ceph: fix cap removal races
[linux-2.6.git] / fs / ceph / mds_client.c
1 #include "ceph_debug.h"
2
3 #include <linux/wait.h>
4 #include <linux/slab.h>
5 #include <linux/sched.h>
6
7 #include "mds_client.h"
8 #include "mon_client.h"
9 #include "super.h"
10 #include "messenger.h"
11 #include "decode.h"
12 #include "auth.h"
13 #include "pagelist.h"
14
15 /*
16  * A cluster of MDS (metadata server) daemons is responsible for
17  * managing the file system namespace (the directory hierarchy and
18  * inodes) and for coordinating shared access to storage.  Metadata is
19  * partitioning hierarchically across a number of servers, and that
20  * partition varies over time as the cluster adjusts the distribution
21  * in order to balance load.
22  *
23  * The MDS client is primarily responsible to managing synchronous
24  * metadata requests for operations like open, unlink, and so forth.
25  * If there is a MDS failure, we find out about it when we (possibly
26  * request and) receive a new MDS map, and can resubmit affected
27  * requests.
28  *
29  * For the most part, though, we take advantage of a lossless
30  * communications channel to the MDS, and do not need to worry about
31  * timing out or resubmitting requests.
32  *
33  * We maintain a stateful "session" with each MDS we interact with.
34  * Within each session, we sent periodic heartbeat messages to ensure
35  * any capabilities or leases we have been issues remain valid.  If
36  * the session times out and goes stale, our leases and capabilities
37  * are no longer valid.
38  */
39
40 static void __wake_requests(struct ceph_mds_client *mdsc,
41                             struct list_head *head);
42
43 const static struct ceph_connection_operations mds_con_ops;
44
45
46 /*
47  * mds reply parsing
48  */
49
50 /*
51  * parse individual inode info
52  */
53 static int parse_reply_info_in(void **p, void *end,
54                                struct ceph_mds_reply_info_in *info)
55 {
56         int err = -EIO;
57
58         info->in = *p;
59         *p += sizeof(struct ceph_mds_reply_inode) +
60                 sizeof(*info->in->fragtree.splits) *
61                 le32_to_cpu(info->in->fragtree.nsplits);
62
63         ceph_decode_32_safe(p, end, info->symlink_len, bad);
64         ceph_decode_need(p, end, info->symlink_len, bad);
65         info->symlink = *p;
66         *p += info->symlink_len;
67
68         ceph_decode_32_safe(p, end, info->xattr_len, bad);
69         ceph_decode_need(p, end, info->xattr_len, bad);
70         info->xattr_data = *p;
71         *p += info->xattr_len;
72         return 0;
73 bad:
74         return err;
75 }
76
77 /*
78  * parse a normal reply, which may contain a (dir+)dentry and/or a
79  * target inode.
80  */
81 static int parse_reply_info_trace(void **p, void *end,
82                                   struct ceph_mds_reply_info_parsed *info)
83 {
84         int err;
85
86         if (info->head->is_dentry) {
87                 err = parse_reply_info_in(p, end, &info->diri);
88                 if (err < 0)
89                         goto out_bad;
90
91                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
92                         goto bad;
93                 info->dirfrag = *p;
94                 *p += sizeof(*info->dirfrag) +
95                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
96                 if (unlikely(*p > end))
97                         goto bad;
98
99                 ceph_decode_32_safe(p, end, info->dname_len, bad);
100                 ceph_decode_need(p, end, info->dname_len, bad);
101                 info->dname = *p;
102                 *p += info->dname_len;
103                 info->dlease = *p;
104                 *p += sizeof(*info->dlease);
105         }
106
107         if (info->head->is_target) {
108                 err = parse_reply_info_in(p, end, &info->targeti);
109                 if (err < 0)
110                         goto out_bad;
111         }
112
113         if (unlikely(*p != end))
114                 goto bad;
115         return 0;
116
117 bad:
118         err = -EIO;
119 out_bad:
120         pr_err("problem parsing mds trace %d\n", err);
121         return err;
122 }
123
124 /*
125  * parse readdir results
126  */
127 static int parse_reply_info_dir(void **p, void *end,
128                                 struct ceph_mds_reply_info_parsed *info)
129 {
130         u32 num, i = 0;
131         int err;
132
133         info->dir_dir = *p;
134         if (*p + sizeof(*info->dir_dir) > end)
135                 goto bad;
136         *p += sizeof(*info->dir_dir) +
137                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
138         if (*p > end)
139                 goto bad;
140
141         ceph_decode_need(p, end, sizeof(num) + 2, bad);
142         num = ceph_decode_32(p);
143         info->dir_end = ceph_decode_8(p);
144         info->dir_complete = ceph_decode_8(p);
145         if (num == 0)
146                 goto done;
147
148         /* alloc large array */
149         info->dir_nr = num;
150         info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
151                                sizeof(*info->dir_dname) +
152                                sizeof(*info->dir_dname_len) +
153                                sizeof(*info->dir_dlease),
154                                GFP_NOFS);
155         if (info->dir_in == NULL) {
156                 err = -ENOMEM;
157                 goto out_bad;
158         }
159         info->dir_dname = (void *)(info->dir_in + num);
160         info->dir_dname_len = (void *)(info->dir_dname + num);
161         info->dir_dlease = (void *)(info->dir_dname_len + num);
162
163         while (num) {
164                 /* dentry */
165                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
166                 info->dir_dname_len[i] = ceph_decode_32(p);
167                 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
168                 info->dir_dname[i] = *p;
169                 *p += info->dir_dname_len[i];
170                 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
171                      info->dir_dname[i]);
172                 info->dir_dlease[i] = *p;
173                 *p += sizeof(struct ceph_mds_reply_lease);
174
175                 /* inode */
176                 err = parse_reply_info_in(p, end, &info->dir_in[i]);
177                 if (err < 0)
178                         goto out_bad;
179                 i++;
180                 num--;
181         }
182
183 done:
184         if (*p != end)
185                 goto bad;
186         return 0;
187
188 bad:
189         err = -EIO;
190 out_bad:
191         pr_err("problem parsing dir contents %d\n", err);
192         return err;
193 }
194
195 /*
196  * parse entire mds reply
197  */
198 static int parse_reply_info(struct ceph_msg *msg,
199                             struct ceph_mds_reply_info_parsed *info)
200 {
201         void *p, *end;
202         u32 len;
203         int err;
204
205         info->head = msg->front.iov_base;
206         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
207         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
208
209         /* trace */
210         ceph_decode_32_safe(&p, end, len, bad);
211         if (len > 0) {
212                 err = parse_reply_info_trace(&p, p+len, info);
213                 if (err < 0)
214                         goto out_bad;
215         }
216
217         /* dir content */
218         ceph_decode_32_safe(&p, end, len, bad);
219         if (len > 0) {
220                 err = parse_reply_info_dir(&p, p+len, info);
221                 if (err < 0)
222                         goto out_bad;
223         }
224
225         /* snap blob */
226         ceph_decode_32_safe(&p, end, len, bad);
227         info->snapblob_len = len;
228         info->snapblob = p;
229         p += len;
230
231         if (p != end)
232                 goto bad;
233         return 0;
234
235 bad:
236         err = -EIO;
237 out_bad:
238         pr_err("mds parse_reply err %d\n", err);
239         return err;
240 }
241
242 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
243 {
244         kfree(info->dir_in);
245 }
246
247
248 /*
249  * sessions
250  */
251 static const char *session_state_name(int s)
252 {
253         switch (s) {
254         case CEPH_MDS_SESSION_NEW: return "new";
255         case CEPH_MDS_SESSION_OPENING: return "opening";
256         case CEPH_MDS_SESSION_OPEN: return "open";
257         case CEPH_MDS_SESSION_HUNG: return "hung";
258         case CEPH_MDS_SESSION_CLOSING: return "closing";
259         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
260         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
261         default: return "???";
262         }
263 }
264
265 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
266 {
267         if (atomic_inc_not_zero(&s->s_ref)) {
268                 dout("mdsc get_session %p %d -> %d\n", s,
269                      atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
270                 return s;
271         } else {
272                 dout("mdsc get_session %p 0 -- FAIL", s);
273                 return NULL;
274         }
275 }
276
277 void ceph_put_mds_session(struct ceph_mds_session *s)
278 {
279         dout("mdsc put_session %p %d -> %d\n", s,
280              atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
281         if (atomic_dec_and_test(&s->s_ref)) {
282                 if (s->s_authorizer)
283                         s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
284                                 s->s_mdsc->client->monc.auth, s->s_authorizer);
285                 kfree(s);
286         }
287 }
288
289 /*
290  * called under mdsc->mutex
291  */
292 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
293                                                    int mds)
294 {
295         struct ceph_mds_session *session;
296
297         if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
298                 return NULL;
299         session = mdsc->sessions[mds];
300         dout("lookup_mds_session %p %d\n", session,
301              atomic_read(&session->s_ref));
302         get_session(session);
303         return session;
304 }
305
306 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
307 {
308         if (mds >= mdsc->max_sessions)
309                 return false;
310         return mdsc->sessions[mds];
311 }
312
313 static int __verify_registered_session(struct ceph_mds_client *mdsc,
314                                        struct ceph_mds_session *s)
315 {
316         if (s->s_mds >= mdsc->max_sessions ||
317             mdsc->sessions[s->s_mds] != s)
318                 return -ENOENT;
319         return 0;
320 }
321
322 /*
323  * create+register a new session for given mds.
324  * called under mdsc->mutex.
325  */
326 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
327                                                  int mds)
328 {
329         struct ceph_mds_session *s;
330
331         s = kzalloc(sizeof(*s), GFP_NOFS);
332         if (!s)
333                 return ERR_PTR(-ENOMEM);
334         s->s_mdsc = mdsc;
335         s->s_mds = mds;
336         s->s_state = CEPH_MDS_SESSION_NEW;
337         s->s_ttl = 0;
338         s->s_seq = 0;
339         mutex_init(&s->s_mutex);
340
341         ceph_con_init(mdsc->client->msgr, &s->s_con);
342         s->s_con.private = s;
343         s->s_con.ops = &mds_con_ops;
344         s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
345         s->s_con.peer_name.num = cpu_to_le64(mds);
346
347         spin_lock_init(&s->s_cap_lock);
348         s->s_cap_gen = 0;
349         s->s_cap_ttl = 0;
350         s->s_renew_requested = 0;
351         s->s_renew_seq = 0;
352         INIT_LIST_HEAD(&s->s_caps);
353         s->s_nr_caps = 0;
354         s->s_trim_caps = 0;
355         atomic_set(&s->s_ref, 1);
356         INIT_LIST_HEAD(&s->s_waiting);
357         INIT_LIST_HEAD(&s->s_unsafe);
358         s->s_num_cap_releases = 0;
359         s->s_cap_iterator = NULL;
360         INIT_LIST_HEAD(&s->s_cap_releases);
361         INIT_LIST_HEAD(&s->s_cap_releases_done);
362         INIT_LIST_HEAD(&s->s_cap_flushing);
363         INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
364
365         dout("register_session mds%d\n", mds);
366         if (mds >= mdsc->max_sessions) {
367                 int newmax = 1 << get_count_order(mds+1);
368                 struct ceph_mds_session **sa;
369
370                 dout("register_session realloc to %d\n", newmax);
371                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
372                 if (sa == NULL)
373                         goto fail_realloc;
374                 if (mdsc->sessions) {
375                         memcpy(sa, mdsc->sessions,
376                                mdsc->max_sessions * sizeof(void *));
377                         kfree(mdsc->sessions);
378                 }
379                 mdsc->sessions = sa;
380                 mdsc->max_sessions = newmax;
381         }
382         mdsc->sessions[mds] = s;
383         atomic_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
384
385         ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
386
387         return s;
388
389 fail_realloc:
390         kfree(s);
391         return ERR_PTR(-ENOMEM);
392 }
393
394 /*
395  * called under mdsc->mutex
396  */
397 static void __unregister_session(struct ceph_mds_client *mdsc,
398                                struct ceph_mds_session *s)
399 {
400         dout("__unregister_session mds%d %p\n", s->s_mds, s);
401         BUG_ON(mdsc->sessions[s->s_mds] != s);
402         mdsc->sessions[s->s_mds] = NULL;
403         ceph_con_close(&s->s_con);
404         ceph_put_mds_session(s);
405 }
406
407 /*
408  * drop session refs in request.
409  *
410  * should be last request ref, or hold mdsc->mutex
411  */
412 static void put_request_session(struct ceph_mds_request *req)
413 {
414         if (req->r_session) {
415                 ceph_put_mds_session(req->r_session);
416                 req->r_session = NULL;
417         }
418 }
419
420 void ceph_mdsc_release_request(struct kref *kref)
421 {
422         struct ceph_mds_request *req = container_of(kref,
423                                                     struct ceph_mds_request,
424                                                     r_kref);
425         if (req->r_request)
426                 ceph_msg_put(req->r_request);
427         if (req->r_reply) {
428                 ceph_msg_put(req->r_reply);
429                 destroy_reply_info(&req->r_reply_info);
430         }
431         if (req->r_inode) {
432                 ceph_put_cap_refs(ceph_inode(req->r_inode),
433                                   CEPH_CAP_PIN);
434                 iput(req->r_inode);
435         }
436         if (req->r_locked_dir)
437                 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
438                                   CEPH_CAP_PIN);
439         if (req->r_target_inode)
440                 iput(req->r_target_inode);
441         if (req->r_dentry)
442                 dput(req->r_dentry);
443         if (req->r_old_dentry) {
444                 ceph_put_cap_refs(
445                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
446                         CEPH_CAP_PIN);
447                 dput(req->r_old_dentry);
448         }
449         kfree(req->r_path1);
450         kfree(req->r_path2);
451         put_request_session(req);
452         ceph_unreserve_caps(&req->r_caps_reservation);
453         kfree(req);
454 }
455
456 /*
457  * lookup session, bump ref if found.
458  *
459  * called under mdsc->mutex.
460  */
461 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
462                                              u64 tid)
463 {
464         struct ceph_mds_request *req;
465         struct rb_node *n = mdsc->request_tree.rb_node;
466
467         while (n) {
468                 req = rb_entry(n, struct ceph_mds_request, r_node);
469                 if (tid < req->r_tid)
470                         n = n->rb_left;
471                 else if (tid > req->r_tid)
472                         n = n->rb_right;
473                 else {
474                         ceph_mdsc_get_request(req);
475                         return req;
476                 }
477         }
478         return NULL;
479 }
480
481 static void __insert_request(struct ceph_mds_client *mdsc,
482                              struct ceph_mds_request *new)
483 {
484         struct rb_node **p = &mdsc->request_tree.rb_node;
485         struct rb_node *parent = NULL;
486         struct ceph_mds_request *req = NULL;
487
488         while (*p) {
489                 parent = *p;
490                 req = rb_entry(parent, struct ceph_mds_request, r_node);
491                 if (new->r_tid < req->r_tid)
492                         p = &(*p)->rb_left;
493                 else if (new->r_tid > req->r_tid)
494                         p = &(*p)->rb_right;
495                 else
496                         BUG();
497         }
498
499         rb_link_node(&new->r_node, parent, p);
500         rb_insert_color(&new->r_node, &mdsc->request_tree);
501 }
502
503 /*
504  * Register an in-flight request, and assign a tid.  Link to directory
505  * are modifying (if any).
506  *
507  * Called under mdsc->mutex.
508  */
509 static void __register_request(struct ceph_mds_client *mdsc,
510                                struct ceph_mds_request *req,
511                                struct inode *dir)
512 {
513         req->r_tid = ++mdsc->last_tid;
514         if (req->r_num_caps)
515                 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
516         dout("__register_request %p tid %lld\n", req, req->r_tid);
517         ceph_mdsc_get_request(req);
518         __insert_request(mdsc, req);
519
520         if (dir) {
521                 struct ceph_inode_info *ci = ceph_inode(dir);
522
523                 spin_lock(&ci->i_unsafe_lock);
524                 req->r_unsafe_dir = dir;
525                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
526                 spin_unlock(&ci->i_unsafe_lock);
527         }
528 }
529
530 static void __unregister_request(struct ceph_mds_client *mdsc,
531                                  struct ceph_mds_request *req)
532 {
533         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
534         rb_erase(&req->r_node, &mdsc->request_tree);
535         RB_CLEAR_NODE(&req->r_node);
536
537         if (req->r_unsafe_dir) {
538                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
539
540                 spin_lock(&ci->i_unsafe_lock);
541                 list_del_init(&req->r_unsafe_dir_item);
542                 spin_unlock(&ci->i_unsafe_lock);
543         }
544
545         ceph_mdsc_put_request(req);
546 }
547
548 /*
549  * Choose mds to send request to next.  If there is a hint set in the
550  * request (e.g., due to a prior forward hint from the mds), use that.
551  * Otherwise, consult frag tree and/or caps to identify the
552  * appropriate mds.  If all else fails, choose randomly.
553  *
554  * Called under mdsc->mutex.
555  */
556 static int __choose_mds(struct ceph_mds_client *mdsc,
557                         struct ceph_mds_request *req)
558 {
559         struct inode *inode;
560         struct ceph_inode_info *ci;
561         struct ceph_cap *cap;
562         int mode = req->r_direct_mode;
563         int mds = -1;
564         u32 hash = req->r_direct_hash;
565         bool is_hash = req->r_direct_is_hash;
566
567         /*
568          * is there a specific mds we should try?  ignore hint if we have
569          * no session and the mds is not up (active or recovering).
570          */
571         if (req->r_resend_mds >= 0 &&
572             (__have_session(mdsc, req->r_resend_mds) ||
573              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
574                 dout("choose_mds using resend_mds mds%d\n",
575                      req->r_resend_mds);
576                 return req->r_resend_mds;
577         }
578
579         if (mode == USE_RANDOM_MDS)
580                 goto random;
581
582         inode = NULL;
583         if (req->r_inode) {
584                 inode = req->r_inode;
585         } else if (req->r_dentry) {
586                 if (req->r_dentry->d_inode) {
587                         inode = req->r_dentry->d_inode;
588                 } else {
589                         inode = req->r_dentry->d_parent->d_inode;
590                         hash = req->r_dentry->d_name.hash;
591                         is_hash = true;
592                 }
593         }
594         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
595              (int)hash, mode);
596         if (!inode)
597                 goto random;
598         ci = ceph_inode(inode);
599
600         if (is_hash && S_ISDIR(inode->i_mode)) {
601                 struct ceph_inode_frag frag;
602                 int found;
603
604                 ceph_choose_frag(ci, hash, &frag, &found);
605                 if (found) {
606                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
607                                 u8 r;
608
609                                 /* choose a random replica */
610                                 get_random_bytes(&r, 1);
611                                 r %= frag.ndist;
612                                 mds = frag.dist[r];
613                                 dout("choose_mds %p %llx.%llx "
614                                      "frag %u mds%d (%d/%d)\n",
615                                      inode, ceph_vinop(inode),
616                                      frag.frag, frag.mds,
617                                      (int)r, frag.ndist);
618                                 return mds;
619                         }
620
621                         /* since this file/dir wasn't known to be
622                          * replicated, then we want to look for the
623                          * authoritative mds. */
624                         mode = USE_AUTH_MDS;
625                         if (frag.mds >= 0) {
626                                 /* choose auth mds */
627                                 mds = frag.mds;
628                                 dout("choose_mds %p %llx.%llx "
629                                      "frag %u mds%d (auth)\n",
630                                      inode, ceph_vinop(inode), frag.frag, mds);
631                                 return mds;
632                         }
633                 }
634         }
635
636         spin_lock(&inode->i_lock);
637         cap = NULL;
638         if (mode == USE_AUTH_MDS)
639                 cap = ci->i_auth_cap;
640         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
641                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
642         if (!cap) {
643                 spin_unlock(&inode->i_lock);
644                 goto random;
645         }
646         mds = cap->session->s_mds;
647         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
648              inode, ceph_vinop(inode), mds,
649              cap == ci->i_auth_cap ? "auth " : "", cap);
650         spin_unlock(&inode->i_lock);
651         return mds;
652
653 random:
654         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
655         dout("choose_mds chose random mds%d\n", mds);
656         return mds;
657 }
658
659
660 /*
661  * session messages
662  */
663 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
664 {
665         struct ceph_msg *msg;
666         struct ceph_mds_session_head *h;
667
668         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
669         if (IS_ERR(msg)) {
670                 pr_err("create_session_msg ENOMEM creating msg\n");
671                 return ERR_PTR(PTR_ERR(msg));
672         }
673         h = msg->front.iov_base;
674         h->op = cpu_to_le32(op);
675         h->seq = cpu_to_le64(seq);
676         return msg;
677 }
678
679 /*
680  * send session open request.
681  *
682  * called under mdsc->mutex
683  */
684 static int __open_session(struct ceph_mds_client *mdsc,
685                           struct ceph_mds_session *session)
686 {
687         struct ceph_msg *msg;
688         int mstate;
689         int mds = session->s_mds;
690         int err = 0;
691
692         /* wait for mds to go active? */
693         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
694         dout("open_session to mds%d (%s)\n", mds,
695              ceph_mds_state_name(mstate));
696         session->s_state = CEPH_MDS_SESSION_OPENING;
697         session->s_renew_requested = jiffies;
698
699         /* send connect message */
700         msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
701         if (IS_ERR(msg)) {
702                 err = PTR_ERR(msg);
703                 goto out;
704         }
705         ceph_con_send(&session->s_con, msg);
706
707 out:
708         return 0;
709 }
710
711 /*
712  * session caps
713  */
714
715 /*
716  * Free preallocated cap messages assigned to this session
717  */
718 static void cleanup_cap_releases(struct ceph_mds_session *session)
719 {
720         struct ceph_msg *msg;
721
722         spin_lock(&session->s_cap_lock);
723         while (!list_empty(&session->s_cap_releases)) {
724                 msg = list_first_entry(&session->s_cap_releases,
725                                        struct ceph_msg, list_head);
726                 list_del_init(&msg->list_head);
727                 ceph_msg_put(msg);
728         }
729         while (!list_empty(&session->s_cap_releases_done)) {
730                 msg = list_first_entry(&session->s_cap_releases_done,
731                                        struct ceph_msg, list_head);
732                 list_del_init(&msg->list_head);
733                 ceph_msg_put(msg);
734         }
735         spin_unlock(&session->s_cap_lock);
736 }
737
738 /*
739  * Helper to safely iterate over all caps associated with a session, with
740  * special care taken to handle a racing __ceph_remove_cap().
741  *
742  * Caller must hold session s_mutex.
743  */
744 static int iterate_session_caps(struct ceph_mds_session *session,
745                                  int (*cb)(struct inode *, struct ceph_cap *,
746                                             void *), void *arg)
747 {
748         struct list_head *p;
749         struct ceph_cap *cap;
750         struct inode *inode, *last_inode = NULL;
751         struct ceph_cap *old_cap = NULL;
752         int ret;
753
754         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
755         spin_lock(&session->s_cap_lock);
756         p = session->s_caps.next;
757         while (p != &session->s_caps) {
758                 cap = list_entry(p, struct ceph_cap, session_caps);
759                 inode = igrab(&cap->ci->vfs_inode);
760                 if (!inode) {
761                         p = p->next;
762                         continue;
763                 }
764                 session->s_cap_iterator = cap;
765                 spin_unlock(&session->s_cap_lock);
766
767                 if (last_inode) {
768                         iput(last_inode);
769                         last_inode = NULL;
770                 }
771                 if (old_cap) {
772                         ceph_put_cap(old_cap);
773                         old_cap = NULL;
774                 }
775
776                 ret = cb(inode, cap, arg);
777                 last_inode = inode;
778
779                 spin_lock(&session->s_cap_lock);
780                 p = p->next;
781                 if (cap->ci == NULL) {
782                         dout("iterate_session_caps  finishing cap %p removal\n",
783                              cap);
784                         BUG_ON(cap->session != session);
785                         list_del_init(&cap->session_caps);
786                         session->s_nr_caps--;
787                         cap->session = NULL;
788                         old_cap = cap;  /* put_cap it w/o locks held */
789                 }
790                 if (ret < 0)
791                         goto out;
792         }
793         ret = 0;
794 out:
795         session->s_cap_iterator = NULL;
796         spin_unlock(&session->s_cap_lock);
797
798         if (last_inode)
799                 iput(last_inode);
800         if (old_cap)
801                 ceph_put_cap(old_cap);
802
803         return ret;
804 }
805
806 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
807                                    void *arg)
808 {
809         struct ceph_inode_info *ci = ceph_inode(inode);
810         dout("removing cap %p, ci is %p, inode is %p\n",
811              cap, ci, &ci->vfs_inode);
812         ceph_remove_cap(cap);
813         return 0;
814 }
815
816 /*
817  * caller must hold session s_mutex
818  */
819 static void remove_session_caps(struct ceph_mds_session *session)
820 {
821         dout("remove_session_caps on %p\n", session);
822         iterate_session_caps(session, remove_session_caps_cb, NULL);
823         BUG_ON(session->s_nr_caps > 0);
824         cleanup_cap_releases(session);
825 }
826
827 /*
828  * wake up any threads waiting on this session's caps.  if the cap is
829  * old (didn't get renewed on the client reconnect), remove it now.
830  *
831  * caller must hold s_mutex.
832  */
833 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
834                               void *arg)
835 {
836         struct ceph_inode_info *ci = ceph_inode(inode);
837
838         wake_up(&ci->i_cap_wq);
839         if (arg) {
840                 spin_lock(&inode->i_lock);
841                 ci->i_wanted_max_size = 0;
842                 ci->i_requested_max_size = 0;
843                 spin_unlock(&inode->i_lock);
844         }
845         return 0;
846 }
847
848 static void wake_up_session_caps(struct ceph_mds_session *session,
849                                  int reconnect)
850 {
851         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
852         iterate_session_caps(session, wake_up_session_cb,
853                              (void *)(unsigned long)reconnect);
854 }
855
856 /*
857  * Send periodic message to MDS renewing all currently held caps.  The
858  * ack will reset the expiration for all caps from this session.
859  *
860  * caller holds s_mutex
861  */
862 static int send_renew_caps(struct ceph_mds_client *mdsc,
863                            struct ceph_mds_session *session)
864 {
865         struct ceph_msg *msg;
866         int state;
867
868         if (time_after_eq(jiffies, session->s_cap_ttl) &&
869             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
870                 pr_info("mds%d caps stale\n", session->s_mds);
871         session->s_renew_requested = jiffies;
872
873         /* do not try to renew caps until a recovering mds has reconnected
874          * with its clients. */
875         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
876         if (state < CEPH_MDS_STATE_RECONNECT) {
877                 dout("send_renew_caps ignoring mds%d (%s)\n",
878                      session->s_mds, ceph_mds_state_name(state));
879                 return 0;
880         }
881
882         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
883                 ceph_mds_state_name(state));
884         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
885                                  ++session->s_renew_seq);
886         if (IS_ERR(msg))
887                 return PTR_ERR(msg);
888         ceph_con_send(&session->s_con, msg);
889         return 0;
890 }
891
892 /*
893  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
894  *
895  * Called under session->s_mutex
896  */
897 static void renewed_caps(struct ceph_mds_client *mdsc,
898                          struct ceph_mds_session *session, int is_renew)
899 {
900         int was_stale;
901         int wake = 0;
902
903         spin_lock(&session->s_cap_lock);
904         was_stale = is_renew && (session->s_cap_ttl == 0 ||
905                                  time_after_eq(jiffies, session->s_cap_ttl));
906
907         session->s_cap_ttl = session->s_renew_requested +
908                 mdsc->mdsmap->m_session_timeout*HZ;
909
910         if (was_stale) {
911                 if (time_before(jiffies, session->s_cap_ttl)) {
912                         pr_info("mds%d caps renewed\n", session->s_mds);
913                         wake = 1;
914                 } else {
915                         pr_info("mds%d caps still stale\n", session->s_mds);
916                 }
917         }
918         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
919              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
920              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
921         spin_unlock(&session->s_cap_lock);
922
923         if (wake)
924                 wake_up_session_caps(session, 0);
925 }
926
927 /*
928  * send a session close request
929  */
930 static int request_close_session(struct ceph_mds_client *mdsc,
931                                  struct ceph_mds_session *session)
932 {
933         struct ceph_msg *msg;
934         int err = 0;
935
936         dout("request_close_session mds%d state %s seq %lld\n",
937              session->s_mds, session_state_name(session->s_state),
938              session->s_seq);
939         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
940         if (IS_ERR(msg))
941                 err = PTR_ERR(msg);
942         else
943                 ceph_con_send(&session->s_con, msg);
944         return err;
945 }
946
947 /*
948  * Called with s_mutex held.
949  */
950 static int __close_session(struct ceph_mds_client *mdsc,
951                          struct ceph_mds_session *session)
952 {
953         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
954                 return 0;
955         session->s_state = CEPH_MDS_SESSION_CLOSING;
956         return request_close_session(mdsc, session);
957 }
958
959 /*
960  * Trim old(er) caps.
961  *
962  * Because we can't cache an inode without one or more caps, we do
963  * this indirectly: if a cap is unused, we prune its aliases, at which
964  * point the inode will hopefully get dropped to.
965  *
966  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
967  * memory pressure from the MDS, though, so it needn't be perfect.
968  */
969 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
970 {
971         struct ceph_mds_session *session = arg;
972         struct ceph_inode_info *ci = ceph_inode(inode);
973         int used, oissued, mine;
974
975         if (session->s_trim_caps <= 0)
976                 return -1;
977
978         spin_lock(&inode->i_lock);
979         mine = cap->issued | cap->implemented;
980         used = __ceph_caps_used(ci);
981         oissued = __ceph_caps_issued_other(ci, cap);
982
983         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
984              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
985              ceph_cap_string(used));
986         if (ci->i_dirty_caps)
987                 goto out;   /* dirty caps */
988         if ((used & ~oissued) & mine)
989                 goto out;   /* we need these caps */
990
991         session->s_trim_caps--;
992         if (oissued) {
993                 /* we aren't the only cap.. just remove us */
994                 __ceph_remove_cap(cap);
995         } else {
996                 /* try to drop referring dentries */
997                 spin_unlock(&inode->i_lock);
998                 d_prune_aliases(inode);
999                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1000                      inode, cap, atomic_read(&inode->i_count));
1001                 return 0;
1002         }
1003
1004 out:
1005         spin_unlock(&inode->i_lock);
1006         return 0;
1007 }
1008
1009 /*
1010  * Trim session cap count down to some max number.
1011  */
1012 static int trim_caps(struct ceph_mds_client *mdsc,
1013                      struct ceph_mds_session *session,
1014                      int max_caps)
1015 {
1016         int trim_caps = session->s_nr_caps - max_caps;
1017
1018         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1019              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1020         if (trim_caps > 0) {
1021                 session->s_trim_caps = trim_caps;
1022                 iterate_session_caps(session, trim_caps_cb, session);
1023                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1024                      session->s_mds, session->s_nr_caps, max_caps,
1025                         trim_caps - session->s_trim_caps);
1026                 session->s_trim_caps = 0;
1027         }
1028         return 0;
1029 }
1030
1031 /*
1032  * Allocate cap_release messages.  If there is a partially full message
1033  * in the queue, try to allocate enough to cover it's remainder, so that
1034  * we can send it immediately.
1035  *
1036  * Called under s_mutex.
1037  */
1038 static int add_cap_releases(struct ceph_mds_client *mdsc,
1039                             struct ceph_mds_session *session,
1040                             int extra)
1041 {
1042         struct ceph_msg *msg;
1043         struct ceph_mds_cap_release *head;
1044         int err = -ENOMEM;
1045
1046         if (extra < 0)
1047                 extra = mdsc->client->mount_args->cap_release_safety;
1048
1049         spin_lock(&session->s_cap_lock);
1050
1051         if (!list_empty(&session->s_cap_releases)) {
1052                 msg = list_first_entry(&session->s_cap_releases,
1053                                        struct ceph_msg,
1054                                  list_head);
1055                 head = msg->front.iov_base;
1056                 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1057         }
1058
1059         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1060                 spin_unlock(&session->s_cap_lock);
1061                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1062                                    0, 0, NULL);
1063                 if (!msg)
1064                         goto out_unlocked;
1065                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1066                      (int)msg->front.iov_len);
1067                 head = msg->front.iov_base;
1068                 head->num = cpu_to_le32(0);
1069                 msg->front.iov_len = sizeof(*head);
1070                 spin_lock(&session->s_cap_lock);
1071                 list_add(&msg->list_head, &session->s_cap_releases);
1072                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1073         }
1074
1075         if (!list_empty(&session->s_cap_releases)) {
1076                 msg = list_first_entry(&session->s_cap_releases,
1077                                        struct ceph_msg,
1078                                        list_head);
1079                 head = msg->front.iov_base;
1080                 if (head->num) {
1081                         dout(" queueing non-full %p (%d)\n", msg,
1082                              le32_to_cpu(head->num));
1083                         list_move_tail(&msg->list_head,
1084                                       &session->s_cap_releases_done);
1085                         session->s_num_cap_releases -=
1086                                 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1087                 }
1088         }
1089         err = 0;
1090         spin_unlock(&session->s_cap_lock);
1091 out_unlocked:
1092         return err;
1093 }
1094
1095 /*
1096  * flush all dirty inode data to disk.
1097  *
1098  * returns true if we've flushed through want_flush_seq
1099  */
1100 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1101 {
1102         int mds, ret = 1;
1103
1104         dout("check_cap_flush want %lld\n", want_flush_seq);
1105         mutex_lock(&mdsc->mutex);
1106         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1107                 struct ceph_mds_session *session = mdsc->sessions[mds];
1108
1109                 if (!session)
1110                         continue;
1111                 get_session(session);
1112                 mutex_unlock(&mdsc->mutex);
1113
1114                 mutex_lock(&session->s_mutex);
1115                 if (!list_empty(&session->s_cap_flushing)) {
1116                         struct ceph_inode_info *ci =
1117                                 list_entry(session->s_cap_flushing.next,
1118                                            struct ceph_inode_info,
1119                                            i_flushing_item);
1120                         struct inode *inode = &ci->vfs_inode;
1121
1122                         spin_lock(&inode->i_lock);
1123                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1124                                 dout("check_cap_flush still flushing %p "
1125                                      "seq %lld <= %lld to mds%d\n", inode,
1126                                      ci->i_cap_flush_seq, want_flush_seq,
1127                                      session->s_mds);
1128                                 ret = 0;
1129                         }
1130                         spin_unlock(&inode->i_lock);
1131                 }
1132                 mutex_unlock(&session->s_mutex);
1133                 ceph_put_mds_session(session);
1134
1135                 if (!ret)
1136                         return ret;
1137                 mutex_lock(&mdsc->mutex);
1138         }
1139
1140         mutex_unlock(&mdsc->mutex);
1141         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1142         return ret;
1143 }
1144
1145 /*
1146  * called under s_mutex
1147  */
1148 static void send_cap_releases(struct ceph_mds_client *mdsc,
1149                        struct ceph_mds_session *session)
1150 {
1151         struct ceph_msg *msg;
1152
1153         dout("send_cap_releases mds%d\n", session->s_mds);
1154         while (1) {
1155                 spin_lock(&session->s_cap_lock);
1156                 if (list_empty(&session->s_cap_releases_done))
1157                         break;
1158                 msg = list_first_entry(&session->s_cap_releases_done,
1159                                  struct ceph_msg, list_head);
1160                 list_del_init(&msg->list_head);
1161                 spin_unlock(&session->s_cap_lock);
1162                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1163                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1164                 ceph_con_send(&session->s_con, msg);
1165         }
1166         spin_unlock(&session->s_cap_lock);
1167 }
1168
1169 /*
1170  * requests
1171  */
1172
1173 /*
1174  * Create an mds request.
1175  */
1176 struct ceph_mds_request *
1177 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1178 {
1179         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1180
1181         if (!req)
1182                 return ERR_PTR(-ENOMEM);
1183
1184         req->r_started = jiffies;
1185         req->r_resend_mds = -1;
1186         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1187         req->r_fmode = -1;
1188         kref_init(&req->r_kref);
1189         INIT_LIST_HEAD(&req->r_wait);
1190         init_completion(&req->r_completion);
1191         init_completion(&req->r_safe_completion);
1192         INIT_LIST_HEAD(&req->r_unsafe_item);
1193
1194         req->r_op = op;
1195         req->r_direct_mode = mode;
1196         return req;
1197 }
1198
1199 /*
1200  * return oldest (lowest) request, tid in request tree, 0 if none.
1201  *
1202  * called under mdsc->mutex.
1203  */
1204 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1205 {
1206         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1207                 return NULL;
1208         return rb_entry(rb_first(&mdsc->request_tree),
1209                         struct ceph_mds_request, r_node);
1210 }
1211
1212 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1213 {
1214         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1215
1216         if (req)
1217                 return req->r_tid;
1218         return 0;
1219 }
1220
1221 /*
1222  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1223  * on build_path_from_dentry in fs/cifs/dir.c.
1224  *
1225  * If @stop_on_nosnap, generate path relative to the first non-snapped
1226  * inode.
1227  *
1228  * Encode hidden .snap dirs as a double /, i.e.
1229  *   foo/.snap/bar -> foo//bar
1230  */
1231 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1232                            int stop_on_nosnap)
1233 {
1234         struct dentry *temp;
1235         char *path;
1236         int len, pos;
1237
1238         if (dentry == NULL)
1239                 return ERR_PTR(-EINVAL);
1240
1241 retry:
1242         len = 0;
1243         for (temp = dentry; !IS_ROOT(temp);) {
1244                 struct inode *inode = temp->d_inode;
1245                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1246                         len++;  /* slash only */
1247                 else if (stop_on_nosnap && inode &&
1248                          ceph_snap(inode) == CEPH_NOSNAP)
1249                         break;
1250                 else
1251                         len += 1 + temp->d_name.len;
1252                 temp = temp->d_parent;
1253                 if (temp == NULL) {
1254                         pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1255                         return ERR_PTR(-EINVAL);
1256                 }
1257         }
1258         if (len)
1259                 len--;  /* no leading '/' */
1260
1261         path = kmalloc(len+1, GFP_NOFS);
1262         if (path == NULL)
1263                 return ERR_PTR(-ENOMEM);
1264         pos = len;
1265         path[pos] = 0;  /* trailing null */
1266         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1267                 struct inode *inode = temp->d_inode;
1268
1269                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1270                         dout("build_path_dentry path+%d: %p SNAPDIR\n",
1271                              pos, temp);
1272                 } else if (stop_on_nosnap && inode &&
1273                            ceph_snap(inode) == CEPH_NOSNAP) {
1274                         break;
1275                 } else {
1276                         pos -= temp->d_name.len;
1277                         if (pos < 0)
1278                                 break;
1279                         strncpy(path + pos, temp->d_name.name,
1280                                 temp->d_name.len);
1281                         dout("build_path_dentry path+%d: %p '%.*s'\n",
1282                              pos, temp, temp->d_name.len, path + pos);
1283                 }
1284                 if (pos)
1285                         path[--pos] = '/';
1286                 temp = temp->d_parent;
1287                 if (temp == NULL) {
1288                         pr_err("build_path_dentry corrupt dentry\n");
1289                         kfree(path);
1290                         return ERR_PTR(-EINVAL);
1291                 }
1292         }
1293         if (pos != 0) {
1294                 pr_err("build_path_dentry did not end path lookup where "
1295                        "expected, namelen is %d, pos is %d\n", len, pos);
1296                 /* presumably this is only possible if racing with a
1297                    rename of one of the parent directories (we can not
1298                    lock the dentries above us to prevent this, but
1299                    retrying should be harmless) */
1300                 kfree(path);
1301                 goto retry;
1302         }
1303
1304         *base = ceph_ino(temp->d_inode);
1305         *plen = len;
1306         dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1307              dentry, atomic_read(&dentry->d_count), *base, len, path);
1308         return path;
1309 }
1310
1311 static int build_dentry_path(struct dentry *dentry,
1312                              const char **ppath, int *ppathlen, u64 *pino,
1313                              int *pfreepath)
1314 {
1315         char *path;
1316
1317         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1318                 *pino = ceph_ino(dentry->d_parent->d_inode);
1319                 *ppath = dentry->d_name.name;
1320                 *ppathlen = dentry->d_name.len;
1321                 return 0;
1322         }
1323         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1324         if (IS_ERR(path))
1325                 return PTR_ERR(path);
1326         *ppath = path;
1327         *pfreepath = 1;
1328         return 0;
1329 }
1330
1331 static int build_inode_path(struct inode *inode,
1332                             const char **ppath, int *ppathlen, u64 *pino,
1333                             int *pfreepath)
1334 {
1335         struct dentry *dentry;
1336         char *path;
1337
1338         if (ceph_snap(inode) == CEPH_NOSNAP) {
1339                 *pino = ceph_ino(inode);
1340                 *ppathlen = 0;
1341                 return 0;
1342         }
1343         dentry = d_find_alias(inode);
1344         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1345         dput(dentry);
1346         if (IS_ERR(path))
1347                 return PTR_ERR(path);
1348         *ppath = path;
1349         *pfreepath = 1;
1350         return 0;
1351 }
1352
1353 /*
1354  * request arguments may be specified via an inode *, a dentry *, or
1355  * an explicit ino+path.
1356  */
1357 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1358                                   const char *rpath, u64 rino,
1359                                   const char **ppath, int *pathlen,
1360                                   u64 *ino, int *freepath)
1361 {
1362         int r = 0;
1363
1364         if (rinode) {
1365                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1366                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1367                      ceph_snap(rinode));
1368         } else if (rdentry) {
1369                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1370                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1371                      *ppath);
1372         } else if (rpath) {
1373                 *ino = rino;
1374                 *ppath = rpath;
1375                 *pathlen = strlen(rpath);
1376                 dout(" path %.*s\n", *pathlen, rpath);
1377         }
1378
1379         return r;
1380 }
1381
1382 /*
1383  * called under mdsc->mutex
1384  */
1385 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1386                                                struct ceph_mds_request *req,
1387                                                int mds)
1388 {
1389         struct ceph_msg *msg;
1390         struct ceph_mds_request_head *head;
1391         const char *path1 = NULL;
1392         const char *path2 = NULL;
1393         u64 ino1 = 0, ino2 = 0;
1394         int pathlen1 = 0, pathlen2 = 0;
1395         int freepath1 = 0, freepath2 = 0;
1396         int len;
1397         u16 releases;
1398         void *p, *end;
1399         int ret;
1400
1401         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1402                               req->r_path1, req->r_ino1.ino,
1403                               &path1, &pathlen1, &ino1, &freepath1);
1404         if (ret < 0) {
1405                 msg = ERR_PTR(ret);
1406                 goto out;
1407         }
1408
1409         ret = set_request_path_attr(NULL, req->r_old_dentry,
1410                               req->r_path2, req->r_ino2.ino,
1411                               &path2, &pathlen2, &ino2, &freepath2);
1412         if (ret < 0) {
1413                 msg = ERR_PTR(ret);
1414                 goto out_free1;
1415         }
1416
1417         len = sizeof(*head) +
1418                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1419
1420         /* calculate (max) length for cap releases */
1421         len += sizeof(struct ceph_mds_request_release) *
1422                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1423                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1424         if (req->r_dentry_drop)
1425                 len += req->r_dentry->d_name.len;
1426         if (req->r_old_dentry_drop)
1427                 len += req->r_old_dentry->d_name.len;
1428
1429         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1430         if (IS_ERR(msg))
1431                 goto out_free2;
1432
1433         msg->hdr.tid = cpu_to_le64(req->r_tid);
1434
1435         head = msg->front.iov_base;
1436         p = msg->front.iov_base + sizeof(*head);
1437         end = msg->front.iov_base + msg->front.iov_len;
1438
1439         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1440         head->op = cpu_to_le32(req->r_op);
1441         head->caller_uid = cpu_to_le32(current_fsuid());
1442         head->caller_gid = cpu_to_le32(current_fsgid());
1443         head->args = req->r_args;
1444
1445         ceph_encode_filepath(&p, end, ino1, path1);
1446         ceph_encode_filepath(&p, end, ino2, path2);
1447
1448         /* cap releases */
1449         releases = 0;
1450         if (req->r_inode_drop)
1451                 releases += ceph_encode_inode_release(&p,
1452                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1453                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1454         if (req->r_dentry_drop)
1455                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1456                        mds, req->r_dentry_drop, req->r_dentry_unless);
1457         if (req->r_old_dentry_drop)
1458                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1459                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1460         if (req->r_old_inode_drop)
1461                 releases += ceph_encode_inode_release(&p,
1462                       req->r_old_dentry->d_inode,
1463                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1464         head->num_releases = cpu_to_le16(releases);
1465
1466         BUG_ON(p > end);
1467         msg->front.iov_len = p - msg->front.iov_base;
1468         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1469
1470         msg->pages = req->r_pages;
1471         msg->nr_pages = req->r_num_pages;
1472         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1473         msg->hdr.data_off = cpu_to_le16(0);
1474
1475 out_free2:
1476         if (freepath2)
1477                 kfree((char *)path2);
1478 out_free1:
1479         if (freepath1)
1480                 kfree((char *)path1);
1481 out:
1482         return msg;
1483 }
1484
1485 /*
1486  * called under mdsc->mutex if error, under no mutex if
1487  * success.
1488  */
1489 static void complete_request(struct ceph_mds_client *mdsc,
1490                              struct ceph_mds_request *req)
1491 {
1492         if (req->r_callback)
1493                 req->r_callback(mdsc, req);
1494         else
1495                 complete(&req->r_completion);
1496 }
1497
1498 /*
1499  * called under mdsc->mutex
1500  */
1501 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1502                                   struct ceph_mds_request *req,
1503                                   int mds)
1504 {
1505         struct ceph_mds_request_head *rhead;
1506         struct ceph_msg *msg;
1507         int flags = 0;
1508
1509         req->r_mds = mds;
1510         req->r_attempts++;
1511         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1512              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1513
1514         if (req->r_request) {
1515                 ceph_msg_put(req->r_request);
1516                 req->r_request = NULL;
1517         }
1518         msg = create_request_message(mdsc, req, mds);
1519         if (IS_ERR(msg)) {
1520                 req->r_reply = ERR_PTR(PTR_ERR(msg));
1521                 complete_request(mdsc, req);
1522                 return -PTR_ERR(msg);
1523         }
1524         req->r_request = msg;
1525
1526         rhead = msg->front.iov_base;
1527         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1528         if (req->r_got_unsafe)
1529                 flags |= CEPH_MDS_FLAG_REPLAY;
1530         if (req->r_locked_dir)
1531                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1532         rhead->flags = cpu_to_le32(flags);
1533         rhead->num_fwd = req->r_num_fwd;
1534         rhead->num_retry = req->r_attempts - 1;
1535
1536         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1537
1538         if (req->r_target_inode && req->r_got_unsafe)
1539                 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1540         else
1541                 rhead->ino = 0;
1542         return 0;
1543 }
1544
1545 /*
1546  * send request, or put it on the appropriate wait list.
1547  */
1548 static int __do_request(struct ceph_mds_client *mdsc,
1549                         struct ceph_mds_request *req)
1550 {
1551         struct ceph_mds_session *session = NULL;
1552         int mds = -1;
1553         int err = -EAGAIN;
1554
1555         if (req->r_reply)
1556                 goto out;
1557
1558         if (req->r_timeout &&
1559             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1560                 dout("do_request timed out\n");
1561                 err = -EIO;
1562                 goto finish;
1563         }
1564
1565         mds = __choose_mds(mdsc, req);
1566         if (mds < 0 ||
1567             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1568                 dout("do_request no mds or not active, waiting for map\n");
1569                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1570                 goto out;
1571         }
1572
1573         /* get, open session */
1574         session = __ceph_lookup_mds_session(mdsc, mds);
1575         if (!session) {
1576                 session = register_session(mdsc, mds);
1577                 if (IS_ERR(session)) {
1578                         err = PTR_ERR(session);
1579                         goto finish;
1580                 }
1581         }
1582         dout("do_request mds%d session %p state %s\n", mds, session,
1583              session_state_name(session->s_state));
1584         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1585             session->s_state != CEPH_MDS_SESSION_HUNG) {
1586                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1587                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1588                         __open_session(mdsc, session);
1589                 list_add(&req->r_wait, &session->s_waiting);
1590                 goto out_session;
1591         }
1592
1593         /* send request */
1594         req->r_session = get_session(session);
1595         req->r_resend_mds = -1;   /* forget any previous mds hint */
1596
1597         if (req->r_request_started == 0)   /* note request start time */
1598                 req->r_request_started = jiffies;
1599
1600         err = __prepare_send_request(mdsc, req, mds);
1601         if (!err) {
1602                 ceph_msg_get(req->r_request);
1603                 ceph_con_send(&session->s_con, req->r_request);
1604         }
1605
1606 out_session:
1607         ceph_put_mds_session(session);
1608 out:
1609         return err;
1610
1611 finish:
1612         req->r_reply = ERR_PTR(err);
1613         complete_request(mdsc, req);
1614         goto out;
1615 }
1616
1617 /*
1618  * called under mdsc->mutex
1619  */
1620 static void __wake_requests(struct ceph_mds_client *mdsc,
1621                             struct list_head *head)
1622 {
1623         struct ceph_mds_request *req, *nreq;
1624
1625         list_for_each_entry_safe(req, nreq, head, r_wait) {
1626                 list_del_init(&req->r_wait);
1627                 __do_request(mdsc, req);
1628         }
1629 }
1630
1631 /*
1632  * Wake up threads with requests pending for @mds, so that they can
1633  * resubmit their requests to a possibly different mds.  If @all is set,
1634  * wake up if their requests has been forwarded to @mds, too.
1635  */
1636 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1637 {
1638         struct ceph_mds_request *req;
1639         struct rb_node *p;
1640
1641         dout("kick_requests mds%d\n", mds);
1642         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1643                 req = rb_entry(p, struct ceph_mds_request, r_node);
1644                 if (req->r_got_unsafe)
1645                         continue;
1646                 if (req->r_session &&
1647                     req->r_session->s_mds == mds) {
1648                         dout(" kicking tid %llu\n", req->r_tid);
1649                         put_request_session(req);
1650                         __do_request(mdsc, req);
1651                 }
1652         }
1653 }
1654
1655 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1656                               struct ceph_mds_request *req)
1657 {
1658         dout("submit_request on %p\n", req);
1659         mutex_lock(&mdsc->mutex);
1660         __register_request(mdsc, req, NULL);
1661         __do_request(mdsc, req);
1662         mutex_unlock(&mdsc->mutex);
1663 }
1664
1665 /*
1666  * Synchrously perform an mds request.  Take care of all of the
1667  * session setup, forwarding, retry details.
1668  */
1669 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1670                          struct inode *dir,
1671                          struct ceph_mds_request *req)
1672 {
1673         int err;
1674
1675         dout("do_request on %p\n", req);
1676
1677         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1678         if (req->r_inode)
1679                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1680         if (req->r_locked_dir)
1681                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1682         if (req->r_old_dentry)
1683                 ceph_get_cap_refs(
1684                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1685                         CEPH_CAP_PIN);
1686
1687         /* issue */
1688         mutex_lock(&mdsc->mutex);
1689         __register_request(mdsc, req, dir);
1690         __do_request(mdsc, req);
1691
1692         /* wait */
1693         if (!req->r_reply) {
1694                 mutex_unlock(&mdsc->mutex);
1695                 if (req->r_timeout) {
1696                         err = (long)wait_for_completion_interruptible_timeout(
1697                                 &req->r_completion, req->r_timeout);
1698                         if (err == 0)
1699                                 req->r_reply = ERR_PTR(-EIO);
1700                         else if (err < 0)
1701                                 req->r_reply = ERR_PTR(err);
1702                 } else {
1703                         err = wait_for_completion_interruptible(
1704                                 &req->r_completion);
1705                         if (err)
1706                                 req->r_reply = ERR_PTR(err);
1707                 }
1708                 mutex_lock(&mdsc->mutex);
1709         }
1710
1711         if (IS_ERR(req->r_reply)) {
1712                 err = PTR_ERR(req->r_reply);
1713                 req->r_reply = NULL;
1714
1715                 if (err == -ERESTARTSYS) {
1716                         /* aborted */
1717                         req->r_aborted = true;
1718
1719                         if (req->r_locked_dir &&
1720                             (req->r_op & CEPH_MDS_OP_WRITE)) {
1721                                 struct ceph_inode_info *ci =
1722                                         ceph_inode(req->r_locked_dir);
1723
1724                                 dout("aborted, clearing I_COMPLETE on %p\n", 
1725                                      req->r_locked_dir);
1726                                 spin_lock(&req->r_locked_dir->i_lock);
1727                                 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1728                                 ci->i_release_count++;
1729                                 spin_unlock(&req->r_locked_dir->i_lock);
1730                         }
1731                 } else {
1732                         /* clean up this request */
1733                         __unregister_request(mdsc, req);
1734                         if (!list_empty(&req->r_unsafe_item))
1735                                 list_del_init(&req->r_unsafe_item);
1736                         complete(&req->r_safe_completion);
1737                 }
1738         } else if (req->r_err) {
1739                 err = req->r_err;
1740         } else {
1741                 err = le32_to_cpu(req->r_reply_info.head->result);
1742         }
1743         mutex_unlock(&mdsc->mutex);
1744
1745         dout("do_request %p done, result %d\n", req, err);
1746         return err;
1747 }
1748
1749 /*
1750  * Handle mds reply.
1751  *
1752  * We take the session mutex and parse and process the reply immediately.
1753  * This preserves the logical ordering of replies, capabilities, etc., sent
1754  * by the MDS as they are applied to our local cache.
1755  */
1756 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1757 {
1758         struct ceph_mds_client *mdsc = session->s_mdsc;
1759         struct ceph_mds_request *req;
1760         struct ceph_mds_reply_head *head = msg->front.iov_base;
1761         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1762         u64 tid;
1763         int err, result;
1764         int mds = session->s_mds;
1765
1766         if (msg->front.iov_len < sizeof(*head)) {
1767                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1768                 ceph_msg_dump(msg);
1769                 return;
1770         }
1771
1772         /* get request, session */
1773         tid = le64_to_cpu(msg->hdr.tid);
1774         mutex_lock(&mdsc->mutex);
1775         req = __lookup_request(mdsc, tid);
1776         if (!req) {
1777                 dout("handle_reply on unknown tid %llu\n", tid);
1778                 mutex_unlock(&mdsc->mutex);
1779                 return;
1780         }
1781         dout("handle_reply %p\n", req);
1782
1783         /* correct session? */
1784         if (req->r_session != session) {
1785                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1786                        " not mds%d\n", tid, session->s_mds,
1787                        req->r_session ? req->r_session->s_mds : -1);
1788                 mutex_unlock(&mdsc->mutex);
1789                 goto out;
1790         }
1791
1792         /* dup? */
1793         if ((req->r_got_unsafe && !head->safe) ||
1794             (req->r_got_safe && head->safe)) {
1795                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1796                            head->safe ? "safe" : "unsafe", tid, mds);
1797                 mutex_unlock(&mdsc->mutex);
1798                 goto out;
1799         }
1800
1801         result = le32_to_cpu(head->result);
1802
1803         /*
1804          * Tolerate 2 consecutive ESTALEs from the same mds.
1805          * FIXME: we should be looking at the cap migrate_seq.
1806          */
1807         if (result == -ESTALE) {
1808                 req->r_direct_mode = USE_AUTH_MDS;
1809                 req->r_num_stale++;
1810                 if (req->r_num_stale <= 2) {
1811                         __do_request(mdsc, req);
1812                         mutex_unlock(&mdsc->mutex);
1813                         goto out;
1814                 }
1815         } else {
1816                 req->r_num_stale = 0;
1817         }
1818
1819         if (head->safe) {
1820                 req->r_got_safe = true;
1821                 __unregister_request(mdsc, req);
1822                 complete(&req->r_safe_completion);
1823
1824                 if (req->r_got_unsafe) {
1825                         /*
1826                          * We already handled the unsafe response, now do the
1827                          * cleanup.  No need to examine the response; the MDS
1828                          * doesn't include any result info in the safe
1829                          * response.  And even if it did, there is nothing
1830                          * useful we could do with a revised return value.
1831                          */
1832                         dout("got safe reply %llu, mds%d\n", tid, mds);
1833                         list_del_init(&req->r_unsafe_item);
1834
1835                         /* last unsafe request during umount? */
1836                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1837                                 complete(&mdsc->safe_umount_waiters);
1838                         mutex_unlock(&mdsc->mutex);
1839                         goto out;
1840                 }
1841         }
1842
1843         BUG_ON(req->r_reply);
1844
1845         if (!head->safe) {
1846                 req->r_got_unsafe = true;
1847                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1848         }
1849
1850         dout("handle_reply tid %lld result %d\n", tid, result);
1851         rinfo = &req->r_reply_info;
1852         err = parse_reply_info(msg, rinfo);
1853         mutex_unlock(&mdsc->mutex);
1854
1855         mutex_lock(&session->s_mutex);
1856         if (err < 0) {
1857                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1858                 ceph_msg_dump(msg);
1859                 goto out_err;
1860         }
1861
1862         /* snap trace */
1863         if (rinfo->snapblob_len) {
1864                 down_write(&mdsc->snap_rwsem);
1865                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1866                                rinfo->snapblob + rinfo->snapblob_len,
1867                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1868                 downgrade_write(&mdsc->snap_rwsem);
1869         } else {
1870                 down_read(&mdsc->snap_rwsem);
1871         }
1872
1873         /* insert trace into our cache */
1874         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1875         if (err == 0) {
1876                 if (result == 0 && rinfo->dir_nr)
1877                         ceph_readdir_prepopulate(req, req->r_session);
1878                 ceph_unreserve_caps(&req->r_caps_reservation);
1879         }
1880
1881         up_read(&mdsc->snap_rwsem);
1882 out_err:
1883         if (err) {
1884                 req->r_err = err;
1885         } else {
1886                 req->r_reply = msg;
1887                 ceph_msg_get(msg);
1888         }
1889
1890         add_cap_releases(mdsc, req->r_session, -1);
1891         mutex_unlock(&session->s_mutex);
1892
1893         /* kick calling process */
1894         complete_request(mdsc, req);
1895 out:
1896         ceph_mdsc_put_request(req);
1897         return;
1898 }
1899
1900
1901
1902 /*
1903  * handle mds notification that our request has been forwarded.
1904  */
1905 static void handle_forward(struct ceph_mds_client *mdsc,
1906                            struct ceph_mds_session *session,
1907                            struct ceph_msg *msg)
1908 {
1909         struct ceph_mds_request *req;
1910         u64 tid = le64_to_cpu(msg->hdr.tid);
1911         u32 next_mds;
1912         u32 fwd_seq;
1913         int err = -EINVAL;
1914         void *p = msg->front.iov_base;
1915         void *end = p + msg->front.iov_len;
1916
1917         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1918         next_mds = ceph_decode_32(&p);
1919         fwd_seq = ceph_decode_32(&p);
1920
1921         mutex_lock(&mdsc->mutex);
1922         req = __lookup_request(mdsc, tid);
1923         if (!req) {
1924                 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1925                 goto out;  /* dup reply? */
1926         }
1927
1928         if (fwd_seq <= req->r_num_fwd) {
1929                 dout("forward %llu to mds%d - old seq %d <= %d\n",
1930                      tid, next_mds, req->r_num_fwd, fwd_seq);
1931         } else {
1932                 /* resend. forward race not possible; mds would drop */
1933                 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1934                 req->r_num_fwd = fwd_seq;
1935                 req->r_resend_mds = next_mds;
1936                 put_request_session(req);
1937                 __do_request(mdsc, req);
1938         }
1939         ceph_mdsc_put_request(req);
1940 out:
1941         mutex_unlock(&mdsc->mutex);
1942         return;
1943
1944 bad:
1945         pr_err("mdsc_handle_forward decode error err=%d\n", err);
1946 }
1947
1948 /*
1949  * handle a mds session control message
1950  */
1951 static void handle_session(struct ceph_mds_session *session,
1952                            struct ceph_msg *msg)
1953 {
1954         struct ceph_mds_client *mdsc = session->s_mdsc;
1955         u32 op;
1956         u64 seq;
1957         int mds = session->s_mds;
1958         struct ceph_mds_session_head *h = msg->front.iov_base;
1959         int wake = 0;
1960
1961         /* decode */
1962         if (msg->front.iov_len != sizeof(*h))
1963                 goto bad;
1964         op = le32_to_cpu(h->op);
1965         seq = le64_to_cpu(h->seq);
1966
1967         mutex_lock(&mdsc->mutex);
1968         if (op == CEPH_SESSION_CLOSE)
1969                 __unregister_session(mdsc, session);
1970         /* FIXME: this ttl calculation is generous */
1971         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1972         mutex_unlock(&mdsc->mutex);
1973
1974         mutex_lock(&session->s_mutex);
1975
1976         dout("handle_session mds%d %s %p state %s seq %llu\n",
1977              mds, ceph_session_op_name(op), session,
1978              session_state_name(session->s_state), seq);
1979
1980         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1981                 session->s_state = CEPH_MDS_SESSION_OPEN;
1982                 pr_info("mds%d came back\n", session->s_mds);
1983         }
1984
1985         switch (op) {
1986         case CEPH_SESSION_OPEN:
1987                 session->s_state = CEPH_MDS_SESSION_OPEN;
1988                 renewed_caps(mdsc, session, 0);
1989                 wake = 1;
1990                 if (mdsc->stopping)
1991                         __close_session(mdsc, session);
1992                 break;
1993
1994         case CEPH_SESSION_RENEWCAPS:
1995                 if (session->s_renew_seq == seq)
1996                         renewed_caps(mdsc, session, 1);
1997                 break;
1998
1999         case CEPH_SESSION_CLOSE:
2000                 remove_session_caps(session);
2001                 wake = 1; /* for good measure */
2002                 complete(&mdsc->session_close_waiters);
2003                 kick_requests(mdsc, mds, 0);      /* cur only */
2004                 break;
2005
2006         case CEPH_SESSION_STALE:
2007                 pr_info("mds%d caps went stale, renewing\n",
2008                         session->s_mds);
2009                 spin_lock(&session->s_cap_lock);
2010                 session->s_cap_gen++;
2011                 session->s_cap_ttl = 0;
2012                 spin_unlock(&session->s_cap_lock);
2013                 send_renew_caps(mdsc, session);
2014                 break;
2015
2016         case CEPH_SESSION_RECALL_STATE:
2017                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2018                 break;
2019
2020         default:
2021                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2022                 WARN_ON(1);
2023         }
2024
2025         mutex_unlock(&session->s_mutex);
2026         if (wake) {
2027                 mutex_lock(&mdsc->mutex);
2028                 __wake_requests(mdsc, &session->s_waiting);
2029                 mutex_unlock(&mdsc->mutex);
2030         }
2031         return;
2032
2033 bad:
2034         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2035                (int)msg->front.iov_len);
2036         ceph_msg_dump(msg);
2037         return;
2038 }
2039
2040
2041 /*
2042  * called under session->mutex.
2043  */
2044 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2045                                    struct ceph_mds_session *session)
2046 {
2047         struct ceph_mds_request *req, *nreq;
2048         int err;
2049
2050         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2051
2052         mutex_lock(&mdsc->mutex);
2053         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2054                 err = __prepare_send_request(mdsc, req, session->s_mds);
2055                 if (!err) {
2056                         ceph_msg_get(req->r_request);
2057                         ceph_con_send(&session->s_con, req->r_request);
2058                 }
2059         }
2060         mutex_unlock(&mdsc->mutex);
2061 }
2062
2063 /*
2064  * Encode information about a cap for a reconnect with the MDS.
2065  */
2066 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2067                           void *arg)
2068 {
2069         struct ceph_mds_cap_reconnect rec;
2070         struct ceph_inode_info *ci;
2071         struct ceph_pagelist *pagelist = arg;
2072         char *path;
2073         int pathlen, err;
2074         u64 pathbase;
2075         struct dentry *dentry;
2076
2077         ci = cap->ci;
2078
2079         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2080              inode, ceph_vinop(inode), cap, cap->cap_id,
2081              ceph_cap_string(cap->issued));
2082         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2083         if (err)
2084                 return err;
2085
2086         dentry = d_find_alias(inode);
2087         if (dentry) {
2088                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2089                 if (IS_ERR(path)) {
2090                         err = PTR_ERR(path);
2091                         BUG_ON(err);
2092                 }
2093         } else {
2094                 path = NULL;
2095                 pathlen = 0;
2096         }
2097         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2098         if (err)
2099                 goto out;
2100
2101         spin_lock(&inode->i_lock);
2102         cap->seq = 0;        /* reset cap seq */
2103         cap->issue_seq = 0;  /* and issue_seq */
2104         rec.cap_id = cpu_to_le64(cap->cap_id);
2105         rec.pathbase = cpu_to_le64(pathbase);
2106         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2107         rec.issued = cpu_to_le32(cap->issued);
2108         rec.size = cpu_to_le64(inode->i_size);
2109         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2110         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2111         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2112         spin_unlock(&inode->i_lock);
2113
2114         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2115
2116 out:
2117         kfree(path);
2118         dput(dentry);
2119         return err;
2120 }
2121
2122
2123 /*
2124  * If an MDS fails and recovers, clients need to reconnect in order to
2125  * reestablish shared state.  This includes all caps issued through
2126  * this session _and_ the snap_realm hierarchy.  Because it's not
2127  * clear which snap realms the mds cares about, we send everything we
2128  * know about.. that ensures we'll then get any new info the
2129  * recovering MDS might have.
2130  *
2131  * This is a relatively heavyweight operation, but it's rare.
2132  *
2133  * called with mdsc->mutex held.
2134  */
2135 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2136 {
2137         struct ceph_mds_session *session = NULL;
2138         struct ceph_msg *reply;
2139         struct rb_node *p;
2140         int err = -ENOMEM;
2141         struct ceph_pagelist *pagelist;
2142
2143         pr_info("reconnect to recovering mds%d\n", mds);
2144
2145         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2146         if (!pagelist)
2147                 goto fail_nopagelist;
2148         ceph_pagelist_init(pagelist);
2149
2150         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2151         if (IS_ERR(reply)) {
2152                 err = PTR_ERR(reply);
2153                 goto fail_nomsg;
2154         }
2155
2156         /* find session */
2157         session = __ceph_lookup_mds_session(mdsc, mds);
2158         mutex_unlock(&mdsc->mutex);    /* drop lock for duration */
2159
2160         if (session) {
2161                 mutex_lock(&session->s_mutex);
2162
2163                 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2164                 session->s_seq = 0;
2165
2166                 ceph_con_open(&session->s_con,
2167                               ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2168
2169                 /* replay unsafe requests */
2170                 replay_unsafe_requests(mdsc, session);
2171         } else {
2172                 dout("no session for mds%d, will send short reconnect\n",
2173                      mds);
2174         }
2175
2176         down_read(&mdsc->snap_rwsem);
2177
2178         if (!session)
2179                 goto send;
2180         dout("session %p state %s\n", session,
2181              session_state_name(session->s_state));
2182
2183         /* traverse this session's caps */
2184         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2185         if (err)
2186                 goto fail;
2187         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2188         if (err < 0)
2189                 goto fail;
2190
2191         /*
2192          * snaprealms.  we provide mds with the ino, seq (version), and
2193          * parent for all of our realms.  If the mds has any newer info,
2194          * it will tell us.
2195          */
2196         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2197                 struct ceph_snap_realm *realm =
2198                         rb_entry(p, struct ceph_snap_realm, node);
2199                 struct ceph_mds_snaprealm_reconnect sr_rec;
2200
2201                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2202                      realm->ino, realm->seq, realm->parent_ino);
2203                 sr_rec.ino = cpu_to_le64(realm->ino);
2204                 sr_rec.seq = cpu_to_le64(realm->seq);
2205                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2206                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2207                 if (err)
2208                         goto fail;
2209         }
2210
2211 send:
2212         reply->pagelist = pagelist;
2213         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2214         reply->nr_pages = calc_pages_for(0, pagelist->length);
2215         ceph_con_send(&session->s_con, reply);
2216
2217         session->s_state = CEPH_MDS_SESSION_OPEN;
2218         mutex_unlock(&session->s_mutex);
2219
2220         mutex_lock(&mdsc->mutex);
2221         __wake_requests(mdsc, &session->s_waiting);
2222         mutex_unlock(&mdsc->mutex);
2223
2224         ceph_put_mds_session(session);
2225
2226         up_read(&mdsc->snap_rwsem);
2227         mutex_lock(&mdsc->mutex);
2228         return;
2229
2230 fail:
2231         ceph_msg_put(reply);
2232         up_read(&mdsc->snap_rwsem);
2233         mutex_unlock(&session->s_mutex);
2234         ceph_put_mds_session(session);
2235 fail_nomsg:
2236         ceph_pagelist_release(pagelist);
2237         kfree(pagelist);
2238 fail_nopagelist:
2239         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2240         mutex_lock(&mdsc->mutex);
2241         return;
2242 }
2243
2244
2245 /*
2246  * compare old and new mdsmaps, kicking requests
2247  * and closing out old connections as necessary
2248  *
2249  * called under mdsc->mutex.
2250  */
2251 static void check_new_map(struct ceph_mds_client *mdsc,
2252                           struct ceph_mdsmap *newmap,
2253                           struct ceph_mdsmap *oldmap)
2254 {
2255         int i;
2256         int oldstate, newstate;
2257         struct ceph_mds_session *s;
2258
2259         dout("check_new_map new %u old %u\n",
2260              newmap->m_epoch, oldmap->m_epoch);
2261
2262         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2263                 if (mdsc->sessions[i] == NULL)
2264                         continue;
2265                 s = mdsc->sessions[i];
2266                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2267                 newstate = ceph_mdsmap_get_state(newmap, i);
2268
2269                 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2270                      i, ceph_mds_state_name(oldstate),
2271                      ceph_mds_state_name(newstate),
2272                      session_state_name(s->s_state));
2273
2274                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2275                            ceph_mdsmap_get_addr(newmap, i),
2276                            sizeof(struct ceph_entity_addr))) {
2277                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2278                                 /* the session never opened, just close it
2279                                  * out now */
2280                                 __wake_requests(mdsc, &s->s_waiting);
2281                                 __unregister_session(mdsc, s);
2282                         } else {
2283                                 /* just close it */
2284                                 mutex_unlock(&mdsc->mutex);
2285                                 mutex_lock(&s->s_mutex);
2286                                 mutex_lock(&mdsc->mutex);
2287                                 ceph_con_close(&s->s_con);
2288                                 mutex_unlock(&s->s_mutex);
2289                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2290                         }
2291
2292                         /* kick any requests waiting on the recovering mds */
2293                         kick_requests(mdsc, i, 1);
2294                 } else if (oldstate == newstate) {
2295                         continue;  /* nothing new with this mds */
2296                 }
2297
2298                 /*
2299                  * send reconnect?
2300                  */
2301                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2302                     newstate >= CEPH_MDS_STATE_RECONNECT)
2303                         send_mds_reconnect(mdsc, i);
2304
2305                 /*
2306                  * kick requests on any mds that has gone active.
2307                  *
2308                  * kick requests on cur or forwarder: we may have sent
2309                  * the request to mds1, mds1 told us it forwarded it
2310                  * to mds2, but then we learn mds1 failed and can't be
2311                  * sure it successfully forwarded our request before
2312                  * it died.
2313                  */
2314                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2315                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2316                         pr_info("mds%d reconnect completed\n", s->s_mds);
2317                         kick_requests(mdsc, i, 1);
2318                         ceph_kick_flushing_caps(mdsc, s);
2319                         wake_up_session_caps(s, 1);
2320                 }
2321         }
2322 }
2323
2324
2325
2326 /*
2327  * leases
2328  */
2329
2330 /*
2331  * caller must hold session s_mutex, dentry->d_lock
2332  */
2333 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2334 {
2335         struct ceph_dentry_info *di = ceph_dentry(dentry);
2336
2337         ceph_put_mds_session(di->lease_session);
2338         di->lease_session = NULL;
2339 }
2340
2341 static void handle_lease(struct ceph_mds_client *mdsc,
2342                          struct ceph_mds_session *session,
2343                          struct ceph_msg *msg)
2344 {
2345         struct super_block *sb = mdsc->client->sb;
2346         struct inode *inode;
2347         struct ceph_inode_info *ci;
2348         struct dentry *parent, *dentry;
2349         struct ceph_dentry_info *di;
2350         int mds = session->s_mds;
2351         struct ceph_mds_lease *h = msg->front.iov_base;
2352         struct ceph_vino vino;
2353         int mask;
2354         struct qstr dname;
2355         int release = 0;
2356
2357         dout("handle_lease from mds%d\n", mds);
2358
2359         /* decode */
2360         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2361                 goto bad;
2362         vino.ino = le64_to_cpu(h->ino);
2363         vino.snap = CEPH_NOSNAP;
2364         mask = le16_to_cpu(h->mask);
2365         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2366         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2367         if (dname.len != get_unaligned_le32(h+1))
2368                 goto bad;
2369
2370         mutex_lock(&session->s_mutex);
2371         session->s_seq++;
2372
2373         /* lookup inode */
2374         inode = ceph_find_inode(sb, vino);
2375         dout("handle_lease '%s', mask %d, ino %llx %p\n",
2376              ceph_lease_op_name(h->action), mask, vino.ino, inode);
2377         if (inode == NULL) {
2378                 dout("handle_lease no inode %llx\n", vino.ino);
2379                 goto release;
2380         }
2381         ci = ceph_inode(inode);
2382
2383         /* dentry */
2384         parent = d_find_alias(inode);
2385         if (!parent) {
2386                 dout("no parent dentry on inode %p\n", inode);
2387                 WARN_ON(1);
2388                 goto release;  /* hrm... */
2389         }
2390         dname.hash = full_name_hash(dname.name, dname.len);
2391         dentry = d_lookup(parent, &dname);
2392         dput(parent);
2393         if (!dentry)
2394                 goto release;
2395
2396         spin_lock(&dentry->d_lock);
2397         di = ceph_dentry(dentry);
2398         switch (h->action) {
2399         case CEPH_MDS_LEASE_REVOKE:
2400                 if (di && di->lease_session == session) {
2401                         h->seq = cpu_to_le32(di->lease_seq);
2402                         __ceph_mdsc_drop_dentry_lease(dentry);
2403                 }
2404                 release = 1;
2405                 break;
2406
2407         case CEPH_MDS_LEASE_RENEW:
2408                 if (di && di->lease_session == session &&
2409                     di->lease_gen == session->s_cap_gen &&
2410                     di->lease_renew_from &&
2411                     di->lease_renew_after == 0) {
2412                         unsigned long duration =
2413                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2414
2415                         di->lease_seq = le32_to_cpu(h->seq);
2416                         dentry->d_time = di->lease_renew_from + duration;
2417                         di->lease_renew_after = di->lease_renew_from +
2418                                 (duration >> 1);
2419                         di->lease_renew_from = 0;
2420                 }
2421                 break;
2422         }
2423         spin_unlock(&dentry->d_lock);
2424         dput(dentry);
2425
2426         if (!release)
2427                 goto out;
2428
2429 release:
2430         /* let's just reuse the same message */
2431         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2432         ceph_msg_get(msg);
2433         ceph_con_send(&session->s_con, msg);
2434
2435 out:
2436         iput(inode);
2437         mutex_unlock(&session->s_mutex);
2438         return;
2439
2440 bad:
2441         pr_err("corrupt lease message\n");
2442         ceph_msg_dump(msg);
2443 }
2444
2445 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2446                               struct inode *inode,
2447                               struct dentry *dentry, char action,
2448                               u32 seq)
2449 {
2450         struct ceph_msg *msg;
2451         struct ceph_mds_lease *lease;
2452         int len = sizeof(*lease) + sizeof(u32);
2453         int dnamelen = 0;
2454
2455         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2456              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2457         dnamelen = dentry->d_name.len;
2458         len += dnamelen;
2459
2460         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2461         if (IS_ERR(msg))
2462                 return;
2463         lease = msg->front.iov_base;
2464         lease->action = action;
2465         lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2466         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2467         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2468         lease->seq = cpu_to_le32(seq);
2469         put_unaligned_le32(dnamelen, lease + 1);
2470         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2471
2472         /*
2473          * if this is a preemptive lease RELEASE, no need to
2474          * flush request stream, since the actual request will
2475          * soon follow.
2476          */
2477         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2478
2479         ceph_con_send(&session->s_con, msg);
2480 }
2481
2482 /*
2483  * Preemptively release a lease we expect to invalidate anyway.
2484  * Pass @inode always, @dentry is optional.
2485  */
2486 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2487                              struct dentry *dentry, int mask)
2488 {
2489         struct ceph_dentry_info *di;
2490         struct ceph_mds_session *session;
2491         u32 seq;
2492
2493         BUG_ON(inode == NULL);
2494         BUG_ON(dentry == NULL);
2495         BUG_ON(mask != CEPH_LOCK_DN);
2496
2497         /* is dentry lease valid? */
2498         spin_lock(&dentry->d_lock);
2499         di = ceph_dentry(dentry);
2500         if (!di || !di->lease_session ||
2501             di->lease_session->s_mds < 0 ||
2502             di->lease_gen != di->lease_session->s_cap_gen ||
2503             !time_before(jiffies, dentry->d_time)) {
2504                 dout("lease_release inode %p dentry %p -- "
2505                      "no lease on %d\n",
2506                      inode, dentry, mask);
2507                 spin_unlock(&dentry->d_lock);
2508                 return;
2509         }
2510
2511         /* we do have a lease on this dentry; note mds and seq */
2512         session = ceph_get_mds_session(di->lease_session);
2513         seq = di->lease_seq;
2514         __ceph_mdsc_drop_dentry_lease(dentry);
2515         spin_unlock(&dentry->d_lock);
2516
2517         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2518              inode, dentry, mask, session->s_mds);
2519         ceph_mdsc_lease_send_msg(session, inode, dentry,
2520                                  CEPH_MDS_LEASE_RELEASE, seq);
2521         ceph_put_mds_session(session);
2522 }
2523
2524 /*
2525  * drop all leases (and dentry refs) in preparation for umount
2526  */
2527 static void drop_leases(struct ceph_mds_client *mdsc)
2528 {
2529         int i;
2530
2531         dout("drop_leases\n");
2532         mutex_lock(&mdsc->mutex);
2533         for (i = 0; i < mdsc->max_sessions; i++) {
2534                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2535                 if (!s)
2536                         continue;
2537                 mutex_unlock(&mdsc->mutex);
2538                 mutex_lock(&s->s_mutex);
2539                 mutex_unlock(&s->s_mutex);
2540                 ceph_put_mds_session(s);
2541                 mutex_lock(&mdsc->mutex);
2542         }
2543         mutex_unlock(&mdsc->mutex);
2544 }
2545
2546
2547
2548 /*
2549  * delayed work -- periodically trim expired leases, renew caps with mds
2550  */
2551 static void schedule_delayed(struct ceph_mds_client *mdsc)
2552 {
2553         int delay = 5;
2554         unsigned hz = round_jiffies_relative(HZ * delay);
2555         schedule_delayed_work(&mdsc->delayed_work, hz);
2556 }
2557
2558 static void delayed_work(struct work_struct *work)
2559 {
2560         int i;
2561         struct ceph_mds_client *mdsc =
2562                 container_of(work, struct ceph_mds_client, delayed_work.work);
2563         int renew_interval;
2564         int renew_caps;
2565
2566         dout("mdsc delayed_work\n");
2567         ceph_check_delayed_caps(mdsc);
2568
2569         mutex_lock(&mdsc->mutex);
2570         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2571         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2572                                    mdsc->last_renew_caps);
2573         if (renew_caps)
2574                 mdsc->last_renew_caps = jiffies;
2575
2576         for (i = 0; i < mdsc->max_sessions; i++) {
2577                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2578                 if (s == NULL)
2579                         continue;
2580                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2581                         dout("resending session close request for mds%d\n",
2582                              s->s_mds);
2583                         request_close_session(mdsc, s);
2584                         ceph_put_mds_session(s);
2585                         continue;
2586                 }
2587                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2588                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2589                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2590                                 pr_info("mds%d hung\n", s->s_mds);
2591                         }
2592                 }
2593                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2594                         /* this mds is failed or recovering, just wait */
2595                         ceph_put_mds_session(s);
2596                         continue;
2597                 }
2598                 mutex_unlock(&mdsc->mutex);
2599
2600                 mutex_lock(&s->s_mutex);
2601                 if (renew_caps)
2602                         send_renew_caps(mdsc, s);
2603                 else
2604                         ceph_con_keepalive(&s->s_con);
2605                 add_cap_releases(mdsc, s, -1);
2606                 send_cap_releases(mdsc, s);
2607                 mutex_unlock(&s->s_mutex);
2608                 ceph_put_mds_session(s);
2609
2610                 mutex_lock(&mdsc->mutex);
2611         }
2612         mutex_unlock(&mdsc->mutex);
2613
2614         schedule_delayed(mdsc);
2615 }
2616
2617
2618 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2619 {
2620         mdsc->client = client;
2621         mutex_init(&mdsc->mutex);
2622         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2623         init_completion(&mdsc->safe_umount_waiters);
2624         init_completion(&mdsc->session_close_waiters);
2625         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2626         mdsc->sessions = NULL;
2627         mdsc->max_sessions = 0;
2628         mdsc->stopping = 0;
2629         init_rwsem(&mdsc->snap_rwsem);
2630         mdsc->snap_realms = RB_ROOT;
2631         INIT_LIST_HEAD(&mdsc->snap_empty);
2632         spin_lock_init(&mdsc->snap_empty_lock);
2633         mdsc->last_tid = 0;
2634         mdsc->request_tree = RB_ROOT;
2635         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2636         mdsc->last_renew_caps = jiffies;
2637         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2638         spin_lock_init(&mdsc->cap_delay_lock);
2639         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2640         spin_lock_init(&mdsc->snap_flush_lock);
2641         mdsc->cap_flush_seq = 0;
2642         INIT_LIST_HEAD(&mdsc->cap_dirty);
2643         mdsc->num_cap_flushing = 0;
2644         spin_lock_init(&mdsc->cap_dirty_lock);
2645         init_waitqueue_head(&mdsc->cap_flushing_wq);
2646         spin_lock_init(&mdsc->dentry_lru_lock);
2647         INIT_LIST_HEAD(&mdsc->dentry_lru);
2648         return 0;
2649 }
2650
2651 /*
2652  * Wait for safe replies on open mds requests.  If we time out, drop
2653  * all requests from the tree to avoid dangling dentry refs.
2654  */
2655 static void wait_requests(struct ceph_mds_client *mdsc)
2656 {
2657         struct ceph_mds_request *req;
2658         struct ceph_client *client = mdsc->client;
2659
2660         mutex_lock(&mdsc->mutex);
2661         if (__get_oldest_req(mdsc)) {
2662                 mutex_unlock(&mdsc->mutex);
2663
2664                 dout("wait_requests waiting for requests\n");
2665                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2666                                     client->mount_args->mount_timeout * HZ);
2667
2668                 /* tear down remaining requests */
2669                 mutex_lock(&mdsc->mutex);
2670                 while ((req = __get_oldest_req(mdsc))) {
2671                         dout("wait_requests timed out on tid %llu\n",
2672                              req->r_tid);
2673                         __unregister_request(mdsc, req);
2674                 }
2675         }
2676         mutex_unlock(&mdsc->mutex);
2677         dout("wait_requests done\n");
2678 }
2679
2680 /*
2681  * called before mount is ro, and before dentries are torn down.
2682  * (hmm, does this still race with new lookups?)
2683  */
2684 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2685 {
2686         dout("pre_umount\n");
2687         mdsc->stopping = 1;
2688
2689         drop_leases(mdsc);
2690         ceph_flush_dirty_caps(mdsc);
2691         wait_requests(mdsc);
2692 }
2693
2694 /*
2695  * wait for all write mds requests to flush.
2696  */
2697 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2698 {
2699         struct ceph_mds_request *req = NULL, *nextreq;
2700         struct rb_node *n;
2701
2702         mutex_lock(&mdsc->mutex);
2703         dout("wait_unsafe_requests want %lld\n", want_tid);
2704 restart:
2705         req = __get_oldest_req(mdsc);
2706         while (req && req->r_tid <= want_tid) {
2707                 /* find next request */
2708                 n = rb_next(&req->r_node);
2709                 if (n)
2710                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2711                 else
2712                         nextreq = NULL;
2713                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2714                         /* write op */
2715                         ceph_mdsc_get_request(req);
2716                         if (nextreq)
2717                                 ceph_mdsc_get_request(nextreq);
2718                         mutex_unlock(&mdsc->mutex);
2719                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2720                              req->r_tid, want_tid);
2721                         wait_for_completion(&req->r_safe_completion);
2722                         mutex_lock(&mdsc->mutex);
2723                         ceph_mdsc_put_request(req);
2724                         if (!nextreq)
2725                                 break;  /* next dne before, so we're done! */
2726                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2727                                 /* next request was removed from tree */
2728                                 ceph_mdsc_put_request(nextreq);
2729                                 goto restart;
2730                         }
2731                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2732                 }
2733                 req = nextreq;
2734         }
2735         mutex_unlock(&mdsc->mutex);
2736         dout("wait_unsafe_requests done\n");
2737 }
2738
2739 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2740 {
2741         u64 want_tid, want_flush;
2742
2743         dout("sync\n");
2744         mutex_lock(&mdsc->mutex);
2745         want_tid = mdsc->last_tid;
2746         want_flush = mdsc->cap_flush_seq;
2747         mutex_unlock(&mdsc->mutex);
2748         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2749
2750         ceph_flush_dirty_caps(mdsc);
2751
2752         wait_unsafe_requests(mdsc, want_tid);
2753         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2754 }
2755
2756
2757 /*
2758  * called after sb is ro.
2759  */
2760 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2761 {
2762         struct ceph_mds_session *session;
2763         int i;
2764         int n;
2765         struct ceph_client *client = mdsc->client;
2766         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2767
2768         dout("close_sessions\n");
2769
2770         mutex_lock(&mdsc->mutex);
2771
2772         /* close sessions */
2773         started = jiffies;
2774         while (time_before(jiffies, started + timeout)) {
2775                 dout("closing sessions\n");
2776                 n = 0;
2777                 for (i = 0; i < mdsc->max_sessions; i++) {
2778                         session = __ceph_lookup_mds_session(mdsc, i);
2779                         if (!session)
2780                                 continue;
2781                         mutex_unlock(&mdsc->mutex);
2782                         mutex_lock(&session->s_mutex);
2783                         __close_session(mdsc, session);
2784                         mutex_unlock(&session->s_mutex);
2785                         ceph_put_mds_session(session);
2786                         mutex_lock(&mdsc->mutex);
2787                         n++;
2788                 }
2789                 if (n == 0)
2790                         break;
2791
2792                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2793                         break;
2794
2795                 dout("waiting for sessions to close\n");
2796                 mutex_unlock(&mdsc->mutex);
2797                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2798                                             timeout);
2799                 mutex_lock(&mdsc->mutex);
2800         }
2801
2802         /* tear down remaining sessions */
2803         for (i = 0; i < mdsc->max_sessions; i++) {
2804                 if (mdsc->sessions[i]) {
2805                         session = get_session(mdsc->sessions[i]);
2806                         __unregister_session(mdsc, session);
2807                         mutex_unlock(&mdsc->mutex);
2808                         mutex_lock(&session->s_mutex);
2809                         remove_session_caps(session);
2810                         mutex_unlock(&session->s_mutex);
2811                         ceph_put_mds_session(session);
2812                         mutex_lock(&mdsc->mutex);
2813                 }
2814         }
2815
2816         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2817
2818         mutex_unlock(&mdsc->mutex);
2819
2820         ceph_cleanup_empty_realms(mdsc);
2821
2822         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2823
2824         dout("stopped\n");
2825 }
2826
2827 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2828 {
2829         dout("stop\n");
2830         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2831         if (mdsc->mdsmap)
2832                 ceph_mdsmap_destroy(mdsc->mdsmap);
2833         kfree(mdsc->sessions);
2834 }
2835
2836
2837 /*
2838  * handle mds map update.
2839  */
2840 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2841 {
2842         u32 epoch;
2843         u32 maplen;
2844         void *p = msg->front.iov_base;
2845         void *end = p + msg->front.iov_len;
2846         struct ceph_mdsmap *newmap, *oldmap;
2847         struct ceph_fsid fsid;
2848         int err = -EINVAL;
2849
2850         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2851         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2852         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2853                 return;
2854         epoch = ceph_decode_32(&p);
2855         maplen = ceph_decode_32(&p);
2856         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2857
2858         /* do we need it? */
2859         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2860         mutex_lock(&mdsc->mutex);
2861         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2862                 dout("handle_map epoch %u <= our %u\n",
2863                      epoch, mdsc->mdsmap->m_epoch);
2864                 mutex_unlock(&mdsc->mutex);
2865                 return;
2866         }
2867
2868         newmap = ceph_mdsmap_decode(&p, end);
2869         if (IS_ERR(newmap)) {
2870                 err = PTR_ERR(newmap);
2871                 goto bad_unlock;
2872         }
2873
2874         /* swap into place */
2875         if (mdsc->mdsmap) {
2876                 oldmap = mdsc->mdsmap;
2877                 mdsc->mdsmap = newmap;
2878                 check_new_map(mdsc, newmap, oldmap);
2879                 ceph_mdsmap_destroy(oldmap);
2880         } else {
2881                 mdsc->mdsmap = newmap;  /* first mds map */
2882         }
2883         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2884
2885         __wake_requests(mdsc, &mdsc->waiting_for_map);
2886
2887         mutex_unlock(&mdsc->mutex);
2888         schedule_delayed(mdsc);
2889         return;
2890
2891 bad_unlock:
2892         mutex_unlock(&mdsc->mutex);
2893 bad:
2894         pr_err("error decoding mdsmap %d\n", err);
2895         return;
2896 }
2897
2898 static struct ceph_connection *con_get(struct ceph_connection *con)
2899 {
2900         struct ceph_mds_session *s = con->private;
2901
2902         if (get_session(s)) {
2903                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2904                 return con;
2905         }
2906         dout("mdsc con_get %p FAIL\n", s);
2907         return NULL;
2908 }
2909
2910 static void con_put(struct ceph_connection *con)
2911 {
2912         struct ceph_mds_session *s = con->private;
2913
2914         ceph_put_mds_session(s);
2915         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2916 }
2917
2918 /*
2919  * if the client is unresponsive for long enough, the mds will kill
2920  * the session entirely.
2921  */
2922 static void peer_reset(struct ceph_connection *con)
2923 {
2924         struct ceph_mds_session *s = con->private;
2925
2926         pr_err("mds%d gave us the boot.  IMPLEMENT RECONNECT.\n",
2927                s->s_mds);
2928 }
2929
2930 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2931 {
2932         struct ceph_mds_session *s = con->private;
2933         struct ceph_mds_client *mdsc = s->s_mdsc;
2934         int type = le16_to_cpu(msg->hdr.type);
2935
2936         mutex_lock(&mdsc->mutex);
2937         if (__verify_registered_session(mdsc, s) < 0) {
2938                 mutex_unlock(&mdsc->mutex);
2939                 goto out;
2940         }
2941         mutex_unlock(&mdsc->mutex);
2942
2943         switch (type) {
2944         case CEPH_MSG_MDS_MAP:
2945                 ceph_mdsc_handle_map(mdsc, msg);
2946                 break;
2947         case CEPH_MSG_CLIENT_SESSION:
2948                 handle_session(s, msg);
2949                 break;
2950         case CEPH_MSG_CLIENT_REPLY:
2951                 handle_reply(s, msg);
2952                 break;
2953         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2954                 handle_forward(mdsc, s, msg);
2955                 break;
2956         case CEPH_MSG_CLIENT_CAPS:
2957                 ceph_handle_caps(s, msg);
2958                 break;
2959         case CEPH_MSG_CLIENT_SNAP:
2960                 ceph_handle_snap(mdsc, s, msg);
2961                 break;
2962         case CEPH_MSG_CLIENT_LEASE:
2963                 handle_lease(mdsc, s, msg);
2964                 break;
2965
2966         default:
2967                 pr_err("received unknown message type %d %s\n", type,
2968                        ceph_msg_type_name(type));
2969         }
2970 out:
2971         ceph_msg_put(msg);
2972 }
2973
2974 /*
2975  * authentication
2976  */
2977 static int get_authorizer(struct ceph_connection *con,
2978                           void **buf, int *len, int *proto,
2979                           void **reply_buf, int *reply_len, int force_new)
2980 {
2981         struct ceph_mds_session *s = con->private;
2982         struct ceph_mds_client *mdsc = s->s_mdsc;
2983         struct ceph_auth_client *ac = mdsc->client->monc.auth;
2984         int ret = 0;
2985
2986         if (force_new && s->s_authorizer) {
2987                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2988                 s->s_authorizer = NULL;
2989         }
2990         if (s->s_authorizer == NULL) {
2991                 if (ac->ops->create_authorizer) {
2992                         ret = ac->ops->create_authorizer(
2993                                 ac, CEPH_ENTITY_TYPE_MDS,
2994                                 &s->s_authorizer,
2995                                 &s->s_authorizer_buf,
2996                                 &s->s_authorizer_buf_len,
2997                                 &s->s_authorizer_reply_buf,
2998                                 &s->s_authorizer_reply_buf_len);
2999                         if (ret)
3000                                 return ret;
3001                 }
3002         }
3003
3004         *proto = ac->protocol;
3005         *buf = s->s_authorizer_buf;
3006         *len = s->s_authorizer_buf_len;
3007         *reply_buf = s->s_authorizer_reply_buf;
3008         *reply_len = s->s_authorizer_reply_buf_len;
3009         return 0;
3010 }
3011
3012
3013 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3014 {
3015         struct ceph_mds_session *s = con->private;
3016         struct ceph_mds_client *mdsc = s->s_mdsc;
3017         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3018
3019         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3020 }
3021
3022 static int invalidate_authorizer(struct ceph_connection *con)
3023 {
3024         struct ceph_mds_session *s = con->private;
3025         struct ceph_mds_client *mdsc = s->s_mdsc;
3026         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3027
3028         if (ac->ops->invalidate_authorizer)
3029                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3030
3031         return ceph_monc_validate_auth(&mdsc->client->monc);
3032 }
3033
3034 const static struct ceph_connection_operations mds_con_ops = {
3035         .get = con_get,
3036         .put = con_put,
3037         .dispatch = dispatch,
3038         .get_authorizer = get_authorizer,
3039         .verify_authorizer_reply = verify_authorizer_reply,
3040         .invalidate_authorizer = invalidate_authorizer,
3041         .peer_reset = peer_reset,
3042 };
3043
3044
3045
3046
3047 /* eof */