xref: /linux/fs/ceph/mds_client.c (revision 0d456bad36d42d16022be045c8a53ddbb59ee478)
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/sched.h>
7 #include <linux/debugfs.h>
8 #include <linux/seq_file.h>
9 
10 #include "super.h"
11 #include "mds_client.h"
12 
13 #include <linux/ceph/ceph_features.h>
14 #include <linux/ceph/messenger.h>
15 #include <linux/ceph/decode.h>
16 #include <linux/ceph/pagelist.h>
17 #include <linux/ceph/auth.h>
18 #include <linux/ceph/debugfs.h>
19 
20 /*
21  * A cluster of MDS (metadata server) daemons is responsible for
22  * managing the file system namespace (the directory hierarchy and
23  * inodes) and for coordinating shared access to storage.  Metadata is
24  * partitioning hierarchically across a number of servers, and that
25  * partition varies over time as the cluster adjusts the distribution
26  * in order to balance load.
27  *
28  * The MDS client is primarily responsible to managing synchronous
29  * metadata requests for operations like open, unlink, and so forth.
30  * If there is a MDS failure, we find out about it when we (possibly
31  * request and) receive a new MDS map, and can resubmit affected
32  * requests.
33  *
34  * For the most part, though, we take advantage of a lossless
35  * communications channel to the MDS, and do not need to worry about
36  * timing out or resubmitting requests.
37  *
38  * We maintain a stateful "session" with each MDS we interact with.
39  * Within each session, we sent periodic heartbeat messages to ensure
40  * any capabilities or leases we have been issues remain valid.  If
41  * the session times out and goes stale, our leases and capabilities
42  * are no longer valid.
43  */
44 
45 struct ceph_reconnect_state {
46 	struct ceph_pagelist *pagelist;
47 	bool flock;
48 };
49 
50 static void __wake_requests(struct ceph_mds_client *mdsc,
51 			    struct list_head *head);
52 
53 static const struct ceph_connection_operations mds_con_ops;
54 
55 
56 /*
57  * mds reply parsing
58  */
59 
60 /*
61  * parse individual inode info
62  */
63 static int parse_reply_info_in(void **p, void *end,
64 			       struct ceph_mds_reply_info_in *info,
65 			       int features)
66 {
67 	int err = -EIO;
68 
69 	info->in = *p;
70 	*p += sizeof(struct ceph_mds_reply_inode) +
71 		sizeof(*info->in->fragtree.splits) *
72 		le32_to_cpu(info->in->fragtree.nsplits);
73 
74 	ceph_decode_32_safe(p, end, info->symlink_len, bad);
75 	ceph_decode_need(p, end, info->symlink_len, bad);
76 	info->symlink = *p;
77 	*p += info->symlink_len;
78 
79 	if (features & CEPH_FEATURE_DIRLAYOUTHASH)
80 		ceph_decode_copy_safe(p, end, &info->dir_layout,
81 				      sizeof(info->dir_layout), bad);
82 	else
83 		memset(&info->dir_layout, 0, sizeof(info->dir_layout));
84 
85 	ceph_decode_32_safe(p, end, info->xattr_len, bad);
86 	ceph_decode_need(p, end, info->xattr_len, bad);
87 	info->xattr_data = *p;
88 	*p += info->xattr_len;
89 	return 0;
90 bad:
91 	return err;
92 }
93 
94 /*
95  * parse a normal reply, which may contain a (dir+)dentry and/or a
96  * target inode.
97  */
98 static int parse_reply_info_trace(void **p, void *end,
99 				  struct ceph_mds_reply_info_parsed *info,
100 				  int features)
101 {
102 	int err;
103 
104 	if (info->head->is_dentry) {
105 		err = parse_reply_info_in(p, end, &info->diri, features);
106 		if (err < 0)
107 			goto out_bad;
108 
109 		if (unlikely(*p + sizeof(*info->dirfrag) > end))
110 			goto bad;
111 		info->dirfrag = *p;
112 		*p += sizeof(*info->dirfrag) +
113 			sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
114 		if (unlikely(*p > end))
115 			goto bad;
116 
117 		ceph_decode_32_safe(p, end, info->dname_len, bad);
118 		ceph_decode_need(p, end, info->dname_len, bad);
119 		info->dname = *p;
120 		*p += info->dname_len;
121 		info->dlease = *p;
122 		*p += sizeof(*info->dlease);
123 	}
124 
125 	if (info->head->is_target) {
126 		err = parse_reply_info_in(p, end, &info->targeti, features);
127 		if (err < 0)
128 			goto out_bad;
129 	}
130 
131 	if (unlikely(*p != end))
132 		goto bad;
133 	return 0;
134 
135 bad:
136 	err = -EIO;
137 out_bad:
138 	pr_err("problem parsing mds trace %d\n", err);
139 	return err;
140 }
141 
142 /*
143  * parse readdir results
144  */
145 static int parse_reply_info_dir(void **p, void *end,
146 				struct ceph_mds_reply_info_parsed *info,
147 				int features)
148 {
149 	u32 num, i = 0;
150 	int err;
151 
152 	info->dir_dir = *p;
153 	if (*p + sizeof(*info->dir_dir) > end)
154 		goto bad;
155 	*p += sizeof(*info->dir_dir) +
156 		sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
157 	if (*p > end)
158 		goto bad;
159 
160 	ceph_decode_need(p, end, sizeof(num) + 2, bad);
161 	num = ceph_decode_32(p);
162 	info->dir_end = ceph_decode_8(p);
163 	info->dir_complete = ceph_decode_8(p);
164 	if (num == 0)
165 		goto done;
166 
167 	/* alloc large array */
168 	info->dir_nr = num;
169 	info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
170 			       sizeof(*info->dir_dname) +
171 			       sizeof(*info->dir_dname_len) +
172 			       sizeof(*info->dir_dlease),
173 			       GFP_NOFS);
174 	if (info->dir_in == NULL) {
175 		err = -ENOMEM;
176 		goto out_bad;
177 	}
178 	info->dir_dname = (void *)(info->dir_in + num);
179 	info->dir_dname_len = (void *)(info->dir_dname + num);
180 	info->dir_dlease = (void *)(info->dir_dname_len + num);
181 
182 	while (num) {
183 		/* dentry */
184 		ceph_decode_need(p, end, sizeof(u32)*2, bad);
185 		info->dir_dname_len[i] = ceph_decode_32(p);
186 		ceph_decode_need(p, end, info->dir_dname_len[i], bad);
187 		info->dir_dname[i] = *p;
188 		*p += info->dir_dname_len[i];
189 		dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
190 		     info->dir_dname[i]);
191 		info->dir_dlease[i] = *p;
192 		*p += sizeof(struct ceph_mds_reply_lease);
193 
194 		/* inode */
195 		err = parse_reply_info_in(p, end, &info->dir_in[i], features);
196 		if (err < 0)
197 			goto out_bad;
198 		i++;
199 		num--;
200 	}
201 
202 done:
203 	if (*p != end)
204 		goto bad;
205 	return 0;
206 
207 bad:
208 	err = -EIO;
209 out_bad:
210 	pr_err("problem parsing dir contents %d\n", err);
211 	return err;
212 }
213 
214 /*
215  * parse fcntl F_GETLK results
216  */
217 static int parse_reply_info_filelock(void **p, void *end,
218 				     struct ceph_mds_reply_info_parsed *info,
219 				     int features)
220 {
221 	if (*p + sizeof(*info->filelock_reply) > end)
222 		goto bad;
223 
224 	info->filelock_reply = *p;
225 	*p += sizeof(*info->filelock_reply);
226 
227 	if (unlikely(*p != end))
228 		goto bad;
229 	return 0;
230 
231 bad:
232 	return -EIO;
233 }
234 
235 /*
236  * parse extra results
237  */
238 static int parse_reply_info_extra(void **p, void *end,
239 				  struct ceph_mds_reply_info_parsed *info,
240 				  int features)
241 {
242 	if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
243 		return parse_reply_info_filelock(p, end, info, features);
244 	else
245 		return parse_reply_info_dir(p, end, info, features);
246 }
247 
248 /*
249  * parse entire mds reply
250  */
251 static int parse_reply_info(struct ceph_msg *msg,
252 			    struct ceph_mds_reply_info_parsed *info,
253 			    int features)
254 {
255 	void *p, *end;
256 	u32 len;
257 	int err;
258 
259 	info->head = msg->front.iov_base;
260 	p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
261 	end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
262 
263 	/* trace */
264 	ceph_decode_32_safe(&p, end, len, bad);
265 	if (len > 0) {
266 		ceph_decode_need(&p, end, len, bad);
267 		err = parse_reply_info_trace(&p, p+len, info, features);
268 		if (err < 0)
269 			goto out_bad;
270 	}
271 
272 	/* extra */
273 	ceph_decode_32_safe(&p, end, len, bad);
274 	if (len > 0) {
275 		ceph_decode_need(&p, end, len, bad);
276 		err = parse_reply_info_extra(&p, p+len, info, features);
277 		if (err < 0)
278 			goto out_bad;
279 	}
280 
281 	/* snap blob */
282 	ceph_decode_32_safe(&p, end, len, bad);
283 	info->snapblob_len = len;
284 	info->snapblob = p;
285 	p += len;
286 
287 	if (p != end)
288 		goto bad;
289 	return 0;
290 
291 bad:
292 	err = -EIO;
293 out_bad:
294 	pr_err("mds parse_reply err %d\n", err);
295 	return err;
296 }
297 
298 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
299 {
300 	kfree(info->dir_in);
301 }
302 
303 
304 /*
305  * sessions
306  */
307 static const char *session_state_name(int s)
308 {
309 	switch (s) {
310 	case CEPH_MDS_SESSION_NEW: return "new";
311 	case CEPH_MDS_SESSION_OPENING: return "opening";
312 	case CEPH_MDS_SESSION_OPEN: return "open";
313 	case CEPH_MDS_SESSION_HUNG: return "hung";
314 	case CEPH_MDS_SESSION_CLOSING: return "closing";
315 	case CEPH_MDS_SESSION_RESTARTING: return "restarting";
316 	case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
317 	default: return "???";
318 	}
319 }
320 
321 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
322 {
323 	if (atomic_inc_not_zero(&s->s_ref)) {
324 		dout("mdsc get_session %p %d -> %d\n", s,
325 		     atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
326 		return s;
327 	} else {
328 		dout("mdsc get_session %p 0 -- FAIL", s);
329 		return NULL;
330 	}
331 }
332 
333 void ceph_put_mds_session(struct ceph_mds_session *s)
334 {
335 	dout("mdsc put_session %p %d -> %d\n", s,
336 	     atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
337 	if (atomic_dec_and_test(&s->s_ref)) {
338 		if (s->s_auth.authorizer)
339 		     s->s_mdsc->fsc->client->monc.auth->ops->destroy_authorizer(
340 			     s->s_mdsc->fsc->client->monc.auth,
341 			     s->s_auth.authorizer);
342 		kfree(s);
343 	}
344 }
345 
346 /*
347  * called under mdsc->mutex
348  */
349 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
350 						   int mds)
351 {
352 	struct ceph_mds_session *session;
353 
354 	if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
355 		return NULL;
356 	session = mdsc->sessions[mds];
357 	dout("lookup_mds_session %p %d\n", session,
358 	     atomic_read(&session->s_ref));
359 	get_session(session);
360 	return session;
361 }
362 
363 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
364 {
365 	if (mds >= mdsc->max_sessions)
366 		return false;
367 	return mdsc->sessions[mds];
368 }
369 
370 static int __verify_registered_session(struct ceph_mds_client *mdsc,
371 				       struct ceph_mds_session *s)
372 {
373 	if (s->s_mds >= mdsc->max_sessions ||
374 	    mdsc->sessions[s->s_mds] != s)
375 		return -ENOENT;
376 	return 0;
377 }
378 
379 /*
380  * create+register a new session for given mds.
381  * called under mdsc->mutex.
382  */
383 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
384 						 int mds)
385 {
386 	struct ceph_mds_session *s;
387 
388 	s = kzalloc(sizeof(*s), GFP_NOFS);
389 	if (!s)
390 		return ERR_PTR(-ENOMEM);
391 	s->s_mdsc = mdsc;
392 	s->s_mds = mds;
393 	s->s_state = CEPH_MDS_SESSION_NEW;
394 	s->s_ttl = 0;
395 	s->s_seq = 0;
396 	mutex_init(&s->s_mutex);
397 
398 	ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
399 
400 	spin_lock_init(&s->s_gen_ttl_lock);
401 	s->s_cap_gen = 0;
402 	s->s_cap_ttl = jiffies - 1;
403 
404 	spin_lock_init(&s->s_cap_lock);
405 	s->s_renew_requested = 0;
406 	s->s_renew_seq = 0;
407 	INIT_LIST_HEAD(&s->s_caps);
408 	s->s_nr_caps = 0;
409 	s->s_trim_caps = 0;
410 	atomic_set(&s->s_ref, 1);
411 	INIT_LIST_HEAD(&s->s_waiting);
412 	INIT_LIST_HEAD(&s->s_unsafe);
413 	s->s_num_cap_releases = 0;
414 	s->s_cap_iterator = NULL;
415 	INIT_LIST_HEAD(&s->s_cap_releases);
416 	INIT_LIST_HEAD(&s->s_cap_releases_done);
417 	INIT_LIST_HEAD(&s->s_cap_flushing);
418 	INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
419 
420 	dout("register_session mds%d\n", mds);
421 	if (mds >= mdsc->max_sessions) {
422 		int newmax = 1 << get_count_order(mds+1);
423 		struct ceph_mds_session **sa;
424 
425 		dout("register_session realloc to %d\n", newmax);
426 		sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
427 		if (sa == NULL)
428 			goto fail_realloc;
429 		if (mdsc->sessions) {
430 			memcpy(sa, mdsc->sessions,
431 			       mdsc->max_sessions * sizeof(void *));
432 			kfree(mdsc->sessions);
433 		}
434 		mdsc->sessions = sa;
435 		mdsc->max_sessions = newmax;
436 	}
437 	mdsc->sessions[mds] = s;
438 	atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
439 
440 	ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
441 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
442 
443 	return s;
444 
445 fail_realloc:
446 	kfree(s);
447 	return ERR_PTR(-ENOMEM);
448 }
449 
450 /*
451  * called under mdsc->mutex
452  */
453 static void __unregister_session(struct ceph_mds_client *mdsc,
454 			       struct ceph_mds_session *s)
455 {
456 	dout("__unregister_session mds%d %p\n", s->s_mds, s);
457 	BUG_ON(mdsc->sessions[s->s_mds] != s);
458 	mdsc->sessions[s->s_mds] = NULL;
459 	ceph_con_close(&s->s_con);
460 	ceph_put_mds_session(s);
461 }
462 
463 /*
464  * drop session refs in request.
465  *
466  * should be last request ref, or hold mdsc->mutex
467  */
468 static void put_request_session(struct ceph_mds_request *req)
469 {
470 	if (req->r_session) {
471 		ceph_put_mds_session(req->r_session);
472 		req->r_session = NULL;
473 	}
474 }
475 
476 void ceph_mdsc_release_request(struct kref *kref)
477 {
478 	struct ceph_mds_request *req = container_of(kref,
479 						    struct ceph_mds_request,
480 						    r_kref);
481 	if (req->r_request)
482 		ceph_msg_put(req->r_request);
483 	if (req->r_reply) {
484 		ceph_msg_put(req->r_reply);
485 		destroy_reply_info(&req->r_reply_info);
486 	}
487 	if (req->r_inode) {
488 		ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
489 		iput(req->r_inode);
490 	}
491 	if (req->r_locked_dir)
492 		ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
493 	if (req->r_target_inode)
494 		iput(req->r_target_inode);
495 	if (req->r_dentry)
496 		dput(req->r_dentry);
497 	if (req->r_old_dentry) {
498 		/*
499 		 * track (and drop pins for) r_old_dentry_dir
500 		 * separately, since r_old_dentry's d_parent may have
501 		 * changed between the dir mutex being dropped and
502 		 * this request being freed.
503 		 */
504 		ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
505 				  CEPH_CAP_PIN);
506 		dput(req->r_old_dentry);
507 		iput(req->r_old_dentry_dir);
508 	}
509 	kfree(req->r_path1);
510 	kfree(req->r_path2);
511 	put_request_session(req);
512 	ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
513 	kfree(req);
514 }
515 
516 /*
517  * lookup session, bump ref if found.
518  *
519  * called under mdsc->mutex.
520  */
521 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
522 					     u64 tid)
523 {
524 	struct ceph_mds_request *req;
525 	struct rb_node *n = mdsc->request_tree.rb_node;
526 
527 	while (n) {
528 		req = rb_entry(n, struct ceph_mds_request, r_node);
529 		if (tid < req->r_tid)
530 			n = n->rb_left;
531 		else if (tid > req->r_tid)
532 			n = n->rb_right;
533 		else {
534 			ceph_mdsc_get_request(req);
535 			return req;
536 		}
537 	}
538 	return NULL;
539 }
540 
541 static void __insert_request(struct ceph_mds_client *mdsc,
542 			     struct ceph_mds_request *new)
543 {
544 	struct rb_node **p = &mdsc->request_tree.rb_node;
545 	struct rb_node *parent = NULL;
546 	struct ceph_mds_request *req = NULL;
547 
548 	while (*p) {
549 		parent = *p;
550 		req = rb_entry(parent, struct ceph_mds_request, r_node);
551 		if (new->r_tid < req->r_tid)
552 			p = &(*p)->rb_left;
553 		else if (new->r_tid > req->r_tid)
554 			p = &(*p)->rb_right;
555 		else
556 			BUG();
557 	}
558 
559 	rb_link_node(&new->r_node, parent, p);
560 	rb_insert_color(&new->r_node, &mdsc->request_tree);
561 }
562 
563 /*
564  * Register an in-flight request, and assign a tid.  Link to directory
565  * are modifying (if any).
566  *
567  * Called under mdsc->mutex.
568  */
569 static void __register_request(struct ceph_mds_client *mdsc,
570 			       struct ceph_mds_request *req,
571 			       struct inode *dir)
572 {
573 	req->r_tid = ++mdsc->last_tid;
574 	if (req->r_num_caps)
575 		ceph_reserve_caps(mdsc, &req->r_caps_reservation,
576 				  req->r_num_caps);
577 	dout("__register_request %p tid %lld\n", req, req->r_tid);
578 	ceph_mdsc_get_request(req);
579 	__insert_request(mdsc, req);
580 
581 	req->r_uid = current_fsuid();
582 	req->r_gid = current_fsgid();
583 
584 	if (dir) {
585 		struct ceph_inode_info *ci = ceph_inode(dir);
586 
587 		ihold(dir);
588 		spin_lock(&ci->i_unsafe_lock);
589 		req->r_unsafe_dir = dir;
590 		list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
591 		spin_unlock(&ci->i_unsafe_lock);
592 	}
593 }
594 
595 static void __unregister_request(struct ceph_mds_client *mdsc,
596 				 struct ceph_mds_request *req)
597 {
598 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
599 	rb_erase(&req->r_node, &mdsc->request_tree);
600 	RB_CLEAR_NODE(&req->r_node);
601 
602 	if (req->r_unsafe_dir) {
603 		struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
604 
605 		spin_lock(&ci->i_unsafe_lock);
606 		list_del_init(&req->r_unsafe_dir_item);
607 		spin_unlock(&ci->i_unsafe_lock);
608 
609 		iput(req->r_unsafe_dir);
610 		req->r_unsafe_dir = NULL;
611 	}
612 
613 	ceph_mdsc_put_request(req);
614 }
615 
616 /*
617  * Choose mds to send request to next.  If there is a hint set in the
618  * request (e.g., due to a prior forward hint from the mds), use that.
619  * Otherwise, consult frag tree and/or caps to identify the
620  * appropriate mds.  If all else fails, choose randomly.
621  *
622  * Called under mdsc->mutex.
623  */
624 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
625 {
626 	/*
627 	 * we don't need to worry about protecting the d_parent access
628 	 * here because we never renaming inside the snapped namespace
629 	 * except to resplice to another snapdir, and either the old or new
630 	 * result is a valid result.
631 	 */
632 	while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
633 		dentry = dentry->d_parent;
634 	return dentry;
635 }
636 
637 static int __choose_mds(struct ceph_mds_client *mdsc,
638 			struct ceph_mds_request *req)
639 {
640 	struct inode *inode;
641 	struct ceph_inode_info *ci;
642 	struct ceph_cap *cap;
643 	int mode = req->r_direct_mode;
644 	int mds = -1;
645 	u32 hash = req->r_direct_hash;
646 	bool is_hash = req->r_direct_is_hash;
647 
648 	/*
649 	 * is there a specific mds we should try?  ignore hint if we have
650 	 * no session and the mds is not up (active or recovering).
651 	 */
652 	if (req->r_resend_mds >= 0 &&
653 	    (__have_session(mdsc, req->r_resend_mds) ||
654 	     ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
655 		dout("choose_mds using resend_mds mds%d\n",
656 		     req->r_resend_mds);
657 		return req->r_resend_mds;
658 	}
659 
660 	if (mode == USE_RANDOM_MDS)
661 		goto random;
662 
663 	inode = NULL;
664 	if (req->r_inode) {
665 		inode = req->r_inode;
666 	} else if (req->r_dentry) {
667 		/* ignore race with rename; old or new d_parent is okay */
668 		struct dentry *parent = req->r_dentry->d_parent;
669 		struct inode *dir = parent->d_inode;
670 
671 		if (dir->i_sb != mdsc->fsc->sb) {
672 			/* not this fs! */
673 			inode = req->r_dentry->d_inode;
674 		} else if (ceph_snap(dir) != CEPH_NOSNAP) {
675 			/* direct snapped/virtual snapdir requests
676 			 * based on parent dir inode */
677 			struct dentry *dn = get_nonsnap_parent(parent);
678 			inode = dn->d_inode;
679 			dout("__choose_mds using nonsnap parent %p\n", inode);
680 		} else if (req->r_dentry->d_inode) {
681 			/* dentry target */
682 			inode = req->r_dentry->d_inode;
683 		} else {
684 			/* dir + name */
685 			inode = dir;
686 			hash = ceph_dentry_hash(dir, req->r_dentry);
687 			is_hash = true;
688 		}
689 	}
690 
691 	dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
692 	     (int)hash, mode);
693 	if (!inode)
694 		goto random;
695 	ci = ceph_inode(inode);
696 
697 	if (is_hash && S_ISDIR(inode->i_mode)) {
698 		struct ceph_inode_frag frag;
699 		int found;
700 
701 		ceph_choose_frag(ci, hash, &frag, &found);
702 		if (found) {
703 			if (mode == USE_ANY_MDS && frag.ndist > 0) {
704 				u8 r;
705 
706 				/* choose a random replica */
707 				get_random_bytes(&r, 1);
708 				r %= frag.ndist;
709 				mds = frag.dist[r];
710 				dout("choose_mds %p %llx.%llx "
711 				     "frag %u mds%d (%d/%d)\n",
712 				     inode, ceph_vinop(inode),
713 				     frag.frag, mds,
714 				     (int)r, frag.ndist);
715 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
716 				    CEPH_MDS_STATE_ACTIVE)
717 					return mds;
718 			}
719 
720 			/* since this file/dir wasn't known to be
721 			 * replicated, then we want to look for the
722 			 * authoritative mds. */
723 			mode = USE_AUTH_MDS;
724 			if (frag.mds >= 0) {
725 				/* choose auth mds */
726 				mds = frag.mds;
727 				dout("choose_mds %p %llx.%llx "
728 				     "frag %u mds%d (auth)\n",
729 				     inode, ceph_vinop(inode), frag.frag, mds);
730 				if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
731 				    CEPH_MDS_STATE_ACTIVE)
732 					return mds;
733 			}
734 		}
735 	}
736 
737 	spin_lock(&ci->i_ceph_lock);
738 	cap = NULL;
739 	if (mode == USE_AUTH_MDS)
740 		cap = ci->i_auth_cap;
741 	if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
742 		cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
743 	if (!cap) {
744 		spin_unlock(&ci->i_ceph_lock);
745 		goto random;
746 	}
747 	mds = cap->session->s_mds;
748 	dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
749 	     inode, ceph_vinop(inode), mds,
750 	     cap == ci->i_auth_cap ? "auth " : "", cap);
751 	spin_unlock(&ci->i_ceph_lock);
752 	return mds;
753 
754 random:
755 	mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
756 	dout("choose_mds chose random mds%d\n", mds);
757 	return mds;
758 }
759 
760 
761 /*
762  * session messages
763  */
764 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
765 {
766 	struct ceph_msg *msg;
767 	struct ceph_mds_session_head *h;
768 
769 	msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
770 			   false);
771 	if (!msg) {
772 		pr_err("create_session_msg ENOMEM creating msg\n");
773 		return NULL;
774 	}
775 	h = msg->front.iov_base;
776 	h->op = cpu_to_le32(op);
777 	h->seq = cpu_to_le64(seq);
778 	return msg;
779 }
780 
781 /*
782  * send session open request.
783  *
784  * called under mdsc->mutex
785  */
786 static int __open_session(struct ceph_mds_client *mdsc,
787 			  struct ceph_mds_session *session)
788 {
789 	struct ceph_msg *msg;
790 	int mstate;
791 	int mds = session->s_mds;
792 
793 	/* wait for mds to go active? */
794 	mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
795 	dout("open_session to mds%d (%s)\n", mds,
796 	     ceph_mds_state_name(mstate));
797 	session->s_state = CEPH_MDS_SESSION_OPENING;
798 	session->s_renew_requested = jiffies;
799 
800 	/* send connect message */
801 	msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
802 	if (!msg)
803 		return -ENOMEM;
804 	ceph_con_send(&session->s_con, msg);
805 	return 0;
806 }
807 
808 /*
809  * open sessions for any export targets for the given mds
810  *
811  * called under mdsc->mutex
812  */
813 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
814 					  struct ceph_mds_session *session)
815 {
816 	struct ceph_mds_info *mi;
817 	struct ceph_mds_session *ts;
818 	int i, mds = session->s_mds;
819 	int target;
820 
821 	if (mds >= mdsc->mdsmap->m_max_mds)
822 		return;
823 	mi = &mdsc->mdsmap->m_info[mds];
824 	dout("open_export_target_sessions for mds%d (%d targets)\n",
825 	     session->s_mds, mi->num_export_targets);
826 
827 	for (i = 0; i < mi->num_export_targets; i++) {
828 		target = mi->export_targets[i];
829 		ts = __ceph_lookup_mds_session(mdsc, target);
830 		if (!ts) {
831 			ts = register_session(mdsc, target);
832 			if (IS_ERR(ts))
833 				return;
834 		}
835 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
836 		    session->s_state == CEPH_MDS_SESSION_CLOSING)
837 			__open_session(mdsc, session);
838 		else
839 			dout(" mds%d target mds%d %p is %s\n", session->s_mds,
840 			     i, ts, session_state_name(ts->s_state));
841 		ceph_put_mds_session(ts);
842 	}
843 }
844 
845 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
846 					   struct ceph_mds_session *session)
847 {
848 	mutex_lock(&mdsc->mutex);
849 	__open_export_target_sessions(mdsc, session);
850 	mutex_unlock(&mdsc->mutex);
851 }
852 
853 /*
854  * session caps
855  */
856 
857 /*
858  * Free preallocated cap messages assigned to this session
859  */
860 static void cleanup_cap_releases(struct ceph_mds_session *session)
861 {
862 	struct ceph_msg *msg;
863 
864 	spin_lock(&session->s_cap_lock);
865 	while (!list_empty(&session->s_cap_releases)) {
866 		msg = list_first_entry(&session->s_cap_releases,
867 				       struct ceph_msg, list_head);
868 		list_del_init(&msg->list_head);
869 		ceph_msg_put(msg);
870 	}
871 	while (!list_empty(&session->s_cap_releases_done)) {
872 		msg = list_first_entry(&session->s_cap_releases_done,
873 				       struct ceph_msg, list_head);
874 		list_del_init(&msg->list_head);
875 		ceph_msg_put(msg);
876 	}
877 	spin_unlock(&session->s_cap_lock);
878 }
879 
880 /*
881  * Helper to safely iterate over all caps associated with a session, with
882  * special care taken to handle a racing __ceph_remove_cap().
883  *
884  * Caller must hold session s_mutex.
885  */
886 static int iterate_session_caps(struct ceph_mds_session *session,
887 				 int (*cb)(struct inode *, struct ceph_cap *,
888 					    void *), void *arg)
889 {
890 	struct list_head *p;
891 	struct ceph_cap *cap;
892 	struct inode *inode, *last_inode = NULL;
893 	struct ceph_cap *old_cap = NULL;
894 	int ret;
895 
896 	dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
897 	spin_lock(&session->s_cap_lock);
898 	p = session->s_caps.next;
899 	while (p != &session->s_caps) {
900 		cap = list_entry(p, struct ceph_cap, session_caps);
901 		inode = igrab(&cap->ci->vfs_inode);
902 		if (!inode) {
903 			p = p->next;
904 			continue;
905 		}
906 		session->s_cap_iterator = cap;
907 		spin_unlock(&session->s_cap_lock);
908 
909 		if (last_inode) {
910 			iput(last_inode);
911 			last_inode = NULL;
912 		}
913 		if (old_cap) {
914 			ceph_put_cap(session->s_mdsc, old_cap);
915 			old_cap = NULL;
916 		}
917 
918 		ret = cb(inode, cap, arg);
919 		last_inode = inode;
920 
921 		spin_lock(&session->s_cap_lock);
922 		p = p->next;
923 		if (cap->ci == NULL) {
924 			dout("iterate_session_caps  finishing cap %p removal\n",
925 			     cap);
926 			BUG_ON(cap->session != session);
927 			list_del_init(&cap->session_caps);
928 			session->s_nr_caps--;
929 			cap->session = NULL;
930 			old_cap = cap;  /* put_cap it w/o locks held */
931 		}
932 		if (ret < 0)
933 			goto out;
934 	}
935 	ret = 0;
936 out:
937 	session->s_cap_iterator = NULL;
938 	spin_unlock(&session->s_cap_lock);
939 
940 	if (last_inode)
941 		iput(last_inode);
942 	if (old_cap)
943 		ceph_put_cap(session->s_mdsc, old_cap);
944 
945 	return ret;
946 }
947 
948 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
949 				  void *arg)
950 {
951 	struct ceph_inode_info *ci = ceph_inode(inode);
952 	int drop = 0;
953 
954 	dout("removing cap %p, ci is %p, inode is %p\n",
955 	     cap, ci, &ci->vfs_inode);
956 	spin_lock(&ci->i_ceph_lock);
957 	__ceph_remove_cap(cap);
958 	if (!__ceph_is_any_real_caps(ci)) {
959 		struct ceph_mds_client *mdsc =
960 			ceph_sb_to_client(inode->i_sb)->mdsc;
961 
962 		spin_lock(&mdsc->cap_dirty_lock);
963 		if (!list_empty(&ci->i_dirty_item)) {
964 			pr_info(" dropping dirty %s state for %p %lld\n",
965 				ceph_cap_string(ci->i_dirty_caps),
966 				inode, ceph_ino(inode));
967 			ci->i_dirty_caps = 0;
968 			list_del_init(&ci->i_dirty_item);
969 			drop = 1;
970 		}
971 		if (!list_empty(&ci->i_flushing_item)) {
972 			pr_info(" dropping dirty+flushing %s state for %p %lld\n",
973 				ceph_cap_string(ci->i_flushing_caps),
974 				inode, ceph_ino(inode));
975 			ci->i_flushing_caps = 0;
976 			list_del_init(&ci->i_flushing_item);
977 			mdsc->num_cap_flushing--;
978 			drop = 1;
979 		}
980 		if (drop && ci->i_wrbuffer_ref) {
981 			pr_info(" dropping dirty data for %p %lld\n",
982 				inode, ceph_ino(inode));
983 			ci->i_wrbuffer_ref = 0;
984 			ci->i_wrbuffer_ref_head = 0;
985 			drop++;
986 		}
987 		spin_unlock(&mdsc->cap_dirty_lock);
988 	}
989 	spin_unlock(&ci->i_ceph_lock);
990 	while (drop--)
991 		iput(inode);
992 	return 0;
993 }
994 
995 /*
996  * caller must hold session s_mutex
997  */
998 static void remove_session_caps(struct ceph_mds_session *session)
999 {
1000 	dout("remove_session_caps on %p\n", session);
1001 	iterate_session_caps(session, remove_session_caps_cb, NULL);
1002 	BUG_ON(session->s_nr_caps > 0);
1003 	BUG_ON(!list_empty(&session->s_cap_flushing));
1004 	cleanup_cap_releases(session);
1005 }
1006 
1007 /*
1008  * wake up any threads waiting on this session's caps.  if the cap is
1009  * old (didn't get renewed on the client reconnect), remove it now.
1010  *
1011  * caller must hold s_mutex.
1012  */
1013 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1014 			      void *arg)
1015 {
1016 	struct ceph_inode_info *ci = ceph_inode(inode);
1017 
1018 	wake_up_all(&ci->i_cap_wq);
1019 	if (arg) {
1020 		spin_lock(&ci->i_ceph_lock);
1021 		ci->i_wanted_max_size = 0;
1022 		ci->i_requested_max_size = 0;
1023 		spin_unlock(&ci->i_ceph_lock);
1024 	}
1025 	return 0;
1026 }
1027 
1028 static void wake_up_session_caps(struct ceph_mds_session *session,
1029 				 int reconnect)
1030 {
1031 	dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1032 	iterate_session_caps(session, wake_up_session_cb,
1033 			     (void *)(unsigned long)reconnect);
1034 }
1035 
1036 /*
1037  * Send periodic message to MDS renewing all currently held caps.  The
1038  * ack will reset the expiration for all caps from this session.
1039  *
1040  * caller holds s_mutex
1041  */
1042 static int send_renew_caps(struct ceph_mds_client *mdsc,
1043 			   struct ceph_mds_session *session)
1044 {
1045 	struct ceph_msg *msg;
1046 	int state;
1047 
1048 	if (time_after_eq(jiffies, session->s_cap_ttl) &&
1049 	    time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1050 		pr_info("mds%d caps stale\n", session->s_mds);
1051 	session->s_renew_requested = jiffies;
1052 
1053 	/* do not try to renew caps until a recovering mds has reconnected
1054 	 * with its clients. */
1055 	state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1056 	if (state < CEPH_MDS_STATE_RECONNECT) {
1057 		dout("send_renew_caps ignoring mds%d (%s)\n",
1058 		     session->s_mds, ceph_mds_state_name(state));
1059 		return 0;
1060 	}
1061 
1062 	dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1063 		ceph_mds_state_name(state));
1064 	msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1065 				 ++session->s_renew_seq);
1066 	if (!msg)
1067 		return -ENOMEM;
1068 	ceph_con_send(&session->s_con, msg);
1069 	return 0;
1070 }
1071 
1072 /*
1073  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1074  *
1075  * Called under session->s_mutex
1076  */
1077 static void renewed_caps(struct ceph_mds_client *mdsc,
1078 			 struct ceph_mds_session *session, int is_renew)
1079 {
1080 	int was_stale;
1081 	int wake = 0;
1082 
1083 	spin_lock(&session->s_cap_lock);
1084 	was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1085 
1086 	session->s_cap_ttl = session->s_renew_requested +
1087 		mdsc->mdsmap->m_session_timeout*HZ;
1088 
1089 	if (was_stale) {
1090 		if (time_before(jiffies, session->s_cap_ttl)) {
1091 			pr_info("mds%d caps renewed\n", session->s_mds);
1092 			wake = 1;
1093 		} else {
1094 			pr_info("mds%d caps still stale\n", session->s_mds);
1095 		}
1096 	}
1097 	dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1098 	     session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1099 	     time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1100 	spin_unlock(&session->s_cap_lock);
1101 
1102 	if (wake)
1103 		wake_up_session_caps(session, 0);
1104 }
1105 
1106 /*
1107  * send a session close request
1108  */
1109 static int request_close_session(struct ceph_mds_client *mdsc,
1110 				 struct ceph_mds_session *session)
1111 {
1112 	struct ceph_msg *msg;
1113 
1114 	dout("request_close_session mds%d state %s seq %lld\n",
1115 	     session->s_mds, session_state_name(session->s_state),
1116 	     session->s_seq);
1117 	msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1118 	if (!msg)
1119 		return -ENOMEM;
1120 	ceph_con_send(&session->s_con, msg);
1121 	return 0;
1122 }
1123 
1124 /*
1125  * Called with s_mutex held.
1126  */
1127 static int __close_session(struct ceph_mds_client *mdsc,
1128 			 struct ceph_mds_session *session)
1129 {
1130 	if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1131 		return 0;
1132 	session->s_state = CEPH_MDS_SESSION_CLOSING;
1133 	return request_close_session(mdsc, session);
1134 }
1135 
1136 /*
1137  * Trim old(er) caps.
1138  *
1139  * Because we can't cache an inode without one or more caps, we do
1140  * this indirectly: if a cap is unused, we prune its aliases, at which
1141  * point the inode will hopefully get dropped to.
1142  *
1143  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1144  * memory pressure from the MDS, though, so it needn't be perfect.
1145  */
1146 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1147 {
1148 	struct ceph_mds_session *session = arg;
1149 	struct ceph_inode_info *ci = ceph_inode(inode);
1150 	int used, oissued, mine;
1151 
1152 	if (session->s_trim_caps <= 0)
1153 		return -1;
1154 
1155 	spin_lock(&ci->i_ceph_lock);
1156 	mine = cap->issued | cap->implemented;
1157 	used = __ceph_caps_used(ci);
1158 	oissued = __ceph_caps_issued_other(ci, cap);
1159 
1160 	dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1161 	     inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1162 	     ceph_cap_string(used));
1163 	if (ci->i_dirty_caps)
1164 		goto out;   /* dirty caps */
1165 	if ((used & ~oissued) & mine)
1166 		goto out;   /* we need these caps */
1167 
1168 	session->s_trim_caps--;
1169 	if (oissued) {
1170 		/* we aren't the only cap.. just remove us */
1171 		__ceph_remove_cap(cap);
1172 	} else {
1173 		/* try to drop referring dentries */
1174 		spin_unlock(&ci->i_ceph_lock);
1175 		d_prune_aliases(inode);
1176 		dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1177 		     inode, cap, atomic_read(&inode->i_count));
1178 		return 0;
1179 	}
1180 
1181 out:
1182 	spin_unlock(&ci->i_ceph_lock);
1183 	return 0;
1184 }
1185 
1186 /*
1187  * Trim session cap count down to some max number.
1188  */
1189 static int trim_caps(struct ceph_mds_client *mdsc,
1190 		     struct ceph_mds_session *session,
1191 		     int max_caps)
1192 {
1193 	int trim_caps = session->s_nr_caps - max_caps;
1194 
1195 	dout("trim_caps mds%d start: %d / %d, trim %d\n",
1196 	     session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1197 	if (trim_caps > 0) {
1198 		session->s_trim_caps = trim_caps;
1199 		iterate_session_caps(session, trim_caps_cb, session);
1200 		dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1201 		     session->s_mds, session->s_nr_caps, max_caps,
1202 			trim_caps - session->s_trim_caps);
1203 		session->s_trim_caps = 0;
1204 	}
1205 	return 0;
1206 }
1207 
1208 /*
1209  * Allocate cap_release messages.  If there is a partially full message
1210  * in the queue, try to allocate enough to cover it's remainder, so that
1211  * we can send it immediately.
1212  *
1213  * Called under s_mutex.
1214  */
1215 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1216 			  struct ceph_mds_session *session)
1217 {
1218 	struct ceph_msg *msg, *partial = NULL;
1219 	struct ceph_mds_cap_release *head;
1220 	int err = -ENOMEM;
1221 	int extra = mdsc->fsc->mount_options->cap_release_safety;
1222 	int num;
1223 
1224 	dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1225 	     extra);
1226 
1227 	spin_lock(&session->s_cap_lock);
1228 
1229 	if (!list_empty(&session->s_cap_releases)) {
1230 		msg = list_first_entry(&session->s_cap_releases,
1231 				       struct ceph_msg,
1232 				 list_head);
1233 		head = msg->front.iov_base;
1234 		num = le32_to_cpu(head->num);
1235 		if (num) {
1236 			dout(" partial %p with (%d/%d)\n", msg, num,
1237 			     (int)CEPH_CAPS_PER_RELEASE);
1238 			extra += CEPH_CAPS_PER_RELEASE - num;
1239 			partial = msg;
1240 		}
1241 	}
1242 	while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1243 		spin_unlock(&session->s_cap_lock);
1244 		msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1245 				   GFP_NOFS, false);
1246 		if (!msg)
1247 			goto out_unlocked;
1248 		dout("add_cap_releases %p msg %p now %d\n", session, msg,
1249 		     (int)msg->front.iov_len);
1250 		head = msg->front.iov_base;
1251 		head->num = cpu_to_le32(0);
1252 		msg->front.iov_len = sizeof(*head);
1253 		spin_lock(&session->s_cap_lock);
1254 		list_add(&msg->list_head, &session->s_cap_releases);
1255 		session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1256 	}
1257 
1258 	if (partial) {
1259 		head = partial->front.iov_base;
1260 		num = le32_to_cpu(head->num);
1261 		dout(" queueing partial %p with %d/%d\n", partial, num,
1262 		     (int)CEPH_CAPS_PER_RELEASE);
1263 		list_move_tail(&partial->list_head,
1264 			       &session->s_cap_releases_done);
1265 		session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1266 	}
1267 	err = 0;
1268 	spin_unlock(&session->s_cap_lock);
1269 out_unlocked:
1270 	return err;
1271 }
1272 
1273 /*
1274  * flush all dirty inode data to disk.
1275  *
1276  * returns true if we've flushed through want_flush_seq
1277  */
1278 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1279 {
1280 	int mds, ret = 1;
1281 
1282 	dout("check_cap_flush want %lld\n", want_flush_seq);
1283 	mutex_lock(&mdsc->mutex);
1284 	for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1285 		struct ceph_mds_session *session = mdsc->sessions[mds];
1286 
1287 		if (!session)
1288 			continue;
1289 		get_session(session);
1290 		mutex_unlock(&mdsc->mutex);
1291 
1292 		mutex_lock(&session->s_mutex);
1293 		if (!list_empty(&session->s_cap_flushing)) {
1294 			struct ceph_inode_info *ci =
1295 				list_entry(session->s_cap_flushing.next,
1296 					   struct ceph_inode_info,
1297 					   i_flushing_item);
1298 			struct inode *inode = &ci->vfs_inode;
1299 
1300 			spin_lock(&ci->i_ceph_lock);
1301 			if (ci->i_cap_flush_seq <= want_flush_seq) {
1302 				dout("check_cap_flush still flushing %p "
1303 				     "seq %lld <= %lld to mds%d\n", inode,
1304 				     ci->i_cap_flush_seq, want_flush_seq,
1305 				     session->s_mds);
1306 				ret = 0;
1307 			}
1308 			spin_unlock(&ci->i_ceph_lock);
1309 		}
1310 		mutex_unlock(&session->s_mutex);
1311 		ceph_put_mds_session(session);
1312 
1313 		if (!ret)
1314 			return ret;
1315 		mutex_lock(&mdsc->mutex);
1316 	}
1317 
1318 	mutex_unlock(&mdsc->mutex);
1319 	dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1320 	return ret;
1321 }
1322 
1323 /*
1324  * called under s_mutex
1325  */
1326 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1327 			    struct ceph_mds_session *session)
1328 {
1329 	struct ceph_msg *msg;
1330 
1331 	dout("send_cap_releases mds%d\n", session->s_mds);
1332 	spin_lock(&session->s_cap_lock);
1333 	while (!list_empty(&session->s_cap_releases_done)) {
1334 		msg = list_first_entry(&session->s_cap_releases_done,
1335 				 struct ceph_msg, list_head);
1336 		list_del_init(&msg->list_head);
1337 		spin_unlock(&session->s_cap_lock);
1338 		msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1339 		dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1340 		ceph_con_send(&session->s_con, msg);
1341 		spin_lock(&session->s_cap_lock);
1342 	}
1343 	spin_unlock(&session->s_cap_lock);
1344 }
1345 
1346 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1347 				 struct ceph_mds_session *session)
1348 {
1349 	struct ceph_msg *msg;
1350 	struct ceph_mds_cap_release *head;
1351 	unsigned num;
1352 
1353 	dout("discard_cap_releases mds%d\n", session->s_mds);
1354 	spin_lock(&session->s_cap_lock);
1355 
1356 	/* zero out the in-progress message */
1357 	msg = list_first_entry(&session->s_cap_releases,
1358 			       struct ceph_msg, list_head);
1359 	head = msg->front.iov_base;
1360 	num = le32_to_cpu(head->num);
1361 	dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1362 	head->num = cpu_to_le32(0);
1363 	session->s_num_cap_releases += num;
1364 
1365 	/* requeue completed messages */
1366 	while (!list_empty(&session->s_cap_releases_done)) {
1367 		msg = list_first_entry(&session->s_cap_releases_done,
1368 				 struct ceph_msg, list_head);
1369 		list_del_init(&msg->list_head);
1370 
1371 		head = msg->front.iov_base;
1372 		num = le32_to_cpu(head->num);
1373 		dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1374 		     num);
1375 		session->s_num_cap_releases += num;
1376 		head->num = cpu_to_le32(0);
1377 		msg->front.iov_len = sizeof(*head);
1378 		list_add(&msg->list_head, &session->s_cap_releases);
1379 	}
1380 
1381 	spin_unlock(&session->s_cap_lock);
1382 }
1383 
1384 /*
1385  * requests
1386  */
1387 
1388 /*
1389  * Create an mds request.
1390  */
1391 struct ceph_mds_request *
1392 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1393 {
1394 	struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1395 
1396 	if (!req)
1397 		return ERR_PTR(-ENOMEM);
1398 
1399 	mutex_init(&req->r_fill_mutex);
1400 	req->r_mdsc = mdsc;
1401 	req->r_started = jiffies;
1402 	req->r_resend_mds = -1;
1403 	INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1404 	req->r_fmode = -1;
1405 	kref_init(&req->r_kref);
1406 	INIT_LIST_HEAD(&req->r_wait);
1407 	init_completion(&req->r_completion);
1408 	init_completion(&req->r_safe_completion);
1409 	INIT_LIST_HEAD(&req->r_unsafe_item);
1410 
1411 	req->r_op = op;
1412 	req->r_direct_mode = mode;
1413 	return req;
1414 }
1415 
1416 /*
1417  * return oldest (lowest) request, tid in request tree, 0 if none.
1418  *
1419  * called under mdsc->mutex.
1420  */
1421 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1422 {
1423 	if (RB_EMPTY_ROOT(&mdsc->request_tree))
1424 		return NULL;
1425 	return rb_entry(rb_first(&mdsc->request_tree),
1426 			struct ceph_mds_request, r_node);
1427 }
1428 
1429 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1430 {
1431 	struct ceph_mds_request *req = __get_oldest_req(mdsc);
1432 
1433 	if (req)
1434 		return req->r_tid;
1435 	return 0;
1436 }
1437 
1438 /*
1439  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1440  * on build_path_from_dentry in fs/cifs/dir.c.
1441  *
1442  * If @stop_on_nosnap, generate path relative to the first non-snapped
1443  * inode.
1444  *
1445  * Encode hidden .snap dirs as a double /, i.e.
1446  *   foo/.snap/bar -> foo//bar
1447  */
1448 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1449 			   int stop_on_nosnap)
1450 {
1451 	struct dentry *temp;
1452 	char *path;
1453 	int len, pos;
1454 	unsigned seq;
1455 
1456 	if (dentry == NULL)
1457 		return ERR_PTR(-EINVAL);
1458 
1459 retry:
1460 	len = 0;
1461 	seq = read_seqbegin(&rename_lock);
1462 	rcu_read_lock();
1463 	for (temp = dentry; !IS_ROOT(temp);) {
1464 		struct inode *inode = temp->d_inode;
1465 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1466 			len++;  /* slash only */
1467 		else if (stop_on_nosnap && inode &&
1468 			 ceph_snap(inode) == CEPH_NOSNAP)
1469 			break;
1470 		else
1471 			len += 1 + temp->d_name.len;
1472 		temp = temp->d_parent;
1473 	}
1474 	rcu_read_unlock();
1475 	if (len)
1476 		len--;  /* no leading '/' */
1477 
1478 	path = kmalloc(len+1, GFP_NOFS);
1479 	if (path == NULL)
1480 		return ERR_PTR(-ENOMEM);
1481 	pos = len;
1482 	path[pos] = 0;	/* trailing null */
1483 	rcu_read_lock();
1484 	for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1485 		struct inode *inode;
1486 
1487 		spin_lock(&temp->d_lock);
1488 		inode = temp->d_inode;
1489 		if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1490 			dout("build_path path+%d: %p SNAPDIR\n",
1491 			     pos, temp);
1492 		} else if (stop_on_nosnap && inode &&
1493 			   ceph_snap(inode) == CEPH_NOSNAP) {
1494 			spin_unlock(&temp->d_lock);
1495 			break;
1496 		} else {
1497 			pos -= temp->d_name.len;
1498 			if (pos < 0) {
1499 				spin_unlock(&temp->d_lock);
1500 				break;
1501 			}
1502 			strncpy(path + pos, temp->d_name.name,
1503 				temp->d_name.len);
1504 		}
1505 		spin_unlock(&temp->d_lock);
1506 		if (pos)
1507 			path[--pos] = '/';
1508 		temp = temp->d_parent;
1509 	}
1510 	rcu_read_unlock();
1511 	if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1512 		pr_err("build_path did not end path lookup where "
1513 		       "expected, namelen is %d, pos is %d\n", len, pos);
1514 		/* presumably this is only possible if racing with a
1515 		   rename of one of the parent directories (we can not
1516 		   lock the dentries above us to prevent this, but
1517 		   retrying should be harmless) */
1518 		kfree(path);
1519 		goto retry;
1520 	}
1521 
1522 	*base = ceph_ino(temp->d_inode);
1523 	*plen = len;
1524 	dout("build_path on %p %d built %llx '%.*s'\n",
1525 	     dentry, dentry->d_count, *base, len, path);
1526 	return path;
1527 }
1528 
1529 static int build_dentry_path(struct dentry *dentry,
1530 			     const char **ppath, int *ppathlen, u64 *pino,
1531 			     int *pfreepath)
1532 {
1533 	char *path;
1534 
1535 	if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1536 		*pino = ceph_ino(dentry->d_parent->d_inode);
1537 		*ppath = dentry->d_name.name;
1538 		*ppathlen = dentry->d_name.len;
1539 		return 0;
1540 	}
1541 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1542 	if (IS_ERR(path))
1543 		return PTR_ERR(path);
1544 	*ppath = path;
1545 	*pfreepath = 1;
1546 	return 0;
1547 }
1548 
1549 static int build_inode_path(struct inode *inode,
1550 			    const char **ppath, int *ppathlen, u64 *pino,
1551 			    int *pfreepath)
1552 {
1553 	struct dentry *dentry;
1554 	char *path;
1555 
1556 	if (ceph_snap(inode) == CEPH_NOSNAP) {
1557 		*pino = ceph_ino(inode);
1558 		*ppathlen = 0;
1559 		return 0;
1560 	}
1561 	dentry = d_find_alias(inode);
1562 	path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1563 	dput(dentry);
1564 	if (IS_ERR(path))
1565 		return PTR_ERR(path);
1566 	*ppath = path;
1567 	*pfreepath = 1;
1568 	return 0;
1569 }
1570 
1571 /*
1572  * request arguments may be specified via an inode *, a dentry *, or
1573  * an explicit ino+path.
1574  */
1575 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1576 				  const char *rpath, u64 rino,
1577 				  const char **ppath, int *pathlen,
1578 				  u64 *ino, int *freepath)
1579 {
1580 	int r = 0;
1581 
1582 	if (rinode) {
1583 		r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1584 		dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1585 		     ceph_snap(rinode));
1586 	} else if (rdentry) {
1587 		r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1588 		dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1589 		     *ppath);
1590 	} else if (rpath || rino) {
1591 		*ino = rino;
1592 		*ppath = rpath;
1593 		*pathlen = strlen(rpath);
1594 		dout(" path %.*s\n", *pathlen, rpath);
1595 	}
1596 
1597 	return r;
1598 }
1599 
1600 /*
1601  * called under mdsc->mutex
1602  */
1603 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1604 					       struct ceph_mds_request *req,
1605 					       int mds)
1606 {
1607 	struct ceph_msg *msg;
1608 	struct ceph_mds_request_head *head;
1609 	const char *path1 = NULL;
1610 	const char *path2 = NULL;
1611 	u64 ino1 = 0, ino2 = 0;
1612 	int pathlen1 = 0, pathlen2 = 0;
1613 	int freepath1 = 0, freepath2 = 0;
1614 	int len;
1615 	u16 releases;
1616 	void *p, *end;
1617 	int ret;
1618 
1619 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
1620 			      req->r_path1, req->r_ino1.ino,
1621 			      &path1, &pathlen1, &ino1, &freepath1);
1622 	if (ret < 0) {
1623 		msg = ERR_PTR(ret);
1624 		goto out;
1625 	}
1626 
1627 	ret = set_request_path_attr(NULL, req->r_old_dentry,
1628 			      req->r_path2, req->r_ino2.ino,
1629 			      &path2, &pathlen2, &ino2, &freepath2);
1630 	if (ret < 0) {
1631 		msg = ERR_PTR(ret);
1632 		goto out_free1;
1633 	}
1634 
1635 	len = sizeof(*head) +
1636 		pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1637 
1638 	/* calculate (max) length for cap releases */
1639 	len += sizeof(struct ceph_mds_request_release) *
1640 		(!!req->r_inode_drop + !!req->r_dentry_drop +
1641 		 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1642 	if (req->r_dentry_drop)
1643 		len += req->r_dentry->d_name.len;
1644 	if (req->r_old_dentry_drop)
1645 		len += req->r_old_dentry->d_name.len;
1646 
1647 	msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1648 	if (!msg) {
1649 		msg = ERR_PTR(-ENOMEM);
1650 		goto out_free2;
1651 	}
1652 
1653 	msg->hdr.tid = cpu_to_le64(req->r_tid);
1654 
1655 	head = msg->front.iov_base;
1656 	p = msg->front.iov_base + sizeof(*head);
1657 	end = msg->front.iov_base + msg->front.iov_len;
1658 
1659 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1660 	head->op = cpu_to_le32(req->r_op);
1661 	head->caller_uid = cpu_to_le32(req->r_uid);
1662 	head->caller_gid = cpu_to_le32(req->r_gid);
1663 	head->args = req->r_args;
1664 
1665 	ceph_encode_filepath(&p, end, ino1, path1);
1666 	ceph_encode_filepath(&p, end, ino2, path2);
1667 
1668 	/* make note of release offset, in case we need to replay */
1669 	req->r_request_release_offset = p - msg->front.iov_base;
1670 
1671 	/* cap releases */
1672 	releases = 0;
1673 	if (req->r_inode_drop)
1674 		releases += ceph_encode_inode_release(&p,
1675 		      req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1676 		      mds, req->r_inode_drop, req->r_inode_unless, 0);
1677 	if (req->r_dentry_drop)
1678 		releases += ceph_encode_dentry_release(&p, req->r_dentry,
1679 		       mds, req->r_dentry_drop, req->r_dentry_unless);
1680 	if (req->r_old_dentry_drop)
1681 		releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1682 		       mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1683 	if (req->r_old_inode_drop)
1684 		releases += ceph_encode_inode_release(&p,
1685 		      req->r_old_dentry->d_inode,
1686 		      mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1687 	head->num_releases = cpu_to_le16(releases);
1688 
1689 	BUG_ON(p > end);
1690 	msg->front.iov_len = p - msg->front.iov_base;
1691 	msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1692 
1693 	msg->pages = req->r_pages;
1694 	msg->nr_pages = req->r_num_pages;
1695 	msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1696 	msg->hdr.data_off = cpu_to_le16(0);
1697 
1698 out_free2:
1699 	if (freepath2)
1700 		kfree((char *)path2);
1701 out_free1:
1702 	if (freepath1)
1703 		kfree((char *)path1);
1704 out:
1705 	return msg;
1706 }
1707 
1708 /*
1709  * called under mdsc->mutex if error, under no mutex if
1710  * success.
1711  */
1712 static void complete_request(struct ceph_mds_client *mdsc,
1713 			     struct ceph_mds_request *req)
1714 {
1715 	if (req->r_callback)
1716 		req->r_callback(mdsc, req);
1717 	else
1718 		complete_all(&req->r_completion);
1719 }
1720 
1721 /*
1722  * called under mdsc->mutex
1723  */
1724 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1725 				  struct ceph_mds_request *req,
1726 				  int mds)
1727 {
1728 	struct ceph_mds_request_head *rhead;
1729 	struct ceph_msg *msg;
1730 	int flags = 0;
1731 
1732 	req->r_attempts++;
1733 	if (req->r_inode) {
1734 		struct ceph_cap *cap =
1735 			ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1736 
1737 		if (cap)
1738 			req->r_sent_on_mseq = cap->mseq;
1739 		else
1740 			req->r_sent_on_mseq = -1;
1741 	}
1742 	dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1743 	     req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1744 
1745 	if (req->r_got_unsafe) {
1746 		/*
1747 		 * Replay.  Do not regenerate message (and rebuild
1748 		 * paths, etc.); just use the original message.
1749 		 * Rebuilding paths will break for renames because
1750 		 * d_move mangles the src name.
1751 		 */
1752 		msg = req->r_request;
1753 		rhead = msg->front.iov_base;
1754 
1755 		flags = le32_to_cpu(rhead->flags);
1756 		flags |= CEPH_MDS_FLAG_REPLAY;
1757 		rhead->flags = cpu_to_le32(flags);
1758 
1759 		if (req->r_target_inode)
1760 			rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1761 
1762 		rhead->num_retry = req->r_attempts - 1;
1763 
1764 		/* remove cap/dentry releases from message */
1765 		rhead->num_releases = 0;
1766 		msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1767 		msg->front.iov_len = req->r_request_release_offset;
1768 		return 0;
1769 	}
1770 
1771 	if (req->r_request) {
1772 		ceph_msg_put(req->r_request);
1773 		req->r_request = NULL;
1774 	}
1775 	msg = create_request_message(mdsc, req, mds);
1776 	if (IS_ERR(msg)) {
1777 		req->r_err = PTR_ERR(msg);
1778 		complete_request(mdsc, req);
1779 		return PTR_ERR(msg);
1780 	}
1781 	req->r_request = msg;
1782 
1783 	rhead = msg->front.iov_base;
1784 	rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1785 	if (req->r_got_unsafe)
1786 		flags |= CEPH_MDS_FLAG_REPLAY;
1787 	if (req->r_locked_dir)
1788 		flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1789 	rhead->flags = cpu_to_le32(flags);
1790 	rhead->num_fwd = req->r_num_fwd;
1791 	rhead->num_retry = req->r_attempts - 1;
1792 	rhead->ino = 0;
1793 
1794 	dout(" r_locked_dir = %p\n", req->r_locked_dir);
1795 	return 0;
1796 }
1797 
1798 /*
1799  * send request, or put it on the appropriate wait list.
1800  */
1801 static int __do_request(struct ceph_mds_client *mdsc,
1802 			struct ceph_mds_request *req)
1803 {
1804 	struct ceph_mds_session *session = NULL;
1805 	int mds = -1;
1806 	int err = -EAGAIN;
1807 
1808 	if (req->r_err || req->r_got_result)
1809 		goto out;
1810 
1811 	if (req->r_timeout &&
1812 	    time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1813 		dout("do_request timed out\n");
1814 		err = -EIO;
1815 		goto finish;
1816 	}
1817 
1818 	put_request_session(req);
1819 
1820 	mds = __choose_mds(mdsc, req);
1821 	if (mds < 0 ||
1822 	    ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1823 		dout("do_request no mds or not active, waiting for map\n");
1824 		list_add(&req->r_wait, &mdsc->waiting_for_map);
1825 		goto out;
1826 	}
1827 
1828 	/* get, open session */
1829 	session = __ceph_lookup_mds_session(mdsc, mds);
1830 	if (!session) {
1831 		session = register_session(mdsc, mds);
1832 		if (IS_ERR(session)) {
1833 			err = PTR_ERR(session);
1834 			goto finish;
1835 		}
1836 	}
1837 	req->r_session = get_session(session);
1838 
1839 	dout("do_request mds%d session %p state %s\n", mds, session,
1840 	     session_state_name(session->s_state));
1841 	if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1842 	    session->s_state != CEPH_MDS_SESSION_HUNG) {
1843 		if (session->s_state == CEPH_MDS_SESSION_NEW ||
1844 		    session->s_state == CEPH_MDS_SESSION_CLOSING)
1845 			__open_session(mdsc, session);
1846 		list_add(&req->r_wait, &session->s_waiting);
1847 		goto out_session;
1848 	}
1849 
1850 	/* send request */
1851 	req->r_resend_mds = -1;   /* forget any previous mds hint */
1852 
1853 	if (req->r_request_started == 0)   /* note request start time */
1854 		req->r_request_started = jiffies;
1855 
1856 	err = __prepare_send_request(mdsc, req, mds);
1857 	if (!err) {
1858 		ceph_msg_get(req->r_request);
1859 		ceph_con_send(&session->s_con, req->r_request);
1860 	}
1861 
1862 out_session:
1863 	ceph_put_mds_session(session);
1864 out:
1865 	return err;
1866 
1867 finish:
1868 	req->r_err = err;
1869 	complete_request(mdsc, req);
1870 	goto out;
1871 }
1872 
1873 /*
1874  * called under mdsc->mutex
1875  */
1876 static void __wake_requests(struct ceph_mds_client *mdsc,
1877 			    struct list_head *head)
1878 {
1879 	struct ceph_mds_request *req, *nreq;
1880 
1881 	list_for_each_entry_safe(req, nreq, head, r_wait) {
1882 		list_del_init(&req->r_wait);
1883 		__do_request(mdsc, req);
1884 	}
1885 }
1886 
1887 /*
1888  * Wake up threads with requests pending for @mds, so that they can
1889  * resubmit their requests to a possibly different mds.
1890  */
1891 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1892 {
1893 	struct ceph_mds_request *req;
1894 	struct rb_node *p;
1895 
1896 	dout("kick_requests mds%d\n", mds);
1897 	for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1898 		req = rb_entry(p, struct ceph_mds_request, r_node);
1899 		if (req->r_got_unsafe)
1900 			continue;
1901 		if (req->r_session &&
1902 		    req->r_session->s_mds == mds) {
1903 			dout(" kicking tid %llu\n", req->r_tid);
1904 			__do_request(mdsc, req);
1905 		}
1906 	}
1907 }
1908 
1909 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1910 			      struct ceph_mds_request *req)
1911 {
1912 	dout("submit_request on %p\n", req);
1913 	mutex_lock(&mdsc->mutex);
1914 	__register_request(mdsc, req, NULL);
1915 	__do_request(mdsc, req);
1916 	mutex_unlock(&mdsc->mutex);
1917 }
1918 
1919 /*
1920  * Synchrously perform an mds request.  Take care of all of the
1921  * session setup, forwarding, retry details.
1922  */
1923 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1924 			 struct inode *dir,
1925 			 struct ceph_mds_request *req)
1926 {
1927 	int err;
1928 
1929 	dout("do_request on %p\n", req);
1930 
1931 	/* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1932 	if (req->r_inode)
1933 		ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1934 	if (req->r_locked_dir)
1935 		ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1936 	if (req->r_old_dentry)
1937 		ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
1938 				  CEPH_CAP_PIN);
1939 
1940 	/* issue */
1941 	mutex_lock(&mdsc->mutex);
1942 	__register_request(mdsc, req, dir);
1943 	__do_request(mdsc, req);
1944 
1945 	if (req->r_err) {
1946 		err = req->r_err;
1947 		__unregister_request(mdsc, req);
1948 		dout("do_request early error %d\n", err);
1949 		goto out;
1950 	}
1951 
1952 	/* wait */
1953 	mutex_unlock(&mdsc->mutex);
1954 	dout("do_request waiting\n");
1955 	if (req->r_timeout) {
1956 		err = (long)wait_for_completion_killable_timeout(
1957 			&req->r_completion, req->r_timeout);
1958 		if (err == 0)
1959 			err = -EIO;
1960 	} else {
1961 		err = wait_for_completion_killable(&req->r_completion);
1962 	}
1963 	dout("do_request waited, got %d\n", err);
1964 	mutex_lock(&mdsc->mutex);
1965 
1966 	/* only abort if we didn't race with a real reply */
1967 	if (req->r_got_result) {
1968 		err = le32_to_cpu(req->r_reply_info.head->result);
1969 	} else if (err < 0) {
1970 		dout("aborted request %lld with %d\n", req->r_tid, err);
1971 
1972 		/*
1973 		 * ensure we aren't running concurrently with
1974 		 * ceph_fill_trace or ceph_readdir_prepopulate, which
1975 		 * rely on locks (dir mutex) held by our caller.
1976 		 */
1977 		mutex_lock(&req->r_fill_mutex);
1978 		req->r_err = err;
1979 		req->r_aborted = true;
1980 		mutex_unlock(&req->r_fill_mutex);
1981 
1982 		if (req->r_locked_dir &&
1983 		    (req->r_op & CEPH_MDS_OP_WRITE))
1984 			ceph_invalidate_dir_request(req);
1985 	} else {
1986 		err = req->r_err;
1987 	}
1988 
1989 out:
1990 	mutex_unlock(&mdsc->mutex);
1991 	dout("do_request %p done, result %d\n", req, err);
1992 	return err;
1993 }
1994 
1995 /*
1996  * Invalidate dir D_COMPLETE, dentry lease state on an aborted MDS
1997  * namespace request.
1998  */
1999 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2000 {
2001 	struct inode *inode = req->r_locked_dir;
2002 	struct ceph_inode_info *ci = ceph_inode(inode);
2003 
2004 	dout("invalidate_dir_request %p (D_COMPLETE, lease(s))\n", inode);
2005 	spin_lock(&ci->i_ceph_lock);
2006 	ceph_dir_clear_complete(inode);
2007 	ci->i_release_count++;
2008 	spin_unlock(&ci->i_ceph_lock);
2009 
2010 	if (req->r_dentry)
2011 		ceph_invalidate_dentry_lease(req->r_dentry);
2012 	if (req->r_old_dentry)
2013 		ceph_invalidate_dentry_lease(req->r_old_dentry);
2014 }
2015 
2016 /*
2017  * Handle mds reply.
2018  *
2019  * We take the session mutex and parse and process the reply immediately.
2020  * This preserves the logical ordering of replies, capabilities, etc., sent
2021  * by the MDS as they are applied to our local cache.
2022  */
2023 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2024 {
2025 	struct ceph_mds_client *mdsc = session->s_mdsc;
2026 	struct ceph_mds_request *req;
2027 	struct ceph_mds_reply_head *head = msg->front.iov_base;
2028 	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2029 	u64 tid;
2030 	int err, result;
2031 	int mds = session->s_mds;
2032 
2033 	if (msg->front.iov_len < sizeof(*head)) {
2034 		pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2035 		ceph_msg_dump(msg);
2036 		return;
2037 	}
2038 
2039 	/* get request, session */
2040 	tid = le64_to_cpu(msg->hdr.tid);
2041 	mutex_lock(&mdsc->mutex);
2042 	req = __lookup_request(mdsc, tid);
2043 	if (!req) {
2044 		dout("handle_reply on unknown tid %llu\n", tid);
2045 		mutex_unlock(&mdsc->mutex);
2046 		return;
2047 	}
2048 	dout("handle_reply %p\n", req);
2049 
2050 	/* correct session? */
2051 	if (req->r_session != session) {
2052 		pr_err("mdsc_handle_reply got %llu on session mds%d"
2053 		       " not mds%d\n", tid, session->s_mds,
2054 		       req->r_session ? req->r_session->s_mds : -1);
2055 		mutex_unlock(&mdsc->mutex);
2056 		goto out;
2057 	}
2058 
2059 	/* dup? */
2060 	if ((req->r_got_unsafe && !head->safe) ||
2061 	    (req->r_got_safe && head->safe)) {
2062 		pr_warning("got a dup %s reply on %llu from mds%d\n",
2063 			   head->safe ? "safe" : "unsafe", tid, mds);
2064 		mutex_unlock(&mdsc->mutex);
2065 		goto out;
2066 	}
2067 	if (req->r_got_safe && !head->safe) {
2068 		pr_warning("got unsafe after safe on %llu from mds%d\n",
2069 			   tid, mds);
2070 		mutex_unlock(&mdsc->mutex);
2071 		goto out;
2072 	}
2073 
2074 	result = le32_to_cpu(head->result);
2075 
2076 	/*
2077 	 * Handle an ESTALE
2078 	 * if we're not talking to the authority, send to them
2079 	 * if the authority has changed while we weren't looking,
2080 	 * send to new authority
2081 	 * Otherwise we just have to return an ESTALE
2082 	 */
2083 	if (result == -ESTALE) {
2084 		dout("got ESTALE on request %llu", req->r_tid);
2085 		if (!req->r_inode) {
2086 			/* do nothing; not an authority problem */
2087 		} else if (req->r_direct_mode != USE_AUTH_MDS) {
2088 			dout("not using auth, setting for that now");
2089 			req->r_direct_mode = USE_AUTH_MDS;
2090 			__do_request(mdsc, req);
2091 			mutex_unlock(&mdsc->mutex);
2092 			goto out;
2093 		} else  {
2094 			struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2095 			struct ceph_cap *cap = NULL;
2096 
2097 			if (req->r_session)
2098 				cap = ceph_get_cap_for_mds(ci,
2099 						   req->r_session->s_mds);
2100 
2101 			dout("already using auth");
2102 			if ((!cap || cap != ci->i_auth_cap) ||
2103 			    (cap->mseq != req->r_sent_on_mseq)) {
2104 				dout("but cap changed, so resending");
2105 				__do_request(mdsc, req);
2106 				mutex_unlock(&mdsc->mutex);
2107 				goto out;
2108 			}
2109 		}
2110 		dout("have to return ESTALE on request %llu", req->r_tid);
2111 	}
2112 
2113 
2114 	if (head->safe) {
2115 		req->r_got_safe = true;
2116 		__unregister_request(mdsc, req);
2117 		complete_all(&req->r_safe_completion);
2118 
2119 		if (req->r_got_unsafe) {
2120 			/*
2121 			 * We already handled the unsafe response, now do the
2122 			 * cleanup.  No need to examine the response; the MDS
2123 			 * doesn't include any result info in the safe
2124 			 * response.  And even if it did, there is nothing
2125 			 * useful we could do with a revised return value.
2126 			 */
2127 			dout("got safe reply %llu, mds%d\n", tid, mds);
2128 			list_del_init(&req->r_unsafe_item);
2129 
2130 			/* last unsafe request during umount? */
2131 			if (mdsc->stopping && !__get_oldest_req(mdsc))
2132 				complete_all(&mdsc->safe_umount_waiters);
2133 			mutex_unlock(&mdsc->mutex);
2134 			goto out;
2135 		}
2136 	} else {
2137 		req->r_got_unsafe = true;
2138 		list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2139 	}
2140 
2141 	dout("handle_reply tid %lld result %d\n", tid, result);
2142 	rinfo = &req->r_reply_info;
2143 	err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2144 	mutex_unlock(&mdsc->mutex);
2145 
2146 	mutex_lock(&session->s_mutex);
2147 	if (err < 0) {
2148 		pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2149 		ceph_msg_dump(msg);
2150 		goto out_err;
2151 	}
2152 
2153 	/* snap trace */
2154 	if (rinfo->snapblob_len) {
2155 		down_write(&mdsc->snap_rwsem);
2156 		ceph_update_snap_trace(mdsc, rinfo->snapblob,
2157 			       rinfo->snapblob + rinfo->snapblob_len,
2158 			       le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2159 		downgrade_write(&mdsc->snap_rwsem);
2160 	} else {
2161 		down_read(&mdsc->snap_rwsem);
2162 	}
2163 
2164 	/* insert trace into our cache */
2165 	mutex_lock(&req->r_fill_mutex);
2166 	err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2167 	if (err == 0) {
2168 		if (result == 0 && req->r_op != CEPH_MDS_OP_GETFILELOCK &&
2169 		    rinfo->dir_nr)
2170 			ceph_readdir_prepopulate(req, req->r_session);
2171 		ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2172 	}
2173 	mutex_unlock(&req->r_fill_mutex);
2174 
2175 	up_read(&mdsc->snap_rwsem);
2176 out_err:
2177 	mutex_lock(&mdsc->mutex);
2178 	if (!req->r_aborted) {
2179 		if (err) {
2180 			req->r_err = err;
2181 		} else {
2182 			req->r_reply = msg;
2183 			ceph_msg_get(msg);
2184 			req->r_got_result = true;
2185 		}
2186 	} else {
2187 		dout("reply arrived after request %lld was aborted\n", tid);
2188 	}
2189 	mutex_unlock(&mdsc->mutex);
2190 
2191 	ceph_add_cap_releases(mdsc, req->r_session);
2192 	mutex_unlock(&session->s_mutex);
2193 
2194 	/* kick calling process */
2195 	complete_request(mdsc, req);
2196 out:
2197 	ceph_mdsc_put_request(req);
2198 	return;
2199 }
2200 
2201 
2202 
2203 /*
2204  * handle mds notification that our request has been forwarded.
2205  */
2206 static void handle_forward(struct ceph_mds_client *mdsc,
2207 			   struct ceph_mds_session *session,
2208 			   struct ceph_msg *msg)
2209 {
2210 	struct ceph_mds_request *req;
2211 	u64 tid = le64_to_cpu(msg->hdr.tid);
2212 	u32 next_mds;
2213 	u32 fwd_seq;
2214 	int err = -EINVAL;
2215 	void *p = msg->front.iov_base;
2216 	void *end = p + msg->front.iov_len;
2217 
2218 	ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2219 	next_mds = ceph_decode_32(&p);
2220 	fwd_seq = ceph_decode_32(&p);
2221 
2222 	mutex_lock(&mdsc->mutex);
2223 	req = __lookup_request(mdsc, tid);
2224 	if (!req) {
2225 		dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2226 		goto out;  /* dup reply? */
2227 	}
2228 
2229 	if (req->r_aborted) {
2230 		dout("forward tid %llu aborted, unregistering\n", tid);
2231 		__unregister_request(mdsc, req);
2232 	} else if (fwd_seq <= req->r_num_fwd) {
2233 		dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2234 		     tid, next_mds, req->r_num_fwd, fwd_seq);
2235 	} else {
2236 		/* resend. forward race not possible; mds would drop */
2237 		dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2238 		BUG_ON(req->r_err);
2239 		BUG_ON(req->r_got_result);
2240 		req->r_num_fwd = fwd_seq;
2241 		req->r_resend_mds = next_mds;
2242 		put_request_session(req);
2243 		__do_request(mdsc, req);
2244 	}
2245 	ceph_mdsc_put_request(req);
2246 out:
2247 	mutex_unlock(&mdsc->mutex);
2248 	return;
2249 
2250 bad:
2251 	pr_err("mdsc_handle_forward decode error err=%d\n", err);
2252 }
2253 
2254 /*
2255  * handle a mds session control message
2256  */
2257 static void handle_session(struct ceph_mds_session *session,
2258 			   struct ceph_msg *msg)
2259 {
2260 	struct ceph_mds_client *mdsc = session->s_mdsc;
2261 	u32 op;
2262 	u64 seq;
2263 	int mds = session->s_mds;
2264 	struct ceph_mds_session_head *h = msg->front.iov_base;
2265 	int wake = 0;
2266 
2267 	/* decode */
2268 	if (msg->front.iov_len != sizeof(*h))
2269 		goto bad;
2270 	op = le32_to_cpu(h->op);
2271 	seq = le64_to_cpu(h->seq);
2272 
2273 	mutex_lock(&mdsc->mutex);
2274 	if (op == CEPH_SESSION_CLOSE)
2275 		__unregister_session(mdsc, session);
2276 	/* FIXME: this ttl calculation is generous */
2277 	session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2278 	mutex_unlock(&mdsc->mutex);
2279 
2280 	mutex_lock(&session->s_mutex);
2281 
2282 	dout("handle_session mds%d %s %p state %s seq %llu\n",
2283 	     mds, ceph_session_op_name(op), session,
2284 	     session_state_name(session->s_state), seq);
2285 
2286 	if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2287 		session->s_state = CEPH_MDS_SESSION_OPEN;
2288 		pr_info("mds%d came back\n", session->s_mds);
2289 	}
2290 
2291 	switch (op) {
2292 	case CEPH_SESSION_OPEN:
2293 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2294 			pr_info("mds%d reconnect success\n", session->s_mds);
2295 		session->s_state = CEPH_MDS_SESSION_OPEN;
2296 		renewed_caps(mdsc, session, 0);
2297 		wake = 1;
2298 		if (mdsc->stopping)
2299 			__close_session(mdsc, session);
2300 		break;
2301 
2302 	case CEPH_SESSION_RENEWCAPS:
2303 		if (session->s_renew_seq == seq)
2304 			renewed_caps(mdsc, session, 1);
2305 		break;
2306 
2307 	case CEPH_SESSION_CLOSE:
2308 		if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2309 			pr_info("mds%d reconnect denied\n", session->s_mds);
2310 		remove_session_caps(session);
2311 		wake = 1; /* for good measure */
2312 		wake_up_all(&mdsc->session_close_wq);
2313 		kick_requests(mdsc, mds);
2314 		break;
2315 
2316 	case CEPH_SESSION_STALE:
2317 		pr_info("mds%d caps went stale, renewing\n",
2318 			session->s_mds);
2319 		spin_lock(&session->s_gen_ttl_lock);
2320 		session->s_cap_gen++;
2321 		session->s_cap_ttl = jiffies - 1;
2322 		spin_unlock(&session->s_gen_ttl_lock);
2323 		send_renew_caps(mdsc, session);
2324 		break;
2325 
2326 	case CEPH_SESSION_RECALL_STATE:
2327 		trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2328 		break;
2329 
2330 	default:
2331 		pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2332 		WARN_ON(1);
2333 	}
2334 
2335 	mutex_unlock(&session->s_mutex);
2336 	if (wake) {
2337 		mutex_lock(&mdsc->mutex);
2338 		__wake_requests(mdsc, &session->s_waiting);
2339 		mutex_unlock(&mdsc->mutex);
2340 	}
2341 	return;
2342 
2343 bad:
2344 	pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2345 	       (int)msg->front.iov_len);
2346 	ceph_msg_dump(msg);
2347 	return;
2348 }
2349 
2350 
2351 /*
2352  * called under session->mutex.
2353  */
2354 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2355 				   struct ceph_mds_session *session)
2356 {
2357 	struct ceph_mds_request *req, *nreq;
2358 	int err;
2359 
2360 	dout("replay_unsafe_requests mds%d\n", session->s_mds);
2361 
2362 	mutex_lock(&mdsc->mutex);
2363 	list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2364 		err = __prepare_send_request(mdsc, req, session->s_mds);
2365 		if (!err) {
2366 			ceph_msg_get(req->r_request);
2367 			ceph_con_send(&session->s_con, req->r_request);
2368 		}
2369 	}
2370 	mutex_unlock(&mdsc->mutex);
2371 }
2372 
2373 /*
2374  * Encode information about a cap for a reconnect with the MDS.
2375  */
2376 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2377 			  void *arg)
2378 {
2379 	union {
2380 		struct ceph_mds_cap_reconnect v2;
2381 		struct ceph_mds_cap_reconnect_v1 v1;
2382 	} rec;
2383 	size_t reclen;
2384 	struct ceph_inode_info *ci;
2385 	struct ceph_reconnect_state *recon_state = arg;
2386 	struct ceph_pagelist *pagelist = recon_state->pagelist;
2387 	char *path;
2388 	int pathlen, err;
2389 	u64 pathbase;
2390 	struct dentry *dentry;
2391 
2392 	ci = cap->ci;
2393 
2394 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2395 	     inode, ceph_vinop(inode), cap, cap->cap_id,
2396 	     ceph_cap_string(cap->issued));
2397 	err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2398 	if (err)
2399 		return err;
2400 
2401 	dentry = d_find_alias(inode);
2402 	if (dentry) {
2403 		path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2404 		if (IS_ERR(path)) {
2405 			err = PTR_ERR(path);
2406 			goto out_dput;
2407 		}
2408 	} else {
2409 		path = NULL;
2410 		pathlen = 0;
2411 	}
2412 	err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2413 	if (err)
2414 		goto out_free;
2415 
2416 	spin_lock(&ci->i_ceph_lock);
2417 	cap->seq = 0;        /* reset cap seq */
2418 	cap->issue_seq = 0;  /* and issue_seq */
2419 
2420 	if (recon_state->flock) {
2421 		rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2422 		rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2423 		rec.v2.issued = cpu_to_le32(cap->issued);
2424 		rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2425 		rec.v2.pathbase = cpu_to_le64(pathbase);
2426 		rec.v2.flock_len = 0;
2427 		reclen = sizeof(rec.v2);
2428 	} else {
2429 		rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2430 		rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2431 		rec.v1.issued = cpu_to_le32(cap->issued);
2432 		rec.v1.size = cpu_to_le64(inode->i_size);
2433 		ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2434 		ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2435 		rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2436 		rec.v1.pathbase = cpu_to_le64(pathbase);
2437 		reclen = sizeof(rec.v1);
2438 	}
2439 	spin_unlock(&ci->i_ceph_lock);
2440 
2441 	if (recon_state->flock) {
2442 		int num_fcntl_locks, num_flock_locks;
2443 		struct ceph_pagelist_cursor trunc_point;
2444 
2445 		ceph_pagelist_set_cursor(pagelist, &trunc_point);
2446 		do {
2447 			lock_flocks();
2448 			ceph_count_locks(inode, &num_fcntl_locks,
2449 					 &num_flock_locks);
2450 			rec.v2.flock_len = (2*sizeof(u32) +
2451 					    (num_fcntl_locks+num_flock_locks) *
2452 					    sizeof(struct ceph_filelock));
2453 			unlock_flocks();
2454 
2455 			/* pre-alloc pagelist */
2456 			ceph_pagelist_truncate(pagelist, &trunc_point);
2457 			err = ceph_pagelist_append(pagelist, &rec, reclen);
2458 			if (!err)
2459 				err = ceph_pagelist_reserve(pagelist,
2460 							    rec.v2.flock_len);
2461 
2462 			/* encode locks */
2463 			if (!err) {
2464 				lock_flocks();
2465 				err = ceph_encode_locks(inode,
2466 							pagelist,
2467 							num_fcntl_locks,
2468 							num_flock_locks);
2469 				unlock_flocks();
2470 			}
2471 		} while (err == -ENOSPC);
2472 	} else {
2473 		err = ceph_pagelist_append(pagelist, &rec, reclen);
2474 	}
2475 
2476 out_free:
2477 	kfree(path);
2478 out_dput:
2479 	dput(dentry);
2480 	return err;
2481 }
2482 
2483 
2484 /*
2485  * If an MDS fails and recovers, clients need to reconnect in order to
2486  * reestablish shared state.  This includes all caps issued through
2487  * this session _and_ the snap_realm hierarchy.  Because it's not
2488  * clear which snap realms the mds cares about, we send everything we
2489  * know about.. that ensures we'll then get any new info the
2490  * recovering MDS might have.
2491  *
2492  * This is a relatively heavyweight operation, but it's rare.
2493  *
2494  * called with mdsc->mutex held.
2495  */
2496 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2497 			       struct ceph_mds_session *session)
2498 {
2499 	struct ceph_msg *reply;
2500 	struct rb_node *p;
2501 	int mds = session->s_mds;
2502 	int err = -ENOMEM;
2503 	struct ceph_pagelist *pagelist;
2504 	struct ceph_reconnect_state recon_state;
2505 
2506 	pr_info("mds%d reconnect start\n", mds);
2507 
2508 	pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2509 	if (!pagelist)
2510 		goto fail_nopagelist;
2511 	ceph_pagelist_init(pagelist);
2512 
2513 	reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2514 	if (!reply)
2515 		goto fail_nomsg;
2516 
2517 	mutex_lock(&session->s_mutex);
2518 	session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2519 	session->s_seq = 0;
2520 
2521 	ceph_con_close(&session->s_con);
2522 	ceph_con_open(&session->s_con,
2523 		      CEPH_ENTITY_TYPE_MDS, mds,
2524 		      ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2525 
2526 	/* replay unsafe requests */
2527 	replay_unsafe_requests(mdsc, session);
2528 
2529 	down_read(&mdsc->snap_rwsem);
2530 
2531 	dout("session %p state %s\n", session,
2532 	     session_state_name(session->s_state));
2533 
2534 	/* drop old cap expires; we're about to reestablish that state */
2535 	discard_cap_releases(mdsc, session);
2536 
2537 	/* traverse this session's caps */
2538 	err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2539 	if (err)
2540 		goto fail;
2541 
2542 	recon_state.pagelist = pagelist;
2543 	recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2544 	err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2545 	if (err < 0)
2546 		goto fail;
2547 
2548 	/*
2549 	 * snaprealms.  we provide mds with the ino, seq (version), and
2550 	 * parent for all of our realms.  If the mds has any newer info,
2551 	 * it will tell us.
2552 	 */
2553 	for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2554 		struct ceph_snap_realm *realm =
2555 			rb_entry(p, struct ceph_snap_realm, node);
2556 		struct ceph_mds_snaprealm_reconnect sr_rec;
2557 
2558 		dout(" adding snap realm %llx seq %lld parent %llx\n",
2559 		     realm->ino, realm->seq, realm->parent_ino);
2560 		sr_rec.ino = cpu_to_le64(realm->ino);
2561 		sr_rec.seq = cpu_to_le64(realm->seq);
2562 		sr_rec.parent = cpu_to_le64(realm->parent_ino);
2563 		err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2564 		if (err)
2565 			goto fail;
2566 	}
2567 
2568 	reply->pagelist = pagelist;
2569 	if (recon_state.flock)
2570 		reply->hdr.version = cpu_to_le16(2);
2571 	reply->hdr.data_len = cpu_to_le32(pagelist->length);
2572 	reply->nr_pages = calc_pages_for(0, pagelist->length);
2573 	ceph_con_send(&session->s_con, reply);
2574 
2575 	mutex_unlock(&session->s_mutex);
2576 
2577 	mutex_lock(&mdsc->mutex);
2578 	__wake_requests(mdsc, &session->s_waiting);
2579 	mutex_unlock(&mdsc->mutex);
2580 
2581 	up_read(&mdsc->snap_rwsem);
2582 	return;
2583 
2584 fail:
2585 	ceph_msg_put(reply);
2586 	up_read(&mdsc->snap_rwsem);
2587 	mutex_unlock(&session->s_mutex);
2588 fail_nomsg:
2589 	ceph_pagelist_release(pagelist);
2590 	kfree(pagelist);
2591 fail_nopagelist:
2592 	pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2593 	return;
2594 }
2595 
2596 
2597 /*
2598  * compare old and new mdsmaps, kicking requests
2599  * and closing out old connections as necessary
2600  *
2601  * called under mdsc->mutex.
2602  */
2603 static void check_new_map(struct ceph_mds_client *mdsc,
2604 			  struct ceph_mdsmap *newmap,
2605 			  struct ceph_mdsmap *oldmap)
2606 {
2607 	int i;
2608 	int oldstate, newstate;
2609 	struct ceph_mds_session *s;
2610 
2611 	dout("check_new_map new %u old %u\n",
2612 	     newmap->m_epoch, oldmap->m_epoch);
2613 
2614 	for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2615 		if (mdsc->sessions[i] == NULL)
2616 			continue;
2617 		s = mdsc->sessions[i];
2618 		oldstate = ceph_mdsmap_get_state(oldmap, i);
2619 		newstate = ceph_mdsmap_get_state(newmap, i);
2620 
2621 		dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2622 		     i, ceph_mds_state_name(oldstate),
2623 		     ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2624 		     ceph_mds_state_name(newstate),
2625 		     ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2626 		     session_state_name(s->s_state));
2627 
2628 		if (i >= newmap->m_max_mds ||
2629 		    memcmp(ceph_mdsmap_get_addr(oldmap, i),
2630 			   ceph_mdsmap_get_addr(newmap, i),
2631 			   sizeof(struct ceph_entity_addr))) {
2632 			if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2633 				/* the session never opened, just close it
2634 				 * out now */
2635 				__wake_requests(mdsc, &s->s_waiting);
2636 				__unregister_session(mdsc, s);
2637 			} else {
2638 				/* just close it */
2639 				mutex_unlock(&mdsc->mutex);
2640 				mutex_lock(&s->s_mutex);
2641 				mutex_lock(&mdsc->mutex);
2642 				ceph_con_close(&s->s_con);
2643 				mutex_unlock(&s->s_mutex);
2644 				s->s_state = CEPH_MDS_SESSION_RESTARTING;
2645 			}
2646 
2647 			/* kick any requests waiting on the recovering mds */
2648 			kick_requests(mdsc, i);
2649 		} else if (oldstate == newstate) {
2650 			continue;  /* nothing new with this mds */
2651 		}
2652 
2653 		/*
2654 		 * send reconnect?
2655 		 */
2656 		if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2657 		    newstate >= CEPH_MDS_STATE_RECONNECT) {
2658 			mutex_unlock(&mdsc->mutex);
2659 			send_mds_reconnect(mdsc, s);
2660 			mutex_lock(&mdsc->mutex);
2661 		}
2662 
2663 		/*
2664 		 * kick request on any mds that has gone active.
2665 		 */
2666 		if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2667 		    newstate >= CEPH_MDS_STATE_ACTIVE) {
2668 			if (oldstate != CEPH_MDS_STATE_CREATING &&
2669 			    oldstate != CEPH_MDS_STATE_STARTING)
2670 				pr_info("mds%d recovery completed\n", s->s_mds);
2671 			kick_requests(mdsc, i);
2672 			ceph_kick_flushing_caps(mdsc, s);
2673 			wake_up_session_caps(s, 1);
2674 		}
2675 	}
2676 
2677 	for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2678 		s = mdsc->sessions[i];
2679 		if (!s)
2680 			continue;
2681 		if (!ceph_mdsmap_is_laggy(newmap, i))
2682 			continue;
2683 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2684 		    s->s_state == CEPH_MDS_SESSION_HUNG ||
2685 		    s->s_state == CEPH_MDS_SESSION_CLOSING) {
2686 			dout(" connecting to export targets of laggy mds%d\n",
2687 			     i);
2688 			__open_export_target_sessions(mdsc, s);
2689 		}
2690 	}
2691 }
2692 
2693 
2694 
2695 /*
2696  * leases
2697  */
2698 
2699 /*
2700  * caller must hold session s_mutex, dentry->d_lock
2701  */
2702 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2703 {
2704 	struct ceph_dentry_info *di = ceph_dentry(dentry);
2705 
2706 	ceph_put_mds_session(di->lease_session);
2707 	di->lease_session = NULL;
2708 }
2709 
2710 static void handle_lease(struct ceph_mds_client *mdsc,
2711 			 struct ceph_mds_session *session,
2712 			 struct ceph_msg *msg)
2713 {
2714 	struct super_block *sb = mdsc->fsc->sb;
2715 	struct inode *inode;
2716 	struct dentry *parent, *dentry;
2717 	struct ceph_dentry_info *di;
2718 	int mds = session->s_mds;
2719 	struct ceph_mds_lease *h = msg->front.iov_base;
2720 	u32 seq;
2721 	struct ceph_vino vino;
2722 	struct qstr dname;
2723 	int release = 0;
2724 
2725 	dout("handle_lease from mds%d\n", mds);
2726 
2727 	/* decode */
2728 	if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2729 		goto bad;
2730 	vino.ino = le64_to_cpu(h->ino);
2731 	vino.snap = CEPH_NOSNAP;
2732 	seq = le32_to_cpu(h->seq);
2733 	dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2734 	dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2735 	if (dname.len != get_unaligned_le32(h+1))
2736 		goto bad;
2737 
2738 	mutex_lock(&session->s_mutex);
2739 	session->s_seq++;
2740 
2741 	/* lookup inode */
2742 	inode = ceph_find_inode(sb, vino);
2743 	dout("handle_lease %s, ino %llx %p %.*s\n",
2744 	     ceph_lease_op_name(h->action), vino.ino, inode,
2745 	     dname.len, dname.name);
2746 	if (inode == NULL) {
2747 		dout("handle_lease no inode %llx\n", vino.ino);
2748 		goto release;
2749 	}
2750 
2751 	/* dentry */
2752 	parent = d_find_alias(inode);
2753 	if (!parent) {
2754 		dout("no parent dentry on inode %p\n", inode);
2755 		WARN_ON(1);
2756 		goto release;  /* hrm... */
2757 	}
2758 	dname.hash = full_name_hash(dname.name, dname.len);
2759 	dentry = d_lookup(parent, &dname);
2760 	dput(parent);
2761 	if (!dentry)
2762 		goto release;
2763 
2764 	spin_lock(&dentry->d_lock);
2765 	di = ceph_dentry(dentry);
2766 	switch (h->action) {
2767 	case CEPH_MDS_LEASE_REVOKE:
2768 		if (di->lease_session == session) {
2769 			if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2770 				h->seq = cpu_to_le32(di->lease_seq);
2771 			__ceph_mdsc_drop_dentry_lease(dentry);
2772 		}
2773 		release = 1;
2774 		break;
2775 
2776 	case CEPH_MDS_LEASE_RENEW:
2777 		if (di->lease_session == session &&
2778 		    di->lease_gen == session->s_cap_gen &&
2779 		    di->lease_renew_from &&
2780 		    di->lease_renew_after == 0) {
2781 			unsigned long duration =
2782 				le32_to_cpu(h->duration_ms) * HZ / 1000;
2783 
2784 			di->lease_seq = seq;
2785 			dentry->d_time = di->lease_renew_from + duration;
2786 			di->lease_renew_after = di->lease_renew_from +
2787 				(duration >> 1);
2788 			di->lease_renew_from = 0;
2789 		}
2790 		break;
2791 	}
2792 	spin_unlock(&dentry->d_lock);
2793 	dput(dentry);
2794 
2795 	if (!release)
2796 		goto out;
2797 
2798 release:
2799 	/* let's just reuse the same message */
2800 	h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2801 	ceph_msg_get(msg);
2802 	ceph_con_send(&session->s_con, msg);
2803 
2804 out:
2805 	iput(inode);
2806 	mutex_unlock(&session->s_mutex);
2807 	return;
2808 
2809 bad:
2810 	pr_err("corrupt lease message\n");
2811 	ceph_msg_dump(msg);
2812 }
2813 
2814 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2815 			      struct inode *inode,
2816 			      struct dentry *dentry, char action,
2817 			      u32 seq)
2818 {
2819 	struct ceph_msg *msg;
2820 	struct ceph_mds_lease *lease;
2821 	int len = sizeof(*lease) + sizeof(u32);
2822 	int dnamelen = 0;
2823 
2824 	dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2825 	     inode, dentry, ceph_lease_op_name(action), session->s_mds);
2826 	dnamelen = dentry->d_name.len;
2827 	len += dnamelen;
2828 
2829 	msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2830 	if (!msg)
2831 		return;
2832 	lease = msg->front.iov_base;
2833 	lease->action = action;
2834 	lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2835 	lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2836 	lease->seq = cpu_to_le32(seq);
2837 	put_unaligned_le32(dnamelen, lease + 1);
2838 	memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2839 
2840 	/*
2841 	 * if this is a preemptive lease RELEASE, no need to
2842 	 * flush request stream, since the actual request will
2843 	 * soon follow.
2844 	 */
2845 	msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2846 
2847 	ceph_con_send(&session->s_con, msg);
2848 }
2849 
2850 /*
2851  * Preemptively release a lease we expect to invalidate anyway.
2852  * Pass @inode always, @dentry is optional.
2853  */
2854 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2855 			     struct dentry *dentry)
2856 {
2857 	struct ceph_dentry_info *di;
2858 	struct ceph_mds_session *session;
2859 	u32 seq;
2860 
2861 	BUG_ON(inode == NULL);
2862 	BUG_ON(dentry == NULL);
2863 
2864 	/* is dentry lease valid? */
2865 	spin_lock(&dentry->d_lock);
2866 	di = ceph_dentry(dentry);
2867 	if (!di || !di->lease_session ||
2868 	    di->lease_session->s_mds < 0 ||
2869 	    di->lease_gen != di->lease_session->s_cap_gen ||
2870 	    !time_before(jiffies, dentry->d_time)) {
2871 		dout("lease_release inode %p dentry %p -- "
2872 		     "no lease\n",
2873 		     inode, dentry);
2874 		spin_unlock(&dentry->d_lock);
2875 		return;
2876 	}
2877 
2878 	/* we do have a lease on this dentry; note mds and seq */
2879 	session = ceph_get_mds_session(di->lease_session);
2880 	seq = di->lease_seq;
2881 	__ceph_mdsc_drop_dentry_lease(dentry);
2882 	spin_unlock(&dentry->d_lock);
2883 
2884 	dout("lease_release inode %p dentry %p to mds%d\n",
2885 	     inode, dentry, session->s_mds);
2886 	ceph_mdsc_lease_send_msg(session, inode, dentry,
2887 				 CEPH_MDS_LEASE_RELEASE, seq);
2888 	ceph_put_mds_session(session);
2889 }
2890 
2891 /*
2892  * drop all leases (and dentry refs) in preparation for umount
2893  */
2894 static void drop_leases(struct ceph_mds_client *mdsc)
2895 {
2896 	int i;
2897 
2898 	dout("drop_leases\n");
2899 	mutex_lock(&mdsc->mutex);
2900 	for (i = 0; i < mdsc->max_sessions; i++) {
2901 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2902 		if (!s)
2903 			continue;
2904 		mutex_unlock(&mdsc->mutex);
2905 		mutex_lock(&s->s_mutex);
2906 		mutex_unlock(&s->s_mutex);
2907 		ceph_put_mds_session(s);
2908 		mutex_lock(&mdsc->mutex);
2909 	}
2910 	mutex_unlock(&mdsc->mutex);
2911 }
2912 
2913 
2914 
2915 /*
2916  * delayed work -- periodically trim expired leases, renew caps with mds
2917  */
2918 static void schedule_delayed(struct ceph_mds_client *mdsc)
2919 {
2920 	int delay = 5;
2921 	unsigned hz = round_jiffies_relative(HZ * delay);
2922 	schedule_delayed_work(&mdsc->delayed_work, hz);
2923 }
2924 
2925 static void delayed_work(struct work_struct *work)
2926 {
2927 	int i;
2928 	struct ceph_mds_client *mdsc =
2929 		container_of(work, struct ceph_mds_client, delayed_work.work);
2930 	int renew_interval;
2931 	int renew_caps;
2932 
2933 	dout("mdsc delayed_work\n");
2934 	ceph_check_delayed_caps(mdsc);
2935 
2936 	mutex_lock(&mdsc->mutex);
2937 	renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2938 	renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2939 				   mdsc->last_renew_caps);
2940 	if (renew_caps)
2941 		mdsc->last_renew_caps = jiffies;
2942 
2943 	for (i = 0; i < mdsc->max_sessions; i++) {
2944 		struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2945 		if (s == NULL)
2946 			continue;
2947 		if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2948 			dout("resending session close request for mds%d\n",
2949 			     s->s_mds);
2950 			request_close_session(mdsc, s);
2951 			ceph_put_mds_session(s);
2952 			continue;
2953 		}
2954 		if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2955 			if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2956 				s->s_state = CEPH_MDS_SESSION_HUNG;
2957 				pr_info("mds%d hung\n", s->s_mds);
2958 			}
2959 		}
2960 		if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2961 			/* this mds is failed or recovering, just wait */
2962 			ceph_put_mds_session(s);
2963 			continue;
2964 		}
2965 		mutex_unlock(&mdsc->mutex);
2966 
2967 		mutex_lock(&s->s_mutex);
2968 		if (renew_caps)
2969 			send_renew_caps(mdsc, s);
2970 		else
2971 			ceph_con_keepalive(&s->s_con);
2972 		ceph_add_cap_releases(mdsc, s);
2973 		if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2974 		    s->s_state == CEPH_MDS_SESSION_HUNG)
2975 			ceph_send_cap_releases(mdsc, s);
2976 		mutex_unlock(&s->s_mutex);
2977 		ceph_put_mds_session(s);
2978 
2979 		mutex_lock(&mdsc->mutex);
2980 	}
2981 	mutex_unlock(&mdsc->mutex);
2982 
2983 	schedule_delayed(mdsc);
2984 }
2985 
2986 int ceph_mdsc_init(struct ceph_fs_client *fsc)
2987 
2988 {
2989 	struct ceph_mds_client *mdsc;
2990 
2991 	mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
2992 	if (!mdsc)
2993 		return -ENOMEM;
2994 	mdsc->fsc = fsc;
2995 	fsc->mdsc = mdsc;
2996 	mutex_init(&mdsc->mutex);
2997 	mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2998 	if (mdsc->mdsmap == NULL)
2999 		return -ENOMEM;
3000 
3001 	init_completion(&mdsc->safe_umount_waiters);
3002 	init_waitqueue_head(&mdsc->session_close_wq);
3003 	INIT_LIST_HEAD(&mdsc->waiting_for_map);
3004 	mdsc->sessions = NULL;
3005 	mdsc->max_sessions = 0;
3006 	mdsc->stopping = 0;
3007 	init_rwsem(&mdsc->snap_rwsem);
3008 	mdsc->snap_realms = RB_ROOT;
3009 	INIT_LIST_HEAD(&mdsc->snap_empty);
3010 	spin_lock_init(&mdsc->snap_empty_lock);
3011 	mdsc->last_tid = 0;
3012 	mdsc->request_tree = RB_ROOT;
3013 	INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3014 	mdsc->last_renew_caps = jiffies;
3015 	INIT_LIST_HEAD(&mdsc->cap_delay_list);
3016 	spin_lock_init(&mdsc->cap_delay_lock);
3017 	INIT_LIST_HEAD(&mdsc->snap_flush_list);
3018 	spin_lock_init(&mdsc->snap_flush_lock);
3019 	mdsc->cap_flush_seq = 0;
3020 	INIT_LIST_HEAD(&mdsc->cap_dirty);
3021 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3022 	mdsc->num_cap_flushing = 0;
3023 	spin_lock_init(&mdsc->cap_dirty_lock);
3024 	init_waitqueue_head(&mdsc->cap_flushing_wq);
3025 	spin_lock_init(&mdsc->dentry_lru_lock);
3026 	INIT_LIST_HEAD(&mdsc->dentry_lru);
3027 
3028 	ceph_caps_init(mdsc);
3029 	ceph_adjust_min_caps(mdsc, fsc->min_caps);
3030 
3031 	return 0;
3032 }
3033 
3034 /*
3035  * Wait for safe replies on open mds requests.  If we time out, drop
3036  * all requests from the tree to avoid dangling dentry refs.
3037  */
3038 static void wait_requests(struct ceph_mds_client *mdsc)
3039 {
3040 	struct ceph_mds_request *req;
3041 	struct ceph_fs_client *fsc = mdsc->fsc;
3042 
3043 	mutex_lock(&mdsc->mutex);
3044 	if (__get_oldest_req(mdsc)) {
3045 		mutex_unlock(&mdsc->mutex);
3046 
3047 		dout("wait_requests waiting for requests\n");
3048 		wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3049 				    fsc->client->options->mount_timeout * HZ);
3050 
3051 		/* tear down remaining requests */
3052 		mutex_lock(&mdsc->mutex);
3053 		while ((req = __get_oldest_req(mdsc))) {
3054 			dout("wait_requests timed out on tid %llu\n",
3055 			     req->r_tid);
3056 			__unregister_request(mdsc, req);
3057 		}
3058 	}
3059 	mutex_unlock(&mdsc->mutex);
3060 	dout("wait_requests done\n");
3061 }
3062 
3063 /*
3064  * called before mount is ro, and before dentries are torn down.
3065  * (hmm, does this still race with new lookups?)
3066  */
3067 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3068 {
3069 	dout("pre_umount\n");
3070 	mdsc->stopping = 1;
3071 
3072 	drop_leases(mdsc);
3073 	ceph_flush_dirty_caps(mdsc);
3074 	wait_requests(mdsc);
3075 
3076 	/*
3077 	 * wait for reply handlers to drop their request refs and
3078 	 * their inode/dcache refs
3079 	 */
3080 	ceph_msgr_flush();
3081 }
3082 
3083 /*
3084  * wait for all write mds requests to flush.
3085  */
3086 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3087 {
3088 	struct ceph_mds_request *req = NULL, *nextreq;
3089 	struct rb_node *n;
3090 
3091 	mutex_lock(&mdsc->mutex);
3092 	dout("wait_unsafe_requests want %lld\n", want_tid);
3093 restart:
3094 	req = __get_oldest_req(mdsc);
3095 	while (req && req->r_tid <= want_tid) {
3096 		/* find next request */
3097 		n = rb_next(&req->r_node);
3098 		if (n)
3099 			nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3100 		else
3101 			nextreq = NULL;
3102 		if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3103 			/* write op */
3104 			ceph_mdsc_get_request(req);
3105 			if (nextreq)
3106 				ceph_mdsc_get_request(nextreq);
3107 			mutex_unlock(&mdsc->mutex);
3108 			dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3109 			     req->r_tid, want_tid);
3110 			wait_for_completion(&req->r_safe_completion);
3111 			mutex_lock(&mdsc->mutex);
3112 			ceph_mdsc_put_request(req);
3113 			if (!nextreq)
3114 				break;  /* next dne before, so we're done! */
3115 			if (RB_EMPTY_NODE(&nextreq->r_node)) {
3116 				/* next request was removed from tree */
3117 				ceph_mdsc_put_request(nextreq);
3118 				goto restart;
3119 			}
3120 			ceph_mdsc_put_request(nextreq);  /* won't go away */
3121 		}
3122 		req = nextreq;
3123 	}
3124 	mutex_unlock(&mdsc->mutex);
3125 	dout("wait_unsafe_requests done\n");
3126 }
3127 
3128 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3129 {
3130 	u64 want_tid, want_flush;
3131 
3132 	if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3133 		return;
3134 
3135 	dout("sync\n");
3136 	mutex_lock(&mdsc->mutex);
3137 	want_tid = mdsc->last_tid;
3138 	want_flush = mdsc->cap_flush_seq;
3139 	mutex_unlock(&mdsc->mutex);
3140 	dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3141 
3142 	ceph_flush_dirty_caps(mdsc);
3143 
3144 	wait_unsafe_requests(mdsc, want_tid);
3145 	wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3146 }
3147 
3148 /*
3149  * true if all sessions are closed, or we force unmount
3150  */
3151 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3152 {
3153 	int i, n = 0;
3154 
3155 	if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3156 		return true;
3157 
3158 	mutex_lock(&mdsc->mutex);
3159 	for (i = 0; i < mdsc->max_sessions; i++)
3160 		if (mdsc->sessions[i])
3161 			n++;
3162 	mutex_unlock(&mdsc->mutex);
3163 	return n == 0;
3164 }
3165 
3166 /*
3167  * called after sb is ro.
3168  */
3169 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3170 {
3171 	struct ceph_mds_session *session;
3172 	int i;
3173 	struct ceph_fs_client *fsc = mdsc->fsc;
3174 	unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3175 
3176 	dout("close_sessions\n");
3177 
3178 	/* close sessions */
3179 	mutex_lock(&mdsc->mutex);
3180 	for (i = 0; i < mdsc->max_sessions; i++) {
3181 		session = __ceph_lookup_mds_session(mdsc, i);
3182 		if (!session)
3183 			continue;
3184 		mutex_unlock(&mdsc->mutex);
3185 		mutex_lock(&session->s_mutex);
3186 		__close_session(mdsc, session);
3187 		mutex_unlock(&session->s_mutex);
3188 		ceph_put_mds_session(session);
3189 		mutex_lock(&mdsc->mutex);
3190 	}
3191 	mutex_unlock(&mdsc->mutex);
3192 
3193 	dout("waiting for sessions to close\n");
3194 	wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3195 			   timeout);
3196 
3197 	/* tear down remaining sessions */
3198 	mutex_lock(&mdsc->mutex);
3199 	for (i = 0; i < mdsc->max_sessions; i++) {
3200 		if (mdsc->sessions[i]) {
3201 			session = get_session(mdsc->sessions[i]);
3202 			__unregister_session(mdsc, session);
3203 			mutex_unlock(&mdsc->mutex);
3204 			mutex_lock(&session->s_mutex);
3205 			remove_session_caps(session);
3206 			mutex_unlock(&session->s_mutex);
3207 			ceph_put_mds_session(session);
3208 			mutex_lock(&mdsc->mutex);
3209 		}
3210 	}
3211 	WARN_ON(!list_empty(&mdsc->cap_delay_list));
3212 	mutex_unlock(&mdsc->mutex);
3213 
3214 	ceph_cleanup_empty_realms(mdsc);
3215 
3216 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3217 
3218 	dout("stopped\n");
3219 }
3220 
3221 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3222 {
3223 	dout("stop\n");
3224 	cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3225 	if (mdsc->mdsmap)
3226 		ceph_mdsmap_destroy(mdsc->mdsmap);
3227 	kfree(mdsc->sessions);
3228 	ceph_caps_finalize(mdsc);
3229 }
3230 
3231 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3232 {
3233 	struct ceph_mds_client *mdsc = fsc->mdsc;
3234 
3235 	dout("mdsc_destroy %p\n", mdsc);
3236 	ceph_mdsc_stop(mdsc);
3237 
3238 	/* flush out any connection work with references to us */
3239 	ceph_msgr_flush();
3240 
3241 	fsc->mdsc = NULL;
3242 	kfree(mdsc);
3243 	dout("mdsc_destroy %p done\n", mdsc);
3244 }
3245 
3246 
3247 /*
3248  * handle mds map update.
3249  */
3250 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3251 {
3252 	u32 epoch;
3253 	u32 maplen;
3254 	void *p = msg->front.iov_base;
3255 	void *end = p + msg->front.iov_len;
3256 	struct ceph_mdsmap *newmap, *oldmap;
3257 	struct ceph_fsid fsid;
3258 	int err = -EINVAL;
3259 
3260 	ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3261 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
3262 	if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3263 		return;
3264 	epoch = ceph_decode_32(&p);
3265 	maplen = ceph_decode_32(&p);
3266 	dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3267 
3268 	/* do we need it? */
3269 	ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3270 	mutex_lock(&mdsc->mutex);
3271 	if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3272 		dout("handle_map epoch %u <= our %u\n",
3273 		     epoch, mdsc->mdsmap->m_epoch);
3274 		mutex_unlock(&mdsc->mutex);
3275 		return;
3276 	}
3277 
3278 	newmap = ceph_mdsmap_decode(&p, end);
3279 	if (IS_ERR(newmap)) {
3280 		err = PTR_ERR(newmap);
3281 		goto bad_unlock;
3282 	}
3283 
3284 	/* swap into place */
3285 	if (mdsc->mdsmap) {
3286 		oldmap = mdsc->mdsmap;
3287 		mdsc->mdsmap = newmap;
3288 		check_new_map(mdsc, newmap, oldmap);
3289 		ceph_mdsmap_destroy(oldmap);
3290 	} else {
3291 		mdsc->mdsmap = newmap;  /* first mds map */
3292 	}
3293 	mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3294 
3295 	__wake_requests(mdsc, &mdsc->waiting_for_map);
3296 
3297 	mutex_unlock(&mdsc->mutex);
3298 	schedule_delayed(mdsc);
3299 	return;
3300 
3301 bad_unlock:
3302 	mutex_unlock(&mdsc->mutex);
3303 bad:
3304 	pr_err("error decoding mdsmap %d\n", err);
3305 	return;
3306 }
3307 
3308 static struct ceph_connection *con_get(struct ceph_connection *con)
3309 {
3310 	struct ceph_mds_session *s = con->private;
3311 
3312 	if (get_session(s)) {
3313 		dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3314 		return con;
3315 	}
3316 	dout("mdsc con_get %p FAIL\n", s);
3317 	return NULL;
3318 }
3319 
3320 static void con_put(struct ceph_connection *con)
3321 {
3322 	struct ceph_mds_session *s = con->private;
3323 
3324 	dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3325 	ceph_put_mds_session(s);
3326 }
3327 
3328 /*
3329  * if the client is unresponsive for long enough, the mds will kill
3330  * the session entirely.
3331  */
3332 static void peer_reset(struct ceph_connection *con)
3333 {
3334 	struct ceph_mds_session *s = con->private;
3335 	struct ceph_mds_client *mdsc = s->s_mdsc;
3336 
3337 	pr_warning("mds%d closed our session\n", s->s_mds);
3338 	send_mds_reconnect(mdsc, s);
3339 }
3340 
3341 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3342 {
3343 	struct ceph_mds_session *s = con->private;
3344 	struct ceph_mds_client *mdsc = s->s_mdsc;
3345 	int type = le16_to_cpu(msg->hdr.type);
3346 
3347 	mutex_lock(&mdsc->mutex);
3348 	if (__verify_registered_session(mdsc, s) < 0) {
3349 		mutex_unlock(&mdsc->mutex);
3350 		goto out;
3351 	}
3352 	mutex_unlock(&mdsc->mutex);
3353 
3354 	switch (type) {
3355 	case CEPH_MSG_MDS_MAP:
3356 		ceph_mdsc_handle_map(mdsc, msg);
3357 		break;
3358 	case CEPH_MSG_CLIENT_SESSION:
3359 		handle_session(s, msg);
3360 		break;
3361 	case CEPH_MSG_CLIENT_REPLY:
3362 		handle_reply(s, msg);
3363 		break;
3364 	case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3365 		handle_forward(mdsc, s, msg);
3366 		break;
3367 	case CEPH_MSG_CLIENT_CAPS:
3368 		ceph_handle_caps(s, msg);
3369 		break;
3370 	case CEPH_MSG_CLIENT_SNAP:
3371 		ceph_handle_snap(mdsc, s, msg);
3372 		break;
3373 	case CEPH_MSG_CLIENT_LEASE:
3374 		handle_lease(mdsc, s, msg);
3375 		break;
3376 
3377 	default:
3378 		pr_err("received unknown message type %d %s\n", type,
3379 		       ceph_msg_type_name(type));
3380 	}
3381 out:
3382 	ceph_msg_put(msg);
3383 }
3384 
3385 /*
3386  * authentication
3387  */
3388 
3389 /*
3390  * Note: returned pointer is the address of a structure that's
3391  * managed separately.  Caller must *not* attempt to free it.
3392  */
3393 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3394 					int *proto, int force_new)
3395 {
3396 	struct ceph_mds_session *s = con->private;
3397 	struct ceph_mds_client *mdsc = s->s_mdsc;
3398 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3399 	struct ceph_auth_handshake *auth = &s->s_auth;
3400 
3401 	if (force_new && auth->authorizer) {
3402 		if (ac->ops && ac->ops->destroy_authorizer)
3403 			ac->ops->destroy_authorizer(ac, auth->authorizer);
3404 		auth->authorizer = NULL;
3405 	}
3406 	if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
3407 		int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3408 							auth);
3409 		if (ret)
3410 			return ERR_PTR(ret);
3411 	}
3412 	*proto = ac->protocol;
3413 
3414 	return auth;
3415 }
3416 
3417 
3418 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3419 {
3420 	struct ceph_mds_session *s = con->private;
3421 	struct ceph_mds_client *mdsc = s->s_mdsc;
3422 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3423 
3424 	return ac->ops->verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3425 }
3426 
3427 static int invalidate_authorizer(struct ceph_connection *con)
3428 {
3429 	struct ceph_mds_session *s = con->private;
3430 	struct ceph_mds_client *mdsc = s->s_mdsc;
3431 	struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3432 
3433 	if (ac->ops->invalidate_authorizer)
3434 		ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3435 
3436 	return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3437 }
3438 
3439 static const struct ceph_connection_operations mds_con_ops = {
3440 	.get = con_get,
3441 	.put = con_put,
3442 	.dispatch = dispatch,
3443 	.get_authorizer = get_authorizer,
3444 	.verify_authorizer_reply = verify_authorizer_reply,
3445 	.invalidate_authorizer = invalidate_authorizer,
3446 	.peer_reset = peer_reset,
3447 };
3448 
3449 /* eof */
3450