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36  * are no longer valid.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