627cf0326b97b2e00f05eed620382f0d627992b0
[platform/kernel/linux-rpi.git] / fs / ceph / mds_client.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3
4 #include <linux/fs.h>
5 #include <linux/wait.h>
6 #include <linux/slab.h>
7 #include <linux/gfp.h>
8 #include <linux/sched.h>
9 #include <linux/debugfs.h>
10 #include <linux/seq_file.h>
11 #include <linux/ratelimit.h>
12
13 #include "super.h"
14 #include "mds_client.h"
15
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
22
23 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
24
25 /*
26  * A cluster of MDS (metadata server) daemons is responsible for
27  * managing the file system namespace (the directory hierarchy and
28  * inodes) and for coordinating shared access to storage.  Metadata is
29  * partitioning hierarchically across a number of servers, and that
30  * partition varies over time as the cluster adjusts the distribution
31  * in order to balance load.
32  *
33  * The MDS client is primarily responsible to managing synchronous
34  * metadata requests for operations like open, unlink, and so forth.
35  * If there is a MDS failure, we find out about it when we (possibly
36  * request and) receive a new MDS map, and can resubmit affected
37  * requests.
38  *
39  * For the most part, though, we take advantage of a lossless
40  * communications channel to the MDS, and do not need to worry about
41  * timing out or resubmitting requests.
42  *
43  * We maintain a stateful "session" with each MDS we interact with.
44  * Within each session, we sent periodic heartbeat messages to ensure
45  * any capabilities or leases we have been issues remain valid.  If
46  * the session times out and goes stale, our leases and capabilities
47  * are no longer valid.
48  */
49
50 struct ceph_reconnect_state {
51         struct ceph_mds_session *session;
52         int nr_caps, nr_realms;
53         struct ceph_pagelist *pagelist;
54         unsigned msg_version;
55         bool allow_multi;
56 };
57
58 static void __wake_requests(struct ceph_mds_client *mdsc,
59                             struct list_head *head);
60 static void ceph_cap_release_work(struct work_struct *work);
61 static void ceph_cap_reclaim_work(struct work_struct *work);
62
63 static const struct ceph_connection_operations mds_con_ops;
64
65
66 /*
67  * mds reply parsing
68  */
69
70 static int parse_reply_info_quota(void **p, void *end,
71                                   struct ceph_mds_reply_info_in *info)
72 {
73         u8 struct_v, struct_compat;
74         u32 struct_len;
75
76         ceph_decode_8_safe(p, end, struct_v, bad);
77         ceph_decode_8_safe(p, end, struct_compat, bad);
78         /* struct_v is expected to be >= 1. we only
79          * understand encoding with struct_compat == 1. */
80         if (!struct_v || struct_compat != 1)
81                 goto bad;
82         ceph_decode_32_safe(p, end, struct_len, bad);
83         ceph_decode_need(p, end, struct_len, bad);
84         end = *p + struct_len;
85         ceph_decode_64_safe(p, end, info->max_bytes, bad);
86         ceph_decode_64_safe(p, end, info->max_files, bad);
87         *p = end;
88         return 0;
89 bad:
90         return -EIO;
91 }
92
93 /*
94  * parse individual inode info
95  */
96 static int parse_reply_info_in(void **p, void *end,
97                                struct ceph_mds_reply_info_in *info,
98                                u64 features)
99 {
100         int err = 0;
101         u8 struct_v = 0;
102
103         if (features == (u64)-1) {
104                 u32 struct_len;
105                 u8 struct_compat;
106                 ceph_decode_8_safe(p, end, struct_v, bad);
107                 ceph_decode_8_safe(p, end, struct_compat, bad);
108                 /* struct_v is expected to be >= 1. we only understand
109                  * encoding with struct_compat == 1. */
110                 if (!struct_v || struct_compat != 1)
111                         goto bad;
112                 ceph_decode_32_safe(p, end, struct_len, bad);
113                 ceph_decode_need(p, end, struct_len, bad);
114                 end = *p + struct_len;
115         }
116
117         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
118         info->in = *p;
119         *p += sizeof(struct ceph_mds_reply_inode) +
120                 sizeof(*info->in->fragtree.splits) *
121                 le32_to_cpu(info->in->fragtree.nsplits);
122
123         ceph_decode_32_safe(p, end, info->symlink_len, bad);
124         ceph_decode_need(p, end, info->symlink_len, bad);
125         info->symlink = *p;
126         *p += info->symlink_len;
127
128         ceph_decode_copy_safe(p, end, &info->dir_layout,
129                               sizeof(info->dir_layout), bad);
130         ceph_decode_32_safe(p, end, info->xattr_len, bad);
131         ceph_decode_need(p, end, info->xattr_len, bad);
132         info->xattr_data = *p;
133         *p += info->xattr_len;
134
135         if (features == (u64)-1) {
136                 /* inline data */
137                 ceph_decode_64_safe(p, end, info->inline_version, bad);
138                 ceph_decode_32_safe(p, end, info->inline_len, bad);
139                 ceph_decode_need(p, end, info->inline_len, bad);
140                 info->inline_data = *p;
141                 *p += info->inline_len;
142                 /* quota */
143                 err = parse_reply_info_quota(p, end, info);
144                 if (err < 0)
145                         goto out_bad;
146                 /* pool namespace */
147                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
148                 if (info->pool_ns_len > 0) {
149                         ceph_decode_need(p, end, info->pool_ns_len, bad);
150                         info->pool_ns_data = *p;
151                         *p += info->pool_ns_len;
152                 }
153
154                 /* btime */
155                 ceph_decode_need(p, end, sizeof(info->btime), bad);
156                 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
157
158                 /* change attribute */
159                 ceph_decode_64_safe(p, end, info->change_attr, bad);
160
161                 /* dir pin */
162                 if (struct_v >= 2) {
163                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
164                 } else {
165                         info->dir_pin = -ENODATA;
166                 }
167
168                 /* snapshot birth time, remains zero for v<=2 */
169                 if (struct_v >= 3) {
170                         ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
171                         ceph_decode_copy(p, &info->snap_btime,
172                                          sizeof(info->snap_btime));
173                 } else {
174                         memset(&info->snap_btime, 0, sizeof(info->snap_btime));
175                 }
176
177                 *p = end;
178         } else {
179                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
180                         ceph_decode_64_safe(p, end, info->inline_version, bad);
181                         ceph_decode_32_safe(p, end, info->inline_len, bad);
182                         ceph_decode_need(p, end, info->inline_len, bad);
183                         info->inline_data = *p;
184                         *p += info->inline_len;
185                 } else
186                         info->inline_version = CEPH_INLINE_NONE;
187
188                 if (features & CEPH_FEATURE_MDS_QUOTA) {
189                         err = parse_reply_info_quota(p, end, info);
190                         if (err < 0)
191                                 goto out_bad;
192                 } else {
193                         info->max_bytes = 0;
194                         info->max_files = 0;
195                 }
196
197                 info->pool_ns_len = 0;
198                 info->pool_ns_data = NULL;
199                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
200                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
201                         if (info->pool_ns_len > 0) {
202                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
203                                 info->pool_ns_data = *p;
204                                 *p += info->pool_ns_len;
205                         }
206                 }
207
208                 if (features & CEPH_FEATURE_FS_BTIME) {
209                         ceph_decode_need(p, end, sizeof(info->btime), bad);
210                         ceph_decode_copy(p, &info->btime, sizeof(info->btime));
211                         ceph_decode_64_safe(p, end, info->change_attr, bad);
212                 }
213
214                 info->dir_pin = -ENODATA;
215                 /* info->snap_btime remains zero */
216         }
217         return 0;
218 bad:
219         err = -EIO;
220 out_bad:
221         return err;
222 }
223
224 static int parse_reply_info_dir(void **p, void *end,
225                                 struct ceph_mds_reply_dirfrag **dirfrag,
226                                 u64 features)
227 {
228         if (features == (u64)-1) {
229                 u8 struct_v, struct_compat;
230                 u32 struct_len;
231                 ceph_decode_8_safe(p, end, struct_v, bad);
232                 ceph_decode_8_safe(p, end, struct_compat, bad);
233                 /* struct_v is expected to be >= 1. we only understand
234                  * encoding whose struct_compat == 1. */
235                 if (!struct_v || struct_compat != 1)
236                         goto bad;
237                 ceph_decode_32_safe(p, end, struct_len, bad);
238                 ceph_decode_need(p, end, struct_len, bad);
239                 end = *p + struct_len;
240         }
241
242         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
243         *dirfrag = *p;
244         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
245         if (unlikely(*p > end))
246                 goto bad;
247         if (features == (u64)-1)
248                 *p = end;
249         return 0;
250 bad:
251         return -EIO;
252 }
253
254 static int parse_reply_info_lease(void **p, void *end,
255                                   struct ceph_mds_reply_lease **lease,
256                                   u64 features)
257 {
258         if (features == (u64)-1) {
259                 u8 struct_v, struct_compat;
260                 u32 struct_len;
261                 ceph_decode_8_safe(p, end, struct_v, bad);
262                 ceph_decode_8_safe(p, end, struct_compat, bad);
263                 /* struct_v is expected to be >= 1. we only understand
264                  * encoding whose struct_compat == 1. */
265                 if (!struct_v || struct_compat != 1)
266                         goto bad;
267                 ceph_decode_32_safe(p, end, struct_len, bad);
268                 ceph_decode_need(p, end, struct_len, bad);
269                 end = *p + struct_len;
270         }
271
272         ceph_decode_need(p, end, sizeof(**lease), bad);
273         *lease = *p;
274         *p += sizeof(**lease);
275         if (features == (u64)-1)
276                 *p = end;
277         return 0;
278 bad:
279         return -EIO;
280 }
281
282 /*
283  * parse a normal reply, which may contain a (dir+)dentry and/or a
284  * target inode.
285  */
286 static int parse_reply_info_trace(void **p, void *end,
287                                   struct ceph_mds_reply_info_parsed *info,
288                                   u64 features)
289 {
290         int err;
291
292         if (info->head->is_dentry) {
293                 err = parse_reply_info_in(p, end, &info->diri, features);
294                 if (err < 0)
295                         goto out_bad;
296
297                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
298                 if (err < 0)
299                         goto out_bad;
300
301                 ceph_decode_32_safe(p, end, info->dname_len, bad);
302                 ceph_decode_need(p, end, info->dname_len, bad);
303                 info->dname = *p;
304                 *p += info->dname_len;
305
306                 err = parse_reply_info_lease(p, end, &info->dlease, features);
307                 if (err < 0)
308                         goto out_bad;
309         }
310
311         if (info->head->is_target) {
312                 err = parse_reply_info_in(p, end, &info->targeti, features);
313                 if (err < 0)
314                         goto out_bad;
315         }
316
317         if (unlikely(*p != end))
318                 goto bad;
319         return 0;
320
321 bad:
322         err = -EIO;
323 out_bad:
324         pr_err("problem parsing mds trace %d\n", err);
325         return err;
326 }
327
328 /*
329  * parse readdir results
330  */
331 static int parse_reply_info_readdir(void **p, void *end,
332                                 struct ceph_mds_reply_info_parsed *info,
333                                 u64 features)
334 {
335         u32 num, i = 0;
336         int err;
337
338         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
339         if (err < 0)
340                 goto out_bad;
341
342         ceph_decode_need(p, end, sizeof(num) + 2, bad);
343         num = ceph_decode_32(p);
344         {
345                 u16 flags = ceph_decode_16(p);
346                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
347                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
348                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
349                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
350         }
351         if (num == 0)
352                 goto done;
353
354         BUG_ON(!info->dir_entries);
355         if ((unsigned long)(info->dir_entries + num) >
356             (unsigned long)info->dir_entries + info->dir_buf_size) {
357                 pr_err("dir contents are larger than expected\n");
358                 WARN_ON(1);
359                 goto bad;
360         }
361
362         info->dir_nr = num;
363         while (num) {
364                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
365                 /* dentry */
366                 ceph_decode_32_safe(p, end, rde->name_len, bad);
367                 ceph_decode_need(p, end, rde->name_len, bad);
368                 rde->name = *p;
369                 *p += rde->name_len;
370                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
371
372                 /* dentry lease */
373                 err = parse_reply_info_lease(p, end, &rde->lease, features);
374                 if (err)
375                         goto out_bad;
376                 /* inode */
377                 err = parse_reply_info_in(p, end, &rde->inode, features);
378                 if (err < 0)
379                         goto out_bad;
380                 /* ceph_readdir_prepopulate() will update it */
381                 rde->offset = 0;
382                 i++;
383                 num--;
384         }
385
386 done:
387         /* Skip over any unrecognized fields */
388         *p = end;
389         return 0;
390
391 bad:
392         err = -EIO;
393 out_bad:
394         pr_err("problem parsing dir contents %d\n", err);
395         return err;
396 }
397
398 /*
399  * parse fcntl F_GETLK results
400  */
401 static int parse_reply_info_filelock(void **p, void *end,
402                                      struct ceph_mds_reply_info_parsed *info,
403                                      u64 features)
404 {
405         if (*p + sizeof(*info->filelock_reply) > end)
406                 goto bad;
407
408         info->filelock_reply = *p;
409
410         /* Skip over any unrecognized fields */
411         *p = end;
412         return 0;
413 bad:
414         return -EIO;
415 }
416
417 /*
418  * parse create results
419  */
420 static int parse_reply_info_create(void **p, void *end,
421                                   struct ceph_mds_reply_info_parsed *info,
422                                   u64 features)
423 {
424         if (features == (u64)-1 ||
425             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
426                 /* Malformed reply? */
427                 if (*p == end) {
428                         info->has_create_ino = false;
429                 } else {
430                         info->has_create_ino = true;
431                         ceph_decode_64_safe(p, end, info->ino, bad);
432                 }
433         } else {
434                 if (*p != end)
435                         goto bad;
436         }
437
438         /* Skip over any unrecognized fields */
439         *p = end;
440         return 0;
441 bad:
442         return -EIO;
443 }
444
445 /*
446  * parse extra results
447  */
448 static int parse_reply_info_extra(void **p, void *end,
449                                   struct ceph_mds_reply_info_parsed *info,
450                                   u64 features)
451 {
452         u32 op = le32_to_cpu(info->head->op);
453
454         if (op == CEPH_MDS_OP_GETFILELOCK)
455                 return parse_reply_info_filelock(p, end, info, features);
456         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
457                 return parse_reply_info_readdir(p, end, info, features);
458         else if (op == CEPH_MDS_OP_CREATE)
459                 return parse_reply_info_create(p, end, info, features);
460         else
461                 return -EIO;
462 }
463
464 /*
465  * parse entire mds reply
466  */
467 static int parse_reply_info(struct ceph_msg *msg,
468                             struct ceph_mds_reply_info_parsed *info,
469                             u64 features)
470 {
471         void *p, *end;
472         u32 len;
473         int err;
474
475         info->head = msg->front.iov_base;
476         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
477         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
478
479         /* trace */
480         ceph_decode_32_safe(&p, end, len, bad);
481         if (len > 0) {
482                 ceph_decode_need(&p, end, len, bad);
483                 err = parse_reply_info_trace(&p, p+len, info, features);
484                 if (err < 0)
485                         goto out_bad;
486         }
487
488         /* extra */
489         ceph_decode_32_safe(&p, end, len, bad);
490         if (len > 0) {
491                 ceph_decode_need(&p, end, len, bad);
492                 err = parse_reply_info_extra(&p, p+len, info, features);
493                 if (err < 0)
494                         goto out_bad;
495         }
496
497         /* snap blob */
498         ceph_decode_32_safe(&p, end, len, bad);
499         info->snapblob_len = len;
500         info->snapblob = p;
501         p += len;
502
503         if (p != end)
504                 goto bad;
505         return 0;
506
507 bad:
508         err = -EIO;
509 out_bad:
510         pr_err("mds parse_reply err %d\n", err);
511         return err;
512 }
513
514 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
515 {
516         if (!info->dir_entries)
517                 return;
518         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
519 }
520
521
522 /*
523  * sessions
524  */
525 const char *ceph_session_state_name(int s)
526 {
527         switch (s) {
528         case CEPH_MDS_SESSION_NEW: return "new";
529         case CEPH_MDS_SESSION_OPENING: return "opening";
530         case CEPH_MDS_SESSION_OPEN: return "open";
531         case CEPH_MDS_SESSION_HUNG: return "hung";
532         case CEPH_MDS_SESSION_CLOSING: return "closing";
533         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
534         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
535         case CEPH_MDS_SESSION_REJECTED: return "rejected";
536         default: return "???";
537         }
538 }
539
540 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
541 {
542         if (refcount_inc_not_zero(&s->s_ref)) {
543                 dout("mdsc get_session %p %d -> %d\n", s,
544                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
545                 return s;
546         } else {
547                 dout("mdsc get_session %p 0 -- FAIL\n", s);
548                 return NULL;
549         }
550 }
551
552 void ceph_put_mds_session(struct ceph_mds_session *s)
553 {
554         dout("mdsc put_session %p %d -> %d\n", s,
555              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
556         if (refcount_dec_and_test(&s->s_ref)) {
557                 if (s->s_auth.authorizer)
558                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
559                 kfree(s);
560         }
561 }
562
563 /*
564  * called under mdsc->mutex
565  */
566 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
567                                                    int mds)
568 {
569         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
570                 return NULL;
571         return get_session(mdsc->sessions[mds]);
572 }
573
574 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
575 {
576         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
577                 return false;
578         else
579                 return true;
580 }
581
582 static int __verify_registered_session(struct ceph_mds_client *mdsc,
583                                        struct ceph_mds_session *s)
584 {
585         if (s->s_mds >= mdsc->max_sessions ||
586             mdsc->sessions[s->s_mds] != s)
587                 return -ENOENT;
588         return 0;
589 }
590
591 /*
592  * create+register a new session for given mds.
593  * called under mdsc->mutex.
594  */
595 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
596                                                  int mds)
597 {
598         struct ceph_mds_session *s;
599
600         if (mds >= mdsc->mdsmap->m_num_mds)
601                 return ERR_PTR(-EINVAL);
602
603         s = kzalloc(sizeof(*s), GFP_NOFS);
604         if (!s)
605                 return ERR_PTR(-ENOMEM);
606
607         if (mds >= mdsc->max_sessions) {
608                 int newmax = 1 << get_count_order(mds + 1);
609                 struct ceph_mds_session **sa;
610
611                 dout("%s: realloc to %d\n", __func__, newmax);
612                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
613                 if (!sa)
614                         goto fail_realloc;
615                 if (mdsc->sessions) {
616                         memcpy(sa, mdsc->sessions,
617                                mdsc->max_sessions * sizeof(void *));
618                         kfree(mdsc->sessions);
619                 }
620                 mdsc->sessions = sa;
621                 mdsc->max_sessions = newmax;
622         }
623
624         dout("%s: mds%d\n", __func__, mds);
625         s->s_mdsc = mdsc;
626         s->s_mds = mds;
627         s->s_state = CEPH_MDS_SESSION_NEW;
628         s->s_ttl = 0;
629         s->s_seq = 0;
630         mutex_init(&s->s_mutex);
631
632         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
633
634         spin_lock_init(&s->s_gen_ttl_lock);
635         s->s_cap_gen = 1;
636         s->s_cap_ttl = jiffies - 1;
637
638         spin_lock_init(&s->s_cap_lock);
639         s->s_renew_requested = 0;
640         s->s_renew_seq = 0;
641         INIT_LIST_HEAD(&s->s_caps);
642         s->s_nr_caps = 0;
643         refcount_set(&s->s_ref, 1);
644         INIT_LIST_HEAD(&s->s_waiting);
645         INIT_LIST_HEAD(&s->s_unsafe);
646         s->s_num_cap_releases = 0;
647         s->s_cap_reconnect = 0;
648         s->s_cap_iterator = NULL;
649         INIT_LIST_HEAD(&s->s_cap_releases);
650         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
651
652         INIT_LIST_HEAD(&s->s_cap_flushing);
653
654         mdsc->sessions[mds] = s;
655         atomic_inc(&mdsc->num_sessions);
656         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
657
658         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
659                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
660
661         return s;
662
663 fail_realloc:
664         kfree(s);
665         return ERR_PTR(-ENOMEM);
666 }
667
668 /*
669  * called under mdsc->mutex
670  */
671 static void __unregister_session(struct ceph_mds_client *mdsc,
672                                struct ceph_mds_session *s)
673 {
674         dout("__unregister_session mds%d %p\n", s->s_mds, s);
675         BUG_ON(mdsc->sessions[s->s_mds] != s);
676         mdsc->sessions[s->s_mds] = NULL;
677         s->s_state = 0;
678         ceph_con_close(&s->s_con);
679         ceph_put_mds_session(s);
680         atomic_dec(&mdsc->num_sessions);
681 }
682
683 /*
684  * drop session refs in request.
685  *
686  * should be last request ref, or hold mdsc->mutex
687  */
688 static void put_request_session(struct ceph_mds_request *req)
689 {
690         if (req->r_session) {
691                 ceph_put_mds_session(req->r_session);
692                 req->r_session = NULL;
693         }
694 }
695
696 void ceph_mdsc_release_request(struct kref *kref)
697 {
698         struct ceph_mds_request *req = container_of(kref,
699                                                     struct ceph_mds_request,
700                                                     r_kref);
701         destroy_reply_info(&req->r_reply_info);
702         if (req->r_request)
703                 ceph_msg_put(req->r_request);
704         if (req->r_reply)
705                 ceph_msg_put(req->r_reply);
706         if (req->r_inode) {
707                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
708                 /* avoid calling iput_final() in mds dispatch threads */
709                 ceph_async_iput(req->r_inode);
710         }
711         if (req->r_parent) {
712                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
713                 ceph_async_iput(req->r_parent);
714         }
715         ceph_async_iput(req->r_target_inode);
716         if (req->r_dentry)
717                 dput(req->r_dentry);
718         if (req->r_old_dentry)
719                 dput(req->r_old_dentry);
720         if (req->r_old_dentry_dir) {
721                 /*
722                  * track (and drop pins for) r_old_dentry_dir
723                  * separately, since r_old_dentry's d_parent may have
724                  * changed between the dir mutex being dropped and
725                  * this request being freed.
726                  */
727                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
728                                   CEPH_CAP_PIN);
729                 ceph_async_iput(req->r_old_dentry_dir);
730         }
731         kfree(req->r_path1);
732         kfree(req->r_path2);
733         if (req->r_pagelist)
734                 ceph_pagelist_release(req->r_pagelist);
735         put_request_session(req);
736         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
737         WARN_ON_ONCE(!list_empty(&req->r_wait));
738         kfree(req);
739 }
740
741 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
742
743 /*
744  * lookup session, bump ref if found.
745  *
746  * called under mdsc->mutex.
747  */
748 static struct ceph_mds_request *
749 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
750 {
751         struct ceph_mds_request *req;
752
753         req = lookup_request(&mdsc->request_tree, tid);
754         if (req)
755                 ceph_mdsc_get_request(req);
756
757         return req;
758 }
759
760 /*
761  * Register an in-flight request, and assign a tid.  Link to directory
762  * are modifying (if any).
763  *
764  * Called under mdsc->mutex.
765  */
766 static void __register_request(struct ceph_mds_client *mdsc,
767                                struct ceph_mds_request *req,
768                                struct inode *dir)
769 {
770         int ret = 0;
771
772         req->r_tid = ++mdsc->last_tid;
773         if (req->r_num_caps) {
774                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
775                                         req->r_num_caps);
776                 if (ret < 0) {
777                         pr_err("__register_request %p "
778                                "failed to reserve caps: %d\n", req, ret);
779                         /* set req->r_err to fail early from __do_request */
780                         req->r_err = ret;
781                         return;
782                 }
783         }
784         dout("__register_request %p tid %lld\n", req, req->r_tid);
785         ceph_mdsc_get_request(req);
786         insert_request(&mdsc->request_tree, req);
787
788         req->r_uid = current_fsuid();
789         req->r_gid = current_fsgid();
790
791         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
792                 mdsc->oldest_tid = req->r_tid;
793
794         if (dir) {
795                 ihold(dir);
796                 req->r_unsafe_dir = dir;
797         }
798 }
799
800 static void __unregister_request(struct ceph_mds_client *mdsc,
801                                  struct ceph_mds_request *req)
802 {
803         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
804
805         /* Never leave an unregistered request on an unsafe list! */
806         list_del_init(&req->r_unsafe_item);
807
808         if (req->r_tid == mdsc->oldest_tid) {
809                 struct rb_node *p = rb_next(&req->r_node);
810                 mdsc->oldest_tid = 0;
811                 while (p) {
812                         struct ceph_mds_request *next_req =
813                                 rb_entry(p, struct ceph_mds_request, r_node);
814                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
815                                 mdsc->oldest_tid = next_req->r_tid;
816                                 break;
817                         }
818                         p = rb_next(p);
819                 }
820         }
821
822         erase_request(&mdsc->request_tree, req);
823
824         if (req->r_unsafe_dir  &&
825             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
826                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
827                 spin_lock(&ci->i_unsafe_lock);
828                 list_del_init(&req->r_unsafe_dir_item);
829                 spin_unlock(&ci->i_unsafe_lock);
830         }
831         if (req->r_target_inode &&
832             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
833                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
834                 spin_lock(&ci->i_unsafe_lock);
835                 list_del_init(&req->r_unsafe_target_item);
836                 spin_unlock(&ci->i_unsafe_lock);
837         }
838
839         if (req->r_unsafe_dir) {
840                 /* avoid calling iput_final() in mds dispatch threads */
841                 ceph_async_iput(req->r_unsafe_dir);
842                 req->r_unsafe_dir = NULL;
843         }
844
845         complete_all(&req->r_safe_completion);
846
847         ceph_mdsc_put_request(req);
848 }
849
850 /*
851  * Walk back up the dentry tree until we hit a dentry representing a
852  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
853  * when calling this) to ensure that the objects won't disappear while we're
854  * working with them. Once we hit a candidate dentry, we attempt to take a
855  * reference to it, and return that as the result.
856  */
857 static struct inode *get_nonsnap_parent(struct dentry *dentry)
858 {
859         struct inode *inode = NULL;
860
861         while (dentry && !IS_ROOT(dentry)) {
862                 inode = d_inode_rcu(dentry);
863                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
864                         break;
865                 dentry = dentry->d_parent;
866         }
867         if (inode)
868                 inode = igrab(inode);
869         return inode;
870 }
871
872 /*
873  * Choose mds to send request to next.  If there is a hint set in the
874  * request (e.g., due to a prior forward hint from the mds), use that.
875  * Otherwise, consult frag tree and/or caps to identify the
876  * appropriate mds.  If all else fails, choose randomly.
877  *
878  * Called under mdsc->mutex.
879  */
880 static int __choose_mds(struct ceph_mds_client *mdsc,
881                         struct ceph_mds_request *req)
882 {
883         struct inode *inode;
884         struct ceph_inode_info *ci;
885         struct ceph_cap *cap;
886         int mode = req->r_direct_mode;
887         int mds = -1;
888         u32 hash = req->r_direct_hash;
889         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
890
891         /*
892          * is there a specific mds we should try?  ignore hint if we have
893          * no session and the mds is not up (active or recovering).
894          */
895         if (req->r_resend_mds >= 0 &&
896             (__have_session(mdsc, req->r_resend_mds) ||
897              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
898                 dout("choose_mds using resend_mds mds%d\n",
899                      req->r_resend_mds);
900                 return req->r_resend_mds;
901         }
902
903         if (mode == USE_RANDOM_MDS)
904                 goto random;
905
906         inode = NULL;
907         if (req->r_inode) {
908                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
909                         inode = req->r_inode;
910                         ihold(inode);
911                 } else {
912                         /* req->r_dentry is non-null for LSSNAP request */
913                         rcu_read_lock();
914                         inode = get_nonsnap_parent(req->r_dentry);
915                         rcu_read_unlock();
916                         dout("__choose_mds using snapdir's parent %p\n", inode);
917                 }
918         } else if (req->r_dentry) {
919                 /* ignore race with rename; old or new d_parent is okay */
920                 struct dentry *parent;
921                 struct inode *dir;
922
923                 rcu_read_lock();
924                 parent = READ_ONCE(req->r_dentry->d_parent);
925                 dir = req->r_parent ? : d_inode_rcu(parent);
926
927                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
928                         /*  not this fs or parent went negative */
929                         inode = d_inode(req->r_dentry);
930                         if (inode)
931                                 ihold(inode);
932                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
933                         /* direct snapped/virtual snapdir requests
934                          * based on parent dir inode */
935                         inode = get_nonsnap_parent(parent);
936                         dout("__choose_mds using nonsnap parent %p\n", inode);
937                 } else {
938                         /* dentry target */
939                         inode = d_inode(req->r_dentry);
940                         if (!inode || mode == USE_AUTH_MDS) {
941                                 /* dir + name */
942                                 inode = igrab(dir);
943                                 hash = ceph_dentry_hash(dir, req->r_dentry);
944                                 is_hash = true;
945                         } else {
946                                 ihold(inode);
947                         }
948                 }
949                 rcu_read_unlock();
950         }
951
952         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
953              (int)hash, mode);
954         if (!inode)
955                 goto random;
956         ci = ceph_inode(inode);
957
958         if (is_hash && S_ISDIR(inode->i_mode)) {
959                 struct ceph_inode_frag frag;
960                 int found;
961
962                 ceph_choose_frag(ci, hash, &frag, &found);
963                 if (found) {
964                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
965                                 u8 r;
966
967                                 /* choose a random replica */
968                                 get_random_bytes(&r, 1);
969                                 r %= frag.ndist;
970                                 mds = frag.dist[r];
971                                 dout("choose_mds %p %llx.%llx "
972                                      "frag %u mds%d (%d/%d)\n",
973                                      inode, ceph_vinop(inode),
974                                      frag.frag, mds,
975                                      (int)r, frag.ndist);
976                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
977                                     CEPH_MDS_STATE_ACTIVE &&
978                                     !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
979                                         goto out;
980                         }
981
982                         /* since this file/dir wasn't known to be
983                          * replicated, then we want to look for the
984                          * authoritative mds. */
985                         if (frag.mds >= 0) {
986                                 /* choose auth mds */
987                                 mds = frag.mds;
988                                 dout("choose_mds %p %llx.%llx "
989                                      "frag %u mds%d (auth)\n",
990                                      inode, ceph_vinop(inode), frag.frag, mds);
991                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
992                                     CEPH_MDS_STATE_ACTIVE) {
993                                         if (mode == USE_ANY_MDS &&
994                                             !ceph_mdsmap_is_laggy(mdsc->mdsmap,
995                                                                   mds))
996                                                 goto out;
997                                 }
998                         }
999                         mode = USE_AUTH_MDS;
1000                 }
1001         }
1002
1003         spin_lock(&ci->i_ceph_lock);
1004         cap = NULL;
1005         if (mode == USE_AUTH_MDS)
1006                 cap = ci->i_auth_cap;
1007         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1008                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1009         if (!cap) {
1010                 spin_unlock(&ci->i_ceph_lock);
1011                 ceph_async_iput(inode);
1012                 goto random;
1013         }
1014         mds = cap->session->s_mds;
1015         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
1016              inode, ceph_vinop(inode), mds,
1017              cap == ci->i_auth_cap ? "auth " : "", cap);
1018         spin_unlock(&ci->i_ceph_lock);
1019 out:
1020         /* avoid calling iput_final() while holding mdsc->mutex or
1021          * in mds dispatch threads */
1022         ceph_async_iput(inode);
1023         return mds;
1024
1025 random:
1026         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1027         dout("choose_mds chose random mds%d\n", mds);
1028         return mds;
1029 }
1030
1031
1032 /*
1033  * session messages
1034  */
1035 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1036 {
1037         struct ceph_msg *msg;
1038         struct ceph_mds_session_head *h;
1039
1040         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1041                            false);
1042         if (!msg) {
1043                 pr_err("create_session_msg ENOMEM creating msg\n");
1044                 return NULL;
1045         }
1046         h = msg->front.iov_base;
1047         h->op = cpu_to_le32(op);
1048         h->seq = cpu_to_le64(seq);
1049
1050         return msg;
1051 }
1052
1053 static void encode_supported_features(void **p, void *end)
1054 {
1055         static const unsigned char bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1056         static const size_t count = ARRAY_SIZE(bits);
1057
1058         if (count > 0) {
1059                 size_t i;
1060                 size_t size = ((size_t)bits[count - 1] + 64) / 64 * 8;
1061
1062                 BUG_ON(*p + 4 + size > end);
1063                 ceph_encode_32(p, size);
1064                 memset(*p, 0, size);
1065                 for (i = 0; i < count; i++)
1066                         ((unsigned char*)(*p))[i / 8] |= 1 << (bits[i] % 8);
1067                 *p += size;
1068         } else {
1069                 BUG_ON(*p + 4 > end);
1070                 ceph_encode_32(p, 0);
1071         }
1072 }
1073
1074 /*
1075  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1076  * to include additional client metadata fields.
1077  */
1078 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1079 {
1080         struct ceph_msg *msg;
1081         struct ceph_mds_session_head *h;
1082         int i = -1;
1083         int extra_bytes = 0;
1084         int metadata_key_count = 0;
1085         struct ceph_options *opt = mdsc->fsc->client->options;
1086         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1087         void *p, *end;
1088
1089         const char* metadata[][2] = {
1090                 {"hostname", mdsc->nodename},
1091                 {"kernel_version", init_utsname()->release},
1092                 {"entity_id", opt->name ? : ""},
1093                 {"root", fsopt->server_path ? : "/"},
1094                 {NULL, NULL}
1095         };
1096
1097         /* Calculate serialized length of metadata */
1098         extra_bytes = 4;  /* map length */
1099         for (i = 0; metadata[i][0]; ++i) {
1100                 extra_bytes += 8 + strlen(metadata[i][0]) +
1101                         strlen(metadata[i][1]);
1102                 metadata_key_count++;
1103         }
1104         /* supported feature */
1105         extra_bytes += 4 + 8;
1106
1107         /* Allocate the message */
1108         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1109                            GFP_NOFS, false);
1110         if (!msg) {
1111                 pr_err("create_session_msg ENOMEM creating msg\n");
1112                 return NULL;
1113         }
1114         p = msg->front.iov_base;
1115         end = p + msg->front.iov_len;
1116
1117         h = p;
1118         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1119         h->seq = cpu_to_le64(seq);
1120
1121         /*
1122          * Serialize client metadata into waiting buffer space, using
1123          * the format that userspace expects for map<string, string>
1124          *
1125          * ClientSession messages with metadata are v2
1126          */
1127         msg->hdr.version = cpu_to_le16(3);
1128         msg->hdr.compat_version = cpu_to_le16(1);
1129
1130         /* The write pointer, following the session_head structure */
1131         p += sizeof(*h);
1132
1133         /* Number of entries in the map */
1134         ceph_encode_32(&p, metadata_key_count);
1135
1136         /* Two length-prefixed strings for each entry in the map */
1137         for (i = 0; metadata[i][0]; ++i) {
1138                 size_t const key_len = strlen(metadata[i][0]);
1139                 size_t const val_len = strlen(metadata[i][1]);
1140
1141                 ceph_encode_32(&p, key_len);
1142                 memcpy(p, metadata[i][0], key_len);
1143                 p += key_len;
1144                 ceph_encode_32(&p, val_len);
1145                 memcpy(p, metadata[i][1], val_len);
1146                 p += val_len;
1147         }
1148
1149         encode_supported_features(&p, end);
1150         msg->front.iov_len = p - msg->front.iov_base;
1151         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1152
1153         return msg;
1154 }
1155
1156 /*
1157  * send session open request.
1158  *
1159  * called under mdsc->mutex
1160  */
1161 static int __open_session(struct ceph_mds_client *mdsc,
1162                           struct ceph_mds_session *session)
1163 {
1164         struct ceph_msg *msg;
1165         int mstate;
1166         int mds = session->s_mds;
1167
1168         /* wait for mds to go active? */
1169         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1170         dout("open_session to mds%d (%s)\n", mds,
1171              ceph_mds_state_name(mstate));
1172         session->s_state = CEPH_MDS_SESSION_OPENING;
1173         session->s_renew_requested = jiffies;
1174
1175         /* send connect message */
1176         msg = create_session_open_msg(mdsc, session->s_seq);
1177         if (!msg)
1178                 return -ENOMEM;
1179         ceph_con_send(&session->s_con, msg);
1180         return 0;
1181 }
1182
1183 /*
1184  * open sessions for any export targets for the given mds
1185  *
1186  * called under mdsc->mutex
1187  */
1188 static struct ceph_mds_session *
1189 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1190 {
1191         struct ceph_mds_session *session;
1192
1193         session = __ceph_lookup_mds_session(mdsc, target);
1194         if (!session) {
1195                 session = register_session(mdsc, target);
1196                 if (IS_ERR(session))
1197                         return session;
1198         }
1199         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1200             session->s_state == CEPH_MDS_SESSION_CLOSING)
1201                 __open_session(mdsc, session);
1202
1203         return session;
1204 }
1205
1206 struct ceph_mds_session *
1207 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1208 {
1209         struct ceph_mds_session *session;
1210
1211         dout("open_export_target_session to mds%d\n", target);
1212
1213         mutex_lock(&mdsc->mutex);
1214         session = __open_export_target_session(mdsc, target);
1215         mutex_unlock(&mdsc->mutex);
1216
1217         return session;
1218 }
1219
1220 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1221                                           struct ceph_mds_session *session)
1222 {
1223         struct ceph_mds_info *mi;
1224         struct ceph_mds_session *ts;
1225         int i, mds = session->s_mds;
1226
1227         if (mds >= mdsc->mdsmap->m_num_mds)
1228                 return;
1229
1230         mi = &mdsc->mdsmap->m_info[mds];
1231         dout("open_export_target_sessions for mds%d (%d targets)\n",
1232              session->s_mds, mi->num_export_targets);
1233
1234         for (i = 0; i < mi->num_export_targets; i++) {
1235                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1236                 if (!IS_ERR(ts))
1237                         ceph_put_mds_session(ts);
1238         }
1239 }
1240
1241 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1242                                            struct ceph_mds_session *session)
1243 {
1244         mutex_lock(&mdsc->mutex);
1245         __open_export_target_sessions(mdsc, session);
1246         mutex_unlock(&mdsc->mutex);
1247 }
1248
1249 /*
1250  * session caps
1251  */
1252
1253 static void detach_cap_releases(struct ceph_mds_session *session,
1254                                 struct list_head *target)
1255 {
1256         lockdep_assert_held(&session->s_cap_lock);
1257
1258         list_splice_init(&session->s_cap_releases, target);
1259         session->s_num_cap_releases = 0;
1260         dout("dispose_cap_releases mds%d\n", session->s_mds);
1261 }
1262
1263 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1264                                  struct list_head *dispose)
1265 {
1266         while (!list_empty(dispose)) {
1267                 struct ceph_cap *cap;
1268                 /* zero out the in-progress message */
1269                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1270                 list_del(&cap->session_caps);
1271                 ceph_put_cap(mdsc, cap);
1272         }
1273 }
1274
1275 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1276                                      struct ceph_mds_session *session)
1277 {
1278         struct ceph_mds_request *req;
1279         struct rb_node *p;
1280         struct ceph_inode_info *ci;
1281
1282         dout("cleanup_session_requests mds%d\n", session->s_mds);
1283         mutex_lock(&mdsc->mutex);
1284         while (!list_empty(&session->s_unsafe)) {
1285                 req = list_first_entry(&session->s_unsafe,
1286                                        struct ceph_mds_request, r_unsafe_item);
1287                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1288                                     req->r_tid);
1289                 if (req->r_target_inode) {
1290                         /* dropping unsafe change of inode's attributes */
1291                         ci = ceph_inode(req->r_target_inode);
1292                         errseq_set(&ci->i_meta_err, -EIO);
1293                 }
1294                 if (req->r_unsafe_dir) {
1295                         /* dropping unsafe directory operation */
1296                         ci = ceph_inode(req->r_unsafe_dir);
1297                         errseq_set(&ci->i_meta_err, -EIO);
1298                 }
1299                 __unregister_request(mdsc, req);
1300         }
1301         /* zero r_attempts, so kick_requests() will re-send requests */
1302         p = rb_first(&mdsc->request_tree);
1303         while (p) {
1304                 req = rb_entry(p, struct ceph_mds_request, r_node);
1305                 p = rb_next(p);
1306                 if (req->r_session &&
1307                     req->r_session->s_mds == session->s_mds)
1308                         req->r_attempts = 0;
1309         }
1310         mutex_unlock(&mdsc->mutex);
1311 }
1312
1313 /*
1314  * Helper to safely iterate over all caps associated with a session, with
1315  * special care taken to handle a racing __ceph_remove_cap().
1316  *
1317  * Caller must hold session s_mutex.
1318  */
1319 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1320                               int (*cb)(struct inode *, struct ceph_cap *,
1321                                         void *), void *arg)
1322 {
1323         struct list_head *p;
1324         struct ceph_cap *cap;
1325         struct inode *inode, *last_inode = NULL;
1326         struct ceph_cap *old_cap = NULL;
1327         int ret;
1328
1329         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1330         spin_lock(&session->s_cap_lock);
1331         p = session->s_caps.next;
1332         while (p != &session->s_caps) {
1333                 cap = list_entry(p, struct ceph_cap, session_caps);
1334                 inode = igrab(&cap->ci->vfs_inode);
1335                 if (!inode) {
1336                         p = p->next;
1337                         continue;
1338                 }
1339                 session->s_cap_iterator = cap;
1340                 spin_unlock(&session->s_cap_lock);
1341
1342                 if (last_inode) {
1343                         /* avoid calling iput_final() while holding
1344                          * s_mutex or in mds dispatch threads */
1345                         ceph_async_iput(last_inode);
1346                         last_inode = NULL;
1347                 }
1348                 if (old_cap) {
1349                         ceph_put_cap(session->s_mdsc, old_cap);
1350                         old_cap = NULL;
1351                 }
1352
1353                 ret = cb(inode, cap, arg);
1354                 last_inode = inode;
1355
1356                 spin_lock(&session->s_cap_lock);
1357                 p = p->next;
1358                 if (!cap->ci) {
1359                         dout("iterate_session_caps  finishing cap %p removal\n",
1360                              cap);
1361                         BUG_ON(cap->session != session);
1362                         cap->session = NULL;
1363                         list_del_init(&cap->session_caps);
1364                         session->s_nr_caps--;
1365                         if (cap->queue_release)
1366                                 __ceph_queue_cap_release(session, cap);
1367                         else
1368                                 old_cap = cap;  /* put_cap it w/o locks held */
1369                 }
1370                 if (ret < 0)
1371                         goto out;
1372         }
1373         ret = 0;
1374 out:
1375         session->s_cap_iterator = NULL;
1376         spin_unlock(&session->s_cap_lock);
1377
1378         ceph_async_iput(last_inode);
1379         if (old_cap)
1380                 ceph_put_cap(session->s_mdsc, old_cap);
1381
1382         return ret;
1383 }
1384
1385 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1386                                   void *arg)
1387 {
1388         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1389         struct ceph_inode_info *ci = ceph_inode(inode);
1390         LIST_HEAD(to_remove);
1391         bool dirty_dropped = false;
1392         bool invalidate = false;
1393
1394         dout("removing cap %p, ci is %p, inode is %p\n",
1395              cap, ci, &ci->vfs_inode);
1396         spin_lock(&ci->i_ceph_lock);
1397         if (cap->mds_wanted | cap->issued)
1398                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1399         __ceph_remove_cap(cap, false);
1400         if (!ci->i_auth_cap) {
1401                 struct ceph_cap_flush *cf;
1402                 struct ceph_mds_client *mdsc = fsc->mdsc;
1403
1404                 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1405                         if (inode->i_data.nrpages > 0)
1406                                 invalidate = true;
1407                         if (ci->i_wrbuffer_ref > 0)
1408                                 mapping_set_error(&inode->i_data, -EIO);
1409                 }
1410
1411                 while (!list_empty(&ci->i_cap_flush_list)) {
1412                         cf = list_first_entry(&ci->i_cap_flush_list,
1413                                               struct ceph_cap_flush, i_list);
1414                         list_move(&cf->i_list, &to_remove);
1415                 }
1416
1417                 spin_lock(&mdsc->cap_dirty_lock);
1418
1419                 list_for_each_entry(cf, &to_remove, i_list)
1420                         list_del(&cf->g_list);
1421
1422                 if (!list_empty(&ci->i_dirty_item)) {
1423                         pr_warn_ratelimited(
1424                                 " dropping dirty %s state for %p %lld\n",
1425                                 ceph_cap_string(ci->i_dirty_caps),
1426                                 inode, ceph_ino(inode));
1427                         ci->i_dirty_caps = 0;
1428                         list_del_init(&ci->i_dirty_item);
1429                         dirty_dropped = true;
1430                 }
1431                 if (!list_empty(&ci->i_flushing_item)) {
1432                         pr_warn_ratelimited(
1433                                 " dropping dirty+flushing %s state for %p %lld\n",
1434                                 ceph_cap_string(ci->i_flushing_caps),
1435                                 inode, ceph_ino(inode));
1436                         ci->i_flushing_caps = 0;
1437                         list_del_init(&ci->i_flushing_item);
1438                         mdsc->num_cap_flushing--;
1439                         dirty_dropped = true;
1440                 }
1441                 spin_unlock(&mdsc->cap_dirty_lock);
1442
1443                 if (dirty_dropped) {
1444                         errseq_set(&ci->i_meta_err, -EIO);
1445
1446                         if (ci->i_wrbuffer_ref_head == 0 &&
1447                             ci->i_wr_ref == 0 &&
1448                             ci->i_dirty_caps == 0 &&
1449                             ci->i_flushing_caps == 0) {
1450                                 ceph_put_snap_context(ci->i_head_snapc);
1451                                 ci->i_head_snapc = NULL;
1452                         }
1453                 }
1454
1455                 if (atomic_read(&ci->i_filelock_ref) > 0) {
1456                         /* make further file lock syscall return -EIO */
1457                         ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1458                         pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1459                                             inode, ceph_ino(inode));
1460                 }
1461
1462                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1463                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1464                         ci->i_prealloc_cap_flush = NULL;
1465                 }
1466         }
1467         spin_unlock(&ci->i_ceph_lock);
1468         while (!list_empty(&to_remove)) {
1469                 struct ceph_cap_flush *cf;
1470                 cf = list_first_entry(&to_remove,
1471                                       struct ceph_cap_flush, i_list);
1472                 list_del(&cf->i_list);
1473                 ceph_free_cap_flush(cf);
1474         }
1475
1476         wake_up_all(&ci->i_cap_wq);
1477         if (invalidate)
1478                 ceph_queue_invalidate(inode);
1479         if (dirty_dropped)
1480                 iput(inode);
1481         return 0;
1482 }
1483
1484 /*
1485  * caller must hold session s_mutex
1486  */
1487 static void remove_session_caps(struct ceph_mds_session *session)
1488 {
1489         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1490         struct super_block *sb = fsc->sb;
1491         LIST_HEAD(dispose);
1492
1493         dout("remove_session_caps on %p\n", session);
1494         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1495
1496         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1497
1498         spin_lock(&session->s_cap_lock);
1499         if (session->s_nr_caps > 0) {
1500                 struct inode *inode;
1501                 struct ceph_cap *cap, *prev = NULL;
1502                 struct ceph_vino vino;
1503                 /*
1504                  * iterate_session_caps() skips inodes that are being
1505                  * deleted, we need to wait until deletions are complete.
1506                  * __wait_on_freeing_inode() is designed for the job,
1507                  * but it is not exported, so use lookup inode function
1508                  * to access it.
1509                  */
1510                 while (!list_empty(&session->s_caps)) {
1511                         cap = list_entry(session->s_caps.next,
1512                                          struct ceph_cap, session_caps);
1513                         if (cap == prev)
1514                                 break;
1515                         prev = cap;
1516                         vino = cap->ci->i_vino;
1517                         spin_unlock(&session->s_cap_lock);
1518
1519                         inode = ceph_find_inode(sb, vino);
1520                          /* avoid calling iput_final() while holding s_mutex */
1521                         ceph_async_iput(inode);
1522
1523                         spin_lock(&session->s_cap_lock);
1524                 }
1525         }
1526
1527         // drop cap expires and unlock s_cap_lock
1528         detach_cap_releases(session, &dispose);
1529
1530         BUG_ON(session->s_nr_caps > 0);
1531         BUG_ON(!list_empty(&session->s_cap_flushing));
1532         spin_unlock(&session->s_cap_lock);
1533         dispose_cap_releases(session->s_mdsc, &dispose);
1534 }
1535
1536 enum {
1537         RECONNECT,
1538         RENEWCAPS,
1539         FORCE_RO,
1540 };
1541
1542 /*
1543  * wake up any threads waiting on this session's caps.  if the cap is
1544  * old (didn't get renewed on the client reconnect), remove it now.
1545  *
1546  * caller must hold s_mutex.
1547  */
1548 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1549                               void *arg)
1550 {
1551         struct ceph_inode_info *ci = ceph_inode(inode);
1552         unsigned long ev = (unsigned long)arg;
1553
1554         if (ev == RECONNECT) {
1555                 spin_lock(&ci->i_ceph_lock);
1556                 ci->i_wanted_max_size = 0;
1557                 ci->i_requested_max_size = 0;
1558                 spin_unlock(&ci->i_ceph_lock);
1559         } else if (ev == RENEWCAPS) {
1560                 if (cap->cap_gen < cap->session->s_cap_gen) {
1561                         /* mds did not re-issue stale cap */
1562                         spin_lock(&ci->i_ceph_lock);
1563                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1564                         /* make sure mds knows what we want */
1565                         if (__ceph_caps_file_wanted(ci) & ~cap->mds_wanted)
1566                                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1567                         spin_unlock(&ci->i_ceph_lock);
1568                 }
1569         } else if (ev == FORCE_RO) {
1570         }
1571         wake_up_all(&ci->i_cap_wq);
1572         return 0;
1573 }
1574
1575 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1576 {
1577         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1578         ceph_iterate_session_caps(session, wake_up_session_cb,
1579                                   (void *)(unsigned long)ev);
1580 }
1581
1582 /*
1583  * Send periodic message to MDS renewing all currently held caps.  The
1584  * ack will reset the expiration for all caps from this session.
1585  *
1586  * caller holds s_mutex
1587  */
1588 static int send_renew_caps(struct ceph_mds_client *mdsc,
1589                            struct ceph_mds_session *session)
1590 {
1591         struct ceph_msg *msg;
1592         int state;
1593
1594         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1595             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1596                 pr_info("mds%d caps stale\n", session->s_mds);
1597         session->s_renew_requested = jiffies;
1598
1599         /* do not try to renew caps until a recovering mds has reconnected
1600          * with its clients. */
1601         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1602         if (state < CEPH_MDS_STATE_RECONNECT) {
1603                 dout("send_renew_caps ignoring mds%d (%s)\n",
1604                      session->s_mds, ceph_mds_state_name(state));
1605                 return 0;
1606         }
1607
1608         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1609                 ceph_mds_state_name(state));
1610         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1611                                  ++session->s_renew_seq);
1612         if (!msg)
1613                 return -ENOMEM;
1614         ceph_con_send(&session->s_con, msg);
1615         return 0;
1616 }
1617
1618 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1619                              struct ceph_mds_session *session, u64 seq)
1620 {
1621         struct ceph_msg *msg;
1622
1623         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1624              session->s_mds, ceph_session_state_name(session->s_state), seq);
1625         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1626         if (!msg)
1627                 return -ENOMEM;
1628         ceph_con_send(&session->s_con, msg);
1629         return 0;
1630 }
1631
1632
1633 /*
1634  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1635  *
1636  * Called under session->s_mutex
1637  */
1638 static void renewed_caps(struct ceph_mds_client *mdsc,
1639                          struct ceph_mds_session *session, int is_renew)
1640 {
1641         int was_stale;
1642         int wake = 0;
1643
1644         spin_lock(&session->s_cap_lock);
1645         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1646
1647         session->s_cap_ttl = session->s_renew_requested +
1648                 mdsc->mdsmap->m_session_timeout*HZ;
1649
1650         if (was_stale) {
1651                 if (time_before(jiffies, session->s_cap_ttl)) {
1652                         pr_info("mds%d caps renewed\n", session->s_mds);
1653                         wake = 1;
1654                 } else {
1655                         pr_info("mds%d caps still stale\n", session->s_mds);
1656                 }
1657         }
1658         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1659              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1660              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1661         spin_unlock(&session->s_cap_lock);
1662
1663         if (wake)
1664                 wake_up_session_caps(session, RENEWCAPS);
1665 }
1666
1667 /*
1668  * send a session close request
1669  */
1670 static int request_close_session(struct ceph_mds_client *mdsc,
1671                                  struct ceph_mds_session *session)
1672 {
1673         struct ceph_msg *msg;
1674
1675         dout("request_close_session mds%d state %s seq %lld\n",
1676              session->s_mds, ceph_session_state_name(session->s_state),
1677              session->s_seq);
1678         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1679         if (!msg)
1680                 return -ENOMEM;
1681         ceph_con_send(&session->s_con, msg);
1682         return 1;
1683 }
1684
1685 /*
1686  * Called with s_mutex held.
1687  */
1688 static int __close_session(struct ceph_mds_client *mdsc,
1689                          struct ceph_mds_session *session)
1690 {
1691         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1692                 return 0;
1693         session->s_state = CEPH_MDS_SESSION_CLOSING;
1694         return request_close_session(mdsc, session);
1695 }
1696
1697 static bool drop_negative_children(struct dentry *dentry)
1698 {
1699         struct dentry *child;
1700         bool all_negative = true;
1701
1702         if (!d_is_dir(dentry))
1703                 goto out;
1704
1705         spin_lock(&dentry->d_lock);
1706         list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1707                 if (d_really_is_positive(child)) {
1708                         all_negative = false;
1709                         break;
1710                 }
1711         }
1712         spin_unlock(&dentry->d_lock);
1713
1714         if (all_negative)
1715                 shrink_dcache_parent(dentry);
1716 out:
1717         return all_negative;
1718 }
1719
1720 /*
1721  * Trim old(er) caps.
1722  *
1723  * Because we can't cache an inode without one or more caps, we do
1724  * this indirectly: if a cap is unused, we prune its aliases, at which
1725  * point the inode will hopefully get dropped to.
1726  *
1727  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1728  * memory pressure from the MDS, though, so it needn't be perfect.
1729  */
1730 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1731 {
1732         int *remaining = arg;
1733         struct ceph_inode_info *ci = ceph_inode(inode);
1734         int used, wanted, oissued, mine;
1735
1736         if (*remaining <= 0)
1737                 return -1;
1738
1739         spin_lock(&ci->i_ceph_lock);
1740         mine = cap->issued | cap->implemented;
1741         used = __ceph_caps_used(ci);
1742         wanted = __ceph_caps_file_wanted(ci);
1743         oissued = __ceph_caps_issued_other(ci, cap);
1744
1745         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1746              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1747              ceph_cap_string(used), ceph_cap_string(wanted));
1748         if (cap == ci->i_auth_cap) {
1749                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1750                     !list_empty(&ci->i_cap_snaps))
1751                         goto out;
1752                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1753                         goto out;
1754                 /* Note: it's possible that i_filelock_ref becomes non-zero
1755                  * after dropping auth caps. It doesn't hurt because reply
1756                  * of lock mds request will re-add auth caps. */
1757                 if (atomic_read(&ci->i_filelock_ref) > 0)
1758                         goto out;
1759         }
1760         /* The inode has cached pages, but it's no longer used.
1761          * we can safely drop it */
1762         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1763             !(oissued & CEPH_CAP_FILE_CACHE)) {
1764           used = 0;
1765           oissued = 0;
1766         }
1767         if ((used | wanted) & ~oissued & mine)
1768                 goto out;   /* we need these caps */
1769
1770         if (oissued) {
1771                 /* we aren't the only cap.. just remove us */
1772                 __ceph_remove_cap(cap, true);
1773                 (*remaining)--;
1774         } else {
1775                 struct dentry *dentry;
1776                 /* try dropping referring dentries */
1777                 spin_unlock(&ci->i_ceph_lock);
1778                 dentry = d_find_any_alias(inode);
1779                 if (dentry && drop_negative_children(dentry)) {
1780                         int count;
1781                         dput(dentry);
1782                         d_prune_aliases(inode);
1783                         count = atomic_read(&inode->i_count);
1784                         if (count == 1)
1785                                 (*remaining)--;
1786                         dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1787                              inode, cap, count);
1788                 } else {
1789                         dput(dentry);
1790                 }
1791                 return 0;
1792         }
1793
1794 out:
1795         spin_unlock(&ci->i_ceph_lock);
1796         return 0;
1797 }
1798
1799 /*
1800  * Trim session cap count down to some max number.
1801  */
1802 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1803                    struct ceph_mds_session *session,
1804                    int max_caps)
1805 {
1806         int trim_caps = session->s_nr_caps - max_caps;
1807
1808         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1809              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1810         if (trim_caps > 0) {
1811                 int remaining = trim_caps;
1812
1813                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1814                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1815                      session->s_mds, session->s_nr_caps, max_caps,
1816                         trim_caps - remaining);
1817         }
1818
1819         ceph_flush_cap_releases(mdsc, session);
1820         return 0;
1821 }
1822
1823 static int check_caps_flush(struct ceph_mds_client *mdsc,
1824                             u64 want_flush_tid)
1825 {
1826         int ret = 1;
1827
1828         spin_lock(&mdsc->cap_dirty_lock);
1829         if (!list_empty(&mdsc->cap_flush_list)) {
1830                 struct ceph_cap_flush *cf =
1831                         list_first_entry(&mdsc->cap_flush_list,
1832                                          struct ceph_cap_flush, g_list);
1833                 if (cf->tid <= want_flush_tid) {
1834                         dout("check_caps_flush still flushing tid "
1835                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1836                         ret = 0;
1837                 }
1838         }
1839         spin_unlock(&mdsc->cap_dirty_lock);
1840         return ret;
1841 }
1842
1843 /*
1844  * flush all dirty inode data to disk.
1845  *
1846  * returns true if we've flushed through want_flush_tid
1847  */
1848 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1849                             u64 want_flush_tid)
1850 {
1851         dout("check_caps_flush want %llu\n", want_flush_tid);
1852
1853         wait_event(mdsc->cap_flushing_wq,
1854                    check_caps_flush(mdsc, want_flush_tid));
1855
1856         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1857 }
1858
1859 /*
1860  * called under s_mutex
1861  */
1862 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1863                                    struct ceph_mds_session *session)
1864 {
1865         struct ceph_msg *msg = NULL;
1866         struct ceph_mds_cap_release *head;
1867         struct ceph_mds_cap_item *item;
1868         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1869         struct ceph_cap *cap;
1870         LIST_HEAD(tmp_list);
1871         int num_cap_releases;
1872         __le32  barrier, *cap_barrier;
1873
1874         down_read(&osdc->lock);
1875         barrier = cpu_to_le32(osdc->epoch_barrier);
1876         up_read(&osdc->lock);
1877
1878         spin_lock(&session->s_cap_lock);
1879 again:
1880         list_splice_init(&session->s_cap_releases, &tmp_list);
1881         num_cap_releases = session->s_num_cap_releases;
1882         session->s_num_cap_releases = 0;
1883         spin_unlock(&session->s_cap_lock);
1884
1885         while (!list_empty(&tmp_list)) {
1886                 if (!msg) {
1887                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1888                                         PAGE_SIZE, GFP_NOFS, false);
1889                         if (!msg)
1890                                 goto out_err;
1891                         head = msg->front.iov_base;
1892                         head->num = cpu_to_le32(0);
1893                         msg->front.iov_len = sizeof(*head);
1894
1895                         msg->hdr.version = cpu_to_le16(2);
1896                         msg->hdr.compat_version = cpu_to_le16(1);
1897                 }
1898
1899                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1900                                         session_caps);
1901                 list_del(&cap->session_caps);
1902                 num_cap_releases--;
1903
1904                 head = msg->front.iov_base;
1905                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
1906                                    &head->num);
1907                 item = msg->front.iov_base + msg->front.iov_len;
1908                 item->ino = cpu_to_le64(cap->cap_ino);
1909                 item->cap_id = cpu_to_le64(cap->cap_id);
1910                 item->migrate_seq = cpu_to_le32(cap->mseq);
1911                 item->seq = cpu_to_le32(cap->issue_seq);
1912                 msg->front.iov_len += sizeof(*item);
1913
1914                 ceph_put_cap(mdsc, cap);
1915
1916                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1917                         // Append cap_barrier field
1918                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1919                         *cap_barrier = barrier;
1920                         msg->front.iov_len += sizeof(*cap_barrier);
1921
1922                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1923                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1924                         ceph_con_send(&session->s_con, msg);
1925                         msg = NULL;
1926                 }
1927         }
1928
1929         BUG_ON(num_cap_releases != 0);
1930
1931         spin_lock(&session->s_cap_lock);
1932         if (!list_empty(&session->s_cap_releases))
1933                 goto again;
1934         spin_unlock(&session->s_cap_lock);
1935
1936         if (msg) {
1937                 // Append cap_barrier field
1938                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1939                 *cap_barrier = barrier;
1940                 msg->front.iov_len += sizeof(*cap_barrier);
1941
1942                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1943                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1944                 ceph_con_send(&session->s_con, msg);
1945         }
1946         return;
1947 out_err:
1948         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1949                 session->s_mds);
1950         spin_lock(&session->s_cap_lock);
1951         list_splice(&tmp_list, &session->s_cap_releases);
1952         session->s_num_cap_releases += num_cap_releases;
1953         spin_unlock(&session->s_cap_lock);
1954 }
1955
1956 static void ceph_cap_release_work(struct work_struct *work)
1957 {
1958         struct ceph_mds_session *session =
1959                 container_of(work, struct ceph_mds_session, s_cap_release_work);
1960
1961         mutex_lock(&session->s_mutex);
1962         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
1963             session->s_state == CEPH_MDS_SESSION_HUNG)
1964                 ceph_send_cap_releases(session->s_mdsc, session);
1965         mutex_unlock(&session->s_mutex);
1966         ceph_put_mds_session(session);
1967 }
1968
1969 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
1970                              struct ceph_mds_session *session)
1971 {
1972         if (mdsc->stopping)
1973                 return;
1974
1975         get_session(session);
1976         if (queue_work(mdsc->fsc->cap_wq,
1977                        &session->s_cap_release_work)) {
1978                 dout("cap release work queued\n");
1979         } else {
1980                 ceph_put_mds_session(session);
1981                 dout("failed to queue cap release work\n");
1982         }
1983 }
1984
1985 /*
1986  * caller holds session->s_cap_lock
1987  */
1988 void __ceph_queue_cap_release(struct ceph_mds_session *session,
1989                               struct ceph_cap *cap)
1990 {
1991         list_add_tail(&cap->session_caps, &session->s_cap_releases);
1992         session->s_num_cap_releases++;
1993
1994         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
1995                 ceph_flush_cap_releases(session->s_mdsc, session);
1996 }
1997
1998 static void ceph_cap_reclaim_work(struct work_struct *work)
1999 {
2000         struct ceph_mds_client *mdsc =
2001                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2002         int ret = ceph_trim_dentries(mdsc);
2003         if (ret == -EAGAIN)
2004                 ceph_queue_cap_reclaim_work(mdsc);
2005 }
2006
2007 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2008 {
2009         if (mdsc->stopping)
2010                 return;
2011
2012         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2013                 dout("caps reclaim work queued\n");
2014         } else {
2015                 dout("failed to queue caps release work\n");
2016         }
2017 }
2018
2019 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2020 {
2021         int val;
2022         if (!nr)
2023                 return;
2024         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2025         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2026                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2027                 ceph_queue_cap_reclaim_work(mdsc);
2028         }
2029 }
2030
2031 /*
2032  * requests
2033  */
2034
2035 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2036                                     struct inode *dir)
2037 {
2038         struct ceph_inode_info *ci = ceph_inode(dir);
2039         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2040         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2041         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2042         unsigned int num_entries;
2043         int order;
2044
2045         spin_lock(&ci->i_ceph_lock);
2046         num_entries = ci->i_files + ci->i_subdirs;
2047         spin_unlock(&ci->i_ceph_lock);
2048         num_entries = max(num_entries, 1U);
2049         num_entries = min(num_entries, opt->max_readdir);
2050
2051         order = get_order(size * num_entries);
2052         while (order >= 0) {
2053                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2054                                                              __GFP_NOWARN,
2055                                                              order);
2056                 if (rinfo->dir_entries)
2057                         break;
2058                 order--;
2059         }
2060         if (!rinfo->dir_entries)
2061                 return -ENOMEM;
2062
2063         num_entries = (PAGE_SIZE << order) / size;
2064         num_entries = min(num_entries, opt->max_readdir);
2065
2066         rinfo->dir_buf_size = PAGE_SIZE << order;
2067         req->r_num_caps = num_entries + 1;
2068         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2069         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2070         return 0;
2071 }
2072
2073 /*
2074  * Create an mds request.
2075  */
2076 struct ceph_mds_request *
2077 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2078 {
2079         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
2080         struct timespec64 ts;
2081
2082         if (!req)
2083                 return ERR_PTR(-ENOMEM);
2084
2085         mutex_init(&req->r_fill_mutex);
2086         req->r_mdsc = mdsc;
2087         req->r_started = jiffies;
2088         req->r_resend_mds = -1;
2089         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2090         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2091         req->r_fmode = -1;
2092         kref_init(&req->r_kref);
2093         RB_CLEAR_NODE(&req->r_node);
2094         INIT_LIST_HEAD(&req->r_wait);
2095         init_completion(&req->r_completion);
2096         init_completion(&req->r_safe_completion);
2097         INIT_LIST_HEAD(&req->r_unsafe_item);
2098
2099         ktime_get_coarse_real_ts64(&ts);
2100         req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran);
2101
2102         req->r_op = op;
2103         req->r_direct_mode = mode;
2104         return req;
2105 }
2106
2107 /*
2108  * return oldest (lowest) request, tid in request tree, 0 if none.
2109  *
2110  * called under mdsc->mutex.
2111  */
2112 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2113 {
2114         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2115                 return NULL;
2116         return rb_entry(rb_first(&mdsc->request_tree),
2117                         struct ceph_mds_request, r_node);
2118 }
2119
2120 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2121 {
2122         return mdsc->oldest_tid;
2123 }
2124
2125 /*
2126  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
2127  * on build_path_from_dentry in fs/cifs/dir.c.
2128  *
2129  * If @stop_on_nosnap, generate path relative to the first non-snapped
2130  * inode.
2131  *
2132  * Encode hidden .snap dirs as a double /, i.e.
2133  *   foo/.snap/bar -> foo//bar
2134  */
2135 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2136                            int stop_on_nosnap)
2137 {
2138         struct dentry *temp;
2139         char *path;
2140         int pos;
2141         unsigned seq;
2142         u64 base;
2143
2144         if (!dentry)
2145                 return ERR_PTR(-EINVAL);
2146
2147         path = __getname();
2148         if (!path)
2149                 return ERR_PTR(-ENOMEM);
2150 retry:
2151         pos = PATH_MAX - 1;
2152         path[pos] = '\0';
2153
2154         seq = read_seqbegin(&rename_lock);
2155         rcu_read_lock();
2156         temp = dentry;
2157         for (;;) {
2158                 struct inode *inode;
2159
2160                 spin_lock(&temp->d_lock);
2161                 inode = d_inode(temp);
2162                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2163                         dout("build_path path+%d: %p SNAPDIR\n",
2164                              pos, temp);
2165                 } else if (stop_on_nosnap && inode && dentry != temp &&
2166                            ceph_snap(inode) == CEPH_NOSNAP) {
2167                         spin_unlock(&temp->d_lock);
2168                         pos++; /* get rid of any prepended '/' */
2169                         break;
2170                 } else {
2171                         pos -= temp->d_name.len;
2172                         if (pos < 0) {
2173                                 spin_unlock(&temp->d_lock);
2174                                 break;
2175                         }
2176                         memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2177                 }
2178                 spin_unlock(&temp->d_lock);
2179                 temp = READ_ONCE(temp->d_parent);
2180
2181                 /* Are we at the root? */
2182                 if (IS_ROOT(temp))
2183                         break;
2184
2185                 /* Are we out of buffer? */
2186                 if (--pos < 0)
2187                         break;
2188
2189                 path[pos] = '/';
2190         }
2191         base = ceph_ino(d_inode(temp));
2192         rcu_read_unlock();
2193
2194         if (read_seqretry(&rename_lock, seq))
2195                 goto retry;
2196
2197         if (pos < 0) {
2198                 /*
2199                  * A rename didn't occur, but somehow we didn't end up where
2200                  * we thought we would. Throw a warning and try again.
2201                  */
2202                 pr_warn("build_path did not end path lookup where "
2203                         "expected, pos is %d\n", pos);
2204                 goto retry;
2205         }
2206
2207         *pbase = base;
2208         *plen = PATH_MAX - 1 - pos;
2209         dout("build_path on %p %d built %llx '%.*s'\n",
2210              dentry, d_count(dentry), base, *plen, path + pos);
2211         return path + pos;
2212 }
2213
2214 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2215                              const char **ppath, int *ppathlen, u64 *pino,
2216                              bool *pfreepath, bool parent_locked)
2217 {
2218         char *path;
2219
2220         rcu_read_lock();
2221         if (!dir)
2222                 dir = d_inode_rcu(dentry->d_parent);
2223         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2224                 *pino = ceph_ino(dir);
2225                 rcu_read_unlock();
2226                 *ppath = dentry->d_name.name;
2227                 *ppathlen = dentry->d_name.len;
2228                 return 0;
2229         }
2230         rcu_read_unlock();
2231         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2232         if (IS_ERR(path))
2233                 return PTR_ERR(path);
2234         *ppath = path;
2235         *pfreepath = true;
2236         return 0;
2237 }
2238
2239 static int build_inode_path(struct inode *inode,
2240                             const char **ppath, int *ppathlen, u64 *pino,
2241                             bool *pfreepath)
2242 {
2243         struct dentry *dentry;
2244         char *path;
2245
2246         if (ceph_snap(inode) == CEPH_NOSNAP) {
2247                 *pino = ceph_ino(inode);
2248                 *ppathlen = 0;
2249                 return 0;
2250         }
2251         dentry = d_find_alias(inode);
2252         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2253         dput(dentry);
2254         if (IS_ERR(path))
2255                 return PTR_ERR(path);
2256         *ppath = path;
2257         *pfreepath = true;
2258         return 0;
2259 }
2260
2261 /*
2262  * request arguments may be specified via an inode *, a dentry *, or
2263  * an explicit ino+path.
2264  */
2265 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2266                                   struct inode *rdiri, const char *rpath,
2267                                   u64 rino, const char **ppath, int *pathlen,
2268                                   u64 *ino, bool *freepath, bool parent_locked)
2269 {
2270         int r = 0;
2271
2272         if (rinode) {
2273                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2274                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2275                      ceph_snap(rinode));
2276         } else if (rdentry) {
2277                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2278                                         freepath, parent_locked);
2279                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2280                      *ppath);
2281         } else if (rpath || rino) {
2282                 *ino = rino;
2283                 *ppath = rpath;
2284                 *pathlen = rpath ? strlen(rpath) : 0;
2285                 dout(" path %.*s\n", *pathlen, rpath);
2286         }
2287
2288         return r;
2289 }
2290
2291 /*
2292  * called under mdsc->mutex
2293  */
2294 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2295                                                struct ceph_mds_request *req,
2296                                                int mds, bool drop_cap_releases)
2297 {
2298         struct ceph_msg *msg;
2299         struct ceph_mds_request_head *head;
2300         const char *path1 = NULL;
2301         const char *path2 = NULL;
2302         u64 ino1 = 0, ino2 = 0;
2303         int pathlen1 = 0, pathlen2 = 0;
2304         bool freepath1 = false, freepath2 = false;
2305         int len;
2306         u16 releases;
2307         void *p, *end;
2308         int ret;
2309
2310         ret = set_request_path_attr(req->r_inode, req->r_dentry,
2311                               req->r_parent, req->r_path1, req->r_ino1.ino,
2312                               &path1, &pathlen1, &ino1, &freepath1,
2313                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2314                                         &req->r_req_flags));
2315         if (ret < 0) {
2316                 msg = ERR_PTR(ret);
2317                 goto out;
2318         }
2319
2320         /* If r_old_dentry is set, then assume that its parent is locked */
2321         ret = set_request_path_attr(NULL, req->r_old_dentry,
2322                               req->r_old_dentry_dir,
2323                               req->r_path2, req->r_ino2.ino,
2324                               &path2, &pathlen2, &ino2, &freepath2, true);
2325         if (ret < 0) {
2326                 msg = ERR_PTR(ret);
2327                 goto out_free1;
2328         }
2329
2330         len = sizeof(*head) +
2331                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2332                 sizeof(struct ceph_timespec);
2333
2334         /* calculate (max) length for cap releases */
2335         len += sizeof(struct ceph_mds_request_release) *
2336                 (!!req->r_inode_drop + !!req->r_dentry_drop +
2337                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2338         if (req->r_dentry_drop)
2339                 len += pathlen1;
2340         if (req->r_old_dentry_drop)
2341                 len += pathlen2;
2342
2343         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2344         if (!msg) {
2345                 msg = ERR_PTR(-ENOMEM);
2346                 goto out_free2;
2347         }
2348
2349         msg->hdr.version = cpu_to_le16(2);
2350         msg->hdr.tid = cpu_to_le64(req->r_tid);
2351
2352         head = msg->front.iov_base;
2353         p = msg->front.iov_base + sizeof(*head);
2354         end = msg->front.iov_base + msg->front.iov_len;
2355
2356         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2357         head->op = cpu_to_le32(req->r_op);
2358         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2359         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2360         head->ino = 0;
2361         head->args = req->r_args;
2362
2363         ceph_encode_filepath(&p, end, ino1, path1);
2364         ceph_encode_filepath(&p, end, ino2, path2);
2365
2366         /* make note of release offset, in case we need to replay */
2367         req->r_request_release_offset = p - msg->front.iov_base;
2368
2369         /* cap releases */
2370         releases = 0;
2371         if (req->r_inode_drop)
2372                 releases += ceph_encode_inode_release(&p,
2373                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2374                       mds, req->r_inode_drop, req->r_inode_unless, 0);
2375         if (req->r_dentry_drop)
2376                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2377                                 req->r_parent, mds, req->r_dentry_drop,
2378                                 req->r_dentry_unless);
2379         if (req->r_old_dentry_drop)
2380                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2381                                 req->r_old_dentry_dir, mds,
2382                                 req->r_old_dentry_drop,
2383                                 req->r_old_dentry_unless);
2384         if (req->r_old_inode_drop)
2385                 releases += ceph_encode_inode_release(&p,
2386                       d_inode(req->r_old_dentry),
2387                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2388
2389         if (drop_cap_releases) {
2390                 releases = 0;
2391                 p = msg->front.iov_base + req->r_request_release_offset;
2392         }
2393
2394         head->num_releases = cpu_to_le16(releases);
2395
2396         /* time stamp */
2397         {
2398                 struct ceph_timespec ts;
2399                 ceph_encode_timespec64(&ts, &req->r_stamp);
2400                 ceph_encode_copy(&p, &ts, sizeof(ts));
2401         }
2402
2403         BUG_ON(p > end);
2404         msg->front.iov_len = p - msg->front.iov_base;
2405         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2406
2407         if (req->r_pagelist) {
2408                 struct ceph_pagelist *pagelist = req->r_pagelist;
2409                 ceph_msg_data_add_pagelist(msg, pagelist);
2410                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2411         } else {
2412                 msg->hdr.data_len = 0;
2413         }
2414
2415         msg->hdr.data_off = cpu_to_le16(0);
2416
2417 out_free2:
2418         if (freepath2)
2419                 ceph_mdsc_free_path((char *)path2, pathlen2);
2420 out_free1:
2421         if (freepath1)
2422                 ceph_mdsc_free_path((char *)path1, pathlen1);
2423 out:
2424         return msg;
2425 }
2426
2427 /*
2428  * called under mdsc->mutex if error, under no mutex if
2429  * success.
2430  */
2431 static void complete_request(struct ceph_mds_client *mdsc,
2432                              struct ceph_mds_request *req)
2433 {
2434         if (req->r_callback)
2435                 req->r_callback(mdsc, req);
2436         complete_all(&req->r_completion);
2437 }
2438
2439 /*
2440  * called under mdsc->mutex
2441  */
2442 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2443                                   struct ceph_mds_request *req,
2444                                   int mds, bool drop_cap_releases)
2445 {
2446         struct ceph_mds_request_head *rhead;
2447         struct ceph_msg *msg;
2448         int flags = 0;
2449
2450         req->r_attempts++;
2451         if (req->r_inode) {
2452                 struct ceph_cap *cap =
2453                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2454
2455                 if (cap)
2456                         req->r_sent_on_mseq = cap->mseq;
2457                 else
2458                         req->r_sent_on_mseq = -1;
2459         }
2460         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2461              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2462
2463         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2464                 void *p;
2465                 /*
2466                  * Replay.  Do not regenerate message (and rebuild
2467                  * paths, etc.); just use the original message.
2468                  * Rebuilding paths will break for renames because
2469                  * d_move mangles the src name.
2470                  */
2471                 msg = req->r_request;
2472                 rhead = msg->front.iov_base;
2473
2474                 flags = le32_to_cpu(rhead->flags);
2475                 flags |= CEPH_MDS_FLAG_REPLAY;
2476                 rhead->flags = cpu_to_le32(flags);
2477
2478                 if (req->r_target_inode)
2479                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2480
2481                 rhead->num_retry = req->r_attempts - 1;
2482
2483                 /* remove cap/dentry releases from message */
2484                 rhead->num_releases = 0;
2485
2486                 /* time stamp */
2487                 p = msg->front.iov_base + req->r_request_release_offset;
2488                 {
2489                         struct ceph_timespec ts;
2490                         ceph_encode_timespec64(&ts, &req->r_stamp);
2491                         ceph_encode_copy(&p, &ts, sizeof(ts));
2492                 }
2493
2494                 msg->front.iov_len = p - msg->front.iov_base;
2495                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2496                 return 0;
2497         }
2498
2499         if (req->r_request) {
2500                 ceph_msg_put(req->r_request);
2501                 req->r_request = NULL;
2502         }
2503         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2504         if (IS_ERR(msg)) {
2505                 req->r_err = PTR_ERR(msg);
2506                 return PTR_ERR(msg);
2507         }
2508         req->r_request = msg;
2509
2510         rhead = msg->front.iov_base;
2511         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2512         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2513                 flags |= CEPH_MDS_FLAG_REPLAY;
2514         if (req->r_parent)
2515                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2516         rhead->flags = cpu_to_le32(flags);
2517         rhead->num_fwd = req->r_num_fwd;
2518         rhead->num_retry = req->r_attempts - 1;
2519         rhead->ino = 0;
2520
2521         dout(" r_parent = %p\n", req->r_parent);
2522         return 0;
2523 }
2524
2525 /*
2526  * called under mdsc->mutex
2527  */
2528 static int __send_request(struct ceph_mds_client *mdsc,
2529                           struct ceph_mds_session *session,
2530                           struct ceph_mds_request *req,
2531                           bool drop_cap_releases)
2532 {
2533         int err;
2534
2535         err = __prepare_send_request(mdsc, req, session->s_mds,
2536                                      drop_cap_releases);
2537         if (!err) {
2538                 ceph_msg_get(req->r_request);
2539                 ceph_con_send(&session->s_con, req->r_request);
2540         }
2541
2542         return err;
2543 }
2544
2545 /*
2546  * send request, or put it on the appropriate wait list.
2547  */
2548 static void __do_request(struct ceph_mds_client *mdsc,
2549                         struct ceph_mds_request *req)
2550 {
2551         struct ceph_mds_session *session = NULL;
2552         int mds = -1;
2553         int err = 0;
2554
2555         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2556                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2557                         __unregister_request(mdsc, req);
2558                 return;
2559         }
2560
2561         if (req->r_timeout &&
2562             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2563                 dout("do_request timed out\n");
2564                 err = -EIO;
2565                 goto finish;
2566         }
2567         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2568                 dout("do_request forced umount\n");
2569                 err = -EIO;
2570                 goto finish;
2571         }
2572         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2573                 if (mdsc->mdsmap_err) {
2574                         err = mdsc->mdsmap_err;
2575                         dout("do_request mdsmap err %d\n", err);
2576                         goto finish;
2577                 }
2578                 if (mdsc->mdsmap->m_epoch == 0) {
2579                         dout("do_request no mdsmap, waiting for map\n");
2580                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2581                         return;
2582                 }
2583                 if (!(mdsc->fsc->mount_options->flags &
2584                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2585                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2586                         err = -ENOENT;
2587                         pr_info("probably no mds server is up\n");
2588                         goto finish;
2589                 }
2590         }
2591
2592         put_request_session(req);
2593
2594         mds = __choose_mds(mdsc, req);
2595         if (mds < 0 ||
2596             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2597                 dout("do_request no mds or not active, waiting for map\n");
2598                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2599                 return;
2600         }
2601
2602         /* get, open session */
2603         session = __ceph_lookup_mds_session(mdsc, mds);
2604         if (!session) {
2605                 session = register_session(mdsc, mds);
2606                 if (IS_ERR(session)) {
2607                         err = PTR_ERR(session);
2608                         goto finish;
2609                 }
2610         }
2611         req->r_session = get_session(session);
2612
2613         dout("do_request mds%d session %p state %s\n", mds, session,
2614              ceph_session_state_name(session->s_state));
2615         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2616             session->s_state != CEPH_MDS_SESSION_HUNG) {
2617                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2618                         err = -EACCES;
2619                         goto out_session;
2620                 }
2621                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2622                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2623                         __open_session(mdsc, session);
2624                 list_add(&req->r_wait, &session->s_waiting);
2625                 goto out_session;
2626         }
2627
2628         /* send request */
2629         req->r_resend_mds = -1;   /* forget any previous mds hint */
2630
2631         if (req->r_request_started == 0)   /* note request start time */
2632                 req->r_request_started = jiffies;
2633
2634         err = __send_request(mdsc, session, req, false);
2635
2636 out_session:
2637         ceph_put_mds_session(session);
2638 finish:
2639         if (err) {
2640                 dout("__do_request early error %d\n", err);
2641                 req->r_err = err;
2642                 complete_request(mdsc, req);
2643                 __unregister_request(mdsc, req);
2644         }
2645         return;
2646 }
2647
2648 /*
2649  * called under mdsc->mutex
2650  */
2651 static void __wake_requests(struct ceph_mds_client *mdsc,
2652                             struct list_head *head)
2653 {
2654         struct ceph_mds_request *req;
2655         LIST_HEAD(tmp_list);
2656
2657         list_splice_init(head, &tmp_list);
2658
2659         while (!list_empty(&tmp_list)) {
2660                 req = list_entry(tmp_list.next,
2661                                  struct ceph_mds_request, r_wait);
2662                 list_del_init(&req->r_wait);
2663                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2664                 __do_request(mdsc, req);
2665         }
2666 }
2667
2668 /*
2669  * Wake up threads with requests pending for @mds, so that they can
2670  * resubmit their requests to a possibly different mds.
2671  */
2672 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2673 {
2674         struct ceph_mds_request *req;
2675         struct rb_node *p = rb_first(&mdsc->request_tree);
2676
2677         dout("kick_requests mds%d\n", mds);
2678         while (p) {
2679                 req = rb_entry(p, struct ceph_mds_request, r_node);
2680                 p = rb_next(p);
2681                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2682                         continue;
2683                 if (req->r_attempts > 0)
2684                         continue; /* only new requests */
2685                 if (req->r_session &&
2686                     req->r_session->s_mds == mds) {
2687                         dout(" kicking tid %llu\n", req->r_tid);
2688                         list_del_init(&req->r_wait);
2689                         __do_request(mdsc, req);
2690                 }
2691         }
2692 }
2693
2694 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2695                               struct ceph_mds_request *req)
2696 {
2697         int err;
2698
2699         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2700         if (req->r_inode)
2701                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2702         if (req->r_parent) {
2703                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2704                 ihold(req->r_parent);
2705         }
2706         if (req->r_old_dentry_dir)
2707                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2708                                   CEPH_CAP_PIN);
2709
2710         dout("submit_request on %p for inode %p\n", req, dir);
2711         mutex_lock(&mdsc->mutex);
2712         __register_request(mdsc, req, dir);
2713         __do_request(mdsc, req);
2714         err = req->r_err;
2715         mutex_unlock(&mdsc->mutex);
2716         return err;
2717 }
2718
2719 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2720                                   struct ceph_mds_request *req)
2721 {
2722         int err;
2723
2724         /* wait */
2725         dout("do_request waiting\n");
2726         if (!req->r_timeout && req->r_wait_for_completion) {
2727                 err = req->r_wait_for_completion(mdsc, req);
2728         } else {
2729                 long timeleft = wait_for_completion_killable_timeout(
2730                                         &req->r_completion,
2731                                         ceph_timeout_jiffies(req->r_timeout));
2732                 if (timeleft > 0)
2733                         err = 0;
2734                 else if (!timeleft)
2735                         err = -EIO;  /* timed out */
2736                 else
2737                         err = timeleft;  /* killed */
2738         }
2739         dout("do_request waited, got %d\n", err);
2740         mutex_lock(&mdsc->mutex);
2741
2742         /* only abort if we didn't race with a real reply */
2743         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2744                 err = le32_to_cpu(req->r_reply_info.head->result);
2745         } else if (err < 0) {
2746                 dout("aborted request %lld with %d\n", req->r_tid, err);
2747
2748                 /*
2749                  * ensure we aren't running concurrently with
2750                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2751                  * rely on locks (dir mutex) held by our caller.
2752                  */
2753                 mutex_lock(&req->r_fill_mutex);
2754                 req->r_err = err;
2755                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2756                 mutex_unlock(&req->r_fill_mutex);
2757
2758                 if (req->r_parent &&
2759                     (req->r_op & CEPH_MDS_OP_WRITE))
2760                         ceph_invalidate_dir_request(req);
2761         } else {
2762                 err = req->r_err;
2763         }
2764
2765         mutex_unlock(&mdsc->mutex);
2766         return err;
2767 }
2768
2769 /*
2770  * Synchrously perform an mds request.  Take care of all of the
2771  * session setup, forwarding, retry details.
2772  */
2773 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2774                          struct inode *dir,
2775                          struct ceph_mds_request *req)
2776 {
2777         int err;
2778
2779         dout("do_request on %p\n", req);
2780
2781         /* issue */
2782         err = ceph_mdsc_submit_request(mdsc, dir, req);
2783         if (!err)
2784                 err = ceph_mdsc_wait_request(mdsc, req);
2785         dout("do_request %p done, result %d\n", req, err);
2786         return err;
2787 }
2788
2789 /*
2790  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2791  * namespace request.
2792  */
2793 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2794 {
2795         struct inode *dir = req->r_parent;
2796         struct inode *old_dir = req->r_old_dentry_dir;
2797
2798         dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2799
2800         ceph_dir_clear_complete(dir);
2801         if (old_dir)
2802                 ceph_dir_clear_complete(old_dir);
2803         if (req->r_dentry)
2804                 ceph_invalidate_dentry_lease(req->r_dentry);
2805         if (req->r_old_dentry)
2806                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2807 }
2808
2809 /*
2810  * Handle mds reply.
2811  *
2812  * We take the session mutex and parse and process the reply immediately.
2813  * This preserves the logical ordering of replies, capabilities, etc., sent
2814  * by the MDS as they are applied to our local cache.
2815  */
2816 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2817 {
2818         struct ceph_mds_client *mdsc = session->s_mdsc;
2819         struct ceph_mds_request *req;
2820         struct ceph_mds_reply_head *head = msg->front.iov_base;
2821         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2822         struct ceph_snap_realm *realm;
2823         u64 tid;
2824         int err, result;
2825         int mds = session->s_mds;
2826
2827         if (msg->front.iov_len < sizeof(*head)) {
2828                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2829                 ceph_msg_dump(msg);
2830                 return;
2831         }
2832
2833         /* get request, session */
2834         tid = le64_to_cpu(msg->hdr.tid);
2835         mutex_lock(&mdsc->mutex);
2836         req = lookup_get_request(mdsc, tid);
2837         if (!req) {
2838                 dout("handle_reply on unknown tid %llu\n", tid);
2839                 mutex_unlock(&mdsc->mutex);
2840                 return;
2841         }
2842         dout("handle_reply %p\n", req);
2843
2844         /* correct session? */
2845         if (req->r_session != session) {
2846                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2847                        " not mds%d\n", tid, session->s_mds,
2848                        req->r_session ? req->r_session->s_mds : -1);
2849                 mutex_unlock(&mdsc->mutex);
2850                 goto out;
2851         }
2852
2853         /* dup? */
2854         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2855             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2856                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2857                            head->safe ? "safe" : "unsafe", tid, mds);
2858                 mutex_unlock(&mdsc->mutex);
2859                 goto out;
2860         }
2861         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2862                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2863                            tid, mds);
2864                 mutex_unlock(&mdsc->mutex);
2865                 goto out;
2866         }
2867
2868         result = le32_to_cpu(head->result);
2869
2870         /*
2871          * Handle an ESTALE
2872          * if we're not talking to the authority, send to them
2873          * if the authority has changed while we weren't looking,
2874          * send to new authority
2875          * Otherwise we just have to return an ESTALE
2876          */
2877         if (result == -ESTALE) {
2878                 dout("got ESTALE on request %llu\n", req->r_tid);
2879                 req->r_resend_mds = -1;
2880                 if (req->r_direct_mode != USE_AUTH_MDS) {
2881                         dout("not using auth, setting for that now\n");
2882                         req->r_direct_mode = USE_AUTH_MDS;
2883                         __do_request(mdsc, req);
2884                         mutex_unlock(&mdsc->mutex);
2885                         goto out;
2886                 } else  {
2887                         int mds = __choose_mds(mdsc, req);
2888                         if (mds >= 0 && mds != req->r_session->s_mds) {
2889                                 dout("but auth changed, so resending\n");
2890                                 __do_request(mdsc, req);
2891                                 mutex_unlock(&mdsc->mutex);
2892                                 goto out;
2893                         }
2894                 }
2895                 dout("have to return ESTALE on request %llu\n", req->r_tid);
2896         }
2897
2898
2899         if (head->safe) {
2900                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2901                 __unregister_request(mdsc, req);
2902
2903                 /* last request during umount? */
2904                 if (mdsc->stopping && !__get_oldest_req(mdsc))
2905                         complete_all(&mdsc->safe_umount_waiters);
2906
2907                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2908                         /*
2909                          * We already handled the unsafe response, now do the
2910                          * cleanup.  No need to examine the response; the MDS
2911                          * doesn't include any result info in the safe
2912                          * response.  And even if it did, there is nothing
2913                          * useful we could do with a revised return value.
2914                          */
2915                         dout("got safe reply %llu, mds%d\n", tid, mds);
2916
2917                         mutex_unlock(&mdsc->mutex);
2918                         goto out;
2919                 }
2920         } else {
2921                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2922                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2923                 if (req->r_unsafe_dir) {
2924                         struct ceph_inode_info *ci =
2925                                         ceph_inode(req->r_unsafe_dir);
2926                         spin_lock(&ci->i_unsafe_lock);
2927                         list_add_tail(&req->r_unsafe_dir_item,
2928                                       &ci->i_unsafe_dirops);
2929                         spin_unlock(&ci->i_unsafe_lock);
2930                 }
2931         }
2932
2933         dout("handle_reply tid %lld result %d\n", tid, result);
2934         rinfo = &req->r_reply_info;
2935         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
2936                 err = parse_reply_info(msg, rinfo, (u64)-1);
2937         else
2938                 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2939         mutex_unlock(&mdsc->mutex);
2940
2941         mutex_lock(&session->s_mutex);
2942         if (err < 0) {
2943                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2944                 ceph_msg_dump(msg);
2945                 goto out_err;
2946         }
2947
2948         /* snap trace */
2949         realm = NULL;
2950         if (rinfo->snapblob_len) {
2951                 down_write(&mdsc->snap_rwsem);
2952                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2953                                 rinfo->snapblob + rinfo->snapblob_len,
2954                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2955                                 &realm);
2956                 downgrade_write(&mdsc->snap_rwsem);
2957         } else {
2958                 down_read(&mdsc->snap_rwsem);
2959         }
2960
2961         /* insert trace into our cache */
2962         mutex_lock(&req->r_fill_mutex);
2963         current->journal_info = req;
2964         err = ceph_fill_trace(mdsc->fsc->sb, req);
2965         if (err == 0) {
2966                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2967                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2968                         ceph_readdir_prepopulate(req, req->r_session);
2969         }
2970         current->journal_info = NULL;
2971         mutex_unlock(&req->r_fill_mutex);
2972
2973         up_read(&mdsc->snap_rwsem);
2974         if (realm)
2975                 ceph_put_snap_realm(mdsc, realm);
2976
2977         if (err == 0) {
2978                 if (req->r_target_inode &&
2979                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2980                         struct ceph_inode_info *ci =
2981                                 ceph_inode(req->r_target_inode);
2982                         spin_lock(&ci->i_unsafe_lock);
2983                         list_add_tail(&req->r_unsafe_target_item,
2984                                       &ci->i_unsafe_iops);
2985                         spin_unlock(&ci->i_unsafe_lock);
2986                 }
2987
2988                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2989         }
2990 out_err:
2991         mutex_lock(&mdsc->mutex);
2992         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2993                 if (err) {
2994                         req->r_err = err;
2995                 } else {
2996                         req->r_reply =  ceph_msg_get(msg);
2997                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2998                 }
2999         } else {
3000                 dout("reply arrived after request %lld was aborted\n", tid);
3001         }
3002         mutex_unlock(&mdsc->mutex);
3003
3004         mutex_unlock(&session->s_mutex);
3005
3006         /* kick calling process */
3007         complete_request(mdsc, req);
3008 out:
3009         ceph_mdsc_put_request(req);
3010         return;
3011 }
3012
3013
3014
3015 /*
3016  * handle mds notification that our request has been forwarded.
3017  */
3018 static void handle_forward(struct ceph_mds_client *mdsc,
3019                            struct ceph_mds_session *session,
3020                            struct ceph_msg *msg)
3021 {
3022         struct ceph_mds_request *req;
3023         u64 tid = le64_to_cpu(msg->hdr.tid);
3024         u32 next_mds;
3025         u32 fwd_seq;
3026         int err = -EINVAL;
3027         void *p = msg->front.iov_base;
3028         void *end = p + msg->front.iov_len;
3029
3030         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3031         next_mds = ceph_decode_32(&p);
3032         fwd_seq = ceph_decode_32(&p);
3033
3034         mutex_lock(&mdsc->mutex);
3035         req = lookup_get_request(mdsc, tid);
3036         if (!req) {
3037                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3038                 goto out;  /* dup reply? */
3039         }
3040
3041         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3042                 dout("forward tid %llu aborted, unregistering\n", tid);
3043                 __unregister_request(mdsc, req);
3044         } else if (fwd_seq <= req->r_num_fwd) {
3045                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3046                      tid, next_mds, req->r_num_fwd, fwd_seq);
3047         } else {
3048                 /* resend. forward race not possible; mds would drop */
3049                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3050                 BUG_ON(req->r_err);
3051                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3052                 req->r_attempts = 0;
3053                 req->r_num_fwd = fwd_seq;
3054                 req->r_resend_mds = next_mds;
3055                 put_request_session(req);
3056                 __do_request(mdsc, req);
3057         }
3058         ceph_mdsc_put_request(req);
3059 out:
3060         mutex_unlock(&mdsc->mutex);
3061         return;
3062
3063 bad:
3064         pr_err("mdsc_handle_forward decode error err=%d\n", err);
3065 }
3066
3067 static int __decode_session_metadata(void **p, void *end,
3068                                      bool *blacklisted)
3069 {
3070         /* map<string,string> */
3071         u32 n;
3072         bool err_str;
3073         ceph_decode_32_safe(p, end, n, bad);
3074         while (n-- > 0) {
3075                 u32 len;
3076                 ceph_decode_32_safe(p, end, len, bad);
3077                 ceph_decode_need(p, end, len, bad);
3078                 err_str = !strncmp(*p, "error_string", len);
3079                 *p += len;
3080                 ceph_decode_32_safe(p, end, len, bad);
3081                 ceph_decode_need(p, end, len, bad);
3082                 if (err_str && strnstr(*p, "blacklisted", len))
3083                         *blacklisted = true;
3084                 *p += len;
3085         }
3086         return 0;
3087 bad:
3088         return -1;
3089 }
3090
3091 /*
3092  * handle a mds session control message
3093  */
3094 static void handle_session(struct ceph_mds_session *session,
3095                            struct ceph_msg *msg)
3096 {
3097         struct ceph_mds_client *mdsc = session->s_mdsc;
3098         int mds = session->s_mds;
3099         int msg_version = le16_to_cpu(msg->hdr.version);
3100         void *p = msg->front.iov_base;
3101         void *end = p + msg->front.iov_len;
3102         struct ceph_mds_session_head *h;
3103         u32 op;
3104         u64 seq;
3105         unsigned long features = 0;
3106         int wake = 0;
3107         bool blacklisted = false;
3108
3109         /* decode */
3110         ceph_decode_need(&p, end, sizeof(*h), bad);
3111         h = p;
3112         p += sizeof(*h);
3113
3114         op = le32_to_cpu(h->op);
3115         seq = le64_to_cpu(h->seq);
3116
3117         if (msg_version >= 3) {
3118                 u32 len;
3119                 /* version >= 2, metadata */
3120                 if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3121                         goto bad;
3122                 /* version >= 3, feature bits */
3123                 ceph_decode_32_safe(&p, end, len, bad);
3124                 ceph_decode_need(&p, end, len, bad);
3125                 memcpy(&features, p, min_t(size_t, len, sizeof(features)));
3126                 p += len;
3127         }
3128
3129         mutex_lock(&mdsc->mutex);
3130         if (op == CEPH_SESSION_CLOSE) {
3131                 get_session(session);
3132                 __unregister_session(mdsc, session);
3133         }
3134         /* FIXME: this ttl calculation is generous */
3135         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3136         mutex_unlock(&mdsc->mutex);
3137
3138         mutex_lock(&session->s_mutex);
3139
3140         dout("handle_session mds%d %s %p state %s seq %llu\n",
3141              mds, ceph_session_op_name(op), session,
3142              ceph_session_state_name(session->s_state), seq);
3143
3144         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3145                 session->s_state = CEPH_MDS_SESSION_OPEN;
3146                 pr_info("mds%d came back\n", session->s_mds);
3147         }
3148
3149         switch (op) {
3150         case CEPH_SESSION_OPEN:
3151                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3152                         pr_info("mds%d reconnect success\n", session->s_mds);
3153                 session->s_state = CEPH_MDS_SESSION_OPEN;
3154                 session->s_features = features;
3155                 renewed_caps(mdsc, session, 0);
3156                 wake = 1;
3157                 if (mdsc->stopping)
3158                         __close_session(mdsc, session);
3159                 break;
3160
3161         case CEPH_SESSION_RENEWCAPS:
3162                 if (session->s_renew_seq == seq)
3163                         renewed_caps(mdsc, session, 1);
3164                 break;
3165
3166         case CEPH_SESSION_CLOSE:
3167                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3168                         pr_info("mds%d reconnect denied\n", session->s_mds);
3169                 cleanup_session_requests(mdsc, session);
3170                 remove_session_caps(session);
3171                 wake = 2; /* for good measure */
3172                 wake_up_all(&mdsc->session_close_wq);
3173                 break;
3174
3175         case CEPH_SESSION_STALE:
3176                 pr_info("mds%d caps went stale, renewing\n",
3177                         session->s_mds);
3178                 spin_lock(&session->s_gen_ttl_lock);
3179                 session->s_cap_gen++;
3180                 session->s_cap_ttl = jiffies - 1;
3181                 spin_unlock(&session->s_gen_ttl_lock);
3182                 send_renew_caps(mdsc, session);
3183                 break;
3184
3185         case CEPH_SESSION_RECALL_STATE:
3186                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3187                 break;
3188
3189         case CEPH_SESSION_FLUSHMSG:
3190                 send_flushmsg_ack(mdsc, session, seq);
3191                 break;
3192
3193         case CEPH_SESSION_FORCE_RO:
3194                 dout("force_session_readonly %p\n", session);
3195                 spin_lock(&session->s_cap_lock);
3196                 session->s_readonly = true;
3197                 spin_unlock(&session->s_cap_lock);
3198                 wake_up_session_caps(session, FORCE_RO);
3199                 break;
3200
3201         case CEPH_SESSION_REJECT:
3202                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3203                 pr_info("mds%d rejected session\n", session->s_mds);
3204                 session->s_state = CEPH_MDS_SESSION_REJECTED;
3205                 cleanup_session_requests(mdsc, session);
3206                 remove_session_caps(session);
3207                 if (blacklisted)
3208                         mdsc->fsc->blacklisted = true;
3209                 wake = 2; /* for good measure */
3210                 break;
3211
3212         default:
3213                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3214                 WARN_ON(1);
3215         }
3216
3217         mutex_unlock(&session->s_mutex);
3218         if (wake) {
3219                 mutex_lock(&mdsc->mutex);
3220                 __wake_requests(mdsc, &session->s_waiting);
3221                 if (wake == 2)
3222                         kick_requests(mdsc, mds);
3223                 mutex_unlock(&mdsc->mutex);
3224         }
3225         if (op == CEPH_SESSION_CLOSE)
3226                 ceph_put_mds_session(session);
3227         return;
3228
3229 bad:
3230         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3231                (int)msg->front.iov_len);
3232         ceph_msg_dump(msg);
3233         return;
3234 }
3235
3236 /*
3237  * called under session->mutex.
3238  */
3239 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3240                                    struct ceph_mds_session *session)
3241 {
3242         struct ceph_mds_request *req, *nreq;
3243         struct rb_node *p;
3244
3245         dout("replay_unsafe_requests mds%d\n", session->s_mds);
3246
3247         mutex_lock(&mdsc->mutex);
3248         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3249                 __send_request(mdsc, session, req, true);
3250
3251         /*
3252          * also re-send old requests when MDS enters reconnect stage. So that MDS
3253          * can process completed request in clientreplay stage.
3254          */
3255         p = rb_first(&mdsc->request_tree);
3256         while (p) {
3257                 req = rb_entry(p, struct ceph_mds_request, r_node);
3258                 p = rb_next(p);
3259                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3260                         continue;
3261                 if (req->r_attempts == 0)
3262                         continue; /* only old requests */
3263                 if (req->r_session &&
3264                     req->r_session->s_mds == session->s_mds)
3265                         __send_request(mdsc, session, req, true);
3266         }
3267         mutex_unlock(&mdsc->mutex);
3268 }
3269
3270 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3271 {
3272         struct ceph_msg *reply;
3273         struct ceph_pagelist *_pagelist;
3274         struct page *page;
3275         __le32 *addr;
3276         int err = -ENOMEM;
3277
3278         if (!recon_state->allow_multi)
3279                 return -ENOSPC;
3280
3281         /* can't handle message that contains both caps and realm */
3282         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3283
3284         /* pre-allocate new pagelist */
3285         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3286         if (!_pagelist)
3287                 return -ENOMEM;
3288
3289         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3290         if (!reply)
3291                 goto fail_msg;
3292
3293         /* placeholder for nr_caps */
3294         err = ceph_pagelist_encode_32(_pagelist, 0);
3295         if (err < 0)
3296                 goto fail;
3297
3298         if (recon_state->nr_caps) {
3299                 /* currently encoding caps */
3300                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3301                 if (err)
3302                         goto fail;
3303         } else {
3304                 /* placeholder for nr_realms (currently encoding relams) */
3305                 err = ceph_pagelist_encode_32(_pagelist, 0);
3306                 if (err < 0)
3307                         goto fail;
3308         }
3309
3310         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3311         if (err)
3312                 goto fail;
3313
3314         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3315         addr = kmap_atomic(page);
3316         if (recon_state->nr_caps) {
3317                 /* currently encoding caps */
3318                 *addr = cpu_to_le32(recon_state->nr_caps);
3319         } else {
3320                 /* currently encoding relams */
3321                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3322         }
3323         kunmap_atomic(addr);
3324
3325         reply->hdr.version = cpu_to_le16(5);
3326         reply->hdr.compat_version = cpu_to_le16(4);
3327
3328         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3329         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3330
3331         ceph_con_send(&recon_state->session->s_con, reply);
3332         ceph_pagelist_release(recon_state->pagelist);
3333
3334         recon_state->pagelist = _pagelist;
3335         recon_state->nr_caps = 0;
3336         recon_state->nr_realms = 0;
3337         recon_state->msg_version = 5;
3338         return 0;
3339 fail:
3340         ceph_msg_put(reply);
3341 fail_msg:
3342         ceph_pagelist_release(_pagelist);
3343         return err;
3344 }
3345
3346 /*
3347  * Encode information about a cap for a reconnect with the MDS.
3348  */
3349 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
3350                           void *arg)
3351 {
3352         union {
3353                 struct ceph_mds_cap_reconnect v2;
3354                 struct ceph_mds_cap_reconnect_v1 v1;
3355         } rec;
3356         struct ceph_inode_info *ci = cap->ci;
3357         struct ceph_reconnect_state *recon_state = arg;
3358         struct ceph_pagelist *pagelist = recon_state->pagelist;
3359         int err;
3360         u64 snap_follows;
3361
3362         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3363              inode, ceph_vinop(inode), cap, cap->cap_id,
3364              ceph_cap_string(cap->issued));
3365
3366         spin_lock(&ci->i_ceph_lock);
3367         cap->seq = 0;        /* reset cap seq */
3368         cap->issue_seq = 0;  /* and issue_seq */
3369         cap->mseq = 0;       /* and migrate_seq */
3370         cap->cap_gen = cap->session->s_cap_gen;
3371
3372         if (recon_state->msg_version >= 2) {
3373                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3374                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3375                 rec.v2.issued = cpu_to_le32(cap->issued);
3376                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3377                 rec.v2.pathbase = 0;
3378                 rec.v2.flock_len = (__force __le32)
3379                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3380         } else {
3381                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3382                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3383                 rec.v1.issued = cpu_to_le32(cap->issued);
3384                 rec.v1.size = cpu_to_le64(inode->i_size);
3385                 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3386                 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3387                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3388                 rec.v1.pathbase = 0;
3389         }
3390
3391         if (list_empty(&ci->i_cap_snaps)) {
3392                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3393         } else {
3394                 struct ceph_cap_snap *capsnap =
3395                         list_first_entry(&ci->i_cap_snaps,
3396                                          struct ceph_cap_snap, ci_item);
3397                 snap_follows = capsnap->follows;
3398         }
3399         spin_unlock(&ci->i_ceph_lock);
3400
3401         if (recon_state->msg_version >= 2) {
3402                 int num_fcntl_locks, num_flock_locks;
3403                 struct ceph_filelock *flocks = NULL;
3404                 size_t struct_len, total_len = sizeof(u64);
3405                 u8 struct_v = 0;
3406
3407 encode_again:
3408                 if (rec.v2.flock_len) {
3409                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3410                 } else {
3411                         num_fcntl_locks = 0;
3412                         num_flock_locks = 0;
3413                 }
3414                 if (num_fcntl_locks + num_flock_locks > 0) {
3415                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3416                                                sizeof(struct ceph_filelock),
3417                                                GFP_NOFS);
3418                         if (!flocks) {
3419                                 err = -ENOMEM;
3420                                 goto out_err;
3421                         }
3422                         err = ceph_encode_locks_to_buffer(inode, flocks,
3423                                                           num_fcntl_locks,
3424                                                           num_flock_locks);
3425                         if (err) {
3426                                 kfree(flocks);
3427                                 flocks = NULL;
3428                                 if (err == -ENOSPC)
3429                                         goto encode_again;
3430                                 goto out_err;
3431                         }
3432                 } else {
3433                         kfree(flocks);
3434                         flocks = NULL;
3435                 }
3436
3437                 if (recon_state->msg_version >= 3) {
3438                         /* version, compat_version and struct_len */
3439                         total_len += 2 * sizeof(u8) + sizeof(u32);
3440                         struct_v = 2;
3441                 }
3442                 /*
3443                  * number of encoded locks is stable, so copy to pagelist
3444                  */
3445                 struct_len = 2 * sizeof(u32) +
3446                             (num_fcntl_locks + num_flock_locks) *
3447                             sizeof(struct ceph_filelock);
3448                 rec.v2.flock_len = cpu_to_le32(struct_len);
3449
3450                 struct_len += sizeof(u32) + sizeof(rec.v2);
3451
3452                 if (struct_v >= 2)
3453                         struct_len += sizeof(u64); /* snap_follows */
3454
3455                 total_len += struct_len;
3456
3457                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3458                         err = send_reconnect_partial(recon_state);
3459                         if (err)
3460                                 goto out_freeflocks;
3461                         pagelist = recon_state->pagelist;
3462                 }
3463
3464                 err = ceph_pagelist_reserve(pagelist, total_len);
3465                 if (err)
3466                         goto out_freeflocks;
3467
3468                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3469                 if (recon_state->msg_version >= 3) {
3470                         ceph_pagelist_encode_8(pagelist, struct_v);
3471                         ceph_pagelist_encode_8(pagelist, 1);
3472                         ceph_pagelist_encode_32(pagelist, struct_len);
3473                 }
3474                 ceph_pagelist_encode_string(pagelist, NULL, 0);
3475                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3476                 ceph_locks_to_pagelist(flocks, pagelist,
3477                                        num_fcntl_locks, num_flock_locks);
3478                 if (struct_v >= 2)
3479                         ceph_pagelist_encode_64(pagelist, snap_follows);
3480 out_freeflocks:
3481                 kfree(flocks);
3482         } else {
3483                 u64 pathbase = 0;
3484                 int pathlen = 0;
3485                 char *path = NULL;
3486                 struct dentry *dentry;
3487
3488                 dentry = d_find_alias(inode);
3489                 if (dentry) {
3490                         path = ceph_mdsc_build_path(dentry,
3491                                                 &pathlen, &pathbase, 0);
3492                         dput(dentry);
3493                         if (IS_ERR(path)) {
3494                                 err = PTR_ERR(path);
3495                                 goto out_err;
3496                         }
3497                         rec.v1.pathbase = cpu_to_le64(pathbase);
3498                 }
3499
3500                 err = ceph_pagelist_reserve(pagelist,
3501                                             sizeof(u64) + sizeof(u32) +
3502                                             pathlen + sizeof(rec.v1));
3503                 if (err) {
3504                         goto out_freepath;
3505                 }
3506
3507                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3508                 ceph_pagelist_encode_string(pagelist, path, pathlen);
3509                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3510 out_freepath:
3511                 ceph_mdsc_free_path(path, pathlen);
3512         }
3513
3514 out_err:
3515         if (err >= 0)
3516                 recon_state->nr_caps++;
3517         return err;
3518 }
3519
3520 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3521                               struct ceph_reconnect_state *recon_state)
3522 {
3523         struct rb_node *p;
3524         struct ceph_pagelist *pagelist = recon_state->pagelist;
3525         int err = 0;
3526
3527         if (recon_state->msg_version >= 4) {
3528                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3529                 if (err < 0)
3530                         goto fail;
3531         }
3532
3533         /*
3534          * snaprealms.  we provide mds with the ino, seq (version), and
3535          * parent for all of our realms.  If the mds has any newer info,
3536          * it will tell us.
3537          */
3538         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3539                 struct ceph_snap_realm *realm =
3540                        rb_entry(p, struct ceph_snap_realm, node);
3541                 struct ceph_mds_snaprealm_reconnect sr_rec;
3542
3543                 if (recon_state->msg_version >= 4) {
3544                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
3545                                       sizeof(sr_rec);
3546
3547                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3548                                 err = send_reconnect_partial(recon_state);
3549                                 if (err)
3550                                         goto fail;
3551                                 pagelist = recon_state->pagelist;
3552                         }
3553
3554                         err = ceph_pagelist_reserve(pagelist, need);
3555                         if (err)
3556                                 goto fail;
3557
3558                         ceph_pagelist_encode_8(pagelist, 1);
3559                         ceph_pagelist_encode_8(pagelist, 1);
3560                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3561                 }
3562
3563                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3564                      realm->ino, realm->seq, realm->parent_ino);
3565                 sr_rec.ino = cpu_to_le64(realm->ino);
3566                 sr_rec.seq = cpu_to_le64(realm->seq);
3567                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3568
3569                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3570                 if (err)
3571                         goto fail;
3572
3573                 recon_state->nr_realms++;
3574         }
3575 fail:
3576         return err;
3577 }
3578
3579
3580 /*
3581  * If an MDS fails and recovers, clients need to reconnect in order to
3582  * reestablish shared state.  This includes all caps issued through
3583  * this session _and_ the snap_realm hierarchy.  Because it's not
3584  * clear which snap realms the mds cares about, we send everything we
3585  * know about.. that ensures we'll then get any new info the
3586  * recovering MDS might have.
3587  *
3588  * This is a relatively heavyweight operation, but it's rare.
3589  *
3590  * called with mdsc->mutex held.
3591  */
3592 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3593                                struct ceph_mds_session *session)
3594 {
3595         struct ceph_msg *reply;
3596         int mds = session->s_mds;
3597         int err = -ENOMEM;
3598         struct ceph_reconnect_state recon_state = {
3599                 .session = session,
3600         };
3601         LIST_HEAD(dispose);
3602
3603         pr_info("mds%d reconnect start\n", mds);
3604
3605         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3606         if (!recon_state.pagelist)
3607                 goto fail_nopagelist;
3608
3609         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3610         if (!reply)
3611                 goto fail_nomsg;
3612
3613         mutex_lock(&session->s_mutex);
3614         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3615         session->s_seq = 0;
3616
3617         dout("session %p state %s\n", session,
3618              ceph_session_state_name(session->s_state));
3619
3620         spin_lock(&session->s_gen_ttl_lock);
3621         session->s_cap_gen++;
3622         spin_unlock(&session->s_gen_ttl_lock);
3623
3624         spin_lock(&session->s_cap_lock);
3625         /* don't know if session is readonly */
3626         session->s_readonly = 0;
3627         /*
3628          * notify __ceph_remove_cap() that we are composing cap reconnect.
3629          * If a cap get released before being added to the cap reconnect,
3630          * __ceph_remove_cap() should skip queuing cap release.
3631          */
3632         session->s_cap_reconnect = 1;
3633         /* drop old cap expires; we're about to reestablish that state */
3634         detach_cap_releases(session, &dispose);
3635         spin_unlock(&session->s_cap_lock);
3636         dispose_cap_releases(mdsc, &dispose);
3637
3638         /* trim unused caps to reduce MDS's cache rejoin time */
3639         if (mdsc->fsc->sb->s_root)
3640                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3641
3642         ceph_con_close(&session->s_con);
3643         ceph_con_open(&session->s_con,
3644                       CEPH_ENTITY_TYPE_MDS, mds,
3645                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3646
3647         /* replay unsafe requests */
3648         replay_unsafe_requests(mdsc, session);
3649
3650         ceph_early_kick_flushing_caps(mdsc, session);
3651
3652         down_read(&mdsc->snap_rwsem);
3653
3654         /* placeholder for nr_caps */
3655         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3656         if (err)
3657                 goto fail;
3658
3659         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3660                 recon_state.msg_version = 3;
3661                 recon_state.allow_multi = true;
3662         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3663                 recon_state.msg_version = 3;
3664         } else {
3665                 recon_state.msg_version = 2;
3666         }
3667         /* trsaverse this session's caps */
3668         err = ceph_iterate_session_caps(session, encode_caps_cb, &recon_state);
3669
3670         spin_lock(&session->s_cap_lock);
3671         session->s_cap_reconnect = 0;
3672         spin_unlock(&session->s_cap_lock);
3673
3674         if (err < 0)
3675                 goto fail;
3676
3677         /* check if all realms can be encoded into current message */
3678         if (mdsc->num_snap_realms) {
3679                 size_t total_len =
3680                         recon_state.pagelist->length +
3681                         mdsc->num_snap_realms *
3682                         sizeof(struct ceph_mds_snaprealm_reconnect);
3683                 if (recon_state.msg_version >= 4) {
3684                         /* number of realms */
3685                         total_len += sizeof(u32);
3686                         /* version, compat_version and struct_len */
3687                         total_len += mdsc->num_snap_realms *
3688                                      (2 * sizeof(u8) + sizeof(u32));
3689                 }
3690                 if (total_len > RECONNECT_MAX_SIZE) {
3691                         if (!recon_state.allow_multi) {
3692                                 err = -ENOSPC;
3693                                 goto fail;
3694                         }
3695                         if (recon_state.nr_caps) {
3696                                 err = send_reconnect_partial(&recon_state);
3697                                 if (err)
3698                                         goto fail;
3699                         }
3700                         recon_state.msg_version = 5;
3701                 }
3702         }
3703
3704         err = encode_snap_realms(mdsc, &recon_state);
3705         if (err < 0)
3706                 goto fail;
3707
3708         if (recon_state.msg_version >= 5) {
3709                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3710                 if (err < 0)
3711                         goto fail;
3712         }
3713
3714         if (recon_state.nr_caps || recon_state.nr_realms) {
3715                 struct page *page =
3716                         list_first_entry(&recon_state.pagelist->head,
3717                                         struct page, lru);
3718                 __le32 *addr = kmap_atomic(page);
3719                 if (recon_state.nr_caps) {
3720                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3721                         *addr = cpu_to_le32(recon_state.nr_caps);
3722                 } else if (recon_state.msg_version >= 4) {
3723                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3724                 }
3725                 kunmap_atomic(addr);
3726         }
3727
3728         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3729         if (recon_state.msg_version >= 4)
3730                 reply->hdr.compat_version = cpu_to_le16(4);
3731
3732         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3733         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3734
3735         ceph_con_send(&session->s_con, reply);
3736
3737         mutex_unlock(&session->s_mutex);
3738
3739         mutex_lock(&mdsc->mutex);
3740         __wake_requests(mdsc, &session->s_waiting);
3741         mutex_unlock(&mdsc->mutex);
3742
3743         up_read(&mdsc->snap_rwsem);
3744         ceph_pagelist_release(recon_state.pagelist);
3745         return;
3746
3747 fail:
3748         ceph_msg_put(reply);
3749         up_read(&mdsc->snap_rwsem);
3750         mutex_unlock(&session->s_mutex);
3751 fail_nomsg:
3752         ceph_pagelist_release(recon_state.pagelist);
3753 fail_nopagelist:
3754         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3755         return;
3756 }
3757
3758
3759 /*
3760  * compare old and new mdsmaps, kicking requests
3761  * and closing out old connections as necessary
3762  *
3763  * called under mdsc->mutex.
3764  */
3765 static void check_new_map(struct ceph_mds_client *mdsc,
3766                           struct ceph_mdsmap *newmap,
3767                           struct ceph_mdsmap *oldmap)
3768 {
3769         int i;
3770         int oldstate, newstate;
3771         struct ceph_mds_session *s;
3772
3773         dout("check_new_map new %u old %u\n",
3774              newmap->m_epoch, oldmap->m_epoch);
3775
3776         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3777                 if (!mdsc->sessions[i])
3778                         continue;
3779                 s = mdsc->sessions[i];
3780                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3781                 newstate = ceph_mdsmap_get_state(newmap, i);
3782
3783                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3784                      i, ceph_mds_state_name(oldstate),
3785                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3786                      ceph_mds_state_name(newstate),
3787                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3788                      ceph_session_state_name(s->s_state));
3789
3790                 if (i >= newmap->m_num_mds) {
3791                         /* force close session for stopped mds */
3792                         get_session(s);
3793                         __unregister_session(mdsc, s);
3794                         __wake_requests(mdsc, &s->s_waiting);
3795                         mutex_unlock(&mdsc->mutex);
3796
3797                         mutex_lock(&s->s_mutex);
3798                         cleanup_session_requests(mdsc, s);
3799                         remove_session_caps(s);
3800                         mutex_unlock(&s->s_mutex);
3801
3802                         ceph_put_mds_session(s);
3803
3804                         mutex_lock(&mdsc->mutex);
3805                         kick_requests(mdsc, i);
3806                         continue;
3807                 }
3808
3809                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
3810                            ceph_mdsmap_get_addr(newmap, i),
3811                            sizeof(struct ceph_entity_addr))) {
3812                         /* just close it */
3813                         mutex_unlock(&mdsc->mutex);
3814                         mutex_lock(&s->s_mutex);
3815                         mutex_lock(&mdsc->mutex);
3816                         ceph_con_close(&s->s_con);
3817                         mutex_unlock(&s->s_mutex);
3818                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
3819                 } else if (oldstate == newstate) {
3820                         continue;  /* nothing new with this mds */
3821                 }
3822
3823                 /*
3824                  * send reconnect?
3825                  */
3826                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3827                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3828                         mutex_unlock(&mdsc->mutex);
3829                         send_mds_reconnect(mdsc, s);
3830                         mutex_lock(&mdsc->mutex);
3831                 }
3832
3833                 /*
3834                  * kick request on any mds that has gone active.
3835                  */
3836                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3837                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3838                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3839                             oldstate != CEPH_MDS_STATE_STARTING)
3840                                 pr_info("mds%d recovery completed\n", s->s_mds);
3841                         kick_requests(mdsc, i);
3842                         ceph_kick_flushing_caps(mdsc, s);
3843                         wake_up_session_caps(s, RECONNECT);
3844                 }
3845         }
3846
3847         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3848                 s = mdsc->sessions[i];
3849                 if (!s)
3850                         continue;
3851                 if (!ceph_mdsmap_is_laggy(newmap, i))
3852                         continue;
3853                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3854                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3855                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3856                         dout(" connecting to export targets of laggy mds%d\n",
3857                              i);
3858                         __open_export_target_sessions(mdsc, s);
3859                 }
3860         }
3861 }
3862
3863
3864
3865 /*
3866  * leases
3867  */
3868
3869 /*
3870  * caller must hold session s_mutex, dentry->d_lock
3871  */
3872 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3873 {
3874         struct ceph_dentry_info *di = ceph_dentry(dentry);
3875
3876         ceph_put_mds_session(di->lease_session);
3877         di->lease_session = NULL;
3878 }
3879
3880 static void handle_lease(struct ceph_mds_client *mdsc,
3881                          struct ceph_mds_session *session,
3882                          struct ceph_msg *msg)
3883 {
3884         struct super_block *sb = mdsc->fsc->sb;
3885         struct inode *inode;
3886         struct dentry *parent, *dentry;
3887         struct ceph_dentry_info *di;
3888         int mds = session->s_mds;
3889         struct ceph_mds_lease *h = msg->front.iov_base;
3890         u32 seq;
3891         struct ceph_vino vino;
3892         struct qstr dname;
3893         int release = 0;
3894
3895         dout("handle_lease from mds%d\n", mds);
3896
3897         /* decode */
3898         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3899                 goto bad;
3900         vino.ino = le64_to_cpu(h->ino);
3901         vino.snap = CEPH_NOSNAP;
3902         seq = le32_to_cpu(h->seq);
3903         dname.len = get_unaligned_le32(h + 1);
3904         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
3905                 goto bad;
3906         dname.name = (void *)(h + 1) + sizeof(u32);
3907
3908         /* lookup inode */
3909         inode = ceph_find_inode(sb, vino);
3910         dout("handle_lease %s, ino %llx %p %.*s\n",
3911              ceph_lease_op_name(h->action), vino.ino, inode,
3912              dname.len, dname.name);
3913
3914         mutex_lock(&session->s_mutex);
3915         session->s_seq++;
3916
3917         if (!inode) {
3918                 dout("handle_lease no inode %llx\n", vino.ino);
3919                 goto release;
3920         }
3921
3922         /* dentry */
3923         parent = d_find_alias(inode);
3924         if (!parent) {
3925                 dout("no parent dentry on inode %p\n", inode);
3926                 WARN_ON(1);
3927                 goto release;  /* hrm... */
3928         }
3929         dname.hash = full_name_hash(parent, dname.name, dname.len);
3930         dentry = d_lookup(parent, &dname);
3931         dput(parent);
3932         if (!dentry)
3933                 goto release;
3934
3935         spin_lock(&dentry->d_lock);
3936         di = ceph_dentry(dentry);
3937         switch (h->action) {
3938         case CEPH_MDS_LEASE_REVOKE:
3939                 if (di->lease_session == session) {
3940                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3941                                 h->seq = cpu_to_le32(di->lease_seq);
3942                         __ceph_mdsc_drop_dentry_lease(dentry);
3943                 }
3944                 release = 1;
3945                 break;
3946
3947         case CEPH_MDS_LEASE_RENEW:
3948                 if (di->lease_session == session &&
3949                     di->lease_gen == session->s_cap_gen &&
3950                     di->lease_renew_from &&
3951                     di->lease_renew_after == 0) {
3952                         unsigned long duration =
3953                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3954
3955                         di->lease_seq = seq;
3956                         di->time = di->lease_renew_from + duration;
3957                         di->lease_renew_after = di->lease_renew_from +
3958                                 (duration >> 1);
3959                         di->lease_renew_from = 0;
3960                 }
3961                 break;
3962         }
3963         spin_unlock(&dentry->d_lock);
3964         dput(dentry);
3965
3966         if (!release)
3967                 goto out;
3968
3969 release:
3970         /* let's just reuse the same message */
3971         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3972         ceph_msg_get(msg);
3973         ceph_con_send(&session->s_con, msg);
3974
3975 out:
3976         mutex_unlock(&session->s_mutex);
3977         /* avoid calling iput_final() in mds dispatch threads */
3978         ceph_async_iput(inode);
3979         return;
3980
3981 bad:
3982         pr_err("corrupt lease message\n");
3983         ceph_msg_dump(msg);
3984 }
3985
3986 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3987                               struct dentry *dentry, char action,
3988                               u32 seq)
3989 {
3990         struct ceph_msg *msg;
3991         struct ceph_mds_lease *lease;
3992         struct inode *dir;
3993         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
3994
3995         dout("lease_send_msg identry %p %s to mds%d\n",
3996              dentry, ceph_lease_op_name(action), session->s_mds);
3997
3998         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3999         if (!msg)
4000                 return;
4001         lease = msg->front.iov_base;
4002         lease->action = action;
4003         lease->seq = cpu_to_le32(seq);
4004
4005         spin_lock(&dentry->d_lock);
4006         dir = d_inode(dentry->d_parent);
4007         lease->ino = cpu_to_le64(ceph_ino(dir));
4008         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4009
4010         put_unaligned_le32(dentry->d_name.len, lease + 1);
4011         memcpy((void *)(lease + 1) + 4,
4012                dentry->d_name.name, dentry->d_name.len);
4013         spin_unlock(&dentry->d_lock);
4014         /*
4015          * if this is a preemptive lease RELEASE, no need to
4016          * flush request stream, since the actual request will
4017          * soon follow.
4018          */
4019         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4020
4021         ceph_con_send(&session->s_con, msg);
4022 }
4023
4024 /*
4025  * lock unlock sessions, to wait ongoing session activities
4026  */
4027 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4028 {
4029         int i;
4030
4031         mutex_lock(&mdsc->mutex);
4032         for (i = 0; i < mdsc->max_sessions; i++) {
4033                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4034                 if (!s)
4035                         continue;
4036                 mutex_unlock(&mdsc->mutex);
4037                 mutex_lock(&s->s_mutex);
4038                 mutex_unlock(&s->s_mutex);
4039                 ceph_put_mds_session(s);
4040                 mutex_lock(&mdsc->mutex);
4041         }
4042         mutex_unlock(&mdsc->mutex);
4043 }
4044
4045 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4046 {
4047         struct ceph_fs_client *fsc = mdsc->fsc;
4048
4049         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4050                 return;
4051
4052         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4053                 return;
4054
4055         if (!READ_ONCE(fsc->blacklisted))
4056                 return;
4057
4058         if (fsc->last_auto_reconnect &&
4059             time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4060                 return;
4061
4062         pr_info("auto reconnect after blacklisted\n");
4063         fsc->last_auto_reconnect = jiffies;
4064         ceph_force_reconnect(fsc->sb);
4065 }
4066
4067 /*
4068  * delayed work -- periodically trim expired leases, renew caps with mds
4069  */
4070 static void schedule_delayed(struct ceph_mds_client *mdsc)
4071 {
4072         int delay = 5;
4073         unsigned hz = round_jiffies_relative(HZ * delay);
4074         schedule_delayed_work(&mdsc->delayed_work, hz);
4075 }
4076
4077 static void delayed_work(struct work_struct *work)
4078 {
4079         int i;
4080         struct ceph_mds_client *mdsc =
4081                 container_of(work, struct ceph_mds_client, delayed_work.work);
4082         int renew_interval;
4083         int renew_caps;
4084
4085         dout("mdsc delayed_work\n");
4086
4087         mutex_lock(&mdsc->mutex);
4088         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4089         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4090                                    mdsc->last_renew_caps);
4091         if (renew_caps)
4092                 mdsc->last_renew_caps = jiffies;
4093
4094         for (i = 0; i < mdsc->max_sessions; i++) {
4095                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4096                 if (!s)
4097                         continue;
4098                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4099                         dout("resending session close request for mds%d\n",
4100                              s->s_mds);
4101                         request_close_session(mdsc, s);
4102                         ceph_put_mds_session(s);
4103                         continue;
4104                 }
4105                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4106                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4107                                 s->s_state = CEPH_MDS_SESSION_HUNG;
4108                                 pr_info("mds%d hung\n", s->s_mds);
4109                         }
4110                 }
4111                 if (s->s_state == CEPH_MDS_SESSION_NEW ||
4112                     s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4113                     s->s_state == CEPH_MDS_SESSION_REJECTED) {
4114                         /* this mds is failed or recovering, just wait */
4115                         ceph_put_mds_session(s);
4116                         continue;
4117                 }
4118                 mutex_unlock(&mdsc->mutex);
4119
4120                 mutex_lock(&s->s_mutex);
4121                 if (renew_caps)
4122                         send_renew_caps(mdsc, s);
4123                 else
4124                         ceph_con_keepalive(&s->s_con);
4125                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4126                     s->s_state == CEPH_MDS_SESSION_HUNG)
4127                         ceph_send_cap_releases(mdsc, s);
4128                 mutex_unlock(&s->s_mutex);
4129                 ceph_put_mds_session(s);
4130
4131                 mutex_lock(&mdsc->mutex);
4132         }
4133         mutex_unlock(&mdsc->mutex);
4134
4135         ceph_check_delayed_caps(mdsc);
4136
4137         ceph_queue_cap_reclaim_work(mdsc);
4138
4139         ceph_trim_snapid_map(mdsc);
4140
4141         maybe_recover_session(mdsc);
4142
4143         schedule_delayed(mdsc);
4144 }
4145
4146 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4147
4148 {
4149         struct ceph_mds_client *mdsc;
4150
4151         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4152         if (!mdsc)
4153                 return -ENOMEM;
4154         mdsc->fsc = fsc;
4155         mutex_init(&mdsc->mutex);
4156         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4157         if (!mdsc->mdsmap) {
4158                 kfree(mdsc);
4159                 return -ENOMEM;
4160         }
4161
4162         fsc->mdsc = mdsc;
4163         init_completion(&mdsc->safe_umount_waiters);
4164         init_waitqueue_head(&mdsc->session_close_wq);
4165         INIT_LIST_HEAD(&mdsc->waiting_for_map);
4166         mdsc->sessions = NULL;
4167         atomic_set(&mdsc->num_sessions, 0);
4168         mdsc->max_sessions = 0;
4169         mdsc->stopping = 0;
4170         atomic64_set(&mdsc->quotarealms_count, 0);
4171         mdsc->quotarealms_inodes = RB_ROOT;
4172         mutex_init(&mdsc->quotarealms_inodes_mutex);
4173         mdsc->last_snap_seq = 0;
4174         init_rwsem(&mdsc->snap_rwsem);
4175         mdsc->snap_realms = RB_ROOT;
4176         INIT_LIST_HEAD(&mdsc->snap_empty);
4177         mdsc->num_snap_realms = 0;
4178         spin_lock_init(&mdsc->snap_empty_lock);
4179         mdsc->last_tid = 0;
4180         mdsc->oldest_tid = 0;
4181         mdsc->request_tree = RB_ROOT;
4182         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4183         mdsc->last_renew_caps = jiffies;
4184         INIT_LIST_HEAD(&mdsc->cap_delay_list);
4185         INIT_LIST_HEAD(&mdsc->cap_wait_list);
4186         spin_lock_init(&mdsc->cap_delay_lock);
4187         INIT_LIST_HEAD(&mdsc->snap_flush_list);
4188         spin_lock_init(&mdsc->snap_flush_lock);
4189         mdsc->last_cap_flush_tid = 1;
4190         INIT_LIST_HEAD(&mdsc->cap_flush_list);
4191         INIT_LIST_HEAD(&mdsc->cap_dirty);
4192         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4193         mdsc->num_cap_flushing = 0;
4194         spin_lock_init(&mdsc->cap_dirty_lock);
4195         init_waitqueue_head(&mdsc->cap_flushing_wq);
4196         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4197         atomic_set(&mdsc->cap_reclaim_pending, 0);
4198
4199         spin_lock_init(&mdsc->dentry_list_lock);
4200         INIT_LIST_HEAD(&mdsc->dentry_leases);
4201         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4202
4203         ceph_caps_init(mdsc);
4204         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4205
4206         spin_lock_init(&mdsc->snapid_map_lock);
4207         mdsc->snapid_map_tree = RB_ROOT;
4208         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4209
4210         init_rwsem(&mdsc->pool_perm_rwsem);
4211         mdsc->pool_perm_tree = RB_ROOT;
4212
4213         strscpy(mdsc->nodename, utsname()->nodename,
4214                 sizeof(mdsc->nodename));
4215         return 0;
4216 }
4217
4218 /*
4219  * Wait for safe replies on open mds requests.  If we time out, drop
4220  * all requests from the tree to avoid dangling dentry refs.
4221  */
4222 static void wait_requests(struct ceph_mds_client *mdsc)
4223 {
4224         struct ceph_options *opts = mdsc->fsc->client->options;
4225         struct ceph_mds_request *req;
4226
4227         mutex_lock(&mdsc->mutex);
4228         if (__get_oldest_req(mdsc)) {
4229                 mutex_unlock(&mdsc->mutex);
4230
4231                 dout("wait_requests waiting for requests\n");
4232                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4233                                     ceph_timeout_jiffies(opts->mount_timeout));
4234
4235                 /* tear down remaining requests */
4236                 mutex_lock(&mdsc->mutex);
4237                 while ((req = __get_oldest_req(mdsc))) {
4238                         dout("wait_requests timed out on tid %llu\n",
4239                              req->r_tid);
4240                         list_del_init(&req->r_wait);
4241                         __unregister_request(mdsc, req);
4242                 }
4243         }
4244         mutex_unlock(&mdsc->mutex);
4245         dout("wait_requests done\n");
4246 }
4247
4248 /*
4249  * called before mount is ro, and before dentries are torn down.
4250  * (hmm, does this still race with new lookups?)
4251  */
4252 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4253 {
4254         dout("pre_umount\n");
4255         mdsc->stopping = 1;
4256
4257         lock_unlock_sessions(mdsc);
4258         ceph_flush_dirty_caps(mdsc);
4259         wait_requests(mdsc);
4260
4261         /*
4262          * wait for reply handlers to drop their request refs and
4263          * their inode/dcache refs
4264          */
4265         ceph_msgr_flush();
4266
4267         ceph_cleanup_quotarealms_inodes(mdsc);
4268 }
4269
4270 /*
4271  * wait for all write mds requests to flush.
4272  */
4273 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4274 {
4275         struct ceph_mds_request *req = NULL, *nextreq;
4276         struct rb_node *n;
4277
4278         mutex_lock(&mdsc->mutex);
4279         dout("wait_unsafe_requests want %lld\n", want_tid);
4280 restart:
4281         req = __get_oldest_req(mdsc);
4282         while (req && req->r_tid <= want_tid) {
4283                 /* find next request */
4284                 n = rb_next(&req->r_node);
4285                 if (n)
4286                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4287                 else
4288                         nextreq = NULL;
4289                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4290                     (req->r_op & CEPH_MDS_OP_WRITE)) {
4291                         /* write op */
4292                         ceph_mdsc_get_request(req);
4293                         if (nextreq)
4294                                 ceph_mdsc_get_request(nextreq);
4295                         mutex_unlock(&mdsc->mutex);
4296                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
4297                              req->r_tid, want_tid);
4298                         wait_for_completion(&req->r_safe_completion);
4299                         mutex_lock(&mdsc->mutex);
4300                         ceph_mdsc_put_request(req);
4301                         if (!nextreq)
4302                                 break;  /* next dne before, so we're done! */
4303                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
4304                                 /* next request was removed from tree */
4305                                 ceph_mdsc_put_request(nextreq);
4306                                 goto restart;
4307                         }
4308                         ceph_mdsc_put_request(nextreq);  /* won't go away */
4309                 }
4310                 req = nextreq;
4311         }
4312         mutex_unlock(&mdsc->mutex);
4313         dout("wait_unsafe_requests done\n");
4314 }
4315
4316 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4317 {
4318         u64 want_tid, want_flush;
4319
4320         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4321                 return;
4322
4323         dout("sync\n");
4324         mutex_lock(&mdsc->mutex);
4325         want_tid = mdsc->last_tid;
4326         mutex_unlock(&mdsc->mutex);
4327
4328         ceph_flush_dirty_caps(mdsc);
4329         spin_lock(&mdsc->cap_dirty_lock);
4330         want_flush = mdsc->last_cap_flush_tid;
4331         if (!list_empty(&mdsc->cap_flush_list)) {
4332                 struct ceph_cap_flush *cf =
4333                         list_last_entry(&mdsc->cap_flush_list,
4334                                         struct ceph_cap_flush, g_list);
4335                 cf->wake = true;
4336         }
4337         spin_unlock(&mdsc->cap_dirty_lock);
4338
4339         dout("sync want tid %lld flush_seq %lld\n",
4340              want_tid, want_flush);
4341
4342         wait_unsafe_requests(mdsc, want_tid);
4343         wait_caps_flush(mdsc, want_flush);
4344 }
4345
4346 /*
4347  * true if all sessions are closed, or we force unmount
4348  */
4349 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4350 {
4351         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4352                 return true;
4353         return atomic_read(&mdsc->num_sessions) <= skipped;
4354 }
4355
4356 /*
4357  * called after sb is ro.
4358  */
4359 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4360 {
4361         struct ceph_options *opts = mdsc->fsc->client->options;
4362         struct ceph_mds_session *session;
4363         int i;
4364         int skipped = 0;
4365
4366         dout("close_sessions\n");
4367
4368         /* close sessions */
4369         mutex_lock(&mdsc->mutex);
4370         for (i = 0; i < mdsc->max_sessions; i++) {
4371                 session = __ceph_lookup_mds_session(mdsc, i);
4372                 if (!session)
4373                         continue;
4374                 mutex_unlock(&mdsc->mutex);
4375                 mutex_lock(&session->s_mutex);
4376                 if (__close_session(mdsc, session) <= 0)
4377                         skipped++;
4378                 mutex_unlock(&session->s_mutex);
4379                 ceph_put_mds_session(session);
4380                 mutex_lock(&mdsc->mutex);
4381         }
4382         mutex_unlock(&mdsc->mutex);
4383
4384         dout("waiting for sessions to close\n");
4385         wait_event_timeout(mdsc->session_close_wq,
4386                            done_closing_sessions(mdsc, skipped),
4387                            ceph_timeout_jiffies(opts->mount_timeout));
4388
4389         /* tear down remaining sessions */
4390         mutex_lock(&mdsc->mutex);
4391         for (i = 0; i < mdsc->max_sessions; i++) {
4392                 if (mdsc->sessions[i]) {
4393                         session = get_session(mdsc->sessions[i]);
4394                         __unregister_session(mdsc, session);
4395                         mutex_unlock(&mdsc->mutex);
4396                         mutex_lock(&session->s_mutex);
4397                         remove_session_caps(session);
4398                         mutex_unlock(&session->s_mutex);
4399                         ceph_put_mds_session(session);
4400                         mutex_lock(&mdsc->mutex);
4401                 }
4402         }
4403         WARN_ON(!list_empty(&mdsc->cap_delay_list));
4404         mutex_unlock(&mdsc->mutex);
4405
4406         ceph_cleanup_snapid_map(mdsc);
4407         ceph_cleanup_empty_realms(mdsc);
4408
4409         cancel_work_sync(&mdsc->cap_reclaim_work);
4410         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4411
4412         dout("stopped\n");
4413 }
4414
4415 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4416 {
4417         struct ceph_mds_session *session;
4418         int mds;
4419
4420         dout("force umount\n");
4421
4422         mutex_lock(&mdsc->mutex);
4423         for (mds = 0; mds < mdsc->max_sessions; mds++) {
4424                 session = __ceph_lookup_mds_session(mdsc, mds);
4425                 if (!session)
4426                         continue;
4427
4428                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4429                         __unregister_session(mdsc, session);
4430                 __wake_requests(mdsc, &session->s_waiting);
4431                 mutex_unlock(&mdsc->mutex);
4432
4433                 mutex_lock(&session->s_mutex);
4434                 __close_session(mdsc, session);
4435                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4436                         cleanup_session_requests(mdsc, session);
4437                         remove_session_caps(session);
4438                 }
4439                 mutex_unlock(&session->s_mutex);
4440                 ceph_put_mds_session(session);
4441
4442                 mutex_lock(&mdsc->mutex);
4443                 kick_requests(mdsc, mds);
4444         }
4445         __wake_requests(mdsc, &mdsc->waiting_for_map);
4446         mutex_unlock(&mdsc->mutex);
4447 }
4448
4449 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4450 {
4451         dout("stop\n");
4452         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4453         if (mdsc->mdsmap)
4454                 ceph_mdsmap_destroy(mdsc->mdsmap);
4455         kfree(mdsc->sessions);
4456         ceph_caps_finalize(mdsc);
4457         ceph_pool_perm_destroy(mdsc);
4458 }
4459
4460 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4461 {
4462         struct ceph_mds_client *mdsc = fsc->mdsc;
4463         dout("mdsc_destroy %p\n", mdsc);
4464
4465         if (!mdsc)
4466                 return;
4467
4468         /* flush out any connection work with references to us */
4469         ceph_msgr_flush();
4470
4471         ceph_mdsc_stop(mdsc);
4472
4473         fsc->mdsc = NULL;
4474         kfree(mdsc);
4475         dout("mdsc_destroy %p done\n", mdsc);
4476 }
4477
4478 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4479 {
4480         struct ceph_fs_client *fsc = mdsc->fsc;
4481         const char *mds_namespace = fsc->mount_options->mds_namespace;
4482         void *p = msg->front.iov_base;
4483         void *end = p + msg->front.iov_len;
4484         u32 epoch;
4485         u32 map_len;
4486         u32 num_fs;
4487         u32 mount_fscid = (u32)-1;
4488         u8 struct_v, struct_cv;
4489         int err = -EINVAL;
4490
4491         ceph_decode_need(&p, end, sizeof(u32), bad);
4492         epoch = ceph_decode_32(&p);
4493
4494         dout("handle_fsmap epoch %u\n", epoch);
4495
4496         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4497         struct_v = ceph_decode_8(&p);
4498         struct_cv = ceph_decode_8(&p);
4499         map_len = ceph_decode_32(&p);
4500
4501         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4502         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4503
4504         num_fs = ceph_decode_32(&p);
4505         while (num_fs-- > 0) {
4506                 void *info_p, *info_end;
4507                 u32 info_len;
4508                 u8 info_v, info_cv;
4509                 u32 fscid, namelen;
4510
4511                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4512                 info_v = ceph_decode_8(&p);
4513                 info_cv = ceph_decode_8(&p);
4514                 info_len = ceph_decode_32(&p);
4515                 ceph_decode_need(&p, end, info_len, bad);
4516                 info_p = p;
4517                 info_end = p + info_len;
4518                 p = info_end;
4519
4520                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4521                 fscid = ceph_decode_32(&info_p);
4522                 namelen = ceph_decode_32(&info_p);
4523                 ceph_decode_need(&info_p, info_end, namelen, bad);
4524
4525                 if (mds_namespace &&
4526                     strlen(mds_namespace) == namelen &&
4527                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
4528                         mount_fscid = fscid;
4529                         break;
4530                 }
4531         }
4532
4533         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4534         if (mount_fscid != (u32)-1) {
4535                 fsc->client->monc.fs_cluster_id = mount_fscid;
4536                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4537                                    0, true);
4538                 ceph_monc_renew_subs(&fsc->client->monc);
4539         } else {
4540                 err = -ENOENT;
4541                 goto err_out;
4542         }
4543         return;
4544
4545 bad:
4546         pr_err("error decoding fsmap\n");
4547 err_out:
4548         mutex_lock(&mdsc->mutex);
4549         mdsc->mdsmap_err = err;
4550         __wake_requests(mdsc, &mdsc->waiting_for_map);
4551         mutex_unlock(&mdsc->mutex);
4552 }
4553
4554 /*
4555  * handle mds map update.
4556  */
4557 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4558 {
4559         u32 epoch;
4560         u32 maplen;
4561         void *p = msg->front.iov_base;
4562         void *end = p + msg->front.iov_len;
4563         struct ceph_mdsmap *newmap, *oldmap;
4564         struct ceph_fsid fsid;
4565         int err = -EINVAL;
4566
4567         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4568         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4569         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4570                 return;
4571         epoch = ceph_decode_32(&p);
4572         maplen = ceph_decode_32(&p);
4573         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4574
4575         /* do we need it? */
4576         mutex_lock(&mdsc->mutex);
4577         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4578                 dout("handle_map epoch %u <= our %u\n",
4579                      epoch, mdsc->mdsmap->m_epoch);
4580                 mutex_unlock(&mdsc->mutex);
4581                 return;
4582         }
4583
4584         newmap = ceph_mdsmap_decode(&p, end);
4585         if (IS_ERR(newmap)) {
4586                 err = PTR_ERR(newmap);
4587                 goto bad_unlock;
4588         }
4589
4590         /* swap into place */
4591         if (mdsc->mdsmap) {
4592                 oldmap = mdsc->mdsmap;
4593                 mdsc->mdsmap = newmap;
4594                 check_new_map(mdsc, newmap, oldmap);
4595                 ceph_mdsmap_destroy(oldmap);
4596         } else {
4597                 mdsc->mdsmap = newmap;  /* first mds map */
4598         }
4599         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4600                                         MAX_LFS_FILESIZE);
4601
4602         __wake_requests(mdsc, &mdsc->waiting_for_map);
4603         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4604                           mdsc->mdsmap->m_epoch);
4605
4606         mutex_unlock(&mdsc->mutex);
4607         schedule_delayed(mdsc);
4608         return;
4609
4610 bad_unlock:
4611         mutex_unlock(&mdsc->mutex);
4612 bad:
4613         pr_err("error decoding mdsmap %d\n", err);
4614         return;
4615 }
4616
4617 static struct ceph_connection *con_get(struct ceph_connection *con)
4618 {
4619         struct ceph_mds_session *s = con->private;
4620
4621         if (get_session(s))
4622                 return con;
4623         return NULL;
4624 }
4625
4626 static void con_put(struct ceph_connection *con)
4627 {
4628         struct ceph_mds_session *s = con->private;
4629
4630         ceph_put_mds_session(s);
4631 }
4632
4633 /*
4634  * if the client is unresponsive for long enough, the mds will kill
4635  * the session entirely.
4636  */
4637 static void peer_reset(struct ceph_connection *con)
4638 {
4639         struct ceph_mds_session *s = con->private;
4640         struct ceph_mds_client *mdsc = s->s_mdsc;
4641
4642         pr_warn("mds%d closed our session\n", s->s_mds);
4643         send_mds_reconnect(mdsc, s);
4644 }
4645
4646 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4647 {
4648         struct ceph_mds_session *s = con->private;
4649         struct ceph_mds_client *mdsc = s->s_mdsc;
4650         int type = le16_to_cpu(msg->hdr.type);
4651
4652         mutex_lock(&mdsc->mutex);
4653         if (__verify_registered_session(mdsc, s) < 0) {
4654                 mutex_unlock(&mdsc->mutex);
4655                 goto out;
4656         }
4657         mutex_unlock(&mdsc->mutex);
4658
4659         switch (type) {
4660         case CEPH_MSG_MDS_MAP:
4661                 ceph_mdsc_handle_mdsmap(mdsc, msg);
4662                 break;
4663         case CEPH_MSG_FS_MAP_USER:
4664                 ceph_mdsc_handle_fsmap(mdsc, msg);
4665                 break;
4666         case CEPH_MSG_CLIENT_SESSION:
4667                 handle_session(s, msg);
4668                 break;
4669         case CEPH_MSG_CLIENT_REPLY:
4670                 handle_reply(s, msg);
4671                 break;
4672         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4673                 handle_forward(mdsc, s, msg);
4674                 break;
4675         case CEPH_MSG_CLIENT_CAPS:
4676                 ceph_handle_caps(s, msg);
4677                 break;
4678         case CEPH_MSG_CLIENT_SNAP:
4679                 ceph_handle_snap(mdsc, s, msg);
4680                 break;
4681         case CEPH_MSG_CLIENT_LEASE:
4682                 handle_lease(mdsc, s, msg);
4683                 break;
4684         case CEPH_MSG_CLIENT_QUOTA:
4685                 ceph_handle_quota(mdsc, s, msg);
4686                 break;
4687
4688         default:
4689                 pr_err("received unknown message type %d %s\n", type,
4690                        ceph_msg_type_name(type));
4691         }
4692 out:
4693         ceph_msg_put(msg);
4694 }
4695
4696 /*
4697  * authentication
4698  */
4699
4700 /*
4701  * Note: returned pointer is the address of a structure that's
4702  * managed separately.  Caller must *not* attempt to free it.
4703  */
4704 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4705                                         int *proto, int force_new)
4706 {
4707         struct ceph_mds_session *s = con->private;
4708         struct ceph_mds_client *mdsc = s->s_mdsc;
4709         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4710         struct ceph_auth_handshake *auth = &s->s_auth;
4711
4712         if (force_new && auth->authorizer) {
4713                 ceph_auth_destroy_authorizer(auth->authorizer);
4714                 auth->authorizer = NULL;
4715         }
4716         if (!auth->authorizer) {
4717                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4718                                                       auth);
4719                 if (ret)
4720                         return ERR_PTR(ret);
4721         } else {
4722                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4723                                                       auth);
4724                 if (ret)
4725                         return ERR_PTR(ret);
4726         }
4727         *proto = ac->protocol;
4728
4729         return auth;
4730 }
4731
4732 static int add_authorizer_challenge(struct ceph_connection *con,
4733                                     void *challenge_buf, int challenge_buf_len)
4734 {
4735         struct ceph_mds_session *s = con->private;
4736         struct ceph_mds_client *mdsc = s->s_mdsc;
4737         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4738
4739         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4740                                             challenge_buf, challenge_buf_len);
4741 }
4742
4743 static int verify_authorizer_reply(struct ceph_connection *con)
4744 {
4745         struct ceph_mds_session *s = con->private;
4746         struct ceph_mds_client *mdsc = s->s_mdsc;
4747         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4748
4749         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4750 }
4751
4752 static int invalidate_authorizer(struct ceph_connection *con)
4753 {
4754         struct ceph_mds_session *s = con->private;
4755         struct ceph_mds_client *mdsc = s->s_mdsc;
4756         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4757
4758         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4759
4760         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4761 }
4762
4763 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4764                                 struct ceph_msg_header *hdr, int *skip)
4765 {
4766         struct ceph_msg *msg;
4767         int type = (int) le16_to_cpu(hdr->type);
4768         int front_len = (int) le32_to_cpu(hdr->front_len);
4769
4770         if (con->in_msg)
4771                 return con->in_msg;
4772
4773         *skip = 0;
4774         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4775         if (!msg) {
4776                 pr_err("unable to allocate msg type %d len %d\n",
4777                        type, front_len);
4778                 return NULL;
4779         }
4780
4781         return msg;
4782 }
4783
4784 static int mds_sign_message(struct ceph_msg *msg)
4785 {
4786        struct ceph_mds_session *s = msg->con->private;
4787        struct ceph_auth_handshake *auth = &s->s_auth;
4788
4789        return ceph_auth_sign_message(auth, msg);
4790 }
4791
4792 static int mds_check_message_signature(struct ceph_msg *msg)
4793 {
4794        struct ceph_mds_session *s = msg->con->private;
4795        struct ceph_auth_handshake *auth = &s->s_auth;
4796
4797        return ceph_auth_check_message_signature(auth, msg);
4798 }
4799
4800 static const struct ceph_connection_operations mds_con_ops = {
4801         .get = con_get,
4802         .put = con_put,
4803         .dispatch = dispatch,
4804         .get_authorizer = get_authorizer,
4805         .add_authorizer_challenge = add_authorizer_challenge,
4806         .verify_authorizer_reply = verify_authorizer_reply,
4807         .invalidate_authorizer = invalidate_authorizer,
4808         .peer_reset = peer_reset,
4809         .alloc_msg = mds_alloc_msg,
4810         .sign_message = mds_sign_message,
4811         .check_message_signature = mds_check_message_signature,
4812 };
4813
4814 /* eof */