xref: /linux/fs/ceph/mds_client.c (revision 4f3c8320c78cdd11c8fdd23c33787407f719322e)
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 #include <linux/ktime.h>
14 
15 #include "super.h"
16 #include "mds_client.h"
17 
18 #include <linux/ceph/ceph_features.h>
19 #include <linux/ceph/messenger.h>
20 #include <linux/ceph/decode.h>
21 #include <linux/ceph/pagelist.h>
22 #include <linux/ceph/auth.h>
23 #include <linux/ceph/debugfs.h>
24 
25 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
26 
27 /*
28  * A cluster of MDS (metadata server) daemons is responsible for
29  * managing the file system namespace (the directory hierarchy and
30  * inodes) and for coordinating shared access to storage.  Metadata is
31  * partitioning hierarchically across a number of servers, and that
32  * partition varies over time as the cluster adjusts the distribution
33  * in order to balance load.
34  *
35  * The MDS client is primarily responsible to managing synchronous
36  * metadata requests for operations like open, unlink, and so forth.
37  * If there is a MDS failure, we find out about it when we (possibly
38  * request and) receive a new MDS map, and can resubmit affected
39  * requests.
40  *
41  * For the most part, though, we take advantage of a lossless
42  * communications channel to the MDS, and do not need to worry about
43  * timing out or resubmitting requests.
44  *
45  * We maintain a stateful "session" with each MDS we interact with.
46  * Within each session, we sent periodic heartbeat messages to ensure
47  * any capabilities or leases we have been issues remain valid.  If
48  * the session times out and goes stale, our leases and capabilities
49  * are no longer valid.
50  */
51 
52 struct ceph_reconnect_state {
53 	struct ceph_mds_session *session;
54 	int nr_caps, nr_realms;
55 	struct ceph_pagelist *pagelist;
56 	unsigned msg_version;
57 	bool allow_multi;
58 };
59 
60 static void __wake_requests(struct ceph_mds_client *mdsc,
61 			    struct list_head *head);
62 static void ceph_cap_release_work(struct work_struct *work);
63 static void ceph_cap_reclaim_work(struct work_struct *work);
64 
65 static const struct ceph_connection_operations mds_con_ops;
66 
67 
68 /*
69  * mds reply parsing
70  */
71 
72 static int parse_reply_info_quota(void **p, void *end,
73 				  struct ceph_mds_reply_info_in *info)
74 {
75 	u8 struct_v, struct_compat;
76 	u32 struct_len;
77 
78 	ceph_decode_8_safe(p, end, struct_v, bad);
79 	ceph_decode_8_safe(p, end, struct_compat, bad);
80 	/* struct_v is expected to be >= 1. we only
81 	 * understand encoding with struct_compat == 1. */
82 	if (!struct_v || struct_compat != 1)
83 		goto bad;
84 	ceph_decode_32_safe(p, end, struct_len, bad);
85 	ceph_decode_need(p, end, struct_len, bad);
86 	end = *p + struct_len;
87 	ceph_decode_64_safe(p, end, info->max_bytes, bad);
88 	ceph_decode_64_safe(p, end, info->max_files, bad);
89 	*p = end;
90 	return 0;
91 bad:
92 	return -EIO;
93 }
94 
95 /*
96  * parse individual inode info
97  */
98 static int parse_reply_info_in(void **p, void *end,
99 			       struct ceph_mds_reply_info_in *info,
100 			       u64 features)
101 {
102 	int err = 0;
103 	u8 struct_v = 0;
104 
105 	if (features == (u64)-1) {
106 		u32 struct_len;
107 		u8 struct_compat;
108 		ceph_decode_8_safe(p, end, struct_v, bad);
109 		ceph_decode_8_safe(p, end, struct_compat, bad);
110 		/* struct_v is expected to be >= 1. we only understand
111 		 * encoding with struct_compat == 1. */
112 		if (!struct_v || struct_compat != 1)
113 			goto bad;
114 		ceph_decode_32_safe(p, end, struct_len, bad);
115 		ceph_decode_need(p, end, struct_len, bad);
116 		end = *p + struct_len;
117 	}
118 
119 	ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
120 	info->in = *p;
121 	*p += sizeof(struct ceph_mds_reply_inode) +
122 		sizeof(*info->in->fragtree.splits) *
123 		le32_to_cpu(info->in->fragtree.nsplits);
124 
125 	ceph_decode_32_safe(p, end, info->symlink_len, bad);
126 	ceph_decode_need(p, end, info->symlink_len, bad);
127 	info->symlink = *p;
128 	*p += info->symlink_len;
129 
130 	ceph_decode_copy_safe(p, end, &info->dir_layout,
131 			      sizeof(info->dir_layout), bad);
132 	ceph_decode_32_safe(p, end, info->xattr_len, bad);
133 	ceph_decode_need(p, end, info->xattr_len, bad);
134 	info->xattr_data = *p;
135 	*p += info->xattr_len;
136 
137 	if (features == (u64)-1) {
138 		/* inline data */
139 		ceph_decode_64_safe(p, end, info->inline_version, bad);
140 		ceph_decode_32_safe(p, end, info->inline_len, bad);
141 		ceph_decode_need(p, end, info->inline_len, bad);
142 		info->inline_data = *p;
143 		*p += info->inline_len;
144 		/* quota */
145 		err = parse_reply_info_quota(p, end, info);
146 		if (err < 0)
147 			goto out_bad;
148 		/* pool namespace */
149 		ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
150 		if (info->pool_ns_len > 0) {
151 			ceph_decode_need(p, end, info->pool_ns_len, bad);
152 			info->pool_ns_data = *p;
153 			*p += info->pool_ns_len;
154 		}
155 
156 		/* btime */
157 		ceph_decode_need(p, end, sizeof(info->btime), bad);
158 		ceph_decode_copy(p, &info->btime, sizeof(info->btime));
159 
160 		/* change attribute */
161 		ceph_decode_64_safe(p, end, info->change_attr, bad);
162 
163 		/* dir pin */
164 		if (struct_v >= 2) {
165 			ceph_decode_32_safe(p, end, info->dir_pin, bad);
166 		} else {
167 			info->dir_pin = -ENODATA;
168 		}
169 
170 		/* snapshot birth time, remains zero for v<=2 */
171 		if (struct_v >= 3) {
172 			ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
173 			ceph_decode_copy(p, &info->snap_btime,
174 					 sizeof(info->snap_btime));
175 		} else {
176 			memset(&info->snap_btime, 0, sizeof(info->snap_btime));
177 		}
178 
179 		*p = end;
180 	} else {
181 		if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
182 			ceph_decode_64_safe(p, end, info->inline_version, bad);
183 			ceph_decode_32_safe(p, end, info->inline_len, bad);
184 			ceph_decode_need(p, end, info->inline_len, bad);
185 			info->inline_data = *p;
186 			*p += info->inline_len;
187 		} else
188 			info->inline_version = CEPH_INLINE_NONE;
189 
190 		if (features & CEPH_FEATURE_MDS_QUOTA) {
191 			err = parse_reply_info_quota(p, end, info);
192 			if (err < 0)
193 				goto out_bad;
194 		} else {
195 			info->max_bytes = 0;
196 			info->max_files = 0;
197 		}
198 
199 		info->pool_ns_len = 0;
200 		info->pool_ns_data = NULL;
201 		if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
202 			ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
203 			if (info->pool_ns_len > 0) {
204 				ceph_decode_need(p, end, info->pool_ns_len, bad);
205 				info->pool_ns_data = *p;
206 				*p += info->pool_ns_len;
207 			}
208 		}
209 
210 		if (features & CEPH_FEATURE_FS_BTIME) {
211 			ceph_decode_need(p, end, sizeof(info->btime), bad);
212 			ceph_decode_copy(p, &info->btime, sizeof(info->btime));
213 			ceph_decode_64_safe(p, end, info->change_attr, bad);
214 		}
215 
216 		info->dir_pin = -ENODATA;
217 		/* info->snap_btime remains zero */
218 	}
219 	return 0;
220 bad:
221 	err = -EIO;
222 out_bad:
223 	return err;
224 }
225 
226 static int parse_reply_info_dir(void **p, void *end,
227 				struct ceph_mds_reply_dirfrag **dirfrag,
228 				u64 features)
229 {
230 	if (features == (u64)-1) {
231 		u8 struct_v, struct_compat;
232 		u32 struct_len;
233 		ceph_decode_8_safe(p, end, struct_v, bad);
234 		ceph_decode_8_safe(p, end, struct_compat, bad);
235 		/* struct_v is expected to be >= 1. we only understand
236 		 * encoding whose struct_compat == 1. */
237 		if (!struct_v || struct_compat != 1)
238 			goto bad;
239 		ceph_decode_32_safe(p, end, struct_len, bad);
240 		ceph_decode_need(p, end, struct_len, bad);
241 		end = *p + struct_len;
242 	}
243 
244 	ceph_decode_need(p, end, sizeof(**dirfrag), bad);
245 	*dirfrag = *p;
246 	*p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
247 	if (unlikely(*p > end))
248 		goto bad;
249 	if (features == (u64)-1)
250 		*p = end;
251 	return 0;
252 bad:
253 	return -EIO;
254 }
255 
256 static int parse_reply_info_lease(void **p, void *end,
257 				  struct ceph_mds_reply_lease **lease,
258 				  u64 features)
259 {
260 	if (features == (u64)-1) {
261 		u8 struct_v, struct_compat;
262 		u32 struct_len;
263 		ceph_decode_8_safe(p, end, struct_v, bad);
264 		ceph_decode_8_safe(p, end, struct_compat, bad);
265 		/* struct_v is expected to be >= 1. we only understand
266 		 * encoding whose struct_compat == 1. */
267 		if (!struct_v || struct_compat != 1)
268 			goto bad;
269 		ceph_decode_32_safe(p, end, struct_len, bad);
270 		ceph_decode_need(p, end, struct_len, bad);
271 		end = *p + struct_len;
272 	}
273 
274 	ceph_decode_need(p, end, sizeof(**lease), bad);
275 	*lease = *p;
276 	*p += sizeof(**lease);
277 	if (features == (u64)-1)
278 		*p = end;
279 	return 0;
280 bad:
281 	return -EIO;
282 }
283 
284 /*
285  * parse a normal reply, which may contain a (dir+)dentry and/or a
286  * target inode.
287  */
288 static int parse_reply_info_trace(void **p, void *end,
289 				  struct ceph_mds_reply_info_parsed *info,
290 				  u64 features)
291 {
292 	int err;
293 
294 	if (info->head->is_dentry) {
295 		err = parse_reply_info_in(p, end, &info->diri, features);
296 		if (err < 0)
297 			goto out_bad;
298 
299 		err = parse_reply_info_dir(p, end, &info->dirfrag, features);
300 		if (err < 0)
301 			goto out_bad;
302 
303 		ceph_decode_32_safe(p, end, info->dname_len, bad);
304 		ceph_decode_need(p, end, info->dname_len, bad);
305 		info->dname = *p;
306 		*p += info->dname_len;
307 
308 		err = parse_reply_info_lease(p, end, &info->dlease, features);
309 		if (err < 0)
310 			goto out_bad;
311 	}
312 
313 	if (info->head->is_target) {
314 		err = parse_reply_info_in(p, end, &info->targeti, features);
315 		if (err < 0)
316 			goto out_bad;
317 	}
318 
319 	if (unlikely(*p != end))
320 		goto bad;
321 	return 0;
322 
323 bad:
324 	err = -EIO;
325 out_bad:
326 	pr_err("problem parsing mds trace %d\n", err);
327 	return err;
328 }
329 
330 /*
331  * parse readdir results
332  */
333 static int parse_reply_info_readdir(void **p, void *end,
334 				struct ceph_mds_reply_info_parsed *info,
335 				u64 features)
336 {
337 	u32 num, i = 0;
338 	int err;
339 
340 	err = parse_reply_info_dir(p, end, &info->dir_dir, features);
341 	if (err < 0)
342 		goto out_bad;
343 
344 	ceph_decode_need(p, end, sizeof(num) + 2, bad);
345 	num = ceph_decode_32(p);
346 	{
347 		u16 flags = ceph_decode_16(p);
348 		info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
349 		info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
350 		info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
351 		info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
352 	}
353 	if (num == 0)
354 		goto done;
355 
356 	BUG_ON(!info->dir_entries);
357 	if ((unsigned long)(info->dir_entries + num) >
358 	    (unsigned long)info->dir_entries + info->dir_buf_size) {
359 		pr_err("dir contents are larger than expected\n");
360 		WARN_ON(1);
361 		goto bad;
362 	}
363 
364 	info->dir_nr = num;
365 	while (num) {
366 		struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
367 		/* dentry */
368 		ceph_decode_32_safe(p, end, rde->name_len, bad);
369 		ceph_decode_need(p, end, rde->name_len, bad);
370 		rde->name = *p;
371 		*p += rde->name_len;
372 		dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
373 
374 		/* dentry lease */
375 		err = parse_reply_info_lease(p, end, &rde->lease, features);
376 		if (err)
377 			goto out_bad;
378 		/* inode */
379 		err = parse_reply_info_in(p, end, &rde->inode, features);
380 		if (err < 0)
381 			goto out_bad;
382 		/* ceph_readdir_prepopulate() will update it */
383 		rde->offset = 0;
384 		i++;
385 		num--;
386 	}
387 
388 done:
389 	/* Skip over any unrecognized fields */
390 	*p = end;
391 	return 0;
392 
393 bad:
394 	err = -EIO;
395 out_bad:
396 	pr_err("problem parsing dir contents %d\n", err);
397 	return err;
398 }
399 
400 /*
401  * parse fcntl F_GETLK results
402  */
403 static int parse_reply_info_filelock(void **p, void *end,
404 				     struct ceph_mds_reply_info_parsed *info,
405 				     u64 features)
406 {
407 	if (*p + sizeof(*info->filelock_reply) > end)
408 		goto bad;
409 
410 	info->filelock_reply = *p;
411 
412 	/* Skip over any unrecognized fields */
413 	*p = end;
414 	return 0;
415 bad:
416 	return -EIO;
417 }
418 
419 
420 #if BITS_PER_LONG == 64
421 
422 #define DELEGATED_INO_AVAILABLE		xa_mk_value(1)
423 
424 static int ceph_parse_deleg_inos(void **p, void *end,
425 				 struct ceph_mds_session *s)
426 {
427 	u32 sets;
428 
429 	ceph_decode_32_safe(p, end, sets, bad);
430 	dout("got %u sets of delegated inodes\n", sets);
431 	while (sets--) {
432 		u64 start, len, ino;
433 
434 		ceph_decode_64_safe(p, end, start, bad);
435 		ceph_decode_64_safe(p, end, len, bad);
436 		while (len--) {
437 			int err = xa_insert(&s->s_delegated_inos, ino = start++,
438 					    DELEGATED_INO_AVAILABLE,
439 					    GFP_KERNEL);
440 			if (!err) {
441 				dout("added delegated inode 0x%llx\n",
442 				     start - 1);
443 			} else if (err == -EBUSY) {
444 				pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
445 					start - 1);
446 			} else {
447 				return err;
448 			}
449 		}
450 	}
451 	return 0;
452 bad:
453 	return -EIO;
454 }
455 
456 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
457 {
458 	unsigned long ino;
459 	void *val;
460 
461 	xa_for_each(&s->s_delegated_inos, ino, val) {
462 		val = xa_erase(&s->s_delegated_inos, ino);
463 		if (val == DELEGATED_INO_AVAILABLE)
464 			return ino;
465 	}
466 	return 0;
467 }
468 
469 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
470 {
471 	return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
472 			 GFP_KERNEL);
473 }
474 #else /* BITS_PER_LONG == 64 */
475 /*
476  * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
477  * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
478  * and bottom words?
479  */
480 static int ceph_parse_deleg_inos(void **p, void *end,
481 				 struct ceph_mds_session *s)
482 {
483 	u32 sets;
484 
485 	ceph_decode_32_safe(p, end, sets, bad);
486 	if (sets)
487 		ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
488 	return 0;
489 bad:
490 	return -EIO;
491 }
492 
493 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
494 {
495 	return 0;
496 }
497 
498 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
499 {
500 	return 0;
501 }
502 #endif /* BITS_PER_LONG == 64 */
503 
504 /*
505  * parse create results
506  */
507 static int parse_reply_info_create(void **p, void *end,
508 				  struct ceph_mds_reply_info_parsed *info,
509 				  u64 features, struct ceph_mds_session *s)
510 {
511 	int ret;
512 
513 	if (features == (u64)-1 ||
514 	    (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
515 		if (*p == end) {
516 			/* Malformed reply? */
517 			info->has_create_ino = false;
518 		} else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
519 			info->has_create_ino = true;
520 			/* struct_v, struct_compat, and len */
521 			ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
522 			ceph_decode_64_safe(p, end, info->ino, bad);
523 			ret = ceph_parse_deleg_inos(p, end, s);
524 			if (ret)
525 				return ret;
526 		} else {
527 			/* legacy */
528 			ceph_decode_64_safe(p, end, info->ino, bad);
529 			info->has_create_ino = true;
530 		}
531 	} else {
532 		if (*p != end)
533 			goto bad;
534 	}
535 
536 	/* Skip over any unrecognized fields */
537 	*p = end;
538 	return 0;
539 bad:
540 	return -EIO;
541 }
542 
543 /*
544  * parse extra results
545  */
546 static int parse_reply_info_extra(void **p, void *end,
547 				  struct ceph_mds_reply_info_parsed *info,
548 				  u64 features, struct ceph_mds_session *s)
549 {
550 	u32 op = le32_to_cpu(info->head->op);
551 
552 	if (op == CEPH_MDS_OP_GETFILELOCK)
553 		return parse_reply_info_filelock(p, end, info, features);
554 	else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
555 		return parse_reply_info_readdir(p, end, info, features);
556 	else if (op == CEPH_MDS_OP_CREATE)
557 		return parse_reply_info_create(p, end, info, features, s);
558 	else
559 		return -EIO;
560 }
561 
562 /*
563  * parse entire mds reply
564  */
565 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
566 			    struct ceph_mds_reply_info_parsed *info,
567 			    u64 features)
568 {
569 	void *p, *end;
570 	u32 len;
571 	int err;
572 
573 	info->head = msg->front.iov_base;
574 	p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
575 	end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
576 
577 	/* trace */
578 	ceph_decode_32_safe(&p, end, len, bad);
579 	if (len > 0) {
580 		ceph_decode_need(&p, end, len, bad);
581 		err = parse_reply_info_trace(&p, p+len, info, features);
582 		if (err < 0)
583 			goto out_bad;
584 	}
585 
586 	/* extra */
587 	ceph_decode_32_safe(&p, end, len, bad);
588 	if (len > 0) {
589 		ceph_decode_need(&p, end, len, bad);
590 		err = parse_reply_info_extra(&p, p+len, info, features, s);
591 		if (err < 0)
592 			goto out_bad;
593 	}
594 
595 	/* snap blob */
596 	ceph_decode_32_safe(&p, end, len, bad);
597 	info->snapblob_len = len;
598 	info->snapblob = p;
599 	p += len;
600 
601 	if (p != end)
602 		goto bad;
603 	return 0;
604 
605 bad:
606 	err = -EIO;
607 out_bad:
608 	pr_err("mds parse_reply err %d\n", err);
609 	return err;
610 }
611 
612 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
613 {
614 	if (!info->dir_entries)
615 		return;
616 	free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
617 }
618 
619 
620 /*
621  * sessions
622  */
623 const char *ceph_session_state_name(int s)
624 {
625 	switch (s) {
626 	case CEPH_MDS_SESSION_NEW: return "new";
627 	case CEPH_MDS_SESSION_OPENING: return "opening";
628 	case CEPH_MDS_SESSION_OPEN: return "open";
629 	case CEPH_MDS_SESSION_HUNG: return "hung";
630 	case CEPH_MDS_SESSION_CLOSING: return "closing";
631 	case CEPH_MDS_SESSION_CLOSED: return "closed";
632 	case CEPH_MDS_SESSION_RESTARTING: return "restarting";
633 	case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
634 	case CEPH_MDS_SESSION_REJECTED: return "rejected";
635 	default: return "???";
636 	}
637 }
638 
639 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
640 {
641 	if (refcount_inc_not_zero(&s->s_ref)) {
642 		dout("mdsc get_session %p %d -> %d\n", s,
643 		     refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
644 		return s;
645 	} else {
646 		dout("mdsc get_session %p 0 -- FAIL\n", s);
647 		return NULL;
648 	}
649 }
650 
651 void ceph_put_mds_session(struct ceph_mds_session *s)
652 {
653 	dout("mdsc put_session %p %d -> %d\n", s,
654 	     refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
655 	if (refcount_dec_and_test(&s->s_ref)) {
656 		if (s->s_auth.authorizer)
657 			ceph_auth_destroy_authorizer(s->s_auth.authorizer);
658 		WARN_ON(mutex_is_locked(&s->s_mutex));
659 		xa_destroy(&s->s_delegated_inos);
660 		kfree(s);
661 	}
662 }
663 
664 /*
665  * called under mdsc->mutex
666  */
667 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
668 						   int mds)
669 {
670 	if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
671 		return NULL;
672 	return ceph_get_mds_session(mdsc->sessions[mds]);
673 }
674 
675 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
676 {
677 	if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
678 		return false;
679 	else
680 		return true;
681 }
682 
683 static int __verify_registered_session(struct ceph_mds_client *mdsc,
684 				       struct ceph_mds_session *s)
685 {
686 	if (s->s_mds >= mdsc->max_sessions ||
687 	    mdsc->sessions[s->s_mds] != s)
688 		return -ENOENT;
689 	return 0;
690 }
691 
692 /*
693  * create+register a new session for given mds.
694  * called under mdsc->mutex.
695  */
696 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
697 						 int mds)
698 {
699 	struct ceph_mds_session *s;
700 
701 	if (mds >= mdsc->mdsmap->possible_max_rank)
702 		return ERR_PTR(-EINVAL);
703 
704 	s = kzalloc(sizeof(*s), GFP_NOFS);
705 	if (!s)
706 		return ERR_PTR(-ENOMEM);
707 
708 	if (mds >= mdsc->max_sessions) {
709 		int newmax = 1 << get_count_order(mds + 1);
710 		struct ceph_mds_session **sa;
711 
712 		dout("%s: realloc to %d\n", __func__, newmax);
713 		sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
714 		if (!sa)
715 			goto fail_realloc;
716 		if (mdsc->sessions) {
717 			memcpy(sa, mdsc->sessions,
718 			       mdsc->max_sessions * sizeof(void *));
719 			kfree(mdsc->sessions);
720 		}
721 		mdsc->sessions = sa;
722 		mdsc->max_sessions = newmax;
723 	}
724 
725 	dout("%s: mds%d\n", __func__, mds);
726 	s->s_mdsc = mdsc;
727 	s->s_mds = mds;
728 	s->s_state = CEPH_MDS_SESSION_NEW;
729 	s->s_ttl = 0;
730 	s->s_seq = 0;
731 	mutex_init(&s->s_mutex);
732 
733 	ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
734 
735 	spin_lock_init(&s->s_gen_ttl_lock);
736 	s->s_cap_gen = 1;
737 	s->s_cap_ttl = jiffies - 1;
738 
739 	spin_lock_init(&s->s_cap_lock);
740 	s->s_renew_requested = 0;
741 	s->s_renew_seq = 0;
742 	INIT_LIST_HEAD(&s->s_caps);
743 	s->s_nr_caps = 0;
744 	refcount_set(&s->s_ref, 1);
745 	INIT_LIST_HEAD(&s->s_waiting);
746 	INIT_LIST_HEAD(&s->s_unsafe);
747 	xa_init(&s->s_delegated_inos);
748 	s->s_num_cap_releases = 0;
749 	s->s_cap_reconnect = 0;
750 	s->s_cap_iterator = NULL;
751 	INIT_LIST_HEAD(&s->s_cap_releases);
752 	INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
753 
754 	INIT_LIST_HEAD(&s->s_cap_dirty);
755 	INIT_LIST_HEAD(&s->s_cap_flushing);
756 
757 	mdsc->sessions[mds] = s;
758 	atomic_inc(&mdsc->num_sessions);
759 	refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
760 
761 	ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
762 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
763 
764 	return s;
765 
766 fail_realloc:
767 	kfree(s);
768 	return ERR_PTR(-ENOMEM);
769 }
770 
771 /*
772  * called under mdsc->mutex
773  */
774 static void __unregister_session(struct ceph_mds_client *mdsc,
775 			       struct ceph_mds_session *s)
776 {
777 	dout("__unregister_session mds%d %p\n", s->s_mds, s);
778 	BUG_ON(mdsc->sessions[s->s_mds] != s);
779 	mdsc->sessions[s->s_mds] = NULL;
780 	ceph_con_close(&s->s_con);
781 	ceph_put_mds_session(s);
782 	atomic_dec(&mdsc->num_sessions);
783 }
784 
785 /*
786  * drop session refs in request.
787  *
788  * should be last request ref, or hold mdsc->mutex
789  */
790 static void put_request_session(struct ceph_mds_request *req)
791 {
792 	if (req->r_session) {
793 		ceph_put_mds_session(req->r_session);
794 		req->r_session = NULL;
795 	}
796 }
797 
798 void ceph_mdsc_release_request(struct kref *kref)
799 {
800 	struct ceph_mds_request *req = container_of(kref,
801 						    struct ceph_mds_request,
802 						    r_kref);
803 	ceph_mdsc_release_dir_caps_no_check(req);
804 	destroy_reply_info(&req->r_reply_info);
805 	if (req->r_request)
806 		ceph_msg_put(req->r_request);
807 	if (req->r_reply)
808 		ceph_msg_put(req->r_reply);
809 	if (req->r_inode) {
810 		ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
811 		/* avoid calling iput_final() in mds dispatch threads */
812 		ceph_async_iput(req->r_inode);
813 	}
814 	if (req->r_parent) {
815 		ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
816 		ceph_async_iput(req->r_parent);
817 	}
818 	ceph_async_iput(req->r_target_inode);
819 	if (req->r_dentry)
820 		dput(req->r_dentry);
821 	if (req->r_old_dentry)
822 		dput(req->r_old_dentry);
823 	if (req->r_old_dentry_dir) {
824 		/*
825 		 * track (and drop pins for) r_old_dentry_dir
826 		 * separately, since r_old_dentry's d_parent may have
827 		 * changed between the dir mutex being dropped and
828 		 * this request being freed.
829 		 */
830 		ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
831 				  CEPH_CAP_PIN);
832 		ceph_async_iput(req->r_old_dentry_dir);
833 	}
834 	kfree(req->r_path1);
835 	kfree(req->r_path2);
836 	put_cred(req->r_cred);
837 	if (req->r_pagelist)
838 		ceph_pagelist_release(req->r_pagelist);
839 	put_request_session(req);
840 	ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
841 	WARN_ON_ONCE(!list_empty(&req->r_wait));
842 	kmem_cache_free(ceph_mds_request_cachep, req);
843 }
844 
845 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
846 
847 /*
848  * lookup session, bump ref if found.
849  *
850  * called under mdsc->mutex.
851  */
852 static struct ceph_mds_request *
853 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
854 {
855 	struct ceph_mds_request *req;
856 
857 	req = lookup_request(&mdsc->request_tree, tid);
858 	if (req)
859 		ceph_mdsc_get_request(req);
860 
861 	return req;
862 }
863 
864 /*
865  * Register an in-flight request, and assign a tid.  Link to directory
866  * are modifying (if any).
867  *
868  * Called under mdsc->mutex.
869  */
870 static void __register_request(struct ceph_mds_client *mdsc,
871 			       struct ceph_mds_request *req,
872 			       struct inode *dir)
873 {
874 	int ret = 0;
875 
876 	req->r_tid = ++mdsc->last_tid;
877 	if (req->r_num_caps) {
878 		ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
879 					req->r_num_caps);
880 		if (ret < 0) {
881 			pr_err("__register_request %p "
882 			       "failed to reserve caps: %d\n", req, ret);
883 			/* set req->r_err to fail early from __do_request */
884 			req->r_err = ret;
885 			return;
886 		}
887 	}
888 	dout("__register_request %p tid %lld\n", req, req->r_tid);
889 	ceph_mdsc_get_request(req);
890 	insert_request(&mdsc->request_tree, req);
891 
892 	req->r_cred = get_current_cred();
893 
894 	if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
895 		mdsc->oldest_tid = req->r_tid;
896 
897 	if (dir) {
898 		struct ceph_inode_info *ci = ceph_inode(dir);
899 
900 		ihold(dir);
901 		req->r_unsafe_dir = dir;
902 		spin_lock(&ci->i_unsafe_lock);
903 		list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
904 		spin_unlock(&ci->i_unsafe_lock);
905 	}
906 }
907 
908 static void __unregister_request(struct ceph_mds_client *mdsc,
909 				 struct ceph_mds_request *req)
910 {
911 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
912 
913 	/* Never leave an unregistered request on an unsafe list! */
914 	list_del_init(&req->r_unsafe_item);
915 
916 	if (req->r_tid == mdsc->oldest_tid) {
917 		struct rb_node *p = rb_next(&req->r_node);
918 		mdsc->oldest_tid = 0;
919 		while (p) {
920 			struct ceph_mds_request *next_req =
921 				rb_entry(p, struct ceph_mds_request, r_node);
922 			if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
923 				mdsc->oldest_tid = next_req->r_tid;
924 				break;
925 			}
926 			p = rb_next(p);
927 		}
928 	}
929 
930 	erase_request(&mdsc->request_tree, req);
931 
932 	if (req->r_unsafe_dir) {
933 		struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
934 		spin_lock(&ci->i_unsafe_lock);
935 		list_del_init(&req->r_unsafe_dir_item);
936 		spin_unlock(&ci->i_unsafe_lock);
937 	}
938 	if (req->r_target_inode &&
939 	    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
940 		struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
941 		spin_lock(&ci->i_unsafe_lock);
942 		list_del_init(&req->r_unsafe_target_item);
943 		spin_unlock(&ci->i_unsafe_lock);
944 	}
945 
946 	if (req->r_unsafe_dir) {
947 		/* avoid calling iput_final() in mds dispatch threads */
948 		ceph_async_iput(req->r_unsafe_dir);
949 		req->r_unsafe_dir = NULL;
950 	}
951 
952 	complete_all(&req->r_safe_completion);
953 
954 	ceph_mdsc_put_request(req);
955 }
956 
957 /*
958  * Walk back up the dentry tree until we hit a dentry representing a
959  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
960  * when calling this) to ensure that the objects won't disappear while we're
961  * working with them. Once we hit a candidate dentry, we attempt to take a
962  * reference to it, and return that as the result.
963  */
964 static struct inode *get_nonsnap_parent(struct dentry *dentry)
965 {
966 	struct inode *inode = NULL;
967 
968 	while (dentry && !IS_ROOT(dentry)) {
969 		inode = d_inode_rcu(dentry);
970 		if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
971 			break;
972 		dentry = dentry->d_parent;
973 	}
974 	if (inode)
975 		inode = igrab(inode);
976 	return inode;
977 }
978 
979 /*
980  * Choose mds to send request to next.  If there is a hint set in the
981  * request (e.g., due to a prior forward hint from the mds), use that.
982  * Otherwise, consult frag tree and/or caps to identify the
983  * appropriate mds.  If all else fails, choose randomly.
984  *
985  * Called under mdsc->mutex.
986  */
987 static int __choose_mds(struct ceph_mds_client *mdsc,
988 			struct ceph_mds_request *req,
989 			bool *random)
990 {
991 	struct inode *inode;
992 	struct ceph_inode_info *ci;
993 	struct ceph_cap *cap;
994 	int mode = req->r_direct_mode;
995 	int mds = -1;
996 	u32 hash = req->r_direct_hash;
997 	bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
998 
999 	if (random)
1000 		*random = false;
1001 
1002 	/*
1003 	 * is there a specific mds we should try?  ignore hint if we have
1004 	 * no session and the mds is not up (active or recovering).
1005 	 */
1006 	if (req->r_resend_mds >= 0 &&
1007 	    (__have_session(mdsc, req->r_resend_mds) ||
1008 	     ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1009 		dout("%s using resend_mds mds%d\n", __func__,
1010 		     req->r_resend_mds);
1011 		return req->r_resend_mds;
1012 	}
1013 
1014 	if (mode == USE_RANDOM_MDS)
1015 		goto random;
1016 
1017 	inode = NULL;
1018 	if (req->r_inode) {
1019 		if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1020 			inode = req->r_inode;
1021 			ihold(inode);
1022 		} else {
1023 			/* req->r_dentry is non-null for LSSNAP request */
1024 			rcu_read_lock();
1025 			inode = get_nonsnap_parent(req->r_dentry);
1026 			rcu_read_unlock();
1027 			dout("%s using snapdir's parent %p\n", __func__, inode);
1028 		}
1029 	} else if (req->r_dentry) {
1030 		/* ignore race with rename; old or new d_parent is okay */
1031 		struct dentry *parent;
1032 		struct inode *dir;
1033 
1034 		rcu_read_lock();
1035 		parent = READ_ONCE(req->r_dentry->d_parent);
1036 		dir = req->r_parent ? : d_inode_rcu(parent);
1037 
1038 		if (!dir || dir->i_sb != mdsc->fsc->sb) {
1039 			/*  not this fs or parent went negative */
1040 			inode = d_inode(req->r_dentry);
1041 			if (inode)
1042 				ihold(inode);
1043 		} else if (ceph_snap(dir) != CEPH_NOSNAP) {
1044 			/* direct snapped/virtual snapdir requests
1045 			 * based on parent dir inode */
1046 			inode = get_nonsnap_parent(parent);
1047 			dout("%s using nonsnap parent %p\n", __func__, inode);
1048 		} else {
1049 			/* dentry target */
1050 			inode = d_inode(req->r_dentry);
1051 			if (!inode || mode == USE_AUTH_MDS) {
1052 				/* dir + name */
1053 				inode = igrab(dir);
1054 				hash = ceph_dentry_hash(dir, req->r_dentry);
1055 				is_hash = true;
1056 			} else {
1057 				ihold(inode);
1058 			}
1059 		}
1060 		rcu_read_unlock();
1061 	}
1062 
1063 	dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1064 	     hash, mode);
1065 	if (!inode)
1066 		goto random;
1067 	ci = ceph_inode(inode);
1068 
1069 	if (is_hash && S_ISDIR(inode->i_mode)) {
1070 		struct ceph_inode_frag frag;
1071 		int found;
1072 
1073 		ceph_choose_frag(ci, hash, &frag, &found);
1074 		if (found) {
1075 			if (mode == USE_ANY_MDS && frag.ndist > 0) {
1076 				u8 r;
1077 
1078 				/* choose a random replica */
1079 				get_random_bytes(&r, 1);
1080 				r %= frag.ndist;
1081 				mds = frag.dist[r];
1082 				dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1083 				     __func__, inode, ceph_vinop(inode),
1084 				     frag.frag, mds, (int)r, frag.ndist);
1085 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1086 				    CEPH_MDS_STATE_ACTIVE &&
1087 				    !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1088 					goto out;
1089 			}
1090 
1091 			/* since this file/dir wasn't known to be
1092 			 * replicated, then we want to look for the
1093 			 * authoritative mds. */
1094 			if (frag.mds >= 0) {
1095 				/* choose auth mds */
1096 				mds = frag.mds;
1097 				dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1098 				     __func__, inode, ceph_vinop(inode),
1099 				     frag.frag, mds);
1100 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1101 				    CEPH_MDS_STATE_ACTIVE) {
1102 					if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1103 								  mds))
1104 						goto out;
1105 				}
1106 			}
1107 			mode = USE_AUTH_MDS;
1108 		}
1109 	}
1110 
1111 	spin_lock(&ci->i_ceph_lock);
1112 	cap = NULL;
1113 	if (mode == USE_AUTH_MDS)
1114 		cap = ci->i_auth_cap;
1115 	if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1116 		cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1117 	if (!cap) {
1118 		spin_unlock(&ci->i_ceph_lock);
1119 		ceph_async_iput(inode);
1120 		goto random;
1121 	}
1122 	mds = cap->session->s_mds;
1123 	dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1124 	     inode, ceph_vinop(inode), mds,
1125 	     cap == ci->i_auth_cap ? "auth " : "", cap);
1126 	spin_unlock(&ci->i_ceph_lock);
1127 out:
1128 	/* avoid calling iput_final() while holding mdsc->mutex or
1129 	 * in mds dispatch threads */
1130 	ceph_async_iput(inode);
1131 	return mds;
1132 
1133 random:
1134 	if (random)
1135 		*random = true;
1136 
1137 	mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1138 	dout("%s chose random mds%d\n", __func__, mds);
1139 	return mds;
1140 }
1141 
1142 
1143 /*
1144  * session messages
1145  */
1146 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1147 {
1148 	struct ceph_msg *msg;
1149 	struct ceph_mds_session_head *h;
1150 
1151 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1152 			   false);
1153 	if (!msg) {
1154 		pr_err("create_session_msg ENOMEM creating msg\n");
1155 		return NULL;
1156 	}
1157 	h = msg->front.iov_base;
1158 	h->op = cpu_to_le32(op);
1159 	h->seq = cpu_to_le64(seq);
1160 
1161 	return msg;
1162 }
1163 
1164 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1165 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1166 static int encode_supported_features(void **p, void *end)
1167 {
1168 	static const size_t count = ARRAY_SIZE(feature_bits);
1169 
1170 	if (count > 0) {
1171 		size_t i;
1172 		size_t size = FEATURE_BYTES(count);
1173 
1174 		if (WARN_ON_ONCE(*p + 4 + size > end))
1175 			return -ERANGE;
1176 
1177 		ceph_encode_32(p, size);
1178 		memset(*p, 0, size);
1179 		for (i = 0; i < count; i++)
1180 			((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1181 		*p += size;
1182 	} else {
1183 		if (WARN_ON_ONCE(*p + 4 > end))
1184 			return -ERANGE;
1185 
1186 		ceph_encode_32(p, 0);
1187 	}
1188 
1189 	return 0;
1190 }
1191 
1192 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1193 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1194 static int encode_metric_spec(void **p, void *end)
1195 {
1196 	static const size_t count = ARRAY_SIZE(metric_bits);
1197 
1198 	/* header */
1199 	if (WARN_ON_ONCE(*p + 2 > end))
1200 		return -ERANGE;
1201 
1202 	ceph_encode_8(p, 1); /* version */
1203 	ceph_encode_8(p, 1); /* compat */
1204 
1205 	if (count > 0) {
1206 		size_t i;
1207 		size_t size = METRIC_BYTES(count);
1208 
1209 		if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1210 			return -ERANGE;
1211 
1212 		/* metric spec info length */
1213 		ceph_encode_32(p, 4 + size);
1214 
1215 		/* metric spec */
1216 		ceph_encode_32(p, size);
1217 		memset(*p, 0, size);
1218 		for (i = 0; i < count; i++)
1219 			((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1220 		*p += size;
1221 	} else {
1222 		if (WARN_ON_ONCE(*p + 4 + 4 > end))
1223 			return -ERANGE;
1224 
1225 		/* metric spec info length */
1226 		ceph_encode_32(p, 4);
1227 		/* metric spec */
1228 		ceph_encode_32(p, 0);
1229 	}
1230 
1231 	return 0;
1232 }
1233 
1234 /*
1235  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1236  * to include additional client metadata fields.
1237  */
1238 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1239 {
1240 	struct ceph_msg *msg;
1241 	struct ceph_mds_session_head *h;
1242 	int i;
1243 	int extra_bytes = 0;
1244 	int metadata_key_count = 0;
1245 	struct ceph_options *opt = mdsc->fsc->client->options;
1246 	struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1247 	size_t size, count;
1248 	void *p, *end;
1249 	int ret;
1250 
1251 	const char* metadata[][2] = {
1252 		{"hostname", mdsc->nodename},
1253 		{"kernel_version", init_utsname()->release},
1254 		{"entity_id", opt->name ? : ""},
1255 		{"root", fsopt->server_path ? : "/"},
1256 		{NULL, NULL}
1257 	};
1258 
1259 	/* Calculate serialized length of metadata */
1260 	extra_bytes = 4;  /* map length */
1261 	for (i = 0; metadata[i][0]; ++i) {
1262 		extra_bytes += 8 + strlen(metadata[i][0]) +
1263 			strlen(metadata[i][1]);
1264 		metadata_key_count++;
1265 	}
1266 
1267 	/* supported feature */
1268 	size = 0;
1269 	count = ARRAY_SIZE(feature_bits);
1270 	if (count > 0)
1271 		size = FEATURE_BYTES(count);
1272 	extra_bytes += 4 + size;
1273 
1274 	/* metric spec */
1275 	size = 0;
1276 	count = ARRAY_SIZE(metric_bits);
1277 	if (count > 0)
1278 		size = METRIC_BYTES(count);
1279 	extra_bytes += 2 + 4 + 4 + size;
1280 
1281 	/* Allocate the message */
1282 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1283 			   GFP_NOFS, false);
1284 	if (!msg) {
1285 		pr_err("create_session_msg ENOMEM creating msg\n");
1286 		return ERR_PTR(-ENOMEM);
1287 	}
1288 	p = msg->front.iov_base;
1289 	end = p + msg->front.iov_len;
1290 
1291 	h = p;
1292 	h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1293 	h->seq = cpu_to_le64(seq);
1294 
1295 	/*
1296 	 * Serialize client metadata into waiting buffer space, using
1297 	 * the format that userspace expects for map<string, string>
1298 	 *
1299 	 * ClientSession messages with metadata are v4
1300 	 */
1301 	msg->hdr.version = cpu_to_le16(4);
1302 	msg->hdr.compat_version = cpu_to_le16(1);
1303 
1304 	/* The write pointer, following the session_head structure */
1305 	p += sizeof(*h);
1306 
1307 	/* Number of entries in the map */
1308 	ceph_encode_32(&p, metadata_key_count);
1309 
1310 	/* Two length-prefixed strings for each entry in the map */
1311 	for (i = 0; metadata[i][0]; ++i) {
1312 		size_t const key_len = strlen(metadata[i][0]);
1313 		size_t const val_len = strlen(metadata[i][1]);
1314 
1315 		ceph_encode_32(&p, key_len);
1316 		memcpy(p, metadata[i][0], key_len);
1317 		p += key_len;
1318 		ceph_encode_32(&p, val_len);
1319 		memcpy(p, metadata[i][1], val_len);
1320 		p += val_len;
1321 	}
1322 
1323 	ret = encode_supported_features(&p, end);
1324 	if (ret) {
1325 		pr_err("encode_supported_features failed!\n");
1326 		ceph_msg_put(msg);
1327 		return ERR_PTR(ret);
1328 	}
1329 
1330 	ret = encode_metric_spec(&p, end);
1331 	if (ret) {
1332 		pr_err("encode_metric_spec failed!\n");
1333 		ceph_msg_put(msg);
1334 		return ERR_PTR(ret);
1335 	}
1336 
1337 	msg->front.iov_len = p - msg->front.iov_base;
1338 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1339 
1340 	return msg;
1341 }
1342 
1343 /*
1344  * send session open request.
1345  *
1346  * called under mdsc->mutex
1347  */
1348 static int __open_session(struct ceph_mds_client *mdsc,
1349 			  struct ceph_mds_session *session)
1350 {
1351 	struct ceph_msg *msg;
1352 	int mstate;
1353 	int mds = session->s_mds;
1354 
1355 	/* wait for mds to go active? */
1356 	mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1357 	dout("open_session to mds%d (%s)\n", mds,
1358 	     ceph_mds_state_name(mstate));
1359 	session->s_state = CEPH_MDS_SESSION_OPENING;
1360 	session->s_renew_requested = jiffies;
1361 
1362 	/* send connect message */
1363 	msg = create_session_open_msg(mdsc, session->s_seq);
1364 	if (IS_ERR(msg))
1365 		return PTR_ERR(msg);
1366 	ceph_con_send(&session->s_con, msg);
1367 	return 0;
1368 }
1369 
1370 /*
1371  * open sessions for any export targets for the given mds
1372  *
1373  * called under mdsc->mutex
1374  */
1375 static struct ceph_mds_session *
1376 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1377 {
1378 	struct ceph_mds_session *session;
1379 	int ret;
1380 
1381 	session = __ceph_lookup_mds_session(mdsc, target);
1382 	if (!session) {
1383 		session = register_session(mdsc, target);
1384 		if (IS_ERR(session))
1385 			return session;
1386 	}
1387 	if (session->s_state == CEPH_MDS_SESSION_NEW ||
1388 	    session->s_state == CEPH_MDS_SESSION_CLOSING) {
1389 		ret = __open_session(mdsc, session);
1390 		if (ret)
1391 			return ERR_PTR(ret);
1392 	}
1393 
1394 	return session;
1395 }
1396 
1397 struct ceph_mds_session *
1398 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1399 {
1400 	struct ceph_mds_session *session;
1401 
1402 	dout("open_export_target_session to mds%d\n", target);
1403 
1404 	mutex_lock(&mdsc->mutex);
1405 	session = __open_export_target_session(mdsc, target);
1406 	mutex_unlock(&mdsc->mutex);
1407 
1408 	return session;
1409 }
1410 
1411 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1412 					  struct ceph_mds_session *session)
1413 {
1414 	struct ceph_mds_info *mi;
1415 	struct ceph_mds_session *ts;
1416 	int i, mds = session->s_mds;
1417 
1418 	if (mds >= mdsc->mdsmap->possible_max_rank)
1419 		return;
1420 
1421 	mi = &mdsc->mdsmap->m_info[mds];
1422 	dout("open_export_target_sessions for mds%d (%d targets)\n",
1423 	     session->s_mds, mi->num_export_targets);
1424 
1425 	for (i = 0; i < mi->num_export_targets; i++) {
1426 		ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1427 		if (!IS_ERR(ts))
1428 			ceph_put_mds_session(ts);
1429 	}
1430 }
1431 
1432 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1433 					   struct ceph_mds_session *session)
1434 {
1435 	mutex_lock(&mdsc->mutex);
1436 	__open_export_target_sessions(mdsc, session);
1437 	mutex_unlock(&mdsc->mutex);
1438 }
1439 
1440 /*
1441  * session caps
1442  */
1443 
1444 static void detach_cap_releases(struct ceph_mds_session *session,
1445 				struct list_head *target)
1446 {
1447 	lockdep_assert_held(&session->s_cap_lock);
1448 
1449 	list_splice_init(&session->s_cap_releases, target);
1450 	session->s_num_cap_releases = 0;
1451 	dout("dispose_cap_releases mds%d\n", session->s_mds);
1452 }
1453 
1454 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1455 				 struct list_head *dispose)
1456 {
1457 	while (!list_empty(dispose)) {
1458 		struct ceph_cap *cap;
1459 		/* zero out the in-progress message */
1460 		cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1461 		list_del(&cap->session_caps);
1462 		ceph_put_cap(mdsc, cap);
1463 	}
1464 }
1465 
1466 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1467 				     struct ceph_mds_session *session)
1468 {
1469 	struct ceph_mds_request *req;
1470 	struct rb_node *p;
1471 	struct ceph_inode_info *ci;
1472 
1473 	dout("cleanup_session_requests mds%d\n", session->s_mds);
1474 	mutex_lock(&mdsc->mutex);
1475 	while (!list_empty(&session->s_unsafe)) {
1476 		req = list_first_entry(&session->s_unsafe,
1477 				       struct ceph_mds_request, r_unsafe_item);
1478 		pr_warn_ratelimited(" dropping unsafe request %llu\n",
1479 				    req->r_tid);
1480 		if (req->r_target_inode) {
1481 			/* dropping unsafe change of inode's attributes */
1482 			ci = ceph_inode(req->r_target_inode);
1483 			errseq_set(&ci->i_meta_err, -EIO);
1484 		}
1485 		if (req->r_unsafe_dir) {
1486 			/* dropping unsafe directory operation */
1487 			ci = ceph_inode(req->r_unsafe_dir);
1488 			errseq_set(&ci->i_meta_err, -EIO);
1489 		}
1490 		__unregister_request(mdsc, req);
1491 	}
1492 	/* zero r_attempts, so kick_requests() will re-send requests */
1493 	p = rb_first(&mdsc->request_tree);
1494 	while (p) {
1495 		req = rb_entry(p, struct ceph_mds_request, r_node);
1496 		p = rb_next(p);
1497 		if (req->r_session &&
1498 		    req->r_session->s_mds == session->s_mds)
1499 			req->r_attempts = 0;
1500 	}
1501 	mutex_unlock(&mdsc->mutex);
1502 }
1503 
1504 /*
1505  * Helper to safely iterate over all caps associated with a session, with
1506  * special care taken to handle a racing __ceph_remove_cap().
1507  *
1508  * Caller must hold session s_mutex.
1509  */
1510 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1511 			      int (*cb)(struct inode *, struct ceph_cap *,
1512 					void *), void *arg)
1513 {
1514 	struct list_head *p;
1515 	struct ceph_cap *cap;
1516 	struct inode *inode, *last_inode = NULL;
1517 	struct ceph_cap *old_cap = NULL;
1518 	int ret;
1519 
1520 	dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1521 	spin_lock(&session->s_cap_lock);
1522 	p = session->s_caps.next;
1523 	while (p != &session->s_caps) {
1524 		cap = list_entry(p, struct ceph_cap, session_caps);
1525 		inode = igrab(&cap->ci->vfs_inode);
1526 		if (!inode) {
1527 			p = p->next;
1528 			continue;
1529 		}
1530 		session->s_cap_iterator = cap;
1531 		spin_unlock(&session->s_cap_lock);
1532 
1533 		if (last_inode) {
1534 			/* avoid calling iput_final() while holding
1535 			 * s_mutex or in mds dispatch threads */
1536 			ceph_async_iput(last_inode);
1537 			last_inode = NULL;
1538 		}
1539 		if (old_cap) {
1540 			ceph_put_cap(session->s_mdsc, old_cap);
1541 			old_cap = NULL;
1542 		}
1543 
1544 		ret = cb(inode, cap, arg);
1545 		last_inode = inode;
1546 
1547 		spin_lock(&session->s_cap_lock);
1548 		p = p->next;
1549 		if (!cap->ci) {
1550 			dout("iterate_session_caps  finishing cap %p removal\n",
1551 			     cap);
1552 			BUG_ON(cap->session != session);
1553 			cap->session = NULL;
1554 			list_del_init(&cap->session_caps);
1555 			session->s_nr_caps--;
1556 			atomic64_dec(&session->s_mdsc->metric.total_caps);
1557 			if (cap->queue_release)
1558 				__ceph_queue_cap_release(session, cap);
1559 			else
1560 				old_cap = cap;  /* put_cap it w/o locks held */
1561 		}
1562 		if (ret < 0)
1563 			goto out;
1564 	}
1565 	ret = 0;
1566 out:
1567 	session->s_cap_iterator = NULL;
1568 	spin_unlock(&session->s_cap_lock);
1569 
1570 	ceph_async_iput(last_inode);
1571 	if (old_cap)
1572 		ceph_put_cap(session->s_mdsc, old_cap);
1573 
1574 	return ret;
1575 }
1576 
1577 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1578 				  void *arg)
1579 {
1580 	struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1581 	struct ceph_inode_info *ci = ceph_inode(inode);
1582 	LIST_HEAD(to_remove);
1583 	bool dirty_dropped = false;
1584 	bool invalidate = false;
1585 
1586 	dout("removing cap %p, ci is %p, inode is %p\n",
1587 	     cap, ci, &ci->vfs_inode);
1588 	spin_lock(&ci->i_ceph_lock);
1589 	__ceph_remove_cap(cap, false);
1590 	if (!ci->i_auth_cap) {
1591 		struct ceph_cap_flush *cf;
1592 		struct ceph_mds_client *mdsc = fsc->mdsc;
1593 
1594 		if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1595 			if (inode->i_data.nrpages > 0)
1596 				invalidate = true;
1597 			if (ci->i_wrbuffer_ref > 0)
1598 				mapping_set_error(&inode->i_data, -EIO);
1599 		}
1600 
1601 		while (!list_empty(&ci->i_cap_flush_list)) {
1602 			cf = list_first_entry(&ci->i_cap_flush_list,
1603 					      struct ceph_cap_flush, i_list);
1604 			list_move(&cf->i_list, &to_remove);
1605 		}
1606 
1607 		spin_lock(&mdsc->cap_dirty_lock);
1608 
1609 		list_for_each_entry(cf, &to_remove, i_list)
1610 			list_del(&cf->g_list);
1611 
1612 		if (!list_empty(&ci->i_dirty_item)) {
1613 			pr_warn_ratelimited(
1614 				" dropping dirty %s state for %p %lld\n",
1615 				ceph_cap_string(ci->i_dirty_caps),
1616 				inode, ceph_ino(inode));
1617 			ci->i_dirty_caps = 0;
1618 			list_del_init(&ci->i_dirty_item);
1619 			dirty_dropped = true;
1620 		}
1621 		if (!list_empty(&ci->i_flushing_item)) {
1622 			pr_warn_ratelimited(
1623 				" dropping dirty+flushing %s state for %p %lld\n",
1624 				ceph_cap_string(ci->i_flushing_caps),
1625 				inode, ceph_ino(inode));
1626 			ci->i_flushing_caps = 0;
1627 			list_del_init(&ci->i_flushing_item);
1628 			mdsc->num_cap_flushing--;
1629 			dirty_dropped = true;
1630 		}
1631 		spin_unlock(&mdsc->cap_dirty_lock);
1632 
1633 		if (dirty_dropped) {
1634 			errseq_set(&ci->i_meta_err, -EIO);
1635 
1636 			if (ci->i_wrbuffer_ref_head == 0 &&
1637 			    ci->i_wr_ref == 0 &&
1638 			    ci->i_dirty_caps == 0 &&
1639 			    ci->i_flushing_caps == 0) {
1640 				ceph_put_snap_context(ci->i_head_snapc);
1641 				ci->i_head_snapc = NULL;
1642 			}
1643 		}
1644 
1645 		if (atomic_read(&ci->i_filelock_ref) > 0) {
1646 			/* make further file lock syscall return -EIO */
1647 			ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1648 			pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1649 					    inode, ceph_ino(inode));
1650 		}
1651 
1652 		if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1653 			list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1654 			ci->i_prealloc_cap_flush = NULL;
1655 		}
1656 	}
1657 	spin_unlock(&ci->i_ceph_lock);
1658 	while (!list_empty(&to_remove)) {
1659 		struct ceph_cap_flush *cf;
1660 		cf = list_first_entry(&to_remove,
1661 				      struct ceph_cap_flush, i_list);
1662 		list_del(&cf->i_list);
1663 		ceph_free_cap_flush(cf);
1664 	}
1665 
1666 	wake_up_all(&ci->i_cap_wq);
1667 	if (invalidate)
1668 		ceph_queue_invalidate(inode);
1669 	if (dirty_dropped)
1670 		iput(inode);
1671 	return 0;
1672 }
1673 
1674 /*
1675  * caller must hold session s_mutex
1676  */
1677 static void remove_session_caps(struct ceph_mds_session *session)
1678 {
1679 	struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1680 	struct super_block *sb = fsc->sb;
1681 	LIST_HEAD(dispose);
1682 
1683 	dout("remove_session_caps on %p\n", session);
1684 	ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1685 
1686 	wake_up_all(&fsc->mdsc->cap_flushing_wq);
1687 
1688 	spin_lock(&session->s_cap_lock);
1689 	if (session->s_nr_caps > 0) {
1690 		struct inode *inode;
1691 		struct ceph_cap *cap, *prev = NULL;
1692 		struct ceph_vino vino;
1693 		/*
1694 		 * iterate_session_caps() skips inodes that are being
1695 		 * deleted, we need to wait until deletions are complete.
1696 		 * __wait_on_freeing_inode() is designed for the job,
1697 		 * but it is not exported, so use lookup inode function
1698 		 * to access it.
1699 		 */
1700 		while (!list_empty(&session->s_caps)) {
1701 			cap = list_entry(session->s_caps.next,
1702 					 struct ceph_cap, session_caps);
1703 			if (cap == prev)
1704 				break;
1705 			prev = cap;
1706 			vino = cap->ci->i_vino;
1707 			spin_unlock(&session->s_cap_lock);
1708 
1709 			inode = ceph_find_inode(sb, vino);
1710 			 /* avoid calling iput_final() while holding s_mutex */
1711 			ceph_async_iput(inode);
1712 
1713 			spin_lock(&session->s_cap_lock);
1714 		}
1715 	}
1716 
1717 	// drop cap expires and unlock s_cap_lock
1718 	detach_cap_releases(session, &dispose);
1719 
1720 	BUG_ON(session->s_nr_caps > 0);
1721 	BUG_ON(!list_empty(&session->s_cap_flushing));
1722 	spin_unlock(&session->s_cap_lock);
1723 	dispose_cap_releases(session->s_mdsc, &dispose);
1724 }
1725 
1726 enum {
1727 	RECONNECT,
1728 	RENEWCAPS,
1729 	FORCE_RO,
1730 };
1731 
1732 /*
1733  * wake up any threads waiting on this session's caps.  if the cap is
1734  * old (didn't get renewed on the client reconnect), remove it now.
1735  *
1736  * caller must hold s_mutex.
1737  */
1738 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1739 			      void *arg)
1740 {
1741 	struct ceph_inode_info *ci = ceph_inode(inode);
1742 	unsigned long ev = (unsigned long)arg;
1743 
1744 	if (ev == RECONNECT) {
1745 		spin_lock(&ci->i_ceph_lock);
1746 		ci->i_wanted_max_size = 0;
1747 		ci->i_requested_max_size = 0;
1748 		spin_unlock(&ci->i_ceph_lock);
1749 	} else if (ev == RENEWCAPS) {
1750 		if (cap->cap_gen < cap->session->s_cap_gen) {
1751 			/* mds did not re-issue stale cap */
1752 			spin_lock(&ci->i_ceph_lock);
1753 			cap->issued = cap->implemented = CEPH_CAP_PIN;
1754 			spin_unlock(&ci->i_ceph_lock);
1755 		}
1756 	} else if (ev == FORCE_RO) {
1757 	}
1758 	wake_up_all(&ci->i_cap_wq);
1759 	return 0;
1760 }
1761 
1762 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1763 {
1764 	dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1765 	ceph_iterate_session_caps(session, wake_up_session_cb,
1766 				  (void *)(unsigned long)ev);
1767 }
1768 
1769 /*
1770  * Send periodic message to MDS renewing all currently held caps.  The
1771  * ack will reset the expiration for all caps from this session.
1772  *
1773  * caller holds s_mutex
1774  */
1775 static int send_renew_caps(struct ceph_mds_client *mdsc,
1776 			   struct ceph_mds_session *session)
1777 {
1778 	struct ceph_msg *msg;
1779 	int state;
1780 
1781 	if (time_after_eq(jiffies, session->s_cap_ttl) &&
1782 	    time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1783 		pr_info("mds%d caps stale\n", session->s_mds);
1784 	session->s_renew_requested = jiffies;
1785 
1786 	/* do not try to renew caps until a recovering mds has reconnected
1787 	 * with its clients. */
1788 	state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1789 	if (state < CEPH_MDS_STATE_RECONNECT) {
1790 		dout("send_renew_caps ignoring mds%d (%s)\n",
1791 		     session->s_mds, ceph_mds_state_name(state));
1792 		return 0;
1793 	}
1794 
1795 	dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1796 		ceph_mds_state_name(state));
1797 	msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1798 				 ++session->s_renew_seq);
1799 	if (!msg)
1800 		return -ENOMEM;
1801 	ceph_con_send(&session->s_con, msg);
1802 	return 0;
1803 }
1804 
1805 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1806 			     struct ceph_mds_session *session, u64 seq)
1807 {
1808 	struct ceph_msg *msg;
1809 
1810 	dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1811 	     session->s_mds, ceph_session_state_name(session->s_state), seq);
1812 	msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1813 	if (!msg)
1814 		return -ENOMEM;
1815 	ceph_con_send(&session->s_con, msg);
1816 	return 0;
1817 }
1818 
1819 
1820 /*
1821  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1822  *
1823  * Called under session->s_mutex
1824  */
1825 static void renewed_caps(struct ceph_mds_client *mdsc,
1826 			 struct ceph_mds_session *session, int is_renew)
1827 {
1828 	int was_stale;
1829 	int wake = 0;
1830 
1831 	spin_lock(&session->s_cap_lock);
1832 	was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1833 
1834 	session->s_cap_ttl = session->s_renew_requested +
1835 		mdsc->mdsmap->m_session_timeout*HZ;
1836 
1837 	if (was_stale) {
1838 		if (time_before(jiffies, session->s_cap_ttl)) {
1839 			pr_info("mds%d caps renewed\n", session->s_mds);
1840 			wake = 1;
1841 		} else {
1842 			pr_info("mds%d caps still stale\n", session->s_mds);
1843 		}
1844 	}
1845 	dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1846 	     session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1847 	     time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1848 	spin_unlock(&session->s_cap_lock);
1849 
1850 	if (wake)
1851 		wake_up_session_caps(session, RENEWCAPS);
1852 }
1853 
1854 /*
1855  * send a session close request
1856  */
1857 static int request_close_session(struct ceph_mds_session *session)
1858 {
1859 	struct ceph_msg *msg;
1860 
1861 	dout("request_close_session mds%d state %s seq %lld\n",
1862 	     session->s_mds, ceph_session_state_name(session->s_state),
1863 	     session->s_seq);
1864 	msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1865 	if (!msg)
1866 		return -ENOMEM;
1867 	ceph_con_send(&session->s_con, msg);
1868 	return 1;
1869 }
1870 
1871 /*
1872  * Called with s_mutex held.
1873  */
1874 static int __close_session(struct ceph_mds_client *mdsc,
1875 			 struct ceph_mds_session *session)
1876 {
1877 	if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1878 		return 0;
1879 	session->s_state = CEPH_MDS_SESSION_CLOSING;
1880 	return request_close_session(session);
1881 }
1882 
1883 static bool drop_negative_children(struct dentry *dentry)
1884 {
1885 	struct dentry *child;
1886 	bool all_negative = true;
1887 
1888 	if (!d_is_dir(dentry))
1889 		goto out;
1890 
1891 	spin_lock(&dentry->d_lock);
1892 	list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1893 		if (d_really_is_positive(child)) {
1894 			all_negative = false;
1895 			break;
1896 		}
1897 	}
1898 	spin_unlock(&dentry->d_lock);
1899 
1900 	if (all_negative)
1901 		shrink_dcache_parent(dentry);
1902 out:
1903 	return all_negative;
1904 }
1905 
1906 /*
1907  * Trim old(er) caps.
1908  *
1909  * Because we can't cache an inode without one or more caps, we do
1910  * this indirectly: if a cap is unused, we prune its aliases, at which
1911  * point the inode will hopefully get dropped to.
1912  *
1913  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1914  * memory pressure from the MDS, though, so it needn't be perfect.
1915  */
1916 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1917 {
1918 	int *remaining = arg;
1919 	struct ceph_inode_info *ci = ceph_inode(inode);
1920 	int used, wanted, oissued, mine;
1921 
1922 	if (*remaining <= 0)
1923 		return -1;
1924 
1925 	spin_lock(&ci->i_ceph_lock);
1926 	mine = cap->issued | cap->implemented;
1927 	used = __ceph_caps_used(ci);
1928 	wanted = __ceph_caps_file_wanted(ci);
1929 	oissued = __ceph_caps_issued_other(ci, cap);
1930 
1931 	dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1932 	     inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1933 	     ceph_cap_string(used), ceph_cap_string(wanted));
1934 	if (cap == ci->i_auth_cap) {
1935 		if (ci->i_dirty_caps || ci->i_flushing_caps ||
1936 		    !list_empty(&ci->i_cap_snaps))
1937 			goto out;
1938 		if ((used | wanted) & CEPH_CAP_ANY_WR)
1939 			goto out;
1940 		/* Note: it's possible that i_filelock_ref becomes non-zero
1941 		 * after dropping auth caps. It doesn't hurt because reply
1942 		 * of lock mds request will re-add auth caps. */
1943 		if (atomic_read(&ci->i_filelock_ref) > 0)
1944 			goto out;
1945 	}
1946 	/* The inode has cached pages, but it's no longer used.
1947 	 * we can safely drop it */
1948 	if (S_ISREG(inode->i_mode) &&
1949 	    wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1950 	    !(oissued & CEPH_CAP_FILE_CACHE)) {
1951 	  used = 0;
1952 	  oissued = 0;
1953 	}
1954 	if ((used | wanted) & ~oissued & mine)
1955 		goto out;   /* we need these caps */
1956 
1957 	if (oissued) {
1958 		/* we aren't the only cap.. just remove us */
1959 		__ceph_remove_cap(cap, true);
1960 		(*remaining)--;
1961 	} else {
1962 		struct dentry *dentry;
1963 		/* try dropping referring dentries */
1964 		spin_unlock(&ci->i_ceph_lock);
1965 		dentry = d_find_any_alias(inode);
1966 		if (dentry && drop_negative_children(dentry)) {
1967 			int count;
1968 			dput(dentry);
1969 			d_prune_aliases(inode);
1970 			count = atomic_read(&inode->i_count);
1971 			if (count == 1)
1972 				(*remaining)--;
1973 			dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1974 			     inode, cap, count);
1975 		} else {
1976 			dput(dentry);
1977 		}
1978 		return 0;
1979 	}
1980 
1981 out:
1982 	spin_unlock(&ci->i_ceph_lock);
1983 	return 0;
1984 }
1985 
1986 /*
1987  * Trim session cap count down to some max number.
1988  */
1989 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1990 		   struct ceph_mds_session *session,
1991 		   int max_caps)
1992 {
1993 	int trim_caps = session->s_nr_caps - max_caps;
1994 
1995 	dout("trim_caps mds%d start: %d / %d, trim %d\n",
1996 	     session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1997 	if (trim_caps > 0) {
1998 		int remaining = trim_caps;
1999 
2000 		ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2001 		dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2002 		     session->s_mds, session->s_nr_caps, max_caps,
2003 			trim_caps - remaining);
2004 	}
2005 
2006 	ceph_flush_cap_releases(mdsc, session);
2007 	return 0;
2008 }
2009 
2010 static int check_caps_flush(struct ceph_mds_client *mdsc,
2011 			    u64 want_flush_tid)
2012 {
2013 	int ret = 1;
2014 
2015 	spin_lock(&mdsc->cap_dirty_lock);
2016 	if (!list_empty(&mdsc->cap_flush_list)) {
2017 		struct ceph_cap_flush *cf =
2018 			list_first_entry(&mdsc->cap_flush_list,
2019 					 struct ceph_cap_flush, g_list);
2020 		if (cf->tid <= want_flush_tid) {
2021 			dout("check_caps_flush still flushing tid "
2022 			     "%llu <= %llu\n", cf->tid, want_flush_tid);
2023 			ret = 0;
2024 		}
2025 	}
2026 	spin_unlock(&mdsc->cap_dirty_lock);
2027 	return ret;
2028 }
2029 
2030 /*
2031  * flush all dirty inode data to disk.
2032  *
2033  * returns true if we've flushed through want_flush_tid
2034  */
2035 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2036 			    u64 want_flush_tid)
2037 {
2038 	dout("check_caps_flush want %llu\n", want_flush_tid);
2039 
2040 	wait_event(mdsc->cap_flushing_wq,
2041 		   check_caps_flush(mdsc, want_flush_tid));
2042 
2043 	dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2044 }
2045 
2046 /*
2047  * called under s_mutex
2048  */
2049 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2050 				   struct ceph_mds_session *session)
2051 {
2052 	struct ceph_msg *msg = NULL;
2053 	struct ceph_mds_cap_release *head;
2054 	struct ceph_mds_cap_item *item;
2055 	struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2056 	struct ceph_cap *cap;
2057 	LIST_HEAD(tmp_list);
2058 	int num_cap_releases;
2059 	__le32	barrier, *cap_barrier;
2060 
2061 	down_read(&osdc->lock);
2062 	barrier = cpu_to_le32(osdc->epoch_barrier);
2063 	up_read(&osdc->lock);
2064 
2065 	spin_lock(&session->s_cap_lock);
2066 again:
2067 	list_splice_init(&session->s_cap_releases, &tmp_list);
2068 	num_cap_releases = session->s_num_cap_releases;
2069 	session->s_num_cap_releases = 0;
2070 	spin_unlock(&session->s_cap_lock);
2071 
2072 	while (!list_empty(&tmp_list)) {
2073 		if (!msg) {
2074 			msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2075 					PAGE_SIZE, GFP_NOFS, false);
2076 			if (!msg)
2077 				goto out_err;
2078 			head = msg->front.iov_base;
2079 			head->num = cpu_to_le32(0);
2080 			msg->front.iov_len = sizeof(*head);
2081 
2082 			msg->hdr.version = cpu_to_le16(2);
2083 			msg->hdr.compat_version = cpu_to_le16(1);
2084 		}
2085 
2086 		cap = list_first_entry(&tmp_list, struct ceph_cap,
2087 					session_caps);
2088 		list_del(&cap->session_caps);
2089 		num_cap_releases--;
2090 
2091 		head = msg->front.iov_base;
2092 		put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2093 				   &head->num);
2094 		item = msg->front.iov_base + msg->front.iov_len;
2095 		item->ino = cpu_to_le64(cap->cap_ino);
2096 		item->cap_id = cpu_to_le64(cap->cap_id);
2097 		item->migrate_seq = cpu_to_le32(cap->mseq);
2098 		item->seq = cpu_to_le32(cap->issue_seq);
2099 		msg->front.iov_len += sizeof(*item);
2100 
2101 		ceph_put_cap(mdsc, cap);
2102 
2103 		if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2104 			// Append cap_barrier field
2105 			cap_barrier = msg->front.iov_base + msg->front.iov_len;
2106 			*cap_barrier = barrier;
2107 			msg->front.iov_len += sizeof(*cap_barrier);
2108 
2109 			msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2110 			dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2111 			ceph_con_send(&session->s_con, msg);
2112 			msg = NULL;
2113 		}
2114 	}
2115 
2116 	BUG_ON(num_cap_releases != 0);
2117 
2118 	spin_lock(&session->s_cap_lock);
2119 	if (!list_empty(&session->s_cap_releases))
2120 		goto again;
2121 	spin_unlock(&session->s_cap_lock);
2122 
2123 	if (msg) {
2124 		// Append cap_barrier field
2125 		cap_barrier = msg->front.iov_base + msg->front.iov_len;
2126 		*cap_barrier = barrier;
2127 		msg->front.iov_len += sizeof(*cap_barrier);
2128 
2129 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2130 		dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2131 		ceph_con_send(&session->s_con, msg);
2132 	}
2133 	return;
2134 out_err:
2135 	pr_err("send_cap_releases mds%d, failed to allocate message\n",
2136 		session->s_mds);
2137 	spin_lock(&session->s_cap_lock);
2138 	list_splice(&tmp_list, &session->s_cap_releases);
2139 	session->s_num_cap_releases += num_cap_releases;
2140 	spin_unlock(&session->s_cap_lock);
2141 }
2142 
2143 static void ceph_cap_release_work(struct work_struct *work)
2144 {
2145 	struct ceph_mds_session *session =
2146 		container_of(work, struct ceph_mds_session, s_cap_release_work);
2147 
2148 	mutex_lock(&session->s_mutex);
2149 	if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2150 	    session->s_state == CEPH_MDS_SESSION_HUNG)
2151 		ceph_send_cap_releases(session->s_mdsc, session);
2152 	mutex_unlock(&session->s_mutex);
2153 	ceph_put_mds_session(session);
2154 }
2155 
2156 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2157 		             struct ceph_mds_session *session)
2158 {
2159 	if (mdsc->stopping)
2160 		return;
2161 
2162 	ceph_get_mds_session(session);
2163 	if (queue_work(mdsc->fsc->cap_wq,
2164 		       &session->s_cap_release_work)) {
2165 		dout("cap release work queued\n");
2166 	} else {
2167 		ceph_put_mds_session(session);
2168 		dout("failed to queue cap release work\n");
2169 	}
2170 }
2171 
2172 /*
2173  * caller holds session->s_cap_lock
2174  */
2175 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2176 			      struct ceph_cap *cap)
2177 {
2178 	list_add_tail(&cap->session_caps, &session->s_cap_releases);
2179 	session->s_num_cap_releases++;
2180 
2181 	if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2182 		ceph_flush_cap_releases(session->s_mdsc, session);
2183 }
2184 
2185 static void ceph_cap_reclaim_work(struct work_struct *work)
2186 {
2187 	struct ceph_mds_client *mdsc =
2188 		container_of(work, struct ceph_mds_client, cap_reclaim_work);
2189 	int ret = ceph_trim_dentries(mdsc);
2190 	if (ret == -EAGAIN)
2191 		ceph_queue_cap_reclaim_work(mdsc);
2192 }
2193 
2194 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2195 {
2196 	if (mdsc->stopping)
2197 		return;
2198 
2199         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2200                 dout("caps reclaim work queued\n");
2201         } else {
2202                 dout("failed to queue caps release work\n");
2203         }
2204 }
2205 
2206 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2207 {
2208 	int val;
2209 	if (!nr)
2210 		return;
2211 	val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2212 	if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2213 		atomic_set(&mdsc->cap_reclaim_pending, 0);
2214 		ceph_queue_cap_reclaim_work(mdsc);
2215 	}
2216 }
2217 
2218 /*
2219  * requests
2220  */
2221 
2222 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2223 				    struct inode *dir)
2224 {
2225 	struct ceph_inode_info *ci = ceph_inode(dir);
2226 	struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2227 	struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2228 	size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2229 	unsigned int num_entries;
2230 	int order;
2231 
2232 	spin_lock(&ci->i_ceph_lock);
2233 	num_entries = ci->i_files + ci->i_subdirs;
2234 	spin_unlock(&ci->i_ceph_lock);
2235 	num_entries = max(num_entries, 1U);
2236 	num_entries = min(num_entries, opt->max_readdir);
2237 
2238 	order = get_order(size * num_entries);
2239 	while (order >= 0) {
2240 		rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2241 							     __GFP_NOWARN,
2242 							     order);
2243 		if (rinfo->dir_entries)
2244 			break;
2245 		order--;
2246 	}
2247 	if (!rinfo->dir_entries)
2248 		return -ENOMEM;
2249 
2250 	num_entries = (PAGE_SIZE << order) / size;
2251 	num_entries = min(num_entries, opt->max_readdir);
2252 
2253 	rinfo->dir_buf_size = PAGE_SIZE << order;
2254 	req->r_num_caps = num_entries + 1;
2255 	req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2256 	req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2257 	return 0;
2258 }
2259 
2260 /*
2261  * Create an mds request.
2262  */
2263 struct ceph_mds_request *
2264 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2265 {
2266 	struct ceph_mds_request *req;
2267 
2268 	req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2269 	if (!req)
2270 		return ERR_PTR(-ENOMEM);
2271 
2272 	mutex_init(&req->r_fill_mutex);
2273 	req->r_mdsc = mdsc;
2274 	req->r_started = jiffies;
2275 	req->r_start_latency = ktime_get();
2276 	req->r_resend_mds = -1;
2277 	INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2278 	INIT_LIST_HEAD(&req->r_unsafe_target_item);
2279 	req->r_fmode = -1;
2280 	kref_init(&req->r_kref);
2281 	RB_CLEAR_NODE(&req->r_node);
2282 	INIT_LIST_HEAD(&req->r_wait);
2283 	init_completion(&req->r_completion);
2284 	init_completion(&req->r_safe_completion);
2285 	INIT_LIST_HEAD(&req->r_unsafe_item);
2286 
2287 	ktime_get_coarse_real_ts64(&req->r_stamp);
2288 
2289 	req->r_op = op;
2290 	req->r_direct_mode = mode;
2291 	return req;
2292 }
2293 
2294 /*
2295  * return oldest (lowest) request, tid in request tree, 0 if none.
2296  *
2297  * called under mdsc->mutex.
2298  */
2299 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2300 {
2301 	if (RB_EMPTY_ROOT(&mdsc->request_tree))
2302 		return NULL;
2303 	return rb_entry(rb_first(&mdsc->request_tree),
2304 			struct ceph_mds_request, r_node);
2305 }
2306 
2307 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2308 {
2309 	return mdsc->oldest_tid;
2310 }
2311 
2312 /*
2313  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2314  * on build_path_from_dentry in fs/cifs/dir.c.
2315  *
2316  * If @stop_on_nosnap, generate path relative to the first non-snapped
2317  * inode.
2318  *
2319  * Encode hidden .snap dirs as a double /, i.e.
2320  *   foo/.snap/bar -> foo//bar
2321  */
2322 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2323 			   int stop_on_nosnap)
2324 {
2325 	struct dentry *temp;
2326 	char *path;
2327 	int pos;
2328 	unsigned seq;
2329 	u64 base;
2330 
2331 	if (!dentry)
2332 		return ERR_PTR(-EINVAL);
2333 
2334 	path = __getname();
2335 	if (!path)
2336 		return ERR_PTR(-ENOMEM);
2337 retry:
2338 	pos = PATH_MAX - 1;
2339 	path[pos] = '\0';
2340 
2341 	seq = read_seqbegin(&rename_lock);
2342 	rcu_read_lock();
2343 	temp = dentry;
2344 	for (;;) {
2345 		struct inode *inode;
2346 
2347 		spin_lock(&temp->d_lock);
2348 		inode = d_inode(temp);
2349 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2350 			dout("build_path path+%d: %p SNAPDIR\n",
2351 			     pos, temp);
2352 		} else if (stop_on_nosnap && inode && dentry != temp &&
2353 			   ceph_snap(inode) == CEPH_NOSNAP) {
2354 			spin_unlock(&temp->d_lock);
2355 			pos++; /* get rid of any prepended '/' */
2356 			break;
2357 		} else {
2358 			pos -= temp->d_name.len;
2359 			if (pos < 0) {
2360 				spin_unlock(&temp->d_lock);
2361 				break;
2362 			}
2363 			memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2364 		}
2365 		spin_unlock(&temp->d_lock);
2366 		temp = READ_ONCE(temp->d_parent);
2367 
2368 		/* Are we at the root? */
2369 		if (IS_ROOT(temp))
2370 			break;
2371 
2372 		/* Are we out of buffer? */
2373 		if (--pos < 0)
2374 			break;
2375 
2376 		path[pos] = '/';
2377 	}
2378 	base = ceph_ino(d_inode(temp));
2379 	rcu_read_unlock();
2380 
2381 	if (read_seqretry(&rename_lock, seq))
2382 		goto retry;
2383 
2384 	if (pos < 0) {
2385 		/*
2386 		 * A rename didn't occur, but somehow we didn't end up where
2387 		 * we thought we would. Throw a warning and try again.
2388 		 */
2389 		pr_warn("build_path did not end path lookup where "
2390 			"expected, pos is %d\n", pos);
2391 		goto retry;
2392 	}
2393 
2394 	*pbase = base;
2395 	*plen = PATH_MAX - 1 - pos;
2396 	dout("build_path on %p %d built %llx '%.*s'\n",
2397 	     dentry, d_count(dentry), base, *plen, path + pos);
2398 	return path + pos;
2399 }
2400 
2401 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2402 			     const char **ppath, int *ppathlen, u64 *pino,
2403 			     bool *pfreepath, bool parent_locked)
2404 {
2405 	char *path;
2406 
2407 	rcu_read_lock();
2408 	if (!dir)
2409 		dir = d_inode_rcu(dentry->d_parent);
2410 	if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2411 		*pino = ceph_ino(dir);
2412 		rcu_read_unlock();
2413 		*ppath = dentry->d_name.name;
2414 		*ppathlen = dentry->d_name.len;
2415 		return 0;
2416 	}
2417 	rcu_read_unlock();
2418 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2419 	if (IS_ERR(path))
2420 		return PTR_ERR(path);
2421 	*ppath = path;
2422 	*pfreepath = true;
2423 	return 0;
2424 }
2425 
2426 static int build_inode_path(struct inode *inode,
2427 			    const char **ppath, int *ppathlen, u64 *pino,
2428 			    bool *pfreepath)
2429 {
2430 	struct dentry *dentry;
2431 	char *path;
2432 
2433 	if (ceph_snap(inode) == CEPH_NOSNAP) {
2434 		*pino = ceph_ino(inode);
2435 		*ppathlen = 0;
2436 		return 0;
2437 	}
2438 	dentry = d_find_alias(inode);
2439 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2440 	dput(dentry);
2441 	if (IS_ERR(path))
2442 		return PTR_ERR(path);
2443 	*ppath = path;
2444 	*pfreepath = true;
2445 	return 0;
2446 }
2447 
2448 /*
2449  * request arguments may be specified via an inode *, a dentry *, or
2450  * an explicit ino+path.
2451  */
2452 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2453 				  struct inode *rdiri, const char *rpath,
2454 				  u64 rino, const char **ppath, int *pathlen,
2455 				  u64 *ino, bool *freepath, bool parent_locked)
2456 {
2457 	int r = 0;
2458 
2459 	if (rinode) {
2460 		r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2461 		dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2462 		     ceph_snap(rinode));
2463 	} else if (rdentry) {
2464 		r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2465 					freepath, parent_locked);
2466 		dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2467 		     *ppath);
2468 	} else if (rpath || rino) {
2469 		*ino = rino;
2470 		*ppath = rpath;
2471 		*pathlen = rpath ? strlen(rpath) : 0;
2472 		dout(" path %.*s\n", *pathlen, rpath);
2473 	}
2474 
2475 	return r;
2476 }
2477 
2478 static void encode_timestamp_and_gids(void **p,
2479 				      const struct ceph_mds_request *req)
2480 {
2481 	struct ceph_timespec ts;
2482 	int i;
2483 
2484 	ceph_encode_timespec64(&ts, &req->r_stamp);
2485 	ceph_encode_copy(p, &ts, sizeof(ts));
2486 
2487 	/* gid_list */
2488 	ceph_encode_32(p, req->r_cred->group_info->ngroups);
2489 	for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2490 		ceph_encode_64(p, from_kgid(&init_user_ns,
2491 					    req->r_cred->group_info->gid[i]));
2492 }
2493 
2494 /*
2495  * called under mdsc->mutex
2496  */
2497 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2498 					       struct ceph_mds_request *req,
2499 					       bool drop_cap_releases)
2500 {
2501 	int mds = session->s_mds;
2502 	struct ceph_mds_client *mdsc = session->s_mdsc;
2503 	struct ceph_msg *msg;
2504 	struct ceph_mds_request_head_old *head;
2505 	const char *path1 = NULL;
2506 	const char *path2 = NULL;
2507 	u64 ino1 = 0, ino2 = 0;
2508 	int pathlen1 = 0, pathlen2 = 0;
2509 	bool freepath1 = false, freepath2 = false;
2510 	int len;
2511 	u16 releases;
2512 	void *p, *end;
2513 	int ret;
2514 	bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2515 
2516 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
2517 			      req->r_parent, req->r_path1, req->r_ino1.ino,
2518 			      &path1, &pathlen1, &ino1, &freepath1,
2519 			      test_bit(CEPH_MDS_R_PARENT_LOCKED,
2520 					&req->r_req_flags));
2521 	if (ret < 0) {
2522 		msg = ERR_PTR(ret);
2523 		goto out;
2524 	}
2525 
2526 	/* If r_old_dentry is set, then assume that its parent is locked */
2527 	ret = set_request_path_attr(NULL, req->r_old_dentry,
2528 			      req->r_old_dentry_dir,
2529 			      req->r_path2, req->r_ino2.ino,
2530 			      &path2, &pathlen2, &ino2, &freepath2, true);
2531 	if (ret < 0) {
2532 		msg = ERR_PTR(ret);
2533 		goto out_free1;
2534 	}
2535 
2536 	len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2537 	len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2538 		sizeof(struct ceph_timespec);
2539 	len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2540 
2541 	/* calculate (max) length for cap releases */
2542 	len += sizeof(struct ceph_mds_request_release) *
2543 		(!!req->r_inode_drop + !!req->r_dentry_drop +
2544 		 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2545 
2546 	if (req->r_dentry_drop)
2547 		len += pathlen1;
2548 	if (req->r_old_dentry_drop)
2549 		len += pathlen2;
2550 
2551 	msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2552 	if (!msg) {
2553 		msg = ERR_PTR(-ENOMEM);
2554 		goto out_free2;
2555 	}
2556 
2557 	msg->hdr.tid = cpu_to_le64(req->r_tid);
2558 
2559 	/*
2560 	 * The old ceph_mds_request_head didn't contain a version field, and
2561 	 * one was added when we moved the message version from 3->4.
2562 	 */
2563 	if (legacy) {
2564 		msg->hdr.version = cpu_to_le16(3);
2565 		head = msg->front.iov_base;
2566 		p = msg->front.iov_base + sizeof(*head);
2567 	} else {
2568 		struct ceph_mds_request_head *new_head = msg->front.iov_base;
2569 
2570 		msg->hdr.version = cpu_to_le16(4);
2571 		new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2572 		head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2573 		p = msg->front.iov_base + sizeof(*new_head);
2574 	}
2575 
2576 	end = msg->front.iov_base + msg->front.iov_len;
2577 
2578 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2579 	head->op = cpu_to_le32(req->r_op);
2580 	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2581 						 req->r_cred->fsuid));
2582 	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2583 						 req->r_cred->fsgid));
2584 	head->ino = cpu_to_le64(req->r_deleg_ino);
2585 	head->args = req->r_args;
2586 
2587 	ceph_encode_filepath(&p, end, ino1, path1);
2588 	ceph_encode_filepath(&p, end, ino2, path2);
2589 
2590 	/* make note of release offset, in case we need to replay */
2591 	req->r_request_release_offset = p - msg->front.iov_base;
2592 
2593 	/* cap releases */
2594 	releases = 0;
2595 	if (req->r_inode_drop)
2596 		releases += ceph_encode_inode_release(&p,
2597 		      req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2598 		      mds, req->r_inode_drop, req->r_inode_unless,
2599 		      req->r_op == CEPH_MDS_OP_READDIR);
2600 	if (req->r_dentry_drop)
2601 		releases += ceph_encode_dentry_release(&p, req->r_dentry,
2602 				req->r_parent, mds, req->r_dentry_drop,
2603 				req->r_dentry_unless);
2604 	if (req->r_old_dentry_drop)
2605 		releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2606 				req->r_old_dentry_dir, mds,
2607 				req->r_old_dentry_drop,
2608 				req->r_old_dentry_unless);
2609 	if (req->r_old_inode_drop)
2610 		releases += ceph_encode_inode_release(&p,
2611 		      d_inode(req->r_old_dentry),
2612 		      mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2613 
2614 	if (drop_cap_releases) {
2615 		releases = 0;
2616 		p = msg->front.iov_base + req->r_request_release_offset;
2617 	}
2618 
2619 	head->num_releases = cpu_to_le16(releases);
2620 
2621 	encode_timestamp_and_gids(&p, req);
2622 
2623 	if (WARN_ON_ONCE(p > end)) {
2624 		ceph_msg_put(msg);
2625 		msg = ERR_PTR(-ERANGE);
2626 		goto out_free2;
2627 	}
2628 
2629 	msg->front.iov_len = p - msg->front.iov_base;
2630 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2631 
2632 	if (req->r_pagelist) {
2633 		struct ceph_pagelist *pagelist = req->r_pagelist;
2634 		ceph_msg_data_add_pagelist(msg, pagelist);
2635 		msg->hdr.data_len = cpu_to_le32(pagelist->length);
2636 	} else {
2637 		msg->hdr.data_len = 0;
2638 	}
2639 
2640 	msg->hdr.data_off = cpu_to_le16(0);
2641 
2642 out_free2:
2643 	if (freepath2)
2644 		ceph_mdsc_free_path((char *)path2, pathlen2);
2645 out_free1:
2646 	if (freepath1)
2647 		ceph_mdsc_free_path((char *)path1, pathlen1);
2648 out:
2649 	return msg;
2650 }
2651 
2652 /*
2653  * called under mdsc->mutex if error, under no mutex if
2654  * success.
2655  */
2656 static void complete_request(struct ceph_mds_client *mdsc,
2657 			     struct ceph_mds_request *req)
2658 {
2659 	req->r_end_latency = ktime_get();
2660 
2661 	if (req->r_callback)
2662 		req->r_callback(mdsc, req);
2663 	complete_all(&req->r_completion);
2664 }
2665 
2666 static struct ceph_mds_request_head_old *
2667 find_old_request_head(void *p, u64 features)
2668 {
2669 	bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2670 	struct ceph_mds_request_head *new_head;
2671 
2672 	if (legacy)
2673 		return (struct ceph_mds_request_head_old *)p;
2674 	new_head = (struct ceph_mds_request_head *)p;
2675 	return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2676 }
2677 
2678 /*
2679  * called under mdsc->mutex
2680  */
2681 static int __prepare_send_request(struct ceph_mds_session *session,
2682 				  struct ceph_mds_request *req,
2683 				  bool drop_cap_releases)
2684 {
2685 	int mds = session->s_mds;
2686 	struct ceph_mds_client *mdsc = session->s_mdsc;
2687 	struct ceph_mds_request_head_old *rhead;
2688 	struct ceph_msg *msg;
2689 	int flags = 0;
2690 
2691 	req->r_attempts++;
2692 	if (req->r_inode) {
2693 		struct ceph_cap *cap =
2694 			ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2695 
2696 		if (cap)
2697 			req->r_sent_on_mseq = cap->mseq;
2698 		else
2699 			req->r_sent_on_mseq = -1;
2700 	}
2701 	dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2702 	     req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2703 
2704 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2705 		void *p;
2706 
2707 		/*
2708 		 * Replay.  Do not regenerate message (and rebuild
2709 		 * paths, etc.); just use the original message.
2710 		 * Rebuilding paths will break for renames because
2711 		 * d_move mangles the src name.
2712 		 */
2713 		msg = req->r_request;
2714 		rhead = find_old_request_head(msg->front.iov_base,
2715 					      session->s_con.peer_features);
2716 
2717 		flags = le32_to_cpu(rhead->flags);
2718 		flags |= CEPH_MDS_FLAG_REPLAY;
2719 		rhead->flags = cpu_to_le32(flags);
2720 
2721 		if (req->r_target_inode)
2722 			rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2723 
2724 		rhead->num_retry = req->r_attempts - 1;
2725 
2726 		/* remove cap/dentry releases from message */
2727 		rhead->num_releases = 0;
2728 
2729 		p = msg->front.iov_base + req->r_request_release_offset;
2730 		encode_timestamp_and_gids(&p, req);
2731 
2732 		msg->front.iov_len = p - msg->front.iov_base;
2733 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2734 		return 0;
2735 	}
2736 
2737 	if (req->r_request) {
2738 		ceph_msg_put(req->r_request);
2739 		req->r_request = NULL;
2740 	}
2741 	msg = create_request_message(session, req, drop_cap_releases);
2742 	if (IS_ERR(msg)) {
2743 		req->r_err = PTR_ERR(msg);
2744 		return PTR_ERR(msg);
2745 	}
2746 	req->r_request = msg;
2747 
2748 	rhead = find_old_request_head(msg->front.iov_base,
2749 				      session->s_con.peer_features);
2750 	rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2751 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2752 		flags |= CEPH_MDS_FLAG_REPLAY;
2753 	if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2754 		flags |= CEPH_MDS_FLAG_ASYNC;
2755 	if (req->r_parent)
2756 		flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2757 	rhead->flags = cpu_to_le32(flags);
2758 	rhead->num_fwd = req->r_num_fwd;
2759 	rhead->num_retry = req->r_attempts - 1;
2760 
2761 	dout(" r_parent = %p\n", req->r_parent);
2762 	return 0;
2763 }
2764 
2765 /*
2766  * called under mdsc->mutex
2767  */
2768 static int __send_request(struct ceph_mds_session *session,
2769 			  struct ceph_mds_request *req,
2770 			  bool drop_cap_releases)
2771 {
2772 	int err;
2773 
2774 	err = __prepare_send_request(session, req, drop_cap_releases);
2775 	if (!err) {
2776 		ceph_msg_get(req->r_request);
2777 		ceph_con_send(&session->s_con, req->r_request);
2778 	}
2779 
2780 	return err;
2781 }
2782 
2783 /*
2784  * send request, or put it on the appropriate wait list.
2785  */
2786 static void __do_request(struct ceph_mds_client *mdsc,
2787 			struct ceph_mds_request *req)
2788 {
2789 	struct ceph_mds_session *session = NULL;
2790 	int mds = -1;
2791 	int err = 0;
2792 	bool random;
2793 
2794 	if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2795 		if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2796 			__unregister_request(mdsc, req);
2797 		return;
2798 	}
2799 
2800 	if (req->r_timeout &&
2801 	    time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2802 		dout("do_request timed out\n");
2803 		err = -ETIMEDOUT;
2804 		goto finish;
2805 	}
2806 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2807 		dout("do_request forced umount\n");
2808 		err = -EIO;
2809 		goto finish;
2810 	}
2811 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2812 		if (mdsc->mdsmap_err) {
2813 			err = mdsc->mdsmap_err;
2814 			dout("do_request mdsmap err %d\n", err);
2815 			goto finish;
2816 		}
2817 		if (mdsc->mdsmap->m_epoch == 0) {
2818 			dout("do_request no mdsmap, waiting for map\n");
2819 			list_add(&req->r_wait, &mdsc->waiting_for_map);
2820 			return;
2821 		}
2822 		if (!(mdsc->fsc->mount_options->flags &
2823 		      CEPH_MOUNT_OPT_MOUNTWAIT) &&
2824 		    !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2825 			err = -EHOSTUNREACH;
2826 			goto finish;
2827 		}
2828 	}
2829 
2830 	put_request_session(req);
2831 
2832 	mds = __choose_mds(mdsc, req, &random);
2833 	if (mds < 0 ||
2834 	    ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2835 		if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2836 			err = -EJUKEBOX;
2837 			goto finish;
2838 		}
2839 		dout("do_request no mds or not active, waiting for map\n");
2840 		list_add(&req->r_wait, &mdsc->waiting_for_map);
2841 		return;
2842 	}
2843 
2844 	/* get, open session */
2845 	session = __ceph_lookup_mds_session(mdsc, mds);
2846 	if (!session) {
2847 		session = register_session(mdsc, mds);
2848 		if (IS_ERR(session)) {
2849 			err = PTR_ERR(session);
2850 			goto finish;
2851 		}
2852 	}
2853 	req->r_session = ceph_get_mds_session(session);
2854 
2855 	dout("do_request mds%d session %p state %s\n", mds, session,
2856 	     ceph_session_state_name(session->s_state));
2857 	if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2858 	    session->s_state != CEPH_MDS_SESSION_HUNG) {
2859 		/*
2860 		 * We cannot queue async requests since the caps and delegated
2861 		 * inodes are bound to the session. Just return -EJUKEBOX and
2862 		 * let the caller retry a sync request in that case.
2863 		 */
2864 		if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2865 			err = -EJUKEBOX;
2866 			goto out_session;
2867 		}
2868 
2869 		/*
2870 		 * If the session has been REJECTED, then return a hard error,
2871 		 * unless it's a CLEANRECOVER mount, in which case we'll queue
2872 		 * it to the mdsc queue.
2873 		 */
2874 		if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2875 			if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2876 				list_add(&req->r_wait, &mdsc->waiting_for_map);
2877 			else
2878 				err = -EACCES;
2879 			goto out_session;
2880 		}
2881 
2882 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
2883 		    session->s_state == CEPH_MDS_SESSION_CLOSING) {
2884 			err = __open_session(mdsc, session);
2885 			if (err)
2886 				goto out_session;
2887 			/* retry the same mds later */
2888 			if (random)
2889 				req->r_resend_mds = mds;
2890 		}
2891 		list_add(&req->r_wait, &session->s_waiting);
2892 		goto out_session;
2893 	}
2894 
2895 	/* send request */
2896 	req->r_resend_mds = -1;   /* forget any previous mds hint */
2897 
2898 	if (req->r_request_started == 0)   /* note request start time */
2899 		req->r_request_started = jiffies;
2900 
2901 	err = __send_request(session, req, false);
2902 
2903 out_session:
2904 	ceph_put_mds_session(session);
2905 finish:
2906 	if (err) {
2907 		dout("__do_request early error %d\n", err);
2908 		req->r_err = err;
2909 		complete_request(mdsc, req);
2910 		__unregister_request(mdsc, req);
2911 	}
2912 	return;
2913 }
2914 
2915 /*
2916  * called under mdsc->mutex
2917  */
2918 static void __wake_requests(struct ceph_mds_client *mdsc,
2919 			    struct list_head *head)
2920 {
2921 	struct ceph_mds_request *req;
2922 	LIST_HEAD(tmp_list);
2923 
2924 	list_splice_init(head, &tmp_list);
2925 
2926 	while (!list_empty(&tmp_list)) {
2927 		req = list_entry(tmp_list.next,
2928 				 struct ceph_mds_request, r_wait);
2929 		list_del_init(&req->r_wait);
2930 		dout(" wake request %p tid %llu\n", req, req->r_tid);
2931 		__do_request(mdsc, req);
2932 	}
2933 }
2934 
2935 /*
2936  * Wake up threads with requests pending for @mds, so that they can
2937  * resubmit their requests to a possibly different mds.
2938  */
2939 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2940 {
2941 	struct ceph_mds_request *req;
2942 	struct rb_node *p = rb_first(&mdsc->request_tree);
2943 
2944 	dout("kick_requests mds%d\n", mds);
2945 	while (p) {
2946 		req = rb_entry(p, struct ceph_mds_request, r_node);
2947 		p = rb_next(p);
2948 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2949 			continue;
2950 		if (req->r_attempts > 0)
2951 			continue; /* only new requests */
2952 		if (req->r_session &&
2953 		    req->r_session->s_mds == mds) {
2954 			dout(" kicking tid %llu\n", req->r_tid);
2955 			list_del_init(&req->r_wait);
2956 			__do_request(mdsc, req);
2957 		}
2958 	}
2959 }
2960 
2961 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2962 			      struct ceph_mds_request *req)
2963 {
2964 	int err = 0;
2965 
2966 	/* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2967 	if (req->r_inode)
2968 		ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2969 	if (req->r_parent) {
2970 		struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2971 		int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2972 			    CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2973 		spin_lock(&ci->i_ceph_lock);
2974 		ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2975 		__ceph_touch_fmode(ci, mdsc, fmode);
2976 		spin_unlock(&ci->i_ceph_lock);
2977 		ihold(req->r_parent);
2978 	}
2979 	if (req->r_old_dentry_dir)
2980 		ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2981 				  CEPH_CAP_PIN);
2982 
2983 	if (req->r_inode) {
2984 		err = ceph_wait_on_async_create(req->r_inode);
2985 		if (err) {
2986 			dout("%s: wait for async create returned: %d\n",
2987 			     __func__, err);
2988 			return err;
2989 		}
2990 	}
2991 
2992 	if (!err && req->r_old_inode) {
2993 		err = ceph_wait_on_async_create(req->r_old_inode);
2994 		if (err) {
2995 			dout("%s: wait for async create returned: %d\n",
2996 			     __func__, err);
2997 			return err;
2998 		}
2999 	}
3000 
3001 	dout("submit_request on %p for inode %p\n", req, dir);
3002 	mutex_lock(&mdsc->mutex);
3003 	__register_request(mdsc, req, dir);
3004 	__do_request(mdsc, req);
3005 	err = req->r_err;
3006 	mutex_unlock(&mdsc->mutex);
3007 	return err;
3008 }
3009 
3010 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3011 				  struct ceph_mds_request *req)
3012 {
3013 	int err;
3014 
3015 	/* wait */
3016 	dout("do_request waiting\n");
3017 	if (!req->r_timeout && req->r_wait_for_completion) {
3018 		err = req->r_wait_for_completion(mdsc, req);
3019 	} else {
3020 		long timeleft = wait_for_completion_killable_timeout(
3021 					&req->r_completion,
3022 					ceph_timeout_jiffies(req->r_timeout));
3023 		if (timeleft > 0)
3024 			err = 0;
3025 		else if (!timeleft)
3026 			err = -ETIMEDOUT;  /* timed out */
3027 		else
3028 			err = timeleft;  /* killed */
3029 	}
3030 	dout("do_request waited, got %d\n", err);
3031 	mutex_lock(&mdsc->mutex);
3032 
3033 	/* only abort if we didn't race with a real reply */
3034 	if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3035 		err = le32_to_cpu(req->r_reply_info.head->result);
3036 	} else if (err < 0) {
3037 		dout("aborted request %lld with %d\n", req->r_tid, err);
3038 
3039 		/*
3040 		 * ensure we aren't running concurrently with
3041 		 * ceph_fill_trace or ceph_readdir_prepopulate, which
3042 		 * rely on locks (dir mutex) held by our caller.
3043 		 */
3044 		mutex_lock(&req->r_fill_mutex);
3045 		req->r_err = err;
3046 		set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3047 		mutex_unlock(&req->r_fill_mutex);
3048 
3049 		if (req->r_parent &&
3050 		    (req->r_op & CEPH_MDS_OP_WRITE))
3051 			ceph_invalidate_dir_request(req);
3052 	} else {
3053 		err = req->r_err;
3054 	}
3055 
3056 	mutex_unlock(&mdsc->mutex);
3057 	return err;
3058 }
3059 
3060 /*
3061  * Synchrously perform an mds request.  Take care of all of the
3062  * session setup, forwarding, retry details.
3063  */
3064 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3065 			 struct inode *dir,
3066 			 struct ceph_mds_request *req)
3067 {
3068 	int err;
3069 
3070 	dout("do_request on %p\n", req);
3071 
3072 	/* issue */
3073 	err = ceph_mdsc_submit_request(mdsc, dir, req);
3074 	if (!err)
3075 		err = ceph_mdsc_wait_request(mdsc, req);
3076 	dout("do_request %p done, result %d\n", req, err);
3077 	return err;
3078 }
3079 
3080 /*
3081  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3082  * namespace request.
3083  */
3084 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3085 {
3086 	struct inode *dir = req->r_parent;
3087 	struct inode *old_dir = req->r_old_dentry_dir;
3088 
3089 	dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3090 
3091 	ceph_dir_clear_complete(dir);
3092 	if (old_dir)
3093 		ceph_dir_clear_complete(old_dir);
3094 	if (req->r_dentry)
3095 		ceph_invalidate_dentry_lease(req->r_dentry);
3096 	if (req->r_old_dentry)
3097 		ceph_invalidate_dentry_lease(req->r_old_dentry);
3098 }
3099 
3100 /*
3101  * Handle mds reply.
3102  *
3103  * We take the session mutex and parse and process the reply immediately.
3104  * This preserves the logical ordering of replies, capabilities, etc., sent
3105  * by the MDS as they are applied to our local cache.
3106  */
3107 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3108 {
3109 	struct ceph_mds_client *mdsc = session->s_mdsc;
3110 	struct ceph_mds_request *req;
3111 	struct ceph_mds_reply_head *head = msg->front.iov_base;
3112 	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3113 	struct ceph_snap_realm *realm;
3114 	u64 tid;
3115 	int err, result;
3116 	int mds = session->s_mds;
3117 
3118 	if (msg->front.iov_len < sizeof(*head)) {
3119 		pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3120 		ceph_msg_dump(msg);
3121 		return;
3122 	}
3123 
3124 	/* get request, session */
3125 	tid = le64_to_cpu(msg->hdr.tid);
3126 	mutex_lock(&mdsc->mutex);
3127 	req = lookup_get_request(mdsc, tid);
3128 	if (!req) {
3129 		dout("handle_reply on unknown tid %llu\n", tid);
3130 		mutex_unlock(&mdsc->mutex);
3131 		return;
3132 	}
3133 	dout("handle_reply %p\n", req);
3134 
3135 	/* correct session? */
3136 	if (req->r_session != session) {
3137 		pr_err("mdsc_handle_reply got %llu on session mds%d"
3138 		       " not mds%d\n", tid, session->s_mds,
3139 		       req->r_session ? req->r_session->s_mds : -1);
3140 		mutex_unlock(&mdsc->mutex);
3141 		goto out;
3142 	}
3143 
3144 	/* dup? */
3145 	if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3146 	    (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3147 		pr_warn("got a dup %s reply on %llu from mds%d\n",
3148 			   head->safe ? "safe" : "unsafe", tid, mds);
3149 		mutex_unlock(&mdsc->mutex);
3150 		goto out;
3151 	}
3152 	if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3153 		pr_warn("got unsafe after safe on %llu from mds%d\n",
3154 			   tid, mds);
3155 		mutex_unlock(&mdsc->mutex);
3156 		goto out;
3157 	}
3158 
3159 	result = le32_to_cpu(head->result);
3160 
3161 	/*
3162 	 * Handle an ESTALE
3163 	 * if we're not talking to the authority, send to them
3164 	 * if the authority has changed while we weren't looking,
3165 	 * send to new authority
3166 	 * Otherwise we just have to return an ESTALE
3167 	 */
3168 	if (result == -ESTALE) {
3169 		dout("got ESTALE on request %llu\n", req->r_tid);
3170 		req->r_resend_mds = -1;
3171 		if (req->r_direct_mode != USE_AUTH_MDS) {
3172 			dout("not using auth, setting for that now\n");
3173 			req->r_direct_mode = USE_AUTH_MDS;
3174 			__do_request(mdsc, req);
3175 			mutex_unlock(&mdsc->mutex);
3176 			goto out;
3177 		} else  {
3178 			int mds = __choose_mds(mdsc, req, NULL);
3179 			if (mds >= 0 && mds != req->r_session->s_mds) {
3180 				dout("but auth changed, so resending\n");
3181 				__do_request(mdsc, req);
3182 				mutex_unlock(&mdsc->mutex);
3183 				goto out;
3184 			}
3185 		}
3186 		dout("have to return ESTALE on request %llu\n", req->r_tid);
3187 	}
3188 
3189 
3190 	if (head->safe) {
3191 		set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3192 		__unregister_request(mdsc, req);
3193 
3194 		/* last request during umount? */
3195 		if (mdsc->stopping && !__get_oldest_req(mdsc))
3196 			complete_all(&mdsc->safe_umount_waiters);
3197 
3198 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3199 			/*
3200 			 * We already handled the unsafe response, now do the
3201 			 * cleanup.  No need to examine the response; the MDS
3202 			 * doesn't include any result info in the safe
3203 			 * response.  And even if it did, there is nothing
3204 			 * useful we could do with a revised return value.
3205 			 */
3206 			dout("got safe reply %llu, mds%d\n", tid, mds);
3207 
3208 			mutex_unlock(&mdsc->mutex);
3209 			goto out;
3210 		}
3211 	} else {
3212 		set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3213 		list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3214 	}
3215 
3216 	dout("handle_reply tid %lld result %d\n", tid, result);
3217 	rinfo = &req->r_reply_info;
3218 	if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3219 		err = parse_reply_info(session, msg, rinfo, (u64)-1);
3220 	else
3221 		err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3222 	mutex_unlock(&mdsc->mutex);
3223 
3224 	/* Must find target inode outside of mutexes to avoid deadlocks */
3225 	if ((err >= 0) && rinfo->head->is_target) {
3226 		struct inode *in;
3227 		struct ceph_vino tvino = {
3228 			.ino  = le64_to_cpu(rinfo->targeti.in->ino),
3229 			.snap = le64_to_cpu(rinfo->targeti.in->snapid)
3230 		};
3231 
3232 		in = ceph_get_inode(mdsc->fsc->sb, tvino);
3233 		if (IS_ERR(in)) {
3234 			err = PTR_ERR(in);
3235 			mutex_lock(&session->s_mutex);
3236 			goto out_err;
3237 		}
3238 		req->r_target_inode = in;
3239 	}
3240 
3241 	mutex_lock(&session->s_mutex);
3242 	if (err < 0) {
3243 		pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3244 		ceph_msg_dump(msg);
3245 		goto out_err;
3246 	}
3247 
3248 	/* snap trace */
3249 	realm = NULL;
3250 	if (rinfo->snapblob_len) {
3251 		down_write(&mdsc->snap_rwsem);
3252 		ceph_update_snap_trace(mdsc, rinfo->snapblob,
3253 				rinfo->snapblob + rinfo->snapblob_len,
3254 				le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3255 				&realm);
3256 		downgrade_write(&mdsc->snap_rwsem);
3257 	} else {
3258 		down_read(&mdsc->snap_rwsem);
3259 	}
3260 
3261 	/* insert trace into our cache */
3262 	mutex_lock(&req->r_fill_mutex);
3263 	current->journal_info = req;
3264 	err = ceph_fill_trace(mdsc->fsc->sb, req);
3265 	if (err == 0) {
3266 		if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3267 				    req->r_op == CEPH_MDS_OP_LSSNAP))
3268 			ceph_readdir_prepopulate(req, req->r_session);
3269 	}
3270 	current->journal_info = NULL;
3271 	mutex_unlock(&req->r_fill_mutex);
3272 
3273 	up_read(&mdsc->snap_rwsem);
3274 	if (realm)
3275 		ceph_put_snap_realm(mdsc, realm);
3276 
3277 	if (err == 0) {
3278 		if (req->r_target_inode &&
3279 		    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3280 			struct ceph_inode_info *ci =
3281 				ceph_inode(req->r_target_inode);
3282 			spin_lock(&ci->i_unsafe_lock);
3283 			list_add_tail(&req->r_unsafe_target_item,
3284 				      &ci->i_unsafe_iops);
3285 			spin_unlock(&ci->i_unsafe_lock);
3286 		}
3287 
3288 		ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3289 	}
3290 out_err:
3291 	mutex_lock(&mdsc->mutex);
3292 	if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3293 		if (err) {
3294 			req->r_err = err;
3295 		} else {
3296 			req->r_reply =  ceph_msg_get(msg);
3297 			set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3298 		}
3299 	} else {
3300 		dout("reply arrived after request %lld was aborted\n", tid);
3301 	}
3302 	mutex_unlock(&mdsc->mutex);
3303 
3304 	mutex_unlock(&session->s_mutex);
3305 
3306 	/* kick calling process */
3307 	complete_request(mdsc, req);
3308 
3309 	ceph_update_metadata_latency(&mdsc->metric, req->r_start_latency,
3310 				     req->r_end_latency, err);
3311 out:
3312 	ceph_mdsc_put_request(req);
3313 	return;
3314 }
3315 
3316 
3317 
3318 /*
3319  * handle mds notification that our request has been forwarded.
3320  */
3321 static void handle_forward(struct ceph_mds_client *mdsc,
3322 			   struct ceph_mds_session *session,
3323 			   struct ceph_msg *msg)
3324 {
3325 	struct ceph_mds_request *req;
3326 	u64 tid = le64_to_cpu(msg->hdr.tid);
3327 	u32 next_mds;
3328 	u32 fwd_seq;
3329 	int err = -EINVAL;
3330 	void *p = msg->front.iov_base;
3331 	void *end = p + msg->front.iov_len;
3332 
3333 	ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3334 	next_mds = ceph_decode_32(&p);
3335 	fwd_seq = ceph_decode_32(&p);
3336 
3337 	mutex_lock(&mdsc->mutex);
3338 	req = lookup_get_request(mdsc, tid);
3339 	if (!req) {
3340 		dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3341 		goto out;  /* dup reply? */
3342 	}
3343 
3344 	if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3345 		dout("forward tid %llu aborted, unregistering\n", tid);
3346 		__unregister_request(mdsc, req);
3347 	} else if (fwd_seq <= req->r_num_fwd) {
3348 		dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3349 		     tid, next_mds, req->r_num_fwd, fwd_seq);
3350 	} else {
3351 		/* resend. forward race not possible; mds would drop */
3352 		dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3353 		BUG_ON(req->r_err);
3354 		BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3355 		req->r_attempts = 0;
3356 		req->r_num_fwd = fwd_seq;
3357 		req->r_resend_mds = next_mds;
3358 		put_request_session(req);
3359 		__do_request(mdsc, req);
3360 	}
3361 	ceph_mdsc_put_request(req);
3362 out:
3363 	mutex_unlock(&mdsc->mutex);
3364 	return;
3365 
3366 bad:
3367 	pr_err("mdsc_handle_forward decode error err=%d\n", err);
3368 }
3369 
3370 static int __decode_session_metadata(void **p, void *end,
3371 				     bool *blocklisted)
3372 {
3373 	/* map<string,string> */
3374 	u32 n;
3375 	bool err_str;
3376 	ceph_decode_32_safe(p, end, n, bad);
3377 	while (n-- > 0) {
3378 		u32 len;
3379 		ceph_decode_32_safe(p, end, len, bad);
3380 		ceph_decode_need(p, end, len, bad);
3381 		err_str = !strncmp(*p, "error_string", len);
3382 		*p += len;
3383 		ceph_decode_32_safe(p, end, len, bad);
3384 		ceph_decode_need(p, end, len, bad);
3385 		/*
3386 		 * Match "blocklisted (blacklisted)" from newer MDSes,
3387 		 * or "blacklisted" from older MDSes.
3388 		 */
3389 		if (err_str && strnstr(*p, "blacklisted", len))
3390 			*blocklisted = true;
3391 		*p += len;
3392 	}
3393 	return 0;
3394 bad:
3395 	return -1;
3396 }
3397 
3398 /*
3399  * handle a mds session control message
3400  */
3401 static void handle_session(struct ceph_mds_session *session,
3402 			   struct ceph_msg *msg)
3403 {
3404 	struct ceph_mds_client *mdsc = session->s_mdsc;
3405 	int mds = session->s_mds;
3406 	int msg_version = le16_to_cpu(msg->hdr.version);
3407 	void *p = msg->front.iov_base;
3408 	void *end = p + msg->front.iov_len;
3409 	struct ceph_mds_session_head *h;
3410 	u32 op;
3411 	u64 seq, features = 0;
3412 	int wake = 0;
3413 	bool blocklisted = false;
3414 
3415 	/* decode */
3416 	ceph_decode_need(&p, end, sizeof(*h), bad);
3417 	h = p;
3418 	p += sizeof(*h);
3419 
3420 	op = le32_to_cpu(h->op);
3421 	seq = le64_to_cpu(h->seq);
3422 
3423 	if (msg_version >= 3) {
3424 		u32 len;
3425 		/* version >= 2, metadata */
3426 		if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3427 			goto bad;
3428 		/* version >= 3, feature bits */
3429 		ceph_decode_32_safe(&p, end, len, bad);
3430 		if (len) {
3431 			ceph_decode_64_safe(&p, end, features, bad);
3432 			p += len - sizeof(features);
3433 		}
3434 	}
3435 
3436 	mutex_lock(&mdsc->mutex);
3437 	if (op == CEPH_SESSION_CLOSE) {
3438 		ceph_get_mds_session(session);
3439 		__unregister_session(mdsc, session);
3440 	}
3441 	/* FIXME: this ttl calculation is generous */
3442 	session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3443 	mutex_unlock(&mdsc->mutex);
3444 
3445 	mutex_lock(&session->s_mutex);
3446 
3447 	dout("handle_session mds%d %s %p state %s seq %llu\n",
3448 	     mds, ceph_session_op_name(op), session,
3449 	     ceph_session_state_name(session->s_state), seq);
3450 
3451 	if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3452 		session->s_state = CEPH_MDS_SESSION_OPEN;
3453 		pr_info("mds%d came back\n", session->s_mds);
3454 	}
3455 
3456 	switch (op) {
3457 	case CEPH_SESSION_OPEN:
3458 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3459 			pr_info("mds%d reconnect success\n", session->s_mds);
3460 		session->s_state = CEPH_MDS_SESSION_OPEN;
3461 		session->s_features = features;
3462 		renewed_caps(mdsc, session, 0);
3463 		if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3464 			metric_schedule_delayed(&mdsc->metric);
3465 		wake = 1;
3466 		if (mdsc->stopping)
3467 			__close_session(mdsc, session);
3468 		break;
3469 
3470 	case CEPH_SESSION_RENEWCAPS:
3471 		if (session->s_renew_seq == seq)
3472 			renewed_caps(mdsc, session, 1);
3473 		break;
3474 
3475 	case CEPH_SESSION_CLOSE:
3476 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3477 			pr_info("mds%d reconnect denied\n", session->s_mds);
3478 		session->s_state = CEPH_MDS_SESSION_CLOSED;
3479 		cleanup_session_requests(mdsc, session);
3480 		remove_session_caps(session);
3481 		wake = 2; /* for good measure */
3482 		wake_up_all(&mdsc->session_close_wq);
3483 		break;
3484 
3485 	case CEPH_SESSION_STALE:
3486 		pr_info("mds%d caps went stale, renewing\n",
3487 			session->s_mds);
3488 		spin_lock(&session->s_gen_ttl_lock);
3489 		session->s_cap_gen++;
3490 		session->s_cap_ttl = jiffies - 1;
3491 		spin_unlock(&session->s_gen_ttl_lock);
3492 		send_renew_caps(mdsc, session);
3493 		break;
3494 
3495 	case CEPH_SESSION_RECALL_STATE:
3496 		ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3497 		break;
3498 
3499 	case CEPH_SESSION_FLUSHMSG:
3500 		send_flushmsg_ack(mdsc, session, seq);
3501 		break;
3502 
3503 	case CEPH_SESSION_FORCE_RO:
3504 		dout("force_session_readonly %p\n", session);
3505 		spin_lock(&session->s_cap_lock);
3506 		session->s_readonly = true;
3507 		spin_unlock(&session->s_cap_lock);
3508 		wake_up_session_caps(session, FORCE_RO);
3509 		break;
3510 
3511 	case CEPH_SESSION_REJECT:
3512 		WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3513 		pr_info("mds%d rejected session\n", session->s_mds);
3514 		session->s_state = CEPH_MDS_SESSION_REJECTED;
3515 		cleanup_session_requests(mdsc, session);
3516 		remove_session_caps(session);
3517 		if (blocklisted)
3518 			mdsc->fsc->blocklisted = true;
3519 		wake = 2; /* for good measure */
3520 		break;
3521 
3522 	default:
3523 		pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3524 		WARN_ON(1);
3525 	}
3526 
3527 	mutex_unlock(&session->s_mutex);
3528 	if (wake) {
3529 		mutex_lock(&mdsc->mutex);
3530 		__wake_requests(mdsc, &session->s_waiting);
3531 		if (wake == 2)
3532 			kick_requests(mdsc, mds);
3533 		mutex_unlock(&mdsc->mutex);
3534 	}
3535 	if (op == CEPH_SESSION_CLOSE)
3536 		ceph_put_mds_session(session);
3537 	return;
3538 
3539 bad:
3540 	pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3541 	       (int)msg->front.iov_len);
3542 	ceph_msg_dump(msg);
3543 	return;
3544 }
3545 
3546 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3547 {
3548 	int dcaps;
3549 
3550 	dcaps = xchg(&req->r_dir_caps, 0);
3551 	if (dcaps) {
3552 		dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3553 		ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3554 	}
3555 }
3556 
3557 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3558 {
3559 	int dcaps;
3560 
3561 	dcaps = xchg(&req->r_dir_caps, 0);
3562 	if (dcaps) {
3563 		dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3564 		ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3565 						dcaps);
3566 	}
3567 }
3568 
3569 /*
3570  * called under session->mutex.
3571  */
3572 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3573 				   struct ceph_mds_session *session)
3574 {
3575 	struct ceph_mds_request *req, *nreq;
3576 	struct rb_node *p;
3577 
3578 	dout("replay_unsafe_requests mds%d\n", session->s_mds);
3579 
3580 	mutex_lock(&mdsc->mutex);
3581 	list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3582 		__send_request(session, req, true);
3583 
3584 	/*
3585 	 * also re-send old requests when MDS enters reconnect stage. So that MDS
3586 	 * can process completed request in clientreplay stage.
3587 	 */
3588 	p = rb_first(&mdsc->request_tree);
3589 	while (p) {
3590 		req = rb_entry(p, struct ceph_mds_request, r_node);
3591 		p = rb_next(p);
3592 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3593 			continue;
3594 		if (req->r_attempts == 0)
3595 			continue; /* only old requests */
3596 		if (!req->r_session)
3597 			continue;
3598 		if (req->r_session->s_mds != session->s_mds)
3599 			continue;
3600 
3601 		ceph_mdsc_release_dir_caps_no_check(req);
3602 
3603 		__send_request(session, req, true);
3604 	}
3605 	mutex_unlock(&mdsc->mutex);
3606 }
3607 
3608 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3609 {
3610 	struct ceph_msg *reply;
3611 	struct ceph_pagelist *_pagelist;
3612 	struct page *page;
3613 	__le32 *addr;
3614 	int err = -ENOMEM;
3615 
3616 	if (!recon_state->allow_multi)
3617 		return -ENOSPC;
3618 
3619 	/* can't handle message that contains both caps and realm */
3620 	BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3621 
3622 	/* pre-allocate new pagelist */
3623 	_pagelist = ceph_pagelist_alloc(GFP_NOFS);
3624 	if (!_pagelist)
3625 		return -ENOMEM;
3626 
3627 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3628 	if (!reply)
3629 		goto fail_msg;
3630 
3631 	/* placeholder for nr_caps */
3632 	err = ceph_pagelist_encode_32(_pagelist, 0);
3633 	if (err < 0)
3634 		goto fail;
3635 
3636 	if (recon_state->nr_caps) {
3637 		/* currently encoding caps */
3638 		err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3639 		if (err)
3640 			goto fail;
3641 	} else {
3642 		/* placeholder for nr_realms (currently encoding relams) */
3643 		err = ceph_pagelist_encode_32(_pagelist, 0);
3644 		if (err < 0)
3645 			goto fail;
3646 	}
3647 
3648 	err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3649 	if (err)
3650 		goto fail;
3651 
3652 	page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3653 	addr = kmap_atomic(page);
3654 	if (recon_state->nr_caps) {
3655 		/* currently encoding caps */
3656 		*addr = cpu_to_le32(recon_state->nr_caps);
3657 	} else {
3658 		/* currently encoding relams */
3659 		*(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3660 	}
3661 	kunmap_atomic(addr);
3662 
3663 	reply->hdr.version = cpu_to_le16(5);
3664 	reply->hdr.compat_version = cpu_to_le16(4);
3665 
3666 	reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3667 	ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3668 
3669 	ceph_con_send(&recon_state->session->s_con, reply);
3670 	ceph_pagelist_release(recon_state->pagelist);
3671 
3672 	recon_state->pagelist = _pagelist;
3673 	recon_state->nr_caps = 0;
3674 	recon_state->nr_realms = 0;
3675 	recon_state->msg_version = 5;
3676 	return 0;
3677 fail:
3678 	ceph_msg_put(reply);
3679 fail_msg:
3680 	ceph_pagelist_release(_pagelist);
3681 	return err;
3682 }
3683 
3684 static struct dentry* d_find_primary(struct inode *inode)
3685 {
3686 	struct dentry *alias, *dn = NULL;
3687 
3688 	if (hlist_empty(&inode->i_dentry))
3689 		return NULL;
3690 
3691 	spin_lock(&inode->i_lock);
3692 	if (hlist_empty(&inode->i_dentry))
3693 		goto out_unlock;
3694 
3695 	if (S_ISDIR(inode->i_mode)) {
3696 		alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3697 		if (!IS_ROOT(alias))
3698 			dn = dget(alias);
3699 		goto out_unlock;
3700 	}
3701 
3702 	hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3703 		spin_lock(&alias->d_lock);
3704 		if (!d_unhashed(alias) &&
3705 		    (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3706 			dn = dget_dlock(alias);
3707 		}
3708 		spin_unlock(&alias->d_lock);
3709 		if (dn)
3710 			break;
3711 	}
3712 out_unlock:
3713 	spin_unlock(&inode->i_lock);
3714 	return dn;
3715 }
3716 
3717 /*
3718  * Encode information about a cap for a reconnect with the MDS.
3719  */
3720 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3721 			  void *arg)
3722 {
3723 	union {
3724 		struct ceph_mds_cap_reconnect v2;
3725 		struct ceph_mds_cap_reconnect_v1 v1;
3726 	} rec;
3727 	struct ceph_inode_info *ci = cap->ci;
3728 	struct ceph_reconnect_state *recon_state = arg;
3729 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3730 	struct dentry *dentry;
3731 	char *path;
3732 	int pathlen, err;
3733 	u64 pathbase;
3734 	u64 snap_follows;
3735 
3736 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3737 	     inode, ceph_vinop(inode), cap, cap->cap_id,
3738 	     ceph_cap_string(cap->issued));
3739 
3740 	dentry = d_find_primary(inode);
3741 	if (dentry) {
3742 		/* set pathbase to parent dir when msg_version >= 2 */
3743 		path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3744 					    recon_state->msg_version >= 2);
3745 		dput(dentry);
3746 		if (IS_ERR(path)) {
3747 			err = PTR_ERR(path);
3748 			goto out_err;
3749 		}
3750 	} else {
3751 		path = NULL;
3752 		pathlen = 0;
3753 		pathbase = 0;
3754 	}
3755 
3756 	spin_lock(&ci->i_ceph_lock);
3757 	cap->seq = 0;        /* reset cap seq */
3758 	cap->issue_seq = 0;  /* and issue_seq */
3759 	cap->mseq = 0;       /* and migrate_seq */
3760 	cap->cap_gen = cap->session->s_cap_gen;
3761 
3762 	/* These are lost when the session goes away */
3763 	if (S_ISDIR(inode->i_mode)) {
3764 		if (cap->issued & CEPH_CAP_DIR_CREATE) {
3765 			ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3766 			memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3767 		}
3768 		cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3769 	}
3770 
3771 	if (recon_state->msg_version >= 2) {
3772 		rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3773 		rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3774 		rec.v2.issued = cpu_to_le32(cap->issued);
3775 		rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3776 		rec.v2.pathbase = cpu_to_le64(pathbase);
3777 		rec.v2.flock_len = (__force __le32)
3778 			((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3779 	} else {
3780 		rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3781 		rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3782 		rec.v1.issued = cpu_to_le32(cap->issued);
3783 		rec.v1.size = cpu_to_le64(inode->i_size);
3784 		ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3785 		ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3786 		rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3787 		rec.v1.pathbase = cpu_to_le64(pathbase);
3788 	}
3789 
3790 	if (list_empty(&ci->i_cap_snaps)) {
3791 		snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3792 	} else {
3793 		struct ceph_cap_snap *capsnap =
3794 			list_first_entry(&ci->i_cap_snaps,
3795 					 struct ceph_cap_snap, ci_item);
3796 		snap_follows = capsnap->follows;
3797 	}
3798 	spin_unlock(&ci->i_ceph_lock);
3799 
3800 	if (recon_state->msg_version >= 2) {
3801 		int num_fcntl_locks, num_flock_locks;
3802 		struct ceph_filelock *flocks = NULL;
3803 		size_t struct_len, total_len = sizeof(u64);
3804 		u8 struct_v = 0;
3805 
3806 encode_again:
3807 		if (rec.v2.flock_len) {
3808 			ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3809 		} else {
3810 			num_fcntl_locks = 0;
3811 			num_flock_locks = 0;
3812 		}
3813 		if (num_fcntl_locks + num_flock_locks > 0) {
3814 			flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3815 					       sizeof(struct ceph_filelock),
3816 					       GFP_NOFS);
3817 			if (!flocks) {
3818 				err = -ENOMEM;
3819 				goto out_err;
3820 			}
3821 			err = ceph_encode_locks_to_buffer(inode, flocks,
3822 							  num_fcntl_locks,
3823 							  num_flock_locks);
3824 			if (err) {
3825 				kfree(flocks);
3826 				flocks = NULL;
3827 				if (err == -ENOSPC)
3828 					goto encode_again;
3829 				goto out_err;
3830 			}
3831 		} else {
3832 			kfree(flocks);
3833 			flocks = NULL;
3834 		}
3835 
3836 		if (recon_state->msg_version >= 3) {
3837 			/* version, compat_version and struct_len */
3838 			total_len += 2 * sizeof(u8) + sizeof(u32);
3839 			struct_v = 2;
3840 		}
3841 		/*
3842 		 * number of encoded locks is stable, so copy to pagelist
3843 		 */
3844 		struct_len = 2 * sizeof(u32) +
3845 			    (num_fcntl_locks + num_flock_locks) *
3846 			    sizeof(struct ceph_filelock);
3847 		rec.v2.flock_len = cpu_to_le32(struct_len);
3848 
3849 		struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3850 
3851 		if (struct_v >= 2)
3852 			struct_len += sizeof(u64); /* snap_follows */
3853 
3854 		total_len += struct_len;
3855 
3856 		if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3857 			err = send_reconnect_partial(recon_state);
3858 			if (err)
3859 				goto out_freeflocks;
3860 			pagelist = recon_state->pagelist;
3861 		}
3862 
3863 		err = ceph_pagelist_reserve(pagelist, total_len);
3864 		if (err)
3865 			goto out_freeflocks;
3866 
3867 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3868 		if (recon_state->msg_version >= 3) {
3869 			ceph_pagelist_encode_8(pagelist, struct_v);
3870 			ceph_pagelist_encode_8(pagelist, 1);
3871 			ceph_pagelist_encode_32(pagelist, struct_len);
3872 		}
3873 		ceph_pagelist_encode_string(pagelist, path, pathlen);
3874 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3875 		ceph_locks_to_pagelist(flocks, pagelist,
3876 				       num_fcntl_locks, num_flock_locks);
3877 		if (struct_v >= 2)
3878 			ceph_pagelist_encode_64(pagelist, snap_follows);
3879 out_freeflocks:
3880 		kfree(flocks);
3881 	} else {
3882 		err = ceph_pagelist_reserve(pagelist,
3883 					    sizeof(u64) + sizeof(u32) +
3884 					    pathlen + sizeof(rec.v1));
3885 		if (err)
3886 			goto out_err;
3887 
3888 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3889 		ceph_pagelist_encode_string(pagelist, path, pathlen);
3890 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3891 	}
3892 
3893 out_err:
3894 	ceph_mdsc_free_path(path, pathlen);
3895 	if (!err)
3896 		recon_state->nr_caps++;
3897 	return err;
3898 }
3899 
3900 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3901 			      struct ceph_reconnect_state *recon_state)
3902 {
3903 	struct rb_node *p;
3904 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3905 	int err = 0;
3906 
3907 	if (recon_state->msg_version >= 4) {
3908 		err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3909 		if (err < 0)
3910 			goto fail;
3911 	}
3912 
3913 	/*
3914 	 * snaprealms.  we provide mds with the ino, seq (version), and
3915 	 * parent for all of our realms.  If the mds has any newer info,
3916 	 * it will tell us.
3917 	 */
3918 	for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3919 		struct ceph_snap_realm *realm =
3920 		       rb_entry(p, struct ceph_snap_realm, node);
3921 		struct ceph_mds_snaprealm_reconnect sr_rec;
3922 
3923 		if (recon_state->msg_version >= 4) {
3924 			size_t need = sizeof(u8) * 2 + sizeof(u32) +
3925 				      sizeof(sr_rec);
3926 
3927 			if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3928 				err = send_reconnect_partial(recon_state);
3929 				if (err)
3930 					goto fail;
3931 				pagelist = recon_state->pagelist;
3932 			}
3933 
3934 			err = ceph_pagelist_reserve(pagelist, need);
3935 			if (err)
3936 				goto fail;
3937 
3938 			ceph_pagelist_encode_8(pagelist, 1);
3939 			ceph_pagelist_encode_8(pagelist, 1);
3940 			ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3941 		}
3942 
3943 		dout(" adding snap realm %llx seq %lld parent %llx\n",
3944 		     realm->ino, realm->seq, realm->parent_ino);
3945 		sr_rec.ino = cpu_to_le64(realm->ino);
3946 		sr_rec.seq = cpu_to_le64(realm->seq);
3947 		sr_rec.parent = cpu_to_le64(realm->parent_ino);
3948 
3949 		err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3950 		if (err)
3951 			goto fail;
3952 
3953 		recon_state->nr_realms++;
3954 	}
3955 fail:
3956 	return err;
3957 }
3958 
3959 
3960 /*
3961  * If an MDS fails and recovers, clients need to reconnect in order to
3962  * reestablish shared state.  This includes all caps issued through
3963  * this session _and_ the snap_realm hierarchy.  Because it's not
3964  * clear which snap realms the mds cares about, we send everything we
3965  * know about.. that ensures we'll then get any new info the
3966  * recovering MDS might have.
3967  *
3968  * This is a relatively heavyweight operation, but it's rare.
3969  */
3970 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3971 			       struct ceph_mds_session *session)
3972 {
3973 	struct ceph_msg *reply;
3974 	int mds = session->s_mds;
3975 	int err = -ENOMEM;
3976 	struct ceph_reconnect_state recon_state = {
3977 		.session = session,
3978 	};
3979 	LIST_HEAD(dispose);
3980 
3981 	pr_info("mds%d reconnect start\n", mds);
3982 
3983 	recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3984 	if (!recon_state.pagelist)
3985 		goto fail_nopagelist;
3986 
3987 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3988 	if (!reply)
3989 		goto fail_nomsg;
3990 
3991 	xa_destroy(&session->s_delegated_inos);
3992 
3993 	mutex_lock(&session->s_mutex);
3994 	session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3995 	session->s_seq = 0;
3996 
3997 	dout("session %p state %s\n", session,
3998 	     ceph_session_state_name(session->s_state));
3999 
4000 	spin_lock(&session->s_gen_ttl_lock);
4001 	session->s_cap_gen++;
4002 	spin_unlock(&session->s_gen_ttl_lock);
4003 
4004 	spin_lock(&session->s_cap_lock);
4005 	/* don't know if session is readonly */
4006 	session->s_readonly = 0;
4007 	/*
4008 	 * notify __ceph_remove_cap() that we are composing cap reconnect.
4009 	 * If a cap get released before being added to the cap reconnect,
4010 	 * __ceph_remove_cap() should skip queuing cap release.
4011 	 */
4012 	session->s_cap_reconnect = 1;
4013 	/* drop old cap expires; we're about to reestablish that state */
4014 	detach_cap_releases(session, &dispose);
4015 	spin_unlock(&session->s_cap_lock);
4016 	dispose_cap_releases(mdsc, &dispose);
4017 
4018 	/* trim unused caps to reduce MDS's cache rejoin time */
4019 	if (mdsc->fsc->sb->s_root)
4020 		shrink_dcache_parent(mdsc->fsc->sb->s_root);
4021 
4022 	ceph_con_close(&session->s_con);
4023 	ceph_con_open(&session->s_con,
4024 		      CEPH_ENTITY_TYPE_MDS, mds,
4025 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4026 
4027 	/* replay unsafe requests */
4028 	replay_unsafe_requests(mdsc, session);
4029 
4030 	ceph_early_kick_flushing_caps(mdsc, session);
4031 
4032 	down_read(&mdsc->snap_rwsem);
4033 
4034 	/* placeholder for nr_caps */
4035 	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4036 	if (err)
4037 		goto fail;
4038 
4039 	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4040 		recon_state.msg_version = 3;
4041 		recon_state.allow_multi = true;
4042 	} else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4043 		recon_state.msg_version = 3;
4044 	} else {
4045 		recon_state.msg_version = 2;
4046 	}
4047 	/* trsaverse this session's caps */
4048 	err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4049 
4050 	spin_lock(&session->s_cap_lock);
4051 	session->s_cap_reconnect = 0;
4052 	spin_unlock(&session->s_cap_lock);
4053 
4054 	if (err < 0)
4055 		goto fail;
4056 
4057 	/* check if all realms can be encoded into current message */
4058 	if (mdsc->num_snap_realms) {
4059 		size_t total_len =
4060 			recon_state.pagelist->length +
4061 			mdsc->num_snap_realms *
4062 			sizeof(struct ceph_mds_snaprealm_reconnect);
4063 		if (recon_state.msg_version >= 4) {
4064 			/* number of realms */
4065 			total_len += sizeof(u32);
4066 			/* version, compat_version and struct_len */
4067 			total_len += mdsc->num_snap_realms *
4068 				     (2 * sizeof(u8) + sizeof(u32));
4069 		}
4070 		if (total_len > RECONNECT_MAX_SIZE) {
4071 			if (!recon_state.allow_multi) {
4072 				err = -ENOSPC;
4073 				goto fail;
4074 			}
4075 			if (recon_state.nr_caps) {
4076 				err = send_reconnect_partial(&recon_state);
4077 				if (err)
4078 					goto fail;
4079 			}
4080 			recon_state.msg_version = 5;
4081 		}
4082 	}
4083 
4084 	err = encode_snap_realms(mdsc, &recon_state);
4085 	if (err < 0)
4086 		goto fail;
4087 
4088 	if (recon_state.msg_version >= 5) {
4089 		err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4090 		if (err < 0)
4091 			goto fail;
4092 	}
4093 
4094 	if (recon_state.nr_caps || recon_state.nr_realms) {
4095 		struct page *page =
4096 			list_first_entry(&recon_state.pagelist->head,
4097 					struct page, lru);
4098 		__le32 *addr = kmap_atomic(page);
4099 		if (recon_state.nr_caps) {
4100 			WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4101 			*addr = cpu_to_le32(recon_state.nr_caps);
4102 		} else if (recon_state.msg_version >= 4) {
4103 			*(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4104 		}
4105 		kunmap_atomic(addr);
4106 	}
4107 
4108 	reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4109 	if (recon_state.msg_version >= 4)
4110 		reply->hdr.compat_version = cpu_to_le16(4);
4111 
4112 	reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4113 	ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4114 
4115 	ceph_con_send(&session->s_con, reply);
4116 
4117 	mutex_unlock(&session->s_mutex);
4118 
4119 	mutex_lock(&mdsc->mutex);
4120 	__wake_requests(mdsc, &session->s_waiting);
4121 	mutex_unlock(&mdsc->mutex);
4122 
4123 	up_read(&mdsc->snap_rwsem);
4124 	ceph_pagelist_release(recon_state.pagelist);
4125 	return;
4126 
4127 fail:
4128 	ceph_msg_put(reply);
4129 	up_read(&mdsc->snap_rwsem);
4130 	mutex_unlock(&session->s_mutex);
4131 fail_nomsg:
4132 	ceph_pagelist_release(recon_state.pagelist);
4133 fail_nopagelist:
4134 	pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4135 	return;
4136 }
4137 
4138 
4139 /*
4140  * compare old and new mdsmaps, kicking requests
4141  * and closing out old connections as necessary
4142  *
4143  * called under mdsc->mutex.
4144  */
4145 static void check_new_map(struct ceph_mds_client *mdsc,
4146 			  struct ceph_mdsmap *newmap,
4147 			  struct ceph_mdsmap *oldmap)
4148 {
4149 	int i;
4150 	int oldstate, newstate;
4151 	struct ceph_mds_session *s;
4152 
4153 	dout("check_new_map new %u old %u\n",
4154 	     newmap->m_epoch, oldmap->m_epoch);
4155 
4156 	for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4157 		if (!mdsc->sessions[i])
4158 			continue;
4159 		s = mdsc->sessions[i];
4160 		oldstate = ceph_mdsmap_get_state(oldmap, i);
4161 		newstate = ceph_mdsmap_get_state(newmap, i);
4162 
4163 		dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4164 		     i, ceph_mds_state_name(oldstate),
4165 		     ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4166 		     ceph_mds_state_name(newstate),
4167 		     ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4168 		     ceph_session_state_name(s->s_state));
4169 
4170 		if (i >= newmap->possible_max_rank) {
4171 			/* force close session for stopped mds */
4172 			ceph_get_mds_session(s);
4173 			__unregister_session(mdsc, s);
4174 			__wake_requests(mdsc, &s->s_waiting);
4175 			mutex_unlock(&mdsc->mutex);
4176 
4177 			mutex_lock(&s->s_mutex);
4178 			cleanup_session_requests(mdsc, s);
4179 			remove_session_caps(s);
4180 			mutex_unlock(&s->s_mutex);
4181 
4182 			ceph_put_mds_session(s);
4183 
4184 			mutex_lock(&mdsc->mutex);
4185 			kick_requests(mdsc, i);
4186 			continue;
4187 		}
4188 
4189 		if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4190 			   ceph_mdsmap_get_addr(newmap, i),
4191 			   sizeof(struct ceph_entity_addr))) {
4192 			/* just close it */
4193 			mutex_unlock(&mdsc->mutex);
4194 			mutex_lock(&s->s_mutex);
4195 			mutex_lock(&mdsc->mutex);
4196 			ceph_con_close(&s->s_con);
4197 			mutex_unlock(&s->s_mutex);
4198 			s->s_state = CEPH_MDS_SESSION_RESTARTING;
4199 		} else if (oldstate == newstate) {
4200 			continue;  /* nothing new with this mds */
4201 		}
4202 
4203 		/*
4204 		 * send reconnect?
4205 		 */
4206 		if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4207 		    newstate >= CEPH_MDS_STATE_RECONNECT) {
4208 			mutex_unlock(&mdsc->mutex);
4209 			send_mds_reconnect(mdsc, s);
4210 			mutex_lock(&mdsc->mutex);
4211 		}
4212 
4213 		/*
4214 		 * kick request on any mds that has gone active.
4215 		 */
4216 		if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4217 		    newstate >= CEPH_MDS_STATE_ACTIVE) {
4218 			if (oldstate != CEPH_MDS_STATE_CREATING &&
4219 			    oldstate != CEPH_MDS_STATE_STARTING)
4220 				pr_info("mds%d recovery completed\n", s->s_mds);
4221 			kick_requests(mdsc, i);
4222 			mutex_unlock(&mdsc->mutex);
4223 			mutex_lock(&s->s_mutex);
4224 			mutex_lock(&mdsc->mutex);
4225 			ceph_kick_flushing_caps(mdsc, s);
4226 			mutex_unlock(&s->s_mutex);
4227 			wake_up_session_caps(s, RECONNECT);
4228 		}
4229 	}
4230 
4231 	for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4232 		s = mdsc->sessions[i];
4233 		if (!s)
4234 			continue;
4235 		if (!ceph_mdsmap_is_laggy(newmap, i))
4236 			continue;
4237 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4238 		    s->s_state == CEPH_MDS_SESSION_HUNG ||
4239 		    s->s_state == CEPH_MDS_SESSION_CLOSING) {
4240 			dout(" connecting to export targets of laggy mds%d\n",
4241 			     i);
4242 			__open_export_target_sessions(mdsc, s);
4243 		}
4244 	}
4245 }
4246 
4247 
4248 
4249 /*
4250  * leases
4251  */
4252 
4253 /*
4254  * caller must hold session s_mutex, dentry->d_lock
4255  */
4256 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4257 {
4258 	struct ceph_dentry_info *di = ceph_dentry(dentry);
4259 
4260 	ceph_put_mds_session(di->lease_session);
4261 	di->lease_session = NULL;
4262 }
4263 
4264 static void handle_lease(struct ceph_mds_client *mdsc,
4265 			 struct ceph_mds_session *session,
4266 			 struct ceph_msg *msg)
4267 {
4268 	struct super_block *sb = mdsc->fsc->sb;
4269 	struct inode *inode;
4270 	struct dentry *parent, *dentry;
4271 	struct ceph_dentry_info *di;
4272 	int mds = session->s_mds;
4273 	struct ceph_mds_lease *h = msg->front.iov_base;
4274 	u32 seq;
4275 	struct ceph_vino vino;
4276 	struct qstr dname;
4277 	int release = 0;
4278 
4279 	dout("handle_lease from mds%d\n", mds);
4280 
4281 	/* decode */
4282 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4283 		goto bad;
4284 	vino.ino = le64_to_cpu(h->ino);
4285 	vino.snap = CEPH_NOSNAP;
4286 	seq = le32_to_cpu(h->seq);
4287 	dname.len = get_unaligned_le32(h + 1);
4288 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4289 		goto bad;
4290 	dname.name = (void *)(h + 1) + sizeof(u32);
4291 
4292 	/* lookup inode */
4293 	inode = ceph_find_inode(sb, vino);
4294 	dout("handle_lease %s, ino %llx %p %.*s\n",
4295 	     ceph_lease_op_name(h->action), vino.ino, inode,
4296 	     dname.len, dname.name);
4297 
4298 	mutex_lock(&session->s_mutex);
4299 	inc_session_sequence(session);
4300 
4301 	if (!inode) {
4302 		dout("handle_lease no inode %llx\n", vino.ino);
4303 		goto release;
4304 	}
4305 
4306 	/* dentry */
4307 	parent = d_find_alias(inode);
4308 	if (!parent) {
4309 		dout("no parent dentry on inode %p\n", inode);
4310 		WARN_ON(1);
4311 		goto release;  /* hrm... */
4312 	}
4313 	dname.hash = full_name_hash(parent, dname.name, dname.len);
4314 	dentry = d_lookup(parent, &dname);
4315 	dput(parent);
4316 	if (!dentry)
4317 		goto release;
4318 
4319 	spin_lock(&dentry->d_lock);
4320 	di = ceph_dentry(dentry);
4321 	switch (h->action) {
4322 	case CEPH_MDS_LEASE_REVOKE:
4323 		if (di->lease_session == session) {
4324 			if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4325 				h->seq = cpu_to_le32(di->lease_seq);
4326 			__ceph_mdsc_drop_dentry_lease(dentry);
4327 		}
4328 		release = 1;
4329 		break;
4330 
4331 	case CEPH_MDS_LEASE_RENEW:
4332 		if (di->lease_session == session &&
4333 		    di->lease_gen == session->s_cap_gen &&
4334 		    di->lease_renew_from &&
4335 		    di->lease_renew_after == 0) {
4336 			unsigned long duration =
4337 				msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4338 
4339 			di->lease_seq = seq;
4340 			di->time = di->lease_renew_from + duration;
4341 			di->lease_renew_after = di->lease_renew_from +
4342 				(duration >> 1);
4343 			di->lease_renew_from = 0;
4344 		}
4345 		break;
4346 	}
4347 	spin_unlock(&dentry->d_lock);
4348 	dput(dentry);
4349 
4350 	if (!release)
4351 		goto out;
4352 
4353 release:
4354 	/* let's just reuse the same message */
4355 	h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4356 	ceph_msg_get(msg);
4357 	ceph_con_send(&session->s_con, msg);
4358 
4359 out:
4360 	mutex_unlock(&session->s_mutex);
4361 	/* avoid calling iput_final() in mds dispatch threads */
4362 	ceph_async_iput(inode);
4363 	return;
4364 
4365 bad:
4366 	pr_err("corrupt lease message\n");
4367 	ceph_msg_dump(msg);
4368 }
4369 
4370 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4371 			      struct dentry *dentry, char action,
4372 			      u32 seq)
4373 {
4374 	struct ceph_msg *msg;
4375 	struct ceph_mds_lease *lease;
4376 	struct inode *dir;
4377 	int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4378 
4379 	dout("lease_send_msg identry %p %s to mds%d\n",
4380 	     dentry, ceph_lease_op_name(action), session->s_mds);
4381 
4382 	msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4383 	if (!msg)
4384 		return;
4385 	lease = msg->front.iov_base;
4386 	lease->action = action;
4387 	lease->seq = cpu_to_le32(seq);
4388 
4389 	spin_lock(&dentry->d_lock);
4390 	dir = d_inode(dentry->d_parent);
4391 	lease->ino = cpu_to_le64(ceph_ino(dir));
4392 	lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4393 
4394 	put_unaligned_le32(dentry->d_name.len, lease + 1);
4395 	memcpy((void *)(lease + 1) + 4,
4396 	       dentry->d_name.name, dentry->d_name.len);
4397 	spin_unlock(&dentry->d_lock);
4398 	/*
4399 	 * if this is a preemptive lease RELEASE, no need to
4400 	 * flush request stream, since the actual request will
4401 	 * soon follow.
4402 	 */
4403 	msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4404 
4405 	ceph_con_send(&session->s_con, msg);
4406 }
4407 
4408 /*
4409  * lock unlock sessions, to wait ongoing session activities
4410  */
4411 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4412 {
4413 	int i;
4414 
4415 	mutex_lock(&mdsc->mutex);
4416 	for (i = 0; i < mdsc->max_sessions; i++) {
4417 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4418 		if (!s)
4419 			continue;
4420 		mutex_unlock(&mdsc->mutex);
4421 		mutex_lock(&s->s_mutex);
4422 		mutex_unlock(&s->s_mutex);
4423 		ceph_put_mds_session(s);
4424 		mutex_lock(&mdsc->mutex);
4425 	}
4426 	mutex_unlock(&mdsc->mutex);
4427 }
4428 
4429 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4430 {
4431 	struct ceph_fs_client *fsc = mdsc->fsc;
4432 
4433 	if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4434 		return;
4435 
4436 	if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4437 		return;
4438 
4439 	if (!READ_ONCE(fsc->blocklisted))
4440 		return;
4441 
4442 	pr_info("auto reconnect after blocklisted\n");
4443 	ceph_force_reconnect(fsc->sb);
4444 }
4445 
4446 bool check_session_state(struct ceph_mds_session *s)
4447 {
4448 	switch (s->s_state) {
4449 	case CEPH_MDS_SESSION_OPEN:
4450 		if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4451 			s->s_state = CEPH_MDS_SESSION_HUNG;
4452 			pr_info("mds%d hung\n", s->s_mds);
4453 		}
4454 		break;
4455 	case CEPH_MDS_SESSION_CLOSING:
4456 		/* Should never reach this when we're unmounting */
4457 		WARN_ON_ONCE(true);
4458 		fallthrough;
4459 	case CEPH_MDS_SESSION_NEW:
4460 	case CEPH_MDS_SESSION_RESTARTING:
4461 	case CEPH_MDS_SESSION_CLOSED:
4462 	case CEPH_MDS_SESSION_REJECTED:
4463 		return false;
4464 	}
4465 
4466 	return true;
4467 }
4468 
4469 /*
4470  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4471  * then we need to retransmit that request.
4472  */
4473 void inc_session_sequence(struct ceph_mds_session *s)
4474 {
4475 	lockdep_assert_held(&s->s_mutex);
4476 
4477 	s->s_seq++;
4478 
4479 	if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4480 		int ret;
4481 
4482 		dout("resending session close request for mds%d\n", s->s_mds);
4483 		ret = request_close_session(s);
4484 		if (ret < 0)
4485 			pr_err("unable to close session to mds%d: %d\n",
4486 			       s->s_mds, ret);
4487 	}
4488 }
4489 
4490 /*
4491  * delayed work -- periodically trim expired leases, renew caps with mds
4492  */
4493 static void schedule_delayed(struct ceph_mds_client *mdsc)
4494 {
4495 	int delay = 5;
4496 	unsigned hz = round_jiffies_relative(HZ * delay);
4497 	schedule_delayed_work(&mdsc->delayed_work, hz);
4498 }
4499 
4500 static void delayed_work(struct work_struct *work)
4501 {
4502 	int i;
4503 	struct ceph_mds_client *mdsc =
4504 		container_of(work, struct ceph_mds_client, delayed_work.work);
4505 	int renew_interval;
4506 	int renew_caps;
4507 
4508 	dout("mdsc delayed_work\n");
4509 
4510 	if (mdsc->stopping)
4511 		return;
4512 
4513 	mutex_lock(&mdsc->mutex);
4514 	renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4515 	renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4516 				   mdsc->last_renew_caps);
4517 	if (renew_caps)
4518 		mdsc->last_renew_caps = jiffies;
4519 
4520 	for (i = 0; i < mdsc->max_sessions; i++) {
4521 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4522 		if (!s)
4523 			continue;
4524 
4525 		if (!check_session_state(s)) {
4526 			ceph_put_mds_session(s);
4527 			continue;
4528 		}
4529 		mutex_unlock(&mdsc->mutex);
4530 
4531 		mutex_lock(&s->s_mutex);
4532 		if (renew_caps)
4533 			send_renew_caps(mdsc, s);
4534 		else
4535 			ceph_con_keepalive(&s->s_con);
4536 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4537 		    s->s_state == CEPH_MDS_SESSION_HUNG)
4538 			ceph_send_cap_releases(mdsc, s);
4539 		mutex_unlock(&s->s_mutex);
4540 		ceph_put_mds_session(s);
4541 
4542 		mutex_lock(&mdsc->mutex);
4543 	}
4544 	mutex_unlock(&mdsc->mutex);
4545 
4546 	ceph_check_delayed_caps(mdsc);
4547 
4548 	ceph_queue_cap_reclaim_work(mdsc);
4549 
4550 	ceph_trim_snapid_map(mdsc);
4551 
4552 	maybe_recover_session(mdsc);
4553 
4554 	schedule_delayed(mdsc);
4555 }
4556 
4557 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4558 
4559 {
4560 	struct ceph_mds_client *mdsc;
4561 	int err;
4562 
4563 	mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4564 	if (!mdsc)
4565 		return -ENOMEM;
4566 	mdsc->fsc = fsc;
4567 	mutex_init(&mdsc->mutex);
4568 	mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4569 	if (!mdsc->mdsmap) {
4570 		err = -ENOMEM;
4571 		goto err_mdsc;
4572 	}
4573 
4574 	init_completion(&mdsc->safe_umount_waiters);
4575 	init_waitqueue_head(&mdsc->session_close_wq);
4576 	INIT_LIST_HEAD(&mdsc->waiting_for_map);
4577 	mdsc->sessions = NULL;
4578 	atomic_set(&mdsc->num_sessions, 0);
4579 	mdsc->max_sessions = 0;
4580 	mdsc->stopping = 0;
4581 	atomic64_set(&mdsc->quotarealms_count, 0);
4582 	mdsc->quotarealms_inodes = RB_ROOT;
4583 	mutex_init(&mdsc->quotarealms_inodes_mutex);
4584 	mdsc->last_snap_seq = 0;
4585 	init_rwsem(&mdsc->snap_rwsem);
4586 	mdsc->snap_realms = RB_ROOT;
4587 	INIT_LIST_HEAD(&mdsc->snap_empty);
4588 	mdsc->num_snap_realms = 0;
4589 	spin_lock_init(&mdsc->snap_empty_lock);
4590 	mdsc->last_tid = 0;
4591 	mdsc->oldest_tid = 0;
4592 	mdsc->request_tree = RB_ROOT;
4593 	INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4594 	mdsc->last_renew_caps = jiffies;
4595 	INIT_LIST_HEAD(&mdsc->cap_delay_list);
4596 	INIT_LIST_HEAD(&mdsc->cap_wait_list);
4597 	spin_lock_init(&mdsc->cap_delay_lock);
4598 	INIT_LIST_HEAD(&mdsc->snap_flush_list);
4599 	spin_lock_init(&mdsc->snap_flush_lock);
4600 	mdsc->last_cap_flush_tid = 1;
4601 	INIT_LIST_HEAD(&mdsc->cap_flush_list);
4602 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4603 	mdsc->num_cap_flushing = 0;
4604 	spin_lock_init(&mdsc->cap_dirty_lock);
4605 	init_waitqueue_head(&mdsc->cap_flushing_wq);
4606 	INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4607 	atomic_set(&mdsc->cap_reclaim_pending, 0);
4608 	err = ceph_metric_init(&mdsc->metric);
4609 	if (err)
4610 		goto err_mdsmap;
4611 
4612 	spin_lock_init(&mdsc->dentry_list_lock);
4613 	INIT_LIST_HEAD(&mdsc->dentry_leases);
4614 	INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4615 
4616 	ceph_caps_init(mdsc);
4617 	ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4618 
4619 	spin_lock_init(&mdsc->snapid_map_lock);
4620 	mdsc->snapid_map_tree = RB_ROOT;
4621 	INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4622 
4623 	init_rwsem(&mdsc->pool_perm_rwsem);
4624 	mdsc->pool_perm_tree = RB_ROOT;
4625 
4626 	strscpy(mdsc->nodename, utsname()->nodename,
4627 		sizeof(mdsc->nodename));
4628 
4629 	fsc->mdsc = mdsc;
4630 	return 0;
4631 
4632 err_mdsmap:
4633 	kfree(mdsc->mdsmap);
4634 err_mdsc:
4635 	kfree(mdsc);
4636 	return err;
4637 }
4638 
4639 /*
4640  * Wait for safe replies on open mds requests.  If we time out, drop
4641  * all requests from the tree to avoid dangling dentry refs.
4642  */
4643 static void wait_requests(struct ceph_mds_client *mdsc)
4644 {
4645 	struct ceph_options *opts = mdsc->fsc->client->options;
4646 	struct ceph_mds_request *req;
4647 
4648 	mutex_lock(&mdsc->mutex);
4649 	if (__get_oldest_req(mdsc)) {
4650 		mutex_unlock(&mdsc->mutex);
4651 
4652 		dout("wait_requests waiting for requests\n");
4653 		wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4654 				    ceph_timeout_jiffies(opts->mount_timeout));
4655 
4656 		/* tear down remaining requests */
4657 		mutex_lock(&mdsc->mutex);
4658 		while ((req = __get_oldest_req(mdsc))) {
4659 			dout("wait_requests timed out on tid %llu\n",
4660 			     req->r_tid);
4661 			list_del_init(&req->r_wait);
4662 			__unregister_request(mdsc, req);
4663 		}
4664 	}
4665 	mutex_unlock(&mdsc->mutex);
4666 	dout("wait_requests done\n");
4667 }
4668 
4669 /*
4670  * called before mount is ro, and before dentries are torn down.
4671  * (hmm, does this still race with new lookups?)
4672  */
4673 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4674 {
4675 	dout("pre_umount\n");
4676 	mdsc->stopping = 1;
4677 
4678 	lock_unlock_sessions(mdsc);
4679 	ceph_flush_dirty_caps(mdsc);
4680 	wait_requests(mdsc);
4681 
4682 	/*
4683 	 * wait for reply handlers to drop their request refs and
4684 	 * their inode/dcache refs
4685 	 */
4686 	ceph_msgr_flush();
4687 
4688 	ceph_cleanup_quotarealms_inodes(mdsc);
4689 }
4690 
4691 /*
4692  * wait for all write mds requests to flush.
4693  */
4694 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4695 {
4696 	struct ceph_mds_request *req = NULL, *nextreq;
4697 	struct rb_node *n;
4698 
4699 	mutex_lock(&mdsc->mutex);
4700 	dout("wait_unsafe_requests want %lld\n", want_tid);
4701 restart:
4702 	req = __get_oldest_req(mdsc);
4703 	while (req && req->r_tid <= want_tid) {
4704 		/* find next request */
4705 		n = rb_next(&req->r_node);
4706 		if (n)
4707 			nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4708 		else
4709 			nextreq = NULL;
4710 		if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4711 		    (req->r_op & CEPH_MDS_OP_WRITE)) {
4712 			/* write op */
4713 			ceph_mdsc_get_request(req);
4714 			if (nextreq)
4715 				ceph_mdsc_get_request(nextreq);
4716 			mutex_unlock(&mdsc->mutex);
4717 			dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4718 			     req->r_tid, want_tid);
4719 			wait_for_completion(&req->r_safe_completion);
4720 			mutex_lock(&mdsc->mutex);
4721 			ceph_mdsc_put_request(req);
4722 			if (!nextreq)
4723 				break;  /* next dne before, so we're done! */
4724 			if (RB_EMPTY_NODE(&nextreq->r_node)) {
4725 				/* next request was removed from tree */
4726 				ceph_mdsc_put_request(nextreq);
4727 				goto restart;
4728 			}
4729 			ceph_mdsc_put_request(nextreq);  /* won't go away */
4730 		}
4731 		req = nextreq;
4732 	}
4733 	mutex_unlock(&mdsc->mutex);
4734 	dout("wait_unsafe_requests done\n");
4735 }
4736 
4737 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4738 {
4739 	u64 want_tid, want_flush;
4740 
4741 	if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4742 		return;
4743 
4744 	dout("sync\n");
4745 	mutex_lock(&mdsc->mutex);
4746 	want_tid = mdsc->last_tid;
4747 	mutex_unlock(&mdsc->mutex);
4748 
4749 	ceph_flush_dirty_caps(mdsc);
4750 	spin_lock(&mdsc->cap_dirty_lock);
4751 	want_flush = mdsc->last_cap_flush_tid;
4752 	if (!list_empty(&mdsc->cap_flush_list)) {
4753 		struct ceph_cap_flush *cf =
4754 			list_last_entry(&mdsc->cap_flush_list,
4755 					struct ceph_cap_flush, g_list);
4756 		cf->wake = true;
4757 	}
4758 	spin_unlock(&mdsc->cap_dirty_lock);
4759 
4760 	dout("sync want tid %lld flush_seq %lld\n",
4761 	     want_tid, want_flush);
4762 
4763 	wait_unsafe_requests(mdsc, want_tid);
4764 	wait_caps_flush(mdsc, want_flush);
4765 }
4766 
4767 /*
4768  * true if all sessions are closed, or we force unmount
4769  */
4770 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4771 {
4772 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4773 		return true;
4774 	return atomic_read(&mdsc->num_sessions) <= skipped;
4775 }
4776 
4777 /*
4778  * called after sb is ro.
4779  */
4780 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4781 {
4782 	struct ceph_options *opts = mdsc->fsc->client->options;
4783 	struct ceph_mds_session *session;
4784 	int i;
4785 	int skipped = 0;
4786 
4787 	dout("close_sessions\n");
4788 
4789 	/* close sessions */
4790 	mutex_lock(&mdsc->mutex);
4791 	for (i = 0; i < mdsc->max_sessions; i++) {
4792 		session = __ceph_lookup_mds_session(mdsc, i);
4793 		if (!session)
4794 			continue;
4795 		mutex_unlock(&mdsc->mutex);
4796 		mutex_lock(&session->s_mutex);
4797 		if (__close_session(mdsc, session) <= 0)
4798 			skipped++;
4799 		mutex_unlock(&session->s_mutex);
4800 		ceph_put_mds_session(session);
4801 		mutex_lock(&mdsc->mutex);
4802 	}
4803 	mutex_unlock(&mdsc->mutex);
4804 
4805 	dout("waiting for sessions to close\n");
4806 	wait_event_timeout(mdsc->session_close_wq,
4807 			   done_closing_sessions(mdsc, skipped),
4808 			   ceph_timeout_jiffies(opts->mount_timeout));
4809 
4810 	/* tear down remaining sessions */
4811 	mutex_lock(&mdsc->mutex);
4812 	for (i = 0; i < mdsc->max_sessions; i++) {
4813 		if (mdsc->sessions[i]) {
4814 			session = ceph_get_mds_session(mdsc->sessions[i]);
4815 			__unregister_session(mdsc, session);
4816 			mutex_unlock(&mdsc->mutex);
4817 			mutex_lock(&session->s_mutex);
4818 			remove_session_caps(session);
4819 			mutex_unlock(&session->s_mutex);
4820 			ceph_put_mds_session(session);
4821 			mutex_lock(&mdsc->mutex);
4822 		}
4823 	}
4824 	WARN_ON(!list_empty(&mdsc->cap_delay_list));
4825 	mutex_unlock(&mdsc->mutex);
4826 
4827 	ceph_cleanup_snapid_map(mdsc);
4828 	ceph_cleanup_empty_realms(mdsc);
4829 
4830 	cancel_work_sync(&mdsc->cap_reclaim_work);
4831 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4832 
4833 	dout("stopped\n");
4834 }
4835 
4836 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4837 {
4838 	struct ceph_mds_session *session;
4839 	int mds;
4840 
4841 	dout("force umount\n");
4842 
4843 	mutex_lock(&mdsc->mutex);
4844 	for (mds = 0; mds < mdsc->max_sessions; mds++) {
4845 		session = __ceph_lookup_mds_session(mdsc, mds);
4846 		if (!session)
4847 			continue;
4848 
4849 		if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4850 			__unregister_session(mdsc, session);
4851 		__wake_requests(mdsc, &session->s_waiting);
4852 		mutex_unlock(&mdsc->mutex);
4853 
4854 		mutex_lock(&session->s_mutex);
4855 		__close_session(mdsc, session);
4856 		if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4857 			cleanup_session_requests(mdsc, session);
4858 			remove_session_caps(session);
4859 		}
4860 		mutex_unlock(&session->s_mutex);
4861 		ceph_put_mds_session(session);
4862 
4863 		mutex_lock(&mdsc->mutex);
4864 		kick_requests(mdsc, mds);
4865 	}
4866 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4867 	mutex_unlock(&mdsc->mutex);
4868 }
4869 
4870 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4871 {
4872 	dout("stop\n");
4873 	/*
4874 	 * Make sure the delayed work stopped before releasing
4875 	 * the resources.
4876 	 *
4877 	 * Because the cancel_delayed_work_sync() will only
4878 	 * guarantee that the work finishes executing. But the
4879 	 * delayed work will re-arm itself again after that.
4880 	 */
4881 	flush_delayed_work(&mdsc->delayed_work);
4882 
4883 	if (mdsc->mdsmap)
4884 		ceph_mdsmap_destroy(mdsc->mdsmap);
4885 	kfree(mdsc->sessions);
4886 	ceph_caps_finalize(mdsc);
4887 	ceph_pool_perm_destroy(mdsc);
4888 }
4889 
4890 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4891 {
4892 	struct ceph_mds_client *mdsc = fsc->mdsc;
4893 	dout("mdsc_destroy %p\n", mdsc);
4894 
4895 	if (!mdsc)
4896 		return;
4897 
4898 	/* flush out any connection work with references to us */
4899 	ceph_msgr_flush();
4900 
4901 	ceph_mdsc_stop(mdsc);
4902 
4903 	ceph_metric_destroy(&mdsc->metric);
4904 
4905 	flush_delayed_work(&mdsc->metric.delayed_work);
4906 	fsc->mdsc = NULL;
4907 	kfree(mdsc);
4908 	dout("mdsc_destroy %p done\n", mdsc);
4909 }
4910 
4911 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4912 {
4913 	struct ceph_fs_client *fsc = mdsc->fsc;
4914 	const char *mds_namespace = fsc->mount_options->mds_namespace;
4915 	void *p = msg->front.iov_base;
4916 	void *end = p + msg->front.iov_len;
4917 	u32 epoch;
4918 	u32 num_fs;
4919 	u32 mount_fscid = (u32)-1;
4920 	int err = -EINVAL;
4921 
4922 	ceph_decode_need(&p, end, sizeof(u32), bad);
4923 	epoch = ceph_decode_32(&p);
4924 
4925 	dout("handle_fsmap epoch %u\n", epoch);
4926 
4927 	/* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
4928 	ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
4929 
4930 	ceph_decode_32_safe(&p, end, num_fs, bad);
4931 	while (num_fs-- > 0) {
4932 		void *info_p, *info_end;
4933 		u32 info_len;
4934 		u32 fscid, namelen;
4935 
4936 		ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4937 		p += 2;		// info_v, info_cv
4938 		info_len = ceph_decode_32(&p);
4939 		ceph_decode_need(&p, end, info_len, bad);
4940 		info_p = p;
4941 		info_end = p + info_len;
4942 		p = info_end;
4943 
4944 		ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4945 		fscid = ceph_decode_32(&info_p);
4946 		namelen = ceph_decode_32(&info_p);
4947 		ceph_decode_need(&info_p, info_end, namelen, bad);
4948 
4949 		if (mds_namespace &&
4950 		    strlen(mds_namespace) == namelen &&
4951 		    !strncmp(mds_namespace, (char *)info_p, namelen)) {
4952 			mount_fscid = fscid;
4953 			break;
4954 		}
4955 	}
4956 
4957 	ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4958 	if (mount_fscid != (u32)-1) {
4959 		fsc->client->monc.fs_cluster_id = mount_fscid;
4960 		ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4961 				   0, true);
4962 		ceph_monc_renew_subs(&fsc->client->monc);
4963 	} else {
4964 		err = -ENOENT;
4965 		goto err_out;
4966 	}
4967 	return;
4968 
4969 bad:
4970 	pr_err("error decoding fsmap\n");
4971 err_out:
4972 	mutex_lock(&mdsc->mutex);
4973 	mdsc->mdsmap_err = err;
4974 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4975 	mutex_unlock(&mdsc->mutex);
4976 }
4977 
4978 /*
4979  * handle mds map update.
4980  */
4981 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4982 {
4983 	u32 epoch;
4984 	u32 maplen;
4985 	void *p = msg->front.iov_base;
4986 	void *end = p + msg->front.iov_len;
4987 	struct ceph_mdsmap *newmap, *oldmap;
4988 	struct ceph_fsid fsid;
4989 	int err = -EINVAL;
4990 
4991 	ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4992 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
4993 	if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4994 		return;
4995 	epoch = ceph_decode_32(&p);
4996 	maplen = ceph_decode_32(&p);
4997 	dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4998 
4999 	/* do we need it? */
5000 	mutex_lock(&mdsc->mutex);
5001 	if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5002 		dout("handle_map epoch %u <= our %u\n",
5003 		     epoch, mdsc->mdsmap->m_epoch);
5004 		mutex_unlock(&mdsc->mutex);
5005 		return;
5006 	}
5007 
5008 	newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5009 	if (IS_ERR(newmap)) {
5010 		err = PTR_ERR(newmap);
5011 		goto bad_unlock;
5012 	}
5013 
5014 	/* swap into place */
5015 	if (mdsc->mdsmap) {
5016 		oldmap = mdsc->mdsmap;
5017 		mdsc->mdsmap = newmap;
5018 		check_new_map(mdsc, newmap, oldmap);
5019 		ceph_mdsmap_destroy(oldmap);
5020 	} else {
5021 		mdsc->mdsmap = newmap;  /* first mds map */
5022 	}
5023 	mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5024 					MAX_LFS_FILESIZE);
5025 
5026 	__wake_requests(mdsc, &mdsc->waiting_for_map);
5027 	ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5028 			  mdsc->mdsmap->m_epoch);
5029 
5030 	mutex_unlock(&mdsc->mutex);
5031 	schedule_delayed(mdsc);
5032 	return;
5033 
5034 bad_unlock:
5035 	mutex_unlock(&mdsc->mutex);
5036 bad:
5037 	pr_err("error decoding mdsmap %d\n", err);
5038 	return;
5039 }
5040 
5041 static struct ceph_connection *con_get(struct ceph_connection *con)
5042 {
5043 	struct ceph_mds_session *s = con->private;
5044 
5045 	if (ceph_get_mds_session(s))
5046 		return con;
5047 	return NULL;
5048 }
5049 
5050 static void con_put(struct ceph_connection *con)
5051 {
5052 	struct ceph_mds_session *s = con->private;
5053 
5054 	ceph_put_mds_session(s);
5055 }
5056 
5057 /*
5058  * if the client is unresponsive for long enough, the mds will kill
5059  * the session entirely.
5060  */
5061 static void peer_reset(struct ceph_connection *con)
5062 {
5063 	struct ceph_mds_session *s = con->private;
5064 	struct ceph_mds_client *mdsc = s->s_mdsc;
5065 
5066 	pr_warn("mds%d closed our session\n", s->s_mds);
5067 	send_mds_reconnect(mdsc, s);
5068 }
5069 
5070 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5071 {
5072 	struct ceph_mds_session *s = con->private;
5073 	struct ceph_mds_client *mdsc = s->s_mdsc;
5074 	int type = le16_to_cpu(msg->hdr.type);
5075 
5076 	mutex_lock(&mdsc->mutex);
5077 	if (__verify_registered_session(mdsc, s) < 0) {
5078 		mutex_unlock(&mdsc->mutex);
5079 		goto out;
5080 	}
5081 	mutex_unlock(&mdsc->mutex);
5082 
5083 	switch (type) {
5084 	case CEPH_MSG_MDS_MAP:
5085 		ceph_mdsc_handle_mdsmap(mdsc, msg);
5086 		break;
5087 	case CEPH_MSG_FS_MAP_USER:
5088 		ceph_mdsc_handle_fsmap(mdsc, msg);
5089 		break;
5090 	case CEPH_MSG_CLIENT_SESSION:
5091 		handle_session(s, msg);
5092 		break;
5093 	case CEPH_MSG_CLIENT_REPLY:
5094 		handle_reply(s, msg);
5095 		break;
5096 	case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5097 		handle_forward(mdsc, s, msg);
5098 		break;
5099 	case CEPH_MSG_CLIENT_CAPS:
5100 		ceph_handle_caps(s, msg);
5101 		break;
5102 	case CEPH_MSG_CLIENT_SNAP:
5103 		ceph_handle_snap(mdsc, s, msg);
5104 		break;
5105 	case CEPH_MSG_CLIENT_LEASE:
5106 		handle_lease(mdsc, s, msg);
5107 		break;
5108 	case CEPH_MSG_CLIENT_QUOTA:
5109 		ceph_handle_quota(mdsc, s, msg);
5110 		break;
5111 
5112 	default:
5113 		pr_err("received unknown message type %d %s\n", type,
5114 		       ceph_msg_type_name(type));
5115 	}
5116 out:
5117 	ceph_msg_put(msg);
5118 }
5119 
5120 /*
5121  * authentication
5122  */
5123 
5124 /*
5125  * Note: returned pointer is the address of a structure that's
5126  * managed separately.  Caller must *not* attempt to free it.
5127  */
5128 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
5129 					int *proto, int force_new)
5130 {
5131 	struct ceph_mds_session *s = con->private;
5132 	struct ceph_mds_client *mdsc = s->s_mdsc;
5133 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5134 	struct ceph_auth_handshake *auth = &s->s_auth;
5135 	int ret;
5136 
5137 	ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5138 					 force_new, proto, NULL, NULL);
5139 	if (ret)
5140 		return ERR_PTR(ret);
5141 
5142 	return auth;
5143 }
5144 
5145 static int add_authorizer_challenge(struct ceph_connection *con,
5146 				    void *challenge_buf, int challenge_buf_len)
5147 {
5148 	struct ceph_mds_session *s = con->private;
5149 	struct ceph_mds_client *mdsc = s->s_mdsc;
5150 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5151 
5152 	return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5153 					    challenge_buf, challenge_buf_len);
5154 }
5155 
5156 static int verify_authorizer_reply(struct ceph_connection *con)
5157 {
5158 	struct ceph_mds_session *s = con->private;
5159 	struct ceph_mds_client *mdsc = s->s_mdsc;
5160 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5161 	struct ceph_auth_handshake *auth = &s->s_auth;
5162 
5163 	return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5164 		auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5165 		NULL, NULL, NULL, NULL);
5166 }
5167 
5168 static int invalidate_authorizer(struct ceph_connection *con)
5169 {
5170 	struct ceph_mds_session *s = con->private;
5171 	struct ceph_mds_client *mdsc = s->s_mdsc;
5172 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5173 
5174 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5175 
5176 	return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5177 }
5178 
5179 static int mds_get_auth_request(struct ceph_connection *con,
5180 				void *buf, int *buf_len,
5181 				void **authorizer, int *authorizer_len)
5182 {
5183 	struct ceph_mds_session *s = con->private;
5184 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5185 	struct ceph_auth_handshake *auth = &s->s_auth;
5186 	int ret;
5187 
5188 	ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5189 				       buf, buf_len);
5190 	if (ret)
5191 		return ret;
5192 
5193 	*authorizer = auth->authorizer_buf;
5194 	*authorizer_len = auth->authorizer_buf_len;
5195 	return 0;
5196 }
5197 
5198 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5199 				      void *reply, int reply_len,
5200 				      void *buf, int *buf_len,
5201 				      void **authorizer, int *authorizer_len)
5202 {
5203 	struct ceph_mds_session *s = con->private;
5204 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5205 	struct ceph_auth_handshake *auth = &s->s_auth;
5206 	int ret;
5207 
5208 	ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5209 					      buf, buf_len);
5210 	if (ret)
5211 		return ret;
5212 
5213 	*authorizer = auth->authorizer_buf;
5214 	*authorizer_len = auth->authorizer_buf_len;
5215 	return 0;
5216 }
5217 
5218 static int mds_handle_auth_done(struct ceph_connection *con,
5219 				u64 global_id, void *reply, int reply_len,
5220 				u8 *session_key, int *session_key_len,
5221 				u8 *con_secret, int *con_secret_len)
5222 {
5223 	struct ceph_mds_session *s = con->private;
5224 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5225 	struct ceph_auth_handshake *auth = &s->s_auth;
5226 
5227 	return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5228 					       session_key, session_key_len,
5229 					       con_secret, con_secret_len);
5230 }
5231 
5232 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5233 				      int used_proto, int result,
5234 				      const int *allowed_protos, int proto_cnt,
5235 				      const int *allowed_modes, int mode_cnt)
5236 {
5237 	struct ceph_mds_session *s = con->private;
5238 	struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5239 	int ret;
5240 
5241 	if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5242 					    used_proto, result,
5243 					    allowed_protos, proto_cnt,
5244 					    allowed_modes, mode_cnt)) {
5245 		ret = ceph_monc_validate_auth(monc);
5246 		if (ret)
5247 			return ret;
5248 	}
5249 
5250 	return -EACCES;
5251 }
5252 
5253 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5254 				struct ceph_msg_header *hdr, int *skip)
5255 {
5256 	struct ceph_msg *msg;
5257 	int type = (int) le16_to_cpu(hdr->type);
5258 	int front_len = (int) le32_to_cpu(hdr->front_len);
5259 
5260 	if (con->in_msg)
5261 		return con->in_msg;
5262 
5263 	*skip = 0;
5264 	msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5265 	if (!msg) {
5266 		pr_err("unable to allocate msg type %d len %d\n",
5267 		       type, front_len);
5268 		return NULL;
5269 	}
5270 
5271 	return msg;
5272 }
5273 
5274 static int mds_sign_message(struct ceph_msg *msg)
5275 {
5276        struct ceph_mds_session *s = msg->con->private;
5277        struct ceph_auth_handshake *auth = &s->s_auth;
5278 
5279        return ceph_auth_sign_message(auth, msg);
5280 }
5281 
5282 static int mds_check_message_signature(struct ceph_msg *msg)
5283 {
5284        struct ceph_mds_session *s = msg->con->private;
5285        struct ceph_auth_handshake *auth = &s->s_auth;
5286 
5287        return ceph_auth_check_message_signature(auth, msg);
5288 }
5289 
5290 static const struct ceph_connection_operations mds_con_ops = {
5291 	.get = con_get,
5292 	.put = con_put,
5293 	.dispatch = dispatch,
5294 	.get_authorizer = get_authorizer,
5295 	.add_authorizer_challenge = add_authorizer_challenge,
5296 	.verify_authorizer_reply = verify_authorizer_reply,
5297 	.invalidate_authorizer = invalidate_authorizer,
5298 	.peer_reset = peer_reset,
5299 	.alloc_msg = mds_alloc_msg,
5300 	.sign_message = mds_sign_message,
5301 	.check_message_signature = mds_check_message_signature,
5302 	.get_auth_request = mds_get_auth_request,
5303 	.handle_auth_reply_more = mds_handle_auth_reply_more,
5304 	.handle_auth_done = mds_handle_auth_done,
5305 	.handle_auth_bad_method = mds_handle_auth_bad_method,
5306 };
5307 
5308 /* eof */
5309