
1297 	p = rb_first(&mdsc->request_tree);
1298 	while (p) {
1299 		req = rb_entry(p, struct ceph_mds_request, r_node);
1300 		p = rb_next(p);
1301 		if (req->r_session &&
1302 		    req->r_session->s_mds == session->s_mds)
1303 			req->r_attempts = 0;
1304 	}
1305 	mutex_unlock(&mdsc->mutex);
1306 }
1307 
1308 /*
1309  * Helper to safely iterate over all caps associated with a session, with
1310  * special care taken to handle a racing __ceph_remove_cap().
1311  *
1312  * Caller must hold session s_mutex.
1313  */
1314 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1315 			      int (*cb)(struct inode *, struct ceph_cap *,
1316 					void *), void *arg)
1317 {
1318 	struct list_head *p;
1319 	struct ceph_cap *cap;
1320 	struct inode *inode, *last_inode = NULL;
1321 	struct ceph_cap *old_cap = NULL;
1322 	int ret;
1323 
1324 	dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1325 	spin_lock(&session->s_cap_lock);
1326 	p = session->s_caps.next;
1327 	while (p != &session->s_caps) {
1328 		cap = list_entry(p, struct ceph_cap, session_caps);
1329 		inode = igrab(&cap->ci->vfs_inode);
1330 		if (!inode) {
1331 			p = p->next;
1332 			continue;
1333 		}
1334 		session->s_cap_iterator = cap;
1335 		spin_unlock(&session->s_cap_lock);
1336 
1337 		if (last_inode) {
1338 			/* avoid calling iput_final() while holding
1339 			 * s_mutex or in mds dispatch threads */
1340 			ceph_async_iput(last_inode);
1341 			last_inode = NULL;
1342 		}
1343 		if (old_cap) {
1344 			ceph_put_cap(session->s_mdsc, old_cap);
1345 			old_cap = NULL;
1346 		}
1347 
1348 		ret = cb(inode, cap, arg);
1349 		last_inode = inode;
1350 
1351 		spin_lock(&session->s_cap_lock);
1352 		p = p->next;
1353 		if (!cap->ci) {
1354 			dout("iterate_session_caps  finishing cap %p removal\n",
1355 			     cap);
1356 			BUG_ON(cap->session != session);
1357 			cap->session = NULL;
1358 			list_del_init(&cap->session_caps);
1359 			session->s_nr_caps--;
1360 			if (cap->queue_release)
1361 				__ceph_queue_cap_release(session, cap);
1362 			else
1363 				old_cap = cap;  /* put_cap it w/o locks held */
1364 		}
1365 		if (ret < 0)
1366 			goto out;
1367 	}
1368 	ret = 0;
1369 out:
1370 	session->s_cap_iterator = NULL;
1371 	spin_unlock(&session->s_cap_lock);
1372 
1373 	ceph_async_iput(last_inode);
1374 	if (old_cap)
1375 		ceph_put_cap(session->s_mdsc, old_cap);
1376 
1377 	return ret;
1378 }
1379 
1380 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1381 				  void *arg)
1382 {
1383 	struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1384 	struct ceph_inode_info *ci = ceph_inode(inode);
1385 	LIST_HEAD(to_remove);
1386 	bool dirty_dropped = false;
1387 	bool invalidate = false;
1388 
1389 	dout("removing cap %p, ci is %p, inode is %p\n",
1390 	     cap, ci, &ci->vfs_inode);
1391 	spin_lock(&ci->i_ceph_lock);
1392 	if (cap->mds_wanted | cap->issued)
1393 		ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1394 	__ceph_remove_cap(cap, false);
1395 	if (!ci->i_auth_cap) {
1396 		struct ceph_cap_flush *cf;
1397 		struct ceph_mds_client *mdsc = fsc->mdsc;
1398 
1399 		if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1400 			if (inode->i_data.nrpages > 0)
1401 				invalidate = true;
1402 			if (ci->i_wrbuffer_ref > 0)
1403 				mapping_set_error(&inode->i_data, -EIO);
1404 		}
1405 
1406 		while (!list_empty(&ci->i_cap_flush_list)) {
1407 			cf = list_first_entry(&ci->i_cap_flush_list,
1408 					      struct ceph_cap_flush, i_list);
1409 			list_move(&cf->i_list, &to_remove);
1410 		}
1411 
1412 		spin_lock(&mdsc->cap_dirty_lock);
1413 
1414 		list_for_each_entry(cf, &to_remove, i_list)
1415 			list_del(&cf->g_list);
1416 
1417 		if (!list_empty(&ci->i_dirty_item)) {
1418 			pr_warn_ratelimited(
1419 				" dropping dirty %s state for %p %lld\n",
1420 				ceph_cap_string(ci->i_dirty_caps),
1421 				inode, ceph_ino(inode));
1422 			ci->i_dirty_caps = 0;
1423 			list_del_init(&ci->i_dirty_item);
1424 			dirty_dropped = true;
1425 		}
1426 		if (!list_empty(&ci->i_flushing_item)) {
1427 			pr_warn_ratelimited(
1428 				" dropping dirty+flushing %s state for %p %lld\n",
1429 				ceph_cap_string(ci->i_flushing_caps),
1430 				inode, ceph_ino(inode));
1431 			ci->i_flushing_caps = 0;
1432 			list_del_init(&ci->i_flushing_item);
1433 			mdsc->num_cap_flushing--;
1434 			dirty_dropped = true;
1435 		}
1436 		spin_unlock(&mdsc->cap_dirty_lock);
1437 
1438 		if (dirty_dropped) {
1439 			errseq_set(&ci->i_meta_err, -EIO);
1440 
1441 			if (ci->i_wrbuffer_ref_head == 0 &&
1442 			    ci->i_wr_ref == 0 &&
1443 			    ci->i_dirty_caps == 0 &&
1444 			    ci->i_flushing_caps == 0) {
1445 				ceph_put_snap_context(ci->i_head_snapc);
1446 				ci->i_head_snapc = NULL;
1447 			}
1448 		}
1449 
1450 		if (atomic_read(&ci->i_filelock_ref) > 0) {
1451 			/* make further file lock syscall return -EIO */
1452 			ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1453 			pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1454 					    inode, ceph_ino(inode));
1455 		}
1456 
1457 		if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1458 			list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1459 			ci->i_prealloc_cap_flush = NULL;
1460 		}
1461 	}
1462 	spin_unlock(&ci->i_ceph_lock);
1463 	while (!list_empty(&to_remove)) {
1464 		struct ceph_cap_flush *cf;
1465 		cf = list_first_entry(&to_remove,
1466 				      struct ceph_cap_flush, i_list);
1467 		list_del(&cf->i_list);
1468 		ceph_free_cap_flush(cf);
1469 	}
1470 
1471 	wake_up_all(&ci->i_cap_wq);
1472 	if (invalidate)
1473 		ceph_queue_invalidate(inode);
1474 	if (dirty_dropped)
1475 		iput(inode);
1476 	return 0;
1477 }
1478 
1479 /*
1480  * caller must hold session s_mutex
1481  */
1482 static void remove_session_caps(struct ceph_mds_session *session)
1483 {
1484 	struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1485 	struct super_block *sb = fsc->sb;
1486 	LIST_HEAD(dispose);
1487 
1488 	dout("remove_session_caps on %p\n", session);
1489 	ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1490 
1491 	wake_up_all(&fsc->mdsc->cap_flushing_wq);
1492 
1493 	spin_lock(&session->s_cap_lock);
1494 	if (session->s_nr_caps > 0) {
1495 		struct inode *inode;
1496 		struct ceph_cap *cap, *prev = NULL;
1497 		struct ceph_vino vino;
1498 		/*
1499 		 * iterate_session_caps() skips inodes that are being
1500 		 * deleted, we need to wait until deletions are complete.
1501 		 * __wait_on_freeing_inode() is designed for the job,
1502 		 * but it is not exported, so use lookup inode function
1503 		 * to access it.
1504 		 */
1505 		while (!list_empty(&session->s_caps)) {
1506 			cap = list_entry(session->s_caps.next,
1507 					 struct ceph_cap, session_caps);
1508 			if (cap == prev)
1509 				break;
1510 			prev = cap;
1511 			vino = cap->ci->i_vino;
1512 			spin_unlock(&session->s_cap_lock);
1513 
1514 			inode = ceph_find_inode(sb, vino);
1515 			 /* avoid calling iput_final() while holding s_mutex */
1516 			ceph_async_iput(inode);
1517 
1518 			spin_lock(&session->s_cap_lock);
1519 		}
1520 	}
1521 
1522 	// drop cap expires and unlock s_cap_lock
1523 	detach_cap_releases(session, &dispose);
1524 
1525 	BUG_ON(session->s_nr_caps > 0);
1526 	BUG_ON(!list_empty(&session->s_cap_flushing));
1527 	spin_unlock(&session->s_cap_lock);
1528 	dispose_cap_releases(session->s_mdsc, &dispose);
1529 }
1530 
1531 enum {
1532 	RECONNECT,
1533 	RENEWCAPS,
1534 	FORCE_RO,
1535 };
1536 
1537 /*
1538  * wake up any threads waiting on this session's caps.  if the cap is
1539  * old (didn't get renewed on the client reconnect), remove it now.
1540  *
1541  * caller must hold s_mutex.
1542  */
1543 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1544 			      void *arg)
1545 {
1546 	struct ceph_inode_info *ci = ceph_inode(inode);
1547 	unsigned long ev = (unsigned long)arg;
1548 
1549 	if (ev == RECONNECT) {
1550 		spin_lock(&ci->i_ceph_lock);
1551 		ci->i_wanted_max_size = 0;
1552 		ci->i_requested_max_size = 0;
1553 		spin_unlock(&ci->i_ceph_lock);
1554 	} else if (ev == RENEWCAPS) {
1555 		if (cap->cap_gen < cap->session->s_cap_gen) {
1556 			/* mds did not re-issue stale cap */
1557 			spin_lock(&ci->i_ceph_lock);
1558 			cap->issued = cap->implemented = CEPH_CAP_PIN;
1559 			/* make sure mds knows what we want */
1560 			if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1561 				ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1562 			spin_unlock(&ci->i_ceph_lock);
1563 		}
1564 	} else if (ev == FORCE_RO) {
1565 	}
1566 	wake_up_all(&ci->i_cap_wq);
1567 	return 0;
1568 }
1569 
1570 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1571 {
1572 	dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1573 	ceph_iterate_session_caps(session, wake_up_session_cb,
1574 				  (void *)(unsigned long)ev);
1575 }
1576 
1577 /*
1578  * Send periodic message to MDS renewing all currently held caps.  The
1579  * ack will reset the expiration for all caps from this session.
1580  *
1581  * caller holds s_mutex
1582  */
1583 static int send_renew_caps(struct ceph_mds_client *mdsc,
1584 			   struct ceph_mds_session *session)
1585 {
1586 	struct ceph_msg *msg;
1587 	int state;
1588 
1589 	if (time_after_eq(jiffies, session->s_cap_ttl) &&
1590 	    time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1591 		pr_info("mds%d caps stale\n", session->s_mds);
1592 	session->s_renew_requested = jiffies;
1593 
1594 	/* do not try to renew caps until a recovering mds has reconnected
1595 	 * with its clients. */
1596 	state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1597 	if (state < CEPH_MDS_STATE_RECONNECT) {
1598 		dout("send_renew_caps ignoring mds%d (%s)\n",
1599 		     session->s_mds, ceph_mds_state_name(state));
1600 		return 0;
1601 	}
1602 
1603 	dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1604 		ceph_mds_state_name(state));
1605 	msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1606 				 ++session->s_renew_seq);
1607 	if (!msg)
1608 		return -ENOMEM;
1609 	ceph_con_send(&session->s_con, msg);
1610 	return 0;
1611 }
1612 
1613 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1614 			     struct ceph_mds_session *session, u64 seq)
1615 {
1616 	struct ceph_msg *msg;
1617 
1618 	dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1619 	     session->s_mds, ceph_session_state_name(session->s_state), seq);
1620 	msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1621 	if (!msg)
1622 		return -ENOMEM;
1623 	ceph_con_send(&session->s_con, msg);
1624 	return 0;
1625 }
1626 
1627 
1628 /*
1629  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1630  *
1631  * Called under session->s_mutex
1632  */
1633 static void renewed_caps(struct ceph_mds_client *mdsc,
1634 			 struct ceph_mds_session *session, int is_renew)
1635 {
1636 	int was_stale;
1637 	int wake = 0;
1638 
1639 	spin_lock(&session->s_cap_lock);
1640 	was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1641 
1642 	session->s_cap_ttl = session->s_renew_requested +
1643 		mdsc->mdsmap->m_session_timeout*HZ;
1644 
1645 	if (was_stale) {
1646 		if (time_before(jiffies, session->s_cap_ttl)) {
1647 			pr_info("mds%d caps renewed\n", session->s_mds);
1648 			wake = 1;
1649 		} else {
1650 			pr_info("mds%d caps still stale\n", session->s_mds);
1651 		}
1652 	}
1653 	dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1654 	     session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1655 	     time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1656 	spin_unlock(&session->s_cap_lock);
1657 
1658 	if (wake)
1659 		wake_up_session_caps(session, RENEWCAPS);
1660 }
1661 
1662 /*
1663  * send a session close request
1664  */
1665 static int request_close_session(struct ceph_mds_client *mdsc,
1666 				 struct ceph_mds_session *session)
1667 {
1668 	struct ceph_msg *msg;
1669 
1670 	dout("request_close_session mds%d state %s seq %lld\n",
1671 	     session->s_mds, ceph_session_state_name(session->s_state),
1672 	     session->s_seq);
1673 	msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1674 	if (!msg)
1675 		return -ENOMEM;
1676 	ceph_con_send(&session->s_con, msg);
1677 	return 1;
1678 }
1679 
1680 /*
1681  * Called with s_mutex held.
1682  */
1683 static int __close_session(struct ceph_mds_client *mdsc,
1684 			 struct ceph_mds_session *session)
1685 {
1686 	if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1687 		return 0;
1688 	session->s_state = CEPH_MDS_SESSION_CLOSING;
1689 	return request_close_session(mdsc, session);
1690 }
1691 
1692 static bool drop_negative_children(struct dentry *dentry)
1693 {
1694 	struct dentry *child;
1695 	bool all_negative = true;
1696 
1697 	if (!d_is_dir(dentry))
1698 		goto out;
1699 
1700 	spin_lock(&dentry->d_lock);
1701 	list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1702 		if (d_really_is_positive(child)) {
1703 			all_negative = false;
1704 			break;
1705 		}
1706 	}
1707 	spin_unlock(&dentry->d_lock);
1708 
1709 	if (all_negative)
1710 		shrink_dcache_parent(dentry);
1711 out:
1712 	return all_negative;
1713 }
1714 
1715 /*
1716  * Trim old(er) caps.
1717  *
1718  * Because we can't cache an inode without one or more caps, we do
1719  * this indirectly: if a cap is unused, we prune its aliases, at which
1720  * point the inode will hopefully get dropped to.
1721  *
1722  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1723  * memory pressure from the MDS, though, so it needn't be perfect.
1724  */
1725 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1726 {
1727 	int *remaining = arg;
1728 	struct ceph_inode_info *ci = ceph_inode(inode);
1729 	int used, wanted, oissued, mine;
1730 
1731 	if (*remaining <= 0)
1732 		return -1;
1733 
1734 	spin_lock(&ci->i_ceph_lock);
1735 	mine = cap->issued | cap->implemented;
1736 	used = __ceph_caps_used(ci);
1737 	wanted = __ceph_caps_file_wanted(ci);
1738 	oissued = __ceph_caps_issued_other(ci, cap);
1739 
1740 	dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1741 	     inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1742 	     ceph_cap_string(used), ceph_cap_string(wanted));
1743 	if (cap == ci->i_auth_cap) {
1744 		if (ci->i_dirty_caps || ci->i_flushing_caps ||
1745 		    !list_empty(&ci->i_cap_snaps))
1746 			goto out;
1747 		if ((used | wanted) & CEPH_CAP_ANY_WR)
1748 			goto out;
1749 		/* Note: it's possible that i_filelock_ref becomes non-zero
1750 		 * after dropping auth caps. It doesn't hurt because reply
1751 		 * of lock mds request will re-add auth caps. */
1752 		if (atomic_read(&ci->i_filelock_ref) > 0)
1753 			goto out;
1754 	}
1755 	/* The inode has cached pages, but it's no longer used.
1756 	 * we can safely drop it */
1757 	if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1758 	    !(oissued & CEPH_CAP_FILE_CACHE)) {
1759 	  used = 0;
1760 	  oissued = 0;
1761 	}
1762 	if ((used | wanted) & ~oissued & mine)
1763 		goto out;   /* we need these caps */
1764 
1765 	if (oissued) {
1766 		/* we aren't the only cap.. just remove us */
1767 		__ceph_remove_cap(cap, true);
1768 		(*remaining)--;
1769 	} else {
1770 		struct dentry *dentry;
1771 		/* try dropping referring dentries */
1772 		spin_unlock(&ci->i_ceph_lock);
1773 		dentry = d_find_any_alias(inode);
1774 		if (dentry && drop_negative_children(dentry)) {
1775 			int count;
1776 			dput(dentry);
1777 			d_prune_aliases(inode);
1778 			count = atomic_read(&inode->i_count);
1779 			if (count == 1)
1780 				(*remaining)--;
1781 			dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1782 			     inode, cap, count);
1783 		} else {
1784 			dput(dentry);
1785 		}
1786 		return 0;
1787 	}
1788 
1789 out:
1790 	spin_unlock(&ci->i_ceph_lock);
1791 	return 0;
1792 }
1793 
1794 /*
1795  * Trim session cap count down to some max number.
1796  */
1797 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1798 		   struct ceph_mds_session *session,
1799 		   int max_caps)
1800 {
1801 	int trim_caps = session->s_nr_caps - max_caps;
1802 
1803 	dout("trim_caps mds%d start: %d / %d, trim %d\n",
1804 	     session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1805 	if (trim_caps > 0) {
1806 		int remaining = trim_caps;
1807 
1808 		ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1809 		dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1810 		     session->s_mds, session->s_nr_caps, max_caps,
1811 			trim_caps - remaining);
1812 	}
1813 
1814 	ceph_flush_cap_releases(mdsc, session);
1815 	return 0;
1816 }
1817 
1818 static int check_caps_flush(struct ceph_mds_client *mdsc,
1819 			    u64 want_flush_tid)
1820 {
1821 	int ret = 1;
1822 
1823 	spin_lock(&mdsc->cap_dirty_lock);
1824 	if (!list_empty(&mdsc->cap_flush_list)) {
1825 		struct ceph_cap_flush *cf =
1826 			list_first_entry(&mdsc->cap_flush_list,
1827 					 struct ceph_cap_flush, g_list);
1828 		if (cf->tid <= want_flush_tid) {
1829 			dout("check_caps_flush still flushing tid "
1830 			     "%llu <= %llu\n", cf->tid, want_flush_tid);
1831 			ret = 0;
1832 		}
1833 	}
1834 	spin_unlock(&mdsc->cap_dirty_lock);
1835 	return ret;
1836 }
1837 
1838 /*
1839  * flush all dirty inode data to disk.
1840  *
1841  * returns true if we've flushed through want_flush_tid
1842  */
1843 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1844 			    u64 want_flush_tid)
1845 {
1846 	dout("check_caps_flush want %llu\n", want_flush_tid);
1847 
1848 	wait_event(mdsc->cap_flushing_wq,
1849 		   check_caps_flush(mdsc, want_flush_tid));
1850 
1851 	dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1852 }
1853 
1854 /*
1855  * called under s_mutex
1856  */
1857 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1858 				   struct ceph_mds_session *session)
1859 {
1860 	struct ceph_msg *msg = NULL;
1861 	struct ceph_mds_cap_release *head;
1862 	struct ceph_mds_cap_item *item;
1863 	struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1864 	struct ceph_cap *cap;
1865 	LIST_HEAD(tmp_list);
1866 	int num_cap_releases;
1867 	__le32	barrier, *cap_barrier;
1868 
1869 	down_read(&osdc->lock);
1870 	barrier = cpu_to_le32(osdc->epoch_barrier);
1871 	up_read(&osdc->lock);
1872 
1873 	spin_lock(&session->s_cap_lock);
1874 again:
1875 	list_splice_init(&session->s_cap_releases, &tmp_list);
1876 	num_cap_releases = session->s_num_cap_releases;
1877 	session->s_num_cap_releases = 0;
1878 	spin_unlock(&session->s_cap_lock);
1879 
1880 	while (!list_empty(&tmp_list)) {
1881 		if (!msg) {
1882 			msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1883 					PAGE_SIZE, GFP_NOFS, false);
1884 			if (!msg)
1885 				goto out_err;
1886 			head = msg->front.iov_base;
1887 			head->num = cpu_to_le32(0);
1888 			msg->front.iov_len = sizeof(*head);
1889 
1890 			msg->hdr.version = cpu_to_le16(2);
1891 			msg->hdr.compat_version = cpu_to_le16(1);
1892 		}
1893 
1894 		cap = list_first_entry(&tmp_list, struct ceph_cap,
1895 					session_caps);
1896 		list_del(&cap->session_caps);
1897 		num_cap_releases--;
1898 
1899 		head = msg->front.iov_base;
1900 		put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1901 				   &head->num);
1902 		item = msg->front.iov_base + msg->front.iov_len;
1903 		item->ino = cpu_to_le64(cap->cap_ino);
1904 		item->cap_id = cpu_to_le64(cap->cap_id);
1905 		item->migrate_seq = cpu_to_le32(cap->mseq);
1906 		item->seq = cpu_to_le32(cap->issue_seq);
1907 		msg->front.iov_len += sizeof(*item);
1908 
1909 		ceph_put_cap(mdsc, cap);
1910 
1911 		if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1912 			// Append cap_barrier field
1913 			cap_barrier = msg->front.iov_base + msg->front.iov_len;
1914 			*cap_barrier = barrier;
1915 			msg->front.iov_len += sizeof(*cap_barrier);
1916 
1917 			msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1918 			dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1919 			ceph_con_send(&session->s_con, msg);
1920 			msg = NULL;
1921 		}
1922 	}
1923 
1924 	BUG_ON(num_cap_releases != 0);
1925 
1926 	spin_lock(&session->s_cap_lock);
1927 	if (!list_empty(&session->s_cap_releases))
1928 		goto again;
1929 	spin_unlock(&session->s_cap_lock);
1930 
1931 	if (msg) {
1932 		// Append cap_barrier field
1933 		cap_barrier = msg->front.iov_base + msg->front.iov_len;
1934 		*cap_barrier = barrier;
1935 		msg->front.iov_len += sizeof(*cap_barrier);
1936 
1937 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1938 		dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1939 		ceph_con_send(&session->s_con, msg);
1940 	}
1941 	return;
1942 out_err:
1943 	pr_err("send_cap_releases mds%d, failed to allocate message\n",
1944 		session->s_mds);
1945 	spin_lock(&session->s_cap_lock);
1946 	list_splice(&tmp_list, &session->s_cap_releases);
1947 	session->s_num_cap_releases += num_cap_releases;
1948 	spin_unlock(&session->s_cap_lock);
1949 }
1950 
1951 static void ceph_cap_release_work(struct work_struct *work)
1952 {
1953 	struct ceph_mds_session *session =
1954 		container_of(work, struct ceph_mds_session, s_cap_release_work);
1955 
1956 	mutex_lock(&session->s_mutex);
1957 	if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1958 	    session->s_state == CEPH_MDS_SESSION_HUNG)
1959 		ceph_send_cap_releases(session->s_mdsc, session);
1960 	mutex_unlock(&session->s_mutex);
1961 	ceph_put_mds_session(session);
1962 }
1963 
1964 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1965 		             struct ceph_mds_session *session)
1966 {
1967 	if (mdsc->stopping)
1968 		return;
1969 
1970 	get_session(session);
1971 	if (queue_work(mdsc->fsc->cap_wq,
1972 		       &session->s_cap_release_work)) {
1973 		dout("cap release work queued\n");
1974 	} else {
1975 		ceph_put_mds_session(session);
1976 		dout("failed to queue cap release work\n");
1977 	}
1978 }
1979 
1980 /*
1981  * caller holds session->s_cap_lock
1982  */
1983 void __ceph_queue_cap_release(struct ceph_mds_session *session,
1984 			      struct ceph_cap *cap)
1985 {
1986 	list_add_tail(&cap->session_caps, &session->s_cap_releases);
1987 	session->s_num_cap_releases++;
1988 
1989 	if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
1990 		ceph_flush_cap_releases(session->s_mdsc, session);
1991 }
1992 
1993 static void ceph_cap_reclaim_work(struct work_struct *work)
1994 {
1995 	struct ceph_mds_client *mdsc =
1996 		container_of(work, struct ceph_mds_client, cap_reclaim_work);
1997 	int ret = ceph_trim_dentries(mdsc);
1998 	if (ret == -EAGAIN)
1999 		ceph_queue_cap_reclaim_work(mdsc);
2000 }
2001 
2002 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2003 {
2004 	if (mdsc->stopping)
2005 		return;
2006 
2007         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2008                 dout("caps reclaim work queued\n");
2009         } else {
2010                 dout("failed to queue caps release work\n");
2011         }
2012 }
2013 
2014 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2015 {
2016 	int val;
2017 	if (!nr)
2018 		return;
2019 	val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2020 	if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2021 		atomic_set(&mdsc->cap_reclaim_pending, 0);
2022 		ceph_queue_cap_reclaim_work(mdsc);
2023 	}
2024 }
2025 
2026 /*
2027  * requests
2028  */
2029 
2030 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2031 				    struct inode *dir)
2032 {
2033 	struct ceph_inode_info *ci = ceph_inode(dir);
2034 	struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2035 	struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2036 	size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2037 	unsigned int num_entries;
2038 	int order;
2039 
2040 	spin_lock(&ci->i_ceph_lock);
2041 	num_entries = ci->i_files + ci->i_subdirs;
2042 	spin_unlock(&ci->i_ceph_lock);
2043 	num_entries = max(num_entries, 1U);
2044 	num_entries = min(num_entries, opt->max_readdir);
2045 
2046 	order = get_order(size * num_entries);
2047 	while (order >= 0) {
2048 		rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2049 							     __GFP_NOWARN,
2050 							     order);
2051 		if (rinfo->dir_entries)
2052 			break;
2053 		order--;
2054 	}
2055 	if (!rinfo->dir_entries)
2056 		return -ENOMEM;
2057 
2058 	num_entries = (PAGE_SIZE << order) / size;
2059 	num_entries = min(num_entries, opt->max_readdir);
2060 
2061 	rinfo->dir_buf_size = PAGE_SIZE << order;
2062 	req->r_num_caps = num_entries + 1;
2063 	req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2064 	req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2065 	return 0;
2066 }
2067 
2068 /*
2069  * Create an mds request.
2070  */
2071 struct ceph_mds_request *
2072 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2073 {
2074 	struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2075 	struct timespec64 ts;
2076 
2077 	if (!req)
2078 		return ERR_PTR(-ENOMEM);
2079 
2080 	mutex_init(&req->r_fill_mutex);
2081 	req->r_mdsc = mdsc;
2082 	req->r_started = jiffies;
2083 	req->r_resend_mds = -1;
2084 	INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2085 	INIT_LIST_HEAD(&req->r_unsafe_target_item);
2086 	req->r_fmode = -1;
2087 	kref_init(&req->r_kref);
2088 	RB_CLEAR_NODE(&req->r_node);
2089 	INIT_LIST_HEAD(&req->r_wait);
2090 	init_completion(&req->r_completion);
2091 	init_completion(&req->r_safe_completion);
2092 	INIT_LIST_HEAD(&req->r_unsafe_item);
2093 
2094 	ktime_get_coarse_real_ts64(&ts);
2095 	req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
2096 
2097 	req->r_op = op;
2098 	req->r_direct_mode = mode;
2099 	return req;
2100 }
2101 
2102 /*
2103  * return oldest (lowest) request, tid in request tree, 0 if none.
2104  *
2105  * called under mdsc->mutex.
2106  */
2107 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2108 {
2109 	if (RB_EMPTY_ROOT(&mdsc->request_tree))
2110 		return NULL;
2111 	return rb_entry(rb_first(&mdsc->request_tree),
2112 			struct ceph_mds_request, r_node);
2113 }
2114 
2115 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2116 {
2117 	return mdsc->oldest_tid;
2118 }
2119 
2120 /*
2121  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2122  * on build_path_from_dentry in fs/cifs/dir.c.
2123  *
2124  * If @stop_on_nosnap, generate path relative to the first non-snapped
2125  * inode.
2126  *
2127  * Encode hidden .snap dirs as a double /, i.e.
2128  *   foo/.snap/bar -> foo//bar
2129  */
2130 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2131 			   int stop_on_nosnap)
2132 {
2133 	struct dentry *temp;
2134 	char *path;
2135 	int pos;
2136 	unsigned seq;
2137 	u64 base;
2138 
2139 	if (!dentry)
2140 		return ERR_PTR(-EINVAL);
2141 
2142 	path = __getname();
2143 	if (!path)
2144 		return ERR_PTR(-ENOMEM);
2145 retry:
2146 	pos = PATH_MAX - 1;
2147 	path[pos] = '\0';
2148 
2149 	seq = read_seqbegin(&rename_lock);
2150 	rcu_read_lock();
2151 	temp = dentry;
2152 	for (;;) {
2153 		struct inode *inode;
2154 
2155 		spin_lock(&temp->d_lock);
2156 		inode = d_inode(temp);
2157 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2158 			dout("build_path path+%d: %p SNAPDIR\n",
2159 			     pos, temp);
2160 		} else if (stop_on_nosnap && inode && dentry != temp &&
2161 			   ceph_snap(inode) == CEPH_NOSNAP) {
2162 			spin_unlock(&temp->d_lock);
2163 			pos++; /* get rid of any prepended '/' */
2164 			break;
2165 		} else {
2166 			pos -= temp->d_name.len;
2167 			if (pos < 0) {
2168 				spin_unlock(&temp->d_lock);
2169 				break;
2170 			}
2171 			memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2172 		}
2173 		spin_unlock(&temp->d_lock);
2174 		temp = READ_ONCE(temp->d_parent);
2175 
2176 		/* Are we at the root? */
2177 		if (IS_ROOT(temp))
2178 			break;
2179 
2180 		/* Are we out of buffer? */
2181 		if (--pos < 0)
2182 			break;
2183 
2184 		path[pos] = '/';
2185 	}
2186 	base = ceph_ino(d_inode(temp));
2187 	rcu_read_unlock();
2188 
2189 	if (read_seqretry(&rename_lock, seq))
2190 		goto retry;
2191 
2192 	if (pos < 0) {
2193 		/*
2194 		 * A rename didn't occur, but somehow we didn't end up where
2195 		 * we thought we would. Throw a warning and try again.
2196 		 */
2197 		pr_warn("build_path did not end path lookup where "
2198 			"expected, pos is %d\n", pos);
2199 		goto retry;
2200 	}
2201 
2202 	*pbase = base;
2203 	*plen = PATH_MAX - 1 - pos;
2204 	dout("build_path on %p %d built %llx '%.*s'\n",
2205 	     dentry, d_count(dentry), base, *plen, path + pos);
2206 	return path + pos;
2207 }
2208 
2209 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2210 			     const char **ppath, int *ppathlen, u64 *pino,
2211 			     bool *pfreepath, bool parent_locked)
2212 {
2213 	char *path;
2214 
2215 	rcu_read_lock();
2216 	if (!dir)
2217 		dir = d_inode_rcu(dentry->d_parent);
2218 	if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2219 		*pino = ceph_ino(dir);
2220 		rcu_read_unlock();
2221 		*ppath = dentry->d_name.name;
2222 		*ppathlen = dentry->d_name.len;
2223 		return 0;
2224 	}
2225 	rcu_read_unlock();
2226 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2227 	if (IS_ERR(path))
2228 		return PTR_ERR(path);
2229 	*ppath = path;
2230 	*pfreepath = true;
2231 	return 0;
2232 }
2233 
2234 static int build_inode_path(struct inode *inode,
2235 			    const char **ppath, int *ppathlen, u64 *pino,
2236 			    bool *pfreepath)
2237 {
2238 	struct dentry *dentry;
2239 	char *path;
2240 
2241 	if (ceph_snap(inode) == CEPH_NOSNAP) {
2242 		*pino = ceph_ino(inode);
2243 		*ppathlen = 0;
2244 		return 0;
2245 	}
2246 	dentry = d_find_alias(inode);
2247 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2248 	dput(dentry);
2249 	if (IS_ERR(path))
2250 		return PTR_ERR(path);
2251 	*ppath = path;
2252 	*pfreepath = true;
2253 	return 0;
2254 }
2255 
2256 /*
2257  * request arguments may be specified via an inode *, a dentry *, or
2258  * an explicit ino+path.
2259  */
2260 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2261 				  struct inode *rdiri, const char *rpath,
2262 				  u64 rino, const char **ppath, int *pathlen,
2263 				  u64 *ino, bool *freepath, bool parent_locked)
2264 {
2265 	int r = 0;
2266 
2267 	if (rinode) {
2268 		r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2269 		dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2270 		     ceph_snap(rinode));
2271 	} else if (rdentry) {
2272 		r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2273 					freepath, parent_locked);
2274 		dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2275 		     *ppath);
2276 	} else if (rpath || rino) {
2277 		*ino = rino;
2278 		*ppath = rpath;
2279 		*pathlen = rpath ? strlen(rpath) : 0;
2280 		dout(" path %.*s\n", *pathlen, rpath);
2281 	}
2282 
2283 	return r;
2284 }
2285 
2286 /*
2287  * called under mdsc->mutex
2288  */
2289 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2290 					       struct ceph_mds_request *req,
2291 					       int mds, bool drop_cap_releases)
2292 {
2293 	struct ceph_msg *msg;
2294 	struct ceph_mds_request_head *head;
2295 	const char *path1 = NULL;
2296 	const char *path2 = NULL;
2297 	u64 ino1 = 0, ino2 = 0;
2298 	int pathlen1 = 0, pathlen2 = 0;
2299 	bool freepath1 = false, freepath2 = false;
2300 	int len;
2301 	u16 releases;
2302 	void *p, *end;
2303 	int ret;
2304 
2305 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
2306 			      req->r_parent, req->r_path1, req->r_ino1.ino,
2307 			      &path1, &pathlen1, &ino1, &freepath1,
2308 			      test_bit(CEPH_MDS_R_PARENT_LOCKED,
2309 					&req->r_req_flags));
2310 	if (ret < 0) {
2311 		msg = ERR_PTR(ret);
2312 		goto out;
2313 	}
2314 
2315 	/* If r_old_dentry is set, then assume that its parent is locked */
2316 	ret = set_request_path_attr(NULL, req->r_old_dentry,
2317 			      req->r_old_dentry_dir,
2318 			      req->r_path2, req->r_ino2.ino,
2319 			      &path2, &pathlen2, &ino2, &freepath2, true);
2320 	if (ret < 0) {
2321 		msg = ERR_PTR(ret);
2322 		goto out_free1;
2323 	}
2324 
2325 	len = sizeof(*head) +
2326 		pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2327 		sizeof(struct ceph_timespec);
2328 
2329 	/* calculate (max) length for cap releases */
2330 	len += sizeof(struct ceph_mds_request_release) *
2331 		(!!req->r_inode_drop + !!req->r_dentry_drop +
2332 		 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2333 	if (req->r_dentry_drop)
2334 		len += pathlen1;
2335 	if (req->r_old_dentry_drop)
2336 		len += pathlen2;
2337 
2338 	msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2339 	if (!msg) {
2340 		msg = ERR_PTR(-ENOMEM);
2341 		goto out_free2;
2342 	}
2343 
2344 	msg->hdr.version = cpu_to_le16(2);
2345 	msg->hdr.tid = cpu_to_le64(req->r_tid);
2346 
2347 	head = msg->front.iov_base;
2348 	p = msg->front.iov_base + sizeof(*head);
2349 	end = msg->front.iov_base + msg->front.iov_len;
2350 
2351 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2352 	head->op = cpu_to_le32(req->r_op);
2353 	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2354 	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2355 	head->ino = 0;
2356 	head->args = req->r_args;
2357 
2358 	ceph_encode_filepath(&p, end, ino1, path1);
2359 	ceph_encode_filepath(&p, end, ino2, path2);
2360 
2361 	/* make note of release offset, in case we need to replay */
2362 	req->r_request_release_offset = p - msg->front.iov_base;
2363 
2364 	/* cap releases */
2365 	releases = 0;
2366 	if (req->r_inode_drop)
2367 		releases += ceph_encode_inode_release(&p,
2368 		      req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2369 		      mds, req->r_inode_drop, req->r_inode_unless, 0);
2370 	if (req->r_dentry_drop)
2371 		releases += ceph_encode_dentry_release(&p, req->r_dentry,
2372 				req->r_parent, mds, req->r_dentry_drop,
2373 				req->r_dentry_unless);
2374 	if (req->r_old_dentry_drop)
2375 		releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2376 				req->r_old_dentry_dir, mds,
2377 				req->r_old_dentry_drop,
2378 				req->r_old_dentry_unless);
2379 	if (req->r_old_inode_drop)
2380 		releases += ceph_encode_inode_release(&p,
2381 		      d_inode(req->r_old_dentry),
2382 		      mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2383 
2384 	if (drop_cap_releases) {
2385 		releases = 0;
2386 		p = msg->front.iov_base + req->r_request_release_offset;
2387 	}
2388 
2389 	head->num_releases = cpu_to_le16(releases);
2390 
2391 	/* time stamp */
2392 	{
2393 		struct ceph_timespec ts;
2394 		ceph_encode_timespec64(&ts, &req->r_stamp);
2395 		ceph_encode_copy(&p, &ts, sizeof(ts));
2396 	}
2397 
2398 	BUG_ON(p > end);
2399 	msg->front.iov_len = p - msg->front.iov_base;
2400 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2401 
2402 	if (req->r_pagelist) {
2403 		struct ceph_pagelist *pagelist = req->r_pagelist;
2404 		ceph_msg_data_add_pagelist(msg, pagelist);
2405 		msg->hdr.data_len = cpu_to_le32(pagelist->length);
2406 	} else {
2407 		msg->hdr.data_len = 0;
2408 	}
2409 
2410 	msg->hdr.data_off = cpu_to_le16(0);
2411 
2412 out_free2:
2413 	if (freepath2)
2414 		ceph_mdsc_free_path((char *)path2, pathlen2);
2415 out_free1:
2416 	if (freepath1)
2417 		ceph_mdsc_free_path((char *)path1, pathlen1);
2418 out:
2419 	return msg;
2420 }
2421 
2422 /*
2423  * called under mdsc->mutex if error, under no mutex if
2424  * success.
2425  */
2426 static void complete_request(struct ceph_mds_client *mdsc,
2427 			     struct ceph_mds_request *req)
2428 {
2429 	if (req->r_callback)
2430 		req->r_callback(mdsc, req);
2431 	complete_all(&req->r_completion);
2432 }
2433 
2434 /*
2435  * called under mdsc->mutex
2436  */
2437 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2438 				  struct ceph_mds_request *req,
2439 				  int mds, bool drop_cap_releases)
2440 {
2441 	struct ceph_mds_request_head *rhead;
2442 	struct ceph_msg *msg;
2443 	int flags = 0;
2444 
2445 	req->r_attempts++;
2446 	if (req->r_inode) {
2447 		struct ceph_cap *cap =
2448 			ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2449 
2450 		if (cap)
2451 			req->r_sent_on_mseq = cap->mseq;
2452 		else
2453 			req->r_sent_on_mseq = -1;
2454 	}
2455 	dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2456 	     req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2457 
2458 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2459 		void *p;
2460 		/*
2461 		 * Replay.  Do not regenerate message (and rebuild
2462 		 * paths, etc.); just use the original message.
2463 		 * Rebuilding paths will break for renames because
2464 		 * d_move mangles the src name.
2465 		 */
2466 		msg = req->r_request;
2467 		rhead = msg->front.iov_base;
2468 
2469 		flags = le32_to_cpu(rhead->flags);
2470 		flags |= CEPH_MDS_FLAG_REPLAY;
2471 		rhead->flags = cpu_to_le32(flags);
2472 
2473 		if (req->r_target_inode)
2474 			rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2475 
2476 		rhead->num_retry = req->r_attempts - 1;
2477 
2478 		/* remove cap/dentry releases from message */
2479 		rhead->num_releases = 0;
2480 
2481 		/* time stamp */
2482 		p = msg->front.iov_base + req->r_request_release_offset;
2483 		{
2484 			struct ceph_timespec ts;
2485 			ceph_encode_timespec64(&ts, &req->r_stamp);
2486 			ceph_encode_copy(&p, &ts, sizeof(ts));
2487 		}
2488 
2489 		msg->front.iov_len = p - msg->front.iov_base;
2490 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2491 		return 0;
2492 	}
2493 
2494 	if (req->r_request) {
2495 		ceph_msg_put(req->r_request);
2496 		req->r_request = NULL;
2497 	}
2498 	msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2499 	if (IS_ERR(msg)) {
2500 		req->r_err = PTR_ERR(msg);
2501 		return PTR_ERR(msg);
2502 	}
2503 	req->r_request = msg;
2504 
2505 	rhead = msg->front.iov_base;
2506 	rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2507 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2508 		flags |= CEPH_MDS_FLAG_REPLAY;
2509 	if (req->r_parent)
2510 		flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2511 	rhead->flags = cpu_to_le32(flags);
2512 	rhead->num_fwd = req->r_num_fwd;
2513 	rhead->num_retry = req->r_attempts - 1;
2514 	rhead->ino = 0;
2515 
2516 	dout(" r_parent = %p\n", req->r_parent);
2517 	return 0;
2518 }
2519 
2520 /*
2521  * send request, or put it on the appropriate wait list.
2522  */
2523 static void __do_request(struct ceph_mds_client *mdsc,
2524 			struct ceph_mds_request *req)
2525 {
2526 	struct ceph_mds_session *session = NULL;
2527 	int mds = -1;
2528 	int err = 0;
2529 
2530 	if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2531 		if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2532 			__unregister_request(mdsc, req);
2533 		return;
2534 	}
2535 
2536 	if (req->r_timeout &&
2537 	    time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2538 		dout("do_request timed out\n");
2539 		err = -EIO;
2540 		goto finish;
2541 	}
2542 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2543 		dout("do_request forced umount\n");
2544 		err = -EIO;
2545 		goto finish;
2546 	}
2547 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2548 		if (mdsc->mdsmap_err) {
2549 			err = mdsc->mdsmap_err;
2550 			dout("do_request mdsmap err %d\n", err);
2551 			goto finish;
2552 		}
2553 		if (mdsc->mdsmap->m_epoch == 0) {
2554 			dout("do_request no mdsmap, waiting for map\n");
2555 			list_add(&req->r_wait, &mdsc->waiting_for_map);
2556 			return;
2557 		}
2558 		if (!(mdsc->fsc->mount_options->flags &
2559 		      CEPH_MOUNT_OPT_MOUNTWAIT) &&
2560 		    !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2561 			err = -ENOENT;
2562 			pr_info("probably no mds server is up\n");
2563 			goto finish;
2564 		}
2565 	}
2566 
2567 	put_request_session(req);
2568 
2569 	mds = __choose_mds(mdsc, req);
2570 	if (mds < 0 ||
2571 	    ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2572 		dout("do_request no mds or not active, waiting for map\n");
2573 		list_add(&req->r_wait, &mdsc->waiting_for_map);
2574 		return;
2575 	}
2576 
2577 	/* get, open session */
2578 	session = __ceph_lookup_mds_session(mdsc, mds);
2579 	if (!session) {
2580 		session = register_session(mdsc, mds);
2581 		if (IS_ERR(session)) {
2582 			err = PTR_ERR(session);
2583 			goto finish;
2584 		}
2585 	}
2586 	req->r_session = get_session(session);
2587 
2588 	dout("do_request mds%d session %p state %s\n", mds, session,
2589 	     ceph_session_state_name(session->s_state));
2590 	if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2591 	    session->s_state != CEPH_MDS_SESSION_HUNG) {
2592 		if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2593 			err = -EACCES;
2594 			goto out_session;
2595 		}
2596 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
2597 		    session->s_state == CEPH_MDS_SESSION_CLOSING)
2598 			__open_session(mdsc, session);
2599 		list_add(&req->r_wait, &session->s_waiting);
2600 		goto out_session;
2601 	}
2602 
2603 	/* send request */
2604 	req->r_resend_mds = -1;   /* forget any previous mds hint */
2605 
2606 	if (req->r_request_started == 0)   /* note request start time */
2607 		req->r_request_started = jiffies;
2608 
2609 	err = __prepare_send_request(mdsc, req, mds, false);
2610 	if (!err) {
2611 		ceph_msg_get(req->r_request);
2612 		ceph_con_send(&session->s_con, req->r_request);
2613 	}
2614 
2615 out_session:
2616 	ceph_put_mds_session(session);
2617 finish:
2618 	if (err) {
2619 		dout("__do_request early error %d\n", err);
2620 		req->r_err = err;
2621 		complete_request(mdsc, req);
2622 		__unregister_request(mdsc, req);
2623 	}
2624 	return;
2625 }
2626 
2627 /*
2628  * called under mdsc->mutex
2629  */
2630 static void __wake_requests(struct ceph_mds_client *mdsc,
2631 			    struct list_head *head)
2632 {
2633 	struct ceph_mds_request *req;
2634 	LIST_HEAD(tmp_list);
2635 
2636 	list_splice_init(head, &tmp_list);
2637 
2638 	while (!list_empty(&tmp_list)) {
2639 		req = list_entry(tmp_list.next,
2640 				 struct ceph_mds_request, r_wait);
2641 		list_del_init(&req->r_wait);
2642 		dout(" wake request %p tid %llu\n", req, req->r_tid);
2643 		__do_request(mdsc, req);
2644 	}
2645 }
2646 
2647 /*
2648  * Wake up threads with requests pending for @mds, so that they can
2649  * resubmit their requests to a possibly different mds.
2650  */
2651 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2652 {
2653 	struct ceph_mds_request *req;
2654 	struct rb_node *p = rb_first(&mdsc->request_tree);
2655 
2656 	dout("kick_requests mds%d\n", mds);
2657 	while (p) {
2658 		req = rb_entry(p, struct ceph_mds_request, r_node);
2659 		p = rb_next(p);
2660 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2661 			continue;
2662 		if (req->r_attempts > 0)
2663 			continue; /* only new requests */
2664 		if (req->r_session &&
2665 		    req->r_session->s_mds == mds) {
2666 			dout(" kicking tid %llu\n", req->r_tid);
2667 			list_del_init(&req->r_wait);
2668 			__do_request(mdsc, req);
2669 		}
2670 	}
2671 }
2672 
2673 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2674 			      struct ceph_mds_request *req)
2675 {
2676 	int err;
2677 
2678 	/* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2679 	if (req->r_inode)
2680 		ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2681 	if (req->r_parent) {
2682 		ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2683 		ihold(req->r_parent);
2684 	}
2685 	if (req->r_old_dentry_dir)
2686 		ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2687 				  CEPH_CAP_PIN);
2688 
2689 	dout("submit_request on %p for inode %p\n", req, dir);
2690 	mutex_lock(&mdsc->mutex);
2691 	__register_request(mdsc, req, dir);
2692 	__do_request(mdsc, req);
2693 	err = req->r_err;
2694 	mutex_unlock(&mdsc->mutex);
2695 	return err;
2696 }
2697 
2698 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2699 				  struct ceph_mds_request *req)
2700 {
2701 	int err;
2702 
2703 	/* wait */
2704 	dout("do_request waiting\n");
2705 	if (!req->r_timeout && req->r_wait_for_completion) {
2706 		err = req->r_wait_for_completion(mdsc, req);
2707 	} else {
2708 		long timeleft = wait_for_completion_killable_timeout(
2709 					&req->r_completion,
2710 					ceph_timeout_jiffies(req->r_timeout));
2711 		if (timeleft > 0)
2712 			err = 0;
2713 		else if (!timeleft)
2714 			err = -EIO;  /* timed out */
2715 		else
2716 			err = timeleft;  /* killed */
2717 	}
2718 	dout("do_request waited, got %d\n", err);
2719 	mutex_lock(&mdsc->mutex);
2720 
2721 	/* only abort if we didn't race with a real reply */
2722 	if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2723 		err = le32_to_cpu(req->r_reply_info.head->result);
2724 	} else if (err < 0) {
2725 		dout("aborted request %lld with %d\n", req->r_tid, err);
2726 
2727 		/*
2728 		 * ensure we aren't running concurrently with
2729 		 * ceph_fill_trace or ceph_readdir_prepopulate, which
2730 		 * rely on locks (dir mutex) held by our caller.
2731 		 */
2732 		mutex_lock(&req->r_fill_mutex);
2733 		req->r_err = err;
2734 		set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2735 		mutex_unlock(&req->r_fill_mutex);
2736 
2737 		if (req->r_parent &&
2738 		    (req->r_op & CEPH_MDS_OP_WRITE))
2739 			ceph_invalidate_dir_request(req);
2740 	} else {
2741 		err = req->r_err;
2742 	}
2743 
2744 	mutex_unlock(&mdsc->mutex);
2745 	return err;
2746 }
2747 
2748 /*
2749  * Synchrously perform an mds request.  Take care of all of the
2750  * session setup, forwarding, retry details.
2751  */
2752 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2753 			 struct inode *dir,
2754 			 struct ceph_mds_request *req)
2755 {
2756 	int err;
2757 
2758 	dout("do_request on %p\n", req);
2759 
2760 	/* issue */
2761 	err = ceph_mdsc_submit_request(mdsc, dir, req);
2762 	if (!err)
2763 		err = ceph_mdsc_wait_request(mdsc, req);
2764 	dout("do_request %p done, result %d\n", req, err);
2765 	return err;
2766 }
2767 
2768 /*
2769  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2770  * namespace request.
2771  */
2772 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2773 {
2774 	struct inode *dir = req->r_parent;
2775 	struct inode *old_dir = req->r_old_dentry_dir;
2776 
2777 	dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2778 
2779 	ceph_dir_clear_complete(dir);
2780 	if (old_dir)
2781 		ceph_dir_clear_complete(old_dir);
2782 	if (req->r_dentry)
2783 		ceph_invalidate_dentry_lease(req->r_dentry);
2784 	if (req->r_old_dentry)
2785 		ceph_invalidate_dentry_lease(req->r_old_dentry);
2786 }
2787 
2788 /*
2789  * Handle mds reply.
2790  *
2791  * We take the session mutex and parse and process the reply immediately.
2792  * This preserves the logical ordering of replies, capabilities, etc., sent
2793  * by the MDS as they are applied to our local cache.
2794  */
2795 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2796 {
2797 	struct ceph_mds_client *mdsc = session->s_mdsc;
2798 	struct ceph_mds_request *req;
2799 	struct ceph_mds_reply_head *head = msg->front.iov_base;
2800 	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2801 	struct ceph_snap_realm *realm;
2802 	u64 tid;
2803 	int err, result;
2804 	int mds = session->s_mds;
2805 
2806 	if (msg->front.iov_len < sizeof(*head)) {
2807 		pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2808 		ceph_msg_dump(msg);
2809 		return;
2810 	}
2811 
2812 	/* get request, session */
2813 	tid = le64_to_cpu(msg->hdr.tid);
2814 	mutex_lock(&mdsc->mutex);
2815 	req = lookup_get_request(mdsc, tid);
2816 	if (!req) {
2817 		dout("handle_reply on unknown tid %llu\n", tid);
2818 		mutex_unlock(&mdsc->mutex);
2819 		return;
2820 	}
2821 	dout("handle_reply %p\n", req);
2822 
2823 	/* correct session? */
2824 	if (req->r_session != session) {
2825 		pr_err("mdsc_handle_reply got %llu on session mds%d"
2826 		       " not mds%d\n", tid, session->s_mds,
2827 		       req->r_session ? req->r_session->s_mds : -1);
2828 		mutex_unlock(&mdsc->mutex);
2829 		goto out;
2830 	}
2831 
2832 	/* dup? */
2833 	if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2834 	    (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2835 		pr_warn("got a dup %s reply on %llu from mds%d\n",
2836 			   head->safe ? "safe" : "unsafe", tid, mds);
2837 		mutex_unlock(&mdsc->mutex);
2838 		goto out;
2839 	}
2840 	if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2841 		pr_warn("got unsafe after safe on %llu from mds%d\n",
2842 			   tid, mds);
2843 		mutex_unlock(&mdsc->mutex);
2844 		goto out;
2845 	}
2846 
2847 	result = le32_to_cpu(head->result);
2848 
2849 	/*
2850 	 * Handle an ESTALE
2851 	 * if we're not talking to the authority, send to them
2852 	 * if the authority has changed while we weren't looking,
2853 	 * send to new authority
2854 	 * Otherwise we just have to return an ESTALE
2855 	 */
2856 	if (result == -ESTALE) {
2857 		dout("got ESTALE on request %llu\n", req->r_tid);
2858 		req->r_resend_mds = -1;
2859 		if (req->r_direct_mode != USE_AUTH_MDS) {
2860 			dout("not using auth, setting for that now\n");
2861 			req->r_direct_mode = USE_AUTH_MDS;
2862 			__do_request(mdsc, req);
2863 			mutex_unlock(&mdsc->mutex);
2864 			goto out;
2865 		} else  {
2866 			int mds = __choose_mds(mdsc, req);
2867 			if (mds >= 0 && mds != req->r_session->s_mds) {
2868 				dout("but auth changed, so resending\n");
2869 				__do_request(mdsc, req);
2870 				mutex_unlock(&mdsc->mutex);
2871 				goto out;
2872 			}
2873 		}
2874 		dout("have to return ESTALE on request %llu\n", req->r_tid);
2875 	}
2876 
2877 
2878 	if (head->safe) {
2879 		set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2880 		__unregister_request(mdsc, req);
2881 
2882 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2883 			/*
2884 			 * We already handled the unsafe response, now do the
2885 			 * cleanup.  No need to examine the response; the MDS
2886 			 * doesn't include any result info in the safe
2887 			 * response.  And even if it did, there is nothing
2888 			 * useful we could do with a revised return value.
2889 			 */
2890 			dout("got safe reply %llu, mds%d\n", tid, mds);
2891 
2892 			/* last unsafe request during umount? */
2893 			if (mdsc->stopping && !__get_oldest_req(mdsc))
2894 				complete_all(&mdsc->safe_umount_waiters);
2895 			mutex_unlock(&mdsc->mutex);
2896 			goto out;
2897 		}
2898 	} else {
2899 		set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2900 		list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2901 		if (req->r_unsafe_dir) {
2902 			struct ceph_inode_info *ci =
2903 					ceph_inode(req->r_unsafe_dir);
2904 			spin_lock(&ci->i_unsafe_lock);
2905 			list_add_tail(&req->r_unsafe_dir_item,
2906 				      &ci->i_unsafe_dirops);
2907 			spin_unlock(&ci->i_unsafe_lock);
2908 		}
2909 	}
2910 
2911 	dout("handle_reply tid %lld result %d\n", tid, result);
2912 	rinfo = &req->r_reply_info;
2913 	if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2914 		err = parse_reply_info(msg, rinfo, (u64)-1);
2915 	else
2916 		err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2917 	mutex_unlock(&mdsc->mutex);
2918 
2919 	mutex_lock(&session->s_mutex);
2920 	if (err < 0) {
2921 		pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2922 		ceph_msg_dump(msg);
2923 		goto out_err;
2924 	}
2925 
2926 	/* snap trace */
2927 	realm = NULL;
2928 	if (rinfo->snapblob_len) {
2929 		down_write(&mdsc->snap_rwsem);
2930 		ceph_update_snap_trace(mdsc, rinfo->snapblob,
2931 				rinfo->snapblob + rinfo->snapblob_len,
2932 				le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2933 				&realm);
2934 		downgrade_write(&mdsc->snap_rwsem);
2935 	} else {
2936 		down_read(&mdsc->snap_rwsem);
2937 	}
2938 
2939 	/* insert trace into our cache */
2940 	mutex_lock(&req->r_fill_mutex);
2941 	current->journal_info = req;
2942 	err = ceph_fill_trace(mdsc->fsc->sb, req);
2943 	if (err == 0) {
2944 		if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2945 				    req->r_op == CEPH_MDS_OP_LSSNAP))
2946 			ceph_readdir_prepopulate(req, req->r_session);
2947 	}
2948 	current->journal_info = NULL;
2949 	mutex_unlock(&req->r_fill_mutex);
2950 
2951 	up_read(&mdsc->snap_rwsem);
2952 	if (realm)
2953 		ceph_put_snap_realm(mdsc, realm);
2954 
2955 	if (err == 0) {
2956 		if (req->r_target_inode &&
2957 		    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2958 			struct ceph_inode_info *ci =
2959 				ceph_inode(req->r_target_inode);
2960 			spin_lock(&ci->i_unsafe_lock);
2961 			list_add_tail(&req->r_unsafe_target_item,
2962 				      &ci->i_unsafe_iops);
2963 			spin_unlock(&ci->i_unsafe_lock);
2964 		}
2965 
2966 		ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2967 	}
2968 out_err:
2969 	mutex_lock(&mdsc->mutex);
2970 	if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2971 		if (err) {
2972 			req->r_err = err;
2973 		} else {
2974 			req->r_reply =  ceph_msg_get(msg);
2975 			set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2976 		}
2977 	} else {
2978 		dout("reply arrived after request %lld was aborted\n", tid);
2979 	}
2980 	mutex_unlock(&mdsc->mutex);
2981 
2982 	mutex_unlock(&session->s_mutex);
2983 
2984 	/* kick calling process */
2985 	complete_request(mdsc, req);
2986 out:
2987 	ceph_mdsc_put_request(req);
2988 	return;
2989 }
2990 
2991 
2992 
2993 /*
2994  * handle mds notification that our request has been forwarded.
2995  */
2996 static void handle_forward(struct ceph_mds_client *mdsc,
2997 			   struct ceph_mds_session *session,
2998 			   struct ceph_msg *msg)
2999 {
3000 	struct ceph_mds_request *req;
3001 	u64 tid = le64_to_cpu(msg->hdr.tid);
3002 	u32 next_mds;
3003 	u32 fwd_seq;
3004 	int err = -EINVAL;
3005 	void *p = msg->front.iov_base;
3006 	void *end = p + msg->front.iov_len;
3007 
3008 	ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3009 	next_mds = ceph_decode_32(&p);
3010 	fwd_seq = ceph_decode_32(&p);
3011 
3012 	mutex_lock(&mdsc->mutex);
3013 	req = lookup_get_request(mdsc, tid);
3014 	if (!req) {
3015 		dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3016 		goto out;  /* dup reply? */
3017 	}
3018 
3019 	if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3020 		dout("forward tid %llu aborted, unregistering\n", tid);
3021 		__unregister_request(mdsc, req);
3022 	} else if (fwd_seq <= req->r_num_fwd) {
3023 		dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3024 		     tid, next_mds, req->r_num_fwd, fwd_seq);
3025 	} else {
3026 		/* resend. forward race not possible; mds would drop */
3027 		dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3028 		BUG_ON(req->r_err);
3029 		BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3030 		req->r_attempts = 0;
3031 		req->r_num_fwd = fwd_seq;
3032 		req->r_resend_mds = next_mds;
3033 		put_request_session(req);
3034 		__do_request(mdsc, req);
3035 	}
3036 	ceph_mdsc_put_request(req);
3037 out:
3038 	mutex_unlock(&mdsc->mutex);
3039 	return;
3040 
3041 bad:
3042 	pr_err("mdsc_handle_forward decode error err=%d\n", err);
3043 }
3044 
3045 static int __decode_session_metadata(void **p, void *end,
3046 				     bool *blacklisted)
3047 {
3048 	/* map<string,string> */
3049 	u32 n;
3050 	bool err_str;
3051 	ceph_decode_32_safe(p, end, n, bad);
3052 	while (n-- > 0) {
3053 		u32 len;
3054 		ceph_decode_32_safe(p, end, len, bad);
3055 		ceph_decode_need(p, end, len, bad);
3056 		err_str = !strncmp(*p, "error_string", len);
3057 		*p += len;
3058 		ceph_decode_32_safe(p, end, len, bad);
3059 		ceph_decode_need(p, end, len, bad);
3060 		if (err_str && strnstr(*p, "blacklisted", len))
3061 			*blacklisted = true;
3062 		*p += len;
3063 	}
3064 	return 0;
3065 bad:
3066 	return -1;
3067 }
3068 
3069 /*
3070  * handle a mds session control message
3071  */
3072 static void handle_session(struct ceph_mds_session *session,
3073 			   struct ceph_msg *msg)
3074 {
3075 	struct ceph_mds_client *mdsc = session->s_mdsc;
3076 	int mds = session->s_mds;
3077 	int msg_version = le16_to_cpu(msg->hdr.version);
3078 	void *p = msg->front.iov_base;
3079 	void *end = p + msg->front.iov_len;
3080 	struct ceph_mds_session_head *h;
3081 	u32 op;
3082 	u64 seq;
3083 	unsigned long features = 0;
3084 	int wake = 0;
3085 	bool blacklisted = false;
3086 
3087 	/* decode */
3088 	ceph_decode_need(&p, end, sizeof(*h), bad);
3089 	h = p;
3090 	p += sizeof(*h);
3091 
3092 	op = le32_to_cpu(h->op);
3093 	seq = le64_to_cpu(h->seq);
3094 
3095 	if (msg_version >= 3) {
3096 		u32 len;
3097 		/* version >= 2, metadata */
3098 		if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3099 			goto bad;
3100 		/* version >= 3, feature bits */
3101 		ceph_decode_32_safe(&p, end, len, bad);
3102 		ceph_decode_need(&p, end, len, bad);
3103 		memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3104 		p += len;
3105 	}
3106 
3107 	mutex_lock(&mdsc->mutex);
3108 	if (op == CEPH_SESSION_CLOSE) {
3109 		get_session(session);
3110 		__unregister_session(mdsc, session);
3111 	}
3112 	/* FIXME: this ttl calculation is generous */
3113 	session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3114 	mutex_unlock(&mdsc->mutex);
3115 
3116 	mutex_lock(&session->s_mutex);
3117 
3118 	dout("handle_session mds%d %s %p state %s seq %llu\n",
3119 	     mds, ceph_session_op_name(op), session,
3120 	     ceph_session_state_name(session->s_state), seq);
3121 
3122 	if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3123 		session->s_state = CEPH_MDS_SESSION_OPEN;
3124 		pr_info("mds%d came back\n", session->s_mds);
3125 	}
3126 
3127 	switch (op) {
3128 	case CEPH_SESSION_OPEN:
3129 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3130 			pr_info("mds%d reconnect success\n", session->s_mds);
3131 		session->s_state = CEPH_MDS_SESSION_OPEN;
3132 		session->s_features = features;
3133 		renewed_caps(mdsc, session, 0);
3134 		wake = 1;
3135 		if (mdsc->stopping)
3136 			__close_session(mdsc, session);
3137 		break;
3138 
3139 	case CEPH_SESSION_RENEWCAPS:
3140 		if (session->s_renew_seq == seq)
3141 			renewed_caps(mdsc, session, 1);
3142 		break;
3143 
3144 	case CEPH_SESSION_CLOSE:
3145 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3146 			pr_info("mds%d reconnect denied\n", session->s_mds);
3147 		cleanup_session_requests(mdsc, session);
3148 		remove_session_caps(session);
3149 		wake = 2; /* for good measure */
3150 		wake_up_all(&mdsc->session_close_wq);
3151 		break;
3152 
3153 	case CEPH_SESSION_STALE:
3154 		pr_info("mds%d caps went stale, renewing\n",
3155 			session->s_mds);
3156 		spin_lock(&session->s_gen_ttl_lock);
3157 		session->s_cap_gen++;
3158 		session->s_cap_ttl = jiffies - 1;
3159 		spin_unlock(&session->s_gen_ttl_lock);
3160 		send_renew_caps(mdsc, session);
3161 		break;
3162 
3163 	case CEPH_SESSION_RECALL_STATE:
3164 		ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3165 		break;
3166 
3167 	case CEPH_SESSION_FLUSHMSG:
3168 		send_flushmsg_ack(mdsc, session, seq);
3169 		break;
3170 
3171 	case CEPH_SESSION_FORCE_RO:
3172 		dout("force_session_readonly %p\n", session);
3173 		spin_lock(&session->s_cap_lock);
3174 		session->s_readonly = true;
3175 		spin_unlock(&session->s_cap_lock);
3176 		wake_up_session_caps(session, FORCE_RO);
3177 		break;
3178 
3179 	case CEPH_SESSION_REJECT:
3180 		WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3181 		pr_info("mds%d rejected session\n", session->s_mds);
3182 		session->s_state = CEPH_MDS_SESSION_REJECTED;
3183 		cleanup_session_requests(mdsc, session);
3184 		remove_session_caps(session);
3185 		if (blacklisted)
3186 			mdsc->fsc->blacklisted = true;
3187 		wake = 2; /* for good measure */
3188 		break;
3189 
3190 	default:
3191 		pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3192 		WARN_ON(1);
3193 	}
3194 
3195 	mutex_unlock(&session->s_mutex);
3196 	if (wake) {
3197 		mutex_lock(&mdsc->mutex);
3198 		__wake_requests(mdsc, &session->s_waiting);
3199 		if (wake == 2)
3200 			kick_requests(mdsc, mds);
3201 		mutex_unlock(&mdsc->mutex);
3202 	}
3203 	if (op == CEPH_SESSION_CLOSE)
3204 		ceph_put_mds_session(session);
3205 	return;
3206 
3207 bad:
3208 	pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3209 	       (int)msg->front.iov_len);
3210 	ceph_msg_dump(msg);
3211 	return;
3212 }
3213 
3214 
3215 /*
3216  * called under session->mutex.
3217  */
3218 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3219 				   struct ceph_mds_session *session)
3220 {
3221 	struct ceph_mds_request *req, *nreq;
3222 	struct rb_node *p;
3223 	int err;
3224 
3225 	dout("replay_unsafe_requests mds%d\n", session->s_mds);
3226 
3227 	mutex_lock(&mdsc->mutex);
3228 	list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
3229 		err = __prepare_send_request(mdsc, req, session->s_mds, true);
3230 		if (!err) {
3231 			ceph_msg_get(req->r_request);
3232 			ceph_con_send(&session->s_con, req->r_request);
3233 		}
3234 	}
3235 
3236 	/*
3237 	 * also re-send old requests when MDS enters reconnect stage. So that MDS
3238 	 * can process completed request in clientreplay stage.
3239 	 */
3240 	p = rb_first(&mdsc->request_tree);
3241 	while (p) {
3242 		req = rb_entry(p, struct ceph_mds_request, r_node);
3243 		p = rb_next(p);
3244 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3245 			continue;
3246 		if (req->r_attempts == 0)
3247 			continue; /* only old requests */
3248 		if (req->r_session &&
3249 		    req->r_session->s_mds == session->s_mds) {
3250 			err = __prepare_send_request(mdsc, req,
3251 						     session->s_mds, true);
3252 			if (!err) {
3253 				ceph_msg_get(req->r_request);
3254 				ceph_con_send(&session->s_con, req->r_request);
3255 			}
3256 		}
3257 	}
3258 	mutex_unlock(&mdsc->mutex);
3259 }
3260 
3261 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3262 {
3263 	struct ceph_msg *reply;
3264 	struct ceph_pagelist *_pagelist;
3265 	struct page *page;
3266 	__le32 *addr;
3267 	int err = -ENOMEM;
3268 
3269 	if (!recon_state->allow_multi)
3270 		return -ENOSPC;
3271 
3272 	/* can't handle message that contains both caps and realm */
3273 	BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3274 
3275 	/* pre-allocate new pagelist */
3276 	_pagelist = ceph_pagelist_alloc(GFP_NOFS);
3277 	if (!_pagelist)
3278 		return -ENOMEM;
3279 
3280 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3281 	if (!reply)
3282 		goto fail_msg;
3283 
3284 	/* placeholder for nr_caps */
3285 	err = ceph_pagelist_encode_32(_pagelist, 0);
3286 	if (err < 0)
3287 		goto fail;
3288 
3289 	if (recon_state->nr_caps) {
3290 		/* currently encoding caps */
3291 		err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3292 		if (err)
3293 			goto fail;
3294 	} else {
3295 		/* placeholder for nr_realms (currently encoding relams) */
3296 		err = ceph_pagelist_encode_32(_pagelist, 0);
3297 		if (err < 0)
3298 			goto fail;
3299 	}
3300 
3301 	err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3302 	if (err)
3303 		goto fail;
3304 
3305 	page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3306 	addr = kmap_atomic(page);
3307 	if (recon_state->nr_caps) {
3308 		/* currently encoding caps */
3309 		*addr = cpu_to_le32(recon_state->nr_caps);
3310 	} else {
3311 		/* currently encoding relams */
3312 		*(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3313 	}
3314 	kunmap_atomic(addr);
3315 
3316 	reply->hdr.version = cpu_to_le16(5);
3317 	reply->hdr.compat_version = cpu_to_le16(4);
3318 
3319 	reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3320 	ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3321 
3322 	ceph_con_send(&recon_state->session->s_con, reply);
3323 	ceph_pagelist_release(recon_state->pagelist);
3324 
3325 	recon_state->pagelist = _pagelist;
3326 	recon_state->nr_caps = 0;
3327 	recon_state->nr_realms = 0;
3328 	recon_state->msg_version = 5;
3329 	return 0;
3330 fail:
3331 	ceph_msg_put(reply);
3332 fail_msg:
3333 	ceph_pagelist_release(_pagelist);
3334 	return err;
3335 }
3336 
3337 /*
3338  * Encode information about a cap for a reconnect with the MDS.
3339  */
3340 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3341 			  void *arg)
3342 {
3343 	union {
3344 		struct ceph_mds_cap_reconnect v2;
3345 		struct ceph_mds_cap_reconnect_v1 v1;
3346 	} rec;
3347 	struct ceph_inode_info *ci = cap->ci;
3348 	struct ceph_reconnect_state *recon_state = arg;
3349 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3350 	int err;
3351 	u64 snap_follows;
3352 
3353 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3354 	     inode, ceph_vinop(inode), cap, cap->cap_id,
3355 	     ceph_cap_string(cap->issued));
3356 
3357 	spin_lock(&ci->i_ceph_lock);
3358 	cap->seq = 0;        /* reset cap seq */
3359 	cap->issue_seq = 0;  /* and issue_seq */
3360 	cap->mseq = 0;       /* and migrate_seq */
3361 	cap->cap_gen = cap->session->s_cap_gen;
3362 
3363 	if (recon_state->msg_version >= 2) {
3364 		rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3365 		rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3366 		rec.v2.issued = cpu_to_le32(cap->issued);
3367 		rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3368 		rec.v2.pathbase = 0;
3369 		rec.v2.flock_len = (__force __le32)
3370 			((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3371 	} else {
3372 		rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3373 		rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3374 		rec.v1.issued = cpu_to_le32(cap->issued);
3375 		rec.v1.size = cpu_to_le64(inode->i_size);
3376 		ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3377 		ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3378 		rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3379 		rec.v1.pathbase = 0;
3380 	}
3381 
3382 	if (list_empty(&ci->i_cap_snaps)) {
3383 		snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3384 	} else {
3385 		struct ceph_cap_snap *capsnap =
3386 			list_first_entry(&ci->i_cap_snaps,
3387 					 struct ceph_cap_snap, ci_item);
3388 		snap_follows = capsnap->follows;
3389 	}
3390 	spin_unlock(&ci->i_ceph_lock);
3391 
3392 	if (recon_state->msg_version >= 2) {
3393 		int num_fcntl_locks, num_flock_locks;
3394 		struct ceph_filelock *flocks = NULL;
3395 		size_t struct_len, total_len = sizeof(u64);
3396 		u8 struct_v = 0;
3397 
3398 encode_again:
3399 		if (rec.v2.flock_len) {
3400 			ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3401 		} else {
3402 			num_fcntl_locks = 0;
3403 			num_flock_locks = 0;
3404 		}
3405 		if (num_fcntl_locks + num_flock_locks > 0) {
3406 			flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3407 					       sizeof(struct ceph_filelock),
3408 					       GFP_NOFS);
3409 			if (!flocks) {
3410 				err = -ENOMEM;
3411 				goto out_err;
3412 			}
3413 			err = ceph_encode_locks_to_buffer(inode, flocks,
3414 							  num_fcntl_locks,
3415 							  num_flock_locks);
3416 			if (err) {
3417 				kfree(flocks);
3418 				flocks = NULL;
3419 				if (err == -ENOSPC)
3420 					goto encode_again;
3421 				goto out_err;
3422 			}
3423 		} else {
3424 			kfree(flocks);
3425 			flocks = NULL;
3426 		}
3427 
3428 		if (recon_state->msg_version >= 3) {
3429 			/* version, compat_version and struct_len */
3430 			total_len += 2 * sizeof(u8) + sizeof(u32);
3431 			struct_v = 2;
3432 		}
3433 		/*
3434 		 * number of encoded locks is stable, so copy to pagelist
3435 		 */
3436 		struct_len = 2 * sizeof(u32) +
3437 			    (num_fcntl_locks + num_flock_locks) *
3438 			    sizeof(struct ceph_filelock);
3439 		rec.v2.flock_len = cpu_to_le32(struct_len);
3440 
3441 		struct_len += sizeof(u32) + sizeof(rec.v2);
3442 
3443 		if (struct_v >= 2)
3444 			struct_len += sizeof(u64); /* snap_follows */
3445 
3446 		total_len += struct_len;
3447 
3448 		if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3449 			err = send_reconnect_partial(recon_state);
3450 			if (err)
3451 				goto out_freeflocks;
3452 			pagelist = recon_state->pagelist;
3453 		}
3454 
3455 		err = ceph_pagelist_reserve(pagelist, total_len);
3456 		if (err)
3457 			goto out_freeflocks;
3458 
3459 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3460 		if (recon_state->msg_version >= 3) {
3461 			ceph_pagelist_encode_8(pagelist, struct_v);
3462 			ceph_pagelist_encode_8(pagelist, 1);
3463 			ceph_pagelist_encode_32(pagelist, struct_len);
3464 		}
3465 		ceph_pagelist_encode_string(pagelist, NULL, 0);
3466 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3467 		ceph_locks_to_pagelist(flocks, pagelist,
3468 				       num_fcntl_locks, num_flock_locks);
3469 		if (struct_v >= 2)
3470 			ceph_pagelist_encode_64(pagelist, snap_follows);
3471 out_freeflocks:
3472 		kfree(flocks);
3473 	} else {
3474 		u64 pathbase = 0;
3475 		int pathlen = 0;
3476 		char *path = NULL;
3477 		struct dentry *dentry;
3478 
3479 		dentry = d_find_alias(inode);
3480 		if (dentry) {
3481 			path = ceph_mdsc_build_path(dentry,
3482 						&pathlen, &pathbase, 0);
3483 			dput(dentry);
3484 			if (IS_ERR(path)) {
3485 				err = PTR_ERR(path);
3486 				goto out_err;
3487 			}
3488 			rec.v1.pathbase = cpu_to_le64(pathbase);
3489 		}
3490 
3491 		err = ceph_pagelist_reserve(pagelist,
3492 					    sizeof(u64) + sizeof(u32) +
3493 					    pathlen + sizeof(rec.v1));
3494 		if (err) {
3495 			goto out_freepath;
3496 		}
3497 
3498 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3499 		ceph_pagelist_encode_string(pagelist, path, pathlen);
3500 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3501 out_freepath:
3502 		ceph_mdsc_free_path(path, pathlen);
3503 	}
3504 
3505 out_err:
3506 	if (err >= 0)
3507 		recon_state->nr_caps++;
3508 	return err;
3509 }
3510 
3511 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3512 			      struct ceph_reconnect_state *recon_state)
3513 {
3514 	struct rb_node *p;
3515 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3516 	int err = 0;
3517 
3518 	if (recon_state->msg_version >= 4) {
3519 		err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3520 		if (err < 0)
3521 			goto fail;
3522 	}
3523 
3524 	/*
3525 	 * snaprealms.  we provide mds with the ino, seq (version), and
3526 	 * parent for all of our realms.  If the mds has any newer info,
3527 	 * it will tell us.
3528 	 */
3529 	for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3530 		struct ceph_snap_realm *realm =
3531 		       rb_entry(p, struct ceph_snap_realm, node);
3532 		struct ceph_mds_snaprealm_reconnect sr_rec;
3533 
3534 		if (recon_state->msg_version >= 4) {
3535 			size_t need = sizeof(u8) * 2 + sizeof(u32) +
3536 				      sizeof(sr_rec);
3537 
3538 			if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3539 				err = send_reconnect_partial(recon_state);
3540 				if (err)
3541 					goto fail;
3542 				pagelist = recon_state->pagelist;
3543 			}
3544 
3545 			err = ceph_pagelist_reserve(pagelist, need);
3546 			if (err)
3547 				goto fail;
3548 
3549 			ceph_pagelist_encode_8(pagelist, 1);
3550 			ceph_pagelist_encode_8(pagelist, 1);
3551 			ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3552 		}
3553 
3554 		dout(" adding snap realm %llx seq %lld parent %llx\n",
3555 		     realm->ino, realm->seq, realm->parent_ino);
3556 		sr_rec.ino = cpu_to_le64(realm->ino);
3557 		sr_rec.seq = cpu_to_le64(realm->seq);
3558 		sr_rec.parent = cpu_to_le64(realm->parent_ino);
3559 
3560 		err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3561 		if (err)
3562 			goto fail;
3563 
3564 		recon_state->nr_realms++;
3565 	}
3566 fail:
3567 	return err;
3568 }
3569 
3570 
3571 /*
3572  * If an MDS fails and recovers, clients need to reconnect in order to
3573  * reestablish shared state.  This includes all caps issued through
3574  * this session _and_ the snap_realm hierarchy.  Because it's not
3575  * clear which snap realms the mds cares about, we send everything we
3576  * know about.. that ensures we'll then get any new info the
3577  * recovering MDS might have.
3578  *
3579  * This is a relatively heavyweight operation, but it's rare.
3580  *
3581  * called with mdsc->mutex held.
3582  */
3583 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3584 			       struct ceph_mds_session *session)
3585 {
3586 	struct ceph_msg *reply;
3587 	int mds = session->s_mds;
3588 	int err = -ENOMEM;
3589 	struct ceph_reconnect_state recon_state = {
3590 		.session = session,
3591 	};
3592 	LIST_HEAD(dispose);
3593 
3594 	pr_info("mds%d reconnect start\n", mds);
3595 
3596 	recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3597 	if (!recon_state.pagelist)
3598 		goto fail_nopagelist;
3599 
3600 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3601 	if (!reply)
3602 		goto fail_nomsg;
3603 
3604 	mutex_lock(&session->s_mutex);
3605 	session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3606 	session->s_seq = 0;
3607 
3608 	dout("session %p state %s\n", session,
3609 	     ceph_session_state_name(session->s_state));
3610 
3611 	spin_lock(&session->s_gen_ttl_lock);
3612 	session->s_cap_gen++;
3613 	spin_unlock(&session->s_gen_ttl_lock);
3614 
3615 	spin_lock(&session->s_cap_lock);
3616 	/* don't know if session is readonly */
3617 	session->s_readonly = 0;
3618 	/*
3619 	 * notify __ceph_remove_cap() that we are composing cap reconnect.
3620 	 * If a cap get released before being added to the cap reconnect,
3621 	 * __ceph_remove_cap() should skip queuing cap release.
3622 	 */
3623 	session->s_cap_reconnect = 1;
3624 	/* drop old cap expires; we're about to reestablish that state */
3625 	detach_cap_releases(session, &dispose);
3626 	spin_unlock(&session->s_cap_lock);
3627 	dispose_cap_releases(mdsc, &dispose);
3628 
3629 	/* trim unused caps to reduce MDS's cache rejoin time */
3630 	if (mdsc->fsc->sb->s_root)
3631 		shrink_dcache_parent(mdsc->fsc->sb->s_root);
3632 
3633 	ceph_con_close(&session->s_con);
3634 	ceph_con_open(&session->s_con,
3635 		      CEPH_ENTITY_TYPE_MDS, mds,
3636 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3637 
3638 	/* replay unsafe requests */
3639 	replay_unsafe_requests(mdsc, session);
3640 
3641 	ceph_early_kick_flushing_caps(mdsc, session);
3642 
3643 	down_read(&mdsc->snap_rwsem);
3644 
3645 	/* placeholder for nr_caps */
3646 	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3647 	if (err)
3648 		goto fail;
3649 
3650 	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3651 		recon_state.msg_version = 3;
3652 		recon_state.allow_multi = true;
3653 	} else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3654 		recon_state.msg_version = 3;
3655 	} else {
3656 		recon_state.msg_version = 2;
3657 	}
3658 	/* trsaverse this session's caps */
3659 	err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3660 
3661 	spin_lock(&session->s_cap_lock);
3662 	session->s_cap_reconnect = 0;
3663 	spin_unlock(&session->s_cap_lock);
3664 
3665 	if (err < 0)
3666 		goto fail;
3667 
3668 	/* check if all realms can be encoded into current message */
3669 	if (mdsc->num_snap_realms) {
3670 		size_t total_len =
3671 			recon_state.pagelist->length +
3672 			mdsc->num_snap_realms *
3673 			sizeof(struct ceph_mds_snaprealm_reconnect);
3674 		if (recon_state.msg_version >= 4) {
3675 			/* number of realms */
3676 			total_len += sizeof(u32);
3677 			/* version, compat_version and struct_len */
3678 			total_len += mdsc->num_snap_realms *
3679 				     (2 * sizeof(u8) + sizeof(u32));
3680 		}
3681 		if (total_len > RECONNECT_MAX_SIZE) {
3682 			if (!recon_state.allow_multi) {
3683 				err = -ENOSPC;
3684 				goto fail;
3685 			}
3686 			if (recon_state.nr_caps) {
3687 				err = send_reconnect_partial(&recon_state);
3688 				if (err)
3689 					goto fail;
3690 			}
3691 			recon_state.msg_version = 5;
3692 		}
3693 	}
3694 
3695 	err = encode_snap_realms(mdsc, &recon_state);
3696 	if (err < 0)
3697 		goto fail;
3698 
3699 	if (recon_state.msg_version >= 5) {
3700 		err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3701 		if (err < 0)
3702 			goto fail;
3703 	}
3704 
3705 	if (recon_state.nr_caps || recon_state.nr_realms) {
3706 		struct page *page =
3707 			list_first_entry(&recon_state.pagelist->head,
3708 					struct page, lru);
3709 		__le32 *addr = kmap_atomic(page);
3710 		if (recon_state.nr_caps) {
3711 			WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3712 			*addr = cpu_to_le32(recon_state.nr_caps);
3713 		} else if (recon_state.msg_version >= 4) {
3714 			*(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3715 		}
3716 		kunmap_atomic(addr);
3717 	}
3718 
3719 	reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3720 	if (recon_state.msg_version >= 4)
3721 		reply->hdr.compat_version = cpu_to_le16(4);
3722 
3723 	reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3724 	ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3725 
3726 	ceph_con_send(&session->s_con, reply);
3727 
3728 	mutex_unlock(&session->s_mutex);
3729 
3730 	mutex_lock(&mdsc->mutex);
3731 	__wake_requests(mdsc, &session->s_waiting);
3732 	mutex_unlock(&mdsc->mutex);
3733 
3734 	up_read(&mdsc->snap_rwsem);
3735 	ceph_pagelist_release(recon_state.pagelist);
3736 	return;
3737 
3738 fail:
3739 	ceph_msg_put(reply);
3740 	up_read(&mdsc->snap_rwsem);
3741 	mutex_unlock(&session->s_mutex);
3742 fail_nomsg:
3743 	ceph_pagelist_release(recon_state.pagelist);
3744 fail_nopagelist:
3745 	pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3746 	return;
3747 }
3748 
3749 
3750 /*
3751  * compare old and new mdsmaps, kicking requests
3752  * and closing out old connections as necessary
3753  *
3754  * called under mdsc->mutex.
3755  */
3756 static void check_new_map(struct ceph_mds_client *mdsc,
3757 			  struct ceph_mdsmap *newmap,
3758 			  struct ceph_mdsmap *oldmap)
3759 {
3760 	int i;
3761 	int oldstate, newstate;
3762 	struct ceph_mds_session *s;
3763 
3764 	dout("check_new_map new %u old %u\n",
3765 	     newmap->m_epoch, oldmap->m_epoch);
3766 
3767 	for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3768 		if (!mdsc->sessions[i])
3769 			continue;
3770 		s = mdsc->sessions[i];
3771 		oldstate = ceph_mdsmap_get_state(oldmap, i);
3772 		newstate = ceph_mdsmap_get_state(newmap, i);
3773 
3774 		dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3775 		     i, ceph_mds_state_name(oldstate),
3776 		     ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3777 		     ceph_mds_state_name(newstate),
3778 		     ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3779 		     ceph_session_state_name(s->s_state));
3780 
3781 		if (i >= newmap->m_num_mds) {
3782 			/* force close session for stopped mds */
3783 			get_session(s);
3784 			__unregister_session(mdsc, s);
3785 			__wake_requests(mdsc, &s->s_waiting);
3786 			mutex_unlock(&mdsc->mutex);
3787 
3788 			mutex_lock(&s->s_mutex);
3789 			cleanup_session_requests(mdsc, s);
3790 			remove_session_caps(s);
3791 			mutex_unlock(&s->s_mutex);
3792 
3793 			ceph_put_mds_session(s);
3794 
3795 			mutex_lock(&mdsc->mutex);
3796 			kick_requests(mdsc, i);
3797 			continue;
3798 		}
3799 
3800 		if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
3801 			   ceph_mdsmap_get_addr(newmap, i),
3802 			   sizeof(struct ceph_entity_addr))) {
3803 			/* just close it */
3804 			mutex_unlock(&mdsc->mutex);
3805 			mutex_lock(&s->s_mutex);
3806 			mutex_lock(&mdsc->mutex);
3807 			ceph_con_close(&s->s_con);
3808 			mutex_unlock(&s->s_mutex);
3809 			s->s_state = CEPH_MDS_SESSION_RESTARTING;
3810 		} else if (oldstate == newstate) {
3811 			continue;  /* nothing new with this mds */
3812 		}
3813 
3814 		/*
3815 		 * send reconnect?
3816 		 */
3817 		if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3818 		    newstate >= CEPH_MDS_STATE_RECONNECT) {
3819 			mutex_unlock(&mdsc->mutex);
3820 			send_mds_reconnect(mdsc, s);
3821 			mutex_lock(&mdsc->mutex);
3822 		}
3823 
3824 		/*
3825 		 * kick request on any mds that has gone active.
3826 		 */
3827 		if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3828 		    newstate >= CEPH_MDS_STATE_ACTIVE) {
3829 			if (oldstate != CEPH_MDS_STATE_CREATING &&
3830 			    oldstate != CEPH_MDS_STATE_STARTING)
3831 				pr_info("mds%d recovery completed\n", s->s_mds);
3832 			kick_requests(mdsc, i);
3833 			ceph_kick_flushing_caps(mdsc, s);
3834 			wake_up_session_caps(s, RECONNECT);
3835 		}
3836 	}
3837 
3838 	for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3839 		s = mdsc->sessions[i];
3840 		if (!s)
3841 			continue;
3842 		if (!ceph_mdsmap_is_laggy(newmap, i))
3843 			continue;
3844 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3845 		    s->s_state == CEPH_MDS_SESSION_HUNG ||
3846 		    s->s_state == CEPH_MDS_SESSION_CLOSING) {
3847 			dout(" connecting to export targets of laggy mds%d\n",
3848 			     i);
3849 			__open_export_target_sessions(mdsc, s);
3850 		}
3851 	}
3852 }
3853 
3854 
3855 
3856 /*
3857  * leases
3858  */
3859 
3860 /*
3861  * caller must hold session s_mutex, dentry->d_lock
3862  */
3863 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3864 {
3865 	struct ceph_dentry_info *di = ceph_dentry(dentry);
3866 
3867 	ceph_put_mds_session(di->lease_session);
3868 	di->lease_session = NULL;
3869 }
3870 
3871 static void handle_lease(struct ceph_mds_client *mdsc,
3872 			 struct ceph_mds_session *session,
3873 			 struct ceph_msg *msg)
3874 {
3875 	struct super_block *sb = mdsc->fsc->sb;
3876 	struct inode *inode;
3877 	struct dentry *parent, *dentry;
3878 	struct ceph_dentry_info *di;
3879 	int mds = session->s_mds;
3880 	struct ceph_mds_lease *h = msg->front.iov_base;
3881 	u32 seq;
3882 	struct ceph_vino vino;
3883 	struct qstr dname;
3884 	int release = 0;
3885 
3886 	dout("handle_lease from mds%d\n", mds);
3887 
3888 	/* decode */
3889 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3890 		goto bad;
3891 	vino.ino = le64_to_cpu(h->ino);
3892 	vino.snap = CEPH_NOSNAP;
3893 	seq = le32_to_cpu(h->seq);
3894 	dname.len = get_unaligned_le32(h + 1);
3895 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3896 		goto bad;
3897 	dname.name = (void *)(h + 1) + sizeof(u32);
3898 
3899 	/* lookup inode */
3900 	inode = ceph_find_inode(sb, vino);
3901 	dout("handle_lease %s, ino %llx %p %.*s\n",
3902 	     ceph_lease_op_name(h->action), vino.ino, inode,
3903 	     dname.len, dname.name);
3904 
3905 	mutex_lock(&session->s_mutex);
3906 	session->s_seq++;
3907 
3908 	if (!inode) {
3909 		dout("handle_lease no inode %llx\n", vino.ino);
3910 		goto release;
3911 	}
3912 
3913 	/* dentry */
3914 	parent = d_find_alias(inode);
3915 	if (!parent) {
3916 		dout("no parent dentry on inode %p\n", inode);
3917 		WARN_ON(1);
3918 		goto release;  /* hrm... */
3919 	}
3920 	dname.hash = full_name_hash(parent, dname.name, dname.len);
3921 	dentry = d_lookup(parent, &dname);
3922 	dput(parent);
3923 	if (!dentry)
3924 		goto release;
3925 
3926 	spin_lock(&dentry->d_lock);
3927 	di = ceph_dentry(dentry);
3928 	switch (h->action) {
3929 	case CEPH_MDS_LEASE_REVOKE:
3930 		if (di->lease_session == session) {
3931 			if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3932 				h->seq = cpu_to_le32(di->lease_seq);
3933 			__ceph_mdsc_drop_dentry_lease(dentry);
3934 		}
3935 		release = 1;
3936 		break;
3937 
3938 	case CEPH_MDS_LEASE_RENEW:
3939 		if (di->lease_session == session &&
3940 		    di->lease_gen == session->s_cap_gen &&
3941 		    di->lease_renew_from &&
3942 		    di->lease_renew_after == 0) {
3943 			unsigned long duration =
3944 				msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3945 
3946 			di->lease_seq = seq;
3947 			di->time = di->lease_renew_from + duration;
3948 			di->lease_renew_after = di->lease_renew_from +
3949 				(duration >> 1);
3950 			di->lease_renew_from = 0;
3951 		}
3952 		break;
3953 	}
3954 	spin_unlock(&dentry->d_lock);
3955 	dput(dentry);
3956 
3957 	if (!release)
3958 		goto out;
3959 
3960 release:
3961 	/* let's just reuse the same message */
3962 	h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3963 	ceph_msg_get(msg);
3964 	ceph_con_send(&session->s_con, msg);
3965 
3966 out:
3967 	mutex_unlock(&session->s_mutex);
3968 	/* avoid calling iput_final() in mds dispatch threads */
3969 	ceph_async_iput(inode);
3970 	return;
3971 
3972 bad:
3973 	pr_err("corrupt lease message\n");
3974 	ceph_msg_dump(msg);
3975 }
3976 
3977 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3978 			      struct dentry *dentry, char action,
3979 			      u32 seq)
3980 {
3981 	struct ceph_msg *msg;
3982 	struct ceph_mds_lease *lease;
3983 	struct inode *dir;
3984 	int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
3985 
3986 	dout("lease_send_msg identry %p %s to mds%d\n",
3987 	     dentry, ceph_lease_op_name(action), session->s_mds);
3988 
3989 	msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3990 	if (!msg)
3991 		return;
3992 	lease = msg->front.iov_base;
3993 	lease->action = action;
3994 	lease->seq = cpu_to_le32(seq);
3995 
3996 	spin_lock(&dentry->d_lock);
3997 	dir = d_inode(dentry->d_parent);
3998 	lease->ino = cpu_to_le64(ceph_ino(dir));
3999 	lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4000 
4001 	put_unaligned_le32(dentry->d_name.len, lease + 1);
4002 	memcpy((void *)(lease + 1) + 4,
4003 	       dentry->d_name.name, dentry->d_name.len);
4004 	spin_unlock(&dentry->d_lock);
4005 	/*
4006 	 * if this is a preemptive lease RELEASE, no need to
4007 	 * flush request stream, since the actual request will
4008 	 * soon follow.
4009 	 */
4010 	msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4011 
4012 	ceph_con_send(&session->s_con, msg);
4013 }
4014 
4015 /*
4016  * lock unlock sessions, to wait ongoing session activities
4017  */
4018 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4019 {
4020 	int i;
4021 
4022 	mutex_lock(&mdsc->mutex);
4023 	for (i = 0; i < mdsc->max_sessions; i++) {
4024 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4025 		if (!s)
4026 			continue;
4027 		mutex_unlock(&mdsc->mutex);
4028 		mutex_lock(&s->s_mutex);
4029 		mutex_unlock(&s->s_mutex);
4030 		ceph_put_mds_session(s);
4031 		mutex_lock(&mdsc->mutex);
4032 	}
4033 	mutex_unlock(&mdsc->mutex);
4034 }
4035 
4036 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4037 {
4038 	struct ceph_fs_client *fsc = mdsc->fsc;
4039 
4040 	if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4041 		return;
4042 
4043 	if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4044 		return;
4045 
4046 	if (!READ_ONCE(fsc->blacklisted))
4047 		return;
4048 
4049 	if (fsc->last_auto_reconnect &&
4050 	    time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4051 		return;
4052 
4053 	pr_info("auto reconnect after blacklisted\n");
4054 	fsc->last_auto_reconnect = jiffies;
4055 	ceph_force_reconnect(fsc->sb);
4056 }
4057 
4058 /*
4059  * delayed work -- periodically trim expired leases, renew caps with mds
4060  */
4061 static void schedule_delayed(struct ceph_mds_client *mdsc)
4062 {
4063 	int delay = 5;
4064 	unsigned hz = round_jiffies_relative(HZ * delay);
4065 	schedule_delayed_work(&mdsc->delayed_work, hz);
4066 }
4067 
4068 static void delayed_work(struct work_struct *work)
4069 {
4070 	int i;
4071 	struct ceph_mds_client *mdsc =
4072 		container_of(work, struct ceph_mds_client, delayed_work.work);
4073 	int renew_interval;
4074 	int renew_caps;
4075 
4076 	dout("mdsc delayed_work\n");
4077 
4078 	mutex_lock(&mdsc->mutex);
4079 	renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4080 	renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4081 				   mdsc->last_renew_caps);
4082 	if (renew_caps)
4083 		mdsc->last_renew_caps = jiffies;
4084 
4085 	for (i = 0; i < mdsc->max_sessions; i++) {
4086 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4087 		if (!s)
4088 			continue;
4089 		if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4090 			dout("resending session close request for mds%d\n",
4091 			     s->s_mds);
4092 			request_close_session(mdsc, s);
4093 			ceph_put_mds_session(s);
4094 			continue;
4095 		}
4096 		if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4097 			if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4098 				s->s_state = CEPH_MDS_SESSION_HUNG;
4099 				pr_info("mds%d hung\n", s->s_mds);
4100 			}
4101 		}
4102 		if (s->s_state == CEPH_MDS_SESSION_NEW ||
4103 		    s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4104 		    s->s_state == CEPH_MDS_SESSION_REJECTED) {
4105 			/* this mds is failed or recovering, just wait */
4106 			ceph_put_mds_session(s);
4107 			continue;
4108 		}
4109 		mutex_unlock(&mdsc->mutex);
4110 
4111 		mutex_lock(&s->s_mutex);
4112 		if (renew_caps)
4113 			send_renew_caps(mdsc, s);
4114 		else
4115 			ceph_con_keepalive(&s->s_con);
4116 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4117 		    s->s_state == CEPH_MDS_SESSION_HUNG)
4118 			ceph_send_cap_releases(mdsc, s);
4119 		mutex_unlock(&s->s_mutex);
4120 		ceph_put_mds_session(s);
4121 
4122 		mutex_lock(&mdsc->mutex);
4123 	}
4124 	mutex_unlock(&mdsc->mutex);
4125 
4126 	ceph_check_delayed_caps(mdsc);
4127 
4128 	ceph_queue_cap_reclaim_work(mdsc);
4129 
4130 	ceph_trim_snapid_map(mdsc);
4131 
4132 	maybe_recover_session(mdsc);
4133 
4134 	schedule_delayed(mdsc);
4135 }
4136 
4137 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4138 
4139 {
4140 	struct ceph_mds_client *mdsc;
4141 
4142 	mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4143 	if (!mdsc)
4144 		return -ENOMEM;
4145 	mdsc->fsc = fsc;
4146 	mutex_init(&mdsc->mutex);
4147 	mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4148 	if (!mdsc->mdsmap) {
4149 		kfree(mdsc);
4150 		return -ENOMEM;
4151 	}
4152 
4153 	fsc->mdsc = mdsc;
4154 	init_completion(&mdsc->safe_umount_waiters);
4155 	init_waitqueue_head(&mdsc->session_close_wq);
4156 	INIT_LIST_HEAD(&mdsc->waiting_for_map);
4157 	mdsc->sessions = NULL;
4158 	atomic_set(&mdsc->num_sessions, 0);
4159 	mdsc->max_sessions = 0;
4160 	mdsc->stopping = 0;
4161 	atomic64_set(&mdsc->quotarealms_count, 0);
4162 	mdsc->quotarealms_inodes = RB_ROOT;
4163 	mutex_init(&mdsc->quotarealms_inodes_mutex);
4164 	mdsc->last_snap_seq = 0;
4165 	init_rwsem(&mdsc->snap_rwsem);
4166 	mdsc->snap_realms = RB_ROOT;
4167 	INIT_LIST_HEAD(&mdsc->snap_empty);
4168 	mdsc->num_snap_realms = 0;
4169 	spin_lock_init(&mdsc->snap_empty_lock);
4170 	mdsc->last_tid = 0;
4171 	mdsc->oldest_tid = 0;
4172 	mdsc->request_tree = RB_ROOT;
4173 	INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4174 	mdsc->last_renew_caps = jiffies;
4175 	INIT_LIST_HEAD(&mdsc->cap_delay_list);
4176 	INIT_LIST_HEAD(&mdsc->cap_wait_list);
4177 	spin_lock_init(&mdsc->cap_delay_lock);
4178 	INIT_LIST_HEAD(&mdsc->snap_flush_list);
4179 	spin_lock_init(&mdsc->snap_flush_lock);
4180 	mdsc->last_cap_flush_tid = 1;
4181 	INIT_LIST_HEAD(&mdsc->cap_flush_list);
4182 	INIT_LIST_HEAD(&mdsc->cap_dirty);
4183 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4184 	mdsc->num_cap_flushing = 0;
4185 	spin_lock_init(&mdsc->cap_dirty_lock);
4186 	init_waitqueue_head(&mdsc->cap_flushing_wq);
4187 	INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4188 	atomic_set(&mdsc->cap_reclaim_pending, 0);
4189 
4190 	spin_lock_init(&mdsc->dentry_list_lock);
4191 	INIT_LIST_HEAD(&mdsc->dentry_leases);
4192 	INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4193 
4194 	ceph_caps_init(mdsc);
4195 	ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4196 
4197 	spin_lock_init(&mdsc->snapid_map_lock);
4198 	mdsc->snapid_map_tree = RB_ROOT;
4199 	INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4200 
4201 	init_rwsem(&mdsc->pool_perm_rwsem);
4202 	mdsc->pool_perm_tree = RB_ROOT;
4203 
4204 	strscpy(mdsc->nodename, utsname()->nodename,
4205 		sizeof(mdsc->nodename));
4206 	return 0;
4207 }
4208 
4209 /*
4210  * Wait for safe replies on open mds requests.  If we time out, drop
4211  * all requests from the tree to avoid dangling dentry refs.
4212  */
4213 static void wait_requests(struct ceph_mds_client *mdsc)
4214 {
4215 	struct ceph_options *opts = mdsc->fsc->client->options;
4216 	struct ceph_mds_request *req;
4217 
4218 	mutex_lock(&mdsc->mutex);
4219 	if (__get_oldest_req(mdsc)) {
4220 		mutex_unlock(&mdsc->mutex);
4221 
4222 		dout("wait_requests waiting for requests\n");
4223 		wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4224 				    ceph_timeout_jiffies(opts->mount_timeout));
4225 
4226 		/* tear down remaining requests */
4227 		mutex_lock(&mdsc->mutex);
4228 		while ((req = __get_oldest_req(mdsc))) {
4229 			dout("wait_requests timed out on tid %llu\n",
4230 			     req->r_tid);
4231 			list_del_init(&req->r_wait);
4232 			__unregister_request(mdsc, req);
4233 		}
4234 	}
4235 	mutex_unlock(&mdsc->mutex);
4236 	dout("wait_requests done\n");
4237 }
4238 
4239 /*
4240  * called before mount is ro, and before dentries are torn down.
4241  * (hmm, does this still race with new lookups?)
4242  */
4243 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4244 {
4245 	dout("pre_umount\n");
4246 	mdsc->stopping = 1;
4247 
4248 	lock_unlock_sessions(mdsc);
4249 	ceph_flush_dirty_caps(mdsc);
4250 	wait_requests(mdsc);
4251 
4252 	/*
4253 	 * wait for reply handlers to drop their request refs and
4254 	 * their inode/dcache refs
4255 	 */
4256 	ceph_msgr_flush();
4257 
4258 	ceph_cleanup_quotarealms_inodes(mdsc);
4259 }
4260 
4261 /*
4262  * wait for all write mds requests to flush.
4263  */
4264 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4265 {
4266 	struct ceph_mds_request *req = NULL, *nextreq;
4267 	struct rb_node *n;
4268 
4269 	mutex_lock(&mdsc->mutex);
4270 	dout("wait_unsafe_requests want %lld\n", want_tid);
4271 restart:
4272 	req = __get_oldest_req(mdsc);
4273 	while (req && req->r_tid <= want_tid) {
4274 		/* find next request */
4275 		n = rb_next(&req->r_node);
4276 		if (n)
4277 			nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4278 		else
4279 			nextreq = NULL;
4280 		if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4281 		    (req->r_op & CEPH_MDS_OP_WRITE)) {
4282 			/* write op */
4283 			ceph_mdsc_get_request(req);
4284 			if (nextreq)
4285 				ceph_mdsc_get_request(nextreq);
4286 			mutex_unlock(&mdsc->mutex);
4287 			dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4288 			     req->r_tid, want_tid);
4289 			wait_for_completion(&req->r_safe_completion);
4290 			mutex_lock(&mdsc->mutex);
4291 			ceph_mdsc_put_request(req);
4292 			if (!nextreq)
4293 				break;  /* next dne before, so we're done! */
4294 			if (RB_EMPTY_NODE(&nextreq->r_node)) {
4295 				/* next request was removed from tree */
4296 				ceph_mdsc_put_request(nextreq);
4297 				goto restart;
4298 			}
4299 			ceph_mdsc_put_request(nextreq);  /* won't go away */
4300 		}
4301 		req = nextreq;
4302 	}
4303 	mutex_unlock(&mdsc->mutex);
4304 	dout("wait_unsafe_requests done\n");
4305 }
4306 
4307 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4308 {
4309 	u64 want_tid, want_flush;
4310 
4311 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4312 		return;
4313 
4314 	dout("sync\n");
4315 	mutex_lock(&mdsc->mutex);
4316 	want_tid = mdsc->last_tid;
4317 	mutex_unlock(&mdsc->mutex);
4318 
4319 	ceph_flush_dirty_caps(mdsc);
4320 	spin_lock(&mdsc->cap_dirty_lock);
4321 	want_flush = mdsc->last_cap_flush_tid;
4322 	if (!list_empty(&mdsc->cap_flush_list)) {
4323 		struct ceph_cap_flush *cf =
4324 			list_last_entry(&mdsc->cap_flush_list,
4325 					struct ceph_cap_flush, g_list);
4326 		cf->wake = true;
4327 	}
4328 	spin_unlock(&mdsc->cap_dirty_lock);
4329 
4330 	dout("sync want tid %lld flush_seq %lld\n",
4331 	     want_tid, want_flush);
4332 
4333 	wait_unsafe_requests(mdsc, want_tid);
4334 	wait_caps_flush(mdsc, want_flush);
4335 }
4336 
4337 /*
4338  * true if all sessions are closed, or we force unmount
4339  */
4340 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4341 {
4342 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4343 		return true;
4344 	return atomic_read(&mdsc->num_sessions) <= skipped;
4345 }
4346 
4347 /*
4348  * called after sb is ro.
4349  */
4350 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4351 {
4352 	struct ceph_options *opts = mdsc->fsc->client->options;
4353 	struct ceph_mds_session *session;
4354 	int i;
4355 	int skipped = 0;
4356 
4357 	dout("close_sessions\n");
4358 
4359 	/* close sessions */
4360 	mutex_lock(&mdsc->mutex);
4361 	for (i = 0; i < mdsc->max_sessions; i++) {
4362 		session = __ceph_lookup_mds_session(mdsc, i);
4363 		if (!session)
4364 			continue;
4365 		mutex_unlock(&mdsc->mutex);
4366 		mutex_lock(&session->s_mutex);
4367 		if (__close_session(mdsc, session) <= 0)
4368 			skipped++;
4369 		mutex_unlock(&session->s_mutex);
4370 		ceph_put_mds_session(session);
4371 		mutex_lock(&mdsc->mutex);
4372 	}
4373 	mutex_unlock(&mdsc->mutex);
4374 
4375 	dout("waiting for sessions to close\n");
4376 	wait_event_timeout(mdsc->session_close_wq,
4377 			   done_closing_sessions(mdsc, skipped),
4378 			   ceph_timeout_jiffies(opts->mount_timeout));
4379 
4380 	/* tear down remaining sessions */
4381 	mutex_lock(&mdsc->mutex);
4382 	for (i = 0; i < mdsc->max_sessions; i++) {
4383 		if (mdsc->sessions[i]) {
4384 			session = get_session(mdsc->sessions[i]);
4385 			__unregister_session(mdsc, session);
4386 			mutex_unlock(&mdsc->mutex);
4387 			mutex_lock(&session->s_mutex);
4388 			remove_session_caps(session);
4389 			mutex_unlock(&session->s_mutex);
4390 			ceph_put_mds_session(session);
4391 			mutex_lock(&mdsc->mutex);
4392 		}
4393 	}
4394 	WARN_ON(!list_empty(&mdsc->cap_delay_list));
4395 	mutex_unlock(&mdsc->mutex);
4396 
4397 	ceph_cleanup_snapid_map(mdsc);
4398 	ceph_cleanup_empty_realms(mdsc);
4399 
4400 	cancel_work_sync(&mdsc->cap_reclaim_work);
4401 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4402 
4403 	dout("stopped\n");
4404 }
4405 
4406 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4407 {
4408 	struct ceph_mds_session *session;
4409 	int mds;
4410 
4411 	dout("force umount\n");
4412 
4413 	mutex_lock(&mdsc->mutex);
4414 	for (mds = 0; mds < mdsc->max_sessions; mds++) {
4415 		session = __ceph_lookup_mds_session(mdsc, mds);
4416 		if (!session)
4417 			continue;
4418 
4419 		if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4420 			__unregister_session(mdsc, session);
4421 		__wake_requests(mdsc, &session->s_waiting);
4422 		mutex_unlock(&mdsc->mutex);
4423 
4424 		mutex_lock(&session->s_mutex);
4425 		__close_session(mdsc, session);
4426 		if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4427 			cleanup_session_requests(mdsc, session);
4428 			remove_session_caps(session);
4429 		}
4430 		mutex_unlock(&session->s_mutex);
4431 		ceph_put_mds_session(session);
4432 
4433 		mutex_lock(&mdsc->mutex);
4434 		kick_requests(mdsc, mds);
4435 	}
4436 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4437 	mutex_unlock(&mdsc->mutex);
4438 }
4439 
4440 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4441 {
4442 	dout("stop\n");
4443 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4444 	if (mdsc->mdsmap)
4445 		ceph_mdsmap_destroy(mdsc->mdsmap);
4446 	kfree(mdsc->sessions);
4447 	ceph_caps_finalize(mdsc);
4448 	ceph_pool_perm_destroy(mdsc);
4449 }
4450 
4451 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4452 {
4453 	struct ceph_mds_client *mdsc = fsc->mdsc;
4454 	dout("mdsc_destroy %p\n", mdsc);
4455 
4456 	if (!mdsc)
4457 		return;
4458 
4459 	/* flush out any connection work with references to us */
4460 	ceph_msgr_flush();
4461 
4462 	ceph_mdsc_stop(mdsc);
4463 
4464 	fsc->mdsc = NULL;
4465 	kfree(mdsc);
4466 	dout("mdsc_destroy %p done\n", mdsc);
4467 }
4468 
4469 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4470 {
4471 	struct ceph_fs_client *fsc = mdsc->fsc;
4472 	const char *mds_namespace = fsc->mount_options->mds_namespace;
4473 	void *p = msg->front.iov_base;
4474 	void *end = p + msg->front.iov_len;
4475 	u32 epoch;
4476 	u32 map_len;
4477 	u32 num_fs;
4478 	u32 mount_fscid = (u32)-1;
4479 	u8 struct_v, struct_cv;
4480 	int err = -EINVAL;
4481 
4482 	ceph_decode_need(&p, end, sizeof(u32), bad);
4483 	epoch = ceph_decode_32(&p);
4484 
4485 	dout("handle_fsmap epoch %u\n", epoch);
4486 
4487 	ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4488 	struct_v = ceph_decode_8(&p);
4489 	struct_cv = ceph_decode_8(&p);
4490 	map_len = ceph_decode_32(&p);
4491 
4492 	ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4493 	p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4494 
4495 	num_fs = ceph_decode_32(&p);
4496 	while (num_fs-- > 0) {
4497 		void *info_p, *info_end;
4498 		u32 info_len;
4499 		u8 info_v, info_cv;
4500 		u32 fscid, namelen;
4501 
4502 		ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4503 		info_v = ceph_decode_8(&p);
4504 		info_cv = ceph_decode_8(&p);
4505 		info_len = ceph_decode_32(&p);
4506 		ceph_decode_need(&p, end, info_len, bad);
4507 		info_p = p;
4508 		info_end = p + info_len;
4509 		p = info_end;
4510 
4511 		ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4512 		fscid = ceph_decode_32(&info_p);
4513 		namelen = ceph_decode_32(&info_p);
4514 		ceph_decode_need(&info_p, info_end, namelen, bad);
4515 
4516 		if (mds_namespace &&
4517 		    strlen(mds_namespace) == namelen &&
4518 		    !strncmp(mds_namespace, (char *)info_p, namelen)) {
4519 			mount_fscid = fscid;
4520 			break;
4521 		}
4522 	}
4523 
4524 	ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4525 	if (mount_fscid != (u32)-1) {
4526 		fsc->client->monc.fs_cluster_id = mount_fscid;
4527 		ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4528 				   0, true);
4529 		ceph_monc_renew_subs(&fsc->client->monc);
4530 	} else {
4531 		err = -ENOENT;
4532 		goto err_out;
4533 	}
4534 	return;
4535 
4536 bad:
4537 	pr_err("error decoding fsmap\n");
4538 err_out:
4539 	mutex_lock(&mdsc->mutex);
4540 	mdsc->mdsmap_err = err;
4541 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4542 	mutex_unlock(&mdsc->mutex);
4543 }
4544 
4545 /*
4546  * handle mds map update.
4547  */
4548 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4549 {
4550 	u32 epoch;
4551 	u32 maplen;
4552 	void *p = msg->front.iov_base;
4553 	void *end = p + msg->front.iov_len;
4554 	struct ceph_mdsmap *newmap, *oldmap;
4555 	struct ceph_fsid fsid;
4556 	int err = -EINVAL;
4557 
4558 	ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4559 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
4560 	if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4561 		return;
4562 	epoch = ceph_decode_32(&p);
4563 	maplen = ceph_decode_32(&p);
4564 	dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4565 
4566 	/* do we need it? */
4567 	mutex_lock(&mdsc->mutex);
4568 	if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4569 		dout("handle_map epoch %u <= our %u\n",
4570 		     epoch, mdsc->mdsmap->m_epoch);
4571 		mutex_unlock(&mdsc->mutex);
4572 		return;
4573 	}
4574 
4575 	newmap = ceph_mdsmap_decode(&p, end);
4576 	if (IS_ERR(newmap)) {
4577 		err = PTR_ERR(newmap);
4578 		goto bad_unlock;
4579 	}
4580 
4581 	/* swap into place */
4582 	if (mdsc->mdsmap) {
4583 		oldmap = mdsc->mdsmap;
4584 		mdsc->mdsmap = newmap;
4585 		check_new_map(mdsc, newmap, oldmap);
4586 		ceph_mdsmap_destroy(oldmap);
4587 	} else {
4588 		mdsc->mdsmap = newmap;  /* first mds map */
4589 	}
4590 	mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4591 					MAX_LFS_FILESIZE);
4592 
4593 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4594 	ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4595 			  mdsc->mdsmap->m_epoch);
4596 
4597 	mutex_unlock(&mdsc->mutex);
4598 	schedule_delayed(mdsc);
4599 	return;
4600 
4601 bad_unlock:
4602 	mutex_unlock(&mdsc->mutex);
4603 bad:
4604 	pr_err("error decoding mdsmap %d\n", err);
4605 	return;
4606 }
4607 
4608 static struct ceph_connection *con_get(struct ceph_connection *con)
4609 {
4610 	struct ceph_mds_session *s = con->private;
4611 
4612 	if (get_session(s)) {
4613 		dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
4614 		return con;
4615 	}
4616 	dout("mdsc con_get %p FAIL\n", s);
4617 	return NULL;
4618 }
4619 
4620 static void con_put(struct ceph_connection *con)
4621 {
4622 	struct ceph_mds_session *s = con->private;
4623 
4624 	dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
4625 	ceph_put_mds_session(s);
4626 }
4627 
4628 /*
4629  * if the client is unresponsive for long enough, the mds will kill
4630  * the session entirely.
4631  */
4632 static void peer_reset(struct ceph_connection *con)
4633 {
4634 	struct ceph_mds_session *s = con->private;
4635 	struct ceph_mds_client *mdsc = s->s_mdsc;
4636 
4637 	pr_warn("mds%d closed our session\n", s->s_mds);
4638 	send_mds_reconnect(mdsc, s);
4639 }
4640 
4641 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4642 {
4643 	struct ceph_mds_session *s = con->private;
4644 	struct ceph_mds_client *mdsc = s->s_mdsc;
4645 	int type = le16_to_cpu(msg->hdr.type);
4646 
4647 	mutex_lock(&mdsc->mutex);
4648 	if (__verify_registered_session(mdsc, s) < 0) {
4649 		mutex_unlock(&mdsc->mutex);
4650 		goto out;
4651 	}
4652 	mutex_unlock(&mdsc->mutex);
4653 
4654 	switch (type) {
4655 	case CEPH_MSG_MDS_MAP:
4656 		ceph_mdsc_handle_mdsmap(mdsc, msg);
4657 		break;
4658 	case CEPH_MSG_FS_MAP_USER:
4659 		ceph_mdsc_handle_fsmap(mdsc, msg);
4660 		break;
4661 	case CEPH_MSG_CLIENT_SESSION:
4662 		handle_session(s, msg);
4663 		break;
4664 	case CEPH_MSG_CLIENT_REPLY:
4665 		handle_reply(s, msg);
4666 		break;
4667 	case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4668 		handle_forward(mdsc, s, msg);
4669 		break;
4670 	case CEPH_MSG_CLIENT_CAPS:
4671 		ceph_handle_caps(s, msg);
4672 		break;
4673 	case CEPH_MSG_CLIENT_SNAP:
4674 		ceph_handle_snap(mdsc, s, msg);
4675 		break;
4676 	case CEPH_MSG_CLIENT_LEASE:
4677 		handle_lease(mdsc, s, msg);
4678 		break;
4679 	case CEPH_MSG_CLIENT_QUOTA:
4680 		ceph_handle_quota(mdsc, s, msg);
4681 		break;
4682 
4683 	default:
4684 		pr_err("received unknown message type %d %s\n", type,
4685 		       ceph_msg_type_name(type));
4686 	}
4687 out:
4688 	ceph_msg_put(msg);
4689 }
4690 
4691 /*
4692  * authentication
4693  */
4694 
4695 /*
4696  * Note: returned pointer is the address of a structure that's
4697  * managed separately.  Caller must *not* attempt to free it.
4698  */
4699 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4700 					int *proto, int force_new)
4701 {
4702 	struct ceph_mds_session *s = con->private;
4703 	struct ceph_mds_client *mdsc = s->s_mdsc;
4704 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4705 	struct ceph_auth_handshake *auth = &s->s_auth;
4706 
4707 	if (force_new && auth->authorizer) {
4708 		ceph_auth_destroy_authorizer(auth->authorizer);
4709 		auth->authorizer = NULL;
4710 	}
4711 	if (!auth->authorizer) {
4712 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4713 						      auth);
4714 		if (ret)
4715 			return ERR_PTR(ret);
4716 	} else {
4717 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4718 						      auth);
4719 		if (ret)
4720 			return ERR_PTR(ret);
4721 	}
4722 	*proto = ac->protocol;
4723 
4724 	return auth;
4725 }
4726 
4727 static int add_authorizer_challenge(struct ceph_connection *con,
4728 				    void *challenge_buf, int challenge_buf_len)
4729 {
4730 	struct ceph_mds_session *s = con->private;
4731 	struct ceph_mds_client *mdsc = s->s_mdsc;
4732 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4733 
4734 	return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4735 					    challenge_buf, challenge_buf_len);
4736 }
4737 
4738 static int verify_authorizer_reply(struct ceph_connection *con)
4739 {
4740 	struct ceph_mds_session *s = con->private;
4741 	struct ceph_mds_client *mdsc = s->s_mdsc;
4742 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4743 
4744 	return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4745 }
4746 
4747 static int invalidate_authorizer(struct ceph_connection *con)
4748 {
4749 	struct ceph_mds_session *s = con->private;
4750 	struct ceph_mds_client *mdsc = s->s_mdsc;
4751 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4752 
4753 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4754 
4755 	return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4756 }
4757 
4758 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4759 				struct ceph_msg_header *hdr, int *skip)
4760 {
4761 	struct ceph_msg *msg;
4762 	int type = (int) le16_to_cpu(hdr->type);
4763 	int front_len = (int) le32_to_cpu(hdr->front_len);
4764 
4765 	if (con->in_msg)
4766 		return con->in_msg;
4767 
4768 	*skip = 0;
4769 	msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4770 	if (!msg) {
4771 		pr_err("unable to allocate msg type %d len %d\n",
4772 		       type, front_len);
4773 		return NULL;
4774 	}
4775 
4776 	return msg;
4777 }
4778 
4779 static int mds_sign_message(struct ceph_msg *msg)
4780 {
4781        struct ceph_mds_session *s = msg->con->private;
4782        struct ceph_auth_handshake *auth = &s->s_auth;
4783 
4784        return ceph_auth_sign_message(auth, msg);
4785 }
4786 
4787 static int mds_check_message_signature(struct ceph_msg *msg)
4788 {
4789        struct ceph_mds_session *s = msg->con->private;
4790        struct ceph_auth_handshake *auth = &s->s_auth;
4791 
4792        return ceph_auth_check_message_signature(auth, msg);
4793 }
4794 
4795 static const struct ceph_connection_operations mds_con_ops = {
4796 	.get = con_get,
4797 	.put = con_put,
4798 	.dispatch = dispatch,
4799 	.get_authorizer = get_authorizer,
4800 	.add_authorizer_challenge = add_authorizer_challenge,
4801 	.verify_authorizer_reply = verify_authorizer_reply,
4802 	.invalidate_authorizer = invalidate_authorizer,
4803 	.peer_reset = peer_reset,
4804 	.alloc_msg = mds_alloc_msg,
4805 	.sign_message = mds_sign_message,
4806 	.check_message_signature = mds_check_message_signature,
4807 };
4808 
4809 /* eof */
4810