]>
Commit | Line | Data |
---|---|---|
2f2dc053 SW |
1 | #include "ceph_debug.h" |
2 | ||
3 | #include <linux/wait.h> | |
5a0e3ad6 | 4 | #include <linux/slab.h> |
2f2dc053 SW |
5 | #include <linux/sched.h> |
6 | ||
7 | #include "mds_client.h" | |
8 | #include "mon_client.h" | |
9 | #include "super.h" | |
10 | #include "messenger.h" | |
11 | #include "decode.h" | |
4e7a5dcd | 12 | #include "auth.h" |
93cea5be | 13 | #include "pagelist.h" |
2f2dc053 SW |
14 | |
15 | /* | |
16 | * A cluster of MDS (metadata server) daemons is responsible for | |
17 | * managing the file system namespace (the directory hierarchy and | |
18 | * inodes) and for coordinating shared access to storage. Metadata is | |
19 | * partitioning hierarchically across a number of servers, and that | |
20 | * partition varies over time as the cluster adjusts the distribution | |
21 | * in order to balance load. | |
22 | * | |
23 | * The MDS client is primarily responsible to managing synchronous | |
24 | * metadata requests for operations like open, unlink, and so forth. | |
25 | * If there is a MDS failure, we find out about it when we (possibly | |
26 | * request and) receive a new MDS map, and can resubmit affected | |
27 | * requests. | |
28 | * | |
29 | * For the most part, though, we take advantage of a lossless | |
30 | * communications channel to the MDS, and do not need to worry about | |
31 | * timing out or resubmitting requests. | |
32 | * | |
33 | * We maintain a stateful "session" with each MDS we interact with. | |
34 | * Within each session, we sent periodic heartbeat messages to ensure | |
35 | * any capabilities or leases we have been issues remain valid. If | |
36 | * the session times out and goes stale, our leases and capabilities | |
37 | * are no longer valid. | |
38 | */ | |
39 | ||
40 | static void __wake_requests(struct ceph_mds_client *mdsc, | |
41 | struct list_head *head); | |
42 | ||
43 | const static struct ceph_connection_operations mds_con_ops; | |
44 | ||
45 | ||
46 | /* | |
47 | * mds reply parsing | |
48 | */ | |
49 | ||
50 | /* | |
51 | * parse individual inode info | |
52 | */ | |
53 | static int parse_reply_info_in(void **p, void *end, | |
54 | struct ceph_mds_reply_info_in *info) | |
55 | { | |
56 | int err = -EIO; | |
57 | ||
58 | info->in = *p; | |
59 | *p += sizeof(struct ceph_mds_reply_inode) + | |
60 | sizeof(*info->in->fragtree.splits) * | |
61 | le32_to_cpu(info->in->fragtree.nsplits); | |
62 | ||
63 | ceph_decode_32_safe(p, end, info->symlink_len, bad); | |
64 | ceph_decode_need(p, end, info->symlink_len, bad); | |
65 | info->symlink = *p; | |
66 | *p += info->symlink_len; | |
67 | ||
68 | ceph_decode_32_safe(p, end, info->xattr_len, bad); | |
69 | ceph_decode_need(p, end, info->xattr_len, bad); | |
70 | info->xattr_data = *p; | |
71 | *p += info->xattr_len; | |
72 | return 0; | |
73 | bad: | |
74 | return err; | |
75 | } | |
76 | ||
77 | /* | |
78 | * parse a normal reply, which may contain a (dir+)dentry and/or a | |
79 | * target inode. | |
80 | */ | |
81 | static int parse_reply_info_trace(void **p, void *end, | |
82 | struct ceph_mds_reply_info_parsed *info) | |
83 | { | |
84 | int err; | |
85 | ||
86 | if (info->head->is_dentry) { | |
87 | err = parse_reply_info_in(p, end, &info->diri); | |
88 | if (err < 0) | |
89 | goto out_bad; | |
90 | ||
91 | if (unlikely(*p + sizeof(*info->dirfrag) > end)) | |
92 | goto bad; | |
93 | info->dirfrag = *p; | |
94 | *p += sizeof(*info->dirfrag) + | |
95 | sizeof(u32)*le32_to_cpu(info->dirfrag->ndist); | |
96 | if (unlikely(*p > end)) | |
97 | goto bad; | |
98 | ||
99 | ceph_decode_32_safe(p, end, info->dname_len, bad); | |
100 | ceph_decode_need(p, end, info->dname_len, bad); | |
101 | info->dname = *p; | |
102 | *p += info->dname_len; | |
103 | info->dlease = *p; | |
104 | *p += sizeof(*info->dlease); | |
105 | } | |
106 | ||
107 | if (info->head->is_target) { | |
108 | err = parse_reply_info_in(p, end, &info->targeti); | |
109 | if (err < 0) | |
110 | goto out_bad; | |
111 | } | |
112 | ||
113 | if (unlikely(*p != end)) | |
114 | goto bad; | |
115 | return 0; | |
116 | ||
117 | bad: | |
118 | err = -EIO; | |
119 | out_bad: | |
120 | pr_err("problem parsing mds trace %d\n", err); | |
121 | return err; | |
122 | } | |
123 | ||
124 | /* | |
125 | * parse readdir results | |
126 | */ | |
127 | static int parse_reply_info_dir(void **p, void *end, | |
128 | struct ceph_mds_reply_info_parsed *info) | |
129 | { | |
130 | u32 num, i = 0; | |
131 | int err; | |
132 | ||
133 | info->dir_dir = *p; | |
134 | if (*p + sizeof(*info->dir_dir) > end) | |
135 | goto bad; | |
136 | *p += sizeof(*info->dir_dir) + | |
137 | sizeof(u32)*le32_to_cpu(info->dir_dir->ndist); | |
138 | if (*p > end) | |
139 | goto bad; | |
140 | ||
141 | ceph_decode_need(p, end, sizeof(num) + 2, bad); | |
c89136ea SW |
142 | num = ceph_decode_32(p); |
143 | info->dir_end = ceph_decode_8(p); | |
144 | info->dir_complete = ceph_decode_8(p); | |
2f2dc053 SW |
145 | if (num == 0) |
146 | goto done; | |
147 | ||
148 | /* alloc large array */ | |
149 | info->dir_nr = num; | |
150 | info->dir_in = kcalloc(num, sizeof(*info->dir_in) + | |
151 | sizeof(*info->dir_dname) + | |
152 | sizeof(*info->dir_dname_len) + | |
153 | sizeof(*info->dir_dlease), | |
154 | GFP_NOFS); | |
155 | if (info->dir_in == NULL) { | |
156 | err = -ENOMEM; | |
157 | goto out_bad; | |
158 | } | |
159 | info->dir_dname = (void *)(info->dir_in + num); | |
160 | info->dir_dname_len = (void *)(info->dir_dname + num); | |
161 | info->dir_dlease = (void *)(info->dir_dname_len + num); | |
162 | ||
163 | while (num) { | |
164 | /* dentry */ | |
165 | ceph_decode_need(p, end, sizeof(u32)*2, bad); | |
c89136ea | 166 | info->dir_dname_len[i] = ceph_decode_32(p); |
2f2dc053 SW |
167 | ceph_decode_need(p, end, info->dir_dname_len[i], bad); |
168 | info->dir_dname[i] = *p; | |
169 | *p += info->dir_dname_len[i]; | |
170 | dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i], | |
171 | info->dir_dname[i]); | |
172 | info->dir_dlease[i] = *p; | |
173 | *p += sizeof(struct ceph_mds_reply_lease); | |
174 | ||
175 | /* inode */ | |
176 | err = parse_reply_info_in(p, end, &info->dir_in[i]); | |
177 | if (err < 0) | |
178 | goto out_bad; | |
179 | i++; | |
180 | num--; | |
181 | } | |
182 | ||
183 | done: | |
184 | if (*p != end) | |
185 | goto bad; | |
186 | return 0; | |
187 | ||
188 | bad: | |
189 | err = -EIO; | |
190 | out_bad: | |
191 | pr_err("problem parsing dir contents %d\n", err); | |
192 | return err; | |
193 | } | |
194 | ||
195 | /* | |
196 | * parse entire mds reply | |
197 | */ | |
198 | static int parse_reply_info(struct ceph_msg *msg, | |
199 | struct ceph_mds_reply_info_parsed *info) | |
200 | { | |
201 | void *p, *end; | |
202 | u32 len; | |
203 | int err; | |
204 | ||
205 | info->head = msg->front.iov_base; | |
206 | p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head); | |
207 | end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head); | |
208 | ||
209 | /* trace */ | |
210 | ceph_decode_32_safe(&p, end, len, bad); | |
211 | if (len > 0) { | |
212 | err = parse_reply_info_trace(&p, p+len, info); | |
213 | if (err < 0) | |
214 | goto out_bad; | |
215 | } | |
216 | ||
217 | /* dir content */ | |
218 | ceph_decode_32_safe(&p, end, len, bad); | |
219 | if (len > 0) { | |
220 | err = parse_reply_info_dir(&p, p+len, info); | |
221 | if (err < 0) | |
222 | goto out_bad; | |
223 | } | |
224 | ||
225 | /* snap blob */ | |
226 | ceph_decode_32_safe(&p, end, len, bad); | |
227 | info->snapblob_len = len; | |
228 | info->snapblob = p; | |
229 | p += len; | |
230 | ||
231 | if (p != end) | |
232 | goto bad; | |
233 | return 0; | |
234 | ||
235 | bad: | |
236 | err = -EIO; | |
237 | out_bad: | |
238 | pr_err("mds parse_reply err %d\n", err); | |
239 | return err; | |
240 | } | |
241 | ||
242 | static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info) | |
243 | { | |
244 | kfree(info->dir_in); | |
245 | } | |
246 | ||
247 | ||
248 | /* | |
249 | * sessions | |
250 | */ | |
251 | static const char *session_state_name(int s) | |
252 | { | |
253 | switch (s) { | |
254 | case CEPH_MDS_SESSION_NEW: return "new"; | |
255 | case CEPH_MDS_SESSION_OPENING: return "opening"; | |
256 | case CEPH_MDS_SESSION_OPEN: return "open"; | |
257 | case CEPH_MDS_SESSION_HUNG: return "hung"; | |
258 | case CEPH_MDS_SESSION_CLOSING: return "closing"; | |
44ca18f2 | 259 | case CEPH_MDS_SESSION_RESTARTING: return "restarting"; |
2f2dc053 SW |
260 | case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting"; |
261 | default: return "???"; | |
262 | } | |
263 | } | |
264 | ||
265 | static struct ceph_mds_session *get_session(struct ceph_mds_session *s) | |
266 | { | |
267 | if (atomic_inc_not_zero(&s->s_ref)) { | |
268 | dout("mdsc get_session %p %d -> %d\n", s, | |
269 | atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref)); | |
270 | return s; | |
271 | } else { | |
272 | dout("mdsc get_session %p 0 -- FAIL", s); | |
273 | return NULL; | |
274 | } | |
275 | } | |
276 | ||
277 | void ceph_put_mds_session(struct ceph_mds_session *s) | |
278 | { | |
279 | dout("mdsc put_session %p %d -> %d\n", s, | |
280 | atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1); | |
4e7a5dcd SW |
281 | if (atomic_dec_and_test(&s->s_ref)) { |
282 | if (s->s_authorizer) | |
283 | s->s_mdsc->client->monc.auth->ops->destroy_authorizer( | |
284 | s->s_mdsc->client->monc.auth, s->s_authorizer); | |
2f2dc053 | 285 | kfree(s); |
4e7a5dcd | 286 | } |
2f2dc053 SW |
287 | } |
288 | ||
289 | /* | |
290 | * called under mdsc->mutex | |
291 | */ | |
292 | struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc, | |
293 | int mds) | |
294 | { | |
295 | struct ceph_mds_session *session; | |
296 | ||
297 | if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL) | |
298 | return NULL; | |
299 | session = mdsc->sessions[mds]; | |
300 | dout("lookup_mds_session %p %d\n", session, | |
301 | atomic_read(&session->s_ref)); | |
302 | get_session(session); | |
303 | return session; | |
304 | } | |
305 | ||
306 | static bool __have_session(struct ceph_mds_client *mdsc, int mds) | |
307 | { | |
308 | if (mds >= mdsc->max_sessions) | |
309 | return false; | |
310 | return mdsc->sessions[mds]; | |
311 | } | |
312 | ||
2600d2dd SW |
313 | static int __verify_registered_session(struct ceph_mds_client *mdsc, |
314 | struct ceph_mds_session *s) | |
315 | { | |
316 | if (s->s_mds >= mdsc->max_sessions || | |
317 | mdsc->sessions[s->s_mds] != s) | |
318 | return -ENOENT; | |
319 | return 0; | |
320 | } | |
321 | ||
2f2dc053 SW |
322 | /* |
323 | * create+register a new session for given mds. | |
324 | * called under mdsc->mutex. | |
325 | */ | |
326 | static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc, | |
327 | int mds) | |
328 | { | |
329 | struct ceph_mds_session *s; | |
330 | ||
331 | s = kzalloc(sizeof(*s), GFP_NOFS); | |
4736b009 DC |
332 | if (!s) |
333 | return ERR_PTR(-ENOMEM); | |
2f2dc053 SW |
334 | s->s_mdsc = mdsc; |
335 | s->s_mds = mds; | |
336 | s->s_state = CEPH_MDS_SESSION_NEW; | |
337 | s->s_ttl = 0; | |
338 | s->s_seq = 0; | |
339 | mutex_init(&s->s_mutex); | |
340 | ||
341 | ceph_con_init(mdsc->client->msgr, &s->s_con); | |
342 | s->s_con.private = s; | |
343 | s->s_con.ops = &mds_con_ops; | |
344 | s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS; | |
345 | s->s_con.peer_name.num = cpu_to_le64(mds); | |
2f2dc053 SW |
346 | |
347 | spin_lock_init(&s->s_cap_lock); | |
348 | s->s_cap_gen = 0; | |
349 | s->s_cap_ttl = 0; | |
350 | s->s_renew_requested = 0; | |
351 | s->s_renew_seq = 0; | |
352 | INIT_LIST_HEAD(&s->s_caps); | |
353 | s->s_nr_caps = 0; | |
5dacf091 | 354 | s->s_trim_caps = 0; |
2f2dc053 SW |
355 | atomic_set(&s->s_ref, 1); |
356 | INIT_LIST_HEAD(&s->s_waiting); | |
357 | INIT_LIST_HEAD(&s->s_unsafe); | |
358 | s->s_num_cap_releases = 0; | |
7c1332b8 | 359 | s->s_cap_iterator = NULL; |
2f2dc053 SW |
360 | INIT_LIST_HEAD(&s->s_cap_releases); |
361 | INIT_LIST_HEAD(&s->s_cap_releases_done); | |
362 | INIT_LIST_HEAD(&s->s_cap_flushing); | |
363 | INIT_LIST_HEAD(&s->s_cap_snaps_flushing); | |
364 | ||
365 | dout("register_session mds%d\n", mds); | |
366 | if (mds >= mdsc->max_sessions) { | |
367 | int newmax = 1 << get_count_order(mds+1); | |
368 | struct ceph_mds_session **sa; | |
369 | ||
370 | dout("register_session realloc to %d\n", newmax); | |
371 | sa = kcalloc(newmax, sizeof(void *), GFP_NOFS); | |
372 | if (sa == NULL) | |
42ce56e5 | 373 | goto fail_realloc; |
2f2dc053 SW |
374 | if (mdsc->sessions) { |
375 | memcpy(sa, mdsc->sessions, | |
376 | mdsc->max_sessions * sizeof(void *)); | |
377 | kfree(mdsc->sessions); | |
378 | } | |
379 | mdsc->sessions = sa; | |
380 | mdsc->max_sessions = newmax; | |
381 | } | |
382 | mdsc->sessions[mds] = s; | |
383 | atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */ | |
42ce56e5 SW |
384 | |
385 | ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds)); | |
386 | ||
2f2dc053 | 387 | return s; |
42ce56e5 SW |
388 | |
389 | fail_realloc: | |
390 | kfree(s); | |
391 | return ERR_PTR(-ENOMEM); | |
2f2dc053 SW |
392 | } |
393 | ||
394 | /* | |
395 | * called under mdsc->mutex | |
396 | */ | |
2600d2dd | 397 | static void __unregister_session(struct ceph_mds_client *mdsc, |
42ce56e5 | 398 | struct ceph_mds_session *s) |
2f2dc053 | 399 | { |
2600d2dd SW |
400 | dout("__unregister_session mds%d %p\n", s->s_mds, s); |
401 | BUG_ON(mdsc->sessions[s->s_mds] != s); | |
42ce56e5 SW |
402 | mdsc->sessions[s->s_mds] = NULL; |
403 | ceph_con_close(&s->s_con); | |
404 | ceph_put_mds_session(s); | |
2f2dc053 SW |
405 | } |
406 | ||
407 | /* | |
408 | * drop session refs in request. | |
409 | * | |
410 | * should be last request ref, or hold mdsc->mutex | |
411 | */ | |
412 | static void put_request_session(struct ceph_mds_request *req) | |
413 | { | |
414 | if (req->r_session) { | |
415 | ceph_put_mds_session(req->r_session); | |
416 | req->r_session = NULL; | |
417 | } | |
418 | } | |
419 | ||
153c8e6b | 420 | void ceph_mdsc_release_request(struct kref *kref) |
2f2dc053 | 421 | { |
153c8e6b SW |
422 | struct ceph_mds_request *req = container_of(kref, |
423 | struct ceph_mds_request, | |
424 | r_kref); | |
425 | if (req->r_request) | |
426 | ceph_msg_put(req->r_request); | |
427 | if (req->r_reply) { | |
428 | ceph_msg_put(req->r_reply); | |
429 | destroy_reply_info(&req->r_reply_info); | |
430 | } | |
431 | if (req->r_inode) { | |
432 | ceph_put_cap_refs(ceph_inode(req->r_inode), | |
433 | CEPH_CAP_PIN); | |
434 | iput(req->r_inode); | |
435 | } | |
436 | if (req->r_locked_dir) | |
437 | ceph_put_cap_refs(ceph_inode(req->r_locked_dir), | |
438 | CEPH_CAP_PIN); | |
439 | if (req->r_target_inode) | |
440 | iput(req->r_target_inode); | |
441 | if (req->r_dentry) | |
442 | dput(req->r_dentry); | |
443 | if (req->r_old_dentry) { | |
444 | ceph_put_cap_refs( | |
445 | ceph_inode(req->r_old_dentry->d_parent->d_inode), | |
446 | CEPH_CAP_PIN); | |
447 | dput(req->r_old_dentry); | |
2f2dc053 | 448 | } |
153c8e6b SW |
449 | kfree(req->r_path1); |
450 | kfree(req->r_path2); | |
451 | put_request_session(req); | |
452 | ceph_unreserve_caps(&req->r_caps_reservation); | |
453 | kfree(req); | |
2f2dc053 SW |
454 | } |
455 | ||
456 | /* | |
457 | * lookup session, bump ref if found. | |
458 | * | |
459 | * called under mdsc->mutex. | |
460 | */ | |
461 | static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc, | |
462 | u64 tid) | |
463 | { | |
464 | struct ceph_mds_request *req; | |
44ca18f2 SW |
465 | struct rb_node *n = mdsc->request_tree.rb_node; |
466 | ||
467 | while (n) { | |
468 | req = rb_entry(n, struct ceph_mds_request, r_node); | |
469 | if (tid < req->r_tid) | |
470 | n = n->rb_left; | |
471 | else if (tid > req->r_tid) | |
472 | n = n->rb_right; | |
473 | else { | |
474 | ceph_mdsc_get_request(req); | |
475 | return req; | |
476 | } | |
477 | } | |
478 | return NULL; | |
479 | } | |
480 | ||
481 | static void __insert_request(struct ceph_mds_client *mdsc, | |
482 | struct ceph_mds_request *new) | |
483 | { | |
484 | struct rb_node **p = &mdsc->request_tree.rb_node; | |
485 | struct rb_node *parent = NULL; | |
486 | struct ceph_mds_request *req = NULL; | |
487 | ||
488 | while (*p) { | |
489 | parent = *p; | |
490 | req = rb_entry(parent, struct ceph_mds_request, r_node); | |
491 | if (new->r_tid < req->r_tid) | |
492 | p = &(*p)->rb_left; | |
493 | else if (new->r_tid > req->r_tid) | |
494 | p = &(*p)->rb_right; | |
495 | else | |
496 | BUG(); | |
497 | } | |
498 | ||
499 | rb_link_node(&new->r_node, parent, p); | |
500 | rb_insert_color(&new->r_node, &mdsc->request_tree); | |
2f2dc053 SW |
501 | } |
502 | ||
503 | /* | |
504 | * Register an in-flight request, and assign a tid. Link to directory | |
505 | * are modifying (if any). | |
506 | * | |
507 | * Called under mdsc->mutex. | |
508 | */ | |
509 | static void __register_request(struct ceph_mds_client *mdsc, | |
510 | struct ceph_mds_request *req, | |
511 | struct inode *dir) | |
512 | { | |
513 | req->r_tid = ++mdsc->last_tid; | |
514 | if (req->r_num_caps) | |
515 | ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps); | |
516 | dout("__register_request %p tid %lld\n", req, req->r_tid); | |
517 | ceph_mdsc_get_request(req); | |
44ca18f2 | 518 | __insert_request(mdsc, req); |
2f2dc053 SW |
519 | |
520 | if (dir) { | |
521 | struct ceph_inode_info *ci = ceph_inode(dir); | |
522 | ||
523 | spin_lock(&ci->i_unsafe_lock); | |
524 | req->r_unsafe_dir = dir; | |
525 | list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops); | |
526 | spin_unlock(&ci->i_unsafe_lock); | |
527 | } | |
528 | } | |
529 | ||
530 | static void __unregister_request(struct ceph_mds_client *mdsc, | |
531 | struct ceph_mds_request *req) | |
532 | { | |
533 | dout("__unregister_request %p tid %lld\n", req, req->r_tid); | |
44ca18f2 | 534 | rb_erase(&req->r_node, &mdsc->request_tree); |
80fc7314 | 535 | RB_CLEAR_NODE(&req->r_node); |
2f2dc053 SW |
536 | |
537 | if (req->r_unsafe_dir) { | |
538 | struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir); | |
539 | ||
540 | spin_lock(&ci->i_unsafe_lock); | |
541 | list_del_init(&req->r_unsafe_dir_item); | |
542 | spin_unlock(&ci->i_unsafe_lock); | |
543 | } | |
94aa8ae1 SW |
544 | |
545 | ceph_mdsc_put_request(req); | |
2f2dc053 SW |
546 | } |
547 | ||
548 | /* | |
549 | * Choose mds to send request to next. If there is a hint set in the | |
550 | * request (e.g., due to a prior forward hint from the mds), use that. | |
551 | * Otherwise, consult frag tree and/or caps to identify the | |
552 | * appropriate mds. If all else fails, choose randomly. | |
553 | * | |
554 | * Called under mdsc->mutex. | |
555 | */ | |
556 | static int __choose_mds(struct ceph_mds_client *mdsc, | |
557 | struct ceph_mds_request *req) | |
558 | { | |
559 | struct inode *inode; | |
560 | struct ceph_inode_info *ci; | |
561 | struct ceph_cap *cap; | |
562 | int mode = req->r_direct_mode; | |
563 | int mds = -1; | |
564 | u32 hash = req->r_direct_hash; | |
565 | bool is_hash = req->r_direct_is_hash; | |
566 | ||
567 | /* | |
568 | * is there a specific mds we should try? ignore hint if we have | |
569 | * no session and the mds is not up (active or recovering). | |
570 | */ | |
571 | if (req->r_resend_mds >= 0 && | |
572 | (__have_session(mdsc, req->r_resend_mds) || | |
573 | ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) { | |
574 | dout("choose_mds using resend_mds mds%d\n", | |
575 | req->r_resend_mds); | |
576 | return req->r_resend_mds; | |
577 | } | |
578 | ||
579 | if (mode == USE_RANDOM_MDS) | |
580 | goto random; | |
581 | ||
582 | inode = NULL; | |
583 | if (req->r_inode) { | |
584 | inode = req->r_inode; | |
585 | } else if (req->r_dentry) { | |
586 | if (req->r_dentry->d_inode) { | |
587 | inode = req->r_dentry->d_inode; | |
588 | } else { | |
589 | inode = req->r_dentry->d_parent->d_inode; | |
590 | hash = req->r_dentry->d_name.hash; | |
591 | is_hash = true; | |
592 | } | |
593 | } | |
594 | dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash, | |
595 | (int)hash, mode); | |
596 | if (!inode) | |
597 | goto random; | |
598 | ci = ceph_inode(inode); | |
599 | ||
600 | if (is_hash && S_ISDIR(inode->i_mode)) { | |
601 | struct ceph_inode_frag frag; | |
602 | int found; | |
603 | ||
604 | ceph_choose_frag(ci, hash, &frag, &found); | |
605 | if (found) { | |
606 | if (mode == USE_ANY_MDS && frag.ndist > 0) { | |
607 | u8 r; | |
608 | ||
609 | /* choose a random replica */ | |
610 | get_random_bytes(&r, 1); | |
611 | r %= frag.ndist; | |
612 | mds = frag.dist[r]; | |
613 | dout("choose_mds %p %llx.%llx " | |
614 | "frag %u mds%d (%d/%d)\n", | |
615 | inode, ceph_vinop(inode), | |
616 | frag.frag, frag.mds, | |
617 | (int)r, frag.ndist); | |
618 | return mds; | |
619 | } | |
620 | ||
621 | /* since this file/dir wasn't known to be | |
622 | * replicated, then we want to look for the | |
623 | * authoritative mds. */ | |
624 | mode = USE_AUTH_MDS; | |
625 | if (frag.mds >= 0) { | |
626 | /* choose auth mds */ | |
627 | mds = frag.mds; | |
628 | dout("choose_mds %p %llx.%llx " | |
629 | "frag %u mds%d (auth)\n", | |
630 | inode, ceph_vinop(inode), frag.frag, mds); | |
631 | return mds; | |
632 | } | |
633 | } | |
634 | } | |
635 | ||
636 | spin_lock(&inode->i_lock); | |
637 | cap = NULL; | |
638 | if (mode == USE_AUTH_MDS) | |
639 | cap = ci->i_auth_cap; | |
640 | if (!cap && !RB_EMPTY_ROOT(&ci->i_caps)) | |
641 | cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node); | |
642 | if (!cap) { | |
643 | spin_unlock(&inode->i_lock); | |
644 | goto random; | |
645 | } | |
646 | mds = cap->session->s_mds; | |
647 | dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n", | |
648 | inode, ceph_vinop(inode), mds, | |
649 | cap == ci->i_auth_cap ? "auth " : "", cap); | |
650 | spin_unlock(&inode->i_lock); | |
651 | return mds; | |
652 | ||
653 | random: | |
654 | mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap); | |
655 | dout("choose_mds chose random mds%d\n", mds); | |
656 | return mds; | |
657 | } | |
658 | ||
659 | ||
660 | /* | |
661 | * session messages | |
662 | */ | |
663 | static struct ceph_msg *create_session_msg(u32 op, u64 seq) | |
664 | { | |
665 | struct ceph_msg *msg; | |
666 | struct ceph_mds_session_head *h; | |
667 | ||
bb257664 | 668 | msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h)); |
a79832f2 | 669 | if (!msg) { |
2f2dc053 | 670 | pr_err("create_session_msg ENOMEM creating msg\n"); |
a79832f2 | 671 | return NULL; |
2f2dc053 SW |
672 | } |
673 | h = msg->front.iov_base; | |
674 | h->op = cpu_to_le32(op); | |
675 | h->seq = cpu_to_le64(seq); | |
676 | return msg; | |
677 | } | |
678 | ||
679 | /* | |
680 | * send session open request. | |
681 | * | |
682 | * called under mdsc->mutex | |
683 | */ | |
684 | static int __open_session(struct ceph_mds_client *mdsc, | |
685 | struct ceph_mds_session *session) | |
686 | { | |
687 | struct ceph_msg *msg; | |
688 | int mstate; | |
689 | int mds = session->s_mds; | |
2f2dc053 SW |
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); | |
a79832f2 SW |
700 | if (!msg) |
701 | return -ENOMEM; | |
2f2dc053 | 702 | ceph_con_send(&session->s_con, msg); |
2f2dc053 SW |
703 | return 0; |
704 | } | |
705 | ||
706 | /* | |
707 | * session caps | |
708 | */ | |
709 | ||
710 | /* | |
711 | * Free preallocated cap messages assigned to this session | |
712 | */ | |
713 | static void cleanup_cap_releases(struct ceph_mds_session *session) | |
714 | { | |
715 | struct ceph_msg *msg; | |
716 | ||
717 | spin_lock(&session->s_cap_lock); | |
718 | while (!list_empty(&session->s_cap_releases)) { | |
719 | msg = list_first_entry(&session->s_cap_releases, | |
720 | struct ceph_msg, list_head); | |
721 | list_del_init(&msg->list_head); | |
722 | ceph_msg_put(msg); | |
723 | } | |
724 | while (!list_empty(&session->s_cap_releases_done)) { | |
725 | msg = list_first_entry(&session->s_cap_releases_done, | |
726 | struct ceph_msg, list_head); | |
727 | list_del_init(&msg->list_head); | |
728 | ceph_msg_put(msg); | |
729 | } | |
730 | spin_unlock(&session->s_cap_lock); | |
731 | } | |
732 | ||
733 | /* | |
f818a736 SW |
734 | * Helper to safely iterate over all caps associated with a session, with |
735 | * special care taken to handle a racing __ceph_remove_cap(). | |
2f2dc053 | 736 | * |
f818a736 | 737 | * Caller must hold session s_mutex. |
2f2dc053 SW |
738 | */ |
739 | static int iterate_session_caps(struct ceph_mds_session *session, | |
740 | int (*cb)(struct inode *, struct ceph_cap *, | |
741 | void *), void *arg) | |
742 | { | |
7c1332b8 SW |
743 | struct list_head *p; |
744 | struct ceph_cap *cap; | |
745 | struct inode *inode, *last_inode = NULL; | |
746 | struct ceph_cap *old_cap = NULL; | |
2f2dc053 SW |
747 | int ret; |
748 | ||
749 | dout("iterate_session_caps %p mds%d\n", session, session->s_mds); | |
750 | spin_lock(&session->s_cap_lock); | |
7c1332b8 SW |
751 | p = session->s_caps.next; |
752 | while (p != &session->s_caps) { | |
753 | cap = list_entry(p, struct ceph_cap, session_caps); | |
2f2dc053 | 754 | inode = igrab(&cap->ci->vfs_inode); |
7c1332b8 SW |
755 | if (!inode) { |
756 | p = p->next; | |
2f2dc053 | 757 | continue; |
7c1332b8 SW |
758 | } |
759 | session->s_cap_iterator = cap; | |
2f2dc053 | 760 | spin_unlock(&session->s_cap_lock); |
7c1332b8 SW |
761 | |
762 | if (last_inode) { | |
763 | iput(last_inode); | |
764 | last_inode = NULL; | |
765 | } | |
766 | if (old_cap) { | |
767 | ceph_put_cap(old_cap); | |
768 | old_cap = NULL; | |
769 | } | |
770 | ||
2f2dc053 | 771 | ret = cb(inode, cap, arg); |
7c1332b8 SW |
772 | last_inode = inode; |
773 | ||
2f2dc053 | 774 | spin_lock(&session->s_cap_lock); |
7c1332b8 SW |
775 | p = p->next; |
776 | if (cap->ci == NULL) { | |
777 | dout("iterate_session_caps finishing cap %p removal\n", | |
778 | cap); | |
779 | BUG_ON(cap->session != session); | |
780 | list_del_init(&cap->session_caps); | |
781 | session->s_nr_caps--; | |
782 | cap->session = NULL; | |
783 | old_cap = cap; /* put_cap it w/o locks held */ | |
784 | } | |
5dacf091 SW |
785 | if (ret < 0) |
786 | goto out; | |
2f2dc053 | 787 | } |
5dacf091 SW |
788 | ret = 0; |
789 | out: | |
7c1332b8 | 790 | session->s_cap_iterator = NULL; |
2f2dc053 | 791 | spin_unlock(&session->s_cap_lock); |
7c1332b8 SW |
792 | |
793 | if (last_inode) | |
794 | iput(last_inode); | |
795 | if (old_cap) | |
796 | ceph_put_cap(old_cap); | |
797 | ||
5dacf091 | 798 | return ret; |
2f2dc053 SW |
799 | } |
800 | ||
801 | static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap, | |
6c99f254 | 802 | void *arg) |
2f2dc053 SW |
803 | { |
804 | struct ceph_inode_info *ci = ceph_inode(inode); | |
6c99f254 SW |
805 | int drop = 0; |
806 | ||
2f2dc053 SW |
807 | dout("removing cap %p, ci is %p, inode is %p\n", |
808 | cap, ci, &ci->vfs_inode); | |
6c99f254 SW |
809 | spin_lock(&inode->i_lock); |
810 | __ceph_remove_cap(cap); | |
811 | if (!__ceph_is_any_real_caps(ci)) { | |
812 | struct ceph_mds_client *mdsc = | |
813 | &ceph_sb_to_client(inode->i_sb)->mdsc; | |
814 | ||
815 | spin_lock(&mdsc->cap_dirty_lock); | |
816 | if (!list_empty(&ci->i_dirty_item)) { | |
817 | pr_info(" dropping dirty %s state for %p %lld\n", | |
818 | ceph_cap_string(ci->i_dirty_caps), | |
819 | inode, ceph_ino(inode)); | |
820 | ci->i_dirty_caps = 0; | |
821 | list_del_init(&ci->i_dirty_item); | |
822 | drop = 1; | |
823 | } | |
824 | if (!list_empty(&ci->i_flushing_item)) { | |
825 | pr_info(" dropping dirty+flushing %s state for %p %lld\n", | |
826 | ceph_cap_string(ci->i_flushing_caps), | |
827 | inode, ceph_ino(inode)); | |
828 | ci->i_flushing_caps = 0; | |
829 | list_del_init(&ci->i_flushing_item); | |
830 | mdsc->num_cap_flushing--; | |
831 | drop = 1; | |
832 | } | |
833 | if (drop && ci->i_wrbuffer_ref) { | |
834 | pr_info(" dropping dirty data for %p %lld\n", | |
835 | inode, ceph_ino(inode)); | |
836 | ci->i_wrbuffer_ref = 0; | |
837 | ci->i_wrbuffer_ref_head = 0; | |
838 | drop++; | |
839 | } | |
840 | spin_unlock(&mdsc->cap_dirty_lock); | |
841 | } | |
842 | spin_unlock(&inode->i_lock); | |
843 | while (drop--) | |
844 | iput(inode); | |
2f2dc053 SW |
845 | return 0; |
846 | } | |
847 | ||
848 | /* | |
849 | * caller must hold session s_mutex | |
850 | */ | |
851 | static void remove_session_caps(struct ceph_mds_session *session) | |
852 | { | |
853 | dout("remove_session_caps on %p\n", session); | |
854 | iterate_session_caps(session, remove_session_caps_cb, NULL); | |
855 | BUG_ON(session->s_nr_caps > 0); | |
6c99f254 | 856 | BUG_ON(!list_empty(&session->s_cap_flushing)); |
2f2dc053 SW |
857 | cleanup_cap_releases(session); |
858 | } | |
859 | ||
860 | /* | |
861 | * wake up any threads waiting on this session's caps. if the cap is | |
862 | * old (didn't get renewed on the client reconnect), remove it now. | |
863 | * | |
864 | * caller must hold s_mutex. | |
865 | */ | |
866 | static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap, | |
867 | void *arg) | |
868 | { | |
0dc2570f SW |
869 | struct ceph_inode_info *ci = ceph_inode(inode); |
870 | ||
871 | wake_up(&ci->i_cap_wq); | |
872 | if (arg) { | |
873 | spin_lock(&inode->i_lock); | |
874 | ci->i_wanted_max_size = 0; | |
875 | ci->i_requested_max_size = 0; | |
876 | spin_unlock(&inode->i_lock); | |
877 | } | |
2f2dc053 SW |
878 | return 0; |
879 | } | |
880 | ||
0dc2570f SW |
881 | static void wake_up_session_caps(struct ceph_mds_session *session, |
882 | int reconnect) | |
2f2dc053 SW |
883 | { |
884 | dout("wake_up_session_caps %p mds%d\n", session, session->s_mds); | |
0dc2570f SW |
885 | iterate_session_caps(session, wake_up_session_cb, |
886 | (void *)(unsigned long)reconnect); | |
2f2dc053 SW |
887 | } |
888 | ||
889 | /* | |
890 | * Send periodic message to MDS renewing all currently held caps. The | |
891 | * ack will reset the expiration for all caps from this session. | |
892 | * | |
893 | * caller holds s_mutex | |
894 | */ | |
895 | static int send_renew_caps(struct ceph_mds_client *mdsc, | |
896 | struct ceph_mds_session *session) | |
897 | { | |
898 | struct ceph_msg *msg; | |
899 | int state; | |
900 | ||
901 | if (time_after_eq(jiffies, session->s_cap_ttl) && | |
902 | time_after_eq(session->s_cap_ttl, session->s_renew_requested)) | |
903 | pr_info("mds%d caps stale\n", session->s_mds); | |
e4cb4cb8 | 904 | session->s_renew_requested = jiffies; |
2f2dc053 SW |
905 | |
906 | /* do not try to renew caps until a recovering mds has reconnected | |
907 | * with its clients. */ | |
908 | state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds); | |
909 | if (state < CEPH_MDS_STATE_RECONNECT) { | |
910 | dout("send_renew_caps ignoring mds%d (%s)\n", | |
911 | session->s_mds, ceph_mds_state_name(state)); | |
912 | return 0; | |
913 | } | |
914 | ||
915 | dout("send_renew_caps to mds%d (%s)\n", session->s_mds, | |
916 | ceph_mds_state_name(state)); | |
2f2dc053 SW |
917 | msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS, |
918 | ++session->s_renew_seq); | |
a79832f2 SW |
919 | if (!msg) |
920 | return -ENOMEM; | |
2f2dc053 SW |
921 | ceph_con_send(&session->s_con, msg); |
922 | return 0; | |
923 | } | |
924 | ||
925 | /* | |
926 | * Note new cap ttl, and any transition from stale -> not stale (fresh?). | |
0dc2570f SW |
927 | * |
928 | * Called under session->s_mutex | |
2f2dc053 SW |
929 | */ |
930 | static void renewed_caps(struct ceph_mds_client *mdsc, | |
931 | struct ceph_mds_session *session, int is_renew) | |
932 | { | |
933 | int was_stale; | |
934 | int wake = 0; | |
935 | ||
936 | spin_lock(&session->s_cap_lock); | |
937 | was_stale = is_renew && (session->s_cap_ttl == 0 || | |
938 | time_after_eq(jiffies, session->s_cap_ttl)); | |
939 | ||
940 | session->s_cap_ttl = session->s_renew_requested + | |
941 | mdsc->mdsmap->m_session_timeout*HZ; | |
942 | ||
943 | if (was_stale) { | |
944 | if (time_before(jiffies, session->s_cap_ttl)) { | |
945 | pr_info("mds%d caps renewed\n", session->s_mds); | |
946 | wake = 1; | |
947 | } else { | |
948 | pr_info("mds%d caps still stale\n", session->s_mds); | |
949 | } | |
950 | } | |
951 | dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n", | |
952 | session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh", | |
953 | time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh"); | |
954 | spin_unlock(&session->s_cap_lock); | |
955 | ||
956 | if (wake) | |
0dc2570f | 957 | wake_up_session_caps(session, 0); |
2f2dc053 SW |
958 | } |
959 | ||
960 | /* | |
961 | * send a session close request | |
962 | */ | |
963 | static int request_close_session(struct ceph_mds_client *mdsc, | |
964 | struct ceph_mds_session *session) | |
965 | { | |
966 | struct ceph_msg *msg; | |
2f2dc053 SW |
967 | |
968 | dout("request_close_session mds%d state %s seq %lld\n", | |
969 | session->s_mds, session_state_name(session->s_state), | |
970 | session->s_seq); | |
971 | msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq); | |
a79832f2 SW |
972 | if (!msg) |
973 | return -ENOMEM; | |
974 | ceph_con_send(&session->s_con, msg); | |
975 | return 0; | |
2f2dc053 SW |
976 | } |
977 | ||
978 | /* | |
979 | * Called with s_mutex held. | |
980 | */ | |
981 | static int __close_session(struct ceph_mds_client *mdsc, | |
982 | struct ceph_mds_session *session) | |
983 | { | |
984 | if (session->s_state >= CEPH_MDS_SESSION_CLOSING) | |
985 | return 0; | |
986 | session->s_state = CEPH_MDS_SESSION_CLOSING; | |
987 | return request_close_session(mdsc, session); | |
988 | } | |
989 | ||
990 | /* | |
991 | * Trim old(er) caps. | |
992 | * | |
993 | * Because we can't cache an inode without one or more caps, we do | |
994 | * this indirectly: if a cap is unused, we prune its aliases, at which | |
995 | * point the inode will hopefully get dropped to. | |
996 | * | |
997 | * Yes, this is a bit sloppy. Our only real goal here is to respond to | |
998 | * memory pressure from the MDS, though, so it needn't be perfect. | |
999 | */ | |
1000 | static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg) | |
1001 | { | |
1002 | struct ceph_mds_session *session = arg; | |
1003 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1004 | int used, oissued, mine; | |
1005 | ||
1006 | if (session->s_trim_caps <= 0) | |
1007 | return -1; | |
1008 | ||
1009 | spin_lock(&inode->i_lock); | |
1010 | mine = cap->issued | cap->implemented; | |
1011 | used = __ceph_caps_used(ci); | |
1012 | oissued = __ceph_caps_issued_other(ci, cap); | |
1013 | ||
1014 | dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n", | |
1015 | inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued), | |
1016 | ceph_cap_string(used)); | |
1017 | if (ci->i_dirty_caps) | |
1018 | goto out; /* dirty caps */ | |
1019 | if ((used & ~oissued) & mine) | |
1020 | goto out; /* we need these caps */ | |
1021 | ||
1022 | session->s_trim_caps--; | |
1023 | if (oissued) { | |
1024 | /* we aren't the only cap.. just remove us */ | |
7c1332b8 | 1025 | __ceph_remove_cap(cap); |
2f2dc053 SW |
1026 | } else { |
1027 | /* try to drop referring dentries */ | |
1028 | spin_unlock(&inode->i_lock); | |
1029 | d_prune_aliases(inode); | |
1030 | dout("trim_caps_cb %p cap %p pruned, count now %d\n", | |
1031 | inode, cap, atomic_read(&inode->i_count)); | |
1032 | return 0; | |
1033 | } | |
1034 | ||
1035 | out: | |
1036 | spin_unlock(&inode->i_lock); | |
1037 | return 0; | |
1038 | } | |
1039 | ||
1040 | /* | |
1041 | * Trim session cap count down to some max number. | |
1042 | */ | |
1043 | static int trim_caps(struct ceph_mds_client *mdsc, | |
1044 | struct ceph_mds_session *session, | |
1045 | int max_caps) | |
1046 | { | |
1047 | int trim_caps = session->s_nr_caps - max_caps; | |
1048 | ||
1049 | dout("trim_caps mds%d start: %d / %d, trim %d\n", | |
1050 | session->s_mds, session->s_nr_caps, max_caps, trim_caps); | |
1051 | if (trim_caps > 0) { | |
1052 | session->s_trim_caps = trim_caps; | |
1053 | iterate_session_caps(session, trim_caps_cb, session); | |
1054 | dout("trim_caps mds%d done: %d / %d, trimmed %d\n", | |
1055 | session->s_mds, session->s_nr_caps, max_caps, | |
1056 | trim_caps - session->s_trim_caps); | |
5dacf091 | 1057 | session->s_trim_caps = 0; |
2f2dc053 SW |
1058 | } |
1059 | return 0; | |
1060 | } | |
1061 | ||
1062 | /* | |
1063 | * Allocate cap_release messages. If there is a partially full message | |
1064 | * in the queue, try to allocate enough to cover it's remainder, so that | |
1065 | * we can send it immediately. | |
1066 | * | |
1067 | * Called under s_mutex. | |
1068 | */ | |
1069 | static int add_cap_releases(struct ceph_mds_client *mdsc, | |
1070 | struct ceph_mds_session *session, | |
1071 | int extra) | |
1072 | { | |
1073 | struct ceph_msg *msg; | |
1074 | struct ceph_mds_cap_release *head; | |
1075 | int err = -ENOMEM; | |
1076 | ||
1077 | if (extra < 0) | |
6b805185 | 1078 | extra = mdsc->client->mount_args->cap_release_safety; |
2f2dc053 SW |
1079 | |
1080 | spin_lock(&session->s_cap_lock); | |
1081 | ||
1082 | if (!list_empty(&session->s_cap_releases)) { | |
1083 | msg = list_first_entry(&session->s_cap_releases, | |
1084 | struct ceph_msg, | |
1085 | list_head); | |
1086 | head = msg->front.iov_base; | |
1087 | extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num); | |
1088 | } | |
1089 | ||
1090 | while (session->s_num_cap_releases < session->s_nr_caps + extra) { | |
1091 | spin_unlock(&session->s_cap_lock); | |
bb257664 | 1092 | msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE); |
2f2dc053 SW |
1093 | if (!msg) |
1094 | goto out_unlocked; | |
1095 | dout("add_cap_releases %p msg %p now %d\n", session, msg, | |
1096 | (int)msg->front.iov_len); | |
1097 | head = msg->front.iov_base; | |
1098 | head->num = cpu_to_le32(0); | |
1099 | msg->front.iov_len = sizeof(*head); | |
1100 | spin_lock(&session->s_cap_lock); | |
1101 | list_add(&msg->list_head, &session->s_cap_releases); | |
1102 | session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE; | |
1103 | } | |
1104 | ||
1105 | if (!list_empty(&session->s_cap_releases)) { | |
1106 | msg = list_first_entry(&session->s_cap_releases, | |
1107 | struct ceph_msg, | |
1108 | list_head); | |
1109 | head = msg->front.iov_base; | |
1110 | if (head->num) { | |
1111 | dout(" queueing non-full %p (%d)\n", msg, | |
1112 | le32_to_cpu(head->num)); | |
1113 | list_move_tail(&msg->list_head, | |
1114 | &session->s_cap_releases_done); | |
1115 | session->s_num_cap_releases -= | |
1116 | CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num); | |
1117 | } | |
1118 | } | |
1119 | err = 0; | |
1120 | spin_unlock(&session->s_cap_lock); | |
1121 | out_unlocked: | |
1122 | return err; | |
1123 | } | |
1124 | ||
1125 | /* | |
1126 | * flush all dirty inode data to disk. | |
1127 | * | |
1128 | * returns true if we've flushed through want_flush_seq | |
1129 | */ | |
1130 | static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq) | |
1131 | { | |
1132 | int mds, ret = 1; | |
1133 | ||
1134 | dout("check_cap_flush want %lld\n", want_flush_seq); | |
1135 | mutex_lock(&mdsc->mutex); | |
1136 | for (mds = 0; ret && mds < mdsc->max_sessions; mds++) { | |
1137 | struct ceph_mds_session *session = mdsc->sessions[mds]; | |
1138 | ||
1139 | if (!session) | |
1140 | continue; | |
1141 | get_session(session); | |
1142 | mutex_unlock(&mdsc->mutex); | |
1143 | ||
1144 | mutex_lock(&session->s_mutex); | |
1145 | if (!list_empty(&session->s_cap_flushing)) { | |
1146 | struct ceph_inode_info *ci = | |
1147 | list_entry(session->s_cap_flushing.next, | |
1148 | struct ceph_inode_info, | |
1149 | i_flushing_item); | |
1150 | struct inode *inode = &ci->vfs_inode; | |
1151 | ||
1152 | spin_lock(&inode->i_lock); | |
1153 | if (ci->i_cap_flush_seq <= want_flush_seq) { | |
1154 | dout("check_cap_flush still flushing %p " | |
1155 | "seq %lld <= %lld to mds%d\n", inode, | |
1156 | ci->i_cap_flush_seq, want_flush_seq, | |
1157 | session->s_mds); | |
1158 | ret = 0; | |
1159 | } | |
1160 | spin_unlock(&inode->i_lock); | |
1161 | } | |
1162 | mutex_unlock(&session->s_mutex); | |
1163 | ceph_put_mds_session(session); | |
1164 | ||
1165 | if (!ret) | |
1166 | return ret; | |
1167 | mutex_lock(&mdsc->mutex); | |
1168 | } | |
1169 | ||
1170 | mutex_unlock(&mdsc->mutex); | |
1171 | dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq); | |
1172 | return ret; | |
1173 | } | |
1174 | ||
1175 | /* | |
1176 | * called under s_mutex | |
1177 | */ | |
1178 | static void send_cap_releases(struct ceph_mds_client *mdsc, | |
1179 | struct ceph_mds_session *session) | |
1180 | { | |
1181 | struct ceph_msg *msg; | |
1182 | ||
1183 | dout("send_cap_releases mds%d\n", session->s_mds); | |
0f8605f2 SW |
1184 | spin_lock(&session->s_cap_lock); |
1185 | while (!list_empty(&session->s_cap_releases_done)) { | |
2f2dc053 SW |
1186 | msg = list_first_entry(&session->s_cap_releases_done, |
1187 | struct ceph_msg, list_head); | |
1188 | list_del_init(&msg->list_head); | |
1189 | spin_unlock(&session->s_cap_lock); | |
1190 | msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); | |
1191 | dout("send_cap_releases mds%d %p\n", session->s_mds, msg); | |
1192 | ceph_con_send(&session->s_con, msg); | |
0f8605f2 | 1193 | spin_lock(&session->s_cap_lock); |
2f2dc053 SW |
1194 | } |
1195 | spin_unlock(&session->s_cap_lock); | |
1196 | } | |
1197 | ||
e01a5946 SW |
1198 | static void discard_cap_releases(struct ceph_mds_client *mdsc, |
1199 | struct ceph_mds_session *session) | |
1200 | { | |
1201 | struct ceph_msg *msg; | |
1202 | struct ceph_mds_cap_release *head; | |
1203 | unsigned num; | |
1204 | ||
1205 | dout("discard_cap_releases mds%d\n", session->s_mds); | |
1206 | spin_lock(&session->s_cap_lock); | |
1207 | ||
1208 | /* zero out the in-progress message */ | |
1209 | msg = list_first_entry(&session->s_cap_releases, | |
1210 | struct ceph_msg, list_head); | |
1211 | head = msg->front.iov_base; | |
1212 | num = le32_to_cpu(head->num); | |
1213 | dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num); | |
1214 | head->num = cpu_to_le32(0); | |
1215 | session->s_num_cap_releases += num; | |
1216 | ||
1217 | /* requeue completed messages */ | |
1218 | while (!list_empty(&session->s_cap_releases_done)) { | |
1219 | msg = list_first_entry(&session->s_cap_releases_done, | |
1220 | struct ceph_msg, list_head); | |
1221 | list_del_init(&msg->list_head); | |
1222 | ||
1223 | head = msg->front.iov_base; | |
1224 | num = le32_to_cpu(head->num); | |
1225 | dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, | |
1226 | num); | |
1227 | session->s_num_cap_releases += num; | |
1228 | head->num = cpu_to_le32(0); | |
1229 | msg->front.iov_len = sizeof(*head); | |
1230 | list_add(&msg->list_head, &session->s_cap_releases); | |
1231 | } | |
1232 | ||
1233 | spin_unlock(&session->s_cap_lock); | |
1234 | } | |
1235 | ||
2f2dc053 SW |
1236 | /* |
1237 | * requests | |
1238 | */ | |
1239 | ||
1240 | /* | |
1241 | * Create an mds request. | |
1242 | */ | |
1243 | struct ceph_mds_request * | |
1244 | ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode) | |
1245 | { | |
1246 | struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS); | |
1247 | ||
1248 | if (!req) | |
1249 | return ERR_PTR(-ENOMEM); | |
1250 | ||
b4556396 | 1251 | mutex_init(&req->r_fill_mutex); |
2f2dc053 SW |
1252 | req->r_started = jiffies; |
1253 | req->r_resend_mds = -1; | |
1254 | INIT_LIST_HEAD(&req->r_unsafe_dir_item); | |
1255 | req->r_fmode = -1; | |
153c8e6b | 1256 | kref_init(&req->r_kref); |
2f2dc053 SW |
1257 | INIT_LIST_HEAD(&req->r_wait); |
1258 | init_completion(&req->r_completion); | |
1259 | init_completion(&req->r_safe_completion); | |
1260 | INIT_LIST_HEAD(&req->r_unsafe_item); | |
1261 | ||
1262 | req->r_op = op; | |
1263 | req->r_direct_mode = mode; | |
1264 | return req; | |
1265 | } | |
1266 | ||
1267 | /* | |
44ca18f2 | 1268 | * return oldest (lowest) request, tid in request tree, 0 if none. |
2f2dc053 SW |
1269 | * |
1270 | * called under mdsc->mutex. | |
1271 | */ | |
44ca18f2 SW |
1272 | static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc) |
1273 | { | |
1274 | if (RB_EMPTY_ROOT(&mdsc->request_tree)) | |
1275 | return NULL; | |
1276 | return rb_entry(rb_first(&mdsc->request_tree), | |
1277 | struct ceph_mds_request, r_node); | |
1278 | } | |
1279 | ||
2f2dc053 SW |
1280 | static u64 __get_oldest_tid(struct ceph_mds_client *mdsc) |
1281 | { | |
44ca18f2 SW |
1282 | struct ceph_mds_request *req = __get_oldest_req(mdsc); |
1283 | ||
1284 | if (req) | |
1285 | return req->r_tid; | |
1286 | return 0; | |
2f2dc053 SW |
1287 | } |
1288 | ||
1289 | /* | |
1290 | * Build a dentry's path. Allocate on heap; caller must kfree. Based | |
1291 | * on build_path_from_dentry in fs/cifs/dir.c. | |
1292 | * | |
1293 | * If @stop_on_nosnap, generate path relative to the first non-snapped | |
1294 | * inode. | |
1295 | * | |
1296 | * Encode hidden .snap dirs as a double /, i.e. | |
1297 | * foo/.snap/bar -> foo//bar | |
1298 | */ | |
1299 | char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base, | |
1300 | int stop_on_nosnap) | |
1301 | { | |
1302 | struct dentry *temp; | |
1303 | char *path; | |
1304 | int len, pos; | |
1305 | ||
1306 | if (dentry == NULL) | |
1307 | return ERR_PTR(-EINVAL); | |
1308 | ||
1309 | retry: | |
1310 | len = 0; | |
1311 | for (temp = dentry; !IS_ROOT(temp);) { | |
1312 | struct inode *inode = temp->d_inode; | |
1313 | if (inode && ceph_snap(inode) == CEPH_SNAPDIR) | |
1314 | len++; /* slash only */ | |
1315 | else if (stop_on_nosnap && inode && | |
1316 | ceph_snap(inode) == CEPH_NOSNAP) | |
1317 | break; | |
1318 | else | |
1319 | len += 1 + temp->d_name.len; | |
1320 | temp = temp->d_parent; | |
1321 | if (temp == NULL) { | |
6c99f254 | 1322 | pr_err("build_path corrupt dentry %p\n", dentry); |
2f2dc053 SW |
1323 | return ERR_PTR(-EINVAL); |
1324 | } | |
1325 | } | |
1326 | if (len) | |
1327 | len--; /* no leading '/' */ | |
1328 | ||
1329 | path = kmalloc(len+1, GFP_NOFS); | |
1330 | if (path == NULL) | |
1331 | return ERR_PTR(-ENOMEM); | |
1332 | pos = len; | |
1333 | path[pos] = 0; /* trailing null */ | |
1334 | for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) { | |
1335 | struct inode *inode = temp->d_inode; | |
1336 | ||
1337 | if (inode && ceph_snap(inode) == CEPH_SNAPDIR) { | |
104648ad | 1338 | dout("build_path path+%d: %p SNAPDIR\n", |
2f2dc053 SW |
1339 | pos, temp); |
1340 | } else if (stop_on_nosnap && inode && | |
1341 | ceph_snap(inode) == CEPH_NOSNAP) { | |
1342 | break; | |
1343 | } else { | |
1344 | pos -= temp->d_name.len; | |
1345 | if (pos < 0) | |
1346 | break; | |
1347 | strncpy(path + pos, temp->d_name.name, | |
1348 | temp->d_name.len); | |
2f2dc053 SW |
1349 | } |
1350 | if (pos) | |
1351 | path[--pos] = '/'; | |
1352 | temp = temp->d_parent; | |
1353 | if (temp == NULL) { | |
104648ad | 1354 | pr_err("build_path corrupt dentry\n"); |
2f2dc053 SW |
1355 | kfree(path); |
1356 | return ERR_PTR(-EINVAL); | |
1357 | } | |
1358 | } | |
1359 | if (pos != 0) { | |
104648ad | 1360 | pr_err("build_path did not end path lookup where " |
2f2dc053 SW |
1361 | "expected, namelen is %d, pos is %d\n", len, pos); |
1362 | /* presumably this is only possible if racing with a | |
1363 | rename of one of the parent directories (we can not | |
1364 | lock the dentries above us to prevent this, but | |
1365 | retrying should be harmless) */ | |
1366 | kfree(path); | |
1367 | goto retry; | |
1368 | } | |
1369 | ||
1370 | *base = ceph_ino(temp->d_inode); | |
1371 | *plen = len; | |
104648ad | 1372 | dout("build_path on %p %d built %llx '%.*s'\n", |
2f2dc053 SW |
1373 | dentry, atomic_read(&dentry->d_count), *base, len, path); |
1374 | return path; | |
1375 | } | |
1376 | ||
1377 | static int build_dentry_path(struct dentry *dentry, | |
1378 | const char **ppath, int *ppathlen, u64 *pino, | |
1379 | int *pfreepath) | |
1380 | { | |
1381 | char *path; | |
1382 | ||
1383 | if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) { | |
1384 | *pino = ceph_ino(dentry->d_parent->d_inode); | |
1385 | *ppath = dentry->d_name.name; | |
1386 | *ppathlen = dentry->d_name.len; | |
1387 | return 0; | |
1388 | } | |
1389 | path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1); | |
1390 | if (IS_ERR(path)) | |
1391 | return PTR_ERR(path); | |
1392 | *ppath = path; | |
1393 | *pfreepath = 1; | |
1394 | return 0; | |
1395 | } | |
1396 | ||
1397 | static int build_inode_path(struct inode *inode, | |
1398 | const char **ppath, int *ppathlen, u64 *pino, | |
1399 | int *pfreepath) | |
1400 | { | |
1401 | struct dentry *dentry; | |
1402 | char *path; | |
1403 | ||
1404 | if (ceph_snap(inode) == CEPH_NOSNAP) { | |
1405 | *pino = ceph_ino(inode); | |
1406 | *ppathlen = 0; | |
1407 | return 0; | |
1408 | } | |
1409 | dentry = d_find_alias(inode); | |
1410 | path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1); | |
1411 | dput(dentry); | |
1412 | if (IS_ERR(path)) | |
1413 | return PTR_ERR(path); | |
1414 | *ppath = path; | |
1415 | *pfreepath = 1; | |
1416 | return 0; | |
1417 | } | |
1418 | ||
1419 | /* | |
1420 | * request arguments may be specified via an inode *, a dentry *, or | |
1421 | * an explicit ino+path. | |
1422 | */ | |
1423 | static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry, | |
1424 | const char *rpath, u64 rino, | |
1425 | const char **ppath, int *pathlen, | |
1426 | u64 *ino, int *freepath) | |
1427 | { | |
1428 | int r = 0; | |
1429 | ||
1430 | if (rinode) { | |
1431 | r = build_inode_path(rinode, ppath, pathlen, ino, freepath); | |
1432 | dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode), | |
1433 | ceph_snap(rinode)); | |
1434 | } else if (rdentry) { | |
1435 | r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath); | |
1436 | dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen, | |
1437 | *ppath); | |
1438 | } else if (rpath) { | |
1439 | *ino = rino; | |
1440 | *ppath = rpath; | |
1441 | *pathlen = strlen(rpath); | |
1442 | dout(" path %.*s\n", *pathlen, rpath); | |
1443 | } | |
1444 | ||
1445 | return r; | |
1446 | } | |
1447 | ||
1448 | /* | |
1449 | * called under mdsc->mutex | |
1450 | */ | |
1451 | static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc, | |
1452 | struct ceph_mds_request *req, | |
1453 | int mds) | |
1454 | { | |
1455 | struct ceph_msg *msg; | |
1456 | struct ceph_mds_request_head *head; | |
1457 | const char *path1 = NULL; | |
1458 | const char *path2 = NULL; | |
1459 | u64 ino1 = 0, ino2 = 0; | |
1460 | int pathlen1 = 0, pathlen2 = 0; | |
1461 | int freepath1 = 0, freepath2 = 0; | |
1462 | int len; | |
1463 | u16 releases; | |
1464 | void *p, *end; | |
1465 | int ret; | |
1466 | ||
1467 | ret = set_request_path_attr(req->r_inode, req->r_dentry, | |
1468 | req->r_path1, req->r_ino1.ino, | |
1469 | &path1, &pathlen1, &ino1, &freepath1); | |
1470 | if (ret < 0) { | |
1471 | msg = ERR_PTR(ret); | |
1472 | goto out; | |
1473 | } | |
1474 | ||
1475 | ret = set_request_path_attr(NULL, req->r_old_dentry, | |
1476 | req->r_path2, req->r_ino2.ino, | |
1477 | &path2, &pathlen2, &ino2, &freepath2); | |
1478 | if (ret < 0) { | |
1479 | msg = ERR_PTR(ret); | |
1480 | goto out_free1; | |
1481 | } | |
1482 | ||
1483 | len = sizeof(*head) + | |
ac8839d7 | 1484 | pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)); |
2f2dc053 SW |
1485 | |
1486 | /* calculate (max) length for cap releases */ | |
1487 | len += sizeof(struct ceph_mds_request_release) * | |
1488 | (!!req->r_inode_drop + !!req->r_dentry_drop + | |
1489 | !!req->r_old_inode_drop + !!req->r_old_dentry_drop); | |
1490 | if (req->r_dentry_drop) | |
1491 | len += req->r_dentry->d_name.len; | |
1492 | if (req->r_old_dentry_drop) | |
1493 | len += req->r_old_dentry->d_name.len; | |
1494 | ||
bb257664 | 1495 | msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len); |
a79832f2 SW |
1496 | if (!msg) { |
1497 | msg = ERR_PTR(-ENOMEM); | |
2f2dc053 | 1498 | goto out_free2; |
a79832f2 | 1499 | } |
2f2dc053 | 1500 | |
6df058c0 SW |
1501 | msg->hdr.tid = cpu_to_le64(req->r_tid); |
1502 | ||
2f2dc053 SW |
1503 | head = msg->front.iov_base; |
1504 | p = msg->front.iov_base + sizeof(*head); | |
1505 | end = msg->front.iov_base + msg->front.iov_len; | |
1506 | ||
1507 | head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch); | |
1508 | head->op = cpu_to_le32(req->r_op); | |
1509 | head->caller_uid = cpu_to_le32(current_fsuid()); | |
1510 | head->caller_gid = cpu_to_le32(current_fsgid()); | |
1511 | head->args = req->r_args; | |
1512 | ||
1513 | ceph_encode_filepath(&p, end, ino1, path1); | |
1514 | ceph_encode_filepath(&p, end, ino2, path2); | |
1515 | ||
1516 | /* cap releases */ | |
1517 | releases = 0; | |
1518 | if (req->r_inode_drop) | |
1519 | releases += ceph_encode_inode_release(&p, | |
1520 | req->r_inode ? req->r_inode : req->r_dentry->d_inode, | |
1521 | mds, req->r_inode_drop, req->r_inode_unless, 0); | |
1522 | if (req->r_dentry_drop) | |
1523 | releases += ceph_encode_dentry_release(&p, req->r_dentry, | |
1524 | mds, req->r_dentry_drop, req->r_dentry_unless); | |
1525 | if (req->r_old_dentry_drop) | |
1526 | releases += ceph_encode_dentry_release(&p, req->r_old_dentry, | |
1527 | mds, req->r_old_dentry_drop, req->r_old_dentry_unless); | |
1528 | if (req->r_old_inode_drop) | |
1529 | releases += ceph_encode_inode_release(&p, | |
1530 | req->r_old_dentry->d_inode, | |
1531 | mds, req->r_old_inode_drop, req->r_old_inode_unless, 0); | |
1532 | head->num_releases = cpu_to_le16(releases); | |
1533 | ||
1534 | BUG_ON(p > end); | |
1535 | msg->front.iov_len = p - msg->front.iov_base; | |
1536 | msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); | |
1537 | ||
1538 | msg->pages = req->r_pages; | |
1539 | msg->nr_pages = req->r_num_pages; | |
1540 | msg->hdr.data_len = cpu_to_le32(req->r_data_len); | |
1541 | msg->hdr.data_off = cpu_to_le16(0); | |
1542 | ||
1543 | out_free2: | |
1544 | if (freepath2) | |
1545 | kfree((char *)path2); | |
1546 | out_free1: | |
1547 | if (freepath1) | |
1548 | kfree((char *)path1); | |
1549 | out: | |
1550 | return msg; | |
1551 | } | |
1552 | ||
1553 | /* | |
1554 | * called under mdsc->mutex if error, under no mutex if | |
1555 | * success. | |
1556 | */ | |
1557 | static void complete_request(struct ceph_mds_client *mdsc, | |
1558 | struct ceph_mds_request *req) | |
1559 | { | |
1560 | if (req->r_callback) | |
1561 | req->r_callback(mdsc, req); | |
1562 | else | |
1563 | complete(&req->r_completion); | |
1564 | } | |
1565 | ||
1566 | /* | |
1567 | * called under mdsc->mutex | |
1568 | */ | |
1569 | static int __prepare_send_request(struct ceph_mds_client *mdsc, | |
1570 | struct ceph_mds_request *req, | |
1571 | int mds) | |
1572 | { | |
1573 | struct ceph_mds_request_head *rhead; | |
1574 | struct ceph_msg *msg; | |
1575 | int flags = 0; | |
1576 | ||
1577 | req->r_mds = mds; | |
1578 | req->r_attempts++; | |
1579 | dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req, | |
1580 | req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts); | |
1581 | ||
1582 | if (req->r_request) { | |
1583 | ceph_msg_put(req->r_request); | |
1584 | req->r_request = NULL; | |
1585 | } | |
1586 | msg = create_request_message(mdsc, req, mds); | |
1587 | if (IS_ERR(msg)) { | |
e1518c7c | 1588 | req->r_err = PTR_ERR(msg); |
2f2dc053 | 1589 | complete_request(mdsc, req); |
a79832f2 | 1590 | return PTR_ERR(msg); |
2f2dc053 SW |
1591 | } |
1592 | req->r_request = msg; | |
1593 | ||
1594 | rhead = msg->front.iov_base; | |
2f2dc053 SW |
1595 | rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc)); |
1596 | if (req->r_got_unsafe) | |
1597 | flags |= CEPH_MDS_FLAG_REPLAY; | |
1598 | if (req->r_locked_dir) | |
1599 | flags |= CEPH_MDS_FLAG_WANT_DENTRY; | |
1600 | rhead->flags = cpu_to_le32(flags); | |
1601 | rhead->num_fwd = req->r_num_fwd; | |
1602 | rhead->num_retry = req->r_attempts - 1; | |
1603 | ||
1604 | dout(" r_locked_dir = %p\n", req->r_locked_dir); | |
1605 | ||
1606 | if (req->r_target_inode && req->r_got_unsafe) | |
1607 | rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode)); | |
1608 | else | |
1609 | rhead->ino = 0; | |
1610 | return 0; | |
1611 | } | |
1612 | ||
1613 | /* | |
1614 | * send request, or put it on the appropriate wait list. | |
1615 | */ | |
1616 | static int __do_request(struct ceph_mds_client *mdsc, | |
1617 | struct ceph_mds_request *req) | |
1618 | { | |
1619 | struct ceph_mds_session *session = NULL; | |
1620 | int mds = -1; | |
1621 | int err = -EAGAIN; | |
1622 | ||
e1518c7c | 1623 | if (req->r_err || req->r_got_result) |
2f2dc053 SW |
1624 | goto out; |
1625 | ||
1626 | if (req->r_timeout && | |
1627 | time_after_eq(jiffies, req->r_started + req->r_timeout)) { | |
1628 | dout("do_request timed out\n"); | |
1629 | err = -EIO; | |
1630 | goto finish; | |
1631 | } | |
1632 | ||
1633 | mds = __choose_mds(mdsc, req); | |
1634 | if (mds < 0 || | |
1635 | ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) { | |
1636 | dout("do_request no mds or not active, waiting for map\n"); | |
1637 | list_add(&req->r_wait, &mdsc->waiting_for_map); | |
1638 | goto out; | |
1639 | } | |
1640 | ||
1641 | /* get, open session */ | |
1642 | session = __ceph_lookup_mds_session(mdsc, mds); | |
9c423956 | 1643 | if (!session) { |
2f2dc053 | 1644 | session = register_session(mdsc, mds); |
9c423956 SW |
1645 | if (IS_ERR(session)) { |
1646 | err = PTR_ERR(session); | |
1647 | goto finish; | |
1648 | } | |
1649 | } | |
2f2dc053 SW |
1650 | dout("do_request mds%d session %p state %s\n", mds, session, |
1651 | session_state_name(session->s_state)); | |
1652 | if (session->s_state != CEPH_MDS_SESSION_OPEN && | |
1653 | session->s_state != CEPH_MDS_SESSION_HUNG) { | |
1654 | if (session->s_state == CEPH_MDS_SESSION_NEW || | |
1655 | session->s_state == CEPH_MDS_SESSION_CLOSING) | |
1656 | __open_session(mdsc, session); | |
1657 | list_add(&req->r_wait, &session->s_waiting); | |
1658 | goto out_session; | |
1659 | } | |
1660 | ||
1661 | /* send request */ | |
1662 | req->r_session = get_session(session); | |
1663 | req->r_resend_mds = -1; /* forget any previous mds hint */ | |
1664 | ||
1665 | if (req->r_request_started == 0) /* note request start time */ | |
1666 | req->r_request_started = jiffies; | |
1667 | ||
1668 | err = __prepare_send_request(mdsc, req, mds); | |
1669 | if (!err) { | |
1670 | ceph_msg_get(req->r_request); | |
1671 | ceph_con_send(&session->s_con, req->r_request); | |
1672 | } | |
1673 | ||
1674 | out_session: | |
1675 | ceph_put_mds_session(session); | |
1676 | out: | |
1677 | return err; | |
1678 | ||
1679 | finish: | |
e1518c7c | 1680 | req->r_err = err; |
2f2dc053 SW |
1681 | complete_request(mdsc, req); |
1682 | goto out; | |
1683 | } | |
1684 | ||
1685 | /* | |
1686 | * called under mdsc->mutex | |
1687 | */ | |
1688 | static void __wake_requests(struct ceph_mds_client *mdsc, | |
1689 | struct list_head *head) | |
1690 | { | |
1691 | struct ceph_mds_request *req, *nreq; | |
1692 | ||
1693 | list_for_each_entry_safe(req, nreq, head, r_wait) { | |
1694 | list_del_init(&req->r_wait); | |
1695 | __do_request(mdsc, req); | |
1696 | } | |
1697 | } | |
1698 | ||
1699 | /* | |
1700 | * Wake up threads with requests pending for @mds, so that they can | |
29790f26 | 1701 | * resubmit their requests to a possibly different mds. |
2f2dc053 | 1702 | */ |
29790f26 | 1703 | static void kick_requests(struct ceph_mds_client *mdsc, int mds) |
2f2dc053 | 1704 | { |
44ca18f2 SW |
1705 | struct ceph_mds_request *req; |
1706 | struct rb_node *p; | |
2f2dc053 SW |
1707 | |
1708 | dout("kick_requests mds%d\n", mds); | |
44ca18f2 SW |
1709 | for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) { |
1710 | req = rb_entry(p, struct ceph_mds_request, r_node); | |
1711 | if (req->r_got_unsafe) | |
1712 | continue; | |
1713 | if (req->r_session && | |
1714 | req->r_session->s_mds == mds) { | |
1715 | dout(" kicking tid %llu\n", req->r_tid); | |
1716 | put_request_session(req); | |
1717 | __do_request(mdsc, req); | |
2f2dc053 SW |
1718 | } |
1719 | } | |
1720 | } | |
1721 | ||
1722 | void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, | |
1723 | struct ceph_mds_request *req) | |
1724 | { | |
1725 | dout("submit_request on %p\n", req); | |
1726 | mutex_lock(&mdsc->mutex); | |
1727 | __register_request(mdsc, req, NULL); | |
1728 | __do_request(mdsc, req); | |
1729 | mutex_unlock(&mdsc->mutex); | |
1730 | } | |
1731 | ||
1732 | /* | |
1733 | * Synchrously perform an mds request. Take care of all of the | |
1734 | * session setup, forwarding, retry details. | |
1735 | */ | |
1736 | int ceph_mdsc_do_request(struct ceph_mds_client *mdsc, | |
1737 | struct inode *dir, | |
1738 | struct ceph_mds_request *req) | |
1739 | { | |
1740 | int err; | |
1741 | ||
1742 | dout("do_request on %p\n", req); | |
1743 | ||
1744 | /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */ | |
1745 | if (req->r_inode) | |
1746 | ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN); | |
1747 | if (req->r_locked_dir) | |
1748 | ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN); | |
1749 | if (req->r_old_dentry) | |
1750 | ceph_get_cap_refs( | |
1751 | ceph_inode(req->r_old_dentry->d_parent->d_inode), | |
1752 | CEPH_CAP_PIN); | |
1753 | ||
1754 | /* issue */ | |
1755 | mutex_lock(&mdsc->mutex); | |
1756 | __register_request(mdsc, req, dir); | |
1757 | __do_request(mdsc, req); | |
1758 | ||
e1518c7c SW |
1759 | if (req->r_err) { |
1760 | err = req->r_err; | |
1761 | __unregister_request(mdsc, req); | |
1762 | dout("do_request early error %d\n", err); | |
1763 | goto out; | |
2f2dc053 SW |
1764 | } |
1765 | ||
e1518c7c SW |
1766 | /* wait */ |
1767 | mutex_unlock(&mdsc->mutex); | |
1768 | dout("do_request waiting\n"); | |
1769 | if (req->r_timeout) { | |
1770 | err = (long)wait_for_completion_interruptible_timeout( | |
1771 | &req->r_completion, req->r_timeout); | |
1772 | if (err == 0) | |
1773 | err = -EIO; | |
1774 | } else { | |
1775 | err = wait_for_completion_interruptible(&req->r_completion); | |
1776 | } | |
1777 | dout("do_request waited, got %d\n", err); | |
1778 | mutex_lock(&mdsc->mutex); | |
5b1daecd | 1779 | |
e1518c7c SW |
1780 | /* only abort if we didn't race with a real reply */ |
1781 | if (req->r_got_result) { | |
1782 | err = le32_to_cpu(req->r_reply_info.head->result); | |
1783 | } else if (err < 0) { | |
1784 | dout("aborted request %lld with %d\n", req->r_tid, err); | |
b4556396 SW |
1785 | |
1786 | /* | |
1787 | * ensure we aren't running concurrently with | |
1788 | * ceph_fill_trace or ceph_readdir_prepopulate, which | |
1789 | * rely on locks (dir mutex) held by our caller. | |
1790 | */ | |
1791 | mutex_lock(&req->r_fill_mutex); | |
e1518c7c SW |
1792 | req->r_err = err; |
1793 | req->r_aborted = true; | |
b4556396 | 1794 | mutex_unlock(&req->r_fill_mutex); |
5b1daecd | 1795 | |
e1518c7c SW |
1796 | if (req->r_locked_dir && |
1797 | (req->r_op & CEPH_MDS_OP_WRITE)) { | |
1798 | struct ceph_inode_info *ci = | |
1799 | ceph_inode(req->r_locked_dir); | |
1800 | ||
81a6cf2d | 1801 | dout("aborted, clearing I_COMPLETE on %p, leases\n", |
e1518c7c SW |
1802 | req->r_locked_dir); |
1803 | spin_lock(&req->r_locked_dir->i_lock); | |
1804 | ci->i_ceph_flags &= ~CEPH_I_COMPLETE; | |
1805 | ci->i_release_count++; | |
1806 | spin_unlock(&req->r_locked_dir->i_lock); | |
81a6cf2d SW |
1807 | |
1808 | if (req->r_dentry) | |
1809 | ceph_invalidate_dentry_lease(req->r_dentry); | |
1810 | if (req->r_old_dentry) | |
1811 | ceph_invalidate_dentry_lease(req->r_old_dentry); | |
5b1daecd | 1812 | } |
2f2dc053 | 1813 | } else { |
e1518c7c | 1814 | err = req->r_err; |
2f2dc053 | 1815 | } |
2f2dc053 | 1816 | |
e1518c7c SW |
1817 | out: |
1818 | mutex_unlock(&mdsc->mutex); | |
2f2dc053 SW |
1819 | dout("do_request %p done, result %d\n", req, err); |
1820 | return err; | |
1821 | } | |
1822 | ||
1823 | /* | |
1824 | * Handle mds reply. | |
1825 | * | |
1826 | * We take the session mutex and parse and process the reply immediately. | |
1827 | * This preserves the logical ordering of replies, capabilities, etc., sent | |
1828 | * by the MDS as they are applied to our local cache. | |
1829 | */ | |
1830 | static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg) | |
1831 | { | |
1832 | struct ceph_mds_client *mdsc = session->s_mdsc; | |
1833 | struct ceph_mds_request *req; | |
1834 | struct ceph_mds_reply_head *head = msg->front.iov_base; | |
1835 | struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */ | |
1836 | u64 tid; | |
1837 | int err, result; | |
2600d2dd | 1838 | int mds = session->s_mds; |
2f2dc053 | 1839 | |
2f2dc053 SW |
1840 | if (msg->front.iov_len < sizeof(*head)) { |
1841 | pr_err("mdsc_handle_reply got corrupt (short) reply\n"); | |
9ec7cab1 | 1842 | ceph_msg_dump(msg); |
2f2dc053 SW |
1843 | return; |
1844 | } | |
1845 | ||
1846 | /* get request, session */ | |
6df058c0 | 1847 | tid = le64_to_cpu(msg->hdr.tid); |
2f2dc053 SW |
1848 | mutex_lock(&mdsc->mutex); |
1849 | req = __lookup_request(mdsc, tid); | |
1850 | if (!req) { | |
1851 | dout("handle_reply on unknown tid %llu\n", tid); | |
1852 | mutex_unlock(&mdsc->mutex); | |
1853 | return; | |
1854 | } | |
1855 | dout("handle_reply %p\n", req); | |
2f2dc053 SW |
1856 | |
1857 | /* correct session? */ | |
d96d6049 | 1858 | if (req->r_session != session) { |
2f2dc053 SW |
1859 | pr_err("mdsc_handle_reply got %llu on session mds%d" |
1860 | " not mds%d\n", tid, session->s_mds, | |
1861 | req->r_session ? req->r_session->s_mds : -1); | |
1862 | mutex_unlock(&mdsc->mutex); | |
1863 | goto out; | |
1864 | } | |
1865 | ||
1866 | /* dup? */ | |
1867 | if ((req->r_got_unsafe && !head->safe) || | |
1868 | (req->r_got_safe && head->safe)) { | |
1869 | pr_warning("got a dup %s reply on %llu from mds%d\n", | |
1870 | head->safe ? "safe" : "unsafe", tid, mds); | |
1871 | mutex_unlock(&mdsc->mutex); | |
1872 | goto out; | |
1873 | } | |
85792d0d SW |
1874 | if (req->r_got_safe && !head->safe) { |
1875 | pr_warning("got unsafe after safe on %llu from mds%d\n", | |
1876 | tid, mds); | |
1877 | mutex_unlock(&mdsc->mutex); | |
1878 | goto out; | |
1879 | } | |
2f2dc053 SW |
1880 | |
1881 | result = le32_to_cpu(head->result); | |
1882 | ||
1883 | /* | |
1884 | * Tolerate 2 consecutive ESTALEs from the same mds. | |
1885 | * FIXME: we should be looking at the cap migrate_seq. | |
1886 | */ | |
1887 | if (result == -ESTALE) { | |
1888 | req->r_direct_mode = USE_AUTH_MDS; | |
1889 | req->r_num_stale++; | |
1890 | if (req->r_num_stale <= 2) { | |
1891 | __do_request(mdsc, req); | |
1892 | mutex_unlock(&mdsc->mutex); | |
1893 | goto out; | |
1894 | } | |
1895 | } else { | |
1896 | req->r_num_stale = 0; | |
1897 | } | |
1898 | ||
1899 | if (head->safe) { | |
1900 | req->r_got_safe = true; | |
1901 | __unregister_request(mdsc, req); | |
1902 | complete(&req->r_safe_completion); | |
1903 | ||
1904 | if (req->r_got_unsafe) { | |
1905 | /* | |
1906 | * We already handled the unsafe response, now do the | |
1907 | * cleanup. No need to examine the response; the MDS | |
1908 | * doesn't include any result info in the safe | |
1909 | * response. And even if it did, there is nothing | |
1910 | * useful we could do with a revised return value. | |
1911 | */ | |
1912 | dout("got safe reply %llu, mds%d\n", tid, mds); | |
1913 | list_del_init(&req->r_unsafe_item); | |
1914 | ||
1915 | /* last unsafe request during umount? */ | |
44ca18f2 | 1916 | if (mdsc->stopping && !__get_oldest_req(mdsc)) |
2f2dc053 SW |
1917 | complete(&mdsc->safe_umount_waiters); |
1918 | mutex_unlock(&mdsc->mutex); | |
1919 | goto out; | |
1920 | } | |
e1518c7c | 1921 | } else { |
2f2dc053 SW |
1922 | req->r_got_unsafe = true; |
1923 | list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe); | |
1924 | } | |
1925 | ||
1926 | dout("handle_reply tid %lld result %d\n", tid, result); | |
1927 | rinfo = &req->r_reply_info; | |
1928 | err = parse_reply_info(msg, rinfo); | |
1929 | mutex_unlock(&mdsc->mutex); | |
1930 | ||
1931 | mutex_lock(&session->s_mutex); | |
1932 | if (err < 0) { | |
1933 | pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds); | |
9ec7cab1 | 1934 | ceph_msg_dump(msg); |
2f2dc053 SW |
1935 | goto out_err; |
1936 | } | |
1937 | ||
1938 | /* snap trace */ | |
1939 | if (rinfo->snapblob_len) { | |
1940 | down_write(&mdsc->snap_rwsem); | |
1941 | ceph_update_snap_trace(mdsc, rinfo->snapblob, | |
1942 | rinfo->snapblob + rinfo->snapblob_len, | |
1943 | le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP); | |
1944 | downgrade_write(&mdsc->snap_rwsem); | |
1945 | } else { | |
1946 | down_read(&mdsc->snap_rwsem); | |
1947 | } | |
1948 | ||
1949 | /* insert trace into our cache */ | |
b4556396 | 1950 | mutex_lock(&req->r_fill_mutex); |
2f2dc053 SW |
1951 | err = ceph_fill_trace(mdsc->client->sb, req, req->r_session); |
1952 | if (err == 0) { | |
1953 | if (result == 0 && rinfo->dir_nr) | |
1954 | ceph_readdir_prepopulate(req, req->r_session); | |
1955 | ceph_unreserve_caps(&req->r_caps_reservation); | |
1956 | } | |
b4556396 | 1957 | mutex_unlock(&req->r_fill_mutex); |
2f2dc053 SW |
1958 | |
1959 | up_read(&mdsc->snap_rwsem); | |
1960 | out_err: | |
e1518c7c SW |
1961 | mutex_lock(&mdsc->mutex); |
1962 | if (!req->r_aborted) { | |
1963 | if (err) { | |
1964 | req->r_err = err; | |
1965 | } else { | |
1966 | req->r_reply = msg; | |
1967 | ceph_msg_get(msg); | |
1968 | req->r_got_result = true; | |
1969 | } | |
2f2dc053 | 1970 | } else { |
e1518c7c | 1971 | dout("reply arrived after request %lld was aborted\n", tid); |
2f2dc053 | 1972 | } |
e1518c7c | 1973 | mutex_unlock(&mdsc->mutex); |
2f2dc053 SW |
1974 | |
1975 | add_cap_releases(mdsc, req->r_session, -1); | |
1976 | mutex_unlock(&session->s_mutex); | |
1977 | ||
1978 | /* kick calling process */ | |
1979 | complete_request(mdsc, req); | |
1980 | out: | |
1981 | ceph_mdsc_put_request(req); | |
1982 | return; | |
1983 | } | |
1984 | ||
1985 | ||
1986 | ||
1987 | /* | |
1988 | * handle mds notification that our request has been forwarded. | |
1989 | */ | |
2600d2dd SW |
1990 | static void handle_forward(struct ceph_mds_client *mdsc, |
1991 | struct ceph_mds_session *session, | |
1992 | struct ceph_msg *msg) | |
2f2dc053 SW |
1993 | { |
1994 | struct ceph_mds_request *req; | |
a1ea787c | 1995 | u64 tid = le64_to_cpu(msg->hdr.tid); |
2f2dc053 SW |
1996 | u32 next_mds; |
1997 | u32 fwd_seq; | |
2f2dc053 SW |
1998 | int err = -EINVAL; |
1999 | void *p = msg->front.iov_base; | |
2000 | void *end = p + msg->front.iov_len; | |
2f2dc053 | 2001 | |
a1ea787c | 2002 | ceph_decode_need(&p, end, 2*sizeof(u32), bad); |
c89136ea SW |
2003 | next_mds = ceph_decode_32(&p); |
2004 | fwd_seq = ceph_decode_32(&p); | |
2f2dc053 SW |
2005 | |
2006 | mutex_lock(&mdsc->mutex); | |
2007 | req = __lookup_request(mdsc, tid); | |
2008 | if (!req) { | |
080af17e | 2009 | dout("forward %llu to mds%d - req dne\n", tid, next_mds); |
2f2dc053 SW |
2010 | goto out; /* dup reply? */ |
2011 | } | |
2012 | ||
2f2dc053 SW |
2013 | if (fwd_seq <= req->r_num_fwd) { |
2014 | dout("forward %llu to mds%d - old seq %d <= %d\n", | |
2015 | tid, next_mds, req->r_num_fwd, fwd_seq); | |
2016 | } else { | |
2017 | /* resend. forward race not possible; mds would drop */ | |
2018 | dout("forward %llu to mds%d (we resend)\n", tid, next_mds); | |
2019 | req->r_num_fwd = fwd_seq; | |
2020 | req->r_resend_mds = next_mds; | |
2021 | put_request_session(req); | |
2022 | __do_request(mdsc, req); | |
2023 | } | |
2024 | ceph_mdsc_put_request(req); | |
2025 | out: | |
2026 | mutex_unlock(&mdsc->mutex); | |
2027 | return; | |
2028 | ||
2029 | bad: | |
2030 | pr_err("mdsc_handle_forward decode error err=%d\n", err); | |
2031 | } | |
2032 | ||
2033 | /* | |
2034 | * handle a mds session control message | |
2035 | */ | |
2036 | static void handle_session(struct ceph_mds_session *session, | |
2037 | struct ceph_msg *msg) | |
2038 | { | |
2039 | struct ceph_mds_client *mdsc = session->s_mdsc; | |
2040 | u32 op; | |
2041 | u64 seq; | |
2600d2dd | 2042 | int mds = session->s_mds; |
2f2dc053 SW |
2043 | struct ceph_mds_session_head *h = msg->front.iov_base; |
2044 | int wake = 0; | |
2045 | ||
2f2dc053 SW |
2046 | /* decode */ |
2047 | if (msg->front.iov_len != sizeof(*h)) | |
2048 | goto bad; | |
2049 | op = le32_to_cpu(h->op); | |
2050 | seq = le64_to_cpu(h->seq); | |
2051 | ||
2052 | mutex_lock(&mdsc->mutex); | |
2600d2dd SW |
2053 | if (op == CEPH_SESSION_CLOSE) |
2054 | __unregister_session(mdsc, session); | |
2f2dc053 SW |
2055 | /* FIXME: this ttl calculation is generous */ |
2056 | session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose; | |
2057 | mutex_unlock(&mdsc->mutex); | |
2058 | ||
2059 | mutex_lock(&session->s_mutex); | |
2060 | ||
2061 | dout("handle_session mds%d %s %p state %s seq %llu\n", | |
2062 | mds, ceph_session_op_name(op), session, | |
2063 | session_state_name(session->s_state), seq); | |
2064 | ||
2065 | if (session->s_state == CEPH_MDS_SESSION_HUNG) { | |
2066 | session->s_state = CEPH_MDS_SESSION_OPEN; | |
2067 | pr_info("mds%d came back\n", session->s_mds); | |
2068 | } | |
2069 | ||
2070 | switch (op) { | |
2071 | case CEPH_SESSION_OPEN: | |
29790f26 SW |
2072 | if (session->s_state == CEPH_MDS_SESSION_RECONNECTING) |
2073 | pr_info("mds%d reconnect success\n", session->s_mds); | |
2f2dc053 SW |
2074 | session->s_state = CEPH_MDS_SESSION_OPEN; |
2075 | renewed_caps(mdsc, session, 0); | |
2076 | wake = 1; | |
2077 | if (mdsc->stopping) | |
2078 | __close_session(mdsc, session); | |
2079 | break; | |
2080 | ||
2081 | case CEPH_SESSION_RENEWCAPS: | |
2082 | if (session->s_renew_seq == seq) | |
2083 | renewed_caps(mdsc, session, 1); | |
2084 | break; | |
2085 | ||
2086 | case CEPH_SESSION_CLOSE: | |
29790f26 SW |
2087 | if (session->s_state == CEPH_MDS_SESSION_RECONNECTING) |
2088 | pr_info("mds%d reconnect denied\n", session->s_mds); | |
2f2dc053 SW |
2089 | remove_session_caps(session); |
2090 | wake = 1; /* for good measure */ | |
2091 | complete(&mdsc->session_close_waiters); | |
29790f26 | 2092 | kick_requests(mdsc, mds); |
2f2dc053 SW |
2093 | break; |
2094 | ||
2095 | case CEPH_SESSION_STALE: | |
2096 | pr_info("mds%d caps went stale, renewing\n", | |
2097 | session->s_mds); | |
2098 | spin_lock(&session->s_cap_lock); | |
2099 | session->s_cap_gen++; | |
2100 | session->s_cap_ttl = 0; | |
2101 | spin_unlock(&session->s_cap_lock); | |
2102 | send_renew_caps(mdsc, session); | |
2103 | break; | |
2104 | ||
2105 | case CEPH_SESSION_RECALL_STATE: | |
2106 | trim_caps(mdsc, session, le32_to_cpu(h->max_caps)); | |
2107 | break; | |
2108 | ||
2109 | default: | |
2110 | pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds); | |
2111 | WARN_ON(1); | |
2112 | } | |
2113 | ||
2114 | mutex_unlock(&session->s_mutex); | |
2115 | if (wake) { | |
2116 | mutex_lock(&mdsc->mutex); | |
2117 | __wake_requests(mdsc, &session->s_waiting); | |
2118 | mutex_unlock(&mdsc->mutex); | |
2119 | } | |
2120 | return; | |
2121 | ||
2122 | bad: | |
2123 | pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds, | |
2124 | (int)msg->front.iov_len); | |
9ec7cab1 | 2125 | ceph_msg_dump(msg); |
2f2dc053 SW |
2126 | return; |
2127 | } | |
2128 | ||
2129 | ||
2130 | /* | |
2131 | * called under session->mutex. | |
2132 | */ | |
2133 | static void replay_unsafe_requests(struct ceph_mds_client *mdsc, | |
2134 | struct ceph_mds_session *session) | |
2135 | { | |
2136 | struct ceph_mds_request *req, *nreq; | |
2137 | int err; | |
2138 | ||
2139 | dout("replay_unsafe_requests mds%d\n", session->s_mds); | |
2140 | ||
2141 | mutex_lock(&mdsc->mutex); | |
2142 | list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) { | |
2143 | err = __prepare_send_request(mdsc, req, session->s_mds); | |
2144 | if (!err) { | |
2145 | ceph_msg_get(req->r_request); | |
2146 | ceph_con_send(&session->s_con, req->r_request); | |
2147 | } | |
2148 | } | |
2149 | mutex_unlock(&mdsc->mutex); | |
2150 | } | |
2151 | ||
2152 | /* | |
2153 | * Encode information about a cap for a reconnect with the MDS. | |
2154 | */ | |
2f2dc053 SW |
2155 | static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap, |
2156 | void *arg) | |
2157 | { | |
93cea5be | 2158 | struct ceph_mds_cap_reconnect rec; |
2f2dc053 | 2159 | struct ceph_inode_info *ci; |
93cea5be | 2160 | struct ceph_pagelist *pagelist = arg; |
2f2dc053 SW |
2161 | char *path; |
2162 | int pathlen, err; | |
2163 | u64 pathbase; | |
2164 | struct dentry *dentry; | |
2165 | ||
2166 | ci = cap->ci; | |
2167 | ||
2168 | dout(" adding %p ino %llx.%llx cap %p %lld %s\n", | |
2169 | inode, ceph_vinop(inode), cap, cap->cap_id, | |
2170 | ceph_cap_string(cap->issued)); | |
93cea5be SW |
2171 | err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode)); |
2172 | if (err) | |
2173 | return err; | |
2f2dc053 SW |
2174 | |
2175 | dentry = d_find_alias(inode); | |
2176 | if (dentry) { | |
2177 | path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0); | |
2178 | if (IS_ERR(path)) { | |
2179 | err = PTR_ERR(path); | |
2180 | BUG_ON(err); | |
2181 | } | |
2182 | } else { | |
2183 | path = NULL; | |
2184 | pathlen = 0; | |
2185 | } | |
93cea5be SW |
2186 | err = ceph_pagelist_encode_string(pagelist, path, pathlen); |
2187 | if (err) | |
2188 | goto out; | |
2f2dc053 | 2189 | |
2f2dc053 SW |
2190 | spin_lock(&inode->i_lock); |
2191 | cap->seq = 0; /* reset cap seq */ | |
2192 | cap->issue_seq = 0; /* and issue_seq */ | |
93cea5be SW |
2193 | rec.cap_id = cpu_to_le64(cap->cap_id); |
2194 | rec.pathbase = cpu_to_le64(pathbase); | |
2195 | rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci)); | |
2196 | rec.issued = cpu_to_le32(cap->issued); | |
2197 | rec.size = cpu_to_le64(inode->i_size); | |
2198 | ceph_encode_timespec(&rec.mtime, &inode->i_mtime); | |
2199 | ceph_encode_timespec(&rec.atime, &inode->i_atime); | |
2200 | rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino); | |
2f2dc053 SW |
2201 | spin_unlock(&inode->i_lock); |
2202 | ||
93cea5be SW |
2203 | err = ceph_pagelist_append(pagelist, &rec, sizeof(rec)); |
2204 | ||
2205 | out: | |
2f2dc053 SW |
2206 | kfree(path); |
2207 | dput(dentry); | |
93cea5be | 2208 | return err; |
2f2dc053 SW |
2209 | } |
2210 | ||
2211 | ||
2212 | /* | |
2213 | * If an MDS fails and recovers, clients need to reconnect in order to | |
2214 | * reestablish shared state. This includes all caps issued through | |
2215 | * this session _and_ the snap_realm hierarchy. Because it's not | |
2216 | * clear which snap realms the mds cares about, we send everything we | |
2217 | * know about.. that ensures we'll then get any new info the | |
2218 | * recovering MDS might have. | |
2219 | * | |
2220 | * This is a relatively heavyweight operation, but it's rare. | |
2221 | * | |
2222 | * called with mdsc->mutex held. | |
2223 | */ | |
34b6c855 SW |
2224 | static void send_mds_reconnect(struct ceph_mds_client *mdsc, |
2225 | struct ceph_mds_session *session) | |
2f2dc053 | 2226 | { |
2f2dc053 | 2227 | struct ceph_msg *reply; |
a105f00c | 2228 | struct rb_node *p; |
34b6c855 | 2229 | int mds = session->s_mds; |
9abf82b8 | 2230 | int err = -ENOMEM; |
93cea5be | 2231 | struct ceph_pagelist *pagelist; |
2f2dc053 | 2232 | |
34b6c855 | 2233 | pr_info("mds%d reconnect start\n", mds); |
2f2dc053 | 2234 | |
93cea5be SW |
2235 | pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS); |
2236 | if (!pagelist) | |
2237 | goto fail_nopagelist; | |
2238 | ceph_pagelist_init(pagelist); | |
2239 | ||
bb257664 | 2240 | reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0); |
a79832f2 | 2241 | if (!reply) |
93cea5be | 2242 | goto fail_nomsg; |
93cea5be | 2243 | |
34b6c855 SW |
2244 | mutex_lock(&session->s_mutex); |
2245 | session->s_state = CEPH_MDS_SESSION_RECONNECTING; | |
2246 | session->s_seq = 0; | |
2f2dc053 | 2247 | |
34b6c855 SW |
2248 | ceph_con_open(&session->s_con, |
2249 | ceph_mdsmap_get_addr(mdsc->mdsmap, mds)); | |
2f2dc053 | 2250 | |
34b6c855 SW |
2251 | /* replay unsafe requests */ |
2252 | replay_unsafe_requests(mdsc, session); | |
2f2dc053 SW |
2253 | |
2254 | down_read(&mdsc->snap_rwsem); | |
2255 | ||
2f2dc053 SW |
2256 | dout("session %p state %s\n", session, |
2257 | session_state_name(session->s_state)); | |
2258 | ||
e01a5946 SW |
2259 | /* drop old cap expires; we're about to reestablish that state */ |
2260 | discard_cap_releases(mdsc, session); | |
2261 | ||
2f2dc053 | 2262 | /* traverse this session's caps */ |
93cea5be SW |
2263 | err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps); |
2264 | if (err) | |
2265 | goto fail; | |
2266 | err = iterate_session_caps(session, encode_caps_cb, pagelist); | |
2f2dc053 | 2267 | if (err < 0) |
9abf82b8 | 2268 | goto fail; |
2f2dc053 SW |
2269 | |
2270 | /* | |
2271 | * snaprealms. we provide mds with the ino, seq (version), and | |
2272 | * parent for all of our realms. If the mds has any newer info, | |
2273 | * it will tell us. | |
2274 | */ | |
a105f00c SW |
2275 | for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) { |
2276 | struct ceph_snap_realm *realm = | |
2277 | rb_entry(p, struct ceph_snap_realm, node); | |
93cea5be | 2278 | struct ceph_mds_snaprealm_reconnect sr_rec; |
2f2dc053 SW |
2279 | |
2280 | dout(" adding snap realm %llx seq %lld parent %llx\n", | |
2281 | realm->ino, realm->seq, realm->parent_ino); | |
93cea5be SW |
2282 | sr_rec.ino = cpu_to_le64(realm->ino); |
2283 | sr_rec.seq = cpu_to_le64(realm->seq); | |
2284 | sr_rec.parent = cpu_to_le64(realm->parent_ino); | |
2285 | err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec)); | |
2286 | if (err) | |
2287 | goto fail; | |
2f2dc053 | 2288 | } |
2f2dc053 | 2289 | |
93cea5be SW |
2290 | reply->pagelist = pagelist; |
2291 | reply->hdr.data_len = cpu_to_le32(pagelist->length); | |
2292 | reply->nr_pages = calc_pages_for(0, pagelist->length); | |
2f2dc053 SW |
2293 | ceph_con_send(&session->s_con, reply); |
2294 | ||
9abf82b8 SW |
2295 | mutex_unlock(&session->s_mutex); |
2296 | ||
2297 | mutex_lock(&mdsc->mutex); | |
2298 | __wake_requests(mdsc, &session->s_waiting); | |
2299 | mutex_unlock(&mdsc->mutex); | |
2300 | ||
2f2dc053 | 2301 | up_read(&mdsc->snap_rwsem); |
2f2dc053 SW |
2302 | return; |
2303 | ||
93cea5be | 2304 | fail: |
2f2dc053 | 2305 | ceph_msg_put(reply); |
9abf82b8 SW |
2306 | up_read(&mdsc->snap_rwsem); |
2307 | mutex_unlock(&session->s_mutex); | |
93cea5be SW |
2308 | fail_nomsg: |
2309 | ceph_pagelist_release(pagelist); | |
2310 | kfree(pagelist); | |
2311 | fail_nopagelist: | |
9abf82b8 | 2312 | pr_err("error %d preparing reconnect for mds%d\n", err, mds); |
9abf82b8 | 2313 | return; |
2f2dc053 SW |
2314 | } |
2315 | ||
2316 | ||
2317 | /* | |
2318 | * compare old and new mdsmaps, kicking requests | |
2319 | * and closing out old connections as necessary | |
2320 | * | |
2321 | * called under mdsc->mutex. | |
2322 | */ | |
2323 | static void check_new_map(struct ceph_mds_client *mdsc, | |
2324 | struct ceph_mdsmap *newmap, | |
2325 | struct ceph_mdsmap *oldmap) | |
2326 | { | |
2327 | int i; | |
2328 | int oldstate, newstate; | |
2329 | struct ceph_mds_session *s; | |
2330 | ||
2331 | dout("check_new_map new %u old %u\n", | |
2332 | newmap->m_epoch, oldmap->m_epoch); | |
2333 | ||
2334 | for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) { | |
2335 | if (mdsc->sessions[i] == NULL) | |
2336 | continue; | |
2337 | s = mdsc->sessions[i]; | |
2338 | oldstate = ceph_mdsmap_get_state(oldmap, i); | |
2339 | newstate = ceph_mdsmap_get_state(newmap, i); | |
2340 | ||
2341 | dout("check_new_map mds%d state %s -> %s (session %s)\n", | |
2342 | i, ceph_mds_state_name(oldstate), | |
2343 | ceph_mds_state_name(newstate), | |
2344 | session_state_name(s->s_state)); | |
2345 | ||
2346 | if (memcmp(ceph_mdsmap_get_addr(oldmap, i), | |
2347 | ceph_mdsmap_get_addr(newmap, i), | |
2348 | sizeof(struct ceph_entity_addr))) { | |
2349 | if (s->s_state == CEPH_MDS_SESSION_OPENING) { | |
2350 | /* the session never opened, just close it | |
2351 | * out now */ | |
2352 | __wake_requests(mdsc, &s->s_waiting); | |
2600d2dd | 2353 | __unregister_session(mdsc, s); |
2f2dc053 SW |
2354 | } else { |
2355 | /* just close it */ | |
2356 | mutex_unlock(&mdsc->mutex); | |
2357 | mutex_lock(&s->s_mutex); | |
2358 | mutex_lock(&mdsc->mutex); | |
2359 | ceph_con_close(&s->s_con); | |
2360 | mutex_unlock(&s->s_mutex); | |
2361 | s->s_state = CEPH_MDS_SESSION_RESTARTING; | |
2362 | } | |
2363 | ||
2364 | /* kick any requests waiting on the recovering mds */ | |
29790f26 | 2365 | kick_requests(mdsc, i); |
2f2dc053 SW |
2366 | } else if (oldstate == newstate) { |
2367 | continue; /* nothing new with this mds */ | |
2368 | } | |
2369 | ||
2370 | /* | |
2371 | * send reconnect? | |
2372 | */ | |
2373 | if (s->s_state == CEPH_MDS_SESSION_RESTARTING && | |
34b6c855 SW |
2374 | newstate >= CEPH_MDS_STATE_RECONNECT) { |
2375 | mutex_unlock(&mdsc->mutex); | |
2376 | send_mds_reconnect(mdsc, s); | |
2377 | mutex_lock(&mdsc->mutex); | |
2378 | } | |
2f2dc053 SW |
2379 | |
2380 | /* | |
29790f26 | 2381 | * kick request on any mds that has gone active. |
2f2dc053 SW |
2382 | */ |
2383 | if (oldstate < CEPH_MDS_STATE_ACTIVE && | |
2384 | newstate >= CEPH_MDS_STATE_ACTIVE) { | |
29790f26 SW |
2385 | if (oldstate != CEPH_MDS_STATE_CREATING && |
2386 | oldstate != CEPH_MDS_STATE_STARTING) | |
2387 | pr_info("mds%d recovery completed\n", s->s_mds); | |
2388 | kick_requests(mdsc, i); | |
2f2dc053 | 2389 | ceph_kick_flushing_caps(mdsc, s); |
0dc2570f | 2390 | wake_up_session_caps(s, 1); |
2f2dc053 SW |
2391 | } |
2392 | } | |
2393 | } | |
2394 | ||
2395 | ||
2396 | ||
2397 | /* | |
2398 | * leases | |
2399 | */ | |
2400 | ||
2401 | /* | |
2402 | * caller must hold session s_mutex, dentry->d_lock | |
2403 | */ | |
2404 | void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry) | |
2405 | { | |
2406 | struct ceph_dentry_info *di = ceph_dentry(dentry); | |
2407 | ||
2408 | ceph_put_mds_session(di->lease_session); | |
2409 | di->lease_session = NULL; | |
2410 | } | |
2411 | ||
2600d2dd SW |
2412 | static void handle_lease(struct ceph_mds_client *mdsc, |
2413 | struct ceph_mds_session *session, | |
2414 | struct ceph_msg *msg) | |
2f2dc053 SW |
2415 | { |
2416 | struct super_block *sb = mdsc->client->sb; | |
2417 | struct inode *inode; | |
2f2dc053 SW |
2418 | struct ceph_inode_info *ci; |
2419 | struct dentry *parent, *dentry; | |
2420 | struct ceph_dentry_info *di; | |
2600d2dd | 2421 | int mds = session->s_mds; |
2f2dc053 SW |
2422 | struct ceph_mds_lease *h = msg->front.iov_base; |
2423 | struct ceph_vino vino; | |
2424 | int mask; | |
2425 | struct qstr dname; | |
2426 | int release = 0; | |
2427 | ||
2f2dc053 SW |
2428 | dout("handle_lease from mds%d\n", mds); |
2429 | ||
2430 | /* decode */ | |
2431 | if (msg->front.iov_len < sizeof(*h) + sizeof(u32)) | |
2432 | goto bad; | |
2433 | vino.ino = le64_to_cpu(h->ino); | |
2434 | vino.snap = CEPH_NOSNAP; | |
2435 | mask = le16_to_cpu(h->mask); | |
2436 | dname.name = (void *)h + sizeof(*h) + sizeof(u32); | |
2437 | dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32); | |
2438 | if (dname.len != get_unaligned_le32(h+1)) | |
2439 | goto bad; | |
2440 | ||
2f2dc053 SW |
2441 | mutex_lock(&session->s_mutex); |
2442 | session->s_seq++; | |
2443 | ||
2444 | /* lookup inode */ | |
2445 | inode = ceph_find_inode(sb, vino); | |
2446 | dout("handle_lease '%s', mask %d, ino %llx %p\n", | |
2447 | ceph_lease_op_name(h->action), mask, vino.ino, inode); | |
2448 | if (inode == NULL) { | |
2449 | dout("handle_lease no inode %llx\n", vino.ino); | |
2450 | goto release; | |
2451 | } | |
2452 | ci = ceph_inode(inode); | |
2453 | ||
2454 | /* dentry */ | |
2455 | parent = d_find_alias(inode); | |
2456 | if (!parent) { | |
2457 | dout("no parent dentry on inode %p\n", inode); | |
2458 | WARN_ON(1); | |
2459 | goto release; /* hrm... */ | |
2460 | } | |
2461 | dname.hash = full_name_hash(dname.name, dname.len); | |
2462 | dentry = d_lookup(parent, &dname); | |
2463 | dput(parent); | |
2464 | if (!dentry) | |
2465 | goto release; | |
2466 | ||
2467 | spin_lock(&dentry->d_lock); | |
2468 | di = ceph_dentry(dentry); | |
2469 | switch (h->action) { | |
2470 | case CEPH_MDS_LEASE_REVOKE: | |
2471 | if (di && di->lease_session == session) { | |
2472 | h->seq = cpu_to_le32(di->lease_seq); | |
2473 | __ceph_mdsc_drop_dentry_lease(dentry); | |
2474 | } | |
2475 | release = 1; | |
2476 | break; | |
2477 | ||
2478 | case CEPH_MDS_LEASE_RENEW: | |
2479 | if (di && di->lease_session == session && | |
2480 | di->lease_gen == session->s_cap_gen && | |
2481 | di->lease_renew_from && | |
2482 | di->lease_renew_after == 0) { | |
2483 | unsigned long duration = | |
2484 | le32_to_cpu(h->duration_ms) * HZ / 1000; | |
2485 | ||
2486 | di->lease_seq = le32_to_cpu(h->seq); | |
2487 | dentry->d_time = di->lease_renew_from + duration; | |
2488 | di->lease_renew_after = di->lease_renew_from + | |
2489 | (duration >> 1); | |
2490 | di->lease_renew_from = 0; | |
2491 | } | |
2492 | break; | |
2493 | } | |
2494 | spin_unlock(&dentry->d_lock); | |
2495 | dput(dentry); | |
2496 | ||
2497 | if (!release) | |
2498 | goto out; | |
2499 | ||
2500 | release: | |
2501 | /* let's just reuse the same message */ | |
2502 | h->action = CEPH_MDS_LEASE_REVOKE_ACK; | |
2503 | ceph_msg_get(msg); | |
2504 | ceph_con_send(&session->s_con, msg); | |
2505 | ||
2506 | out: | |
2507 | iput(inode); | |
2508 | mutex_unlock(&session->s_mutex); | |
2f2dc053 SW |
2509 | return; |
2510 | ||
2511 | bad: | |
2512 | pr_err("corrupt lease message\n"); | |
9ec7cab1 | 2513 | ceph_msg_dump(msg); |
2f2dc053 SW |
2514 | } |
2515 | ||
2516 | void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session, | |
2517 | struct inode *inode, | |
2518 | struct dentry *dentry, char action, | |
2519 | u32 seq) | |
2520 | { | |
2521 | struct ceph_msg *msg; | |
2522 | struct ceph_mds_lease *lease; | |
2523 | int len = sizeof(*lease) + sizeof(u32); | |
2524 | int dnamelen = 0; | |
2525 | ||
2526 | dout("lease_send_msg inode %p dentry %p %s to mds%d\n", | |
2527 | inode, dentry, ceph_lease_op_name(action), session->s_mds); | |
2528 | dnamelen = dentry->d_name.len; | |
2529 | len += dnamelen; | |
2530 | ||
bb257664 | 2531 | msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len); |
a79832f2 | 2532 | if (!msg) |
2f2dc053 SW |
2533 | return; |
2534 | lease = msg->front.iov_base; | |
2535 | lease->action = action; | |
2536 | lease->mask = cpu_to_le16(CEPH_LOCK_DN); | |
2537 | lease->ino = cpu_to_le64(ceph_vino(inode).ino); | |
2538 | lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap); | |
2539 | lease->seq = cpu_to_le32(seq); | |
2540 | put_unaligned_le32(dnamelen, lease + 1); | |
2541 | memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen); | |
2542 | ||
2543 | /* | |
2544 | * if this is a preemptive lease RELEASE, no need to | |
2545 | * flush request stream, since the actual request will | |
2546 | * soon follow. | |
2547 | */ | |
2548 | msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE); | |
2549 | ||
2550 | ceph_con_send(&session->s_con, msg); | |
2551 | } | |
2552 | ||
2553 | /* | |
2554 | * Preemptively release a lease we expect to invalidate anyway. | |
2555 | * Pass @inode always, @dentry is optional. | |
2556 | */ | |
2557 | void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode, | |
2558 | struct dentry *dentry, int mask) | |
2559 | { | |
2560 | struct ceph_dentry_info *di; | |
2561 | struct ceph_mds_session *session; | |
2562 | u32 seq; | |
2563 | ||
2564 | BUG_ON(inode == NULL); | |
2565 | BUG_ON(dentry == NULL); | |
2566 | BUG_ON(mask != CEPH_LOCK_DN); | |
2567 | ||
2568 | /* is dentry lease valid? */ | |
2569 | spin_lock(&dentry->d_lock); | |
2570 | di = ceph_dentry(dentry); | |
2571 | if (!di || !di->lease_session || | |
2572 | di->lease_session->s_mds < 0 || | |
2573 | di->lease_gen != di->lease_session->s_cap_gen || | |
2574 | !time_before(jiffies, dentry->d_time)) { | |
2575 | dout("lease_release inode %p dentry %p -- " | |
2576 | "no lease on %d\n", | |
2577 | inode, dentry, mask); | |
2578 | spin_unlock(&dentry->d_lock); | |
2579 | return; | |
2580 | } | |
2581 | ||
2582 | /* we do have a lease on this dentry; note mds and seq */ | |
2583 | session = ceph_get_mds_session(di->lease_session); | |
2584 | seq = di->lease_seq; | |
2585 | __ceph_mdsc_drop_dentry_lease(dentry); | |
2586 | spin_unlock(&dentry->d_lock); | |
2587 | ||
2588 | dout("lease_release inode %p dentry %p mask %d to mds%d\n", | |
2589 | inode, dentry, mask, session->s_mds); | |
2590 | ceph_mdsc_lease_send_msg(session, inode, dentry, | |
2591 | CEPH_MDS_LEASE_RELEASE, seq); | |
2592 | ceph_put_mds_session(session); | |
2593 | } | |
2594 | ||
2595 | /* | |
2596 | * drop all leases (and dentry refs) in preparation for umount | |
2597 | */ | |
2598 | static void drop_leases(struct ceph_mds_client *mdsc) | |
2599 | { | |
2600 | int i; | |
2601 | ||
2602 | dout("drop_leases\n"); | |
2603 | mutex_lock(&mdsc->mutex); | |
2604 | for (i = 0; i < mdsc->max_sessions; i++) { | |
2605 | struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i); | |
2606 | if (!s) | |
2607 | continue; | |
2608 | mutex_unlock(&mdsc->mutex); | |
2609 | mutex_lock(&s->s_mutex); | |
2610 | mutex_unlock(&s->s_mutex); | |
2611 | ceph_put_mds_session(s); | |
2612 | mutex_lock(&mdsc->mutex); | |
2613 | } | |
2614 | mutex_unlock(&mdsc->mutex); | |
2615 | } | |
2616 | ||
2617 | ||
2618 | ||
2619 | /* | |
2620 | * delayed work -- periodically trim expired leases, renew caps with mds | |
2621 | */ | |
2622 | static void schedule_delayed(struct ceph_mds_client *mdsc) | |
2623 | { | |
2624 | int delay = 5; | |
2625 | unsigned hz = round_jiffies_relative(HZ * delay); | |
2626 | schedule_delayed_work(&mdsc->delayed_work, hz); | |
2627 | } | |
2628 | ||
2629 | static void delayed_work(struct work_struct *work) | |
2630 | { | |
2631 | int i; | |
2632 | struct ceph_mds_client *mdsc = | |
2633 | container_of(work, struct ceph_mds_client, delayed_work.work); | |
2634 | int renew_interval; | |
2635 | int renew_caps; | |
2636 | ||
2637 | dout("mdsc delayed_work\n"); | |
afcdaea3 | 2638 | ceph_check_delayed_caps(mdsc); |
2f2dc053 SW |
2639 | |
2640 | mutex_lock(&mdsc->mutex); | |
2641 | renew_interval = mdsc->mdsmap->m_session_timeout >> 2; | |
2642 | renew_caps = time_after_eq(jiffies, HZ*renew_interval + | |
2643 | mdsc->last_renew_caps); | |
2644 | if (renew_caps) | |
2645 | mdsc->last_renew_caps = jiffies; | |
2646 | ||
2647 | for (i = 0; i < mdsc->max_sessions; i++) { | |
2648 | struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i); | |
2649 | if (s == NULL) | |
2650 | continue; | |
2651 | if (s->s_state == CEPH_MDS_SESSION_CLOSING) { | |
2652 | dout("resending session close request for mds%d\n", | |
2653 | s->s_mds); | |
2654 | request_close_session(mdsc, s); | |
2655 | ceph_put_mds_session(s); | |
2656 | continue; | |
2657 | } | |
2658 | if (s->s_ttl && time_after(jiffies, s->s_ttl)) { | |
2659 | if (s->s_state == CEPH_MDS_SESSION_OPEN) { | |
2660 | s->s_state = CEPH_MDS_SESSION_HUNG; | |
2661 | pr_info("mds%d hung\n", s->s_mds); | |
2662 | } | |
2663 | } | |
2664 | if (s->s_state < CEPH_MDS_SESSION_OPEN) { | |
2665 | /* this mds is failed or recovering, just wait */ | |
2666 | ceph_put_mds_session(s); | |
2667 | continue; | |
2668 | } | |
2669 | mutex_unlock(&mdsc->mutex); | |
2670 | ||
2671 | mutex_lock(&s->s_mutex); | |
2672 | if (renew_caps) | |
2673 | send_renew_caps(mdsc, s); | |
2674 | else | |
2675 | ceph_con_keepalive(&s->s_con); | |
2676 | add_cap_releases(mdsc, s, -1); | |
aab53dd9 SW |
2677 | if (s->s_state == CEPH_MDS_SESSION_OPEN || |
2678 | s->s_state == CEPH_MDS_SESSION_HUNG) | |
2679 | send_cap_releases(mdsc, s); | |
2f2dc053 SW |
2680 | mutex_unlock(&s->s_mutex); |
2681 | ceph_put_mds_session(s); | |
2682 | ||
2683 | mutex_lock(&mdsc->mutex); | |
2684 | } | |
2685 | mutex_unlock(&mdsc->mutex); | |
2686 | ||
2687 | schedule_delayed(mdsc); | |
2688 | } | |
2689 | ||
2690 | ||
5f44f142 | 2691 | int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client) |
2f2dc053 SW |
2692 | { |
2693 | mdsc->client = client; | |
2694 | mutex_init(&mdsc->mutex); | |
2695 | mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS); | |
2d06eeb8 CR |
2696 | if (mdsc->mdsmap == NULL) |
2697 | return -ENOMEM; | |
2698 | ||
2f2dc053 SW |
2699 | init_completion(&mdsc->safe_umount_waiters); |
2700 | init_completion(&mdsc->session_close_waiters); | |
2701 | INIT_LIST_HEAD(&mdsc->waiting_for_map); | |
2702 | mdsc->sessions = NULL; | |
2703 | mdsc->max_sessions = 0; | |
2704 | mdsc->stopping = 0; | |
2705 | init_rwsem(&mdsc->snap_rwsem); | |
a105f00c | 2706 | mdsc->snap_realms = RB_ROOT; |
2f2dc053 SW |
2707 | INIT_LIST_HEAD(&mdsc->snap_empty); |
2708 | spin_lock_init(&mdsc->snap_empty_lock); | |
2709 | mdsc->last_tid = 0; | |
44ca18f2 | 2710 | mdsc->request_tree = RB_ROOT; |
2f2dc053 SW |
2711 | INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work); |
2712 | mdsc->last_renew_caps = jiffies; | |
2713 | INIT_LIST_HEAD(&mdsc->cap_delay_list); | |
2714 | spin_lock_init(&mdsc->cap_delay_lock); | |
2715 | INIT_LIST_HEAD(&mdsc->snap_flush_list); | |
2716 | spin_lock_init(&mdsc->snap_flush_lock); | |
2717 | mdsc->cap_flush_seq = 0; | |
2718 | INIT_LIST_HEAD(&mdsc->cap_dirty); | |
2719 | mdsc->num_cap_flushing = 0; | |
2720 | spin_lock_init(&mdsc->cap_dirty_lock); | |
2721 | init_waitqueue_head(&mdsc->cap_flushing_wq); | |
2722 | spin_lock_init(&mdsc->dentry_lru_lock); | |
2723 | INIT_LIST_HEAD(&mdsc->dentry_lru); | |
2d06eeb8 | 2724 | |
5f44f142 | 2725 | return 0; |
2f2dc053 SW |
2726 | } |
2727 | ||
2728 | /* | |
2729 | * Wait for safe replies on open mds requests. If we time out, drop | |
2730 | * all requests from the tree to avoid dangling dentry refs. | |
2731 | */ | |
2732 | static void wait_requests(struct ceph_mds_client *mdsc) | |
2733 | { | |
2734 | struct ceph_mds_request *req; | |
2735 | struct ceph_client *client = mdsc->client; | |
2736 | ||
2737 | mutex_lock(&mdsc->mutex); | |
44ca18f2 | 2738 | if (__get_oldest_req(mdsc)) { |
2f2dc053 | 2739 | mutex_unlock(&mdsc->mutex); |
44ca18f2 | 2740 | |
2f2dc053 SW |
2741 | dout("wait_requests waiting for requests\n"); |
2742 | wait_for_completion_timeout(&mdsc->safe_umount_waiters, | |
6b805185 | 2743 | client->mount_args->mount_timeout * HZ); |
2f2dc053 SW |
2744 | |
2745 | /* tear down remaining requests */ | |
44ca18f2 SW |
2746 | mutex_lock(&mdsc->mutex); |
2747 | while ((req = __get_oldest_req(mdsc))) { | |
2f2dc053 SW |
2748 | dout("wait_requests timed out on tid %llu\n", |
2749 | req->r_tid); | |
44ca18f2 | 2750 | __unregister_request(mdsc, req); |
2f2dc053 SW |
2751 | } |
2752 | } | |
2753 | mutex_unlock(&mdsc->mutex); | |
2754 | dout("wait_requests done\n"); | |
2755 | } | |
2756 | ||
2757 | /* | |
2758 | * called before mount is ro, and before dentries are torn down. | |
2759 | * (hmm, does this still race with new lookups?) | |
2760 | */ | |
2761 | void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc) | |
2762 | { | |
2763 | dout("pre_umount\n"); | |
2764 | mdsc->stopping = 1; | |
2765 | ||
2766 | drop_leases(mdsc); | |
afcdaea3 | 2767 | ceph_flush_dirty_caps(mdsc); |
2f2dc053 SW |
2768 | wait_requests(mdsc); |
2769 | } | |
2770 | ||
2771 | /* | |
2772 | * wait for all write mds requests to flush. | |
2773 | */ | |
2774 | static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid) | |
2775 | { | |
80fc7314 | 2776 | struct ceph_mds_request *req = NULL, *nextreq; |
44ca18f2 | 2777 | struct rb_node *n; |
2f2dc053 SW |
2778 | |
2779 | mutex_lock(&mdsc->mutex); | |
2780 | dout("wait_unsafe_requests want %lld\n", want_tid); | |
80fc7314 | 2781 | restart: |
44ca18f2 SW |
2782 | req = __get_oldest_req(mdsc); |
2783 | while (req && req->r_tid <= want_tid) { | |
80fc7314 SW |
2784 | /* find next request */ |
2785 | n = rb_next(&req->r_node); | |
2786 | if (n) | |
2787 | nextreq = rb_entry(n, struct ceph_mds_request, r_node); | |
2788 | else | |
2789 | nextreq = NULL; | |
44ca18f2 SW |
2790 | if ((req->r_op & CEPH_MDS_OP_WRITE)) { |
2791 | /* write op */ | |
2792 | ceph_mdsc_get_request(req); | |
80fc7314 SW |
2793 | if (nextreq) |
2794 | ceph_mdsc_get_request(nextreq); | |
44ca18f2 SW |
2795 | mutex_unlock(&mdsc->mutex); |
2796 | dout("wait_unsafe_requests wait on %llu (want %llu)\n", | |
2797 | req->r_tid, want_tid); | |
2798 | wait_for_completion(&req->r_safe_completion); | |
2799 | mutex_lock(&mdsc->mutex); | |
44ca18f2 | 2800 | ceph_mdsc_put_request(req); |
80fc7314 SW |
2801 | if (!nextreq) |
2802 | break; /* next dne before, so we're done! */ | |
2803 | if (RB_EMPTY_NODE(&nextreq->r_node)) { | |
2804 | /* next request was removed from tree */ | |
2805 | ceph_mdsc_put_request(nextreq); | |
2806 | goto restart; | |
2807 | } | |
2808 | ceph_mdsc_put_request(nextreq); /* won't go away */ | |
44ca18f2 | 2809 | } |
80fc7314 | 2810 | req = nextreq; |
2f2dc053 SW |
2811 | } |
2812 | mutex_unlock(&mdsc->mutex); | |
2813 | dout("wait_unsafe_requests done\n"); | |
2814 | } | |
2815 | ||
2816 | void ceph_mdsc_sync(struct ceph_mds_client *mdsc) | |
2817 | { | |
2818 | u64 want_tid, want_flush; | |
2819 | ||
56b7cf95 SW |
2820 | if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN) |
2821 | return; | |
2822 | ||
2f2dc053 SW |
2823 | dout("sync\n"); |
2824 | mutex_lock(&mdsc->mutex); | |
2825 | want_tid = mdsc->last_tid; | |
2826 | want_flush = mdsc->cap_flush_seq; | |
2827 | mutex_unlock(&mdsc->mutex); | |
2828 | dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush); | |
2829 | ||
afcdaea3 | 2830 | ceph_flush_dirty_caps(mdsc); |
2f2dc053 SW |
2831 | |
2832 | wait_unsafe_requests(mdsc, want_tid); | |
2833 | wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush)); | |
2834 | } | |
2835 | ||
2836 | ||
2837 | /* | |
2838 | * called after sb is ro. | |
2839 | */ | |
2840 | void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc) | |
2841 | { | |
2842 | struct ceph_mds_session *session; | |
2843 | int i; | |
2844 | int n; | |
2845 | struct ceph_client *client = mdsc->client; | |
6b805185 | 2846 | unsigned long started, timeout = client->mount_args->mount_timeout * HZ; |
2f2dc053 SW |
2847 | |
2848 | dout("close_sessions\n"); | |
2849 | ||
2850 | mutex_lock(&mdsc->mutex); | |
2851 | ||
2852 | /* close sessions */ | |
2853 | started = jiffies; | |
2854 | while (time_before(jiffies, started + timeout)) { | |
2855 | dout("closing sessions\n"); | |
2856 | n = 0; | |
2857 | for (i = 0; i < mdsc->max_sessions; i++) { | |
2858 | session = __ceph_lookup_mds_session(mdsc, i); | |
2859 | if (!session) | |
2860 | continue; | |
2861 | mutex_unlock(&mdsc->mutex); | |
2862 | mutex_lock(&session->s_mutex); | |
2863 | __close_session(mdsc, session); | |
2864 | mutex_unlock(&session->s_mutex); | |
2865 | ceph_put_mds_session(session); | |
2866 | mutex_lock(&mdsc->mutex); | |
2867 | n++; | |
2868 | } | |
2869 | if (n == 0) | |
2870 | break; | |
2871 | ||
2872 | if (client->mount_state == CEPH_MOUNT_SHUTDOWN) | |
2873 | break; | |
2874 | ||
2875 | dout("waiting for sessions to close\n"); | |
2876 | mutex_unlock(&mdsc->mutex); | |
2877 | wait_for_completion_timeout(&mdsc->session_close_waiters, | |
2878 | timeout); | |
2879 | mutex_lock(&mdsc->mutex); | |
2880 | } | |
2881 | ||
2882 | /* tear down remaining sessions */ | |
2883 | for (i = 0; i < mdsc->max_sessions; i++) { | |
2884 | if (mdsc->sessions[i]) { | |
2885 | session = get_session(mdsc->sessions[i]); | |
2600d2dd | 2886 | __unregister_session(mdsc, session); |
2f2dc053 SW |
2887 | mutex_unlock(&mdsc->mutex); |
2888 | mutex_lock(&session->s_mutex); | |
2889 | remove_session_caps(session); | |
2890 | mutex_unlock(&session->s_mutex); | |
2891 | ceph_put_mds_session(session); | |
2892 | mutex_lock(&mdsc->mutex); | |
2893 | } | |
2894 | } | |
2895 | ||
2896 | WARN_ON(!list_empty(&mdsc->cap_delay_list)); | |
2897 | ||
2898 | mutex_unlock(&mdsc->mutex); | |
2899 | ||
2900 | ceph_cleanup_empty_realms(mdsc); | |
2901 | ||
2902 | cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */ | |
2903 | ||
2904 | dout("stopped\n"); | |
2905 | } | |
2906 | ||
2907 | void ceph_mdsc_stop(struct ceph_mds_client *mdsc) | |
2908 | { | |
2909 | dout("stop\n"); | |
2910 | cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */ | |
2911 | if (mdsc->mdsmap) | |
2912 | ceph_mdsmap_destroy(mdsc->mdsmap); | |
2913 | kfree(mdsc->sessions); | |
2914 | } | |
2915 | ||
2916 | ||
2917 | /* | |
2918 | * handle mds map update. | |
2919 | */ | |
2920 | void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg) | |
2921 | { | |
2922 | u32 epoch; | |
2923 | u32 maplen; | |
2924 | void *p = msg->front.iov_base; | |
2925 | void *end = p + msg->front.iov_len; | |
2926 | struct ceph_mdsmap *newmap, *oldmap; | |
2927 | struct ceph_fsid fsid; | |
2928 | int err = -EINVAL; | |
2929 | ||
2930 | ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad); | |
2931 | ceph_decode_copy(&p, &fsid, sizeof(fsid)); | |
0743304d SW |
2932 | if (ceph_check_fsid(mdsc->client, &fsid) < 0) |
2933 | return; | |
c89136ea SW |
2934 | epoch = ceph_decode_32(&p); |
2935 | maplen = ceph_decode_32(&p); | |
2f2dc053 SW |
2936 | dout("handle_map epoch %u len %d\n", epoch, (int)maplen); |
2937 | ||
2938 | /* do we need it? */ | |
2939 | ceph_monc_got_mdsmap(&mdsc->client->monc, epoch); | |
2940 | mutex_lock(&mdsc->mutex); | |
2941 | if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) { | |
2942 | dout("handle_map epoch %u <= our %u\n", | |
2943 | epoch, mdsc->mdsmap->m_epoch); | |
2944 | mutex_unlock(&mdsc->mutex); | |
2945 | return; | |
2946 | } | |
2947 | ||
2948 | newmap = ceph_mdsmap_decode(&p, end); | |
2949 | if (IS_ERR(newmap)) { | |
2950 | err = PTR_ERR(newmap); | |
2951 | goto bad_unlock; | |
2952 | } | |
2953 | ||
2954 | /* swap into place */ | |
2955 | if (mdsc->mdsmap) { | |
2956 | oldmap = mdsc->mdsmap; | |
2957 | mdsc->mdsmap = newmap; | |
2958 | check_new_map(mdsc, newmap, oldmap); | |
2959 | ceph_mdsmap_destroy(oldmap); | |
2960 | } else { | |
2961 | mdsc->mdsmap = newmap; /* first mds map */ | |
2962 | } | |
2963 | mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size; | |
2964 | ||
2965 | __wake_requests(mdsc, &mdsc->waiting_for_map); | |
2966 | ||
2967 | mutex_unlock(&mdsc->mutex); | |
2968 | schedule_delayed(mdsc); | |
2969 | return; | |
2970 | ||
2971 | bad_unlock: | |
2972 | mutex_unlock(&mdsc->mutex); | |
2973 | bad: | |
2974 | pr_err("error decoding mdsmap %d\n", err); | |
2975 | return; | |
2976 | } | |
2977 | ||
2978 | static struct ceph_connection *con_get(struct ceph_connection *con) | |
2979 | { | |
2980 | struct ceph_mds_session *s = con->private; | |
2981 | ||
2982 | if (get_session(s)) { | |
2600d2dd | 2983 | dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref)); |
2f2dc053 SW |
2984 | return con; |
2985 | } | |
2986 | dout("mdsc con_get %p FAIL\n", s); | |
2987 | return NULL; | |
2988 | } | |
2989 | ||
2990 | static void con_put(struct ceph_connection *con) | |
2991 | { | |
2992 | struct ceph_mds_session *s = con->private; | |
2993 | ||
2f2dc053 | 2994 | ceph_put_mds_session(s); |
2600d2dd | 2995 | dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref)); |
2f2dc053 SW |
2996 | } |
2997 | ||
2998 | /* | |
2999 | * if the client is unresponsive for long enough, the mds will kill | |
3000 | * the session entirely. | |
3001 | */ | |
3002 | static void peer_reset(struct ceph_connection *con) | |
3003 | { | |
3004 | struct ceph_mds_session *s = con->private; | |
7e70f0ed | 3005 | struct ceph_mds_client *mdsc = s->s_mdsc; |
2f2dc053 | 3006 | |
7e70f0ed SW |
3007 | pr_warning("mds%d closed our session\n", s->s_mds); |
3008 | send_mds_reconnect(mdsc, s); | |
2f2dc053 SW |
3009 | } |
3010 | ||
3011 | static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) | |
3012 | { | |
3013 | struct ceph_mds_session *s = con->private; | |
3014 | struct ceph_mds_client *mdsc = s->s_mdsc; | |
3015 | int type = le16_to_cpu(msg->hdr.type); | |
3016 | ||
2600d2dd SW |
3017 | mutex_lock(&mdsc->mutex); |
3018 | if (__verify_registered_session(mdsc, s) < 0) { | |
3019 | mutex_unlock(&mdsc->mutex); | |
3020 | goto out; | |
3021 | } | |
3022 | mutex_unlock(&mdsc->mutex); | |
3023 | ||
2f2dc053 SW |
3024 | switch (type) { |
3025 | case CEPH_MSG_MDS_MAP: | |
3026 | ceph_mdsc_handle_map(mdsc, msg); | |
3027 | break; | |
3028 | case CEPH_MSG_CLIENT_SESSION: | |
3029 | handle_session(s, msg); | |
3030 | break; | |
3031 | case CEPH_MSG_CLIENT_REPLY: | |
3032 | handle_reply(s, msg); | |
3033 | break; | |
3034 | case CEPH_MSG_CLIENT_REQUEST_FORWARD: | |
2600d2dd | 3035 | handle_forward(mdsc, s, msg); |
2f2dc053 SW |
3036 | break; |
3037 | case CEPH_MSG_CLIENT_CAPS: | |
3038 | ceph_handle_caps(s, msg); | |
3039 | break; | |
3040 | case CEPH_MSG_CLIENT_SNAP: | |
2600d2dd | 3041 | ceph_handle_snap(mdsc, s, msg); |
2f2dc053 SW |
3042 | break; |
3043 | case CEPH_MSG_CLIENT_LEASE: | |
2600d2dd | 3044 | handle_lease(mdsc, s, msg); |
2f2dc053 SW |
3045 | break; |
3046 | ||
3047 | default: | |
3048 | pr_err("received unknown message type %d %s\n", type, | |
3049 | ceph_msg_type_name(type)); | |
3050 | } | |
2600d2dd | 3051 | out: |
2f2dc053 SW |
3052 | ceph_msg_put(msg); |
3053 | } | |
3054 | ||
4e7a5dcd SW |
3055 | /* |
3056 | * authentication | |
3057 | */ | |
3058 | static int get_authorizer(struct ceph_connection *con, | |
3059 | void **buf, int *len, int *proto, | |
3060 | void **reply_buf, int *reply_len, int force_new) | |
3061 | { | |
3062 | struct ceph_mds_session *s = con->private; | |
3063 | struct ceph_mds_client *mdsc = s->s_mdsc; | |
3064 | struct ceph_auth_client *ac = mdsc->client->monc.auth; | |
3065 | int ret = 0; | |
3066 | ||
3067 | if (force_new && s->s_authorizer) { | |
3068 | ac->ops->destroy_authorizer(ac, s->s_authorizer); | |
3069 | s->s_authorizer = NULL; | |
3070 | } | |
3071 | if (s->s_authorizer == NULL) { | |
3072 | if (ac->ops->create_authorizer) { | |
3073 | ret = ac->ops->create_authorizer( | |
3074 | ac, CEPH_ENTITY_TYPE_MDS, | |
3075 | &s->s_authorizer, | |
3076 | &s->s_authorizer_buf, | |
3077 | &s->s_authorizer_buf_len, | |
3078 | &s->s_authorizer_reply_buf, | |
3079 | &s->s_authorizer_reply_buf_len); | |
3080 | if (ret) | |
3081 | return ret; | |
3082 | } | |
3083 | } | |
3084 | ||
3085 | *proto = ac->protocol; | |
3086 | *buf = s->s_authorizer_buf; | |
3087 | *len = s->s_authorizer_buf_len; | |
3088 | *reply_buf = s->s_authorizer_reply_buf; | |
3089 | *reply_len = s->s_authorizer_reply_buf_len; | |
3090 | return 0; | |
3091 | } | |
3092 | ||
3093 | ||
3094 | static int verify_authorizer_reply(struct ceph_connection *con, int len) | |
3095 | { | |
3096 | struct ceph_mds_session *s = con->private; | |
3097 | struct ceph_mds_client *mdsc = s->s_mdsc; | |
3098 | struct ceph_auth_client *ac = mdsc->client->monc.auth; | |
3099 | ||
3100 | return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len); | |
3101 | } | |
3102 | ||
9bd2e6f8 SW |
3103 | static int invalidate_authorizer(struct ceph_connection *con) |
3104 | { | |
3105 | struct ceph_mds_session *s = con->private; | |
3106 | struct ceph_mds_client *mdsc = s->s_mdsc; | |
3107 | struct ceph_auth_client *ac = mdsc->client->monc.auth; | |
3108 | ||
3109 | if (ac->ops->invalidate_authorizer) | |
3110 | ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS); | |
3111 | ||
3112 | return ceph_monc_validate_auth(&mdsc->client->monc); | |
3113 | } | |
3114 | ||
2f2dc053 SW |
3115 | const static struct ceph_connection_operations mds_con_ops = { |
3116 | .get = con_get, | |
3117 | .put = con_put, | |
3118 | .dispatch = dispatch, | |
4e7a5dcd SW |
3119 | .get_authorizer = get_authorizer, |
3120 | .verify_authorizer_reply = verify_authorizer_reply, | |
9bd2e6f8 | 3121 | .invalidate_authorizer = invalidate_authorizer, |
2f2dc053 | 3122 | .peer_reset = peer_reset, |
2f2dc053 SW |
3123 | }; |
3124 | ||
3125 | ||
3126 | ||
3127 | ||
3128 | /* eof */ |