ceph: do caps accounting per mds_client
[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37  * are no longer valid.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 static const 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_mdsc, &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(mdsc, &req->r_caps_reservation,
516                                   req->r_num_caps);
517         dout("__register_request %p tid %lld\n", req, req->r_tid);
518         ceph_mdsc_get_request(req);
519         __insert_request(mdsc, req);
520
521         if (dir) {
522                 struct ceph_inode_info *ci = ceph_inode(dir);
523
524                 spin_lock(&ci->i_unsafe_lock);
525                 req->r_unsafe_dir = dir;
526                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
527                 spin_unlock(&ci->i_unsafe_lock);
528         }
529 }
530
531 static void __unregister_request(struct ceph_mds_client *mdsc,
532                                  struct ceph_mds_request *req)
533 {
534         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
535         rb_erase(&req->r_node, &mdsc->request_tree);
536         RB_CLEAR_NODE(&req->r_node);
537
538         if (req->r_unsafe_dir) {
539                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
540
541                 spin_lock(&ci->i_unsafe_lock);
542                 list_del_init(&req->r_unsafe_dir_item);
543                 spin_unlock(&ci->i_unsafe_lock);
544         }
545
546         ceph_mdsc_put_request(req);
547 }
548
549 /*
550  * Choose mds to send request to next.  If there is a hint set in the
551  * request (e.g., due to a prior forward hint from the mds), use that.
552  * Otherwise, consult frag tree and/or caps to identify the
553  * appropriate mds.  If all else fails, choose randomly.
554  *
555  * Called under mdsc->mutex.
556  */
557 static int __choose_mds(struct ceph_mds_client *mdsc,
558                         struct ceph_mds_request *req)
559 {
560         struct inode *inode;
561         struct ceph_inode_info *ci;
562         struct ceph_cap *cap;
563         int mode = req->r_direct_mode;
564         int mds = -1;
565         u32 hash = req->r_direct_hash;
566         bool is_hash = req->r_direct_is_hash;
567
568         /*
569          * is there a specific mds we should try?  ignore hint if we have
570          * no session and the mds is not up (active or recovering).
571          */
572         if (req->r_resend_mds >= 0 &&
573             (__have_session(mdsc, req->r_resend_mds) ||
574              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
575                 dout("choose_mds using resend_mds mds%d\n",
576                      req->r_resend_mds);
577                 return req->r_resend_mds;
578         }
579
580         if (mode == USE_RANDOM_MDS)
581                 goto random;
582
583         inode = NULL;
584         if (req->r_inode) {
585                 inode = req->r_inode;
586         } else if (req->r_dentry) {
587                 if (req->r_dentry->d_inode) {
588                         inode = req->r_dentry->d_inode;
589                 } else {
590                         inode = req->r_dentry->d_parent->d_inode;
591                         hash = req->r_dentry->d_name.hash;
592                         is_hash = true;
593                 }
594         }
595         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
596              (int)hash, mode);
597         if (!inode)
598                 goto random;
599         ci = ceph_inode(inode);
600
601         if (is_hash && S_ISDIR(inode->i_mode)) {
602                 struct ceph_inode_frag frag;
603                 int found;
604
605                 ceph_choose_frag(ci, hash, &frag, &found);
606                 if (found) {
607                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
608                                 u8 r;
609
610                                 /* choose a random replica */
611                                 get_random_bytes(&r, 1);
612                                 r %= frag.ndist;
613                                 mds = frag.dist[r];
614                                 dout("choose_mds %p %llx.%llx "
615                                      "frag %u mds%d (%d/%d)\n",
616                                      inode, ceph_vinop(inode),
617                                      frag.frag, frag.mds,
618                                      (int)r, frag.ndist);
619                                 return mds;
620                         }
621
622                         /* since this file/dir wasn't known to be
623                          * replicated, then we want to look for the
624                          * authoritative mds. */
625                         mode = USE_AUTH_MDS;
626                         if (frag.mds >= 0) {
627                                 /* choose auth mds */
628                                 mds = frag.mds;
629                                 dout("choose_mds %p %llx.%llx "
630                                      "frag %u mds%d (auth)\n",
631                                      inode, ceph_vinop(inode), frag.frag, mds);
632                                 return mds;
633                         }
634                 }
635         }
636
637         spin_lock(&inode->i_lock);
638         cap = NULL;
639         if (mode == USE_AUTH_MDS)
640                 cap = ci->i_auth_cap;
641         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
642                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
643         if (!cap) {
644                 spin_unlock(&inode->i_lock);
645                 goto random;
646         }
647         mds = cap->session->s_mds;
648         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
649              inode, ceph_vinop(inode), mds,
650              cap == ci->i_auth_cap ? "auth " : "", cap);
651         spin_unlock(&inode->i_lock);
652         return mds;
653
654 random:
655         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
656         dout("choose_mds chose random mds%d\n", mds);
657         return mds;
658 }
659
660
661 /*
662  * session messages
663  */
664 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
665 {
666         struct ceph_msg *msg;
667         struct ceph_mds_session_head *h;
668
669         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS);
670         if (!msg) {
671                 pr_err("create_session_msg ENOMEM creating msg\n");
672                 return NULL;
673         }
674         h = msg->front.iov_base;
675         h->op = cpu_to_le32(op);
676         h->seq = cpu_to_le64(seq);
677         return msg;
678 }
679
680 /*
681  * send session open request.
682  *
683  * called under mdsc->mutex
684  */
685 static int __open_session(struct ceph_mds_client *mdsc,
686                           struct ceph_mds_session *session)
687 {
688         struct ceph_msg *msg;
689         int mstate;
690         int mds = session->s_mds;
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 (!msg)
702                 return -ENOMEM;
703         ceph_con_send(&session->s_con, msg);
704         return 0;
705 }
706
707 /*
708  * session caps
709  */
710
711 /*
712  * Free preallocated cap messages assigned to this session
713  */
714 static void cleanup_cap_releases(struct ceph_mds_session *session)
715 {
716         struct ceph_msg *msg;
717
718         spin_lock(&session->s_cap_lock);
719         while (!list_empty(&session->s_cap_releases)) {
720                 msg = list_first_entry(&session->s_cap_releases,
721                                        struct ceph_msg, list_head);
722                 list_del_init(&msg->list_head);
723                 ceph_msg_put(msg);
724         }
725         while (!list_empty(&session->s_cap_releases_done)) {
726                 msg = list_first_entry(&session->s_cap_releases_done,
727                                        struct ceph_msg, list_head);
728                 list_del_init(&msg->list_head);
729                 ceph_msg_put(msg);
730         }
731         spin_unlock(&session->s_cap_lock);
732 }
733
734 /*
735  * Helper to safely iterate over all caps associated with a session, with
736  * special care taken to handle a racing __ceph_remove_cap().
737  *
738  * Caller must hold session s_mutex.
739  */
740 static int iterate_session_caps(struct ceph_mds_session *session,
741                                  int (*cb)(struct inode *, struct ceph_cap *,
742                                             void *), void *arg)
743 {
744         struct list_head *p;
745         struct ceph_cap *cap;
746         struct inode *inode, *last_inode = NULL;
747         struct ceph_cap *old_cap = NULL;
748         int ret;
749
750         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
751         spin_lock(&session->s_cap_lock);
752         p = session->s_caps.next;
753         while (p != &session->s_caps) {
754                 cap = list_entry(p, struct ceph_cap, session_caps);
755                 inode = igrab(&cap->ci->vfs_inode);
756                 if (!inode) {
757                         p = p->next;
758                         continue;
759                 }
760                 session->s_cap_iterator = cap;
761                 spin_unlock(&session->s_cap_lock);
762
763                 if (last_inode) {
764                         iput(last_inode);
765                         last_inode = NULL;
766                 }
767                 if (old_cap) {
768                         ceph_put_cap(session->s_mdsc, old_cap);
769                         old_cap = NULL;
770                 }
771
772                 ret = cb(inode, cap, arg);
773                 last_inode = inode;
774
775                 spin_lock(&session->s_cap_lock);
776                 p = p->next;
777                 if (cap->ci == NULL) {
778                         dout("iterate_session_caps  finishing cap %p removal\n",
779                              cap);
780                         BUG_ON(cap->session != session);
781                         list_del_init(&cap->session_caps);
782                         session->s_nr_caps--;
783                         cap->session = NULL;
784                         old_cap = cap;  /* put_cap it w/o locks held */
785                 }
786                 if (ret < 0)
787                         goto out;
788         }
789         ret = 0;
790 out:
791         session->s_cap_iterator = NULL;
792         spin_unlock(&session->s_cap_lock);
793
794         if (last_inode)
795                 iput(last_inode);
796         if (old_cap)
797                 ceph_put_cap(session->s_mdsc, old_cap);
798
799         return ret;
800 }
801
802 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
803                                   void *arg)
804 {
805         struct ceph_inode_info *ci = ceph_inode(inode);
806         int drop = 0;
807
808         dout("removing cap %p, ci is %p, inode is %p\n",
809              cap, ci, &ci->vfs_inode);
810         spin_lock(&inode->i_lock);
811         __ceph_remove_cap(cap);
812         if (!__ceph_is_any_real_caps(ci)) {
813                 struct ceph_mds_client *mdsc =
814                         &ceph_sb_to_client(inode->i_sb)->mdsc;
815
816                 spin_lock(&mdsc->cap_dirty_lock);
817                 if (!list_empty(&ci->i_dirty_item)) {
818                         pr_info(" dropping dirty %s state for %p %lld\n",
819                                 ceph_cap_string(ci->i_dirty_caps),
820                                 inode, ceph_ino(inode));
821                         ci->i_dirty_caps = 0;
822                         list_del_init(&ci->i_dirty_item);
823                         drop = 1;
824                 }
825                 if (!list_empty(&ci->i_flushing_item)) {
826                         pr_info(" dropping dirty+flushing %s state for %p %lld\n",
827                                 ceph_cap_string(ci->i_flushing_caps),
828                                 inode, ceph_ino(inode));
829                         ci->i_flushing_caps = 0;
830                         list_del_init(&ci->i_flushing_item);
831                         mdsc->num_cap_flushing--;
832                         drop = 1;
833                 }
834                 if (drop && ci->i_wrbuffer_ref) {
835                         pr_info(" dropping dirty data for %p %lld\n",
836                                 inode, ceph_ino(inode));
837                         ci->i_wrbuffer_ref = 0;
838                         ci->i_wrbuffer_ref_head = 0;
839                         drop++;
840                 }
841                 spin_unlock(&mdsc->cap_dirty_lock);
842         }
843         spin_unlock(&inode->i_lock);
844         while (drop--)
845                 iput(inode);
846         return 0;
847 }
848
849 /*
850  * caller must hold session s_mutex
851  */
852 static void remove_session_caps(struct ceph_mds_session *session)
853 {
854         dout("remove_session_caps on %p\n", session);
855         iterate_session_caps(session, remove_session_caps_cb, NULL);
856         BUG_ON(session->s_nr_caps > 0);
857         BUG_ON(!list_empty(&session->s_cap_flushing));
858         cleanup_cap_releases(session);
859 }
860
861 /*
862  * wake up any threads waiting on this session's caps.  if the cap is
863  * old (didn't get renewed on the client reconnect), remove it now.
864  *
865  * caller must hold s_mutex.
866  */
867 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
868                               void *arg)
869 {
870         struct ceph_inode_info *ci = ceph_inode(inode);
871
872         wake_up_all(&ci->i_cap_wq);
873         if (arg) {
874                 spin_lock(&inode->i_lock);
875                 ci->i_wanted_max_size = 0;
876                 ci->i_requested_max_size = 0;
877                 spin_unlock(&inode->i_lock);
878         }
879         return 0;
880 }
881
882 static void wake_up_session_caps(struct ceph_mds_session *session,
883                                  int reconnect)
884 {
885         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
886         iterate_session_caps(session, wake_up_session_cb,
887                              (void *)(unsigned long)reconnect);
888 }
889
890 /*
891  * Send periodic message to MDS renewing all currently held caps.  The
892  * ack will reset the expiration for all caps from this session.
893  *
894  * caller holds s_mutex
895  */
896 static int send_renew_caps(struct ceph_mds_client *mdsc,
897                            struct ceph_mds_session *session)
898 {
899         struct ceph_msg *msg;
900         int state;
901
902         if (time_after_eq(jiffies, session->s_cap_ttl) &&
903             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
904                 pr_info("mds%d caps stale\n", session->s_mds);
905         session->s_renew_requested = jiffies;
906
907         /* do not try to renew caps until a recovering mds has reconnected
908          * with its clients. */
909         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
910         if (state < CEPH_MDS_STATE_RECONNECT) {
911                 dout("send_renew_caps ignoring mds%d (%s)\n",
912                      session->s_mds, ceph_mds_state_name(state));
913                 return 0;
914         }
915
916         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
917                 ceph_mds_state_name(state));
918         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
919                                  ++session->s_renew_seq);
920         if (!msg)
921                 return -ENOMEM;
922         ceph_con_send(&session->s_con, msg);
923         return 0;
924 }
925
926 /*
927  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
928  *
929  * Called under session->s_mutex
930  */
931 static void renewed_caps(struct ceph_mds_client *mdsc,
932                          struct ceph_mds_session *session, int is_renew)
933 {
934         int was_stale;
935         int wake = 0;
936
937         spin_lock(&session->s_cap_lock);
938         was_stale = is_renew && (session->s_cap_ttl == 0 ||
939                                  time_after_eq(jiffies, session->s_cap_ttl));
940
941         session->s_cap_ttl = session->s_renew_requested +
942                 mdsc->mdsmap->m_session_timeout*HZ;
943
944         if (was_stale) {
945                 if (time_before(jiffies, session->s_cap_ttl)) {
946                         pr_info("mds%d caps renewed\n", session->s_mds);
947                         wake = 1;
948                 } else {
949                         pr_info("mds%d caps still stale\n", session->s_mds);
950                 }
951         }
952         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
953              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
954              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
955         spin_unlock(&session->s_cap_lock);
956
957         if (wake)
958                 wake_up_session_caps(session, 0);
959 }
960
961 /*
962  * send a session close request
963  */
964 static int request_close_session(struct ceph_mds_client *mdsc,
965                                  struct ceph_mds_session *session)
966 {
967         struct ceph_msg *msg;
968
969         dout("request_close_session mds%d state %s seq %lld\n",
970              session->s_mds, session_state_name(session->s_state),
971              session->s_seq);
972         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
973         if (!msg)
974                 return -ENOMEM;
975         ceph_con_send(&session->s_con, msg);
976         return 0;
977 }
978
979 /*
980  * Called with s_mutex held.
981  */
982 static int __close_session(struct ceph_mds_client *mdsc,
983                          struct ceph_mds_session *session)
984 {
985         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
986                 return 0;
987         session->s_state = CEPH_MDS_SESSION_CLOSING;
988         return request_close_session(mdsc, session);
989 }
990
991 /*
992  * Trim old(er) caps.
993  *
994  * Because we can't cache an inode without one or more caps, we do
995  * this indirectly: if a cap is unused, we prune its aliases, at which
996  * point the inode will hopefully get dropped to.
997  *
998  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
999  * memory pressure from the MDS, though, so it needn't be perfect.
1000  */
1001 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1002 {
1003         struct ceph_mds_session *session = arg;
1004         struct ceph_inode_info *ci = ceph_inode(inode);
1005         int used, oissued, mine;
1006
1007         if (session->s_trim_caps <= 0)
1008                 return -1;
1009
1010         spin_lock(&inode->i_lock);
1011         mine = cap->issued | cap->implemented;
1012         used = __ceph_caps_used(ci);
1013         oissued = __ceph_caps_issued_other(ci, cap);
1014
1015         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1016              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1017              ceph_cap_string(used));
1018         if (ci->i_dirty_caps)
1019                 goto out;   /* dirty caps */
1020         if ((used & ~oissued) & mine)
1021                 goto out;   /* we need these caps */
1022
1023         session->s_trim_caps--;
1024         if (oissued) {
1025                 /* we aren't the only cap.. just remove us */
1026                 __ceph_remove_cap(cap);
1027         } else {
1028                 /* try to drop referring dentries */
1029                 spin_unlock(&inode->i_lock);
1030                 d_prune_aliases(inode);
1031                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1032                      inode, cap, atomic_read(&inode->i_count));
1033                 return 0;
1034         }
1035
1036 out:
1037         spin_unlock(&inode->i_lock);
1038         return 0;
1039 }
1040
1041 /*
1042  * Trim session cap count down to some max number.
1043  */
1044 static int trim_caps(struct ceph_mds_client *mdsc,
1045                      struct ceph_mds_session *session,
1046                      int max_caps)
1047 {
1048         int trim_caps = session->s_nr_caps - max_caps;
1049
1050         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1051              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1052         if (trim_caps > 0) {
1053                 session->s_trim_caps = trim_caps;
1054                 iterate_session_caps(session, trim_caps_cb, session);
1055                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1056                      session->s_mds, session->s_nr_caps, max_caps,
1057                         trim_caps - session->s_trim_caps);
1058                 session->s_trim_caps = 0;
1059         }
1060         return 0;
1061 }
1062
1063 /*
1064  * Allocate cap_release messages.  If there is a partially full message
1065  * in the queue, try to allocate enough to cover it's remainder, so that
1066  * we can send it immediately.
1067  *
1068  * Called under s_mutex.
1069  */
1070 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1071                           struct ceph_mds_session *session)
1072 {
1073         struct ceph_msg *msg, *partial = NULL;
1074         struct ceph_mds_cap_release *head;
1075         int err = -ENOMEM;
1076         int extra = mdsc->client->mount_args->cap_release_safety;
1077         int num;
1078
1079         dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1080              extra);
1081
1082         spin_lock(&session->s_cap_lock);
1083
1084         if (!list_empty(&session->s_cap_releases)) {
1085                 msg = list_first_entry(&session->s_cap_releases,
1086                                        struct ceph_msg,
1087                                  list_head);
1088                 head = msg->front.iov_base;
1089                 num = le32_to_cpu(head->num);
1090                 if (num) {
1091                         dout(" partial %p with (%d/%d)\n", msg, num,
1092                              (int)CEPH_CAPS_PER_RELEASE);
1093                         extra += CEPH_CAPS_PER_RELEASE - num;
1094                         partial = msg;
1095                 }
1096         }
1097         while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1098                 spin_unlock(&session->s_cap_lock);
1099                 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1100                                    GFP_NOFS);
1101                 if (!msg)
1102                         goto out_unlocked;
1103                 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1104                      (int)msg->front.iov_len);
1105                 head = msg->front.iov_base;
1106                 head->num = cpu_to_le32(0);
1107                 msg->front.iov_len = sizeof(*head);
1108                 spin_lock(&session->s_cap_lock);
1109                 list_add(&msg->list_head, &session->s_cap_releases);
1110                 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1111         }
1112
1113         if (partial) {
1114                 head = partial->front.iov_base;
1115                 num = le32_to_cpu(head->num);
1116                 dout(" queueing partial %p with %d/%d\n", partial, num,
1117                      (int)CEPH_CAPS_PER_RELEASE);
1118                 list_move_tail(&partial->list_head,
1119                                &session->s_cap_releases_done);
1120                 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1121         }
1122         err = 0;
1123         spin_unlock(&session->s_cap_lock);
1124 out_unlocked:
1125         return err;
1126 }
1127
1128 /*
1129  * flush all dirty inode data to disk.
1130  *
1131  * returns true if we've flushed through want_flush_seq
1132  */
1133 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1134 {
1135         int mds, ret = 1;
1136
1137         dout("check_cap_flush want %lld\n", want_flush_seq);
1138         mutex_lock(&mdsc->mutex);
1139         for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1140                 struct ceph_mds_session *session = mdsc->sessions[mds];
1141
1142                 if (!session)
1143                         continue;
1144                 get_session(session);
1145                 mutex_unlock(&mdsc->mutex);
1146
1147                 mutex_lock(&session->s_mutex);
1148                 if (!list_empty(&session->s_cap_flushing)) {
1149                         struct ceph_inode_info *ci =
1150                                 list_entry(session->s_cap_flushing.next,
1151                                            struct ceph_inode_info,
1152                                            i_flushing_item);
1153                         struct inode *inode = &ci->vfs_inode;
1154
1155                         spin_lock(&inode->i_lock);
1156                         if (ci->i_cap_flush_seq <= want_flush_seq) {
1157                                 dout("check_cap_flush still flushing %p "
1158                                      "seq %lld <= %lld to mds%d\n", inode,
1159                                      ci->i_cap_flush_seq, want_flush_seq,
1160                                      session->s_mds);
1161                                 ret = 0;
1162                         }
1163                         spin_unlock(&inode->i_lock);
1164                 }
1165                 mutex_unlock(&session->s_mutex);
1166                 ceph_put_mds_session(session);
1167
1168                 if (!ret)
1169                         return ret;
1170                 mutex_lock(&mdsc->mutex);
1171         }
1172
1173         mutex_unlock(&mdsc->mutex);
1174         dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1175         return ret;
1176 }
1177
1178 /*
1179  * called under s_mutex
1180  */
1181 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1182                             struct ceph_mds_session *session)
1183 {
1184         struct ceph_msg *msg;
1185
1186         dout("send_cap_releases mds%d\n", session->s_mds);
1187         spin_lock(&session->s_cap_lock);
1188         while (!list_empty(&session->s_cap_releases_done)) {
1189                 msg = list_first_entry(&session->s_cap_releases_done,
1190                                  struct ceph_msg, list_head);
1191                 list_del_init(&msg->list_head);
1192                 spin_unlock(&session->s_cap_lock);
1193                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1194                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1195                 ceph_con_send(&session->s_con, msg);
1196                 spin_lock(&session->s_cap_lock);
1197         }
1198         spin_unlock(&session->s_cap_lock);
1199 }
1200
1201 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1202                                  struct ceph_mds_session *session)
1203 {
1204         struct ceph_msg *msg;
1205         struct ceph_mds_cap_release *head;
1206         unsigned num;
1207
1208         dout("discard_cap_releases mds%d\n", session->s_mds);
1209         spin_lock(&session->s_cap_lock);
1210
1211         /* zero out the in-progress message */
1212         msg = list_first_entry(&session->s_cap_releases,
1213                                struct ceph_msg, list_head);
1214         head = msg->front.iov_base;
1215         num = le32_to_cpu(head->num);
1216         dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1217         head->num = cpu_to_le32(0);
1218         session->s_num_cap_releases += num;
1219
1220         /* requeue completed messages */
1221         while (!list_empty(&session->s_cap_releases_done)) {
1222                 msg = list_first_entry(&session->s_cap_releases_done,
1223                                  struct ceph_msg, list_head);
1224                 list_del_init(&msg->list_head);
1225
1226                 head = msg->front.iov_base;
1227                 num = le32_to_cpu(head->num);
1228                 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1229                      num);
1230                 session->s_num_cap_releases += num;
1231                 head->num = cpu_to_le32(0);
1232                 msg->front.iov_len = sizeof(*head);
1233                 list_add(&msg->list_head, &session->s_cap_releases);
1234         }
1235
1236         spin_unlock(&session->s_cap_lock);
1237 }
1238
1239 /*
1240  * requests
1241  */
1242
1243 /*
1244  * Create an mds request.
1245  */
1246 struct ceph_mds_request *
1247 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1248 {
1249         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1250
1251         if (!req)
1252                 return ERR_PTR(-ENOMEM);
1253
1254         mutex_init(&req->r_fill_mutex);
1255         req->r_mdsc = mdsc;
1256         req->r_started = jiffies;
1257         req->r_resend_mds = -1;
1258         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1259         req->r_fmode = -1;
1260         kref_init(&req->r_kref);
1261         INIT_LIST_HEAD(&req->r_wait);
1262         init_completion(&req->r_completion);
1263         init_completion(&req->r_safe_completion);
1264         INIT_LIST_HEAD(&req->r_unsafe_item);
1265
1266         req->r_op = op;
1267         req->r_direct_mode = mode;
1268         return req;
1269 }
1270
1271 /*
1272  * return oldest (lowest) request, tid in request tree, 0 if none.
1273  *
1274  * called under mdsc->mutex.
1275  */
1276 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1277 {
1278         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1279                 return NULL;
1280         return rb_entry(rb_first(&mdsc->request_tree),
1281                         struct ceph_mds_request, r_node);
1282 }
1283
1284 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1285 {
1286         struct ceph_mds_request *req = __get_oldest_req(mdsc);
1287
1288         if (req)
1289                 return req->r_tid;
1290         return 0;
1291 }
1292
1293 /*
1294  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1295  * on build_path_from_dentry in fs/cifs/dir.c.
1296  *
1297  * If @stop_on_nosnap, generate path relative to the first non-snapped
1298  * inode.
1299  *
1300  * Encode hidden .snap dirs as a double /, i.e.
1301  *   foo/.snap/bar -> foo//bar
1302  */
1303 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1304                            int stop_on_nosnap)
1305 {
1306         struct dentry *temp;
1307         char *path;
1308         int len, pos;
1309
1310         if (dentry == NULL)
1311                 return ERR_PTR(-EINVAL);
1312
1313 retry:
1314         len = 0;
1315         for (temp = dentry; !IS_ROOT(temp);) {
1316                 struct inode *inode = temp->d_inode;
1317                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1318                         len++;  /* slash only */
1319                 else if (stop_on_nosnap && inode &&
1320                          ceph_snap(inode) == CEPH_NOSNAP)
1321                         break;
1322                 else
1323                         len += 1 + temp->d_name.len;
1324                 temp = temp->d_parent;
1325                 if (temp == NULL) {
1326                         pr_err("build_path corrupt dentry %p\n", dentry);
1327                         return ERR_PTR(-EINVAL);
1328                 }
1329         }
1330         if (len)
1331                 len--;  /* no leading '/' */
1332
1333         path = kmalloc(len+1, GFP_NOFS);
1334         if (path == NULL)
1335                 return ERR_PTR(-ENOMEM);
1336         pos = len;
1337         path[pos] = 0;  /* trailing null */
1338         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1339                 struct inode *inode = temp->d_inode;
1340
1341                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1342                         dout("build_path path+%d: %p SNAPDIR\n",
1343                              pos, temp);
1344                 } else if (stop_on_nosnap && inode &&
1345                            ceph_snap(inode) == CEPH_NOSNAP) {
1346                         break;
1347                 } else {
1348                         pos -= temp->d_name.len;
1349                         if (pos < 0)
1350                                 break;
1351                         strncpy(path + pos, temp->d_name.name,
1352                                 temp->d_name.len);
1353                 }
1354                 if (pos)
1355                         path[--pos] = '/';
1356                 temp = temp->d_parent;
1357                 if (temp == NULL) {
1358                         pr_err("build_path corrupt dentry\n");
1359                         kfree(path);
1360                         return ERR_PTR(-EINVAL);
1361                 }
1362         }
1363         if (pos != 0) {
1364                 pr_err("build_path did not end path lookup where "
1365                        "expected, namelen is %d, pos is %d\n", len, pos);
1366                 /* presumably this is only possible if racing with a
1367                    rename of one of the parent directories (we can not
1368                    lock the dentries above us to prevent this, but
1369                    retrying should be harmless) */
1370                 kfree(path);
1371                 goto retry;
1372         }
1373
1374         *base = ceph_ino(temp->d_inode);
1375         *plen = len;
1376         dout("build_path on %p %d built %llx '%.*s'\n",
1377              dentry, atomic_read(&dentry->d_count), *base, len, path);
1378         return path;
1379 }
1380
1381 static int build_dentry_path(struct dentry *dentry,
1382                              const char **ppath, int *ppathlen, u64 *pino,
1383                              int *pfreepath)
1384 {
1385         char *path;
1386
1387         if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1388                 *pino = ceph_ino(dentry->d_parent->d_inode);
1389                 *ppath = dentry->d_name.name;
1390                 *ppathlen = dentry->d_name.len;
1391                 return 0;
1392         }
1393         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1394         if (IS_ERR(path))
1395                 return PTR_ERR(path);
1396         *ppath = path;
1397         *pfreepath = 1;
1398         return 0;
1399 }
1400
1401 static int build_inode_path(struct inode *inode,
1402                             const char **ppath, int *ppathlen, u64 *pino,
1403                             int *pfreepath)
1404 {
1405         struct dentry *dentry;
1406         char *path;
1407
1408         if (ceph_snap(inode) == CEPH_NOSNAP) {
1409                 *pino = ceph_ino(inode);
1410                 *ppathlen = 0;
1411                 return 0;
1412         }
1413         dentry = d_find_alias(inode);
1414         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1415         dput(dentry);
1416         if (IS_ERR(path))
1417                 return PTR_ERR(path);
1418         *ppath = path;
1419         *pfreepath = 1;
1420         return 0;
1421 }
1422
1423 /*
1424  * request arguments may be specified via an inode *, a dentry *, or
1425  * an explicit ino+path.
1426  */
1427 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1428                                   const char *rpath, u64 rino,
1429                                   const char **ppath, int *pathlen,
1430                                   u64 *ino, int *freepath)
1431 {
1432         int r = 0;
1433
1434         if (rinode) {
1435                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1436                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1437                      ceph_snap(rinode));
1438         } else if (rdentry) {
1439                 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1440                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1441                      *ppath);
1442         } else if (rpath) {
1443                 *ino = rino;
1444                 *ppath = rpath;
1445                 *pathlen = strlen(rpath);
1446                 dout(" path %.*s\n", *pathlen, rpath);
1447         }
1448
1449         return r;
1450 }
1451
1452 /*
1453  * called under mdsc->mutex
1454  */
1455 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1456                                                struct ceph_mds_request *req,
1457                                                int mds)
1458 {
1459         struct ceph_msg *msg;
1460         struct ceph_mds_request_head *head;
1461         const char *path1 = NULL;
1462         const char *path2 = NULL;
1463         u64 ino1 = 0, ino2 = 0;
1464         int pathlen1 = 0, pathlen2 = 0;
1465         int freepath1 = 0, freepath2 = 0;
1466         int len;
1467         u16 releases;
1468         void *p, *end;
1469         int ret;
1470
1471         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1472                               req->r_path1, req->r_ino1.ino,
1473                               &path1, &pathlen1, &ino1, &freepath1);
1474         if (ret < 0) {
1475                 msg = ERR_PTR(ret);
1476                 goto out;
1477         }
1478
1479         ret = set_request_path_attr(NULL, req->r_old_dentry,
1480                               req->r_path2, req->r_ino2.ino,
1481                               &path2, &pathlen2, &ino2, &freepath2);
1482         if (ret < 0) {
1483                 msg = ERR_PTR(ret);
1484                 goto out_free1;
1485         }
1486
1487         len = sizeof(*head) +
1488                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1489
1490         /* calculate (max) length for cap releases */
1491         len += sizeof(struct ceph_mds_request_release) *
1492                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1493                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1494         if (req->r_dentry_drop)
1495                 len += req->r_dentry->d_name.len;
1496         if (req->r_old_dentry_drop)
1497                 len += req->r_old_dentry->d_name.len;
1498
1499         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS);
1500         if (!msg) {
1501                 msg = ERR_PTR(-ENOMEM);
1502                 goto out_free2;
1503         }
1504
1505         msg->hdr.tid = cpu_to_le64(req->r_tid);
1506
1507         head = msg->front.iov_base;
1508         p = msg->front.iov_base + sizeof(*head);
1509         end = msg->front.iov_base + msg->front.iov_len;
1510
1511         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1512         head->op = cpu_to_le32(req->r_op);
1513         head->caller_uid = cpu_to_le32(current_fsuid());
1514         head->caller_gid = cpu_to_le32(current_fsgid());
1515         head->args = req->r_args;
1516
1517         ceph_encode_filepath(&p, end, ino1, path1);
1518         ceph_encode_filepath(&p, end, ino2, path2);
1519
1520         /* make note of release offset, in case we need to replay */
1521         req->r_request_release_offset = p - msg->front.iov_base;
1522
1523         /* cap releases */
1524         releases = 0;
1525         if (req->r_inode_drop)
1526                 releases += ceph_encode_inode_release(&p,
1527                       req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1528                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1529         if (req->r_dentry_drop)
1530                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1531                        mds, req->r_dentry_drop, req->r_dentry_unless);
1532         if (req->r_old_dentry_drop)
1533                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1534                        mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1535         if (req->r_old_inode_drop)
1536                 releases += ceph_encode_inode_release(&p,
1537                       req->r_old_dentry->d_inode,
1538                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1539         head->num_releases = cpu_to_le16(releases);
1540
1541         BUG_ON(p > end);
1542         msg->front.iov_len = p - msg->front.iov_base;
1543         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1544
1545         msg->pages = req->r_pages;
1546         msg->nr_pages = req->r_num_pages;
1547         msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1548         msg->hdr.data_off = cpu_to_le16(0);
1549
1550 out_free2:
1551         if (freepath2)
1552                 kfree((char *)path2);
1553 out_free1:
1554         if (freepath1)
1555                 kfree((char *)path1);
1556 out:
1557         return msg;
1558 }
1559
1560 /*
1561  * called under mdsc->mutex if error, under no mutex if
1562  * success.
1563  */
1564 static void complete_request(struct ceph_mds_client *mdsc,
1565                              struct ceph_mds_request *req)
1566 {
1567         if (req->r_callback)
1568                 req->r_callback(mdsc, req);
1569         else
1570                 complete_all(&req->r_completion);
1571 }
1572
1573 /*
1574  * called under mdsc->mutex
1575  */
1576 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1577                                   struct ceph_mds_request *req,
1578                                   int mds)
1579 {
1580         struct ceph_mds_request_head *rhead;
1581         struct ceph_msg *msg;
1582         int flags = 0;
1583
1584         req->r_mds = mds;
1585         req->r_attempts++;
1586         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1587              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1588
1589         if (req->r_got_unsafe) {
1590                 /*
1591                  * Replay.  Do not regenerate message (and rebuild
1592                  * paths, etc.); just use the original message.
1593                  * Rebuilding paths will break for renames because
1594                  * d_move mangles the src name.
1595                  */
1596                 msg = req->r_request;
1597                 rhead = msg->front.iov_base;
1598
1599                 flags = le32_to_cpu(rhead->flags);
1600                 flags |= CEPH_MDS_FLAG_REPLAY;
1601                 rhead->flags = cpu_to_le32(flags);
1602
1603                 if (req->r_target_inode)
1604                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1605
1606                 rhead->num_retry = req->r_attempts - 1;
1607
1608                 /* remove cap/dentry releases from message */
1609                 rhead->num_releases = 0;
1610                 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1611                 msg->front.iov_len = req->r_request_release_offset;
1612                 return 0;
1613         }
1614
1615         if (req->r_request) {
1616                 ceph_msg_put(req->r_request);
1617                 req->r_request = NULL;
1618         }
1619         msg = create_request_message(mdsc, req, mds);
1620         if (IS_ERR(msg)) {
1621                 req->r_err = PTR_ERR(msg);
1622                 complete_request(mdsc, req);
1623                 return PTR_ERR(msg);
1624         }
1625         req->r_request = msg;
1626
1627         rhead = msg->front.iov_base;
1628         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1629         if (req->r_got_unsafe)
1630                 flags |= CEPH_MDS_FLAG_REPLAY;
1631         if (req->r_locked_dir)
1632                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1633         rhead->flags = cpu_to_le32(flags);
1634         rhead->num_fwd = req->r_num_fwd;
1635         rhead->num_retry = req->r_attempts - 1;
1636         rhead->ino = 0;
1637
1638         dout(" r_locked_dir = %p\n", req->r_locked_dir);
1639         return 0;
1640 }
1641
1642 /*
1643  * send request, or put it on the appropriate wait list.
1644  */
1645 static int __do_request(struct ceph_mds_client *mdsc,
1646                         struct ceph_mds_request *req)
1647 {
1648         struct ceph_mds_session *session = NULL;
1649         int mds = -1;
1650         int err = -EAGAIN;
1651
1652         if (req->r_err || req->r_got_result)
1653                 goto out;
1654
1655         if (req->r_timeout &&
1656             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1657                 dout("do_request timed out\n");
1658                 err = -EIO;
1659                 goto finish;
1660         }
1661
1662         mds = __choose_mds(mdsc, req);
1663         if (mds < 0 ||
1664             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1665                 dout("do_request no mds or not active, waiting for map\n");
1666                 list_add(&req->r_wait, &mdsc->waiting_for_map);
1667                 goto out;
1668         }
1669
1670         /* get, open session */
1671         session = __ceph_lookup_mds_session(mdsc, mds);
1672         if (!session) {
1673                 session = register_session(mdsc, mds);
1674                 if (IS_ERR(session)) {
1675                         err = PTR_ERR(session);
1676                         goto finish;
1677                 }
1678         }
1679         dout("do_request mds%d session %p state %s\n", mds, session,
1680              session_state_name(session->s_state));
1681         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1682             session->s_state != CEPH_MDS_SESSION_HUNG) {
1683                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1684                     session->s_state == CEPH_MDS_SESSION_CLOSING)
1685                         __open_session(mdsc, session);
1686                 list_add(&req->r_wait, &session->s_waiting);
1687                 goto out_session;
1688         }
1689
1690         /* send request */
1691         req->r_session = get_session(session);
1692         req->r_resend_mds = -1;   /* forget any previous mds hint */
1693
1694         if (req->r_request_started == 0)   /* note request start time */
1695                 req->r_request_started = jiffies;
1696
1697         err = __prepare_send_request(mdsc, req, mds);
1698         if (!err) {
1699                 ceph_msg_get(req->r_request);
1700                 ceph_con_send(&session->s_con, req->r_request);
1701         }
1702
1703 out_session:
1704         ceph_put_mds_session(session);
1705 out:
1706         return err;
1707
1708 finish:
1709         req->r_err = err;
1710         complete_request(mdsc, req);
1711         goto out;
1712 }
1713
1714 /*
1715  * called under mdsc->mutex
1716  */
1717 static void __wake_requests(struct ceph_mds_client *mdsc,
1718                             struct list_head *head)
1719 {
1720         struct ceph_mds_request *req, *nreq;
1721
1722         list_for_each_entry_safe(req, nreq, head, r_wait) {
1723                 list_del_init(&req->r_wait);
1724                 __do_request(mdsc, req);
1725         }
1726 }
1727
1728 /*
1729  * Wake up threads with requests pending for @mds, so that they can
1730  * resubmit their requests to a possibly different mds.
1731  */
1732 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1733 {
1734         struct ceph_mds_request *req;
1735         struct rb_node *p;
1736
1737         dout("kick_requests mds%d\n", mds);
1738         for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1739                 req = rb_entry(p, struct ceph_mds_request, r_node);
1740                 if (req->r_got_unsafe)
1741                         continue;
1742                 if (req->r_session &&
1743                     req->r_session->s_mds == mds) {
1744                         dout(" kicking tid %llu\n", req->r_tid);
1745                         put_request_session(req);
1746                         __do_request(mdsc, req);
1747                 }
1748         }
1749 }
1750
1751 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1752                               struct ceph_mds_request *req)
1753 {
1754         dout("submit_request on %p\n", req);
1755         mutex_lock(&mdsc->mutex);
1756         __register_request(mdsc, req, NULL);
1757         __do_request(mdsc, req);
1758         mutex_unlock(&mdsc->mutex);
1759 }
1760
1761 /*
1762  * Synchrously perform an mds request.  Take care of all of the
1763  * session setup, forwarding, retry details.
1764  */
1765 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1766                          struct inode *dir,
1767                          struct ceph_mds_request *req)
1768 {
1769         int err;
1770
1771         dout("do_request on %p\n", req);
1772
1773         /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1774         if (req->r_inode)
1775                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1776         if (req->r_locked_dir)
1777                 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1778         if (req->r_old_dentry)
1779                 ceph_get_cap_refs(
1780                         ceph_inode(req->r_old_dentry->d_parent->d_inode),
1781                         CEPH_CAP_PIN);
1782
1783         /* issue */
1784         mutex_lock(&mdsc->mutex);
1785         __register_request(mdsc, req, dir);
1786         __do_request(mdsc, req);
1787
1788         if (req->r_err) {
1789                 err = req->r_err;
1790                 __unregister_request(mdsc, req);
1791                 dout("do_request early error %d\n", err);
1792                 goto out;
1793         }
1794
1795         /* wait */
1796         mutex_unlock(&mdsc->mutex);
1797         dout("do_request waiting\n");
1798         if (req->r_timeout) {
1799                 err = (long)wait_for_completion_killable_timeout(
1800                         &req->r_completion, req->r_timeout);
1801                 if (err == 0)
1802                         err = -EIO;
1803         } else {
1804                 err = wait_for_completion_killable(&req->r_completion);
1805         }
1806         dout("do_request waited, got %d\n", err);
1807         mutex_lock(&mdsc->mutex);
1808
1809         /* only abort if we didn't race with a real reply */
1810         if (req->r_got_result) {
1811                 err = le32_to_cpu(req->r_reply_info.head->result);
1812         } else if (err < 0) {
1813                 dout("aborted request %lld with %d\n", req->r_tid, err);
1814
1815                 /*
1816                  * ensure we aren't running concurrently with
1817                  * ceph_fill_trace or ceph_readdir_prepopulate, which
1818                  * rely on locks (dir mutex) held by our caller.
1819                  */
1820                 mutex_lock(&req->r_fill_mutex);
1821                 req->r_err = err;
1822                 req->r_aborted = true;
1823                 mutex_unlock(&req->r_fill_mutex);
1824
1825                 if (req->r_locked_dir &&
1826                     (req->r_op & CEPH_MDS_OP_WRITE))
1827                         ceph_invalidate_dir_request(req);
1828         } else {
1829                 err = req->r_err;
1830         }
1831
1832 out:
1833         mutex_unlock(&mdsc->mutex);
1834         dout("do_request %p done, result %d\n", req, err);
1835         return err;
1836 }
1837
1838 /*
1839  * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
1840  * namespace request.
1841  */
1842 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
1843 {
1844         struct inode *inode = req->r_locked_dir;
1845         struct ceph_inode_info *ci = ceph_inode(inode);
1846
1847         dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode);
1848         spin_lock(&inode->i_lock);
1849         ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1850         ci->i_release_count++;
1851         spin_unlock(&inode->i_lock);
1852
1853         if (req->r_dentry)
1854                 ceph_invalidate_dentry_lease(req->r_dentry);
1855         if (req->r_old_dentry)
1856                 ceph_invalidate_dentry_lease(req->r_old_dentry);
1857 }
1858
1859 /*
1860  * Handle mds reply.
1861  *
1862  * We take the session mutex and parse and process the reply immediately.
1863  * This preserves the logical ordering of replies, capabilities, etc., sent
1864  * by the MDS as they are applied to our local cache.
1865  */
1866 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1867 {
1868         struct ceph_mds_client *mdsc = session->s_mdsc;
1869         struct ceph_mds_request *req;
1870         struct ceph_mds_reply_head *head = msg->front.iov_base;
1871         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
1872         u64 tid;
1873         int err, result;
1874         int mds = session->s_mds;
1875
1876         if (msg->front.iov_len < sizeof(*head)) {
1877                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1878                 ceph_msg_dump(msg);
1879                 return;
1880         }
1881
1882         /* get request, session */
1883         tid = le64_to_cpu(msg->hdr.tid);
1884         mutex_lock(&mdsc->mutex);
1885         req = __lookup_request(mdsc, tid);
1886         if (!req) {
1887                 dout("handle_reply on unknown tid %llu\n", tid);
1888                 mutex_unlock(&mdsc->mutex);
1889                 return;
1890         }
1891         dout("handle_reply %p\n", req);
1892
1893         /* correct session? */
1894         if (req->r_session != session) {
1895                 pr_err("mdsc_handle_reply got %llu on session mds%d"
1896                        " not mds%d\n", tid, session->s_mds,
1897                        req->r_session ? req->r_session->s_mds : -1);
1898                 mutex_unlock(&mdsc->mutex);
1899                 goto out;
1900         }
1901
1902         /* dup? */
1903         if ((req->r_got_unsafe && !head->safe) ||
1904             (req->r_got_safe && head->safe)) {
1905                 pr_warning("got a dup %s reply on %llu from mds%d\n",
1906                            head->safe ? "safe" : "unsafe", tid, mds);
1907                 mutex_unlock(&mdsc->mutex);
1908                 goto out;
1909         }
1910         if (req->r_got_safe && !head->safe) {
1911                 pr_warning("got unsafe after safe on %llu from mds%d\n",
1912                            tid, mds);
1913                 mutex_unlock(&mdsc->mutex);
1914                 goto out;
1915         }
1916
1917         result = le32_to_cpu(head->result);
1918
1919         /*
1920          * Tolerate 2 consecutive ESTALEs from the same mds.
1921          * FIXME: we should be looking at the cap migrate_seq.
1922          */
1923         if (result == -ESTALE) {
1924                 req->r_direct_mode = USE_AUTH_MDS;
1925                 req->r_num_stale++;
1926                 if (req->r_num_stale <= 2) {
1927                         __do_request(mdsc, req);
1928                         mutex_unlock(&mdsc->mutex);
1929                         goto out;
1930                 }
1931         } else {
1932                 req->r_num_stale = 0;
1933         }
1934
1935         if (head->safe) {
1936                 req->r_got_safe = true;
1937                 __unregister_request(mdsc, req);
1938                 complete_all(&req->r_safe_completion);
1939
1940                 if (req->r_got_unsafe) {
1941                         /*
1942                          * We already handled the unsafe response, now do the
1943                          * cleanup.  No need to examine the response; the MDS
1944                          * doesn't include any result info in the safe
1945                          * response.  And even if it did, there is nothing
1946                          * useful we could do with a revised return value.
1947                          */
1948                         dout("got safe reply %llu, mds%d\n", tid, mds);
1949                         list_del_init(&req->r_unsafe_item);
1950
1951                         /* last unsafe request during umount? */
1952                         if (mdsc->stopping && !__get_oldest_req(mdsc))
1953                                 complete_all(&mdsc->safe_umount_waiters);
1954                         mutex_unlock(&mdsc->mutex);
1955                         goto out;
1956                 }
1957         } else {
1958                 req->r_got_unsafe = true;
1959                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1960         }
1961
1962         dout("handle_reply tid %lld result %d\n", tid, result);
1963         rinfo = &req->r_reply_info;
1964         err = parse_reply_info(msg, rinfo);
1965         mutex_unlock(&mdsc->mutex);
1966
1967         mutex_lock(&session->s_mutex);
1968         if (err < 0) {
1969                 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1970                 ceph_msg_dump(msg);
1971                 goto out_err;
1972         }
1973
1974         /* snap trace */
1975         if (rinfo->snapblob_len) {
1976                 down_write(&mdsc->snap_rwsem);
1977                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1978                                rinfo->snapblob + rinfo->snapblob_len,
1979                                le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1980                 downgrade_write(&mdsc->snap_rwsem);
1981         } else {
1982                 down_read(&mdsc->snap_rwsem);
1983         }
1984
1985         /* insert trace into our cache */
1986         mutex_lock(&req->r_fill_mutex);
1987         err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1988         if (err == 0) {
1989                 if (result == 0 && rinfo->dir_nr)
1990                         ceph_readdir_prepopulate(req, req->r_session);
1991                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
1992         }
1993         mutex_unlock(&req->r_fill_mutex);
1994
1995         up_read(&mdsc->snap_rwsem);
1996 out_err:
1997         mutex_lock(&mdsc->mutex);
1998         if (!req->r_aborted) {
1999                 if (err) {
2000                         req->r_err = err;
2001                 } else {
2002                         req->r_reply = msg;
2003                         ceph_msg_get(msg);
2004                         req->r_got_result = true;
2005                 }
2006         } else {
2007                 dout("reply arrived after request %lld was aborted\n", tid);
2008         }
2009         mutex_unlock(&mdsc->mutex);
2010
2011         ceph_add_cap_releases(mdsc, req->r_session);
2012         mutex_unlock(&session->s_mutex);
2013
2014         /* kick calling process */
2015         complete_request(mdsc, req);
2016 out:
2017         ceph_mdsc_put_request(req);
2018         return;
2019 }
2020
2021
2022
2023 /*
2024  * handle mds notification that our request has been forwarded.
2025  */
2026 static void handle_forward(struct ceph_mds_client *mdsc,
2027                            struct ceph_mds_session *session,
2028                            struct ceph_msg *msg)
2029 {
2030         struct ceph_mds_request *req;
2031         u64 tid = le64_to_cpu(msg->hdr.tid);
2032         u32 next_mds;
2033         u32 fwd_seq;
2034         int err = -EINVAL;
2035         void *p = msg->front.iov_base;
2036         void *end = p + msg->front.iov_len;
2037
2038         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2039         next_mds = ceph_decode_32(&p);
2040         fwd_seq = ceph_decode_32(&p);
2041
2042         mutex_lock(&mdsc->mutex);
2043         req = __lookup_request(mdsc, tid);
2044         if (!req) {
2045                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2046                 goto out;  /* dup reply? */
2047         }
2048
2049         if (req->r_aborted) {
2050                 dout("forward tid %llu aborted, unregistering\n", tid);
2051                 __unregister_request(mdsc, req);
2052         } else if (fwd_seq <= req->r_num_fwd) {
2053                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2054                      tid, next_mds, req->r_num_fwd, fwd_seq);
2055         } else {
2056                 /* resend. forward race not possible; mds would drop */
2057                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2058                 BUG_ON(req->r_err);
2059                 BUG_ON(req->r_got_result);
2060                 req->r_num_fwd = fwd_seq;
2061                 req->r_resend_mds = next_mds;
2062                 put_request_session(req);
2063                 __do_request(mdsc, req);
2064         }
2065         ceph_mdsc_put_request(req);
2066 out:
2067         mutex_unlock(&mdsc->mutex);
2068         return;
2069
2070 bad:
2071         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2072 }
2073
2074 /*
2075  * handle a mds session control message
2076  */
2077 static void handle_session(struct ceph_mds_session *session,
2078                            struct ceph_msg *msg)
2079 {
2080         struct ceph_mds_client *mdsc = session->s_mdsc;
2081         u32 op;
2082         u64 seq;
2083         int mds = session->s_mds;
2084         struct ceph_mds_session_head *h = msg->front.iov_base;
2085         int wake = 0;
2086
2087         /* decode */
2088         if (msg->front.iov_len != sizeof(*h))
2089                 goto bad;
2090         op = le32_to_cpu(h->op);
2091         seq = le64_to_cpu(h->seq);
2092
2093         mutex_lock(&mdsc->mutex);
2094         if (op == CEPH_SESSION_CLOSE)
2095                 __unregister_session(mdsc, session);
2096         /* FIXME: this ttl calculation is generous */
2097         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2098         mutex_unlock(&mdsc->mutex);
2099
2100         mutex_lock(&session->s_mutex);
2101
2102         dout("handle_session mds%d %s %p state %s seq %llu\n",
2103              mds, ceph_session_op_name(op), session,
2104              session_state_name(session->s_state), seq);
2105
2106         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2107                 session->s_state = CEPH_MDS_SESSION_OPEN;
2108                 pr_info("mds%d came back\n", session->s_mds);
2109         }
2110
2111         switch (op) {
2112         case CEPH_SESSION_OPEN:
2113                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2114                         pr_info("mds%d reconnect success\n", session->s_mds);
2115                 session->s_state = CEPH_MDS_SESSION_OPEN;
2116                 renewed_caps(mdsc, session, 0);
2117                 wake = 1;
2118                 if (mdsc->stopping)
2119                         __close_session(mdsc, session);
2120                 break;
2121
2122         case CEPH_SESSION_RENEWCAPS:
2123                 if (session->s_renew_seq == seq)
2124                         renewed_caps(mdsc, session, 1);
2125                 break;
2126
2127         case CEPH_SESSION_CLOSE:
2128                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2129                         pr_info("mds%d reconnect denied\n", session->s_mds);
2130                 remove_session_caps(session);
2131                 wake = 1; /* for good measure */
2132                 complete_all(&mdsc->session_close_waiters);
2133                 kick_requests(mdsc, mds);
2134                 break;
2135
2136         case CEPH_SESSION_STALE:
2137                 pr_info("mds%d caps went stale, renewing\n",
2138                         session->s_mds);
2139                 spin_lock(&session->s_cap_lock);
2140                 session->s_cap_gen++;
2141                 session->s_cap_ttl = 0;
2142                 spin_unlock(&session->s_cap_lock);
2143                 send_renew_caps(mdsc, session);
2144                 break;
2145
2146         case CEPH_SESSION_RECALL_STATE:
2147                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2148                 break;
2149
2150         default:
2151                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2152                 WARN_ON(1);
2153         }
2154
2155         mutex_unlock(&session->s_mutex);
2156         if (wake) {
2157                 mutex_lock(&mdsc->mutex);
2158                 __wake_requests(mdsc, &session->s_waiting);
2159                 mutex_unlock(&mdsc->mutex);
2160         }
2161         return;
2162
2163 bad:
2164         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2165                (int)msg->front.iov_len);
2166         ceph_msg_dump(msg);
2167         return;
2168 }
2169
2170
2171 /*
2172  * called under session->mutex.
2173  */
2174 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2175                                    struct ceph_mds_session *session)
2176 {
2177         struct ceph_mds_request *req, *nreq;
2178         int err;
2179
2180         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2181
2182         mutex_lock(&mdsc->mutex);
2183         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2184                 err = __prepare_send_request(mdsc, req, session->s_mds);
2185                 if (!err) {
2186                         ceph_msg_get(req->r_request);
2187                         ceph_con_send(&session->s_con, req->r_request);
2188                 }
2189         }
2190         mutex_unlock(&mdsc->mutex);
2191 }
2192
2193 /*
2194  * Encode information about a cap for a reconnect with the MDS.
2195  */
2196 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2197                           void *arg)
2198 {
2199         struct ceph_mds_cap_reconnect rec;
2200         struct ceph_inode_info *ci;
2201         struct ceph_pagelist *pagelist = arg;
2202         char *path;
2203         int pathlen, err;
2204         u64 pathbase;
2205         struct dentry *dentry;
2206
2207         ci = cap->ci;
2208
2209         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2210              inode, ceph_vinop(inode), cap, cap->cap_id,
2211              ceph_cap_string(cap->issued));
2212         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2213         if (err)
2214                 return err;
2215
2216         dentry = d_find_alias(inode);
2217         if (dentry) {
2218                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2219                 if (IS_ERR(path)) {
2220                         err = PTR_ERR(path);
2221                         BUG_ON(err);
2222                 }
2223         } else {
2224                 path = NULL;
2225                 pathlen = 0;
2226         }
2227         err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2228         if (err)
2229                 goto out;
2230
2231         spin_lock(&inode->i_lock);
2232         cap->seq = 0;        /* reset cap seq */
2233         cap->issue_seq = 0;  /* and issue_seq */
2234         rec.cap_id = cpu_to_le64(cap->cap_id);
2235         rec.pathbase = cpu_to_le64(pathbase);
2236         rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2237         rec.issued = cpu_to_le32(cap->issued);
2238         rec.size = cpu_to_le64(inode->i_size);
2239         ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2240         ceph_encode_timespec(&rec.atime, &inode->i_atime);
2241         rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2242         spin_unlock(&inode->i_lock);
2243
2244         err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2245
2246 out:
2247         kfree(path);
2248         dput(dentry);
2249         return err;
2250 }
2251
2252
2253 /*
2254  * If an MDS fails and recovers, clients need to reconnect in order to
2255  * reestablish shared state.  This includes all caps issued through
2256  * this session _and_ the snap_realm hierarchy.  Because it's not
2257  * clear which snap realms the mds cares about, we send everything we
2258  * know about.. that ensures we'll then get any new info the
2259  * recovering MDS might have.
2260  *
2261  * This is a relatively heavyweight operation, but it's rare.
2262  *
2263  * called with mdsc->mutex held.
2264  */
2265 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2266                                struct ceph_mds_session *session)
2267 {
2268         struct ceph_msg *reply;
2269         struct rb_node *p;
2270         int mds = session->s_mds;
2271         int err = -ENOMEM;
2272         struct ceph_pagelist *pagelist;
2273
2274         pr_info("mds%d reconnect start\n", mds);
2275
2276         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2277         if (!pagelist)
2278                 goto fail_nopagelist;
2279         ceph_pagelist_init(pagelist);
2280
2281         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS);
2282         if (!reply)
2283                 goto fail_nomsg;
2284
2285         mutex_lock(&session->s_mutex);
2286         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2287         session->s_seq = 0;
2288
2289         ceph_con_open(&session->s_con,
2290                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2291
2292         /* replay unsafe requests */
2293         replay_unsafe_requests(mdsc, session);
2294
2295         down_read(&mdsc->snap_rwsem);
2296
2297         dout("session %p state %s\n", session,
2298              session_state_name(session->s_state));
2299
2300         /* drop old cap expires; we're about to reestablish that state */
2301         discard_cap_releases(mdsc, session);
2302
2303         /* traverse this session's caps */
2304         err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2305         if (err)
2306                 goto fail;
2307         err = iterate_session_caps(session, encode_caps_cb, pagelist);
2308         if (err < 0)
2309                 goto fail;
2310
2311         /*
2312          * snaprealms.  we provide mds with the ino, seq (version), and
2313          * parent for all of our realms.  If the mds has any newer info,
2314          * it will tell us.
2315          */
2316         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2317                 struct ceph_snap_realm *realm =
2318                         rb_entry(p, struct ceph_snap_realm, node);
2319                 struct ceph_mds_snaprealm_reconnect sr_rec;
2320
2321                 dout(" adding snap realm %llx seq %lld parent %llx\n",
2322                      realm->ino, realm->seq, realm->parent_ino);
2323                 sr_rec.ino = cpu_to_le64(realm->ino);
2324                 sr_rec.seq = cpu_to_le64(realm->seq);
2325                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2326                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2327                 if (err)
2328                         goto fail;
2329         }
2330
2331         reply->pagelist = pagelist;
2332         reply->hdr.data_len = cpu_to_le32(pagelist->length);
2333         reply->nr_pages = calc_pages_for(0, pagelist->length);
2334         ceph_con_send(&session->s_con, reply);
2335
2336         mutex_unlock(&session->s_mutex);
2337
2338         mutex_lock(&mdsc->mutex);
2339         __wake_requests(mdsc, &session->s_waiting);
2340         mutex_unlock(&mdsc->mutex);
2341
2342         up_read(&mdsc->snap_rwsem);
2343         return;
2344
2345 fail:
2346         ceph_msg_put(reply);
2347         up_read(&mdsc->snap_rwsem);
2348         mutex_unlock(&session->s_mutex);
2349 fail_nomsg:
2350         ceph_pagelist_release(pagelist);
2351         kfree(pagelist);
2352 fail_nopagelist:
2353         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2354         return;
2355 }
2356
2357
2358 /*
2359  * compare old and new mdsmaps, kicking requests
2360  * and closing out old connections as necessary
2361  *
2362  * called under mdsc->mutex.
2363  */
2364 static void check_new_map(struct ceph_mds_client *mdsc,
2365                           struct ceph_mdsmap *newmap,
2366                           struct ceph_mdsmap *oldmap)
2367 {
2368         int i;
2369         int oldstate, newstate;
2370         struct ceph_mds_session *s;
2371
2372         dout("check_new_map new %u old %u\n",
2373              newmap->m_epoch, oldmap->m_epoch);
2374
2375         for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2376                 if (mdsc->sessions[i] == NULL)
2377                         continue;
2378                 s = mdsc->sessions[i];
2379                 oldstate = ceph_mdsmap_get_state(oldmap, i);
2380                 newstate = ceph_mdsmap_get_state(newmap, i);
2381
2382                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2383                      i, ceph_mds_state_name(oldstate),
2384                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2385                      ceph_mds_state_name(newstate),
2386                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2387                      session_state_name(s->s_state));
2388
2389                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2390                            ceph_mdsmap_get_addr(newmap, i),
2391                            sizeof(struct ceph_entity_addr))) {
2392                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2393                                 /* the session never opened, just close it
2394                                  * out now */
2395                                 __wake_requests(mdsc, &s->s_waiting);
2396                                 __unregister_session(mdsc, s);
2397                         } else {
2398                                 /* just close it */
2399                                 mutex_unlock(&mdsc->mutex);
2400                                 mutex_lock(&s->s_mutex);
2401                                 mutex_lock(&mdsc->mutex);
2402                                 ceph_con_close(&s->s_con);
2403                                 mutex_unlock(&s->s_mutex);
2404                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2405                         }
2406
2407                         /* kick any requests waiting on the recovering mds */
2408                         kick_requests(mdsc, i);
2409                 } else if (oldstate == newstate) {
2410                         continue;  /* nothing new with this mds */
2411                 }
2412
2413                 /*
2414                  * send reconnect?
2415                  */
2416                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2417                     newstate >= CEPH_MDS_STATE_RECONNECT) {
2418                         mutex_unlock(&mdsc->mutex);
2419                         send_mds_reconnect(mdsc, s);
2420                         mutex_lock(&mdsc->mutex);
2421                 }
2422
2423                 /*
2424                  * kick request on any mds that has gone active.
2425                  */
2426                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2427                     newstate >= CEPH_MDS_STATE_ACTIVE) {
2428                         if (oldstate != CEPH_MDS_STATE_CREATING &&
2429                             oldstate != CEPH_MDS_STATE_STARTING)
2430                                 pr_info("mds%d recovery completed\n", s->s_mds);
2431                         kick_requests(mdsc, i);
2432                         ceph_kick_flushing_caps(mdsc, s);
2433                         wake_up_session_caps(s, 1);
2434                 }
2435         }
2436 }
2437
2438
2439
2440 /*
2441  * leases
2442  */
2443
2444 /*
2445  * caller must hold session s_mutex, dentry->d_lock
2446  */
2447 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2448 {
2449         struct ceph_dentry_info *di = ceph_dentry(dentry);
2450
2451         ceph_put_mds_session(di->lease_session);
2452         di->lease_session = NULL;
2453 }
2454
2455 static void handle_lease(struct ceph_mds_client *mdsc,
2456                          struct ceph_mds_session *session,
2457                          struct ceph_msg *msg)
2458 {
2459         struct super_block *sb = mdsc->client->sb;
2460         struct inode *inode;
2461         struct ceph_inode_info *ci;
2462         struct dentry *parent, *dentry;
2463         struct ceph_dentry_info *di;
2464         int mds = session->s_mds;
2465         struct ceph_mds_lease *h = msg->front.iov_base;
2466         u32 seq;
2467         struct ceph_vino vino;
2468         int mask;
2469         struct qstr dname;
2470         int release = 0;
2471
2472         dout("handle_lease from mds%d\n", mds);
2473
2474         /* decode */
2475         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2476                 goto bad;
2477         vino.ino = le64_to_cpu(h->ino);
2478         vino.snap = CEPH_NOSNAP;
2479         mask = le16_to_cpu(h->mask);
2480         seq = le32_to_cpu(h->seq);
2481         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2482         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2483         if (dname.len != get_unaligned_le32(h+1))
2484                 goto bad;
2485
2486         mutex_lock(&session->s_mutex);
2487         session->s_seq++;
2488
2489         /* lookup inode */
2490         inode = ceph_find_inode(sb, vino);
2491         dout("handle_lease %s, mask %d, ino %llx %p %.*s\n",
2492              ceph_lease_op_name(h->action), mask, vino.ino, inode,
2493              dname.len, dname.name);
2494         if (inode == NULL) {
2495                 dout("handle_lease no inode %llx\n", vino.ino);
2496                 goto release;
2497         }
2498         ci = ceph_inode(inode);
2499
2500         /* dentry */
2501         parent = d_find_alias(inode);
2502         if (!parent) {
2503                 dout("no parent dentry on inode %p\n", inode);
2504                 WARN_ON(1);
2505                 goto release;  /* hrm... */
2506         }
2507         dname.hash = full_name_hash(dname.name, dname.len);
2508         dentry = d_lookup(parent, &dname);
2509         dput(parent);
2510         if (!dentry)
2511                 goto release;
2512
2513         spin_lock(&dentry->d_lock);
2514         di = ceph_dentry(dentry);
2515         switch (h->action) {
2516         case CEPH_MDS_LEASE_REVOKE:
2517                 if (di && di->lease_session == session) {
2518                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2519                                 h->seq = cpu_to_le32(di->lease_seq);
2520                         __ceph_mdsc_drop_dentry_lease(dentry);
2521                 }
2522                 release = 1;
2523                 break;
2524
2525         case CEPH_MDS_LEASE_RENEW:
2526                 if (di && di->lease_session == session &&
2527                     di->lease_gen == session->s_cap_gen &&
2528                     di->lease_renew_from &&
2529                     di->lease_renew_after == 0) {
2530                         unsigned long duration =
2531                                 le32_to_cpu(h->duration_ms) * HZ / 1000;
2532
2533                         di->lease_seq = seq;
2534                         dentry->d_time = di->lease_renew_from + duration;
2535                         di->lease_renew_after = di->lease_renew_from +
2536                                 (duration >> 1);
2537                         di->lease_renew_from = 0;
2538                 }
2539                 break;
2540         }
2541         spin_unlock(&dentry->d_lock);
2542         dput(dentry);
2543
2544         if (!release)
2545                 goto out;
2546
2547 release:
2548         /* let's just reuse the same message */
2549         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2550         ceph_msg_get(msg);
2551         ceph_con_send(&session->s_con, msg);
2552
2553 out:
2554         iput(inode);
2555         mutex_unlock(&session->s_mutex);
2556         return;
2557
2558 bad:
2559         pr_err("corrupt lease message\n");
2560         ceph_msg_dump(msg);
2561 }
2562
2563 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2564                               struct inode *inode,
2565                               struct dentry *dentry, char action,
2566                               u32 seq)
2567 {
2568         struct ceph_msg *msg;
2569         struct ceph_mds_lease *lease;
2570         int len = sizeof(*lease) + sizeof(u32);
2571         int dnamelen = 0;
2572
2573         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2574              inode, dentry, ceph_lease_op_name(action), session->s_mds);
2575         dnamelen = dentry->d_name.len;
2576         len += dnamelen;
2577
2578         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS);
2579         if (!msg)
2580                 return;
2581         lease = msg->front.iov_base;
2582         lease->action = action;
2583         lease->mask = cpu_to_le16(1);
2584         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2585         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2586         lease->seq = cpu_to_le32(seq);
2587         put_unaligned_le32(dnamelen, lease + 1);
2588         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2589
2590         /*
2591          * if this is a preemptive lease RELEASE, no need to
2592          * flush request stream, since the actual request will
2593          * soon follow.
2594          */
2595         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2596
2597         ceph_con_send(&session->s_con, msg);
2598 }
2599
2600 /*
2601  * Preemptively release a lease we expect to invalidate anyway.
2602  * Pass @inode always, @dentry is optional.
2603  */
2604 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2605                              struct dentry *dentry, int mask)
2606 {
2607         struct ceph_dentry_info *di;
2608         struct ceph_mds_session *session;
2609         u32 seq;
2610
2611         BUG_ON(inode == NULL);
2612         BUG_ON(dentry == NULL);
2613         BUG_ON(mask == 0);
2614
2615         /* is dentry lease valid? */
2616         spin_lock(&dentry->d_lock);
2617         di = ceph_dentry(dentry);
2618         if (!di || !di->lease_session ||
2619             di->lease_session->s_mds < 0 ||
2620             di->lease_gen != di->lease_session->s_cap_gen ||
2621             !time_before(jiffies, dentry->d_time)) {
2622                 dout("lease_release inode %p dentry %p -- "
2623                      "no lease on %d\n",
2624                      inode, dentry, mask);
2625                 spin_unlock(&dentry->d_lock);
2626                 return;
2627         }
2628
2629         /* we do have a lease on this dentry; note mds and seq */
2630         session = ceph_get_mds_session(di->lease_session);
2631         seq = di->lease_seq;
2632         __ceph_mdsc_drop_dentry_lease(dentry);
2633         spin_unlock(&dentry->d_lock);
2634
2635         dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2636              inode, dentry, mask, session->s_mds);
2637         ceph_mdsc_lease_send_msg(session, inode, dentry,
2638                                  CEPH_MDS_LEASE_RELEASE, seq);
2639         ceph_put_mds_session(session);
2640 }
2641
2642 /*
2643  * drop all leases (and dentry refs) in preparation for umount
2644  */
2645 static void drop_leases(struct ceph_mds_client *mdsc)
2646 {
2647         int i;
2648
2649         dout("drop_leases\n");
2650         mutex_lock(&mdsc->mutex);
2651         for (i = 0; i < mdsc->max_sessions; i++) {
2652                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2653                 if (!s)
2654                         continue;
2655                 mutex_unlock(&mdsc->mutex);
2656                 mutex_lock(&s->s_mutex);
2657                 mutex_unlock(&s->s_mutex);
2658                 ceph_put_mds_session(s);
2659                 mutex_lock(&mdsc->mutex);
2660         }
2661         mutex_unlock(&mdsc->mutex);
2662 }
2663
2664
2665
2666 /*
2667  * delayed work -- periodically trim expired leases, renew caps with mds
2668  */
2669 static void schedule_delayed(struct ceph_mds_client *mdsc)
2670 {
2671         int delay = 5;
2672         unsigned hz = round_jiffies_relative(HZ * delay);
2673         schedule_delayed_work(&mdsc->delayed_work, hz);
2674 }
2675
2676 static void delayed_work(struct work_struct *work)
2677 {
2678         int i;
2679         struct ceph_mds_client *mdsc =
2680                 container_of(work, struct ceph_mds_client, delayed_work.work);
2681         int renew_interval;
2682         int renew_caps;
2683
2684         dout("mdsc delayed_work\n");
2685         ceph_check_delayed_caps(mdsc);
2686
2687         mutex_lock(&mdsc->mutex);
2688         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2689         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2690                                    mdsc->last_renew_caps);
2691         if (renew_caps)
2692                 mdsc->last_renew_caps = jiffies;
2693
2694         for (i = 0; i < mdsc->max_sessions; i++) {
2695                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2696                 if (s == NULL)
2697                         continue;
2698                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2699                         dout("resending session close request for mds%d\n",
2700                              s->s_mds);
2701                         request_close_session(mdsc, s);
2702                         ceph_put_mds_session(s);
2703                         continue;
2704                 }
2705                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2706                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2707                                 s->s_state = CEPH_MDS_SESSION_HUNG;
2708                                 pr_info("mds%d hung\n", s->s_mds);
2709                         }
2710                 }
2711                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2712                         /* this mds is failed or recovering, just wait */
2713                         ceph_put_mds_session(s);
2714                         continue;
2715                 }
2716                 mutex_unlock(&mdsc->mutex);
2717
2718                 mutex_lock(&s->s_mutex);
2719                 if (renew_caps)
2720                         send_renew_caps(mdsc, s);
2721                 else
2722                         ceph_con_keepalive(&s->s_con);
2723                 ceph_add_cap_releases(mdsc, s);
2724                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2725                     s->s_state == CEPH_MDS_SESSION_HUNG)
2726                         ceph_send_cap_releases(mdsc, s);
2727                 mutex_unlock(&s->s_mutex);
2728                 ceph_put_mds_session(s);
2729
2730                 mutex_lock(&mdsc->mutex);
2731         }
2732         mutex_unlock(&mdsc->mutex);
2733
2734         schedule_delayed(mdsc);
2735 }
2736
2737
2738 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2739 {
2740         mdsc->client = client;
2741         mutex_init(&mdsc->mutex);
2742         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2743         if (mdsc->mdsmap == NULL)
2744                 return -ENOMEM;
2745
2746         init_completion(&mdsc->safe_umount_waiters);
2747         init_completion(&mdsc->session_close_waiters);
2748         INIT_LIST_HEAD(&mdsc->waiting_for_map);
2749         mdsc->sessions = NULL;
2750         mdsc->max_sessions = 0;
2751         mdsc->stopping = 0;
2752         init_rwsem(&mdsc->snap_rwsem);
2753         mdsc->snap_realms = RB_ROOT;
2754         INIT_LIST_HEAD(&mdsc->snap_empty);
2755         spin_lock_init(&mdsc->snap_empty_lock);
2756         mdsc->last_tid = 0;
2757         mdsc->request_tree = RB_ROOT;
2758         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2759         mdsc->last_renew_caps = jiffies;
2760         INIT_LIST_HEAD(&mdsc->cap_delay_list);
2761         spin_lock_init(&mdsc->cap_delay_lock);
2762         INIT_LIST_HEAD(&mdsc->snap_flush_list);
2763         spin_lock_init(&mdsc->snap_flush_lock);
2764         mdsc->cap_flush_seq = 0;
2765         INIT_LIST_HEAD(&mdsc->cap_dirty);
2766         mdsc->num_cap_flushing = 0;
2767         spin_lock_init(&mdsc->cap_dirty_lock);
2768         init_waitqueue_head(&mdsc->cap_flushing_wq);
2769         spin_lock_init(&mdsc->dentry_lru_lock);
2770         INIT_LIST_HEAD(&mdsc->dentry_lru);
2771
2772         ceph_caps_init(mdsc);
2773         ceph_adjust_min_caps(mdsc, client->min_caps);
2774
2775         return 0;
2776 }
2777
2778 /*
2779  * Wait for safe replies on open mds requests.  If we time out, drop
2780  * all requests from the tree to avoid dangling dentry refs.
2781  */
2782 static void wait_requests(struct ceph_mds_client *mdsc)
2783 {
2784         struct ceph_mds_request *req;
2785         struct ceph_client *client = mdsc->client;
2786
2787         mutex_lock(&mdsc->mutex);
2788         if (__get_oldest_req(mdsc)) {
2789                 mutex_unlock(&mdsc->mutex);
2790
2791                 dout("wait_requests waiting for requests\n");
2792                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2793                                     client->mount_args->mount_timeout * HZ);
2794
2795                 /* tear down remaining requests */
2796                 mutex_lock(&mdsc->mutex);
2797                 while ((req = __get_oldest_req(mdsc))) {
2798                         dout("wait_requests timed out on tid %llu\n",
2799                              req->r_tid);
2800                         __unregister_request(mdsc, req);
2801                 }
2802         }
2803         mutex_unlock(&mdsc->mutex);
2804         dout("wait_requests done\n");
2805 }
2806
2807 /*
2808  * called before mount is ro, and before dentries are torn down.
2809  * (hmm, does this still race with new lookups?)
2810  */
2811 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2812 {
2813         dout("pre_umount\n");
2814         mdsc->stopping = 1;
2815
2816         drop_leases(mdsc);
2817         ceph_flush_dirty_caps(mdsc);
2818         wait_requests(mdsc);
2819
2820         /*
2821          * wait for reply handlers to drop their request refs and
2822          * their inode/dcache refs
2823          */
2824         ceph_msgr_flush();
2825 }
2826
2827 /*
2828  * wait for all write mds requests to flush.
2829  */
2830 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2831 {
2832         struct ceph_mds_request *req = NULL, *nextreq;
2833         struct rb_node *n;
2834
2835         mutex_lock(&mdsc->mutex);
2836         dout("wait_unsafe_requests want %lld\n", want_tid);
2837 restart:
2838         req = __get_oldest_req(mdsc);
2839         while (req && req->r_tid <= want_tid) {
2840                 /* find next request */
2841                 n = rb_next(&req->r_node);
2842                 if (n)
2843                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2844                 else
2845                         nextreq = NULL;
2846                 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2847                         /* write op */
2848                         ceph_mdsc_get_request(req);
2849                         if (nextreq)
2850                                 ceph_mdsc_get_request(nextreq);
2851                         mutex_unlock(&mdsc->mutex);
2852                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
2853                              req->r_tid, want_tid);
2854                         wait_for_completion(&req->r_safe_completion);
2855                         mutex_lock(&mdsc->mutex);
2856                         ceph_mdsc_put_request(req);
2857                         if (!nextreq)
2858                                 break;  /* next dne before, so we're done! */
2859                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
2860                                 /* next request was removed from tree */
2861                                 ceph_mdsc_put_request(nextreq);
2862                                 goto restart;
2863                         }
2864                         ceph_mdsc_put_request(nextreq);  /* won't go away */
2865                 }
2866                 req = nextreq;
2867         }
2868         mutex_unlock(&mdsc->mutex);
2869         dout("wait_unsafe_requests done\n");
2870 }
2871
2872 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2873 {
2874         u64 want_tid, want_flush;
2875
2876         if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
2877                 return;
2878
2879         dout("sync\n");
2880         mutex_lock(&mdsc->mutex);
2881         want_tid = mdsc->last_tid;
2882         want_flush = mdsc->cap_flush_seq;
2883         mutex_unlock(&mdsc->mutex);
2884         dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2885
2886         ceph_flush_dirty_caps(mdsc);
2887
2888         wait_unsafe_requests(mdsc, want_tid);
2889         wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2890 }
2891
2892
2893 /*
2894  * called after sb is ro.
2895  */
2896 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2897 {
2898         struct ceph_mds_session *session;
2899         int i;
2900         int n;
2901         struct ceph_client *client = mdsc->client;
2902         unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2903
2904         dout("close_sessions\n");
2905
2906         mutex_lock(&mdsc->mutex);
2907
2908         /* close sessions */
2909         started = jiffies;
2910         while (time_before(jiffies, started + timeout)) {
2911                 dout("closing sessions\n");
2912                 n = 0;
2913                 for (i = 0; i < mdsc->max_sessions; i++) {
2914                         session = __ceph_lookup_mds_session(mdsc, i);
2915                         if (!session)
2916                                 continue;
2917                         mutex_unlock(&mdsc->mutex);
2918                         mutex_lock(&session->s_mutex);
2919                         __close_session(mdsc, session);
2920                         mutex_unlock(&session->s_mutex);
2921                         ceph_put_mds_session(session);
2922                         mutex_lock(&mdsc->mutex);
2923                         n++;
2924                 }
2925                 if (n == 0)
2926                         break;
2927
2928                 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2929                         break;
2930
2931                 dout("waiting for sessions to close\n");
2932                 mutex_unlock(&mdsc->mutex);
2933                 wait_for_completion_timeout(&mdsc->session_close_waiters,
2934                                             timeout);
2935                 mutex_lock(&mdsc->mutex);
2936         }
2937
2938         /* tear down remaining sessions */
2939         for (i = 0; i < mdsc->max_sessions; i++) {
2940                 if (mdsc->sessions[i]) {
2941                         session = get_session(mdsc->sessions[i]);
2942                         __unregister_session(mdsc, session);
2943                         mutex_unlock(&mdsc->mutex);
2944                         mutex_lock(&session->s_mutex);
2945                         remove_session_caps(session);
2946                         mutex_unlock(&session->s_mutex);
2947                         ceph_put_mds_session(session);
2948                         mutex_lock(&mdsc->mutex);
2949                 }
2950         }
2951
2952         WARN_ON(!list_empty(&mdsc->cap_delay_list));
2953
2954         mutex_unlock(&mdsc->mutex);
2955
2956         ceph_cleanup_empty_realms(mdsc);
2957
2958         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2959
2960         dout("stopped\n");
2961 }
2962
2963 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2964 {
2965         dout("stop\n");
2966         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2967         if (mdsc->mdsmap)
2968                 ceph_mdsmap_destroy(mdsc->mdsmap);
2969         kfree(mdsc->sessions);
2970         ceph_caps_finalize(mdsc);
2971 }
2972
2973
2974 /*
2975  * handle mds map update.
2976  */
2977 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2978 {
2979         u32 epoch;
2980         u32 maplen;
2981         void *p = msg->front.iov_base;
2982         void *end = p + msg->front.iov_len;
2983         struct ceph_mdsmap *newmap, *oldmap;
2984         struct ceph_fsid fsid;
2985         int err = -EINVAL;
2986
2987         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2988         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2989         if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2990                 return;
2991         epoch = ceph_decode_32(&p);
2992         maplen = ceph_decode_32(&p);
2993         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2994
2995         /* do we need it? */
2996         ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2997         mutex_lock(&mdsc->mutex);
2998         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2999                 dout("handle_map epoch %u <= our %u\n",
3000                      epoch, mdsc->mdsmap->m_epoch);
3001                 mutex_unlock(&mdsc->mutex);
3002                 return;
3003         }
3004
3005         newmap = ceph_mdsmap_decode(&p, end);
3006         if (IS_ERR(newmap)) {
3007                 err = PTR_ERR(newmap);
3008                 goto bad_unlock;
3009         }
3010
3011         /* swap into place */
3012         if (mdsc->mdsmap) {
3013                 oldmap = mdsc->mdsmap;
3014                 mdsc->mdsmap = newmap;
3015                 check_new_map(mdsc, newmap, oldmap);
3016                 ceph_mdsmap_destroy(oldmap);
3017         } else {
3018                 mdsc->mdsmap = newmap;  /* first mds map */
3019         }
3020         mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3021
3022         __wake_requests(mdsc, &mdsc->waiting_for_map);
3023
3024         mutex_unlock(&mdsc->mutex);
3025         schedule_delayed(mdsc);
3026         return;
3027
3028 bad_unlock:
3029         mutex_unlock(&mdsc->mutex);
3030 bad:
3031         pr_err("error decoding mdsmap %d\n", err);
3032         return;
3033 }
3034
3035 static struct ceph_connection *con_get(struct ceph_connection *con)
3036 {
3037         struct ceph_mds_session *s = con->private;
3038
3039         if (get_session(s)) {
3040                 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3041                 return con;
3042         }
3043         dout("mdsc con_get %p FAIL\n", s);
3044         return NULL;
3045 }
3046
3047 static void con_put(struct ceph_connection *con)
3048 {
3049         struct ceph_mds_session *s = con->private;
3050
3051         ceph_put_mds_session(s);
3052         dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
3053 }
3054
3055 /*
3056  * if the client is unresponsive for long enough, the mds will kill
3057  * the session entirely.
3058  */
3059 static void peer_reset(struct ceph_connection *con)
3060 {
3061         struct ceph_mds_session *s = con->private;
3062         struct ceph_mds_client *mdsc = s->s_mdsc;
3063
3064         pr_warning("mds%d closed our session\n", s->s_mds);
3065         send_mds_reconnect(mdsc, s);
3066 }
3067
3068 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3069 {
3070         struct ceph_mds_session *s = con->private;
3071         struct ceph_mds_client *mdsc = s->s_mdsc;
3072         int type = le16_to_cpu(msg->hdr.type);
3073
3074         mutex_lock(&mdsc->mutex);
3075         if (__verify_registered_session(mdsc, s) < 0) {
3076                 mutex_unlock(&mdsc->mutex);
3077                 goto out;
3078         }
3079         mutex_unlock(&mdsc->mutex);
3080
3081         switch (type) {
3082         case CEPH_MSG_MDS_MAP:
3083                 ceph_mdsc_handle_map(mdsc, msg);
3084                 break;
3085         case CEPH_MSG_CLIENT_SESSION:
3086                 handle_session(s, msg);
3087                 break;
3088         case CEPH_MSG_CLIENT_REPLY:
3089                 handle_reply(s, msg);
3090                 break;
3091         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3092                 handle_forward(mdsc, s, msg);
3093                 break;
3094         case CEPH_MSG_CLIENT_CAPS:
3095                 ceph_handle_caps(s, msg);
3096                 break;
3097         case CEPH_MSG_CLIENT_SNAP:
3098                 ceph_handle_snap(mdsc, s, msg);
3099                 break;
3100         case CEPH_MSG_CLIENT_LEASE:
3101                 handle_lease(mdsc, s, msg);
3102                 break;
3103
3104         default:
3105                 pr_err("received unknown message type %d %s\n", type,
3106                        ceph_msg_type_name(type));
3107         }
3108 out:
3109         ceph_msg_put(msg);
3110 }
3111
3112 /*
3113  * authentication
3114  */
3115 static int get_authorizer(struct ceph_connection *con,
3116                           void **buf, int *len, int *proto,
3117                           void **reply_buf, int *reply_len, int force_new)
3118 {
3119         struct ceph_mds_session *s = con->private;
3120         struct ceph_mds_client *mdsc = s->s_mdsc;
3121         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3122         int ret = 0;
3123
3124         if (force_new && s->s_authorizer) {
3125                 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3126                 s->s_authorizer = NULL;
3127         }
3128         if (s->s_authorizer == NULL) {
3129                 if (ac->ops->create_authorizer) {
3130                         ret = ac->ops->create_authorizer(
3131                                 ac, CEPH_ENTITY_TYPE_MDS,
3132                                 &s->s_authorizer,
3133                                 &s->s_authorizer_buf,
3134                                 &s->s_authorizer_buf_len,
3135                                 &s->s_authorizer_reply_buf,
3136                                 &s->s_authorizer_reply_buf_len);
3137                         if (ret)
3138                                 return ret;
3139                 }
3140         }
3141
3142         *proto = ac->protocol;
3143         *buf = s->s_authorizer_buf;
3144         *len = s->s_authorizer_buf_len;
3145         *reply_buf = s->s_authorizer_reply_buf;
3146         *reply_len = s->s_authorizer_reply_buf_len;
3147         return 0;
3148 }
3149
3150
3151 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3152 {
3153         struct ceph_mds_session *s = con->private;
3154         struct ceph_mds_client *mdsc = s->s_mdsc;
3155         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3156
3157         return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3158 }
3159
3160 static int invalidate_authorizer(struct ceph_connection *con)
3161 {
3162         struct ceph_mds_session *s = con->private;
3163         struct ceph_mds_client *mdsc = s->s_mdsc;
3164         struct ceph_auth_client *ac = mdsc->client->monc.auth;
3165
3166         if (ac->ops->invalidate_authorizer)
3167                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3168
3169         return ceph_monc_validate_auth(&mdsc->client->monc);
3170 }
3171
3172 static const struct ceph_connection_operations mds_con_ops = {
3173         .get = con_get,
3174         .put = con_put,
3175         .dispatch = dispatch,
3176         .get_authorizer = get_authorizer,
3177         .verify_authorizer_reply = verify_authorizer_reply,
3178         .invalidate_authorizer = invalidate_authorizer,
3179         .peer_reset = peer_reset,
3180 };
3181
3182
3183
3184
3185 /* eof */