xref: /openbmc/linux/fs/ceph/mds_client.c (revision 034f90b3)
1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/fs.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
6 #include <linux/gfp.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11 
12 #include "super.h"
13 #include "mds_client.h"
14 
15 #include <linux/ceph/ceph_features.h>
16 #include <linux/ceph/messenger.h>
17 #include <linux/ceph/decode.h>
18 #include <linux/ceph/pagelist.h>
19 #include <linux/ceph/auth.h>
20 #include <linux/ceph/debugfs.h>
21 
22 /*
23  * A cluster of MDS (metadata server) daemons is responsible for
24  * managing the file system namespace (the directory hierarchy and
25  * inodes) and for coordinating shared access to storage.  Metadata is
26  * partitioning hierarchically across a number of servers, and that
27  * partition varies over time as the cluster adjusts the distribution
28  * in order to balance load.
29  *
30  * The MDS client is primarily responsible to managing synchronous
31  * metadata requests for operations like open, unlink, and so forth.
32  * If there is a MDS failure, we find out about it when we (possibly
33  * request and) receive a new MDS map, and can resubmit affected
34  * requests.
35  *
36  * For the most part, though, we take advantage of a lossless
37  * communications channel to the MDS, and do not need to worry about
38  * timing out or resubmitting requests.
39  *
40  * We maintain a stateful "session" with each MDS we interact with.
41  * Within each session, we sent periodic heartbeat messages to ensure
42  * any capabilities or leases we have been issues remain valid.  If
43  * the session times out and goes stale, our leases and capabilities
44  * are no longer valid.
45  */
46 
47 struct ceph_reconnect_state {
48 	int nr_caps;
49 	struct ceph_pagelist *pagelist;
50 	bool flock;
51 };
52 
53 static void __wake_requests(struct ceph_mds_client *mdsc,
54 			    struct list_head *head);
55 
56 static const struct ceph_connection_operations mds_con_ops;
57 
58 
59 /*
60  * mds reply parsing
61  */
62 
63 /*
64  * parse individual inode info
65  */
66 static int parse_reply_info_in(void **p, void *end,
67 			       struct ceph_mds_reply_info_in *info,
68 			       u64 features)
69 {
70 	int err = -EIO;
71 
72 	info->in = *p;
73 	*p += sizeof(struct ceph_mds_reply_inode) +
74 		sizeof(*info->in->fragtree.splits) *
75 		le32_to_cpu(info->in->fragtree.nsplits);
76 
77 	ceph_decode_32_safe(p, end, info->symlink_len, bad);
78 	ceph_decode_need(p, end, info->symlink_len, bad);
79 	info->symlink = *p;
80 	*p += info->symlink_len;
81 
82 	if (features & CEPH_FEATURE_DIRLAYOUTHASH)
83 		ceph_decode_copy_safe(p, end, &info->dir_layout,
84 				      sizeof(info->dir_layout), bad);
85 	else
86 		memset(&info->dir_layout, 0, sizeof(info->dir_layout));
87 
88 	ceph_decode_32_safe(p, end, info->xattr_len, bad);
89 	ceph_decode_need(p, end, info->xattr_len, bad);
90 	info->xattr_data = *p;
91 	*p += info->xattr_len;
92 
93 	if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
94 		ceph_decode_64_safe(p, end, info->inline_version, bad);
95 		ceph_decode_32_safe(p, end, info->inline_len, bad);
96 		ceph_decode_need(p, end, info->inline_len, bad);
97 		info->inline_data = *p;
98 		*p += info->inline_len;
99 	} else
100 		info->inline_version = CEPH_INLINE_NONE;
101 
102 	return 0;
103 bad:
104 	return err;
105 }
106 
107 /*
108  * parse a normal reply, which may contain a (dir+)dentry and/or a
109  * target inode.
110  */
111 static int parse_reply_info_trace(void **p, void *end,
112 				  struct ceph_mds_reply_info_parsed *info,
113 				  u64 features)
114 {
115 	int err;
116 
117 	if (info->head->is_dentry) {
118 		err = parse_reply_info_in(p, end, &info->diri, features);
119 		if (err < 0)
120 			goto out_bad;
121 
122 		if (unlikely(*p + sizeof(*info->dirfrag) > end))
123 			goto bad;
124 		info->dirfrag = *p;
125 		*p += sizeof(*info->dirfrag) +
126 			sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
127 		if (unlikely(*p > end))
128 			goto bad;
129 
130 		ceph_decode_32_safe(p, end, info->dname_len, bad);
131 		ceph_decode_need(p, end, info->dname_len, bad);
132 		info->dname = *p;
133 		*p += info->dname_len;
134 		info->dlease = *p;
135 		*p += sizeof(*info->dlease);
136 	}
137 
138 	if (info->head->is_target) {
139 		err = parse_reply_info_in(p, end, &info->targeti, features);
140 		if (err < 0)
141 			goto out_bad;
142 	}
143 
144 	if (unlikely(*p != end))
145 		goto bad;
146 	return 0;
147 
148 bad:
149 	err = -EIO;
150 out_bad:
151 	pr_err("problem parsing mds trace %d\n", err);
152 	return err;
153 }
154 
155 /*
156  * parse readdir results
157  */
158 static int parse_reply_info_dir(void **p, void *end,
159 				struct ceph_mds_reply_info_parsed *info,
160 				u64 features)
161 {
162 	u32 num, i = 0;
163 	int err;
164 
165 	info->dir_dir = *p;
166 	if (*p + sizeof(*info->dir_dir) > end)
167 		goto bad;
168 	*p += sizeof(*info->dir_dir) +
169 		sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
170 	if (*p > end)
171 		goto bad;
172 
173 	ceph_decode_need(p, end, sizeof(num) + 2, bad);
174 	num = ceph_decode_32(p);
175 	info->dir_end = ceph_decode_8(p);
176 	info->dir_complete = ceph_decode_8(p);
177 	if (num == 0)
178 		goto done;
179 
180 	BUG_ON(!info->dir_in);
181 	info->dir_dname = (void *)(info->dir_in + num);
182 	info->dir_dname_len = (void *)(info->dir_dname + num);
183 	info->dir_dlease = (void *)(info->dir_dname_len + num);
184 	if ((unsigned long)(info->dir_dlease + num) >
185 	    (unsigned long)info->dir_in + info->dir_buf_size) {
186 		pr_err("dir contents are larger than expected\n");
187 		WARN_ON(1);
188 		goto bad;
189 	}
190 
191 	info->dir_nr = num;
192 	while (num) {
193 		/* dentry */
194 		ceph_decode_need(p, end, sizeof(u32)*2, bad);
195 		info->dir_dname_len[i] = ceph_decode_32(p);
196 		ceph_decode_need(p, end, info->dir_dname_len[i], bad);
197 		info->dir_dname[i] = *p;
198 		*p += info->dir_dname_len[i];
199 		dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
200 		     info->dir_dname[i]);
201 		info->dir_dlease[i] = *p;
202 		*p += sizeof(struct ceph_mds_reply_lease);
203 
204 		/* inode */
205 		err = parse_reply_info_in(p, end, &info->dir_in[i], features);
206 		if (err < 0)
207 			goto out_bad;
208 		i++;
209 		num--;
210 	}
211 
212 done:
213 	if (*p != end)
214 		goto bad;
215 	return 0;
216 
217 bad:
218 	err = -EIO;
219 out_bad:
220 	pr_err("problem parsing dir contents %d\n", err);
221 	return err;
222 }
223 
224 /*
225  * parse fcntl F_GETLK results
226  */
227 static int parse_reply_info_filelock(void **p, void *end,
228 				     struct ceph_mds_reply_info_parsed *info,
229 				     u64 features)
230 {
231 	if (*p + sizeof(*info->filelock_reply) > end)
232 		goto bad;
233 
234 	info->filelock_reply = *p;
235 	*p += sizeof(*info->filelock_reply);
236 
237 	if (unlikely(*p != end))
238 		goto bad;
239 	return 0;
240 
241 bad:
242 	return -EIO;
243 }
244 
245 /*
246  * parse create results
247  */
248 static int parse_reply_info_create(void **p, void *end,
249 				  struct ceph_mds_reply_info_parsed *info,
250 				  u64 features)
251 {
252 	if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
253 		if (*p == end) {
254 			info->has_create_ino = false;
255 		} else {
256 			info->has_create_ino = true;
257 			info->ino = ceph_decode_64(p);
258 		}
259 	}
260 
261 	if (unlikely(*p != end))
262 		goto bad;
263 	return 0;
264 
265 bad:
266 	return -EIO;
267 }
268 
269 /*
270  * parse extra results
271  */
272 static int parse_reply_info_extra(void **p, void *end,
273 				  struct ceph_mds_reply_info_parsed *info,
274 				  u64 features)
275 {
276 	if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
277 		return parse_reply_info_filelock(p, end, info, features);
278 	else if (info->head->op == CEPH_MDS_OP_READDIR ||
279 		 info->head->op == CEPH_MDS_OP_LSSNAP)
280 		return parse_reply_info_dir(p, end, info, features);
281 	else if (info->head->op == CEPH_MDS_OP_CREATE)
282 		return parse_reply_info_create(p, end, info, features);
283 	else
284 		return -EIO;
285 }
286 
287 /*
288  * parse entire mds reply
289  */
290 static int parse_reply_info(struct ceph_msg *msg,
291 			    struct ceph_mds_reply_info_parsed *info,
292 			    u64 features)
293 {
294 	void *p, *end;
295 	u32 len;
296 	int err;
297 
298 	info->head = msg->front.iov_base;
299 	p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
300 	end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
301 
302 	/* trace */
303 	ceph_decode_32_safe(&p, end, len, bad);
304 	if (len > 0) {
305 		ceph_decode_need(&p, end, len, bad);
306 		err = parse_reply_info_trace(&p, p+len, info, features);
307 		if (err < 0)
308 			goto out_bad;
309 	}
310 
311 	/* extra */
312 	ceph_decode_32_safe(&p, end, len, bad);
313 	if (len > 0) {
314 		ceph_decode_need(&p, end, len, bad);
315 		err = parse_reply_info_extra(&p, p+len, info, features);
316 		if (err < 0)
317 			goto out_bad;
318 	}
319 
320 	/* snap blob */
321 	ceph_decode_32_safe(&p, end, len, bad);
322 	info->snapblob_len = len;
323 	info->snapblob = p;
324 	p += len;
325 
326 	if (p != end)
327 		goto bad;
328 	return 0;
329 
330 bad:
331 	err = -EIO;
332 out_bad:
333 	pr_err("mds parse_reply err %d\n", err);
334 	return err;
335 }
336 
337 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
338 {
339 	if (!info->dir_in)
340 		return;
341 	free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
342 }
343 
344 
345 /*
346  * sessions
347  */
348 const char *ceph_session_state_name(int s)
349 {
350 	switch (s) {
351 	case CEPH_MDS_SESSION_NEW: return "new";
352 	case CEPH_MDS_SESSION_OPENING: return "opening";
353 	case CEPH_MDS_SESSION_OPEN: return "open";
354 	case CEPH_MDS_SESSION_HUNG: return "hung";
355 	case CEPH_MDS_SESSION_CLOSING: return "closing";
356 	case CEPH_MDS_SESSION_RESTARTING: return "restarting";
357 	case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
358 	default: return "???";
359 	}
360 }
361 
362 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
363 {
364 	if (atomic_inc_not_zero(&s->s_ref)) {
365 		dout("mdsc get_session %p %d -> %d\n", s,
366 		     atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
367 		return s;
368 	} else {
369 		dout("mdsc get_session %p 0 -- FAIL", s);
370 		return NULL;
371 	}
372 }
373 
374 void ceph_put_mds_session(struct ceph_mds_session *s)
375 {
376 	dout("mdsc put_session %p %d -> %d\n", s,
377 	     atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
378 	if (atomic_dec_and_test(&s->s_ref)) {
379 		if (s->s_auth.authorizer)
380 			ceph_auth_destroy_authorizer(
381 				s->s_mdsc->fsc->client->monc.auth,
382 				s->s_auth.authorizer);
383 		kfree(s);
384 	}
385 }
386 
387 /*
388  * called under mdsc->mutex
389  */
390 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
391 						   int mds)
392 {
393 	struct ceph_mds_session *session;
394 
395 	if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
396 		return NULL;
397 	session = mdsc->sessions[mds];
398 	dout("lookup_mds_session %p %d\n", session,
399 	     atomic_read(&session->s_ref));
400 	get_session(session);
401 	return session;
402 }
403 
404 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
405 {
406 	if (mds >= mdsc->max_sessions)
407 		return false;
408 	return mdsc->sessions[mds];
409 }
410 
411 static int __verify_registered_session(struct ceph_mds_client *mdsc,
412 				       struct ceph_mds_session *s)
413 {
414 	if (s->s_mds >= mdsc->max_sessions ||
415 	    mdsc->sessions[s->s_mds] != s)
416 		return -ENOENT;
417 	return 0;
418 }
419 
420 /*
421  * create+register a new session for given mds.
422  * called under mdsc->mutex.
423  */
424 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
425 						 int mds)
426 {
427 	struct ceph_mds_session *s;
428 
429 	if (mds >= mdsc->mdsmap->m_max_mds)
430 		return ERR_PTR(-EINVAL);
431 
432 	s = kzalloc(sizeof(*s), GFP_NOFS);
433 	if (!s)
434 		return ERR_PTR(-ENOMEM);
435 	s->s_mdsc = mdsc;
436 	s->s_mds = mds;
437 	s->s_state = CEPH_MDS_SESSION_NEW;
438 	s->s_ttl = 0;
439 	s->s_seq = 0;
440 	mutex_init(&s->s_mutex);
441 
442 	ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
443 
444 	spin_lock_init(&s->s_gen_ttl_lock);
445 	s->s_cap_gen = 0;
446 	s->s_cap_ttl = jiffies - 1;
447 
448 	spin_lock_init(&s->s_cap_lock);
449 	s->s_renew_requested = 0;
450 	s->s_renew_seq = 0;
451 	INIT_LIST_HEAD(&s->s_caps);
452 	s->s_nr_caps = 0;
453 	s->s_trim_caps = 0;
454 	atomic_set(&s->s_ref, 1);
455 	INIT_LIST_HEAD(&s->s_waiting);
456 	INIT_LIST_HEAD(&s->s_unsafe);
457 	s->s_num_cap_releases = 0;
458 	s->s_cap_reconnect = 0;
459 	s->s_cap_iterator = NULL;
460 	INIT_LIST_HEAD(&s->s_cap_releases);
461 	INIT_LIST_HEAD(&s->s_cap_releases_done);
462 	INIT_LIST_HEAD(&s->s_cap_flushing);
463 	INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
464 
465 	dout("register_session mds%d\n", mds);
466 	if (mds >= mdsc->max_sessions) {
467 		int newmax = 1 << get_count_order(mds+1);
468 		struct ceph_mds_session **sa;
469 
470 		dout("register_session realloc to %d\n", newmax);
471 		sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
472 		if (sa == NULL)
473 			goto fail_realloc;
474 		if (mdsc->sessions) {
475 			memcpy(sa, mdsc->sessions,
476 			       mdsc->max_sessions * sizeof(void *));
477 			kfree(mdsc->sessions);
478 		}
479 		mdsc->sessions = sa;
480 		mdsc->max_sessions = newmax;
481 	}
482 	mdsc->sessions[mds] = s;
483 	atomic_inc(&mdsc->num_sessions);
484 	atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
485 
486 	ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
487 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
488 
489 	return s;
490 
491 fail_realloc:
492 	kfree(s);
493 	return ERR_PTR(-ENOMEM);
494 }
495 
496 /*
497  * called under mdsc->mutex
498  */
499 static void __unregister_session(struct ceph_mds_client *mdsc,
500 			       struct ceph_mds_session *s)
501 {
502 	dout("__unregister_session mds%d %p\n", s->s_mds, s);
503 	BUG_ON(mdsc->sessions[s->s_mds] != s);
504 	mdsc->sessions[s->s_mds] = NULL;
505 	ceph_con_close(&s->s_con);
506 	ceph_put_mds_session(s);
507 	atomic_dec(&mdsc->num_sessions);
508 }
509 
510 /*
511  * drop session refs in request.
512  *
513  * should be last request ref, or hold mdsc->mutex
514  */
515 static void put_request_session(struct ceph_mds_request *req)
516 {
517 	if (req->r_session) {
518 		ceph_put_mds_session(req->r_session);
519 		req->r_session = NULL;
520 	}
521 }
522 
523 void ceph_mdsc_release_request(struct kref *kref)
524 {
525 	struct ceph_mds_request *req = container_of(kref,
526 						    struct ceph_mds_request,
527 						    r_kref);
528 	destroy_reply_info(&req->r_reply_info);
529 	if (req->r_request)
530 		ceph_msg_put(req->r_request);
531 	if (req->r_reply)
532 		ceph_msg_put(req->r_reply);
533 	if (req->r_inode) {
534 		ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
535 		iput(req->r_inode);
536 	}
537 	if (req->r_locked_dir)
538 		ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
539 	iput(req->r_target_inode);
540 	if (req->r_dentry)
541 		dput(req->r_dentry);
542 	if (req->r_old_dentry)
543 		dput(req->r_old_dentry);
544 	if (req->r_old_dentry_dir) {
545 		/*
546 		 * track (and drop pins for) r_old_dentry_dir
547 		 * separately, since r_old_dentry's d_parent may have
548 		 * changed between the dir mutex being dropped and
549 		 * this request being freed.
550 		 */
551 		ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
552 				  CEPH_CAP_PIN);
553 		iput(req->r_old_dentry_dir);
554 	}
555 	kfree(req->r_path1);
556 	kfree(req->r_path2);
557 	if (req->r_pagelist)
558 		ceph_pagelist_release(req->r_pagelist);
559 	put_request_session(req);
560 	ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
561 	kfree(req);
562 }
563 
564 /*
565  * lookup session, bump ref if found.
566  *
567  * called under mdsc->mutex.
568  */
569 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
570 					     u64 tid)
571 {
572 	struct ceph_mds_request *req;
573 	struct rb_node *n = mdsc->request_tree.rb_node;
574 
575 	while (n) {
576 		req = rb_entry(n, struct ceph_mds_request, r_node);
577 		if (tid < req->r_tid)
578 			n = n->rb_left;
579 		else if (tid > req->r_tid)
580 			n = n->rb_right;
581 		else {
582 			ceph_mdsc_get_request(req);
583 			return req;
584 		}
585 	}
586 	return NULL;
587 }
588 
589 static void __insert_request(struct ceph_mds_client *mdsc,
590 			     struct ceph_mds_request *new)
591 {
592 	struct rb_node **p = &mdsc->request_tree.rb_node;
593 	struct rb_node *parent = NULL;
594 	struct ceph_mds_request *req = NULL;
595 
596 	while (*p) {
597 		parent = *p;
598 		req = rb_entry(parent, struct ceph_mds_request, r_node);
599 		if (new->r_tid < req->r_tid)
600 			p = &(*p)->rb_left;
601 		else if (new->r_tid > req->r_tid)
602 			p = &(*p)->rb_right;
603 		else
604 			BUG();
605 	}
606 
607 	rb_link_node(&new->r_node, parent, p);
608 	rb_insert_color(&new->r_node, &mdsc->request_tree);
609 }
610 
611 /*
612  * Register an in-flight request, and assign a tid.  Link to directory
613  * are modifying (if any).
614  *
615  * Called under mdsc->mutex.
616  */
617 static void __register_request(struct ceph_mds_client *mdsc,
618 			       struct ceph_mds_request *req,
619 			       struct inode *dir)
620 {
621 	req->r_tid = ++mdsc->last_tid;
622 	if (req->r_num_caps)
623 		ceph_reserve_caps(mdsc, &req->r_caps_reservation,
624 				  req->r_num_caps);
625 	dout("__register_request %p tid %lld\n", req, req->r_tid);
626 	ceph_mdsc_get_request(req);
627 	__insert_request(mdsc, req);
628 
629 	req->r_uid = current_fsuid();
630 	req->r_gid = current_fsgid();
631 
632 	if (dir) {
633 		struct ceph_inode_info *ci = ceph_inode(dir);
634 
635 		ihold(dir);
636 		spin_lock(&ci->i_unsafe_lock);
637 		req->r_unsafe_dir = dir;
638 		list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
639 		spin_unlock(&ci->i_unsafe_lock);
640 	}
641 }
642 
643 static void __unregister_request(struct ceph_mds_client *mdsc,
644 				 struct ceph_mds_request *req)
645 {
646 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
647 	rb_erase(&req->r_node, &mdsc->request_tree);
648 	RB_CLEAR_NODE(&req->r_node);
649 
650 	if (req->r_unsafe_dir) {
651 		struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
652 
653 		spin_lock(&ci->i_unsafe_lock);
654 		list_del_init(&req->r_unsafe_dir_item);
655 		spin_unlock(&ci->i_unsafe_lock);
656 
657 		iput(req->r_unsafe_dir);
658 		req->r_unsafe_dir = NULL;
659 	}
660 
661 	complete_all(&req->r_safe_completion);
662 
663 	ceph_mdsc_put_request(req);
664 }
665 
666 /*
667  * Choose mds to send request to next.  If there is a hint set in the
668  * request (e.g., due to a prior forward hint from the mds), use that.
669  * Otherwise, consult frag tree and/or caps to identify the
670  * appropriate mds.  If all else fails, choose randomly.
671  *
672  * Called under mdsc->mutex.
673  */
674 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
675 {
676 	/*
677 	 * we don't need to worry about protecting the d_parent access
678 	 * here because we never renaming inside the snapped namespace
679 	 * except to resplice to another snapdir, and either the old or new
680 	 * result is a valid result.
681 	 */
682 	while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
683 		dentry = dentry->d_parent;
684 	return dentry;
685 }
686 
687 static int __choose_mds(struct ceph_mds_client *mdsc,
688 			struct ceph_mds_request *req)
689 {
690 	struct inode *inode;
691 	struct ceph_inode_info *ci;
692 	struct ceph_cap *cap;
693 	int mode = req->r_direct_mode;
694 	int mds = -1;
695 	u32 hash = req->r_direct_hash;
696 	bool is_hash = req->r_direct_is_hash;
697 
698 	/*
699 	 * is there a specific mds we should try?  ignore hint if we have
700 	 * no session and the mds is not up (active or recovering).
701 	 */
702 	if (req->r_resend_mds >= 0 &&
703 	    (__have_session(mdsc, req->r_resend_mds) ||
704 	     ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
705 		dout("choose_mds using resend_mds mds%d\n",
706 		     req->r_resend_mds);
707 		return req->r_resend_mds;
708 	}
709 
710 	if (mode == USE_RANDOM_MDS)
711 		goto random;
712 
713 	inode = NULL;
714 	if (req->r_inode) {
715 		inode = req->r_inode;
716 	} else if (req->r_dentry) {
717 		/* ignore race with rename; old or new d_parent is okay */
718 		struct dentry *parent = req->r_dentry->d_parent;
719 		struct inode *dir = parent->d_inode;
720 
721 		if (dir->i_sb != mdsc->fsc->sb) {
722 			/* not this fs! */
723 			inode = req->r_dentry->d_inode;
724 		} else if (ceph_snap(dir) != CEPH_NOSNAP) {
725 			/* direct snapped/virtual snapdir requests
726 			 * based on parent dir inode */
727 			struct dentry *dn = get_nonsnap_parent(parent);
728 			inode = dn->d_inode;
729 			dout("__choose_mds using nonsnap parent %p\n", inode);
730 		} else {
731 			/* dentry target */
732 			inode = req->r_dentry->d_inode;
733 			if (!inode || mode == USE_AUTH_MDS) {
734 				/* dir + name */
735 				inode = dir;
736 				hash = ceph_dentry_hash(dir, req->r_dentry);
737 				is_hash = true;
738 			}
739 		}
740 	}
741 
742 	dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
743 	     (int)hash, mode);
744 	if (!inode)
745 		goto random;
746 	ci = ceph_inode(inode);
747 
748 	if (is_hash && S_ISDIR(inode->i_mode)) {
749 		struct ceph_inode_frag frag;
750 		int found;
751 
752 		ceph_choose_frag(ci, hash, &frag, &found);
753 		if (found) {
754 			if (mode == USE_ANY_MDS && frag.ndist > 0) {
755 				u8 r;
756 
757 				/* choose a random replica */
758 				get_random_bytes(&r, 1);
759 				r %= frag.ndist;
760 				mds = frag.dist[r];
761 				dout("choose_mds %p %llx.%llx "
762 				     "frag %u mds%d (%d/%d)\n",
763 				     inode, ceph_vinop(inode),
764 				     frag.frag, mds,
765 				     (int)r, frag.ndist);
766 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
767 				    CEPH_MDS_STATE_ACTIVE)
768 					return mds;
769 			}
770 
771 			/* since this file/dir wasn't known to be
772 			 * replicated, then we want to look for the
773 			 * authoritative mds. */
774 			mode = USE_AUTH_MDS;
775 			if (frag.mds >= 0) {
776 				/* choose auth mds */
777 				mds = frag.mds;
778 				dout("choose_mds %p %llx.%llx "
779 				     "frag %u mds%d (auth)\n",
780 				     inode, ceph_vinop(inode), frag.frag, mds);
781 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
782 				    CEPH_MDS_STATE_ACTIVE)
783 					return mds;
784 			}
785 		}
786 	}
787 
788 	spin_lock(&ci->i_ceph_lock);
789 	cap = NULL;
790 	if (mode == USE_AUTH_MDS)
791 		cap = ci->i_auth_cap;
792 	if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
793 		cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
794 	if (!cap) {
795 		spin_unlock(&ci->i_ceph_lock);
796 		goto random;
797 	}
798 	mds = cap->session->s_mds;
799 	dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
800 	     inode, ceph_vinop(inode), mds,
801 	     cap == ci->i_auth_cap ? "auth " : "", cap);
802 	spin_unlock(&ci->i_ceph_lock);
803 	return mds;
804 
805 random:
806 	mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
807 	dout("choose_mds chose random mds%d\n", mds);
808 	return mds;
809 }
810 
811 
812 /*
813  * session messages
814  */
815 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
816 {
817 	struct ceph_msg *msg;
818 	struct ceph_mds_session_head *h;
819 
820 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
821 			   false);
822 	if (!msg) {
823 		pr_err("create_session_msg ENOMEM creating msg\n");
824 		return NULL;
825 	}
826 	h = msg->front.iov_base;
827 	h->op = cpu_to_le32(op);
828 	h->seq = cpu_to_le64(seq);
829 
830 	return msg;
831 }
832 
833 /*
834  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
835  * to include additional client metadata fields.
836  */
837 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
838 {
839 	struct ceph_msg *msg;
840 	struct ceph_mds_session_head *h;
841 	int i = -1;
842 	int metadata_bytes = 0;
843 	int metadata_key_count = 0;
844 	struct ceph_options *opt = mdsc->fsc->client->options;
845 	void *p;
846 
847 	const char* metadata[][2] = {
848 		{"hostname", utsname()->nodename},
849 		{"kernel_version", utsname()->release},
850 		{"entity_id", opt->name ? opt->name : ""},
851 		{NULL, NULL}
852 	};
853 
854 	/* Calculate serialized length of metadata */
855 	metadata_bytes = 4;  /* map length */
856 	for (i = 0; metadata[i][0] != NULL; ++i) {
857 		metadata_bytes += 8 + strlen(metadata[i][0]) +
858 			strlen(metadata[i][1]);
859 		metadata_key_count++;
860 	}
861 
862 	/* Allocate the message */
863 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
864 			   GFP_NOFS, false);
865 	if (!msg) {
866 		pr_err("create_session_msg ENOMEM creating msg\n");
867 		return NULL;
868 	}
869 	h = msg->front.iov_base;
870 	h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
871 	h->seq = cpu_to_le64(seq);
872 
873 	/*
874 	 * Serialize client metadata into waiting buffer space, using
875 	 * the format that userspace expects for map<string, string>
876 	 *
877 	 * ClientSession messages with metadata are v2
878 	 */
879 	msg->hdr.version = cpu_to_le16(2);
880 	msg->hdr.compat_version = cpu_to_le16(1);
881 
882 	/* The write pointer, following the session_head structure */
883 	p = msg->front.iov_base + sizeof(*h);
884 
885 	/* Number of entries in the map */
886 	ceph_encode_32(&p, metadata_key_count);
887 
888 	/* Two length-prefixed strings for each entry in the map */
889 	for (i = 0; metadata[i][0] != NULL; ++i) {
890 		size_t const key_len = strlen(metadata[i][0]);
891 		size_t const val_len = strlen(metadata[i][1]);
892 
893 		ceph_encode_32(&p, key_len);
894 		memcpy(p, metadata[i][0], key_len);
895 		p += key_len;
896 		ceph_encode_32(&p, val_len);
897 		memcpy(p, metadata[i][1], val_len);
898 		p += val_len;
899 	}
900 
901 	return msg;
902 }
903 
904 /*
905  * send session open request.
906  *
907  * called under mdsc->mutex
908  */
909 static int __open_session(struct ceph_mds_client *mdsc,
910 			  struct ceph_mds_session *session)
911 {
912 	struct ceph_msg *msg;
913 	int mstate;
914 	int mds = session->s_mds;
915 
916 	/* wait for mds to go active? */
917 	mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
918 	dout("open_session to mds%d (%s)\n", mds,
919 	     ceph_mds_state_name(mstate));
920 	session->s_state = CEPH_MDS_SESSION_OPENING;
921 	session->s_renew_requested = jiffies;
922 
923 	/* send connect message */
924 	msg = create_session_open_msg(mdsc, session->s_seq);
925 	if (!msg)
926 		return -ENOMEM;
927 	ceph_con_send(&session->s_con, msg);
928 	return 0;
929 }
930 
931 /*
932  * open sessions for any export targets for the given mds
933  *
934  * called under mdsc->mutex
935  */
936 static struct ceph_mds_session *
937 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
938 {
939 	struct ceph_mds_session *session;
940 
941 	session = __ceph_lookup_mds_session(mdsc, target);
942 	if (!session) {
943 		session = register_session(mdsc, target);
944 		if (IS_ERR(session))
945 			return session;
946 	}
947 	if (session->s_state == CEPH_MDS_SESSION_NEW ||
948 	    session->s_state == CEPH_MDS_SESSION_CLOSING)
949 		__open_session(mdsc, session);
950 
951 	return session;
952 }
953 
954 struct ceph_mds_session *
955 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
956 {
957 	struct ceph_mds_session *session;
958 
959 	dout("open_export_target_session to mds%d\n", target);
960 
961 	mutex_lock(&mdsc->mutex);
962 	session = __open_export_target_session(mdsc, target);
963 	mutex_unlock(&mdsc->mutex);
964 
965 	return session;
966 }
967 
968 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
969 					  struct ceph_mds_session *session)
970 {
971 	struct ceph_mds_info *mi;
972 	struct ceph_mds_session *ts;
973 	int i, mds = session->s_mds;
974 
975 	if (mds >= mdsc->mdsmap->m_max_mds)
976 		return;
977 
978 	mi = &mdsc->mdsmap->m_info[mds];
979 	dout("open_export_target_sessions for mds%d (%d targets)\n",
980 	     session->s_mds, mi->num_export_targets);
981 
982 	for (i = 0; i < mi->num_export_targets; i++) {
983 		ts = __open_export_target_session(mdsc, mi->export_targets[i]);
984 		if (!IS_ERR(ts))
985 			ceph_put_mds_session(ts);
986 	}
987 }
988 
989 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
990 					   struct ceph_mds_session *session)
991 {
992 	mutex_lock(&mdsc->mutex);
993 	__open_export_target_sessions(mdsc, session);
994 	mutex_unlock(&mdsc->mutex);
995 }
996 
997 /*
998  * session caps
999  */
1000 
1001 /*
1002  * Free preallocated cap messages assigned to this session
1003  */
1004 static void cleanup_cap_releases(struct ceph_mds_session *session)
1005 {
1006 	struct ceph_msg *msg;
1007 
1008 	spin_lock(&session->s_cap_lock);
1009 	while (!list_empty(&session->s_cap_releases)) {
1010 		msg = list_first_entry(&session->s_cap_releases,
1011 				       struct ceph_msg, list_head);
1012 		list_del_init(&msg->list_head);
1013 		ceph_msg_put(msg);
1014 	}
1015 	while (!list_empty(&session->s_cap_releases_done)) {
1016 		msg = list_first_entry(&session->s_cap_releases_done,
1017 				       struct ceph_msg, list_head);
1018 		list_del_init(&msg->list_head);
1019 		ceph_msg_put(msg);
1020 	}
1021 	spin_unlock(&session->s_cap_lock);
1022 }
1023 
1024 /*
1025  * Helper to safely iterate over all caps associated with a session, with
1026  * special care taken to handle a racing __ceph_remove_cap().
1027  *
1028  * Caller must hold session s_mutex.
1029  */
1030 static int iterate_session_caps(struct ceph_mds_session *session,
1031 				 int (*cb)(struct inode *, struct ceph_cap *,
1032 					    void *), void *arg)
1033 {
1034 	struct list_head *p;
1035 	struct ceph_cap *cap;
1036 	struct inode *inode, *last_inode = NULL;
1037 	struct ceph_cap *old_cap = NULL;
1038 	int ret;
1039 
1040 	dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1041 	spin_lock(&session->s_cap_lock);
1042 	p = session->s_caps.next;
1043 	while (p != &session->s_caps) {
1044 		cap = list_entry(p, struct ceph_cap, session_caps);
1045 		inode = igrab(&cap->ci->vfs_inode);
1046 		if (!inode) {
1047 			p = p->next;
1048 			continue;
1049 		}
1050 		session->s_cap_iterator = cap;
1051 		spin_unlock(&session->s_cap_lock);
1052 
1053 		if (last_inode) {
1054 			iput(last_inode);
1055 			last_inode = NULL;
1056 		}
1057 		if (old_cap) {
1058 			ceph_put_cap(session->s_mdsc, old_cap);
1059 			old_cap = NULL;
1060 		}
1061 
1062 		ret = cb(inode, cap, arg);
1063 		last_inode = inode;
1064 
1065 		spin_lock(&session->s_cap_lock);
1066 		p = p->next;
1067 		if (cap->ci == NULL) {
1068 			dout("iterate_session_caps  finishing cap %p removal\n",
1069 			     cap);
1070 			BUG_ON(cap->session != session);
1071 			list_del_init(&cap->session_caps);
1072 			session->s_nr_caps--;
1073 			cap->session = NULL;
1074 			old_cap = cap;  /* put_cap it w/o locks held */
1075 		}
1076 		if (ret < 0)
1077 			goto out;
1078 	}
1079 	ret = 0;
1080 out:
1081 	session->s_cap_iterator = NULL;
1082 	spin_unlock(&session->s_cap_lock);
1083 
1084 	iput(last_inode);
1085 	if (old_cap)
1086 		ceph_put_cap(session->s_mdsc, old_cap);
1087 
1088 	return ret;
1089 }
1090 
1091 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1092 				  void *arg)
1093 {
1094 	struct ceph_inode_info *ci = ceph_inode(inode);
1095 	int drop = 0;
1096 
1097 	dout("removing cap %p, ci is %p, inode is %p\n",
1098 	     cap, ci, &ci->vfs_inode);
1099 	spin_lock(&ci->i_ceph_lock);
1100 	__ceph_remove_cap(cap, false);
1101 	if (!__ceph_is_any_real_caps(ci)) {
1102 		struct ceph_mds_client *mdsc =
1103 			ceph_sb_to_client(inode->i_sb)->mdsc;
1104 
1105 		spin_lock(&mdsc->cap_dirty_lock);
1106 		if (!list_empty(&ci->i_dirty_item)) {
1107 			pr_info(" dropping dirty %s state for %p %lld\n",
1108 				ceph_cap_string(ci->i_dirty_caps),
1109 				inode, ceph_ino(inode));
1110 			ci->i_dirty_caps = 0;
1111 			list_del_init(&ci->i_dirty_item);
1112 			drop = 1;
1113 		}
1114 		if (!list_empty(&ci->i_flushing_item)) {
1115 			pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1116 				ceph_cap_string(ci->i_flushing_caps),
1117 				inode, ceph_ino(inode));
1118 			ci->i_flushing_caps = 0;
1119 			list_del_init(&ci->i_flushing_item);
1120 			mdsc->num_cap_flushing--;
1121 			drop = 1;
1122 		}
1123 		if (drop && ci->i_wrbuffer_ref) {
1124 			pr_info(" dropping dirty data for %p %lld\n",
1125 				inode, ceph_ino(inode));
1126 			ci->i_wrbuffer_ref = 0;
1127 			ci->i_wrbuffer_ref_head = 0;
1128 			drop++;
1129 		}
1130 		spin_unlock(&mdsc->cap_dirty_lock);
1131 	}
1132 	spin_unlock(&ci->i_ceph_lock);
1133 	while (drop--)
1134 		iput(inode);
1135 	return 0;
1136 }
1137 
1138 /*
1139  * caller must hold session s_mutex
1140  */
1141 static void remove_session_caps(struct ceph_mds_session *session)
1142 {
1143 	dout("remove_session_caps on %p\n", session);
1144 	iterate_session_caps(session, remove_session_caps_cb, NULL);
1145 
1146 	spin_lock(&session->s_cap_lock);
1147 	if (session->s_nr_caps > 0) {
1148 		struct super_block *sb = session->s_mdsc->fsc->sb;
1149 		struct inode *inode;
1150 		struct ceph_cap *cap, *prev = NULL;
1151 		struct ceph_vino vino;
1152 		/*
1153 		 * iterate_session_caps() skips inodes that are being
1154 		 * deleted, we need to wait until deletions are complete.
1155 		 * __wait_on_freeing_inode() is designed for the job,
1156 		 * but it is not exported, so use lookup inode function
1157 		 * to access it.
1158 		 */
1159 		while (!list_empty(&session->s_caps)) {
1160 			cap = list_entry(session->s_caps.next,
1161 					 struct ceph_cap, session_caps);
1162 			if (cap == prev)
1163 				break;
1164 			prev = cap;
1165 			vino = cap->ci->i_vino;
1166 			spin_unlock(&session->s_cap_lock);
1167 
1168 			inode = ceph_find_inode(sb, vino);
1169 			iput(inode);
1170 
1171 			spin_lock(&session->s_cap_lock);
1172 		}
1173 	}
1174 	spin_unlock(&session->s_cap_lock);
1175 
1176 	BUG_ON(session->s_nr_caps > 0);
1177 	BUG_ON(!list_empty(&session->s_cap_flushing));
1178 	cleanup_cap_releases(session);
1179 }
1180 
1181 /*
1182  * wake up any threads waiting on this session's caps.  if the cap is
1183  * old (didn't get renewed on the client reconnect), remove it now.
1184  *
1185  * caller must hold s_mutex.
1186  */
1187 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1188 			      void *arg)
1189 {
1190 	struct ceph_inode_info *ci = ceph_inode(inode);
1191 
1192 	wake_up_all(&ci->i_cap_wq);
1193 	if (arg) {
1194 		spin_lock(&ci->i_ceph_lock);
1195 		ci->i_wanted_max_size = 0;
1196 		ci->i_requested_max_size = 0;
1197 		spin_unlock(&ci->i_ceph_lock);
1198 	}
1199 	return 0;
1200 }
1201 
1202 static void wake_up_session_caps(struct ceph_mds_session *session,
1203 				 int reconnect)
1204 {
1205 	dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1206 	iterate_session_caps(session, wake_up_session_cb,
1207 			     (void *)(unsigned long)reconnect);
1208 }
1209 
1210 /*
1211  * Send periodic message to MDS renewing all currently held caps.  The
1212  * ack will reset the expiration for all caps from this session.
1213  *
1214  * caller holds s_mutex
1215  */
1216 static int send_renew_caps(struct ceph_mds_client *mdsc,
1217 			   struct ceph_mds_session *session)
1218 {
1219 	struct ceph_msg *msg;
1220 	int state;
1221 
1222 	if (time_after_eq(jiffies, session->s_cap_ttl) &&
1223 	    time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1224 		pr_info("mds%d caps stale\n", session->s_mds);
1225 	session->s_renew_requested = jiffies;
1226 
1227 	/* do not try to renew caps until a recovering mds has reconnected
1228 	 * with its clients. */
1229 	state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1230 	if (state < CEPH_MDS_STATE_RECONNECT) {
1231 		dout("send_renew_caps ignoring mds%d (%s)\n",
1232 		     session->s_mds, ceph_mds_state_name(state));
1233 		return 0;
1234 	}
1235 
1236 	dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1237 		ceph_mds_state_name(state));
1238 	msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1239 				 ++session->s_renew_seq);
1240 	if (!msg)
1241 		return -ENOMEM;
1242 	ceph_con_send(&session->s_con, msg);
1243 	return 0;
1244 }
1245 
1246 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1247 			     struct ceph_mds_session *session, u64 seq)
1248 {
1249 	struct ceph_msg *msg;
1250 
1251 	dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1252 	     session->s_mds, ceph_session_state_name(session->s_state), seq);
1253 	msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1254 	if (!msg)
1255 		return -ENOMEM;
1256 	ceph_con_send(&session->s_con, msg);
1257 	return 0;
1258 }
1259 
1260 
1261 /*
1262  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1263  *
1264  * Called under session->s_mutex
1265  */
1266 static void renewed_caps(struct ceph_mds_client *mdsc,
1267 			 struct ceph_mds_session *session, int is_renew)
1268 {
1269 	int was_stale;
1270 	int wake = 0;
1271 
1272 	spin_lock(&session->s_cap_lock);
1273 	was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1274 
1275 	session->s_cap_ttl = session->s_renew_requested +
1276 		mdsc->mdsmap->m_session_timeout*HZ;
1277 
1278 	if (was_stale) {
1279 		if (time_before(jiffies, session->s_cap_ttl)) {
1280 			pr_info("mds%d caps renewed\n", session->s_mds);
1281 			wake = 1;
1282 		} else {
1283 			pr_info("mds%d caps still stale\n", session->s_mds);
1284 		}
1285 	}
1286 	dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1287 	     session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1288 	     time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1289 	spin_unlock(&session->s_cap_lock);
1290 
1291 	if (wake)
1292 		wake_up_session_caps(session, 0);
1293 }
1294 
1295 /*
1296  * send a session close request
1297  */
1298 static int request_close_session(struct ceph_mds_client *mdsc,
1299 				 struct ceph_mds_session *session)
1300 {
1301 	struct ceph_msg *msg;
1302 
1303 	dout("request_close_session mds%d state %s seq %lld\n",
1304 	     session->s_mds, ceph_session_state_name(session->s_state),
1305 	     session->s_seq);
1306 	msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1307 	if (!msg)
1308 		return -ENOMEM;
1309 	ceph_con_send(&session->s_con, msg);
1310 	return 0;
1311 }
1312 
1313 /*
1314  * Called with s_mutex held.
1315  */
1316 static int __close_session(struct ceph_mds_client *mdsc,
1317 			 struct ceph_mds_session *session)
1318 {
1319 	if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1320 		return 0;
1321 	session->s_state = CEPH_MDS_SESSION_CLOSING;
1322 	return request_close_session(mdsc, session);
1323 }
1324 
1325 /*
1326  * Trim old(er) caps.
1327  *
1328  * Because we can't cache an inode without one or more caps, we do
1329  * this indirectly: if a cap is unused, we prune its aliases, at which
1330  * point the inode will hopefully get dropped to.
1331  *
1332  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1333  * memory pressure from the MDS, though, so it needn't be perfect.
1334  */
1335 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1336 {
1337 	struct ceph_mds_session *session = arg;
1338 	struct ceph_inode_info *ci = ceph_inode(inode);
1339 	int used, wanted, oissued, mine;
1340 
1341 	if (session->s_trim_caps <= 0)
1342 		return -1;
1343 
1344 	spin_lock(&ci->i_ceph_lock);
1345 	mine = cap->issued | cap->implemented;
1346 	used = __ceph_caps_used(ci);
1347 	wanted = __ceph_caps_file_wanted(ci);
1348 	oissued = __ceph_caps_issued_other(ci, cap);
1349 
1350 	dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1351 	     inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1352 	     ceph_cap_string(used), ceph_cap_string(wanted));
1353 	if (cap == ci->i_auth_cap) {
1354 		if (ci->i_dirty_caps | ci->i_flushing_caps)
1355 			goto out;
1356 		if ((used | wanted) & CEPH_CAP_ANY_WR)
1357 			goto out;
1358 	}
1359 	if ((used | wanted) & ~oissued & mine)
1360 		goto out;   /* we need these caps */
1361 
1362 	session->s_trim_caps--;
1363 	if (oissued) {
1364 		/* we aren't the only cap.. just remove us */
1365 		__ceph_remove_cap(cap, true);
1366 	} else {
1367 		/* try to drop referring dentries */
1368 		spin_unlock(&ci->i_ceph_lock);
1369 		d_prune_aliases(inode);
1370 		dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1371 		     inode, cap, atomic_read(&inode->i_count));
1372 		return 0;
1373 	}
1374 
1375 out:
1376 	spin_unlock(&ci->i_ceph_lock);
1377 	return 0;
1378 }
1379 
1380 /*
1381  * Trim session cap count down to some max number.
1382  */
1383 static int trim_caps(struct ceph_mds_client *mdsc,
1384 		     struct ceph_mds_session *session,
1385 		     int max_caps)
1386 {
1387 	int trim_caps = session->s_nr_caps - max_caps;
1388 
1389 	dout("trim_caps mds%d start: %d / %d, trim %d\n",
1390 	     session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1391 	if (trim_caps > 0) {
1392 		session->s_trim_caps = trim_caps;
1393 		iterate_session_caps(session, trim_caps_cb, session);
1394 		dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1395 		     session->s_mds, session->s_nr_caps, max_caps,
1396 			trim_caps - session->s_trim_caps);
1397 		session->s_trim_caps = 0;
1398 	}
1399 
1400 	ceph_add_cap_releases(mdsc, session);
1401 	ceph_send_cap_releases(mdsc, session);
1402 	return 0;
1403 }
1404 
1405 /*
1406  * Allocate cap_release messages.  If there is a partially full message
1407  * in the queue, try to allocate enough to cover it's remainder, so that
1408  * we can send it immediately.
1409  *
1410  * Called under s_mutex.
1411  */
1412 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1413 			  struct ceph_mds_session *session)
1414 {
1415 	struct ceph_msg *msg, *partial = NULL;
1416 	struct ceph_mds_cap_release *head;
1417 	int err = -ENOMEM;
1418 	int extra = mdsc->fsc->mount_options->cap_release_safety;
1419 	int num;
1420 
1421 	dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1422 	     extra);
1423 
1424 	spin_lock(&session->s_cap_lock);
1425 
1426 	if (!list_empty(&session->s_cap_releases)) {
1427 		msg = list_first_entry(&session->s_cap_releases,
1428 				       struct ceph_msg,
1429 				 list_head);
1430 		head = msg->front.iov_base;
1431 		num = le32_to_cpu(head->num);
1432 		if (num) {
1433 			dout(" partial %p with (%d/%d)\n", msg, num,
1434 			     (int)CEPH_CAPS_PER_RELEASE);
1435 			extra += CEPH_CAPS_PER_RELEASE - num;
1436 			partial = msg;
1437 		}
1438 	}
1439 	while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1440 		spin_unlock(&session->s_cap_lock);
1441 		msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1442 				   GFP_NOFS, false);
1443 		if (!msg)
1444 			goto out_unlocked;
1445 		dout("add_cap_releases %p msg %p now %d\n", session, msg,
1446 		     (int)msg->front.iov_len);
1447 		head = msg->front.iov_base;
1448 		head->num = cpu_to_le32(0);
1449 		msg->front.iov_len = sizeof(*head);
1450 		spin_lock(&session->s_cap_lock);
1451 		list_add(&msg->list_head, &session->s_cap_releases);
1452 		session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1453 	}
1454 
1455 	if (partial) {
1456 		head = partial->front.iov_base;
1457 		num = le32_to_cpu(head->num);
1458 		dout(" queueing partial %p with %d/%d\n", partial, num,
1459 		     (int)CEPH_CAPS_PER_RELEASE);
1460 		list_move_tail(&partial->list_head,
1461 			       &session->s_cap_releases_done);
1462 		session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1463 	}
1464 	err = 0;
1465 	spin_unlock(&session->s_cap_lock);
1466 out_unlocked:
1467 	return err;
1468 }
1469 
1470 static int check_cap_flush(struct inode *inode, u64 want_flush_seq)
1471 {
1472 	struct ceph_inode_info *ci = ceph_inode(inode);
1473 	int ret;
1474 	spin_lock(&ci->i_ceph_lock);
1475 	if (ci->i_flushing_caps)
1476 		ret = ci->i_cap_flush_seq >= want_flush_seq;
1477 	else
1478 		ret = 1;
1479 	spin_unlock(&ci->i_ceph_lock);
1480 	return ret;
1481 }
1482 
1483 /*
1484  * flush all dirty inode data to disk.
1485  *
1486  * returns true if we've flushed through want_flush_seq
1487  */
1488 static void wait_caps_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1489 {
1490 	int mds;
1491 
1492 	dout("check_cap_flush want %lld\n", want_flush_seq);
1493 	mutex_lock(&mdsc->mutex);
1494 	for (mds = 0; mds < mdsc->max_sessions; mds++) {
1495 		struct ceph_mds_session *session = mdsc->sessions[mds];
1496 		struct inode *inode = NULL;
1497 
1498 		if (!session)
1499 			continue;
1500 		get_session(session);
1501 		mutex_unlock(&mdsc->mutex);
1502 
1503 		mutex_lock(&session->s_mutex);
1504 		if (!list_empty(&session->s_cap_flushing)) {
1505 			struct ceph_inode_info *ci =
1506 				list_entry(session->s_cap_flushing.next,
1507 					   struct ceph_inode_info,
1508 					   i_flushing_item);
1509 
1510 			if (!check_cap_flush(&ci->vfs_inode, want_flush_seq)) {
1511 				dout("check_cap_flush still flushing %p "
1512 				     "seq %lld <= %lld to mds%d\n",
1513 				     &ci->vfs_inode, ci->i_cap_flush_seq,
1514 				     want_flush_seq, session->s_mds);
1515 				inode = igrab(&ci->vfs_inode);
1516 			}
1517 		}
1518 		mutex_unlock(&session->s_mutex);
1519 		ceph_put_mds_session(session);
1520 
1521 		if (inode) {
1522 			wait_event(mdsc->cap_flushing_wq,
1523 				   check_cap_flush(inode, want_flush_seq));
1524 			iput(inode);
1525 		}
1526 
1527 		mutex_lock(&mdsc->mutex);
1528 	}
1529 
1530 	mutex_unlock(&mdsc->mutex);
1531 	dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1532 }
1533 
1534 /*
1535  * called under s_mutex
1536  */
1537 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1538 			    struct ceph_mds_session *session)
1539 {
1540 	struct ceph_msg *msg;
1541 
1542 	dout("send_cap_releases mds%d\n", session->s_mds);
1543 	spin_lock(&session->s_cap_lock);
1544 	while (!list_empty(&session->s_cap_releases_done)) {
1545 		msg = list_first_entry(&session->s_cap_releases_done,
1546 				 struct ceph_msg, list_head);
1547 		list_del_init(&msg->list_head);
1548 		spin_unlock(&session->s_cap_lock);
1549 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1550 		dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1551 		ceph_con_send(&session->s_con, msg);
1552 		spin_lock(&session->s_cap_lock);
1553 	}
1554 	spin_unlock(&session->s_cap_lock);
1555 }
1556 
1557 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1558 				 struct ceph_mds_session *session)
1559 {
1560 	struct ceph_msg *msg;
1561 	struct ceph_mds_cap_release *head;
1562 	unsigned num;
1563 
1564 	dout("discard_cap_releases mds%d\n", session->s_mds);
1565 
1566 	if (!list_empty(&session->s_cap_releases)) {
1567 		/* zero out the in-progress message */
1568 		msg = list_first_entry(&session->s_cap_releases,
1569 					struct ceph_msg, list_head);
1570 		head = msg->front.iov_base;
1571 		num = le32_to_cpu(head->num);
1572 		dout("discard_cap_releases mds%d %p %u\n",
1573 		     session->s_mds, msg, num);
1574 		head->num = cpu_to_le32(0);
1575 		msg->front.iov_len = sizeof(*head);
1576 		session->s_num_cap_releases += num;
1577 	}
1578 
1579 	/* requeue completed messages */
1580 	while (!list_empty(&session->s_cap_releases_done)) {
1581 		msg = list_first_entry(&session->s_cap_releases_done,
1582 				 struct ceph_msg, list_head);
1583 		list_del_init(&msg->list_head);
1584 
1585 		head = msg->front.iov_base;
1586 		num = le32_to_cpu(head->num);
1587 		dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1588 		     num);
1589 		session->s_num_cap_releases += num;
1590 		head->num = cpu_to_le32(0);
1591 		msg->front.iov_len = sizeof(*head);
1592 		list_add(&msg->list_head, &session->s_cap_releases);
1593 	}
1594 }
1595 
1596 /*
1597  * requests
1598  */
1599 
1600 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1601 				    struct inode *dir)
1602 {
1603 	struct ceph_inode_info *ci = ceph_inode(dir);
1604 	struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1605 	struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1606 	size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1607 		      sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1608 	int order, num_entries;
1609 
1610 	spin_lock(&ci->i_ceph_lock);
1611 	num_entries = ci->i_files + ci->i_subdirs;
1612 	spin_unlock(&ci->i_ceph_lock);
1613 	num_entries = max(num_entries, 1);
1614 	num_entries = min(num_entries, opt->max_readdir);
1615 
1616 	order = get_order(size * num_entries);
1617 	while (order >= 0) {
1618 		rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN,
1619 							order);
1620 		if (rinfo->dir_in)
1621 			break;
1622 		order--;
1623 	}
1624 	if (!rinfo->dir_in)
1625 		return -ENOMEM;
1626 
1627 	num_entries = (PAGE_SIZE << order) / size;
1628 	num_entries = min(num_entries, opt->max_readdir);
1629 
1630 	rinfo->dir_buf_size = PAGE_SIZE << order;
1631 	req->r_num_caps = num_entries + 1;
1632 	req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1633 	req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1634 	return 0;
1635 }
1636 
1637 /*
1638  * Create an mds request.
1639  */
1640 struct ceph_mds_request *
1641 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1642 {
1643 	struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1644 
1645 	if (!req)
1646 		return ERR_PTR(-ENOMEM);
1647 
1648 	mutex_init(&req->r_fill_mutex);
1649 	req->r_mdsc = mdsc;
1650 	req->r_started = jiffies;
1651 	req->r_resend_mds = -1;
1652 	INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1653 	req->r_fmode = -1;
1654 	kref_init(&req->r_kref);
1655 	INIT_LIST_HEAD(&req->r_wait);
1656 	init_completion(&req->r_completion);
1657 	init_completion(&req->r_safe_completion);
1658 	INIT_LIST_HEAD(&req->r_unsafe_item);
1659 
1660 	req->r_stamp = CURRENT_TIME;
1661 
1662 	req->r_op = op;
1663 	req->r_direct_mode = mode;
1664 	return req;
1665 }
1666 
1667 /*
1668  * return oldest (lowest) request, tid in request tree, 0 if none.
1669  *
1670  * called under mdsc->mutex.
1671  */
1672 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1673 {
1674 	if (RB_EMPTY_ROOT(&mdsc->request_tree))
1675 		return NULL;
1676 	return rb_entry(rb_first(&mdsc->request_tree),
1677 			struct ceph_mds_request, r_node);
1678 }
1679 
1680 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1681 {
1682 	struct ceph_mds_request *req = __get_oldest_req(mdsc);
1683 
1684 	if (req)
1685 		return req->r_tid;
1686 	return 0;
1687 }
1688 
1689 /*
1690  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1691  * on build_path_from_dentry in fs/cifs/dir.c.
1692  *
1693  * If @stop_on_nosnap, generate path relative to the first non-snapped
1694  * inode.
1695  *
1696  * Encode hidden .snap dirs as a double /, i.e.
1697  *   foo/.snap/bar -> foo//bar
1698  */
1699 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1700 			   int stop_on_nosnap)
1701 {
1702 	struct dentry *temp;
1703 	char *path;
1704 	int len, pos;
1705 	unsigned seq;
1706 
1707 	if (dentry == NULL)
1708 		return ERR_PTR(-EINVAL);
1709 
1710 retry:
1711 	len = 0;
1712 	seq = read_seqbegin(&rename_lock);
1713 	rcu_read_lock();
1714 	for (temp = dentry; !IS_ROOT(temp);) {
1715 		struct inode *inode = temp->d_inode;
1716 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1717 			len++;  /* slash only */
1718 		else if (stop_on_nosnap && inode &&
1719 			 ceph_snap(inode) == CEPH_NOSNAP)
1720 			break;
1721 		else
1722 			len += 1 + temp->d_name.len;
1723 		temp = temp->d_parent;
1724 	}
1725 	rcu_read_unlock();
1726 	if (len)
1727 		len--;  /* no leading '/' */
1728 
1729 	path = kmalloc(len+1, GFP_NOFS);
1730 	if (path == NULL)
1731 		return ERR_PTR(-ENOMEM);
1732 	pos = len;
1733 	path[pos] = 0;	/* trailing null */
1734 	rcu_read_lock();
1735 	for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1736 		struct inode *inode;
1737 
1738 		spin_lock(&temp->d_lock);
1739 		inode = temp->d_inode;
1740 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1741 			dout("build_path path+%d: %p SNAPDIR\n",
1742 			     pos, temp);
1743 		} else if (stop_on_nosnap && inode &&
1744 			   ceph_snap(inode) == CEPH_NOSNAP) {
1745 			spin_unlock(&temp->d_lock);
1746 			break;
1747 		} else {
1748 			pos -= temp->d_name.len;
1749 			if (pos < 0) {
1750 				spin_unlock(&temp->d_lock);
1751 				break;
1752 			}
1753 			strncpy(path + pos, temp->d_name.name,
1754 				temp->d_name.len);
1755 		}
1756 		spin_unlock(&temp->d_lock);
1757 		if (pos)
1758 			path[--pos] = '/';
1759 		temp = temp->d_parent;
1760 	}
1761 	rcu_read_unlock();
1762 	if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1763 		pr_err("build_path did not end path lookup where "
1764 		       "expected, namelen is %d, pos is %d\n", len, pos);
1765 		/* presumably this is only possible if racing with a
1766 		   rename of one of the parent directories (we can not
1767 		   lock the dentries above us to prevent this, but
1768 		   retrying should be harmless) */
1769 		kfree(path);
1770 		goto retry;
1771 	}
1772 
1773 	*base = ceph_ino(temp->d_inode);
1774 	*plen = len;
1775 	dout("build_path on %p %d built %llx '%.*s'\n",
1776 	     dentry, d_count(dentry), *base, len, path);
1777 	return path;
1778 }
1779 
1780 static int build_dentry_path(struct dentry *dentry,
1781 			     const char **ppath, int *ppathlen, u64 *pino,
1782 			     int *pfreepath)
1783 {
1784 	char *path;
1785 
1786 	if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1787 		*pino = ceph_ino(dentry->d_parent->d_inode);
1788 		*ppath = dentry->d_name.name;
1789 		*ppathlen = dentry->d_name.len;
1790 		return 0;
1791 	}
1792 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1793 	if (IS_ERR(path))
1794 		return PTR_ERR(path);
1795 	*ppath = path;
1796 	*pfreepath = 1;
1797 	return 0;
1798 }
1799 
1800 static int build_inode_path(struct inode *inode,
1801 			    const char **ppath, int *ppathlen, u64 *pino,
1802 			    int *pfreepath)
1803 {
1804 	struct dentry *dentry;
1805 	char *path;
1806 
1807 	if (ceph_snap(inode) == CEPH_NOSNAP) {
1808 		*pino = ceph_ino(inode);
1809 		*ppathlen = 0;
1810 		return 0;
1811 	}
1812 	dentry = d_find_alias(inode);
1813 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1814 	dput(dentry);
1815 	if (IS_ERR(path))
1816 		return PTR_ERR(path);
1817 	*ppath = path;
1818 	*pfreepath = 1;
1819 	return 0;
1820 }
1821 
1822 /*
1823  * request arguments may be specified via an inode *, a dentry *, or
1824  * an explicit ino+path.
1825  */
1826 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1827 				  const char *rpath, u64 rino,
1828 				  const char **ppath, int *pathlen,
1829 				  u64 *ino, int *freepath)
1830 {
1831 	int r = 0;
1832 
1833 	if (rinode) {
1834 		r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1835 		dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1836 		     ceph_snap(rinode));
1837 	} else if (rdentry) {
1838 		r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1839 		dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1840 		     *ppath);
1841 	} else if (rpath || rino) {
1842 		*ino = rino;
1843 		*ppath = rpath;
1844 		*pathlen = rpath ? strlen(rpath) : 0;
1845 		dout(" path %.*s\n", *pathlen, rpath);
1846 	}
1847 
1848 	return r;
1849 }
1850 
1851 /*
1852  * called under mdsc->mutex
1853  */
1854 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1855 					       struct ceph_mds_request *req,
1856 					       int mds)
1857 {
1858 	struct ceph_msg *msg;
1859 	struct ceph_mds_request_head *head;
1860 	const char *path1 = NULL;
1861 	const char *path2 = NULL;
1862 	u64 ino1 = 0, ino2 = 0;
1863 	int pathlen1 = 0, pathlen2 = 0;
1864 	int freepath1 = 0, freepath2 = 0;
1865 	int len;
1866 	u16 releases;
1867 	void *p, *end;
1868 	int ret;
1869 
1870 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
1871 			      req->r_path1, req->r_ino1.ino,
1872 			      &path1, &pathlen1, &ino1, &freepath1);
1873 	if (ret < 0) {
1874 		msg = ERR_PTR(ret);
1875 		goto out;
1876 	}
1877 
1878 	ret = set_request_path_attr(NULL, req->r_old_dentry,
1879 			      req->r_path2, req->r_ino2.ino,
1880 			      &path2, &pathlen2, &ino2, &freepath2);
1881 	if (ret < 0) {
1882 		msg = ERR_PTR(ret);
1883 		goto out_free1;
1884 	}
1885 
1886 	len = sizeof(*head) +
1887 		pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1888 		sizeof(struct timespec);
1889 
1890 	/* calculate (max) length for cap releases */
1891 	len += sizeof(struct ceph_mds_request_release) *
1892 		(!!req->r_inode_drop + !!req->r_dentry_drop +
1893 		 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1894 	if (req->r_dentry_drop)
1895 		len += req->r_dentry->d_name.len;
1896 	if (req->r_old_dentry_drop)
1897 		len += req->r_old_dentry->d_name.len;
1898 
1899 	msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1900 	if (!msg) {
1901 		msg = ERR_PTR(-ENOMEM);
1902 		goto out_free2;
1903 	}
1904 
1905 	msg->hdr.version = cpu_to_le16(2);
1906 	msg->hdr.tid = cpu_to_le64(req->r_tid);
1907 
1908 	head = msg->front.iov_base;
1909 	p = msg->front.iov_base + sizeof(*head);
1910 	end = msg->front.iov_base + msg->front.iov_len;
1911 
1912 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1913 	head->op = cpu_to_le32(req->r_op);
1914 	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1915 	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1916 	head->args = req->r_args;
1917 
1918 	ceph_encode_filepath(&p, end, ino1, path1);
1919 	ceph_encode_filepath(&p, end, ino2, path2);
1920 
1921 	/* make note of release offset, in case we need to replay */
1922 	req->r_request_release_offset = p - msg->front.iov_base;
1923 
1924 	/* cap releases */
1925 	releases = 0;
1926 	if (req->r_inode_drop)
1927 		releases += ceph_encode_inode_release(&p,
1928 		      req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1929 		      mds, req->r_inode_drop, req->r_inode_unless, 0);
1930 	if (req->r_dentry_drop)
1931 		releases += ceph_encode_dentry_release(&p, req->r_dentry,
1932 		       mds, req->r_dentry_drop, req->r_dentry_unless);
1933 	if (req->r_old_dentry_drop)
1934 		releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1935 		       mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1936 	if (req->r_old_inode_drop)
1937 		releases += ceph_encode_inode_release(&p,
1938 		      req->r_old_dentry->d_inode,
1939 		      mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1940 	head->num_releases = cpu_to_le16(releases);
1941 
1942 	/* time stamp */
1943 	{
1944 		struct ceph_timespec ts;
1945 		ceph_encode_timespec(&ts, &req->r_stamp);
1946 		ceph_encode_copy(&p, &ts, sizeof(ts));
1947 	}
1948 
1949 	BUG_ON(p > end);
1950 	msg->front.iov_len = p - msg->front.iov_base;
1951 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1952 
1953 	if (req->r_pagelist) {
1954 		struct ceph_pagelist *pagelist = req->r_pagelist;
1955 		atomic_inc(&pagelist->refcnt);
1956 		ceph_msg_data_add_pagelist(msg, pagelist);
1957 		msg->hdr.data_len = cpu_to_le32(pagelist->length);
1958 	} else {
1959 		msg->hdr.data_len = 0;
1960 	}
1961 
1962 	msg->hdr.data_off = cpu_to_le16(0);
1963 
1964 out_free2:
1965 	if (freepath2)
1966 		kfree((char *)path2);
1967 out_free1:
1968 	if (freepath1)
1969 		kfree((char *)path1);
1970 out:
1971 	return msg;
1972 }
1973 
1974 /*
1975  * called under mdsc->mutex if error, under no mutex if
1976  * success.
1977  */
1978 static void complete_request(struct ceph_mds_client *mdsc,
1979 			     struct ceph_mds_request *req)
1980 {
1981 	if (req->r_callback)
1982 		req->r_callback(mdsc, req);
1983 	else
1984 		complete_all(&req->r_completion);
1985 }
1986 
1987 /*
1988  * called under mdsc->mutex
1989  */
1990 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1991 				  struct ceph_mds_request *req,
1992 				  int mds)
1993 {
1994 	struct ceph_mds_request_head *rhead;
1995 	struct ceph_msg *msg;
1996 	int flags = 0;
1997 
1998 	req->r_attempts++;
1999 	if (req->r_inode) {
2000 		struct ceph_cap *cap =
2001 			ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2002 
2003 		if (cap)
2004 			req->r_sent_on_mseq = cap->mseq;
2005 		else
2006 			req->r_sent_on_mseq = -1;
2007 	}
2008 	dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2009 	     req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2010 
2011 	if (req->r_got_unsafe) {
2012 		void *p;
2013 		/*
2014 		 * Replay.  Do not regenerate message (and rebuild
2015 		 * paths, etc.); just use the original message.
2016 		 * Rebuilding paths will break for renames because
2017 		 * d_move mangles the src name.
2018 		 */
2019 		msg = req->r_request;
2020 		rhead = msg->front.iov_base;
2021 
2022 		flags = le32_to_cpu(rhead->flags);
2023 		flags |= CEPH_MDS_FLAG_REPLAY;
2024 		rhead->flags = cpu_to_le32(flags);
2025 
2026 		if (req->r_target_inode)
2027 			rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2028 
2029 		rhead->num_retry = req->r_attempts - 1;
2030 
2031 		/* remove cap/dentry releases from message */
2032 		rhead->num_releases = 0;
2033 
2034 		/* time stamp */
2035 		p = msg->front.iov_base + req->r_request_release_offset;
2036 		{
2037 			struct ceph_timespec ts;
2038 			ceph_encode_timespec(&ts, &req->r_stamp);
2039 			ceph_encode_copy(&p, &ts, sizeof(ts));
2040 		}
2041 
2042 		msg->front.iov_len = p - msg->front.iov_base;
2043 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2044 		return 0;
2045 	}
2046 
2047 	if (req->r_request) {
2048 		ceph_msg_put(req->r_request);
2049 		req->r_request = NULL;
2050 	}
2051 	msg = create_request_message(mdsc, req, mds);
2052 	if (IS_ERR(msg)) {
2053 		req->r_err = PTR_ERR(msg);
2054 		complete_request(mdsc, req);
2055 		return PTR_ERR(msg);
2056 	}
2057 	req->r_request = msg;
2058 
2059 	rhead = msg->front.iov_base;
2060 	rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2061 	if (req->r_got_unsafe)
2062 		flags |= CEPH_MDS_FLAG_REPLAY;
2063 	if (req->r_locked_dir)
2064 		flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2065 	rhead->flags = cpu_to_le32(flags);
2066 	rhead->num_fwd = req->r_num_fwd;
2067 	rhead->num_retry = req->r_attempts - 1;
2068 	rhead->ino = 0;
2069 
2070 	dout(" r_locked_dir = %p\n", req->r_locked_dir);
2071 	return 0;
2072 }
2073 
2074 /*
2075  * send request, or put it on the appropriate wait list.
2076  */
2077 static int __do_request(struct ceph_mds_client *mdsc,
2078 			struct ceph_mds_request *req)
2079 {
2080 	struct ceph_mds_session *session = NULL;
2081 	int mds = -1;
2082 	int err = -EAGAIN;
2083 
2084 	if (req->r_err || req->r_got_result) {
2085 		if (req->r_aborted)
2086 			__unregister_request(mdsc, req);
2087 		goto out;
2088 	}
2089 
2090 	if (req->r_timeout &&
2091 	    time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2092 		dout("do_request timed out\n");
2093 		err = -EIO;
2094 		goto finish;
2095 	}
2096 
2097 	put_request_session(req);
2098 
2099 	mds = __choose_mds(mdsc, req);
2100 	if (mds < 0 ||
2101 	    ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2102 		dout("do_request no mds or not active, waiting for map\n");
2103 		list_add(&req->r_wait, &mdsc->waiting_for_map);
2104 		goto out;
2105 	}
2106 
2107 	/* get, open session */
2108 	session = __ceph_lookup_mds_session(mdsc, mds);
2109 	if (!session) {
2110 		session = register_session(mdsc, mds);
2111 		if (IS_ERR(session)) {
2112 			err = PTR_ERR(session);
2113 			goto finish;
2114 		}
2115 	}
2116 	req->r_session = get_session(session);
2117 
2118 	dout("do_request mds%d session %p state %s\n", mds, session,
2119 	     ceph_session_state_name(session->s_state));
2120 	if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2121 	    session->s_state != CEPH_MDS_SESSION_HUNG) {
2122 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
2123 		    session->s_state == CEPH_MDS_SESSION_CLOSING)
2124 			__open_session(mdsc, session);
2125 		list_add(&req->r_wait, &session->s_waiting);
2126 		goto out_session;
2127 	}
2128 
2129 	/* send request */
2130 	req->r_resend_mds = -1;   /* forget any previous mds hint */
2131 
2132 	if (req->r_request_started == 0)   /* note request start time */
2133 		req->r_request_started = jiffies;
2134 
2135 	err = __prepare_send_request(mdsc, req, mds);
2136 	if (!err) {
2137 		ceph_msg_get(req->r_request);
2138 		ceph_con_send(&session->s_con, req->r_request);
2139 	}
2140 
2141 out_session:
2142 	ceph_put_mds_session(session);
2143 out:
2144 	return err;
2145 
2146 finish:
2147 	req->r_err = err;
2148 	complete_request(mdsc, req);
2149 	goto out;
2150 }
2151 
2152 /*
2153  * called under mdsc->mutex
2154  */
2155 static void __wake_requests(struct ceph_mds_client *mdsc,
2156 			    struct list_head *head)
2157 {
2158 	struct ceph_mds_request *req;
2159 	LIST_HEAD(tmp_list);
2160 
2161 	list_splice_init(head, &tmp_list);
2162 
2163 	while (!list_empty(&tmp_list)) {
2164 		req = list_entry(tmp_list.next,
2165 				 struct ceph_mds_request, r_wait);
2166 		list_del_init(&req->r_wait);
2167 		dout(" wake request %p tid %llu\n", req, req->r_tid);
2168 		__do_request(mdsc, req);
2169 	}
2170 }
2171 
2172 /*
2173  * Wake up threads with requests pending for @mds, so that they can
2174  * resubmit their requests to a possibly different mds.
2175  */
2176 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2177 {
2178 	struct ceph_mds_request *req;
2179 	struct rb_node *p = rb_first(&mdsc->request_tree);
2180 
2181 	dout("kick_requests mds%d\n", mds);
2182 	while (p) {
2183 		req = rb_entry(p, struct ceph_mds_request, r_node);
2184 		p = rb_next(p);
2185 		if (req->r_got_unsafe)
2186 			continue;
2187 		if (req->r_attempts > 0)
2188 			continue; /* only new requests */
2189 		if (req->r_session &&
2190 		    req->r_session->s_mds == mds) {
2191 			dout(" kicking tid %llu\n", req->r_tid);
2192 			list_del_init(&req->r_wait);
2193 			__do_request(mdsc, req);
2194 		}
2195 	}
2196 }
2197 
2198 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2199 			      struct ceph_mds_request *req)
2200 {
2201 	dout("submit_request on %p\n", req);
2202 	mutex_lock(&mdsc->mutex);
2203 	__register_request(mdsc, req, NULL);
2204 	__do_request(mdsc, req);
2205 	mutex_unlock(&mdsc->mutex);
2206 }
2207 
2208 /*
2209  * Synchrously perform an mds request.  Take care of all of the
2210  * session setup, forwarding, retry details.
2211  */
2212 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2213 			 struct inode *dir,
2214 			 struct ceph_mds_request *req)
2215 {
2216 	int err;
2217 
2218 	dout("do_request on %p\n", req);
2219 
2220 	/* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2221 	if (req->r_inode)
2222 		ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2223 	if (req->r_locked_dir)
2224 		ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2225 	if (req->r_old_dentry_dir)
2226 		ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2227 				  CEPH_CAP_PIN);
2228 
2229 	/* issue */
2230 	mutex_lock(&mdsc->mutex);
2231 	__register_request(mdsc, req, dir);
2232 	__do_request(mdsc, req);
2233 
2234 	if (req->r_err) {
2235 		err = req->r_err;
2236 		__unregister_request(mdsc, req);
2237 		dout("do_request early error %d\n", err);
2238 		goto out;
2239 	}
2240 
2241 	/* wait */
2242 	mutex_unlock(&mdsc->mutex);
2243 	dout("do_request waiting\n");
2244 	if (req->r_timeout) {
2245 		err = (long)wait_for_completion_killable_timeout(
2246 			&req->r_completion, req->r_timeout);
2247 		if (err == 0)
2248 			err = -EIO;
2249 	} else if (req->r_wait_for_completion) {
2250 		err = req->r_wait_for_completion(mdsc, req);
2251 	} else {
2252 		err = wait_for_completion_killable(&req->r_completion);
2253 	}
2254 	dout("do_request waited, got %d\n", err);
2255 	mutex_lock(&mdsc->mutex);
2256 
2257 	/* only abort if we didn't race with a real reply */
2258 	if (req->r_got_result) {
2259 		err = le32_to_cpu(req->r_reply_info.head->result);
2260 	} else if (err < 0) {
2261 		dout("aborted request %lld with %d\n", req->r_tid, err);
2262 
2263 		/*
2264 		 * ensure we aren't running concurrently with
2265 		 * ceph_fill_trace or ceph_readdir_prepopulate, which
2266 		 * rely on locks (dir mutex) held by our caller.
2267 		 */
2268 		mutex_lock(&req->r_fill_mutex);
2269 		req->r_err = err;
2270 		req->r_aborted = true;
2271 		mutex_unlock(&req->r_fill_mutex);
2272 
2273 		if (req->r_locked_dir &&
2274 		    (req->r_op & CEPH_MDS_OP_WRITE))
2275 			ceph_invalidate_dir_request(req);
2276 	} else {
2277 		err = req->r_err;
2278 	}
2279 
2280 out:
2281 	mutex_unlock(&mdsc->mutex);
2282 	dout("do_request %p done, result %d\n", req, err);
2283 	return err;
2284 }
2285 
2286 /*
2287  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2288  * namespace request.
2289  */
2290 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2291 {
2292 	struct inode *inode = req->r_locked_dir;
2293 
2294 	dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2295 
2296 	ceph_dir_clear_complete(inode);
2297 	if (req->r_dentry)
2298 		ceph_invalidate_dentry_lease(req->r_dentry);
2299 	if (req->r_old_dentry)
2300 		ceph_invalidate_dentry_lease(req->r_old_dentry);
2301 }
2302 
2303 /*
2304  * Handle mds reply.
2305  *
2306  * We take the session mutex and parse and process the reply immediately.
2307  * This preserves the logical ordering of replies, capabilities, etc., sent
2308  * by the MDS as they are applied to our local cache.
2309  */
2310 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2311 {
2312 	struct ceph_mds_client *mdsc = session->s_mdsc;
2313 	struct ceph_mds_request *req;
2314 	struct ceph_mds_reply_head *head = msg->front.iov_base;
2315 	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2316 	struct ceph_snap_realm *realm;
2317 	u64 tid;
2318 	int err, result;
2319 	int mds = session->s_mds;
2320 
2321 	if (msg->front.iov_len < sizeof(*head)) {
2322 		pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2323 		ceph_msg_dump(msg);
2324 		return;
2325 	}
2326 
2327 	/* get request, session */
2328 	tid = le64_to_cpu(msg->hdr.tid);
2329 	mutex_lock(&mdsc->mutex);
2330 	req = __lookup_request(mdsc, tid);
2331 	if (!req) {
2332 		dout("handle_reply on unknown tid %llu\n", tid);
2333 		mutex_unlock(&mdsc->mutex);
2334 		return;
2335 	}
2336 	dout("handle_reply %p\n", req);
2337 
2338 	/* correct session? */
2339 	if (req->r_session != session) {
2340 		pr_err("mdsc_handle_reply got %llu on session mds%d"
2341 		       " not mds%d\n", tid, session->s_mds,
2342 		       req->r_session ? req->r_session->s_mds : -1);
2343 		mutex_unlock(&mdsc->mutex);
2344 		goto out;
2345 	}
2346 
2347 	/* dup? */
2348 	if ((req->r_got_unsafe && !head->safe) ||
2349 	    (req->r_got_safe && head->safe)) {
2350 		pr_warn("got a dup %s reply on %llu from mds%d\n",
2351 			   head->safe ? "safe" : "unsafe", tid, mds);
2352 		mutex_unlock(&mdsc->mutex);
2353 		goto out;
2354 	}
2355 	if (req->r_got_safe && !head->safe) {
2356 		pr_warn("got unsafe after safe on %llu from mds%d\n",
2357 			   tid, mds);
2358 		mutex_unlock(&mdsc->mutex);
2359 		goto out;
2360 	}
2361 
2362 	result = le32_to_cpu(head->result);
2363 
2364 	/*
2365 	 * Handle an ESTALE
2366 	 * if we're not talking to the authority, send to them
2367 	 * if the authority has changed while we weren't looking,
2368 	 * send to new authority
2369 	 * Otherwise we just have to return an ESTALE
2370 	 */
2371 	if (result == -ESTALE) {
2372 		dout("got ESTALE on request %llu", req->r_tid);
2373 		req->r_resend_mds = -1;
2374 		if (req->r_direct_mode != USE_AUTH_MDS) {
2375 			dout("not using auth, setting for that now");
2376 			req->r_direct_mode = USE_AUTH_MDS;
2377 			__do_request(mdsc, req);
2378 			mutex_unlock(&mdsc->mutex);
2379 			goto out;
2380 		} else  {
2381 			int mds = __choose_mds(mdsc, req);
2382 			if (mds >= 0 && mds != req->r_session->s_mds) {
2383 				dout("but auth changed, so resending");
2384 				__do_request(mdsc, req);
2385 				mutex_unlock(&mdsc->mutex);
2386 				goto out;
2387 			}
2388 		}
2389 		dout("have to return ESTALE on request %llu", req->r_tid);
2390 	}
2391 
2392 
2393 	if (head->safe) {
2394 		req->r_got_safe = true;
2395 		__unregister_request(mdsc, req);
2396 
2397 		if (req->r_got_unsafe) {
2398 			/*
2399 			 * We already handled the unsafe response, now do the
2400 			 * cleanup.  No need to examine the response; the MDS
2401 			 * doesn't include any result info in the safe
2402 			 * response.  And even if it did, there is nothing
2403 			 * useful we could do with a revised return value.
2404 			 */
2405 			dout("got safe reply %llu, mds%d\n", tid, mds);
2406 			list_del_init(&req->r_unsafe_item);
2407 
2408 			/* last unsafe request during umount? */
2409 			if (mdsc->stopping && !__get_oldest_req(mdsc))
2410 				complete_all(&mdsc->safe_umount_waiters);
2411 			mutex_unlock(&mdsc->mutex);
2412 			goto out;
2413 		}
2414 	} else {
2415 		req->r_got_unsafe = true;
2416 		list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2417 	}
2418 
2419 	dout("handle_reply tid %lld result %d\n", tid, result);
2420 	rinfo = &req->r_reply_info;
2421 	err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2422 	mutex_unlock(&mdsc->mutex);
2423 
2424 	mutex_lock(&session->s_mutex);
2425 	if (err < 0) {
2426 		pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2427 		ceph_msg_dump(msg);
2428 		goto out_err;
2429 	}
2430 
2431 	/* snap trace */
2432 	realm = NULL;
2433 	if (rinfo->snapblob_len) {
2434 		down_write(&mdsc->snap_rwsem);
2435 		ceph_update_snap_trace(mdsc, rinfo->snapblob,
2436 				rinfo->snapblob + rinfo->snapblob_len,
2437 				le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2438 				&realm);
2439 		downgrade_write(&mdsc->snap_rwsem);
2440 	} else {
2441 		down_read(&mdsc->snap_rwsem);
2442 	}
2443 
2444 	/* insert trace into our cache */
2445 	mutex_lock(&req->r_fill_mutex);
2446 	err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2447 	if (err == 0) {
2448 		if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2449 				    req->r_op == CEPH_MDS_OP_LSSNAP))
2450 			ceph_readdir_prepopulate(req, req->r_session);
2451 		ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2452 	}
2453 	mutex_unlock(&req->r_fill_mutex);
2454 
2455 	up_read(&mdsc->snap_rwsem);
2456 	if (realm)
2457 		ceph_put_snap_realm(mdsc, realm);
2458 out_err:
2459 	mutex_lock(&mdsc->mutex);
2460 	if (!req->r_aborted) {
2461 		if (err) {
2462 			req->r_err = err;
2463 		} else {
2464 			req->r_reply = msg;
2465 			ceph_msg_get(msg);
2466 			req->r_got_result = true;
2467 		}
2468 	} else {
2469 		dout("reply arrived after request %lld was aborted\n", tid);
2470 	}
2471 	mutex_unlock(&mdsc->mutex);
2472 
2473 	ceph_add_cap_releases(mdsc, req->r_session);
2474 	mutex_unlock(&session->s_mutex);
2475 
2476 	/* kick calling process */
2477 	complete_request(mdsc, req);
2478 out:
2479 	ceph_mdsc_put_request(req);
2480 	return;
2481 }
2482 
2483 
2484 
2485 /*
2486  * handle mds notification that our request has been forwarded.
2487  */
2488 static void handle_forward(struct ceph_mds_client *mdsc,
2489 			   struct ceph_mds_session *session,
2490 			   struct ceph_msg *msg)
2491 {
2492 	struct ceph_mds_request *req;
2493 	u64 tid = le64_to_cpu(msg->hdr.tid);
2494 	u32 next_mds;
2495 	u32 fwd_seq;
2496 	int err = -EINVAL;
2497 	void *p = msg->front.iov_base;
2498 	void *end = p + msg->front.iov_len;
2499 
2500 	ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2501 	next_mds = ceph_decode_32(&p);
2502 	fwd_seq = ceph_decode_32(&p);
2503 
2504 	mutex_lock(&mdsc->mutex);
2505 	req = __lookup_request(mdsc, tid);
2506 	if (!req) {
2507 		dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2508 		goto out;  /* dup reply? */
2509 	}
2510 
2511 	if (req->r_aborted) {
2512 		dout("forward tid %llu aborted, unregistering\n", tid);
2513 		__unregister_request(mdsc, req);
2514 	} else if (fwd_seq <= req->r_num_fwd) {
2515 		dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2516 		     tid, next_mds, req->r_num_fwd, fwd_seq);
2517 	} else {
2518 		/* resend. forward race not possible; mds would drop */
2519 		dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2520 		BUG_ON(req->r_err);
2521 		BUG_ON(req->r_got_result);
2522 		req->r_attempts = 0;
2523 		req->r_num_fwd = fwd_seq;
2524 		req->r_resend_mds = next_mds;
2525 		put_request_session(req);
2526 		__do_request(mdsc, req);
2527 	}
2528 	ceph_mdsc_put_request(req);
2529 out:
2530 	mutex_unlock(&mdsc->mutex);
2531 	return;
2532 
2533 bad:
2534 	pr_err("mdsc_handle_forward decode error err=%d\n", err);
2535 }
2536 
2537 /*
2538  * handle a mds session control message
2539  */
2540 static void handle_session(struct ceph_mds_session *session,
2541 			   struct ceph_msg *msg)
2542 {
2543 	struct ceph_mds_client *mdsc = session->s_mdsc;
2544 	u32 op;
2545 	u64 seq;
2546 	int mds = session->s_mds;
2547 	struct ceph_mds_session_head *h = msg->front.iov_base;
2548 	int wake = 0;
2549 
2550 	/* decode */
2551 	if (msg->front.iov_len != sizeof(*h))
2552 		goto bad;
2553 	op = le32_to_cpu(h->op);
2554 	seq = le64_to_cpu(h->seq);
2555 
2556 	mutex_lock(&mdsc->mutex);
2557 	if (op == CEPH_SESSION_CLOSE)
2558 		__unregister_session(mdsc, session);
2559 	/* FIXME: this ttl calculation is generous */
2560 	session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2561 	mutex_unlock(&mdsc->mutex);
2562 
2563 	mutex_lock(&session->s_mutex);
2564 
2565 	dout("handle_session mds%d %s %p state %s seq %llu\n",
2566 	     mds, ceph_session_op_name(op), session,
2567 	     ceph_session_state_name(session->s_state), seq);
2568 
2569 	if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2570 		session->s_state = CEPH_MDS_SESSION_OPEN;
2571 		pr_info("mds%d came back\n", session->s_mds);
2572 	}
2573 
2574 	switch (op) {
2575 	case CEPH_SESSION_OPEN:
2576 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2577 			pr_info("mds%d reconnect success\n", session->s_mds);
2578 		session->s_state = CEPH_MDS_SESSION_OPEN;
2579 		renewed_caps(mdsc, session, 0);
2580 		wake = 1;
2581 		if (mdsc->stopping)
2582 			__close_session(mdsc, session);
2583 		break;
2584 
2585 	case CEPH_SESSION_RENEWCAPS:
2586 		if (session->s_renew_seq == seq)
2587 			renewed_caps(mdsc, session, 1);
2588 		break;
2589 
2590 	case CEPH_SESSION_CLOSE:
2591 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2592 			pr_info("mds%d reconnect denied\n", session->s_mds);
2593 		remove_session_caps(session);
2594 		wake = 2; /* for good measure */
2595 		wake_up_all(&mdsc->session_close_wq);
2596 		break;
2597 
2598 	case CEPH_SESSION_STALE:
2599 		pr_info("mds%d caps went stale, renewing\n",
2600 			session->s_mds);
2601 		spin_lock(&session->s_gen_ttl_lock);
2602 		session->s_cap_gen++;
2603 		session->s_cap_ttl = jiffies - 1;
2604 		spin_unlock(&session->s_gen_ttl_lock);
2605 		send_renew_caps(mdsc, session);
2606 		break;
2607 
2608 	case CEPH_SESSION_RECALL_STATE:
2609 		trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2610 		break;
2611 
2612 	case CEPH_SESSION_FLUSHMSG:
2613 		send_flushmsg_ack(mdsc, session, seq);
2614 		break;
2615 
2616 	case CEPH_SESSION_FORCE_RO:
2617 		dout("force_session_readonly %p\n", session);
2618 		spin_lock(&session->s_cap_lock);
2619 		session->s_readonly = true;
2620 		spin_unlock(&session->s_cap_lock);
2621 		wake_up_session_caps(session, 0);
2622 		break;
2623 
2624 	default:
2625 		pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2626 		WARN_ON(1);
2627 	}
2628 
2629 	mutex_unlock(&session->s_mutex);
2630 	if (wake) {
2631 		mutex_lock(&mdsc->mutex);
2632 		__wake_requests(mdsc, &session->s_waiting);
2633 		if (wake == 2)
2634 			kick_requests(mdsc, mds);
2635 		mutex_unlock(&mdsc->mutex);
2636 	}
2637 	return;
2638 
2639 bad:
2640 	pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2641 	       (int)msg->front.iov_len);
2642 	ceph_msg_dump(msg);
2643 	return;
2644 }
2645 
2646 
2647 /*
2648  * called under session->mutex.
2649  */
2650 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2651 				   struct ceph_mds_session *session)
2652 {
2653 	struct ceph_mds_request *req, *nreq;
2654 	struct rb_node *p;
2655 	int err;
2656 
2657 	dout("replay_unsafe_requests mds%d\n", session->s_mds);
2658 
2659 	mutex_lock(&mdsc->mutex);
2660 	list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2661 		err = __prepare_send_request(mdsc, req, session->s_mds);
2662 		if (!err) {
2663 			ceph_msg_get(req->r_request);
2664 			ceph_con_send(&session->s_con, req->r_request);
2665 		}
2666 	}
2667 
2668 	/*
2669 	 * also re-send old requests when MDS enters reconnect stage. So that MDS
2670 	 * can process completed request in clientreplay stage.
2671 	 */
2672 	p = rb_first(&mdsc->request_tree);
2673 	while (p) {
2674 		req = rb_entry(p, struct ceph_mds_request, r_node);
2675 		p = rb_next(p);
2676 		if (req->r_got_unsafe)
2677 			continue;
2678 		if (req->r_attempts == 0)
2679 			continue; /* only old requests */
2680 		if (req->r_session &&
2681 		    req->r_session->s_mds == session->s_mds) {
2682 			err = __prepare_send_request(mdsc, req, session->s_mds);
2683 			if (!err) {
2684 				ceph_msg_get(req->r_request);
2685 				ceph_con_send(&session->s_con, req->r_request);
2686 			}
2687 		}
2688 	}
2689 	mutex_unlock(&mdsc->mutex);
2690 }
2691 
2692 /*
2693  * Encode information about a cap for a reconnect with the MDS.
2694  */
2695 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2696 			  void *arg)
2697 {
2698 	union {
2699 		struct ceph_mds_cap_reconnect v2;
2700 		struct ceph_mds_cap_reconnect_v1 v1;
2701 	} rec;
2702 	size_t reclen;
2703 	struct ceph_inode_info *ci;
2704 	struct ceph_reconnect_state *recon_state = arg;
2705 	struct ceph_pagelist *pagelist = recon_state->pagelist;
2706 	char *path;
2707 	int pathlen, err;
2708 	u64 pathbase;
2709 	struct dentry *dentry;
2710 
2711 	ci = cap->ci;
2712 
2713 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2714 	     inode, ceph_vinop(inode), cap, cap->cap_id,
2715 	     ceph_cap_string(cap->issued));
2716 	err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2717 	if (err)
2718 		return err;
2719 
2720 	dentry = d_find_alias(inode);
2721 	if (dentry) {
2722 		path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2723 		if (IS_ERR(path)) {
2724 			err = PTR_ERR(path);
2725 			goto out_dput;
2726 		}
2727 	} else {
2728 		path = NULL;
2729 		pathlen = 0;
2730 	}
2731 	err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2732 	if (err)
2733 		goto out_free;
2734 
2735 	spin_lock(&ci->i_ceph_lock);
2736 	cap->seq = 0;        /* reset cap seq */
2737 	cap->issue_seq = 0;  /* and issue_seq */
2738 	cap->mseq = 0;       /* and migrate_seq */
2739 	cap->cap_gen = cap->session->s_cap_gen;
2740 
2741 	if (recon_state->flock) {
2742 		rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2743 		rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2744 		rec.v2.issued = cpu_to_le32(cap->issued);
2745 		rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2746 		rec.v2.pathbase = cpu_to_le64(pathbase);
2747 		rec.v2.flock_len = 0;
2748 		reclen = sizeof(rec.v2);
2749 	} else {
2750 		rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2751 		rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2752 		rec.v1.issued = cpu_to_le32(cap->issued);
2753 		rec.v1.size = cpu_to_le64(inode->i_size);
2754 		ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2755 		ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2756 		rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2757 		rec.v1.pathbase = cpu_to_le64(pathbase);
2758 		reclen = sizeof(rec.v1);
2759 	}
2760 	spin_unlock(&ci->i_ceph_lock);
2761 
2762 	if (recon_state->flock) {
2763 		int num_fcntl_locks, num_flock_locks;
2764 		struct ceph_filelock *flocks;
2765 
2766 encode_again:
2767 		ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2768 		flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2769 				 sizeof(struct ceph_filelock), GFP_NOFS);
2770 		if (!flocks) {
2771 			err = -ENOMEM;
2772 			goto out_free;
2773 		}
2774 		err = ceph_encode_locks_to_buffer(inode, flocks,
2775 						  num_fcntl_locks,
2776 						  num_flock_locks);
2777 		if (err) {
2778 			kfree(flocks);
2779 			if (err == -ENOSPC)
2780 				goto encode_again;
2781 			goto out_free;
2782 		}
2783 		/*
2784 		 * number of encoded locks is stable, so copy to pagelist
2785 		 */
2786 		rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2787 				    (num_fcntl_locks+num_flock_locks) *
2788 				    sizeof(struct ceph_filelock));
2789 		err = ceph_pagelist_append(pagelist, &rec, reclen);
2790 		if (!err)
2791 			err = ceph_locks_to_pagelist(flocks, pagelist,
2792 						     num_fcntl_locks,
2793 						     num_flock_locks);
2794 		kfree(flocks);
2795 	} else {
2796 		err = ceph_pagelist_append(pagelist, &rec, reclen);
2797 	}
2798 
2799 	recon_state->nr_caps++;
2800 out_free:
2801 	kfree(path);
2802 out_dput:
2803 	dput(dentry);
2804 	return err;
2805 }
2806 
2807 
2808 /*
2809  * If an MDS fails and recovers, clients need to reconnect in order to
2810  * reestablish shared state.  This includes all caps issued through
2811  * this session _and_ the snap_realm hierarchy.  Because it's not
2812  * clear which snap realms the mds cares about, we send everything we
2813  * know about.. that ensures we'll then get any new info the
2814  * recovering MDS might have.
2815  *
2816  * This is a relatively heavyweight operation, but it's rare.
2817  *
2818  * called with mdsc->mutex held.
2819  */
2820 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2821 			       struct ceph_mds_session *session)
2822 {
2823 	struct ceph_msg *reply;
2824 	struct rb_node *p;
2825 	int mds = session->s_mds;
2826 	int err = -ENOMEM;
2827 	int s_nr_caps;
2828 	struct ceph_pagelist *pagelist;
2829 	struct ceph_reconnect_state recon_state;
2830 
2831 	pr_info("mds%d reconnect start\n", mds);
2832 
2833 	pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2834 	if (!pagelist)
2835 		goto fail_nopagelist;
2836 	ceph_pagelist_init(pagelist);
2837 
2838 	reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2839 	if (!reply)
2840 		goto fail_nomsg;
2841 
2842 	mutex_lock(&session->s_mutex);
2843 	session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2844 	session->s_seq = 0;
2845 
2846 	dout("session %p state %s\n", session,
2847 	     ceph_session_state_name(session->s_state));
2848 
2849 	spin_lock(&session->s_gen_ttl_lock);
2850 	session->s_cap_gen++;
2851 	spin_unlock(&session->s_gen_ttl_lock);
2852 
2853 	spin_lock(&session->s_cap_lock);
2854 	/* don't know if session is readonly */
2855 	session->s_readonly = 0;
2856 	/*
2857 	 * notify __ceph_remove_cap() that we are composing cap reconnect.
2858 	 * If a cap get released before being added to the cap reconnect,
2859 	 * __ceph_remove_cap() should skip queuing cap release.
2860 	 */
2861 	session->s_cap_reconnect = 1;
2862 	/* drop old cap expires; we're about to reestablish that state */
2863 	discard_cap_releases(mdsc, session);
2864 	spin_unlock(&session->s_cap_lock);
2865 
2866 	/* trim unused caps to reduce MDS's cache rejoin time */
2867 	shrink_dcache_parent(mdsc->fsc->sb->s_root);
2868 
2869 	ceph_con_close(&session->s_con);
2870 	ceph_con_open(&session->s_con,
2871 		      CEPH_ENTITY_TYPE_MDS, mds,
2872 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2873 
2874 	/* replay unsafe requests */
2875 	replay_unsafe_requests(mdsc, session);
2876 
2877 	down_read(&mdsc->snap_rwsem);
2878 
2879 	/* traverse this session's caps */
2880 	s_nr_caps = session->s_nr_caps;
2881 	err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2882 	if (err)
2883 		goto fail;
2884 
2885 	recon_state.nr_caps = 0;
2886 	recon_state.pagelist = pagelist;
2887 	recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2888 	err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2889 	if (err < 0)
2890 		goto fail;
2891 
2892 	spin_lock(&session->s_cap_lock);
2893 	session->s_cap_reconnect = 0;
2894 	spin_unlock(&session->s_cap_lock);
2895 
2896 	/*
2897 	 * snaprealms.  we provide mds with the ino, seq (version), and
2898 	 * parent for all of our realms.  If the mds has any newer info,
2899 	 * it will tell us.
2900 	 */
2901 	for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2902 		struct ceph_snap_realm *realm =
2903 			rb_entry(p, struct ceph_snap_realm, node);
2904 		struct ceph_mds_snaprealm_reconnect sr_rec;
2905 
2906 		dout(" adding snap realm %llx seq %lld parent %llx\n",
2907 		     realm->ino, realm->seq, realm->parent_ino);
2908 		sr_rec.ino = cpu_to_le64(realm->ino);
2909 		sr_rec.seq = cpu_to_le64(realm->seq);
2910 		sr_rec.parent = cpu_to_le64(realm->parent_ino);
2911 		err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2912 		if (err)
2913 			goto fail;
2914 	}
2915 
2916 	if (recon_state.flock)
2917 		reply->hdr.version = cpu_to_le16(2);
2918 
2919 	/* raced with cap release? */
2920 	if (s_nr_caps != recon_state.nr_caps) {
2921 		struct page *page = list_first_entry(&pagelist->head,
2922 						     struct page, lru);
2923 		__le32 *addr = kmap_atomic(page);
2924 		*addr = cpu_to_le32(recon_state.nr_caps);
2925 		kunmap_atomic(addr);
2926 	}
2927 
2928 	reply->hdr.data_len = cpu_to_le32(pagelist->length);
2929 	ceph_msg_data_add_pagelist(reply, pagelist);
2930 	ceph_con_send(&session->s_con, reply);
2931 
2932 	mutex_unlock(&session->s_mutex);
2933 
2934 	mutex_lock(&mdsc->mutex);
2935 	__wake_requests(mdsc, &session->s_waiting);
2936 	mutex_unlock(&mdsc->mutex);
2937 
2938 	up_read(&mdsc->snap_rwsem);
2939 	return;
2940 
2941 fail:
2942 	ceph_msg_put(reply);
2943 	up_read(&mdsc->snap_rwsem);
2944 	mutex_unlock(&session->s_mutex);
2945 fail_nomsg:
2946 	ceph_pagelist_release(pagelist);
2947 fail_nopagelist:
2948 	pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2949 	return;
2950 }
2951 
2952 
2953 /*
2954  * compare old and new mdsmaps, kicking requests
2955  * and closing out old connections as necessary
2956  *
2957  * called under mdsc->mutex.
2958  */
2959 static void check_new_map(struct ceph_mds_client *mdsc,
2960 			  struct ceph_mdsmap *newmap,
2961 			  struct ceph_mdsmap *oldmap)
2962 {
2963 	int i;
2964 	int oldstate, newstate;
2965 	struct ceph_mds_session *s;
2966 
2967 	dout("check_new_map new %u old %u\n",
2968 	     newmap->m_epoch, oldmap->m_epoch);
2969 
2970 	for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2971 		if (mdsc->sessions[i] == NULL)
2972 			continue;
2973 		s = mdsc->sessions[i];
2974 		oldstate = ceph_mdsmap_get_state(oldmap, i);
2975 		newstate = ceph_mdsmap_get_state(newmap, i);
2976 
2977 		dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2978 		     i, ceph_mds_state_name(oldstate),
2979 		     ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2980 		     ceph_mds_state_name(newstate),
2981 		     ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2982 		     ceph_session_state_name(s->s_state));
2983 
2984 		if (i >= newmap->m_max_mds ||
2985 		    memcmp(ceph_mdsmap_get_addr(oldmap, i),
2986 			   ceph_mdsmap_get_addr(newmap, i),
2987 			   sizeof(struct ceph_entity_addr))) {
2988 			if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2989 				/* the session never opened, just close it
2990 				 * out now */
2991 				__wake_requests(mdsc, &s->s_waiting);
2992 				__unregister_session(mdsc, s);
2993 			} else {
2994 				/* just close it */
2995 				mutex_unlock(&mdsc->mutex);
2996 				mutex_lock(&s->s_mutex);
2997 				mutex_lock(&mdsc->mutex);
2998 				ceph_con_close(&s->s_con);
2999 				mutex_unlock(&s->s_mutex);
3000 				s->s_state = CEPH_MDS_SESSION_RESTARTING;
3001 			}
3002 		} else if (oldstate == newstate) {
3003 			continue;  /* nothing new with this mds */
3004 		}
3005 
3006 		/*
3007 		 * send reconnect?
3008 		 */
3009 		if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3010 		    newstate >= CEPH_MDS_STATE_RECONNECT) {
3011 			mutex_unlock(&mdsc->mutex);
3012 			send_mds_reconnect(mdsc, s);
3013 			mutex_lock(&mdsc->mutex);
3014 		}
3015 
3016 		/*
3017 		 * kick request on any mds that has gone active.
3018 		 */
3019 		if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3020 		    newstate >= CEPH_MDS_STATE_ACTIVE) {
3021 			if (oldstate != CEPH_MDS_STATE_CREATING &&
3022 			    oldstate != CEPH_MDS_STATE_STARTING)
3023 				pr_info("mds%d recovery completed\n", s->s_mds);
3024 			kick_requests(mdsc, i);
3025 			ceph_kick_flushing_caps(mdsc, s);
3026 			wake_up_session_caps(s, 1);
3027 		}
3028 	}
3029 
3030 	for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3031 		s = mdsc->sessions[i];
3032 		if (!s)
3033 			continue;
3034 		if (!ceph_mdsmap_is_laggy(newmap, i))
3035 			continue;
3036 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3037 		    s->s_state == CEPH_MDS_SESSION_HUNG ||
3038 		    s->s_state == CEPH_MDS_SESSION_CLOSING) {
3039 			dout(" connecting to export targets of laggy mds%d\n",
3040 			     i);
3041 			__open_export_target_sessions(mdsc, s);
3042 		}
3043 	}
3044 }
3045 
3046 
3047 
3048 /*
3049  * leases
3050  */
3051 
3052 /*
3053  * caller must hold session s_mutex, dentry->d_lock
3054  */
3055 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3056 {
3057 	struct ceph_dentry_info *di = ceph_dentry(dentry);
3058 
3059 	ceph_put_mds_session(di->lease_session);
3060 	di->lease_session = NULL;
3061 }
3062 
3063 static void handle_lease(struct ceph_mds_client *mdsc,
3064 			 struct ceph_mds_session *session,
3065 			 struct ceph_msg *msg)
3066 {
3067 	struct super_block *sb = mdsc->fsc->sb;
3068 	struct inode *inode;
3069 	struct dentry *parent, *dentry;
3070 	struct ceph_dentry_info *di;
3071 	int mds = session->s_mds;
3072 	struct ceph_mds_lease *h = msg->front.iov_base;
3073 	u32 seq;
3074 	struct ceph_vino vino;
3075 	struct qstr dname;
3076 	int release = 0;
3077 
3078 	dout("handle_lease from mds%d\n", mds);
3079 
3080 	/* decode */
3081 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3082 		goto bad;
3083 	vino.ino = le64_to_cpu(h->ino);
3084 	vino.snap = CEPH_NOSNAP;
3085 	seq = le32_to_cpu(h->seq);
3086 	dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3087 	dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3088 	if (dname.len != get_unaligned_le32(h+1))
3089 		goto bad;
3090 
3091 	/* lookup inode */
3092 	inode = ceph_find_inode(sb, vino);
3093 	dout("handle_lease %s, ino %llx %p %.*s\n",
3094 	     ceph_lease_op_name(h->action), vino.ino, inode,
3095 	     dname.len, dname.name);
3096 
3097 	mutex_lock(&session->s_mutex);
3098 	session->s_seq++;
3099 
3100 	if (inode == NULL) {
3101 		dout("handle_lease no inode %llx\n", vino.ino);
3102 		goto release;
3103 	}
3104 
3105 	/* dentry */
3106 	parent = d_find_alias(inode);
3107 	if (!parent) {
3108 		dout("no parent dentry on inode %p\n", inode);
3109 		WARN_ON(1);
3110 		goto release;  /* hrm... */
3111 	}
3112 	dname.hash = full_name_hash(dname.name, dname.len);
3113 	dentry = d_lookup(parent, &dname);
3114 	dput(parent);
3115 	if (!dentry)
3116 		goto release;
3117 
3118 	spin_lock(&dentry->d_lock);
3119 	di = ceph_dentry(dentry);
3120 	switch (h->action) {
3121 	case CEPH_MDS_LEASE_REVOKE:
3122 		if (di->lease_session == session) {
3123 			if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3124 				h->seq = cpu_to_le32(di->lease_seq);
3125 			__ceph_mdsc_drop_dentry_lease(dentry);
3126 		}
3127 		release = 1;
3128 		break;
3129 
3130 	case CEPH_MDS_LEASE_RENEW:
3131 		if (di->lease_session == session &&
3132 		    di->lease_gen == session->s_cap_gen &&
3133 		    di->lease_renew_from &&
3134 		    di->lease_renew_after == 0) {
3135 			unsigned long duration =
3136 				le32_to_cpu(h->duration_ms) * HZ / 1000;
3137 
3138 			di->lease_seq = seq;
3139 			dentry->d_time = di->lease_renew_from + duration;
3140 			di->lease_renew_after = di->lease_renew_from +
3141 				(duration >> 1);
3142 			di->lease_renew_from = 0;
3143 		}
3144 		break;
3145 	}
3146 	spin_unlock(&dentry->d_lock);
3147 	dput(dentry);
3148 
3149 	if (!release)
3150 		goto out;
3151 
3152 release:
3153 	/* let's just reuse the same message */
3154 	h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3155 	ceph_msg_get(msg);
3156 	ceph_con_send(&session->s_con, msg);
3157 
3158 out:
3159 	iput(inode);
3160 	mutex_unlock(&session->s_mutex);
3161 	return;
3162 
3163 bad:
3164 	pr_err("corrupt lease message\n");
3165 	ceph_msg_dump(msg);
3166 }
3167 
3168 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3169 			      struct inode *inode,
3170 			      struct dentry *dentry, char action,
3171 			      u32 seq)
3172 {
3173 	struct ceph_msg *msg;
3174 	struct ceph_mds_lease *lease;
3175 	int len = sizeof(*lease) + sizeof(u32);
3176 	int dnamelen = 0;
3177 
3178 	dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3179 	     inode, dentry, ceph_lease_op_name(action), session->s_mds);
3180 	dnamelen = dentry->d_name.len;
3181 	len += dnamelen;
3182 
3183 	msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3184 	if (!msg)
3185 		return;
3186 	lease = msg->front.iov_base;
3187 	lease->action = action;
3188 	lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3189 	lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3190 	lease->seq = cpu_to_le32(seq);
3191 	put_unaligned_le32(dnamelen, lease + 1);
3192 	memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3193 
3194 	/*
3195 	 * if this is a preemptive lease RELEASE, no need to
3196 	 * flush request stream, since the actual request will
3197 	 * soon follow.
3198 	 */
3199 	msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3200 
3201 	ceph_con_send(&session->s_con, msg);
3202 }
3203 
3204 /*
3205  * Preemptively release a lease we expect to invalidate anyway.
3206  * Pass @inode always, @dentry is optional.
3207  */
3208 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3209 			     struct dentry *dentry)
3210 {
3211 	struct ceph_dentry_info *di;
3212 	struct ceph_mds_session *session;
3213 	u32 seq;
3214 
3215 	BUG_ON(inode == NULL);
3216 	BUG_ON(dentry == NULL);
3217 
3218 	/* is dentry lease valid? */
3219 	spin_lock(&dentry->d_lock);
3220 	di = ceph_dentry(dentry);
3221 	if (!di || !di->lease_session ||
3222 	    di->lease_session->s_mds < 0 ||
3223 	    di->lease_gen != di->lease_session->s_cap_gen ||
3224 	    !time_before(jiffies, dentry->d_time)) {
3225 		dout("lease_release inode %p dentry %p -- "
3226 		     "no lease\n",
3227 		     inode, dentry);
3228 		spin_unlock(&dentry->d_lock);
3229 		return;
3230 	}
3231 
3232 	/* we do have a lease on this dentry; note mds and seq */
3233 	session = ceph_get_mds_session(di->lease_session);
3234 	seq = di->lease_seq;
3235 	__ceph_mdsc_drop_dentry_lease(dentry);
3236 	spin_unlock(&dentry->d_lock);
3237 
3238 	dout("lease_release inode %p dentry %p to mds%d\n",
3239 	     inode, dentry, session->s_mds);
3240 	ceph_mdsc_lease_send_msg(session, inode, dentry,
3241 				 CEPH_MDS_LEASE_RELEASE, seq);
3242 	ceph_put_mds_session(session);
3243 }
3244 
3245 /*
3246  * drop all leases (and dentry refs) in preparation for umount
3247  */
3248 static void drop_leases(struct ceph_mds_client *mdsc)
3249 {
3250 	int i;
3251 
3252 	dout("drop_leases\n");
3253 	mutex_lock(&mdsc->mutex);
3254 	for (i = 0; i < mdsc->max_sessions; i++) {
3255 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3256 		if (!s)
3257 			continue;
3258 		mutex_unlock(&mdsc->mutex);
3259 		mutex_lock(&s->s_mutex);
3260 		mutex_unlock(&s->s_mutex);
3261 		ceph_put_mds_session(s);
3262 		mutex_lock(&mdsc->mutex);
3263 	}
3264 	mutex_unlock(&mdsc->mutex);
3265 }
3266 
3267 
3268 
3269 /*
3270  * delayed work -- periodically trim expired leases, renew caps with mds
3271  */
3272 static void schedule_delayed(struct ceph_mds_client *mdsc)
3273 {
3274 	int delay = 5;
3275 	unsigned hz = round_jiffies_relative(HZ * delay);
3276 	schedule_delayed_work(&mdsc->delayed_work, hz);
3277 }
3278 
3279 static void delayed_work(struct work_struct *work)
3280 {
3281 	int i;
3282 	struct ceph_mds_client *mdsc =
3283 		container_of(work, struct ceph_mds_client, delayed_work.work);
3284 	int renew_interval;
3285 	int renew_caps;
3286 
3287 	dout("mdsc delayed_work\n");
3288 	ceph_check_delayed_caps(mdsc);
3289 
3290 	mutex_lock(&mdsc->mutex);
3291 	renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3292 	renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3293 				   mdsc->last_renew_caps);
3294 	if (renew_caps)
3295 		mdsc->last_renew_caps = jiffies;
3296 
3297 	for (i = 0; i < mdsc->max_sessions; i++) {
3298 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3299 		if (s == NULL)
3300 			continue;
3301 		if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3302 			dout("resending session close request for mds%d\n",
3303 			     s->s_mds);
3304 			request_close_session(mdsc, s);
3305 			ceph_put_mds_session(s);
3306 			continue;
3307 		}
3308 		if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3309 			if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3310 				s->s_state = CEPH_MDS_SESSION_HUNG;
3311 				pr_info("mds%d hung\n", s->s_mds);
3312 			}
3313 		}
3314 		if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3315 			/* this mds is failed or recovering, just wait */
3316 			ceph_put_mds_session(s);
3317 			continue;
3318 		}
3319 		mutex_unlock(&mdsc->mutex);
3320 
3321 		mutex_lock(&s->s_mutex);
3322 		if (renew_caps)
3323 			send_renew_caps(mdsc, s);
3324 		else
3325 			ceph_con_keepalive(&s->s_con);
3326 		ceph_add_cap_releases(mdsc, s);
3327 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3328 		    s->s_state == CEPH_MDS_SESSION_HUNG)
3329 			ceph_send_cap_releases(mdsc, s);
3330 		mutex_unlock(&s->s_mutex);
3331 		ceph_put_mds_session(s);
3332 
3333 		mutex_lock(&mdsc->mutex);
3334 	}
3335 	mutex_unlock(&mdsc->mutex);
3336 
3337 	schedule_delayed(mdsc);
3338 }
3339 
3340 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3341 
3342 {
3343 	struct ceph_mds_client *mdsc;
3344 
3345 	mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3346 	if (!mdsc)
3347 		return -ENOMEM;
3348 	mdsc->fsc = fsc;
3349 	fsc->mdsc = mdsc;
3350 	mutex_init(&mdsc->mutex);
3351 	mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3352 	if (mdsc->mdsmap == NULL) {
3353 		kfree(mdsc);
3354 		return -ENOMEM;
3355 	}
3356 
3357 	init_completion(&mdsc->safe_umount_waiters);
3358 	init_waitqueue_head(&mdsc->session_close_wq);
3359 	INIT_LIST_HEAD(&mdsc->waiting_for_map);
3360 	mdsc->sessions = NULL;
3361 	atomic_set(&mdsc->num_sessions, 0);
3362 	mdsc->max_sessions = 0;
3363 	mdsc->stopping = 0;
3364 	init_rwsem(&mdsc->snap_rwsem);
3365 	mdsc->snap_realms = RB_ROOT;
3366 	INIT_LIST_HEAD(&mdsc->snap_empty);
3367 	spin_lock_init(&mdsc->snap_empty_lock);
3368 	mdsc->last_tid = 0;
3369 	mdsc->request_tree = RB_ROOT;
3370 	INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3371 	mdsc->last_renew_caps = jiffies;
3372 	INIT_LIST_HEAD(&mdsc->cap_delay_list);
3373 	spin_lock_init(&mdsc->cap_delay_lock);
3374 	INIT_LIST_HEAD(&mdsc->snap_flush_list);
3375 	spin_lock_init(&mdsc->snap_flush_lock);
3376 	mdsc->cap_flush_seq = 0;
3377 	INIT_LIST_HEAD(&mdsc->cap_dirty);
3378 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3379 	mdsc->num_cap_flushing = 0;
3380 	spin_lock_init(&mdsc->cap_dirty_lock);
3381 	init_waitqueue_head(&mdsc->cap_flushing_wq);
3382 	spin_lock_init(&mdsc->dentry_lru_lock);
3383 	INIT_LIST_HEAD(&mdsc->dentry_lru);
3384 
3385 	ceph_caps_init(mdsc);
3386 	ceph_adjust_min_caps(mdsc, fsc->min_caps);
3387 
3388 	return 0;
3389 }
3390 
3391 /*
3392  * Wait for safe replies on open mds requests.  If we time out, drop
3393  * all requests from the tree to avoid dangling dentry refs.
3394  */
3395 static void wait_requests(struct ceph_mds_client *mdsc)
3396 {
3397 	struct ceph_mds_request *req;
3398 	struct ceph_fs_client *fsc = mdsc->fsc;
3399 
3400 	mutex_lock(&mdsc->mutex);
3401 	if (__get_oldest_req(mdsc)) {
3402 		mutex_unlock(&mdsc->mutex);
3403 
3404 		dout("wait_requests waiting for requests\n");
3405 		wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3406 				    fsc->client->options->mount_timeout * HZ);
3407 
3408 		/* tear down remaining requests */
3409 		mutex_lock(&mdsc->mutex);
3410 		while ((req = __get_oldest_req(mdsc))) {
3411 			dout("wait_requests timed out on tid %llu\n",
3412 			     req->r_tid);
3413 			__unregister_request(mdsc, req);
3414 		}
3415 	}
3416 	mutex_unlock(&mdsc->mutex);
3417 	dout("wait_requests done\n");
3418 }
3419 
3420 /*
3421  * called before mount is ro, and before dentries are torn down.
3422  * (hmm, does this still race with new lookups?)
3423  */
3424 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3425 {
3426 	dout("pre_umount\n");
3427 	mdsc->stopping = 1;
3428 
3429 	drop_leases(mdsc);
3430 	ceph_flush_dirty_caps(mdsc);
3431 	wait_requests(mdsc);
3432 
3433 	/*
3434 	 * wait for reply handlers to drop their request refs and
3435 	 * their inode/dcache refs
3436 	 */
3437 	ceph_msgr_flush();
3438 }
3439 
3440 /*
3441  * wait for all write mds requests to flush.
3442  */
3443 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3444 {
3445 	struct ceph_mds_request *req = NULL, *nextreq;
3446 	struct rb_node *n;
3447 
3448 	mutex_lock(&mdsc->mutex);
3449 	dout("wait_unsafe_requests want %lld\n", want_tid);
3450 restart:
3451 	req = __get_oldest_req(mdsc);
3452 	while (req && req->r_tid <= want_tid) {
3453 		/* find next request */
3454 		n = rb_next(&req->r_node);
3455 		if (n)
3456 			nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3457 		else
3458 			nextreq = NULL;
3459 		if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3460 			/* write op */
3461 			ceph_mdsc_get_request(req);
3462 			if (nextreq)
3463 				ceph_mdsc_get_request(nextreq);
3464 			mutex_unlock(&mdsc->mutex);
3465 			dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3466 			     req->r_tid, want_tid);
3467 			wait_for_completion(&req->r_safe_completion);
3468 			mutex_lock(&mdsc->mutex);
3469 			ceph_mdsc_put_request(req);
3470 			if (!nextreq)
3471 				break;  /* next dne before, so we're done! */
3472 			if (RB_EMPTY_NODE(&nextreq->r_node)) {
3473 				/* next request was removed from tree */
3474 				ceph_mdsc_put_request(nextreq);
3475 				goto restart;
3476 			}
3477 			ceph_mdsc_put_request(nextreq);  /* won't go away */
3478 		}
3479 		req = nextreq;
3480 	}
3481 	mutex_unlock(&mdsc->mutex);
3482 	dout("wait_unsafe_requests done\n");
3483 }
3484 
3485 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3486 {
3487 	u64 want_tid, want_flush;
3488 
3489 	if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3490 		return;
3491 
3492 	dout("sync\n");
3493 	mutex_lock(&mdsc->mutex);
3494 	want_tid = mdsc->last_tid;
3495 	mutex_unlock(&mdsc->mutex);
3496 
3497 	ceph_flush_dirty_caps(mdsc);
3498 	spin_lock(&mdsc->cap_dirty_lock);
3499 	want_flush = mdsc->cap_flush_seq;
3500 	spin_unlock(&mdsc->cap_dirty_lock);
3501 
3502 	dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3503 
3504 	wait_unsafe_requests(mdsc, want_tid);
3505 	wait_caps_flush(mdsc, want_flush);
3506 }
3507 
3508 /*
3509  * true if all sessions are closed, or we force unmount
3510  */
3511 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3512 {
3513 	if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3514 		return true;
3515 	return atomic_read(&mdsc->num_sessions) == 0;
3516 }
3517 
3518 /*
3519  * called after sb is ro.
3520  */
3521 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3522 {
3523 	struct ceph_mds_session *session;
3524 	int i;
3525 	struct ceph_fs_client *fsc = mdsc->fsc;
3526 	unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3527 
3528 	dout("close_sessions\n");
3529 
3530 	/* close sessions */
3531 	mutex_lock(&mdsc->mutex);
3532 	for (i = 0; i < mdsc->max_sessions; i++) {
3533 		session = __ceph_lookup_mds_session(mdsc, i);
3534 		if (!session)
3535 			continue;
3536 		mutex_unlock(&mdsc->mutex);
3537 		mutex_lock(&session->s_mutex);
3538 		__close_session(mdsc, session);
3539 		mutex_unlock(&session->s_mutex);
3540 		ceph_put_mds_session(session);
3541 		mutex_lock(&mdsc->mutex);
3542 	}
3543 	mutex_unlock(&mdsc->mutex);
3544 
3545 	dout("waiting for sessions to close\n");
3546 	wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3547 			   timeout);
3548 
3549 	/* tear down remaining sessions */
3550 	mutex_lock(&mdsc->mutex);
3551 	for (i = 0; i < mdsc->max_sessions; i++) {
3552 		if (mdsc->sessions[i]) {
3553 			session = get_session(mdsc->sessions[i]);
3554 			__unregister_session(mdsc, session);
3555 			mutex_unlock(&mdsc->mutex);
3556 			mutex_lock(&session->s_mutex);
3557 			remove_session_caps(session);
3558 			mutex_unlock(&session->s_mutex);
3559 			ceph_put_mds_session(session);
3560 			mutex_lock(&mdsc->mutex);
3561 		}
3562 	}
3563 	WARN_ON(!list_empty(&mdsc->cap_delay_list));
3564 	mutex_unlock(&mdsc->mutex);
3565 
3566 	ceph_cleanup_empty_realms(mdsc);
3567 
3568 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3569 
3570 	dout("stopped\n");
3571 }
3572 
3573 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3574 {
3575 	dout("stop\n");
3576 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3577 	if (mdsc->mdsmap)
3578 		ceph_mdsmap_destroy(mdsc->mdsmap);
3579 	kfree(mdsc->sessions);
3580 	ceph_caps_finalize(mdsc);
3581 }
3582 
3583 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3584 {
3585 	struct ceph_mds_client *mdsc = fsc->mdsc;
3586 
3587 	dout("mdsc_destroy %p\n", mdsc);
3588 	ceph_mdsc_stop(mdsc);
3589 
3590 	/* flush out any connection work with references to us */
3591 	ceph_msgr_flush();
3592 
3593 	fsc->mdsc = NULL;
3594 	kfree(mdsc);
3595 	dout("mdsc_destroy %p done\n", mdsc);
3596 }
3597 
3598 
3599 /*
3600  * handle mds map update.
3601  */
3602 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3603 {
3604 	u32 epoch;
3605 	u32 maplen;
3606 	void *p = msg->front.iov_base;
3607 	void *end = p + msg->front.iov_len;
3608 	struct ceph_mdsmap *newmap, *oldmap;
3609 	struct ceph_fsid fsid;
3610 	int err = -EINVAL;
3611 
3612 	ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3613 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
3614 	if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3615 		return;
3616 	epoch = ceph_decode_32(&p);
3617 	maplen = ceph_decode_32(&p);
3618 	dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3619 
3620 	/* do we need it? */
3621 	ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3622 	mutex_lock(&mdsc->mutex);
3623 	if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3624 		dout("handle_map epoch %u <= our %u\n",
3625 		     epoch, mdsc->mdsmap->m_epoch);
3626 		mutex_unlock(&mdsc->mutex);
3627 		return;
3628 	}
3629 
3630 	newmap = ceph_mdsmap_decode(&p, end);
3631 	if (IS_ERR(newmap)) {
3632 		err = PTR_ERR(newmap);
3633 		goto bad_unlock;
3634 	}
3635 
3636 	/* swap into place */
3637 	if (mdsc->mdsmap) {
3638 		oldmap = mdsc->mdsmap;
3639 		mdsc->mdsmap = newmap;
3640 		check_new_map(mdsc, newmap, oldmap);
3641 		ceph_mdsmap_destroy(oldmap);
3642 	} else {
3643 		mdsc->mdsmap = newmap;  /* first mds map */
3644 	}
3645 	mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3646 
3647 	__wake_requests(mdsc, &mdsc->waiting_for_map);
3648 
3649 	mutex_unlock(&mdsc->mutex);
3650 	schedule_delayed(mdsc);
3651 	return;
3652 
3653 bad_unlock:
3654 	mutex_unlock(&mdsc->mutex);
3655 bad:
3656 	pr_err("error decoding mdsmap %d\n", err);
3657 	return;
3658 }
3659 
3660 static struct ceph_connection *con_get(struct ceph_connection *con)
3661 {
3662 	struct ceph_mds_session *s = con->private;
3663 
3664 	if (get_session(s)) {
3665 		dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3666 		return con;
3667 	}
3668 	dout("mdsc con_get %p FAIL\n", s);
3669 	return NULL;
3670 }
3671 
3672 static void con_put(struct ceph_connection *con)
3673 {
3674 	struct ceph_mds_session *s = con->private;
3675 
3676 	dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3677 	ceph_put_mds_session(s);
3678 }
3679 
3680 /*
3681  * if the client is unresponsive for long enough, the mds will kill
3682  * the session entirely.
3683  */
3684 static void peer_reset(struct ceph_connection *con)
3685 {
3686 	struct ceph_mds_session *s = con->private;
3687 	struct ceph_mds_client *mdsc = s->s_mdsc;
3688 
3689 	pr_warn("mds%d closed our session\n", s->s_mds);
3690 	send_mds_reconnect(mdsc, s);
3691 }
3692 
3693 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3694 {
3695 	struct ceph_mds_session *s = con->private;
3696 	struct ceph_mds_client *mdsc = s->s_mdsc;
3697 	int type = le16_to_cpu(msg->hdr.type);
3698 
3699 	mutex_lock(&mdsc->mutex);
3700 	if (__verify_registered_session(mdsc, s) < 0) {
3701 		mutex_unlock(&mdsc->mutex);
3702 		goto out;
3703 	}
3704 	mutex_unlock(&mdsc->mutex);
3705 
3706 	switch (type) {
3707 	case CEPH_MSG_MDS_MAP:
3708 		ceph_mdsc_handle_map(mdsc, msg);
3709 		break;
3710 	case CEPH_MSG_CLIENT_SESSION:
3711 		handle_session(s, msg);
3712 		break;
3713 	case CEPH_MSG_CLIENT_REPLY:
3714 		handle_reply(s, msg);
3715 		break;
3716 	case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3717 		handle_forward(mdsc, s, msg);
3718 		break;
3719 	case CEPH_MSG_CLIENT_CAPS:
3720 		ceph_handle_caps(s, msg);
3721 		break;
3722 	case CEPH_MSG_CLIENT_SNAP:
3723 		ceph_handle_snap(mdsc, s, msg);
3724 		break;
3725 	case CEPH_MSG_CLIENT_LEASE:
3726 		handle_lease(mdsc, s, msg);
3727 		break;
3728 
3729 	default:
3730 		pr_err("received unknown message type %d %s\n", type,
3731 		       ceph_msg_type_name(type));
3732 	}
3733 out:
3734 	ceph_msg_put(msg);
3735 }
3736 
3737 /*
3738  * authentication
3739  */
3740 
3741 /*
3742  * Note: returned pointer is the address of a structure that's
3743  * managed separately.  Caller must *not* attempt to free it.
3744  */
3745 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3746 					int *proto, int force_new)
3747 {
3748 	struct ceph_mds_session *s = con->private;
3749 	struct ceph_mds_client *mdsc = s->s_mdsc;
3750 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3751 	struct ceph_auth_handshake *auth = &s->s_auth;
3752 
3753 	if (force_new && auth->authorizer) {
3754 		ceph_auth_destroy_authorizer(ac, auth->authorizer);
3755 		auth->authorizer = NULL;
3756 	}
3757 	if (!auth->authorizer) {
3758 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3759 						      auth);
3760 		if (ret)
3761 			return ERR_PTR(ret);
3762 	} else {
3763 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3764 						      auth);
3765 		if (ret)
3766 			return ERR_PTR(ret);
3767 	}
3768 	*proto = ac->protocol;
3769 
3770 	return auth;
3771 }
3772 
3773 
3774 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3775 {
3776 	struct ceph_mds_session *s = con->private;
3777 	struct ceph_mds_client *mdsc = s->s_mdsc;
3778 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3779 
3780 	return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3781 }
3782 
3783 static int invalidate_authorizer(struct ceph_connection *con)
3784 {
3785 	struct ceph_mds_session *s = con->private;
3786 	struct ceph_mds_client *mdsc = s->s_mdsc;
3787 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3788 
3789 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3790 
3791 	return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3792 }
3793 
3794 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3795 				struct ceph_msg_header *hdr, int *skip)
3796 {
3797 	struct ceph_msg *msg;
3798 	int type = (int) le16_to_cpu(hdr->type);
3799 	int front_len = (int) le32_to_cpu(hdr->front_len);
3800 
3801 	if (con->in_msg)
3802 		return con->in_msg;
3803 
3804 	*skip = 0;
3805 	msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3806 	if (!msg) {
3807 		pr_err("unable to allocate msg type %d len %d\n",
3808 		       type, front_len);
3809 		return NULL;
3810 	}
3811 
3812 	return msg;
3813 }
3814 
3815 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3816 {
3817        struct ceph_mds_session *s = con->private;
3818        struct ceph_auth_handshake *auth = &s->s_auth;
3819        return ceph_auth_sign_message(auth, msg);
3820 }
3821 
3822 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3823 {
3824        struct ceph_mds_session *s = con->private;
3825        struct ceph_auth_handshake *auth = &s->s_auth;
3826        return ceph_auth_check_message_signature(auth, msg);
3827 }
3828 
3829 static const struct ceph_connection_operations mds_con_ops = {
3830 	.get = con_get,
3831 	.put = con_put,
3832 	.dispatch = dispatch,
3833 	.get_authorizer = get_authorizer,
3834 	.verify_authorizer_reply = verify_authorizer_reply,
3835 	.invalidate_authorizer = invalidate_authorizer,
3836 	.peer_reset = peer_reset,
3837 	.alloc_msg = mds_alloc_msg,
3838 	.sign_message = sign_message,
3839 	.check_message_signature = check_message_signature,
3840 };
3841 
3842 /* eof */
3843