xref: /openbmc/linux/fs/ceph/mds_client.c (revision d9679d00)
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 #include <linux/bitmap.h>
15 
16 #include "super.h"
17 #include "mds_client.h"
18 
19 #include <linux/ceph/ceph_features.h>
20 #include <linux/ceph/messenger.h>
21 #include <linux/ceph/decode.h>
22 #include <linux/ceph/pagelist.h>
23 #include <linux/ceph/auth.h>
24 #include <linux/ceph/debugfs.h>
25 
26 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
27 
28 /*
29  * A cluster of MDS (metadata server) daemons is responsible for
30  * managing the file system namespace (the directory hierarchy and
31  * inodes) and for coordinating shared access to storage.  Metadata is
32  * partitioning hierarchically across a number of servers, and that
33  * partition varies over time as the cluster adjusts the distribution
34  * in order to balance load.
35  *
36  * The MDS client is primarily responsible to managing synchronous
37  * metadata requests for operations like open, unlink, and so forth.
38  * If there is a MDS failure, we find out about it when we (possibly
39  * request and) receive a new MDS map, and can resubmit affected
40  * requests.
41  *
42  * For the most part, though, we take advantage of a lossless
43  * communications channel to the MDS, and do not need to worry about
44  * timing out or resubmitting requests.
45  *
46  * We maintain a stateful "session" with each MDS we interact with.
47  * Within each session, we sent periodic heartbeat messages to ensure
48  * any capabilities or leases we have been issues remain valid.  If
49  * the session times out and goes stale, our leases and capabilities
50  * are no longer valid.
51  */
52 
53 struct ceph_reconnect_state {
54 	struct ceph_mds_session *session;
55 	int nr_caps, nr_realms;
56 	struct ceph_pagelist *pagelist;
57 	unsigned msg_version;
58 	bool allow_multi;
59 };
60 
61 static void __wake_requests(struct ceph_mds_client *mdsc,
62 			    struct list_head *head);
63 static void ceph_cap_release_work(struct work_struct *work);
64 static void ceph_cap_reclaim_work(struct work_struct *work);
65 
66 static const struct ceph_connection_operations mds_con_ops;
67 
68 
69 /*
70  * mds reply parsing
71  */
72 
73 static int parse_reply_info_quota(void **p, void *end,
74 				  struct ceph_mds_reply_info_in *info)
75 {
76 	u8 struct_v, struct_compat;
77 	u32 struct_len;
78 
79 	ceph_decode_8_safe(p, end, struct_v, bad);
80 	ceph_decode_8_safe(p, end, struct_compat, bad);
81 	/* struct_v is expected to be >= 1. we only
82 	 * understand encoding with struct_compat == 1. */
83 	if (!struct_v || struct_compat != 1)
84 		goto bad;
85 	ceph_decode_32_safe(p, end, struct_len, bad);
86 	ceph_decode_need(p, end, struct_len, bad);
87 	end = *p + struct_len;
88 	ceph_decode_64_safe(p, end, info->max_bytes, bad);
89 	ceph_decode_64_safe(p, end, info->max_files, bad);
90 	*p = end;
91 	return 0;
92 bad:
93 	return -EIO;
94 }
95 
96 /*
97  * parse individual inode info
98  */
99 static int parse_reply_info_in(void **p, void *end,
100 			       struct ceph_mds_reply_info_in *info,
101 			       u64 features)
102 {
103 	int err = 0;
104 	u8 struct_v = 0;
105 
106 	if (features == (u64)-1) {
107 		u32 struct_len;
108 		u8 struct_compat;
109 		ceph_decode_8_safe(p, end, struct_v, bad);
110 		ceph_decode_8_safe(p, end, struct_compat, bad);
111 		/* struct_v is expected to be >= 1. we only understand
112 		 * encoding with struct_compat == 1. */
113 		if (!struct_v || struct_compat != 1)
114 			goto bad;
115 		ceph_decode_32_safe(p, end, struct_len, bad);
116 		ceph_decode_need(p, end, struct_len, bad);
117 		end = *p + struct_len;
118 	}
119 
120 	ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
121 	info->in = *p;
122 	*p += sizeof(struct ceph_mds_reply_inode) +
123 		sizeof(*info->in->fragtree.splits) *
124 		le32_to_cpu(info->in->fragtree.nsplits);
125 
126 	ceph_decode_32_safe(p, end, info->symlink_len, bad);
127 	ceph_decode_need(p, end, info->symlink_len, bad);
128 	info->symlink = *p;
129 	*p += info->symlink_len;
130 
131 	ceph_decode_copy_safe(p, end, &info->dir_layout,
132 			      sizeof(info->dir_layout), bad);
133 	ceph_decode_32_safe(p, end, info->xattr_len, bad);
134 	ceph_decode_need(p, end, info->xattr_len, bad);
135 	info->xattr_data = *p;
136 	*p += info->xattr_len;
137 
138 	if (features == (u64)-1) {
139 		/* inline data */
140 		ceph_decode_64_safe(p, end, info->inline_version, bad);
141 		ceph_decode_32_safe(p, end, info->inline_len, bad);
142 		ceph_decode_need(p, end, info->inline_len, bad);
143 		info->inline_data = *p;
144 		*p += info->inline_len;
145 		/* quota */
146 		err = parse_reply_info_quota(p, end, info);
147 		if (err < 0)
148 			goto out_bad;
149 		/* pool namespace */
150 		ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
151 		if (info->pool_ns_len > 0) {
152 			ceph_decode_need(p, end, info->pool_ns_len, bad);
153 			info->pool_ns_data = *p;
154 			*p += info->pool_ns_len;
155 		}
156 
157 		/* btime */
158 		ceph_decode_need(p, end, sizeof(info->btime), bad);
159 		ceph_decode_copy(p, &info->btime, sizeof(info->btime));
160 
161 		/* change attribute */
162 		ceph_decode_64_safe(p, end, info->change_attr, bad);
163 
164 		/* dir pin */
165 		if (struct_v >= 2) {
166 			ceph_decode_32_safe(p, end, info->dir_pin, bad);
167 		} else {
168 			info->dir_pin = -ENODATA;
169 		}
170 
171 		/* snapshot birth time, remains zero for v<=2 */
172 		if (struct_v >= 3) {
173 			ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
174 			ceph_decode_copy(p, &info->snap_btime,
175 					 sizeof(info->snap_btime));
176 		} else {
177 			memset(&info->snap_btime, 0, sizeof(info->snap_btime));
178 		}
179 
180 		/* snapshot count, remains zero for v<=3 */
181 		if (struct_v >= 4) {
182 			ceph_decode_64_safe(p, end, info->rsnaps, bad);
183 		} else {
184 			info->rsnaps = 0;
185 		}
186 
187 		*p = end;
188 	} else {
189 		if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
190 			ceph_decode_64_safe(p, end, info->inline_version, bad);
191 			ceph_decode_32_safe(p, end, info->inline_len, bad);
192 			ceph_decode_need(p, end, info->inline_len, bad);
193 			info->inline_data = *p;
194 			*p += info->inline_len;
195 		} else
196 			info->inline_version = CEPH_INLINE_NONE;
197 
198 		if (features & CEPH_FEATURE_MDS_QUOTA) {
199 			err = parse_reply_info_quota(p, end, info);
200 			if (err < 0)
201 				goto out_bad;
202 		} else {
203 			info->max_bytes = 0;
204 			info->max_files = 0;
205 		}
206 
207 		info->pool_ns_len = 0;
208 		info->pool_ns_data = NULL;
209 		if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
210 			ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
211 			if (info->pool_ns_len > 0) {
212 				ceph_decode_need(p, end, info->pool_ns_len, bad);
213 				info->pool_ns_data = *p;
214 				*p += info->pool_ns_len;
215 			}
216 		}
217 
218 		if (features & CEPH_FEATURE_FS_BTIME) {
219 			ceph_decode_need(p, end, sizeof(info->btime), bad);
220 			ceph_decode_copy(p, &info->btime, sizeof(info->btime));
221 			ceph_decode_64_safe(p, end, info->change_attr, bad);
222 		}
223 
224 		info->dir_pin = -ENODATA;
225 		/* info->snap_btime and info->rsnaps remain zero */
226 	}
227 	return 0;
228 bad:
229 	err = -EIO;
230 out_bad:
231 	return err;
232 }
233 
234 static int parse_reply_info_dir(void **p, void *end,
235 				struct ceph_mds_reply_dirfrag **dirfrag,
236 				u64 features)
237 {
238 	if (features == (u64)-1) {
239 		u8 struct_v, struct_compat;
240 		u32 struct_len;
241 		ceph_decode_8_safe(p, end, struct_v, bad);
242 		ceph_decode_8_safe(p, end, struct_compat, bad);
243 		/* struct_v is expected to be >= 1. we only understand
244 		 * encoding whose struct_compat == 1. */
245 		if (!struct_v || struct_compat != 1)
246 			goto bad;
247 		ceph_decode_32_safe(p, end, struct_len, bad);
248 		ceph_decode_need(p, end, struct_len, bad);
249 		end = *p + struct_len;
250 	}
251 
252 	ceph_decode_need(p, end, sizeof(**dirfrag), bad);
253 	*dirfrag = *p;
254 	*p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
255 	if (unlikely(*p > end))
256 		goto bad;
257 	if (features == (u64)-1)
258 		*p = end;
259 	return 0;
260 bad:
261 	return -EIO;
262 }
263 
264 static int parse_reply_info_lease(void **p, void *end,
265 				  struct ceph_mds_reply_lease **lease,
266 				  u64 features)
267 {
268 	if (features == (u64)-1) {
269 		u8 struct_v, struct_compat;
270 		u32 struct_len;
271 		ceph_decode_8_safe(p, end, struct_v, bad);
272 		ceph_decode_8_safe(p, end, struct_compat, bad);
273 		/* struct_v is expected to be >= 1. we only understand
274 		 * encoding whose struct_compat == 1. */
275 		if (!struct_v || struct_compat != 1)
276 			goto bad;
277 		ceph_decode_32_safe(p, end, struct_len, bad);
278 		ceph_decode_need(p, end, struct_len, bad);
279 		end = *p + struct_len;
280 	}
281 
282 	ceph_decode_need(p, end, sizeof(**lease), bad);
283 	*lease = *p;
284 	*p += sizeof(**lease);
285 	if (features == (u64)-1)
286 		*p = end;
287 	return 0;
288 bad:
289 	return -EIO;
290 }
291 
292 /*
293  * parse a normal reply, which may contain a (dir+)dentry and/or a
294  * target inode.
295  */
296 static int parse_reply_info_trace(void **p, void *end,
297 				  struct ceph_mds_reply_info_parsed *info,
298 				  u64 features)
299 {
300 	int err;
301 
302 	if (info->head->is_dentry) {
303 		err = parse_reply_info_in(p, end, &info->diri, features);
304 		if (err < 0)
305 			goto out_bad;
306 
307 		err = parse_reply_info_dir(p, end, &info->dirfrag, features);
308 		if (err < 0)
309 			goto out_bad;
310 
311 		ceph_decode_32_safe(p, end, info->dname_len, bad);
312 		ceph_decode_need(p, end, info->dname_len, bad);
313 		info->dname = *p;
314 		*p += info->dname_len;
315 
316 		err = parse_reply_info_lease(p, end, &info->dlease, features);
317 		if (err < 0)
318 			goto out_bad;
319 	}
320 
321 	if (info->head->is_target) {
322 		err = parse_reply_info_in(p, end, &info->targeti, features);
323 		if (err < 0)
324 			goto out_bad;
325 	}
326 
327 	if (unlikely(*p != end))
328 		goto bad;
329 	return 0;
330 
331 bad:
332 	err = -EIO;
333 out_bad:
334 	pr_err("problem parsing mds trace %d\n", err);
335 	return err;
336 }
337 
338 /*
339  * parse readdir results
340  */
341 static int parse_reply_info_readdir(void **p, void *end,
342 				struct ceph_mds_reply_info_parsed *info,
343 				u64 features)
344 {
345 	u32 num, i = 0;
346 	int err;
347 
348 	err = parse_reply_info_dir(p, end, &info->dir_dir, features);
349 	if (err < 0)
350 		goto out_bad;
351 
352 	ceph_decode_need(p, end, sizeof(num) + 2, bad);
353 	num = ceph_decode_32(p);
354 	{
355 		u16 flags = ceph_decode_16(p);
356 		info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
357 		info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
358 		info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
359 		info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
360 	}
361 	if (num == 0)
362 		goto done;
363 
364 	BUG_ON(!info->dir_entries);
365 	if ((unsigned long)(info->dir_entries + num) >
366 	    (unsigned long)info->dir_entries + info->dir_buf_size) {
367 		pr_err("dir contents are larger than expected\n");
368 		WARN_ON(1);
369 		goto bad;
370 	}
371 
372 	info->dir_nr = num;
373 	while (num) {
374 		struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
375 		/* dentry */
376 		ceph_decode_32_safe(p, end, rde->name_len, bad);
377 		ceph_decode_need(p, end, rde->name_len, bad);
378 		rde->name = *p;
379 		*p += rde->name_len;
380 		dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
381 
382 		/* dentry lease */
383 		err = parse_reply_info_lease(p, end, &rde->lease, features);
384 		if (err)
385 			goto out_bad;
386 		/* inode */
387 		err = parse_reply_info_in(p, end, &rde->inode, features);
388 		if (err < 0)
389 			goto out_bad;
390 		/* ceph_readdir_prepopulate() will update it */
391 		rde->offset = 0;
392 		i++;
393 		num--;
394 	}
395 
396 done:
397 	/* Skip over any unrecognized fields */
398 	*p = end;
399 	return 0;
400 
401 bad:
402 	err = -EIO;
403 out_bad:
404 	pr_err("problem parsing dir contents %d\n", err);
405 	return err;
406 }
407 
408 /*
409  * parse fcntl F_GETLK results
410  */
411 static int parse_reply_info_filelock(void **p, void *end,
412 				     struct ceph_mds_reply_info_parsed *info,
413 				     u64 features)
414 {
415 	if (*p + sizeof(*info->filelock_reply) > end)
416 		goto bad;
417 
418 	info->filelock_reply = *p;
419 
420 	/* Skip over any unrecognized fields */
421 	*p = end;
422 	return 0;
423 bad:
424 	return -EIO;
425 }
426 
427 
428 #if BITS_PER_LONG == 64
429 
430 #define DELEGATED_INO_AVAILABLE		xa_mk_value(1)
431 
432 static int ceph_parse_deleg_inos(void **p, void *end,
433 				 struct ceph_mds_session *s)
434 {
435 	u32 sets;
436 
437 	ceph_decode_32_safe(p, end, sets, bad);
438 	dout("got %u sets of delegated inodes\n", sets);
439 	while (sets--) {
440 		u64 start, len, ino;
441 
442 		ceph_decode_64_safe(p, end, start, bad);
443 		ceph_decode_64_safe(p, end, len, bad);
444 
445 		/* Don't accept a delegation of system inodes */
446 		if (start < CEPH_INO_SYSTEM_BASE) {
447 			pr_warn_ratelimited("ceph: ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
448 					start, len);
449 			continue;
450 		}
451 		while (len--) {
452 			int err = xa_insert(&s->s_delegated_inos, ino = start++,
453 					    DELEGATED_INO_AVAILABLE,
454 					    GFP_KERNEL);
455 			if (!err) {
456 				dout("added delegated inode 0x%llx\n",
457 				     start - 1);
458 			} else if (err == -EBUSY) {
459 				pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
460 					start - 1);
461 			} else {
462 				return err;
463 			}
464 		}
465 	}
466 	return 0;
467 bad:
468 	return -EIO;
469 }
470 
471 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
472 {
473 	unsigned long ino;
474 	void *val;
475 
476 	xa_for_each(&s->s_delegated_inos, ino, val) {
477 		val = xa_erase(&s->s_delegated_inos, ino);
478 		if (val == DELEGATED_INO_AVAILABLE)
479 			return ino;
480 	}
481 	return 0;
482 }
483 
484 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
485 {
486 	return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
487 			 GFP_KERNEL);
488 }
489 #else /* BITS_PER_LONG == 64 */
490 /*
491  * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
492  * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
493  * and bottom words?
494  */
495 static int ceph_parse_deleg_inos(void **p, void *end,
496 				 struct ceph_mds_session *s)
497 {
498 	u32 sets;
499 
500 	ceph_decode_32_safe(p, end, sets, bad);
501 	if (sets)
502 		ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
503 	return 0;
504 bad:
505 	return -EIO;
506 }
507 
508 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
509 {
510 	return 0;
511 }
512 
513 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
514 {
515 	return 0;
516 }
517 #endif /* BITS_PER_LONG == 64 */
518 
519 /*
520  * parse create results
521  */
522 static int parse_reply_info_create(void **p, void *end,
523 				  struct ceph_mds_reply_info_parsed *info,
524 				  u64 features, struct ceph_mds_session *s)
525 {
526 	int ret;
527 
528 	if (features == (u64)-1 ||
529 	    (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
530 		if (*p == end) {
531 			/* Malformed reply? */
532 			info->has_create_ino = false;
533 		} else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
534 			info->has_create_ino = true;
535 			/* struct_v, struct_compat, and len */
536 			ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
537 			ceph_decode_64_safe(p, end, info->ino, bad);
538 			ret = ceph_parse_deleg_inos(p, end, s);
539 			if (ret)
540 				return ret;
541 		} else {
542 			/* legacy */
543 			ceph_decode_64_safe(p, end, info->ino, bad);
544 			info->has_create_ino = true;
545 		}
546 	} else {
547 		if (*p != end)
548 			goto bad;
549 	}
550 
551 	/* Skip over any unrecognized fields */
552 	*p = end;
553 	return 0;
554 bad:
555 	return -EIO;
556 }
557 
558 /*
559  * parse extra results
560  */
561 static int parse_reply_info_extra(void **p, void *end,
562 				  struct ceph_mds_reply_info_parsed *info,
563 				  u64 features, struct ceph_mds_session *s)
564 {
565 	u32 op = le32_to_cpu(info->head->op);
566 
567 	if (op == CEPH_MDS_OP_GETFILELOCK)
568 		return parse_reply_info_filelock(p, end, info, features);
569 	else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
570 		return parse_reply_info_readdir(p, end, info, features);
571 	else if (op == CEPH_MDS_OP_CREATE)
572 		return parse_reply_info_create(p, end, info, features, s);
573 	else
574 		return -EIO;
575 }
576 
577 /*
578  * parse entire mds reply
579  */
580 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
581 			    struct ceph_mds_reply_info_parsed *info,
582 			    u64 features)
583 {
584 	void *p, *end;
585 	u32 len;
586 	int err;
587 
588 	info->head = msg->front.iov_base;
589 	p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
590 	end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
591 
592 	/* trace */
593 	ceph_decode_32_safe(&p, end, len, bad);
594 	if (len > 0) {
595 		ceph_decode_need(&p, end, len, bad);
596 		err = parse_reply_info_trace(&p, p+len, info, features);
597 		if (err < 0)
598 			goto out_bad;
599 	}
600 
601 	/* extra */
602 	ceph_decode_32_safe(&p, end, len, bad);
603 	if (len > 0) {
604 		ceph_decode_need(&p, end, len, bad);
605 		err = parse_reply_info_extra(&p, p+len, info, features, s);
606 		if (err < 0)
607 			goto out_bad;
608 	}
609 
610 	/* snap blob */
611 	ceph_decode_32_safe(&p, end, len, bad);
612 	info->snapblob_len = len;
613 	info->snapblob = p;
614 	p += len;
615 
616 	if (p != end)
617 		goto bad;
618 	return 0;
619 
620 bad:
621 	err = -EIO;
622 out_bad:
623 	pr_err("mds parse_reply err %d\n", err);
624 	return err;
625 }
626 
627 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
628 {
629 	if (!info->dir_entries)
630 		return;
631 	free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
632 }
633 
634 
635 /*
636  * sessions
637  */
638 const char *ceph_session_state_name(int s)
639 {
640 	switch (s) {
641 	case CEPH_MDS_SESSION_NEW: return "new";
642 	case CEPH_MDS_SESSION_OPENING: return "opening";
643 	case CEPH_MDS_SESSION_OPEN: return "open";
644 	case CEPH_MDS_SESSION_HUNG: return "hung";
645 	case CEPH_MDS_SESSION_CLOSING: return "closing";
646 	case CEPH_MDS_SESSION_CLOSED: return "closed";
647 	case CEPH_MDS_SESSION_RESTARTING: return "restarting";
648 	case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
649 	case CEPH_MDS_SESSION_REJECTED: return "rejected";
650 	default: return "???";
651 	}
652 }
653 
654 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
655 {
656 	if (refcount_inc_not_zero(&s->s_ref))
657 		return s;
658 	return NULL;
659 }
660 
661 void ceph_put_mds_session(struct ceph_mds_session *s)
662 {
663 	if (IS_ERR_OR_NULL(s))
664 		return;
665 
666 	if (refcount_dec_and_test(&s->s_ref)) {
667 		if (s->s_auth.authorizer)
668 			ceph_auth_destroy_authorizer(s->s_auth.authorizer);
669 		WARN_ON(mutex_is_locked(&s->s_mutex));
670 		xa_destroy(&s->s_delegated_inos);
671 		kfree(s);
672 	}
673 }
674 
675 /*
676  * called under mdsc->mutex
677  */
678 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
679 						   int mds)
680 {
681 	if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
682 		return NULL;
683 	return ceph_get_mds_session(mdsc->sessions[mds]);
684 }
685 
686 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
687 {
688 	if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
689 		return false;
690 	else
691 		return true;
692 }
693 
694 static int __verify_registered_session(struct ceph_mds_client *mdsc,
695 				       struct ceph_mds_session *s)
696 {
697 	if (s->s_mds >= mdsc->max_sessions ||
698 	    mdsc->sessions[s->s_mds] != s)
699 		return -ENOENT;
700 	return 0;
701 }
702 
703 /*
704  * create+register a new session for given mds.
705  * called under mdsc->mutex.
706  */
707 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
708 						 int mds)
709 {
710 	struct ceph_mds_session *s;
711 
712 	if (mds >= mdsc->mdsmap->possible_max_rank)
713 		return ERR_PTR(-EINVAL);
714 
715 	s = kzalloc(sizeof(*s), GFP_NOFS);
716 	if (!s)
717 		return ERR_PTR(-ENOMEM);
718 
719 	if (mds >= mdsc->max_sessions) {
720 		int newmax = 1 << get_count_order(mds + 1);
721 		struct ceph_mds_session **sa;
722 
723 		dout("%s: realloc to %d\n", __func__, newmax);
724 		sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
725 		if (!sa)
726 			goto fail_realloc;
727 		if (mdsc->sessions) {
728 			memcpy(sa, mdsc->sessions,
729 			       mdsc->max_sessions * sizeof(void *));
730 			kfree(mdsc->sessions);
731 		}
732 		mdsc->sessions = sa;
733 		mdsc->max_sessions = newmax;
734 	}
735 
736 	dout("%s: mds%d\n", __func__, mds);
737 	s->s_mdsc = mdsc;
738 	s->s_mds = mds;
739 	s->s_state = CEPH_MDS_SESSION_NEW;
740 	mutex_init(&s->s_mutex);
741 
742 	ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
743 
744 	atomic_set(&s->s_cap_gen, 1);
745 	s->s_cap_ttl = jiffies - 1;
746 
747 	spin_lock_init(&s->s_cap_lock);
748 	INIT_LIST_HEAD(&s->s_caps);
749 	refcount_set(&s->s_ref, 1);
750 	INIT_LIST_HEAD(&s->s_waiting);
751 	INIT_LIST_HEAD(&s->s_unsafe);
752 	xa_init(&s->s_delegated_inos);
753 	INIT_LIST_HEAD(&s->s_cap_releases);
754 	INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
755 
756 	INIT_LIST_HEAD(&s->s_cap_dirty);
757 	INIT_LIST_HEAD(&s->s_cap_flushing);
758 
759 	mdsc->sessions[mds] = s;
760 	atomic_inc(&mdsc->num_sessions);
761 	refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
762 
763 	ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
764 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
765 
766 	return s;
767 
768 fail_realloc:
769 	kfree(s);
770 	return ERR_PTR(-ENOMEM);
771 }
772 
773 /*
774  * called under mdsc->mutex
775  */
776 static void __unregister_session(struct ceph_mds_client *mdsc,
777 			       struct ceph_mds_session *s)
778 {
779 	dout("__unregister_session mds%d %p\n", s->s_mds, s);
780 	BUG_ON(mdsc->sessions[s->s_mds] != s);
781 	mdsc->sessions[s->s_mds] = NULL;
782 	ceph_con_close(&s->s_con);
783 	ceph_put_mds_session(s);
784 	atomic_dec(&mdsc->num_sessions);
785 }
786 
787 /*
788  * drop session refs in request.
789  *
790  * should be last request ref, or hold mdsc->mutex
791  */
792 static void put_request_session(struct ceph_mds_request *req)
793 {
794 	if (req->r_session) {
795 		ceph_put_mds_session(req->r_session);
796 		req->r_session = NULL;
797 	}
798 }
799 
800 void ceph_mdsc_iterate_sessions(struct ceph_mds_client *mdsc,
801 				void (*cb)(struct ceph_mds_session *),
802 				bool check_state)
803 {
804 	int mds;
805 
806 	mutex_lock(&mdsc->mutex);
807 	for (mds = 0; mds < mdsc->max_sessions; ++mds) {
808 		struct ceph_mds_session *s;
809 
810 		s = __ceph_lookup_mds_session(mdsc, mds);
811 		if (!s)
812 			continue;
813 
814 		if (check_state && !check_session_state(s)) {
815 			ceph_put_mds_session(s);
816 			continue;
817 		}
818 
819 		mutex_unlock(&mdsc->mutex);
820 		cb(s);
821 		ceph_put_mds_session(s);
822 		mutex_lock(&mdsc->mutex);
823 	}
824 	mutex_unlock(&mdsc->mutex);
825 }
826 
827 void ceph_mdsc_release_request(struct kref *kref)
828 {
829 	struct ceph_mds_request *req = container_of(kref,
830 						    struct ceph_mds_request,
831 						    r_kref);
832 	ceph_mdsc_release_dir_caps_no_check(req);
833 	destroy_reply_info(&req->r_reply_info);
834 	if (req->r_request)
835 		ceph_msg_put(req->r_request);
836 	if (req->r_reply)
837 		ceph_msg_put(req->r_reply);
838 	if (req->r_inode) {
839 		ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
840 		iput(req->r_inode);
841 	}
842 	if (req->r_parent) {
843 		ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
844 		iput(req->r_parent);
845 	}
846 	iput(req->r_target_inode);
847 	if (req->r_dentry)
848 		dput(req->r_dentry);
849 	if (req->r_old_dentry)
850 		dput(req->r_old_dentry);
851 	if (req->r_old_dentry_dir) {
852 		/*
853 		 * track (and drop pins for) r_old_dentry_dir
854 		 * separately, since r_old_dentry's d_parent may have
855 		 * changed between the dir mutex being dropped and
856 		 * this request being freed.
857 		 */
858 		ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
859 				  CEPH_CAP_PIN);
860 		iput(req->r_old_dentry_dir);
861 	}
862 	kfree(req->r_path1);
863 	kfree(req->r_path2);
864 	put_cred(req->r_cred);
865 	if (req->r_pagelist)
866 		ceph_pagelist_release(req->r_pagelist);
867 	put_request_session(req);
868 	ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
869 	WARN_ON_ONCE(!list_empty(&req->r_wait));
870 	kmem_cache_free(ceph_mds_request_cachep, req);
871 }
872 
873 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
874 
875 /*
876  * lookup session, bump ref if found.
877  *
878  * called under mdsc->mutex.
879  */
880 static struct ceph_mds_request *
881 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
882 {
883 	struct ceph_mds_request *req;
884 
885 	req = lookup_request(&mdsc->request_tree, tid);
886 	if (req)
887 		ceph_mdsc_get_request(req);
888 
889 	return req;
890 }
891 
892 /*
893  * Register an in-flight request, and assign a tid.  Link to directory
894  * are modifying (if any).
895  *
896  * Called under mdsc->mutex.
897  */
898 static void __register_request(struct ceph_mds_client *mdsc,
899 			       struct ceph_mds_request *req,
900 			       struct inode *dir)
901 {
902 	int ret = 0;
903 
904 	req->r_tid = ++mdsc->last_tid;
905 	if (req->r_num_caps) {
906 		ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
907 					req->r_num_caps);
908 		if (ret < 0) {
909 			pr_err("__register_request %p "
910 			       "failed to reserve caps: %d\n", req, ret);
911 			/* set req->r_err to fail early from __do_request */
912 			req->r_err = ret;
913 			return;
914 		}
915 	}
916 	dout("__register_request %p tid %lld\n", req, req->r_tid);
917 	ceph_mdsc_get_request(req);
918 	insert_request(&mdsc->request_tree, req);
919 
920 	req->r_cred = get_current_cred();
921 
922 	if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
923 		mdsc->oldest_tid = req->r_tid;
924 
925 	if (dir) {
926 		struct ceph_inode_info *ci = ceph_inode(dir);
927 
928 		ihold(dir);
929 		req->r_unsafe_dir = dir;
930 		spin_lock(&ci->i_unsafe_lock);
931 		list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
932 		spin_unlock(&ci->i_unsafe_lock);
933 	}
934 }
935 
936 static void __unregister_request(struct ceph_mds_client *mdsc,
937 				 struct ceph_mds_request *req)
938 {
939 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
940 
941 	/* Never leave an unregistered request on an unsafe list! */
942 	list_del_init(&req->r_unsafe_item);
943 
944 	if (req->r_tid == mdsc->oldest_tid) {
945 		struct rb_node *p = rb_next(&req->r_node);
946 		mdsc->oldest_tid = 0;
947 		while (p) {
948 			struct ceph_mds_request *next_req =
949 				rb_entry(p, struct ceph_mds_request, r_node);
950 			if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
951 				mdsc->oldest_tid = next_req->r_tid;
952 				break;
953 			}
954 			p = rb_next(p);
955 		}
956 	}
957 
958 	erase_request(&mdsc->request_tree, req);
959 
960 	if (req->r_unsafe_dir) {
961 		struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
962 		spin_lock(&ci->i_unsafe_lock);
963 		list_del_init(&req->r_unsafe_dir_item);
964 		spin_unlock(&ci->i_unsafe_lock);
965 	}
966 	if (req->r_target_inode &&
967 	    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
968 		struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
969 		spin_lock(&ci->i_unsafe_lock);
970 		list_del_init(&req->r_unsafe_target_item);
971 		spin_unlock(&ci->i_unsafe_lock);
972 	}
973 
974 	if (req->r_unsafe_dir) {
975 		iput(req->r_unsafe_dir);
976 		req->r_unsafe_dir = NULL;
977 	}
978 
979 	complete_all(&req->r_safe_completion);
980 
981 	ceph_mdsc_put_request(req);
982 }
983 
984 /*
985  * Walk back up the dentry tree until we hit a dentry representing a
986  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
987  * when calling this) to ensure that the objects won't disappear while we're
988  * working with them. Once we hit a candidate dentry, we attempt to take a
989  * reference to it, and return that as the result.
990  */
991 static struct inode *get_nonsnap_parent(struct dentry *dentry)
992 {
993 	struct inode *inode = NULL;
994 
995 	while (dentry && !IS_ROOT(dentry)) {
996 		inode = d_inode_rcu(dentry);
997 		if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
998 			break;
999 		dentry = dentry->d_parent;
1000 	}
1001 	if (inode)
1002 		inode = igrab(inode);
1003 	return inode;
1004 }
1005 
1006 /*
1007  * Choose mds to send request to next.  If there is a hint set in the
1008  * request (e.g., due to a prior forward hint from the mds), use that.
1009  * Otherwise, consult frag tree and/or caps to identify the
1010  * appropriate mds.  If all else fails, choose randomly.
1011  *
1012  * Called under mdsc->mutex.
1013  */
1014 static int __choose_mds(struct ceph_mds_client *mdsc,
1015 			struct ceph_mds_request *req,
1016 			bool *random)
1017 {
1018 	struct inode *inode;
1019 	struct ceph_inode_info *ci;
1020 	struct ceph_cap *cap;
1021 	int mode = req->r_direct_mode;
1022 	int mds = -1;
1023 	u32 hash = req->r_direct_hash;
1024 	bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1025 
1026 	if (random)
1027 		*random = false;
1028 
1029 	/*
1030 	 * is there a specific mds we should try?  ignore hint if we have
1031 	 * no session and the mds is not up (active or recovering).
1032 	 */
1033 	if (req->r_resend_mds >= 0 &&
1034 	    (__have_session(mdsc, req->r_resend_mds) ||
1035 	     ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1036 		dout("%s using resend_mds mds%d\n", __func__,
1037 		     req->r_resend_mds);
1038 		return req->r_resend_mds;
1039 	}
1040 
1041 	if (mode == USE_RANDOM_MDS)
1042 		goto random;
1043 
1044 	inode = NULL;
1045 	if (req->r_inode) {
1046 		if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1047 			inode = req->r_inode;
1048 			ihold(inode);
1049 		} else {
1050 			/* req->r_dentry is non-null for LSSNAP request */
1051 			rcu_read_lock();
1052 			inode = get_nonsnap_parent(req->r_dentry);
1053 			rcu_read_unlock();
1054 			dout("%s using snapdir's parent %p\n", __func__, inode);
1055 		}
1056 	} else if (req->r_dentry) {
1057 		/* ignore race with rename; old or new d_parent is okay */
1058 		struct dentry *parent;
1059 		struct inode *dir;
1060 
1061 		rcu_read_lock();
1062 		parent = READ_ONCE(req->r_dentry->d_parent);
1063 		dir = req->r_parent ? : d_inode_rcu(parent);
1064 
1065 		if (!dir || dir->i_sb != mdsc->fsc->sb) {
1066 			/*  not this fs or parent went negative */
1067 			inode = d_inode(req->r_dentry);
1068 			if (inode)
1069 				ihold(inode);
1070 		} else if (ceph_snap(dir) != CEPH_NOSNAP) {
1071 			/* direct snapped/virtual snapdir requests
1072 			 * based on parent dir inode */
1073 			inode = get_nonsnap_parent(parent);
1074 			dout("%s using nonsnap parent %p\n", __func__, inode);
1075 		} else {
1076 			/* dentry target */
1077 			inode = d_inode(req->r_dentry);
1078 			if (!inode || mode == USE_AUTH_MDS) {
1079 				/* dir + name */
1080 				inode = igrab(dir);
1081 				hash = ceph_dentry_hash(dir, req->r_dentry);
1082 				is_hash = true;
1083 			} else {
1084 				ihold(inode);
1085 			}
1086 		}
1087 		rcu_read_unlock();
1088 	}
1089 
1090 	dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1091 	     hash, mode);
1092 	if (!inode)
1093 		goto random;
1094 	ci = ceph_inode(inode);
1095 
1096 	if (is_hash && S_ISDIR(inode->i_mode)) {
1097 		struct ceph_inode_frag frag;
1098 		int found;
1099 
1100 		ceph_choose_frag(ci, hash, &frag, &found);
1101 		if (found) {
1102 			if (mode == USE_ANY_MDS && frag.ndist > 0) {
1103 				u8 r;
1104 
1105 				/* choose a random replica */
1106 				get_random_bytes(&r, 1);
1107 				r %= frag.ndist;
1108 				mds = frag.dist[r];
1109 				dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1110 				     __func__, inode, ceph_vinop(inode),
1111 				     frag.frag, mds, (int)r, frag.ndist);
1112 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1113 				    CEPH_MDS_STATE_ACTIVE &&
1114 				    !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1115 					goto out;
1116 			}
1117 
1118 			/* since this file/dir wasn't known to be
1119 			 * replicated, then we want to look for the
1120 			 * authoritative mds. */
1121 			if (frag.mds >= 0) {
1122 				/* choose auth mds */
1123 				mds = frag.mds;
1124 				dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1125 				     __func__, inode, ceph_vinop(inode),
1126 				     frag.frag, mds);
1127 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1128 				    CEPH_MDS_STATE_ACTIVE) {
1129 					if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1130 								  mds))
1131 						goto out;
1132 				}
1133 			}
1134 			mode = USE_AUTH_MDS;
1135 		}
1136 	}
1137 
1138 	spin_lock(&ci->i_ceph_lock);
1139 	cap = NULL;
1140 	if (mode == USE_AUTH_MDS)
1141 		cap = ci->i_auth_cap;
1142 	if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1143 		cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1144 	if (!cap) {
1145 		spin_unlock(&ci->i_ceph_lock);
1146 		iput(inode);
1147 		goto random;
1148 	}
1149 	mds = cap->session->s_mds;
1150 	dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1151 	     inode, ceph_vinop(inode), mds,
1152 	     cap == ci->i_auth_cap ? "auth " : "", cap);
1153 	spin_unlock(&ci->i_ceph_lock);
1154 out:
1155 	iput(inode);
1156 	return mds;
1157 
1158 random:
1159 	if (random)
1160 		*random = true;
1161 
1162 	mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1163 	dout("%s chose random mds%d\n", __func__, mds);
1164 	return mds;
1165 }
1166 
1167 
1168 /*
1169  * session messages
1170  */
1171 struct ceph_msg *ceph_create_session_msg(u32 op, u64 seq)
1172 {
1173 	struct ceph_msg *msg;
1174 	struct ceph_mds_session_head *h;
1175 
1176 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1177 			   false);
1178 	if (!msg) {
1179 		pr_err("ENOMEM creating session %s msg\n",
1180 		       ceph_session_op_name(op));
1181 		return NULL;
1182 	}
1183 	h = msg->front.iov_base;
1184 	h->op = cpu_to_le32(op);
1185 	h->seq = cpu_to_le64(seq);
1186 
1187 	return msg;
1188 }
1189 
1190 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1191 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1192 static int encode_supported_features(void **p, void *end)
1193 {
1194 	static const size_t count = ARRAY_SIZE(feature_bits);
1195 
1196 	if (count > 0) {
1197 		size_t i;
1198 		size_t size = FEATURE_BYTES(count);
1199 
1200 		if (WARN_ON_ONCE(*p + 4 + size > end))
1201 			return -ERANGE;
1202 
1203 		ceph_encode_32(p, size);
1204 		memset(*p, 0, size);
1205 		for (i = 0; i < count; i++)
1206 			((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1207 		*p += size;
1208 	} else {
1209 		if (WARN_ON_ONCE(*p + 4 > end))
1210 			return -ERANGE;
1211 
1212 		ceph_encode_32(p, 0);
1213 	}
1214 
1215 	return 0;
1216 }
1217 
1218 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1219 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1220 static int encode_metric_spec(void **p, void *end)
1221 {
1222 	static const size_t count = ARRAY_SIZE(metric_bits);
1223 
1224 	/* header */
1225 	if (WARN_ON_ONCE(*p + 2 > end))
1226 		return -ERANGE;
1227 
1228 	ceph_encode_8(p, 1); /* version */
1229 	ceph_encode_8(p, 1); /* compat */
1230 
1231 	if (count > 0) {
1232 		size_t i;
1233 		size_t size = METRIC_BYTES(count);
1234 
1235 		if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1236 			return -ERANGE;
1237 
1238 		/* metric spec info length */
1239 		ceph_encode_32(p, 4 + size);
1240 
1241 		/* metric spec */
1242 		ceph_encode_32(p, size);
1243 		memset(*p, 0, size);
1244 		for (i = 0; i < count; i++)
1245 			((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1246 		*p += size;
1247 	} else {
1248 		if (WARN_ON_ONCE(*p + 4 + 4 > end))
1249 			return -ERANGE;
1250 
1251 		/* metric spec info length */
1252 		ceph_encode_32(p, 4);
1253 		/* metric spec */
1254 		ceph_encode_32(p, 0);
1255 	}
1256 
1257 	return 0;
1258 }
1259 
1260 /*
1261  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1262  * to include additional client metadata fields.
1263  */
1264 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1265 {
1266 	struct ceph_msg *msg;
1267 	struct ceph_mds_session_head *h;
1268 	int i;
1269 	int extra_bytes = 0;
1270 	int metadata_key_count = 0;
1271 	struct ceph_options *opt = mdsc->fsc->client->options;
1272 	struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1273 	size_t size, count;
1274 	void *p, *end;
1275 	int ret;
1276 
1277 	const char* metadata[][2] = {
1278 		{"hostname", mdsc->nodename},
1279 		{"kernel_version", init_utsname()->release},
1280 		{"entity_id", opt->name ? : ""},
1281 		{"root", fsopt->server_path ? : "/"},
1282 		{NULL, NULL}
1283 	};
1284 
1285 	/* Calculate serialized length of metadata */
1286 	extra_bytes = 4;  /* map length */
1287 	for (i = 0; metadata[i][0]; ++i) {
1288 		extra_bytes += 8 + strlen(metadata[i][0]) +
1289 			strlen(metadata[i][1]);
1290 		metadata_key_count++;
1291 	}
1292 
1293 	/* supported feature */
1294 	size = 0;
1295 	count = ARRAY_SIZE(feature_bits);
1296 	if (count > 0)
1297 		size = FEATURE_BYTES(count);
1298 	extra_bytes += 4 + size;
1299 
1300 	/* metric spec */
1301 	size = 0;
1302 	count = ARRAY_SIZE(metric_bits);
1303 	if (count > 0)
1304 		size = METRIC_BYTES(count);
1305 	extra_bytes += 2 + 4 + 4 + size;
1306 
1307 	/* Allocate the message */
1308 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1309 			   GFP_NOFS, false);
1310 	if (!msg) {
1311 		pr_err("ENOMEM creating session open msg\n");
1312 		return ERR_PTR(-ENOMEM);
1313 	}
1314 	p = msg->front.iov_base;
1315 	end = p + msg->front.iov_len;
1316 
1317 	h = p;
1318 	h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1319 	h->seq = cpu_to_le64(seq);
1320 
1321 	/*
1322 	 * Serialize client metadata into waiting buffer space, using
1323 	 * the format that userspace expects for map<string, string>
1324 	 *
1325 	 * ClientSession messages with metadata are v4
1326 	 */
1327 	msg->hdr.version = cpu_to_le16(4);
1328 	msg->hdr.compat_version = cpu_to_le16(1);
1329 
1330 	/* The write pointer, following the session_head structure */
1331 	p += sizeof(*h);
1332 
1333 	/* Number of entries in the map */
1334 	ceph_encode_32(&p, metadata_key_count);
1335 
1336 	/* Two length-prefixed strings for each entry in the map */
1337 	for (i = 0; metadata[i][0]; ++i) {
1338 		size_t const key_len = strlen(metadata[i][0]);
1339 		size_t const val_len = strlen(metadata[i][1]);
1340 
1341 		ceph_encode_32(&p, key_len);
1342 		memcpy(p, metadata[i][0], key_len);
1343 		p += key_len;
1344 		ceph_encode_32(&p, val_len);
1345 		memcpy(p, metadata[i][1], val_len);
1346 		p += val_len;
1347 	}
1348 
1349 	ret = encode_supported_features(&p, end);
1350 	if (ret) {
1351 		pr_err("encode_supported_features failed!\n");
1352 		ceph_msg_put(msg);
1353 		return ERR_PTR(ret);
1354 	}
1355 
1356 	ret = encode_metric_spec(&p, end);
1357 	if (ret) {
1358 		pr_err("encode_metric_spec failed!\n");
1359 		ceph_msg_put(msg);
1360 		return ERR_PTR(ret);
1361 	}
1362 
1363 	msg->front.iov_len = p - msg->front.iov_base;
1364 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1365 
1366 	return msg;
1367 }
1368 
1369 /*
1370  * send session open request.
1371  *
1372  * called under mdsc->mutex
1373  */
1374 static int __open_session(struct ceph_mds_client *mdsc,
1375 			  struct ceph_mds_session *session)
1376 {
1377 	struct ceph_msg *msg;
1378 	int mstate;
1379 	int mds = session->s_mds;
1380 
1381 	/* wait for mds to go active? */
1382 	mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1383 	dout("open_session to mds%d (%s)\n", mds,
1384 	     ceph_mds_state_name(mstate));
1385 	session->s_state = CEPH_MDS_SESSION_OPENING;
1386 	session->s_renew_requested = jiffies;
1387 
1388 	/* send connect message */
1389 	msg = create_session_open_msg(mdsc, session->s_seq);
1390 	if (IS_ERR(msg))
1391 		return PTR_ERR(msg);
1392 	ceph_con_send(&session->s_con, msg);
1393 	return 0;
1394 }
1395 
1396 /*
1397  * open sessions for any export targets for the given mds
1398  *
1399  * called under mdsc->mutex
1400  */
1401 static struct ceph_mds_session *
1402 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1403 {
1404 	struct ceph_mds_session *session;
1405 	int ret;
1406 
1407 	session = __ceph_lookup_mds_session(mdsc, target);
1408 	if (!session) {
1409 		session = register_session(mdsc, target);
1410 		if (IS_ERR(session))
1411 			return session;
1412 	}
1413 	if (session->s_state == CEPH_MDS_SESSION_NEW ||
1414 	    session->s_state == CEPH_MDS_SESSION_CLOSING) {
1415 		ret = __open_session(mdsc, session);
1416 		if (ret)
1417 			return ERR_PTR(ret);
1418 	}
1419 
1420 	return session;
1421 }
1422 
1423 struct ceph_mds_session *
1424 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1425 {
1426 	struct ceph_mds_session *session;
1427 
1428 	dout("open_export_target_session to mds%d\n", target);
1429 
1430 	mutex_lock(&mdsc->mutex);
1431 	session = __open_export_target_session(mdsc, target);
1432 	mutex_unlock(&mdsc->mutex);
1433 
1434 	return session;
1435 }
1436 
1437 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1438 					  struct ceph_mds_session *session)
1439 {
1440 	struct ceph_mds_info *mi;
1441 	struct ceph_mds_session *ts;
1442 	int i, mds = session->s_mds;
1443 
1444 	if (mds >= mdsc->mdsmap->possible_max_rank)
1445 		return;
1446 
1447 	mi = &mdsc->mdsmap->m_info[mds];
1448 	dout("open_export_target_sessions for mds%d (%d targets)\n",
1449 	     session->s_mds, mi->num_export_targets);
1450 
1451 	for (i = 0; i < mi->num_export_targets; i++) {
1452 		ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1453 		ceph_put_mds_session(ts);
1454 	}
1455 }
1456 
1457 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1458 					   struct ceph_mds_session *session)
1459 {
1460 	mutex_lock(&mdsc->mutex);
1461 	__open_export_target_sessions(mdsc, session);
1462 	mutex_unlock(&mdsc->mutex);
1463 }
1464 
1465 /*
1466  * session caps
1467  */
1468 
1469 static void detach_cap_releases(struct ceph_mds_session *session,
1470 				struct list_head *target)
1471 {
1472 	lockdep_assert_held(&session->s_cap_lock);
1473 
1474 	list_splice_init(&session->s_cap_releases, target);
1475 	session->s_num_cap_releases = 0;
1476 	dout("dispose_cap_releases mds%d\n", session->s_mds);
1477 }
1478 
1479 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1480 				 struct list_head *dispose)
1481 {
1482 	while (!list_empty(dispose)) {
1483 		struct ceph_cap *cap;
1484 		/* zero out the in-progress message */
1485 		cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1486 		list_del(&cap->session_caps);
1487 		ceph_put_cap(mdsc, cap);
1488 	}
1489 }
1490 
1491 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1492 				     struct ceph_mds_session *session)
1493 {
1494 	struct ceph_mds_request *req;
1495 	struct rb_node *p;
1496 
1497 	dout("cleanup_session_requests mds%d\n", session->s_mds);
1498 	mutex_lock(&mdsc->mutex);
1499 	while (!list_empty(&session->s_unsafe)) {
1500 		req = list_first_entry(&session->s_unsafe,
1501 				       struct ceph_mds_request, r_unsafe_item);
1502 		pr_warn_ratelimited(" dropping unsafe request %llu\n",
1503 				    req->r_tid);
1504 		if (req->r_target_inode)
1505 			mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1506 		if (req->r_unsafe_dir)
1507 			mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1508 		__unregister_request(mdsc, req);
1509 	}
1510 	/* zero r_attempts, so kick_requests() will re-send requests */
1511 	p = rb_first(&mdsc->request_tree);
1512 	while (p) {
1513 		req = rb_entry(p, struct ceph_mds_request, r_node);
1514 		p = rb_next(p);
1515 		if (req->r_session &&
1516 		    req->r_session->s_mds == session->s_mds)
1517 			req->r_attempts = 0;
1518 	}
1519 	mutex_unlock(&mdsc->mutex);
1520 }
1521 
1522 /*
1523  * Helper to safely iterate over all caps associated with a session, with
1524  * special care taken to handle a racing __ceph_remove_cap().
1525  *
1526  * Caller must hold session s_mutex.
1527  */
1528 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1529 			      int (*cb)(struct inode *, struct ceph_cap *,
1530 					void *), void *arg)
1531 {
1532 	struct list_head *p;
1533 	struct ceph_cap *cap;
1534 	struct inode *inode, *last_inode = NULL;
1535 	struct ceph_cap *old_cap = NULL;
1536 	int ret;
1537 
1538 	dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1539 	spin_lock(&session->s_cap_lock);
1540 	p = session->s_caps.next;
1541 	while (p != &session->s_caps) {
1542 		cap = list_entry(p, struct ceph_cap, session_caps);
1543 		inode = igrab(&cap->ci->vfs_inode);
1544 		if (!inode) {
1545 			p = p->next;
1546 			continue;
1547 		}
1548 		session->s_cap_iterator = cap;
1549 		spin_unlock(&session->s_cap_lock);
1550 
1551 		if (last_inode) {
1552 			iput(last_inode);
1553 			last_inode = NULL;
1554 		}
1555 		if (old_cap) {
1556 			ceph_put_cap(session->s_mdsc, old_cap);
1557 			old_cap = NULL;
1558 		}
1559 
1560 		ret = cb(inode, cap, arg);
1561 		last_inode = inode;
1562 
1563 		spin_lock(&session->s_cap_lock);
1564 		p = p->next;
1565 		if (!cap->ci) {
1566 			dout("iterate_session_caps  finishing cap %p removal\n",
1567 			     cap);
1568 			BUG_ON(cap->session != session);
1569 			cap->session = NULL;
1570 			list_del_init(&cap->session_caps);
1571 			session->s_nr_caps--;
1572 			atomic64_dec(&session->s_mdsc->metric.total_caps);
1573 			if (cap->queue_release)
1574 				__ceph_queue_cap_release(session, cap);
1575 			else
1576 				old_cap = cap;  /* put_cap it w/o locks held */
1577 		}
1578 		if (ret < 0)
1579 			goto out;
1580 	}
1581 	ret = 0;
1582 out:
1583 	session->s_cap_iterator = NULL;
1584 	spin_unlock(&session->s_cap_lock);
1585 
1586 	iput(last_inode);
1587 	if (old_cap)
1588 		ceph_put_cap(session->s_mdsc, old_cap);
1589 
1590 	return ret;
1591 }
1592 
1593 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1594 				  void *arg)
1595 {
1596 	struct ceph_inode_info *ci = ceph_inode(inode);
1597 	bool invalidate = false;
1598 	int iputs;
1599 
1600 	dout("removing cap %p, ci is %p, inode is %p\n",
1601 	     cap, ci, &ci->vfs_inode);
1602 	spin_lock(&ci->i_ceph_lock);
1603 	iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1604 	spin_unlock(&ci->i_ceph_lock);
1605 
1606 	wake_up_all(&ci->i_cap_wq);
1607 	if (invalidate)
1608 		ceph_queue_invalidate(inode);
1609 	while (iputs--)
1610 		iput(inode);
1611 	return 0;
1612 }
1613 
1614 /*
1615  * caller must hold session s_mutex
1616  */
1617 static void remove_session_caps(struct ceph_mds_session *session)
1618 {
1619 	struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1620 	struct super_block *sb = fsc->sb;
1621 	LIST_HEAD(dispose);
1622 
1623 	dout("remove_session_caps on %p\n", session);
1624 	ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1625 
1626 	wake_up_all(&fsc->mdsc->cap_flushing_wq);
1627 
1628 	spin_lock(&session->s_cap_lock);
1629 	if (session->s_nr_caps > 0) {
1630 		struct inode *inode;
1631 		struct ceph_cap *cap, *prev = NULL;
1632 		struct ceph_vino vino;
1633 		/*
1634 		 * iterate_session_caps() skips inodes that are being
1635 		 * deleted, we need to wait until deletions are complete.
1636 		 * __wait_on_freeing_inode() is designed for the job,
1637 		 * but it is not exported, so use lookup inode function
1638 		 * to access it.
1639 		 */
1640 		while (!list_empty(&session->s_caps)) {
1641 			cap = list_entry(session->s_caps.next,
1642 					 struct ceph_cap, session_caps);
1643 			if (cap == prev)
1644 				break;
1645 			prev = cap;
1646 			vino = cap->ci->i_vino;
1647 			spin_unlock(&session->s_cap_lock);
1648 
1649 			inode = ceph_find_inode(sb, vino);
1650 			iput(inode);
1651 
1652 			spin_lock(&session->s_cap_lock);
1653 		}
1654 	}
1655 
1656 	// drop cap expires and unlock s_cap_lock
1657 	detach_cap_releases(session, &dispose);
1658 
1659 	BUG_ON(session->s_nr_caps > 0);
1660 	BUG_ON(!list_empty(&session->s_cap_flushing));
1661 	spin_unlock(&session->s_cap_lock);
1662 	dispose_cap_releases(session->s_mdsc, &dispose);
1663 }
1664 
1665 enum {
1666 	RECONNECT,
1667 	RENEWCAPS,
1668 	FORCE_RO,
1669 };
1670 
1671 /*
1672  * wake up any threads waiting on this session's caps.  if the cap is
1673  * old (didn't get renewed on the client reconnect), remove it now.
1674  *
1675  * caller must hold s_mutex.
1676  */
1677 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1678 			      void *arg)
1679 {
1680 	struct ceph_inode_info *ci = ceph_inode(inode);
1681 	unsigned long ev = (unsigned long)arg;
1682 
1683 	if (ev == RECONNECT) {
1684 		spin_lock(&ci->i_ceph_lock);
1685 		ci->i_wanted_max_size = 0;
1686 		ci->i_requested_max_size = 0;
1687 		spin_unlock(&ci->i_ceph_lock);
1688 	} else if (ev == RENEWCAPS) {
1689 		if (cap->cap_gen < atomic_read(&cap->session->s_cap_gen)) {
1690 			/* mds did not re-issue stale cap */
1691 			spin_lock(&ci->i_ceph_lock);
1692 			cap->issued = cap->implemented = CEPH_CAP_PIN;
1693 			spin_unlock(&ci->i_ceph_lock);
1694 		}
1695 	} else if (ev == FORCE_RO) {
1696 	}
1697 	wake_up_all(&ci->i_cap_wq);
1698 	return 0;
1699 }
1700 
1701 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1702 {
1703 	dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1704 	ceph_iterate_session_caps(session, wake_up_session_cb,
1705 				  (void *)(unsigned long)ev);
1706 }
1707 
1708 /*
1709  * Send periodic message to MDS renewing all currently held caps.  The
1710  * ack will reset the expiration for all caps from this session.
1711  *
1712  * caller holds s_mutex
1713  */
1714 static int send_renew_caps(struct ceph_mds_client *mdsc,
1715 			   struct ceph_mds_session *session)
1716 {
1717 	struct ceph_msg *msg;
1718 	int state;
1719 
1720 	if (time_after_eq(jiffies, session->s_cap_ttl) &&
1721 	    time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1722 		pr_info("mds%d caps stale\n", session->s_mds);
1723 	session->s_renew_requested = jiffies;
1724 
1725 	/* do not try to renew caps until a recovering mds has reconnected
1726 	 * with its clients. */
1727 	state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1728 	if (state < CEPH_MDS_STATE_RECONNECT) {
1729 		dout("send_renew_caps ignoring mds%d (%s)\n",
1730 		     session->s_mds, ceph_mds_state_name(state));
1731 		return 0;
1732 	}
1733 
1734 	dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1735 		ceph_mds_state_name(state));
1736 	msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1737 				      ++session->s_renew_seq);
1738 	if (!msg)
1739 		return -ENOMEM;
1740 	ceph_con_send(&session->s_con, msg);
1741 	return 0;
1742 }
1743 
1744 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1745 			     struct ceph_mds_session *session, u64 seq)
1746 {
1747 	struct ceph_msg *msg;
1748 
1749 	dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1750 	     session->s_mds, ceph_session_state_name(session->s_state), seq);
1751 	msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1752 	if (!msg)
1753 		return -ENOMEM;
1754 	ceph_con_send(&session->s_con, msg);
1755 	return 0;
1756 }
1757 
1758 
1759 /*
1760  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1761  *
1762  * Called under session->s_mutex
1763  */
1764 static void renewed_caps(struct ceph_mds_client *mdsc,
1765 			 struct ceph_mds_session *session, int is_renew)
1766 {
1767 	int was_stale;
1768 	int wake = 0;
1769 
1770 	spin_lock(&session->s_cap_lock);
1771 	was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1772 
1773 	session->s_cap_ttl = session->s_renew_requested +
1774 		mdsc->mdsmap->m_session_timeout*HZ;
1775 
1776 	if (was_stale) {
1777 		if (time_before(jiffies, session->s_cap_ttl)) {
1778 			pr_info("mds%d caps renewed\n", session->s_mds);
1779 			wake = 1;
1780 		} else {
1781 			pr_info("mds%d caps still stale\n", session->s_mds);
1782 		}
1783 	}
1784 	dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1785 	     session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1786 	     time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1787 	spin_unlock(&session->s_cap_lock);
1788 
1789 	if (wake)
1790 		wake_up_session_caps(session, RENEWCAPS);
1791 }
1792 
1793 /*
1794  * send a session close request
1795  */
1796 static int request_close_session(struct ceph_mds_session *session)
1797 {
1798 	struct ceph_msg *msg;
1799 
1800 	dout("request_close_session mds%d state %s seq %lld\n",
1801 	     session->s_mds, ceph_session_state_name(session->s_state),
1802 	     session->s_seq);
1803 	msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
1804 				      session->s_seq);
1805 	if (!msg)
1806 		return -ENOMEM;
1807 	ceph_con_send(&session->s_con, msg);
1808 	return 1;
1809 }
1810 
1811 /*
1812  * Called with s_mutex held.
1813  */
1814 static int __close_session(struct ceph_mds_client *mdsc,
1815 			 struct ceph_mds_session *session)
1816 {
1817 	if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1818 		return 0;
1819 	session->s_state = CEPH_MDS_SESSION_CLOSING;
1820 	return request_close_session(session);
1821 }
1822 
1823 static bool drop_negative_children(struct dentry *dentry)
1824 {
1825 	struct dentry *child;
1826 	bool all_negative = true;
1827 
1828 	if (!d_is_dir(dentry))
1829 		goto out;
1830 
1831 	spin_lock(&dentry->d_lock);
1832 	list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1833 		if (d_really_is_positive(child)) {
1834 			all_negative = false;
1835 			break;
1836 		}
1837 	}
1838 	spin_unlock(&dentry->d_lock);
1839 
1840 	if (all_negative)
1841 		shrink_dcache_parent(dentry);
1842 out:
1843 	return all_negative;
1844 }
1845 
1846 /*
1847  * Trim old(er) caps.
1848  *
1849  * Because we can't cache an inode without one or more caps, we do
1850  * this indirectly: if a cap is unused, we prune its aliases, at which
1851  * point the inode will hopefully get dropped to.
1852  *
1853  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1854  * memory pressure from the MDS, though, so it needn't be perfect.
1855  */
1856 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1857 {
1858 	int *remaining = arg;
1859 	struct ceph_inode_info *ci = ceph_inode(inode);
1860 	int used, wanted, oissued, mine;
1861 
1862 	if (*remaining <= 0)
1863 		return -1;
1864 
1865 	spin_lock(&ci->i_ceph_lock);
1866 	mine = cap->issued | cap->implemented;
1867 	used = __ceph_caps_used(ci);
1868 	wanted = __ceph_caps_file_wanted(ci);
1869 	oissued = __ceph_caps_issued_other(ci, cap);
1870 
1871 	dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1872 	     inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1873 	     ceph_cap_string(used), ceph_cap_string(wanted));
1874 	if (cap == ci->i_auth_cap) {
1875 		if (ci->i_dirty_caps || ci->i_flushing_caps ||
1876 		    !list_empty(&ci->i_cap_snaps))
1877 			goto out;
1878 		if ((used | wanted) & CEPH_CAP_ANY_WR)
1879 			goto out;
1880 		/* Note: it's possible that i_filelock_ref becomes non-zero
1881 		 * after dropping auth caps. It doesn't hurt because reply
1882 		 * of lock mds request will re-add auth caps. */
1883 		if (atomic_read(&ci->i_filelock_ref) > 0)
1884 			goto out;
1885 	}
1886 	/* The inode has cached pages, but it's no longer used.
1887 	 * we can safely drop it */
1888 	if (S_ISREG(inode->i_mode) &&
1889 	    wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1890 	    !(oissued & CEPH_CAP_FILE_CACHE)) {
1891 	  used = 0;
1892 	  oissued = 0;
1893 	}
1894 	if ((used | wanted) & ~oissued & mine)
1895 		goto out;   /* we need these caps */
1896 
1897 	if (oissued) {
1898 		/* we aren't the only cap.. just remove us */
1899 		ceph_remove_cap(cap, true);
1900 		(*remaining)--;
1901 	} else {
1902 		struct dentry *dentry;
1903 		/* try dropping referring dentries */
1904 		spin_unlock(&ci->i_ceph_lock);
1905 		dentry = d_find_any_alias(inode);
1906 		if (dentry && drop_negative_children(dentry)) {
1907 			int count;
1908 			dput(dentry);
1909 			d_prune_aliases(inode);
1910 			count = atomic_read(&inode->i_count);
1911 			if (count == 1)
1912 				(*remaining)--;
1913 			dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1914 			     inode, cap, count);
1915 		} else {
1916 			dput(dentry);
1917 		}
1918 		return 0;
1919 	}
1920 
1921 out:
1922 	spin_unlock(&ci->i_ceph_lock);
1923 	return 0;
1924 }
1925 
1926 /*
1927  * Trim session cap count down to some max number.
1928  */
1929 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1930 		   struct ceph_mds_session *session,
1931 		   int max_caps)
1932 {
1933 	int trim_caps = session->s_nr_caps - max_caps;
1934 
1935 	dout("trim_caps mds%d start: %d / %d, trim %d\n",
1936 	     session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1937 	if (trim_caps > 0) {
1938 		int remaining = trim_caps;
1939 
1940 		ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1941 		dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1942 		     session->s_mds, session->s_nr_caps, max_caps,
1943 			trim_caps - remaining);
1944 	}
1945 
1946 	ceph_flush_cap_releases(mdsc, session);
1947 	return 0;
1948 }
1949 
1950 static int check_caps_flush(struct ceph_mds_client *mdsc,
1951 			    u64 want_flush_tid)
1952 {
1953 	int ret = 1;
1954 
1955 	spin_lock(&mdsc->cap_dirty_lock);
1956 	if (!list_empty(&mdsc->cap_flush_list)) {
1957 		struct ceph_cap_flush *cf =
1958 			list_first_entry(&mdsc->cap_flush_list,
1959 					 struct ceph_cap_flush, g_list);
1960 		if (cf->tid <= want_flush_tid) {
1961 			dout("check_caps_flush still flushing tid "
1962 			     "%llu <= %llu\n", cf->tid, want_flush_tid);
1963 			ret = 0;
1964 		}
1965 	}
1966 	spin_unlock(&mdsc->cap_dirty_lock);
1967 	return ret;
1968 }
1969 
1970 /*
1971  * flush all dirty inode data to disk.
1972  *
1973  * returns true if we've flushed through want_flush_tid
1974  */
1975 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1976 			    u64 want_flush_tid)
1977 {
1978 	dout("check_caps_flush want %llu\n", want_flush_tid);
1979 
1980 	wait_event(mdsc->cap_flushing_wq,
1981 		   check_caps_flush(mdsc, want_flush_tid));
1982 
1983 	dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1984 }
1985 
1986 /*
1987  * called under s_mutex
1988  */
1989 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1990 				   struct ceph_mds_session *session)
1991 {
1992 	struct ceph_msg *msg = NULL;
1993 	struct ceph_mds_cap_release *head;
1994 	struct ceph_mds_cap_item *item;
1995 	struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1996 	struct ceph_cap *cap;
1997 	LIST_HEAD(tmp_list);
1998 	int num_cap_releases;
1999 	__le32	barrier, *cap_barrier;
2000 
2001 	down_read(&osdc->lock);
2002 	barrier = cpu_to_le32(osdc->epoch_barrier);
2003 	up_read(&osdc->lock);
2004 
2005 	spin_lock(&session->s_cap_lock);
2006 again:
2007 	list_splice_init(&session->s_cap_releases, &tmp_list);
2008 	num_cap_releases = session->s_num_cap_releases;
2009 	session->s_num_cap_releases = 0;
2010 	spin_unlock(&session->s_cap_lock);
2011 
2012 	while (!list_empty(&tmp_list)) {
2013 		if (!msg) {
2014 			msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2015 					PAGE_SIZE, GFP_NOFS, false);
2016 			if (!msg)
2017 				goto out_err;
2018 			head = msg->front.iov_base;
2019 			head->num = cpu_to_le32(0);
2020 			msg->front.iov_len = sizeof(*head);
2021 
2022 			msg->hdr.version = cpu_to_le16(2);
2023 			msg->hdr.compat_version = cpu_to_le16(1);
2024 		}
2025 
2026 		cap = list_first_entry(&tmp_list, struct ceph_cap,
2027 					session_caps);
2028 		list_del(&cap->session_caps);
2029 		num_cap_releases--;
2030 
2031 		head = msg->front.iov_base;
2032 		put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2033 				   &head->num);
2034 		item = msg->front.iov_base + msg->front.iov_len;
2035 		item->ino = cpu_to_le64(cap->cap_ino);
2036 		item->cap_id = cpu_to_le64(cap->cap_id);
2037 		item->migrate_seq = cpu_to_le32(cap->mseq);
2038 		item->seq = cpu_to_le32(cap->issue_seq);
2039 		msg->front.iov_len += sizeof(*item);
2040 
2041 		ceph_put_cap(mdsc, cap);
2042 
2043 		if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2044 			// Append cap_barrier field
2045 			cap_barrier = msg->front.iov_base + msg->front.iov_len;
2046 			*cap_barrier = barrier;
2047 			msg->front.iov_len += sizeof(*cap_barrier);
2048 
2049 			msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2050 			dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2051 			ceph_con_send(&session->s_con, msg);
2052 			msg = NULL;
2053 		}
2054 	}
2055 
2056 	BUG_ON(num_cap_releases != 0);
2057 
2058 	spin_lock(&session->s_cap_lock);
2059 	if (!list_empty(&session->s_cap_releases))
2060 		goto again;
2061 	spin_unlock(&session->s_cap_lock);
2062 
2063 	if (msg) {
2064 		// Append cap_barrier field
2065 		cap_barrier = msg->front.iov_base + msg->front.iov_len;
2066 		*cap_barrier = barrier;
2067 		msg->front.iov_len += sizeof(*cap_barrier);
2068 
2069 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2070 		dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2071 		ceph_con_send(&session->s_con, msg);
2072 	}
2073 	return;
2074 out_err:
2075 	pr_err("send_cap_releases mds%d, failed to allocate message\n",
2076 		session->s_mds);
2077 	spin_lock(&session->s_cap_lock);
2078 	list_splice(&tmp_list, &session->s_cap_releases);
2079 	session->s_num_cap_releases += num_cap_releases;
2080 	spin_unlock(&session->s_cap_lock);
2081 }
2082 
2083 static void ceph_cap_release_work(struct work_struct *work)
2084 {
2085 	struct ceph_mds_session *session =
2086 		container_of(work, struct ceph_mds_session, s_cap_release_work);
2087 
2088 	mutex_lock(&session->s_mutex);
2089 	if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2090 	    session->s_state == CEPH_MDS_SESSION_HUNG)
2091 		ceph_send_cap_releases(session->s_mdsc, session);
2092 	mutex_unlock(&session->s_mutex);
2093 	ceph_put_mds_session(session);
2094 }
2095 
2096 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2097 		             struct ceph_mds_session *session)
2098 {
2099 	if (mdsc->stopping)
2100 		return;
2101 
2102 	ceph_get_mds_session(session);
2103 	if (queue_work(mdsc->fsc->cap_wq,
2104 		       &session->s_cap_release_work)) {
2105 		dout("cap release work queued\n");
2106 	} else {
2107 		ceph_put_mds_session(session);
2108 		dout("failed to queue cap release work\n");
2109 	}
2110 }
2111 
2112 /*
2113  * caller holds session->s_cap_lock
2114  */
2115 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2116 			      struct ceph_cap *cap)
2117 {
2118 	list_add_tail(&cap->session_caps, &session->s_cap_releases);
2119 	session->s_num_cap_releases++;
2120 
2121 	if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2122 		ceph_flush_cap_releases(session->s_mdsc, session);
2123 }
2124 
2125 static void ceph_cap_reclaim_work(struct work_struct *work)
2126 {
2127 	struct ceph_mds_client *mdsc =
2128 		container_of(work, struct ceph_mds_client, cap_reclaim_work);
2129 	int ret = ceph_trim_dentries(mdsc);
2130 	if (ret == -EAGAIN)
2131 		ceph_queue_cap_reclaim_work(mdsc);
2132 }
2133 
2134 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2135 {
2136 	if (mdsc->stopping)
2137 		return;
2138 
2139         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2140                 dout("caps reclaim work queued\n");
2141         } else {
2142                 dout("failed to queue caps release work\n");
2143         }
2144 }
2145 
2146 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2147 {
2148 	int val;
2149 	if (!nr)
2150 		return;
2151 	val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2152 	if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2153 		atomic_set(&mdsc->cap_reclaim_pending, 0);
2154 		ceph_queue_cap_reclaim_work(mdsc);
2155 	}
2156 }
2157 
2158 /*
2159  * requests
2160  */
2161 
2162 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2163 				    struct inode *dir)
2164 {
2165 	struct ceph_inode_info *ci = ceph_inode(dir);
2166 	struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2167 	struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2168 	size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2169 	unsigned int num_entries;
2170 	int order;
2171 
2172 	spin_lock(&ci->i_ceph_lock);
2173 	num_entries = ci->i_files + ci->i_subdirs;
2174 	spin_unlock(&ci->i_ceph_lock);
2175 	num_entries = max(num_entries, 1U);
2176 	num_entries = min(num_entries, opt->max_readdir);
2177 
2178 	order = get_order(size * num_entries);
2179 	while (order >= 0) {
2180 		rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2181 							     __GFP_NOWARN,
2182 							     order);
2183 		if (rinfo->dir_entries)
2184 			break;
2185 		order--;
2186 	}
2187 	if (!rinfo->dir_entries)
2188 		return -ENOMEM;
2189 
2190 	num_entries = (PAGE_SIZE << order) / size;
2191 	num_entries = min(num_entries, opt->max_readdir);
2192 
2193 	rinfo->dir_buf_size = PAGE_SIZE << order;
2194 	req->r_num_caps = num_entries + 1;
2195 	req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2196 	req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2197 	return 0;
2198 }
2199 
2200 /*
2201  * Create an mds request.
2202  */
2203 struct ceph_mds_request *
2204 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2205 {
2206 	struct ceph_mds_request *req;
2207 
2208 	req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2209 	if (!req)
2210 		return ERR_PTR(-ENOMEM);
2211 
2212 	mutex_init(&req->r_fill_mutex);
2213 	req->r_mdsc = mdsc;
2214 	req->r_started = jiffies;
2215 	req->r_start_latency = ktime_get();
2216 	req->r_resend_mds = -1;
2217 	INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2218 	INIT_LIST_HEAD(&req->r_unsafe_target_item);
2219 	req->r_fmode = -1;
2220 	kref_init(&req->r_kref);
2221 	RB_CLEAR_NODE(&req->r_node);
2222 	INIT_LIST_HEAD(&req->r_wait);
2223 	init_completion(&req->r_completion);
2224 	init_completion(&req->r_safe_completion);
2225 	INIT_LIST_HEAD(&req->r_unsafe_item);
2226 
2227 	ktime_get_coarse_real_ts64(&req->r_stamp);
2228 
2229 	req->r_op = op;
2230 	req->r_direct_mode = mode;
2231 	return req;
2232 }
2233 
2234 /*
2235  * return oldest (lowest) request, tid in request tree, 0 if none.
2236  *
2237  * called under mdsc->mutex.
2238  */
2239 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2240 {
2241 	if (RB_EMPTY_ROOT(&mdsc->request_tree))
2242 		return NULL;
2243 	return rb_entry(rb_first(&mdsc->request_tree),
2244 			struct ceph_mds_request, r_node);
2245 }
2246 
2247 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2248 {
2249 	return mdsc->oldest_tid;
2250 }
2251 
2252 /*
2253  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2254  * on build_path_from_dentry in fs/cifs/dir.c.
2255  *
2256  * If @stop_on_nosnap, generate path relative to the first non-snapped
2257  * inode.
2258  *
2259  * Encode hidden .snap dirs as a double /, i.e.
2260  *   foo/.snap/bar -> foo//bar
2261  */
2262 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2263 			   int stop_on_nosnap)
2264 {
2265 	struct dentry *temp;
2266 	char *path;
2267 	int pos;
2268 	unsigned seq;
2269 	u64 base;
2270 
2271 	if (!dentry)
2272 		return ERR_PTR(-EINVAL);
2273 
2274 	path = __getname();
2275 	if (!path)
2276 		return ERR_PTR(-ENOMEM);
2277 retry:
2278 	pos = PATH_MAX - 1;
2279 	path[pos] = '\0';
2280 
2281 	seq = read_seqbegin(&rename_lock);
2282 	rcu_read_lock();
2283 	temp = dentry;
2284 	for (;;) {
2285 		struct inode *inode;
2286 
2287 		spin_lock(&temp->d_lock);
2288 		inode = d_inode(temp);
2289 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2290 			dout("build_path path+%d: %p SNAPDIR\n",
2291 			     pos, temp);
2292 		} else if (stop_on_nosnap && inode && dentry != temp &&
2293 			   ceph_snap(inode) == CEPH_NOSNAP) {
2294 			spin_unlock(&temp->d_lock);
2295 			pos++; /* get rid of any prepended '/' */
2296 			break;
2297 		} else {
2298 			pos -= temp->d_name.len;
2299 			if (pos < 0) {
2300 				spin_unlock(&temp->d_lock);
2301 				break;
2302 			}
2303 			memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2304 		}
2305 		spin_unlock(&temp->d_lock);
2306 		temp = READ_ONCE(temp->d_parent);
2307 
2308 		/* Are we at the root? */
2309 		if (IS_ROOT(temp))
2310 			break;
2311 
2312 		/* Are we out of buffer? */
2313 		if (--pos < 0)
2314 			break;
2315 
2316 		path[pos] = '/';
2317 	}
2318 	base = ceph_ino(d_inode(temp));
2319 	rcu_read_unlock();
2320 
2321 	if (read_seqretry(&rename_lock, seq))
2322 		goto retry;
2323 
2324 	if (pos < 0) {
2325 		/*
2326 		 * A rename didn't occur, but somehow we didn't end up where
2327 		 * we thought we would. Throw a warning and try again.
2328 		 */
2329 		pr_warn("build_path did not end path lookup where "
2330 			"expected, pos is %d\n", pos);
2331 		goto retry;
2332 	}
2333 
2334 	*pbase = base;
2335 	*plen = PATH_MAX - 1 - pos;
2336 	dout("build_path on %p %d built %llx '%.*s'\n",
2337 	     dentry, d_count(dentry), base, *plen, path + pos);
2338 	return path + pos;
2339 }
2340 
2341 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2342 			     const char **ppath, int *ppathlen, u64 *pino,
2343 			     bool *pfreepath, bool parent_locked)
2344 {
2345 	char *path;
2346 
2347 	rcu_read_lock();
2348 	if (!dir)
2349 		dir = d_inode_rcu(dentry->d_parent);
2350 	if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2351 		*pino = ceph_ino(dir);
2352 		rcu_read_unlock();
2353 		*ppath = dentry->d_name.name;
2354 		*ppathlen = dentry->d_name.len;
2355 		return 0;
2356 	}
2357 	rcu_read_unlock();
2358 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2359 	if (IS_ERR(path))
2360 		return PTR_ERR(path);
2361 	*ppath = path;
2362 	*pfreepath = true;
2363 	return 0;
2364 }
2365 
2366 static int build_inode_path(struct inode *inode,
2367 			    const char **ppath, int *ppathlen, u64 *pino,
2368 			    bool *pfreepath)
2369 {
2370 	struct dentry *dentry;
2371 	char *path;
2372 
2373 	if (ceph_snap(inode) == CEPH_NOSNAP) {
2374 		*pino = ceph_ino(inode);
2375 		*ppathlen = 0;
2376 		return 0;
2377 	}
2378 	dentry = d_find_alias(inode);
2379 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2380 	dput(dentry);
2381 	if (IS_ERR(path))
2382 		return PTR_ERR(path);
2383 	*ppath = path;
2384 	*pfreepath = true;
2385 	return 0;
2386 }
2387 
2388 /*
2389  * request arguments may be specified via an inode *, a dentry *, or
2390  * an explicit ino+path.
2391  */
2392 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2393 				  struct inode *rdiri, const char *rpath,
2394 				  u64 rino, const char **ppath, int *pathlen,
2395 				  u64 *ino, bool *freepath, bool parent_locked)
2396 {
2397 	int r = 0;
2398 
2399 	if (rinode) {
2400 		r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2401 		dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2402 		     ceph_snap(rinode));
2403 	} else if (rdentry) {
2404 		r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2405 					freepath, parent_locked);
2406 		dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2407 		     *ppath);
2408 	} else if (rpath || rino) {
2409 		*ino = rino;
2410 		*ppath = rpath;
2411 		*pathlen = rpath ? strlen(rpath) : 0;
2412 		dout(" path %.*s\n", *pathlen, rpath);
2413 	}
2414 
2415 	return r;
2416 }
2417 
2418 static void encode_timestamp_and_gids(void **p,
2419 				      const struct ceph_mds_request *req)
2420 {
2421 	struct ceph_timespec ts;
2422 	int i;
2423 
2424 	ceph_encode_timespec64(&ts, &req->r_stamp);
2425 	ceph_encode_copy(p, &ts, sizeof(ts));
2426 
2427 	/* gid_list */
2428 	ceph_encode_32(p, req->r_cred->group_info->ngroups);
2429 	for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2430 		ceph_encode_64(p, from_kgid(&init_user_ns,
2431 					    req->r_cred->group_info->gid[i]));
2432 }
2433 
2434 /*
2435  * called under mdsc->mutex
2436  */
2437 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2438 					       struct ceph_mds_request *req,
2439 					       bool drop_cap_releases)
2440 {
2441 	int mds = session->s_mds;
2442 	struct ceph_mds_client *mdsc = session->s_mdsc;
2443 	struct ceph_msg *msg;
2444 	struct ceph_mds_request_head_old *head;
2445 	const char *path1 = NULL;
2446 	const char *path2 = NULL;
2447 	u64 ino1 = 0, ino2 = 0;
2448 	int pathlen1 = 0, pathlen2 = 0;
2449 	bool freepath1 = false, freepath2 = false;
2450 	int len;
2451 	u16 releases;
2452 	void *p, *end;
2453 	int ret;
2454 	bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2455 
2456 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
2457 			      req->r_parent, req->r_path1, req->r_ino1.ino,
2458 			      &path1, &pathlen1, &ino1, &freepath1,
2459 			      test_bit(CEPH_MDS_R_PARENT_LOCKED,
2460 					&req->r_req_flags));
2461 	if (ret < 0) {
2462 		msg = ERR_PTR(ret);
2463 		goto out;
2464 	}
2465 
2466 	/* If r_old_dentry is set, then assume that its parent is locked */
2467 	ret = set_request_path_attr(NULL, req->r_old_dentry,
2468 			      req->r_old_dentry_dir,
2469 			      req->r_path2, req->r_ino2.ino,
2470 			      &path2, &pathlen2, &ino2, &freepath2, true);
2471 	if (ret < 0) {
2472 		msg = ERR_PTR(ret);
2473 		goto out_free1;
2474 	}
2475 
2476 	len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2477 	len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2478 		sizeof(struct ceph_timespec);
2479 	len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2480 
2481 	/* calculate (max) length for cap releases */
2482 	len += sizeof(struct ceph_mds_request_release) *
2483 		(!!req->r_inode_drop + !!req->r_dentry_drop +
2484 		 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2485 
2486 	if (req->r_dentry_drop)
2487 		len += pathlen1;
2488 	if (req->r_old_dentry_drop)
2489 		len += pathlen2;
2490 
2491 	msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2492 	if (!msg) {
2493 		msg = ERR_PTR(-ENOMEM);
2494 		goto out_free2;
2495 	}
2496 
2497 	msg->hdr.tid = cpu_to_le64(req->r_tid);
2498 
2499 	/*
2500 	 * The old ceph_mds_request_head didn't contain a version field, and
2501 	 * one was added when we moved the message version from 3->4.
2502 	 */
2503 	if (legacy) {
2504 		msg->hdr.version = cpu_to_le16(3);
2505 		head = msg->front.iov_base;
2506 		p = msg->front.iov_base + sizeof(*head);
2507 	} else {
2508 		struct ceph_mds_request_head *new_head = msg->front.iov_base;
2509 
2510 		msg->hdr.version = cpu_to_le16(4);
2511 		new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2512 		head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2513 		p = msg->front.iov_base + sizeof(*new_head);
2514 	}
2515 
2516 	end = msg->front.iov_base + msg->front.iov_len;
2517 
2518 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2519 	head->op = cpu_to_le32(req->r_op);
2520 	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2521 						 req->r_cred->fsuid));
2522 	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2523 						 req->r_cred->fsgid));
2524 	head->ino = cpu_to_le64(req->r_deleg_ino);
2525 	head->args = req->r_args;
2526 
2527 	ceph_encode_filepath(&p, end, ino1, path1);
2528 	ceph_encode_filepath(&p, end, ino2, path2);
2529 
2530 	/* make note of release offset, in case we need to replay */
2531 	req->r_request_release_offset = p - msg->front.iov_base;
2532 
2533 	/* cap releases */
2534 	releases = 0;
2535 	if (req->r_inode_drop)
2536 		releases += ceph_encode_inode_release(&p,
2537 		      req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2538 		      mds, req->r_inode_drop, req->r_inode_unless,
2539 		      req->r_op == CEPH_MDS_OP_READDIR);
2540 	if (req->r_dentry_drop)
2541 		releases += ceph_encode_dentry_release(&p, req->r_dentry,
2542 				req->r_parent, mds, req->r_dentry_drop,
2543 				req->r_dentry_unless);
2544 	if (req->r_old_dentry_drop)
2545 		releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2546 				req->r_old_dentry_dir, mds,
2547 				req->r_old_dentry_drop,
2548 				req->r_old_dentry_unless);
2549 	if (req->r_old_inode_drop)
2550 		releases += ceph_encode_inode_release(&p,
2551 		      d_inode(req->r_old_dentry),
2552 		      mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2553 
2554 	if (drop_cap_releases) {
2555 		releases = 0;
2556 		p = msg->front.iov_base + req->r_request_release_offset;
2557 	}
2558 
2559 	head->num_releases = cpu_to_le16(releases);
2560 
2561 	encode_timestamp_and_gids(&p, req);
2562 
2563 	if (WARN_ON_ONCE(p > end)) {
2564 		ceph_msg_put(msg);
2565 		msg = ERR_PTR(-ERANGE);
2566 		goto out_free2;
2567 	}
2568 
2569 	msg->front.iov_len = p - msg->front.iov_base;
2570 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2571 
2572 	if (req->r_pagelist) {
2573 		struct ceph_pagelist *pagelist = req->r_pagelist;
2574 		ceph_msg_data_add_pagelist(msg, pagelist);
2575 		msg->hdr.data_len = cpu_to_le32(pagelist->length);
2576 	} else {
2577 		msg->hdr.data_len = 0;
2578 	}
2579 
2580 	msg->hdr.data_off = cpu_to_le16(0);
2581 
2582 out_free2:
2583 	if (freepath2)
2584 		ceph_mdsc_free_path((char *)path2, pathlen2);
2585 out_free1:
2586 	if (freepath1)
2587 		ceph_mdsc_free_path((char *)path1, pathlen1);
2588 out:
2589 	return msg;
2590 }
2591 
2592 /*
2593  * called under mdsc->mutex if error, under no mutex if
2594  * success.
2595  */
2596 static void complete_request(struct ceph_mds_client *mdsc,
2597 			     struct ceph_mds_request *req)
2598 {
2599 	req->r_end_latency = ktime_get();
2600 
2601 	if (req->r_callback)
2602 		req->r_callback(mdsc, req);
2603 	complete_all(&req->r_completion);
2604 }
2605 
2606 static struct ceph_mds_request_head_old *
2607 find_old_request_head(void *p, u64 features)
2608 {
2609 	bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2610 	struct ceph_mds_request_head *new_head;
2611 
2612 	if (legacy)
2613 		return (struct ceph_mds_request_head_old *)p;
2614 	new_head = (struct ceph_mds_request_head *)p;
2615 	return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2616 }
2617 
2618 /*
2619  * called under mdsc->mutex
2620  */
2621 static int __prepare_send_request(struct ceph_mds_session *session,
2622 				  struct ceph_mds_request *req,
2623 				  bool drop_cap_releases)
2624 {
2625 	int mds = session->s_mds;
2626 	struct ceph_mds_client *mdsc = session->s_mdsc;
2627 	struct ceph_mds_request_head_old *rhead;
2628 	struct ceph_msg *msg;
2629 	int flags = 0;
2630 
2631 	req->r_attempts++;
2632 	if (req->r_inode) {
2633 		struct ceph_cap *cap =
2634 			ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2635 
2636 		if (cap)
2637 			req->r_sent_on_mseq = cap->mseq;
2638 		else
2639 			req->r_sent_on_mseq = -1;
2640 	}
2641 	dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2642 	     req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2643 
2644 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2645 		void *p;
2646 
2647 		/*
2648 		 * Replay.  Do not regenerate message (and rebuild
2649 		 * paths, etc.); just use the original message.
2650 		 * Rebuilding paths will break for renames because
2651 		 * d_move mangles the src name.
2652 		 */
2653 		msg = req->r_request;
2654 		rhead = find_old_request_head(msg->front.iov_base,
2655 					      session->s_con.peer_features);
2656 
2657 		flags = le32_to_cpu(rhead->flags);
2658 		flags |= CEPH_MDS_FLAG_REPLAY;
2659 		rhead->flags = cpu_to_le32(flags);
2660 
2661 		if (req->r_target_inode)
2662 			rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2663 
2664 		rhead->num_retry = req->r_attempts - 1;
2665 
2666 		/* remove cap/dentry releases from message */
2667 		rhead->num_releases = 0;
2668 
2669 		p = msg->front.iov_base + req->r_request_release_offset;
2670 		encode_timestamp_and_gids(&p, req);
2671 
2672 		msg->front.iov_len = p - msg->front.iov_base;
2673 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2674 		return 0;
2675 	}
2676 
2677 	if (req->r_request) {
2678 		ceph_msg_put(req->r_request);
2679 		req->r_request = NULL;
2680 	}
2681 	msg = create_request_message(session, req, drop_cap_releases);
2682 	if (IS_ERR(msg)) {
2683 		req->r_err = PTR_ERR(msg);
2684 		return PTR_ERR(msg);
2685 	}
2686 	req->r_request = msg;
2687 
2688 	rhead = find_old_request_head(msg->front.iov_base,
2689 				      session->s_con.peer_features);
2690 	rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2691 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2692 		flags |= CEPH_MDS_FLAG_REPLAY;
2693 	if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2694 		flags |= CEPH_MDS_FLAG_ASYNC;
2695 	if (req->r_parent)
2696 		flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2697 	rhead->flags = cpu_to_le32(flags);
2698 	rhead->num_fwd = req->r_num_fwd;
2699 	rhead->num_retry = req->r_attempts - 1;
2700 
2701 	dout(" r_parent = %p\n", req->r_parent);
2702 	return 0;
2703 }
2704 
2705 /*
2706  * called under mdsc->mutex
2707  */
2708 static int __send_request(struct ceph_mds_session *session,
2709 			  struct ceph_mds_request *req,
2710 			  bool drop_cap_releases)
2711 {
2712 	int err;
2713 
2714 	err = __prepare_send_request(session, req, drop_cap_releases);
2715 	if (!err) {
2716 		ceph_msg_get(req->r_request);
2717 		ceph_con_send(&session->s_con, req->r_request);
2718 	}
2719 
2720 	return err;
2721 }
2722 
2723 /*
2724  * send request, or put it on the appropriate wait list.
2725  */
2726 static void __do_request(struct ceph_mds_client *mdsc,
2727 			struct ceph_mds_request *req)
2728 {
2729 	struct ceph_mds_session *session = NULL;
2730 	int mds = -1;
2731 	int err = 0;
2732 	bool random;
2733 
2734 	if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2735 		if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2736 			__unregister_request(mdsc, req);
2737 		return;
2738 	}
2739 
2740 	if (req->r_timeout &&
2741 	    time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2742 		dout("do_request timed out\n");
2743 		err = -ETIMEDOUT;
2744 		goto finish;
2745 	}
2746 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2747 		dout("do_request forced umount\n");
2748 		err = -EIO;
2749 		goto finish;
2750 	}
2751 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2752 		if (mdsc->mdsmap_err) {
2753 			err = mdsc->mdsmap_err;
2754 			dout("do_request mdsmap err %d\n", err);
2755 			goto finish;
2756 		}
2757 		if (mdsc->mdsmap->m_epoch == 0) {
2758 			dout("do_request no mdsmap, waiting for map\n");
2759 			list_add(&req->r_wait, &mdsc->waiting_for_map);
2760 			return;
2761 		}
2762 		if (!(mdsc->fsc->mount_options->flags &
2763 		      CEPH_MOUNT_OPT_MOUNTWAIT) &&
2764 		    !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2765 			err = -EHOSTUNREACH;
2766 			goto finish;
2767 		}
2768 	}
2769 
2770 	put_request_session(req);
2771 
2772 	mds = __choose_mds(mdsc, req, &random);
2773 	if (mds < 0 ||
2774 	    ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2775 		if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2776 			err = -EJUKEBOX;
2777 			goto finish;
2778 		}
2779 		dout("do_request no mds or not active, waiting for map\n");
2780 		list_add(&req->r_wait, &mdsc->waiting_for_map);
2781 		return;
2782 	}
2783 
2784 	/* get, open session */
2785 	session = __ceph_lookup_mds_session(mdsc, mds);
2786 	if (!session) {
2787 		session = register_session(mdsc, mds);
2788 		if (IS_ERR(session)) {
2789 			err = PTR_ERR(session);
2790 			goto finish;
2791 		}
2792 	}
2793 	req->r_session = ceph_get_mds_session(session);
2794 
2795 	dout("do_request mds%d session %p state %s\n", mds, session,
2796 	     ceph_session_state_name(session->s_state));
2797 	if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2798 	    session->s_state != CEPH_MDS_SESSION_HUNG) {
2799 		/*
2800 		 * We cannot queue async requests since the caps and delegated
2801 		 * inodes are bound to the session. Just return -EJUKEBOX and
2802 		 * let the caller retry a sync request in that case.
2803 		 */
2804 		if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2805 			err = -EJUKEBOX;
2806 			goto out_session;
2807 		}
2808 
2809 		/*
2810 		 * If the session has been REJECTED, then return a hard error,
2811 		 * unless it's a CLEANRECOVER mount, in which case we'll queue
2812 		 * it to the mdsc queue.
2813 		 */
2814 		if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2815 			if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2816 				list_add(&req->r_wait, &mdsc->waiting_for_map);
2817 			else
2818 				err = -EACCES;
2819 			goto out_session;
2820 		}
2821 
2822 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
2823 		    session->s_state == CEPH_MDS_SESSION_CLOSING) {
2824 			err = __open_session(mdsc, session);
2825 			if (err)
2826 				goto out_session;
2827 			/* retry the same mds later */
2828 			if (random)
2829 				req->r_resend_mds = mds;
2830 		}
2831 		list_add(&req->r_wait, &session->s_waiting);
2832 		goto out_session;
2833 	}
2834 
2835 	/* send request */
2836 	req->r_resend_mds = -1;   /* forget any previous mds hint */
2837 
2838 	if (req->r_request_started == 0)   /* note request start time */
2839 		req->r_request_started = jiffies;
2840 
2841 	err = __send_request(session, req, false);
2842 
2843 out_session:
2844 	ceph_put_mds_session(session);
2845 finish:
2846 	if (err) {
2847 		dout("__do_request early error %d\n", err);
2848 		req->r_err = err;
2849 		complete_request(mdsc, req);
2850 		__unregister_request(mdsc, req);
2851 	}
2852 	return;
2853 }
2854 
2855 /*
2856  * called under mdsc->mutex
2857  */
2858 static void __wake_requests(struct ceph_mds_client *mdsc,
2859 			    struct list_head *head)
2860 {
2861 	struct ceph_mds_request *req;
2862 	LIST_HEAD(tmp_list);
2863 
2864 	list_splice_init(head, &tmp_list);
2865 
2866 	while (!list_empty(&tmp_list)) {
2867 		req = list_entry(tmp_list.next,
2868 				 struct ceph_mds_request, r_wait);
2869 		list_del_init(&req->r_wait);
2870 		dout(" wake request %p tid %llu\n", req, req->r_tid);
2871 		__do_request(mdsc, req);
2872 	}
2873 }
2874 
2875 /*
2876  * Wake up threads with requests pending for @mds, so that they can
2877  * resubmit their requests to a possibly different mds.
2878  */
2879 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2880 {
2881 	struct ceph_mds_request *req;
2882 	struct rb_node *p = rb_first(&mdsc->request_tree);
2883 
2884 	dout("kick_requests mds%d\n", mds);
2885 	while (p) {
2886 		req = rb_entry(p, struct ceph_mds_request, r_node);
2887 		p = rb_next(p);
2888 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2889 			continue;
2890 		if (req->r_attempts > 0)
2891 			continue; /* only new requests */
2892 		if (req->r_session &&
2893 		    req->r_session->s_mds == mds) {
2894 			dout(" kicking tid %llu\n", req->r_tid);
2895 			list_del_init(&req->r_wait);
2896 			__do_request(mdsc, req);
2897 		}
2898 	}
2899 }
2900 
2901 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2902 			      struct ceph_mds_request *req)
2903 {
2904 	int err = 0;
2905 
2906 	/* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2907 	if (req->r_inode)
2908 		ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2909 	if (req->r_parent) {
2910 		struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2911 		int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2912 			    CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2913 		spin_lock(&ci->i_ceph_lock);
2914 		ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2915 		__ceph_touch_fmode(ci, mdsc, fmode);
2916 		spin_unlock(&ci->i_ceph_lock);
2917 	}
2918 	if (req->r_old_dentry_dir)
2919 		ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2920 				  CEPH_CAP_PIN);
2921 
2922 	if (req->r_inode) {
2923 		err = ceph_wait_on_async_create(req->r_inode);
2924 		if (err) {
2925 			dout("%s: wait for async create returned: %d\n",
2926 			     __func__, err);
2927 			return err;
2928 		}
2929 	}
2930 
2931 	if (!err && req->r_old_inode) {
2932 		err = ceph_wait_on_async_create(req->r_old_inode);
2933 		if (err) {
2934 			dout("%s: wait for async create returned: %d\n",
2935 			     __func__, err);
2936 			return err;
2937 		}
2938 	}
2939 
2940 	dout("submit_request on %p for inode %p\n", req, dir);
2941 	mutex_lock(&mdsc->mutex);
2942 	__register_request(mdsc, req, dir);
2943 	__do_request(mdsc, req);
2944 	err = req->r_err;
2945 	mutex_unlock(&mdsc->mutex);
2946 	return err;
2947 }
2948 
2949 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2950 				  struct ceph_mds_request *req)
2951 {
2952 	int err;
2953 
2954 	/* wait */
2955 	dout("do_request waiting\n");
2956 	if (!req->r_timeout && req->r_wait_for_completion) {
2957 		err = req->r_wait_for_completion(mdsc, req);
2958 	} else {
2959 		long timeleft = wait_for_completion_killable_timeout(
2960 					&req->r_completion,
2961 					ceph_timeout_jiffies(req->r_timeout));
2962 		if (timeleft > 0)
2963 			err = 0;
2964 		else if (!timeleft)
2965 			err = -ETIMEDOUT;  /* timed out */
2966 		else
2967 			err = timeleft;  /* killed */
2968 	}
2969 	dout("do_request waited, got %d\n", err);
2970 	mutex_lock(&mdsc->mutex);
2971 
2972 	/* only abort if we didn't race with a real reply */
2973 	if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2974 		err = le32_to_cpu(req->r_reply_info.head->result);
2975 	} else if (err < 0) {
2976 		dout("aborted request %lld with %d\n", req->r_tid, err);
2977 
2978 		/*
2979 		 * ensure we aren't running concurrently with
2980 		 * ceph_fill_trace or ceph_readdir_prepopulate, which
2981 		 * rely on locks (dir mutex) held by our caller.
2982 		 */
2983 		mutex_lock(&req->r_fill_mutex);
2984 		req->r_err = err;
2985 		set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2986 		mutex_unlock(&req->r_fill_mutex);
2987 
2988 		if (req->r_parent &&
2989 		    (req->r_op & CEPH_MDS_OP_WRITE))
2990 			ceph_invalidate_dir_request(req);
2991 	} else {
2992 		err = req->r_err;
2993 	}
2994 
2995 	mutex_unlock(&mdsc->mutex);
2996 	return err;
2997 }
2998 
2999 /*
3000  * Synchrously perform an mds request.  Take care of all of the
3001  * session setup, forwarding, retry details.
3002  */
3003 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3004 			 struct inode *dir,
3005 			 struct ceph_mds_request *req)
3006 {
3007 	int err;
3008 
3009 	dout("do_request on %p\n", req);
3010 
3011 	/* issue */
3012 	err = ceph_mdsc_submit_request(mdsc, dir, req);
3013 	if (!err)
3014 		err = ceph_mdsc_wait_request(mdsc, req);
3015 	dout("do_request %p done, result %d\n", req, err);
3016 	return err;
3017 }
3018 
3019 /*
3020  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3021  * namespace request.
3022  */
3023 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3024 {
3025 	struct inode *dir = req->r_parent;
3026 	struct inode *old_dir = req->r_old_dentry_dir;
3027 
3028 	dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3029 
3030 	ceph_dir_clear_complete(dir);
3031 	if (old_dir)
3032 		ceph_dir_clear_complete(old_dir);
3033 	if (req->r_dentry)
3034 		ceph_invalidate_dentry_lease(req->r_dentry);
3035 	if (req->r_old_dentry)
3036 		ceph_invalidate_dentry_lease(req->r_old_dentry);
3037 }
3038 
3039 /*
3040  * Handle mds reply.
3041  *
3042  * We take the session mutex and parse and process the reply immediately.
3043  * This preserves the logical ordering of replies, capabilities, etc., sent
3044  * by the MDS as they are applied to our local cache.
3045  */
3046 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3047 {
3048 	struct ceph_mds_client *mdsc = session->s_mdsc;
3049 	struct ceph_mds_request *req;
3050 	struct ceph_mds_reply_head *head = msg->front.iov_base;
3051 	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3052 	struct ceph_snap_realm *realm;
3053 	u64 tid;
3054 	int err, result;
3055 	int mds = session->s_mds;
3056 
3057 	if (msg->front.iov_len < sizeof(*head)) {
3058 		pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3059 		ceph_msg_dump(msg);
3060 		return;
3061 	}
3062 
3063 	/* get request, session */
3064 	tid = le64_to_cpu(msg->hdr.tid);
3065 	mutex_lock(&mdsc->mutex);
3066 	req = lookup_get_request(mdsc, tid);
3067 	if (!req) {
3068 		dout("handle_reply on unknown tid %llu\n", tid);
3069 		mutex_unlock(&mdsc->mutex);
3070 		return;
3071 	}
3072 	dout("handle_reply %p\n", req);
3073 
3074 	/* correct session? */
3075 	if (req->r_session != session) {
3076 		pr_err("mdsc_handle_reply got %llu on session mds%d"
3077 		       " not mds%d\n", tid, session->s_mds,
3078 		       req->r_session ? req->r_session->s_mds : -1);
3079 		mutex_unlock(&mdsc->mutex);
3080 		goto out;
3081 	}
3082 
3083 	/* dup? */
3084 	if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3085 	    (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3086 		pr_warn("got a dup %s reply on %llu from mds%d\n",
3087 			   head->safe ? "safe" : "unsafe", tid, mds);
3088 		mutex_unlock(&mdsc->mutex);
3089 		goto out;
3090 	}
3091 	if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3092 		pr_warn("got unsafe after safe on %llu from mds%d\n",
3093 			   tid, mds);
3094 		mutex_unlock(&mdsc->mutex);
3095 		goto out;
3096 	}
3097 
3098 	result = le32_to_cpu(head->result);
3099 
3100 	/*
3101 	 * Handle an ESTALE
3102 	 * if we're not talking to the authority, send to them
3103 	 * if the authority has changed while we weren't looking,
3104 	 * send to new authority
3105 	 * Otherwise we just have to return an ESTALE
3106 	 */
3107 	if (result == -ESTALE) {
3108 		dout("got ESTALE on request %llu\n", req->r_tid);
3109 		req->r_resend_mds = -1;
3110 		if (req->r_direct_mode != USE_AUTH_MDS) {
3111 			dout("not using auth, setting for that now\n");
3112 			req->r_direct_mode = USE_AUTH_MDS;
3113 			__do_request(mdsc, req);
3114 			mutex_unlock(&mdsc->mutex);
3115 			goto out;
3116 		} else  {
3117 			int mds = __choose_mds(mdsc, req, NULL);
3118 			if (mds >= 0 && mds != req->r_session->s_mds) {
3119 				dout("but auth changed, so resending\n");
3120 				__do_request(mdsc, req);
3121 				mutex_unlock(&mdsc->mutex);
3122 				goto out;
3123 			}
3124 		}
3125 		dout("have to return ESTALE on request %llu\n", req->r_tid);
3126 	}
3127 
3128 
3129 	if (head->safe) {
3130 		set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3131 		__unregister_request(mdsc, req);
3132 
3133 		/* last request during umount? */
3134 		if (mdsc->stopping && !__get_oldest_req(mdsc))
3135 			complete_all(&mdsc->safe_umount_waiters);
3136 
3137 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3138 			/*
3139 			 * We already handled the unsafe response, now do the
3140 			 * cleanup.  No need to examine the response; the MDS
3141 			 * doesn't include any result info in the safe
3142 			 * response.  And even if it did, there is nothing
3143 			 * useful we could do with a revised return value.
3144 			 */
3145 			dout("got safe reply %llu, mds%d\n", tid, mds);
3146 
3147 			mutex_unlock(&mdsc->mutex);
3148 			goto out;
3149 		}
3150 	} else {
3151 		set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3152 		list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3153 	}
3154 
3155 	dout("handle_reply tid %lld result %d\n", tid, result);
3156 	rinfo = &req->r_reply_info;
3157 	if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3158 		err = parse_reply_info(session, msg, rinfo, (u64)-1);
3159 	else
3160 		err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3161 	mutex_unlock(&mdsc->mutex);
3162 
3163 	/* Must find target inode outside of mutexes to avoid deadlocks */
3164 	if ((err >= 0) && rinfo->head->is_target) {
3165 		struct inode *in;
3166 		struct ceph_vino tvino = {
3167 			.ino  = le64_to_cpu(rinfo->targeti.in->ino),
3168 			.snap = le64_to_cpu(rinfo->targeti.in->snapid)
3169 		};
3170 
3171 		in = ceph_get_inode(mdsc->fsc->sb, tvino);
3172 		if (IS_ERR(in)) {
3173 			err = PTR_ERR(in);
3174 			mutex_lock(&session->s_mutex);
3175 			goto out_err;
3176 		}
3177 		req->r_target_inode = in;
3178 	}
3179 
3180 	mutex_lock(&session->s_mutex);
3181 	if (err < 0) {
3182 		pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3183 		ceph_msg_dump(msg);
3184 		goto out_err;
3185 	}
3186 
3187 	/* snap trace */
3188 	realm = NULL;
3189 	if (rinfo->snapblob_len) {
3190 		down_write(&mdsc->snap_rwsem);
3191 		ceph_update_snap_trace(mdsc, rinfo->snapblob,
3192 				rinfo->snapblob + rinfo->snapblob_len,
3193 				le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3194 				&realm);
3195 		downgrade_write(&mdsc->snap_rwsem);
3196 	} else {
3197 		down_read(&mdsc->snap_rwsem);
3198 	}
3199 
3200 	/* insert trace into our cache */
3201 	mutex_lock(&req->r_fill_mutex);
3202 	current->journal_info = req;
3203 	err = ceph_fill_trace(mdsc->fsc->sb, req);
3204 	if (err == 0) {
3205 		if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3206 				    req->r_op == CEPH_MDS_OP_LSSNAP))
3207 			ceph_readdir_prepopulate(req, req->r_session);
3208 	}
3209 	current->journal_info = NULL;
3210 	mutex_unlock(&req->r_fill_mutex);
3211 
3212 	up_read(&mdsc->snap_rwsem);
3213 	if (realm)
3214 		ceph_put_snap_realm(mdsc, realm);
3215 
3216 	if (err == 0) {
3217 		if (req->r_target_inode &&
3218 		    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3219 			struct ceph_inode_info *ci =
3220 				ceph_inode(req->r_target_inode);
3221 			spin_lock(&ci->i_unsafe_lock);
3222 			list_add_tail(&req->r_unsafe_target_item,
3223 				      &ci->i_unsafe_iops);
3224 			spin_unlock(&ci->i_unsafe_lock);
3225 		}
3226 
3227 		ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3228 	}
3229 out_err:
3230 	mutex_lock(&mdsc->mutex);
3231 	if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3232 		if (err) {
3233 			req->r_err = err;
3234 		} else {
3235 			req->r_reply =  ceph_msg_get(msg);
3236 			set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3237 		}
3238 	} else {
3239 		dout("reply arrived after request %lld was aborted\n", tid);
3240 	}
3241 	mutex_unlock(&mdsc->mutex);
3242 
3243 	mutex_unlock(&session->s_mutex);
3244 
3245 	/* kick calling process */
3246 	complete_request(mdsc, req);
3247 
3248 	ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3249 				     req->r_end_latency, err);
3250 out:
3251 	ceph_mdsc_put_request(req);
3252 	return;
3253 }
3254 
3255 
3256 
3257 /*
3258  * handle mds notification that our request has been forwarded.
3259  */
3260 static void handle_forward(struct ceph_mds_client *mdsc,
3261 			   struct ceph_mds_session *session,
3262 			   struct ceph_msg *msg)
3263 {
3264 	struct ceph_mds_request *req;
3265 	u64 tid = le64_to_cpu(msg->hdr.tid);
3266 	u32 next_mds;
3267 	u32 fwd_seq;
3268 	int err = -EINVAL;
3269 	void *p = msg->front.iov_base;
3270 	void *end = p + msg->front.iov_len;
3271 
3272 	ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3273 	next_mds = ceph_decode_32(&p);
3274 	fwd_seq = ceph_decode_32(&p);
3275 
3276 	mutex_lock(&mdsc->mutex);
3277 	req = lookup_get_request(mdsc, tid);
3278 	if (!req) {
3279 		dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3280 		goto out;  /* dup reply? */
3281 	}
3282 
3283 	if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3284 		dout("forward tid %llu aborted, unregistering\n", tid);
3285 		__unregister_request(mdsc, req);
3286 	} else if (fwd_seq <= req->r_num_fwd) {
3287 		dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3288 		     tid, next_mds, req->r_num_fwd, fwd_seq);
3289 	} else {
3290 		/* resend. forward race not possible; mds would drop */
3291 		dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3292 		BUG_ON(req->r_err);
3293 		BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3294 		req->r_attempts = 0;
3295 		req->r_num_fwd = fwd_seq;
3296 		req->r_resend_mds = next_mds;
3297 		put_request_session(req);
3298 		__do_request(mdsc, req);
3299 	}
3300 	ceph_mdsc_put_request(req);
3301 out:
3302 	mutex_unlock(&mdsc->mutex);
3303 	return;
3304 
3305 bad:
3306 	pr_err("mdsc_handle_forward decode error err=%d\n", err);
3307 }
3308 
3309 static int __decode_session_metadata(void **p, void *end,
3310 				     bool *blocklisted)
3311 {
3312 	/* map<string,string> */
3313 	u32 n;
3314 	bool err_str;
3315 	ceph_decode_32_safe(p, end, n, bad);
3316 	while (n-- > 0) {
3317 		u32 len;
3318 		ceph_decode_32_safe(p, end, len, bad);
3319 		ceph_decode_need(p, end, len, bad);
3320 		err_str = !strncmp(*p, "error_string", len);
3321 		*p += len;
3322 		ceph_decode_32_safe(p, end, len, bad);
3323 		ceph_decode_need(p, end, len, bad);
3324 		/*
3325 		 * Match "blocklisted (blacklisted)" from newer MDSes,
3326 		 * or "blacklisted" from older MDSes.
3327 		 */
3328 		if (err_str && strnstr(*p, "blacklisted", len))
3329 			*blocklisted = true;
3330 		*p += len;
3331 	}
3332 	return 0;
3333 bad:
3334 	return -1;
3335 }
3336 
3337 /*
3338  * handle a mds session control message
3339  */
3340 static void handle_session(struct ceph_mds_session *session,
3341 			   struct ceph_msg *msg)
3342 {
3343 	struct ceph_mds_client *mdsc = session->s_mdsc;
3344 	int mds = session->s_mds;
3345 	int msg_version = le16_to_cpu(msg->hdr.version);
3346 	void *p = msg->front.iov_base;
3347 	void *end = p + msg->front.iov_len;
3348 	struct ceph_mds_session_head *h;
3349 	u32 op;
3350 	u64 seq, features = 0;
3351 	int wake = 0;
3352 	bool blocklisted = false;
3353 
3354 	/* decode */
3355 	ceph_decode_need(&p, end, sizeof(*h), bad);
3356 	h = p;
3357 	p += sizeof(*h);
3358 
3359 	op = le32_to_cpu(h->op);
3360 	seq = le64_to_cpu(h->seq);
3361 
3362 	if (msg_version >= 3) {
3363 		u32 len;
3364 		/* version >= 2 and < 5, decode metadata, skip otherwise
3365 		 * as it's handled via flags.
3366 		 */
3367 		if (msg_version >= 5)
3368 			ceph_decode_skip_map(&p, end, string, string, bad);
3369 		else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3370 			goto bad;
3371 
3372 		/* version >= 3, feature bits */
3373 		ceph_decode_32_safe(&p, end, len, bad);
3374 		if (len) {
3375 			ceph_decode_64_safe(&p, end, features, bad);
3376 			p += len - sizeof(features);
3377 		}
3378 	}
3379 
3380 	if (msg_version >= 5) {
3381 		u32 flags;
3382 		/* version >= 4, struct_v, struct_cv, len, metric_spec */
3383 	        ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 2, bad);
3384 		/* version >= 5, flags   */
3385                 ceph_decode_32_safe(&p, end, flags, bad);
3386 		if (flags & CEPH_SESSION_BLOCKLISTED) {
3387 		        pr_warn("mds%d session blocklisted\n", session->s_mds);
3388 			blocklisted = true;
3389 		}
3390 	}
3391 
3392 	mutex_lock(&mdsc->mutex);
3393 	if (op == CEPH_SESSION_CLOSE) {
3394 		ceph_get_mds_session(session);
3395 		__unregister_session(mdsc, session);
3396 	}
3397 	/* FIXME: this ttl calculation is generous */
3398 	session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3399 	mutex_unlock(&mdsc->mutex);
3400 
3401 	mutex_lock(&session->s_mutex);
3402 
3403 	dout("handle_session mds%d %s %p state %s seq %llu\n",
3404 	     mds, ceph_session_op_name(op), session,
3405 	     ceph_session_state_name(session->s_state), seq);
3406 
3407 	if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3408 		session->s_state = CEPH_MDS_SESSION_OPEN;
3409 		pr_info("mds%d came back\n", session->s_mds);
3410 	}
3411 
3412 	switch (op) {
3413 	case CEPH_SESSION_OPEN:
3414 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3415 			pr_info("mds%d reconnect success\n", session->s_mds);
3416 		session->s_state = CEPH_MDS_SESSION_OPEN;
3417 		session->s_features = features;
3418 		renewed_caps(mdsc, session, 0);
3419 		if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3420 			metric_schedule_delayed(&mdsc->metric);
3421 		wake = 1;
3422 		if (mdsc->stopping)
3423 			__close_session(mdsc, session);
3424 		break;
3425 
3426 	case CEPH_SESSION_RENEWCAPS:
3427 		if (session->s_renew_seq == seq)
3428 			renewed_caps(mdsc, session, 1);
3429 		break;
3430 
3431 	case CEPH_SESSION_CLOSE:
3432 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3433 			pr_info("mds%d reconnect denied\n", session->s_mds);
3434 		session->s_state = CEPH_MDS_SESSION_CLOSED;
3435 		cleanup_session_requests(mdsc, session);
3436 		remove_session_caps(session);
3437 		wake = 2; /* for good measure */
3438 		wake_up_all(&mdsc->session_close_wq);
3439 		break;
3440 
3441 	case CEPH_SESSION_STALE:
3442 		pr_info("mds%d caps went stale, renewing\n",
3443 			session->s_mds);
3444 		atomic_inc(&session->s_cap_gen);
3445 		session->s_cap_ttl = jiffies - 1;
3446 		send_renew_caps(mdsc, session);
3447 		break;
3448 
3449 	case CEPH_SESSION_RECALL_STATE:
3450 		ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3451 		break;
3452 
3453 	case CEPH_SESSION_FLUSHMSG:
3454 		send_flushmsg_ack(mdsc, session, seq);
3455 		break;
3456 
3457 	case CEPH_SESSION_FORCE_RO:
3458 		dout("force_session_readonly %p\n", session);
3459 		spin_lock(&session->s_cap_lock);
3460 		session->s_readonly = true;
3461 		spin_unlock(&session->s_cap_lock);
3462 		wake_up_session_caps(session, FORCE_RO);
3463 		break;
3464 
3465 	case CEPH_SESSION_REJECT:
3466 		WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3467 		pr_info("mds%d rejected session\n", session->s_mds);
3468 		session->s_state = CEPH_MDS_SESSION_REJECTED;
3469 		cleanup_session_requests(mdsc, session);
3470 		remove_session_caps(session);
3471 		if (blocklisted)
3472 			mdsc->fsc->blocklisted = true;
3473 		wake = 2; /* for good measure */
3474 		break;
3475 
3476 	default:
3477 		pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3478 		WARN_ON(1);
3479 	}
3480 
3481 	mutex_unlock(&session->s_mutex);
3482 	if (wake) {
3483 		mutex_lock(&mdsc->mutex);
3484 		__wake_requests(mdsc, &session->s_waiting);
3485 		if (wake == 2)
3486 			kick_requests(mdsc, mds);
3487 		mutex_unlock(&mdsc->mutex);
3488 	}
3489 	if (op == CEPH_SESSION_CLOSE)
3490 		ceph_put_mds_session(session);
3491 	return;
3492 
3493 bad:
3494 	pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3495 	       (int)msg->front.iov_len);
3496 	ceph_msg_dump(msg);
3497 	return;
3498 }
3499 
3500 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3501 {
3502 	int dcaps;
3503 
3504 	dcaps = xchg(&req->r_dir_caps, 0);
3505 	if (dcaps) {
3506 		dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3507 		ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3508 	}
3509 }
3510 
3511 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3512 {
3513 	int dcaps;
3514 
3515 	dcaps = xchg(&req->r_dir_caps, 0);
3516 	if (dcaps) {
3517 		dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3518 		ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3519 						dcaps);
3520 	}
3521 }
3522 
3523 /*
3524  * called under session->mutex.
3525  */
3526 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3527 				   struct ceph_mds_session *session)
3528 {
3529 	struct ceph_mds_request *req, *nreq;
3530 	struct rb_node *p;
3531 
3532 	dout("replay_unsafe_requests mds%d\n", session->s_mds);
3533 
3534 	mutex_lock(&mdsc->mutex);
3535 	list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3536 		__send_request(session, req, true);
3537 
3538 	/*
3539 	 * also re-send old requests when MDS enters reconnect stage. So that MDS
3540 	 * can process completed request in clientreplay stage.
3541 	 */
3542 	p = rb_first(&mdsc->request_tree);
3543 	while (p) {
3544 		req = rb_entry(p, struct ceph_mds_request, r_node);
3545 		p = rb_next(p);
3546 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3547 			continue;
3548 		if (req->r_attempts == 0)
3549 			continue; /* only old requests */
3550 		if (!req->r_session)
3551 			continue;
3552 		if (req->r_session->s_mds != session->s_mds)
3553 			continue;
3554 
3555 		ceph_mdsc_release_dir_caps_no_check(req);
3556 
3557 		__send_request(session, req, true);
3558 	}
3559 	mutex_unlock(&mdsc->mutex);
3560 }
3561 
3562 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3563 {
3564 	struct ceph_msg *reply;
3565 	struct ceph_pagelist *_pagelist;
3566 	struct page *page;
3567 	__le32 *addr;
3568 	int err = -ENOMEM;
3569 
3570 	if (!recon_state->allow_multi)
3571 		return -ENOSPC;
3572 
3573 	/* can't handle message that contains both caps and realm */
3574 	BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3575 
3576 	/* pre-allocate new pagelist */
3577 	_pagelist = ceph_pagelist_alloc(GFP_NOFS);
3578 	if (!_pagelist)
3579 		return -ENOMEM;
3580 
3581 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3582 	if (!reply)
3583 		goto fail_msg;
3584 
3585 	/* placeholder for nr_caps */
3586 	err = ceph_pagelist_encode_32(_pagelist, 0);
3587 	if (err < 0)
3588 		goto fail;
3589 
3590 	if (recon_state->nr_caps) {
3591 		/* currently encoding caps */
3592 		err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3593 		if (err)
3594 			goto fail;
3595 	} else {
3596 		/* placeholder for nr_realms (currently encoding relams) */
3597 		err = ceph_pagelist_encode_32(_pagelist, 0);
3598 		if (err < 0)
3599 			goto fail;
3600 	}
3601 
3602 	err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3603 	if (err)
3604 		goto fail;
3605 
3606 	page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3607 	addr = kmap_atomic(page);
3608 	if (recon_state->nr_caps) {
3609 		/* currently encoding caps */
3610 		*addr = cpu_to_le32(recon_state->nr_caps);
3611 	} else {
3612 		/* currently encoding relams */
3613 		*(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3614 	}
3615 	kunmap_atomic(addr);
3616 
3617 	reply->hdr.version = cpu_to_le16(5);
3618 	reply->hdr.compat_version = cpu_to_le16(4);
3619 
3620 	reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3621 	ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3622 
3623 	ceph_con_send(&recon_state->session->s_con, reply);
3624 	ceph_pagelist_release(recon_state->pagelist);
3625 
3626 	recon_state->pagelist = _pagelist;
3627 	recon_state->nr_caps = 0;
3628 	recon_state->nr_realms = 0;
3629 	recon_state->msg_version = 5;
3630 	return 0;
3631 fail:
3632 	ceph_msg_put(reply);
3633 fail_msg:
3634 	ceph_pagelist_release(_pagelist);
3635 	return err;
3636 }
3637 
3638 static struct dentry* d_find_primary(struct inode *inode)
3639 {
3640 	struct dentry *alias, *dn = NULL;
3641 
3642 	if (hlist_empty(&inode->i_dentry))
3643 		return NULL;
3644 
3645 	spin_lock(&inode->i_lock);
3646 	if (hlist_empty(&inode->i_dentry))
3647 		goto out_unlock;
3648 
3649 	if (S_ISDIR(inode->i_mode)) {
3650 		alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3651 		if (!IS_ROOT(alias))
3652 			dn = dget(alias);
3653 		goto out_unlock;
3654 	}
3655 
3656 	hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3657 		spin_lock(&alias->d_lock);
3658 		if (!d_unhashed(alias) &&
3659 		    (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3660 			dn = dget_dlock(alias);
3661 		}
3662 		spin_unlock(&alias->d_lock);
3663 		if (dn)
3664 			break;
3665 	}
3666 out_unlock:
3667 	spin_unlock(&inode->i_lock);
3668 	return dn;
3669 }
3670 
3671 /*
3672  * Encode information about a cap for a reconnect with the MDS.
3673  */
3674 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3675 			  void *arg)
3676 {
3677 	union {
3678 		struct ceph_mds_cap_reconnect v2;
3679 		struct ceph_mds_cap_reconnect_v1 v1;
3680 	} rec;
3681 	struct ceph_inode_info *ci = cap->ci;
3682 	struct ceph_reconnect_state *recon_state = arg;
3683 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3684 	struct dentry *dentry;
3685 	char *path;
3686 	int pathlen = 0, err;
3687 	u64 pathbase;
3688 	u64 snap_follows;
3689 
3690 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3691 	     inode, ceph_vinop(inode), cap, cap->cap_id,
3692 	     ceph_cap_string(cap->issued));
3693 
3694 	dentry = d_find_primary(inode);
3695 	if (dentry) {
3696 		/* set pathbase to parent dir when msg_version >= 2 */
3697 		path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3698 					    recon_state->msg_version >= 2);
3699 		dput(dentry);
3700 		if (IS_ERR(path)) {
3701 			err = PTR_ERR(path);
3702 			goto out_err;
3703 		}
3704 	} else {
3705 		path = NULL;
3706 		pathbase = 0;
3707 	}
3708 
3709 	spin_lock(&ci->i_ceph_lock);
3710 	cap->seq = 0;        /* reset cap seq */
3711 	cap->issue_seq = 0;  /* and issue_seq */
3712 	cap->mseq = 0;       /* and migrate_seq */
3713 	cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
3714 
3715 	/* These are lost when the session goes away */
3716 	if (S_ISDIR(inode->i_mode)) {
3717 		if (cap->issued & CEPH_CAP_DIR_CREATE) {
3718 			ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3719 			memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3720 		}
3721 		cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3722 	}
3723 
3724 	if (recon_state->msg_version >= 2) {
3725 		rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3726 		rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3727 		rec.v2.issued = cpu_to_le32(cap->issued);
3728 		rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3729 		rec.v2.pathbase = cpu_to_le64(pathbase);
3730 		rec.v2.flock_len = (__force __le32)
3731 			((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3732 	} else {
3733 		rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3734 		rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3735 		rec.v1.issued = cpu_to_le32(cap->issued);
3736 		rec.v1.size = cpu_to_le64(i_size_read(inode));
3737 		ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3738 		ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3739 		rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3740 		rec.v1.pathbase = cpu_to_le64(pathbase);
3741 	}
3742 
3743 	if (list_empty(&ci->i_cap_snaps)) {
3744 		snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3745 	} else {
3746 		struct ceph_cap_snap *capsnap =
3747 			list_first_entry(&ci->i_cap_snaps,
3748 					 struct ceph_cap_snap, ci_item);
3749 		snap_follows = capsnap->follows;
3750 	}
3751 	spin_unlock(&ci->i_ceph_lock);
3752 
3753 	if (recon_state->msg_version >= 2) {
3754 		int num_fcntl_locks, num_flock_locks;
3755 		struct ceph_filelock *flocks = NULL;
3756 		size_t struct_len, total_len = sizeof(u64);
3757 		u8 struct_v = 0;
3758 
3759 encode_again:
3760 		if (rec.v2.flock_len) {
3761 			ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3762 		} else {
3763 			num_fcntl_locks = 0;
3764 			num_flock_locks = 0;
3765 		}
3766 		if (num_fcntl_locks + num_flock_locks > 0) {
3767 			flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3768 					       sizeof(struct ceph_filelock),
3769 					       GFP_NOFS);
3770 			if (!flocks) {
3771 				err = -ENOMEM;
3772 				goto out_err;
3773 			}
3774 			err = ceph_encode_locks_to_buffer(inode, flocks,
3775 							  num_fcntl_locks,
3776 							  num_flock_locks);
3777 			if (err) {
3778 				kfree(flocks);
3779 				flocks = NULL;
3780 				if (err == -ENOSPC)
3781 					goto encode_again;
3782 				goto out_err;
3783 			}
3784 		} else {
3785 			kfree(flocks);
3786 			flocks = NULL;
3787 		}
3788 
3789 		if (recon_state->msg_version >= 3) {
3790 			/* version, compat_version and struct_len */
3791 			total_len += 2 * sizeof(u8) + sizeof(u32);
3792 			struct_v = 2;
3793 		}
3794 		/*
3795 		 * number of encoded locks is stable, so copy to pagelist
3796 		 */
3797 		struct_len = 2 * sizeof(u32) +
3798 			    (num_fcntl_locks + num_flock_locks) *
3799 			    sizeof(struct ceph_filelock);
3800 		rec.v2.flock_len = cpu_to_le32(struct_len);
3801 
3802 		struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3803 
3804 		if (struct_v >= 2)
3805 			struct_len += sizeof(u64); /* snap_follows */
3806 
3807 		total_len += struct_len;
3808 
3809 		if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3810 			err = send_reconnect_partial(recon_state);
3811 			if (err)
3812 				goto out_freeflocks;
3813 			pagelist = recon_state->pagelist;
3814 		}
3815 
3816 		err = ceph_pagelist_reserve(pagelist, total_len);
3817 		if (err)
3818 			goto out_freeflocks;
3819 
3820 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3821 		if (recon_state->msg_version >= 3) {
3822 			ceph_pagelist_encode_8(pagelist, struct_v);
3823 			ceph_pagelist_encode_8(pagelist, 1);
3824 			ceph_pagelist_encode_32(pagelist, struct_len);
3825 		}
3826 		ceph_pagelist_encode_string(pagelist, path, pathlen);
3827 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3828 		ceph_locks_to_pagelist(flocks, pagelist,
3829 				       num_fcntl_locks, num_flock_locks);
3830 		if (struct_v >= 2)
3831 			ceph_pagelist_encode_64(pagelist, snap_follows);
3832 out_freeflocks:
3833 		kfree(flocks);
3834 	} else {
3835 		err = ceph_pagelist_reserve(pagelist,
3836 					    sizeof(u64) + sizeof(u32) +
3837 					    pathlen + sizeof(rec.v1));
3838 		if (err)
3839 			goto out_err;
3840 
3841 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3842 		ceph_pagelist_encode_string(pagelist, path, pathlen);
3843 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3844 	}
3845 
3846 out_err:
3847 	ceph_mdsc_free_path(path, pathlen);
3848 	if (!err)
3849 		recon_state->nr_caps++;
3850 	return err;
3851 }
3852 
3853 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3854 			      struct ceph_reconnect_state *recon_state)
3855 {
3856 	struct rb_node *p;
3857 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3858 	int err = 0;
3859 
3860 	if (recon_state->msg_version >= 4) {
3861 		err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3862 		if (err < 0)
3863 			goto fail;
3864 	}
3865 
3866 	/*
3867 	 * snaprealms.  we provide mds with the ino, seq (version), and
3868 	 * parent for all of our realms.  If the mds has any newer info,
3869 	 * it will tell us.
3870 	 */
3871 	for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3872 		struct ceph_snap_realm *realm =
3873 		       rb_entry(p, struct ceph_snap_realm, node);
3874 		struct ceph_mds_snaprealm_reconnect sr_rec;
3875 
3876 		if (recon_state->msg_version >= 4) {
3877 			size_t need = sizeof(u8) * 2 + sizeof(u32) +
3878 				      sizeof(sr_rec);
3879 
3880 			if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3881 				err = send_reconnect_partial(recon_state);
3882 				if (err)
3883 					goto fail;
3884 				pagelist = recon_state->pagelist;
3885 			}
3886 
3887 			err = ceph_pagelist_reserve(pagelist, need);
3888 			if (err)
3889 				goto fail;
3890 
3891 			ceph_pagelist_encode_8(pagelist, 1);
3892 			ceph_pagelist_encode_8(pagelist, 1);
3893 			ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3894 		}
3895 
3896 		dout(" adding snap realm %llx seq %lld parent %llx\n",
3897 		     realm->ino, realm->seq, realm->parent_ino);
3898 		sr_rec.ino = cpu_to_le64(realm->ino);
3899 		sr_rec.seq = cpu_to_le64(realm->seq);
3900 		sr_rec.parent = cpu_to_le64(realm->parent_ino);
3901 
3902 		err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3903 		if (err)
3904 			goto fail;
3905 
3906 		recon_state->nr_realms++;
3907 	}
3908 fail:
3909 	return err;
3910 }
3911 
3912 
3913 /*
3914  * If an MDS fails and recovers, clients need to reconnect in order to
3915  * reestablish shared state.  This includes all caps issued through
3916  * this session _and_ the snap_realm hierarchy.  Because it's not
3917  * clear which snap realms the mds cares about, we send everything we
3918  * know about.. that ensures we'll then get any new info the
3919  * recovering MDS might have.
3920  *
3921  * This is a relatively heavyweight operation, but it's rare.
3922  */
3923 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3924 			       struct ceph_mds_session *session)
3925 {
3926 	struct ceph_msg *reply;
3927 	int mds = session->s_mds;
3928 	int err = -ENOMEM;
3929 	struct ceph_reconnect_state recon_state = {
3930 		.session = session,
3931 	};
3932 	LIST_HEAD(dispose);
3933 
3934 	pr_info("mds%d reconnect start\n", mds);
3935 
3936 	recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3937 	if (!recon_state.pagelist)
3938 		goto fail_nopagelist;
3939 
3940 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3941 	if (!reply)
3942 		goto fail_nomsg;
3943 
3944 	xa_destroy(&session->s_delegated_inos);
3945 
3946 	mutex_lock(&session->s_mutex);
3947 	session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3948 	session->s_seq = 0;
3949 
3950 	dout("session %p state %s\n", session,
3951 	     ceph_session_state_name(session->s_state));
3952 
3953 	atomic_inc(&session->s_cap_gen);
3954 
3955 	spin_lock(&session->s_cap_lock);
3956 	/* don't know if session is readonly */
3957 	session->s_readonly = 0;
3958 	/*
3959 	 * notify __ceph_remove_cap() that we are composing cap reconnect.
3960 	 * If a cap get released before being added to the cap reconnect,
3961 	 * __ceph_remove_cap() should skip queuing cap release.
3962 	 */
3963 	session->s_cap_reconnect = 1;
3964 	/* drop old cap expires; we're about to reestablish that state */
3965 	detach_cap_releases(session, &dispose);
3966 	spin_unlock(&session->s_cap_lock);
3967 	dispose_cap_releases(mdsc, &dispose);
3968 
3969 	/* trim unused caps to reduce MDS's cache rejoin time */
3970 	if (mdsc->fsc->sb->s_root)
3971 		shrink_dcache_parent(mdsc->fsc->sb->s_root);
3972 
3973 	ceph_con_close(&session->s_con);
3974 	ceph_con_open(&session->s_con,
3975 		      CEPH_ENTITY_TYPE_MDS, mds,
3976 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3977 
3978 	/* replay unsafe requests */
3979 	replay_unsafe_requests(mdsc, session);
3980 
3981 	ceph_early_kick_flushing_caps(mdsc, session);
3982 
3983 	down_read(&mdsc->snap_rwsem);
3984 
3985 	/* placeholder for nr_caps */
3986 	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3987 	if (err)
3988 		goto fail;
3989 
3990 	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3991 		recon_state.msg_version = 3;
3992 		recon_state.allow_multi = true;
3993 	} else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3994 		recon_state.msg_version = 3;
3995 	} else {
3996 		recon_state.msg_version = 2;
3997 	}
3998 	/* trsaverse this session's caps */
3999 	err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4000 
4001 	spin_lock(&session->s_cap_lock);
4002 	session->s_cap_reconnect = 0;
4003 	spin_unlock(&session->s_cap_lock);
4004 
4005 	if (err < 0)
4006 		goto fail;
4007 
4008 	/* check if all realms can be encoded into current message */
4009 	if (mdsc->num_snap_realms) {
4010 		size_t total_len =
4011 			recon_state.pagelist->length +
4012 			mdsc->num_snap_realms *
4013 			sizeof(struct ceph_mds_snaprealm_reconnect);
4014 		if (recon_state.msg_version >= 4) {
4015 			/* number of realms */
4016 			total_len += sizeof(u32);
4017 			/* version, compat_version and struct_len */
4018 			total_len += mdsc->num_snap_realms *
4019 				     (2 * sizeof(u8) + sizeof(u32));
4020 		}
4021 		if (total_len > RECONNECT_MAX_SIZE) {
4022 			if (!recon_state.allow_multi) {
4023 				err = -ENOSPC;
4024 				goto fail;
4025 			}
4026 			if (recon_state.nr_caps) {
4027 				err = send_reconnect_partial(&recon_state);
4028 				if (err)
4029 					goto fail;
4030 			}
4031 			recon_state.msg_version = 5;
4032 		}
4033 	}
4034 
4035 	err = encode_snap_realms(mdsc, &recon_state);
4036 	if (err < 0)
4037 		goto fail;
4038 
4039 	if (recon_state.msg_version >= 5) {
4040 		err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4041 		if (err < 0)
4042 			goto fail;
4043 	}
4044 
4045 	if (recon_state.nr_caps || recon_state.nr_realms) {
4046 		struct page *page =
4047 			list_first_entry(&recon_state.pagelist->head,
4048 					struct page, lru);
4049 		__le32 *addr = kmap_atomic(page);
4050 		if (recon_state.nr_caps) {
4051 			WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4052 			*addr = cpu_to_le32(recon_state.nr_caps);
4053 		} else if (recon_state.msg_version >= 4) {
4054 			*(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4055 		}
4056 		kunmap_atomic(addr);
4057 	}
4058 
4059 	reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4060 	if (recon_state.msg_version >= 4)
4061 		reply->hdr.compat_version = cpu_to_le16(4);
4062 
4063 	reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4064 	ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4065 
4066 	ceph_con_send(&session->s_con, reply);
4067 
4068 	mutex_unlock(&session->s_mutex);
4069 
4070 	mutex_lock(&mdsc->mutex);
4071 	__wake_requests(mdsc, &session->s_waiting);
4072 	mutex_unlock(&mdsc->mutex);
4073 
4074 	up_read(&mdsc->snap_rwsem);
4075 	ceph_pagelist_release(recon_state.pagelist);
4076 	return;
4077 
4078 fail:
4079 	ceph_msg_put(reply);
4080 	up_read(&mdsc->snap_rwsem);
4081 	mutex_unlock(&session->s_mutex);
4082 fail_nomsg:
4083 	ceph_pagelist_release(recon_state.pagelist);
4084 fail_nopagelist:
4085 	pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4086 	return;
4087 }
4088 
4089 
4090 /*
4091  * compare old and new mdsmaps, kicking requests
4092  * and closing out old connections as necessary
4093  *
4094  * called under mdsc->mutex.
4095  */
4096 static void check_new_map(struct ceph_mds_client *mdsc,
4097 			  struct ceph_mdsmap *newmap,
4098 			  struct ceph_mdsmap *oldmap)
4099 {
4100 	int i, j, err;
4101 	int oldstate, newstate;
4102 	struct ceph_mds_session *s;
4103 	unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
4104 
4105 	dout("check_new_map new %u old %u\n",
4106 	     newmap->m_epoch, oldmap->m_epoch);
4107 
4108 	if (newmap->m_info) {
4109 		for (i = 0; i < newmap->possible_max_rank; i++) {
4110 			for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
4111 				set_bit(newmap->m_info[i].export_targets[j], targets);
4112 		}
4113 	}
4114 
4115 	for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4116 		if (!mdsc->sessions[i])
4117 			continue;
4118 		s = mdsc->sessions[i];
4119 		oldstate = ceph_mdsmap_get_state(oldmap, i);
4120 		newstate = ceph_mdsmap_get_state(newmap, i);
4121 
4122 		dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4123 		     i, ceph_mds_state_name(oldstate),
4124 		     ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4125 		     ceph_mds_state_name(newstate),
4126 		     ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4127 		     ceph_session_state_name(s->s_state));
4128 
4129 		if (i >= newmap->possible_max_rank) {
4130 			/* force close session for stopped mds */
4131 			ceph_get_mds_session(s);
4132 			__unregister_session(mdsc, s);
4133 			__wake_requests(mdsc, &s->s_waiting);
4134 			mutex_unlock(&mdsc->mutex);
4135 
4136 			mutex_lock(&s->s_mutex);
4137 			cleanup_session_requests(mdsc, s);
4138 			remove_session_caps(s);
4139 			mutex_unlock(&s->s_mutex);
4140 
4141 			ceph_put_mds_session(s);
4142 
4143 			mutex_lock(&mdsc->mutex);
4144 			kick_requests(mdsc, i);
4145 			continue;
4146 		}
4147 
4148 		if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4149 			   ceph_mdsmap_get_addr(newmap, i),
4150 			   sizeof(struct ceph_entity_addr))) {
4151 			/* just close it */
4152 			mutex_unlock(&mdsc->mutex);
4153 			mutex_lock(&s->s_mutex);
4154 			mutex_lock(&mdsc->mutex);
4155 			ceph_con_close(&s->s_con);
4156 			mutex_unlock(&s->s_mutex);
4157 			s->s_state = CEPH_MDS_SESSION_RESTARTING;
4158 		} else if (oldstate == newstate) {
4159 			continue;  /* nothing new with this mds */
4160 		}
4161 
4162 		/*
4163 		 * send reconnect?
4164 		 */
4165 		if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4166 		    newstate >= CEPH_MDS_STATE_RECONNECT) {
4167 			mutex_unlock(&mdsc->mutex);
4168 			clear_bit(i, targets);
4169 			send_mds_reconnect(mdsc, s);
4170 			mutex_lock(&mdsc->mutex);
4171 		}
4172 
4173 		/*
4174 		 * kick request on any mds that has gone active.
4175 		 */
4176 		if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4177 		    newstate >= CEPH_MDS_STATE_ACTIVE) {
4178 			if (oldstate != CEPH_MDS_STATE_CREATING &&
4179 			    oldstate != CEPH_MDS_STATE_STARTING)
4180 				pr_info("mds%d recovery completed\n", s->s_mds);
4181 			kick_requests(mdsc, i);
4182 			mutex_unlock(&mdsc->mutex);
4183 			mutex_lock(&s->s_mutex);
4184 			mutex_lock(&mdsc->mutex);
4185 			ceph_kick_flushing_caps(mdsc, s);
4186 			mutex_unlock(&s->s_mutex);
4187 			wake_up_session_caps(s, RECONNECT);
4188 		}
4189 	}
4190 
4191 	/*
4192 	 * Only open and reconnect sessions that don't exist yet.
4193 	 */
4194 	for (i = 0; i < newmap->possible_max_rank; i++) {
4195 		/*
4196 		 * In case the import MDS is crashed just after
4197 		 * the EImportStart journal is flushed, so when
4198 		 * a standby MDS takes over it and is replaying
4199 		 * the EImportStart journal the new MDS daemon
4200 		 * will wait the client to reconnect it, but the
4201 		 * client may never register/open the session yet.
4202 		 *
4203 		 * Will try to reconnect that MDS daemon if the
4204 		 * rank number is in the export targets array and
4205 		 * is the up:reconnect state.
4206 		 */
4207 		newstate = ceph_mdsmap_get_state(newmap, i);
4208 		if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
4209 			continue;
4210 
4211 		/*
4212 		 * The session maybe registered and opened by some
4213 		 * requests which were choosing random MDSes during
4214 		 * the mdsc->mutex's unlock/lock gap below in rare
4215 		 * case. But the related MDS daemon will just queue
4216 		 * that requests and be still waiting for the client's
4217 		 * reconnection request in up:reconnect state.
4218 		 */
4219 		s = __ceph_lookup_mds_session(mdsc, i);
4220 		if (likely(!s)) {
4221 			s = __open_export_target_session(mdsc, i);
4222 			if (IS_ERR(s)) {
4223 				err = PTR_ERR(s);
4224 				pr_err("failed to open export target session, err %d\n",
4225 				       err);
4226 				continue;
4227 			}
4228 		}
4229 		dout("send reconnect to export target mds.%d\n", i);
4230 		mutex_unlock(&mdsc->mutex);
4231 		send_mds_reconnect(mdsc, s);
4232 		ceph_put_mds_session(s);
4233 		mutex_lock(&mdsc->mutex);
4234 	}
4235 
4236 	for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4237 		s = mdsc->sessions[i];
4238 		if (!s)
4239 			continue;
4240 		if (!ceph_mdsmap_is_laggy(newmap, i))
4241 			continue;
4242 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4243 		    s->s_state == CEPH_MDS_SESSION_HUNG ||
4244 		    s->s_state == CEPH_MDS_SESSION_CLOSING) {
4245 			dout(" connecting to export targets of laggy mds%d\n",
4246 			     i);
4247 			__open_export_target_sessions(mdsc, s);
4248 		}
4249 	}
4250 }
4251 
4252 
4253 
4254 /*
4255  * leases
4256  */
4257 
4258 /*
4259  * caller must hold session s_mutex, dentry->d_lock
4260  */
4261 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4262 {
4263 	struct ceph_dentry_info *di = ceph_dentry(dentry);
4264 
4265 	ceph_put_mds_session(di->lease_session);
4266 	di->lease_session = NULL;
4267 }
4268 
4269 static void handle_lease(struct ceph_mds_client *mdsc,
4270 			 struct ceph_mds_session *session,
4271 			 struct ceph_msg *msg)
4272 {
4273 	struct super_block *sb = mdsc->fsc->sb;
4274 	struct inode *inode;
4275 	struct dentry *parent, *dentry;
4276 	struct ceph_dentry_info *di;
4277 	int mds = session->s_mds;
4278 	struct ceph_mds_lease *h = msg->front.iov_base;
4279 	u32 seq;
4280 	struct ceph_vino vino;
4281 	struct qstr dname;
4282 	int release = 0;
4283 
4284 	dout("handle_lease from mds%d\n", mds);
4285 
4286 	/* decode */
4287 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4288 		goto bad;
4289 	vino.ino = le64_to_cpu(h->ino);
4290 	vino.snap = CEPH_NOSNAP;
4291 	seq = le32_to_cpu(h->seq);
4292 	dname.len = get_unaligned_le32(h + 1);
4293 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4294 		goto bad;
4295 	dname.name = (void *)(h + 1) + sizeof(u32);
4296 
4297 	/* lookup inode */
4298 	inode = ceph_find_inode(sb, vino);
4299 	dout("handle_lease %s, ino %llx %p %.*s\n",
4300 	     ceph_lease_op_name(h->action), vino.ino, inode,
4301 	     dname.len, dname.name);
4302 
4303 	mutex_lock(&session->s_mutex);
4304 	inc_session_sequence(session);
4305 
4306 	if (!inode) {
4307 		dout("handle_lease no inode %llx\n", vino.ino);
4308 		goto release;
4309 	}
4310 
4311 	/* dentry */
4312 	parent = d_find_alias(inode);
4313 	if (!parent) {
4314 		dout("no parent dentry on inode %p\n", inode);
4315 		WARN_ON(1);
4316 		goto release;  /* hrm... */
4317 	}
4318 	dname.hash = full_name_hash(parent, dname.name, dname.len);
4319 	dentry = d_lookup(parent, &dname);
4320 	dput(parent);
4321 	if (!dentry)
4322 		goto release;
4323 
4324 	spin_lock(&dentry->d_lock);
4325 	di = ceph_dentry(dentry);
4326 	switch (h->action) {
4327 	case CEPH_MDS_LEASE_REVOKE:
4328 		if (di->lease_session == session) {
4329 			if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4330 				h->seq = cpu_to_le32(di->lease_seq);
4331 			__ceph_mdsc_drop_dentry_lease(dentry);
4332 		}
4333 		release = 1;
4334 		break;
4335 
4336 	case CEPH_MDS_LEASE_RENEW:
4337 		if (di->lease_session == session &&
4338 		    di->lease_gen == atomic_read(&session->s_cap_gen) &&
4339 		    di->lease_renew_from &&
4340 		    di->lease_renew_after == 0) {
4341 			unsigned long duration =
4342 				msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4343 
4344 			di->lease_seq = seq;
4345 			di->time = di->lease_renew_from + duration;
4346 			di->lease_renew_after = di->lease_renew_from +
4347 				(duration >> 1);
4348 			di->lease_renew_from = 0;
4349 		}
4350 		break;
4351 	}
4352 	spin_unlock(&dentry->d_lock);
4353 	dput(dentry);
4354 
4355 	if (!release)
4356 		goto out;
4357 
4358 release:
4359 	/* let's just reuse the same message */
4360 	h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4361 	ceph_msg_get(msg);
4362 	ceph_con_send(&session->s_con, msg);
4363 
4364 out:
4365 	mutex_unlock(&session->s_mutex);
4366 	iput(inode);
4367 	return;
4368 
4369 bad:
4370 	pr_err("corrupt lease message\n");
4371 	ceph_msg_dump(msg);
4372 }
4373 
4374 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4375 			      struct dentry *dentry, char action,
4376 			      u32 seq)
4377 {
4378 	struct ceph_msg *msg;
4379 	struct ceph_mds_lease *lease;
4380 	struct inode *dir;
4381 	int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4382 
4383 	dout("lease_send_msg identry %p %s to mds%d\n",
4384 	     dentry, ceph_lease_op_name(action), session->s_mds);
4385 
4386 	msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4387 	if (!msg)
4388 		return;
4389 	lease = msg->front.iov_base;
4390 	lease->action = action;
4391 	lease->seq = cpu_to_le32(seq);
4392 
4393 	spin_lock(&dentry->d_lock);
4394 	dir = d_inode(dentry->d_parent);
4395 	lease->ino = cpu_to_le64(ceph_ino(dir));
4396 	lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4397 
4398 	put_unaligned_le32(dentry->d_name.len, lease + 1);
4399 	memcpy((void *)(lease + 1) + 4,
4400 	       dentry->d_name.name, dentry->d_name.len);
4401 	spin_unlock(&dentry->d_lock);
4402 	/*
4403 	 * if this is a preemptive lease RELEASE, no need to
4404 	 * flush request stream, since the actual request will
4405 	 * soon follow.
4406 	 */
4407 	msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4408 
4409 	ceph_con_send(&session->s_con, msg);
4410 }
4411 
4412 /*
4413  * lock unlock the session, to wait ongoing session activities
4414  */
4415 static void lock_unlock_session(struct ceph_mds_session *s)
4416 {
4417 	mutex_lock(&s->s_mutex);
4418 	mutex_unlock(&s->s_mutex);
4419 }
4420 
4421 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4422 {
4423 	struct ceph_fs_client *fsc = mdsc->fsc;
4424 
4425 	if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4426 		return;
4427 
4428 	if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4429 		return;
4430 
4431 	if (!READ_ONCE(fsc->blocklisted))
4432 		return;
4433 
4434 	pr_info("auto reconnect after blocklisted\n");
4435 	ceph_force_reconnect(fsc->sb);
4436 }
4437 
4438 bool check_session_state(struct ceph_mds_session *s)
4439 {
4440 	struct ceph_fs_client *fsc = s->s_mdsc->fsc;
4441 
4442 	switch (s->s_state) {
4443 	case CEPH_MDS_SESSION_OPEN:
4444 		if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4445 			s->s_state = CEPH_MDS_SESSION_HUNG;
4446 			pr_info("mds%d hung\n", s->s_mds);
4447 		}
4448 		break;
4449 	case CEPH_MDS_SESSION_CLOSING:
4450 		/* Should never reach this when not force unmounting */
4451 		WARN_ON_ONCE(s->s_ttl &&
4452 			     READ_ONCE(fsc->mount_state) != CEPH_MOUNT_SHUTDOWN);
4453 		fallthrough;
4454 	case CEPH_MDS_SESSION_NEW:
4455 	case CEPH_MDS_SESSION_RESTARTING:
4456 	case CEPH_MDS_SESSION_CLOSED:
4457 	case CEPH_MDS_SESSION_REJECTED:
4458 		return false;
4459 	}
4460 
4461 	return true;
4462 }
4463 
4464 /*
4465  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4466  * then we need to retransmit that request.
4467  */
4468 void inc_session_sequence(struct ceph_mds_session *s)
4469 {
4470 	lockdep_assert_held(&s->s_mutex);
4471 
4472 	s->s_seq++;
4473 
4474 	if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4475 		int ret;
4476 
4477 		dout("resending session close request for mds%d\n", s->s_mds);
4478 		ret = request_close_session(s);
4479 		if (ret < 0)
4480 			pr_err("unable to close session to mds%d: %d\n",
4481 			       s->s_mds, ret);
4482 	}
4483 }
4484 
4485 /*
4486  * delayed work -- periodically trim expired leases, renew caps with mds.  If
4487  * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4488  * workqueue delay value of 5 secs will be used.
4489  */
4490 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4491 {
4492 	unsigned long max_delay = HZ * 5;
4493 
4494 	/* 5 secs default delay */
4495 	if (!delay || (delay > max_delay))
4496 		delay = max_delay;
4497 	schedule_delayed_work(&mdsc->delayed_work,
4498 			      round_jiffies_relative(delay));
4499 }
4500 
4501 static void delayed_work(struct work_struct *work)
4502 {
4503 	struct ceph_mds_client *mdsc =
4504 		container_of(work, struct ceph_mds_client, delayed_work.work);
4505 	unsigned long delay;
4506 	int renew_interval;
4507 	int renew_caps;
4508 	int i;
4509 
4510 	dout("mdsc delayed_work\n");
4511 
4512 	if (mdsc->stopping)
4513 		return;
4514 
4515 	mutex_lock(&mdsc->mutex);
4516 	renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4517 	renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4518 				   mdsc->last_renew_caps);
4519 	if (renew_caps)
4520 		mdsc->last_renew_caps = jiffies;
4521 
4522 	for (i = 0; i < mdsc->max_sessions; i++) {
4523 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4524 		if (!s)
4525 			continue;
4526 
4527 		if (!check_session_state(s)) {
4528 			ceph_put_mds_session(s);
4529 			continue;
4530 		}
4531 		mutex_unlock(&mdsc->mutex);
4532 
4533 		mutex_lock(&s->s_mutex);
4534 		if (renew_caps)
4535 			send_renew_caps(mdsc, s);
4536 		else
4537 			ceph_con_keepalive(&s->s_con);
4538 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4539 		    s->s_state == CEPH_MDS_SESSION_HUNG)
4540 			ceph_send_cap_releases(mdsc, s);
4541 		mutex_unlock(&s->s_mutex);
4542 		ceph_put_mds_session(s);
4543 
4544 		mutex_lock(&mdsc->mutex);
4545 	}
4546 	mutex_unlock(&mdsc->mutex);
4547 
4548 	delay = ceph_check_delayed_caps(mdsc);
4549 
4550 	ceph_queue_cap_reclaim_work(mdsc);
4551 
4552 	ceph_trim_snapid_map(mdsc);
4553 
4554 	maybe_recover_session(mdsc);
4555 
4556 	schedule_delayed(mdsc, delay);
4557 }
4558 
4559 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4560 
4561 {
4562 	struct ceph_mds_client *mdsc;
4563 	int err;
4564 
4565 	mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4566 	if (!mdsc)
4567 		return -ENOMEM;
4568 	mdsc->fsc = fsc;
4569 	mutex_init(&mdsc->mutex);
4570 	mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4571 	if (!mdsc->mdsmap) {
4572 		err = -ENOMEM;
4573 		goto err_mdsc;
4574 	}
4575 
4576 	init_completion(&mdsc->safe_umount_waiters);
4577 	init_waitqueue_head(&mdsc->session_close_wq);
4578 	INIT_LIST_HEAD(&mdsc->waiting_for_map);
4579 	mdsc->quotarealms_inodes = RB_ROOT;
4580 	mutex_init(&mdsc->quotarealms_inodes_mutex);
4581 	init_rwsem(&mdsc->snap_rwsem);
4582 	mdsc->snap_realms = RB_ROOT;
4583 	INIT_LIST_HEAD(&mdsc->snap_empty);
4584 	spin_lock_init(&mdsc->snap_empty_lock);
4585 	mdsc->request_tree = RB_ROOT;
4586 	INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4587 	mdsc->last_renew_caps = jiffies;
4588 	INIT_LIST_HEAD(&mdsc->cap_delay_list);
4589 	INIT_LIST_HEAD(&mdsc->cap_wait_list);
4590 	spin_lock_init(&mdsc->cap_delay_lock);
4591 	INIT_LIST_HEAD(&mdsc->snap_flush_list);
4592 	spin_lock_init(&mdsc->snap_flush_lock);
4593 	mdsc->last_cap_flush_tid = 1;
4594 	INIT_LIST_HEAD(&mdsc->cap_flush_list);
4595 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4596 	spin_lock_init(&mdsc->cap_dirty_lock);
4597 	init_waitqueue_head(&mdsc->cap_flushing_wq);
4598 	INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4599 	err = ceph_metric_init(&mdsc->metric);
4600 	if (err)
4601 		goto err_mdsmap;
4602 
4603 	spin_lock_init(&mdsc->dentry_list_lock);
4604 	INIT_LIST_HEAD(&mdsc->dentry_leases);
4605 	INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4606 
4607 	ceph_caps_init(mdsc);
4608 	ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4609 
4610 	spin_lock_init(&mdsc->snapid_map_lock);
4611 	mdsc->snapid_map_tree = RB_ROOT;
4612 	INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4613 
4614 	init_rwsem(&mdsc->pool_perm_rwsem);
4615 	mdsc->pool_perm_tree = RB_ROOT;
4616 
4617 	strscpy(mdsc->nodename, utsname()->nodename,
4618 		sizeof(mdsc->nodename));
4619 
4620 	fsc->mdsc = mdsc;
4621 	return 0;
4622 
4623 err_mdsmap:
4624 	kfree(mdsc->mdsmap);
4625 err_mdsc:
4626 	kfree(mdsc);
4627 	return err;
4628 }
4629 
4630 /*
4631  * Wait for safe replies on open mds requests.  If we time out, drop
4632  * all requests from the tree to avoid dangling dentry refs.
4633  */
4634 static void wait_requests(struct ceph_mds_client *mdsc)
4635 {
4636 	struct ceph_options *opts = mdsc->fsc->client->options;
4637 	struct ceph_mds_request *req;
4638 
4639 	mutex_lock(&mdsc->mutex);
4640 	if (__get_oldest_req(mdsc)) {
4641 		mutex_unlock(&mdsc->mutex);
4642 
4643 		dout("wait_requests waiting for requests\n");
4644 		wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4645 				    ceph_timeout_jiffies(opts->mount_timeout));
4646 
4647 		/* tear down remaining requests */
4648 		mutex_lock(&mdsc->mutex);
4649 		while ((req = __get_oldest_req(mdsc))) {
4650 			dout("wait_requests timed out on tid %llu\n",
4651 			     req->r_tid);
4652 			list_del_init(&req->r_wait);
4653 			__unregister_request(mdsc, req);
4654 		}
4655 	}
4656 	mutex_unlock(&mdsc->mutex);
4657 	dout("wait_requests done\n");
4658 }
4659 
4660 void send_flush_mdlog(struct ceph_mds_session *s)
4661 {
4662 	struct ceph_msg *msg;
4663 
4664 	/*
4665 	 * Pre-luminous MDS crashes when it sees an unknown session request
4666 	 */
4667 	if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
4668 		return;
4669 
4670 	mutex_lock(&s->s_mutex);
4671 	dout("request mdlog flush to mds%d (%s)s seq %lld\n", s->s_mds,
4672 	     ceph_session_state_name(s->s_state), s->s_seq);
4673 	msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
4674 				      s->s_seq);
4675 	if (!msg) {
4676 		pr_err("failed to request mdlog flush to mds%d (%s) seq %lld\n",
4677 		       s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
4678 	} else {
4679 		ceph_con_send(&s->s_con, msg);
4680 	}
4681 	mutex_unlock(&s->s_mutex);
4682 }
4683 
4684 /*
4685  * called before mount is ro, and before dentries are torn down.
4686  * (hmm, does this still race with new lookups?)
4687  */
4688 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4689 {
4690 	dout("pre_umount\n");
4691 	mdsc->stopping = 1;
4692 
4693 	ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
4694 	ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
4695 	ceph_flush_dirty_caps(mdsc);
4696 	wait_requests(mdsc);
4697 
4698 	/*
4699 	 * wait for reply handlers to drop their request refs and
4700 	 * their inode/dcache refs
4701 	 */
4702 	ceph_msgr_flush();
4703 
4704 	ceph_cleanup_quotarealms_inodes(mdsc);
4705 }
4706 
4707 /*
4708  * wait for all write mds requests to flush.
4709  */
4710 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4711 {
4712 	struct ceph_mds_request *req = NULL, *nextreq;
4713 	struct rb_node *n;
4714 
4715 	mutex_lock(&mdsc->mutex);
4716 	dout("wait_unsafe_requests want %lld\n", want_tid);
4717 restart:
4718 	req = __get_oldest_req(mdsc);
4719 	while (req && req->r_tid <= want_tid) {
4720 		/* find next request */
4721 		n = rb_next(&req->r_node);
4722 		if (n)
4723 			nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4724 		else
4725 			nextreq = NULL;
4726 		if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4727 		    (req->r_op & CEPH_MDS_OP_WRITE)) {
4728 			/* write op */
4729 			ceph_mdsc_get_request(req);
4730 			if (nextreq)
4731 				ceph_mdsc_get_request(nextreq);
4732 			mutex_unlock(&mdsc->mutex);
4733 			dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4734 			     req->r_tid, want_tid);
4735 			wait_for_completion(&req->r_safe_completion);
4736 			mutex_lock(&mdsc->mutex);
4737 			ceph_mdsc_put_request(req);
4738 			if (!nextreq)
4739 				break;  /* next dne before, so we're done! */
4740 			if (RB_EMPTY_NODE(&nextreq->r_node)) {
4741 				/* next request was removed from tree */
4742 				ceph_mdsc_put_request(nextreq);
4743 				goto restart;
4744 			}
4745 			ceph_mdsc_put_request(nextreq);  /* won't go away */
4746 		}
4747 		req = nextreq;
4748 	}
4749 	mutex_unlock(&mdsc->mutex);
4750 	dout("wait_unsafe_requests done\n");
4751 }
4752 
4753 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4754 {
4755 	u64 want_tid, want_flush;
4756 
4757 	if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4758 		return;
4759 
4760 	dout("sync\n");
4761 	mutex_lock(&mdsc->mutex);
4762 	want_tid = mdsc->last_tid;
4763 	mutex_unlock(&mdsc->mutex);
4764 
4765 	ceph_flush_dirty_caps(mdsc);
4766 	spin_lock(&mdsc->cap_dirty_lock);
4767 	want_flush = mdsc->last_cap_flush_tid;
4768 	if (!list_empty(&mdsc->cap_flush_list)) {
4769 		struct ceph_cap_flush *cf =
4770 			list_last_entry(&mdsc->cap_flush_list,
4771 					struct ceph_cap_flush, g_list);
4772 		cf->wake = true;
4773 	}
4774 	spin_unlock(&mdsc->cap_dirty_lock);
4775 
4776 	dout("sync want tid %lld flush_seq %lld\n",
4777 	     want_tid, want_flush);
4778 
4779 	wait_unsafe_requests(mdsc, want_tid);
4780 	wait_caps_flush(mdsc, want_flush);
4781 }
4782 
4783 /*
4784  * true if all sessions are closed, or we force unmount
4785  */
4786 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4787 {
4788 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4789 		return true;
4790 	return atomic_read(&mdsc->num_sessions) <= skipped;
4791 }
4792 
4793 /*
4794  * called after sb is ro.
4795  */
4796 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4797 {
4798 	struct ceph_options *opts = mdsc->fsc->client->options;
4799 	struct ceph_mds_session *session;
4800 	int i;
4801 	int skipped = 0;
4802 
4803 	dout("close_sessions\n");
4804 
4805 	/* close sessions */
4806 	mutex_lock(&mdsc->mutex);
4807 	for (i = 0; i < mdsc->max_sessions; i++) {
4808 		session = __ceph_lookup_mds_session(mdsc, i);
4809 		if (!session)
4810 			continue;
4811 		mutex_unlock(&mdsc->mutex);
4812 		mutex_lock(&session->s_mutex);
4813 		if (__close_session(mdsc, session) <= 0)
4814 			skipped++;
4815 		mutex_unlock(&session->s_mutex);
4816 		ceph_put_mds_session(session);
4817 		mutex_lock(&mdsc->mutex);
4818 	}
4819 	mutex_unlock(&mdsc->mutex);
4820 
4821 	dout("waiting for sessions to close\n");
4822 	wait_event_timeout(mdsc->session_close_wq,
4823 			   done_closing_sessions(mdsc, skipped),
4824 			   ceph_timeout_jiffies(opts->mount_timeout));
4825 
4826 	/* tear down remaining sessions */
4827 	mutex_lock(&mdsc->mutex);
4828 	for (i = 0; i < mdsc->max_sessions; i++) {
4829 		if (mdsc->sessions[i]) {
4830 			session = ceph_get_mds_session(mdsc->sessions[i]);
4831 			__unregister_session(mdsc, session);
4832 			mutex_unlock(&mdsc->mutex);
4833 			mutex_lock(&session->s_mutex);
4834 			remove_session_caps(session);
4835 			mutex_unlock(&session->s_mutex);
4836 			ceph_put_mds_session(session);
4837 			mutex_lock(&mdsc->mutex);
4838 		}
4839 	}
4840 	WARN_ON(!list_empty(&mdsc->cap_delay_list));
4841 	mutex_unlock(&mdsc->mutex);
4842 
4843 	ceph_cleanup_snapid_map(mdsc);
4844 	ceph_cleanup_empty_realms(mdsc);
4845 
4846 	cancel_work_sync(&mdsc->cap_reclaim_work);
4847 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4848 
4849 	dout("stopped\n");
4850 }
4851 
4852 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4853 {
4854 	struct ceph_mds_session *session;
4855 	int mds;
4856 
4857 	dout("force umount\n");
4858 
4859 	mutex_lock(&mdsc->mutex);
4860 	for (mds = 0; mds < mdsc->max_sessions; mds++) {
4861 		session = __ceph_lookup_mds_session(mdsc, mds);
4862 		if (!session)
4863 			continue;
4864 
4865 		if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4866 			__unregister_session(mdsc, session);
4867 		__wake_requests(mdsc, &session->s_waiting);
4868 		mutex_unlock(&mdsc->mutex);
4869 
4870 		mutex_lock(&session->s_mutex);
4871 		__close_session(mdsc, session);
4872 		if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4873 			cleanup_session_requests(mdsc, session);
4874 			remove_session_caps(session);
4875 		}
4876 		mutex_unlock(&session->s_mutex);
4877 		ceph_put_mds_session(session);
4878 
4879 		mutex_lock(&mdsc->mutex);
4880 		kick_requests(mdsc, mds);
4881 	}
4882 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4883 	mutex_unlock(&mdsc->mutex);
4884 }
4885 
4886 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4887 {
4888 	dout("stop\n");
4889 	/*
4890 	 * Make sure the delayed work stopped before releasing
4891 	 * the resources.
4892 	 *
4893 	 * Because the cancel_delayed_work_sync() will only
4894 	 * guarantee that the work finishes executing. But the
4895 	 * delayed work will re-arm itself again after that.
4896 	 */
4897 	flush_delayed_work(&mdsc->delayed_work);
4898 
4899 	if (mdsc->mdsmap)
4900 		ceph_mdsmap_destroy(mdsc->mdsmap);
4901 	kfree(mdsc->sessions);
4902 	ceph_caps_finalize(mdsc);
4903 	ceph_pool_perm_destroy(mdsc);
4904 }
4905 
4906 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4907 {
4908 	struct ceph_mds_client *mdsc = fsc->mdsc;
4909 	dout("mdsc_destroy %p\n", mdsc);
4910 
4911 	if (!mdsc)
4912 		return;
4913 
4914 	/* flush out any connection work with references to us */
4915 	ceph_msgr_flush();
4916 
4917 	ceph_mdsc_stop(mdsc);
4918 
4919 	ceph_metric_destroy(&mdsc->metric);
4920 
4921 	fsc->mdsc = NULL;
4922 	kfree(mdsc);
4923 	dout("mdsc_destroy %p done\n", mdsc);
4924 }
4925 
4926 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4927 {
4928 	struct ceph_fs_client *fsc = mdsc->fsc;
4929 	const char *mds_namespace = fsc->mount_options->mds_namespace;
4930 	void *p = msg->front.iov_base;
4931 	void *end = p + msg->front.iov_len;
4932 	u32 epoch;
4933 	u32 num_fs;
4934 	u32 mount_fscid = (u32)-1;
4935 	int err = -EINVAL;
4936 
4937 	ceph_decode_need(&p, end, sizeof(u32), bad);
4938 	epoch = ceph_decode_32(&p);
4939 
4940 	dout("handle_fsmap epoch %u\n", epoch);
4941 
4942 	/* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
4943 	ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
4944 
4945 	ceph_decode_32_safe(&p, end, num_fs, bad);
4946 	while (num_fs-- > 0) {
4947 		void *info_p, *info_end;
4948 		u32 info_len;
4949 		u32 fscid, namelen;
4950 
4951 		ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4952 		p += 2;		// info_v, info_cv
4953 		info_len = ceph_decode_32(&p);
4954 		ceph_decode_need(&p, end, info_len, bad);
4955 		info_p = p;
4956 		info_end = p + info_len;
4957 		p = info_end;
4958 
4959 		ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4960 		fscid = ceph_decode_32(&info_p);
4961 		namelen = ceph_decode_32(&info_p);
4962 		ceph_decode_need(&info_p, info_end, namelen, bad);
4963 
4964 		if (mds_namespace &&
4965 		    strlen(mds_namespace) == namelen &&
4966 		    !strncmp(mds_namespace, (char *)info_p, namelen)) {
4967 			mount_fscid = fscid;
4968 			break;
4969 		}
4970 	}
4971 
4972 	ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4973 	if (mount_fscid != (u32)-1) {
4974 		fsc->client->monc.fs_cluster_id = mount_fscid;
4975 		ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4976 				   0, true);
4977 		ceph_monc_renew_subs(&fsc->client->monc);
4978 	} else {
4979 		err = -ENOENT;
4980 		goto err_out;
4981 	}
4982 	return;
4983 
4984 bad:
4985 	pr_err("error decoding fsmap %d. Shutting down mount.\n", err);
4986 	ceph_umount_begin(mdsc->fsc->sb);
4987 err_out:
4988 	mutex_lock(&mdsc->mutex);
4989 	mdsc->mdsmap_err = err;
4990 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4991 	mutex_unlock(&mdsc->mutex);
4992 }
4993 
4994 /*
4995  * handle mds map update.
4996  */
4997 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4998 {
4999 	u32 epoch;
5000 	u32 maplen;
5001 	void *p = msg->front.iov_base;
5002 	void *end = p + msg->front.iov_len;
5003 	struct ceph_mdsmap *newmap, *oldmap;
5004 	struct ceph_fsid fsid;
5005 	int err = -EINVAL;
5006 
5007 	ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5008 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
5009 	if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5010 		return;
5011 	epoch = ceph_decode_32(&p);
5012 	maplen = ceph_decode_32(&p);
5013 	dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5014 
5015 	/* do we need it? */
5016 	mutex_lock(&mdsc->mutex);
5017 	if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5018 		dout("handle_map epoch %u <= our %u\n",
5019 		     epoch, mdsc->mdsmap->m_epoch);
5020 		mutex_unlock(&mdsc->mutex);
5021 		return;
5022 	}
5023 
5024 	newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5025 	if (IS_ERR(newmap)) {
5026 		err = PTR_ERR(newmap);
5027 		goto bad_unlock;
5028 	}
5029 
5030 	/* swap into place */
5031 	if (mdsc->mdsmap) {
5032 		oldmap = mdsc->mdsmap;
5033 		mdsc->mdsmap = newmap;
5034 		check_new_map(mdsc, newmap, oldmap);
5035 		ceph_mdsmap_destroy(oldmap);
5036 	} else {
5037 		mdsc->mdsmap = newmap;  /* first mds map */
5038 	}
5039 	mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5040 					MAX_LFS_FILESIZE);
5041 
5042 	__wake_requests(mdsc, &mdsc->waiting_for_map);
5043 	ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5044 			  mdsc->mdsmap->m_epoch);
5045 
5046 	mutex_unlock(&mdsc->mutex);
5047 	schedule_delayed(mdsc, 0);
5048 	return;
5049 
5050 bad_unlock:
5051 	mutex_unlock(&mdsc->mutex);
5052 bad:
5053 	pr_err("error decoding mdsmap %d. Shutting down mount.\n", err);
5054 	ceph_umount_begin(mdsc->fsc->sb);
5055 	return;
5056 }
5057 
5058 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5059 {
5060 	struct ceph_mds_session *s = con->private;
5061 
5062 	if (ceph_get_mds_session(s))
5063 		return con;
5064 	return NULL;
5065 }
5066 
5067 static void mds_put_con(struct ceph_connection *con)
5068 {
5069 	struct ceph_mds_session *s = con->private;
5070 
5071 	ceph_put_mds_session(s);
5072 }
5073 
5074 /*
5075  * if the client is unresponsive for long enough, the mds will kill
5076  * the session entirely.
5077  */
5078 static void mds_peer_reset(struct ceph_connection *con)
5079 {
5080 	struct ceph_mds_session *s = con->private;
5081 	struct ceph_mds_client *mdsc = s->s_mdsc;
5082 
5083 	pr_warn("mds%d closed our session\n", s->s_mds);
5084 	send_mds_reconnect(mdsc, s);
5085 }
5086 
5087 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5088 {
5089 	struct ceph_mds_session *s = con->private;
5090 	struct ceph_mds_client *mdsc = s->s_mdsc;
5091 	int type = le16_to_cpu(msg->hdr.type);
5092 
5093 	mutex_lock(&mdsc->mutex);
5094 	if (__verify_registered_session(mdsc, s) < 0) {
5095 		mutex_unlock(&mdsc->mutex);
5096 		goto out;
5097 	}
5098 	mutex_unlock(&mdsc->mutex);
5099 
5100 	switch (type) {
5101 	case CEPH_MSG_MDS_MAP:
5102 		ceph_mdsc_handle_mdsmap(mdsc, msg);
5103 		break;
5104 	case CEPH_MSG_FS_MAP_USER:
5105 		ceph_mdsc_handle_fsmap(mdsc, msg);
5106 		break;
5107 	case CEPH_MSG_CLIENT_SESSION:
5108 		handle_session(s, msg);
5109 		break;
5110 	case CEPH_MSG_CLIENT_REPLY:
5111 		handle_reply(s, msg);
5112 		break;
5113 	case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5114 		handle_forward(mdsc, s, msg);
5115 		break;
5116 	case CEPH_MSG_CLIENT_CAPS:
5117 		ceph_handle_caps(s, msg);
5118 		break;
5119 	case CEPH_MSG_CLIENT_SNAP:
5120 		ceph_handle_snap(mdsc, s, msg);
5121 		break;
5122 	case CEPH_MSG_CLIENT_LEASE:
5123 		handle_lease(mdsc, s, msg);
5124 		break;
5125 	case CEPH_MSG_CLIENT_QUOTA:
5126 		ceph_handle_quota(mdsc, s, msg);
5127 		break;
5128 
5129 	default:
5130 		pr_err("received unknown message type %d %s\n", type,
5131 		       ceph_msg_type_name(type));
5132 	}
5133 out:
5134 	ceph_msg_put(msg);
5135 }
5136 
5137 /*
5138  * authentication
5139  */
5140 
5141 /*
5142  * Note: returned pointer is the address of a structure that's
5143  * managed separately.  Caller must *not* attempt to free it.
5144  */
5145 static struct ceph_auth_handshake *
5146 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
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 	struct ceph_auth_handshake *auth = &s->s_auth;
5152 	int ret;
5153 
5154 	ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5155 					 force_new, proto, NULL, NULL);
5156 	if (ret)
5157 		return ERR_PTR(ret);
5158 
5159 	return auth;
5160 }
5161 
5162 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5163 				    void *challenge_buf, int challenge_buf_len)
5164 {
5165 	struct ceph_mds_session *s = con->private;
5166 	struct ceph_mds_client *mdsc = s->s_mdsc;
5167 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5168 
5169 	return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5170 					    challenge_buf, challenge_buf_len);
5171 }
5172 
5173 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5174 {
5175 	struct ceph_mds_session *s = con->private;
5176 	struct ceph_mds_client *mdsc = s->s_mdsc;
5177 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5178 	struct ceph_auth_handshake *auth = &s->s_auth;
5179 
5180 	return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5181 		auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5182 		NULL, NULL, NULL, NULL);
5183 }
5184 
5185 static int mds_invalidate_authorizer(struct ceph_connection *con)
5186 {
5187 	struct ceph_mds_session *s = con->private;
5188 	struct ceph_mds_client *mdsc = s->s_mdsc;
5189 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5190 
5191 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5192 
5193 	return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5194 }
5195 
5196 static int mds_get_auth_request(struct ceph_connection *con,
5197 				void *buf, int *buf_len,
5198 				void **authorizer, int *authorizer_len)
5199 {
5200 	struct ceph_mds_session *s = con->private;
5201 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5202 	struct ceph_auth_handshake *auth = &s->s_auth;
5203 	int ret;
5204 
5205 	ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5206 				       buf, buf_len);
5207 	if (ret)
5208 		return ret;
5209 
5210 	*authorizer = auth->authorizer_buf;
5211 	*authorizer_len = auth->authorizer_buf_len;
5212 	return 0;
5213 }
5214 
5215 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5216 				      void *reply, int reply_len,
5217 				      void *buf, int *buf_len,
5218 				      void **authorizer, int *authorizer_len)
5219 {
5220 	struct ceph_mds_session *s = con->private;
5221 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5222 	struct ceph_auth_handshake *auth = &s->s_auth;
5223 	int ret;
5224 
5225 	ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5226 					      buf, buf_len);
5227 	if (ret)
5228 		return ret;
5229 
5230 	*authorizer = auth->authorizer_buf;
5231 	*authorizer_len = auth->authorizer_buf_len;
5232 	return 0;
5233 }
5234 
5235 static int mds_handle_auth_done(struct ceph_connection *con,
5236 				u64 global_id, void *reply, int reply_len,
5237 				u8 *session_key, int *session_key_len,
5238 				u8 *con_secret, int *con_secret_len)
5239 {
5240 	struct ceph_mds_session *s = con->private;
5241 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5242 	struct ceph_auth_handshake *auth = &s->s_auth;
5243 
5244 	return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5245 					       session_key, session_key_len,
5246 					       con_secret, con_secret_len);
5247 }
5248 
5249 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5250 				      int used_proto, int result,
5251 				      const int *allowed_protos, int proto_cnt,
5252 				      const int *allowed_modes, int mode_cnt)
5253 {
5254 	struct ceph_mds_session *s = con->private;
5255 	struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5256 	int ret;
5257 
5258 	if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5259 					    used_proto, result,
5260 					    allowed_protos, proto_cnt,
5261 					    allowed_modes, mode_cnt)) {
5262 		ret = ceph_monc_validate_auth(monc);
5263 		if (ret)
5264 			return ret;
5265 	}
5266 
5267 	return -EACCES;
5268 }
5269 
5270 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5271 				struct ceph_msg_header *hdr, int *skip)
5272 {
5273 	struct ceph_msg *msg;
5274 	int type = (int) le16_to_cpu(hdr->type);
5275 	int front_len = (int) le32_to_cpu(hdr->front_len);
5276 
5277 	if (con->in_msg)
5278 		return con->in_msg;
5279 
5280 	*skip = 0;
5281 	msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5282 	if (!msg) {
5283 		pr_err("unable to allocate msg type %d len %d\n",
5284 		       type, front_len);
5285 		return NULL;
5286 	}
5287 
5288 	return msg;
5289 }
5290 
5291 static int mds_sign_message(struct ceph_msg *msg)
5292 {
5293        struct ceph_mds_session *s = msg->con->private;
5294        struct ceph_auth_handshake *auth = &s->s_auth;
5295 
5296        return ceph_auth_sign_message(auth, msg);
5297 }
5298 
5299 static int mds_check_message_signature(struct ceph_msg *msg)
5300 {
5301        struct ceph_mds_session *s = msg->con->private;
5302        struct ceph_auth_handshake *auth = &s->s_auth;
5303 
5304        return ceph_auth_check_message_signature(auth, msg);
5305 }
5306 
5307 static const struct ceph_connection_operations mds_con_ops = {
5308 	.get = mds_get_con,
5309 	.put = mds_put_con,
5310 	.alloc_msg = mds_alloc_msg,
5311 	.dispatch = mds_dispatch,
5312 	.peer_reset = mds_peer_reset,
5313 	.get_authorizer = mds_get_authorizer,
5314 	.add_authorizer_challenge = mds_add_authorizer_challenge,
5315 	.verify_authorizer_reply = mds_verify_authorizer_reply,
5316 	.invalidate_authorizer = mds_invalidate_authorizer,
5317 	.sign_message = mds_sign_message,
5318 	.check_message_signature = mds_check_message_signature,
5319 	.get_auth_request = mds_get_auth_request,
5320 	.handle_auth_reply_more = mds_handle_auth_reply_more,
5321 	.handle_auth_done = mds_handle_auth_done,
5322 	.handle_auth_bad_method = mds_handle_auth_bad_method,
5323 };
5324 
5325 /* eof */
5326