xref: /openbmc/linux/fs/ceph/mds_client.c (revision c4a11bf4)
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50  * are no longer valid.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_capsnaps(struct ceph_mds_client *mdsc, struct inode *inode)
1594 {
1595 	struct ceph_inode_info *ci = ceph_inode(inode);
1596 	struct ceph_cap_snap *capsnap;
1597 	int capsnap_release = 0;
1598 
1599 	lockdep_assert_held(&ci->i_ceph_lock);
1600 
1601 	dout("removing capsnaps, ci is %p, inode is %p\n", ci, inode);
1602 
1603 	while (!list_empty(&ci->i_cap_snaps)) {
1604 		capsnap = list_first_entry(&ci->i_cap_snaps,
1605 					   struct ceph_cap_snap, ci_item);
1606 		__ceph_remove_capsnap(inode, capsnap, NULL, NULL);
1607 		ceph_put_snap_context(capsnap->context);
1608 		ceph_put_cap_snap(capsnap);
1609 		capsnap_release++;
1610 	}
1611 	wake_up_all(&ci->i_cap_wq);
1612 	wake_up_all(&mdsc->cap_flushing_wq);
1613 	return capsnap_release;
1614 }
1615 
1616 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1617 				  void *arg)
1618 {
1619 	struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1620 	struct ceph_mds_client *mdsc = fsc->mdsc;
1621 	struct ceph_inode_info *ci = ceph_inode(inode);
1622 	LIST_HEAD(to_remove);
1623 	bool dirty_dropped = false;
1624 	bool invalidate = false;
1625 	int capsnap_release = 0;
1626 
1627 	dout("removing cap %p, ci is %p, inode is %p\n",
1628 	     cap, ci, &ci->vfs_inode);
1629 	spin_lock(&ci->i_ceph_lock);
1630 	__ceph_remove_cap(cap, false);
1631 	if (!ci->i_auth_cap) {
1632 		struct ceph_cap_flush *cf;
1633 
1634 		if (READ_ONCE(fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
1635 			if (inode->i_data.nrpages > 0)
1636 				invalidate = true;
1637 			if (ci->i_wrbuffer_ref > 0)
1638 				mapping_set_error(&inode->i_data, -EIO);
1639 		}
1640 
1641 		while (!list_empty(&ci->i_cap_flush_list)) {
1642 			cf = list_first_entry(&ci->i_cap_flush_list,
1643 					      struct ceph_cap_flush, i_list);
1644 			list_move(&cf->i_list, &to_remove);
1645 		}
1646 
1647 		spin_lock(&mdsc->cap_dirty_lock);
1648 
1649 		list_for_each_entry(cf, &to_remove, i_list)
1650 			list_del_init(&cf->g_list);
1651 
1652 		if (!list_empty(&ci->i_dirty_item)) {
1653 			pr_warn_ratelimited(
1654 				" dropping dirty %s state for %p %lld\n",
1655 				ceph_cap_string(ci->i_dirty_caps),
1656 				inode, ceph_ino(inode));
1657 			ci->i_dirty_caps = 0;
1658 			list_del_init(&ci->i_dirty_item);
1659 			dirty_dropped = true;
1660 		}
1661 		if (!list_empty(&ci->i_flushing_item)) {
1662 			pr_warn_ratelimited(
1663 				" dropping dirty+flushing %s state for %p %lld\n",
1664 				ceph_cap_string(ci->i_flushing_caps),
1665 				inode, ceph_ino(inode));
1666 			ci->i_flushing_caps = 0;
1667 			list_del_init(&ci->i_flushing_item);
1668 			mdsc->num_cap_flushing--;
1669 			dirty_dropped = true;
1670 		}
1671 		spin_unlock(&mdsc->cap_dirty_lock);
1672 
1673 		if (dirty_dropped) {
1674 			mapping_set_error(inode->i_mapping, -EIO);
1675 
1676 			if (ci->i_wrbuffer_ref_head == 0 &&
1677 			    ci->i_wr_ref == 0 &&
1678 			    ci->i_dirty_caps == 0 &&
1679 			    ci->i_flushing_caps == 0) {
1680 				ceph_put_snap_context(ci->i_head_snapc);
1681 				ci->i_head_snapc = NULL;
1682 			}
1683 		}
1684 
1685 		if (atomic_read(&ci->i_filelock_ref) > 0) {
1686 			/* make further file lock syscall return -EIO */
1687 			ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1688 			pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1689 					    inode, ceph_ino(inode));
1690 		}
1691 
1692 		if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1693 			list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1694 			ci->i_prealloc_cap_flush = NULL;
1695 		}
1696 
1697 		if (!list_empty(&ci->i_cap_snaps))
1698 			capsnap_release = remove_capsnaps(mdsc, inode);
1699 	}
1700 	spin_unlock(&ci->i_ceph_lock);
1701 	while (!list_empty(&to_remove)) {
1702 		struct ceph_cap_flush *cf;
1703 		cf = list_first_entry(&to_remove,
1704 				      struct ceph_cap_flush, i_list);
1705 		list_del_init(&cf->i_list);
1706 		if (!cf->is_capsnap)
1707 			ceph_free_cap_flush(cf);
1708 	}
1709 
1710 	wake_up_all(&ci->i_cap_wq);
1711 	if (invalidate)
1712 		ceph_queue_invalidate(inode);
1713 	if (dirty_dropped)
1714 		iput(inode);
1715 	while (capsnap_release--)
1716 		iput(inode);
1717 	return 0;
1718 }
1719 
1720 /*
1721  * caller must hold session s_mutex
1722  */
1723 static void remove_session_caps(struct ceph_mds_session *session)
1724 {
1725 	struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1726 	struct super_block *sb = fsc->sb;
1727 	LIST_HEAD(dispose);
1728 
1729 	dout("remove_session_caps on %p\n", session);
1730 	ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1731 
1732 	wake_up_all(&fsc->mdsc->cap_flushing_wq);
1733 
1734 	spin_lock(&session->s_cap_lock);
1735 	if (session->s_nr_caps > 0) {
1736 		struct inode *inode;
1737 		struct ceph_cap *cap, *prev = NULL;
1738 		struct ceph_vino vino;
1739 		/*
1740 		 * iterate_session_caps() skips inodes that are being
1741 		 * deleted, we need to wait until deletions are complete.
1742 		 * __wait_on_freeing_inode() is designed for the job,
1743 		 * but it is not exported, so use lookup inode function
1744 		 * to access it.
1745 		 */
1746 		while (!list_empty(&session->s_caps)) {
1747 			cap = list_entry(session->s_caps.next,
1748 					 struct ceph_cap, session_caps);
1749 			if (cap == prev)
1750 				break;
1751 			prev = cap;
1752 			vino = cap->ci->i_vino;
1753 			spin_unlock(&session->s_cap_lock);
1754 
1755 			inode = ceph_find_inode(sb, vino);
1756 			iput(inode);
1757 
1758 			spin_lock(&session->s_cap_lock);
1759 		}
1760 	}
1761 
1762 	// drop cap expires and unlock s_cap_lock
1763 	detach_cap_releases(session, &dispose);
1764 
1765 	BUG_ON(session->s_nr_caps > 0);
1766 	BUG_ON(!list_empty(&session->s_cap_flushing));
1767 	spin_unlock(&session->s_cap_lock);
1768 	dispose_cap_releases(session->s_mdsc, &dispose);
1769 }
1770 
1771 enum {
1772 	RECONNECT,
1773 	RENEWCAPS,
1774 	FORCE_RO,
1775 };
1776 
1777 /*
1778  * wake up any threads waiting on this session's caps.  if the cap is
1779  * old (didn't get renewed on the client reconnect), remove it now.
1780  *
1781  * caller must hold s_mutex.
1782  */
1783 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1784 			      void *arg)
1785 {
1786 	struct ceph_inode_info *ci = ceph_inode(inode);
1787 	unsigned long ev = (unsigned long)arg;
1788 
1789 	if (ev == RECONNECT) {
1790 		spin_lock(&ci->i_ceph_lock);
1791 		ci->i_wanted_max_size = 0;
1792 		ci->i_requested_max_size = 0;
1793 		spin_unlock(&ci->i_ceph_lock);
1794 	} else if (ev == RENEWCAPS) {
1795 		if (cap->cap_gen < atomic_read(&cap->session->s_cap_gen)) {
1796 			/* mds did not re-issue stale cap */
1797 			spin_lock(&ci->i_ceph_lock);
1798 			cap->issued = cap->implemented = CEPH_CAP_PIN;
1799 			spin_unlock(&ci->i_ceph_lock);
1800 		}
1801 	} else if (ev == FORCE_RO) {
1802 	}
1803 	wake_up_all(&ci->i_cap_wq);
1804 	return 0;
1805 }
1806 
1807 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1808 {
1809 	dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1810 	ceph_iterate_session_caps(session, wake_up_session_cb,
1811 				  (void *)(unsigned long)ev);
1812 }
1813 
1814 /*
1815  * Send periodic message to MDS renewing all currently held caps.  The
1816  * ack will reset the expiration for all caps from this session.
1817  *
1818  * caller holds s_mutex
1819  */
1820 static int send_renew_caps(struct ceph_mds_client *mdsc,
1821 			   struct ceph_mds_session *session)
1822 {
1823 	struct ceph_msg *msg;
1824 	int state;
1825 
1826 	if (time_after_eq(jiffies, session->s_cap_ttl) &&
1827 	    time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1828 		pr_info("mds%d caps stale\n", session->s_mds);
1829 	session->s_renew_requested = jiffies;
1830 
1831 	/* do not try to renew caps until a recovering mds has reconnected
1832 	 * with its clients. */
1833 	state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1834 	if (state < CEPH_MDS_STATE_RECONNECT) {
1835 		dout("send_renew_caps ignoring mds%d (%s)\n",
1836 		     session->s_mds, ceph_mds_state_name(state));
1837 		return 0;
1838 	}
1839 
1840 	dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1841 		ceph_mds_state_name(state));
1842 	msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1843 				      ++session->s_renew_seq);
1844 	if (!msg)
1845 		return -ENOMEM;
1846 	ceph_con_send(&session->s_con, msg);
1847 	return 0;
1848 }
1849 
1850 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1851 			     struct ceph_mds_session *session, u64 seq)
1852 {
1853 	struct ceph_msg *msg;
1854 
1855 	dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1856 	     session->s_mds, ceph_session_state_name(session->s_state), seq);
1857 	msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1858 	if (!msg)
1859 		return -ENOMEM;
1860 	ceph_con_send(&session->s_con, msg);
1861 	return 0;
1862 }
1863 
1864 
1865 /*
1866  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1867  *
1868  * Called under session->s_mutex
1869  */
1870 static void renewed_caps(struct ceph_mds_client *mdsc,
1871 			 struct ceph_mds_session *session, int is_renew)
1872 {
1873 	int was_stale;
1874 	int wake = 0;
1875 
1876 	spin_lock(&session->s_cap_lock);
1877 	was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1878 
1879 	session->s_cap_ttl = session->s_renew_requested +
1880 		mdsc->mdsmap->m_session_timeout*HZ;
1881 
1882 	if (was_stale) {
1883 		if (time_before(jiffies, session->s_cap_ttl)) {
1884 			pr_info("mds%d caps renewed\n", session->s_mds);
1885 			wake = 1;
1886 		} else {
1887 			pr_info("mds%d caps still stale\n", session->s_mds);
1888 		}
1889 	}
1890 	dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1891 	     session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1892 	     time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1893 	spin_unlock(&session->s_cap_lock);
1894 
1895 	if (wake)
1896 		wake_up_session_caps(session, RENEWCAPS);
1897 }
1898 
1899 /*
1900  * send a session close request
1901  */
1902 static int request_close_session(struct ceph_mds_session *session)
1903 {
1904 	struct ceph_msg *msg;
1905 
1906 	dout("request_close_session mds%d state %s seq %lld\n",
1907 	     session->s_mds, ceph_session_state_name(session->s_state),
1908 	     session->s_seq);
1909 	msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
1910 				      session->s_seq);
1911 	if (!msg)
1912 		return -ENOMEM;
1913 	ceph_con_send(&session->s_con, msg);
1914 	return 1;
1915 }
1916 
1917 /*
1918  * Called with s_mutex held.
1919  */
1920 static int __close_session(struct ceph_mds_client *mdsc,
1921 			 struct ceph_mds_session *session)
1922 {
1923 	if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1924 		return 0;
1925 	session->s_state = CEPH_MDS_SESSION_CLOSING;
1926 	return request_close_session(session);
1927 }
1928 
1929 static bool drop_negative_children(struct dentry *dentry)
1930 {
1931 	struct dentry *child;
1932 	bool all_negative = true;
1933 
1934 	if (!d_is_dir(dentry))
1935 		goto out;
1936 
1937 	spin_lock(&dentry->d_lock);
1938 	list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1939 		if (d_really_is_positive(child)) {
1940 			all_negative = false;
1941 			break;
1942 		}
1943 	}
1944 	spin_unlock(&dentry->d_lock);
1945 
1946 	if (all_negative)
1947 		shrink_dcache_parent(dentry);
1948 out:
1949 	return all_negative;
1950 }
1951 
1952 /*
1953  * Trim old(er) caps.
1954  *
1955  * Because we can't cache an inode without one or more caps, we do
1956  * this indirectly: if a cap is unused, we prune its aliases, at which
1957  * point the inode will hopefully get dropped to.
1958  *
1959  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1960  * memory pressure from the MDS, though, so it needn't be perfect.
1961  */
1962 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1963 {
1964 	int *remaining = arg;
1965 	struct ceph_inode_info *ci = ceph_inode(inode);
1966 	int used, wanted, oissued, mine;
1967 
1968 	if (*remaining <= 0)
1969 		return -1;
1970 
1971 	spin_lock(&ci->i_ceph_lock);
1972 	mine = cap->issued | cap->implemented;
1973 	used = __ceph_caps_used(ci);
1974 	wanted = __ceph_caps_file_wanted(ci);
1975 	oissued = __ceph_caps_issued_other(ci, cap);
1976 
1977 	dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1978 	     inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1979 	     ceph_cap_string(used), ceph_cap_string(wanted));
1980 	if (cap == ci->i_auth_cap) {
1981 		if (ci->i_dirty_caps || ci->i_flushing_caps ||
1982 		    !list_empty(&ci->i_cap_snaps))
1983 			goto out;
1984 		if ((used | wanted) & CEPH_CAP_ANY_WR)
1985 			goto out;
1986 		/* Note: it's possible that i_filelock_ref becomes non-zero
1987 		 * after dropping auth caps. It doesn't hurt because reply
1988 		 * of lock mds request will re-add auth caps. */
1989 		if (atomic_read(&ci->i_filelock_ref) > 0)
1990 			goto out;
1991 	}
1992 	/* The inode has cached pages, but it's no longer used.
1993 	 * we can safely drop it */
1994 	if (S_ISREG(inode->i_mode) &&
1995 	    wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1996 	    !(oissued & CEPH_CAP_FILE_CACHE)) {
1997 	  used = 0;
1998 	  oissued = 0;
1999 	}
2000 	if ((used | wanted) & ~oissued & mine)
2001 		goto out;   /* we need these caps */
2002 
2003 	if (oissued) {
2004 		/* we aren't the only cap.. just remove us */
2005 		ceph_remove_cap(cap, true);
2006 		(*remaining)--;
2007 	} else {
2008 		struct dentry *dentry;
2009 		/* try dropping referring dentries */
2010 		spin_unlock(&ci->i_ceph_lock);
2011 		dentry = d_find_any_alias(inode);
2012 		if (dentry && drop_negative_children(dentry)) {
2013 			int count;
2014 			dput(dentry);
2015 			d_prune_aliases(inode);
2016 			count = atomic_read(&inode->i_count);
2017 			if (count == 1)
2018 				(*remaining)--;
2019 			dout("trim_caps_cb %p cap %p pruned, count now %d\n",
2020 			     inode, cap, count);
2021 		} else {
2022 			dput(dentry);
2023 		}
2024 		return 0;
2025 	}
2026 
2027 out:
2028 	spin_unlock(&ci->i_ceph_lock);
2029 	return 0;
2030 }
2031 
2032 /*
2033  * Trim session cap count down to some max number.
2034  */
2035 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2036 		   struct ceph_mds_session *session,
2037 		   int max_caps)
2038 {
2039 	int trim_caps = session->s_nr_caps - max_caps;
2040 
2041 	dout("trim_caps mds%d start: %d / %d, trim %d\n",
2042 	     session->s_mds, session->s_nr_caps, max_caps, trim_caps);
2043 	if (trim_caps > 0) {
2044 		int remaining = trim_caps;
2045 
2046 		ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2047 		dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2048 		     session->s_mds, session->s_nr_caps, max_caps,
2049 			trim_caps - remaining);
2050 	}
2051 
2052 	ceph_flush_cap_releases(mdsc, session);
2053 	return 0;
2054 }
2055 
2056 static int check_caps_flush(struct ceph_mds_client *mdsc,
2057 			    u64 want_flush_tid)
2058 {
2059 	int ret = 1;
2060 
2061 	spin_lock(&mdsc->cap_dirty_lock);
2062 	if (!list_empty(&mdsc->cap_flush_list)) {
2063 		struct ceph_cap_flush *cf =
2064 			list_first_entry(&mdsc->cap_flush_list,
2065 					 struct ceph_cap_flush, g_list);
2066 		if (cf->tid <= want_flush_tid) {
2067 			dout("check_caps_flush still flushing tid "
2068 			     "%llu <= %llu\n", cf->tid, want_flush_tid);
2069 			ret = 0;
2070 		}
2071 	}
2072 	spin_unlock(&mdsc->cap_dirty_lock);
2073 	return ret;
2074 }
2075 
2076 /*
2077  * flush all dirty inode data to disk.
2078  *
2079  * returns true if we've flushed through want_flush_tid
2080  */
2081 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2082 			    u64 want_flush_tid)
2083 {
2084 	dout("check_caps_flush want %llu\n", want_flush_tid);
2085 
2086 	wait_event(mdsc->cap_flushing_wq,
2087 		   check_caps_flush(mdsc, want_flush_tid));
2088 
2089 	dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2090 }
2091 
2092 /*
2093  * called under s_mutex
2094  */
2095 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2096 				   struct ceph_mds_session *session)
2097 {
2098 	struct ceph_msg *msg = NULL;
2099 	struct ceph_mds_cap_release *head;
2100 	struct ceph_mds_cap_item *item;
2101 	struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2102 	struct ceph_cap *cap;
2103 	LIST_HEAD(tmp_list);
2104 	int num_cap_releases;
2105 	__le32	barrier, *cap_barrier;
2106 
2107 	down_read(&osdc->lock);
2108 	barrier = cpu_to_le32(osdc->epoch_barrier);
2109 	up_read(&osdc->lock);
2110 
2111 	spin_lock(&session->s_cap_lock);
2112 again:
2113 	list_splice_init(&session->s_cap_releases, &tmp_list);
2114 	num_cap_releases = session->s_num_cap_releases;
2115 	session->s_num_cap_releases = 0;
2116 	spin_unlock(&session->s_cap_lock);
2117 
2118 	while (!list_empty(&tmp_list)) {
2119 		if (!msg) {
2120 			msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2121 					PAGE_SIZE, GFP_NOFS, false);
2122 			if (!msg)
2123 				goto out_err;
2124 			head = msg->front.iov_base;
2125 			head->num = cpu_to_le32(0);
2126 			msg->front.iov_len = sizeof(*head);
2127 
2128 			msg->hdr.version = cpu_to_le16(2);
2129 			msg->hdr.compat_version = cpu_to_le16(1);
2130 		}
2131 
2132 		cap = list_first_entry(&tmp_list, struct ceph_cap,
2133 					session_caps);
2134 		list_del(&cap->session_caps);
2135 		num_cap_releases--;
2136 
2137 		head = msg->front.iov_base;
2138 		put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2139 				   &head->num);
2140 		item = msg->front.iov_base + msg->front.iov_len;
2141 		item->ino = cpu_to_le64(cap->cap_ino);
2142 		item->cap_id = cpu_to_le64(cap->cap_id);
2143 		item->migrate_seq = cpu_to_le32(cap->mseq);
2144 		item->seq = cpu_to_le32(cap->issue_seq);
2145 		msg->front.iov_len += sizeof(*item);
2146 
2147 		ceph_put_cap(mdsc, cap);
2148 
2149 		if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2150 			// Append cap_barrier field
2151 			cap_barrier = msg->front.iov_base + msg->front.iov_len;
2152 			*cap_barrier = barrier;
2153 			msg->front.iov_len += sizeof(*cap_barrier);
2154 
2155 			msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2156 			dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2157 			ceph_con_send(&session->s_con, msg);
2158 			msg = NULL;
2159 		}
2160 	}
2161 
2162 	BUG_ON(num_cap_releases != 0);
2163 
2164 	spin_lock(&session->s_cap_lock);
2165 	if (!list_empty(&session->s_cap_releases))
2166 		goto again;
2167 	spin_unlock(&session->s_cap_lock);
2168 
2169 	if (msg) {
2170 		// Append cap_barrier field
2171 		cap_barrier = msg->front.iov_base + msg->front.iov_len;
2172 		*cap_barrier = barrier;
2173 		msg->front.iov_len += sizeof(*cap_barrier);
2174 
2175 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2176 		dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2177 		ceph_con_send(&session->s_con, msg);
2178 	}
2179 	return;
2180 out_err:
2181 	pr_err("send_cap_releases mds%d, failed to allocate message\n",
2182 		session->s_mds);
2183 	spin_lock(&session->s_cap_lock);
2184 	list_splice(&tmp_list, &session->s_cap_releases);
2185 	session->s_num_cap_releases += num_cap_releases;
2186 	spin_unlock(&session->s_cap_lock);
2187 }
2188 
2189 static void ceph_cap_release_work(struct work_struct *work)
2190 {
2191 	struct ceph_mds_session *session =
2192 		container_of(work, struct ceph_mds_session, s_cap_release_work);
2193 
2194 	mutex_lock(&session->s_mutex);
2195 	if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2196 	    session->s_state == CEPH_MDS_SESSION_HUNG)
2197 		ceph_send_cap_releases(session->s_mdsc, session);
2198 	mutex_unlock(&session->s_mutex);
2199 	ceph_put_mds_session(session);
2200 }
2201 
2202 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2203 		             struct ceph_mds_session *session)
2204 {
2205 	if (mdsc->stopping)
2206 		return;
2207 
2208 	ceph_get_mds_session(session);
2209 	if (queue_work(mdsc->fsc->cap_wq,
2210 		       &session->s_cap_release_work)) {
2211 		dout("cap release work queued\n");
2212 	} else {
2213 		ceph_put_mds_session(session);
2214 		dout("failed to queue cap release work\n");
2215 	}
2216 }
2217 
2218 /*
2219  * caller holds session->s_cap_lock
2220  */
2221 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2222 			      struct ceph_cap *cap)
2223 {
2224 	list_add_tail(&cap->session_caps, &session->s_cap_releases);
2225 	session->s_num_cap_releases++;
2226 
2227 	if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2228 		ceph_flush_cap_releases(session->s_mdsc, session);
2229 }
2230 
2231 static void ceph_cap_reclaim_work(struct work_struct *work)
2232 {
2233 	struct ceph_mds_client *mdsc =
2234 		container_of(work, struct ceph_mds_client, cap_reclaim_work);
2235 	int ret = ceph_trim_dentries(mdsc);
2236 	if (ret == -EAGAIN)
2237 		ceph_queue_cap_reclaim_work(mdsc);
2238 }
2239 
2240 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2241 {
2242 	if (mdsc->stopping)
2243 		return;
2244 
2245         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2246                 dout("caps reclaim work queued\n");
2247         } else {
2248                 dout("failed to queue caps release work\n");
2249         }
2250 }
2251 
2252 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2253 {
2254 	int val;
2255 	if (!nr)
2256 		return;
2257 	val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2258 	if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2259 		atomic_set(&mdsc->cap_reclaim_pending, 0);
2260 		ceph_queue_cap_reclaim_work(mdsc);
2261 	}
2262 }
2263 
2264 /*
2265  * requests
2266  */
2267 
2268 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2269 				    struct inode *dir)
2270 {
2271 	struct ceph_inode_info *ci = ceph_inode(dir);
2272 	struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2273 	struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2274 	size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2275 	unsigned int num_entries;
2276 	int order;
2277 
2278 	spin_lock(&ci->i_ceph_lock);
2279 	num_entries = ci->i_files + ci->i_subdirs;
2280 	spin_unlock(&ci->i_ceph_lock);
2281 	num_entries = max(num_entries, 1U);
2282 	num_entries = min(num_entries, opt->max_readdir);
2283 
2284 	order = get_order(size * num_entries);
2285 	while (order >= 0) {
2286 		rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2287 							     __GFP_NOWARN,
2288 							     order);
2289 		if (rinfo->dir_entries)
2290 			break;
2291 		order--;
2292 	}
2293 	if (!rinfo->dir_entries)
2294 		return -ENOMEM;
2295 
2296 	num_entries = (PAGE_SIZE << order) / size;
2297 	num_entries = min(num_entries, opt->max_readdir);
2298 
2299 	rinfo->dir_buf_size = PAGE_SIZE << order;
2300 	req->r_num_caps = num_entries + 1;
2301 	req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2302 	req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2303 	return 0;
2304 }
2305 
2306 /*
2307  * Create an mds request.
2308  */
2309 struct ceph_mds_request *
2310 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2311 {
2312 	struct ceph_mds_request *req;
2313 
2314 	req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2315 	if (!req)
2316 		return ERR_PTR(-ENOMEM);
2317 
2318 	mutex_init(&req->r_fill_mutex);
2319 	req->r_mdsc = mdsc;
2320 	req->r_started = jiffies;
2321 	req->r_start_latency = ktime_get();
2322 	req->r_resend_mds = -1;
2323 	INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2324 	INIT_LIST_HEAD(&req->r_unsafe_target_item);
2325 	req->r_fmode = -1;
2326 	kref_init(&req->r_kref);
2327 	RB_CLEAR_NODE(&req->r_node);
2328 	INIT_LIST_HEAD(&req->r_wait);
2329 	init_completion(&req->r_completion);
2330 	init_completion(&req->r_safe_completion);
2331 	INIT_LIST_HEAD(&req->r_unsafe_item);
2332 
2333 	ktime_get_coarse_real_ts64(&req->r_stamp);
2334 
2335 	req->r_op = op;
2336 	req->r_direct_mode = mode;
2337 	return req;
2338 }
2339 
2340 /*
2341  * return oldest (lowest) request, tid in request tree, 0 if none.
2342  *
2343  * called under mdsc->mutex.
2344  */
2345 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2346 {
2347 	if (RB_EMPTY_ROOT(&mdsc->request_tree))
2348 		return NULL;
2349 	return rb_entry(rb_first(&mdsc->request_tree),
2350 			struct ceph_mds_request, r_node);
2351 }
2352 
2353 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2354 {
2355 	return mdsc->oldest_tid;
2356 }
2357 
2358 /*
2359  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2360  * on build_path_from_dentry in fs/cifs/dir.c.
2361  *
2362  * If @stop_on_nosnap, generate path relative to the first non-snapped
2363  * inode.
2364  *
2365  * Encode hidden .snap dirs as a double /, i.e.
2366  *   foo/.snap/bar -> foo//bar
2367  */
2368 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2369 			   int stop_on_nosnap)
2370 {
2371 	struct dentry *temp;
2372 	char *path;
2373 	int pos;
2374 	unsigned seq;
2375 	u64 base;
2376 
2377 	if (!dentry)
2378 		return ERR_PTR(-EINVAL);
2379 
2380 	path = __getname();
2381 	if (!path)
2382 		return ERR_PTR(-ENOMEM);
2383 retry:
2384 	pos = PATH_MAX - 1;
2385 	path[pos] = '\0';
2386 
2387 	seq = read_seqbegin(&rename_lock);
2388 	rcu_read_lock();
2389 	temp = dentry;
2390 	for (;;) {
2391 		struct inode *inode;
2392 
2393 		spin_lock(&temp->d_lock);
2394 		inode = d_inode(temp);
2395 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2396 			dout("build_path path+%d: %p SNAPDIR\n",
2397 			     pos, temp);
2398 		} else if (stop_on_nosnap && inode && dentry != temp &&
2399 			   ceph_snap(inode) == CEPH_NOSNAP) {
2400 			spin_unlock(&temp->d_lock);
2401 			pos++; /* get rid of any prepended '/' */
2402 			break;
2403 		} else {
2404 			pos -= temp->d_name.len;
2405 			if (pos < 0) {
2406 				spin_unlock(&temp->d_lock);
2407 				break;
2408 			}
2409 			memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2410 		}
2411 		spin_unlock(&temp->d_lock);
2412 		temp = READ_ONCE(temp->d_parent);
2413 
2414 		/* Are we at the root? */
2415 		if (IS_ROOT(temp))
2416 			break;
2417 
2418 		/* Are we out of buffer? */
2419 		if (--pos < 0)
2420 			break;
2421 
2422 		path[pos] = '/';
2423 	}
2424 	base = ceph_ino(d_inode(temp));
2425 	rcu_read_unlock();
2426 
2427 	if (read_seqretry(&rename_lock, seq))
2428 		goto retry;
2429 
2430 	if (pos < 0) {
2431 		/*
2432 		 * A rename didn't occur, but somehow we didn't end up where
2433 		 * we thought we would. Throw a warning and try again.
2434 		 */
2435 		pr_warn("build_path did not end path lookup where "
2436 			"expected, pos is %d\n", pos);
2437 		goto retry;
2438 	}
2439 
2440 	*pbase = base;
2441 	*plen = PATH_MAX - 1 - pos;
2442 	dout("build_path on %p %d built %llx '%.*s'\n",
2443 	     dentry, d_count(dentry), base, *plen, path + pos);
2444 	return path + pos;
2445 }
2446 
2447 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2448 			     const char **ppath, int *ppathlen, u64 *pino,
2449 			     bool *pfreepath, bool parent_locked)
2450 {
2451 	char *path;
2452 
2453 	rcu_read_lock();
2454 	if (!dir)
2455 		dir = d_inode_rcu(dentry->d_parent);
2456 	if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2457 		*pino = ceph_ino(dir);
2458 		rcu_read_unlock();
2459 		*ppath = dentry->d_name.name;
2460 		*ppathlen = dentry->d_name.len;
2461 		return 0;
2462 	}
2463 	rcu_read_unlock();
2464 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2465 	if (IS_ERR(path))
2466 		return PTR_ERR(path);
2467 	*ppath = path;
2468 	*pfreepath = true;
2469 	return 0;
2470 }
2471 
2472 static int build_inode_path(struct inode *inode,
2473 			    const char **ppath, int *ppathlen, u64 *pino,
2474 			    bool *pfreepath)
2475 {
2476 	struct dentry *dentry;
2477 	char *path;
2478 
2479 	if (ceph_snap(inode) == CEPH_NOSNAP) {
2480 		*pino = ceph_ino(inode);
2481 		*ppathlen = 0;
2482 		return 0;
2483 	}
2484 	dentry = d_find_alias(inode);
2485 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2486 	dput(dentry);
2487 	if (IS_ERR(path))
2488 		return PTR_ERR(path);
2489 	*ppath = path;
2490 	*pfreepath = true;
2491 	return 0;
2492 }
2493 
2494 /*
2495  * request arguments may be specified via an inode *, a dentry *, or
2496  * an explicit ino+path.
2497  */
2498 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2499 				  struct inode *rdiri, const char *rpath,
2500 				  u64 rino, const char **ppath, int *pathlen,
2501 				  u64 *ino, bool *freepath, bool parent_locked)
2502 {
2503 	int r = 0;
2504 
2505 	if (rinode) {
2506 		r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2507 		dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2508 		     ceph_snap(rinode));
2509 	} else if (rdentry) {
2510 		r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2511 					freepath, parent_locked);
2512 		dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2513 		     *ppath);
2514 	} else if (rpath || rino) {
2515 		*ino = rino;
2516 		*ppath = rpath;
2517 		*pathlen = rpath ? strlen(rpath) : 0;
2518 		dout(" path %.*s\n", *pathlen, rpath);
2519 	}
2520 
2521 	return r;
2522 }
2523 
2524 static void encode_timestamp_and_gids(void **p,
2525 				      const struct ceph_mds_request *req)
2526 {
2527 	struct ceph_timespec ts;
2528 	int i;
2529 
2530 	ceph_encode_timespec64(&ts, &req->r_stamp);
2531 	ceph_encode_copy(p, &ts, sizeof(ts));
2532 
2533 	/* gid_list */
2534 	ceph_encode_32(p, req->r_cred->group_info->ngroups);
2535 	for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2536 		ceph_encode_64(p, from_kgid(&init_user_ns,
2537 					    req->r_cred->group_info->gid[i]));
2538 }
2539 
2540 /*
2541  * called under mdsc->mutex
2542  */
2543 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2544 					       struct ceph_mds_request *req,
2545 					       bool drop_cap_releases)
2546 {
2547 	int mds = session->s_mds;
2548 	struct ceph_mds_client *mdsc = session->s_mdsc;
2549 	struct ceph_msg *msg;
2550 	struct ceph_mds_request_head_old *head;
2551 	const char *path1 = NULL;
2552 	const char *path2 = NULL;
2553 	u64 ino1 = 0, ino2 = 0;
2554 	int pathlen1 = 0, pathlen2 = 0;
2555 	bool freepath1 = false, freepath2 = false;
2556 	int len;
2557 	u16 releases;
2558 	void *p, *end;
2559 	int ret;
2560 	bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2561 
2562 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
2563 			      req->r_parent, req->r_path1, req->r_ino1.ino,
2564 			      &path1, &pathlen1, &ino1, &freepath1,
2565 			      test_bit(CEPH_MDS_R_PARENT_LOCKED,
2566 					&req->r_req_flags));
2567 	if (ret < 0) {
2568 		msg = ERR_PTR(ret);
2569 		goto out;
2570 	}
2571 
2572 	/* If r_old_dentry is set, then assume that its parent is locked */
2573 	ret = set_request_path_attr(NULL, req->r_old_dentry,
2574 			      req->r_old_dentry_dir,
2575 			      req->r_path2, req->r_ino2.ino,
2576 			      &path2, &pathlen2, &ino2, &freepath2, true);
2577 	if (ret < 0) {
2578 		msg = ERR_PTR(ret);
2579 		goto out_free1;
2580 	}
2581 
2582 	len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2583 	len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2584 		sizeof(struct ceph_timespec);
2585 	len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2586 
2587 	/* calculate (max) length for cap releases */
2588 	len += sizeof(struct ceph_mds_request_release) *
2589 		(!!req->r_inode_drop + !!req->r_dentry_drop +
2590 		 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2591 
2592 	if (req->r_dentry_drop)
2593 		len += pathlen1;
2594 	if (req->r_old_dentry_drop)
2595 		len += pathlen2;
2596 
2597 	msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2598 	if (!msg) {
2599 		msg = ERR_PTR(-ENOMEM);
2600 		goto out_free2;
2601 	}
2602 
2603 	msg->hdr.tid = cpu_to_le64(req->r_tid);
2604 
2605 	/*
2606 	 * The old ceph_mds_request_head didn't contain a version field, and
2607 	 * one was added when we moved the message version from 3->4.
2608 	 */
2609 	if (legacy) {
2610 		msg->hdr.version = cpu_to_le16(3);
2611 		head = msg->front.iov_base;
2612 		p = msg->front.iov_base + sizeof(*head);
2613 	} else {
2614 		struct ceph_mds_request_head *new_head = msg->front.iov_base;
2615 
2616 		msg->hdr.version = cpu_to_le16(4);
2617 		new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2618 		head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2619 		p = msg->front.iov_base + sizeof(*new_head);
2620 	}
2621 
2622 	end = msg->front.iov_base + msg->front.iov_len;
2623 
2624 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2625 	head->op = cpu_to_le32(req->r_op);
2626 	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2627 						 req->r_cred->fsuid));
2628 	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2629 						 req->r_cred->fsgid));
2630 	head->ino = cpu_to_le64(req->r_deleg_ino);
2631 	head->args = req->r_args;
2632 
2633 	ceph_encode_filepath(&p, end, ino1, path1);
2634 	ceph_encode_filepath(&p, end, ino2, path2);
2635 
2636 	/* make note of release offset, in case we need to replay */
2637 	req->r_request_release_offset = p - msg->front.iov_base;
2638 
2639 	/* cap releases */
2640 	releases = 0;
2641 	if (req->r_inode_drop)
2642 		releases += ceph_encode_inode_release(&p,
2643 		      req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2644 		      mds, req->r_inode_drop, req->r_inode_unless,
2645 		      req->r_op == CEPH_MDS_OP_READDIR);
2646 	if (req->r_dentry_drop)
2647 		releases += ceph_encode_dentry_release(&p, req->r_dentry,
2648 				req->r_parent, mds, req->r_dentry_drop,
2649 				req->r_dentry_unless);
2650 	if (req->r_old_dentry_drop)
2651 		releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2652 				req->r_old_dentry_dir, mds,
2653 				req->r_old_dentry_drop,
2654 				req->r_old_dentry_unless);
2655 	if (req->r_old_inode_drop)
2656 		releases += ceph_encode_inode_release(&p,
2657 		      d_inode(req->r_old_dentry),
2658 		      mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2659 
2660 	if (drop_cap_releases) {
2661 		releases = 0;
2662 		p = msg->front.iov_base + req->r_request_release_offset;
2663 	}
2664 
2665 	head->num_releases = cpu_to_le16(releases);
2666 
2667 	encode_timestamp_and_gids(&p, req);
2668 
2669 	if (WARN_ON_ONCE(p > end)) {
2670 		ceph_msg_put(msg);
2671 		msg = ERR_PTR(-ERANGE);
2672 		goto out_free2;
2673 	}
2674 
2675 	msg->front.iov_len = p - msg->front.iov_base;
2676 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2677 
2678 	if (req->r_pagelist) {
2679 		struct ceph_pagelist *pagelist = req->r_pagelist;
2680 		ceph_msg_data_add_pagelist(msg, pagelist);
2681 		msg->hdr.data_len = cpu_to_le32(pagelist->length);
2682 	} else {
2683 		msg->hdr.data_len = 0;
2684 	}
2685 
2686 	msg->hdr.data_off = cpu_to_le16(0);
2687 
2688 out_free2:
2689 	if (freepath2)
2690 		ceph_mdsc_free_path((char *)path2, pathlen2);
2691 out_free1:
2692 	if (freepath1)
2693 		ceph_mdsc_free_path((char *)path1, pathlen1);
2694 out:
2695 	return msg;
2696 }
2697 
2698 /*
2699  * called under mdsc->mutex if error, under no mutex if
2700  * success.
2701  */
2702 static void complete_request(struct ceph_mds_client *mdsc,
2703 			     struct ceph_mds_request *req)
2704 {
2705 	req->r_end_latency = ktime_get();
2706 
2707 	if (req->r_callback)
2708 		req->r_callback(mdsc, req);
2709 	complete_all(&req->r_completion);
2710 }
2711 
2712 static struct ceph_mds_request_head_old *
2713 find_old_request_head(void *p, u64 features)
2714 {
2715 	bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2716 	struct ceph_mds_request_head *new_head;
2717 
2718 	if (legacy)
2719 		return (struct ceph_mds_request_head_old *)p;
2720 	new_head = (struct ceph_mds_request_head *)p;
2721 	return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2722 }
2723 
2724 /*
2725  * called under mdsc->mutex
2726  */
2727 static int __prepare_send_request(struct ceph_mds_session *session,
2728 				  struct ceph_mds_request *req,
2729 				  bool drop_cap_releases)
2730 {
2731 	int mds = session->s_mds;
2732 	struct ceph_mds_client *mdsc = session->s_mdsc;
2733 	struct ceph_mds_request_head_old *rhead;
2734 	struct ceph_msg *msg;
2735 	int flags = 0;
2736 
2737 	req->r_attempts++;
2738 	if (req->r_inode) {
2739 		struct ceph_cap *cap =
2740 			ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2741 
2742 		if (cap)
2743 			req->r_sent_on_mseq = cap->mseq;
2744 		else
2745 			req->r_sent_on_mseq = -1;
2746 	}
2747 	dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2748 	     req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2749 
2750 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2751 		void *p;
2752 
2753 		/*
2754 		 * Replay.  Do not regenerate message (and rebuild
2755 		 * paths, etc.); just use the original message.
2756 		 * Rebuilding paths will break for renames because
2757 		 * d_move mangles the src name.
2758 		 */
2759 		msg = req->r_request;
2760 		rhead = find_old_request_head(msg->front.iov_base,
2761 					      session->s_con.peer_features);
2762 
2763 		flags = le32_to_cpu(rhead->flags);
2764 		flags |= CEPH_MDS_FLAG_REPLAY;
2765 		rhead->flags = cpu_to_le32(flags);
2766 
2767 		if (req->r_target_inode)
2768 			rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2769 
2770 		rhead->num_retry = req->r_attempts - 1;
2771 
2772 		/* remove cap/dentry releases from message */
2773 		rhead->num_releases = 0;
2774 
2775 		p = msg->front.iov_base + req->r_request_release_offset;
2776 		encode_timestamp_and_gids(&p, req);
2777 
2778 		msg->front.iov_len = p - msg->front.iov_base;
2779 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2780 		return 0;
2781 	}
2782 
2783 	if (req->r_request) {
2784 		ceph_msg_put(req->r_request);
2785 		req->r_request = NULL;
2786 	}
2787 	msg = create_request_message(session, req, drop_cap_releases);
2788 	if (IS_ERR(msg)) {
2789 		req->r_err = PTR_ERR(msg);
2790 		return PTR_ERR(msg);
2791 	}
2792 	req->r_request = msg;
2793 
2794 	rhead = find_old_request_head(msg->front.iov_base,
2795 				      session->s_con.peer_features);
2796 	rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2797 	if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2798 		flags |= CEPH_MDS_FLAG_REPLAY;
2799 	if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2800 		flags |= CEPH_MDS_FLAG_ASYNC;
2801 	if (req->r_parent)
2802 		flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2803 	rhead->flags = cpu_to_le32(flags);
2804 	rhead->num_fwd = req->r_num_fwd;
2805 	rhead->num_retry = req->r_attempts - 1;
2806 
2807 	dout(" r_parent = %p\n", req->r_parent);
2808 	return 0;
2809 }
2810 
2811 /*
2812  * called under mdsc->mutex
2813  */
2814 static int __send_request(struct ceph_mds_session *session,
2815 			  struct ceph_mds_request *req,
2816 			  bool drop_cap_releases)
2817 {
2818 	int err;
2819 
2820 	err = __prepare_send_request(session, req, drop_cap_releases);
2821 	if (!err) {
2822 		ceph_msg_get(req->r_request);
2823 		ceph_con_send(&session->s_con, req->r_request);
2824 	}
2825 
2826 	return err;
2827 }
2828 
2829 /*
2830  * send request, or put it on the appropriate wait list.
2831  */
2832 static void __do_request(struct ceph_mds_client *mdsc,
2833 			struct ceph_mds_request *req)
2834 {
2835 	struct ceph_mds_session *session = NULL;
2836 	int mds = -1;
2837 	int err = 0;
2838 	bool random;
2839 
2840 	if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2841 		if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2842 			__unregister_request(mdsc, req);
2843 		return;
2844 	}
2845 
2846 	if (req->r_timeout &&
2847 	    time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2848 		dout("do_request timed out\n");
2849 		err = -ETIMEDOUT;
2850 		goto finish;
2851 	}
2852 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2853 		dout("do_request forced umount\n");
2854 		err = -EIO;
2855 		goto finish;
2856 	}
2857 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2858 		if (mdsc->mdsmap_err) {
2859 			err = mdsc->mdsmap_err;
2860 			dout("do_request mdsmap err %d\n", err);
2861 			goto finish;
2862 		}
2863 		if (mdsc->mdsmap->m_epoch == 0) {
2864 			dout("do_request no mdsmap, waiting for map\n");
2865 			list_add(&req->r_wait, &mdsc->waiting_for_map);
2866 			return;
2867 		}
2868 		if (!(mdsc->fsc->mount_options->flags &
2869 		      CEPH_MOUNT_OPT_MOUNTWAIT) &&
2870 		    !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2871 			err = -EHOSTUNREACH;
2872 			goto finish;
2873 		}
2874 	}
2875 
2876 	put_request_session(req);
2877 
2878 	mds = __choose_mds(mdsc, req, &random);
2879 	if (mds < 0 ||
2880 	    ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2881 		if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2882 			err = -EJUKEBOX;
2883 			goto finish;
2884 		}
2885 		dout("do_request no mds or not active, waiting for map\n");
2886 		list_add(&req->r_wait, &mdsc->waiting_for_map);
2887 		return;
2888 	}
2889 
2890 	/* get, open session */
2891 	session = __ceph_lookup_mds_session(mdsc, mds);
2892 	if (!session) {
2893 		session = register_session(mdsc, mds);
2894 		if (IS_ERR(session)) {
2895 			err = PTR_ERR(session);
2896 			goto finish;
2897 		}
2898 	}
2899 	req->r_session = ceph_get_mds_session(session);
2900 
2901 	dout("do_request mds%d session %p state %s\n", mds, session,
2902 	     ceph_session_state_name(session->s_state));
2903 	if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2904 	    session->s_state != CEPH_MDS_SESSION_HUNG) {
2905 		/*
2906 		 * We cannot queue async requests since the caps and delegated
2907 		 * inodes are bound to the session. Just return -EJUKEBOX and
2908 		 * let the caller retry a sync request in that case.
2909 		 */
2910 		if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2911 			err = -EJUKEBOX;
2912 			goto out_session;
2913 		}
2914 
2915 		/*
2916 		 * If the session has been REJECTED, then return a hard error,
2917 		 * unless it's a CLEANRECOVER mount, in which case we'll queue
2918 		 * it to the mdsc queue.
2919 		 */
2920 		if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2921 			if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2922 				list_add(&req->r_wait, &mdsc->waiting_for_map);
2923 			else
2924 				err = -EACCES;
2925 			goto out_session;
2926 		}
2927 
2928 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
2929 		    session->s_state == CEPH_MDS_SESSION_CLOSING) {
2930 			err = __open_session(mdsc, session);
2931 			if (err)
2932 				goto out_session;
2933 			/* retry the same mds later */
2934 			if (random)
2935 				req->r_resend_mds = mds;
2936 		}
2937 		list_add(&req->r_wait, &session->s_waiting);
2938 		goto out_session;
2939 	}
2940 
2941 	/* send request */
2942 	req->r_resend_mds = -1;   /* forget any previous mds hint */
2943 
2944 	if (req->r_request_started == 0)   /* note request start time */
2945 		req->r_request_started = jiffies;
2946 
2947 	err = __send_request(session, req, false);
2948 
2949 out_session:
2950 	ceph_put_mds_session(session);
2951 finish:
2952 	if (err) {
2953 		dout("__do_request early error %d\n", err);
2954 		req->r_err = err;
2955 		complete_request(mdsc, req);
2956 		__unregister_request(mdsc, req);
2957 	}
2958 	return;
2959 }
2960 
2961 /*
2962  * called under mdsc->mutex
2963  */
2964 static void __wake_requests(struct ceph_mds_client *mdsc,
2965 			    struct list_head *head)
2966 {
2967 	struct ceph_mds_request *req;
2968 	LIST_HEAD(tmp_list);
2969 
2970 	list_splice_init(head, &tmp_list);
2971 
2972 	while (!list_empty(&tmp_list)) {
2973 		req = list_entry(tmp_list.next,
2974 				 struct ceph_mds_request, r_wait);
2975 		list_del_init(&req->r_wait);
2976 		dout(" wake request %p tid %llu\n", req, req->r_tid);
2977 		__do_request(mdsc, req);
2978 	}
2979 }
2980 
2981 /*
2982  * Wake up threads with requests pending for @mds, so that they can
2983  * resubmit their requests to a possibly different mds.
2984  */
2985 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2986 {
2987 	struct ceph_mds_request *req;
2988 	struct rb_node *p = rb_first(&mdsc->request_tree);
2989 
2990 	dout("kick_requests mds%d\n", mds);
2991 	while (p) {
2992 		req = rb_entry(p, struct ceph_mds_request, r_node);
2993 		p = rb_next(p);
2994 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2995 			continue;
2996 		if (req->r_attempts > 0)
2997 			continue; /* only new requests */
2998 		if (req->r_session &&
2999 		    req->r_session->s_mds == mds) {
3000 			dout(" kicking tid %llu\n", req->r_tid);
3001 			list_del_init(&req->r_wait);
3002 			__do_request(mdsc, req);
3003 		}
3004 	}
3005 }
3006 
3007 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
3008 			      struct ceph_mds_request *req)
3009 {
3010 	int err = 0;
3011 
3012 	/* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
3013 	if (req->r_inode)
3014 		ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3015 	if (req->r_parent) {
3016 		struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3017 		int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3018 			    CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3019 		spin_lock(&ci->i_ceph_lock);
3020 		ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3021 		__ceph_touch_fmode(ci, mdsc, fmode);
3022 		spin_unlock(&ci->i_ceph_lock);
3023 	}
3024 	if (req->r_old_dentry_dir)
3025 		ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3026 				  CEPH_CAP_PIN);
3027 
3028 	if (req->r_inode) {
3029 		err = ceph_wait_on_async_create(req->r_inode);
3030 		if (err) {
3031 			dout("%s: wait for async create returned: %d\n",
3032 			     __func__, err);
3033 			return err;
3034 		}
3035 	}
3036 
3037 	if (!err && req->r_old_inode) {
3038 		err = ceph_wait_on_async_create(req->r_old_inode);
3039 		if (err) {
3040 			dout("%s: wait for async create returned: %d\n",
3041 			     __func__, err);
3042 			return err;
3043 		}
3044 	}
3045 
3046 	dout("submit_request on %p for inode %p\n", req, dir);
3047 	mutex_lock(&mdsc->mutex);
3048 	__register_request(mdsc, req, dir);
3049 	__do_request(mdsc, req);
3050 	err = req->r_err;
3051 	mutex_unlock(&mdsc->mutex);
3052 	return err;
3053 }
3054 
3055 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3056 				  struct ceph_mds_request *req)
3057 {
3058 	int err;
3059 
3060 	/* wait */
3061 	dout("do_request waiting\n");
3062 	if (!req->r_timeout && req->r_wait_for_completion) {
3063 		err = req->r_wait_for_completion(mdsc, req);
3064 	} else {
3065 		long timeleft = wait_for_completion_killable_timeout(
3066 					&req->r_completion,
3067 					ceph_timeout_jiffies(req->r_timeout));
3068 		if (timeleft > 0)
3069 			err = 0;
3070 		else if (!timeleft)
3071 			err = -ETIMEDOUT;  /* timed out */
3072 		else
3073 			err = timeleft;  /* killed */
3074 	}
3075 	dout("do_request waited, got %d\n", err);
3076 	mutex_lock(&mdsc->mutex);
3077 
3078 	/* only abort if we didn't race with a real reply */
3079 	if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3080 		err = le32_to_cpu(req->r_reply_info.head->result);
3081 	} else if (err < 0) {
3082 		dout("aborted request %lld with %d\n", req->r_tid, err);
3083 
3084 		/*
3085 		 * ensure we aren't running concurrently with
3086 		 * ceph_fill_trace or ceph_readdir_prepopulate, which
3087 		 * rely on locks (dir mutex) held by our caller.
3088 		 */
3089 		mutex_lock(&req->r_fill_mutex);
3090 		req->r_err = err;
3091 		set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3092 		mutex_unlock(&req->r_fill_mutex);
3093 
3094 		if (req->r_parent &&
3095 		    (req->r_op & CEPH_MDS_OP_WRITE))
3096 			ceph_invalidate_dir_request(req);
3097 	} else {
3098 		err = req->r_err;
3099 	}
3100 
3101 	mutex_unlock(&mdsc->mutex);
3102 	return err;
3103 }
3104 
3105 /*
3106  * Synchrously perform an mds request.  Take care of all of the
3107  * session setup, forwarding, retry details.
3108  */
3109 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3110 			 struct inode *dir,
3111 			 struct ceph_mds_request *req)
3112 {
3113 	int err;
3114 
3115 	dout("do_request on %p\n", req);
3116 
3117 	/* issue */
3118 	err = ceph_mdsc_submit_request(mdsc, dir, req);
3119 	if (!err)
3120 		err = ceph_mdsc_wait_request(mdsc, req);
3121 	dout("do_request %p done, result %d\n", req, err);
3122 	return err;
3123 }
3124 
3125 /*
3126  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3127  * namespace request.
3128  */
3129 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3130 {
3131 	struct inode *dir = req->r_parent;
3132 	struct inode *old_dir = req->r_old_dentry_dir;
3133 
3134 	dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3135 
3136 	ceph_dir_clear_complete(dir);
3137 	if (old_dir)
3138 		ceph_dir_clear_complete(old_dir);
3139 	if (req->r_dentry)
3140 		ceph_invalidate_dentry_lease(req->r_dentry);
3141 	if (req->r_old_dentry)
3142 		ceph_invalidate_dentry_lease(req->r_old_dentry);
3143 }
3144 
3145 /*
3146  * Handle mds reply.
3147  *
3148  * We take the session mutex and parse and process the reply immediately.
3149  * This preserves the logical ordering of replies, capabilities, etc., sent
3150  * by the MDS as they are applied to our local cache.
3151  */
3152 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3153 {
3154 	struct ceph_mds_client *mdsc = session->s_mdsc;
3155 	struct ceph_mds_request *req;
3156 	struct ceph_mds_reply_head *head = msg->front.iov_base;
3157 	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3158 	struct ceph_snap_realm *realm;
3159 	u64 tid;
3160 	int err, result;
3161 	int mds = session->s_mds;
3162 
3163 	if (msg->front.iov_len < sizeof(*head)) {
3164 		pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3165 		ceph_msg_dump(msg);
3166 		return;
3167 	}
3168 
3169 	/* get request, session */
3170 	tid = le64_to_cpu(msg->hdr.tid);
3171 	mutex_lock(&mdsc->mutex);
3172 	req = lookup_get_request(mdsc, tid);
3173 	if (!req) {
3174 		dout("handle_reply on unknown tid %llu\n", tid);
3175 		mutex_unlock(&mdsc->mutex);
3176 		return;
3177 	}
3178 	dout("handle_reply %p\n", req);
3179 
3180 	/* correct session? */
3181 	if (req->r_session != session) {
3182 		pr_err("mdsc_handle_reply got %llu on session mds%d"
3183 		       " not mds%d\n", tid, session->s_mds,
3184 		       req->r_session ? req->r_session->s_mds : -1);
3185 		mutex_unlock(&mdsc->mutex);
3186 		goto out;
3187 	}
3188 
3189 	/* dup? */
3190 	if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3191 	    (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3192 		pr_warn("got a dup %s reply on %llu from mds%d\n",
3193 			   head->safe ? "safe" : "unsafe", tid, mds);
3194 		mutex_unlock(&mdsc->mutex);
3195 		goto out;
3196 	}
3197 	if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3198 		pr_warn("got unsafe after safe on %llu from mds%d\n",
3199 			   tid, mds);
3200 		mutex_unlock(&mdsc->mutex);
3201 		goto out;
3202 	}
3203 
3204 	result = le32_to_cpu(head->result);
3205 
3206 	/*
3207 	 * Handle an ESTALE
3208 	 * if we're not talking to the authority, send to them
3209 	 * if the authority has changed while we weren't looking,
3210 	 * send to new authority
3211 	 * Otherwise we just have to return an ESTALE
3212 	 */
3213 	if (result == -ESTALE) {
3214 		dout("got ESTALE on request %llu\n", req->r_tid);
3215 		req->r_resend_mds = -1;
3216 		if (req->r_direct_mode != USE_AUTH_MDS) {
3217 			dout("not using auth, setting for that now\n");
3218 			req->r_direct_mode = USE_AUTH_MDS;
3219 			__do_request(mdsc, req);
3220 			mutex_unlock(&mdsc->mutex);
3221 			goto out;
3222 		} else  {
3223 			int mds = __choose_mds(mdsc, req, NULL);
3224 			if (mds >= 0 && mds != req->r_session->s_mds) {
3225 				dout("but auth changed, so resending\n");
3226 				__do_request(mdsc, req);
3227 				mutex_unlock(&mdsc->mutex);
3228 				goto out;
3229 			}
3230 		}
3231 		dout("have to return ESTALE on request %llu\n", req->r_tid);
3232 	}
3233 
3234 
3235 	if (head->safe) {
3236 		set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3237 		__unregister_request(mdsc, req);
3238 
3239 		/* last request during umount? */
3240 		if (mdsc->stopping && !__get_oldest_req(mdsc))
3241 			complete_all(&mdsc->safe_umount_waiters);
3242 
3243 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3244 			/*
3245 			 * We already handled the unsafe response, now do the
3246 			 * cleanup.  No need to examine the response; the MDS
3247 			 * doesn't include any result info in the safe
3248 			 * response.  And even if it did, there is nothing
3249 			 * useful we could do with a revised return value.
3250 			 */
3251 			dout("got safe reply %llu, mds%d\n", tid, mds);
3252 
3253 			mutex_unlock(&mdsc->mutex);
3254 			goto out;
3255 		}
3256 	} else {
3257 		set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3258 		list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3259 	}
3260 
3261 	dout("handle_reply tid %lld result %d\n", tid, result);
3262 	rinfo = &req->r_reply_info;
3263 	if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3264 		err = parse_reply_info(session, msg, rinfo, (u64)-1);
3265 	else
3266 		err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3267 	mutex_unlock(&mdsc->mutex);
3268 
3269 	/* Must find target inode outside of mutexes to avoid deadlocks */
3270 	if ((err >= 0) && rinfo->head->is_target) {
3271 		struct inode *in;
3272 		struct ceph_vino tvino = {
3273 			.ino  = le64_to_cpu(rinfo->targeti.in->ino),
3274 			.snap = le64_to_cpu(rinfo->targeti.in->snapid)
3275 		};
3276 
3277 		in = ceph_get_inode(mdsc->fsc->sb, tvino);
3278 		if (IS_ERR(in)) {
3279 			err = PTR_ERR(in);
3280 			mutex_lock(&session->s_mutex);
3281 			goto out_err;
3282 		}
3283 		req->r_target_inode = in;
3284 	}
3285 
3286 	mutex_lock(&session->s_mutex);
3287 	if (err < 0) {
3288 		pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3289 		ceph_msg_dump(msg);
3290 		goto out_err;
3291 	}
3292 
3293 	/* snap trace */
3294 	realm = NULL;
3295 	if (rinfo->snapblob_len) {
3296 		down_write(&mdsc->snap_rwsem);
3297 		ceph_update_snap_trace(mdsc, rinfo->snapblob,
3298 				rinfo->snapblob + rinfo->snapblob_len,
3299 				le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3300 				&realm);
3301 		downgrade_write(&mdsc->snap_rwsem);
3302 	} else {
3303 		down_read(&mdsc->snap_rwsem);
3304 	}
3305 
3306 	/* insert trace into our cache */
3307 	mutex_lock(&req->r_fill_mutex);
3308 	current->journal_info = req;
3309 	err = ceph_fill_trace(mdsc->fsc->sb, req);
3310 	if (err == 0) {
3311 		if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3312 				    req->r_op == CEPH_MDS_OP_LSSNAP))
3313 			ceph_readdir_prepopulate(req, req->r_session);
3314 	}
3315 	current->journal_info = NULL;
3316 	mutex_unlock(&req->r_fill_mutex);
3317 
3318 	up_read(&mdsc->snap_rwsem);
3319 	if (realm)
3320 		ceph_put_snap_realm(mdsc, realm);
3321 
3322 	if (err == 0) {
3323 		if (req->r_target_inode &&
3324 		    test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3325 			struct ceph_inode_info *ci =
3326 				ceph_inode(req->r_target_inode);
3327 			spin_lock(&ci->i_unsafe_lock);
3328 			list_add_tail(&req->r_unsafe_target_item,
3329 				      &ci->i_unsafe_iops);
3330 			spin_unlock(&ci->i_unsafe_lock);
3331 		}
3332 
3333 		ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3334 	}
3335 out_err:
3336 	mutex_lock(&mdsc->mutex);
3337 	if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3338 		if (err) {
3339 			req->r_err = err;
3340 		} else {
3341 			req->r_reply =  ceph_msg_get(msg);
3342 			set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3343 		}
3344 	} else {
3345 		dout("reply arrived after request %lld was aborted\n", tid);
3346 	}
3347 	mutex_unlock(&mdsc->mutex);
3348 
3349 	mutex_unlock(&session->s_mutex);
3350 
3351 	/* kick calling process */
3352 	complete_request(mdsc, req);
3353 
3354 	ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3355 				     req->r_end_latency, err);
3356 out:
3357 	ceph_mdsc_put_request(req);
3358 	return;
3359 }
3360 
3361 
3362 
3363 /*
3364  * handle mds notification that our request has been forwarded.
3365  */
3366 static void handle_forward(struct ceph_mds_client *mdsc,
3367 			   struct ceph_mds_session *session,
3368 			   struct ceph_msg *msg)
3369 {
3370 	struct ceph_mds_request *req;
3371 	u64 tid = le64_to_cpu(msg->hdr.tid);
3372 	u32 next_mds;
3373 	u32 fwd_seq;
3374 	int err = -EINVAL;
3375 	void *p = msg->front.iov_base;
3376 	void *end = p + msg->front.iov_len;
3377 
3378 	ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3379 	next_mds = ceph_decode_32(&p);
3380 	fwd_seq = ceph_decode_32(&p);
3381 
3382 	mutex_lock(&mdsc->mutex);
3383 	req = lookup_get_request(mdsc, tid);
3384 	if (!req) {
3385 		dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3386 		goto out;  /* dup reply? */
3387 	}
3388 
3389 	if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3390 		dout("forward tid %llu aborted, unregistering\n", tid);
3391 		__unregister_request(mdsc, req);
3392 	} else if (fwd_seq <= req->r_num_fwd) {
3393 		dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3394 		     tid, next_mds, req->r_num_fwd, fwd_seq);
3395 	} else {
3396 		/* resend. forward race not possible; mds would drop */
3397 		dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3398 		BUG_ON(req->r_err);
3399 		BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3400 		req->r_attempts = 0;
3401 		req->r_num_fwd = fwd_seq;
3402 		req->r_resend_mds = next_mds;
3403 		put_request_session(req);
3404 		__do_request(mdsc, req);
3405 	}
3406 	ceph_mdsc_put_request(req);
3407 out:
3408 	mutex_unlock(&mdsc->mutex);
3409 	return;
3410 
3411 bad:
3412 	pr_err("mdsc_handle_forward decode error err=%d\n", err);
3413 }
3414 
3415 static int __decode_session_metadata(void **p, void *end,
3416 				     bool *blocklisted)
3417 {
3418 	/* map<string,string> */
3419 	u32 n;
3420 	bool err_str;
3421 	ceph_decode_32_safe(p, end, n, bad);
3422 	while (n-- > 0) {
3423 		u32 len;
3424 		ceph_decode_32_safe(p, end, len, bad);
3425 		ceph_decode_need(p, end, len, bad);
3426 		err_str = !strncmp(*p, "error_string", len);
3427 		*p += len;
3428 		ceph_decode_32_safe(p, end, len, bad);
3429 		ceph_decode_need(p, end, len, bad);
3430 		/*
3431 		 * Match "blocklisted (blacklisted)" from newer MDSes,
3432 		 * or "blacklisted" from older MDSes.
3433 		 */
3434 		if (err_str && strnstr(*p, "blacklisted", len))
3435 			*blocklisted = true;
3436 		*p += len;
3437 	}
3438 	return 0;
3439 bad:
3440 	return -1;
3441 }
3442 
3443 /*
3444  * handle a mds session control message
3445  */
3446 static void handle_session(struct ceph_mds_session *session,
3447 			   struct ceph_msg *msg)
3448 {
3449 	struct ceph_mds_client *mdsc = session->s_mdsc;
3450 	int mds = session->s_mds;
3451 	int msg_version = le16_to_cpu(msg->hdr.version);
3452 	void *p = msg->front.iov_base;
3453 	void *end = p + msg->front.iov_len;
3454 	struct ceph_mds_session_head *h;
3455 	u32 op;
3456 	u64 seq, features = 0;
3457 	int wake = 0;
3458 	bool blocklisted = false;
3459 
3460 	/* decode */
3461 	ceph_decode_need(&p, end, sizeof(*h), bad);
3462 	h = p;
3463 	p += sizeof(*h);
3464 
3465 	op = le32_to_cpu(h->op);
3466 	seq = le64_to_cpu(h->seq);
3467 
3468 	if (msg_version >= 3) {
3469 		u32 len;
3470 		/* version >= 2, metadata */
3471 		if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3472 			goto bad;
3473 		/* version >= 3, feature bits */
3474 		ceph_decode_32_safe(&p, end, len, bad);
3475 		if (len) {
3476 			ceph_decode_64_safe(&p, end, features, bad);
3477 			p += len - sizeof(features);
3478 		}
3479 	}
3480 
3481 	mutex_lock(&mdsc->mutex);
3482 	if (op == CEPH_SESSION_CLOSE) {
3483 		ceph_get_mds_session(session);
3484 		__unregister_session(mdsc, session);
3485 	}
3486 	/* FIXME: this ttl calculation is generous */
3487 	session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3488 	mutex_unlock(&mdsc->mutex);
3489 
3490 	mutex_lock(&session->s_mutex);
3491 
3492 	dout("handle_session mds%d %s %p state %s seq %llu\n",
3493 	     mds, ceph_session_op_name(op), session,
3494 	     ceph_session_state_name(session->s_state), seq);
3495 
3496 	if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3497 		session->s_state = CEPH_MDS_SESSION_OPEN;
3498 		pr_info("mds%d came back\n", session->s_mds);
3499 	}
3500 
3501 	switch (op) {
3502 	case CEPH_SESSION_OPEN:
3503 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3504 			pr_info("mds%d reconnect success\n", session->s_mds);
3505 		session->s_state = CEPH_MDS_SESSION_OPEN;
3506 		session->s_features = features;
3507 		renewed_caps(mdsc, session, 0);
3508 		if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3509 			metric_schedule_delayed(&mdsc->metric);
3510 		wake = 1;
3511 		if (mdsc->stopping)
3512 			__close_session(mdsc, session);
3513 		break;
3514 
3515 	case CEPH_SESSION_RENEWCAPS:
3516 		if (session->s_renew_seq == seq)
3517 			renewed_caps(mdsc, session, 1);
3518 		break;
3519 
3520 	case CEPH_SESSION_CLOSE:
3521 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3522 			pr_info("mds%d reconnect denied\n", session->s_mds);
3523 		session->s_state = CEPH_MDS_SESSION_CLOSED;
3524 		cleanup_session_requests(mdsc, session);
3525 		remove_session_caps(session);
3526 		wake = 2; /* for good measure */
3527 		wake_up_all(&mdsc->session_close_wq);
3528 		break;
3529 
3530 	case CEPH_SESSION_STALE:
3531 		pr_info("mds%d caps went stale, renewing\n",
3532 			session->s_mds);
3533 		atomic_inc(&session->s_cap_gen);
3534 		session->s_cap_ttl = jiffies - 1;
3535 		send_renew_caps(mdsc, session);
3536 		break;
3537 
3538 	case CEPH_SESSION_RECALL_STATE:
3539 		ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3540 		break;
3541 
3542 	case CEPH_SESSION_FLUSHMSG:
3543 		send_flushmsg_ack(mdsc, session, seq);
3544 		break;
3545 
3546 	case CEPH_SESSION_FORCE_RO:
3547 		dout("force_session_readonly %p\n", session);
3548 		spin_lock(&session->s_cap_lock);
3549 		session->s_readonly = true;
3550 		spin_unlock(&session->s_cap_lock);
3551 		wake_up_session_caps(session, FORCE_RO);
3552 		break;
3553 
3554 	case CEPH_SESSION_REJECT:
3555 		WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3556 		pr_info("mds%d rejected session\n", session->s_mds);
3557 		session->s_state = CEPH_MDS_SESSION_REJECTED;
3558 		cleanup_session_requests(mdsc, session);
3559 		remove_session_caps(session);
3560 		if (blocklisted)
3561 			mdsc->fsc->blocklisted = true;
3562 		wake = 2; /* for good measure */
3563 		break;
3564 
3565 	default:
3566 		pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3567 		WARN_ON(1);
3568 	}
3569 
3570 	mutex_unlock(&session->s_mutex);
3571 	if (wake) {
3572 		mutex_lock(&mdsc->mutex);
3573 		__wake_requests(mdsc, &session->s_waiting);
3574 		if (wake == 2)
3575 			kick_requests(mdsc, mds);
3576 		mutex_unlock(&mdsc->mutex);
3577 	}
3578 	if (op == CEPH_SESSION_CLOSE)
3579 		ceph_put_mds_session(session);
3580 	return;
3581 
3582 bad:
3583 	pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3584 	       (int)msg->front.iov_len);
3585 	ceph_msg_dump(msg);
3586 	return;
3587 }
3588 
3589 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3590 {
3591 	int dcaps;
3592 
3593 	dcaps = xchg(&req->r_dir_caps, 0);
3594 	if (dcaps) {
3595 		dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3596 		ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3597 	}
3598 }
3599 
3600 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3601 {
3602 	int dcaps;
3603 
3604 	dcaps = xchg(&req->r_dir_caps, 0);
3605 	if (dcaps) {
3606 		dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3607 		ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3608 						dcaps);
3609 	}
3610 }
3611 
3612 /*
3613  * called under session->mutex.
3614  */
3615 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3616 				   struct ceph_mds_session *session)
3617 {
3618 	struct ceph_mds_request *req, *nreq;
3619 	struct rb_node *p;
3620 
3621 	dout("replay_unsafe_requests mds%d\n", session->s_mds);
3622 
3623 	mutex_lock(&mdsc->mutex);
3624 	list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3625 		__send_request(session, req, true);
3626 
3627 	/*
3628 	 * also re-send old requests when MDS enters reconnect stage. So that MDS
3629 	 * can process completed request in clientreplay stage.
3630 	 */
3631 	p = rb_first(&mdsc->request_tree);
3632 	while (p) {
3633 		req = rb_entry(p, struct ceph_mds_request, r_node);
3634 		p = rb_next(p);
3635 		if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3636 			continue;
3637 		if (req->r_attempts == 0)
3638 			continue; /* only old requests */
3639 		if (!req->r_session)
3640 			continue;
3641 		if (req->r_session->s_mds != session->s_mds)
3642 			continue;
3643 
3644 		ceph_mdsc_release_dir_caps_no_check(req);
3645 
3646 		__send_request(session, req, true);
3647 	}
3648 	mutex_unlock(&mdsc->mutex);
3649 }
3650 
3651 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3652 {
3653 	struct ceph_msg *reply;
3654 	struct ceph_pagelist *_pagelist;
3655 	struct page *page;
3656 	__le32 *addr;
3657 	int err = -ENOMEM;
3658 
3659 	if (!recon_state->allow_multi)
3660 		return -ENOSPC;
3661 
3662 	/* can't handle message that contains both caps and realm */
3663 	BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3664 
3665 	/* pre-allocate new pagelist */
3666 	_pagelist = ceph_pagelist_alloc(GFP_NOFS);
3667 	if (!_pagelist)
3668 		return -ENOMEM;
3669 
3670 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3671 	if (!reply)
3672 		goto fail_msg;
3673 
3674 	/* placeholder for nr_caps */
3675 	err = ceph_pagelist_encode_32(_pagelist, 0);
3676 	if (err < 0)
3677 		goto fail;
3678 
3679 	if (recon_state->nr_caps) {
3680 		/* currently encoding caps */
3681 		err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3682 		if (err)
3683 			goto fail;
3684 	} else {
3685 		/* placeholder for nr_realms (currently encoding relams) */
3686 		err = ceph_pagelist_encode_32(_pagelist, 0);
3687 		if (err < 0)
3688 			goto fail;
3689 	}
3690 
3691 	err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3692 	if (err)
3693 		goto fail;
3694 
3695 	page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3696 	addr = kmap_atomic(page);
3697 	if (recon_state->nr_caps) {
3698 		/* currently encoding caps */
3699 		*addr = cpu_to_le32(recon_state->nr_caps);
3700 	} else {
3701 		/* currently encoding relams */
3702 		*(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3703 	}
3704 	kunmap_atomic(addr);
3705 
3706 	reply->hdr.version = cpu_to_le16(5);
3707 	reply->hdr.compat_version = cpu_to_le16(4);
3708 
3709 	reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3710 	ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3711 
3712 	ceph_con_send(&recon_state->session->s_con, reply);
3713 	ceph_pagelist_release(recon_state->pagelist);
3714 
3715 	recon_state->pagelist = _pagelist;
3716 	recon_state->nr_caps = 0;
3717 	recon_state->nr_realms = 0;
3718 	recon_state->msg_version = 5;
3719 	return 0;
3720 fail:
3721 	ceph_msg_put(reply);
3722 fail_msg:
3723 	ceph_pagelist_release(_pagelist);
3724 	return err;
3725 }
3726 
3727 static struct dentry* d_find_primary(struct inode *inode)
3728 {
3729 	struct dentry *alias, *dn = NULL;
3730 
3731 	if (hlist_empty(&inode->i_dentry))
3732 		return NULL;
3733 
3734 	spin_lock(&inode->i_lock);
3735 	if (hlist_empty(&inode->i_dentry))
3736 		goto out_unlock;
3737 
3738 	if (S_ISDIR(inode->i_mode)) {
3739 		alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3740 		if (!IS_ROOT(alias))
3741 			dn = dget(alias);
3742 		goto out_unlock;
3743 	}
3744 
3745 	hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3746 		spin_lock(&alias->d_lock);
3747 		if (!d_unhashed(alias) &&
3748 		    (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3749 			dn = dget_dlock(alias);
3750 		}
3751 		spin_unlock(&alias->d_lock);
3752 		if (dn)
3753 			break;
3754 	}
3755 out_unlock:
3756 	spin_unlock(&inode->i_lock);
3757 	return dn;
3758 }
3759 
3760 /*
3761  * Encode information about a cap for a reconnect with the MDS.
3762  */
3763 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3764 			  void *arg)
3765 {
3766 	union {
3767 		struct ceph_mds_cap_reconnect v2;
3768 		struct ceph_mds_cap_reconnect_v1 v1;
3769 	} rec;
3770 	struct ceph_inode_info *ci = cap->ci;
3771 	struct ceph_reconnect_state *recon_state = arg;
3772 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3773 	struct dentry *dentry;
3774 	char *path;
3775 	int pathlen, err;
3776 	u64 pathbase;
3777 	u64 snap_follows;
3778 
3779 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3780 	     inode, ceph_vinop(inode), cap, cap->cap_id,
3781 	     ceph_cap_string(cap->issued));
3782 
3783 	dentry = d_find_primary(inode);
3784 	if (dentry) {
3785 		/* set pathbase to parent dir when msg_version >= 2 */
3786 		path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3787 					    recon_state->msg_version >= 2);
3788 		dput(dentry);
3789 		if (IS_ERR(path)) {
3790 			err = PTR_ERR(path);
3791 			goto out_err;
3792 		}
3793 	} else {
3794 		path = NULL;
3795 		pathlen = 0;
3796 		pathbase = 0;
3797 	}
3798 
3799 	spin_lock(&ci->i_ceph_lock);
3800 	cap->seq = 0;        /* reset cap seq */
3801 	cap->issue_seq = 0;  /* and issue_seq */
3802 	cap->mseq = 0;       /* and migrate_seq */
3803 	cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
3804 
3805 	/* These are lost when the session goes away */
3806 	if (S_ISDIR(inode->i_mode)) {
3807 		if (cap->issued & CEPH_CAP_DIR_CREATE) {
3808 			ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3809 			memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3810 		}
3811 		cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3812 	}
3813 
3814 	if (recon_state->msg_version >= 2) {
3815 		rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3816 		rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3817 		rec.v2.issued = cpu_to_le32(cap->issued);
3818 		rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3819 		rec.v2.pathbase = cpu_to_le64(pathbase);
3820 		rec.v2.flock_len = (__force __le32)
3821 			((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3822 	} else {
3823 		rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3824 		rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3825 		rec.v1.issued = cpu_to_le32(cap->issued);
3826 		rec.v1.size = cpu_to_le64(i_size_read(inode));
3827 		ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3828 		ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3829 		rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3830 		rec.v1.pathbase = cpu_to_le64(pathbase);
3831 	}
3832 
3833 	if (list_empty(&ci->i_cap_snaps)) {
3834 		snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3835 	} else {
3836 		struct ceph_cap_snap *capsnap =
3837 			list_first_entry(&ci->i_cap_snaps,
3838 					 struct ceph_cap_snap, ci_item);
3839 		snap_follows = capsnap->follows;
3840 	}
3841 	spin_unlock(&ci->i_ceph_lock);
3842 
3843 	if (recon_state->msg_version >= 2) {
3844 		int num_fcntl_locks, num_flock_locks;
3845 		struct ceph_filelock *flocks = NULL;
3846 		size_t struct_len, total_len = sizeof(u64);
3847 		u8 struct_v = 0;
3848 
3849 encode_again:
3850 		if (rec.v2.flock_len) {
3851 			ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3852 		} else {
3853 			num_fcntl_locks = 0;
3854 			num_flock_locks = 0;
3855 		}
3856 		if (num_fcntl_locks + num_flock_locks > 0) {
3857 			flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3858 					       sizeof(struct ceph_filelock),
3859 					       GFP_NOFS);
3860 			if (!flocks) {
3861 				err = -ENOMEM;
3862 				goto out_err;
3863 			}
3864 			err = ceph_encode_locks_to_buffer(inode, flocks,
3865 							  num_fcntl_locks,
3866 							  num_flock_locks);
3867 			if (err) {
3868 				kfree(flocks);
3869 				flocks = NULL;
3870 				if (err == -ENOSPC)
3871 					goto encode_again;
3872 				goto out_err;
3873 			}
3874 		} else {
3875 			kfree(flocks);
3876 			flocks = NULL;
3877 		}
3878 
3879 		if (recon_state->msg_version >= 3) {
3880 			/* version, compat_version and struct_len */
3881 			total_len += 2 * sizeof(u8) + sizeof(u32);
3882 			struct_v = 2;
3883 		}
3884 		/*
3885 		 * number of encoded locks is stable, so copy to pagelist
3886 		 */
3887 		struct_len = 2 * sizeof(u32) +
3888 			    (num_fcntl_locks + num_flock_locks) *
3889 			    sizeof(struct ceph_filelock);
3890 		rec.v2.flock_len = cpu_to_le32(struct_len);
3891 
3892 		struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3893 
3894 		if (struct_v >= 2)
3895 			struct_len += sizeof(u64); /* snap_follows */
3896 
3897 		total_len += struct_len;
3898 
3899 		if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3900 			err = send_reconnect_partial(recon_state);
3901 			if (err)
3902 				goto out_freeflocks;
3903 			pagelist = recon_state->pagelist;
3904 		}
3905 
3906 		err = ceph_pagelist_reserve(pagelist, total_len);
3907 		if (err)
3908 			goto out_freeflocks;
3909 
3910 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3911 		if (recon_state->msg_version >= 3) {
3912 			ceph_pagelist_encode_8(pagelist, struct_v);
3913 			ceph_pagelist_encode_8(pagelist, 1);
3914 			ceph_pagelist_encode_32(pagelist, struct_len);
3915 		}
3916 		ceph_pagelist_encode_string(pagelist, path, pathlen);
3917 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3918 		ceph_locks_to_pagelist(flocks, pagelist,
3919 				       num_fcntl_locks, num_flock_locks);
3920 		if (struct_v >= 2)
3921 			ceph_pagelist_encode_64(pagelist, snap_follows);
3922 out_freeflocks:
3923 		kfree(flocks);
3924 	} else {
3925 		err = ceph_pagelist_reserve(pagelist,
3926 					    sizeof(u64) + sizeof(u32) +
3927 					    pathlen + sizeof(rec.v1));
3928 		if (err)
3929 			goto out_err;
3930 
3931 		ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3932 		ceph_pagelist_encode_string(pagelist, path, pathlen);
3933 		ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3934 	}
3935 
3936 out_err:
3937 	ceph_mdsc_free_path(path, pathlen);
3938 	if (!err)
3939 		recon_state->nr_caps++;
3940 	return err;
3941 }
3942 
3943 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3944 			      struct ceph_reconnect_state *recon_state)
3945 {
3946 	struct rb_node *p;
3947 	struct ceph_pagelist *pagelist = recon_state->pagelist;
3948 	int err = 0;
3949 
3950 	if (recon_state->msg_version >= 4) {
3951 		err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3952 		if (err < 0)
3953 			goto fail;
3954 	}
3955 
3956 	/*
3957 	 * snaprealms.  we provide mds with the ino, seq (version), and
3958 	 * parent for all of our realms.  If the mds has any newer info,
3959 	 * it will tell us.
3960 	 */
3961 	for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3962 		struct ceph_snap_realm *realm =
3963 		       rb_entry(p, struct ceph_snap_realm, node);
3964 		struct ceph_mds_snaprealm_reconnect sr_rec;
3965 
3966 		if (recon_state->msg_version >= 4) {
3967 			size_t need = sizeof(u8) * 2 + sizeof(u32) +
3968 				      sizeof(sr_rec);
3969 
3970 			if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3971 				err = send_reconnect_partial(recon_state);
3972 				if (err)
3973 					goto fail;
3974 				pagelist = recon_state->pagelist;
3975 			}
3976 
3977 			err = ceph_pagelist_reserve(pagelist, need);
3978 			if (err)
3979 				goto fail;
3980 
3981 			ceph_pagelist_encode_8(pagelist, 1);
3982 			ceph_pagelist_encode_8(pagelist, 1);
3983 			ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3984 		}
3985 
3986 		dout(" adding snap realm %llx seq %lld parent %llx\n",
3987 		     realm->ino, realm->seq, realm->parent_ino);
3988 		sr_rec.ino = cpu_to_le64(realm->ino);
3989 		sr_rec.seq = cpu_to_le64(realm->seq);
3990 		sr_rec.parent = cpu_to_le64(realm->parent_ino);
3991 
3992 		err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3993 		if (err)
3994 			goto fail;
3995 
3996 		recon_state->nr_realms++;
3997 	}
3998 fail:
3999 	return err;
4000 }
4001 
4002 
4003 /*
4004  * If an MDS fails and recovers, clients need to reconnect in order to
4005  * reestablish shared state.  This includes all caps issued through
4006  * this session _and_ the snap_realm hierarchy.  Because it's not
4007  * clear which snap realms the mds cares about, we send everything we
4008  * know about.. that ensures we'll then get any new info the
4009  * recovering MDS might have.
4010  *
4011  * This is a relatively heavyweight operation, but it's rare.
4012  */
4013 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
4014 			       struct ceph_mds_session *session)
4015 {
4016 	struct ceph_msg *reply;
4017 	int mds = session->s_mds;
4018 	int err = -ENOMEM;
4019 	struct ceph_reconnect_state recon_state = {
4020 		.session = session,
4021 	};
4022 	LIST_HEAD(dispose);
4023 
4024 	pr_info("mds%d reconnect start\n", mds);
4025 
4026 	recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4027 	if (!recon_state.pagelist)
4028 		goto fail_nopagelist;
4029 
4030 	reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4031 	if (!reply)
4032 		goto fail_nomsg;
4033 
4034 	xa_destroy(&session->s_delegated_inos);
4035 
4036 	mutex_lock(&session->s_mutex);
4037 	session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4038 	session->s_seq = 0;
4039 
4040 	dout("session %p state %s\n", session,
4041 	     ceph_session_state_name(session->s_state));
4042 
4043 	atomic_inc(&session->s_cap_gen);
4044 
4045 	spin_lock(&session->s_cap_lock);
4046 	/* don't know if session is readonly */
4047 	session->s_readonly = 0;
4048 	/*
4049 	 * notify __ceph_remove_cap() that we are composing cap reconnect.
4050 	 * If a cap get released before being added to the cap reconnect,
4051 	 * __ceph_remove_cap() should skip queuing cap release.
4052 	 */
4053 	session->s_cap_reconnect = 1;
4054 	/* drop old cap expires; we're about to reestablish that state */
4055 	detach_cap_releases(session, &dispose);
4056 	spin_unlock(&session->s_cap_lock);
4057 	dispose_cap_releases(mdsc, &dispose);
4058 
4059 	/* trim unused caps to reduce MDS's cache rejoin time */
4060 	if (mdsc->fsc->sb->s_root)
4061 		shrink_dcache_parent(mdsc->fsc->sb->s_root);
4062 
4063 	ceph_con_close(&session->s_con);
4064 	ceph_con_open(&session->s_con,
4065 		      CEPH_ENTITY_TYPE_MDS, mds,
4066 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4067 
4068 	/* replay unsafe requests */
4069 	replay_unsafe_requests(mdsc, session);
4070 
4071 	ceph_early_kick_flushing_caps(mdsc, session);
4072 
4073 	down_read(&mdsc->snap_rwsem);
4074 
4075 	/* placeholder for nr_caps */
4076 	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4077 	if (err)
4078 		goto fail;
4079 
4080 	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4081 		recon_state.msg_version = 3;
4082 		recon_state.allow_multi = true;
4083 	} else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4084 		recon_state.msg_version = 3;
4085 	} else {
4086 		recon_state.msg_version = 2;
4087 	}
4088 	/* trsaverse this session's caps */
4089 	err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4090 
4091 	spin_lock(&session->s_cap_lock);
4092 	session->s_cap_reconnect = 0;
4093 	spin_unlock(&session->s_cap_lock);
4094 
4095 	if (err < 0)
4096 		goto fail;
4097 
4098 	/* check if all realms can be encoded into current message */
4099 	if (mdsc->num_snap_realms) {
4100 		size_t total_len =
4101 			recon_state.pagelist->length +
4102 			mdsc->num_snap_realms *
4103 			sizeof(struct ceph_mds_snaprealm_reconnect);
4104 		if (recon_state.msg_version >= 4) {
4105 			/* number of realms */
4106 			total_len += sizeof(u32);
4107 			/* version, compat_version and struct_len */
4108 			total_len += mdsc->num_snap_realms *
4109 				     (2 * sizeof(u8) + sizeof(u32));
4110 		}
4111 		if (total_len > RECONNECT_MAX_SIZE) {
4112 			if (!recon_state.allow_multi) {
4113 				err = -ENOSPC;
4114 				goto fail;
4115 			}
4116 			if (recon_state.nr_caps) {
4117 				err = send_reconnect_partial(&recon_state);
4118 				if (err)
4119 					goto fail;
4120 			}
4121 			recon_state.msg_version = 5;
4122 		}
4123 	}
4124 
4125 	err = encode_snap_realms(mdsc, &recon_state);
4126 	if (err < 0)
4127 		goto fail;
4128 
4129 	if (recon_state.msg_version >= 5) {
4130 		err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4131 		if (err < 0)
4132 			goto fail;
4133 	}
4134 
4135 	if (recon_state.nr_caps || recon_state.nr_realms) {
4136 		struct page *page =
4137 			list_first_entry(&recon_state.pagelist->head,
4138 					struct page, lru);
4139 		__le32 *addr = kmap_atomic(page);
4140 		if (recon_state.nr_caps) {
4141 			WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4142 			*addr = cpu_to_le32(recon_state.nr_caps);
4143 		} else if (recon_state.msg_version >= 4) {
4144 			*(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4145 		}
4146 		kunmap_atomic(addr);
4147 	}
4148 
4149 	reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4150 	if (recon_state.msg_version >= 4)
4151 		reply->hdr.compat_version = cpu_to_le16(4);
4152 
4153 	reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4154 	ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4155 
4156 	ceph_con_send(&session->s_con, reply);
4157 
4158 	mutex_unlock(&session->s_mutex);
4159 
4160 	mutex_lock(&mdsc->mutex);
4161 	__wake_requests(mdsc, &session->s_waiting);
4162 	mutex_unlock(&mdsc->mutex);
4163 
4164 	up_read(&mdsc->snap_rwsem);
4165 	ceph_pagelist_release(recon_state.pagelist);
4166 	return;
4167 
4168 fail:
4169 	ceph_msg_put(reply);
4170 	up_read(&mdsc->snap_rwsem);
4171 	mutex_unlock(&session->s_mutex);
4172 fail_nomsg:
4173 	ceph_pagelist_release(recon_state.pagelist);
4174 fail_nopagelist:
4175 	pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4176 	return;
4177 }
4178 
4179 
4180 /*
4181  * compare old and new mdsmaps, kicking requests
4182  * and closing out old connections as necessary
4183  *
4184  * called under mdsc->mutex.
4185  */
4186 static void check_new_map(struct ceph_mds_client *mdsc,
4187 			  struct ceph_mdsmap *newmap,
4188 			  struct ceph_mdsmap *oldmap)
4189 {
4190 	int i, j, err;
4191 	int oldstate, newstate;
4192 	struct ceph_mds_session *s;
4193 	unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
4194 
4195 	dout("check_new_map new %u old %u\n",
4196 	     newmap->m_epoch, oldmap->m_epoch);
4197 
4198 	if (newmap->m_info) {
4199 		for (i = 0; i < newmap->possible_max_rank; i++) {
4200 			for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
4201 				set_bit(newmap->m_info[i].export_targets[j], targets);
4202 		}
4203 	}
4204 
4205 	for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4206 		if (!mdsc->sessions[i])
4207 			continue;
4208 		s = mdsc->sessions[i];
4209 		oldstate = ceph_mdsmap_get_state(oldmap, i);
4210 		newstate = ceph_mdsmap_get_state(newmap, i);
4211 
4212 		dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4213 		     i, ceph_mds_state_name(oldstate),
4214 		     ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4215 		     ceph_mds_state_name(newstate),
4216 		     ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4217 		     ceph_session_state_name(s->s_state));
4218 
4219 		if (i >= newmap->possible_max_rank) {
4220 			/* force close session for stopped mds */
4221 			ceph_get_mds_session(s);
4222 			__unregister_session(mdsc, s);
4223 			__wake_requests(mdsc, &s->s_waiting);
4224 			mutex_unlock(&mdsc->mutex);
4225 
4226 			mutex_lock(&s->s_mutex);
4227 			cleanup_session_requests(mdsc, s);
4228 			remove_session_caps(s);
4229 			mutex_unlock(&s->s_mutex);
4230 
4231 			ceph_put_mds_session(s);
4232 
4233 			mutex_lock(&mdsc->mutex);
4234 			kick_requests(mdsc, i);
4235 			continue;
4236 		}
4237 
4238 		if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4239 			   ceph_mdsmap_get_addr(newmap, i),
4240 			   sizeof(struct ceph_entity_addr))) {
4241 			/* just close it */
4242 			mutex_unlock(&mdsc->mutex);
4243 			mutex_lock(&s->s_mutex);
4244 			mutex_lock(&mdsc->mutex);
4245 			ceph_con_close(&s->s_con);
4246 			mutex_unlock(&s->s_mutex);
4247 			s->s_state = CEPH_MDS_SESSION_RESTARTING;
4248 		} else if (oldstate == newstate) {
4249 			continue;  /* nothing new with this mds */
4250 		}
4251 
4252 		/*
4253 		 * send reconnect?
4254 		 */
4255 		if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4256 		    newstate >= CEPH_MDS_STATE_RECONNECT) {
4257 			mutex_unlock(&mdsc->mutex);
4258 			clear_bit(i, targets);
4259 			send_mds_reconnect(mdsc, s);
4260 			mutex_lock(&mdsc->mutex);
4261 		}
4262 
4263 		/*
4264 		 * kick request on any mds that has gone active.
4265 		 */
4266 		if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4267 		    newstate >= CEPH_MDS_STATE_ACTIVE) {
4268 			if (oldstate != CEPH_MDS_STATE_CREATING &&
4269 			    oldstate != CEPH_MDS_STATE_STARTING)
4270 				pr_info("mds%d recovery completed\n", s->s_mds);
4271 			kick_requests(mdsc, i);
4272 			mutex_unlock(&mdsc->mutex);
4273 			mutex_lock(&s->s_mutex);
4274 			mutex_lock(&mdsc->mutex);
4275 			ceph_kick_flushing_caps(mdsc, s);
4276 			mutex_unlock(&s->s_mutex);
4277 			wake_up_session_caps(s, RECONNECT);
4278 		}
4279 	}
4280 
4281 	/*
4282 	 * Only open and reconnect sessions that don't exist yet.
4283 	 */
4284 	for (i = 0; i < newmap->possible_max_rank; i++) {
4285 		/*
4286 		 * In case the import MDS is crashed just after
4287 		 * the EImportStart journal is flushed, so when
4288 		 * a standby MDS takes over it and is replaying
4289 		 * the EImportStart journal the new MDS daemon
4290 		 * will wait the client to reconnect it, but the
4291 		 * client may never register/open the session yet.
4292 		 *
4293 		 * Will try to reconnect that MDS daemon if the
4294 		 * rank number is in the export targets array and
4295 		 * is the up:reconnect state.
4296 		 */
4297 		newstate = ceph_mdsmap_get_state(newmap, i);
4298 		if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
4299 			continue;
4300 
4301 		/*
4302 		 * The session maybe registered and opened by some
4303 		 * requests which were choosing random MDSes during
4304 		 * the mdsc->mutex's unlock/lock gap below in rare
4305 		 * case. But the related MDS daemon will just queue
4306 		 * that requests and be still waiting for the client's
4307 		 * reconnection request in up:reconnect state.
4308 		 */
4309 		s = __ceph_lookup_mds_session(mdsc, i);
4310 		if (likely(!s)) {
4311 			s = __open_export_target_session(mdsc, i);
4312 			if (IS_ERR(s)) {
4313 				err = PTR_ERR(s);
4314 				pr_err("failed to open export target session, err %d\n",
4315 				       err);
4316 				continue;
4317 			}
4318 		}
4319 		dout("send reconnect to export target mds.%d\n", i);
4320 		mutex_unlock(&mdsc->mutex);
4321 		send_mds_reconnect(mdsc, s);
4322 		ceph_put_mds_session(s);
4323 		mutex_lock(&mdsc->mutex);
4324 	}
4325 
4326 	for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4327 		s = mdsc->sessions[i];
4328 		if (!s)
4329 			continue;
4330 		if (!ceph_mdsmap_is_laggy(newmap, i))
4331 			continue;
4332 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4333 		    s->s_state == CEPH_MDS_SESSION_HUNG ||
4334 		    s->s_state == CEPH_MDS_SESSION_CLOSING) {
4335 			dout(" connecting to export targets of laggy mds%d\n",
4336 			     i);
4337 			__open_export_target_sessions(mdsc, s);
4338 		}
4339 	}
4340 }
4341 
4342 
4343 
4344 /*
4345  * leases
4346  */
4347 
4348 /*
4349  * caller must hold session s_mutex, dentry->d_lock
4350  */
4351 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4352 {
4353 	struct ceph_dentry_info *di = ceph_dentry(dentry);
4354 
4355 	ceph_put_mds_session(di->lease_session);
4356 	di->lease_session = NULL;
4357 }
4358 
4359 static void handle_lease(struct ceph_mds_client *mdsc,
4360 			 struct ceph_mds_session *session,
4361 			 struct ceph_msg *msg)
4362 {
4363 	struct super_block *sb = mdsc->fsc->sb;
4364 	struct inode *inode;
4365 	struct dentry *parent, *dentry;
4366 	struct ceph_dentry_info *di;
4367 	int mds = session->s_mds;
4368 	struct ceph_mds_lease *h = msg->front.iov_base;
4369 	u32 seq;
4370 	struct ceph_vino vino;
4371 	struct qstr dname;
4372 	int release = 0;
4373 
4374 	dout("handle_lease from mds%d\n", mds);
4375 
4376 	/* decode */
4377 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4378 		goto bad;
4379 	vino.ino = le64_to_cpu(h->ino);
4380 	vino.snap = CEPH_NOSNAP;
4381 	seq = le32_to_cpu(h->seq);
4382 	dname.len = get_unaligned_le32(h + 1);
4383 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4384 		goto bad;
4385 	dname.name = (void *)(h + 1) + sizeof(u32);
4386 
4387 	/* lookup inode */
4388 	inode = ceph_find_inode(sb, vino);
4389 	dout("handle_lease %s, ino %llx %p %.*s\n",
4390 	     ceph_lease_op_name(h->action), vino.ino, inode,
4391 	     dname.len, dname.name);
4392 
4393 	mutex_lock(&session->s_mutex);
4394 	inc_session_sequence(session);
4395 
4396 	if (!inode) {
4397 		dout("handle_lease no inode %llx\n", vino.ino);
4398 		goto release;
4399 	}
4400 
4401 	/* dentry */
4402 	parent = d_find_alias(inode);
4403 	if (!parent) {
4404 		dout("no parent dentry on inode %p\n", inode);
4405 		WARN_ON(1);
4406 		goto release;  /* hrm... */
4407 	}
4408 	dname.hash = full_name_hash(parent, dname.name, dname.len);
4409 	dentry = d_lookup(parent, &dname);
4410 	dput(parent);
4411 	if (!dentry)
4412 		goto release;
4413 
4414 	spin_lock(&dentry->d_lock);
4415 	di = ceph_dentry(dentry);
4416 	switch (h->action) {
4417 	case CEPH_MDS_LEASE_REVOKE:
4418 		if (di->lease_session == session) {
4419 			if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4420 				h->seq = cpu_to_le32(di->lease_seq);
4421 			__ceph_mdsc_drop_dentry_lease(dentry);
4422 		}
4423 		release = 1;
4424 		break;
4425 
4426 	case CEPH_MDS_LEASE_RENEW:
4427 		if (di->lease_session == session &&
4428 		    di->lease_gen == atomic_read(&session->s_cap_gen) &&
4429 		    di->lease_renew_from &&
4430 		    di->lease_renew_after == 0) {
4431 			unsigned long duration =
4432 				msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4433 
4434 			di->lease_seq = seq;
4435 			di->time = di->lease_renew_from + duration;
4436 			di->lease_renew_after = di->lease_renew_from +
4437 				(duration >> 1);
4438 			di->lease_renew_from = 0;
4439 		}
4440 		break;
4441 	}
4442 	spin_unlock(&dentry->d_lock);
4443 	dput(dentry);
4444 
4445 	if (!release)
4446 		goto out;
4447 
4448 release:
4449 	/* let's just reuse the same message */
4450 	h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4451 	ceph_msg_get(msg);
4452 	ceph_con_send(&session->s_con, msg);
4453 
4454 out:
4455 	mutex_unlock(&session->s_mutex);
4456 	iput(inode);
4457 	return;
4458 
4459 bad:
4460 	pr_err("corrupt lease message\n");
4461 	ceph_msg_dump(msg);
4462 }
4463 
4464 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4465 			      struct dentry *dentry, char action,
4466 			      u32 seq)
4467 {
4468 	struct ceph_msg *msg;
4469 	struct ceph_mds_lease *lease;
4470 	struct inode *dir;
4471 	int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4472 
4473 	dout("lease_send_msg identry %p %s to mds%d\n",
4474 	     dentry, ceph_lease_op_name(action), session->s_mds);
4475 
4476 	msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4477 	if (!msg)
4478 		return;
4479 	lease = msg->front.iov_base;
4480 	lease->action = action;
4481 	lease->seq = cpu_to_le32(seq);
4482 
4483 	spin_lock(&dentry->d_lock);
4484 	dir = d_inode(dentry->d_parent);
4485 	lease->ino = cpu_to_le64(ceph_ino(dir));
4486 	lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4487 
4488 	put_unaligned_le32(dentry->d_name.len, lease + 1);
4489 	memcpy((void *)(lease + 1) + 4,
4490 	       dentry->d_name.name, dentry->d_name.len);
4491 	spin_unlock(&dentry->d_lock);
4492 	/*
4493 	 * if this is a preemptive lease RELEASE, no need to
4494 	 * flush request stream, since the actual request will
4495 	 * soon follow.
4496 	 */
4497 	msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4498 
4499 	ceph_con_send(&session->s_con, msg);
4500 }
4501 
4502 /*
4503  * lock unlock the session, to wait ongoing session activities
4504  */
4505 static void lock_unlock_session(struct ceph_mds_session *s)
4506 {
4507 	mutex_lock(&s->s_mutex);
4508 	mutex_unlock(&s->s_mutex);
4509 }
4510 
4511 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4512 {
4513 	struct ceph_fs_client *fsc = mdsc->fsc;
4514 
4515 	if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4516 		return;
4517 
4518 	if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4519 		return;
4520 
4521 	if (!READ_ONCE(fsc->blocklisted))
4522 		return;
4523 
4524 	pr_info("auto reconnect after blocklisted\n");
4525 	ceph_force_reconnect(fsc->sb);
4526 }
4527 
4528 bool check_session_state(struct ceph_mds_session *s)
4529 {
4530 	struct ceph_fs_client *fsc = s->s_mdsc->fsc;
4531 
4532 	switch (s->s_state) {
4533 	case CEPH_MDS_SESSION_OPEN:
4534 		if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4535 			s->s_state = CEPH_MDS_SESSION_HUNG;
4536 			pr_info("mds%d hung\n", s->s_mds);
4537 		}
4538 		break;
4539 	case CEPH_MDS_SESSION_CLOSING:
4540 		/* Should never reach this when not force unmounting */
4541 		WARN_ON_ONCE(s->s_ttl &&
4542 			     READ_ONCE(fsc->mount_state) != CEPH_MOUNT_SHUTDOWN);
4543 		fallthrough;
4544 	case CEPH_MDS_SESSION_NEW:
4545 	case CEPH_MDS_SESSION_RESTARTING:
4546 	case CEPH_MDS_SESSION_CLOSED:
4547 	case CEPH_MDS_SESSION_REJECTED:
4548 		return false;
4549 	}
4550 
4551 	return true;
4552 }
4553 
4554 /*
4555  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4556  * then we need to retransmit that request.
4557  */
4558 void inc_session_sequence(struct ceph_mds_session *s)
4559 {
4560 	lockdep_assert_held(&s->s_mutex);
4561 
4562 	s->s_seq++;
4563 
4564 	if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4565 		int ret;
4566 
4567 		dout("resending session close request for mds%d\n", s->s_mds);
4568 		ret = request_close_session(s);
4569 		if (ret < 0)
4570 			pr_err("unable to close session to mds%d: %d\n",
4571 			       s->s_mds, ret);
4572 	}
4573 }
4574 
4575 /*
4576  * delayed work -- periodically trim expired leases, renew caps with mds.  If
4577  * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4578  * workqueue delay value of 5 secs will be used.
4579  */
4580 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4581 {
4582 	unsigned long max_delay = HZ * 5;
4583 
4584 	/* 5 secs default delay */
4585 	if (!delay || (delay > max_delay))
4586 		delay = max_delay;
4587 	schedule_delayed_work(&mdsc->delayed_work,
4588 			      round_jiffies_relative(delay));
4589 }
4590 
4591 static void delayed_work(struct work_struct *work)
4592 {
4593 	struct ceph_mds_client *mdsc =
4594 		container_of(work, struct ceph_mds_client, delayed_work.work);
4595 	unsigned long delay;
4596 	int renew_interval;
4597 	int renew_caps;
4598 	int i;
4599 
4600 	dout("mdsc delayed_work\n");
4601 
4602 	if (mdsc->stopping)
4603 		return;
4604 
4605 	mutex_lock(&mdsc->mutex);
4606 	renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4607 	renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4608 				   mdsc->last_renew_caps);
4609 	if (renew_caps)
4610 		mdsc->last_renew_caps = jiffies;
4611 
4612 	for (i = 0; i < mdsc->max_sessions; i++) {
4613 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4614 		if (!s)
4615 			continue;
4616 
4617 		if (!check_session_state(s)) {
4618 			ceph_put_mds_session(s);
4619 			continue;
4620 		}
4621 		mutex_unlock(&mdsc->mutex);
4622 
4623 		mutex_lock(&s->s_mutex);
4624 		if (renew_caps)
4625 			send_renew_caps(mdsc, s);
4626 		else
4627 			ceph_con_keepalive(&s->s_con);
4628 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4629 		    s->s_state == CEPH_MDS_SESSION_HUNG)
4630 			ceph_send_cap_releases(mdsc, s);
4631 		mutex_unlock(&s->s_mutex);
4632 		ceph_put_mds_session(s);
4633 
4634 		mutex_lock(&mdsc->mutex);
4635 	}
4636 	mutex_unlock(&mdsc->mutex);
4637 
4638 	delay = ceph_check_delayed_caps(mdsc);
4639 
4640 	ceph_queue_cap_reclaim_work(mdsc);
4641 
4642 	ceph_trim_snapid_map(mdsc);
4643 
4644 	maybe_recover_session(mdsc);
4645 
4646 	schedule_delayed(mdsc, delay);
4647 }
4648 
4649 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4650 
4651 {
4652 	struct ceph_mds_client *mdsc;
4653 	int err;
4654 
4655 	mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4656 	if (!mdsc)
4657 		return -ENOMEM;
4658 	mdsc->fsc = fsc;
4659 	mutex_init(&mdsc->mutex);
4660 	mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4661 	if (!mdsc->mdsmap) {
4662 		err = -ENOMEM;
4663 		goto err_mdsc;
4664 	}
4665 
4666 	init_completion(&mdsc->safe_umount_waiters);
4667 	init_waitqueue_head(&mdsc->session_close_wq);
4668 	INIT_LIST_HEAD(&mdsc->waiting_for_map);
4669 	mdsc->quotarealms_inodes = RB_ROOT;
4670 	mutex_init(&mdsc->quotarealms_inodes_mutex);
4671 	init_rwsem(&mdsc->snap_rwsem);
4672 	mdsc->snap_realms = RB_ROOT;
4673 	INIT_LIST_HEAD(&mdsc->snap_empty);
4674 	spin_lock_init(&mdsc->snap_empty_lock);
4675 	mdsc->request_tree = RB_ROOT;
4676 	INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4677 	mdsc->last_renew_caps = jiffies;
4678 	INIT_LIST_HEAD(&mdsc->cap_delay_list);
4679 	INIT_LIST_HEAD(&mdsc->cap_wait_list);
4680 	spin_lock_init(&mdsc->cap_delay_lock);
4681 	INIT_LIST_HEAD(&mdsc->snap_flush_list);
4682 	spin_lock_init(&mdsc->snap_flush_lock);
4683 	mdsc->last_cap_flush_tid = 1;
4684 	INIT_LIST_HEAD(&mdsc->cap_flush_list);
4685 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4686 	spin_lock_init(&mdsc->cap_dirty_lock);
4687 	init_waitqueue_head(&mdsc->cap_flushing_wq);
4688 	INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4689 	err = ceph_metric_init(&mdsc->metric);
4690 	if (err)
4691 		goto err_mdsmap;
4692 
4693 	spin_lock_init(&mdsc->dentry_list_lock);
4694 	INIT_LIST_HEAD(&mdsc->dentry_leases);
4695 	INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4696 
4697 	ceph_caps_init(mdsc);
4698 	ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4699 
4700 	spin_lock_init(&mdsc->snapid_map_lock);
4701 	mdsc->snapid_map_tree = RB_ROOT;
4702 	INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4703 
4704 	init_rwsem(&mdsc->pool_perm_rwsem);
4705 	mdsc->pool_perm_tree = RB_ROOT;
4706 
4707 	strscpy(mdsc->nodename, utsname()->nodename,
4708 		sizeof(mdsc->nodename));
4709 
4710 	fsc->mdsc = mdsc;
4711 	return 0;
4712 
4713 err_mdsmap:
4714 	kfree(mdsc->mdsmap);
4715 err_mdsc:
4716 	kfree(mdsc);
4717 	return err;
4718 }
4719 
4720 /*
4721  * Wait for safe replies on open mds requests.  If we time out, drop
4722  * all requests from the tree to avoid dangling dentry refs.
4723  */
4724 static void wait_requests(struct ceph_mds_client *mdsc)
4725 {
4726 	struct ceph_options *opts = mdsc->fsc->client->options;
4727 	struct ceph_mds_request *req;
4728 
4729 	mutex_lock(&mdsc->mutex);
4730 	if (__get_oldest_req(mdsc)) {
4731 		mutex_unlock(&mdsc->mutex);
4732 
4733 		dout("wait_requests waiting for requests\n");
4734 		wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4735 				    ceph_timeout_jiffies(opts->mount_timeout));
4736 
4737 		/* tear down remaining requests */
4738 		mutex_lock(&mdsc->mutex);
4739 		while ((req = __get_oldest_req(mdsc))) {
4740 			dout("wait_requests timed out on tid %llu\n",
4741 			     req->r_tid);
4742 			list_del_init(&req->r_wait);
4743 			__unregister_request(mdsc, req);
4744 		}
4745 	}
4746 	mutex_unlock(&mdsc->mutex);
4747 	dout("wait_requests done\n");
4748 }
4749 
4750 void send_flush_mdlog(struct ceph_mds_session *s)
4751 {
4752 	struct ceph_msg *msg;
4753 
4754 	/*
4755 	 * Pre-luminous MDS crashes when it sees an unknown session request
4756 	 */
4757 	if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
4758 		return;
4759 
4760 	mutex_lock(&s->s_mutex);
4761 	dout("request mdlog flush to mds%d (%s)s seq %lld\n", s->s_mds,
4762 	     ceph_session_state_name(s->s_state), s->s_seq);
4763 	msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
4764 				      s->s_seq);
4765 	if (!msg) {
4766 		pr_err("failed to request mdlog flush to mds%d (%s) seq %lld\n",
4767 		       s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
4768 	} else {
4769 		ceph_con_send(&s->s_con, msg);
4770 	}
4771 	mutex_unlock(&s->s_mutex);
4772 }
4773 
4774 /*
4775  * called before mount is ro, and before dentries are torn down.
4776  * (hmm, does this still race with new lookups?)
4777  */
4778 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4779 {
4780 	dout("pre_umount\n");
4781 	mdsc->stopping = 1;
4782 
4783 	ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
4784 	ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
4785 	ceph_flush_dirty_caps(mdsc);
4786 	wait_requests(mdsc);
4787 
4788 	/*
4789 	 * wait for reply handlers to drop their request refs and
4790 	 * their inode/dcache refs
4791 	 */
4792 	ceph_msgr_flush();
4793 
4794 	ceph_cleanup_quotarealms_inodes(mdsc);
4795 }
4796 
4797 /*
4798  * wait for all write mds requests to flush.
4799  */
4800 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4801 {
4802 	struct ceph_mds_request *req = NULL, *nextreq;
4803 	struct rb_node *n;
4804 
4805 	mutex_lock(&mdsc->mutex);
4806 	dout("wait_unsafe_requests want %lld\n", want_tid);
4807 restart:
4808 	req = __get_oldest_req(mdsc);
4809 	while (req && req->r_tid <= want_tid) {
4810 		/* find next request */
4811 		n = rb_next(&req->r_node);
4812 		if (n)
4813 			nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4814 		else
4815 			nextreq = NULL;
4816 		if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4817 		    (req->r_op & CEPH_MDS_OP_WRITE)) {
4818 			/* write op */
4819 			ceph_mdsc_get_request(req);
4820 			if (nextreq)
4821 				ceph_mdsc_get_request(nextreq);
4822 			mutex_unlock(&mdsc->mutex);
4823 			dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4824 			     req->r_tid, want_tid);
4825 			wait_for_completion(&req->r_safe_completion);
4826 			mutex_lock(&mdsc->mutex);
4827 			ceph_mdsc_put_request(req);
4828 			if (!nextreq)
4829 				break;  /* next dne before, so we're done! */
4830 			if (RB_EMPTY_NODE(&nextreq->r_node)) {
4831 				/* next request was removed from tree */
4832 				ceph_mdsc_put_request(nextreq);
4833 				goto restart;
4834 			}
4835 			ceph_mdsc_put_request(nextreq);  /* won't go away */
4836 		}
4837 		req = nextreq;
4838 	}
4839 	mutex_unlock(&mdsc->mutex);
4840 	dout("wait_unsafe_requests done\n");
4841 }
4842 
4843 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4844 {
4845 	u64 want_tid, want_flush;
4846 
4847 	if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4848 		return;
4849 
4850 	dout("sync\n");
4851 	mutex_lock(&mdsc->mutex);
4852 	want_tid = mdsc->last_tid;
4853 	mutex_unlock(&mdsc->mutex);
4854 
4855 	ceph_flush_dirty_caps(mdsc);
4856 	spin_lock(&mdsc->cap_dirty_lock);
4857 	want_flush = mdsc->last_cap_flush_tid;
4858 	if (!list_empty(&mdsc->cap_flush_list)) {
4859 		struct ceph_cap_flush *cf =
4860 			list_last_entry(&mdsc->cap_flush_list,
4861 					struct ceph_cap_flush, g_list);
4862 		cf->wake = true;
4863 	}
4864 	spin_unlock(&mdsc->cap_dirty_lock);
4865 
4866 	dout("sync want tid %lld flush_seq %lld\n",
4867 	     want_tid, want_flush);
4868 
4869 	wait_unsafe_requests(mdsc, want_tid);
4870 	wait_caps_flush(mdsc, want_flush);
4871 }
4872 
4873 /*
4874  * true if all sessions are closed, or we force unmount
4875  */
4876 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4877 {
4878 	if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4879 		return true;
4880 	return atomic_read(&mdsc->num_sessions) <= skipped;
4881 }
4882 
4883 /*
4884  * called after sb is ro.
4885  */
4886 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4887 {
4888 	struct ceph_options *opts = mdsc->fsc->client->options;
4889 	struct ceph_mds_session *session;
4890 	int i;
4891 	int skipped = 0;
4892 
4893 	dout("close_sessions\n");
4894 
4895 	/* close sessions */
4896 	mutex_lock(&mdsc->mutex);
4897 	for (i = 0; i < mdsc->max_sessions; i++) {
4898 		session = __ceph_lookup_mds_session(mdsc, i);
4899 		if (!session)
4900 			continue;
4901 		mutex_unlock(&mdsc->mutex);
4902 		mutex_lock(&session->s_mutex);
4903 		if (__close_session(mdsc, session) <= 0)
4904 			skipped++;
4905 		mutex_unlock(&session->s_mutex);
4906 		ceph_put_mds_session(session);
4907 		mutex_lock(&mdsc->mutex);
4908 	}
4909 	mutex_unlock(&mdsc->mutex);
4910 
4911 	dout("waiting for sessions to close\n");
4912 	wait_event_timeout(mdsc->session_close_wq,
4913 			   done_closing_sessions(mdsc, skipped),
4914 			   ceph_timeout_jiffies(opts->mount_timeout));
4915 
4916 	/* tear down remaining sessions */
4917 	mutex_lock(&mdsc->mutex);
4918 	for (i = 0; i < mdsc->max_sessions; i++) {
4919 		if (mdsc->sessions[i]) {
4920 			session = ceph_get_mds_session(mdsc->sessions[i]);
4921 			__unregister_session(mdsc, session);
4922 			mutex_unlock(&mdsc->mutex);
4923 			mutex_lock(&session->s_mutex);
4924 			remove_session_caps(session);
4925 			mutex_unlock(&session->s_mutex);
4926 			ceph_put_mds_session(session);
4927 			mutex_lock(&mdsc->mutex);
4928 		}
4929 	}
4930 	WARN_ON(!list_empty(&mdsc->cap_delay_list));
4931 	mutex_unlock(&mdsc->mutex);
4932 
4933 	ceph_cleanup_snapid_map(mdsc);
4934 	ceph_cleanup_empty_realms(mdsc);
4935 
4936 	cancel_work_sync(&mdsc->cap_reclaim_work);
4937 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4938 
4939 	dout("stopped\n");
4940 }
4941 
4942 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4943 {
4944 	struct ceph_mds_session *session;
4945 	int mds;
4946 
4947 	dout("force umount\n");
4948 
4949 	mutex_lock(&mdsc->mutex);
4950 	for (mds = 0; mds < mdsc->max_sessions; mds++) {
4951 		session = __ceph_lookup_mds_session(mdsc, mds);
4952 		if (!session)
4953 			continue;
4954 
4955 		if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4956 			__unregister_session(mdsc, session);
4957 		__wake_requests(mdsc, &session->s_waiting);
4958 		mutex_unlock(&mdsc->mutex);
4959 
4960 		mutex_lock(&session->s_mutex);
4961 		__close_session(mdsc, session);
4962 		if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4963 			cleanup_session_requests(mdsc, session);
4964 			remove_session_caps(session);
4965 		}
4966 		mutex_unlock(&session->s_mutex);
4967 		ceph_put_mds_session(session);
4968 
4969 		mutex_lock(&mdsc->mutex);
4970 		kick_requests(mdsc, mds);
4971 	}
4972 	__wake_requests(mdsc, &mdsc->waiting_for_map);
4973 	mutex_unlock(&mdsc->mutex);
4974 }
4975 
4976 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4977 {
4978 	dout("stop\n");
4979 	/*
4980 	 * Make sure the delayed work stopped before releasing
4981 	 * the resources.
4982 	 *
4983 	 * Because the cancel_delayed_work_sync() will only
4984 	 * guarantee that the work finishes executing. But the
4985 	 * delayed work will re-arm itself again after that.
4986 	 */
4987 	flush_delayed_work(&mdsc->delayed_work);
4988 
4989 	if (mdsc->mdsmap)
4990 		ceph_mdsmap_destroy(mdsc->mdsmap);
4991 	kfree(mdsc->sessions);
4992 	ceph_caps_finalize(mdsc);
4993 	ceph_pool_perm_destroy(mdsc);
4994 }
4995 
4996 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4997 {
4998 	struct ceph_mds_client *mdsc = fsc->mdsc;
4999 	dout("mdsc_destroy %p\n", mdsc);
5000 
5001 	if (!mdsc)
5002 		return;
5003 
5004 	/* flush out any connection work with references to us */
5005 	ceph_msgr_flush();
5006 
5007 	ceph_mdsc_stop(mdsc);
5008 
5009 	ceph_metric_destroy(&mdsc->metric);
5010 
5011 	fsc->mdsc = NULL;
5012 	kfree(mdsc);
5013 	dout("mdsc_destroy %p done\n", mdsc);
5014 }
5015 
5016 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5017 {
5018 	struct ceph_fs_client *fsc = mdsc->fsc;
5019 	const char *mds_namespace = fsc->mount_options->mds_namespace;
5020 	void *p = msg->front.iov_base;
5021 	void *end = p + msg->front.iov_len;
5022 	u32 epoch;
5023 	u32 num_fs;
5024 	u32 mount_fscid = (u32)-1;
5025 	int err = -EINVAL;
5026 
5027 	ceph_decode_need(&p, end, sizeof(u32), bad);
5028 	epoch = ceph_decode_32(&p);
5029 
5030 	dout("handle_fsmap epoch %u\n", epoch);
5031 
5032 	/* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
5033 	ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
5034 
5035 	ceph_decode_32_safe(&p, end, num_fs, bad);
5036 	while (num_fs-- > 0) {
5037 		void *info_p, *info_end;
5038 		u32 info_len;
5039 		u32 fscid, namelen;
5040 
5041 		ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
5042 		p += 2;		// info_v, info_cv
5043 		info_len = ceph_decode_32(&p);
5044 		ceph_decode_need(&p, end, info_len, bad);
5045 		info_p = p;
5046 		info_end = p + info_len;
5047 		p = info_end;
5048 
5049 		ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
5050 		fscid = ceph_decode_32(&info_p);
5051 		namelen = ceph_decode_32(&info_p);
5052 		ceph_decode_need(&info_p, info_end, namelen, bad);
5053 
5054 		if (mds_namespace &&
5055 		    strlen(mds_namespace) == namelen &&
5056 		    !strncmp(mds_namespace, (char *)info_p, namelen)) {
5057 			mount_fscid = fscid;
5058 			break;
5059 		}
5060 	}
5061 
5062 	ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
5063 	if (mount_fscid != (u32)-1) {
5064 		fsc->client->monc.fs_cluster_id = mount_fscid;
5065 		ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
5066 				   0, true);
5067 		ceph_monc_renew_subs(&fsc->client->monc);
5068 	} else {
5069 		err = -ENOENT;
5070 		goto err_out;
5071 	}
5072 	return;
5073 
5074 bad:
5075 	pr_err("error decoding fsmap\n");
5076 err_out:
5077 	mutex_lock(&mdsc->mutex);
5078 	mdsc->mdsmap_err = err;
5079 	__wake_requests(mdsc, &mdsc->waiting_for_map);
5080 	mutex_unlock(&mdsc->mutex);
5081 }
5082 
5083 /*
5084  * handle mds map update.
5085  */
5086 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5087 {
5088 	u32 epoch;
5089 	u32 maplen;
5090 	void *p = msg->front.iov_base;
5091 	void *end = p + msg->front.iov_len;
5092 	struct ceph_mdsmap *newmap, *oldmap;
5093 	struct ceph_fsid fsid;
5094 	int err = -EINVAL;
5095 
5096 	ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5097 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
5098 	if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5099 		return;
5100 	epoch = ceph_decode_32(&p);
5101 	maplen = ceph_decode_32(&p);
5102 	dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5103 
5104 	/* do we need it? */
5105 	mutex_lock(&mdsc->mutex);
5106 	if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5107 		dout("handle_map epoch %u <= our %u\n",
5108 		     epoch, mdsc->mdsmap->m_epoch);
5109 		mutex_unlock(&mdsc->mutex);
5110 		return;
5111 	}
5112 
5113 	newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5114 	if (IS_ERR(newmap)) {
5115 		err = PTR_ERR(newmap);
5116 		goto bad_unlock;
5117 	}
5118 
5119 	/* swap into place */
5120 	if (mdsc->mdsmap) {
5121 		oldmap = mdsc->mdsmap;
5122 		mdsc->mdsmap = newmap;
5123 		check_new_map(mdsc, newmap, oldmap);
5124 		ceph_mdsmap_destroy(oldmap);
5125 	} else {
5126 		mdsc->mdsmap = newmap;  /* first mds map */
5127 	}
5128 	mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5129 					MAX_LFS_FILESIZE);
5130 
5131 	__wake_requests(mdsc, &mdsc->waiting_for_map);
5132 	ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5133 			  mdsc->mdsmap->m_epoch);
5134 
5135 	mutex_unlock(&mdsc->mutex);
5136 	schedule_delayed(mdsc, 0);
5137 	return;
5138 
5139 bad_unlock:
5140 	mutex_unlock(&mdsc->mutex);
5141 bad:
5142 	pr_err("error decoding mdsmap %d\n", err);
5143 	return;
5144 }
5145 
5146 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5147 {
5148 	struct ceph_mds_session *s = con->private;
5149 
5150 	if (ceph_get_mds_session(s))
5151 		return con;
5152 	return NULL;
5153 }
5154 
5155 static void mds_put_con(struct ceph_connection *con)
5156 {
5157 	struct ceph_mds_session *s = con->private;
5158 
5159 	ceph_put_mds_session(s);
5160 }
5161 
5162 /*
5163  * if the client is unresponsive for long enough, the mds will kill
5164  * the session entirely.
5165  */
5166 static void mds_peer_reset(struct ceph_connection *con)
5167 {
5168 	struct ceph_mds_session *s = con->private;
5169 	struct ceph_mds_client *mdsc = s->s_mdsc;
5170 
5171 	pr_warn("mds%d closed our session\n", s->s_mds);
5172 	send_mds_reconnect(mdsc, s);
5173 }
5174 
5175 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5176 {
5177 	struct ceph_mds_session *s = con->private;
5178 	struct ceph_mds_client *mdsc = s->s_mdsc;
5179 	int type = le16_to_cpu(msg->hdr.type);
5180 
5181 	mutex_lock(&mdsc->mutex);
5182 	if (__verify_registered_session(mdsc, s) < 0) {
5183 		mutex_unlock(&mdsc->mutex);
5184 		goto out;
5185 	}
5186 	mutex_unlock(&mdsc->mutex);
5187 
5188 	switch (type) {
5189 	case CEPH_MSG_MDS_MAP:
5190 		ceph_mdsc_handle_mdsmap(mdsc, msg);
5191 		break;
5192 	case CEPH_MSG_FS_MAP_USER:
5193 		ceph_mdsc_handle_fsmap(mdsc, msg);
5194 		break;
5195 	case CEPH_MSG_CLIENT_SESSION:
5196 		handle_session(s, msg);
5197 		break;
5198 	case CEPH_MSG_CLIENT_REPLY:
5199 		handle_reply(s, msg);
5200 		break;
5201 	case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5202 		handle_forward(mdsc, s, msg);
5203 		break;
5204 	case CEPH_MSG_CLIENT_CAPS:
5205 		ceph_handle_caps(s, msg);
5206 		break;
5207 	case CEPH_MSG_CLIENT_SNAP:
5208 		ceph_handle_snap(mdsc, s, msg);
5209 		break;
5210 	case CEPH_MSG_CLIENT_LEASE:
5211 		handle_lease(mdsc, s, msg);
5212 		break;
5213 	case CEPH_MSG_CLIENT_QUOTA:
5214 		ceph_handle_quota(mdsc, s, msg);
5215 		break;
5216 
5217 	default:
5218 		pr_err("received unknown message type %d %s\n", type,
5219 		       ceph_msg_type_name(type));
5220 	}
5221 out:
5222 	ceph_msg_put(msg);
5223 }
5224 
5225 /*
5226  * authentication
5227  */
5228 
5229 /*
5230  * Note: returned pointer is the address of a structure that's
5231  * managed separately.  Caller must *not* attempt to free it.
5232  */
5233 static struct ceph_auth_handshake *
5234 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5235 {
5236 	struct ceph_mds_session *s = con->private;
5237 	struct ceph_mds_client *mdsc = s->s_mdsc;
5238 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5239 	struct ceph_auth_handshake *auth = &s->s_auth;
5240 	int ret;
5241 
5242 	ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5243 					 force_new, proto, NULL, NULL);
5244 	if (ret)
5245 		return ERR_PTR(ret);
5246 
5247 	return auth;
5248 }
5249 
5250 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5251 				    void *challenge_buf, int challenge_buf_len)
5252 {
5253 	struct ceph_mds_session *s = con->private;
5254 	struct ceph_mds_client *mdsc = s->s_mdsc;
5255 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5256 
5257 	return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5258 					    challenge_buf, challenge_buf_len);
5259 }
5260 
5261 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5262 {
5263 	struct ceph_mds_session *s = con->private;
5264 	struct ceph_mds_client *mdsc = s->s_mdsc;
5265 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5266 	struct ceph_auth_handshake *auth = &s->s_auth;
5267 
5268 	return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5269 		auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5270 		NULL, NULL, NULL, NULL);
5271 }
5272 
5273 static int mds_invalidate_authorizer(struct ceph_connection *con)
5274 {
5275 	struct ceph_mds_session *s = con->private;
5276 	struct ceph_mds_client *mdsc = s->s_mdsc;
5277 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5278 
5279 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5280 
5281 	return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5282 }
5283 
5284 static int mds_get_auth_request(struct ceph_connection *con,
5285 				void *buf, int *buf_len,
5286 				void **authorizer, int *authorizer_len)
5287 {
5288 	struct ceph_mds_session *s = con->private;
5289 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5290 	struct ceph_auth_handshake *auth = &s->s_auth;
5291 	int ret;
5292 
5293 	ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5294 				       buf, buf_len);
5295 	if (ret)
5296 		return ret;
5297 
5298 	*authorizer = auth->authorizer_buf;
5299 	*authorizer_len = auth->authorizer_buf_len;
5300 	return 0;
5301 }
5302 
5303 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5304 				      void *reply, int reply_len,
5305 				      void *buf, int *buf_len,
5306 				      void **authorizer, int *authorizer_len)
5307 {
5308 	struct ceph_mds_session *s = con->private;
5309 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5310 	struct ceph_auth_handshake *auth = &s->s_auth;
5311 	int ret;
5312 
5313 	ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5314 					      buf, buf_len);
5315 	if (ret)
5316 		return ret;
5317 
5318 	*authorizer = auth->authorizer_buf;
5319 	*authorizer_len = auth->authorizer_buf_len;
5320 	return 0;
5321 }
5322 
5323 static int mds_handle_auth_done(struct ceph_connection *con,
5324 				u64 global_id, void *reply, int reply_len,
5325 				u8 *session_key, int *session_key_len,
5326 				u8 *con_secret, int *con_secret_len)
5327 {
5328 	struct ceph_mds_session *s = con->private;
5329 	struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5330 	struct ceph_auth_handshake *auth = &s->s_auth;
5331 
5332 	return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5333 					       session_key, session_key_len,
5334 					       con_secret, con_secret_len);
5335 }
5336 
5337 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5338 				      int used_proto, int result,
5339 				      const int *allowed_protos, int proto_cnt,
5340 				      const int *allowed_modes, int mode_cnt)
5341 {
5342 	struct ceph_mds_session *s = con->private;
5343 	struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5344 	int ret;
5345 
5346 	if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5347 					    used_proto, result,
5348 					    allowed_protos, proto_cnt,
5349 					    allowed_modes, mode_cnt)) {
5350 		ret = ceph_monc_validate_auth(monc);
5351 		if (ret)
5352 			return ret;
5353 	}
5354 
5355 	return -EACCES;
5356 }
5357 
5358 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5359 				struct ceph_msg_header *hdr, int *skip)
5360 {
5361 	struct ceph_msg *msg;
5362 	int type = (int) le16_to_cpu(hdr->type);
5363 	int front_len = (int) le32_to_cpu(hdr->front_len);
5364 
5365 	if (con->in_msg)
5366 		return con->in_msg;
5367 
5368 	*skip = 0;
5369 	msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5370 	if (!msg) {
5371 		pr_err("unable to allocate msg type %d len %d\n",
5372 		       type, front_len);
5373 		return NULL;
5374 	}
5375 
5376 	return msg;
5377 }
5378 
5379 static int mds_sign_message(struct ceph_msg *msg)
5380 {
5381        struct ceph_mds_session *s = msg->con->private;
5382        struct ceph_auth_handshake *auth = &s->s_auth;
5383 
5384        return ceph_auth_sign_message(auth, msg);
5385 }
5386 
5387 static int mds_check_message_signature(struct ceph_msg *msg)
5388 {
5389        struct ceph_mds_session *s = msg->con->private;
5390        struct ceph_auth_handshake *auth = &s->s_auth;
5391 
5392        return ceph_auth_check_message_signature(auth, msg);
5393 }
5394 
5395 static const struct ceph_connection_operations mds_con_ops = {
5396 	.get = mds_get_con,
5397 	.put = mds_put_con,
5398 	.alloc_msg = mds_alloc_msg,
5399 	.dispatch = mds_dispatch,
5400 	.peer_reset = mds_peer_reset,
5401 	.get_authorizer = mds_get_authorizer,
5402 	.add_authorizer_challenge = mds_add_authorizer_challenge,
5403 	.verify_authorizer_reply = mds_verify_authorizer_reply,
5404 	.invalidate_authorizer = mds_invalidate_authorizer,
5405 	.sign_message = mds_sign_message,
5406 	.check_message_signature = mds_check_message_signature,
5407 	.get_auth_request = mds_get_auth_request,
5408 	.handle_auth_reply_more = mds_handle_auth_reply_more,
5409 	.handle_auth_done = mds_handle_auth_done,
5410 	.handle_auth_bad_method = mds_handle_auth_bad_method,
5411 };
5412 
5413 /* eof */
5414