ceph: allow renewal of auth credentials
[linux-2.6.git] / fs / ceph / mds_client.c
1 #include "ceph_debug.h"
2
3 #include <linux/wait.h>
4 #include <linux/sched.h>
5
6 #include "mds_client.h"
7 #include "mon_client.h"
8 #include "super.h"
9 #include "messenger.h"
10 #include "decode.h"
11 #include "auth.h"
12 #include "pagelist.h"
13
14 /*
15  * A cluster of MDS (metadata server) daemons is responsible for
16  * managing the file system namespace (the directory hierarchy and
17  * inodes) and for coordinating shared access to storage.  Metadata is
18  * partitioning hierarchically across a number of servers, and that
19  * partition varies over time as the cluster adjusts the distribution
20  * in order to balance load.
21  *
22  * The MDS client is primarily responsible to managing synchronous
23  * metadata requests for operations like open, unlink, and so forth.
24  * If there is a MDS failure, we find out about it when we (possibly
25  * request and) receive a new MDS map, and can resubmit affected
26  * requests.
27  *
28  * For the most part, though, we take advantage of a lossless
29  * communications channel to the MDS, and do not need to worry about
30  * timing out or resubmitting requests.
31  *
32  * We maintain a stateful "session" with each MDS we interact with.
33  * Within each session, we sent periodic heartbeat messages to ensure
34  * any capabilities or leases we have been issues remain valid.  If
35  * the session times out and goes stale, our leases and capabilities
36  * are no longer valid.
37  */
38
39 static void __wake_requests(struct ceph_mds_client *mdsc,
40                             struct list_head *head);
41
42 const static struct ceph_connection_operations mds_con_ops;
43
44
45 /*
46  * mds reply parsing
47  */
48
49 /*
50  * parse individual inode info
51  */
52 static int parse_reply_info_in(void **p, void *end,
53                                struct ceph_mds_reply_info_in *info)
54 {
55         int err = -EIO;
56
57         info->in = *p;
58         *p += sizeof(struct ceph_mds_reply_inode) +
59                 sizeof(*info->in->fragtree.splits) *
60                 le32_to_cpu(info->in->fragtree.nsplits);
61
62         ceph_decode_32_safe(p, end, info->symlink_len, bad);
63         ceph_decode_need(p, end, info->symlink_len, bad);
64         info->symlink = *p;
65         *p += info->symlink_len;
66
67         ceph_decode_32_safe(p, end, info->xattr_len, bad);
68         ceph_decode_need(p, end, info->xattr_len, bad);
69         info->xattr_data = *p;
70         *p += info->xattr_len;
71         return 0;
72 bad:
73         return err;
74 }
75
76 /*
77  * parse a normal reply, which may contain a (dir+)dentry and/or a
78  * target inode.
79  */
80 static int parse_reply_info_trace(void **p, void *end,
81                                   struct ceph_mds_reply_info_parsed *info)
82 {
83         int err;
84
85         if (info->head->is_dentry) {
86                 err = parse_reply_info_in(p, end, &info->diri);
87                 if (err < 0)
88                         goto out_bad;
89
90                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
91                         goto bad;
92                 info->dirfrag = *p;
93                 *p += sizeof(*info->dirfrag) +
94                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
95                 if (unlikely(*p > end))
96                         goto bad;
97
98                 ceph_decode_32_safe(p, end, info->dname_len, bad);
99                 ceph_decode_need(p, end, info->dname_len, bad);
100                 info->dname = *p;
101                 *p += info->dname_len;
102                 info->dlease = *p;
103                 *p += sizeof(*info->dlease);
104         }
105
106         if (info->head->is_target) {
107                 err = parse_reply_info_in(p, end, &info->targeti);
108                 if (err < 0)
109                         goto out_bad;
110         }
111
112         if (unlikely(*p != end))
113                 goto bad;
114         return 0;
115
116 bad:
117         err = -EIO;
118 out_bad:
119         pr_err("problem parsing mds trace %d\n", err);
120         return err;
121 }
122
123 /*
124  * parse readdir results
125  */
126 static int parse_reply_info_dir(void **p, void *end,
127                                 struct ceph_mds_reply_info_parsed *info)
128 {
129         u32 num, i = 0;
130         int err;
131
132         info->dir_dir = *p;
133         if (*p + sizeof(*info->dir_dir) > end)
134                 goto bad;
135         *p += sizeof(*info->dir_dir) +
136                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
137         if (*p > end)
138                 goto bad;
139
140         ceph_decode_need(p, end, sizeof(num) + 2, bad);
141         num = ceph_decode_32(p);
142         info->dir_end = ceph_decode_8(p);
143         info->dir_complete = ceph_decode_8(p);
144         if (num == 0)
145                 goto done;
146
147         /* alloc large array */
148         info->dir_nr = num;
149         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
150                                sizeof(*info->dir_dname) +
151                                sizeof(*info->dir_dname_len) +
152                                sizeof(*info->dir_dlease),
153                                GFP_NOFS);
154         if (info->dir_in == NULL) {
155                 err = -ENOMEM;
156                 goto out_bad;
157         }
158         info->dir_dname = (void *)(info->dir_in + num);
159         info->dir_dname_len = (void *)(info->dir_dname + num);
160         info->dir_dlease = (void *)(info->dir_dname_len + num);
161
162         while (num) {
163                 /* dentry */
164                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
165                 info->dir_dname_len[i] = ceph_decode_32(p);
166                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
167                 info->dir_dname[i] = *p;
168                 *p += info->dir_dname_len[i];
169                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
170                      info->dir_dname[i]);
171                 info->dir_dlease[i] = *p;
172                 *p += sizeof(struct ceph_mds_reply_lease);
173
174                 /* inode */
175                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
176                 if (err < 0)
177                         goto out_bad;
178                 i++;
179                 num--;
180         }
181
182 done:
183         if (*p != end)
184                 goto bad;
185         return 0;
186
187 bad:
188         err = -EIO;
189 out_bad:
190         pr_err("problem parsing dir contents %d\n", err);
191         return err;
192 }
193
194 /*
195  * parse entire mds reply
196  */
197 static int parse_reply_info(struct ceph_msg *msg,
198                             struct ceph_mds_reply_info_parsed *info)
199 {
200         void *p, *end;
201         u32 len;
202         int err;
203
204         info->head = msg->front.iov_base;
205         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
206         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
207
208         /* trace */
209         ceph_decode_32_safe(&p, end, len, bad);
210         if (len > 0) {
211                 err = parse_reply_info_trace(&p, p+len, info);
212                 if (err < 0)
213                         goto out_bad;
214         }
215
216         /* dir content */
217         ceph_decode_32_safe(&p, end, len, bad);
218         if (len > 0) {
219                 err = parse_reply_info_dir(&p, p+len, info);
220                 if (err < 0)
221                         goto out_bad;
222         }
223
224         /* snap blob */
225         ceph_decode_32_safe(&p, end, len, bad);
226         info->snapblob_len = len;
227         info->snapblob = p;
228         p += len;
229
230         if (p != end)
231                 goto bad;
232         return 0;
233
234 bad:
235         err = -EIO;
236 out_bad:
237         pr_err("mds parse_reply err %d\n", err);
238         return err;
239 }
240
241 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
242 {
243         kfree(info->dir_in);
244 }
245
246
247 /*
248  * sessions
249  */
250 static const char *session_state_name(int s)
251 {
252         switch (s) {
253         case CEPH_MDS_SESSION_NEW: return "new";
254         case CEPH_MDS_SESSION_OPENING: return "opening";
255         case CEPH_MDS_SESSION_OPEN: return "open";
256         case CEPH_MDS_SESSION_HUNG: return "hung";
257         case CEPH_MDS_SESSION_CLOSING: return "closing";
258         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
259         default: return "???";
260         }
261 }
262
263 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
264 {
265         if (atomic_inc_not_zero(&s->s_ref)) {
266                 dout("mdsc get_session %p %d -> %d\n", s,
267                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
268                 return s;
269         } else {
270                 dout("mdsc get_session %p 0 -- FAIL", s);
271                 return NULL;
272         }
273 }
274
275 void ceph_put_mds_session(struct ceph_mds_session *s)
276 {
277         dout("mdsc put_session %p %d -> %d\n", s,
278              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
279         if (atomic_dec_and_test(&s->s_ref)) {
280                 if (s->s_authorizer)
281                         s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
282                                 s->s_mdsc->client->monc.auth, s->s_authorizer);
283                 kfree(s);
284         }
285 }
286
287 /*
288  * called under mdsc->mutex
289  */
290 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
291                                                    int mds)
292 {
293         struct ceph_mds_session *session;
294
295         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
296                 return NULL;
297         session = mdsc->sessions[mds];
298         dout("lookup_mds_session %p %d\n", session,
299              atomic_read(&session->s_ref));
300         get_session(session);
301         return session;
302 }
303
304 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
305 {
306         if (mds >= mdsc->max_sessions)
307                 return false;
308         return mdsc->sessions[mds];
309 }
310
311 /*
312  * create+register a new session for given mds.
313  * called under mdsc->mutex.
314  */
315 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
316                                                  int mds)
317 {
318         struct ceph_mds_session *s;
319
320         s = kzalloc(sizeof(*s), GFP_NOFS);
321         s->s_mdsc = mdsc;
322         s->s_mds = mds;
323         s->s_state = CEPH_MDS_SESSION_NEW;
324         s->s_ttl = 0;
325         s->s_seq = 0;
326         mutex_init(&s->s_mutex);
327
328         ceph_con_init(mdsc->client->msgr, &s->s_con);
329         s->s_con.private = s;
330         s->s_con.ops = &mds_con_ops;
331         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
332         s->s_con.peer_name.num = cpu_to_le64(mds);
333
334         spin_lock_init(&s->s_cap_lock);
335         s->s_cap_gen = 0;
336         s->s_cap_ttl = 0;
337         s->s_renew_requested = 0;
338         s->s_renew_seq = 0;
339         INIT_LIST_HEAD(&s->s_caps);
340         s->s_nr_caps = 0;
341         s->s_trim_caps = 0;
342         atomic_set(&s->s_ref, 1);
343         INIT_LIST_HEAD(&s->s_waiting);
344         INIT_LIST_HEAD(&s->s_unsafe);
345         s->s_num_cap_releases = 0;
346         s->s_iterating_caps = false;
347         INIT_LIST_HEAD(&s->s_cap_releases);
348         INIT_LIST_HEAD(&s->s_cap_releases_done);
349         INIT_LIST_HEAD(&s->s_cap_flushing);
350         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
351
352         dout("register_session mds%d\n", mds);
353         if (mds >= mdsc->max_sessions) {
354                 int newmax = 1 << get_count_order(mds+1);
355                 struct ceph_mds_session **sa;
356
357                 dout("register_session realloc to %d\n", newmax);
358                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
359                 if (sa == NULL)
360                         goto fail_realloc;
361                 if (mdsc->sessions) {
362                         memcpy(sa, mdsc->sessions,
363                                mdsc->max_sessions * sizeof(void *));
364                         kfree(mdsc->sessions);
365                 }
366                 mdsc->sessions = sa;
367                 mdsc->max_sessions = newmax;
368         }
369         mdsc->sessions[mds] = s;
370         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
371
372         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
373
374         return s;
375
376 fail_realloc:
377         kfree(s);
378         return ERR_PTR(-ENOMEM);
379 }
380
381 /*
382  * called under mdsc->mutex
383  */
384 static void unregister_session(struct ceph_mds_client *mdsc,
385                                struct ceph_mds_session *s)
386 {
387         dout("unregister_session mds%d %p\n", s->s_mds, s);
388         mdsc->sessions[s->s_mds] = NULL;
389         ceph_con_close(&s->s_con);
390         ceph_put_mds_session(s);
391 }
392
393 /*
394  * drop session refs in request.
395  *
396  * should be last request ref, or hold mdsc->mutex
397  */
398 static void put_request_session(struct ceph_mds_request *req)
399 {
400         if (req->r_session) {
401                 ceph_put_mds_session(req->r_session);
402                 req->r_session = NULL;
403         }
404 }
405
406 void ceph_mdsc_release_request(struct kref *kref)
407 {
408         struct ceph_mds_request *req = container_of(kref,
409                                                     struct ceph_mds_request,
410                                                     r_kref);
411         if (req->r_request)
412                 ceph_msg_put(req->r_request);
413         if (req->r_reply) {
414                 ceph_msg_put(req->r_reply);
415                 destroy_reply_info(&req->r_reply_info);
416         }
417         if (req->r_inode) {
418                 ceph_put_cap_refs(ceph_inode(req->r_inode),
419                                   CEPH_CAP_PIN);
420                 iput(req->r_inode);
421         }
422         if (req->r_locked_dir)
423                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
424                                   CEPH_CAP_PIN);
425         if (req->r_target_inode)
426                 iput(req->r_target_inode);
427         if (req->r_dentry)
428                 dput(req->r_dentry);
429         if (req->r_old_dentry) {
430                 ceph_put_cap_refs(
431                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
432                         CEPH_CAP_PIN);
433                 dput(req->r_old_dentry);
434         }
435         kfree(req->r_path1);
436         kfree(req->r_path2);
437         put_request_session(req);
438         ceph_unreserve_caps(&req->r_caps_reservation);
439         kfree(req);
440 }
441
442 /*
443  * lookup session, bump ref if found.
444  *
445  * called under mdsc->mutex.
446  */
447 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
448                                              u64 tid)
449 {
450         struct ceph_mds_request *req;
451         req = radix_tree_lookup(&mdsc->request_tree, tid);
452         if (req)
453                 ceph_mdsc_get_request(req);
454         return req;
455 }
456
457 /*
458  * Register an in-flight request, and assign a tid.  Link to directory
459  * are modifying (if any).
460  *
461  * Called under mdsc->mutex.
462  */
463 static void __register_request(struct ceph_mds_client *mdsc,
464                                struct ceph_mds_request *req,
465                                struct inode *dir)
466 {
467         req->r_tid = ++mdsc->last_tid;
468         if (req->r_num_caps)
469                 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
470         dout("__register_request %p tid %lld\n", req, req->r_tid);
471         ceph_mdsc_get_request(req);
472         radix_tree_insert(&mdsc->request_tree, req->r_tid, (void *)req);
473
474         if (dir) {
475                 struct ceph_inode_info *ci = ceph_inode(dir);
476
477                 spin_lock(&ci->i_unsafe_lock);
478                 req->r_unsafe_dir = dir;
479                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
480                 spin_unlock(&ci->i_unsafe_lock);
481         }
482 }
483
484 static void __unregister_request(struct ceph_mds_client *mdsc,
485                                  struct ceph_mds_request *req)
486 {
487         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
488         radix_tree_delete(&mdsc->request_tree, req->r_tid);
489         ceph_mdsc_put_request(req);
490
491         if (req->r_unsafe_dir) {
492                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
493
494                 spin_lock(&ci->i_unsafe_lock);
495                 list_del_init(&req->r_unsafe_dir_item);
496                 spin_unlock(&ci->i_unsafe_lock);
497         }
498 }
499
500 /*
501  * Choose mds to send request to next.  If there is a hint set in the
502  * request (e.g., due to a prior forward hint from the mds), use that.
503  * Otherwise, consult frag tree and/or caps to identify the
504  * appropriate mds.  If all else fails, choose randomly.
505  *
506  * Called under mdsc->mutex.
507  */
508 static int __choose_mds(struct ceph_mds_client *mdsc,
509                         struct ceph_mds_request *req)
510 {
511         struct inode *inode;
512         struct ceph_inode_info *ci;
513         struct ceph_cap *cap;
514         int mode = req->r_direct_mode;
515         int mds = -1;
516         u32 hash = req->r_direct_hash;
517         bool is_hash = req->r_direct_is_hash;
518
519         /*
520          * is there a specific mds we should try?  ignore hint if we have
521          * no session and the mds is not up (active or recovering).
522          */
523         if (req->r_resend_mds >= 0 &&
524             (__have_session(mdsc, req->r_resend_mds) ||
525              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
526                 dout("choose_mds using resend_mds mds%d\n",
527                      req->r_resend_mds);
528                 return req->r_resend_mds;
529         }
530
531         if (mode == USE_RANDOM_MDS)
532                 goto random;
533
534         inode = NULL;
535         if (req->r_inode) {
536                 inode = req->r_inode;
537         } else if (req->r_dentry) {
538                 if (req->r_dentry->d_inode) {
539                         inode = req->r_dentry->d_inode;
540                 } else {
541                         inode = req->r_dentry->d_parent->d_inode;
542                         hash = req->r_dentry->d_name.hash;
543                         is_hash = true;
544                 }
545         }
546         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
547              (int)hash, mode);
548         if (!inode)
549                 goto random;
550         ci = ceph_inode(inode);
551
552         if (is_hash && S_ISDIR(inode->i_mode)) {
553                 struct ceph_inode_frag frag;
554                 int found;
555
556                 ceph_choose_frag(ci, hash, &frag, &found);
557                 if (found) {
558                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
559                                 u8 r;
560
561                                 /* choose a random replica */
562                                 get_random_bytes(&r, 1);
563                                 r %= frag.ndist;
564                                 mds = frag.dist[r];
565                                 dout("choose_mds %p %llx.%llx "
566                                      "frag %u mds%d (%d/%d)\n",
567                                      inode, ceph_vinop(inode),
568                                      frag.frag, frag.mds,
569                                      (int)r, frag.ndist);
570                                 return mds;
571                         }
572
573                         /* since this file/dir wasn't known to be
574                          * replicated, then we want to look for the
575                          * authoritative mds. */
576                         mode = USE_AUTH_MDS;
577                         if (frag.mds >= 0) {
578                                 /* choose auth mds */
579                                 mds = frag.mds;
580                                 dout("choose_mds %p %llx.%llx "
581                                      "frag %u mds%d (auth)\n",
582                                      inode, ceph_vinop(inode), frag.frag, mds);
583                                 return mds;
584                         }
585                 }
586         }
587
588         spin_lock(&inode->i_lock);
589         cap = NULL;
590         if (mode == USE_AUTH_MDS)
591                 cap = ci->i_auth_cap;
592         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
593                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
594         if (!cap) {
595                 spin_unlock(&inode->i_lock);
596                 goto random;
597         }
598         mds = cap->session->s_mds;
599         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
600              inode, ceph_vinop(inode), mds,
601              cap == ci->i_auth_cap ? "auth " : "", cap);
602         spin_unlock(&inode->i_lock);
603         return mds;
604
605 random:
606         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
607         dout("choose_mds chose random mds%d\n", mds);
608         return mds;
609 }
610
611
612 /*
613  * session messages
614  */
615 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
616 {
617         struct ceph_msg *msg;
618         struct ceph_mds_session_head *h;
619
620         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
621         if (IS_ERR(msg)) {
622                 pr_err("create_session_msg ENOMEM creating msg\n");
623                 return ERR_PTR(PTR_ERR(msg));
624         }
625         h = msg->front.iov_base;
626         h->op = cpu_to_le32(op);
627         h->seq = cpu_to_le64(seq);
628         return msg;
629 }
630
631 /*
632  * send session open request.
633  *
634  * called under mdsc->mutex
635  */
636 static int __open_session(struct ceph_mds_client *mdsc,
637                           struct ceph_mds_session *session)
638 {
639         struct ceph_msg *msg;
640         int mstate;
641         int mds = session->s_mds;
642         int err = 0;
643
644         /* wait for mds to go active? */
645         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
646         dout("open_session to mds%d (%s)\n", mds,
647              ceph_mds_state_name(mstate));
648         session->s_state = CEPH_MDS_SESSION_OPENING;
649         session->s_renew_requested = jiffies;
650
651         /* send connect message */
652         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
653         if (IS_ERR(msg)) {
654                 err = PTR_ERR(msg);
655                 goto out;
656         }
657         ceph_con_send(&session->s_con, msg);
658
659 out:
660         return 0;
661 }
662
663 /*
664  * session caps
665  */
666
667 /*
668  * Free preallocated cap messages assigned to this session
669  */
670 static void cleanup_cap_releases(struct ceph_mds_session *session)
671 {
672         struct ceph_msg *msg;
673
674         spin_lock(&session->s_cap_lock);
675         while (!list_empty(&session->s_cap_releases)) {
676                 msg = list_first_entry(&session->s_cap_releases,
677                                        struct ceph_msg, list_head);
678                 list_del_init(&msg->list_head);
679                 ceph_msg_put(msg);
680         }
681         while (!list_empty(&session->s_cap_releases_done)) {
682                 msg = list_first_entry(&session->s_cap_releases_done,
683                                        struct ceph_msg, list_head);
684                 list_del_init(&msg->list_head);
685                 ceph_msg_put(msg);
686         }
687         spin_unlock(&session->s_cap_lock);
688 }
689
690 /*
691  * Helper to safely iterate over all caps associated with a session.
692  *
693  * caller must hold session s_mutex
694  */
695 static int iterate_session_caps(struct ceph_mds_session *session,
696                                  int (*cb)(struct inode *, struct ceph_cap *,
697                                             void *), void *arg)
698 {
699         struct ceph_cap *cap, *ncap;
700         struct inode *inode;
701         int ret;
702
703         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
704         spin_lock(&session->s_cap_lock);
705         session->s_iterating_caps = true;
706         list_for_each_entry_safe(cap, ncap, &session->s_caps, session_caps) {
707                 inode = igrab(&cap->ci->vfs_inode);
708                 if (!inode)
709                         continue;
710                 spin_unlock(&session->s_cap_lock);
711                 ret = cb(inode, cap, arg);
712                 iput(inode);
713                 spin_lock(&session->s_cap_lock);
714                 if (ret < 0)
715                         goto out;
716         }
717         ret = 0;
718 out:
719         session->s_iterating_caps = false;
720         spin_unlock(&session->s_cap_lock);
721         return ret;
722 }
723
724 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
725                                    void *arg)
726 {
727         struct ceph_inode_info *ci = ceph_inode(inode);
728         dout("removing cap %p, ci is %p, inode is %p\n",
729              cap, ci, &ci->vfs_inode);
730         ceph_remove_cap(cap);
731         return 0;
732 }
733
734 /*
735  * caller must hold session s_mutex
736  */
737 static void remove_session_caps(struct ceph_mds_session *session)
738 {
739         dout("remove_session_caps on %p\n", session);
740         iterate_session_caps(session, remove_session_caps_cb, NULL);
741         BUG_ON(session->s_nr_caps > 0);
742         cleanup_cap_releases(session);
743 }
744
745 /*
746  * wake up any threads waiting on this session's caps.  if the cap is
747  * old (didn't get renewed on the client reconnect), remove it now.
748  *
749  * caller must hold s_mutex.
750  */
751 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
752                               void *arg)
753 {
754         struct ceph_inode_info *ci = ceph_inode(inode);
755
756         wake_up(&ci->i_cap_wq);
757         if (arg) {
758                 spin_lock(&inode->i_lock);
759                 ci->i_wanted_max_size = 0;
760                 ci->i_requested_max_size = 0;
761                 spin_unlock(&inode->i_lock);
762         }
763         return 0;
764 }
765
766 static void wake_up_session_caps(struct ceph_mds_session *session,
767                                  int reconnect)
768 {
769         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
770         iterate_session_caps(session, wake_up_session_cb,
771                              (void *)(unsigned long)reconnect);
772 }
773
774 /*
775  * Send periodic message to MDS renewing all currently held caps.  The
776  * ack will reset the expiration for all caps from this session.
777  *
778  * caller holds s_mutex
779  */
780 static int send_renew_caps(struct ceph_mds_client *mdsc,
781                            struct ceph_mds_session *session)
782 {
783         struct ceph_msg *msg;
784         int state;
785
786         if (time_after_eq(jiffies, session->s_cap_ttl) &&
787             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
788                 pr_info("mds%d caps stale\n", session->s_mds);
789
790         /* do not try to renew caps until a recovering mds has reconnected
791          * with its clients. */
792         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
793         if (state < CEPH_MDS_STATE_RECONNECT) {
794                 dout("send_renew_caps ignoring mds%d (%s)\n",
795                      session->s_mds, ceph_mds_state_name(state));
796                 return 0;
797         }
798
799         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
800                 ceph_mds_state_name(state));
801         session->s_renew_requested = jiffies;
802         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
803                                  ++session->s_renew_seq);
804         if (IS_ERR(msg))
805                 return PTR_ERR(msg);
806         ceph_con_send(&session->s_con, msg);
807         return 0;
808 }
809
810 /*
811  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
812  *
813  * Called under session->s_mutex
814  */
815 static void renewed_caps(struct ceph_mds_client *mdsc,
816                          struct ceph_mds_session *session, int is_renew)
817 {
818         int was_stale;
819         int wake = 0;
820
821         spin_lock(&session->s_cap_lock);
822         was_stale = is_renew && (session->s_cap_ttl == 0 ||
823                                  time_after_eq(jiffies, session->s_cap_ttl));
824
825         session->s_cap_ttl = session->s_renew_requested +
826                 mdsc->mdsmap->m_session_timeout*HZ;
827
828         if (was_stale) {
829                 if (time_before(jiffies, session->s_cap_ttl)) {
830                         pr_info("mds%d caps renewed\n", session->s_mds);
831                         wake = 1;
832                 } else {
833                         pr_info("mds%d caps still stale\n", session->s_mds);
834                 }
835         }
836         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
837              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
838              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
839         spin_unlock(&session->s_cap_lock);
840
841         if (wake)
842                 wake_up_session_caps(session, 0);
843 }
844
845 /*
846  * send a session close request
847  */
848 static int request_close_session(struct ceph_mds_client *mdsc,
849                                  struct ceph_mds_session *session)
850 {
851         struct ceph_msg *msg;
852         int err = 0;
853
854         dout("request_close_session mds%d state %s seq %lld\n",
855              session->s_mds, session_state_name(session->s_state),
856              session->s_seq);
857         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
858         if (IS_ERR(msg))
859                 err = PTR_ERR(msg);
860         else
861                 ceph_con_send(&session->s_con, msg);
862         return err;
863 }
864
865 /*
866  * Called with s_mutex held.
867  */
868 static int __close_session(struct ceph_mds_client *mdsc,
869                          struct ceph_mds_session *session)
870 {
871         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
872                 return 0;
873         session->s_state = CEPH_MDS_SESSION_CLOSING;
874         return request_close_session(mdsc, session);
875 }
876
877 /*
878  * Trim old(er) caps.
879  *
880  * Because we can't cache an inode without one or more caps, we do
881  * this indirectly: if a cap is unused, we prune its aliases, at which
882  * point the inode will hopefully get dropped to.
883  *
884  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
885  * memory pressure from the MDS, though, so it needn't be perfect.
886  */
887 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
888 {
889         struct ceph_mds_session *session = arg;
890         struct ceph_inode_info *ci = ceph_inode(inode);
891         int used, oissued, mine;
892
893         if (session->s_trim_caps <= 0)
894                 return -1;
895
896         spin_lock(&inode->i_lock);
897         mine = cap->issued | cap->implemented;
898         used = __ceph_caps_used(ci);
899         oissued = __ceph_caps_issued_other(ci, cap);
900
901         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
902              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
903              ceph_cap_string(used));
904         if (ci->i_dirty_caps)
905                 goto out;   /* dirty caps */
906         if ((used & ~oissued) & mine)
907                 goto out;   /* we need these caps */
908
909         session->s_trim_caps--;
910         if (oissued) {
911                 /* we aren't the only cap.. just remove us */
912                 __ceph_remove_cap(cap, NULL);
913         } else {
914                 /* try to drop referring dentries */
915                 spin_unlock(&inode->i_lock);
916                 d_prune_aliases(inode);
917                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
918                      inode, cap, atomic_read(&inode->i_count));
919                 return 0;
920         }
921
922 out:
923         spin_unlock(&inode->i_lock);
924         return 0;
925 }
926
927 /*
928  * Trim session cap count down to some max number.
929  */
930 static int trim_caps(struct ceph_mds_client *mdsc,
931                      struct ceph_mds_session *session,
932                      int max_caps)
933 {
934         int trim_caps = session->s_nr_caps - max_caps;
935
936         dout("trim_caps mds%d start: %d / %d, trim %d\n",
937              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
938         if (trim_caps > 0) {
939                 session->s_trim_caps = trim_caps;
940                 iterate_session_caps(session, trim_caps_cb, session);
941                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
942                      session->s_mds, session->s_nr_caps, max_caps,
943                         trim_caps - session->s_trim_caps);
944                 session->s_trim_caps = 0;
945         }
946         return 0;
947 }
948
949 /*
950  * Allocate cap_release messages.  If there is a partially full message
951  * in the queue, try to allocate enough to cover it's remainder, so that
952  * we can send it immediately.
953  *
954  * Called under s_mutex.
955  */
956 static int add_cap_releases(struct ceph_mds_client *mdsc,
957                             struct ceph_mds_session *session,
958                             int extra)
959 {
960         struct ceph_msg *msg;
961         struct ceph_mds_cap_release *head;
962         int err = -ENOMEM;
963
964         if (extra < 0)
965                 extra = mdsc->client->mount_args->cap_release_safety;
966
967         spin_lock(&session->s_cap_lock);
968
969         if (!list_empty(&session->s_cap_releases)) {
970                 msg = list_first_entry(&session->s_cap_releases,
971                                        struct ceph_msg,
972                                  list_head);
973                 head = msg->front.iov_base;
974                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
975         }
976
977         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
978                 spin_unlock(&session->s_cap_lock);
979                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
980                                    0, 0, NULL);
981                 if (!msg)
982                         goto out_unlocked;
983                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
984                      (int)msg->front.iov_len);
985                 head = msg->front.iov_base;
986                 head->num = cpu_to_le32(0);
987                 msg->front.iov_len = sizeof(*head);
988                 spin_lock(&session->s_cap_lock);
989                 list_add(&msg->list_head, &session->s_cap_releases);
990                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
991         }
992
993         if (!list_empty(&session->s_cap_releases)) {
994                 msg = list_first_entry(&session->s_cap_releases,
995                                        struct ceph_msg,
996                                        list_head);
997                 head = msg->front.iov_base;
998                 if (head->num) {
999                         dout(" queueing non-full %p (%d)\n", msg,
1000                              le32_to_cpu(head->num));
1001                         list_move_tail(&msg->list_head,
1002                                       &session->s_cap_releases_done);
1003                         session->s_num_cap_releases -=
1004                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1005                 }
1006         }
1007         err = 0;
1008         spin_unlock(&session->s_cap_lock);
1009 out_unlocked:
1010         return err;
1011 }
1012
1013 /*
1014  * flush all dirty inode data to disk.
1015  *
1016  * returns true if we've flushed through want_flush_seq
1017  */
1018 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1019 {
1020         int mds, ret = 1;
1021
1022         dout("check_cap_flush want %lld\n", want_flush_seq);
1023         mutex_lock(&mdsc->mutex);
1024         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1025                 struct ceph_mds_session *session = mdsc->sessions[mds];
1026
1027                 if (!session)
1028                         continue;
1029                 get_session(session);
1030                 mutex_unlock(&mdsc->mutex);
1031
1032                 mutex_lock(&session->s_mutex);
1033                 if (!list_empty(&session->s_cap_flushing)) {
1034                         struct ceph_inode_info *ci =
1035                                 list_entry(session->s_cap_flushing.next,
1036                                            struct ceph_inode_info,
1037                                            i_flushing_item);
1038                         struct inode *inode = &ci->vfs_inode;
1039
1040                         spin_lock(&inode->i_lock);
1041                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1042                                 dout("check_cap_flush still flushing %p "
1043                                      "seq %lld <= %lld to mds%d\n", inode,
1044                                      ci->i_cap_flush_seq, want_flush_seq,
1045                                      session->s_mds);
1046                                 ret = 0;
1047                         }
1048                         spin_unlock(&inode->i_lock);
1049                 }
1050                 mutex_unlock(&session->s_mutex);
1051                 ceph_put_mds_session(session);
1052
1053                 if (!ret)
1054                         return ret;
1055                 mutex_lock(&mdsc->mutex);
1056         }
1057
1058         mutex_unlock(&mdsc->mutex);
1059         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1060         return ret;
1061 }
1062
1063 /*
1064  * called under s_mutex
1065  */
1066 static void send_cap_releases(struct ceph_mds_client *mdsc,
1067                        struct ceph_mds_session *session)
1068 {
1069         struct ceph_msg *msg;
1070
1071         dout("send_cap_releases mds%d\n", session->s_mds);
1072         while (1) {
1073                 spin_lock(&session->s_cap_lock);
1074                 if (list_empty(&session->s_cap_releases_done))
1075                         break;
1076                 msg = list_first_entry(&session->s_cap_releases_done,
1077                                  struct ceph_msg, list_head);
1078                 list_del_init(&msg->list_head);
1079                 spin_unlock(&session->s_cap_lock);
1080                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1081                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1082                 ceph_con_send(&session->s_con, msg);
1083         }
1084         spin_unlock(&session->s_cap_lock);
1085 }
1086
1087 /*
1088  * requests
1089  */
1090
1091 /*
1092  * Create an mds request.
1093  */
1094 struct ceph_mds_request *
1095 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1096 {
1097         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1098
1099         if (!req)
1100                 return ERR_PTR(-ENOMEM);
1101
1102         req->r_started = jiffies;
1103         req->r_resend_mds = -1;
1104         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1105         req->r_fmode = -1;
1106         kref_init(&req->r_kref);
1107         INIT_LIST_HEAD(&req->r_wait);
1108         init_completion(&req->r_completion);
1109         init_completion(&req->r_safe_completion);
1110         INIT_LIST_HEAD(&req->r_unsafe_item);
1111
1112         req->r_op = op;
1113         req->r_direct_mode = mode;
1114         return req;
1115 }
1116
1117 /*
1118  * return oldest (lowest) tid in request tree, 0 if none.
1119  *
1120  * called under mdsc->mutex.
1121  */
1122 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1123 {
1124         struct ceph_mds_request *first;
1125         if (radix_tree_gang_lookup(&mdsc->request_tree,
1126                                    (void **)&first, 0, 1) <= 0)
1127                 return 0;
1128         return first->r_tid;
1129 }
1130
1131 /*
1132  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1133  * on build_path_from_dentry in fs/cifs/dir.c.
1134  *
1135  * If @stop_on_nosnap, generate path relative to the first non-snapped
1136  * inode.
1137  *
1138  * Encode hidden .snap dirs as a double /, i.e.
1139  *   foo/.snap/bar -> foo//bar
1140  */
1141 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1142                            int stop_on_nosnap)
1143 {
1144         struct dentry *temp;
1145         char *path;
1146         int len, pos;
1147
1148         if (dentry == NULL)
1149                 return ERR_PTR(-EINVAL);
1150
1151 retry:
1152         len = 0;
1153         for (temp = dentry; !IS_ROOT(temp);) {
1154                 struct inode *inode = temp->d_inode;
1155                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1156                         len++;  /* slash only */
1157                 else if (stop_on_nosnap && inode &&
1158                          ceph_snap(inode) == CEPH_NOSNAP)
1159                         break;
1160                 else
1161                         len += 1 + temp->d_name.len;
1162                 temp = temp->d_parent;
1163                 if (temp == NULL) {
1164                         pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1165                         return ERR_PTR(-EINVAL);
1166                 }
1167         }
1168         if (len)
1169                 len--;  /* no leading '/' */
1170
1171         path = kmalloc(len+1, GFP_NOFS);
1172         if (path == NULL)
1173                 return ERR_PTR(-ENOMEM);
1174         pos = len;
1175         path[pos] = 0;  /* trailing null */
1176         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1177                 struct inode *inode = temp->d_inode;
1178
1179                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1180                         dout("build_path_dentry path+%d: %p SNAPDIR\n",
1181                              pos, temp);
1182                 } else if (stop_on_nosnap && inode &&
1183                            ceph_snap(inode) == CEPH_NOSNAP) {
1184                         break;
1185                 } else {
1186                         pos -= temp->d_name.len;
1187                         if (pos < 0)
1188                                 break;
1189                         strncpy(path + pos, temp->d_name.name,
1190                                 temp->d_name.len);
1191                         dout("build_path_dentry path+%d: %p '%.*s'\n",
1192                              pos, temp, temp->d_name.len, path + pos);
1193                 }
1194                 if (pos)
1195                         path[--pos] = '/';
1196                 temp = temp->d_parent;
1197                 if (temp == NULL) {
1198                         pr_err("build_path_dentry corrupt dentry\n");
1199                         kfree(path);
1200                         return ERR_PTR(-EINVAL);
1201                 }
1202         }
1203         if (pos != 0) {
1204                 pr_err("build_path_dentry did not end path lookup where "
1205                        "expected, namelen is %d, pos is %d\n", len, pos);
1206                 /* presumably this is only possible if racing with a
1207                    rename of one of the parent directories (we can not
1208                    lock the dentries above us to prevent this, but
1209                    retrying should be harmless) */
1210                 kfree(path);
1211                 goto retry;
1212         }
1213
1214         *base = ceph_ino(temp->d_inode);
1215         *plen = len;
1216         dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1217              dentry, atomic_read(&dentry->d_count), *base, len, path);
1218         return path;
1219 }
1220
1221 static int build_dentry_path(struct dentry *dentry,
1222                              const char **ppath, int *ppathlen, u64 *pino,
1223                              int *pfreepath)
1224 {
1225         char *path;
1226
1227         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1228                 *pino = ceph_ino(dentry->d_parent->d_inode);
1229                 *ppath = dentry->d_name.name;
1230                 *ppathlen = dentry->d_name.len;
1231                 return 0;
1232         }
1233         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1234         if (IS_ERR(path))
1235                 return PTR_ERR(path);
1236         *ppath = path;
1237         *pfreepath = 1;
1238         return 0;
1239 }
1240
1241 static int build_inode_path(struct inode *inode,
1242                             const char **ppath, int *ppathlen, u64 *pino,
1243                             int *pfreepath)
1244 {
1245         struct dentry *dentry;
1246         char *path;
1247
1248         if (ceph_snap(inode) == CEPH_NOSNAP) {
1249                 *pino = ceph_ino(inode);
1250                 *ppathlen = 0;
1251                 return 0;
1252         }
1253         dentry = d_find_alias(inode);
1254         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1255         dput(dentry);
1256         if (IS_ERR(path))
1257                 return PTR_ERR(path);
1258         *ppath = path;
1259         *pfreepath = 1;
1260         return 0;
1261 }
1262
1263 /*
1264  * request arguments may be specified via an inode *, a dentry *, or
1265  * an explicit ino+path.
1266  */
1267 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1268                                   const char *rpath, u64 rino,
1269                                   const char **ppath, int *pathlen,
1270                                   u64 *ino, int *freepath)
1271 {
1272         int r = 0;
1273
1274         if (rinode) {
1275                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1276                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1277                      ceph_snap(rinode));
1278         } else if (rdentry) {
1279                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1280                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1281                      *ppath);
1282         } else if (rpath) {
1283                 *ino = rino;
1284                 *ppath = rpath;
1285                 *pathlen = strlen(rpath);
1286                 dout(" path %.*s\n", *pathlen, rpath);
1287         }
1288
1289         return r;
1290 }
1291
1292 /*
1293  * called under mdsc->mutex
1294  */
1295 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1296                                                struct ceph_mds_request *req,
1297                                                int mds)
1298 {
1299         struct ceph_msg *msg;
1300         struct ceph_mds_request_head *head;
1301         const char *path1 = NULL;
1302         const char *path2 = NULL;
1303         u64 ino1 = 0, ino2 = 0;
1304         int pathlen1 = 0, pathlen2 = 0;
1305         int freepath1 = 0, freepath2 = 0;
1306         int len;
1307         u16 releases;
1308         void *p, *end;
1309         int ret;
1310
1311         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1312                               req->r_path1, req->r_ino1.ino,
1313                               &path1, &pathlen1, &ino1, &freepath1);
1314         if (ret < 0) {
1315                 msg = ERR_PTR(ret);
1316                 goto out;
1317         }
1318
1319         ret = set_request_path_attr(NULL, req->r_old_dentry,
1320                               req->r_path2, req->r_ino2.ino,
1321                               &path2, &pathlen2, &ino2, &freepath2);
1322         if (ret < 0) {
1323                 msg = ERR_PTR(ret);
1324                 goto out_free1;
1325         }
1326
1327         len = sizeof(*head) +
1328                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1329
1330         /* calculate (max) length for cap releases */
1331         len += sizeof(struct ceph_mds_request_release) *
1332                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1333                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1334         if (req->r_dentry_drop)
1335                 len += req->r_dentry->d_name.len;
1336         if (req->r_old_dentry_drop)
1337                 len += req->r_old_dentry->d_name.len;
1338
1339         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1340         if (IS_ERR(msg))
1341                 goto out_free2;
1342
1343         msg->hdr.tid = cpu_to_le64(req->r_tid);
1344
1345         head = msg->front.iov_base;
1346         p = msg->front.iov_base + sizeof(*head);
1347         end = msg->front.iov_base + msg->front.iov_len;
1348
1349         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1350         head->op = cpu_to_le32(req->r_op);
1351         head->caller_uid = cpu_to_le32(current_fsuid());
1352         head->caller_gid = cpu_to_le32(current_fsgid());
1353         head->args = req->r_args;
1354
1355         ceph_encode_filepath(&p, end, ino1, path1);
1356         ceph_encode_filepath(&p, end, ino2, path2);
1357
1358         /* cap releases */
1359         releases = 0;
1360         if (req->r_inode_drop)
1361                 releases += ceph_encode_inode_release(&p,
1362                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1363                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1364         if (req->r_dentry_drop)
1365                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1366                        mds, req->r_dentry_drop, req->r_dentry_unless);
1367         if (req->r_old_dentry_drop)
1368                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1369                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1370         if (req->r_old_inode_drop)
1371                 releases += ceph_encode_inode_release(&p,
1372                       req->r_old_dentry->d_inode,
1373                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1374         head->num_releases = cpu_to_le16(releases);
1375
1376         BUG_ON(p > end);
1377         msg->front.iov_len = p - msg->front.iov_base;
1378         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1379
1380         msg->pages = req->r_pages;
1381         msg->nr_pages = req->r_num_pages;
1382         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1383         msg->hdr.data_off = cpu_to_le16(0);
1384
1385 out_free2:
1386         if (freepath2)
1387                 kfree((char *)path2);
1388 out_free1:
1389         if (freepath1)
1390                 kfree((char *)path1);
1391 out:
1392         return msg;
1393 }
1394
1395 /*
1396  * called under mdsc->mutex if error, under no mutex if
1397  * success.
1398  */
1399 static void complete_request(struct ceph_mds_client *mdsc,
1400                              struct ceph_mds_request *req)
1401 {
1402         if (req->r_callback)
1403                 req->r_callback(mdsc, req);
1404         else
1405                 complete(&req->r_completion);
1406 }
1407
1408 /*
1409  * called under mdsc->mutex
1410  */
1411 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1412                                   struct ceph_mds_request *req,
1413                                   int mds)
1414 {
1415         struct ceph_mds_request_head *rhead;
1416         struct ceph_msg *msg;
1417         int flags = 0;
1418
1419         req->r_mds = mds;
1420         req->r_attempts++;
1421         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1422              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1423
1424         if (req->r_request) {
1425                 ceph_msg_put(req->r_request);
1426                 req->r_request = NULL;
1427         }
1428         msg = create_request_message(mdsc, req, mds);
1429         if (IS_ERR(msg)) {
1430                 req->r_reply = ERR_PTR(PTR_ERR(msg));
1431                 complete_request(mdsc, req);
1432                 return -PTR_ERR(msg);
1433         }
1434         req->r_request = msg;
1435
1436         rhead = msg->front.iov_base;
1437         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1438         if (req->r_got_unsafe)
1439                 flags |= CEPH_MDS_FLAG_REPLAY;
1440         if (req->r_locked_dir)
1441                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1442         rhead->flags = cpu_to_le32(flags);
1443         rhead->num_fwd = req->r_num_fwd;
1444         rhead->num_retry = req->r_attempts - 1;
1445
1446         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1447
1448         if (req->r_target_inode && req->r_got_unsafe)
1449                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1450         else
1451                 rhead->ino = 0;
1452         return 0;
1453 }
1454
1455 /*
1456  * send request, or put it on the appropriate wait list.
1457  */
1458 static int __do_request(struct ceph_mds_client *mdsc,
1459                         struct ceph_mds_request *req)
1460 {
1461         struct ceph_mds_session *session = NULL;
1462         int mds = -1;
1463         int err = -EAGAIN;
1464
1465         if (req->r_reply)
1466                 goto out;
1467
1468         if (req->r_timeout &&
1469             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1470                 dout("do_request timed out\n");
1471                 err = -EIO;
1472                 goto finish;
1473         }
1474
1475         mds = __choose_mds(mdsc, req);
1476         if (mds < 0 ||
1477             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1478                 dout("do_request no mds or not active, waiting for map\n");
1479                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1480                 goto out;
1481         }
1482
1483         /* get, open session */
1484         session = __ceph_lookup_mds_session(mdsc, mds);
1485         if (!session)
1486                 session = register_session(mdsc, mds);
1487         dout("do_request mds%d session %p state %s\n", mds, session,
1488              session_state_name(session->s_state));
1489         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1490             session->s_state != CEPH_MDS_SESSION_HUNG) {
1491                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1492                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1493                         __open_session(mdsc, session);
1494                 list_add(&req->r_wait, &session->s_waiting);
1495                 goto out_session;
1496         }
1497
1498         /* send request */
1499         req->r_session = get_session(session);
1500         req->r_resend_mds = -1;   /* forget any previous mds hint */
1501
1502         if (req->r_request_started == 0)   /* note request start time */
1503                 req->r_request_started = jiffies;
1504
1505         err = __prepare_send_request(mdsc, req, mds);
1506         if (!err) {
1507                 ceph_msg_get(req->r_request);
1508                 ceph_con_send(&session->s_con, req->r_request);
1509         }
1510
1511 out_session:
1512         ceph_put_mds_session(session);
1513 out:
1514         return err;
1515
1516 finish:
1517         req->r_reply = ERR_PTR(err);
1518         complete_request(mdsc, req);
1519         goto out;
1520 }
1521
1522 /*
1523  * called under mdsc->mutex
1524  */
1525 static void __wake_requests(struct ceph_mds_client *mdsc,
1526                             struct list_head *head)
1527 {
1528         struct ceph_mds_request *req, *nreq;
1529
1530         list_for_each_entry_safe(req, nreq, head, r_wait) {
1531                 list_del_init(&req->r_wait);
1532                 __do_request(mdsc, req);
1533         }
1534 }
1535
1536 /*
1537  * Wake up threads with requests pending for @mds, so that they can
1538  * resubmit their requests to a possibly different mds.  If @all is set,
1539  * wake up if their requests has been forwarded to @mds, too.
1540  */
1541 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1542 {
1543         struct ceph_mds_request *reqs[10];
1544         u64 nexttid = 0;
1545         int i, got;
1546
1547         dout("kick_requests mds%d\n", mds);
1548         while (nexttid <= mdsc->last_tid) {
1549                 got = radix_tree_gang_lookup(&mdsc->request_tree,
1550                                              (void **)&reqs, nexttid, 10);
1551                 if (got == 0)
1552                         break;
1553                 nexttid = reqs[got-1]->r_tid + 1;
1554                 for (i = 0; i < got; i++) {
1555                         if (reqs[i]->r_got_unsafe)
1556                                 continue;
1557                         if (reqs[i]->r_session &&
1558                             reqs[i]->r_session->s_mds == mds) {
1559                                 dout(" kicking tid %llu\n", reqs[i]->r_tid);
1560                                 put_request_session(reqs[i]);
1561                                 __do_request(mdsc, reqs[i]);
1562                         }
1563                 }
1564         }
1565 }
1566
1567 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1568                               struct ceph_mds_request *req)
1569 {
1570         dout("submit_request on %p\n", req);
1571         mutex_lock(&mdsc->mutex);
1572         __register_request(mdsc, req, NULL);
1573         __do_request(mdsc, req);
1574         mutex_unlock(&mdsc->mutex);
1575 }
1576
1577 /*
1578  * Synchrously perform an mds request.  Take care of all of the
1579  * session setup, forwarding, retry details.
1580  */
1581 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1582                          struct inode *dir,
1583                          struct ceph_mds_request *req)
1584 {
1585         int err;
1586
1587         dout("do_request on %p\n", req);
1588
1589         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1590         if (req->r_inode)
1591                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1592         if (req->r_locked_dir)
1593                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1594         if (req->r_old_dentry)
1595                 ceph_get_cap_refs(
1596                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1597                         CEPH_CAP_PIN);
1598
1599         /* issue */
1600         mutex_lock(&mdsc->mutex);
1601         __register_request(mdsc, req, dir);
1602         __do_request(mdsc, req);
1603
1604         /* wait */
1605         if (!req->r_reply) {
1606                 mutex_unlock(&mdsc->mutex);
1607                 if (req->r_timeout) {
1608                         err = (long)wait_for_completion_interruptible_timeout(
1609                                 &req->r_completion, req->r_timeout);
1610                         if (err == 0)
1611                                 req->r_reply = ERR_PTR(-EIO);
1612                         else if (err < 0)
1613                                 req->r_reply = ERR_PTR(err);
1614                 } else {
1615                         err = wait_for_completion_interruptible(
1616                                 &req->r_completion);
1617                         if (err)
1618                                 req->r_reply = ERR_PTR(err);
1619                 }
1620                 mutex_lock(&mdsc->mutex);
1621         }
1622
1623         if (IS_ERR(req->r_reply)) {
1624                 err = PTR_ERR(req->r_reply);
1625                 req->r_reply = NULL;
1626
1627                 if (err == -ERESTARTSYS) {
1628                         /* aborted */
1629                         req->r_aborted = true;
1630
1631                         if (req->r_locked_dir &&
1632                             (req->r_op & CEPH_MDS_OP_WRITE)) {
1633                                 struct ceph_inode_info *ci =
1634                                         ceph_inode(req->r_locked_dir);
1635
1636                                 dout("aborted, clearing I_COMPLETE on %p\n", 
1637                                      req->r_locked_dir);
1638                                 spin_lock(&req->r_locked_dir->i_lock);
1639                                 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1640                                 ci->i_release_count++;
1641                                 spin_unlock(&req->r_locked_dir->i_lock);
1642                         }
1643                 } else {
1644                         /* clean up this request */
1645                         __unregister_request(mdsc, req);
1646                         if (!list_empty(&req->r_unsafe_item))
1647                                 list_del_init(&req->r_unsafe_item);
1648                         complete(&req->r_safe_completion);
1649                 }
1650         } else if (req->r_err) {
1651                 err = req->r_err;
1652         } else {
1653                 err = le32_to_cpu(req->r_reply_info.head->result);
1654         }
1655         mutex_unlock(&mdsc->mutex);
1656
1657         dout("do_request %p done, result %d\n", req, err);
1658         return err;
1659 }
1660
1661 /*
1662  * Handle mds reply.
1663  *
1664  * We take the session mutex and parse and process the reply immediately.
1665  * This preserves the logical ordering of replies, capabilities, etc., sent
1666  * by the MDS as they are applied to our local cache.
1667  */
1668 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1669 {
1670         struct ceph_mds_client *mdsc = session->s_mdsc;
1671         struct ceph_mds_request *req;
1672         struct ceph_mds_reply_head *head = msg->front.iov_base;
1673         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1674         u64 tid;
1675         int err, result;
1676         int mds;
1677
1678         if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
1679                 return;
1680         if (msg->front.iov_len < sizeof(*head)) {
1681                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1682                 ceph_msg_dump(msg);
1683                 return;
1684         }
1685
1686         /* get request, session */
1687         tid = le64_to_cpu(msg->hdr.tid);
1688         mutex_lock(&mdsc->mutex);
1689         req = __lookup_request(mdsc, tid);
1690         if (!req) {
1691                 dout("handle_reply on unknown tid %llu\n", tid);
1692                 mutex_unlock(&mdsc->mutex);
1693                 return;
1694         }
1695         dout("handle_reply %p\n", req);
1696         mds = le64_to_cpu(msg->hdr.src.name.num);
1697
1698         /* correct session? */
1699         if (!req->r_session && req->r_session != session) {
1700                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1701                        " not mds%d\n", tid, session->s_mds,
1702                        req->r_session ? req->r_session->s_mds : -1);
1703                 mutex_unlock(&mdsc->mutex);
1704                 goto out;
1705         }
1706
1707         /* dup? */
1708         if ((req->r_got_unsafe && !head->safe) ||
1709             (req->r_got_safe && head->safe)) {
1710                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1711                            head->safe ? "safe" : "unsafe", tid, mds);
1712                 mutex_unlock(&mdsc->mutex);
1713                 goto out;
1714         }
1715
1716         result = le32_to_cpu(head->result);
1717
1718         /*
1719          * Tolerate 2 consecutive ESTALEs from the same mds.
1720          * FIXME: we should be looking at the cap migrate_seq.
1721          */
1722         if (result == -ESTALE) {
1723                 req->r_direct_mode = USE_AUTH_MDS;
1724                 req->r_num_stale++;
1725                 if (req->r_num_stale <= 2) {
1726                         __do_request(mdsc, req);
1727                         mutex_unlock(&mdsc->mutex);
1728                         goto out;
1729                 }
1730         } else {
1731                 req->r_num_stale = 0;
1732         }
1733
1734         if (head->safe) {
1735                 req->r_got_safe = true;
1736                 __unregister_request(mdsc, req);
1737                 complete(&req->r_safe_completion);
1738
1739                 if (req->r_got_unsafe) {
1740                         /*
1741                          * We already handled the unsafe response, now do the
1742                          * cleanup.  No need to examine the response; the MDS
1743                          * doesn't include any result info in the safe
1744                          * response.  And even if it did, there is nothing
1745                          * useful we could do with a revised return value.
1746                          */
1747                         dout("got safe reply %llu, mds%d\n", tid, mds);
1748                         list_del_init(&req->r_unsafe_item);
1749
1750                         /* last unsafe request during umount? */
1751                         if (mdsc->stopping && !__get_oldest_tid(mdsc))
1752                                 complete(&mdsc->safe_umount_waiters);
1753                         mutex_unlock(&mdsc->mutex);
1754                         goto out;
1755                 }
1756         }
1757
1758         BUG_ON(req->r_reply);
1759
1760         if (!head->safe) {
1761                 req->r_got_unsafe = true;
1762                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1763         }
1764
1765         dout("handle_reply tid %lld result %d\n", tid, result);
1766         rinfo = &req->r_reply_info;
1767         err = parse_reply_info(msg, rinfo);
1768         mutex_unlock(&mdsc->mutex);
1769
1770         mutex_lock(&session->s_mutex);
1771         if (err < 0) {
1772                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1773                 ceph_msg_dump(msg);
1774                 goto out_err;
1775         }
1776
1777         /* snap trace */
1778         if (rinfo->snapblob_len) {
1779                 down_write(&mdsc->snap_rwsem);
1780                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1781                                rinfo->snapblob + rinfo->snapblob_len,
1782                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1783                 downgrade_write(&mdsc->snap_rwsem);
1784         } else {
1785                 down_read(&mdsc->snap_rwsem);
1786         }
1787
1788         /* insert trace into our cache */
1789         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1790         if (err == 0) {
1791                 if (result == 0 && rinfo->dir_nr)
1792                         ceph_readdir_prepopulate(req, req->r_session);
1793                 ceph_unreserve_caps(&req->r_caps_reservation);
1794         }
1795
1796         up_read(&mdsc->snap_rwsem);
1797 out_err:
1798         if (err) {
1799                 req->r_err = err;
1800         } else {
1801                 req->r_reply = msg;
1802                 ceph_msg_get(msg);
1803         }
1804
1805         add_cap_releases(mdsc, req->r_session, -1);
1806         mutex_unlock(&session->s_mutex);
1807
1808         /* kick calling process */
1809         complete_request(mdsc, req);
1810 out:
1811         ceph_mdsc_put_request(req);
1812         return;
1813 }
1814
1815
1816
1817 /*
1818  * handle mds notification that our request has been forwarded.
1819  */
1820 static void handle_forward(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
1821 {
1822         struct ceph_mds_request *req;
1823         u64 tid;
1824         u32 next_mds;
1825         u32 fwd_seq;
1826         u8 must_resend;
1827         int err = -EINVAL;
1828         void *p = msg->front.iov_base;
1829         void *end = p + msg->front.iov_len;
1830         int from_mds, state;
1831
1832         if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
1833                 goto bad;
1834         from_mds = le64_to_cpu(msg->hdr.src.name.num);
1835
1836         ceph_decode_need(&p, end, sizeof(u64)+2*sizeof(u32), bad);
1837         tid = ceph_decode_64(&p);
1838         next_mds = ceph_decode_32(&p);
1839         fwd_seq = ceph_decode_32(&p);
1840         must_resend = ceph_decode_8(&p);
1841
1842         WARN_ON(must_resend);  /* shouldn't happen. */
1843
1844         mutex_lock(&mdsc->mutex);
1845         req = __lookup_request(mdsc, tid);
1846         if (!req) {
1847                 dout("forward %llu dne\n", tid);
1848                 goto out;  /* dup reply? */
1849         }
1850
1851         state = mdsc->sessions[next_mds]->s_state;
1852         if (fwd_seq <= req->r_num_fwd) {
1853                 dout("forward %llu to mds%d - old seq %d <= %d\n",
1854                      tid, next_mds, req->r_num_fwd, fwd_seq);
1855         } else {
1856                 /* resend. forward race not possible; mds would drop */
1857                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1858                 req->r_num_fwd = fwd_seq;
1859                 req->r_resend_mds = next_mds;
1860                 put_request_session(req);
1861                 __do_request(mdsc, req);
1862         }
1863         ceph_mdsc_put_request(req);
1864 out:
1865         mutex_unlock(&mdsc->mutex);
1866         return;
1867
1868 bad:
1869         pr_err("mdsc_handle_forward decode error err=%d\n", err);
1870 }
1871
1872 /*
1873  * handle a mds session control message
1874  */
1875 static void handle_session(struct ceph_mds_session *session,
1876                            struct ceph_msg *msg)
1877 {
1878         struct ceph_mds_client *mdsc = session->s_mdsc;
1879         u32 op;
1880         u64 seq;
1881         int mds;
1882         struct ceph_mds_session_head *h = msg->front.iov_base;
1883         int wake = 0;
1884
1885         if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
1886                 return;
1887         mds = le64_to_cpu(msg->hdr.src.name.num);
1888
1889         /* decode */
1890         if (msg->front.iov_len != sizeof(*h))
1891                 goto bad;
1892         op = le32_to_cpu(h->op);
1893         seq = le64_to_cpu(h->seq);
1894
1895         mutex_lock(&mdsc->mutex);
1896         /* FIXME: this ttl calculation is generous */
1897         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1898         mutex_unlock(&mdsc->mutex);
1899
1900         mutex_lock(&session->s_mutex);
1901
1902         dout("handle_session mds%d %s %p state %s seq %llu\n",
1903              mds, ceph_session_op_name(op), session,
1904              session_state_name(session->s_state), seq);
1905
1906         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1907                 session->s_state = CEPH_MDS_SESSION_OPEN;
1908                 pr_info("mds%d came back\n", session->s_mds);
1909         }
1910
1911         switch (op) {
1912         case CEPH_SESSION_OPEN:
1913                 session->s_state = CEPH_MDS_SESSION_OPEN;
1914                 renewed_caps(mdsc, session, 0);
1915                 wake = 1;
1916                 if (mdsc->stopping)
1917                         __close_session(mdsc, session);
1918                 break;
1919
1920         case CEPH_SESSION_RENEWCAPS:
1921                 if (session->s_renew_seq == seq)
1922                         renewed_caps(mdsc, session, 1);
1923                 break;
1924
1925         case CEPH_SESSION_CLOSE:
1926                 unregister_session(mdsc, session);
1927                 remove_session_caps(session);
1928                 wake = 1; /* for good measure */
1929                 complete(&mdsc->session_close_waiters);
1930                 kick_requests(mdsc, mds, 0);      /* cur only */
1931                 break;
1932
1933         case CEPH_SESSION_STALE:
1934                 pr_info("mds%d caps went stale, renewing\n",
1935                         session->s_mds);
1936                 spin_lock(&session->s_cap_lock);
1937                 session->s_cap_gen++;
1938                 session->s_cap_ttl = 0;
1939                 spin_unlock(&session->s_cap_lock);
1940                 send_renew_caps(mdsc, session);
1941                 break;
1942
1943         case CEPH_SESSION_RECALL_STATE:
1944                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
1945                 break;
1946
1947         default:
1948                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
1949                 WARN_ON(1);
1950         }
1951
1952         mutex_unlock(&session->s_mutex);
1953         if (wake) {
1954                 mutex_lock(&mdsc->mutex);
1955                 __wake_requests(mdsc, &session->s_waiting);
1956                 mutex_unlock(&mdsc->mutex);
1957         }
1958         return;
1959
1960 bad:
1961         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
1962                (int)msg->front.iov_len);
1963         ceph_msg_dump(msg);
1964         return;
1965 }
1966
1967
1968 /*
1969  * called under session->mutex.
1970  */
1971 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
1972                                    struct ceph_mds_session *session)
1973 {
1974         struct ceph_mds_request *req, *nreq;
1975         int err;
1976
1977         dout("replay_unsafe_requests mds%d\n", session->s_mds);
1978
1979         mutex_lock(&mdsc->mutex);
1980         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
1981                 err = __prepare_send_request(mdsc, req, session->s_mds);
1982                 if (!err) {
1983                         ceph_msg_get(req->r_request);
1984                         ceph_con_send(&session->s_con, req->r_request);
1985                 }
1986         }
1987         mutex_unlock(&mdsc->mutex);
1988 }
1989
1990 /*
1991  * Encode information about a cap for a reconnect with the MDS.
1992  */
1993 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
1994                           void *arg)
1995 {
1996         struct ceph_mds_cap_reconnect rec;
1997         struct ceph_inode_info *ci;
1998         struct ceph_pagelist *pagelist = arg;
1999         char *path;
2000         int pathlen, err;
2001         u64 pathbase;
2002         struct dentry *dentry;
2003
2004         ci = cap->ci;
2005
2006         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2007              inode, ceph_vinop(inode), cap, cap->cap_id,
2008              ceph_cap_string(cap->issued));
2009         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2010         if (err)
2011                 return err;
2012
2013         dentry = d_find_alias(inode);
2014         if (dentry) {
2015                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2016                 if (IS_ERR(path)) {
2017                         err = PTR_ERR(path);
2018                         BUG_ON(err);
2019                 }
2020         } else {
2021                 path = NULL;
2022                 pathlen = 0;
2023         }
2024         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2025         if (err)
2026                 goto out;
2027
2028         spin_lock(&inode->i_lock);
2029         cap->seq = 0;        /* reset cap seq */
2030         cap->issue_seq = 0;  /* and issue_seq */
2031         rec.cap_id = cpu_to_le64(cap->cap_id);
2032         rec.pathbase = cpu_to_le64(pathbase);
2033         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2034         rec.issued = cpu_to_le32(cap->issued);
2035         rec.size = cpu_to_le64(inode->i_size);
2036         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2037         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2038         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2039         spin_unlock(&inode->i_lock);
2040
2041         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2042
2043 out:
2044         kfree(path);
2045         dput(dentry);
2046         return err;
2047 }
2048
2049
2050 /*
2051  * If an MDS fails and recovers, clients need to reconnect in order to
2052  * reestablish shared state.  This includes all caps issued through
2053  * this session _and_ the snap_realm hierarchy.  Because it's not
2054  * clear which snap realms the mds cares about, we send everything we
2055  * know about.. that ensures we'll then get any new info the
2056  * recovering MDS might have.
2057  *
2058  * This is a relatively heavyweight operation, but it's rare.
2059  *
2060  * called with mdsc->mutex held.
2061  */
2062 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2063 {
2064         struct ceph_mds_session *session = NULL;
2065         struct ceph_msg *reply;
2066         int err;
2067         int got;
2068         u64 next_snap_ino = 0;
2069         struct ceph_pagelist *pagelist;
2070
2071         pr_info("reconnect to recovering mds%d\n", mds);
2072
2073         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2074         if (!pagelist)
2075                 goto fail_nopagelist;
2076         ceph_pagelist_init(pagelist);
2077
2078         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2079         if (IS_ERR(reply)) {
2080                 err = PTR_ERR(reply);
2081                 goto fail_nomsg;
2082         }
2083
2084         /* find session */
2085         session = __ceph_lookup_mds_session(mdsc, mds);
2086         mutex_unlock(&mdsc->mutex);    /* drop lock for duration */
2087
2088         if (session) {
2089                 mutex_lock(&session->s_mutex);
2090
2091                 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2092                 session->s_seq = 0;
2093
2094                 ceph_con_open(&session->s_con,
2095                               ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2096
2097                 /* replay unsafe requests */
2098                 replay_unsafe_requests(mdsc, session);
2099         } else {
2100                 dout("no session for mds%d, will send short reconnect\n",
2101                      mds);
2102         }
2103
2104         down_read(&mdsc->snap_rwsem);
2105
2106         if (!session)
2107                 goto send;
2108         dout("session %p state %s\n", session,
2109              session_state_name(session->s_state));
2110
2111         /* traverse this session's caps */
2112         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2113         if (err)
2114                 goto fail;
2115         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2116         if (err < 0)
2117                 goto out;
2118
2119         /*
2120          * snaprealms.  we provide mds with the ino, seq (version), and
2121          * parent for all of our realms.  If the mds has any newer info,
2122          * it will tell us.
2123          */
2124         next_snap_ino = 0;
2125         while (1) {
2126                 struct ceph_snap_realm *realm;
2127                 struct ceph_mds_snaprealm_reconnect sr_rec;
2128                 got = radix_tree_gang_lookup(&mdsc->snap_realms,
2129                                              (void **)&realm, next_snap_ino, 1);
2130                 if (!got)
2131                         break;
2132
2133                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2134                      realm->ino, realm->seq, realm->parent_ino);
2135                 sr_rec.ino = cpu_to_le64(realm->ino);
2136                 sr_rec.seq = cpu_to_le64(realm->seq);
2137                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2138                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2139                 if (err)
2140                         goto fail;
2141                 next_snap_ino = realm->ino + 1;
2142         }
2143
2144 send:
2145         reply->pagelist = pagelist;
2146         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2147         reply->nr_pages = calc_pages_for(0, pagelist->length);
2148         ceph_con_send(&session->s_con, reply);
2149
2150         if (session) {
2151                 session->s_state = CEPH_MDS_SESSION_OPEN;
2152                 __wake_requests(mdsc, &session->s_waiting);
2153         }
2154
2155 out:
2156         up_read(&mdsc->snap_rwsem);
2157         if (session) {
2158                 mutex_unlock(&session->s_mutex);
2159                 ceph_put_mds_session(session);
2160         }
2161         mutex_lock(&mdsc->mutex);
2162         return;
2163
2164 fail:
2165         ceph_msg_put(reply);
2166 fail_nomsg:
2167         ceph_pagelist_release(pagelist);
2168         kfree(pagelist);
2169 fail_nopagelist:
2170         pr_err("ENOMEM preparing reconnect for mds%d\n", mds);
2171         goto out;
2172 }
2173
2174
2175 /*
2176  * compare old and new mdsmaps, kicking requests
2177  * and closing out old connections as necessary
2178  *
2179  * called under mdsc->mutex.
2180  */
2181 static void check_new_map(struct ceph_mds_client *mdsc,
2182                           struct ceph_mdsmap *newmap,
2183                           struct ceph_mdsmap *oldmap)
2184 {
2185         int i;
2186         int oldstate, newstate;
2187         struct ceph_mds_session *s;
2188
2189         dout("check_new_map new %u old %u\n",
2190              newmap->m_epoch, oldmap->m_epoch);
2191
2192         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2193                 if (mdsc->sessions[i] == NULL)
2194                         continue;
2195                 s = mdsc->sessions[i];
2196                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2197                 newstate = ceph_mdsmap_get_state(newmap, i);
2198
2199                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2200                      i, ceph_mds_state_name(oldstate),
2201                      ceph_mds_state_name(newstate),
2202                      session_state_name(s->s_state));
2203
2204                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2205                            ceph_mdsmap_get_addr(newmap, i),
2206                            sizeof(struct ceph_entity_addr))) {
2207                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2208                                 /* the session never opened, just close it
2209                                  * out now */
2210                                 __wake_requests(mdsc, &s->s_waiting);
2211                                 unregister_session(mdsc, s);
2212                         } else {
2213                                 /* just close it */
2214                                 mutex_unlock(&mdsc->mutex);
2215                                 mutex_lock(&s->s_mutex);
2216                                 mutex_lock(&mdsc->mutex);
2217                                 ceph_con_close(&s->s_con);
2218                                 mutex_unlock(&s->s_mutex);
2219                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2220                         }
2221
2222                         /* kick any requests waiting on the recovering mds */
2223                         kick_requests(mdsc, i, 1);
2224                 } else if (oldstate == newstate) {
2225                         continue;  /* nothing new with this mds */
2226                 }
2227
2228                 /*
2229                  * send reconnect?
2230                  */
2231                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2232                     newstate >= CEPH_MDS_STATE_RECONNECT)
2233                         send_mds_reconnect(mdsc, i);
2234
2235                 /*
2236                  * kick requests on any mds that has gone active.
2237                  *
2238                  * kick requests on cur or forwarder: we may have sent
2239                  * the request to mds1, mds1 told us it forwarded it
2240                  * to mds2, but then we learn mds1 failed and can't be
2241                  * sure it successfully forwarded our request before
2242                  * it died.
2243                  */
2244                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2245                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2246                         pr_info("mds%d reconnect completed\n", s->s_mds);
2247                         kick_requests(mdsc, i, 1);
2248                         ceph_kick_flushing_caps(mdsc, s);
2249                         wake_up_session_caps(s, 1);
2250                 }
2251         }
2252 }
2253
2254
2255
2256 /*
2257  * leases
2258  */
2259
2260 /*
2261  * caller must hold session s_mutex, dentry->d_lock
2262  */
2263 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2264 {
2265         struct ceph_dentry_info *di = ceph_dentry(dentry);
2266
2267         ceph_put_mds_session(di->lease_session);
2268         di->lease_session = NULL;
2269 }
2270
2271 static void handle_lease(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2272 {
2273         struct super_block *sb = mdsc->client->sb;
2274         struct inode *inode;
2275         struct ceph_mds_session *session;
2276         struct ceph_inode_info *ci;
2277         struct dentry *parent, *dentry;
2278         struct ceph_dentry_info *di;
2279         int mds;
2280         struct ceph_mds_lease *h = msg->front.iov_base;
2281         struct ceph_vino vino;
2282         int mask;
2283         struct qstr dname;
2284         int release = 0;
2285
2286         if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
2287                 return;
2288         mds = le64_to_cpu(msg->hdr.src.name.num);
2289         dout("handle_lease from mds%d\n", mds);
2290
2291         /* decode */
2292         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2293                 goto bad;
2294         vino.ino = le64_to_cpu(h->ino);
2295         vino.snap = CEPH_NOSNAP;
2296         mask = le16_to_cpu(h->mask);
2297         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2298         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2299         if (dname.len != get_unaligned_le32(h+1))
2300                 goto bad;
2301
2302         /* find session */
2303         mutex_lock(&mdsc->mutex);
2304         session = __ceph_lookup_mds_session(mdsc, mds);
2305         mutex_unlock(&mdsc->mutex);
2306         if (!session) {
2307                 pr_err("handle_lease got lease but no session mds%d\n", mds);
2308                 return;
2309         }
2310
2311         mutex_lock(&session->s_mutex);
2312         session->s_seq++;
2313
2314         /* lookup inode */
2315         inode = ceph_find_inode(sb, vino);
2316         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2317              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2318         if (inode == NULL) {
2319                 dout("handle_lease no inode %llx\n", vino.ino);
2320                 goto release;
2321         }
2322         ci = ceph_inode(inode);
2323
2324         /* dentry */
2325         parent = d_find_alias(inode);
2326         if (!parent) {
2327                 dout("no parent dentry on inode %p\n", inode);
2328                 WARN_ON(1);
2329                 goto release;  /* hrm... */
2330         }
2331         dname.hash = full_name_hash(dname.name, dname.len);
2332         dentry = d_lookup(parent, &dname);
2333         dput(parent);
2334         if (!dentry)
2335                 goto release;
2336
2337         spin_lock(&dentry->d_lock);
2338         di = ceph_dentry(dentry);
2339         switch (h->action) {
2340         case CEPH_MDS_LEASE_REVOKE:
2341                 if (di && di->lease_session == session) {
2342                         h->seq = cpu_to_le32(di->lease_seq);
2343                         __ceph_mdsc_drop_dentry_lease(dentry);
2344                 }
2345                 release = 1;
2346                 break;
2347
2348         case CEPH_MDS_LEASE_RENEW:
2349                 if (di && di->lease_session == session &&
2350                     di->lease_gen == session->s_cap_gen &&
2351                     di->lease_renew_from &&
2352                     di->lease_renew_after == 0) {
2353                         unsigned long duration =
2354                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2355
2356                         di->lease_seq = le32_to_cpu(h->seq);
2357                         dentry->d_time = di->lease_renew_from + duration;
2358                         di->lease_renew_after = di->lease_renew_from +
2359                                 (duration >> 1);
2360                         di->lease_renew_from = 0;
2361                 }
2362                 break;
2363         }
2364         spin_unlock(&dentry->d_lock);
2365         dput(dentry);
2366
2367         if (!release)
2368                 goto out;
2369
2370 release:
2371         /* let's just reuse the same message */
2372         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2373         ceph_msg_get(msg);
2374         ceph_con_send(&session->s_con, msg);
2375
2376 out:
2377         iput(inode);
2378         mutex_unlock(&session->s_mutex);
2379         ceph_put_mds_session(session);
2380         return;
2381
2382 bad:
2383         pr_err("corrupt lease message\n");
2384         ceph_msg_dump(msg);
2385 }
2386
2387 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2388                               struct inode *inode,
2389                               struct dentry *dentry, char action,
2390                               u32 seq)
2391 {
2392         struct ceph_msg *msg;
2393         struct ceph_mds_lease *lease;
2394         int len = sizeof(*lease) + sizeof(u32);
2395         int dnamelen = 0;
2396
2397         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2398              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2399         dnamelen = dentry->d_name.len;
2400         len += dnamelen;
2401
2402         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2403         if (IS_ERR(msg))
2404                 return;
2405         lease = msg->front.iov_base;
2406         lease->action = action;
2407         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2408         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2409         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2410         lease->seq = cpu_to_le32(seq);
2411         put_unaligned_le32(dnamelen, lease + 1);
2412         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2413
2414         /*
2415          * if this is a preemptive lease RELEASE, no need to
2416          * flush request stream, since the actual request will
2417          * soon follow.
2418          */
2419         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2420
2421         ceph_con_send(&session->s_con, msg);
2422 }
2423
2424 /*
2425  * Preemptively release a lease we expect to invalidate anyway.
2426  * Pass @inode always, @dentry is optional.
2427  */
2428 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2429                              struct dentry *dentry, int mask)
2430 {
2431         struct ceph_dentry_info *di;
2432         struct ceph_mds_session *session;
2433         u32 seq;
2434
2435         BUG_ON(inode == NULL);
2436         BUG_ON(dentry == NULL);
2437         BUG_ON(mask != CEPH_LOCK_DN);
2438
2439         /* is dentry lease valid? */
2440         spin_lock(&dentry->d_lock);
2441         di = ceph_dentry(dentry);
2442         if (!di || !di->lease_session ||
2443             di->lease_session->s_mds < 0 ||
2444             di->lease_gen != di->lease_session->s_cap_gen ||
2445             !time_before(jiffies, dentry->d_time)) {
2446                 dout("lease_release inode %p dentry %p -- "
2447                      "no lease on %d\n",
2448                      inode, dentry, mask);
2449                 spin_unlock(&dentry->d_lock);
2450                 return;
2451         }
2452
2453         /* we do have a lease on this dentry; note mds and seq */
2454         session = ceph_get_mds_session(di->lease_session);
2455         seq = di->lease_seq;
2456         __ceph_mdsc_drop_dentry_lease(dentry);
2457         spin_unlock(&dentry->d_lock);
2458
2459         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2460              inode, dentry, mask, session->s_mds);
2461         ceph_mdsc_lease_send_msg(session, inode, dentry,
2462                                  CEPH_MDS_LEASE_RELEASE, seq);
2463         ceph_put_mds_session(session);
2464 }
2465
2466 /*
2467  * drop all leases (and dentry refs) in preparation for umount
2468  */
2469 static void drop_leases(struct ceph_mds_client *mdsc)
2470 {
2471         int i;
2472
2473         dout("drop_leases\n");
2474         mutex_lock(&mdsc->mutex);
2475         for (i = 0; i < mdsc->max_sessions; i++) {
2476                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2477                 if (!s)
2478                         continue;
2479                 mutex_unlock(&mdsc->mutex);
2480                 mutex_lock(&s->s_mutex);
2481                 mutex_unlock(&s->s_mutex);
2482                 ceph_put_mds_session(s);
2483                 mutex_lock(&mdsc->mutex);
2484         }
2485         mutex_unlock(&mdsc->mutex);
2486 }
2487
2488
2489
2490 /*
2491  * delayed work -- periodically trim expired leases, renew caps with mds
2492  */
2493 static void schedule_delayed(struct ceph_mds_client *mdsc)
2494 {
2495         int delay = 5;
2496         unsigned hz = round_jiffies_relative(HZ * delay);
2497         schedule_delayed_work(&mdsc->delayed_work, hz);
2498 }
2499
2500 static void delayed_work(struct work_struct *work)
2501 {
2502         int i;
2503         struct ceph_mds_client *mdsc =
2504                 container_of(work, struct ceph_mds_client, delayed_work.work);
2505         int renew_interval;
2506         int renew_caps;
2507
2508         dout("mdsc delayed_work\n");
2509         ceph_check_delayed_caps(mdsc);
2510
2511         mutex_lock(&mdsc->mutex);
2512         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2513         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2514                                    mdsc->last_renew_caps);
2515         if (renew_caps)
2516                 mdsc->last_renew_caps = jiffies;
2517
2518         for (i = 0; i < mdsc->max_sessions; i++) {
2519                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2520                 if (s == NULL)
2521                         continue;
2522                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2523                         dout("resending session close request for mds%d\n",
2524                              s->s_mds);
2525                         request_close_session(mdsc, s);
2526                         ceph_put_mds_session(s);
2527                         continue;
2528                 }
2529                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2530                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2531                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2532                                 pr_info("mds%d hung\n", s->s_mds);
2533                         }
2534                 }
2535                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2536                         /* this mds is failed or recovering, just wait */
2537                         ceph_put_mds_session(s);
2538                         continue;
2539                 }
2540                 mutex_unlock(&mdsc->mutex);
2541
2542                 mutex_lock(&s->s_mutex);
2543                 if (renew_caps)
2544                         send_renew_caps(mdsc, s);
2545                 else
2546                         ceph_con_keepalive(&s->s_con);
2547                 add_cap_releases(mdsc, s, -1);
2548                 send_cap_releases(mdsc, s);
2549                 mutex_unlock(&s->s_mutex);
2550                 ceph_put_mds_session(s);
2551
2552                 mutex_lock(&mdsc->mutex);
2553         }
2554         mutex_unlock(&mdsc->mutex);
2555
2556         schedule_delayed(mdsc);
2557 }
2558
2559
2560 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2561 {
2562         mdsc->client = client;
2563         mutex_init(&mdsc->mutex);
2564         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2565         init_completion(&mdsc->safe_umount_waiters);
2566         init_completion(&mdsc->session_close_waiters);
2567         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2568         mdsc->sessions = NULL;
2569         mdsc->max_sessions = 0;
2570         mdsc->stopping = 0;
2571         init_rwsem(&mdsc->snap_rwsem);
2572         INIT_RADIX_TREE(&mdsc->snap_realms, GFP_NOFS);
2573         INIT_LIST_HEAD(&mdsc->snap_empty);
2574         spin_lock_init(&mdsc->snap_empty_lock);
2575         mdsc->last_tid = 0;
2576         INIT_RADIX_TREE(&mdsc->request_tree, GFP_NOFS);
2577         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2578         mdsc->last_renew_caps = jiffies;
2579         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2580         spin_lock_init(&mdsc->cap_delay_lock);
2581         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2582         spin_lock_init(&mdsc->snap_flush_lock);
2583         mdsc->cap_flush_seq = 0;
2584         INIT_LIST_HEAD(&mdsc->cap_dirty);
2585         mdsc->num_cap_flushing = 0;
2586         spin_lock_init(&mdsc->cap_dirty_lock);
2587         init_waitqueue_head(&mdsc->cap_flushing_wq);
2588         spin_lock_init(&mdsc->dentry_lru_lock);
2589         INIT_LIST_HEAD(&mdsc->dentry_lru);
2590         return 0;
2591 }
2592
2593 /*
2594  * Wait for safe replies on open mds requests.  If we time out, drop
2595  * all requests from the tree to avoid dangling dentry refs.
2596  */
2597 static void wait_requests(struct ceph_mds_client *mdsc)
2598 {
2599         struct ceph_mds_request *req;
2600         struct ceph_client *client = mdsc->client;
2601
2602         mutex_lock(&mdsc->mutex);
2603         if (__get_oldest_tid(mdsc)) {
2604                 mutex_unlock(&mdsc->mutex);
2605                 dout("wait_requests waiting for requests\n");
2606                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2607                                     client->mount_args->mount_timeout * HZ);
2608                 mutex_lock(&mdsc->mutex);
2609
2610                 /* tear down remaining requests */
2611                 while (radix_tree_gang_lookup(&mdsc->request_tree,
2612                                               (void **)&req, 0, 1)) {
2613                         dout("wait_requests timed out on tid %llu\n",
2614                              req->r_tid);
2615                         radix_tree_delete(&mdsc->request_tree, req->r_tid);
2616                         ceph_mdsc_put_request(req);
2617                 }
2618         }
2619         mutex_unlock(&mdsc->mutex);
2620         dout("wait_requests done\n");
2621 }
2622
2623 /*
2624  * called before mount is ro, and before dentries are torn down.
2625  * (hmm, does this still race with new lookups?)
2626  */
2627 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2628 {
2629         dout("pre_umount\n");
2630         mdsc->stopping = 1;
2631
2632         drop_leases(mdsc);
2633         ceph_flush_dirty_caps(mdsc);
2634         wait_requests(mdsc);
2635 }
2636
2637 /*
2638  * wait for all write mds requests to flush.
2639  */
2640 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2641 {
2642         struct ceph_mds_request *req;
2643         u64 next_tid = 0;
2644         int got;
2645
2646         mutex_lock(&mdsc->mutex);
2647         dout("wait_unsafe_requests want %lld\n", want_tid);
2648         while (1) {
2649                 got = radix_tree_gang_lookup(&mdsc->request_tree, (void **)&req,
2650                                              next_tid, 1);
2651                 if (!got)
2652                         break;
2653                 if (req->r_tid > want_tid)
2654                         break;
2655
2656                 next_tid = req->r_tid + 1;
2657                 if ((req->r_op & CEPH_MDS_OP_WRITE) == 0)
2658                         continue;  /* not a write op */
2659
2660                 ceph_mdsc_get_request(req);
2661                 mutex_unlock(&mdsc->mutex);
2662                 dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2663                      req->r_tid, want_tid);
2664                 wait_for_completion(&req->r_safe_completion);
2665                 mutex_lock(&mdsc->mutex);
2666                 ceph_mdsc_put_request(req);
2667         }
2668         mutex_unlock(&mdsc->mutex);
2669         dout("wait_unsafe_requests done\n");
2670 }
2671
2672 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2673 {
2674         u64 want_tid, want_flush;
2675
2676         dout("sync\n");
2677         mutex_lock(&mdsc->mutex);
2678         want_tid = mdsc->last_tid;
2679         want_flush = mdsc->cap_flush_seq;
2680         mutex_unlock(&mdsc->mutex);
2681         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2682
2683         ceph_flush_dirty_caps(mdsc);
2684
2685         wait_unsafe_requests(mdsc, want_tid);
2686         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2687 }
2688
2689
2690 /*
2691  * called after sb is ro.
2692  */
2693 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2694 {
2695         struct ceph_mds_session *session;
2696         int i;
2697         int n;
2698         struct ceph_client *client = mdsc->client;
2699         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2700
2701         dout("close_sessions\n");
2702
2703         mutex_lock(&mdsc->mutex);
2704
2705         /* close sessions */
2706         started = jiffies;
2707         while (time_before(jiffies, started + timeout)) {
2708                 dout("closing sessions\n");
2709                 n = 0;
2710                 for (i = 0; i < mdsc->max_sessions; i++) {
2711                         session = __ceph_lookup_mds_session(mdsc, i);
2712                         if (!session)
2713                                 continue;
2714                         mutex_unlock(&mdsc->mutex);
2715                         mutex_lock(&session->s_mutex);
2716                         __close_session(mdsc, session);
2717                         mutex_unlock(&session->s_mutex);
2718                         ceph_put_mds_session(session);
2719                         mutex_lock(&mdsc->mutex);
2720                         n++;
2721                 }
2722                 if (n == 0)
2723                         break;
2724
2725                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2726                         break;
2727
2728                 dout("waiting for sessions to close\n");
2729                 mutex_unlock(&mdsc->mutex);
2730                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2731                                             timeout);
2732                 mutex_lock(&mdsc->mutex);
2733         }
2734
2735         /* tear down remaining sessions */
2736         for (i = 0; i < mdsc->max_sessions; i++) {
2737                 if (mdsc->sessions[i]) {
2738                         session = get_session(mdsc->sessions[i]);
2739                         unregister_session(mdsc, session);
2740                         mutex_unlock(&mdsc->mutex);
2741                         mutex_lock(&session->s_mutex);
2742                         remove_session_caps(session);
2743                         mutex_unlock(&session->s_mutex);
2744                         ceph_put_mds_session(session);
2745                         mutex_lock(&mdsc->mutex);
2746                 }
2747         }
2748
2749         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2750
2751         mutex_unlock(&mdsc->mutex);
2752
2753         ceph_cleanup_empty_realms(mdsc);
2754
2755         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2756
2757         dout("stopped\n");
2758 }
2759
2760 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2761 {
2762         dout("stop\n");
2763         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2764         if (mdsc->mdsmap)
2765                 ceph_mdsmap_destroy(mdsc->mdsmap);
2766         kfree(mdsc->sessions);
2767 }
2768
2769
2770 /*
2771  * handle mds map update.
2772  */
2773 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2774 {
2775         u32 epoch;
2776         u32 maplen;
2777         void *p = msg->front.iov_base;
2778         void *end = p + msg->front.iov_len;
2779         struct ceph_mdsmap *newmap, *oldmap;
2780         struct ceph_fsid fsid;
2781         int err = -EINVAL;
2782
2783         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2784         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2785         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2786                 return;
2787         epoch = ceph_decode_32(&p);
2788         maplen = ceph_decode_32(&p);
2789         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2790
2791         /* do we need it? */
2792         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2793         mutex_lock(&mdsc->mutex);
2794         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2795                 dout("handle_map epoch %u <= our %u\n",
2796                      epoch, mdsc->mdsmap->m_epoch);
2797                 mutex_unlock(&mdsc->mutex);
2798                 return;
2799         }
2800
2801         newmap = ceph_mdsmap_decode(&p, end);
2802         if (IS_ERR(newmap)) {
2803                 err = PTR_ERR(newmap);
2804                 goto bad_unlock;
2805         }
2806
2807         /* swap into place */
2808         if (mdsc->mdsmap) {
2809                 oldmap = mdsc->mdsmap;
2810                 mdsc->mdsmap = newmap;
2811                 check_new_map(mdsc, newmap, oldmap);
2812                 ceph_mdsmap_destroy(oldmap);
2813         } else {
2814                 mdsc->mdsmap = newmap;  /* first mds map */
2815         }
2816         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2817
2818         __wake_requests(mdsc, &mdsc->waiting_for_map);
2819
2820         mutex_unlock(&mdsc->mutex);
2821         schedule_delayed(mdsc);
2822         return;
2823
2824 bad_unlock:
2825         mutex_unlock(&mdsc->mutex);
2826 bad:
2827         pr_err("error decoding mdsmap %d\n", err);
2828         return;
2829 }
2830
2831 static struct ceph_connection *con_get(struct ceph_connection *con)
2832 {
2833         struct ceph_mds_session *s = con->private;
2834
2835         if (get_session(s)) {
2836                 dout("mdsc con_get %p %d -> %d\n", s,
2837                      atomic_read(&s->s_ref) - 1, atomic_read(&s->s_ref));
2838                 return con;
2839         }
2840         dout("mdsc con_get %p FAIL\n", s);
2841         return NULL;
2842 }
2843
2844 static void con_put(struct ceph_connection *con)
2845 {
2846         struct ceph_mds_session *s = con->private;
2847
2848         dout("mdsc con_put %p %d -> %d\n", s, atomic_read(&s->s_ref),
2849              atomic_read(&s->s_ref) - 1);
2850         ceph_put_mds_session(s);
2851 }
2852
2853 /*
2854  * if the client is unresponsive for long enough, the mds will kill
2855  * the session entirely.
2856  */
2857 static void peer_reset(struct ceph_connection *con)
2858 {
2859         struct ceph_mds_session *s = con->private;
2860
2861         pr_err("mds%d gave us the boot.  IMPLEMENT RECONNECT.\n",
2862                s->s_mds);
2863 }
2864
2865 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2866 {
2867         struct ceph_mds_session *s = con->private;
2868         struct ceph_mds_client *mdsc = s->s_mdsc;
2869         int type = le16_to_cpu(msg->hdr.type);
2870
2871         switch (type) {
2872         case CEPH_MSG_MDS_MAP:
2873                 ceph_mdsc_handle_map(mdsc, msg);
2874                 break;
2875         case CEPH_MSG_CLIENT_SESSION:
2876                 handle_session(s, msg);
2877                 break;
2878         case CEPH_MSG_CLIENT_REPLY:
2879                 handle_reply(s, msg);
2880                 break;
2881         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2882                 handle_forward(mdsc, msg);
2883                 break;
2884         case CEPH_MSG_CLIENT_CAPS:
2885                 ceph_handle_caps(s, msg);
2886                 break;
2887         case CEPH_MSG_CLIENT_SNAP:
2888                 ceph_handle_snap(mdsc, msg);
2889                 break;
2890         case CEPH_MSG_CLIENT_LEASE:
2891                 handle_lease(mdsc, msg);
2892                 break;
2893
2894         default:
2895                 pr_err("received unknown message type %d %s\n", type,
2896                        ceph_msg_type_name(type));
2897         }
2898         ceph_msg_put(msg);
2899 }
2900
2901 /*
2902  * authentication
2903  */
2904 static int get_authorizer(struct ceph_connection *con,
2905                           void **buf, int *len, int *proto,
2906                           void **reply_buf, int *reply_len, int force_new)
2907 {
2908         struct ceph_mds_session *s = con->private;
2909         struct ceph_mds_client *mdsc = s->s_mdsc;
2910         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2911         int ret = 0;
2912
2913         if (force_new && s->s_authorizer) {
2914                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2915                 s->s_authorizer = NULL;
2916         }
2917         if (s->s_authorizer == NULL) {
2918                 if (ac->ops->create_authorizer) {
2919                         ret = ac->ops->create_authorizer(
2920                                 ac, CEPH_ENTITY_TYPE_MDS,
2921                                 &s->s_authorizer,
2922                                 &s->s_authorizer_buf,
2923                                 &s->s_authorizer_buf_len,
2924                                 &s->s_authorizer_reply_buf,
2925                                 &s->s_authorizer_reply_buf_len);
2926                         if (ret)
2927                                 return ret;
2928                 }
2929         }
2930
2931         *proto = ac->protocol;
2932         *buf = s->s_authorizer_buf;
2933         *len = s->s_authorizer_buf_len;
2934         *reply_buf = s->s_authorizer_reply_buf;
2935         *reply_len = s->s_authorizer_reply_buf_len;
2936         return 0;
2937 }
2938
2939
2940 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2941 {
2942         struct ceph_mds_session *s = con->private;
2943         struct ceph_mds_client *mdsc = s->s_mdsc;
2944         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2945
2946         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
2947 }
2948
2949 static int invalidate_authorizer(struct ceph_connection *con)
2950 {
2951         struct ceph_mds_session *s = con->private;
2952         struct ceph_mds_client *mdsc = s->s_mdsc;
2953         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2954
2955         if (ac->ops->invalidate_authorizer)
2956                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
2957
2958         return ceph_monc_validate_auth(&mdsc->client->monc);
2959 }
2960
2961 const static struct ceph_connection_operations mds_con_ops = {
2962         .get = con_get,
2963         .put = con_put,
2964         .dispatch = dispatch,
2965         .get_authorizer = get_authorizer,
2966         .verify_authorizer_reply = verify_authorizer_reply,
2967         .invalidate_authorizer = invalidate_authorizer,
2968         .peer_reset = peer_reset,
2969 };
2970
2971
2972
2973
2974 /* eof */