]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - fs/ceph/mds_client.c
ceph: drop cap releases in requests composed before cap reconnect
[karo-tx-linux.git] / fs / ceph / mds_client.c
1 #include <linux/ceph/ceph_debug.h>
2
3 #include <linux/fs.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
6 #include <linux/gfp.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11
12 #include "super.h"
13 #include "mds_client.h"
14
15 #include <linux/ceph/ceph_features.h>
16 #include <linux/ceph/messenger.h>
17 #include <linux/ceph/decode.h>
18 #include <linux/ceph/pagelist.h>
19 #include <linux/ceph/auth.h>
20 #include <linux/ceph/debugfs.h>
21
22 /*
23  * A cluster of MDS (metadata server) daemons is responsible for
24  * managing the file system namespace (the directory hierarchy and
25  * inodes) and for coordinating shared access to storage.  Metadata is
26  * partitioning hierarchically across a number of servers, and that
27  * partition varies over time as the cluster adjusts the distribution
28  * in order to balance load.
29  *
30  * The MDS client is primarily responsible to managing synchronous
31  * metadata requests for operations like open, unlink, and so forth.
32  * If there is a MDS failure, we find out about it when we (possibly
33  * request and) receive a new MDS map, and can resubmit affected
34  * requests.
35  *
36  * For the most part, though, we take advantage of a lossless
37  * communications channel to the MDS, and do not need to worry about
38  * timing out or resubmitting requests.
39  *
40  * We maintain a stateful "session" with each MDS we interact with.
41  * Within each session, we sent periodic heartbeat messages to ensure
42  * any capabilities or leases we have been issues remain valid.  If
43  * the session times out and goes stale, our leases and capabilities
44  * are no longer valid.
45  */
46
47 struct ceph_reconnect_state {
48         int nr_caps;
49         struct ceph_pagelist *pagelist;
50         bool flock;
51 };
52
53 static void __wake_requests(struct ceph_mds_client *mdsc,
54                             struct list_head *head);
55
56 static const struct ceph_connection_operations mds_con_ops;
57
58
59 /*
60  * mds reply parsing
61  */
62
63 /*
64  * parse individual inode info
65  */
66 static int parse_reply_info_in(void **p, void *end,
67                                struct ceph_mds_reply_info_in *info,
68                                u64 features)
69 {
70         int err = -EIO;
71
72         info->in = *p;
73         *p += sizeof(struct ceph_mds_reply_inode) +
74                 sizeof(*info->in->fragtree.splits) *
75                 le32_to_cpu(info->in->fragtree.nsplits);
76
77         ceph_decode_32_safe(p, end, info->symlink_len, bad);
78         ceph_decode_need(p, end, info->symlink_len, bad);
79         info->symlink = *p;
80         *p += info->symlink_len;
81
82         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
83                 ceph_decode_copy_safe(p, end, &info->dir_layout,
84                                       sizeof(info->dir_layout), bad);
85         else
86                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
87
88         ceph_decode_32_safe(p, end, info->xattr_len, bad);
89         ceph_decode_need(p, end, info->xattr_len, bad);
90         info->xattr_data = *p;
91         *p += info->xattr_len;
92
93         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
94                 ceph_decode_64_safe(p, end, info->inline_version, bad);
95                 ceph_decode_32_safe(p, end, info->inline_len, bad);
96                 ceph_decode_need(p, end, info->inline_len, bad);
97                 info->inline_data = *p;
98                 *p += info->inline_len;
99         } else
100                 info->inline_version = CEPH_INLINE_NONE;
101
102         return 0;
103 bad:
104         return err;
105 }
106
107 /*
108  * parse a normal reply, which may contain a (dir+)dentry and/or a
109  * target inode.
110  */
111 static int parse_reply_info_trace(void **p, void *end,
112                                   struct ceph_mds_reply_info_parsed *info,
113                                   u64 features)
114 {
115         int err;
116
117         if (info->head->is_dentry) {
118                 err = parse_reply_info_in(p, end, &info->diri, features);
119                 if (err < 0)
120                         goto out_bad;
121
122                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
123                         goto bad;
124                 info->dirfrag = *p;
125                 *p += sizeof(*info->dirfrag) +
126                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
127                 if (unlikely(*p > end))
128                         goto bad;
129
130                 ceph_decode_32_safe(p, end, info->dname_len, bad);
131                 ceph_decode_need(p, end, info->dname_len, bad);
132                 info->dname = *p;
133                 *p += info->dname_len;
134                 info->dlease = *p;
135                 *p += sizeof(*info->dlease);
136         }
137
138         if (info->head->is_target) {
139                 err = parse_reply_info_in(p, end, &info->targeti, features);
140                 if (err < 0)
141                         goto out_bad;
142         }
143
144         if (unlikely(*p != end))
145                 goto bad;
146         return 0;
147
148 bad:
149         err = -EIO;
150 out_bad:
151         pr_err("problem parsing mds trace %d\n", err);
152         return err;
153 }
154
155 /*
156  * parse readdir results
157  */
158 static int parse_reply_info_dir(void **p, void *end,
159                                 struct ceph_mds_reply_info_parsed *info,
160                                 u64 features)
161 {
162         u32 num, i = 0;
163         int err;
164
165         info->dir_dir = *p;
166         if (*p + sizeof(*info->dir_dir) > end)
167                 goto bad;
168         *p += sizeof(*info->dir_dir) +
169                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
170         if (*p > end)
171                 goto bad;
172
173         ceph_decode_need(p, end, sizeof(num) + 2, bad);
174         num = ceph_decode_32(p);
175         info->dir_end = ceph_decode_8(p);
176         info->dir_complete = ceph_decode_8(p);
177         if (num == 0)
178                 goto done;
179
180         BUG_ON(!info->dir_in);
181         info->dir_dname = (void *)(info->dir_in + num);
182         info->dir_dname_len = (void *)(info->dir_dname + num);
183         info->dir_dlease = (void *)(info->dir_dname_len + num);
184         if ((unsigned long)(info->dir_dlease + num) >
185             (unsigned long)info->dir_in + info->dir_buf_size) {
186                 pr_err("dir contents are larger than expected\n");
187                 WARN_ON(1);
188                 goto bad;
189         }
190
191         info->dir_nr = num;
192         while (num) {
193                 /* dentry */
194                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
195                 info->dir_dname_len[i] = ceph_decode_32(p);
196                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
197                 info->dir_dname[i] = *p;
198                 *p += info->dir_dname_len[i];
199                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
200                      info->dir_dname[i]);
201                 info->dir_dlease[i] = *p;
202                 *p += sizeof(struct ceph_mds_reply_lease);
203
204                 /* inode */
205                 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
206                 if (err < 0)
207                         goto out_bad;
208                 i++;
209                 num--;
210         }
211
212 done:
213         if (*p != end)
214                 goto bad;
215         return 0;
216
217 bad:
218         err = -EIO;
219 out_bad:
220         pr_err("problem parsing dir contents %d\n", err);
221         return err;
222 }
223
224 /*
225  * parse fcntl F_GETLK results
226  */
227 static int parse_reply_info_filelock(void **p, void *end,
228                                      struct ceph_mds_reply_info_parsed *info,
229                                      u64 features)
230 {
231         if (*p + sizeof(*info->filelock_reply) > end)
232                 goto bad;
233
234         info->filelock_reply = *p;
235         *p += sizeof(*info->filelock_reply);
236
237         if (unlikely(*p != end))
238                 goto bad;
239         return 0;
240
241 bad:
242         return -EIO;
243 }
244
245 /*
246  * parse create results
247  */
248 static int parse_reply_info_create(void **p, void *end,
249                                   struct ceph_mds_reply_info_parsed *info,
250                                   u64 features)
251 {
252         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
253                 if (*p == end) {
254                         info->has_create_ino = false;
255                 } else {
256                         info->has_create_ino = true;
257                         info->ino = ceph_decode_64(p);
258                 }
259         }
260
261         if (unlikely(*p != end))
262                 goto bad;
263         return 0;
264
265 bad:
266         return -EIO;
267 }
268
269 /*
270  * parse extra results
271  */
272 static int parse_reply_info_extra(void **p, void *end,
273                                   struct ceph_mds_reply_info_parsed *info,
274                                   u64 features)
275 {
276         if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
277                 return parse_reply_info_filelock(p, end, info, features);
278         else if (info->head->op == CEPH_MDS_OP_READDIR ||
279                  info->head->op == CEPH_MDS_OP_LSSNAP)
280                 return parse_reply_info_dir(p, end, info, features);
281         else if (info->head->op == CEPH_MDS_OP_CREATE)
282                 return parse_reply_info_create(p, end, info, features);
283         else
284                 return -EIO;
285 }
286
287 /*
288  * parse entire mds reply
289  */
290 static int parse_reply_info(struct ceph_msg *msg,
291                             struct ceph_mds_reply_info_parsed *info,
292                             u64 features)
293 {
294         void *p, *end;
295         u32 len;
296         int err;
297
298         info->head = msg->front.iov_base;
299         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
300         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
301
302         /* trace */
303         ceph_decode_32_safe(&p, end, len, bad);
304         if (len > 0) {
305                 ceph_decode_need(&p, end, len, bad);
306                 err = parse_reply_info_trace(&p, p+len, info, features);
307                 if (err < 0)
308                         goto out_bad;
309         }
310
311         /* extra */
312         ceph_decode_32_safe(&p, end, len, bad);
313         if (len > 0) {
314                 ceph_decode_need(&p, end, len, bad);
315                 err = parse_reply_info_extra(&p, p+len, info, features);
316                 if (err < 0)
317                         goto out_bad;
318         }
319
320         /* snap blob */
321         ceph_decode_32_safe(&p, end, len, bad);
322         info->snapblob_len = len;
323         info->snapblob = p;
324         p += len;
325
326         if (p != end)
327                 goto bad;
328         return 0;
329
330 bad:
331         err = -EIO;
332 out_bad:
333         pr_err("mds parse_reply err %d\n", err);
334         return err;
335 }
336
337 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
338 {
339         if (!info->dir_in)
340                 return;
341         free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
342 }
343
344
345 /*
346  * sessions
347  */
348 const char *ceph_session_state_name(int s)
349 {
350         switch (s) {
351         case CEPH_MDS_SESSION_NEW: return "new";
352         case CEPH_MDS_SESSION_OPENING: return "opening";
353         case CEPH_MDS_SESSION_OPEN: return "open";
354         case CEPH_MDS_SESSION_HUNG: return "hung";
355         case CEPH_MDS_SESSION_CLOSING: return "closing";
356         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
357         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
358         default: return "???";
359         }
360 }
361
362 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
363 {
364         if (atomic_inc_not_zero(&s->s_ref)) {
365                 dout("mdsc get_session %p %d -> %d\n", s,
366                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
367                 return s;
368         } else {
369                 dout("mdsc get_session %p 0 -- FAIL", s);
370                 return NULL;
371         }
372 }
373
374 void ceph_put_mds_session(struct ceph_mds_session *s)
375 {
376         dout("mdsc put_session %p %d -> %d\n", s,
377              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
378         if (atomic_dec_and_test(&s->s_ref)) {
379                 if (s->s_auth.authorizer)
380                         ceph_auth_destroy_authorizer(
381                                 s->s_mdsc->fsc->client->monc.auth,
382                                 s->s_auth.authorizer);
383                 kfree(s);
384         }
385 }
386
387 /*
388  * called under mdsc->mutex
389  */
390 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
391                                                    int mds)
392 {
393         struct ceph_mds_session *session;
394
395         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
396                 return NULL;
397         session = mdsc->sessions[mds];
398         dout("lookup_mds_session %p %d\n", session,
399              atomic_read(&session->s_ref));
400         get_session(session);
401         return session;
402 }
403
404 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
405 {
406         if (mds >= mdsc->max_sessions)
407                 return false;
408         return mdsc->sessions[mds];
409 }
410
411 static int __verify_registered_session(struct ceph_mds_client *mdsc,
412                                        struct ceph_mds_session *s)
413 {
414         if (s->s_mds >= mdsc->max_sessions ||
415             mdsc->sessions[s->s_mds] != s)
416                 return -ENOENT;
417         return 0;
418 }
419
420 /*
421  * create+register a new session for given mds.
422  * called under mdsc->mutex.
423  */
424 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
425                                                  int mds)
426 {
427         struct ceph_mds_session *s;
428
429         if (mds >= mdsc->mdsmap->m_max_mds)
430                 return ERR_PTR(-EINVAL);
431
432         s = kzalloc(sizeof(*s), GFP_NOFS);
433         if (!s)
434                 return ERR_PTR(-ENOMEM);
435         s->s_mdsc = mdsc;
436         s->s_mds = mds;
437         s->s_state = CEPH_MDS_SESSION_NEW;
438         s->s_ttl = 0;
439         s->s_seq = 0;
440         mutex_init(&s->s_mutex);
441
442         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
443
444         spin_lock_init(&s->s_gen_ttl_lock);
445         s->s_cap_gen = 0;
446         s->s_cap_ttl = jiffies - 1;
447
448         spin_lock_init(&s->s_cap_lock);
449         s->s_renew_requested = 0;
450         s->s_renew_seq = 0;
451         INIT_LIST_HEAD(&s->s_caps);
452         s->s_nr_caps = 0;
453         s->s_trim_caps = 0;
454         atomic_set(&s->s_ref, 1);
455         INIT_LIST_HEAD(&s->s_waiting);
456         INIT_LIST_HEAD(&s->s_unsafe);
457         s->s_num_cap_releases = 0;
458         s->s_cap_reconnect = 0;
459         s->s_cap_iterator = NULL;
460         INIT_LIST_HEAD(&s->s_cap_releases);
461         INIT_LIST_HEAD(&s->s_cap_releases_done);
462         INIT_LIST_HEAD(&s->s_cap_flushing);
463         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
464
465         dout("register_session mds%d\n", mds);
466         if (mds >= mdsc->max_sessions) {
467                 int newmax = 1 << get_count_order(mds+1);
468                 struct ceph_mds_session **sa;
469
470                 dout("register_session realloc to %d\n", newmax);
471                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
472                 if (sa == NULL)
473                         goto fail_realloc;
474                 if (mdsc->sessions) {
475                         memcpy(sa, mdsc->sessions,
476                                mdsc->max_sessions * sizeof(void *));
477                         kfree(mdsc->sessions);
478                 }
479                 mdsc->sessions = sa;
480                 mdsc->max_sessions = newmax;
481         }
482         mdsc->sessions[mds] = s;
483         atomic_inc(&mdsc->num_sessions);
484         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
485
486         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
487                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
488
489         return s;
490
491 fail_realloc:
492         kfree(s);
493         return ERR_PTR(-ENOMEM);
494 }
495
496 /*
497  * called under mdsc->mutex
498  */
499 static void __unregister_session(struct ceph_mds_client *mdsc,
500                                struct ceph_mds_session *s)
501 {
502         dout("__unregister_session mds%d %p\n", s->s_mds, s);
503         BUG_ON(mdsc->sessions[s->s_mds] != s);
504         mdsc->sessions[s->s_mds] = NULL;
505         ceph_con_close(&s->s_con);
506         ceph_put_mds_session(s);
507         atomic_dec(&mdsc->num_sessions);
508 }
509
510 /*
511  * drop session refs in request.
512  *
513  * should be last request ref, or hold mdsc->mutex
514  */
515 static void put_request_session(struct ceph_mds_request *req)
516 {
517         if (req->r_session) {
518                 ceph_put_mds_session(req->r_session);
519                 req->r_session = NULL;
520         }
521 }
522
523 void ceph_mdsc_release_request(struct kref *kref)
524 {
525         struct ceph_mds_request *req = container_of(kref,
526                                                     struct ceph_mds_request,
527                                                     r_kref);
528         destroy_reply_info(&req->r_reply_info);
529         if (req->r_request)
530                 ceph_msg_put(req->r_request);
531         if (req->r_reply)
532                 ceph_msg_put(req->r_reply);
533         if (req->r_inode) {
534                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
535                 iput(req->r_inode);
536         }
537         if (req->r_locked_dir)
538                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
539         iput(req->r_target_inode);
540         if (req->r_dentry)
541                 dput(req->r_dentry);
542         if (req->r_old_dentry)
543                 dput(req->r_old_dentry);
544         if (req->r_old_dentry_dir) {
545                 /*
546                  * track (and drop pins for) r_old_dentry_dir
547                  * separately, since r_old_dentry's d_parent may have
548                  * changed between the dir mutex being dropped and
549                  * this request being freed.
550                  */
551                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
552                                   CEPH_CAP_PIN);
553                 iput(req->r_old_dentry_dir);
554         }
555         kfree(req->r_path1);
556         kfree(req->r_path2);
557         if (req->r_pagelist)
558                 ceph_pagelist_release(req->r_pagelist);
559         put_request_session(req);
560         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
561         kfree(req);
562 }
563
564 /*
565  * lookup session, bump ref if found.
566  *
567  * called under mdsc->mutex.
568  */
569 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
570                                              u64 tid)
571 {
572         struct ceph_mds_request *req;
573         struct rb_node *n = mdsc->request_tree.rb_node;
574
575         while (n) {
576                 req = rb_entry(n, struct ceph_mds_request, r_node);
577                 if (tid < req->r_tid)
578                         n = n->rb_left;
579                 else if (tid > req->r_tid)
580                         n = n->rb_right;
581                 else {
582                         ceph_mdsc_get_request(req);
583                         return req;
584                 }
585         }
586         return NULL;
587 }
588
589 static void __insert_request(struct ceph_mds_client *mdsc,
590                              struct ceph_mds_request *new)
591 {
592         struct rb_node **p = &mdsc->request_tree.rb_node;
593         struct rb_node *parent = NULL;
594         struct ceph_mds_request *req = NULL;
595
596         while (*p) {
597                 parent = *p;
598                 req = rb_entry(parent, struct ceph_mds_request, r_node);
599                 if (new->r_tid < req->r_tid)
600                         p = &(*p)->rb_left;
601                 else if (new->r_tid > req->r_tid)
602                         p = &(*p)->rb_right;
603                 else
604                         BUG();
605         }
606
607         rb_link_node(&new->r_node, parent, p);
608         rb_insert_color(&new->r_node, &mdsc->request_tree);
609 }
610
611 /*
612  * Register an in-flight request, and assign a tid.  Link to directory
613  * are modifying (if any).
614  *
615  * Called under mdsc->mutex.
616  */
617 static void __register_request(struct ceph_mds_client *mdsc,
618                                struct ceph_mds_request *req,
619                                struct inode *dir)
620 {
621         req->r_tid = ++mdsc->last_tid;
622         if (req->r_num_caps)
623                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
624                                   req->r_num_caps);
625         dout("__register_request %p tid %lld\n", req, req->r_tid);
626         ceph_mdsc_get_request(req);
627         __insert_request(mdsc, req);
628
629         req->r_uid = current_fsuid();
630         req->r_gid = current_fsgid();
631
632         if (dir) {
633                 struct ceph_inode_info *ci = ceph_inode(dir);
634
635                 ihold(dir);
636                 spin_lock(&ci->i_unsafe_lock);
637                 req->r_unsafe_dir = dir;
638                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
639                 spin_unlock(&ci->i_unsafe_lock);
640         }
641 }
642
643 static void __unregister_request(struct ceph_mds_client *mdsc,
644                                  struct ceph_mds_request *req)
645 {
646         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
647         rb_erase(&req->r_node, &mdsc->request_tree);
648         RB_CLEAR_NODE(&req->r_node);
649
650         if (req->r_unsafe_dir) {
651                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
652
653                 spin_lock(&ci->i_unsafe_lock);
654                 list_del_init(&req->r_unsafe_dir_item);
655                 spin_unlock(&ci->i_unsafe_lock);
656
657                 iput(req->r_unsafe_dir);
658                 req->r_unsafe_dir = NULL;
659         }
660
661         complete_all(&req->r_safe_completion);
662
663         ceph_mdsc_put_request(req);
664 }
665
666 /*
667  * Choose mds to send request to next.  If there is a hint set in the
668  * request (e.g., due to a prior forward hint from the mds), use that.
669  * Otherwise, consult frag tree and/or caps to identify the
670  * appropriate mds.  If all else fails, choose randomly.
671  *
672  * Called under mdsc->mutex.
673  */
674 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
675 {
676         /*
677          * we don't need to worry about protecting the d_parent access
678          * here because we never renaming inside the snapped namespace
679          * except to resplice to another snapdir, and either the old or new
680          * result is a valid result.
681          */
682         while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
683                 dentry = dentry->d_parent;
684         return dentry;
685 }
686
687 static int __choose_mds(struct ceph_mds_client *mdsc,
688                         struct ceph_mds_request *req)
689 {
690         struct inode *inode;
691         struct ceph_inode_info *ci;
692         struct ceph_cap *cap;
693         int mode = req->r_direct_mode;
694         int mds = -1;
695         u32 hash = req->r_direct_hash;
696         bool is_hash = req->r_direct_is_hash;
697
698         /*
699          * is there a specific mds we should try?  ignore hint if we have
700          * no session and the mds is not up (active or recovering).
701          */
702         if (req->r_resend_mds >= 0 &&
703             (__have_session(mdsc, req->r_resend_mds) ||
704              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
705                 dout("choose_mds using resend_mds mds%d\n",
706                      req->r_resend_mds);
707                 return req->r_resend_mds;
708         }
709
710         if (mode == USE_RANDOM_MDS)
711                 goto random;
712
713         inode = NULL;
714         if (req->r_inode) {
715                 inode = req->r_inode;
716         } else if (req->r_dentry) {
717                 /* ignore race with rename; old or new d_parent is okay */
718                 struct dentry *parent = req->r_dentry->d_parent;
719                 struct inode *dir = parent->d_inode;
720
721                 if (dir->i_sb != mdsc->fsc->sb) {
722                         /* not this fs! */
723                         inode = req->r_dentry->d_inode;
724                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
725                         /* direct snapped/virtual snapdir requests
726                          * based on parent dir inode */
727                         struct dentry *dn = get_nonsnap_parent(parent);
728                         inode = dn->d_inode;
729                         dout("__choose_mds using nonsnap parent %p\n", inode);
730                 } else {
731                         /* dentry target */
732                         inode = req->r_dentry->d_inode;
733                         if (!inode || mode == USE_AUTH_MDS) {
734                                 /* dir + name */
735                                 inode = dir;
736                                 hash = ceph_dentry_hash(dir, req->r_dentry);
737                                 is_hash = true;
738                         }
739                 }
740         }
741
742         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
743              (int)hash, mode);
744         if (!inode)
745                 goto random;
746         ci = ceph_inode(inode);
747
748         if (is_hash && S_ISDIR(inode->i_mode)) {
749                 struct ceph_inode_frag frag;
750                 int found;
751
752                 ceph_choose_frag(ci, hash, &frag, &found);
753                 if (found) {
754                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
755                                 u8 r;
756
757                                 /* choose a random replica */
758                                 get_random_bytes(&r, 1);
759                                 r %= frag.ndist;
760                                 mds = frag.dist[r];
761                                 dout("choose_mds %p %llx.%llx "
762                                      "frag %u mds%d (%d/%d)\n",
763                                      inode, ceph_vinop(inode),
764                                      frag.frag, mds,
765                                      (int)r, frag.ndist);
766                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
767                                     CEPH_MDS_STATE_ACTIVE)
768                                         return mds;
769                         }
770
771                         /* since this file/dir wasn't known to be
772                          * replicated, then we want to look for the
773                          * authoritative mds. */
774                         mode = USE_AUTH_MDS;
775                         if (frag.mds >= 0) {
776                                 /* choose auth mds */
777                                 mds = frag.mds;
778                                 dout("choose_mds %p %llx.%llx "
779                                      "frag %u mds%d (auth)\n",
780                                      inode, ceph_vinop(inode), frag.frag, mds);
781                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
782                                     CEPH_MDS_STATE_ACTIVE)
783                                         return mds;
784                         }
785                 }
786         }
787
788         spin_lock(&ci->i_ceph_lock);
789         cap = NULL;
790         if (mode == USE_AUTH_MDS)
791                 cap = ci->i_auth_cap;
792         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
793                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
794         if (!cap) {
795                 spin_unlock(&ci->i_ceph_lock);
796                 goto random;
797         }
798         mds = cap->session->s_mds;
799         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
800              inode, ceph_vinop(inode), mds,
801              cap == ci->i_auth_cap ? "auth " : "", cap);
802         spin_unlock(&ci->i_ceph_lock);
803         return mds;
804
805 random:
806         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
807         dout("choose_mds chose random mds%d\n", mds);
808         return mds;
809 }
810
811
812 /*
813  * session messages
814  */
815 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
816 {
817         struct ceph_msg *msg;
818         struct ceph_mds_session_head *h;
819
820         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
821                            false);
822         if (!msg) {
823                 pr_err("create_session_msg ENOMEM creating msg\n");
824                 return NULL;
825         }
826         h = msg->front.iov_base;
827         h->op = cpu_to_le32(op);
828         h->seq = cpu_to_le64(seq);
829
830         return msg;
831 }
832
833 /*
834  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
835  * to include additional client metadata fields.
836  */
837 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
838 {
839         struct ceph_msg *msg;
840         struct ceph_mds_session_head *h;
841         int i = -1;
842         int metadata_bytes = 0;
843         int metadata_key_count = 0;
844         struct ceph_options *opt = mdsc->fsc->client->options;
845         void *p;
846
847         const char* metadata[][2] = {
848                 {"hostname", utsname()->nodename},
849                 {"kernel_version", utsname()->release},
850                 {"entity_id", opt->name ? opt->name : ""},
851                 {NULL, NULL}
852         };
853
854         /* Calculate serialized length of metadata */
855         metadata_bytes = 4;  /* map length */
856         for (i = 0; metadata[i][0] != NULL; ++i) {
857                 metadata_bytes += 8 + strlen(metadata[i][0]) +
858                         strlen(metadata[i][1]);
859                 metadata_key_count++;
860         }
861
862         /* Allocate the message */
863         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
864                            GFP_NOFS, false);
865         if (!msg) {
866                 pr_err("create_session_msg ENOMEM creating msg\n");
867                 return NULL;
868         }
869         h = msg->front.iov_base;
870         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
871         h->seq = cpu_to_le64(seq);
872
873         /*
874          * Serialize client metadata into waiting buffer space, using
875          * the format that userspace expects for map<string, string>
876          *
877          * ClientSession messages with metadata are v2
878          */
879         msg->hdr.version = cpu_to_le16(2);
880         msg->hdr.compat_version = cpu_to_le16(1);
881
882         /* The write pointer, following the session_head structure */
883         p = msg->front.iov_base + sizeof(*h);
884
885         /* Number of entries in the map */
886         ceph_encode_32(&p, metadata_key_count);
887
888         /* Two length-prefixed strings for each entry in the map */
889         for (i = 0; metadata[i][0] != NULL; ++i) {
890                 size_t const key_len = strlen(metadata[i][0]);
891                 size_t const val_len = strlen(metadata[i][1]);
892
893                 ceph_encode_32(&p, key_len);
894                 memcpy(p, metadata[i][0], key_len);
895                 p += key_len;
896                 ceph_encode_32(&p, val_len);
897                 memcpy(p, metadata[i][1], val_len);
898                 p += val_len;
899         }
900
901         return msg;
902 }
903
904 /*
905  * send session open request.
906  *
907  * called under mdsc->mutex
908  */
909 static int __open_session(struct ceph_mds_client *mdsc,
910                           struct ceph_mds_session *session)
911 {
912         struct ceph_msg *msg;
913         int mstate;
914         int mds = session->s_mds;
915
916         /* wait for mds to go active? */
917         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
918         dout("open_session to mds%d (%s)\n", mds,
919              ceph_mds_state_name(mstate));
920         session->s_state = CEPH_MDS_SESSION_OPENING;
921         session->s_renew_requested = jiffies;
922
923         /* send connect message */
924         msg = create_session_open_msg(mdsc, session->s_seq);
925         if (!msg)
926                 return -ENOMEM;
927         ceph_con_send(&session->s_con, msg);
928         return 0;
929 }
930
931 /*
932  * open sessions for any export targets for the given mds
933  *
934  * called under mdsc->mutex
935  */
936 static struct ceph_mds_session *
937 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
938 {
939         struct ceph_mds_session *session;
940
941         session = __ceph_lookup_mds_session(mdsc, target);
942         if (!session) {
943                 session = register_session(mdsc, target);
944                 if (IS_ERR(session))
945                         return session;
946         }
947         if (session->s_state == CEPH_MDS_SESSION_NEW ||
948             session->s_state == CEPH_MDS_SESSION_CLOSING)
949                 __open_session(mdsc, session);
950
951         return session;
952 }
953
954 struct ceph_mds_session *
955 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
956 {
957         struct ceph_mds_session *session;
958
959         dout("open_export_target_session to mds%d\n", target);
960
961         mutex_lock(&mdsc->mutex);
962         session = __open_export_target_session(mdsc, target);
963         mutex_unlock(&mdsc->mutex);
964
965         return session;
966 }
967
968 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
969                                           struct ceph_mds_session *session)
970 {
971         struct ceph_mds_info *mi;
972         struct ceph_mds_session *ts;
973         int i, mds = session->s_mds;
974
975         if (mds >= mdsc->mdsmap->m_max_mds)
976                 return;
977
978         mi = &mdsc->mdsmap->m_info[mds];
979         dout("open_export_target_sessions for mds%d (%d targets)\n",
980              session->s_mds, mi->num_export_targets);
981
982         for (i = 0; i < mi->num_export_targets; i++) {
983                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
984                 if (!IS_ERR(ts))
985                         ceph_put_mds_session(ts);
986         }
987 }
988
989 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
990                                            struct ceph_mds_session *session)
991 {
992         mutex_lock(&mdsc->mutex);
993         __open_export_target_sessions(mdsc, session);
994         mutex_unlock(&mdsc->mutex);
995 }
996
997 /*
998  * session caps
999  */
1000
1001 /*
1002  * Free preallocated cap messages assigned to this session
1003  */
1004 static void cleanup_cap_releases(struct ceph_mds_session *session)
1005 {
1006         struct ceph_msg *msg;
1007
1008         spin_lock(&session->s_cap_lock);
1009         while (!list_empty(&session->s_cap_releases)) {
1010                 msg = list_first_entry(&session->s_cap_releases,
1011                                        struct ceph_msg, list_head);
1012                 list_del_init(&msg->list_head);
1013                 ceph_msg_put(msg);
1014         }
1015         while (!list_empty(&session->s_cap_releases_done)) {
1016                 msg = list_first_entry(&session->s_cap_releases_done,
1017                                        struct ceph_msg, list_head);
1018                 list_del_init(&msg->list_head);
1019                 ceph_msg_put(msg);
1020         }
1021         spin_unlock(&session->s_cap_lock);
1022 }
1023
1024 /*
1025  * Helper to safely iterate over all caps associated with a session, with
1026  * special care taken to handle a racing __ceph_remove_cap().
1027  *
1028  * Caller must hold session s_mutex.
1029  */
1030 static int iterate_session_caps(struct ceph_mds_session *session,
1031                                  int (*cb)(struct inode *, struct ceph_cap *,
1032                                             void *), void *arg)
1033 {
1034         struct list_head *p;
1035         struct ceph_cap *cap;
1036         struct inode *inode, *last_inode = NULL;
1037         struct ceph_cap *old_cap = NULL;
1038         int ret;
1039
1040         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1041         spin_lock(&session->s_cap_lock);
1042         p = session->s_caps.next;
1043         while (p != &session->s_caps) {
1044                 cap = list_entry(p, struct ceph_cap, session_caps);
1045                 inode = igrab(&cap->ci->vfs_inode);
1046                 if (!inode) {
1047                         p = p->next;
1048                         continue;
1049                 }
1050                 session->s_cap_iterator = cap;
1051                 spin_unlock(&session->s_cap_lock);
1052
1053                 if (last_inode) {
1054                         iput(last_inode);
1055                         last_inode = NULL;
1056                 }
1057                 if (old_cap) {
1058                         ceph_put_cap(session->s_mdsc, old_cap);
1059                         old_cap = NULL;
1060                 }
1061
1062                 ret = cb(inode, cap, arg);
1063                 last_inode = inode;
1064
1065                 spin_lock(&session->s_cap_lock);
1066                 p = p->next;
1067                 if (cap->ci == NULL) {
1068                         dout("iterate_session_caps  finishing cap %p removal\n",
1069                              cap);
1070                         BUG_ON(cap->session != session);
1071                         list_del_init(&cap->session_caps);
1072                         session->s_nr_caps--;
1073                         cap->session = NULL;
1074                         old_cap = cap;  /* put_cap it w/o locks held */
1075                 }
1076                 if (ret < 0)
1077                         goto out;
1078         }
1079         ret = 0;
1080 out:
1081         session->s_cap_iterator = NULL;
1082         spin_unlock(&session->s_cap_lock);
1083
1084         iput(last_inode);
1085         if (old_cap)
1086                 ceph_put_cap(session->s_mdsc, old_cap);
1087
1088         return ret;
1089 }
1090
1091 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1092                                   void *arg)
1093 {
1094         struct ceph_inode_info *ci = ceph_inode(inode);
1095         int drop = 0;
1096
1097         dout("removing cap %p, ci is %p, inode is %p\n",
1098              cap, ci, &ci->vfs_inode);
1099         spin_lock(&ci->i_ceph_lock);
1100         __ceph_remove_cap(cap, false);
1101         if (!__ceph_is_any_real_caps(ci)) {
1102                 struct ceph_mds_client *mdsc =
1103                         ceph_sb_to_client(inode->i_sb)->mdsc;
1104
1105                 spin_lock(&mdsc->cap_dirty_lock);
1106                 if (!list_empty(&ci->i_dirty_item)) {
1107                         pr_info(" dropping dirty %s state for %p %lld\n",
1108                                 ceph_cap_string(ci->i_dirty_caps),
1109                                 inode, ceph_ino(inode));
1110                         ci->i_dirty_caps = 0;
1111                         list_del_init(&ci->i_dirty_item);
1112                         drop = 1;
1113                 }
1114                 if (!list_empty(&ci->i_flushing_item)) {
1115                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1116                                 ceph_cap_string(ci->i_flushing_caps),
1117                                 inode, ceph_ino(inode));
1118                         ci->i_flushing_caps = 0;
1119                         list_del_init(&ci->i_flushing_item);
1120                         mdsc->num_cap_flushing--;
1121                         drop = 1;
1122                 }
1123                 if (drop && ci->i_wrbuffer_ref) {
1124                         pr_info(" dropping dirty data for %p %lld\n",
1125                                 inode, ceph_ino(inode));
1126                         ci->i_wrbuffer_ref = 0;
1127                         ci->i_wrbuffer_ref_head = 0;
1128                         drop++;
1129                 }
1130                 spin_unlock(&mdsc->cap_dirty_lock);
1131         }
1132         spin_unlock(&ci->i_ceph_lock);
1133         while (drop--)
1134                 iput(inode);
1135         return 0;
1136 }
1137
1138 /*
1139  * caller must hold session s_mutex
1140  */
1141 static void remove_session_caps(struct ceph_mds_session *session)
1142 {
1143         dout("remove_session_caps on %p\n", session);
1144         iterate_session_caps(session, remove_session_caps_cb, NULL);
1145
1146         spin_lock(&session->s_cap_lock);
1147         if (session->s_nr_caps > 0) {
1148                 struct super_block *sb = session->s_mdsc->fsc->sb;
1149                 struct inode *inode;
1150                 struct ceph_cap *cap, *prev = NULL;
1151                 struct ceph_vino vino;
1152                 /*
1153                  * iterate_session_caps() skips inodes that are being
1154                  * deleted, we need to wait until deletions are complete.
1155                  * __wait_on_freeing_inode() is designed for the job,
1156                  * but it is not exported, so use lookup inode function
1157                  * to access it.
1158                  */
1159                 while (!list_empty(&session->s_caps)) {
1160                         cap = list_entry(session->s_caps.next,
1161                                          struct ceph_cap, session_caps);
1162                         if (cap == prev)
1163                                 break;
1164                         prev = cap;
1165                         vino = cap->ci->i_vino;
1166                         spin_unlock(&session->s_cap_lock);
1167
1168                         inode = ceph_find_inode(sb, vino);
1169                         iput(inode);
1170
1171                         spin_lock(&session->s_cap_lock);
1172                 }
1173         }
1174         spin_unlock(&session->s_cap_lock);
1175
1176         BUG_ON(session->s_nr_caps > 0);
1177         BUG_ON(!list_empty(&session->s_cap_flushing));
1178         cleanup_cap_releases(session);
1179 }
1180
1181 /*
1182  * wake up any threads waiting on this session's caps.  if the cap is
1183  * old (didn't get renewed on the client reconnect), remove it now.
1184  *
1185  * caller must hold s_mutex.
1186  */
1187 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1188                               void *arg)
1189 {
1190         struct ceph_inode_info *ci = ceph_inode(inode);
1191
1192         wake_up_all(&ci->i_cap_wq);
1193         if (arg) {
1194                 spin_lock(&ci->i_ceph_lock);
1195                 ci->i_wanted_max_size = 0;
1196                 ci->i_requested_max_size = 0;
1197                 spin_unlock(&ci->i_ceph_lock);
1198         }
1199         return 0;
1200 }
1201
1202 static void wake_up_session_caps(struct ceph_mds_session *session,
1203                                  int reconnect)
1204 {
1205         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1206         iterate_session_caps(session, wake_up_session_cb,
1207                              (void *)(unsigned long)reconnect);
1208 }
1209
1210 /*
1211  * Send periodic message to MDS renewing all currently held caps.  The
1212  * ack will reset the expiration for all caps from this session.
1213  *
1214  * caller holds s_mutex
1215  */
1216 static int send_renew_caps(struct ceph_mds_client *mdsc,
1217                            struct ceph_mds_session *session)
1218 {
1219         struct ceph_msg *msg;
1220         int state;
1221
1222         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1223             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1224                 pr_info("mds%d caps stale\n", session->s_mds);
1225         session->s_renew_requested = jiffies;
1226
1227         /* do not try to renew caps until a recovering mds has reconnected
1228          * with its clients. */
1229         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1230         if (state < CEPH_MDS_STATE_RECONNECT) {
1231                 dout("send_renew_caps ignoring mds%d (%s)\n",
1232                      session->s_mds, ceph_mds_state_name(state));
1233                 return 0;
1234         }
1235
1236         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1237                 ceph_mds_state_name(state));
1238         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1239                                  ++session->s_renew_seq);
1240         if (!msg)
1241                 return -ENOMEM;
1242         ceph_con_send(&session->s_con, msg);
1243         return 0;
1244 }
1245
1246 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1247                              struct ceph_mds_session *session, u64 seq)
1248 {
1249         struct ceph_msg *msg;
1250
1251         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1252              session->s_mds, ceph_session_state_name(session->s_state), seq);
1253         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1254         if (!msg)
1255                 return -ENOMEM;
1256         ceph_con_send(&session->s_con, msg);
1257         return 0;
1258 }
1259
1260
1261 /*
1262  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1263  *
1264  * Called under session->s_mutex
1265  */
1266 static void renewed_caps(struct ceph_mds_client *mdsc,
1267                          struct ceph_mds_session *session, int is_renew)
1268 {
1269         int was_stale;
1270         int wake = 0;
1271
1272         spin_lock(&session->s_cap_lock);
1273         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1274
1275         session->s_cap_ttl = session->s_renew_requested +
1276                 mdsc->mdsmap->m_session_timeout*HZ;
1277
1278         if (was_stale) {
1279                 if (time_before(jiffies, session->s_cap_ttl)) {
1280                         pr_info("mds%d caps renewed\n", session->s_mds);
1281                         wake = 1;
1282                 } else {
1283                         pr_info("mds%d caps still stale\n", session->s_mds);
1284                 }
1285         }
1286         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1287              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1288              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1289         spin_unlock(&session->s_cap_lock);
1290
1291         if (wake)
1292                 wake_up_session_caps(session, 0);
1293 }
1294
1295 /*
1296  * send a session close request
1297  */
1298 static int request_close_session(struct ceph_mds_client *mdsc,
1299                                  struct ceph_mds_session *session)
1300 {
1301         struct ceph_msg *msg;
1302
1303         dout("request_close_session mds%d state %s seq %lld\n",
1304              session->s_mds, ceph_session_state_name(session->s_state),
1305              session->s_seq);
1306         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1307         if (!msg)
1308                 return -ENOMEM;
1309         ceph_con_send(&session->s_con, msg);
1310         return 0;
1311 }
1312
1313 /*
1314  * Called with s_mutex held.
1315  */
1316 static int __close_session(struct ceph_mds_client *mdsc,
1317                          struct ceph_mds_session *session)
1318 {
1319         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1320                 return 0;
1321         session->s_state = CEPH_MDS_SESSION_CLOSING;
1322         return request_close_session(mdsc, session);
1323 }
1324
1325 /*
1326  * Trim old(er) caps.
1327  *
1328  * Because we can't cache an inode without one or more caps, we do
1329  * this indirectly: if a cap is unused, we prune its aliases, at which
1330  * point the inode will hopefully get dropped to.
1331  *
1332  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1333  * memory pressure from the MDS, though, so it needn't be perfect.
1334  */
1335 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1336 {
1337         struct ceph_mds_session *session = arg;
1338         struct ceph_inode_info *ci = ceph_inode(inode);
1339         int used, wanted, oissued, mine;
1340
1341         if (session->s_trim_caps <= 0)
1342                 return -1;
1343
1344         spin_lock(&ci->i_ceph_lock);
1345         mine = cap->issued | cap->implemented;
1346         used = __ceph_caps_used(ci);
1347         wanted = __ceph_caps_file_wanted(ci);
1348         oissued = __ceph_caps_issued_other(ci, cap);
1349
1350         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1351              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1352              ceph_cap_string(used), ceph_cap_string(wanted));
1353         if (cap == ci->i_auth_cap) {
1354                 if (ci->i_dirty_caps | ci->i_flushing_caps)
1355                         goto out;
1356                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1357                         goto out;
1358         }
1359         if ((used | wanted) & ~oissued & mine)
1360                 goto out;   /* we need these caps */
1361
1362         session->s_trim_caps--;
1363         if (oissued) {
1364                 /* we aren't the only cap.. just remove us */
1365                 __ceph_remove_cap(cap, true);
1366         } else {
1367                 /* try to drop referring dentries */
1368                 spin_unlock(&ci->i_ceph_lock);
1369                 d_prune_aliases(inode);
1370                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1371                      inode, cap, atomic_read(&inode->i_count));
1372                 return 0;
1373         }
1374
1375 out:
1376         spin_unlock(&ci->i_ceph_lock);
1377         return 0;
1378 }
1379
1380 /*
1381  * Trim session cap count down to some max number.
1382  */
1383 static int trim_caps(struct ceph_mds_client *mdsc,
1384                      struct ceph_mds_session *session,
1385                      int max_caps)
1386 {
1387         int trim_caps = session->s_nr_caps - max_caps;
1388
1389         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1390              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1391         if (trim_caps > 0) {
1392                 session->s_trim_caps = trim_caps;
1393                 iterate_session_caps(session, trim_caps_cb, session);
1394                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1395                      session->s_mds, session->s_nr_caps, max_caps,
1396                         trim_caps - session->s_trim_caps);
1397                 session->s_trim_caps = 0;
1398         }
1399
1400         ceph_add_cap_releases(mdsc, session);
1401         ceph_send_cap_releases(mdsc, session);
1402         return 0;
1403 }
1404
1405 /*
1406  * Allocate cap_release messages.  If there is a partially full message
1407  * in the queue, try to allocate enough to cover it's remainder, so that
1408  * we can send it immediately.
1409  *
1410  * Called under s_mutex.
1411  */
1412 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1413                           struct ceph_mds_session *session)
1414 {
1415         struct ceph_msg *msg, *partial = NULL;
1416         struct ceph_mds_cap_release *head;
1417         int err = -ENOMEM;
1418         int extra = mdsc->fsc->mount_options->cap_release_safety;
1419         int num;
1420
1421         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1422              extra);
1423
1424         spin_lock(&session->s_cap_lock);
1425
1426         if (!list_empty(&session->s_cap_releases)) {
1427                 msg = list_first_entry(&session->s_cap_releases,
1428                                        struct ceph_msg,
1429                                  list_head);
1430                 head = msg->front.iov_base;
1431                 num = le32_to_cpu(head->num);
1432                 if (num) {
1433                         dout(" partial %p with (%d/%d)\n", msg, num,
1434                              (int)CEPH_CAPS_PER_RELEASE);
1435                         extra += CEPH_CAPS_PER_RELEASE - num;
1436                         partial = msg;
1437                 }
1438         }
1439         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1440                 spin_unlock(&session->s_cap_lock);
1441                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1442                                    GFP_NOFS, false);
1443                 if (!msg)
1444                         goto out_unlocked;
1445                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1446                      (int)msg->front.iov_len);
1447                 head = msg->front.iov_base;
1448                 head->num = cpu_to_le32(0);
1449                 msg->front.iov_len = sizeof(*head);
1450                 spin_lock(&session->s_cap_lock);
1451                 list_add(&msg->list_head, &session->s_cap_releases);
1452                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1453         }
1454
1455         if (partial) {
1456                 head = partial->front.iov_base;
1457                 num = le32_to_cpu(head->num);
1458                 dout(" queueing partial %p with %d/%d\n", partial, num,
1459                      (int)CEPH_CAPS_PER_RELEASE);
1460                 list_move_tail(&partial->list_head,
1461                                &session->s_cap_releases_done);
1462                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1463         }
1464         err = 0;
1465         spin_unlock(&session->s_cap_lock);
1466 out_unlocked:
1467         return err;
1468 }
1469
1470 static int check_cap_flush(struct inode *inode, u64 want_flush_seq)
1471 {
1472         struct ceph_inode_info *ci = ceph_inode(inode);
1473         int ret;
1474         spin_lock(&ci->i_ceph_lock);
1475         if (ci->i_flushing_caps)
1476                 ret = ci->i_cap_flush_seq >= want_flush_seq;
1477         else
1478                 ret = 1;
1479         spin_unlock(&ci->i_ceph_lock);
1480         return ret;
1481 }
1482
1483 /*
1484  * flush all dirty inode data to disk.
1485  *
1486  * returns true if we've flushed through want_flush_seq
1487  */
1488 static void wait_caps_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1489 {
1490         int mds;
1491
1492         dout("check_cap_flush want %lld\n", want_flush_seq);
1493         mutex_lock(&mdsc->mutex);
1494         for (mds = 0; mds < mdsc->max_sessions; mds++) {
1495                 struct ceph_mds_session *session = mdsc->sessions[mds];
1496                 struct inode *inode = NULL;
1497
1498                 if (!session)
1499                         continue;
1500                 get_session(session);
1501                 mutex_unlock(&mdsc->mutex);
1502
1503                 mutex_lock(&session->s_mutex);
1504                 if (!list_empty(&session->s_cap_flushing)) {
1505                         struct ceph_inode_info *ci =
1506                                 list_entry(session->s_cap_flushing.next,
1507                                            struct ceph_inode_info,
1508                                            i_flushing_item);
1509
1510                         if (!check_cap_flush(&ci->vfs_inode, want_flush_seq)) {
1511                                 dout("check_cap_flush still flushing %p "
1512                                      "seq %lld <= %lld to mds%d\n",
1513                                      &ci->vfs_inode, ci->i_cap_flush_seq,
1514                                      want_flush_seq, session->s_mds);
1515                                 inode = igrab(&ci->vfs_inode);
1516                         }
1517                 }
1518                 mutex_unlock(&session->s_mutex);
1519                 ceph_put_mds_session(session);
1520
1521                 if (inode) {
1522                         wait_event(mdsc->cap_flushing_wq,
1523                                    check_cap_flush(inode, want_flush_seq));
1524                         iput(inode);
1525                 }
1526
1527                 mutex_lock(&mdsc->mutex);
1528         }
1529
1530         mutex_unlock(&mdsc->mutex);
1531         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1532 }
1533
1534 /*
1535  * called under s_mutex
1536  */
1537 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1538                             struct ceph_mds_session *session)
1539 {
1540         struct ceph_msg *msg;
1541
1542         dout("send_cap_releases mds%d\n", session->s_mds);
1543         spin_lock(&session->s_cap_lock);
1544         while (!list_empty(&session->s_cap_releases_done)) {
1545                 msg = list_first_entry(&session->s_cap_releases_done,
1546                                  struct ceph_msg, list_head);
1547                 list_del_init(&msg->list_head);
1548                 spin_unlock(&session->s_cap_lock);
1549                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1550                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1551                 ceph_con_send(&session->s_con, msg);
1552                 spin_lock(&session->s_cap_lock);
1553         }
1554         spin_unlock(&session->s_cap_lock);
1555 }
1556
1557 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1558                                  struct ceph_mds_session *session)
1559 {
1560         struct ceph_msg *msg;
1561         struct ceph_mds_cap_release *head;
1562         unsigned num;
1563
1564         dout("discard_cap_releases mds%d\n", session->s_mds);
1565
1566         if (!list_empty(&session->s_cap_releases)) {
1567                 /* zero out the in-progress message */
1568                 msg = list_first_entry(&session->s_cap_releases,
1569                                         struct ceph_msg, list_head);
1570                 head = msg->front.iov_base;
1571                 num = le32_to_cpu(head->num);
1572                 dout("discard_cap_releases mds%d %p %u\n",
1573                      session->s_mds, msg, num);
1574                 head->num = cpu_to_le32(0);
1575                 msg->front.iov_len = sizeof(*head);
1576                 session->s_num_cap_releases += num;
1577         }
1578
1579         /* requeue completed messages */
1580         while (!list_empty(&session->s_cap_releases_done)) {
1581                 msg = list_first_entry(&session->s_cap_releases_done,
1582                                  struct ceph_msg, list_head);
1583                 list_del_init(&msg->list_head);
1584
1585                 head = msg->front.iov_base;
1586                 num = le32_to_cpu(head->num);
1587                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1588                      num);
1589                 session->s_num_cap_releases += num;
1590                 head->num = cpu_to_le32(0);
1591                 msg->front.iov_len = sizeof(*head);
1592                 list_add(&msg->list_head, &session->s_cap_releases);
1593         }
1594 }
1595
1596 /*
1597  * requests
1598  */
1599
1600 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1601                                     struct inode *dir)
1602 {
1603         struct ceph_inode_info *ci = ceph_inode(dir);
1604         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1605         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1606         size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1607                       sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1608         int order, num_entries;
1609
1610         spin_lock(&ci->i_ceph_lock);
1611         num_entries = ci->i_files + ci->i_subdirs;
1612         spin_unlock(&ci->i_ceph_lock);
1613         num_entries = max(num_entries, 1);
1614         num_entries = min(num_entries, opt->max_readdir);
1615
1616         order = get_order(size * num_entries);
1617         while (order >= 0) {
1618                 rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN,
1619                                                         order);
1620                 if (rinfo->dir_in)
1621                         break;
1622                 order--;
1623         }
1624         if (!rinfo->dir_in)
1625                 return -ENOMEM;
1626
1627         num_entries = (PAGE_SIZE << order) / size;
1628         num_entries = min(num_entries, opt->max_readdir);
1629
1630         rinfo->dir_buf_size = PAGE_SIZE << order;
1631         req->r_num_caps = num_entries + 1;
1632         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1633         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1634         return 0;
1635 }
1636
1637 /*
1638  * Create an mds request.
1639  */
1640 struct ceph_mds_request *
1641 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1642 {
1643         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1644
1645         if (!req)
1646                 return ERR_PTR(-ENOMEM);
1647
1648         mutex_init(&req->r_fill_mutex);
1649         req->r_mdsc = mdsc;
1650         req->r_started = jiffies;
1651         req->r_resend_mds = -1;
1652         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1653         req->r_fmode = -1;
1654         kref_init(&req->r_kref);
1655         INIT_LIST_HEAD(&req->r_wait);
1656         init_completion(&req->r_completion);
1657         init_completion(&req->r_safe_completion);
1658         INIT_LIST_HEAD(&req->r_unsafe_item);
1659
1660         req->r_stamp = CURRENT_TIME;
1661
1662         req->r_op = op;
1663         req->r_direct_mode = mode;
1664         return req;
1665 }
1666
1667 /*
1668  * return oldest (lowest) request, tid in request tree, 0 if none.
1669  *
1670  * called under mdsc->mutex.
1671  */
1672 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1673 {
1674         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1675                 return NULL;
1676         return rb_entry(rb_first(&mdsc->request_tree),
1677                         struct ceph_mds_request, r_node);
1678 }
1679
1680 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1681 {
1682         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1683
1684         if (req)
1685                 return req->r_tid;
1686         return 0;
1687 }
1688
1689 /*
1690  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1691  * on build_path_from_dentry in fs/cifs/dir.c.
1692  *
1693  * If @stop_on_nosnap, generate path relative to the first non-snapped
1694  * inode.
1695  *
1696  * Encode hidden .snap dirs as a double /, i.e.
1697  *   foo/.snap/bar -> foo//bar
1698  */
1699 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1700                            int stop_on_nosnap)
1701 {
1702         struct dentry *temp;
1703         char *path;
1704         int len, pos;
1705         unsigned seq;
1706
1707         if (dentry == NULL)
1708                 return ERR_PTR(-EINVAL);
1709
1710 retry:
1711         len = 0;
1712         seq = read_seqbegin(&rename_lock);
1713         rcu_read_lock();
1714         for (temp = dentry; !IS_ROOT(temp);) {
1715                 struct inode *inode = temp->d_inode;
1716                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1717                         len++;  /* slash only */
1718                 else if (stop_on_nosnap && inode &&
1719                          ceph_snap(inode) == CEPH_NOSNAP)
1720                         break;
1721                 else
1722                         len += 1 + temp->d_name.len;
1723                 temp = temp->d_parent;
1724         }
1725         rcu_read_unlock();
1726         if (len)
1727                 len--;  /* no leading '/' */
1728
1729         path = kmalloc(len+1, GFP_NOFS);
1730         if (path == NULL)
1731                 return ERR_PTR(-ENOMEM);
1732         pos = len;
1733         path[pos] = 0;  /* trailing null */
1734         rcu_read_lock();
1735         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1736                 struct inode *inode;
1737
1738                 spin_lock(&temp->d_lock);
1739                 inode = temp->d_inode;
1740                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1741                         dout("build_path path+%d: %p SNAPDIR\n",
1742                              pos, temp);
1743                 } else if (stop_on_nosnap && inode &&
1744                            ceph_snap(inode) == CEPH_NOSNAP) {
1745                         spin_unlock(&temp->d_lock);
1746                         break;
1747                 } else {
1748                         pos -= temp->d_name.len;
1749                         if (pos < 0) {
1750                                 spin_unlock(&temp->d_lock);
1751                                 break;
1752                         }
1753                         strncpy(path + pos, temp->d_name.name,
1754                                 temp->d_name.len);
1755                 }
1756                 spin_unlock(&temp->d_lock);
1757                 if (pos)
1758                         path[--pos] = '/';
1759                 temp = temp->d_parent;
1760         }
1761         rcu_read_unlock();
1762         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1763                 pr_err("build_path did not end path lookup where "
1764                        "expected, namelen is %d, pos is %d\n", len, pos);
1765                 /* presumably this is only possible if racing with a
1766                    rename of one of the parent directories (we can not
1767                    lock the dentries above us to prevent this, but
1768                    retrying should be harmless) */
1769                 kfree(path);
1770                 goto retry;
1771         }
1772
1773         *base = ceph_ino(temp->d_inode);
1774         *plen = len;
1775         dout("build_path on %p %d built %llx '%.*s'\n",
1776              dentry, d_count(dentry), *base, len, path);
1777         return path;
1778 }
1779
1780 static int build_dentry_path(struct dentry *dentry,
1781                              const char **ppath, int *ppathlen, u64 *pino,
1782                              int *pfreepath)
1783 {
1784         char *path;
1785
1786         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1787                 *pino = ceph_ino(dentry->d_parent->d_inode);
1788                 *ppath = dentry->d_name.name;
1789                 *ppathlen = dentry->d_name.len;
1790                 return 0;
1791         }
1792         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1793         if (IS_ERR(path))
1794                 return PTR_ERR(path);
1795         *ppath = path;
1796         *pfreepath = 1;
1797         return 0;
1798 }
1799
1800 static int build_inode_path(struct inode *inode,
1801                             const char **ppath, int *ppathlen, u64 *pino,
1802                             int *pfreepath)
1803 {
1804         struct dentry *dentry;
1805         char *path;
1806
1807         if (ceph_snap(inode) == CEPH_NOSNAP) {
1808                 *pino = ceph_ino(inode);
1809                 *ppathlen = 0;
1810                 return 0;
1811         }
1812         dentry = d_find_alias(inode);
1813         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1814         dput(dentry);
1815         if (IS_ERR(path))
1816                 return PTR_ERR(path);
1817         *ppath = path;
1818         *pfreepath = 1;
1819         return 0;
1820 }
1821
1822 /*
1823  * request arguments may be specified via an inode *, a dentry *, or
1824  * an explicit ino+path.
1825  */
1826 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1827                                   const char *rpath, u64 rino,
1828                                   const char **ppath, int *pathlen,
1829                                   u64 *ino, int *freepath)
1830 {
1831         int r = 0;
1832
1833         if (rinode) {
1834                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1835                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1836                      ceph_snap(rinode));
1837         } else if (rdentry) {
1838                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1839                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1840                      *ppath);
1841         } else if (rpath || rino) {
1842                 *ino = rino;
1843                 *ppath = rpath;
1844                 *pathlen = rpath ? strlen(rpath) : 0;
1845                 dout(" path %.*s\n", *pathlen, rpath);
1846         }
1847
1848         return r;
1849 }
1850
1851 /*
1852  * called under mdsc->mutex
1853  */
1854 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1855                                                struct ceph_mds_request *req,
1856                                                int mds, bool drop_cap_releases)
1857 {
1858         struct ceph_msg *msg;
1859         struct ceph_mds_request_head *head;
1860         const char *path1 = NULL;
1861         const char *path2 = NULL;
1862         u64 ino1 = 0, ino2 = 0;
1863         int pathlen1 = 0, pathlen2 = 0;
1864         int freepath1 = 0, freepath2 = 0;
1865         int len;
1866         u16 releases;
1867         void *p, *end;
1868         int ret;
1869
1870         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1871                               req->r_path1, req->r_ino1.ino,
1872                               &path1, &pathlen1, &ino1, &freepath1);
1873         if (ret < 0) {
1874                 msg = ERR_PTR(ret);
1875                 goto out;
1876         }
1877
1878         ret = set_request_path_attr(NULL, req->r_old_dentry,
1879                               req->r_path2, req->r_ino2.ino,
1880                               &path2, &pathlen2, &ino2, &freepath2);
1881         if (ret < 0) {
1882                 msg = ERR_PTR(ret);
1883                 goto out_free1;
1884         }
1885
1886         len = sizeof(*head) +
1887                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1888                 sizeof(struct timespec);
1889
1890         /* calculate (max) length for cap releases */
1891         len += sizeof(struct ceph_mds_request_release) *
1892                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1893                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1894         if (req->r_dentry_drop)
1895                 len += req->r_dentry->d_name.len;
1896         if (req->r_old_dentry_drop)
1897                 len += req->r_old_dentry->d_name.len;
1898
1899         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1900         if (!msg) {
1901                 msg = ERR_PTR(-ENOMEM);
1902                 goto out_free2;
1903         }
1904
1905         msg->hdr.version = cpu_to_le16(2);
1906         msg->hdr.tid = cpu_to_le64(req->r_tid);
1907
1908         head = msg->front.iov_base;
1909         p = msg->front.iov_base + sizeof(*head);
1910         end = msg->front.iov_base + msg->front.iov_len;
1911
1912         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1913         head->op = cpu_to_le32(req->r_op);
1914         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1915         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1916         head->args = req->r_args;
1917
1918         ceph_encode_filepath(&p, end, ino1, path1);
1919         ceph_encode_filepath(&p, end, ino2, path2);
1920
1921         /* make note of release offset, in case we need to replay */
1922         req->r_request_release_offset = p - msg->front.iov_base;
1923
1924         /* cap releases */
1925         releases = 0;
1926         if (req->r_inode_drop)
1927                 releases += ceph_encode_inode_release(&p,
1928                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1929                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1930         if (req->r_dentry_drop)
1931                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1932                        mds, req->r_dentry_drop, req->r_dentry_unless);
1933         if (req->r_old_dentry_drop)
1934                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1935                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1936         if (req->r_old_inode_drop)
1937                 releases += ceph_encode_inode_release(&p,
1938                       req->r_old_dentry->d_inode,
1939                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1940
1941         if (drop_cap_releases) {
1942                 releases = 0;
1943                 p = msg->front.iov_base + req->r_request_release_offset;
1944         }
1945
1946         head->num_releases = cpu_to_le16(releases);
1947
1948         /* time stamp */
1949         {
1950                 struct ceph_timespec ts;
1951                 ceph_encode_timespec(&ts, &req->r_stamp);
1952                 ceph_encode_copy(&p, &ts, sizeof(ts));
1953         }
1954
1955         BUG_ON(p > end);
1956         msg->front.iov_len = p - msg->front.iov_base;
1957         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1958
1959         if (req->r_pagelist) {
1960                 struct ceph_pagelist *pagelist = req->r_pagelist;
1961                 atomic_inc(&pagelist->refcnt);
1962                 ceph_msg_data_add_pagelist(msg, pagelist);
1963                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1964         } else {
1965                 msg->hdr.data_len = 0;
1966         }
1967
1968         msg->hdr.data_off = cpu_to_le16(0);
1969
1970 out_free2:
1971         if (freepath2)
1972                 kfree((char *)path2);
1973 out_free1:
1974         if (freepath1)
1975                 kfree((char *)path1);
1976 out:
1977         return msg;
1978 }
1979
1980 /*
1981  * called under mdsc->mutex if error, under no mutex if
1982  * success.
1983  */
1984 static void complete_request(struct ceph_mds_client *mdsc,
1985                              struct ceph_mds_request *req)
1986 {
1987         if (req->r_callback)
1988                 req->r_callback(mdsc, req);
1989         else
1990                 complete_all(&req->r_completion);
1991 }
1992
1993 /*
1994  * called under mdsc->mutex
1995  */
1996 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1997                                   struct ceph_mds_request *req,
1998                                   int mds, bool drop_cap_releases)
1999 {
2000         struct ceph_mds_request_head *rhead;
2001         struct ceph_msg *msg;
2002         int flags = 0;
2003
2004         req->r_attempts++;
2005         if (req->r_inode) {
2006                 struct ceph_cap *cap =
2007                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2008
2009                 if (cap)
2010                         req->r_sent_on_mseq = cap->mseq;
2011                 else
2012                         req->r_sent_on_mseq = -1;
2013         }
2014         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2015              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2016
2017         if (req->r_got_unsafe) {
2018                 void *p;
2019                 /*
2020                  * Replay.  Do not regenerate message (and rebuild
2021                  * paths, etc.); just use the original message.
2022                  * Rebuilding paths will break for renames because
2023                  * d_move mangles the src name.
2024                  */
2025                 msg = req->r_request;
2026                 rhead = msg->front.iov_base;
2027
2028                 flags = le32_to_cpu(rhead->flags);
2029                 flags |= CEPH_MDS_FLAG_REPLAY;
2030                 rhead->flags = cpu_to_le32(flags);
2031
2032                 if (req->r_target_inode)
2033                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2034
2035                 rhead->num_retry = req->r_attempts - 1;
2036
2037                 /* remove cap/dentry releases from message */
2038                 rhead->num_releases = 0;
2039
2040                 /* time stamp */
2041                 p = msg->front.iov_base + req->r_request_release_offset;
2042                 {
2043                         struct ceph_timespec ts;
2044                         ceph_encode_timespec(&ts, &req->r_stamp);
2045                         ceph_encode_copy(&p, &ts, sizeof(ts));
2046                 }
2047
2048                 msg->front.iov_len = p - msg->front.iov_base;
2049                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2050                 return 0;
2051         }
2052
2053         if (req->r_request) {
2054                 ceph_msg_put(req->r_request);
2055                 req->r_request = NULL;
2056         }
2057         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2058         if (IS_ERR(msg)) {
2059                 req->r_err = PTR_ERR(msg);
2060                 complete_request(mdsc, req);
2061                 return PTR_ERR(msg);
2062         }
2063         req->r_request = msg;
2064
2065         rhead = msg->front.iov_base;
2066         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2067         if (req->r_got_unsafe)
2068                 flags |= CEPH_MDS_FLAG_REPLAY;
2069         if (req->r_locked_dir)
2070                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2071         rhead->flags = cpu_to_le32(flags);
2072         rhead->num_fwd = req->r_num_fwd;
2073         rhead->num_retry = req->r_attempts - 1;
2074         rhead->ino = 0;
2075
2076         dout(" r_locked_dir = %p\n", req->r_locked_dir);
2077         return 0;
2078 }
2079
2080 /*
2081  * send request, or put it on the appropriate wait list.
2082  */
2083 static int __do_request(struct ceph_mds_client *mdsc,
2084                         struct ceph_mds_request *req)
2085 {
2086         struct ceph_mds_session *session = NULL;
2087         int mds = -1;
2088         int err = -EAGAIN;
2089
2090         if (req->r_err || req->r_got_result) {
2091                 if (req->r_aborted)
2092                         __unregister_request(mdsc, req);
2093                 goto out;
2094         }
2095
2096         if (req->r_timeout &&
2097             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2098                 dout("do_request timed out\n");
2099                 err = -EIO;
2100                 goto finish;
2101         }
2102
2103         put_request_session(req);
2104
2105         mds = __choose_mds(mdsc, req);
2106         if (mds < 0 ||
2107             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2108                 dout("do_request no mds or not active, waiting for map\n");
2109                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2110                 goto out;
2111         }
2112
2113         /* get, open session */
2114         session = __ceph_lookup_mds_session(mdsc, mds);
2115         if (!session) {
2116                 session = register_session(mdsc, mds);
2117                 if (IS_ERR(session)) {
2118                         err = PTR_ERR(session);
2119                         goto finish;
2120                 }
2121         }
2122         req->r_session = get_session(session);
2123
2124         dout("do_request mds%d session %p state %s\n", mds, session,
2125              ceph_session_state_name(session->s_state));
2126         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2127             session->s_state != CEPH_MDS_SESSION_HUNG) {
2128                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2129                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2130                         __open_session(mdsc, session);
2131                 list_add(&req->r_wait, &session->s_waiting);
2132                 goto out_session;
2133         }
2134
2135         /* send request */
2136         req->r_resend_mds = -1;   /* forget any previous mds hint */
2137
2138         if (req->r_request_started == 0)   /* note request start time */
2139                 req->r_request_started = jiffies;
2140
2141         err = __prepare_send_request(mdsc, req, mds, false);
2142         if (!err) {
2143                 ceph_msg_get(req->r_request);
2144                 ceph_con_send(&session->s_con, req->r_request);
2145         }
2146
2147 out_session:
2148         ceph_put_mds_session(session);
2149 out:
2150         return err;
2151
2152 finish:
2153         req->r_err = err;
2154         complete_request(mdsc, req);
2155         goto out;
2156 }
2157
2158 /*
2159  * called under mdsc->mutex
2160  */
2161 static void __wake_requests(struct ceph_mds_client *mdsc,
2162                             struct list_head *head)
2163 {
2164         struct ceph_mds_request *req;
2165         LIST_HEAD(tmp_list);
2166
2167         list_splice_init(head, &tmp_list);
2168
2169         while (!list_empty(&tmp_list)) {
2170                 req = list_entry(tmp_list.next,
2171                                  struct ceph_mds_request, r_wait);
2172                 list_del_init(&req->r_wait);
2173                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2174                 __do_request(mdsc, req);
2175         }
2176 }
2177
2178 /*
2179  * Wake up threads with requests pending for @mds, so that they can
2180  * resubmit their requests to a possibly different mds.
2181  */
2182 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2183 {
2184         struct ceph_mds_request *req;
2185         struct rb_node *p = rb_first(&mdsc->request_tree);
2186
2187         dout("kick_requests mds%d\n", mds);
2188         while (p) {
2189                 req = rb_entry(p, struct ceph_mds_request, r_node);
2190                 p = rb_next(p);
2191                 if (req->r_got_unsafe)
2192                         continue;
2193                 if (req->r_attempts > 0)
2194                         continue; /* only new requests */
2195                 if (req->r_session &&
2196                     req->r_session->s_mds == mds) {
2197                         dout(" kicking tid %llu\n", req->r_tid);
2198                         list_del_init(&req->r_wait);
2199                         __do_request(mdsc, req);
2200                 }
2201         }
2202 }
2203
2204 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2205                               struct ceph_mds_request *req)
2206 {
2207         dout("submit_request on %p\n", req);
2208         mutex_lock(&mdsc->mutex);
2209         __register_request(mdsc, req, NULL);
2210         __do_request(mdsc, req);
2211         mutex_unlock(&mdsc->mutex);
2212 }
2213
2214 /*
2215  * Synchrously perform an mds request.  Take care of all of the
2216  * session setup, forwarding, retry details.
2217  */
2218 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2219                          struct inode *dir,
2220                          struct ceph_mds_request *req)
2221 {
2222         int err;
2223
2224         dout("do_request on %p\n", req);
2225
2226         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2227         if (req->r_inode)
2228                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2229         if (req->r_locked_dir)
2230                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2231         if (req->r_old_dentry_dir)
2232                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2233                                   CEPH_CAP_PIN);
2234
2235         /* issue */
2236         mutex_lock(&mdsc->mutex);
2237         __register_request(mdsc, req, dir);
2238         __do_request(mdsc, req);
2239
2240         if (req->r_err) {
2241                 err = req->r_err;
2242                 __unregister_request(mdsc, req);
2243                 dout("do_request early error %d\n", err);
2244                 goto out;
2245         }
2246
2247         /* wait */
2248         mutex_unlock(&mdsc->mutex);
2249         dout("do_request waiting\n");
2250         if (req->r_timeout) {
2251                 err = (long)wait_for_completion_killable_timeout(
2252                         &req->r_completion, req->r_timeout);
2253                 if (err == 0)
2254                         err = -EIO;
2255         } else if (req->r_wait_for_completion) {
2256                 err = req->r_wait_for_completion(mdsc, req);
2257         } else {
2258                 err = wait_for_completion_killable(&req->r_completion);
2259         }
2260         dout("do_request waited, got %d\n", err);
2261         mutex_lock(&mdsc->mutex);
2262
2263         /* only abort if we didn't race with a real reply */
2264         if (req->r_got_result) {
2265                 err = le32_to_cpu(req->r_reply_info.head->result);
2266         } else if (err < 0) {
2267                 dout("aborted request %lld with %d\n", req->r_tid, err);
2268
2269                 /*
2270                  * ensure we aren't running concurrently with
2271                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2272                  * rely on locks (dir mutex) held by our caller.
2273                  */
2274                 mutex_lock(&req->r_fill_mutex);
2275                 req->r_err = err;
2276                 req->r_aborted = true;
2277                 mutex_unlock(&req->r_fill_mutex);
2278
2279                 if (req->r_locked_dir &&
2280                     (req->r_op & CEPH_MDS_OP_WRITE))
2281                         ceph_invalidate_dir_request(req);
2282         } else {
2283                 err = req->r_err;
2284         }
2285
2286 out:
2287         mutex_unlock(&mdsc->mutex);
2288         dout("do_request %p done, result %d\n", req, err);
2289         return err;
2290 }
2291
2292 /*
2293  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2294  * namespace request.
2295  */
2296 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2297 {
2298         struct inode *inode = req->r_locked_dir;
2299
2300         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2301
2302         ceph_dir_clear_complete(inode);
2303         if (req->r_dentry)
2304                 ceph_invalidate_dentry_lease(req->r_dentry);
2305         if (req->r_old_dentry)
2306                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2307 }
2308
2309 /*
2310  * Handle mds reply.
2311  *
2312  * We take the session mutex and parse and process the reply immediately.
2313  * This preserves the logical ordering of replies, capabilities, etc., sent
2314  * by the MDS as they are applied to our local cache.
2315  */
2316 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2317 {
2318         struct ceph_mds_client *mdsc = session->s_mdsc;
2319         struct ceph_mds_request *req;
2320         struct ceph_mds_reply_head *head = msg->front.iov_base;
2321         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2322         struct ceph_snap_realm *realm;
2323         u64 tid;
2324         int err, result;
2325         int mds = session->s_mds;
2326
2327         if (msg->front.iov_len < sizeof(*head)) {
2328                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2329                 ceph_msg_dump(msg);
2330                 return;
2331         }
2332
2333         /* get request, session */
2334         tid = le64_to_cpu(msg->hdr.tid);
2335         mutex_lock(&mdsc->mutex);
2336         req = __lookup_request(mdsc, tid);
2337         if (!req) {
2338                 dout("handle_reply on unknown tid %llu\n", tid);
2339                 mutex_unlock(&mdsc->mutex);
2340                 return;
2341         }
2342         dout("handle_reply %p\n", req);
2343
2344         /* correct session? */
2345         if (req->r_session != session) {
2346                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2347                        " not mds%d\n", tid, session->s_mds,
2348                        req->r_session ? req->r_session->s_mds : -1);
2349                 mutex_unlock(&mdsc->mutex);
2350                 goto out;
2351         }
2352
2353         /* dup? */
2354         if ((req->r_got_unsafe && !head->safe) ||
2355             (req->r_got_safe && head->safe)) {
2356                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2357                            head->safe ? "safe" : "unsafe", tid, mds);
2358                 mutex_unlock(&mdsc->mutex);
2359                 goto out;
2360         }
2361         if (req->r_got_safe && !head->safe) {
2362                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2363                            tid, mds);
2364                 mutex_unlock(&mdsc->mutex);
2365                 goto out;
2366         }
2367
2368         result = le32_to_cpu(head->result);
2369
2370         /*
2371          * Handle an ESTALE
2372          * if we're not talking to the authority, send to them
2373          * if the authority has changed while we weren't looking,
2374          * send to new authority
2375          * Otherwise we just have to return an ESTALE
2376          */
2377         if (result == -ESTALE) {
2378                 dout("got ESTALE on request %llu", req->r_tid);
2379                 req->r_resend_mds = -1;
2380                 if (req->r_direct_mode != USE_AUTH_MDS) {
2381                         dout("not using auth, setting for that now");
2382                         req->r_direct_mode = USE_AUTH_MDS;
2383                         __do_request(mdsc, req);
2384                         mutex_unlock(&mdsc->mutex);
2385                         goto out;
2386                 } else  {
2387                         int mds = __choose_mds(mdsc, req);
2388                         if (mds >= 0 && mds != req->r_session->s_mds) {
2389                                 dout("but auth changed, so resending");
2390                                 __do_request(mdsc, req);
2391                                 mutex_unlock(&mdsc->mutex);
2392                                 goto out;
2393                         }
2394                 }
2395                 dout("have to return ESTALE on request %llu", req->r_tid);
2396         }
2397
2398
2399         if (head->safe) {
2400                 req->r_got_safe = true;
2401                 __unregister_request(mdsc, req);
2402
2403                 if (req->r_got_unsafe) {
2404                         /*
2405                          * We already handled the unsafe response, now do the
2406                          * cleanup.  No need to examine the response; the MDS
2407                          * doesn't include any result info in the safe
2408                          * response.  And even if it did, there is nothing
2409                          * useful we could do with a revised return value.
2410                          */
2411                         dout("got safe reply %llu, mds%d\n", tid, mds);
2412                         list_del_init(&req->r_unsafe_item);
2413
2414                         /* last unsafe request during umount? */
2415                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2416                                 complete_all(&mdsc->safe_umount_waiters);
2417                         mutex_unlock(&mdsc->mutex);
2418                         goto out;
2419                 }
2420         } else {
2421                 req->r_got_unsafe = true;
2422                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2423         }
2424
2425         dout("handle_reply tid %lld result %d\n", tid, result);
2426         rinfo = &req->r_reply_info;
2427         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2428         mutex_unlock(&mdsc->mutex);
2429
2430         mutex_lock(&session->s_mutex);
2431         if (err < 0) {
2432                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2433                 ceph_msg_dump(msg);
2434                 goto out_err;
2435         }
2436
2437         /* snap trace */
2438         realm = NULL;
2439         if (rinfo->snapblob_len) {
2440                 down_write(&mdsc->snap_rwsem);
2441                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2442                                 rinfo->snapblob + rinfo->snapblob_len,
2443                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2444                                 &realm);
2445                 downgrade_write(&mdsc->snap_rwsem);
2446         } else {
2447                 down_read(&mdsc->snap_rwsem);
2448         }
2449
2450         /* insert trace into our cache */
2451         mutex_lock(&req->r_fill_mutex);
2452         err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2453         if (err == 0) {
2454                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2455                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2456                         ceph_readdir_prepopulate(req, req->r_session);
2457                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2458         }
2459         mutex_unlock(&req->r_fill_mutex);
2460
2461         up_read(&mdsc->snap_rwsem);
2462         if (realm)
2463                 ceph_put_snap_realm(mdsc, realm);
2464 out_err:
2465         mutex_lock(&mdsc->mutex);
2466         if (!req->r_aborted) {
2467                 if (err) {
2468                         req->r_err = err;
2469                 } else {
2470                         req->r_reply = msg;
2471                         ceph_msg_get(msg);
2472                         req->r_got_result = true;
2473                 }
2474         } else {
2475                 dout("reply arrived after request %lld was aborted\n", tid);
2476         }
2477         mutex_unlock(&mdsc->mutex);
2478
2479         ceph_add_cap_releases(mdsc, req->r_session);
2480         mutex_unlock(&session->s_mutex);
2481
2482         /* kick calling process */
2483         complete_request(mdsc, req);
2484 out:
2485         ceph_mdsc_put_request(req);
2486         return;
2487 }
2488
2489
2490
2491 /*
2492  * handle mds notification that our request has been forwarded.
2493  */
2494 static void handle_forward(struct ceph_mds_client *mdsc,
2495                            struct ceph_mds_session *session,
2496                            struct ceph_msg *msg)
2497 {
2498         struct ceph_mds_request *req;
2499         u64 tid = le64_to_cpu(msg->hdr.tid);
2500         u32 next_mds;
2501         u32 fwd_seq;
2502         int err = -EINVAL;
2503         void *p = msg->front.iov_base;
2504         void *end = p + msg->front.iov_len;
2505
2506         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2507         next_mds = ceph_decode_32(&p);
2508         fwd_seq = ceph_decode_32(&p);
2509
2510         mutex_lock(&mdsc->mutex);
2511         req = __lookup_request(mdsc, tid);
2512         if (!req) {
2513                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2514                 goto out;  /* dup reply? */
2515         }
2516
2517         if (req->r_aborted) {
2518                 dout("forward tid %llu aborted, unregistering\n", tid);
2519                 __unregister_request(mdsc, req);
2520         } else if (fwd_seq <= req->r_num_fwd) {
2521                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2522                      tid, next_mds, req->r_num_fwd, fwd_seq);
2523         } else {
2524                 /* resend. forward race not possible; mds would drop */
2525                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2526                 BUG_ON(req->r_err);
2527                 BUG_ON(req->r_got_result);
2528                 req->r_attempts = 0;
2529                 req->r_num_fwd = fwd_seq;
2530                 req->r_resend_mds = next_mds;
2531                 put_request_session(req);
2532                 __do_request(mdsc, req);
2533         }
2534         ceph_mdsc_put_request(req);
2535 out:
2536         mutex_unlock(&mdsc->mutex);
2537         return;
2538
2539 bad:
2540         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2541 }
2542
2543 /*
2544  * handle a mds session control message
2545  */
2546 static void handle_session(struct ceph_mds_session *session,
2547                            struct ceph_msg *msg)
2548 {
2549         struct ceph_mds_client *mdsc = session->s_mdsc;
2550         u32 op;
2551         u64 seq;
2552         int mds = session->s_mds;
2553         struct ceph_mds_session_head *h = msg->front.iov_base;
2554         int wake = 0;
2555
2556         /* decode */
2557         if (msg->front.iov_len != sizeof(*h))
2558                 goto bad;
2559         op = le32_to_cpu(h->op);
2560         seq = le64_to_cpu(h->seq);
2561
2562         mutex_lock(&mdsc->mutex);
2563         if (op == CEPH_SESSION_CLOSE)
2564                 __unregister_session(mdsc, session);
2565         /* FIXME: this ttl calculation is generous */
2566         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2567         mutex_unlock(&mdsc->mutex);
2568
2569         mutex_lock(&session->s_mutex);
2570
2571         dout("handle_session mds%d %s %p state %s seq %llu\n",
2572              mds, ceph_session_op_name(op), session,
2573              ceph_session_state_name(session->s_state), seq);
2574
2575         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2576                 session->s_state = CEPH_MDS_SESSION_OPEN;
2577                 pr_info("mds%d came back\n", session->s_mds);
2578         }
2579
2580         switch (op) {
2581         case CEPH_SESSION_OPEN:
2582                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2583                         pr_info("mds%d reconnect success\n", session->s_mds);
2584                 session->s_state = CEPH_MDS_SESSION_OPEN;
2585                 renewed_caps(mdsc, session, 0);
2586                 wake = 1;
2587                 if (mdsc->stopping)
2588                         __close_session(mdsc, session);
2589                 break;
2590
2591         case CEPH_SESSION_RENEWCAPS:
2592                 if (session->s_renew_seq == seq)
2593                         renewed_caps(mdsc, session, 1);
2594                 break;
2595
2596         case CEPH_SESSION_CLOSE:
2597                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2598                         pr_info("mds%d reconnect denied\n", session->s_mds);
2599                 remove_session_caps(session);
2600                 wake = 2; /* for good measure */
2601                 wake_up_all(&mdsc->session_close_wq);
2602                 break;
2603
2604         case CEPH_SESSION_STALE:
2605                 pr_info("mds%d caps went stale, renewing\n",
2606                         session->s_mds);
2607                 spin_lock(&session->s_gen_ttl_lock);
2608                 session->s_cap_gen++;
2609                 session->s_cap_ttl = jiffies - 1;
2610                 spin_unlock(&session->s_gen_ttl_lock);
2611                 send_renew_caps(mdsc, session);
2612                 break;
2613
2614         case CEPH_SESSION_RECALL_STATE:
2615                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2616                 break;
2617
2618         case CEPH_SESSION_FLUSHMSG:
2619                 send_flushmsg_ack(mdsc, session, seq);
2620                 break;
2621
2622         case CEPH_SESSION_FORCE_RO:
2623                 dout("force_session_readonly %p\n", session);
2624                 spin_lock(&session->s_cap_lock);
2625                 session->s_readonly = true;
2626                 spin_unlock(&session->s_cap_lock);
2627                 wake_up_session_caps(session, 0);
2628                 break;
2629
2630         default:
2631                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2632                 WARN_ON(1);
2633         }
2634
2635         mutex_unlock(&session->s_mutex);
2636         if (wake) {
2637                 mutex_lock(&mdsc->mutex);
2638                 __wake_requests(mdsc, &session->s_waiting);
2639                 if (wake == 2)
2640                         kick_requests(mdsc, mds);
2641                 mutex_unlock(&mdsc->mutex);
2642         }
2643         return;
2644
2645 bad:
2646         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2647                (int)msg->front.iov_len);
2648         ceph_msg_dump(msg);
2649         return;
2650 }
2651
2652
2653 /*
2654  * called under session->mutex.
2655  */
2656 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2657                                    struct ceph_mds_session *session)
2658 {
2659         struct ceph_mds_request *req, *nreq;
2660         struct rb_node *p;
2661         int err;
2662
2663         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2664
2665         mutex_lock(&mdsc->mutex);
2666         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2667                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2668                 if (!err) {
2669                         ceph_msg_get(req->r_request);
2670                         ceph_con_send(&session->s_con, req->r_request);
2671                 }
2672         }
2673
2674         /*
2675          * also re-send old requests when MDS enters reconnect stage. So that MDS
2676          * can process completed request in clientreplay stage.
2677          */
2678         p = rb_first(&mdsc->request_tree);
2679         while (p) {
2680                 req = rb_entry(p, struct ceph_mds_request, r_node);
2681                 p = rb_next(p);
2682                 if (req->r_got_unsafe)
2683                         continue;
2684                 if (req->r_attempts == 0)
2685                         continue; /* only old requests */
2686                 if (req->r_session &&
2687                     req->r_session->s_mds == session->s_mds) {
2688                         err = __prepare_send_request(mdsc, req,
2689                                                      session->s_mds, true);
2690                         if (!err) {
2691                                 ceph_msg_get(req->r_request);
2692                                 ceph_con_send(&session->s_con, req->r_request);
2693                         }
2694                 }
2695         }
2696         mutex_unlock(&mdsc->mutex);
2697 }
2698
2699 /*
2700  * Encode information about a cap for a reconnect with the MDS.
2701  */
2702 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2703                           void *arg)
2704 {
2705         union {
2706                 struct ceph_mds_cap_reconnect v2;
2707                 struct ceph_mds_cap_reconnect_v1 v1;
2708         } rec;
2709         size_t reclen;
2710         struct ceph_inode_info *ci;
2711         struct ceph_reconnect_state *recon_state = arg;
2712         struct ceph_pagelist *pagelist = recon_state->pagelist;
2713         char *path;
2714         int pathlen, err;
2715         u64 pathbase;
2716         struct dentry *dentry;
2717
2718         ci = cap->ci;
2719
2720         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2721              inode, ceph_vinop(inode), cap, cap->cap_id,
2722              ceph_cap_string(cap->issued));
2723         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2724         if (err)
2725                 return err;
2726
2727         dentry = d_find_alias(inode);
2728         if (dentry) {
2729                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2730                 if (IS_ERR(path)) {
2731                         err = PTR_ERR(path);
2732                         goto out_dput;
2733                 }
2734         } else {
2735                 path = NULL;
2736                 pathlen = 0;
2737         }
2738         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2739         if (err)
2740                 goto out_free;
2741
2742         spin_lock(&ci->i_ceph_lock);
2743         cap->seq = 0;        /* reset cap seq */
2744         cap->issue_seq = 0;  /* and issue_seq */
2745         cap->mseq = 0;       /* and migrate_seq */
2746         cap->cap_gen = cap->session->s_cap_gen;
2747
2748         if (recon_state->flock) {
2749                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2750                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2751                 rec.v2.issued = cpu_to_le32(cap->issued);
2752                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2753                 rec.v2.pathbase = cpu_to_le64(pathbase);
2754                 rec.v2.flock_len = 0;
2755                 reclen = sizeof(rec.v2);
2756         } else {
2757                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2758                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2759                 rec.v1.issued = cpu_to_le32(cap->issued);
2760                 rec.v1.size = cpu_to_le64(inode->i_size);
2761                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2762                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2763                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2764                 rec.v1.pathbase = cpu_to_le64(pathbase);
2765                 reclen = sizeof(rec.v1);
2766         }
2767         spin_unlock(&ci->i_ceph_lock);
2768
2769         if (recon_state->flock) {
2770                 int num_fcntl_locks, num_flock_locks;
2771                 struct ceph_filelock *flocks;
2772
2773 encode_again:
2774                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2775                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2776                                  sizeof(struct ceph_filelock), GFP_NOFS);
2777                 if (!flocks) {
2778                         err = -ENOMEM;
2779                         goto out_free;
2780                 }
2781                 err = ceph_encode_locks_to_buffer(inode, flocks,
2782                                                   num_fcntl_locks,
2783                                                   num_flock_locks);
2784                 if (err) {
2785                         kfree(flocks);
2786                         if (err == -ENOSPC)
2787                                 goto encode_again;
2788                         goto out_free;
2789                 }
2790                 /*
2791                  * number of encoded locks is stable, so copy to pagelist
2792                  */
2793                 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2794                                     (num_fcntl_locks+num_flock_locks) *
2795                                     sizeof(struct ceph_filelock));
2796                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2797                 if (!err)
2798                         err = ceph_locks_to_pagelist(flocks, pagelist,
2799                                                      num_fcntl_locks,
2800                                                      num_flock_locks);
2801                 kfree(flocks);
2802         } else {
2803                 err = ceph_pagelist_append(pagelist, &rec, reclen);
2804         }
2805
2806         recon_state->nr_caps++;
2807 out_free:
2808         kfree(path);
2809 out_dput:
2810         dput(dentry);
2811         return err;
2812 }
2813
2814
2815 /*
2816  * If an MDS fails and recovers, clients need to reconnect in order to
2817  * reestablish shared state.  This includes all caps issued through
2818  * this session _and_ the snap_realm hierarchy.  Because it's not
2819  * clear which snap realms the mds cares about, we send everything we
2820  * know about.. that ensures we'll then get any new info the
2821  * recovering MDS might have.
2822  *
2823  * This is a relatively heavyweight operation, but it's rare.
2824  *
2825  * called with mdsc->mutex held.
2826  */
2827 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2828                                struct ceph_mds_session *session)
2829 {
2830         struct ceph_msg *reply;
2831         struct rb_node *p;
2832         int mds = session->s_mds;
2833         int err = -ENOMEM;
2834         int s_nr_caps;
2835         struct ceph_pagelist *pagelist;
2836         struct ceph_reconnect_state recon_state;
2837
2838         pr_info("mds%d reconnect start\n", mds);
2839
2840         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2841         if (!pagelist)
2842                 goto fail_nopagelist;
2843         ceph_pagelist_init(pagelist);
2844
2845         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2846         if (!reply)
2847                 goto fail_nomsg;
2848
2849         mutex_lock(&session->s_mutex);
2850         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2851         session->s_seq = 0;
2852
2853         dout("session %p state %s\n", session,
2854              ceph_session_state_name(session->s_state));
2855
2856         spin_lock(&session->s_gen_ttl_lock);
2857         session->s_cap_gen++;
2858         spin_unlock(&session->s_gen_ttl_lock);
2859
2860         spin_lock(&session->s_cap_lock);
2861         /* don't know if session is readonly */
2862         session->s_readonly = 0;
2863         /*
2864          * notify __ceph_remove_cap() that we are composing cap reconnect.
2865          * If a cap get released before being added to the cap reconnect,
2866          * __ceph_remove_cap() should skip queuing cap release.
2867          */
2868         session->s_cap_reconnect = 1;
2869         /* drop old cap expires; we're about to reestablish that state */
2870         discard_cap_releases(mdsc, session);
2871         spin_unlock(&session->s_cap_lock);
2872
2873         /* trim unused caps to reduce MDS's cache rejoin time */
2874         shrink_dcache_parent(mdsc->fsc->sb->s_root);
2875
2876         ceph_con_close(&session->s_con);
2877         ceph_con_open(&session->s_con,
2878                       CEPH_ENTITY_TYPE_MDS, mds,
2879                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2880
2881         /* replay unsafe requests */
2882         replay_unsafe_requests(mdsc, session);
2883
2884         down_read(&mdsc->snap_rwsem);
2885
2886         /* traverse this session's caps */
2887         s_nr_caps = session->s_nr_caps;
2888         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2889         if (err)
2890                 goto fail;
2891
2892         recon_state.nr_caps = 0;
2893         recon_state.pagelist = pagelist;
2894         recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2895         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2896         if (err < 0)
2897                 goto fail;
2898
2899         spin_lock(&session->s_cap_lock);
2900         session->s_cap_reconnect = 0;
2901         spin_unlock(&session->s_cap_lock);
2902
2903         /*
2904          * snaprealms.  we provide mds with the ino, seq (version), and
2905          * parent for all of our realms.  If the mds has any newer info,
2906          * it will tell us.
2907          */
2908         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2909                 struct ceph_snap_realm *realm =
2910                         rb_entry(p, struct ceph_snap_realm, node);
2911                 struct ceph_mds_snaprealm_reconnect sr_rec;
2912
2913                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2914                      realm->ino, realm->seq, realm->parent_ino);
2915                 sr_rec.ino = cpu_to_le64(realm->ino);
2916                 sr_rec.seq = cpu_to_le64(realm->seq);
2917                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2918                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2919                 if (err)
2920                         goto fail;
2921         }
2922
2923         if (recon_state.flock)
2924                 reply->hdr.version = cpu_to_le16(2);
2925
2926         /* raced with cap release? */
2927         if (s_nr_caps != recon_state.nr_caps) {
2928                 struct page *page = list_first_entry(&pagelist->head,
2929                                                      struct page, lru);
2930                 __le32 *addr = kmap_atomic(page);
2931                 *addr = cpu_to_le32(recon_state.nr_caps);
2932                 kunmap_atomic(addr);
2933         }
2934
2935         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2936         ceph_msg_data_add_pagelist(reply, pagelist);
2937         ceph_con_send(&session->s_con, reply);
2938
2939         mutex_unlock(&session->s_mutex);
2940
2941         mutex_lock(&mdsc->mutex);
2942         __wake_requests(mdsc, &session->s_waiting);
2943         mutex_unlock(&mdsc->mutex);
2944
2945         up_read(&mdsc->snap_rwsem);
2946         return;
2947
2948 fail:
2949         ceph_msg_put(reply);
2950         up_read(&mdsc->snap_rwsem);
2951         mutex_unlock(&session->s_mutex);
2952 fail_nomsg:
2953         ceph_pagelist_release(pagelist);
2954 fail_nopagelist:
2955         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2956         return;
2957 }
2958
2959
2960 /*
2961  * compare old and new mdsmaps, kicking requests
2962  * and closing out old connections as necessary
2963  *
2964  * called under mdsc->mutex.
2965  */
2966 static void check_new_map(struct ceph_mds_client *mdsc,
2967                           struct ceph_mdsmap *newmap,
2968                           struct ceph_mdsmap *oldmap)
2969 {
2970         int i;
2971         int oldstate, newstate;
2972         struct ceph_mds_session *s;
2973
2974         dout("check_new_map new %u old %u\n",
2975              newmap->m_epoch, oldmap->m_epoch);
2976
2977         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2978                 if (mdsc->sessions[i] == NULL)
2979                         continue;
2980                 s = mdsc->sessions[i];
2981                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2982                 newstate = ceph_mdsmap_get_state(newmap, i);
2983
2984                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2985                      i, ceph_mds_state_name(oldstate),
2986                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2987                      ceph_mds_state_name(newstate),
2988                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2989                      ceph_session_state_name(s->s_state));
2990
2991                 if (i >= newmap->m_max_mds ||
2992                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
2993                            ceph_mdsmap_get_addr(newmap, i),
2994                            sizeof(struct ceph_entity_addr))) {
2995                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2996                                 /* the session never opened, just close it
2997                                  * out now */
2998                                 __wake_requests(mdsc, &s->s_waiting);
2999                                 __unregister_session(mdsc, s);
3000                         } else {
3001                                 /* just close it */
3002                                 mutex_unlock(&mdsc->mutex);
3003                                 mutex_lock(&s->s_mutex);
3004                                 mutex_lock(&mdsc->mutex);
3005                                 ceph_con_close(&s->s_con);
3006                                 mutex_unlock(&s->s_mutex);
3007                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3008                         }
3009                 } else if (oldstate == newstate) {
3010                         continue;  /* nothing new with this mds */
3011                 }
3012
3013                 /*
3014                  * send reconnect?
3015                  */
3016                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3017                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3018                         mutex_unlock(&mdsc->mutex);
3019                         send_mds_reconnect(mdsc, s);
3020                         mutex_lock(&mdsc->mutex);
3021                 }
3022
3023                 /*
3024                  * kick request on any mds that has gone active.
3025                  */
3026                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3027                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3028                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3029                             oldstate != CEPH_MDS_STATE_STARTING)
3030                                 pr_info("mds%d recovery completed\n", s->s_mds);
3031                         kick_requests(mdsc, i);
3032                         ceph_kick_flushing_caps(mdsc, s);
3033                         wake_up_session_caps(s, 1);
3034                 }
3035         }
3036
3037         for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3038                 s = mdsc->sessions[i];
3039                 if (!s)
3040                         continue;
3041                 if (!ceph_mdsmap_is_laggy(newmap, i))
3042                         continue;
3043                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3044                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3045                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3046                         dout(" connecting to export targets of laggy mds%d\n",
3047                              i);
3048                         __open_export_target_sessions(mdsc, s);
3049                 }
3050         }
3051 }
3052
3053
3054
3055 /*
3056  * leases
3057  */
3058
3059 /*
3060  * caller must hold session s_mutex, dentry->d_lock
3061  */
3062 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3063 {
3064         struct ceph_dentry_info *di = ceph_dentry(dentry);
3065
3066         ceph_put_mds_session(di->lease_session);
3067         di->lease_session = NULL;
3068 }
3069
3070 static void handle_lease(struct ceph_mds_client *mdsc,
3071                          struct ceph_mds_session *session,
3072                          struct ceph_msg *msg)
3073 {
3074         struct super_block *sb = mdsc->fsc->sb;
3075         struct inode *inode;
3076         struct dentry *parent, *dentry;
3077         struct ceph_dentry_info *di;
3078         int mds = session->s_mds;
3079         struct ceph_mds_lease *h = msg->front.iov_base;
3080         u32 seq;
3081         struct ceph_vino vino;
3082         struct qstr dname;
3083         int release = 0;
3084
3085         dout("handle_lease from mds%d\n", mds);
3086
3087         /* decode */
3088         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3089                 goto bad;
3090         vino.ino = le64_to_cpu(h->ino);
3091         vino.snap = CEPH_NOSNAP;
3092         seq = le32_to_cpu(h->seq);
3093         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3094         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3095         if (dname.len != get_unaligned_le32(h+1))
3096                 goto bad;
3097
3098         /* lookup inode */
3099         inode = ceph_find_inode(sb, vino);
3100         dout("handle_lease %s, ino %llx %p %.*s\n",
3101              ceph_lease_op_name(h->action), vino.ino, inode,
3102              dname.len, dname.name);
3103
3104         mutex_lock(&session->s_mutex);
3105         session->s_seq++;
3106
3107         if (inode == NULL) {
3108                 dout("handle_lease no inode %llx\n", vino.ino);
3109                 goto release;
3110         }
3111
3112         /* dentry */
3113         parent = d_find_alias(inode);
3114         if (!parent) {
3115                 dout("no parent dentry on inode %p\n", inode);
3116                 WARN_ON(1);
3117                 goto release;  /* hrm... */
3118         }
3119         dname.hash = full_name_hash(dname.name, dname.len);
3120         dentry = d_lookup(parent, &dname);
3121         dput(parent);
3122         if (!dentry)
3123                 goto release;
3124
3125         spin_lock(&dentry->d_lock);
3126         di = ceph_dentry(dentry);
3127         switch (h->action) {
3128         case CEPH_MDS_LEASE_REVOKE:
3129                 if (di->lease_session == session) {
3130                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3131                                 h->seq = cpu_to_le32(di->lease_seq);
3132                         __ceph_mdsc_drop_dentry_lease(dentry);
3133                 }
3134                 release = 1;
3135                 break;
3136
3137         case CEPH_MDS_LEASE_RENEW:
3138                 if (di->lease_session == session &&
3139                     di->lease_gen == session->s_cap_gen &&
3140                     di->lease_renew_from &&
3141                     di->lease_renew_after == 0) {
3142                         unsigned long duration =
3143                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
3144
3145                         di->lease_seq = seq;
3146                         dentry->d_time = di->lease_renew_from + duration;
3147                         di->lease_renew_after = di->lease_renew_from +
3148                                 (duration >> 1);
3149                         di->lease_renew_from = 0;
3150                 }
3151                 break;
3152         }
3153         spin_unlock(&dentry->d_lock);
3154         dput(dentry);
3155
3156         if (!release)
3157                 goto out;
3158
3159 release:
3160         /* let's just reuse the same message */
3161         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3162         ceph_msg_get(msg);
3163         ceph_con_send(&session->s_con, msg);
3164
3165 out:
3166         iput(inode);
3167         mutex_unlock(&session->s_mutex);
3168         return;
3169
3170 bad:
3171         pr_err("corrupt lease message\n");
3172         ceph_msg_dump(msg);
3173 }
3174
3175 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3176                               struct inode *inode,
3177                               struct dentry *dentry, char action,
3178                               u32 seq)
3179 {
3180         struct ceph_msg *msg;
3181         struct ceph_mds_lease *lease;
3182         int len = sizeof(*lease) + sizeof(u32);
3183         int dnamelen = 0;
3184
3185         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3186              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3187         dnamelen = dentry->d_name.len;
3188         len += dnamelen;
3189
3190         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3191         if (!msg)
3192                 return;
3193         lease = msg->front.iov_base;
3194         lease->action = action;
3195         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3196         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3197         lease->seq = cpu_to_le32(seq);
3198         put_unaligned_le32(dnamelen, lease + 1);
3199         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3200
3201         /*
3202          * if this is a preemptive lease RELEASE, no need to
3203          * flush request stream, since the actual request will
3204          * soon follow.
3205          */
3206         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3207
3208         ceph_con_send(&session->s_con, msg);
3209 }
3210
3211 /*
3212  * Preemptively release a lease we expect to invalidate anyway.
3213  * Pass @inode always, @dentry is optional.
3214  */
3215 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3216                              struct dentry *dentry)
3217 {
3218         struct ceph_dentry_info *di;
3219         struct ceph_mds_session *session;
3220         u32 seq;
3221
3222         BUG_ON(inode == NULL);
3223         BUG_ON(dentry == NULL);
3224
3225         /* is dentry lease valid? */
3226         spin_lock(&dentry->d_lock);
3227         di = ceph_dentry(dentry);
3228         if (!di || !di->lease_session ||
3229             di->lease_session->s_mds < 0 ||
3230             di->lease_gen != di->lease_session->s_cap_gen ||
3231             !time_before(jiffies, dentry->d_time)) {
3232                 dout("lease_release inode %p dentry %p -- "
3233                      "no lease\n",
3234                      inode, dentry);
3235                 spin_unlock(&dentry->d_lock);
3236                 return;
3237         }
3238
3239         /* we do have a lease on this dentry; note mds and seq */
3240         session = ceph_get_mds_session(di->lease_session);
3241         seq = di->lease_seq;
3242         __ceph_mdsc_drop_dentry_lease(dentry);
3243         spin_unlock(&dentry->d_lock);
3244
3245         dout("lease_release inode %p dentry %p to mds%d\n",
3246              inode, dentry, session->s_mds);
3247         ceph_mdsc_lease_send_msg(session, inode, dentry,
3248                                  CEPH_MDS_LEASE_RELEASE, seq);
3249         ceph_put_mds_session(session);
3250 }
3251
3252 /*
3253  * drop all leases (and dentry refs) in preparation for umount
3254  */
3255 static void drop_leases(struct ceph_mds_client *mdsc)
3256 {
3257         int i;
3258
3259         dout("drop_leases\n");
3260         mutex_lock(&mdsc->mutex);
3261         for (i = 0; i < mdsc->max_sessions; i++) {
3262                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3263                 if (!s)
3264                         continue;
3265                 mutex_unlock(&mdsc->mutex);
3266                 mutex_lock(&s->s_mutex);
3267                 mutex_unlock(&s->s_mutex);
3268                 ceph_put_mds_session(s);
3269                 mutex_lock(&mdsc->mutex);
3270         }
3271         mutex_unlock(&mdsc->mutex);
3272 }
3273
3274
3275
3276 /*
3277  * delayed work -- periodically trim expired leases, renew caps with mds
3278  */
3279 static void schedule_delayed(struct ceph_mds_client *mdsc)
3280 {
3281         int delay = 5;
3282         unsigned hz = round_jiffies_relative(HZ * delay);
3283         schedule_delayed_work(&mdsc->delayed_work, hz);
3284 }
3285
3286 static void delayed_work(struct work_struct *work)
3287 {
3288         int i;
3289         struct ceph_mds_client *mdsc =
3290                 container_of(work, struct ceph_mds_client, delayed_work.work);
3291         int renew_interval;
3292         int renew_caps;
3293
3294         dout("mdsc delayed_work\n");
3295         ceph_check_delayed_caps(mdsc);
3296
3297         mutex_lock(&mdsc->mutex);
3298         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3299         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3300                                    mdsc->last_renew_caps);
3301         if (renew_caps)
3302                 mdsc->last_renew_caps = jiffies;
3303
3304         for (i = 0; i < mdsc->max_sessions; i++) {
3305                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3306                 if (s == NULL)
3307                         continue;
3308                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3309                         dout("resending session close request for mds%d\n",
3310                              s->s_mds);
3311                         request_close_session(mdsc, s);
3312                         ceph_put_mds_session(s);
3313                         continue;
3314                 }
3315                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3316                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3317                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3318                                 pr_info("mds%d hung\n", s->s_mds);
3319                         }
3320                 }
3321                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3322                         /* this mds is failed or recovering, just wait */
3323                         ceph_put_mds_session(s);
3324                         continue;
3325                 }
3326                 mutex_unlock(&mdsc->mutex);
3327
3328                 mutex_lock(&s->s_mutex);
3329                 if (renew_caps)
3330                         send_renew_caps(mdsc, s);
3331                 else
3332                         ceph_con_keepalive(&s->s_con);
3333                 ceph_add_cap_releases(mdsc, s);
3334                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3335                     s->s_state == CEPH_MDS_SESSION_HUNG)
3336                         ceph_send_cap_releases(mdsc, s);
3337                 mutex_unlock(&s->s_mutex);
3338                 ceph_put_mds_session(s);
3339
3340                 mutex_lock(&mdsc->mutex);
3341         }
3342         mutex_unlock(&mdsc->mutex);
3343
3344         schedule_delayed(mdsc);
3345 }
3346
3347 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3348
3349 {
3350         struct ceph_mds_client *mdsc;
3351
3352         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3353         if (!mdsc)
3354                 return -ENOMEM;
3355         mdsc->fsc = fsc;
3356         fsc->mdsc = mdsc;
3357         mutex_init(&mdsc->mutex);
3358         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3359         if (mdsc->mdsmap == NULL) {
3360                 kfree(mdsc);
3361                 return -ENOMEM;
3362         }
3363
3364         init_completion(&mdsc->safe_umount_waiters);
3365         init_waitqueue_head(&mdsc->session_close_wq);
3366         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3367         mdsc->sessions = NULL;
3368         atomic_set(&mdsc->num_sessions, 0);
3369         mdsc->max_sessions = 0;
3370         mdsc->stopping = 0;
3371         init_rwsem(&mdsc->snap_rwsem);
3372         mdsc->snap_realms = RB_ROOT;
3373         INIT_LIST_HEAD(&mdsc->snap_empty);
3374         spin_lock_init(&mdsc->snap_empty_lock);
3375         mdsc->last_tid = 0;
3376         mdsc->request_tree = RB_ROOT;
3377         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3378         mdsc->last_renew_caps = jiffies;
3379         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3380         spin_lock_init(&mdsc->cap_delay_lock);
3381         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3382         spin_lock_init(&mdsc->snap_flush_lock);
3383         mdsc->cap_flush_seq = 0;
3384         INIT_LIST_HEAD(&mdsc->cap_dirty);
3385         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3386         mdsc->num_cap_flushing = 0;
3387         spin_lock_init(&mdsc->cap_dirty_lock);
3388         init_waitqueue_head(&mdsc->cap_flushing_wq);
3389         spin_lock_init(&mdsc->dentry_lru_lock);
3390         INIT_LIST_HEAD(&mdsc->dentry_lru);
3391
3392         ceph_caps_init(mdsc);
3393         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3394
3395         return 0;
3396 }
3397
3398 /*
3399  * Wait for safe replies on open mds requests.  If we time out, drop
3400  * all requests from the tree to avoid dangling dentry refs.
3401  */
3402 static void wait_requests(struct ceph_mds_client *mdsc)
3403 {
3404         struct ceph_mds_request *req;
3405         struct ceph_fs_client *fsc = mdsc->fsc;
3406
3407         mutex_lock(&mdsc->mutex);
3408         if (__get_oldest_req(mdsc)) {
3409                 mutex_unlock(&mdsc->mutex);
3410
3411                 dout("wait_requests waiting for requests\n");
3412                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3413                                     fsc->client->options->mount_timeout * HZ);
3414
3415                 /* tear down remaining requests */
3416                 mutex_lock(&mdsc->mutex);
3417                 while ((req = __get_oldest_req(mdsc))) {
3418                         dout("wait_requests timed out on tid %llu\n",
3419                              req->r_tid);
3420                         __unregister_request(mdsc, req);
3421                 }
3422         }
3423         mutex_unlock(&mdsc->mutex);
3424         dout("wait_requests done\n");
3425 }
3426
3427 /*
3428  * called before mount is ro, and before dentries are torn down.
3429  * (hmm, does this still race with new lookups?)
3430  */
3431 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3432 {
3433         dout("pre_umount\n");
3434         mdsc->stopping = 1;
3435
3436         drop_leases(mdsc);
3437         ceph_flush_dirty_caps(mdsc);
3438         wait_requests(mdsc);
3439
3440         /*
3441          * wait for reply handlers to drop their request refs and
3442          * their inode/dcache refs
3443          */
3444         ceph_msgr_flush();
3445 }
3446
3447 /*
3448  * wait for all write mds requests to flush.
3449  */
3450 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3451 {
3452         struct ceph_mds_request *req = NULL, *nextreq;
3453         struct rb_node *n;
3454
3455         mutex_lock(&mdsc->mutex);
3456         dout("wait_unsafe_requests want %lld\n", want_tid);
3457 restart:
3458         req = __get_oldest_req(mdsc);
3459         while (req && req->r_tid <= want_tid) {
3460                 /* find next request */
3461                 n = rb_next(&req->r_node);
3462                 if (n)
3463                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3464                 else
3465                         nextreq = NULL;
3466                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3467                         /* write op */
3468                         ceph_mdsc_get_request(req);
3469                         if (nextreq)
3470                                 ceph_mdsc_get_request(nextreq);
3471                         mutex_unlock(&mdsc->mutex);
3472                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3473                              req->r_tid, want_tid);
3474                         wait_for_completion(&req->r_safe_completion);
3475                         mutex_lock(&mdsc->mutex);
3476                         ceph_mdsc_put_request(req);
3477                         if (!nextreq)
3478                                 break;  /* next dne before, so we're done! */
3479                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3480                                 /* next request was removed from tree */
3481                                 ceph_mdsc_put_request(nextreq);
3482                                 goto restart;
3483                         }
3484                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3485                 }
3486                 req = nextreq;
3487         }
3488         mutex_unlock(&mdsc->mutex);
3489         dout("wait_unsafe_requests done\n");
3490 }
3491
3492 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3493 {
3494         u64 want_tid, want_flush;
3495
3496         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3497                 return;
3498
3499         dout("sync\n");
3500         mutex_lock(&mdsc->mutex);
3501         want_tid = mdsc->last_tid;
3502         mutex_unlock(&mdsc->mutex);
3503
3504         ceph_flush_dirty_caps(mdsc);
3505         spin_lock(&mdsc->cap_dirty_lock);
3506         want_flush = mdsc->cap_flush_seq;
3507         spin_unlock(&mdsc->cap_dirty_lock);
3508
3509         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3510
3511         wait_unsafe_requests(mdsc, want_tid);
3512         wait_caps_flush(mdsc, want_flush);
3513 }
3514
3515 /*
3516  * true if all sessions are closed, or we force unmount
3517  */
3518 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3519 {
3520         if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3521                 return true;
3522         return atomic_read(&mdsc->num_sessions) == 0;
3523 }
3524
3525 /*
3526  * called after sb is ro.
3527  */
3528 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3529 {
3530         struct ceph_mds_session *session;
3531         int i;
3532         struct ceph_fs_client *fsc = mdsc->fsc;
3533         unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3534
3535         dout("close_sessions\n");
3536
3537         /* close sessions */
3538         mutex_lock(&mdsc->mutex);
3539         for (i = 0; i < mdsc->max_sessions; i++) {
3540                 session = __ceph_lookup_mds_session(mdsc, i);
3541                 if (!session)
3542                         continue;
3543                 mutex_unlock(&mdsc->mutex);
3544                 mutex_lock(&session->s_mutex);
3545                 __close_session(mdsc, session);
3546                 mutex_unlock(&session->s_mutex);
3547                 ceph_put_mds_session(session);
3548                 mutex_lock(&mdsc->mutex);
3549         }
3550         mutex_unlock(&mdsc->mutex);
3551
3552         dout("waiting for sessions to close\n");
3553         wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3554                            timeout);
3555
3556         /* tear down remaining sessions */
3557         mutex_lock(&mdsc->mutex);
3558         for (i = 0; i < mdsc->max_sessions; i++) {
3559                 if (mdsc->sessions[i]) {
3560                         session = get_session(mdsc->sessions[i]);
3561                         __unregister_session(mdsc, session);
3562                         mutex_unlock(&mdsc->mutex);
3563                         mutex_lock(&session->s_mutex);
3564                         remove_session_caps(session);
3565                         mutex_unlock(&session->s_mutex);
3566                         ceph_put_mds_session(session);
3567                         mutex_lock(&mdsc->mutex);
3568                 }
3569         }
3570         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3571         mutex_unlock(&mdsc->mutex);
3572
3573         ceph_cleanup_empty_realms(mdsc);
3574
3575         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3576
3577         dout("stopped\n");
3578 }
3579
3580 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3581 {
3582         dout("stop\n");
3583         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3584         if (mdsc->mdsmap)
3585                 ceph_mdsmap_destroy(mdsc->mdsmap);
3586         kfree(mdsc->sessions);
3587         ceph_caps_finalize(mdsc);
3588 }
3589
3590 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3591 {
3592         struct ceph_mds_client *mdsc = fsc->mdsc;
3593
3594         dout("mdsc_destroy %p\n", mdsc);
3595         ceph_mdsc_stop(mdsc);
3596
3597         /* flush out any connection work with references to us */
3598         ceph_msgr_flush();
3599
3600         fsc->mdsc = NULL;
3601         kfree(mdsc);
3602         dout("mdsc_destroy %p done\n", mdsc);
3603 }
3604
3605
3606 /*
3607  * handle mds map update.
3608  */
3609 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3610 {
3611         u32 epoch;
3612         u32 maplen;
3613         void *p = msg->front.iov_base;
3614         void *end = p + msg->front.iov_len;
3615         struct ceph_mdsmap *newmap, *oldmap;
3616         struct ceph_fsid fsid;
3617         int err = -EINVAL;
3618
3619         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3620         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3621         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3622                 return;
3623         epoch = ceph_decode_32(&p);
3624         maplen = ceph_decode_32(&p);
3625         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3626
3627         /* do we need it? */
3628         ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3629         mutex_lock(&mdsc->mutex);
3630         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3631                 dout("handle_map epoch %u <= our %u\n",
3632                      epoch, mdsc->mdsmap->m_epoch);
3633                 mutex_unlock(&mdsc->mutex);
3634                 return;
3635         }
3636
3637         newmap = ceph_mdsmap_decode(&p, end);
3638         if (IS_ERR(newmap)) {
3639                 err = PTR_ERR(newmap);
3640                 goto bad_unlock;
3641         }
3642
3643         /* swap into place */
3644         if (mdsc->mdsmap) {
3645                 oldmap = mdsc->mdsmap;
3646                 mdsc->mdsmap = newmap;
3647                 check_new_map(mdsc, newmap, oldmap);
3648                 ceph_mdsmap_destroy(oldmap);
3649         } else {
3650                 mdsc->mdsmap = newmap;  /* first mds map */
3651         }
3652         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3653
3654         __wake_requests(mdsc, &mdsc->waiting_for_map);
3655
3656         mutex_unlock(&mdsc->mutex);
3657         schedule_delayed(mdsc);
3658         return;
3659
3660 bad_unlock:
3661         mutex_unlock(&mdsc->mutex);
3662 bad:
3663         pr_err("error decoding mdsmap %d\n", err);
3664         return;
3665 }
3666
3667 static struct ceph_connection *con_get(struct ceph_connection *con)
3668 {
3669         struct ceph_mds_session *s = con->private;
3670
3671         if (get_session(s)) {
3672                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3673                 return con;
3674         }
3675         dout("mdsc con_get %p FAIL\n", s);
3676         return NULL;
3677 }
3678
3679 static void con_put(struct ceph_connection *con)
3680 {
3681         struct ceph_mds_session *s = con->private;
3682
3683         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3684         ceph_put_mds_session(s);
3685 }
3686
3687 /*
3688  * if the client is unresponsive for long enough, the mds will kill
3689  * the session entirely.
3690  */
3691 static void peer_reset(struct ceph_connection *con)
3692 {
3693         struct ceph_mds_session *s = con->private;
3694         struct ceph_mds_client *mdsc = s->s_mdsc;
3695
3696         pr_warn("mds%d closed our session\n", s->s_mds);
3697         send_mds_reconnect(mdsc, s);
3698 }
3699
3700 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3701 {
3702         struct ceph_mds_session *s = con->private;
3703         struct ceph_mds_client *mdsc = s->s_mdsc;
3704         int type = le16_to_cpu(msg->hdr.type);
3705
3706         mutex_lock(&mdsc->mutex);
3707         if (__verify_registered_session(mdsc, s) < 0) {
3708                 mutex_unlock(&mdsc->mutex);
3709                 goto out;
3710         }
3711         mutex_unlock(&mdsc->mutex);
3712
3713         switch (type) {
3714         case CEPH_MSG_MDS_MAP:
3715                 ceph_mdsc_handle_map(mdsc, msg);
3716                 break;
3717         case CEPH_MSG_CLIENT_SESSION:
3718                 handle_session(s, msg);
3719                 break;
3720         case CEPH_MSG_CLIENT_REPLY:
3721                 handle_reply(s, msg);
3722                 break;
3723         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3724                 handle_forward(mdsc, s, msg);
3725                 break;
3726         case CEPH_MSG_CLIENT_CAPS:
3727                 ceph_handle_caps(s, msg);
3728                 break;
3729         case CEPH_MSG_CLIENT_SNAP:
3730                 ceph_handle_snap(mdsc, s, msg);
3731                 break;
3732         case CEPH_MSG_CLIENT_LEASE:
3733                 handle_lease(mdsc, s, msg);
3734                 break;
3735
3736         default:
3737                 pr_err("received unknown message type %d %s\n", type,
3738                        ceph_msg_type_name(type));
3739         }
3740 out:
3741         ceph_msg_put(msg);
3742 }
3743
3744 /*
3745  * authentication
3746  */
3747
3748 /*
3749  * Note: returned pointer is the address of a structure that's
3750  * managed separately.  Caller must *not* attempt to free it.
3751  */
3752 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3753                                         int *proto, int force_new)
3754 {
3755         struct ceph_mds_session *s = con->private;
3756         struct ceph_mds_client *mdsc = s->s_mdsc;
3757         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3758         struct ceph_auth_handshake *auth = &s->s_auth;
3759
3760         if (force_new && auth->authorizer) {
3761                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3762                 auth->authorizer = NULL;
3763         }
3764         if (!auth->authorizer) {
3765                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3766                                                       auth);
3767                 if (ret)
3768                         return ERR_PTR(ret);
3769         } else {
3770                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3771                                                       auth);
3772                 if (ret)
3773                         return ERR_PTR(ret);
3774         }
3775         *proto = ac->protocol;
3776
3777         return auth;
3778 }
3779
3780
3781 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3782 {
3783         struct ceph_mds_session *s = con->private;
3784         struct ceph_mds_client *mdsc = s->s_mdsc;
3785         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3786
3787         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3788 }
3789
3790 static int invalidate_authorizer(struct ceph_connection *con)
3791 {
3792         struct ceph_mds_session *s = con->private;
3793         struct ceph_mds_client *mdsc = s->s_mdsc;
3794         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3795
3796         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3797
3798         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3799 }
3800
3801 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3802                                 struct ceph_msg_header *hdr, int *skip)
3803 {
3804         struct ceph_msg *msg;
3805         int type = (int) le16_to_cpu(hdr->type);
3806         int front_len = (int) le32_to_cpu(hdr->front_len);
3807
3808         if (con->in_msg)
3809                 return con->in_msg;
3810
3811         *skip = 0;
3812         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3813         if (!msg) {
3814                 pr_err("unable to allocate msg type %d len %d\n",
3815                        type, front_len);
3816                 return NULL;
3817         }
3818
3819         return msg;
3820 }
3821
3822 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3823 {
3824        struct ceph_mds_session *s = con->private;
3825        struct ceph_auth_handshake *auth = &s->s_auth;
3826        return ceph_auth_sign_message(auth, msg);
3827 }
3828
3829 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3830 {
3831        struct ceph_mds_session *s = con->private;
3832        struct ceph_auth_handshake *auth = &s->s_auth;
3833        return ceph_auth_check_message_signature(auth, msg);
3834 }
3835
3836 static const struct ceph_connection_operations mds_con_ops = {
3837         .get = con_get,
3838         .put = con_put,
3839         .dispatch = dispatch,
3840         .get_authorizer = get_authorizer,
3841         .verify_authorizer_reply = verify_authorizer_reply,
3842         .invalidate_authorizer = invalidate_authorizer,
3843         .peer_reset = peer_reset,
3844         .alloc_msg = mds_alloc_msg,
3845         .sign_message = sign_message,
3846         .check_message_signature = check_message_signature,
3847 };
3848
3849 /* eof */