xref: /openbmc/linux/fs/ceph/mds_client.c (revision f9834f18)
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 #include <linux/bits.h>
13 
14 #include "super.h"
15 #include "mds_client.h"
16 
17 #include <linux/ceph/ceph_features.h>
18 #include <linux/ceph/messenger.h>
19 #include <linux/ceph/decode.h>
20 #include <linux/ceph/pagelist.h>
21 #include <linux/ceph/auth.h>
22 #include <linux/ceph/debugfs.h>
23 
24 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
25 
26 /*
27  * A cluster of MDS (metadata server) daemons is responsible for
28  * managing the file system namespace (the directory hierarchy and
29  * inodes) and for coordinating shared access to storage.  Metadata is
30  * partitioning hierarchically across a number of servers, and that
31  * partition varies over time as the cluster adjusts the distribution
32  * in order to balance load.
33  *
34  * The MDS client is primarily responsible to managing synchronous
35  * metadata requests for operations like open, unlink, and so forth.
36  * If there is a MDS failure, we find out about it when we (possibly
37  * request and) receive a new MDS map, and can resubmit affected
38  * requests.
39  *
40  * For the most part, though, we take advantage of a lossless
41  * communications channel to the MDS, and do not need to worry about
42  * timing out or resubmitting requests.
43  *
44  * We maintain a stateful "session" with each MDS we interact with.
45  * Within each session, we sent periodic heartbeat messages to ensure
46  * any capabilities or leases we have been issues remain valid.  If
47  * the session times out and goes stale, our leases and capabilities
48  * are no longer valid.
49  */
50 
51 struct ceph_reconnect_state {
52 	struct ceph_mds_session *session;
53 	int nr_caps, nr_realms;
54 	struct ceph_pagelist *pagelist;
55 	unsigned msg_version;
56 	bool allow_multi;
57 };
58 
59 static void __wake_requests(struct ceph_mds_client *mdsc,
60 			    struct list_head *head);
61 static void ceph_cap_release_work(struct work_struct *work);
62 static void ceph_cap_reclaim_work(struct work_struct *work);
63 
64 static const struct ceph_connection_operations mds_con_ops;
65 
66 
67 /*
68  * mds reply parsing
69  */
70 
71 static int parse_reply_info_quota(void **p, void *end,
72 				  struct ceph_mds_reply_info_in *info)
73 {
74 	u8 struct_v, struct_compat;
75 	u32 struct_len;
76 
77 	ceph_decode_8_safe(p, end, struct_v, bad);
78 	ceph_decode_8_safe(p, end, struct_compat, bad);
79 	/* struct_v is expected to be >= 1. we only
80 	 * understand encoding with struct_compat == 1. */
81 	if (!struct_v || struct_compat != 1)
82 		goto bad;
83 	ceph_decode_32_safe(p, end, struct_len, bad);
84 	ceph_decode_need(p, end, struct_len, bad);
85 	end = *p + struct_len;
86 	ceph_decode_64_safe(p, end, info->max_bytes, bad);
87 	ceph_decode_64_safe(p, end, info->max_files, bad);
88 	*p = end;
89 	return 0;
90 bad:
91 	return -EIO;
92 }
93 
94 /*
95  * parse individual inode info
96  */
97 static int parse_reply_info_in(void **p, void *end,
98 			       struct ceph_mds_reply_info_in *info,
99 			       u64 features)
100 {
101 	int err = 0;
102 	u8 struct_v = 0;
103 
104 	if (features == (u64)-1) {
105 		u32 struct_len;
106 		u8 struct_compat;
107 		ceph_decode_8_safe(p, end, struct_v, bad);
108 		ceph_decode_8_safe(p, end, struct_compat, bad);
109 		/* struct_v is expected to be >= 1. we only understand
110 		 * encoding with struct_compat == 1. */
111 		if (!struct_v || struct_compat != 1)
112 			goto bad;
113 		ceph_decode_32_safe(p, end, struct_len, bad);
114 		ceph_decode_need(p, end, struct_len, bad);
115 		end = *p + struct_len;
116 	}
117 
118 	ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
119 	info->in = *p;
120 	*p += sizeof(struct ceph_mds_reply_inode) +
121 		sizeof(*info->in->fragtree.splits) *
122 		le32_to_cpu(info->in->fragtree.nsplits);
123 
124 	ceph_decode_32_safe(p, end, info->symlink_len, bad);
125 	ceph_decode_need(p, end, info->symlink_len, bad);
126 	info->symlink = *p;
127 	*p += info->symlink_len;
128 
129 	ceph_decode_copy_safe(p, end, &info->dir_layout,
130 			      sizeof(info->dir_layout), bad);
131 	ceph_decode_32_safe(p, end, info->xattr_len, bad);
132 	ceph_decode_need(p, end, info->xattr_len, bad);
133 	info->xattr_data = *p;
134 	*p += info->xattr_len;
135 
136 	if (features == (u64)-1) {
137 		/* inline data */
138 		ceph_decode_64_safe(p, end, info->inline_version, bad);
139 		ceph_decode_32_safe(p, end, info->inline_len, bad);
140 		ceph_decode_need(p, end, info->inline_len, bad);
141 		info->inline_data = *p;
142 		*p += info->inline_len;
143 		/* quota */
144 		err = parse_reply_info_quota(p, end, info);
145 		if (err < 0)
146 			goto out_bad;
147 		/* pool namespace */
148 		ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
149 		if (info->pool_ns_len > 0) {
150 			ceph_decode_need(p, end, info->pool_ns_len, bad);
151 			info->pool_ns_data = *p;
152 			*p += info->pool_ns_len;
153 		}
154 
155 		/* btime */
156 		ceph_decode_need(p, end, sizeof(info->btime), bad);
157 		ceph_decode_copy(p, &info->btime, sizeof(info->btime));
158 
159 		/* change attribute */
160 		ceph_decode_64_safe(p, end, info->change_attr, bad);
161 
162 		/* dir pin */
163 		if (struct_v >= 2) {
164 			ceph_decode_32_safe(p, end, info->dir_pin, bad);
165 		} else {
166 			info->dir_pin = -ENODATA;
167 		}
168 
169 		/* snapshot birth time, remains zero for v<=2 */
170 		if (struct_v >= 3) {
171 			ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
172 			ceph_decode_copy(p, &info->snap_btime,
173 					 sizeof(info->snap_btime));
174 		} else {
175 			memset(&info->snap_btime, 0, sizeof(info->snap_btime));
176 		}
177 
178 		*p = end;
179 	} else {
180 		if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
181 			ceph_decode_64_safe(p, end, info->inline_version, bad);
182 			ceph_decode_32_safe(p, end, info->inline_len, bad);
183 			ceph_decode_need(p, end, info->inline_len, bad);
184 			info->inline_data = *p;
185 			*p += info->inline_len;
186 		} else
187 			info->inline_version = CEPH_INLINE_NONE;
188 
189 		if (features & CEPH_FEATURE_MDS_QUOTA) {
190 			err = parse_reply_info_quota(p, end, info);
191 			if (err < 0)
192 				goto out_bad;
193 		} else {
194 			info->max_bytes = 0;
195 			info->max_files = 0;
196 		}
197 
198 		info->pool_ns_len = 0;
199 		info->pool_ns_data = NULL;
200 		if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
201 			ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
202 			if (info->pool_ns_len > 0) {
203 				ceph_decode_need(p, end, info->pool_ns_len, bad);
204 				info->pool_ns_data = *p;
205 				*p += info->pool_ns_len;
206 			}
207 		}
208 
209 		if (features & CEPH_FEATURE_FS_BTIME) {
210 			ceph_decode_need(p, end, sizeof(info->btime), bad);
211 			ceph_decode_copy(p, &info->btime, sizeof(info->btime));
212 			ceph_decode_64_safe(p, end, info->change_attr, bad);
213 		}
214 
215 		info->dir_pin = -ENODATA;
216 		/* info->snap_btime remains zero */
217 	}
218 	return 0;
219 bad:
220 	err = -EIO;
221 out_bad:
222 	return err;
223 }
224 
225 static int parse_reply_info_dir(void **p, void *end,
226 				struct ceph_mds_reply_dirfrag **dirfrag,
227 				u64 features)
228 {
229 	if (features == (u64)-1) {
230 		u8 struct_v, struct_compat;
231 		u32 struct_len;
232 		ceph_decode_8_safe(p, end, struct_v, bad);
233 		ceph_decode_8_safe(p, end, struct_compat, bad);
234 		/* struct_v is expected to be >= 1. we only understand
235 		 * encoding whose struct_compat == 1. */
236 		if (!struct_v || struct_compat != 1)
237 			goto bad;
238 		ceph_decode_32_safe(p, end, struct_len, bad);
239 		ceph_decode_need(p, end, struct_len, bad);
240 		end = *p + struct_len;
241 	}
242 
243 	ceph_decode_need(p, end, sizeof(**dirfrag), bad);
244 	*dirfrag = *p;
245 	*p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
246 	if (unlikely(*p > end))
247 		goto bad;
248 	if (features == (u64)-1)
249 		*p = end;
250 	return 0;
251 bad:
252 	return -EIO;
253 }
254 
255 static int parse_reply_info_lease(void **p, void *end,
256 				  struct ceph_mds_reply_lease **lease,
257 				  u64 features)
258 {
259 	if (features == (u64)-1) {
260 		u8 struct_v, struct_compat;
261 		u32 struct_len;
262 		ceph_decode_8_safe(p, end, struct_v, bad);
263 		ceph_decode_8_safe(p, end, struct_compat, bad);
264 		/* struct_v is expected to be >= 1. we only understand
265 		 * encoding whose struct_compat == 1. */
266 		if (!struct_v || struct_compat != 1)
267 			goto bad;
268 		ceph_decode_32_safe(p, end, struct_len, bad);
269 		ceph_decode_need(p, end, struct_len, bad);
270 		end = *p + struct_len;
271 	}
272 
273 	ceph_decode_need(p, end, sizeof(**lease), bad);
274 	*lease = *p;
275 	*p += sizeof(**lease);
276 	if (features == (u64)-1)
277 		*p = end;
278 	return 0;
279 bad:
280 	return -EIO;
281 }
282 
283 /*
284  * parse a normal reply, which may contain a (dir+)dentry and/or a
285  * target inode.
286  */
287 static int parse_reply_info_trace(void **p, void *end,
288 				  struct ceph_mds_reply_info_parsed *info,
289 				  u64 features)
290 {
291 	int err;
292 
293 	if (info->head->is_dentry) {
294 		err = parse_reply_info_in(p, end, &info->diri, features);
295 		if (err < 0)
296 			goto out_bad;
297 
298 		err = parse_reply_info_dir(p, end, &info->dirfrag, features);
299 		if (err < 0)
300 			goto out_bad;
301 
302 		ceph_decode_32_safe(p, end, info->dname_len, bad);
303 		ceph_decode_need(p, end, info->dname_len, bad);
304 		info->dname = *p;
305 		*p += info->dname_len;
306 
307 		err = parse_reply_info_lease(p, end, &info->dlease, features);
308 		if (err < 0)
309 			goto out_bad;
310 	}
311 
312 	if (info->head->is_target) {
313 		err = parse_reply_info_in(p, end, &info->targeti, features);
314 		if (err < 0)
315 			goto out_bad;
316 	}
317 
318 	if (unlikely(*p != end))
319 		goto bad;
320 	return 0;
321 
322 bad:
323 	err = -EIO;
324 out_bad:
325 	pr_err("problem parsing mds trace %d\n", err);
326 	return err;
327 }
328 
329 /*
330  * parse readdir results
331  */
332 static int parse_reply_info_readdir(void **p, void *end,
333 				struct ceph_mds_reply_info_parsed *info,
334 				u64 features)
335 {
336 	u32 num, i = 0;
337 	int err;
338 
339 	err = parse_reply_info_dir(p, end, &info->dir_dir, features);
340 	if (err < 0)
341 		goto out_bad;
342 
343 	ceph_decode_need(p, end, sizeof(num) + 2, bad);
344 	num = ceph_decode_32(p);
345 	{
346 		u16 flags = ceph_decode_16(p);
347 		info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
348 		info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
349 		info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
350 		info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
351 	}
352 	if (num == 0)
353 		goto done;
354 
355 	BUG_ON(!info->dir_entries);
356 	if ((unsigned long)(info->dir_entries + num) >
357 	    (unsigned long)info->dir_entries + info->dir_buf_size) {
358 		pr_err("dir contents are larger than expected\n");
359 		WARN_ON(1);
360 		goto bad;
361 	}
362 
363 	info->dir_nr = num;
364 	while (num) {
365 		struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
366 		/* dentry */
367 		ceph_decode_32_safe(p, end, rde->name_len, bad);
368 		ceph_decode_need(p, end, rde->name_len, bad);
369 		rde->name = *p;
370 		*p += rde->name_len;
371 		dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
372 
373 		/* dentry lease */
374 		err = parse_reply_info_lease(p, end, &rde->lease, features);
375 		if (err)
376 			goto out_bad;
377 		/* inode */
378 		err = parse_reply_info_in(p, end, &rde->inode, features);
379 		if (err < 0)
380 			goto out_bad;
381 		/* ceph_readdir_prepopulate() will update it */
382 		rde->offset = 0;
383 		i++;
384 		num--;
385 	}
386 
387 done:
388 	/* Skip over any unrecognized fields */
389 	*p = end;
390 	return 0;
391 
392 bad:
393 	err = -EIO;
394 out_bad:
395 	pr_err("problem parsing dir contents %d\n", err);
396 	return err;
397 }
398 
399 /*
400  * parse fcntl F_GETLK results
401  */
402 static int parse_reply_info_filelock(void **p, void *end,
403 				     struct ceph_mds_reply_info_parsed *info,
404 				     u64 features)
405 {
406 	if (*p + sizeof(*info->filelock_reply) > end)
407 		goto bad;
408 
409 	info->filelock_reply = *p;
410 
411 	/* Skip over any unrecognized fields */
412 	*p = end;
413 	return 0;
414 bad:
415 	return -EIO;
416 }
417 
418 /*
419  * parse create results
420  */
421 static int parse_reply_info_create(void **p, void *end,
422 				  struct ceph_mds_reply_info_parsed *info,
423 				  u64 features)
424 {
425 	if (features == (u64)-1 ||
426 	    (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
427 		/* Malformed reply? */
428 		if (*p == end) {
429 			info->has_create_ino = false;
430 		} else {
431 			info->has_create_ino = true;
432 			ceph_decode_64_safe(p, end, info->ino, bad);
433 		}
434 	} else {
435 		if (*p != end)
436 			goto bad;
437 	}
438 
439 	/* Skip over any unrecognized fields */
440 	*p = end;
441 	return 0;
442 bad:
443 	return -EIO;
444 }
445 
446 /*
447  * parse extra results
448  */
449 static int parse_reply_info_extra(void **p, void *end,
450 				  struct ceph_mds_reply_info_parsed *info,
451 				  u64 features)
452 {
453 	u32 op = le32_to_cpu(info->head->op);
454 
455 	if (op == CEPH_MDS_OP_GETFILELOCK)
456 		return parse_reply_info_filelock(p, end, info, features);
457 	else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
458 		return parse_reply_info_readdir(p, end, info, features);
459 	else if (op == CEPH_MDS_OP_CREATE)
460 		return parse_reply_info_create(p, end, info, features);
461 	else
462 		return -EIO;
463 }
464 
465 /*
466  * parse entire mds reply
467  */
468 static int parse_reply_info(struct ceph_msg *msg,
469 			    struct ceph_mds_reply_info_parsed *info,
470 			    u64 features)
471 {
472 	void *p, *end;
473 	u32 len;
474 	int err;
475 
476 	info->head = msg->front.iov_base;
477 	p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
478 	end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
479 
480 	/* trace */
481 	ceph_decode_32_safe(&p, end, len, bad);
482 	if (len > 0) {
483 		ceph_decode_need(&p, end, len, bad);
484 		err = parse_reply_info_trace(&p, p+len, info, features);
485 		if (err < 0)
486 			goto out_bad;
487 	}
488 
489 	/* extra */
490 	ceph_decode_32_safe(&p, end, len, bad);
491 	if (len > 0) {
492 		ceph_decode_need(&p, end, len, bad);
493 		err = parse_reply_info_extra(&p, p+len, info, features);
494 		if (err < 0)
495 			goto out_bad;
496 	}
497 
498 	/* snap blob */
499 	ceph_decode_32_safe(&p, end, len, bad);
500 	info->snapblob_len = len;
501 	info->snapblob = p;
502 	p += len;
503 
504 	if (p != end)
505 		goto bad;
506 	return 0;
507 
508 bad:
509 	err = -EIO;
510 out_bad:
511 	pr_err("mds parse_reply err %d\n", err);
512 	return err;
513 }
514 
515 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
516 {
517 	if (!info->dir_entries)
518 		return;
519 	free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
520 }
521 
522 
523 /*
524  * sessions
525  */
526 const char *ceph_session_state_name(int s)
527 {
528 	switch (s) {
529 	case CEPH_MDS_SESSION_NEW: return "new";
530 	case CEPH_MDS_SESSION_OPENING: return "opening";
531 	case CEPH_MDS_SESSION_OPEN: return "open";
532 	case CEPH_MDS_SESSION_HUNG: return "hung";
533 	case CEPH_MDS_SESSION_CLOSING: return "closing";
534 	case CEPH_MDS_SESSION_CLOSED: return "closed";
535 	case CEPH_MDS_SESSION_RESTARTING: return "restarting";
536 	case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
537 	case CEPH_MDS_SESSION_REJECTED: return "rejected";
538 	default: return "???";
539 	}
540 }
541 
542 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
543 {
544 	if (refcount_inc_not_zero(&s->s_ref)) {
545 		dout("mdsc get_session %p %d -> %d\n", s,
546 		     refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
547 		return s;
548 	} else {
549 		dout("mdsc get_session %p 0 -- FAIL\n", s);
550 		return NULL;
551 	}
552 }
553 
554 void ceph_put_mds_session(struct ceph_mds_session *s)
555 {
556 	dout("mdsc put_session %p %d -> %d\n", s,
557 	     refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
558 	if (refcount_dec_and_test(&s->s_ref)) {
559 		if (s->s_auth.authorizer)
560 			ceph_auth_destroy_authorizer(s->s_auth.authorizer);
561 		kfree(s);
562 	}
563 }
564 
565 /*
566  * called under mdsc->mutex
567  */
568 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
569 						   int mds)
570 {
571 	if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
572 		return NULL;
573 	return ceph_get_mds_session(mdsc->sessions[mds]);
574 }
575 
576 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
577 {
578 	if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
579 		return false;
580 	else
581 		return true;
582 }
583 
584 static int __verify_registered_session(struct ceph_mds_client *mdsc,
585 				       struct ceph_mds_session *s)
586 {
587 	if (s->s_mds >= mdsc->max_sessions ||
588 	    mdsc->sessions[s->s_mds] != s)
589 		return -ENOENT;
590 	return 0;
591 }
592 
593 /*
594  * create+register a new session for given mds.
595  * called under mdsc->mutex.
596  */
597 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
598 						 int mds)
599 {
600 	struct ceph_mds_session *s;
601 
602 	if (mds >= mdsc->mdsmap->possible_max_rank)
603 		return ERR_PTR(-EINVAL);
604 
605 	s = kzalloc(sizeof(*s), GFP_NOFS);
606 	if (!s)
607 		return ERR_PTR(-ENOMEM);
608 
609 	if (mds >= mdsc->max_sessions) {
610 		int newmax = 1 << get_count_order(mds + 1);
611 		struct ceph_mds_session **sa;
612 
613 		dout("%s: realloc to %d\n", __func__, newmax);
614 		sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
615 		if (!sa)
616 			goto fail_realloc;
617 		if (mdsc->sessions) {
618 			memcpy(sa, mdsc->sessions,
619 			       mdsc->max_sessions * sizeof(void *));
620 			kfree(mdsc->sessions);
621 		}
622 		mdsc->sessions = sa;
623 		mdsc->max_sessions = newmax;
624 	}
625 
626 	dout("%s: mds%d\n", __func__, mds);
627 	s->s_mdsc = mdsc;
628 	s->s_mds = mds;
629 	s->s_state = CEPH_MDS_SESSION_NEW;
630 	s->s_ttl = 0;
631 	s->s_seq = 0;
632 	mutex_init(&s->s_mutex);
633 
634 	ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
635 
636 	spin_lock_init(&s->s_gen_ttl_lock);
637 	s->s_cap_gen = 1;
638 	s->s_cap_ttl = jiffies - 1;
639 
640 	spin_lock_init(&s->s_cap_lock);
641 	s->s_renew_requested = 0;
642 	s->s_renew_seq = 0;
643 	INIT_LIST_HEAD(&s->s_caps);
644 	s->s_nr_caps = 0;
645 	refcount_set(&s->s_ref, 1);
646 	INIT_LIST_HEAD(&s->s_waiting);
647 	INIT_LIST_HEAD(&s->s_unsafe);
648 	s->s_num_cap_releases = 0;
649 	s->s_cap_reconnect = 0;
650 	s->s_cap_iterator = NULL;
651 	INIT_LIST_HEAD(&s->s_cap_releases);
652 	INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
653 
654 	INIT_LIST_HEAD(&s->s_cap_flushing);
655 
656 	mdsc->sessions[mds] = s;
657 	atomic_inc(&mdsc->num_sessions);
658 	refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
659 
660 	ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
661 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
662 
663 	return s;
664 
665 fail_realloc:
666 	kfree(s);
667 	return ERR_PTR(-ENOMEM);
668 }
669 
670 /*
671  * called under mdsc->mutex
672  */
673 static void __unregister_session(struct ceph_mds_client *mdsc,
674 			       struct ceph_mds_session *s)
675 {
676 	dout("__unregister_session mds%d %p\n", s->s_mds, s);
677 	BUG_ON(mdsc->sessions[s->s_mds] != s);
678 	mdsc->sessions[s->s_mds] = NULL;
679 	ceph_con_close(&s->s_con);
680 	ceph_put_mds_session(s);
681 	atomic_dec(&mdsc->num_sessions);
682 }
683 
684 /*
685  * drop session refs in request.
686  *
687  * should be last request ref, or hold mdsc->mutex
688  */
689 static void put_request_session(struct ceph_mds_request *req)
690 {
691 	if (req->r_session) {
692 		ceph_put_mds_session(req->r_session);
693 		req->r_session = NULL;
694 	}
695 }
696 
697 void ceph_mdsc_release_request(struct kref *kref)
698 {
699 	struct ceph_mds_request *req = container_of(kref,
700 						    struct ceph_mds_request,
701 						    r_kref);
702 	destroy_reply_info(&req->r_reply_info);
703 	if (req->r_request)
704 		ceph_msg_put(req->r_request);
705 	if (req->r_reply)
706 		ceph_msg_put(req->r_reply);
707 	if (req->r_inode) {
708 		ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
709 		/* avoid calling iput_final() in mds dispatch threads */
710 		ceph_async_iput(req->r_inode);
711 	}
712 	if (req->r_parent) {
713 		ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
714 		ceph_async_iput(req->r_parent);
715 	}
716 	ceph_async_iput(req->r_target_inode);
717 	if (req->r_dentry)
718 		dput(req->r_dentry);
719 	if (req->r_old_dentry)
720 		dput(req->r_old_dentry);
721 	if (req->r_old_dentry_dir) {
722 		/*
723 		 * track (and drop pins for) r_old_dentry_dir
724 		 * separately, since r_old_dentry's d_parent may have
725 		 * changed between the dir mutex being dropped and
726 		 * this request being freed.
727 		 */
728 		ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
729 				  CEPH_CAP_PIN);
730 		ceph_async_iput(req->r_old_dentry_dir);
731 	}
732 	kfree(req->r_path1);
733 	kfree(req->r_path2);
734 	if (req->r_pagelist)
735 		ceph_pagelist_release(req->r_pagelist);
736 	put_request_session(req);
737 	ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
738 	WARN_ON_ONCE(!list_empty(&req->r_wait));
739 	kfree(req);
740 }
741 
742 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
743 
744 /*
745  * lookup session, bump ref if found.
746  *
747  * called under mdsc->mutex.
748  */
749 static struct ceph_mds_request *
750 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
751 {
752 	struct ceph_mds_request *req;
753 
754 	req = lookup_request(&mdsc->request_tree, tid);
755 	if (req)
756 		ceph_mdsc_get_request(req);
757 
758 	return req;
759 }
760 
761 /*
762  * Register an in-flight request, and assign a tid.  Link to directory
763  * are modifying (if any).
764  *
765  * Called under mdsc->mutex.
766  */
767 static void __register_request(struct ceph_mds_client *mdsc,
768 			       struct ceph_mds_request *req,
769 			       struct inode *dir)
770 {
771 	int ret = 0;
772 
773 	req->r_tid = ++mdsc->last_tid;
774 	if (req->r_num_caps) {
775 		ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
776 					req->r_num_caps);
777 		if (ret < 0) {
778 			pr_err("__register_request %p "
779 			       "failed to reserve caps: %d\n", req, ret);
780 			/* set req->r_err to fail early from __do_request */
781 			req->r_err = ret;
782 			return;
783 		}
784 	}
785 	dout("__register_request %p tid %lld\n", req, req->r_tid);
786 	ceph_mdsc_get_request(req);
787 	insert_request(&mdsc->request_tree, req);
788 
789 	req->r_uid = current_fsuid();
790 	req->r_gid = current_fsgid();
791 
792 	if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
793 		mdsc->oldest_tid = req->r_tid;
794 
795 	if (dir) {
796 		ihold(dir);
797 		req->r_unsafe_dir = dir;
798 	}
799 }
800 
801 static void __unregister_request(struct ceph_mds_client *mdsc,
802 				 struct ceph_mds_request *req)
803 {
804 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
805 
806 	/* Never leave an unregistered request on an unsafe list! */
807 	list_del_init(&req->r_unsafe_item);
808 
809 	if (req->r_tid == mdsc->oldest_tid) {
810 		struct rb_node *p = rb_next(&req->r_node);
811 		mdsc->oldest_tid = 0;
812 		while (p) {
813 			struct ceph_mds_request *next_req =
814 				rb_entry(p, struct ceph_mds_request, r_node);
815 			if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
816 				mdsc->oldest_tid = next_req->r_tid;
817 				break;
818 			}
819 			p = rb_next(p);
820 		}
821 	}
822 
823 	erase_request(&mdsc->request_tree, req);
824 
825 	if (req->r_unsafe_dir  &&
826 	    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
827 		struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
828 		spin_lock(&ci->i_unsafe_lock);
829 		list_del_init(&req->r_unsafe_dir_item);
830 		spin_unlock(&ci->i_unsafe_lock);
831 	}
832 	if (req->r_target_inode &&
833 	    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
834 		struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
835 		spin_lock(&ci->i_unsafe_lock);
836 		list_del_init(&req->r_unsafe_target_item);
837 		spin_unlock(&ci->i_unsafe_lock);
838 	}
839 
840 	if (req->r_unsafe_dir) {
841 		/* avoid calling iput_final() in mds dispatch threads */
842 		ceph_async_iput(req->r_unsafe_dir);
843 		req->r_unsafe_dir = NULL;
844 	}
845 
846 	complete_all(&req->r_safe_completion);
847 
848 	ceph_mdsc_put_request(req);
849 }
850 
851 /*
852  * Walk back up the dentry tree until we hit a dentry representing a
853  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
854  * when calling this) to ensure that the objects won't disappear while we're
855  * working with them. Once we hit a candidate dentry, we attempt to take a
856  * reference to it, and return that as the result.
857  */
858 static struct inode *get_nonsnap_parent(struct dentry *dentry)
859 {
860 	struct inode *inode = NULL;
861 
862 	while (dentry && !IS_ROOT(dentry)) {
863 		inode = d_inode_rcu(dentry);
864 		if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
865 			break;
866 		dentry = dentry->d_parent;
867 	}
868 	if (inode)
869 		inode = igrab(inode);
870 	return inode;
871 }
872 
873 /*
874  * Choose mds to send request to next.  If there is a hint set in the
875  * request (e.g., due to a prior forward hint from the mds), use that.
876  * Otherwise, consult frag tree and/or caps to identify the
877  * appropriate mds.  If all else fails, choose randomly.
878  *
879  * Called under mdsc->mutex.
880  */
881 static int __choose_mds(struct ceph_mds_client *mdsc,
882 			struct ceph_mds_request *req,
883 			bool *random)
884 {
885 	struct inode *inode;
886 	struct ceph_inode_info *ci;
887 	struct ceph_cap *cap;
888 	int mode = req->r_direct_mode;
889 	int mds = -1;
890 	u32 hash = req->r_direct_hash;
891 	bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
892 
893 	if (random)
894 		*random = false;
895 
896 	/*
897 	 * is there a specific mds we should try?  ignore hint if we have
898 	 * no session and the mds is not up (active or recovering).
899 	 */
900 	if (req->r_resend_mds >= 0 &&
901 	    (__have_session(mdsc, req->r_resend_mds) ||
902 	     ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
903 		dout("%s using resend_mds mds%d\n", __func__,
904 		     req->r_resend_mds);
905 		return req->r_resend_mds;
906 	}
907 
908 	if (mode == USE_RANDOM_MDS)
909 		goto random;
910 
911 	inode = NULL;
912 	if (req->r_inode) {
913 		if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
914 			inode = req->r_inode;
915 			ihold(inode);
916 		} else {
917 			/* req->r_dentry is non-null for LSSNAP request */
918 			rcu_read_lock();
919 			inode = get_nonsnap_parent(req->r_dentry);
920 			rcu_read_unlock();
921 			dout("%s using snapdir's parent %p\n", __func__, inode);
922 		}
923 	} else if (req->r_dentry) {
924 		/* ignore race with rename; old or new d_parent is okay */
925 		struct dentry *parent;
926 		struct inode *dir;
927 
928 		rcu_read_lock();
929 		parent = READ_ONCE(req->r_dentry->d_parent);
930 		dir = req->r_parent ? : d_inode_rcu(parent);
931 
932 		if (!dir || dir->i_sb != mdsc->fsc->sb) {
933 			/*  not this fs or parent went negative */
934 			inode = d_inode(req->r_dentry);
935 			if (inode)
936 				ihold(inode);
937 		} else if (ceph_snap(dir) != CEPH_NOSNAP) {
938 			/* direct snapped/virtual snapdir requests
939 			 * based on parent dir inode */
940 			inode = get_nonsnap_parent(parent);
941 			dout("%s using nonsnap parent %p\n", __func__, inode);
942 		} else {
943 			/* dentry target */
944 			inode = d_inode(req->r_dentry);
945 			if (!inode || mode == USE_AUTH_MDS) {
946 				/* dir + name */
947 				inode = igrab(dir);
948 				hash = ceph_dentry_hash(dir, req->r_dentry);
949 				is_hash = true;
950 			} else {
951 				ihold(inode);
952 			}
953 		}
954 		rcu_read_unlock();
955 	}
956 
957 	dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
958 	     hash, mode);
959 	if (!inode)
960 		goto random;
961 	ci = ceph_inode(inode);
962 
963 	if (is_hash && S_ISDIR(inode->i_mode)) {
964 		struct ceph_inode_frag frag;
965 		int found;
966 
967 		ceph_choose_frag(ci, hash, &frag, &found);
968 		if (found) {
969 			if (mode == USE_ANY_MDS && frag.ndist > 0) {
970 				u8 r;
971 
972 				/* choose a random replica */
973 				get_random_bytes(&r, 1);
974 				r %= frag.ndist;
975 				mds = frag.dist[r];
976 				dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
977 				     __func__, inode, ceph_vinop(inode),
978 				     frag.frag, mds, (int)r, frag.ndist);
979 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
980 				    CEPH_MDS_STATE_ACTIVE &&
981 				    !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
982 					goto out;
983 			}
984 
985 			/* since this file/dir wasn't known to be
986 			 * replicated, then we want to look for the
987 			 * authoritative mds. */
988 			if (frag.mds >= 0) {
989 				/* choose auth mds */
990 				mds = frag.mds;
991 				dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
992 				     __func__, inode, ceph_vinop(inode),
993 				     frag.frag, mds);
994 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
995 				    CEPH_MDS_STATE_ACTIVE) {
996 					if (mode == USE_ANY_MDS &&
997 					    !ceph_mdsmap_is_laggy(mdsc->mdsmap,
998 								  mds))
999 						goto out;
1000 				}
1001 			}
1002 			mode = USE_AUTH_MDS;
1003 		}
1004 	}
1005 
1006 	spin_lock(&ci->i_ceph_lock);
1007 	cap = NULL;
1008 	if (mode == USE_AUTH_MDS)
1009 		cap = ci->i_auth_cap;
1010 	if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1011 		cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1012 	if (!cap) {
1013 		spin_unlock(&ci->i_ceph_lock);
1014 		ceph_async_iput(inode);
1015 		goto random;
1016 	}
1017 	mds = cap->session->s_mds;
1018 	dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1019 	     inode, ceph_vinop(inode), mds,
1020 	     cap == ci->i_auth_cap ? "auth " : "", cap);
1021 	spin_unlock(&ci->i_ceph_lock);
1022 out:
1023 	/* avoid calling iput_final() while holding mdsc->mutex or
1024 	 * in mds dispatch threads */
1025 	ceph_async_iput(inode);
1026 	return mds;
1027 
1028 random:
1029 	if (random)
1030 		*random = true;
1031 
1032 	mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1033 	dout("%s chose random mds%d\n", __func__, mds);
1034 	return mds;
1035 }
1036 
1037 
1038 /*
1039  * session messages
1040  */
1041 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1042 {
1043 	struct ceph_msg *msg;
1044 	struct ceph_mds_session_head *h;
1045 
1046 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1047 			   false);
1048 	if (!msg) {
1049 		pr_err("create_session_msg ENOMEM creating msg\n");
1050 		return NULL;
1051 	}
1052 	h = msg->front.iov_base;
1053 	h->op = cpu_to_le32(op);
1054 	h->seq = cpu_to_le64(seq);
1055 
1056 	return msg;
1057 }
1058 
1059 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1060 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1061 static void encode_supported_features(void **p, void *end)
1062 {
1063 	static const size_t count = ARRAY_SIZE(feature_bits);
1064 
1065 	if (count > 0) {
1066 		size_t i;
1067 		size_t size = FEATURE_BYTES(count);
1068 
1069 		BUG_ON(*p + 4 + size > end);
1070 		ceph_encode_32(p, size);
1071 		memset(*p, 0, size);
1072 		for (i = 0; i < count; i++)
1073 			((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1074 		*p += size;
1075 	} else {
1076 		BUG_ON(*p + 4 > end);
1077 		ceph_encode_32(p, 0);
1078 	}
1079 }
1080 
1081 /*
1082  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1083  * to include additional client metadata fields.
1084  */
1085 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1086 {
1087 	struct ceph_msg *msg;
1088 	struct ceph_mds_session_head *h;
1089 	int i = -1;
1090 	int extra_bytes = 0;
1091 	int metadata_key_count = 0;
1092 	struct ceph_options *opt = mdsc->fsc->client->options;
1093 	struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1094 	size_t size, count;
1095 	void *p, *end;
1096 
1097 	const char* metadata[][2] = {
1098 		{"hostname", mdsc->nodename},
1099 		{"kernel_version", init_utsname()->release},
1100 		{"entity_id", opt->name ? : ""},
1101 		{"root", fsopt->server_path ? : "/"},
1102 		{NULL, NULL}
1103 	};
1104 
1105 	/* Calculate serialized length of metadata */
1106 	extra_bytes = 4;  /* map length */
1107 	for (i = 0; metadata[i][0]; ++i) {
1108 		extra_bytes += 8 + strlen(metadata[i][0]) +
1109 			strlen(metadata[i][1]);
1110 		metadata_key_count++;
1111 	}
1112 
1113 	/* supported feature */
1114 	size = 0;
1115 	count = ARRAY_SIZE(feature_bits);
1116 	if (count > 0)
1117 		size = FEATURE_BYTES(count);
1118 	extra_bytes += 4 + size;
1119 
1120 	/* Allocate the message */
1121 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1122 			   GFP_NOFS, false);
1123 	if (!msg) {
1124 		pr_err("create_session_msg ENOMEM creating msg\n");
1125 		return NULL;
1126 	}
1127 	p = msg->front.iov_base;
1128 	end = p + msg->front.iov_len;
1129 
1130 	h = p;
1131 	h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1132 	h->seq = cpu_to_le64(seq);
1133 
1134 	/*
1135 	 * Serialize client metadata into waiting buffer space, using
1136 	 * the format that userspace expects for map<string, string>
1137 	 *
1138 	 * ClientSession messages with metadata are v3
1139 	 */
1140 	msg->hdr.version = cpu_to_le16(3);
1141 	msg->hdr.compat_version = cpu_to_le16(1);
1142 
1143 	/* The write pointer, following the session_head structure */
1144 	p += sizeof(*h);
1145 
1146 	/* Number of entries in the map */
1147 	ceph_encode_32(&p, metadata_key_count);
1148 
1149 	/* Two length-prefixed strings for each entry in the map */
1150 	for (i = 0; metadata[i][0]; ++i) {
1151 		size_t const key_len = strlen(metadata[i][0]);
1152 		size_t const val_len = strlen(metadata[i][1]);
1153 
1154 		ceph_encode_32(&p, key_len);
1155 		memcpy(p, metadata[i][0], key_len);
1156 		p += key_len;
1157 		ceph_encode_32(&p, val_len);
1158 		memcpy(p, metadata[i][1], val_len);
1159 		p += val_len;
1160 	}
1161 
1162 	encode_supported_features(&p, end);
1163 	msg->front.iov_len = p - msg->front.iov_base;
1164 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1165 
1166 	return msg;
1167 }
1168 
1169 /*
1170  * send session open request.
1171  *
1172  * called under mdsc->mutex
1173  */
1174 static int __open_session(struct ceph_mds_client *mdsc,
1175 			  struct ceph_mds_session *session)
1176 {
1177 	struct ceph_msg *msg;
1178 	int mstate;
1179 	int mds = session->s_mds;
1180 
1181 	/* wait for mds to go active? */
1182 	mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1183 	dout("open_session to mds%d (%s)\n", mds,
1184 	     ceph_mds_state_name(mstate));
1185 	session->s_state = CEPH_MDS_SESSION_OPENING;
1186 	session->s_renew_requested = jiffies;
1187 
1188 	/* send connect message */
1189 	msg = create_session_open_msg(mdsc, session->s_seq);
1190 	if (!msg)
1191 		return -ENOMEM;
1192 	ceph_con_send(&session->s_con, msg);
1193 	return 0;
1194 }
1195 
1196 /*
1197  * open sessions for any export targets for the given mds
1198  *
1199  * called under mdsc->mutex
1200  */
1201 static struct ceph_mds_session *
1202 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1203 {
1204 	struct ceph_mds_session *session;
1205 
1206 	session = __ceph_lookup_mds_session(mdsc, target);
1207 	if (!session) {
1208 		session = register_session(mdsc, target);
1209 		if (IS_ERR(session))
1210 			return session;
1211 	}
1212 	if (session->s_state == CEPH_MDS_SESSION_NEW ||
1213 	    session->s_state == CEPH_MDS_SESSION_CLOSING)
1214 		__open_session(mdsc, session);
1215 
1216 	return session;
1217 }
1218 
1219 struct ceph_mds_session *
1220 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1221 {
1222 	struct ceph_mds_session *session;
1223 
1224 	dout("open_export_target_session to mds%d\n", target);
1225 
1226 	mutex_lock(&mdsc->mutex);
1227 	session = __open_export_target_session(mdsc, target);
1228 	mutex_unlock(&mdsc->mutex);
1229 
1230 	return session;
1231 }
1232 
1233 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1234 					  struct ceph_mds_session *session)
1235 {
1236 	struct ceph_mds_info *mi;
1237 	struct ceph_mds_session *ts;
1238 	int i, mds = session->s_mds;
1239 
1240 	if (mds >= mdsc->mdsmap->possible_max_rank)
1241 		return;
1242 
1243 	mi = &mdsc->mdsmap->m_info[mds];
1244 	dout("open_export_target_sessions for mds%d (%d targets)\n",
1245 	     session->s_mds, mi->num_export_targets);
1246 
1247 	for (i = 0; i < mi->num_export_targets; i++) {
1248 		ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1249 		if (!IS_ERR(ts))
1250 			ceph_put_mds_session(ts);
1251 	}
1252 }
1253 
1254 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1255 					   struct ceph_mds_session *session)
1256 {
1257 	mutex_lock(&mdsc->mutex);
1258 	__open_export_target_sessions(mdsc, session);
1259 	mutex_unlock(&mdsc->mutex);
1260 }
1261 
1262 /*
1263  * session caps
1264  */
1265 
1266 static void detach_cap_releases(struct ceph_mds_session *session,
1267 				struct list_head *target)
1268 {
1269 	lockdep_assert_held(&session->s_cap_lock);
1270 
1271 	list_splice_init(&session->s_cap_releases, target);
1272 	session->s_num_cap_releases = 0;
1273 	dout("dispose_cap_releases mds%d\n", session->s_mds);
1274 }
1275 
1276 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1277 				 struct list_head *dispose)
1278 {
1279 	while (!list_empty(dispose)) {
1280 		struct ceph_cap *cap;
1281 		/* zero out the in-progress message */
1282 		cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1283 		list_del(&cap->session_caps);
1284 		ceph_put_cap(mdsc, cap);
1285 	}
1286 }
1287 
1288 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1289 				     struct ceph_mds_session *session)
1290 {
1291 	struct ceph_mds_request *req;
1292 	struct rb_node *p;
1293 	struct ceph_inode_info *ci;
1294 
1295 	dout("cleanup_session_requests mds%d\n", session->s_mds);
1296 	mutex_lock(&mdsc->mutex);
1297 	while (!list_empty(&session->s_unsafe)) {
1298 		req = list_first_entry(&session->s_unsafe,
1299 				       struct ceph_mds_request, r_unsafe_item);
1300 		pr_warn_ratelimited(" dropping unsafe request %llu\n",
1301 				    req->r_tid);
1302 		if (req->r_target_inode) {
1303 			/* dropping unsafe change of inode's attributes */
1304 			ci = ceph_inode(req->r_target_inode);
1305 			errseq_set(&ci->i_meta_err, -EIO);
1306 		}
1307 		if (req->r_unsafe_dir) {
1308 			/* dropping unsafe directory operation */
1309 			ci = ceph_inode(req->r_unsafe_dir);
1310 			errseq_set(&ci->i_meta_err, -EIO);
1311 		}
1312 		__unregister_request(mdsc, req);
1313 	}
1314 	/* zero r_attempts, so kick_requests() will re-send requests */
1315 	p = rb_first(&mdsc->request_tree);
1316 	while (p) {
1317 		req = rb_entry(p, struct ceph_mds_request, r_node);
1318 		p = rb_next(p);
1319 		if (req->r_session &&
1320 		    req->r_session->s_mds == session->s_mds)
1321 			req->r_attempts = 0;
1322 	}
1323 	mutex_unlock(&mdsc->mutex);
1324 }
1325 
1326 /*
1327  * Helper to safely iterate over all caps associated with a session, with
1328  * special care taken to handle a racing __ceph_remove_cap().
1329  *
1330  * Caller must hold session s_mutex.
1331  */
1332 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1333 			      int (*cb)(struct inode *, struct ceph_cap *,
1334 					void *), void *arg)
1335 {
1336 	struct list_head *p;
1337 	struct ceph_cap *cap;
1338 	struct inode *inode, *last_inode = NULL;
1339 	struct ceph_cap *old_cap = NULL;
1340 	int ret;
1341 
1342 	dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1343 	spin_lock(&session->s_cap_lock);
1344 	p = session->s_caps.next;
1345 	while (p != &session->s_caps) {
1346 		cap = list_entry(p, struct ceph_cap, session_caps);
1347 		inode = igrab(&cap->ci->vfs_inode);
1348 		if (!inode) {
1349 			p = p->next;
1350 			continue;
1351 		}
1352 		session->s_cap_iterator = cap;
1353 		spin_unlock(&session->s_cap_lock);
1354 
1355 		if (last_inode) {
1356 			/* avoid calling iput_final() while holding
1357 			 * s_mutex or in mds dispatch threads */
1358 			ceph_async_iput(last_inode);
1359 			last_inode = NULL;
1360 		}
1361 		if (old_cap) {
1362 			ceph_put_cap(session->s_mdsc, old_cap);
1363 			old_cap = NULL;
1364 		}
1365 
1366 		ret = cb(inode, cap, arg);
1367 		last_inode = inode;
1368 
1369 		spin_lock(&session->s_cap_lock);
1370 		p = p->next;
1371 		if (!cap->ci) {
1372 			dout("iterate_session_caps  finishing cap %p removal\n",
1373 			     cap);
1374 			BUG_ON(cap->session != session);
1375 			cap->session = NULL;
1376 			list_del_init(&cap->session_caps);
1377 			session->s_nr_caps--;
1378 			if (cap->queue_release)
1379 				__ceph_queue_cap_release(session, cap);
1380 			else
1381 				old_cap = cap;  /* put_cap it w/o locks held */
1382 		}
1383 		if (ret < 0)
1384 			goto out;
1385 	}
1386 	ret = 0;
1387 out:
1388 	session->s_cap_iterator = NULL;
1389 	spin_unlock(&session->s_cap_lock);
1390 
1391 	ceph_async_iput(last_inode);
1392 	if (old_cap)
1393 		ceph_put_cap(session->s_mdsc, old_cap);
1394 
1395 	return ret;
1396 }
1397 
1398 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1399 				  void *arg)
1400 {
1401 	struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1402 	struct ceph_inode_info *ci = ceph_inode(inode);
1403 	LIST_HEAD(to_remove);
1404 	bool dirty_dropped = false;
1405 	bool invalidate = false;
1406 
1407 	dout("removing cap %p, ci is %p, inode is %p\n",
1408 	     cap, ci, &ci->vfs_inode);
1409 	spin_lock(&ci->i_ceph_lock);
1410 	if (cap->mds_wanted | cap->issued)
1411 		ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1412 	__ceph_remove_cap(cap, false);
1413 	if (!ci->i_auth_cap) {
1414 		struct ceph_cap_flush *cf;
1415 		struct ceph_mds_client *mdsc = fsc->mdsc;
1416 
1417 		if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1418 			if (inode->i_data.nrpages > 0)
1419 				invalidate = true;
1420 			if (ci->i_wrbuffer_ref > 0)
1421 				mapping_set_error(&inode->i_data, -EIO);
1422 		}
1423 
1424 		while (!list_empty(&ci->i_cap_flush_list)) {
1425 			cf = list_first_entry(&ci->i_cap_flush_list,
1426 					      struct ceph_cap_flush, i_list);
1427 			list_move(&cf->i_list, &to_remove);
1428 		}
1429 
1430 		spin_lock(&mdsc->cap_dirty_lock);
1431 
1432 		list_for_each_entry(cf, &to_remove, i_list)
1433 			list_del(&cf->g_list);
1434 
1435 		if (!list_empty(&ci->i_dirty_item)) {
1436 			pr_warn_ratelimited(
1437 				" dropping dirty %s state for %p %lld\n",
1438 				ceph_cap_string(ci->i_dirty_caps),
1439 				inode, ceph_ino(inode));
1440 			ci->i_dirty_caps = 0;
1441 			list_del_init(&ci->i_dirty_item);
1442 			dirty_dropped = true;
1443 		}
1444 		if (!list_empty(&ci->i_flushing_item)) {
1445 			pr_warn_ratelimited(
1446 				" dropping dirty+flushing %s state for %p %lld\n",
1447 				ceph_cap_string(ci->i_flushing_caps),
1448 				inode, ceph_ino(inode));
1449 			ci->i_flushing_caps = 0;
1450 			list_del_init(&ci->i_flushing_item);
1451 			mdsc->num_cap_flushing--;
1452 			dirty_dropped = true;
1453 		}
1454 		spin_unlock(&mdsc->cap_dirty_lock);
1455 
1456 		if (dirty_dropped) {
1457 			errseq_set(&ci->i_meta_err, -EIO);
1458 
1459 			if (ci->i_wrbuffer_ref_head == 0 &&
1460 			    ci->i_wr_ref == 0 &&
1461 			    ci->i_dirty_caps == 0 &&
1462 			    ci->i_flushing_caps == 0) {
1463 				ceph_put_snap_context(ci->i_head_snapc);
1464 				ci->i_head_snapc = NULL;
1465 			}
1466 		}
1467 
1468 		if (atomic_read(&ci->i_filelock_ref) > 0) {
1469 			/* make further file lock syscall return -EIO */
1470 			ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1471 			pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1472 					    inode, ceph_ino(inode));
1473 		}
1474 
1475 		if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1476 			list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1477 			ci->i_prealloc_cap_flush = NULL;
1478 		}
1479 	}
1480 	spin_unlock(&ci->i_ceph_lock);
1481 	while (!list_empty(&to_remove)) {
1482 		struct ceph_cap_flush *cf;
1483 		cf = list_first_entry(&to_remove,
1484 				      struct ceph_cap_flush, i_list);
1485 		list_del(&cf->i_list);
1486 		ceph_free_cap_flush(cf);
1487 	}
1488 
1489 	wake_up_all(&ci->i_cap_wq);
1490 	if (invalidate)
1491 		ceph_queue_invalidate(inode);
1492 	if (dirty_dropped)
1493 		iput(inode);
1494 	return 0;
1495 }
1496 
1497 /*
1498  * caller must hold session s_mutex
1499  */
1500 static void remove_session_caps(struct ceph_mds_session *session)
1501 {
1502 	struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1503 	struct super_block *sb = fsc->sb;
1504 	LIST_HEAD(dispose);
1505 
1506 	dout("remove_session_caps on %p\n", session);
1507 	ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1508 
1509 	wake_up_all(&fsc->mdsc->cap_flushing_wq);
1510 
1511 	spin_lock(&session->s_cap_lock);
1512 	if (session->s_nr_caps > 0) {
1513 		struct inode *inode;
1514 		struct ceph_cap *cap, *prev = NULL;
1515 		struct ceph_vino vino;
1516 		/*
1517 		 * iterate_session_caps() skips inodes that are being
1518 		 * deleted, we need to wait until deletions are complete.
1519 		 * __wait_on_freeing_inode() is designed for the job,
1520 		 * but it is not exported, so use lookup inode function
1521 		 * to access it.
1522 		 */
1523 		while (!list_empty(&session->s_caps)) {
1524 			cap = list_entry(session->s_caps.next,
1525 					 struct ceph_cap, session_caps);
1526 			if (cap == prev)
1527 				break;
1528 			prev = cap;
1529 			vino = cap->ci->i_vino;
1530 			spin_unlock(&session->s_cap_lock);
1531 
1532 			inode = ceph_find_inode(sb, vino);
1533 			 /* avoid calling iput_final() while holding s_mutex */
1534 			ceph_async_iput(inode);
1535 
1536 			spin_lock(&session->s_cap_lock);
1537 		}
1538 	}
1539 
1540 	// drop cap expires and unlock s_cap_lock
1541 	detach_cap_releases(session, &dispose);
1542 
1543 	BUG_ON(session->s_nr_caps > 0);
1544 	BUG_ON(!list_empty(&session->s_cap_flushing));
1545 	spin_unlock(&session->s_cap_lock);
1546 	dispose_cap_releases(session->s_mdsc, &dispose);
1547 }
1548 
1549 enum {
1550 	RECONNECT,
1551 	RENEWCAPS,
1552 	FORCE_RO,
1553 };
1554 
1555 /*
1556  * wake up any threads waiting on this session's caps.  if the cap is
1557  * old (didn't get renewed on the client reconnect), remove it now.
1558  *
1559  * caller must hold s_mutex.
1560  */
1561 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1562 			      void *arg)
1563 {
1564 	struct ceph_inode_info *ci = ceph_inode(inode);
1565 	unsigned long ev = (unsigned long)arg;
1566 
1567 	if (ev == RECONNECT) {
1568 		spin_lock(&ci->i_ceph_lock);
1569 		ci->i_wanted_max_size = 0;
1570 		ci->i_requested_max_size = 0;
1571 		spin_unlock(&ci->i_ceph_lock);
1572 	} else if (ev == RENEWCAPS) {
1573 		if (cap->cap_gen < cap->session->s_cap_gen) {
1574 			/* mds did not re-issue stale cap */
1575 			spin_lock(&ci->i_ceph_lock);
1576 			cap->issued = cap->implemented = CEPH_CAP_PIN;
1577 			/* make sure mds knows what we want */
1578 			if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1579 				ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1580 			spin_unlock(&ci->i_ceph_lock);
1581 		}
1582 	} else if (ev == FORCE_RO) {
1583 	}
1584 	wake_up_all(&ci->i_cap_wq);
1585 	return 0;
1586 }
1587 
1588 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1589 {
1590 	dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1591 	ceph_iterate_session_caps(session, wake_up_session_cb,
1592 				  (void *)(unsigned long)ev);
1593 }
1594 
1595 /*
1596  * Send periodic message to MDS renewing all currently held caps.  The
1597  * ack will reset the expiration for all caps from this session.
1598  *
1599  * caller holds s_mutex
1600  */
1601 static int send_renew_caps(struct ceph_mds_client *mdsc,
1602 			   struct ceph_mds_session *session)
1603 {
1604 	struct ceph_msg *msg;
1605 	int state;
1606 
1607 	if (time_after_eq(jiffies, session->s_cap_ttl) &&
1608 	    time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1609 		pr_info("mds%d caps stale\n", session->s_mds);
1610 	session->s_renew_requested = jiffies;
1611 
1612 	/* do not try to renew caps until a recovering mds has reconnected
1613 	 * with its clients. */
1614 	state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1615 	if (state < CEPH_MDS_STATE_RECONNECT) {
1616 		dout("send_renew_caps ignoring mds%d (%s)\n",
1617 		     session->s_mds, ceph_mds_state_name(state));
1618 		return 0;
1619 	}
1620 
1621 	dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1622 		ceph_mds_state_name(state));
1623 	msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1624 				 ++session->s_renew_seq);
1625 	if (!msg)
1626 		return -ENOMEM;
1627 	ceph_con_send(&session->s_con, msg);
1628 	return 0;
1629 }
1630 
1631 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1632 			     struct ceph_mds_session *session, u64 seq)
1633 {
1634 	struct ceph_msg *msg;
1635 
1636 	dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1637 	     session->s_mds, ceph_session_state_name(session->s_state), seq);
1638 	msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1639 	if (!msg)
1640 		return -ENOMEM;
1641 	ceph_con_send(&session->s_con, msg);
1642 	return 0;
1643 }
1644 
1645 
1646 /*
1647  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1648  *
1649  * Called under session->s_mutex
1650  */
1651 static void renewed_caps(struct ceph_mds_client *mdsc,
1652 			 struct ceph_mds_session *session, int is_renew)
1653 {
1654 	int was_stale;
1655 	int wake = 0;
1656 
1657 	spin_lock(&session->s_cap_lock);
1658 	was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1659 
1660 	session->s_cap_ttl = session->s_renew_requested +
1661 		mdsc->mdsmap->m_session_timeout*HZ;
1662 
1663 	if (was_stale) {
1664 		if (time_before(jiffies, session->s_cap_ttl)) {
1665 			pr_info("mds%d caps renewed\n", session->s_mds);
1666 			wake = 1;
1667 		} else {
1668 			pr_info("mds%d caps still stale\n", session->s_mds);
1669 		}
1670 	}
1671 	dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1672 	     session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1673 	     time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1674 	spin_unlock(&session->s_cap_lock);
1675 
1676 	if (wake)
1677 		wake_up_session_caps(session, RENEWCAPS);
1678 }
1679 
1680 /*
1681  * send a session close request
1682  */
1683 static int request_close_session(struct ceph_mds_client *mdsc,
1684 				 struct ceph_mds_session *session)
1685 {
1686 	struct ceph_msg *msg;
1687 
1688 	dout("request_close_session mds%d state %s seq %lld\n",
1689 	     session->s_mds, ceph_session_state_name(session->s_state),
1690 	     session->s_seq);
1691 	msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1692 	if (!msg)
1693 		return -ENOMEM;
1694 	ceph_con_send(&session->s_con, msg);
1695 	return 1;
1696 }
1697 
1698 /*
1699  * Called with s_mutex held.
1700  */
1701 static int __close_session(struct ceph_mds_client *mdsc,
1702 			 struct ceph_mds_session *session)
1703 {
1704 	if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1705 		return 0;
1706 	session->s_state = CEPH_MDS_SESSION_CLOSING;
1707 	return request_close_session(mdsc, session);
1708 }
1709 
1710 static bool drop_negative_children(struct dentry *dentry)
1711 {
1712 	struct dentry *child;
1713 	bool all_negative = true;
1714 
1715 	if (!d_is_dir(dentry))
1716 		goto out;
1717 
1718 	spin_lock(&dentry->d_lock);
1719 	list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1720 		if (d_really_is_positive(child)) {
1721 			all_negative = false;
1722 			break;
1723 		}
1724 	}
1725 	spin_unlock(&dentry->d_lock);
1726 
1727 	if (all_negative)
1728 		shrink_dcache_parent(dentry);
1729 out:
1730 	return all_negative;
1731 }
1732 
1733 /*
1734  * Trim old(er) caps.
1735  *
1736  * Because we can't cache an inode without one or more caps, we do
1737  * this indirectly: if a cap is unused, we prune its aliases, at which
1738  * point the inode will hopefully get dropped to.
1739  *
1740  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1741  * memory pressure from the MDS, though, so it needn't be perfect.
1742  */
1743 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1744 {
1745 	int *remaining = arg;
1746 	struct ceph_inode_info *ci = ceph_inode(inode);
1747 	int used, wanted, oissued, mine;
1748 
1749 	if (*remaining <= 0)
1750 		return -1;
1751 
1752 	spin_lock(&ci->i_ceph_lock);
1753 	mine = cap->issued | cap->implemented;
1754 	used = __ceph_caps_used(ci);
1755 	wanted = __ceph_caps_file_wanted(ci);
1756 	oissued = __ceph_caps_issued_other(ci, cap);
1757 
1758 	dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1759 	     inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1760 	     ceph_cap_string(used), ceph_cap_string(wanted));
1761 	if (cap == ci->i_auth_cap) {
1762 		if (ci->i_dirty_caps || ci->i_flushing_caps ||
1763 		    !list_empty(&ci->i_cap_snaps))
1764 			goto out;
1765 		if ((used | wanted) & CEPH_CAP_ANY_WR)
1766 			goto out;
1767 		/* Note: it's possible that i_filelock_ref becomes non-zero
1768 		 * after dropping auth caps. It doesn't hurt because reply
1769 		 * of lock mds request will re-add auth caps. */
1770 		if (atomic_read(&ci->i_filelock_ref) > 0)
1771 			goto out;
1772 	}
1773 	/* The inode has cached pages, but it's no longer used.
1774 	 * we can safely drop it */
1775 	if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1776 	    !(oissued & CEPH_CAP_FILE_CACHE)) {
1777 	  used = 0;
1778 	  oissued = 0;
1779 	}
1780 	if ((used | wanted) & ~oissued & mine)
1781 		goto out;   /* we need these caps */
1782 
1783 	if (oissued) {
1784 		/* we aren't the only cap.. just remove us */
1785 		__ceph_remove_cap(cap, true);
1786 		(*remaining)--;
1787 	} else {
1788 		struct dentry *dentry;
1789 		/* try dropping referring dentries */
1790 		spin_unlock(&ci->i_ceph_lock);
1791 		dentry = d_find_any_alias(inode);
1792 		if (dentry && drop_negative_children(dentry)) {
1793 			int count;
1794 			dput(dentry);
1795 			d_prune_aliases(inode);
1796 			count = atomic_read(&inode->i_count);
1797 			if (count == 1)
1798 				(*remaining)--;
1799 			dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1800 			     inode, cap, count);
1801 		} else {
1802 			dput(dentry);
1803 		}
1804 		return 0;
1805 	}
1806 
1807 out:
1808 	spin_unlock(&ci->i_ceph_lock);
1809 	return 0;
1810 }
1811 
1812 /*
1813  * Trim session cap count down to some max number.
1814  */
1815 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1816 		   struct ceph_mds_session *session,
1817 		   int max_caps)
1818 {
1819 	int trim_caps = session->s_nr_caps - max_caps;
1820 
1821 	dout("trim_caps mds%d start: %d / %d, trim %d\n",
1822 	     session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1823 	if (trim_caps > 0) {
1824 		int remaining = trim_caps;
1825 
1826 		ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1827 		dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1828 		     session->s_mds, session->s_nr_caps, max_caps,
1829 			trim_caps - remaining);
1830 	}
1831 
1832 	ceph_flush_cap_releases(mdsc, session);
1833 	return 0;
1834 }
1835 
1836 static int check_caps_flush(struct ceph_mds_client *mdsc,
1837 			    u64 want_flush_tid)
1838 {
1839 	int ret = 1;
1840 
1841 	spin_lock(&mdsc->cap_dirty_lock);
1842 	if (!list_empty(&mdsc->cap_flush_list)) {
1843 		struct ceph_cap_flush *cf =
1844 			list_first_entry(&mdsc->cap_flush_list,
1845 					 struct ceph_cap_flush, g_list);
1846 		if (cf->tid <= want_flush_tid) {
1847 			dout("check_caps_flush still flushing tid "
1848 			     "%llu <= %llu\n", cf->tid, want_flush_tid);
1849 			ret = 0;
1850 		}
1851 	}
1852 	spin_unlock(&mdsc->cap_dirty_lock);
1853 	return ret;
1854 }
1855 
1856 /*
1857  * flush all dirty inode data to disk.
1858  *
1859  * returns true if we've flushed through want_flush_tid
1860  */
1861 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1862 			    u64 want_flush_tid)
1863 {
1864 	dout("check_caps_flush want %llu\n", want_flush_tid);
1865 
1866 	wait_event(mdsc->cap_flushing_wq,
1867 		   check_caps_flush(mdsc, want_flush_tid));
1868 
1869 	dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1870 }
1871 
1872 /*
1873  * called under s_mutex
1874  */
1875 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1876 				   struct ceph_mds_session *session)
1877 {
1878 	struct ceph_msg *msg = NULL;
1879 	struct ceph_mds_cap_release *head;
1880 	struct ceph_mds_cap_item *item;
1881 	struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1882 	struct ceph_cap *cap;
1883 	LIST_HEAD(tmp_list);
1884 	int num_cap_releases;
1885 	__le32	barrier, *cap_barrier;
1886 
1887 	down_read(&osdc->lock);
1888 	barrier = cpu_to_le32(osdc->epoch_barrier);
1889 	up_read(&osdc->lock);
1890 
1891 	spin_lock(&session->s_cap_lock);
1892 again:
1893 	list_splice_init(&session->s_cap_releases, &tmp_list);
1894 	num_cap_releases = session->s_num_cap_releases;
1895 	session->s_num_cap_releases = 0;
1896 	spin_unlock(&session->s_cap_lock);
1897 
1898 	while (!list_empty(&tmp_list)) {
1899 		if (!msg) {
1900 			msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1901 					PAGE_SIZE, GFP_NOFS, false);
1902 			if (!msg)
1903 				goto out_err;
1904 			head = msg->front.iov_base;
1905 			head->num = cpu_to_le32(0);
1906 			msg->front.iov_len = sizeof(*head);
1907 
1908 			msg->hdr.version = cpu_to_le16(2);
1909 			msg->hdr.compat_version = cpu_to_le16(1);
1910 		}
1911 
1912 		cap = list_first_entry(&tmp_list, struct ceph_cap,
1913 					session_caps);
1914 		list_del(&cap->session_caps);
1915 		num_cap_releases--;
1916 
1917 		head = msg->front.iov_base;
1918 		put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1919 				   &head->num);
1920 		item = msg->front.iov_base + msg->front.iov_len;
1921 		item->ino = cpu_to_le64(cap->cap_ino);
1922 		item->cap_id = cpu_to_le64(cap->cap_id);
1923 		item->migrate_seq = cpu_to_le32(cap->mseq);
1924 		item->seq = cpu_to_le32(cap->issue_seq);
1925 		msg->front.iov_len += sizeof(*item);
1926 
1927 		ceph_put_cap(mdsc, cap);
1928 
1929 		if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1930 			// Append cap_barrier field
1931 			cap_barrier = msg->front.iov_base + msg->front.iov_len;
1932 			*cap_barrier = barrier;
1933 			msg->front.iov_len += sizeof(*cap_barrier);
1934 
1935 			msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1936 			dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1937 			ceph_con_send(&session->s_con, msg);
1938 			msg = NULL;
1939 		}
1940 	}
1941 
1942 	BUG_ON(num_cap_releases != 0);
1943 
1944 	spin_lock(&session->s_cap_lock);
1945 	if (!list_empty(&session->s_cap_releases))
1946 		goto again;
1947 	spin_unlock(&session->s_cap_lock);
1948 
1949 	if (msg) {
1950 		// Append cap_barrier field
1951 		cap_barrier = msg->front.iov_base + msg->front.iov_len;
1952 		*cap_barrier = barrier;
1953 		msg->front.iov_len += sizeof(*cap_barrier);
1954 
1955 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1956 		dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1957 		ceph_con_send(&session->s_con, msg);
1958 	}
1959 	return;
1960 out_err:
1961 	pr_err("send_cap_releases mds%d, failed to allocate message\n",
1962 		session->s_mds);
1963 	spin_lock(&session->s_cap_lock);
1964 	list_splice(&tmp_list, &session->s_cap_releases);
1965 	session->s_num_cap_releases += num_cap_releases;
1966 	spin_unlock(&session->s_cap_lock);
1967 }
1968 
1969 static void ceph_cap_release_work(struct work_struct *work)
1970 {
1971 	struct ceph_mds_session *session =
1972 		container_of(work, struct ceph_mds_session, s_cap_release_work);
1973 
1974 	mutex_lock(&session->s_mutex);
1975 	if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1976 	    session->s_state == CEPH_MDS_SESSION_HUNG)
1977 		ceph_send_cap_releases(session->s_mdsc, session);
1978 	mutex_unlock(&session->s_mutex);
1979 	ceph_put_mds_session(session);
1980 }
1981 
1982 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1983 		             struct ceph_mds_session *session)
1984 {
1985 	if (mdsc->stopping)
1986 		return;
1987 
1988 	ceph_get_mds_session(session);
1989 	if (queue_work(mdsc->fsc->cap_wq,
1990 		       &session->s_cap_release_work)) {
1991 		dout("cap release work queued\n");
1992 	} else {
1993 		ceph_put_mds_session(session);
1994 		dout("failed to queue cap release work\n");
1995 	}
1996 }
1997 
1998 /*
1999  * caller holds session->s_cap_lock
2000  */
2001 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2002 			      struct ceph_cap *cap)
2003 {
2004 	list_add_tail(&cap->session_caps, &session->s_cap_releases);
2005 	session->s_num_cap_releases++;
2006 
2007 	if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2008 		ceph_flush_cap_releases(session->s_mdsc, session);
2009 }
2010 
2011 static void ceph_cap_reclaim_work(struct work_struct *work)
2012 {
2013 	struct ceph_mds_client *mdsc =
2014 		container_of(work, struct ceph_mds_client, cap_reclaim_work);
2015 	int ret = ceph_trim_dentries(mdsc);
2016 	if (ret == -EAGAIN)
2017 		ceph_queue_cap_reclaim_work(mdsc);
2018 }
2019 
2020 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2021 {
2022 	if (mdsc->stopping)
2023 		return;
2024 
2025         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2026                 dout("caps reclaim work queued\n");
2027         } else {
2028                 dout("failed to queue caps release work\n");
2029         }
2030 }
2031 
2032 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2033 {
2034 	int val;
2035 	if (!nr)
2036 		return;
2037 	val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2038 	if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2039 		atomic_set(&mdsc->cap_reclaim_pending, 0);
2040 		ceph_queue_cap_reclaim_work(mdsc);
2041 	}
2042 }
2043 
2044 /*
2045  * requests
2046  */
2047 
2048 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2049 				    struct inode *dir)
2050 {
2051 	struct ceph_inode_info *ci = ceph_inode(dir);
2052 	struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2053 	struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2054 	size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2055 	unsigned int num_entries;
2056 	int order;
2057 
2058 	spin_lock(&ci->i_ceph_lock);
2059 	num_entries = ci->i_files + ci->i_subdirs;
2060 	spin_unlock(&ci->i_ceph_lock);
2061 	num_entries = max(num_entries, 1U);
2062 	num_entries = min(num_entries, opt->max_readdir);
2063 
2064 	order = get_order(size * num_entries);
2065 	while (order >= 0) {
2066 		rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2067 							     __GFP_NOWARN,
2068 							     order);
2069 		if (rinfo->dir_entries)
2070 			break;
2071 		order--;
2072 	}
2073 	if (!rinfo->dir_entries)
2074 		return -ENOMEM;
2075 
2076 	num_entries = (PAGE_SIZE << order) / size;
2077 	num_entries = min(num_entries, opt->max_readdir);
2078 
2079 	rinfo->dir_buf_size = PAGE_SIZE << order;
2080 	req->r_num_caps = num_entries + 1;
2081 	req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2082 	req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2083 	return 0;
2084 }
2085 
2086 /*
2087  * Create an mds request.
2088  */
2089 struct ceph_mds_request *
2090 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2091 {
2092 	struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2093 
2094 	if (!req)
2095 		return ERR_PTR(-ENOMEM);
2096 
2097 	mutex_init(&req->r_fill_mutex);
2098 	req->r_mdsc = mdsc;
2099 	req->r_started = jiffies;
2100 	req->r_resend_mds = -1;
2101 	INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2102 	INIT_LIST_HEAD(&req->r_unsafe_target_item);
2103 	req->r_fmode = -1;
2104 	kref_init(&req->r_kref);
2105 	RB_CLEAR_NODE(&req->r_node);
2106 	INIT_LIST_HEAD(&req->r_wait);
2107 	init_completion(&req->r_completion);
2108 	init_completion(&req->r_safe_completion);
2109 	INIT_LIST_HEAD(&req->r_unsafe_item);
2110 
2111 	ktime_get_coarse_real_ts64(&req->r_stamp);
2112 
2113 	req->r_op = op;
2114 	req->r_direct_mode = mode;
2115 	return req;
2116 }
2117 
2118 /*
2119  * return oldest (lowest) request, tid in request tree, 0 if none.
2120  *
2121  * called under mdsc->mutex.
2122  */
2123 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2124 {
2125 	if (RB_EMPTY_ROOT(&mdsc->request_tree))
2126 		return NULL;
2127 	return rb_entry(rb_first(&mdsc->request_tree),
2128 			struct ceph_mds_request, r_node);
2129 }
2130 
2131 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2132 {
2133 	return mdsc->oldest_tid;
2134 }
2135 
2136 /*
2137  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2138  * on build_path_from_dentry in fs/cifs/dir.c.
2139  *
2140  * If @stop_on_nosnap, generate path relative to the first non-snapped
2141  * inode.
2142  *
2143  * Encode hidden .snap dirs as a double /, i.e.
2144  *   foo/.snap/bar -> foo//bar
2145  */
2146 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2147 			   int stop_on_nosnap)
2148 {
2149 	struct dentry *temp;
2150 	char *path;
2151 	int pos;
2152 	unsigned seq;
2153 	u64 base;
2154 
2155 	if (!dentry)
2156 		return ERR_PTR(-EINVAL);
2157 
2158 	path = __getname();
2159 	if (!path)
2160 		return ERR_PTR(-ENOMEM);
2161 retry:
2162 	pos = PATH_MAX - 1;
2163 	path[pos] = '\0';
2164 
2165 	seq = read_seqbegin(&rename_lock);
2166 	rcu_read_lock();
2167 	temp = dentry;
2168 	for (;;) {
2169 		struct inode *inode;
2170 
2171 		spin_lock(&temp->d_lock);
2172 		inode = d_inode(temp);
2173 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2174 			dout("build_path path+%d: %p SNAPDIR\n",
2175 			     pos, temp);
2176 		} else if (stop_on_nosnap && inode && dentry != temp &&
2177 			   ceph_snap(inode) == CEPH_NOSNAP) {
2178 			spin_unlock(&temp->d_lock);
2179 			pos++; /* get rid of any prepended '/' */
2180 			break;
2181 		} else {
2182 			pos -= temp->d_name.len;
2183 			if (pos < 0) {
2184 				spin_unlock(&temp->d_lock);
2185 				break;
2186 			}
2187 			memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2188 		}
2189 		spin_unlock(&temp->d_lock);
2190 		temp = READ_ONCE(temp->d_parent);
2191 
2192 		/* Are we at the root? */
2193 		if (IS_ROOT(temp))
2194 			break;
2195 
2196 		/* Are we out of buffer? */
2197 		if (--pos < 0)
2198 			break;
2199 
2200 		path[pos] = '/';
2201 	}
2202 	base = ceph_ino(d_inode(temp));
2203 	rcu_read_unlock();
2204 
2205 	if (read_seqretry(&rename_lock, seq))
2206 		goto retry;
2207 
2208 	if (pos < 0) {
2209 		/*
2210 		 * A rename didn't occur, but somehow we didn't end up where
2211 		 * we thought we would. Throw a warning and try again.
2212 		 */
2213 		pr_warn("build_path did not end path lookup where "
2214 			"expected, pos is %d\n", pos);
2215 		goto retry;
2216 	}
2217 
2218 	*pbase = base;
2219 	*plen = PATH_MAX - 1 - pos;
2220 	dout("build_path on %p %d built %llx '%.*s'\n",
2221 	     dentry, d_count(dentry), base, *plen, path + pos);
2222 	return path + pos;
2223 }
2224 
2225 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2226 			     const char **ppath, int *ppathlen, u64 *pino,
2227 			     bool *pfreepath, bool parent_locked)
2228 {
2229 	char *path;
2230 
2231 	rcu_read_lock();
2232 	if (!dir)
2233 		dir = d_inode_rcu(dentry->d_parent);
2234 	if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2235 		*pino = ceph_ino(dir);
2236 		rcu_read_unlock();
2237 		*ppath = dentry->d_name.name;
2238 		*ppathlen = dentry->d_name.len;
2239 		return 0;
2240 	}
2241 	rcu_read_unlock();
2242 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2243 	if (IS_ERR(path))
2244 		return PTR_ERR(path);
2245 	*ppath = path;
2246 	*pfreepath = true;
2247 	return 0;
2248 }
2249 
2250 static int build_inode_path(struct inode *inode,
2251 			    const char **ppath, int *ppathlen, u64 *pino,
2252 			    bool *pfreepath)
2253 {
2254 	struct dentry *dentry;
2255 	char *path;
2256 
2257 	if (ceph_snap(inode) == CEPH_NOSNAP) {
2258 		*pino = ceph_ino(inode);
2259 		*ppathlen = 0;
2260 		return 0;
2261 	}
2262 	dentry = d_find_alias(inode);
2263 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2264 	dput(dentry);
2265 	if (IS_ERR(path))
2266 		return PTR_ERR(path);
2267 	*ppath = path;
2268 	*pfreepath = true;
2269 	return 0;
2270 }
2271 
2272 /*
2273  * request arguments may be specified via an inode *, a dentry *, or
2274  * an explicit ino+path.
2275  */
2276 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2277 				  struct inode *rdiri, const char *rpath,
2278 				  u64 rino, const char **ppath, int *pathlen,
2279 				  u64 *ino, bool *freepath, bool parent_locked)
2280 {
2281 	int r = 0;
2282 
2283 	if (rinode) {
2284 		r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2285 		dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2286 		     ceph_snap(rinode));
2287 	} else if (rdentry) {
2288 		r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2289 					freepath, parent_locked);
2290 		dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2291 		     *ppath);
2292 	} else if (rpath || rino) {
2293 		*ino = rino;
2294 		*ppath = rpath;
2295 		*pathlen = rpath ? strlen(rpath) : 0;
2296 		dout(" path %.*s\n", *pathlen, rpath);
2297 	}
2298 
2299 	return r;
2300 }
2301 
2302 /*
2303  * called under mdsc->mutex
2304  */
2305 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2306 					       struct ceph_mds_request *req,
2307 					       int mds, bool drop_cap_releases)
2308 {
2309 	struct ceph_msg *msg;
2310 	struct ceph_mds_request_head *head;
2311 	const char *path1 = NULL;
2312 	const char *path2 = NULL;
2313 	u64 ino1 = 0, ino2 = 0;
2314 	int pathlen1 = 0, pathlen2 = 0;
2315 	bool freepath1 = false, freepath2 = false;
2316 	int len;
2317 	u16 releases;
2318 	void *p, *end;
2319 	int ret;
2320 
2321 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
2322 			      req->r_parent, req->r_path1, req->r_ino1.ino,
2323 			      &path1, &pathlen1, &ino1, &freepath1,
2324 			      test_bit(CEPH_MDS_R_PARENT_LOCKED,
2325 					&req->r_req_flags));
2326 	if (ret < 0) {
2327 		msg = ERR_PTR(ret);
2328 		goto out;
2329 	}
2330 
2331 	/* If r_old_dentry is set, then assume that its parent is locked */
2332 	ret = set_request_path_attr(NULL, req->r_old_dentry,
2333 			      req->r_old_dentry_dir,
2334 			      req->r_path2, req->r_ino2.ino,
2335 			      &path2, &pathlen2, &ino2, &freepath2, true);
2336 	if (ret < 0) {
2337 		msg = ERR_PTR(ret);
2338 		goto out_free1;
2339 	}
2340 
2341 	len = sizeof(*head) +
2342 		pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2343 		sizeof(struct ceph_timespec);
2344 
2345 	/* calculate (max) length for cap releases */
2346 	len += sizeof(struct ceph_mds_request_release) *
2347 		(!!req->r_inode_drop + !!req->r_dentry_drop +
2348 		 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2349 	if (req->r_dentry_drop)
2350 		len += pathlen1;
2351 	if (req->r_old_dentry_drop)
2352 		len += pathlen2;
2353 
2354 	msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2355 	if (!msg) {
2356 		msg = ERR_PTR(-ENOMEM);
2357 		goto out_free2;
2358 	}
2359 
2360 	msg->hdr.version = cpu_to_le16(2);
2361 	msg->hdr.tid = cpu_to_le64(req->r_tid);
2362 
2363 	head = msg->front.iov_base;
2364 	p = msg->front.iov_base + sizeof(*head);
2365 	end = msg->front.iov_base + msg->front.iov_len;
2366 
2367 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2368 	head->op = cpu_to_le32(req->r_op);
2369 	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2370 	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2371 	head->ino = 0;
2372 	head->args = req->r_args;
2373 
2374 	ceph_encode_filepath(&p, end, ino1, path1);
2375 	ceph_encode_filepath(&p, end, ino2, path2);
2376 
2377 	/* make note of release offset, in case we need to replay */
2378 	req->r_request_release_offset = p - msg->front.iov_base;
2379 
2380 	/* cap releases */
2381 	releases = 0;
2382 	if (req->r_inode_drop)
2383 		releases += ceph_encode_inode_release(&p,
2384 		      req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2385 		      mds, req->r_inode_drop, req->r_inode_unless, 0);
2386 	if (req->r_dentry_drop)
2387 		releases += ceph_encode_dentry_release(&p, req->r_dentry,
2388 				req->r_parent, mds, req->r_dentry_drop,
2389 				req->r_dentry_unless);
2390 	if (req->r_old_dentry_drop)
2391 		releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2392 				req->r_old_dentry_dir, mds,
2393 				req->r_old_dentry_drop,
2394 				req->r_old_dentry_unless);
2395 	if (req->r_old_inode_drop)
2396 		releases += ceph_encode_inode_release(&p,
2397 		      d_inode(req->r_old_dentry),
2398 		      mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2399 
2400 	if (drop_cap_releases) {
2401 		releases = 0;
2402 		p = msg->front.iov_base + req->r_request_release_offset;
2403 	}
2404 
2405 	head->num_releases = cpu_to_le16(releases);
2406 
2407 	/* time stamp */
2408 	{
2409 		struct ceph_timespec ts;
2410 		ceph_encode_timespec64(&ts, &req->r_stamp);
2411 		ceph_encode_copy(&p, &ts, sizeof(ts));
2412 	}
2413 
2414 	BUG_ON(p > end);
2415 	msg->front.iov_len = p - msg->front.iov_base;
2416 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2417 
2418 	if (req->r_pagelist) {
2419 		struct ceph_pagelist *pagelist = req->r_pagelist;
2420 		ceph_msg_data_add_pagelist(msg, pagelist);
2421 		msg->hdr.data_len = cpu_to_le32(pagelist->length);
2422 	} else {
2423 		msg->hdr.data_len = 0;
2424 	}
2425 
2426 	msg->hdr.data_off = cpu_to_le16(0);
2427 
2428 out_free2:
2429 	if (freepath2)
2430 		ceph_mdsc_free_path((char *)path2, pathlen2);
2431 out_free1:
2432 	if (freepath1)
2433 		ceph_mdsc_free_path((char *)path1, pathlen1);
2434 out:
2435 	return msg;
2436 }
2437 
2438 /*
2439  * called under mdsc->mutex if error, under no mutex if
2440  * success.
2441  */
2442 static void complete_request(struct ceph_mds_client *mdsc,
2443 			     struct ceph_mds_request *req)
2444 {
2445 	if (req->r_callback)
2446 		req->r_callback(mdsc, req);
2447 	complete_all(&req->r_completion);
2448 }
2449 
2450 /*
2451  * called under mdsc->mutex
2452  */
2453 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2454 				  struct ceph_mds_request *req,
2455 				  int mds, bool drop_cap_releases)
2456 {
2457 	struct ceph_mds_request_head *rhead;
2458 	struct ceph_msg *msg;
2459 	int flags = 0;
2460 
2461 	req->r_attempts++;
2462 	if (req->r_inode) {
2463 		struct ceph_cap *cap =
2464 			ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2465 
2466 		if (cap)
2467 			req->r_sent_on_mseq = cap->mseq;
2468 		else
2469 			req->r_sent_on_mseq = -1;
2470 	}
2471 	dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2472 	     req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2473 
2474 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2475 		void *p;
2476 		/*
2477 		 * Replay.  Do not regenerate message (and rebuild
2478 		 * paths, etc.); just use the original message.
2479 		 * Rebuilding paths will break for renames because
2480 		 * d_move mangles the src name.
2481 		 */
2482 		msg = req->r_request;
2483 		rhead = msg->front.iov_base;
2484 
2485 		flags = le32_to_cpu(rhead->flags);
2486 		flags |= CEPH_MDS_FLAG_REPLAY;
2487 		rhead->flags = cpu_to_le32(flags);
2488 
2489 		if (req->r_target_inode)
2490 			rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2491 
2492 		rhead->num_retry = req->r_attempts - 1;
2493 
2494 		/* remove cap/dentry releases from message */
2495 		rhead->num_releases = 0;
2496 
2497 		/* time stamp */
2498 		p = msg->front.iov_base + req->r_request_release_offset;
2499 		{
2500 			struct ceph_timespec ts;
2501 			ceph_encode_timespec64(&ts, &req->r_stamp);
2502 			ceph_encode_copy(&p, &ts, sizeof(ts));
2503 		}
2504 
2505 		msg->front.iov_len = p - msg->front.iov_base;
2506 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2507 		return 0;
2508 	}
2509 
2510 	if (req->r_request) {
2511 		ceph_msg_put(req->r_request);
2512 		req->r_request = NULL;
2513 	}
2514 	msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2515 	if (IS_ERR(msg)) {
2516 		req->r_err = PTR_ERR(msg);
2517 		return PTR_ERR(msg);
2518 	}
2519 	req->r_request = msg;
2520 
2521 	rhead = msg->front.iov_base;
2522 	rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2523 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2524 		flags |= CEPH_MDS_FLAG_REPLAY;
2525 	if (req->r_parent)
2526 		flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2527 	rhead->flags = cpu_to_le32(flags);
2528 	rhead->num_fwd = req->r_num_fwd;
2529 	rhead->num_retry = req->r_attempts - 1;
2530 	rhead->ino = 0;
2531 
2532 	dout(" r_parent = %p\n", req->r_parent);
2533 	return 0;
2534 }
2535 
2536 /*
2537  * called under mdsc->mutex
2538  */
2539 static int __send_request(struct ceph_mds_client *mdsc,
2540 			  struct ceph_mds_session *session,
2541 			  struct ceph_mds_request *req,
2542 			  bool drop_cap_releases)
2543 {
2544 	int err;
2545 
2546 	err = __prepare_send_request(mdsc, req, session->s_mds,
2547 				     drop_cap_releases);
2548 	if (!err) {
2549 		ceph_msg_get(req->r_request);
2550 		ceph_con_send(&session->s_con, req->r_request);
2551 	}
2552 
2553 	return err;
2554 }
2555 
2556 /*
2557  * send request, or put it on the appropriate wait list.
2558  */
2559 static void __do_request(struct ceph_mds_client *mdsc,
2560 			struct ceph_mds_request *req)
2561 {
2562 	struct ceph_mds_session *session = NULL;
2563 	int mds = -1;
2564 	int err = 0;
2565 	bool random;
2566 
2567 	if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2568 		if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2569 			__unregister_request(mdsc, req);
2570 		return;
2571 	}
2572 
2573 	if (req->r_timeout &&
2574 	    time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2575 		dout("do_request timed out\n");
2576 		err = -EIO;
2577 		goto finish;
2578 	}
2579 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2580 		dout("do_request forced umount\n");
2581 		err = -EIO;
2582 		goto finish;
2583 	}
2584 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2585 		if (mdsc->mdsmap_err) {
2586 			err = mdsc->mdsmap_err;
2587 			dout("do_request mdsmap err %d\n", err);
2588 			goto finish;
2589 		}
2590 		if (mdsc->mdsmap->m_epoch == 0) {
2591 			dout("do_request no mdsmap, waiting for map\n");
2592 			list_add(&req->r_wait, &mdsc->waiting_for_map);
2593 			return;
2594 		}
2595 		if (!(mdsc->fsc->mount_options->flags &
2596 		      CEPH_MOUNT_OPT_MOUNTWAIT) &&
2597 		    !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2598 			err = -EHOSTUNREACH;
2599 			goto finish;
2600 		}
2601 	}
2602 
2603 	put_request_session(req);
2604 
2605 	mds = __choose_mds(mdsc, req, &random);
2606 	if (mds < 0 ||
2607 	    ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2608 		dout("do_request no mds or not active, waiting for map\n");
2609 		list_add(&req->r_wait, &mdsc->waiting_for_map);
2610 		return;
2611 	}
2612 
2613 	/* get, open session */
2614 	session = __ceph_lookup_mds_session(mdsc, mds);
2615 	if (!session) {
2616 		session = register_session(mdsc, mds);
2617 		if (IS_ERR(session)) {
2618 			err = PTR_ERR(session);
2619 			goto finish;
2620 		}
2621 	}
2622 	req->r_session = ceph_get_mds_session(session);
2623 
2624 	dout("do_request mds%d session %p state %s\n", mds, session,
2625 	     ceph_session_state_name(session->s_state));
2626 	if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2627 	    session->s_state != CEPH_MDS_SESSION_HUNG) {
2628 		if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2629 			err = -EACCES;
2630 			goto out_session;
2631 		}
2632 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
2633 		    session->s_state == CEPH_MDS_SESSION_CLOSING) {
2634 			__open_session(mdsc, session);
2635 			/* retry the same mds later */
2636 			if (random)
2637 				req->r_resend_mds = mds;
2638 		}
2639 		list_add(&req->r_wait, &session->s_waiting);
2640 		goto out_session;
2641 	}
2642 
2643 	/* send request */
2644 	req->r_resend_mds = -1;   /* forget any previous mds hint */
2645 
2646 	if (req->r_request_started == 0)   /* note request start time */
2647 		req->r_request_started = jiffies;
2648 
2649 	err = __send_request(mdsc, session, req, false);
2650 
2651 out_session:
2652 	ceph_put_mds_session(session);
2653 finish:
2654 	if (err) {
2655 		dout("__do_request early error %d\n", err);
2656 		req->r_err = err;
2657 		complete_request(mdsc, req);
2658 		__unregister_request(mdsc, req);
2659 	}
2660 	return;
2661 }
2662 
2663 /*
2664  * called under mdsc->mutex
2665  */
2666 static void __wake_requests(struct ceph_mds_client *mdsc,
2667 			    struct list_head *head)
2668 {
2669 	struct ceph_mds_request *req;
2670 	LIST_HEAD(tmp_list);
2671 
2672 	list_splice_init(head, &tmp_list);
2673 
2674 	while (!list_empty(&tmp_list)) {
2675 		req = list_entry(tmp_list.next,
2676 				 struct ceph_mds_request, r_wait);
2677 		list_del_init(&req->r_wait);
2678 		dout(" wake request %p tid %llu\n", req, req->r_tid);
2679 		__do_request(mdsc, req);
2680 	}
2681 }
2682 
2683 /*
2684  * Wake up threads with requests pending for @mds, so that they can
2685  * resubmit their requests to a possibly different mds.
2686  */
2687 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2688 {
2689 	struct ceph_mds_request *req;
2690 	struct rb_node *p = rb_first(&mdsc->request_tree);
2691 
2692 	dout("kick_requests mds%d\n", mds);
2693 	while (p) {
2694 		req = rb_entry(p, struct ceph_mds_request, r_node);
2695 		p = rb_next(p);
2696 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2697 			continue;
2698 		if (req->r_attempts > 0)
2699 			continue; /* only new requests */
2700 		if (req->r_session &&
2701 		    req->r_session->s_mds == mds) {
2702 			dout(" kicking tid %llu\n", req->r_tid);
2703 			list_del_init(&req->r_wait);
2704 			__do_request(mdsc, req);
2705 		}
2706 	}
2707 }
2708 
2709 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2710 			      struct ceph_mds_request *req)
2711 {
2712 	int err;
2713 
2714 	/* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2715 	if (req->r_inode)
2716 		ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2717 	if (req->r_parent) {
2718 		ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2719 		ihold(req->r_parent);
2720 	}
2721 	if (req->r_old_dentry_dir)
2722 		ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2723 				  CEPH_CAP_PIN);
2724 
2725 	dout("submit_request on %p for inode %p\n", req, dir);
2726 	mutex_lock(&mdsc->mutex);
2727 	__register_request(mdsc, req, dir);
2728 	__do_request(mdsc, req);
2729 	err = req->r_err;
2730 	mutex_unlock(&mdsc->mutex);
2731 	return err;
2732 }
2733 
2734 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2735 				  struct ceph_mds_request *req)
2736 {
2737 	int err;
2738 
2739 	/* wait */
2740 	dout("do_request waiting\n");
2741 	if (!req->r_timeout && req->r_wait_for_completion) {
2742 		err = req->r_wait_for_completion(mdsc, req);
2743 	} else {
2744 		long timeleft = wait_for_completion_killable_timeout(
2745 					&req->r_completion,
2746 					ceph_timeout_jiffies(req->r_timeout));
2747 		if (timeleft > 0)
2748 			err = 0;
2749 		else if (!timeleft)
2750 			err = -EIO;  /* timed out */
2751 		else
2752 			err = timeleft;  /* killed */
2753 	}
2754 	dout("do_request waited, got %d\n", err);
2755 	mutex_lock(&mdsc->mutex);
2756 
2757 	/* only abort if we didn't race with a real reply */
2758 	if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2759 		err = le32_to_cpu(req->r_reply_info.head->result);
2760 	} else if (err < 0) {
2761 		dout("aborted request %lld with %d\n", req->r_tid, err);
2762 
2763 		/*
2764 		 * ensure we aren't running concurrently with
2765 		 * ceph_fill_trace or ceph_readdir_prepopulate, which
2766 		 * rely on locks (dir mutex) held by our caller.
2767 		 */
2768 		mutex_lock(&req->r_fill_mutex);
2769 		req->r_err = err;
2770 		set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2771 		mutex_unlock(&req->r_fill_mutex);
2772 
2773 		if (req->r_parent &&
2774 		    (req->r_op & CEPH_MDS_OP_WRITE))
2775 			ceph_invalidate_dir_request(req);
2776 	} else {
2777 		err = req->r_err;
2778 	}
2779 
2780 	mutex_unlock(&mdsc->mutex);
2781 	return err;
2782 }
2783 
2784 /*
2785  * Synchrously perform an mds request.  Take care of all of the
2786  * session setup, forwarding, retry details.
2787  */
2788 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2789 			 struct inode *dir,
2790 			 struct ceph_mds_request *req)
2791 {
2792 	int err;
2793 
2794 	dout("do_request on %p\n", req);
2795 
2796 	/* issue */
2797 	err = ceph_mdsc_submit_request(mdsc, dir, req);
2798 	if (!err)
2799 		err = ceph_mdsc_wait_request(mdsc, req);
2800 	dout("do_request %p done, result %d\n", req, err);
2801 	return err;
2802 }
2803 
2804 /*
2805  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2806  * namespace request.
2807  */
2808 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2809 {
2810 	struct inode *dir = req->r_parent;
2811 	struct inode *old_dir = req->r_old_dentry_dir;
2812 
2813 	dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2814 
2815 	ceph_dir_clear_complete(dir);
2816 	if (old_dir)
2817 		ceph_dir_clear_complete(old_dir);
2818 	if (req->r_dentry)
2819 		ceph_invalidate_dentry_lease(req->r_dentry);
2820 	if (req->r_old_dentry)
2821 		ceph_invalidate_dentry_lease(req->r_old_dentry);
2822 }
2823 
2824 /*
2825  * Handle mds reply.
2826  *
2827  * We take the session mutex and parse and process the reply immediately.
2828  * This preserves the logical ordering of replies, capabilities, etc., sent
2829  * by the MDS as they are applied to our local cache.
2830  */
2831 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2832 {
2833 	struct ceph_mds_client *mdsc = session->s_mdsc;
2834 	struct ceph_mds_request *req;
2835 	struct ceph_mds_reply_head *head = msg->front.iov_base;
2836 	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2837 	struct ceph_snap_realm *realm;
2838 	u64 tid;
2839 	int err, result;
2840 	int mds = session->s_mds;
2841 
2842 	if (msg->front.iov_len < sizeof(*head)) {
2843 		pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2844 		ceph_msg_dump(msg);
2845 		return;
2846 	}
2847 
2848 	/* get request, session */
2849 	tid = le64_to_cpu(msg->hdr.tid);
2850 	mutex_lock(&mdsc->mutex);
2851 	req = lookup_get_request(mdsc, tid);
2852 	if (!req) {
2853 		dout("handle_reply on unknown tid %llu\n", tid);
2854 		mutex_unlock(&mdsc->mutex);
2855 		return;
2856 	}
2857 	dout("handle_reply %p\n", req);
2858 
2859 	/* correct session? */
2860 	if (req->r_session != session) {
2861 		pr_err("mdsc_handle_reply got %llu on session mds%d"
2862 		       " not mds%d\n", tid, session->s_mds,
2863 		       req->r_session ? req->r_session->s_mds : -1);
2864 		mutex_unlock(&mdsc->mutex);
2865 		goto out;
2866 	}
2867 
2868 	/* dup? */
2869 	if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2870 	    (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2871 		pr_warn("got a dup %s reply on %llu from mds%d\n",
2872 			   head->safe ? "safe" : "unsafe", tid, mds);
2873 		mutex_unlock(&mdsc->mutex);
2874 		goto out;
2875 	}
2876 	if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2877 		pr_warn("got unsafe after safe on %llu from mds%d\n",
2878 			   tid, mds);
2879 		mutex_unlock(&mdsc->mutex);
2880 		goto out;
2881 	}
2882 
2883 	result = le32_to_cpu(head->result);
2884 
2885 	/*
2886 	 * Handle an ESTALE
2887 	 * if we're not talking to the authority, send to them
2888 	 * if the authority has changed while we weren't looking,
2889 	 * send to new authority
2890 	 * Otherwise we just have to return an ESTALE
2891 	 */
2892 	if (result == -ESTALE) {
2893 		dout("got ESTALE on request %llu\n", req->r_tid);
2894 		req->r_resend_mds = -1;
2895 		if (req->r_direct_mode != USE_AUTH_MDS) {
2896 			dout("not using auth, setting for that now\n");
2897 			req->r_direct_mode = USE_AUTH_MDS;
2898 			__do_request(mdsc, req);
2899 			mutex_unlock(&mdsc->mutex);
2900 			goto out;
2901 		} else  {
2902 			int mds = __choose_mds(mdsc, req, NULL);
2903 			if (mds >= 0 && mds != req->r_session->s_mds) {
2904 				dout("but auth changed, so resending\n");
2905 				__do_request(mdsc, req);
2906 				mutex_unlock(&mdsc->mutex);
2907 				goto out;
2908 			}
2909 		}
2910 		dout("have to return ESTALE on request %llu\n", req->r_tid);
2911 	}
2912 
2913 
2914 	if (head->safe) {
2915 		set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2916 		__unregister_request(mdsc, req);
2917 
2918 		/* last request during umount? */
2919 		if (mdsc->stopping && !__get_oldest_req(mdsc))
2920 			complete_all(&mdsc->safe_umount_waiters);
2921 
2922 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2923 			/*
2924 			 * We already handled the unsafe response, now do the
2925 			 * cleanup.  No need to examine the response; the MDS
2926 			 * doesn't include any result info in the safe
2927 			 * response.  And even if it did, there is nothing
2928 			 * useful we could do with a revised return value.
2929 			 */
2930 			dout("got safe reply %llu, mds%d\n", tid, mds);
2931 
2932 			mutex_unlock(&mdsc->mutex);
2933 			goto out;
2934 		}
2935 	} else {
2936 		set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2937 		list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2938 		if (req->r_unsafe_dir) {
2939 			struct ceph_inode_info *ci =
2940 					ceph_inode(req->r_unsafe_dir);
2941 			spin_lock(&ci->i_unsafe_lock);
2942 			list_add_tail(&req->r_unsafe_dir_item,
2943 				      &ci->i_unsafe_dirops);
2944 			spin_unlock(&ci->i_unsafe_lock);
2945 		}
2946 	}
2947 
2948 	dout("handle_reply tid %lld result %d\n", tid, result);
2949 	rinfo = &req->r_reply_info;
2950 	if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2951 		err = parse_reply_info(msg, rinfo, (u64)-1);
2952 	else
2953 		err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2954 	mutex_unlock(&mdsc->mutex);
2955 
2956 	mutex_lock(&session->s_mutex);
2957 	if (err < 0) {
2958 		pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2959 		ceph_msg_dump(msg);
2960 		goto out_err;
2961 	}
2962 
2963 	/* snap trace */
2964 	realm = NULL;
2965 	if (rinfo->snapblob_len) {
2966 		down_write(&mdsc->snap_rwsem);
2967 		ceph_update_snap_trace(mdsc, rinfo->snapblob,
2968 				rinfo->snapblob + rinfo->snapblob_len,
2969 				le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2970 				&realm);
2971 		downgrade_write(&mdsc->snap_rwsem);
2972 	} else {
2973 		down_read(&mdsc->snap_rwsem);
2974 	}
2975 
2976 	/* insert trace into our cache */
2977 	mutex_lock(&req->r_fill_mutex);
2978 	current->journal_info = req;
2979 	err = ceph_fill_trace(mdsc->fsc->sb, req);
2980 	if (err == 0) {
2981 		if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2982 				    req->r_op == CEPH_MDS_OP_LSSNAP))
2983 			ceph_readdir_prepopulate(req, req->r_session);
2984 	}
2985 	current->journal_info = NULL;
2986 	mutex_unlock(&req->r_fill_mutex);
2987 
2988 	up_read(&mdsc->snap_rwsem);
2989 	if (realm)
2990 		ceph_put_snap_realm(mdsc, realm);
2991 
2992 	if (err == 0) {
2993 		if (req->r_target_inode &&
2994 		    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2995 			struct ceph_inode_info *ci =
2996 				ceph_inode(req->r_target_inode);
2997 			spin_lock(&ci->i_unsafe_lock);
2998 			list_add_tail(&req->r_unsafe_target_item,
2999 				      &ci->i_unsafe_iops);
3000 			spin_unlock(&ci->i_unsafe_lock);
3001 		}
3002 
3003 		ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3004 	}
3005 out_err:
3006 	mutex_lock(&mdsc->mutex);
3007 	if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3008 		if (err) {
3009 			req->r_err = err;
3010 		} else {
3011 			req->r_reply =  ceph_msg_get(msg);
3012 			set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3013 		}
3014 	} else {
3015 		dout("reply arrived after request %lld was aborted\n", tid);
3016 	}
3017 	mutex_unlock(&mdsc->mutex);
3018 
3019 	mutex_unlock(&session->s_mutex);
3020 
3021 	/* kick calling process */
3022 	complete_request(mdsc, req);
3023 out:
3024 	ceph_mdsc_put_request(req);
3025 	return;
3026 }
3027 
3028 
3029 
3030 /*
3031  * handle mds notification that our request has been forwarded.
3032  */
3033 static void handle_forward(struct ceph_mds_client *mdsc,
3034 			   struct ceph_mds_session *session,
3035 			   struct ceph_msg *msg)
3036 {
3037 	struct ceph_mds_request *req;
3038 	u64 tid = le64_to_cpu(msg->hdr.tid);
3039 	u32 next_mds;
3040 	u32 fwd_seq;
3041 	int err = -EINVAL;
3042 	void *p = msg->front.iov_base;
3043 	void *end = p + msg->front.iov_len;
3044 
3045 	ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3046 	next_mds = ceph_decode_32(&p);
3047 	fwd_seq = ceph_decode_32(&p);
3048 
3049 	mutex_lock(&mdsc->mutex);
3050 	req = lookup_get_request(mdsc, tid);
3051 	if (!req) {
3052 		dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3053 		goto out;  /* dup reply? */
3054 	}
3055 
3056 	if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3057 		dout("forward tid %llu aborted, unregistering\n", tid);
3058 		__unregister_request(mdsc, req);
3059 	} else if (fwd_seq <= req->r_num_fwd) {
3060 		dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3061 		     tid, next_mds, req->r_num_fwd, fwd_seq);
3062 	} else {
3063 		/* resend. forward race not possible; mds would drop */
3064 		dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3065 		BUG_ON(req->r_err);
3066 		BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3067 		req->r_attempts = 0;
3068 		req->r_num_fwd = fwd_seq;
3069 		req->r_resend_mds = next_mds;
3070 		put_request_session(req);
3071 		__do_request(mdsc, req);
3072 	}
3073 	ceph_mdsc_put_request(req);
3074 out:
3075 	mutex_unlock(&mdsc->mutex);
3076 	return;
3077 
3078 bad:
3079 	pr_err("mdsc_handle_forward decode error err=%d\n", err);
3080 }
3081 
3082 static int __decode_session_metadata(void **p, void *end,
3083 				     bool *blacklisted)
3084 {
3085 	/* map<string,string> */
3086 	u32 n;
3087 	bool err_str;
3088 	ceph_decode_32_safe(p, end, n, bad);
3089 	while (n-- > 0) {
3090 		u32 len;
3091 		ceph_decode_32_safe(p, end, len, bad);
3092 		ceph_decode_need(p, end, len, bad);
3093 		err_str = !strncmp(*p, "error_string", len);
3094 		*p += len;
3095 		ceph_decode_32_safe(p, end, len, bad);
3096 		ceph_decode_need(p, end, len, bad);
3097 		if (err_str && strnstr(*p, "blacklisted", len))
3098 			*blacklisted = true;
3099 		*p += len;
3100 	}
3101 	return 0;
3102 bad:
3103 	return -1;
3104 }
3105 
3106 /*
3107  * handle a mds session control message
3108  */
3109 static void handle_session(struct ceph_mds_session *session,
3110 			   struct ceph_msg *msg)
3111 {
3112 	struct ceph_mds_client *mdsc = session->s_mdsc;
3113 	int mds = session->s_mds;
3114 	int msg_version = le16_to_cpu(msg->hdr.version);
3115 	void *p = msg->front.iov_base;
3116 	void *end = p + msg->front.iov_len;
3117 	struct ceph_mds_session_head *h;
3118 	u32 op;
3119 	u64 seq;
3120 	unsigned long features = 0;
3121 	int wake = 0;
3122 	bool blacklisted = false;
3123 
3124 	/* decode */
3125 	ceph_decode_need(&p, end, sizeof(*h), bad);
3126 	h = p;
3127 	p += sizeof(*h);
3128 
3129 	op = le32_to_cpu(h->op);
3130 	seq = le64_to_cpu(h->seq);
3131 
3132 	if (msg_version >= 3) {
3133 		u32 len;
3134 		/* version >= 2, metadata */
3135 		if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3136 			goto bad;
3137 		/* version >= 3, feature bits */
3138 		ceph_decode_32_safe(&p, end, len, bad);
3139 		ceph_decode_need(&p, end, len, bad);
3140 		memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3141 		p += len;
3142 	}
3143 
3144 	mutex_lock(&mdsc->mutex);
3145 	if (op == CEPH_SESSION_CLOSE) {
3146 		ceph_get_mds_session(session);
3147 		__unregister_session(mdsc, session);
3148 	}
3149 	/* FIXME: this ttl calculation is generous */
3150 	session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3151 	mutex_unlock(&mdsc->mutex);
3152 
3153 	mutex_lock(&session->s_mutex);
3154 
3155 	dout("handle_session mds%d %s %p state %s seq %llu\n",
3156 	     mds, ceph_session_op_name(op), session,
3157 	     ceph_session_state_name(session->s_state), seq);
3158 
3159 	if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3160 		session->s_state = CEPH_MDS_SESSION_OPEN;
3161 		pr_info("mds%d came back\n", session->s_mds);
3162 	}
3163 
3164 	switch (op) {
3165 	case CEPH_SESSION_OPEN:
3166 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3167 			pr_info("mds%d reconnect success\n", session->s_mds);
3168 		session->s_state = CEPH_MDS_SESSION_OPEN;
3169 		session->s_features = features;
3170 		renewed_caps(mdsc, session, 0);
3171 		wake = 1;
3172 		if (mdsc->stopping)
3173 			__close_session(mdsc, session);
3174 		break;
3175 
3176 	case CEPH_SESSION_RENEWCAPS:
3177 		if (session->s_renew_seq == seq)
3178 			renewed_caps(mdsc, session, 1);
3179 		break;
3180 
3181 	case CEPH_SESSION_CLOSE:
3182 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3183 			pr_info("mds%d reconnect denied\n", session->s_mds);
3184 		session->s_state = CEPH_MDS_SESSION_CLOSED;
3185 		cleanup_session_requests(mdsc, session);
3186 		remove_session_caps(session);
3187 		wake = 2; /* for good measure */
3188 		wake_up_all(&mdsc->session_close_wq);
3189 		break;
3190 
3191 	case CEPH_SESSION_STALE:
3192 		pr_info("mds%d caps went stale, renewing\n",
3193 			session->s_mds);
3194 		spin_lock(&session->s_gen_ttl_lock);
3195 		session->s_cap_gen++;
3196 		session->s_cap_ttl = jiffies - 1;
3197 		spin_unlock(&session->s_gen_ttl_lock);
3198 		send_renew_caps(mdsc, session);
3199 		break;
3200 
3201 	case CEPH_SESSION_RECALL_STATE:
3202 		ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3203 		break;
3204 
3205 	case CEPH_SESSION_FLUSHMSG:
3206 		send_flushmsg_ack(mdsc, session, seq);
3207 		break;
3208 
3209 	case CEPH_SESSION_FORCE_RO:
3210 		dout("force_session_readonly %p\n", session);
3211 		spin_lock(&session->s_cap_lock);
3212 		session->s_readonly = true;
3213 		spin_unlock(&session->s_cap_lock);
3214 		wake_up_session_caps(session, FORCE_RO);
3215 		break;
3216 
3217 	case CEPH_SESSION_REJECT:
3218 		WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3219 		pr_info("mds%d rejected session\n", session->s_mds);
3220 		session->s_state = CEPH_MDS_SESSION_REJECTED;
3221 		cleanup_session_requests(mdsc, session);
3222 		remove_session_caps(session);
3223 		if (blacklisted)
3224 			mdsc->fsc->blacklisted = true;
3225 		wake = 2; /* for good measure */
3226 		break;
3227 
3228 	default:
3229 		pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3230 		WARN_ON(1);
3231 	}
3232 
3233 	mutex_unlock(&session->s_mutex);
3234 	if (wake) {
3235 		mutex_lock(&mdsc->mutex);
3236 		__wake_requests(mdsc, &session->s_waiting);
3237 		if (wake == 2)
3238 			kick_requests(mdsc, mds);
3239 		mutex_unlock(&mdsc->mutex);
3240 	}
3241 	if (op == CEPH_SESSION_CLOSE)
3242 		ceph_put_mds_session(session);
3243 	return;
3244 
3245 bad:
3246 	pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3247 	       (int)msg->front.iov_len);
3248 	ceph_msg_dump(msg);
3249 	return;
3250 }
3251 
3252 /*
3253  * called under session->mutex.
3254  */
3255 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3256 				   struct ceph_mds_session *session)
3257 {
3258 	struct ceph_mds_request *req, *nreq;
3259 	struct rb_node *p;
3260 
3261 	dout("replay_unsafe_requests mds%d\n", session->s_mds);
3262 
3263 	mutex_lock(&mdsc->mutex);
3264 	list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3265 		__send_request(mdsc, session, req, true);
3266 
3267 	/*
3268 	 * also re-send old requests when MDS enters reconnect stage. So that MDS
3269 	 * can process completed request in clientreplay stage.
3270 	 */
3271 	p = rb_first(&mdsc->request_tree);
3272 	while (p) {
3273 		req = rb_entry(p, struct ceph_mds_request, r_node);
3274 		p = rb_next(p);
3275 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3276 			continue;
3277 		if (req->r_attempts == 0)
3278 			continue; /* only old requests */
3279 		if (req->r_session &&
3280 		    req->r_session->s_mds == session->s_mds)
3281 			__send_request(mdsc, session, req, true);
3282 	}
3283 	mutex_unlock(&mdsc->mutex);
3284 }
3285 
3286 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3287 {
3288 	struct ceph_msg *reply;
3289 	struct ceph_pagelist *_pagelist;
3290 	struct page *page;
3291 	__le32 *addr;
3292 	int err = -ENOMEM;
3293 
3294 	if (!recon_state->allow_multi)
3295 		return -ENOSPC;
3296 
3297 	/* can't handle message that contains both caps and realm */
3298 	BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3299 
3300 	/* pre-allocate new pagelist */
3301 	_pagelist = ceph_pagelist_alloc(GFP_NOFS);
3302 	if (!_pagelist)
3303 		return -ENOMEM;
3304 
3305 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3306 	if (!reply)
3307 		goto fail_msg;
3308 
3309 	/* placeholder for nr_caps */
3310 	err = ceph_pagelist_encode_32(_pagelist, 0);
3311 	if (err < 0)
3312 		goto fail;
3313 
3314 	if (recon_state->nr_caps) {
3315 		/* currently encoding caps */
3316 		err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3317 		if (err)
3318 			goto fail;
3319 	} else {
3320 		/* placeholder for nr_realms (currently encoding relams) */
3321 		err = ceph_pagelist_encode_32(_pagelist, 0);
3322 		if (err < 0)
3323 			goto fail;
3324 	}
3325 
3326 	err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3327 	if (err)
3328 		goto fail;
3329 
3330 	page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3331 	addr = kmap_atomic(page);
3332 	if (recon_state->nr_caps) {
3333 		/* currently encoding caps */
3334 		*addr = cpu_to_le32(recon_state->nr_caps);
3335 	} else {
3336 		/* currently encoding relams */
3337 		*(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3338 	}
3339 	kunmap_atomic(addr);
3340 
3341 	reply->hdr.version = cpu_to_le16(5);
3342 	reply->hdr.compat_version = cpu_to_le16(4);
3343 
3344 	reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3345 	ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3346 
3347 	ceph_con_send(&recon_state->session->s_con, reply);
3348 	ceph_pagelist_release(recon_state->pagelist);
3349 
3350 	recon_state->pagelist = _pagelist;
3351 	recon_state->nr_caps = 0;
3352 	recon_state->nr_realms = 0;
3353 	recon_state->msg_version = 5;
3354 	return 0;
3355 fail:
3356 	ceph_msg_put(reply);
3357 fail_msg:
3358 	ceph_pagelist_release(_pagelist);
3359 	return err;
3360 }
3361 
3362 /*
3363  * Encode information about a cap for a reconnect with the MDS.
3364  */
3365 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3366 			  void *arg)
3367 {
3368 	union {
3369 		struct ceph_mds_cap_reconnect v2;
3370 		struct ceph_mds_cap_reconnect_v1 v1;
3371 	} rec;
3372 	struct ceph_inode_info *ci = cap->ci;
3373 	struct ceph_reconnect_state *recon_state = arg;
3374 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3375 	int err;
3376 	u64 snap_follows;
3377 
3378 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3379 	     inode, ceph_vinop(inode), cap, cap->cap_id,
3380 	     ceph_cap_string(cap->issued));
3381 
3382 	spin_lock(&ci->i_ceph_lock);
3383 	cap->seq = 0;        /* reset cap seq */
3384 	cap->issue_seq = 0;  /* and issue_seq */
3385 	cap->mseq = 0;       /* and migrate_seq */
3386 	cap->cap_gen = cap->session->s_cap_gen;
3387 
3388 	if (recon_state->msg_version >= 2) {
3389 		rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3390 		rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3391 		rec.v2.issued = cpu_to_le32(cap->issued);
3392 		rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3393 		rec.v2.pathbase = 0;
3394 		rec.v2.flock_len = (__force __le32)
3395 			((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3396 	} else {
3397 		rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3398 		rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3399 		rec.v1.issued = cpu_to_le32(cap->issued);
3400 		rec.v1.size = cpu_to_le64(inode->i_size);
3401 		ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3402 		ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3403 		rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3404 		rec.v1.pathbase = 0;
3405 	}
3406 
3407 	if (list_empty(&ci->i_cap_snaps)) {
3408 		snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3409 	} else {
3410 		struct ceph_cap_snap *capsnap =
3411 			list_first_entry(&ci->i_cap_snaps,
3412 					 struct ceph_cap_snap, ci_item);
3413 		snap_follows = capsnap->follows;
3414 	}
3415 	spin_unlock(&ci->i_ceph_lock);
3416 
3417 	if (recon_state->msg_version >= 2) {
3418 		int num_fcntl_locks, num_flock_locks;
3419 		struct ceph_filelock *flocks = NULL;
3420 		size_t struct_len, total_len = sizeof(u64);
3421 		u8 struct_v = 0;
3422 
3423 encode_again:
3424 		if (rec.v2.flock_len) {
3425 			ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3426 		} else {
3427 			num_fcntl_locks = 0;
3428 			num_flock_locks = 0;
3429 		}
3430 		if (num_fcntl_locks + num_flock_locks > 0) {
3431 			flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3432 					       sizeof(struct ceph_filelock),
3433 					       GFP_NOFS);
3434 			if (!flocks) {
3435 				err = -ENOMEM;
3436 				goto out_err;
3437 			}
3438 			err = ceph_encode_locks_to_buffer(inode, flocks,
3439 							  num_fcntl_locks,
3440 							  num_flock_locks);
3441 			if (err) {
3442 				kfree(flocks);
3443 				flocks = NULL;
3444 				if (err == -ENOSPC)
3445 					goto encode_again;
3446 				goto out_err;
3447 			}
3448 		} else {
3449 			kfree(flocks);
3450 			flocks = NULL;
3451 		}
3452 
3453 		if (recon_state->msg_version >= 3) {
3454 			/* version, compat_version and struct_len */
3455 			total_len += 2 * sizeof(u8) + sizeof(u32);
3456 			struct_v = 2;
3457 		}
3458 		/*
3459 		 * number of encoded locks is stable, so copy to pagelist
3460 		 */
3461 		struct_len = 2 * sizeof(u32) +
3462 			    (num_fcntl_locks + num_flock_locks) *
3463 			    sizeof(struct ceph_filelock);
3464 		rec.v2.flock_len = cpu_to_le32(struct_len);
3465 
3466 		struct_len += sizeof(u32) + sizeof(rec.v2);
3467 
3468 		if (struct_v >= 2)
3469 			struct_len += sizeof(u64); /* snap_follows */
3470 
3471 		total_len += struct_len;
3472 
3473 		if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3474 			err = send_reconnect_partial(recon_state);
3475 			if (err)
3476 				goto out_freeflocks;
3477 			pagelist = recon_state->pagelist;
3478 		}
3479 
3480 		err = ceph_pagelist_reserve(pagelist, total_len);
3481 		if (err)
3482 			goto out_freeflocks;
3483 
3484 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3485 		if (recon_state->msg_version >= 3) {
3486 			ceph_pagelist_encode_8(pagelist, struct_v);
3487 			ceph_pagelist_encode_8(pagelist, 1);
3488 			ceph_pagelist_encode_32(pagelist, struct_len);
3489 		}
3490 		ceph_pagelist_encode_string(pagelist, NULL, 0);
3491 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3492 		ceph_locks_to_pagelist(flocks, pagelist,
3493 				       num_fcntl_locks, num_flock_locks);
3494 		if (struct_v >= 2)
3495 			ceph_pagelist_encode_64(pagelist, snap_follows);
3496 out_freeflocks:
3497 		kfree(flocks);
3498 	} else {
3499 		u64 pathbase = 0;
3500 		int pathlen = 0;
3501 		char *path = NULL;
3502 		struct dentry *dentry;
3503 
3504 		dentry = d_find_alias(inode);
3505 		if (dentry) {
3506 			path = ceph_mdsc_build_path(dentry,
3507 						&pathlen, &pathbase, 0);
3508 			dput(dentry);
3509 			if (IS_ERR(path)) {
3510 				err = PTR_ERR(path);
3511 				goto out_err;
3512 			}
3513 			rec.v1.pathbase = cpu_to_le64(pathbase);
3514 		}
3515 
3516 		err = ceph_pagelist_reserve(pagelist,
3517 					    sizeof(u64) + sizeof(u32) +
3518 					    pathlen + sizeof(rec.v1));
3519 		if (err) {
3520 			goto out_freepath;
3521 		}
3522 
3523 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3524 		ceph_pagelist_encode_string(pagelist, path, pathlen);
3525 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3526 out_freepath:
3527 		ceph_mdsc_free_path(path, pathlen);
3528 	}
3529 
3530 out_err:
3531 	if (err >= 0)
3532 		recon_state->nr_caps++;
3533 	return err;
3534 }
3535 
3536 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3537 			      struct ceph_reconnect_state *recon_state)
3538 {
3539 	struct rb_node *p;
3540 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3541 	int err = 0;
3542 
3543 	if (recon_state->msg_version >= 4) {
3544 		err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3545 		if (err < 0)
3546 			goto fail;
3547 	}
3548 
3549 	/*
3550 	 * snaprealms.  we provide mds with the ino, seq (version), and
3551 	 * parent for all of our realms.  If the mds has any newer info,
3552 	 * it will tell us.
3553 	 */
3554 	for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3555 		struct ceph_snap_realm *realm =
3556 		       rb_entry(p, struct ceph_snap_realm, node);
3557 		struct ceph_mds_snaprealm_reconnect sr_rec;
3558 
3559 		if (recon_state->msg_version >= 4) {
3560 			size_t need = sizeof(u8) * 2 + sizeof(u32) +
3561 				      sizeof(sr_rec);
3562 
3563 			if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3564 				err = send_reconnect_partial(recon_state);
3565 				if (err)
3566 					goto fail;
3567 				pagelist = recon_state->pagelist;
3568 			}
3569 
3570 			err = ceph_pagelist_reserve(pagelist, need);
3571 			if (err)
3572 				goto fail;
3573 
3574 			ceph_pagelist_encode_8(pagelist, 1);
3575 			ceph_pagelist_encode_8(pagelist, 1);
3576 			ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3577 		}
3578 
3579 		dout(" adding snap realm %llx seq %lld parent %llx\n",
3580 		     realm->ino, realm->seq, realm->parent_ino);
3581 		sr_rec.ino = cpu_to_le64(realm->ino);
3582 		sr_rec.seq = cpu_to_le64(realm->seq);
3583 		sr_rec.parent = cpu_to_le64(realm->parent_ino);
3584 
3585 		err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3586 		if (err)
3587 			goto fail;
3588 
3589 		recon_state->nr_realms++;
3590 	}
3591 fail:
3592 	return err;
3593 }
3594 
3595 
3596 /*
3597  * If an MDS fails and recovers, clients need to reconnect in order to
3598  * reestablish shared state.  This includes all caps issued through
3599  * this session _and_ the snap_realm hierarchy.  Because it's not
3600  * clear which snap realms the mds cares about, we send everything we
3601  * know about.. that ensures we'll then get any new info the
3602  * recovering MDS might have.
3603  *
3604  * This is a relatively heavyweight operation, but it's rare.
3605  *
3606  * called with mdsc->mutex held.
3607  */
3608 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3609 			       struct ceph_mds_session *session)
3610 {
3611 	struct ceph_msg *reply;
3612 	int mds = session->s_mds;
3613 	int err = -ENOMEM;
3614 	struct ceph_reconnect_state recon_state = {
3615 		.session = session,
3616 	};
3617 	LIST_HEAD(dispose);
3618 
3619 	pr_info("mds%d reconnect start\n", mds);
3620 
3621 	recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3622 	if (!recon_state.pagelist)
3623 		goto fail_nopagelist;
3624 
3625 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3626 	if (!reply)
3627 		goto fail_nomsg;
3628 
3629 	mutex_lock(&session->s_mutex);
3630 	session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3631 	session->s_seq = 0;
3632 
3633 	dout("session %p state %s\n", session,
3634 	     ceph_session_state_name(session->s_state));
3635 
3636 	spin_lock(&session->s_gen_ttl_lock);
3637 	session->s_cap_gen++;
3638 	spin_unlock(&session->s_gen_ttl_lock);
3639 
3640 	spin_lock(&session->s_cap_lock);
3641 	/* don't know if session is readonly */
3642 	session->s_readonly = 0;
3643 	/*
3644 	 * notify __ceph_remove_cap() that we are composing cap reconnect.
3645 	 * If a cap get released before being added to the cap reconnect,
3646 	 * __ceph_remove_cap() should skip queuing cap release.
3647 	 */
3648 	session->s_cap_reconnect = 1;
3649 	/* drop old cap expires; we're about to reestablish that state */
3650 	detach_cap_releases(session, &dispose);
3651 	spin_unlock(&session->s_cap_lock);
3652 	dispose_cap_releases(mdsc, &dispose);
3653 
3654 	/* trim unused caps to reduce MDS's cache rejoin time */
3655 	if (mdsc->fsc->sb->s_root)
3656 		shrink_dcache_parent(mdsc->fsc->sb->s_root);
3657 
3658 	ceph_con_close(&session->s_con);
3659 	ceph_con_open(&session->s_con,
3660 		      CEPH_ENTITY_TYPE_MDS, mds,
3661 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3662 
3663 	/* replay unsafe requests */
3664 	replay_unsafe_requests(mdsc, session);
3665 
3666 	ceph_early_kick_flushing_caps(mdsc, session);
3667 
3668 	down_read(&mdsc->snap_rwsem);
3669 
3670 	/* placeholder for nr_caps */
3671 	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3672 	if (err)
3673 		goto fail;
3674 
3675 	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3676 		recon_state.msg_version = 3;
3677 		recon_state.allow_multi = true;
3678 	} else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3679 		recon_state.msg_version = 3;
3680 	} else {
3681 		recon_state.msg_version = 2;
3682 	}
3683 	/* trsaverse this session's caps */
3684 	err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3685 
3686 	spin_lock(&session->s_cap_lock);
3687 	session->s_cap_reconnect = 0;
3688 	spin_unlock(&session->s_cap_lock);
3689 
3690 	if (err < 0)
3691 		goto fail;
3692 
3693 	/* check if all realms can be encoded into current message */
3694 	if (mdsc->num_snap_realms) {
3695 		size_t total_len =
3696 			recon_state.pagelist->length +
3697 			mdsc->num_snap_realms *
3698 			sizeof(struct ceph_mds_snaprealm_reconnect);
3699 		if (recon_state.msg_version >= 4) {
3700 			/* number of realms */
3701 			total_len += sizeof(u32);
3702 			/* version, compat_version and struct_len */
3703 			total_len += mdsc->num_snap_realms *
3704 				     (2 * sizeof(u8) + sizeof(u32));
3705 		}
3706 		if (total_len > RECONNECT_MAX_SIZE) {
3707 			if (!recon_state.allow_multi) {
3708 				err = -ENOSPC;
3709 				goto fail;
3710 			}
3711 			if (recon_state.nr_caps) {
3712 				err = send_reconnect_partial(&recon_state);
3713 				if (err)
3714 					goto fail;
3715 			}
3716 			recon_state.msg_version = 5;
3717 		}
3718 	}
3719 
3720 	err = encode_snap_realms(mdsc, &recon_state);
3721 	if (err < 0)
3722 		goto fail;
3723 
3724 	if (recon_state.msg_version >= 5) {
3725 		err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3726 		if (err < 0)
3727 			goto fail;
3728 	}
3729 
3730 	if (recon_state.nr_caps || recon_state.nr_realms) {
3731 		struct page *page =
3732 			list_first_entry(&recon_state.pagelist->head,
3733 					struct page, lru);
3734 		__le32 *addr = kmap_atomic(page);
3735 		if (recon_state.nr_caps) {
3736 			WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3737 			*addr = cpu_to_le32(recon_state.nr_caps);
3738 		} else if (recon_state.msg_version >= 4) {
3739 			*(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3740 		}
3741 		kunmap_atomic(addr);
3742 	}
3743 
3744 	reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3745 	if (recon_state.msg_version >= 4)
3746 		reply->hdr.compat_version = cpu_to_le16(4);
3747 
3748 	reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3749 	ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3750 
3751 	ceph_con_send(&session->s_con, reply);
3752 
3753 	mutex_unlock(&session->s_mutex);
3754 
3755 	mutex_lock(&mdsc->mutex);
3756 	__wake_requests(mdsc, &session->s_waiting);
3757 	mutex_unlock(&mdsc->mutex);
3758 
3759 	up_read(&mdsc->snap_rwsem);
3760 	ceph_pagelist_release(recon_state.pagelist);
3761 	return;
3762 
3763 fail:
3764 	ceph_msg_put(reply);
3765 	up_read(&mdsc->snap_rwsem);
3766 	mutex_unlock(&session->s_mutex);
3767 fail_nomsg:
3768 	ceph_pagelist_release(recon_state.pagelist);
3769 fail_nopagelist:
3770 	pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3771 	return;
3772 }
3773 
3774 
3775 /*
3776  * compare old and new mdsmaps, kicking requests
3777  * and closing out old connections as necessary
3778  *
3779  * called under mdsc->mutex.
3780  */
3781 static void check_new_map(struct ceph_mds_client *mdsc,
3782 			  struct ceph_mdsmap *newmap,
3783 			  struct ceph_mdsmap *oldmap)
3784 {
3785 	int i;
3786 	int oldstate, newstate;
3787 	struct ceph_mds_session *s;
3788 
3789 	dout("check_new_map new %u old %u\n",
3790 	     newmap->m_epoch, oldmap->m_epoch);
3791 
3792 	for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
3793 		if (!mdsc->sessions[i])
3794 			continue;
3795 		s = mdsc->sessions[i];
3796 		oldstate = ceph_mdsmap_get_state(oldmap, i);
3797 		newstate = ceph_mdsmap_get_state(newmap, i);
3798 
3799 		dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3800 		     i, ceph_mds_state_name(oldstate),
3801 		     ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3802 		     ceph_mds_state_name(newstate),
3803 		     ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3804 		     ceph_session_state_name(s->s_state));
3805 
3806 		if (i >= newmap->possible_max_rank) {
3807 			/* force close session for stopped mds */
3808 			ceph_get_mds_session(s);
3809 			__unregister_session(mdsc, s);
3810 			__wake_requests(mdsc, &s->s_waiting);
3811 			mutex_unlock(&mdsc->mutex);
3812 
3813 			mutex_lock(&s->s_mutex);
3814 			cleanup_session_requests(mdsc, s);
3815 			remove_session_caps(s);
3816 			mutex_unlock(&s->s_mutex);
3817 
3818 			ceph_put_mds_session(s);
3819 
3820 			mutex_lock(&mdsc->mutex);
3821 			kick_requests(mdsc, i);
3822 			continue;
3823 		}
3824 
3825 		if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
3826 			   ceph_mdsmap_get_addr(newmap, i),
3827 			   sizeof(struct ceph_entity_addr))) {
3828 			/* just close it */
3829 			mutex_unlock(&mdsc->mutex);
3830 			mutex_lock(&s->s_mutex);
3831 			mutex_lock(&mdsc->mutex);
3832 			ceph_con_close(&s->s_con);
3833 			mutex_unlock(&s->s_mutex);
3834 			s->s_state = CEPH_MDS_SESSION_RESTARTING;
3835 		} else if (oldstate == newstate) {
3836 			continue;  /* nothing new with this mds */
3837 		}
3838 
3839 		/*
3840 		 * send reconnect?
3841 		 */
3842 		if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3843 		    newstate >= CEPH_MDS_STATE_RECONNECT) {
3844 			mutex_unlock(&mdsc->mutex);
3845 			send_mds_reconnect(mdsc, s);
3846 			mutex_lock(&mdsc->mutex);
3847 		}
3848 
3849 		/*
3850 		 * kick request on any mds that has gone active.
3851 		 */
3852 		if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3853 		    newstate >= CEPH_MDS_STATE_ACTIVE) {
3854 			if (oldstate != CEPH_MDS_STATE_CREATING &&
3855 			    oldstate != CEPH_MDS_STATE_STARTING)
3856 				pr_info("mds%d recovery completed\n", s->s_mds);
3857 			kick_requests(mdsc, i);
3858 			ceph_kick_flushing_caps(mdsc, s);
3859 			wake_up_session_caps(s, RECONNECT);
3860 		}
3861 	}
3862 
3863 	for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
3864 		s = mdsc->sessions[i];
3865 		if (!s)
3866 			continue;
3867 		if (!ceph_mdsmap_is_laggy(newmap, i))
3868 			continue;
3869 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3870 		    s->s_state == CEPH_MDS_SESSION_HUNG ||
3871 		    s->s_state == CEPH_MDS_SESSION_CLOSING) {
3872 			dout(" connecting to export targets of laggy mds%d\n",
3873 			     i);
3874 			__open_export_target_sessions(mdsc, s);
3875 		}
3876 	}
3877 }
3878 
3879 
3880 
3881 /*
3882  * leases
3883  */
3884 
3885 /*
3886  * caller must hold session s_mutex, dentry->d_lock
3887  */
3888 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3889 {
3890 	struct ceph_dentry_info *di = ceph_dentry(dentry);
3891 
3892 	ceph_put_mds_session(di->lease_session);
3893 	di->lease_session = NULL;
3894 }
3895 
3896 static void handle_lease(struct ceph_mds_client *mdsc,
3897 			 struct ceph_mds_session *session,
3898 			 struct ceph_msg *msg)
3899 {
3900 	struct super_block *sb = mdsc->fsc->sb;
3901 	struct inode *inode;
3902 	struct dentry *parent, *dentry;
3903 	struct ceph_dentry_info *di;
3904 	int mds = session->s_mds;
3905 	struct ceph_mds_lease *h = msg->front.iov_base;
3906 	u32 seq;
3907 	struct ceph_vino vino;
3908 	struct qstr dname;
3909 	int release = 0;
3910 
3911 	dout("handle_lease from mds%d\n", mds);
3912 
3913 	/* decode */
3914 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3915 		goto bad;
3916 	vino.ino = le64_to_cpu(h->ino);
3917 	vino.snap = CEPH_NOSNAP;
3918 	seq = le32_to_cpu(h->seq);
3919 	dname.len = get_unaligned_le32(h + 1);
3920 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3921 		goto bad;
3922 	dname.name = (void *)(h + 1) + sizeof(u32);
3923 
3924 	/* lookup inode */
3925 	inode = ceph_find_inode(sb, vino);
3926 	dout("handle_lease %s, ino %llx %p %.*s\n",
3927 	     ceph_lease_op_name(h->action), vino.ino, inode,
3928 	     dname.len, dname.name);
3929 
3930 	mutex_lock(&session->s_mutex);
3931 	session->s_seq++;
3932 
3933 	if (!inode) {
3934 		dout("handle_lease no inode %llx\n", vino.ino);
3935 		goto release;
3936 	}
3937 
3938 	/* dentry */
3939 	parent = d_find_alias(inode);
3940 	if (!parent) {
3941 		dout("no parent dentry on inode %p\n", inode);
3942 		WARN_ON(1);
3943 		goto release;  /* hrm... */
3944 	}
3945 	dname.hash = full_name_hash(parent, dname.name, dname.len);
3946 	dentry = d_lookup(parent, &dname);
3947 	dput(parent);
3948 	if (!dentry)
3949 		goto release;
3950 
3951 	spin_lock(&dentry->d_lock);
3952 	di = ceph_dentry(dentry);
3953 	switch (h->action) {
3954 	case CEPH_MDS_LEASE_REVOKE:
3955 		if (di->lease_session == session) {
3956 			if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3957 				h->seq = cpu_to_le32(di->lease_seq);
3958 			__ceph_mdsc_drop_dentry_lease(dentry);
3959 		}
3960 		release = 1;
3961 		break;
3962 
3963 	case CEPH_MDS_LEASE_RENEW:
3964 		if (di->lease_session == session &&
3965 		    di->lease_gen == session->s_cap_gen &&
3966 		    di->lease_renew_from &&
3967 		    di->lease_renew_after == 0) {
3968 			unsigned long duration =
3969 				msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3970 
3971 			di->lease_seq = seq;
3972 			di->time = di->lease_renew_from + duration;
3973 			di->lease_renew_after = di->lease_renew_from +
3974 				(duration >> 1);
3975 			di->lease_renew_from = 0;
3976 		}
3977 		break;
3978 	}
3979 	spin_unlock(&dentry->d_lock);
3980 	dput(dentry);
3981 
3982 	if (!release)
3983 		goto out;
3984 
3985 release:
3986 	/* let's just reuse the same message */
3987 	h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3988 	ceph_msg_get(msg);
3989 	ceph_con_send(&session->s_con, msg);
3990 
3991 out:
3992 	mutex_unlock(&session->s_mutex);
3993 	/* avoid calling iput_final() in mds dispatch threads */
3994 	ceph_async_iput(inode);
3995 	return;
3996 
3997 bad:
3998 	pr_err("corrupt lease message\n");
3999 	ceph_msg_dump(msg);
4000 }
4001 
4002 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4003 			      struct dentry *dentry, char action,
4004 			      u32 seq)
4005 {
4006 	struct ceph_msg *msg;
4007 	struct ceph_mds_lease *lease;
4008 	struct inode *dir;
4009 	int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4010 
4011 	dout("lease_send_msg identry %p %s to mds%d\n",
4012 	     dentry, ceph_lease_op_name(action), session->s_mds);
4013 
4014 	msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4015 	if (!msg)
4016 		return;
4017 	lease = msg->front.iov_base;
4018 	lease->action = action;
4019 	lease->seq = cpu_to_le32(seq);
4020 
4021 	spin_lock(&dentry->d_lock);
4022 	dir = d_inode(dentry->d_parent);
4023 	lease->ino = cpu_to_le64(ceph_ino(dir));
4024 	lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4025 
4026 	put_unaligned_le32(dentry->d_name.len, lease + 1);
4027 	memcpy((void *)(lease + 1) + 4,
4028 	       dentry->d_name.name, dentry->d_name.len);
4029 	spin_unlock(&dentry->d_lock);
4030 	/*
4031 	 * if this is a preemptive lease RELEASE, no need to
4032 	 * flush request stream, since the actual request will
4033 	 * soon follow.
4034 	 */
4035 	msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4036 
4037 	ceph_con_send(&session->s_con, msg);
4038 }
4039 
4040 /*
4041  * lock unlock sessions, to wait ongoing session activities
4042  */
4043 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4044 {
4045 	int i;
4046 
4047 	mutex_lock(&mdsc->mutex);
4048 	for (i = 0; i < mdsc->max_sessions; i++) {
4049 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4050 		if (!s)
4051 			continue;
4052 		mutex_unlock(&mdsc->mutex);
4053 		mutex_lock(&s->s_mutex);
4054 		mutex_unlock(&s->s_mutex);
4055 		ceph_put_mds_session(s);
4056 		mutex_lock(&mdsc->mutex);
4057 	}
4058 	mutex_unlock(&mdsc->mutex);
4059 }
4060 
4061 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4062 {
4063 	struct ceph_fs_client *fsc = mdsc->fsc;
4064 
4065 	if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4066 		return;
4067 
4068 	if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4069 		return;
4070 
4071 	if (!READ_ONCE(fsc->blacklisted))
4072 		return;
4073 
4074 	if (fsc->last_auto_reconnect &&
4075 	    time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4076 		return;
4077 
4078 	pr_info("auto reconnect after blacklisted\n");
4079 	fsc->last_auto_reconnect = jiffies;
4080 	ceph_force_reconnect(fsc->sb);
4081 }
4082 
4083 /*
4084  * delayed work -- periodically trim expired leases, renew caps with mds
4085  */
4086 static void schedule_delayed(struct ceph_mds_client *mdsc)
4087 {
4088 	int delay = 5;
4089 	unsigned hz = round_jiffies_relative(HZ * delay);
4090 	schedule_delayed_work(&mdsc->delayed_work, hz);
4091 }
4092 
4093 static void delayed_work(struct work_struct *work)
4094 {
4095 	int i;
4096 	struct ceph_mds_client *mdsc =
4097 		container_of(work, struct ceph_mds_client, delayed_work.work);
4098 	int renew_interval;
4099 	int renew_caps;
4100 
4101 	dout("mdsc delayed_work\n");
4102 
4103 	mutex_lock(&mdsc->mutex);
4104 	renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4105 	renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4106 				   mdsc->last_renew_caps);
4107 	if (renew_caps)
4108 		mdsc->last_renew_caps = jiffies;
4109 
4110 	for (i = 0; i < mdsc->max_sessions; i++) {
4111 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4112 		if (!s)
4113 			continue;
4114 		if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4115 			dout("resending session close request for mds%d\n",
4116 			     s->s_mds);
4117 			request_close_session(mdsc, s);
4118 			ceph_put_mds_session(s);
4119 			continue;
4120 		}
4121 		if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4122 			if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4123 				s->s_state = CEPH_MDS_SESSION_HUNG;
4124 				pr_info("mds%d hung\n", s->s_mds);
4125 			}
4126 		}
4127 		if (s->s_state == CEPH_MDS_SESSION_NEW ||
4128 		    s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4129 		    s->s_state == CEPH_MDS_SESSION_REJECTED) {
4130 			/* this mds is failed or recovering, just wait */
4131 			ceph_put_mds_session(s);
4132 			continue;
4133 		}
4134 		mutex_unlock(&mdsc->mutex);
4135 
4136 		mutex_lock(&s->s_mutex);
4137 		if (renew_caps)
4138 			send_renew_caps(mdsc, s);
4139 		else
4140 			ceph_con_keepalive(&s->s_con);
4141 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4142 		    s->s_state == CEPH_MDS_SESSION_HUNG)
4143 			ceph_send_cap_releases(mdsc, s);
4144 		mutex_unlock(&s->s_mutex);
4145 		ceph_put_mds_session(s);
4146 
4147 		mutex_lock(&mdsc->mutex);
4148 	}
4149 	mutex_unlock(&mdsc->mutex);
4150 
4151 	ceph_check_delayed_caps(mdsc);
4152 
4153 	ceph_queue_cap_reclaim_work(mdsc);
4154 
4155 	ceph_trim_snapid_map(mdsc);
4156 
4157 	maybe_recover_session(mdsc);
4158 
4159 	schedule_delayed(mdsc);
4160 }
4161 
4162 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4163 
4164 {
4165 	struct ceph_mds_client *mdsc;
4166 
4167 	mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4168 	if (!mdsc)
4169 		return -ENOMEM;
4170 	mdsc->fsc = fsc;
4171 	mutex_init(&mdsc->mutex);
4172 	mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4173 	if (!mdsc->mdsmap) {
4174 		kfree(mdsc);
4175 		return -ENOMEM;
4176 	}
4177 
4178 	fsc->mdsc = mdsc;
4179 	init_completion(&mdsc->safe_umount_waiters);
4180 	init_waitqueue_head(&mdsc->session_close_wq);
4181 	INIT_LIST_HEAD(&mdsc->waiting_for_map);
4182 	mdsc->sessions = NULL;
4183 	atomic_set(&mdsc->num_sessions, 0);
4184 	mdsc->max_sessions = 0;
4185 	mdsc->stopping = 0;
4186 	atomic64_set(&mdsc->quotarealms_count, 0);
4187 	mdsc->quotarealms_inodes = RB_ROOT;
4188 	mutex_init(&mdsc->quotarealms_inodes_mutex);
4189 	mdsc->last_snap_seq = 0;
4190 	init_rwsem(&mdsc->snap_rwsem);
4191 	mdsc->snap_realms = RB_ROOT;
4192 	INIT_LIST_HEAD(&mdsc->snap_empty);
4193 	mdsc->num_snap_realms = 0;
4194 	spin_lock_init(&mdsc->snap_empty_lock);
4195 	mdsc->last_tid = 0;
4196 	mdsc->oldest_tid = 0;
4197 	mdsc->request_tree = RB_ROOT;
4198 	INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4199 	mdsc->last_renew_caps = jiffies;
4200 	INIT_LIST_HEAD(&mdsc->cap_delay_list);
4201 	INIT_LIST_HEAD(&mdsc->cap_wait_list);
4202 	spin_lock_init(&mdsc->cap_delay_lock);
4203 	INIT_LIST_HEAD(&mdsc->snap_flush_list);
4204 	spin_lock_init(&mdsc->snap_flush_lock);
4205 	mdsc->last_cap_flush_tid = 1;
4206 	INIT_LIST_HEAD(&mdsc->cap_flush_list);
4207 	INIT_LIST_HEAD(&mdsc->cap_dirty);
4208 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4209 	mdsc->num_cap_flushing = 0;
4210 	spin_lock_init(&mdsc->cap_dirty_lock);
4211 	init_waitqueue_head(&mdsc->cap_flushing_wq);
4212 	INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4213 	atomic_set(&mdsc->cap_reclaim_pending, 0);
4214 
4215 	spin_lock_init(&mdsc->dentry_list_lock);
4216 	INIT_LIST_HEAD(&mdsc->dentry_leases);
4217 	INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4218 
4219 	ceph_caps_init(mdsc);
4220 	ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4221 
4222 	spin_lock_init(&mdsc->snapid_map_lock);
4223 	mdsc->snapid_map_tree = RB_ROOT;
4224 	INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4225 
4226 	init_rwsem(&mdsc->pool_perm_rwsem);
4227 	mdsc->pool_perm_tree = RB_ROOT;
4228 
4229 	strscpy(mdsc->nodename, utsname()->nodename,
4230 		sizeof(mdsc->nodename));
4231 	return 0;
4232 }
4233 
4234 /*
4235  * Wait for safe replies on open mds requests.  If we time out, drop
4236  * all requests from the tree to avoid dangling dentry refs.
4237  */
4238 static void wait_requests(struct ceph_mds_client *mdsc)
4239 {
4240 	struct ceph_options *opts = mdsc->fsc->client->options;
4241 	struct ceph_mds_request *req;
4242 
4243 	mutex_lock(&mdsc->mutex);
4244 	if (__get_oldest_req(mdsc)) {
4245 		mutex_unlock(&mdsc->mutex);
4246 
4247 		dout("wait_requests waiting for requests\n");
4248 		wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4249 				    ceph_timeout_jiffies(opts->mount_timeout));
4250 
4251 		/* tear down remaining requests */
4252 		mutex_lock(&mdsc->mutex);
4253 		while ((req = __get_oldest_req(mdsc))) {
4254 			dout("wait_requests timed out on tid %llu\n",
4255 			     req->r_tid);
4256 			list_del_init(&req->r_wait);
4257 			__unregister_request(mdsc, req);
4258 		}
4259 	}
4260 	mutex_unlock(&mdsc->mutex);
4261 	dout("wait_requests done\n");
4262 }
4263 
4264 /*
4265  * called before mount is ro, and before dentries are torn down.
4266  * (hmm, does this still race with new lookups?)
4267  */
4268 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4269 {
4270 	dout("pre_umount\n");
4271 	mdsc->stopping = 1;
4272 
4273 	lock_unlock_sessions(mdsc);
4274 	ceph_flush_dirty_caps(mdsc);
4275 	wait_requests(mdsc);
4276 
4277 	/*
4278 	 * wait for reply handlers to drop their request refs and
4279 	 * their inode/dcache refs
4280 	 */
4281 	ceph_msgr_flush();
4282 
4283 	ceph_cleanup_quotarealms_inodes(mdsc);
4284 }
4285 
4286 /*
4287  * wait for all write mds requests to flush.
4288  */
4289 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4290 {
4291 	struct ceph_mds_request *req = NULL, *nextreq;
4292 	struct rb_node *n;
4293 
4294 	mutex_lock(&mdsc->mutex);
4295 	dout("wait_unsafe_requests want %lld\n", want_tid);
4296 restart:
4297 	req = __get_oldest_req(mdsc);
4298 	while (req && req->r_tid <= want_tid) {
4299 		/* find next request */
4300 		n = rb_next(&req->r_node);
4301 		if (n)
4302 			nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4303 		else
4304 			nextreq = NULL;
4305 		if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4306 		    (req->r_op & CEPH_MDS_OP_WRITE)) {
4307 			/* write op */
4308 			ceph_mdsc_get_request(req);
4309 			if (nextreq)
4310 				ceph_mdsc_get_request(nextreq);
4311 			mutex_unlock(&mdsc->mutex);
4312 			dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4313 			     req->r_tid, want_tid);
4314 			wait_for_completion(&req->r_safe_completion);
4315 			mutex_lock(&mdsc->mutex);
4316 			ceph_mdsc_put_request(req);
4317 			if (!nextreq)
4318 				break;  /* next dne before, so we're done! */
4319 			if (RB_EMPTY_NODE(&nextreq->r_node)) {
4320 				/* next request was removed from tree */
4321 				ceph_mdsc_put_request(nextreq);
4322 				goto restart;
4323 			}
4324 			ceph_mdsc_put_request(nextreq);  /* won't go away */
4325 		}
4326 		req = nextreq;
4327 	}
4328 	mutex_unlock(&mdsc->mutex);
4329 	dout("wait_unsafe_requests done\n");
4330 }
4331 
4332 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4333 {
4334 	u64 want_tid, want_flush;
4335 
4336 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4337 		return;
4338 
4339 	dout("sync\n");
4340 	mutex_lock(&mdsc->mutex);
4341 	want_tid = mdsc->last_tid;
4342 	mutex_unlock(&mdsc->mutex);
4343 
4344 	ceph_flush_dirty_caps(mdsc);
4345 	spin_lock(&mdsc->cap_dirty_lock);
4346 	want_flush = mdsc->last_cap_flush_tid;
4347 	if (!list_empty(&mdsc->cap_flush_list)) {
4348 		struct ceph_cap_flush *cf =
4349 			list_last_entry(&mdsc->cap_flush_list,
4350 					struct ceph_cap_flush, g_list);
4351 		cf->wake = true;
4352 	}
4353 	spin_unlock(&mdsc->cap_dirty_lock);
4354 
4355 	dout("sync want tid %lld flush_seq %lld\n",
4356 	     want_tid, want_flush);
4357 
4358 	wait_unsafe_requests(mdsc, want_tid);
4359 	wait_caps_flush(mdsc, want_flush);
4360 }
4361 
4362 /*
4363  * true if all sessions are closed, or we force unmount
4364  */
4365 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4366 {
4367 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4368 		return true;
4369 	return atomic_read(&mdsc->num_sessions) <= skipped;
4370 }
4371 
4372 /*
4373  * called after sb is ro.
4374  */
4375 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4376 {
4377 	struct ceph_options *opts = mdsc->fsc->client->options;
4378 	struct ceph_mds_session *session;
4379 	int i;
4380 	int skipped = 0;
4381 
4382 	dout("close_sessions\n");
4383 
4384 	/* close sessions */
4385 	mutex_lock(&mdsc->mutex);
4386 	for (i = 0; i < mdsc->max_sessions; i++) {
4387 		session = __ceph_lookup_mds_session(mdsc, i);
4388 		if (!session)
4389 			continue;
4390 		mutex_unlock(&mdsc->mutex);
4391 		mutex_lock(&session->s_mutex);
4392 		if (__close_session(mdsc, session) <= 0)
4393 			skipped++;
4394 		mutex_unlock(&session->s_mutex);
4395 		ceph_put_mds_session(session);
4396 		mutex_lock(&mdsc->mutex);
4397 	}
4398 	mutex_unlock(&mdsc->mutex);
4399 
4400 	dout("waiting for sessions to close\n");
4401 	wait_event_timeout(mdsc->session_close_wq,
4402 			   done_closing_sessions(mdsc, skipped),
4403 			   ceph_timeout_jiffies(opts->mount_timeout));
4404 
4405 	/* tear down remaining sessions */
4406 	mutex_lock(&mdsc->mutex);
4407 	for (i = 0; i < mdsc->max_sessions; i++) {
4408 		if (mdsc->sessions[i]) {
4409 			session = ceph_get_mds_session(mdsc->sessions[i]);
4410 			__unregister_session(mdsc, session);
4411 			mutex_unlock(&mdsc->mutex);
4412 			mutex_lock(&session->s_mutex);
4413 			remove_session_caps(session);
4414 			mutex_unlock(&session->s_mutex);
4415 			ceph_put_mds_session(session);
4416 			mutex_lock(&mdsc->mutex);
4417 		}
4418 	}
4419 	WARN_ON(!list_empty(&mdsc->cap_delay_list));
4420 	mutex_unlock(&mdsc->mutex);
4421 
4422 	ceph_cleanup_snapid_map(mdsc);
4423 	ceph_cleanup_empty_realms(mdsc);
4424 
4425 	cancel_work_sync(&mdsc->cap_reclaim_work);
4426 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4427 
4428 	dout("stopped\n");
4429 }
4430 
4431 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4432 {
4433 	struct ceph_mds_session *session;
4434 	int mds;
4435 
4436 	dout("force umount\n");
4437 
4438 	mutex_lock(&mdsc->mutex);
4439 	for (mds = 0; mds < mdsc->max_sessions; mds++) {
4440 		session = __ceph_lookup_mds_session(mdsc, mds);
4441 		if (!session)
4442 			continue;
4443 
4444 		if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4445 			__unregister_session(mdsc, session);
4446 		__wake_requests(mdsc, &session->s_waiting);
4447 		mutex_unlock(&mdsc->mutex);
4448 
4449 		mutex_lock(&session->s_mutex);
4450 		__close_session(mdsc, session);
4451 		if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4452 			cleanup_session_requests(mdsc, session);
4453 			remove_session_caps(session);
4454 		}
4455 		mutex_unlock(&session->s_mutex);
4456 		ceph_put_mds_session(session);
4457 
4458 		mutex_lock(&mdsc->mutex);
4459 		kick_requests(mdsc, mds);
4460 	}
4461 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4462 	mutex_unlock(&mdsc->mutex);
4463 }
4464 
4465 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4466 {
4467 	dout("stop\n");
4468 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4469 	if (mdsc->mdsmap)
4470 		ceph_mdsmap_destroy(mdsc->mdsmap);
4471 	kfree(mdsc->sessions);
4472 	ceph_caps_finalize(mdsc);
4473 	ceph_pool_perm_destroy(mdsc);
4474 }
4475 
4476 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4477 {
4478 	struct ceph_mds_client *mdsc = fsc->mdsc;
4479 	dout("mdsc_destroy %p\n", mdsc);
4480 
4481 	if (!mdsc)
4482 		return;
4483 
4484 	/* flush out any connection work with references to us */
4485 	ceph_msgr_flush();
4486 
4487 	ceph_mdsc_stop(mdsc);
4488 
4489 	fsc->mdsc = NULL;
4490 	kfree(mdsc);
4491 	dout("mdsc_destroy %p done\n", mdsc);
4492 }
4493 
4494 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4495 {
4496 	struct ceph_fs_client *fsc = mdsc->fsc;
4497 	const char *mds_namespace = fsc->mount_options->mds_namespace;
4498 	void *p = msg->front.iov_base;
4499 	void *end = p + msg->front.iov_len;
4500 	u32 epoch;
4501 	u32 map_len;
4502 	u32 num_fs;
4503 	u32 mount_fscid = (u32)-1;
4504 	u8 struct_v, struct_cv;
4505 	int err = -EINVAL;
4506 
4507 	ceph_decode_need(&p, end, sizeof(u32), bad);
4508 	epoch = ceph_decode_32(&p);
4509 
4510 	dout("handle_fsmap epoch %u\n", epoch);
4511 
4512 	ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4513 	struct_v = ceph_decode_8(&p);
4514 	struct_cv = ceph_decode_8(&p);
4515 	map_len = ceph_decode_32(&p);
4516 
4517 	ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4518 	p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4519 
4520 	num_fs = ceph_decode_32(&p);
4521 	while (num_fs-- > 0) {
4522 		void *info_p, *info_end;
4523 		u32 info_len;
4524 		u8 info_v, info_cv;
4525 		u32 fscid, namelen;
4526 
4527 		ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4528 		info_v = ceph_decode_8(&p);
4529 		info_cv = ceph_decode_8(&p);
4530 		info_len = ceph_decode_32(&p);
4531 		ceph_decode_need(&p, end, info_len, bad);
4532 		info_p = p;
4533 		info_end = p + info_len;
4534 		p = info_end;
4535 
4536 		ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4537 		fscid = ceph_decode_32(&info_p);
4538 		namelen = ceph_decode_32(&info_p);
4539 		ceph_decode_need(&info_p, info_end, namelen, bad);
4540 
4541 		if (mds_namespace &&
4542 		    strlen(mds_namespace) == namelen &&
4543 		    !strncmp(mds_namespace, (char *)info_p, namelen)) {
4544 			mount_fscid = fscid;
4545 			break;
4546 		}
4547 	}
4548 
4549 	ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4550 	if (mount_fscid != (u32)-1) {
4551 		fsc->client->monc.fs_cluster_id = mount_fscid;
4552 		ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4553 				   0, true);
4554 		ceph_monc_renew_subs(&fsc->client->monc);
4555 	} else {
4556 		err = -ENOENT;
4557 		goto err_out;
4558 	}
4559 	return;
4560 
4561 bad:
4562 	pr_err("error decoding fsmap\n");
4563 err_out:
4564 	mutex_lock(&mdsc->mutex);
4565 	mdsc->mdsmap_err = err;
4566 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4567 	mutex_unlock(&mdsc->mutex);
4568 }
4569 
4570 /*
4571  * handle mds map update.
4572  */
4573 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4574 {
4575 	u32 epoch;
4576 	u32 maplen;
4577 	void *p = msg->front.iov_base;
4578 	void *end = p + msg->front.iov_len;
4579 	struct ceph_mdsmap *newmap, *oldmap;
4580 	struct ceph_fsid fsid;
4581 	int err = -EINVAL;
4582 
4583 	ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4584 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
4585 	if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4586 		return;
4587 	epoch = ceph_decode_32(&p);
4588 	maplen = ceph_decode_32(&p);
4589 	dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4590 
4591 	/* do we need it? */
4592 	mutex_lock(&mdsc->mutex);
4593 	if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4594 		dout("handle_map epoch %u <= our %u\n",
4595 		     epoch, mdsc->mdsmap->m_epoch);
4596 		mutex_unlock(&mdsc->mutex);
4597 		return;
4598 	}
4599 
4600 	newmap = ceph_mdsmap_decode(&p, end);
4601 	if (IS_ERR(newmap)) {
4602 		err = PTR_ERR(newmap);
4603 		goto bad_unlock;
4604 	}
4605 
4606 	/* swap into place */
4607 	if (mdsc->mdsmap) {
4608 		oldmap = mdsc->mdsmap;
4609 		mdsc->mdsmap = newmap;
4610 		check_new_map(mdsc, newmap, oldmap);
4611 		ceph_mdsmap_destroy(oldmap);
4612 	} else {
4613 		mdsc->mdsmap = newmap;  /* first mds map */
4614 	}
4615 	mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4616 					MAX_LFS_FILESIZE);
4617 
4618 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4619 	ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4620 			  mdsc->mdsmap->m_epoch);
4621 
4622 	mutex_unlock(&mdsc->mutex);
4623 	schedule_delayed(mdsc);
4624 	return;
4625 
4626 bad_unlock:
4627 	mutex_unlock(&mdsc->mutex);
4628 bad:
4629 	pr_err("error decoding mdsmap %d\n", err);
4630 	return;
4631 }
4632 
4633 static struct ceph_connection *con_get(struct ceph_connection *con)
4634 {
4635 	struct ceph_mds_session *s = con->private;
4636 
4637 	if (ceph_get_mds_session(s))
4638 		return con;
4639 	return NULL;
4640 }
4641 
4642 static void con_put(struct ceph_connection *con)
4643 {
4644 	struct ceph_mds_session *s = con->private;
4645 
4646 	ceph_put_mds_session(s);
4647 }
4648 
4649 /*
4650  * if the client is unresponsive for long enough, the mds will kill
4651  * the session entirely.
4652  */
4653 static void peer_reset(struct ceph_connection *con)
4654 {
4655 	struct ceph_mds_session *s = con->private;
4656 	struct ceph_mds_client *mdsc = s->s_mdsc;
4657 
4658 	pr_warn("mds%d closed our session\n", s->s_mds);
4659 	send_mds_reconnect(mdsc, s);
4660 }
4661 
4662 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4663 {
4664 	struct ceph_mds_session *s = con->private;
4665 	struct ceph_mds_client *mdsc = s->s_mdsc;
4666 	int type = le16_to_cpu(msg->hdr.type);
4667 
4668 	mutex_lock(&mdsc->mutex);
4669 	if (__verify_registered_session(mdsc, s) < 0) {
4670 		mutex_unlock(&mdsc->mutex);
4671 		goto out;
4672 	}
4673 	mutex_unlock(&mdsc->mutex);
4674 
4675 	switch (type) {
4676 	case CEPH_MSG_MDS_MAP:
4677 		ceph_mdsc_handle_mdsmap(mdsc, msg);
4678 		break;
4679 	case CEPH_MSG_FS_MAP_USER:
4680 		ceph_mdsc_handle_fsmap(mdsc, msg);
4681 		break;
4682 	case CEPH_MSG_CLIENT_SESSION:
4683 		handle_session(s, msg);
4684 		break;
4685 	case CEPH_MSG_CLIENT_REPLY:
4686 		handle_reply(s, msg);
4687 		break;
4688 	case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4689 		handle_forward(mdsc, s, msg);
4690 		break;
4691 	case CEPH_MSG_CLIENT_CAPS:
4692 		ceph_handle_caps(s, msg);
4693 		break;
4694 	case CEPH_MSG_CLIENT_SNAP:
4695 		ceph_handle_snap(mdsc, s, msg);
4696 		break;
4697 	case CEPH_MSG_CLIENT_LEASE:
4698 		handle_lease(mdsc, s, msg);
4699 		break;
4700 	case CEPH_MSG_CLIENT_QUOTA:
4701 		ceph_handle_quota(mdsc, s, msg);
4702 		break;
4703 
4704 	default:
4705 		pr_err("received unknown message type %d %s\n", type,
4706 		       ceph_msg_type_name(type));
4707 	}
4708 out:
4709 	ceph_msg_put(msg);
4710 }
4711 
4712 /*
4713  * authentication
4714  */
4715 
4716 /*
4717  * Note: returned pointer is the address of a structure that's
4718  * managed separately.  Caller must *not* attempt to free it.
4719  */
4720 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4721 					int *proto, int force_new)
4722 {
4723 	struct ceph_mds_session *s = con->private;
4724 	struct ceph_mds_client *mdsc = s->s_mdsc;
4725 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4726 	struct ceph_auth_handshake *auth = &s->s_auth;
4727 
4728 	if (force_new && auth->authorizer) {
4729 		ceph_auth_destroy_authorizer(auth->authorizer);
4730 		auth->authorizer = NULL;
4731 	}
4732 	if (!auth->authorizer) {
4733 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4734 						      auth);
4735 		if (ret)
4736 			return ERR_PTR(ret);
4737 	} else {
4738 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4739 						      auth);
4740 		if (ret)
4741 			return ERR_PTR(ret);
4742 	}
4743 	*proto = ac->protocol;
4744 
4745 	return auth;
4746 }
4747 
4748 static int add_authorizer_challenge(struct ceph_connection *con,
4749 				    void *challenge_buf, int challenge_buf_len)
4750 {
4751 	struct ceph_mds_session *s = con->private;
4752 	struct ceph_mds_client *mdsc = s->s_mdsc;
4753 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4754 
4755 	return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4756 					    challenge_buf, challenge_buf_len);
4757 }
4758 
4759 static int verify_authorizer_reply(struct ceph_connection *con)
4760 {
4761 	struct ceph_mds_session *s = con->private;
4762 	struct ceph_mds_client *mdsc = s->s_mdsc;
4763 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4764 
4765 	return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4766 }
4767 
4768 static int invalidate_authorizer(struct ceph_connection *con)
4769 {
4770 	struct ceph_mds_session *s = con->private;
4771 	struct ceph_mds_client *mdsc = s->s_mdsc;
4772 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4773 
4774 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4775 
4776 	return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4777 }
4778 
4779 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4780 				struct ceph_msg_header *hdr, int *skip)
4781 {
4782 	struct ceph_msg *msg;
4783 	int type = (int) le16_to_cpu(hdr->type);
4784 	int front_len = (int) le32_to_cpu(hdr->front_len);
4785 
4786 	if (con->in_msg)
4787 		return con->in_msg;
4788 
4789 	*skip = 0;
4790 	msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4791 	if (!msg) {
4792 		pr_err("unable to allocate msg type %d len %d\n",
4793 		       type, front_len);
4794 		return NULL;
4795 	}
4796 
4797 	return msg;
4798 }
4799 
4800 static int mds_sign_message(struct ceph_msg *msg)
4801 {
4802        struct ceph_mds_session *s = msg->con->private;
4803        struct ceph_auth_handshake *auth = &s->s_auth;
4804 
4805        return ceph_auth_sign_message(auth, msg);
4806 }
4807 
4808 static int mds_check_message_signature(struct ceph_msg *msg)
4809 {
4810        struct ceph_mds_session *s = msg->con->private;
4811        struct ceph_auth_handshake *auth = &s->s_auth;
4812 
4813        return ceph_auth_check_message_signature(auth, msg);
4814 }
4815 
4816 static const struct ceph_connection_operations mds_con_ops = {
4817 	.get = con_get,
4818 	.put = con_put,
4819 	.dispatch = dispatch,
4820 	.get_authorizer = get_authorizer,
4821 	.add_authorizer_challenge = add_authorizer_challenge,
4822 	.verify_authorizer_reply = verify_authorizer_reply,
4823 	.invalidate_authorizer = invalidate_authorizer,
4824 	.peer_reset = peer_reset,
4825 	.alloc_msg = mds_alloc_msg,
4826 	.sign_message = mds_sign_message,
4827 	.check_message_signature = mds_check_message_signature,
4828 };
4829 
4830 /* eof */
4831