]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - fs/ceph/mds_client.c
projects / mirror_ubuntu-bionic-kernel.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ceph: hold on to exclusive caps on complete directories
[mirror_ubuntu-bionic-kernel.git] / fs / ceph / mds_client.c
CommitLineData
3d14c5d2 1#include <linux/ceph/ceph_debug.h>
2f2dc053 2
496e5955 3#include <linux/fs.h>
2f2dc053 4#include <linux/wait.h>
5a0e3ad6 5#include <linux/slab.h>
54008399 6#include <linux/gfp.h>
2f2dc053 7#include <linux/sched.h>
3d14c5d2
YS		 8#include <linux/debugfs.h>
9#include <linux/seq_file.h>
dbd0c8bf 10#include <linux/utsname.h>
2f2dc053 11
2f2dc053 12#include "super.h"
3d14c5d2
YS		 13#include "mds_client.h"
14
1fe60e51 15#include <linux/ceph/ceph_features.h>
3d14c5d2
YS		 16#include <linux/ceph/messenger.h>
17#include <linux/ceph/decode.h>
18#include <linux/ceph/pagelist.h>
19#include <linux/ceph/auth.h>
20#include <linux/ceph/debugfs.h>
2f2dc053
SW		 21
22/*
23 * A cluster of MDS (metadata server) daemons is responsible for
24 * managing the file system namespace (the directory hierarchy and
25 * inodes) and for coordinating shared access to storage. Metadata is
26 * partitioning hierarchically across a number of servers, and that
27 * partition varies over time as the cluster adjusts the distribution
28 * in order to balance load.
29 *
30 * The MDS client is primarily responsible to managing synchronous
31 * metadata requests for operations like open, unlink, and so forth.
32 * If there is a MDS failure, we find out about it when we (possibly
33 * request and) receive a new MDS map, and can resubmit affected
34 * requests.
35 *
36 * For the most part, though, we take advantage of a lossless
37 * communications channel to the MDS, and do not need to worry about
38 * timing out or resubmitting requests.
39 *
40 * We maintain a stateful "session" with each MDS we interact with.
41 * Within each session, we sent periodic heartbeat messages to ensure
42 * any capabilities or leases we have been issues remain valid. If
43 * the session times out and goes stale, our leases and capabilities
44 * are no longer valid.
45 */
46
20cb34ae 47struct ceph_reconnect_state {
44c99757 48 int nr_caps;
20cb34ae
SW		 49 struct ceph_pagelist *pagelist;
50 bool flock;
51};
52
2f2dc053
SW		 53static void __wake_requests(struct ceph_mds_client *mdsc,
54 struct list_head *head);
55
9e32789f 56static const struct ceph_connection_operations mds_con_ops;
2f2dc053
SW		 57
58
59/*
60 * mds reply parsing
61 */
62
63/*
64 * parse individual inode info
65 */
66static int parse_reply_info_in(void **p, void *end,
14303d20 67 struct ceph_mds_reply_info_in *info,
12b4629a 68 u64 features)
2f2dc053
SW		 69{
70 int err = -EIO;
71
72 info->in = *p;
73 *p += sizeof(struct ceph_mds_reply_inode) +
74 sizeof(*info->in->fragtree.splits) *
75 le32_to_cpu(info->in->fragtree.nsplits);
76
77 ceph_decode_32_safe(p, end, info->symlink_len, bad);
78 ceph_decode_need(p, end, info->symlink_len, bad);
79 info->symlink = *p;
80 *p += info->symlink_len;
81
14303d20
SW		 82 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
83 ceph_decode_copy_safe(p, end, &info->dir_layout,
84 sizeof(info->dir_layout), bad);
85 else
86 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
87
2f2dc053
SW		 88 ceph_decode_32_safe(p, end, info->xattr_len, bad);
89 ceph_decode_need(p, end, info->xattr_len, bad);
90 info->xattr_data = *p;
91 *p += info->xattr_len;
fb01d1f8
YZ		 92
93 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
94 ceph_decode_64_safe(p, end, info->inline_version, bad);
95 ceph_decode_32_safe(p, end, info->inline_len, bad);
96 ceph_decode_need(p, end, info->inline_len, bad);
97 info->inline_data = *p;
98 *p += info->inline_len;
99 } else
100 info->inline_version = CEPH_INLINE_NONE;
101
2f2dc053
SW		 102 return 0;
103bad:
104 return err;
105}
106
107/*
108 * parse a normal reply, which may contain a (dir+)dentry and/or a
109 * target inode.
110 */
111static int parse_reply_info_trace(void **p, void *end,
14303d20 112 struct ceph_mds_reply_info_parsed *info,
12b4629a 113 u64 features)
2f2dc053
SW		 114{
115 int err;
116
117 if (info->head->is_dentry) {
14303d20 118 err = parse_reply_info_in(p, end, &info->diri, features);
2f2dc053
SW		 119 if (err < 0)
120 goto out_bad;
121
122 if (unlikely(*p + sizeof(*info->dirfrag) > end))
123 goto bad;
124 info->dirfrag = *p;
125 *p += sizeof(*info->dirfrag) +
126 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
127 if (unlikely(*p > end))
128 goto bad;
129
130 ceph_decode_32_safe(p, end, info->dname_len, bad);
131 ceph_decode_need(p, end, info->dname_len, bad);
132 info->dname = *p;
133 *p += info->dname_len;
134 info->dlease = *p;
135 *p += sizeof(*info->dlease);
136 }
137
138 if (info->head->is_target) {
14303d20 139 err = parse_reply_info_in(p, end, &info->targeti, features);
2f2dc053
SW		 140 if (err < 0)
141 goto out_bad;
142 }
143
144 if (unlikely(*p != end))
145 goto bad;
146 return 0;
147
148bad:
149 err = -EIO;
150out_bad:
151 pr_err("problem parsing mds trace %d\n", err);
152 return err;
153}
154
155/*
156 * parse readdir results
157 */
158static int parse_reply_info_dir(void **p, void *end,
14303d20 159 struct ceph_mds_reply_info_parsed *info,
12b4629a 160 u64 features)
2f2dc053
SW		 161{
162 u32 num, i = 0;
163 int err;
164
165 info->dir_dir = *p;
166 if (*p + sizeof(*info->dir_dir) > end)
167 goto bad;
168 *p += sizeof(*info->dir_dir) +
169 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
170 if (*p > end)
171 goto bad;
172
173 ceph_decode_need(p, end, sizeof(num) + 2, bad);
c89136ea
SW		 174 num = ceph_decode_32(p);
175 info->dir_end = ceph_decode_8(p);
176 info->dir_complete = ceph_decode_8(p);
2f2dc053
SW		 177 if (num == 0)
178 goto done;
179
54008399 180 BUG_ON(!info->dir_in);
2f2dc053
SW		 181 info->dir_dname = (void *)(info->dir_in + num);
182 info->dir_dname_len = (void *)(info->dir_dname + num);
183 info->dir_dlease = (void *)(info->dir_dname_len + num);
54008399
YZ		 184 if ((unsigned long)(info->dir_dlease + num) >
185 (unsigned long)info->dir_in + info->dir_buf_size) {
186 pr_err("dir contents are larger than expected\n");
187 WARN_ON(1);
188 goto bad;
189 }
2f2dc053 190
54008399 191 info->dir_nr = num;
2f2dc053
SW		 192 while (num) {
193 /* dentry */
194 ceph_decode_need(p, end, sizeof(u32)*2, bad);
c89136ea 195 info->dir_dname_len[i] = ceph_decode_32(p);
2f2dc053
SW		 196 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
197 info->dir_dname[i] = *p;
198 *p += info->dir_dname_len[i];
199 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
200 info->dir_dname[i]);
201 info->dir_dlease[i] = *p;
202 *p += sizeof(struct ceph_mds_reply_lease);
203
204 /* inode */
14303d20 205 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
2f2dc053
SW		 206 if (err < 0)
207 goto out_bad;
208 i++;
209 num--;
210 }
211
212done:
213 if (*p != end)
214 goto bad;
215 return 0;
216
217bad:
218 err = -EIO;
219out_bad:
220 pr_err("problem parsing dir contents %d\n", err);
221 return err;
222}
223
25933abd
HS		 224/*
225 * parse fcntl F_GETLK results
226 */
227static int parse_reply_info_filelock(void **p, void *end,
14303d20 228 struct ceph_mds_reply_info_parsed *info,
12b4629a 229 u64 features)
25933abd
HS		 230{
231 if (*p + sizeof(*info->filelock_reply) > end)
232 goto bad;
233
234 info->filelock_reply = *p;
235 *p += sizeof(*info->filelock_reply);
236
237 if (unlikely(*p != end))
238 goto bad;
239 return 0;
240
241bad:
242 return -EIO;
243}
244
6e8575fa
SL		 245/*
246 * parse create results
247 */
248static int parse_reply_info_create(void **p, void *end,
249 struct ceph_mds_reply_info_parsed *info,
12b4629a 250 u64 features)
6e8575fa
SL		 251{
252 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
253 if (*p == end) {
254 info->has_create_ino = false;
255 } else {
256 info->has_create_ino = true;
257 info->ino = ceph_decode_64(p);
258 }
259 }
260
261 if (unlikely(*p != end))
262 goto bad;
263 return 0;
264
265bad:
266 return -EIO;
267}
268
25933abd
HS		 269/*
270 * parse extra results
271 */
272static int parse_reply_info_extra(void **p, void *end,
14303d20 273 struct ceph_mds_reply_info_parsed *info,
12b4629a 274 u64 features)
25933abd
HS		 275{
276 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
14303d20 277 return parse_reply_info_filelock(p, end, info, features);
8a034497
YZ		 278 else if (info->head->op == CEPH_MDS_OP_READDIR ||
279 info->head->op == CEPH_MDS_OP_LSSNAP)
14303d20 280 return parse_reply_info_dir(p, end, info, features);
6e8575fa
SL		 281 else if (info->head->op == CEPH_MDS_OP_CREATE)
282 return parse_reply_info_create(p, end, info, features);
283 else
284 return -EIO;
25933abd
HS		 285}
286
2f2dc053
SW		 287/*
288 * parse entire mds reply
289 */
290static int parse_reply_info(struct ceph_msg *msg,
14303d20 291 struct ceph_mds_reply_info_parsed *info,
12b4629a 292 u64 features)
2f2dc053
SW		 293{
294 void *p, *end;
295 u32 len;
296 int err;
297
298 info->head = msg->front.iov_base;
299 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
300 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
301
302 /* trace */
303 ceph_decode_32_safe(&p, end, len, bad);
304 if (len > 0) {
32852a81 305 ceph_decode_need(&p, end, len, bad);
14303d20 306 err = parse_reply_info_trace(&p, p+len, info, features);
2f2dc053
SW		 307 if (err < 0)
308 goto out_bad;
309 }
310
25933abd 311 /* extra */
2f2dc053
SW		 312 ceph_decode_32_safe(&p, end, len, bad);
313 if (len > 0) {
32852a81 314 ceph_decode_need(&p, end, len, bad);
14303d20 315 err = parse_reply_info_extra(&p, p+len, info, features);
2f2dc053
SW		 316 if (err < 0)
317 goto out_bad;
318 }
319
320 /* snap blob */
321 ceph_decode_32_safe(&p, end, len, bad);
322 info->snapblob_len = len;
323 info->snapblob = p;
324 p += len;
325
326 if (p != end)
327 goto bad;
328 return 0;
329
330bad:
331 err = -EIO;
332out_bad:
333 pr_err("mds parse_reply err %d\n", err);
334 return err;
335}
336
337static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
338{
54008399
YZ		 339 if (!info->dir_in)
340 return;
341 free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
2f2dc053
SW		 342}
343
344
345/*
346 * sessions
347 */
a687ecaf 348const char *ceph_session_state_name(int s)
2f2dc053
SW		 349{
350 switch (s) {
351 case CEPH_MDS_SESSION_NEW: return "new";
352 case CEPH_MDS_SESSION_OPENING: return "opening";
353 case CEPH_MDS_SESSION_OPEN: return "open";
354 case CEPH_MDS_SESSION_HUNG: return "hung";
355 case CEPH_MDS_SESSION_CLOSING: return "closing";
44ca18f2 356 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
2f2dc053
SW		 357 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
358 default: return "???";
359 }
360}
361
362static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
363{
364 if (atomic_inc_not_zero(&s->s_ref)) {
365 dout("mdsc get_session %p %d -> %d\n", s,
366 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
367 return s;
368 } else {
369 dout("mdsc get_session %p 0 -- FAIL", s);
370 return NULL;
371 }
372}
373
374void ceph_put_mds_session(struct ceph_mds_session *s)
375{
376 dout("mdsc put_session %p %d -> %d\n", s,
377 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
4e7a5dcd 378 if (atomic_dec_and_test(&s->s_ref)) {
6c4a1915 379 if (s->s_auth.authorizer)
27859f97
SW		 380 ceph_auth_destroy_authorizer(
381 s->s_mdsc->fsc->client->monc.auth,
382 s->s_auth.authorizer);
2f2dc053 383 kfree(s);
4e7a5dcd 384 }
2f2dc053
SW		 385}
386
387/*
388 * called under mdsc->mutex
389 */
390struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
391 int mds)
392{
393 struct ceph_mds_session *session;
394
395 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
396 return NULL;
397 session = mdsc->sessions[mds];
398 dout("lookup_mds_session %p %d\n", session,
399 atomic_read(&session->s_ref));
400 get_session(session);
401 return session;
402}
403
404static bool __have_session(struct ceph_mds_client *mdsc, int mds)
405{
406 if (mds >= mdsc->max_sessions)
407 return false;
408 return mdsc->sessions[mds];
409}
410
2600d2dd
SW		 411static int __verify_registered_session(struct ceph_mds_client *mdsc,
412 struct ceph_mds_session *s)
413{
414 if (s->s_mds >= mdsc->max_sessions ||
415 mdsc->sessions[s->s_mds] != s)
416 return -ENOENT;
417 return 0;
418}
419
2f2dc053
SW		 420/*
421 * create+register a new session for given mds.
422 * called under mdsc->mutex.
423 */
424static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
425 int mds)
426{
427 struct ceph_mds_session *s;
428
c338c07c
NY		 429 if (mds >= mdsc->mdsmap->m_max_mds)
430 return ERR_PTR(-EINVAL);
431
2f2dc053 432 s = kzalloc(sizeof(*s), GFP_NOFS);
4736b009
DC		 433 if (!s)
434 return ERR_PTR(-ENOMEM);
2f2dc053
SW		 435 s->s_mdsc = mdsc;
436 s->s_mds = mds;
437 s->s_state = CEPH_MDS_SESSION_NEW;
438 s->s_ttl = 0;
439 s->s_seq = 0;
440 mutex_init(&s->s_mutex);
441
b7a9e5dd 442 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
2f2dc053 443
d8fb02ab 444 spin_lock_init(&s->s_gen_ttl_lock);
2f2dc053 445 s->s_cap_gen = 0;
1ce208a6 446 s->s_cap_ttl = jiffies - 1;
d8fb02ab
AE		 447
448 spin_lock_init(&s->s_cap_lock);
2f2dc053
SW		 449 s->s_renew_requested = 0;
450 s->s_renew_seq = 0;
451 INIT_LIST_HEAD(&s->s_caps);
452 s->s_nr_caps = 0;
5dacf091 453 s->s_trim_caps = 0;
2f2dc053
SW		 454 atomic_set(&s->s_ref, 1);
455 INIT_LIST_HEAD(&s->s_waiting);
456 INIT_LIST_HEAD(&s->s_unsafe);
457 s->s_num_cap_releases = 0;
99a9c273 458 s->s_cap_reconnect = 0;
7c1332b8 459 s->s_cap_iterator = NULL;
2f2dc053
SW		 460 INIT_LIST_HEAD(&s->s_cap_releases);
461 INIT_LIST_HEAD(&s->s_cap_releases_done);
462 INIT_LIST_HEAD(&s->s_cap_flushing);
463 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
464
465 dout("register_session mds%d\n", mds);
466 if (mds >= mdsc->max_sessions) {
467 int newmax = 1 << get_count_order(mds+1);
468 struct ceph_mds_session **sa;
469
470 dout("register_session realloc to %d\n", newmax);
471 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
472 if (sa == NULL)
42ce56e5 473 goto fail_realloc;
2f2dc053
SW		 474 if (mdsc->sessions) {
475 memcpy(sa, mdsc->sessions,
476 mdsc->max_sessions * sizeof(void *));
477 kfree(mdsc->sessions);
478 }
479 mdsc->sessions = sa;
480 mdsc->max_sessions = newmax;
481 }
482 mdsc->sessions[mds] = s;
86d8f67b 483 atomic_inc(&mdsc->num_sessions);
2f2dc053 484 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
42ce56e5 485
b7a9e5dd
SW		 486 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
487 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
42ce56e5 488
2f2dc053 489 return s;
42ce56e5
SW		 490
491fail_realloc:
492 kfree(s);
493 return ERR_PTR(-ENOMEM);
2f2dc053
SW		 494}
495
496/*
497 * called under mdsc->mutex
498 */
2600d2dd 499static void __unregister_session(struct ceph_mds_client *mdsc,
42ce56e5 500 struct ceph_mds_session *s)
2f2dc053 501{
2600d2dd
SW		 502 dout("__unregister_session mds%d %p\n", s->s_mds, s);
503 BUG_ON(mdsc->sessions[s->s_mds] != s);
42ce56e5
SW		 504 mdsc->sessions[s->s_mds] = NULL;
505 ceph_con_close(&s->s_con);
506 ceph_put_mds_session(s);
86d8f67b 507 atomic_dec(&mdsc->num_sessions);
2f2dc053
SW		 508}
509
510/*
511 * drop session refs in request.
512 *
513 * should be last request ref, or hold mdsc->mutex
514 */
515static void put_request_session(struct ceph_mds_request *req)
516{
517 if (req->r_session) {
518 ceph_put_mds_session(req->r_session);
519 req->r_session = NULL;
520 }
521}
522
153c8e6b 523void ceph_mdsc_release_request(struct kref *kref)
2f2dc053 524{
153c8e6b
SW		 525 struct ceph_mds_request *req = container_of(kref,
526 struct ceph_mds_request,
527 r_kref);
54008399 528 destroy_reply_info(&req->r_reply_info);
153c8e6b
SW		 529 if (req->r_request)
530 ceph_msg_put(req->r_request);
54008399 531 if (req->r_reply)
153c8e6b 532 ceph_msg_put(req->r_reply);
153c8e6b 533 if (req->r_inode) {
41b02e1f 534 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
153c8e6b
SW		 535 iput(req->r_inode);
536 }
537 if (req->r_locked_dir)
41b02e1f 538 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
e96a650a 539 iput(req->r_target_inode);
153c8e6b
SW		 540 if (req->r_dentry)
541 dput(req->r_dentry);
844d87c3
SW		 542 if (req->r_old_dentry)
543 dput(req->r_old_dentry);
544 if (req->r_old_dentry_dir) {
41b02e1f
SW		 545 /*
546 * track (and drop pins for) r_old_dentry_dir
547 * separately, since r_old_dentry's d_parent may have
548 * changed between the dir mutex being dropped and
549 * this request being freed.
550 */
551 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
552 CEPH_CAP_PIN);
41b02e1f 553 iput(req->r_old_dentry_dir);
2f2dc053 554 }
153c8e6b
SW		 555 kfree(req->r_path1);
556 kfree(req->r_path2);
25e6bae3
YZ		 557 if (req->r_pagelist)
558 ceph_pagelist_release(req->r_pagelist);
153c8e6b 559 put_request_session(req);
37151668 560 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
153c8e6b 561 kfree(req);
2f2dc053
SW		 562}
563
564/*
565 * lookup session, bump ref if found.
566 *
567 * called under mdsc->mutex.
568 */
569static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
570 u64 tid)
571{
572 struct ceph_mds_request *req;
44ca18f2
SW		 573 struct rb_node *n = mdsc->request_tree.rb_node;
574
575 while (n) {
576 req = rb_entry(n, struct ceph_mds_request, r_node);
577 if (tid < req->r_tid)
578 n = n->rb_left;
579 else if (tid > req->r_tid)
580 n = n->rb_right;
581 else {
582 ceph_mdsc_get_request(req);
583 return req;
584 }
585 }
586 return NULL;
587}
588
589static void __insert_request(struct ceph_mds_client *mdsc,
590 struct ceph_mds_request *new)
591{
592 struct rb_node **p = &mdsc->request_tree.rb_node;
593 struct rb_node *parent = NULL;
594 struct ceph_mds_request *req = NULL;
595
596 while (*p) {
597 parent = *p;
598 req = rb_entry(parent, struct ceph_mds_request, r_node);
599 if (new->r_tid < req->r_tid)
600 p = &(*p)->rb_left;
601 else if (new->r_tid > req->r_tid)
602 p = &(*p)->rb_right;
603 else
604 BUG();
605 }
606
607 rb_link_node(&new->r_node, parent, p);
608 rb_insert_color(&new->r_node, &mdsc->request_tree);
2f2dc053
SW		 609}
610
611/*
612 * Register an in-flight request, and assign a tid. Link to directory
613 * are modifying (if any).
614 *
615 * Called under mdsc->mutex.
616 */
617static void __register_request(struct ceph_mds_client *mdsc,
618 struct ceph_mds_request *req,
619 struct inode *dir)
620{
621 req->r_tid = ++mdsc->last_tid;
622 if (req->r_num_caps)
37151668
YS		 623 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
624 req->r_num_caps);
2f2dc053
SW		 625 dout("__register_request %p tid %lld\n", req, req->r_tid);
626 ceph_mdsc_get_request(req);
44ca18f2 627 __insert_request(mdsc, req);
2f2dc053 628
cb4276cc
SW		 629 req->r_uid = current_fsuid();
630 req->r_gid = current_fsgid();
631
2f2dc053
SW		 632 if (dir) {
633 struct ceph_inode_info *ci = ceph_inode(dir);
634
3b663780 635 ihold(dir);
2f2dc053
SW		 636 spin_lock(&ci->i_unsafe_lock);
637 req->r_unsafe_dir = dir;
638 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
639 spin_unlock(&ci->i_unsafe_lock);
640 }
641}
642
643static void __unregister_request(struct ceph_mds_client *mdsc,
644 struct ceph_mds_request *req)
645{
646 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
44ca18f2 647 rb_erase(&req->r_node, &mdsc->request_tree);
80fc7314 648 RB_CLEAR_NODE(&req->r_node);
2f2dc053
SW		 649
650 if (req->r_unsafe_dir) {
651 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
652
653 spin_lock(&ci->i_unsafe_lock);
654 list_del_init(&req->r_unsafe_dir_item);
655 spin_unlock(&ci->i_unsafe_lock);
3b663780
SW		 656
657 iput(req->r_unsafe_dir);
658 req->r_unsafe_dir = NULL;
2f2dc053 659 }
94aa8ae1 660
fc55d2c9
YZ		 661 complete_all(&req->r_safe_completion);
662
94aa8ae1 663 ceph_mdsc_put_request(req);
2f2dc053
SW		 664}
665
666/*
667 * Choose mds to send request to next. If there is a hint set in the
668 * request (e.g., due to a prior forward hint from the mds), use that.
669 * Otherwise, consult frag tree and/or caps to identify the
670 * appropriate mds. If all else fails, choose randomly.
671 *
672 * Called under mdsc->mutex.
673 */
7fd7d101 674static struct dentry *get_nonsnap_parent(struct dentry *dentry)
eb6bb1c5 675{
d79698da
SW		 676 /*
677 * we don't need to worry about protecting the d_parent access
678 * here because we never renaming inside the snapped namespace
679 * except to resplice to another snapdir, and either the old or new
680 * result is a valid result.
681 */
eb6bb1c5
SW		 682 while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
683 dentry = dentry->d_parent;
684 return dentry;
685}
686
2f2dc053
SW		 687static int __choose_mds(struct ceph_mds_client *mdsc,
688 struct ceph_mds_request *req)
689{
690 struct inode *inode;
691 struct ceph_inode_info *ci;
692 struct ceph_cap *cap;
693 int mode = req->r_direct_mode;
694 int mds = -1;
695 u32 hash = req->r_direct_hash;
696 bool is_hash = req->r_direct_is_hash;
697
698 /*
699 * is there a specific mds we should try? ignore hint if we have
700 * no session and the mds is not up (active or recovering).
701 */
702 if (req->r_resend_mds >= 0 &&
703 (__have_session(mdsc, req->r_resend_mds) ||
704 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
705 dout("choose_mds using resend_mds mds%d\n",
706 req->r_resend_mds);
707 return req->r_resend_mds;
708 }
709
710 if (mode == USE_RANDOM_MDS)
711 goto random;
712
713 inode = NULL;
714 if (req->r_inode) {
715 inode = req->r_inode;
716 } else if (req->r_dentry) {
d79698da
SW		 717 /* ignore race with rename; old or new d_parent is okay */
718 struct dentry *parent = req->r_dentry->d_parent;
719 struct inode *dir = parent->d_inode;
eb6bb1c5 720
3d14c5d2 721 if (dir->i_sb != mdsc->fsc->sb) {
eb6bb1c5
SW		 722 /* not this fs! */
723 inode = req->r_dentry->d_inode;
724 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
725 /* direct snapped/virtual snapdir requests
726 * based on parent dir inode */
d79698da 727 struct dentry *dn = get_nonsnap_parent(parent);
eb6bb1c5
SW		 728 inode = dn->d_inode;
729 dout("__choose_mds using nonsnap parent %p\n", inode);
ca18bede 730 } else {
eb6bb1c5 731 /* dentry target */
2f2dc053 732 inode = req->r_dentry->d_inode;
ca18bede
YZ		 733 if (!inode || mode == USE_AUTH_MDS) {
734 /* dir + name */
735 inode = dir;
736 hash = ceph_dentry_hash(dir, req->r_dentry);
737 is_hash = true;
738 }
2f2dc053
SW		 739 }
740 }
eb6bb1c5 741
2f2dc053
SW		 742 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
743 (int)hash, mode);
744 if (!inode)
745 goto random;
746 ci = ceph_inode(inode);
747
748 if (is_hash && S_ISDIR(inode->i_mode)) {
749 struct ceph_inode_frag frag;
750 int found;
751
752 ceph_choose_frag(ci, hash, &frag, &found);
753 if (found) {
754 if (mode == USE_ANY_MDS && frag.ndist > 0) {
755 u8 r;
756
757 /* choose a random replica */
758 get_random_bytes(&r, 1);
759 r %= frag.ndist;
760 mds = frag.dist[r];
761 dout("choose_mds %p %llx.%llx "
762 "frag %u mds%d (%d/%d)\n",
763 inode, ceph_vinop(inode),
d66bbd44 764 frag.frag, mds,
2f2dc053 765 (int)r, frag.ndist);
d66bbd44
SW		 766 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
767 CEPH_MDS_STATE_ACTIVE)
768 return mds;
2f2dc053
SW		 769 }
770
771 /* since this file/dir wasn't known to be
772 * replicated, then we want to look for the
773 * authoritative mds. */
774 mode = USE_AUTH_MDS;
775 if (frag.mds >= 0) {
776 /* choose auth mds */
777 mds = frag.mds;
778 dout("choose_mds %p %llx.%llx "
779 "frag %u mds%d (auth)\n",
780 inode, ceph_vinop(inode), frag.frag, mds);
d66bbd44
SW		 781 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
782 CEPH_MDS_STATE_ACTIVE)
783 return mds;
2f2dc053
SW		 784 }
785 }
786 }
787
be655596 788 spin_lock(&ci->i_ceph_lock);
2f2dc053
SW		 789 cap = NULL;
790 if (mode == USE_AUTH_MDS)
791 cap = ci->i_auth_cap;
792 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
793 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
794 if (!cap) {
be655596 795 spin_unlock(&ci->i_ceph_lock);
2f2dc053
SW		 796 goto random;
797 }
798 mds = cap->session->s_mds;
799 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
800 inode, ceph_vinop(inode), mds,
801 cap == ci->i_auth_cap ? "auth " : "", cap);
be655596 802 spin_unlock(&ci->i_ceph_lock);
2f2dc053
SW		 803 return mds;
804
805random:
806 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
807 dout("choose_mds chose random mds%d\n", mds);
808 return mds;
809}
810
811
812/*
813 * session messages
814 */
815static struct ceph_msg *create_session_msg(u32 op, u64 seq)
816{
817 struct ceph_msg *msg;
818 struct ceph_mds_session_head *h;
819
b61c2763
SW		 820 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
821 false);
a79832f2 822 if (!msg) {
2f2dc053 823 pr_err("create_session_msg ENOMEM creating msg\n");
a79832f2 824 return NULL;
2f2dc053
SW		 825 }
826 h = msg->front.iov_base;
827 h->op = cpu_to_le32(op);
828 h->seq = cpu_to_le64(seq);
dbd0c8bf
JS		 829
830 return msg;
831}
832
833/*
834 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
835 * to include additional client metadata fields.
836 */
837static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
838{
839 struct ceph_msg *msg;
840 struct ceph_mds_session_head *h;
841 int i = -1;
842 int metadata_bytes = 0;
843 int metadata_key_count = 0;
844 struct ceph_options *opt = mdsc->fsc->client->options;
845 void *p;
846
a6a5ce4f 847 const char* metadata[][2] = {
dbd0c8bf 848 {"hostname", utsname()->nodename},
a6a5ce4f 849 {"kernel_version", utsname()->release},
dbd0c8bf
JS		 850 {"entity_id", opt->name ? opt->name : ""},
851 {NULL, NULL}
852 };
853
854 /* Calculate serialized length of metadata */
855 metadata_bytes = 4; /* map length */
856 for (i = 0; metadata[i][0] != NULL; ++i) {
857 metadata_bytes += 8 + strlen(metadata[i][0]) +
858 strlen(metadata[i][1]);
859 metadata_key_count++;
860 }
861
862 /* Allocate the message */
863 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
864 GFP_NOFS, false);
865 if (!msg) {
866 pr_err("create_session_msg ENOMEM creating msg\n");
867 return NULL;
868 }
869 h = msg->front.iov_base;
870 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
871 h->seq = cpu_to_le64(seq);
872
873 /*
874 * Serialize client metadata into waiting buffer space, using
875 * the format that userspace expects for map<string, string>
7cfa0313
JS		 876 *
877 * ClientSession messages with metadata are v2
dbd0c8bf 878 */
7cfa0313
JS		 879 msg->hdr.version = cpu_to_le16(2);
880 msg->hdr.compat_version = cpu_to_le16(1);
dbd0c8bf
JS		 881
882 /* The write pointer, following the session_head structure */
883 p = msg->front.iov_base + sizeof(*h);
884
885 /* Number of entries in the map */
886 ceph_encode_32(&p, metadata_key_count);
887
888 /* Two length-prefixed strings for each entry in the map */
889 for (i = 0; metadata[i][0] != NULL; ++i) {
890 size_t const key_len = strlen(metadata[i][0]);
891 size_t const val_len = strlen(metadata[i][1]);
892
893 ceph_encode_32(&p, key_len);
894 memcpy(p, metadata[i][0], key_len);
895 p += key_len;
896 ceph_encode_32(&p, val_len);
897 memcpy(p, metadata[i][1], val_len);
898 p += val_len;
899 }
900
2f2dc053
SW		 901 return msg;
902}
903
904/*
905 * send session open request.
906 *
907 * called under mdsc->mutex
908 */
909static int __open_session(struct ceph_mds_client *mdsc,
910 struct ceph_mds_session *session)
911{
912 struct ceph_msg *msg;
913 int mstate;
914 int mds = session->s_mds;
2f2dc053
SW		 915
916 /* wait for mds to go active? */
917 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
918 dout("open_session to mds%d (%s)\n", mds,
919 ceph_mds_state_name(mstate));
920 session->s_state = CEPH_MDS_SESSION_OPENING;
921 session->s_renew_requested = jiffies;
922
923 /* send connect message */
dbd0c8bf 924 msg = create_session_open_msg(mdsc, session->s_seq);
a79832f2
SW		 925 if (!msg)
926 return -ENOMEM;
2f2dc053 927 ceph_con_send(&session->s_con, msg);
2f2dc053
SW		 928 return 0;
929}
930
ed0552a1
SW		 931/*
932 * open sessions for any export targets for the given mds
933 *
934 * called under mdsc->mutex
935 */
5d72d13c
YZ		 936static struct ceph_mds_session *
937__open_export_target_session(struct ceph_mds_client *mdsc, int target)
938{
939 struct ceph_mds_session *session;
940
941 session = __ceph_lookup_mds_session(mdsc, target);
942 if (!session) {
943 session = register_session(mdsc, target);
944 if (IS_ERR(session))
945 return session;
946 }
947 if (session->s_state == CEPH_MDS_SESSION_NEW ||
948 session->s_state == CEPH_MDS_SESSION_CLOSING)
949 __open_session(mdsc, session);
950
951 return session;
952}
953
954struct ceph_mds_session *
955ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
956{
957 struct ceph_mds_session *session;
958
959 dout("open_export_target_session to mds%d\n", target);
960
961 mutex_lock(&mdsc->mutex);
962 session = __open_export_target_session(mdsc, target);
963 mutex_unlock(&mdsc->mutex);
964
965 return session;
966}
967
ed0552a1
SW		 968static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
969 struct ceph_mds_session *session)
970{
971 struct ceph_mds_info *mi;
972 struct ceph_mds_session *ts;
973 int i, mds = session->s_mds;
ed0552a1
SW		 974
975 if (mds >= mdsc->mdsmap->m_max_mds)
976 return;
5d72d13c 977
ed0552a1
SW		 978 mi = &mdsc->mdsmap->m_info[mds];
979 dout("open_export_target_sessions for mds%d (%d targets)\n",
980 session->s_mds, mi->num_export_targets);
981
982 for (i = 0; i < mi->num_export_targets; i++) {
5d72d13c
YZ		 983 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
984 if (!IS_ERR(ts))
985 ceph_put_mds_session(ts);
ed0552a1
SW		 986 }
987}
988
154f42c2
SW		 989void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
990 struct ceph_mds_session *session)
991{
992 mutex_lock(&mdsc->mutex);
993 __open_export_target_sessions(mdsc, session);
994 mutex_unlock(&mdsc->mutex);
995}
996
2f2dc053
SW		 997/*
998 * session caps
999 */
1000
1001/*
1002 * Free preallocated cap messages assigned to this session
1003 */
1004static void cleanup_cap_releases(struct ceph_mds_session *session)
1005{
1006 struct ceph_msg *msg;
1007
1008 spin_lock(&session->s_cap_lock);
1009 while (!list_empty(&session->s_cap_releases)) {
1010 msg = list_first_entry(&session->s_cap_releases,
1011 struct ceph_msg, list_head);
1012 list_del_init(&msg->list_head);
1013 ceph_msg_put(msg);
1014 }
1015 while (!list_empty(&session->s_cap_releases_done)) {
1016 msg = list_first_entry(&session->s_cap_releases_done,
1017 struct ceph_msg, list_head);
1018 list_del_init(&msg->list_head);
1019 ceph_msg_put(msg);
1020 }
1021 spin_unlock(&session->s_cap_lock);
1022}
1023
1c841a96
YZ		 1024static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1025 struct ceph_mds_session *session)
1026{
1027 struct ceph_mds_request *req;
1028 struct rb_node *p;
1029
1030 dout("cleanup_session_requests mds%d\n", session->s_mds);
1031 mutex_lock(&mdsc->mutex);
1032 while (!list_empty(&session->s_unsafe)) {
1033 req = list_first_entry(&session->s_unsafe,
1034 struct ceph_mds_request, r_unsafe_item);
1035 list_del_init(&req->r_unsafe_item);
1036 pr_info(" dropping unsafe request %llu\n", req->r_tid);
1037 __unregister_request(mdsc, req);
1038 }
1039 /* zero r_attempts, so kick_requests() will re-send requests */
1040 p = rb_first(&mdsc->request_tree);
1041 while (p) {
1042 req = rb_entry(p, struct ceph_mds_request, r_node);
1043 p = rb_next(p);
1044 if (req->r_session &&
1045 req->r_session->s_mds == session->s_mds)
1046 req->r_attempts = 0;
1047 }
1048 mutex_unlock(&mdsc->mutex);
1049}
1050
2f2dc053 1051/*
f818a736
SW		 1052 * Helper to safely iterate over all caps associated with a session, with
1053 * special care taken to handle a racing __ceph_remove_cap().
2f2dc053 1054 *
f818a736 1055 * Caller must hold session s_mutex.
2f2dc053
SW		 1056 */
1057static int iterate_session_caps(struct ceph_mds_session *session,
1058 int (*cb)(struct inode *, struct ceph_cap *,
1059 void *), void *arg)
1060{
7c1332b8
SW		 1061 struct list_head *p;
1062 struct ceph_cap *cap;
1063 struct inode *inode, *last_inode = NULL;
1064 struct ceph_cap *old_cap = NULL;
2f2dc053
SW		 1065 int ret;
1066
1067 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1068 spin_lock(&session->s_cap_lock);
7c1332b8
SW		 1069 p = session->s_caps.next;
1070 while (p != &session->s_caps) {
1071 cap = list_entry(p, struct ceph_cap, session_caps);
2f2dc053 1072 inode = igrab(&cap->ci->vfs_inode);
7c1332b8
SW		 1073 if (!inode) {
1074 p = p->next;
2f2dc053 1075 continue;
7c1332b8
SW		 1076 }
1077 session->s_cap_iterator = cap;
2f2dc053 1078 spin_unlock(&session->s_cap_lock);
7c1332b8
SW		 1079
1080 if (last_inode) {
1081 iput(last_inode);
1082 last_inode = NULL;
1083 }
1084 if (old_cap) {
37151668 1085 ceph_put_cap(session->s_mdsc, old_cap);
7c1332b8
SW		 1086 old_cap = NULL;
1087 }
1088
2f2dc053 1089 ret = cb(inode, cap, arg);
7c1332b8
SW		 1090 last_inode = inode;
1091
2f2dc053 1092 spin_lock(&session->s_cap_lock);
7c1332b8
SW		 1093 p = p->next;
1094 if (cap->ci == NULL) {
1095 dout("iterate_session_caps finishing cap %p removal\n",
1096 cap);
1097 BUG_ON(cap->session != session);
1098 list_del_init(&cap->session_caps);
1099 session->s_nr_caps--;
1100 cap->session = NULL;
1101 old_cap = cap; /* put_cap it w/o locks held */
1102 }
5dacf091
SW		 1103 if (ret < 0)
1104 goto out;
2f2dc053 1105 }
5dacf091
SW		 1106 ret = 0;
1107out:
7c1332b8 1108 session->s_cap_iterator = NULL;
2f2dc053 1109 spin_unlock(&session->s_cap_lock);
7c1332b8 1110
e96a650a 1111 iput(last_inode);
7c1332b8 1112 if (old_cap)
37151668 1113 ceph_put_cap(session->s_mdsc, old_cap);
7c1332b8 1114
5dacf091 1115 return ret;
2f2dc053
SW		 1116}
1117
1118static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
6c99f254 1119 void *arg)
2f2dc053
SW		 1120{
1121 struct ceph_inode_info *ci = ceph_inode(inode);
6c99f254
SW		 1122 int drop = 0;
1123
2f2dc053
SW		 1124 dout("removing cap %p, ci is %p, inode is %p\n",
1125 cap, ci, &ci->vfs_inode);
be655596 1126 spin_lock(&ci->i_ceph_lock);
a096b09a 1127 __ceph_remove_cap(cap, false);
571ade33 1128 if (!ci->i_auth_cap) {
6c99f254 1129 struct ceph_mds_client *mdsc =
3d14c5d2 1130 ceph_sb_to_client(inode->i_sb)->mdsc;
6c99f254
SW		 1131
1132 spin_lock(&mdsc->cap_dirty_lock);
1133 if (!list_empty(&ci->i_dirty_item)) {
1134 pr_info(" dropping dirty %s state for %p %lld\n",
1135 ceph_cap_string(ci->i_dirty_caps),
1136 inode, ceph_ino(inode));
1137 ci->i_dirty_caps = 0;
1138 list_del_init(&ci->i_dirty_item);
1139 drop = 1;
1140 }
1141 if (!list_empty(&ci->i_flushing_item)) {
1142 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1143 ceph_cap_string(ci->i_flushing_caps),
1144 inode, ceph_ino(inode));
1145 ci->i_flushing_caps = 0;
1146 list_del_init(&ci->i_flushing_item);
1147 mdsc->num_cap_flushing--;
1148 drop = 1;
1149 }
6c99f254
SW		 1150 spin_unlock(&mdsc->cap_dirty_lock);
1151 }
be655596 1152 spin_unlock(&ci->i_ceph_lock);
6c99f254
SW		 1153 while (drop--)
1154 iput(inode);
2f2dc053
SW		 1155 return 0;
1156}
1157
1158/*
1159 * caller must hold session s_mutex
1160 */
1161static void remove_session_caps(struct ceph_mds_session *session)
1162{
1163 dout("remove_session_caps on %p\n", session);
1164 iterate_session_caps(session, remove_session_caps_cb, NULL);
6f60f889
YZ		 1165
1166 spin_lock(&session->s_cap_lock);
1167 if (session->s_nr_caps > 0) {
1168 struct super_block *sb = session->s_mdsc->fsc->sb;
1169 struct inode *inode;
1170 struct ceph_cap *cap, *prev = NULL;
1171 struct ceph_vino vino;
1172 /*
1173 * iterate_session_caps() skips inodes that are being
1174 * deleted, we need to wait until deletions are complete.
1175 * __wait_on_freeing_inode() is designed for the job,
1176 * but it is not exported, so use lookup inode function
1177 * to access it.
1178 */
1179 while (!list_empty(&session->s_caps)) {
1180 cap = list_entry(session->s_caps.next,
1181 struct ceph_cap, session_caps);
1182 if (cap == prev)
1183 break;
1184 prev = cap;
1185 vino = cap->ci->i_vino;
1186 spin_unlock(&session->s_cap_lock);
1187
ed284c49 1188 inode = ceph_find_inode(sb, vino);
6f60f889
YZ		 1189 iput(inode);
1190
1191 spin_lock(&session->s_cap_lock);
1192 }
1193 }
1194 spin_unlock(&session->s_cap_lock);
1195
2f2dc053 1196 BUG_ON(session->s_nr_caps > 0);
6c99f254 1197 BUG_ON(!list_empty(&session->s_cap_flushing));
2f2dc053
SW		 1198 cleanup_cap_releases(session);
1199}
1200
1201/*
1202 * wake up any threads waiting on this session's caps. if the cap is
1203 * old (didn't get renewed on the client reconnect), remove it now.
1204 *
1205 * caller must hold s_mutex.
1206 */
1207static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1208 void *arg)
1209{
0dc2570f
SW		 1210 struct ceph_inode_info *ci = ceph_inode(inode);
1211
03066f23 1212 wake_up_all(&ci->i_cap_wq);
0dc2570f 1213 if (arg) {
be655596 1214 spin_lock(&ci->i_ceph_lock);
0dc2570f
SW		 1215 ci->i_wanted_max_size = 0;
1216 ci->i_requested_max_size = 0;
be655596 1217 spin_unlock(&ci->i_ceph_lock);
0dc2570f 1218 }
2f2dc053
SW		 1219 return 0;
1220}
1221
0dc2570f
SW		 1222static void wake_up_session_caps(struct ceph_mds_session *session,
1223 int reconnect)
2f2dc053
SW		 1224{
1225 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
0dc2570f
SW		 1226 iterate_session_caps(session, wake_up_session_cb,
1227 (void *)(unsigned long)reconnect);
2f2dc053
SW		 1228}
1229
1230/*
1231 * Send periodic message to MDS renewing all currently held caps. The
1232 * ack will reset the expiration for all caps from this session.
1233 *
1234 * caller holds s_mutex
1235 */
1236static int send_renew_caps(struct ceph_mds_client *mdsc,
1237 struct ceph_mds_session *session)
1238{
1239 struct ceph_msg *msg;
1240 int state;
1241
1242 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1243 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1244 pr_info("mds%d caps stale\n", session->s_mds);
e4cb4cb8 1245 session->s_renew_requested = jiffies;
2f2dc053
SW		 1246
1247 /* do not try to renew caps until a recovering mds has reconnected
1248 * with its clients. */
1249 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1250 if (state < CEPH_MDS_STATE_RECONNECT) {
1251 dout("send_renew_caps ignoring mds%d (%s)\n",
1252 session->s_mds, ceph_mds_state_name(state));
1253 return 0;
1254 }
1255
1256 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1257 ceph_mds_state_name(state));
2f2dc053
SW		 1258 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1259 ++session->s_renew_seq);
a79832f2
SW		 1260 if (!msg)
1261 return -ENOMEM;
2f2dc053
SW		 1262 ceph_con_send(&session->s_con, msg);
1263 return 0;
1264}
1265
186e4f7a
YZ		 1266static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1267 struct ceph_mds_session *session, u64 seq)
1268{
1269 struct ceph_msg *msg;
1270
1271 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
a687ecaf 1272 session->s_mds, ceph_session_state_name(session->s_state), seq);
186e4f7a
YZ		 1273 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1274 if (!msg)
1275 return -ENOMEM;
1276 ceph_con_send(&session->s_con, msg);
1277 return 0;
1278}
1279
1280
2f2dc053
SW		 1281/*
1282 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
0dc2570f
SW		 1283 *
1284 * Called under session->s_mutex
2f2dc053
SW		 1285 */
1286static void renewed_caps(struct ceph_mds_client *mdsc,
1287 struct ceph_mds_session *session, int is_renew)
1288{
1289 int was_stale;
1290 int wake = 0;
1291
1292 spin_lock(&session->s_cap_lock);
1ce208a6 1293 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
2f2dc053
SW		 1294
1295 session->s_cap_ttl = session->s_renew_requested +
1296 mdsc->mdsmap->m_session_timeout*HZ;
1297
1298 if (was_stale) {
1299 if (time_before(jiffies, session->s_cap_ttl)) {
1300 pr_info("mds%d caps renewed\n", session->s_mds);
1301 wake = 1;
1302 } else {
1303 pr_info("mds%d caps still stale\n", session->s_mds);
1304 }
1305 }
1306 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1307 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1308 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1309 spin_unlock(&session->s_cap_lock);
1310
1311 if (wake)
0dc2570f 1312 wake_up_session_caps(session, 0);
2f2dc053
SW		 1313}
1314
1315/*
1316 * send a session close request
1317 */
1318static int request_close_session(struct ceph_mds_client *mdsc,
1319 struct ceph_mds_session *session)
1320{
1321 struct ceph_msg *msg;
2f2dc053
SW		 1322
1323 dout("request_close_session mds%d state %s seq %lld\n",
a687ecaf 1324 session->s_mds, ceph_session_state_name(session->s_state),
2f2dc053
SW		 1325 session->s_seq);
1326 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
a79832f2
SW		 1327 if (!msg)
1328 return -ENOMEM;
1329 ceph_con_send(&session->s_con, msg);
1330 return 0;
2f2dc053
SW		 1331}
1332
1333/*
1334 * Called with s_mutex held.
1335 */
1336static int __close_session(struct ceph_mds_client *mdsc,
1337 struct ceph_mds_session *session)
1338{
1339 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1340 return 0;
1341 session->s_state = CEPH_MDS_SESSION_CLOSING;
1342 return request_close_session(mdsc, session);
1343}
1344
1345/*
1346 * Trim old(er) caps.
1347 *
1348 * Because we can't cache an inode without one or more caps, we do
1349 * this indirectly: if a cap is unused, we prune its aliases, at which
1350 * point the inode will hopefully get dropped to.
1351 *
1352 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1353 * memory pressure from the MDS, though, so it needn't be perfect.
1354 */
1355static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1356{
1357 struct ceph_mds_session *session = arg;
1358 struct ceph_inode_info *ci = ceph_inode(inode);
979abfdd 1359 int used, wanted, oissued, mine;
2f2dc053
SW		 1360
1361 if (session->s_trim_caps <= 0)
1362 return -1;
1363
be655596 1364 spin_lock(&ci->i_ceph_lock);
2f2dc053
SW		 1365 mine = cap->issued | cap->implemented;
1366 used = __ceph_caps_used(ci);
979abfdd 1367 wanted = __ceph_caps_file_wanted(ci);
2f2dc053
SW		 1368 oissued = __ceph_caps_issued_other(ci, cap);
1369
979abfdd 1370 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
2f2dc053 1371 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
979abfdd
YZ		 1372 ceph_cap_string(used), ceph_cap_string(wanted));
1373 if (cap == ci->i_auth_cap) {
1374 if (ci->i_dirty_caps | ci->i_flushing_caps)
1375 goto out;
1376 if ((used | wanted) & CEPH_CAP_ANY_WR)
1377 goto out;
1378 }
1379 if ((used | wanted) & ~oissued & mine)
2f2dc053
SW		 1380 goto out; /* we need these caps */
1381
1382 session->s_trim_caps--;
1383 if (oissued) {
1384 /* we aren't the only cap.. just remove us */
a096b09a 1385 __ceph_remove_cap(cap, true);
2f2dc053
SW		 1386 } else {
1387 /* try to drop referring dentries */
be655596 1388 spin_unlock(&ci->i_ceph_lock);
2f2dc053
SW		 1389 d_prune_aliases(inode);
1390 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1391 inode, cap, atomic_read(&inode->i_count));
1392 return 0;
1393 }
1394
1395out:
be655596 1396 spin_unlock(&ci->i_ceph_lock);
2f2dc053
SW		 1397 return 0;
1398}
1399
1400/*
1401 * Trim session cap count down to some max number.
1402 */
1403static int trim_caps(struct ceph_mds_client *mdsc,
1404 struct ceph_mds_session *session,
1405 int max_caps)
1406{
1407 int trim_caps = session->s_nr_caps - max_caps;
1408
1409 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1410 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1411 if (trim_caps > 0) {
1412 session->s_trim_caps = trim_caps;
1413 iterate_session_caps(session, trim_caps_cb, session);
1414 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1415 session->s_mds, session->s_nr_caps, max_caps,
1416 trim_caps - session->s_trim_caps);
5dacf091 1417 session->s_trim_caps = 0;
2f2dc053 1418 }
a56371d9
YZ		 1419
1420 ceph_add_cap_releases(mdsc, session);
1421 ceph_send_cap_releases(mdsc, session);
2f2dc053
SW		 1422 return 0;
1423}
1424
1425/*
1426 * Allocate cap_release messages. If there is a partially full message
1427 * in the queue, try to allocate enough to cover it's remainder, so that
1428 * we can send it immediately.
1429 *
1430 * Called under s_mutex.
1431 */
2b2300d6 1432int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
ee6b272b 1433 struct ceph_mds_session *session)
2f2dc053 1434{
38e8883e 1435 struct ceph_msg *msg, *partial = NULL;
2f2dc053
SW		 1436 struct ceph_mds_cap_release *head;
1437 int err = -ENOMEM;
3d14c5d2 1438 int extra = mdsc->fsc->mount_options->cap_release_safety;
38e8883e 1439 int num;
2f2dc053 1440
38e8883e
SW		 1441 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1442 extra);
2f2dc053
SW		 1443
1444 spin_lock(&session->s_cap_lock);
1445
1446 if (!list_empty(&session->s_cap_releases)) {
1447 msg = list_first_entry(&session->s_cap_releases,
1448 struct ceph_msg,
1449 list_head);
1450 head = msg->front.iov_base;
38e8883e
SW		 1451 num = le32_to_cpu(head->num);
1452 if (num) {
1453 dout(" partial %p with (%d/%d)\n", msg, num,
1454 (int)CEPH_CAPS_PER_RELEASE);
1455 extra += CEPH_CAPS_PER_RELEASE - num;
1456 partial = msg;
1457 }
2f2dc053 1458 }
2f2dc053
SW		 1459 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1460 spin_unlock(&session->s_cap_lock);
34d23762 1461 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
b61c2763 1462 GFP_NOFS, false);
2f2dc053
SW		 1463 if (!msg)
1464 goto out_unlocked;
1465 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1466 (int)msg->front.iov_len);
1467 head = msg->front.iov_base;
1468 head->num = cpu_to_le32(0);
1469 msg->front.iov_len = sizeof(*head);
1470 spin_lock(&session->s_cap_lock);
1471 list_add(&msg->list_head, &session->s_cap_releases);
1472 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1473 }
1474
38e8883e
SW		 1475 if (partial) {
1476 head = partial->front.iov_base;
1477 num = le32_to_cpu(head->num);
1478 dout(" queueing partial %p with %d/%d\n", partial, num,
1479 (int)CEPH_CAPS_PER_RELEASE);
1480 list_move_tail(&partial->list_head,
1481 &session->s_cap_releases_done);
1482 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
2f2dc053
SW		 1483 }
1484 err = 0;
1485 spin_unlock(&session->s_cap_lock);
1486out_unlocked:
1487 return err;
1488}
1489
d3383a8e
YZ		 1490static int check_cap_flush(struct inode *inode, u64 want_flush_seq)
1491{
1492 struct ceph_inode_info *ci = ceph_inode(inode);
1493 int ret;
1494 spin_lock(&ci->i_ceph_lock);
1495 if (ci->i_flushing_caps)
1496 ret = ci->i_cap_flush_seq >= want_flush_seq;
1497 else
1498 ret = 1;
1499 spin_unlock(&ci->i_ceph_lock);
1500 return ret;
1501}
1502
2f2dc053
SW		 1503/*
1504 * flush all dirty inode data to disk.
1505 *
1506 * returns true if we've flushed through want_flush_seq
1507 */
d3383a8e 1508static void wait_caps_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
2f2dc053 1509{
d3383a8e 1510 int mds;
2f2dc053
SW		 1511
1512 dout("check_cap_flush want %lld\n", want_flush_seq);
1513 mutex_lock(&mdsc->mutex);
d3383a8e 1514 for (mds = 0; mds < mdsc->max_sessions; mds++) {
2f2dc053 1515 struct ceph_mds_session *session = mdsc->sessions[mds];
d3383a8e 1516 struct inode *inode = NULL;
2f2dc053
SW		 1517
1518 if (!session)
1519 continue;
1520 get_session(session);
1521 mutex_unlock(&mdsc->mutex);
1522
1523 mutex_lock(&session->s_mutex);
1524 if (!list_empty(&session->s_cap_flushing)) {
1525 struct ceph_inode_info *ci =
1526 list_entry(session->s_cap_flushing.next,
1527 struct ceph_inode_info,
1528 i_flushing_item);
2f2dc053 1529
d3383a8e 1530 if (!check_cap_flush(&ci->vfs_inode, want_flush_seq)) {
2f2dc053 1531 dout("check_cap_flush still flushing %p "
d3383a8e
YZ		 1532 "seq %lld <= %lld to mds%d\n",
1533 &ci->vfs_inode, ci->i_cap_flush_seq,
1534 want_flush_seq, session->s_mds);
1535 inode = igrab(&ci->vfs_inode);
2f2dc053 1536 }
2f2dc053
SW		 1537 }
1538 mutex_unlock(&session->s_mutex);
1539 ceph_put_mds_session(session);
1540
d3383a8e
YZ		 1541 if (inode) {
1542 wait_event(mdsc->cap_flushing_wq,
1543 check_cap_flush(inode, want_flush_seq));
1544 iput(inode);
1545 }
1546
2f2dc053
SW		 1547 mutex_lock(&mdsc->mutex);
1548 }
1549
1550 mutex_unlock(&mdsc->mutex);
1551 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
2f2dc053
SW		 1552}
1553
1554/*
1555 * called under s_mutex
1556 */
3d7ded4d
SW		 1557void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1558 struct ceph_mds_session *session)
2f2dc053
SW		 1559{
1560 struct ceph_msg *msg;
1561
1562 dout("send_cap_releases mds%d\n", session->s_mds);
0f8605f2
SW		 1563 spin_lock(&session->s_cap_lock);
1564 while (!list_empty(&session->s_cap_releases_done)) {
2f2dc053
SW		 1565 msg = list_first_entry(&session->s_cap_releases_done,
1566 struct ceph_msg, list_head);
1567 list_del_init(&msg->list_head);
1568 spin_unlock(&session->s_cap_lock);
1569 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1570 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1571 ceph_con_send(&session->s_con, msg);
0f8605f2 1572 spin_lock(&session->s_cap_lock);
2f2dc053
SW		 1573 }
1574 spin_unlock(&session->s_cap_lock);
1575}
1576
e01a5946
SW		 1577static void discard_cap_releases(struct ceph_mds_client *mdsc,
1578 struct ceph_mds_session *session)
1579{
1580 struct ceph_msg *msg;
1581 struct ceph_mds_cap_release *head;
1582 unsigned num;
1583
1584 dout("discard_cap_releases mds%d\n", session->s_mds);
e01a5946 1585
00bd8edb
YZ		 1586 if (!list_empty(&session->s_cap_releases)) {
1587 /* zero out the in-progress message */
1588 msg = list_first_entry(&session->s_cap_releases,
1589 struct ceph_msg, list_head);
1590 head = msg->front.iov_base;
1591 num = le32_to_cpu(head->num);
1592 dout("discard_cap_releases mds%d %p %u\n",
1593 session->s_mds, msg, num);
1594 head->num = cpu_to_le32(0);
1595 msg->front.iov_len = sizeof(*head);
1596 session->s_num_cap_releases += num;
1597 }
e01a5946
SW		 1598
1599 /* requeue completed messages */
1600 while (!list_empty(&session->s_cap_releases_done)) {
1601 msg = list_first_entry(&session->s_cap_releases_done,
1602 struct ceph_msg, list_head);
1603 list_del_init(&msg->list_head);
1604
1605 head = msg->front.iov_base;
1606 num = le32_to_cpu(head->num);
1607 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1608 num);
1609 session->s_num_cap_releases += num;
1610 head->num = cpu_to_le32(0);
1611 msg->front.iov_len = sizeof(*head);
1612 list_add(&msg->list_head, &session->s_cap_releases);
1613 }
e01a5946
SW		 1614}
1615
2f2dc053
SW		 1616/*
1617 * requests
1618 */
1619
54008399
YZ		 1620int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1621 struct inode *dir)
1622{
1623 struct ceph_inode_info *ci = ceph_inode(dir);
1624 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1625 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1626 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1627 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1628 int order, num_entries;
1629
1630 spin_lock(&ci->i_ceph_lock);
1631 num_entries = ci->i_files + ci->i_subdirs;
1632 spin_unlock(&ci->i_ceph_lock);
1633 num_entries = max(num_entries, 1);
1634 num_entries = min(num_entries, opt->max_readdir);
1635
1636 order = get_order(size * num_entries);
1637 while (order >= 0) {
1638 rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN,
1639 order);
1640 if (rinfo->dir_in)
1641 break;
1642 order--;
1643 }
1644 if (!rinfo->dir_in)
1645 return -ENOMEM;
1646
1647 num_entries = (PAGE_SIZE << order) / size;
1648 num_entries = min(num_entries, opt->max_readdir);
1649
1650 rinfo->dir_buf_size = PAGE_SIZE << order;
1651 req->r_num_caps = num_entries + 1;
1652 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1653 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1654 return 0;
1655}
1656
2f2dc053
SW		 1657/*
1658 * Create an mds request.
1659 */
1660struct ceph_mds_request *
1661ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1662{
1663 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1664
1665 if (!req)
1666 return ERR_PTR(-ENOMEM);
1667
b4556396 1668 mutex_init(&req->r_fill_mutex);
37151668 1669 req->r_mdsc = mdsc;
2f2dc053
SW		 1670 req->r_started = jiffies;
1671 req->r_resend_mds = -1;
1672 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1673 req->r_fmode = -1;
153c8e6b 1674 kref_init(&req->r_kref);
2f2dc053
SW		 1675 INIT_LIST_HEAD(&req->r_wait);
1676 init_completion(&req->r_completion);
1677 init_completion(&req->r_safe_completion);
1678 INIT_LIST_HEAD(&req->r_unsafe_item);
1679
b8e69066
SW		 1680 req->r_stamp = CURRENT_TIME;
1681
2f2dc053
SW		 1682 req->r_op = op;
1683 req->r_direct_mode = mode;
1684 return req;
1685}
1686
1687/*
44ca18f2 1688 * return oldest (lowest) request, tid in request tree, 0 if none.
2f2dc053
SW		 1689 *
1690 * called under mdsc->mutex.
1691 */
44ca18f2
SW		 1692static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1693{
1694 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1695 return NULL;
1696 return rb_entry(rb_first(&mdsc->request_tree),
1697 struct ceph_mds_request, r_node);
1698}
1699
2f2dc053
SW		 1700static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1701{
44ca18f2
SW		 1702 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1703
1704 if (req)
1705 return req->r_tid;
1706 return 0;
2f2dc053
SW		 1707}
1708
1709/*
1710 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1711 * on build_path_from_dentry in fs/cifs/dir.c.
1712 *
1713 * If @stop_on_nosnap, generate path relative to the first non-snapped
1714 * inode.
1715 *
1716 * Encode hidden .snap dirs as a double /, i.e.
1717 * foo/.snap/bar -> foo//bar
1718 */
1719char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1720 int stop_on_nosnap)
1721{
1722 struct dentry *temp;
1723 char *path;
1724 int len, pos;
1b71fe2e 1725 unsigned seq;
2f2dc053
SW		 1726
1727 if (dentry == NULL)
1728 return ERR_PTR(-EINVAL);
1729
1730retry:
1731 len = 0;
1b71fe2e
AV		 1732 seq = read_seqbegin(&rename_lock);
1733 rcu_read_lock();
2f2dc053
SW		 1734 for (temp = dentry; !IS_ROOT(temp);) {
1735 struct inode *inode = temp->d_inode;
1736 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1737 len++; /* slash only */
1738 else if (stop_on_nosnap && inode &&
1739 ceph_snap(inode) == CEPH_NOSNAP)
1740 break;
1741 else
1742 len += 1 + temp->d_name.len;
1743 temp = temp->d_parent;
2f2dc053 1744 }
1b71fe2e 1745 rcu_read_unlock();
2f2dc053
SW		 1746 if (len)
1747 len--; /* no leading '/' */
1748
1749 path = kmalloc(len+1, GFP_NOFS);
1750 if (path == NULL)
1751 return ERR_PTR(-ENOMEM);
1752 pos = len;
1753 path[pos] = 0; /* trailing null */
1b71fe2e 1754 rcu_read_lock();
2f2dc053 1755 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1b71fe2e 1756 struct inode *inode;
2f2dc053 1757
1b71fe2e
AV		 1758 spin_lock(&temp->d_lock);
1759 inode = temp->d_inode;
2f2dc053 1760 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
104648ad 1761 dout("build_path path+%d: %p SNAPDIR\n",
2f2dc053
SW		 1762 pos, temp);
1763 } else if (stop_on_nosnap && inode &&
1764 ceph_snap(inode) == CEPH_NOSNAP) {
9d5a09e6 1765 spin_unlock(&temp->d_lock);
2f2dc053
SW		 1766 break;
1767 } else {
1768 pos -= temp->d_name.len;
1b71fe2e
AV		 1769 if (pos < 0) {
1770 spin_unlock(&temp->d_lock);
2f2dc053 1771 break;
1b71fe2e 1772 }
2f2dc053
SW		 1773 strncpy(path + pos, temp->d_name.name,
1774 temp->d_name.len);
2f2dc053 1775 }
1b71fe2e 1776 spin_unlock(&temp->d_lock);
2f2dc053
SW		 1777 if (pos)
1778 path[--pos] = '/';
1779 temp = temp->d_parent;
2f2dc053 1780 }
1b71fe2e
AV		 1781 rcu_read_unlock();
1782 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
104648ad 1783 pr_err("build_path did not end path lookup where "
2f2dc053
SW		 1784 "expected, namelen is %d, pos is %d\n", len, pos);
1785 /* presumably this is only possible if racing with a
1786 rename of one of the parent directories (we can not
1787 lock the dentries above us to prevent this, but
1788 retrying should be harmless) */
1789 kfree(path);
1790 goto retry;
1791 }
1792
1793 *base = ceph_ino(temp->d_inode);
1794 *plen = len;
104648ad 1795 dout("build_path on %p %d built %llx '%.*s'\n",
84d08fa8 1796 dentry, d_count(dentry), *base, len, path);
2f2dc053
SW		 1797 return path;
1798}
1799
1800static int build_dentry_path(struct dentry *dentry,
1801 const char **ppath, int *ppathlen, u64 *pino,
1802 int *pfreepath)
1803{
1804 char *path;
1805
1806 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1807 *pino = ceph_ino(dentry->d_parent->d_inode);
1808 *ppath = dentry->d_name.name;
1809 *ppathlen = dentry->d_name.len;
1810 return 0;
1811 }
1812 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1813 if (IS_ERR(path))
1814 return PTR_ERR(path);
1815 *ppath = path;
1816 *pfreepath = 1;
1817 return 0;
1818}
1819
1820static int build_inode_path(struct inode *inode,
1821 const char **ppath, int *ppathlen, u64 *pino,
1822 int *pfreepath)
1823{
1824 struct dentry *dentry;
1825 char *path;
1826
1827 if (ceph_snap(inode) == CEPH_NOSNAP) {
1828 *pino = ceph_ino(inode);
1829 *ppathlen = 0;
1830 return 0;
1831 }
1832 dentry = d_find_alias(inode);
1833 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1834 dput(dentry);
1835 if (IS_ERR(path))
1836 return PTR_ERR(path);
1837 *ppath = path;
1838 *pfreepath = 1;
1839 return 0;
1840}
1841
1842/*
1843 * request arguments may be specified via an inode *, a dentry *, or
1844 * an explicit ino+path.
1845 */
1846static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1847 const char *rpath, u64 rino,
1848 const char **ppath, int *pathlen,
1849 u64 *ino, int *freepath)
1850{
1851 int r = 0;
1852
1853 if (rinode) {
1854 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1855 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1856 ceph_snap(rinode));
1857 } else if (rdentry) {
1858 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1859 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1860 *ppath);
795858db 1861 } else if (rpath || rino) {
2f2dc053
SW		 1862 *ino = rino;
1863 *ppath = rpath;
b000056a 1864 *pathlen = rpath ? strlen(rpath) : 0;
2f2dc053
SW		 1865 dout(" path %.*s\n", *pathlen, rpath);
1866 }
1867
1868 return r;
1869}
1870
1871/*
1872 * called under mdsc->mutex
1873 */
1874static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1875 struct ceph_mds_request *req,
6e6f0923 1876 int mds, bool drop_cap_releases)
2f2dc053
SW		 1877{
1878 struct ceph_msg *msg;
1879 struct ceph_mds_request_head *head;
1880 const char *path1 = NULL;
1881 const char *path2 = NULL;
1882 u64 ino1 = 0, ino2 = 0;
1883 int pathlen1 = 0, pathlen2 = 0;
1884 int freepath1 = 0, freepath2 = 0;
1885 int len;
1886 u16 releases;
1887 void *p, *end;
1888 int ret;
1889
1890 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1891 req->r_path1, req->r_ino1.ino,
1892 &path1, &pathlen1, &ino1, &freepath1);
1893 if (ret < 0) {
1894 msg = ERR_PTR(ret);
1895 goto out;
1896 }
1897
1898 ret = set_request_path_attr(NULL, req->r_old_dentry,
1899 req->r_path2, req->r_ino2.ino,
1900 &path2, &pathlen2, &ino2, &freepath2);
1901 if (ret < 0) {
1902 msg = ERR_PTR(ret);
1903 goto out_free1;
1904 }
1905
1906 len = sizeof(*head) +
b8e69066
SW		 1907 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1908 sizeof(struct timespec);
2f2dc053
SW		 1909
1910 /* calculate (max) length for cap releases */
1911 len += sizeof(struct ceph_mds_request_release) *
1912 (!!req->r_inode_drop + !!req->r_dentry_drop +
1913 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1914 if (req->r_dentry_drop)
1915 len += req->r_dentry->d_name.len;
1916 if (req->r_old_dentry_drop)
1917 len += req->r_old_dentry->d_name.len;
1918
b61c2763 1919 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
a79832f2
SW		 1920 if (!msg) {
1921 msg = ERR_PTR(-ENOMEM);
2f2dc053 1922 goto out_free2;
a79832f2 1923 }
2f2dc053 1924
7cfa0313 1925 msg->hdr.version = cpu_to_le16(2);
6df058c0
SW		 1926 msg->hdr.tid = cpu_to_le64(req->r_tid);
1927
2f2dc053
SW		 1928 head = msg->front.iov_base;
1929 p = msg->front.iov_base + sizeof(*head);
1930 end = msg->front.iov_base + msg->front.iov_len;
1931
1932 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1933 head->op = cpu_to_le32(req->r_op);
ff3d0046
EB		 1934 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1935 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2f2dc053
SW		 1936 head->args = req->r_args;
1937
1938 ceph_encode_filepath(&p, end, ino1, path1);
1939 ceph_encode_filepath(&p, end, ino2, path2);
1940
e979cf50
SW		 1941 /* make note of release offset, in case we need to replay */
1942 req->r_request_release_offset = p - msg->front.iov_base;
1943
2f2dc053
SW		 1944 /* cap releases */
1945 releases = 0;
1946 if (req->r_inode_drop)
1947 releases += ceph_encode_inode_release(&p,
1948 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1949 mds, req->r_inode_drop, req->r_inode_unless, 0);
1950 if (req->r_dentry_drop)
1951 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1952 mds, req->r_dentry_drop, req->r_dentry_unless);
1953 if (req->r_old_dentry_drop)
1954 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1955 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1956 if (req->r_old_inode_drop)
1957 releases += ceph_encode_inode_release(&p,
1958 req->r_old_dentry->d_inode,
1959 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
6e6f0923
YZ		 1960
1961 if (drop_cap_releases) {
1962 releases = 0;
1963 p = msg->front.iov_base + req->r_request_release_offset;
1964 }
1965
2f2dc053
SW		 1966 head->num_releases = cpu_to_le16(releases);
1967
b8e69066 1968 /* time stamp */
1f041a89
YZ		 1969 {
1970 struct ceph_timespec ts;
1971 ceph_encode_timespec(&ts, &req->r_stamp);
1972 ceph_encode_copy(&p, &ts, sizeof(ts));
1973 }
b8e69066 1974
2f2dc053
SW		 1975 BUG_ON(p > end);
1976 msg->front.iov_len = p - msg->front.iov_base;
1977 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1978
25e6bae3
YZ		 1979 if (req->r_pagelist) {
1980 struct ceph_pagelist *pagelist = req->r_pagelist;
1981 atomic_inc(&pagelist->refcnt);
1982 ceph_msg_data_add_pagelist(msg, pagelist);
1983 msg->hdr.data_len = cpu_to_le32(pagelist->length);
1984 } else {
1985 msg->hdr.data_len = 0;
ebf18f47 1986 }
02afca6c 1987
2f2dc053
SW		 1988 msg->hdr.data_off = cpu_to_le16(0);
1989
1990out_free2:
1991 if (freepath2)
1992 kfree((char *)path2);
1993out_free1:
1994 if (freepath1)
1995 kfree((char *)path1);
1996out:
1997 return msg;
1998}
1999
2000/*
2001 * called under mdsc->mutex if error, under no mutex if
2002 * success.
2003 */
2004static void complete_request(struct ceph_mds_client *mdsc,
2005 struct ceph_mds_request *req)
2006{
2007 if (req->r_callback)
2008 req->r_callback(mdsc, req);
2009 else
03066f23 2010 complete_all(&req->r_completion);
2f2dc053
SW		 2011}
2012
2013/*
2014 * called under mdsc->mutex
2015 */
2016static int __prepare_send_request(struct ceph_mds_client *mdsc,
2017 struct ceph_mds_request *req,
6e6f0923 2018 int mds, bool drop_cap_releases)
2f2dc053
SW		 2019{
2020 struct ceph_mds_request_head *rhead;
2021 struct ceph_msg *msg;
2022 int flags = 0;
2023
2f2dc053 2024 req->r_attempts++;
e55b71f8
GF		 2025 if (req->r_inode) {
2026 struct ceph_cap *cap =
2027 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2028
2029 if (cap)
2030 req->r_sent_on_mseq = cap->mseq;
2031 else
2032 req->r_sent_on_mseq = -1;
2033 }
2f2dc053
SW		 2034 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2035 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2036
01a92f17 2037 if (req->r_got_unsafe) {
c5c9a0bf 2038 void *p;
01a92f17
SW		 2039 /*
2040 * Replay. Do not regenerate message (and rebuild
2041 * paths, etc.); just use the original message.
2042 * Rebuilding paths will break for renames because
2043 * d_move mangles the src name.
2044 */
2045 msg = req->r_request;
2046 rhead = msg->front.iov_base;
2047
2048 flags = le32_to_cpu(rhead->flags);
2049 flags |= CEPH_MDS_FLAG_REPLAY;
2050 rhead->flags = cpu_to_le32(flags);
2051
2052 if (req->r_target_inode)
2053 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2054
2055 rhead->num_retry = req->r_attempts - 1;
e979cf50
SW		 2056
2057 /* remove cap/dentry releases from message */
2058 rhead->num_releases = 0;
c5c9a0bf
YZ		 2059
2060 /* time stamp */
2061 p = msg->front.iov_base + req->r_request_release_offset;
1f041a89
YZ		 2062 {
2063 struct ceph_timespec ts;
2064 ceph_encode_timespec(&ts, &req->r_stamp);
2065 ceph_encode_copy(&p, &ts, sizeof(ts));
2066 }
c5c9a0bf
YZ		 2067
2068 msg->front.iov_len = p - msg->front.iov_base;
2069 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
01a92f17
SW		 2070 return 0;
2071 }
2072
2f2dc053
SW		 2073 if (req->r_request) {
2074 ceph_msg_put(req->r_request);
2075 req->r_request = NULL;
2076 }
6e6f0923 2077 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2f2dc053 2078 if (IS_ERR(msg)) {
e1518c7c 2079 req->r_err = PTR_ERR(msg);
2f2dc053 2080 complete_request(mdsc, req);
a79832f2 2081 return PTR_ERR(msg);
2f2dc053
SW		 2082 }
2083 req->r_request = msg;
2084
2085 rhead = msg->front.iov_base;
2f2dc053
SW		 2086 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2087 if (req->r_got_unsafe)
2088 flags |= CEPH_MDS_FLAG_REPLAY;
2089 if (req->r_locked_dir)
2090 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2091 rhead->flags = cpu_to_le32(flags);
2092 rhead->num_fwd = req->r_num_fwd;
2093 rhead->num_retry = req->r_attempts - 1;
01a92f17 2094 rhead->ino = 0;
2f2dc053
SW		 2095
2096 dout(" r_locked_dir = %p\n", req->r_locked_dir);
2f2dc053
SW		 2097 return 0;
2098}
2099
2100/*
2101 * send request, or put it on the appropriate wait list.
2102 */
2103static int __do_request(struct ceph_mds_client *mdsc,
2104 struct ceph_mds_request *req)
2105{
2106 struct ceph_mds_session *session = NULL;
2107 int mds = -1;
2108 int err = -EAGAIN;
2109
eb1b8af3
YZ		 2110 if (req->r_err || req->r_got_result) {
2111 if (req->r_aborted)
2112 __unregister_request(mdsc, req);
2f2dc053 2113 goto out;
eb1b8af3 2114 }
2f2dc053
SW		 2115
2116 if (req->r_timeout &&
2117 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2118 dout("do_request timed out\n");
2119 err = -EIO;
2120 goto finish;
2121 }
2122
dc69e2e9
SW		 2123 put_request_session(req);
2124
2f2dc053
SW		 2125 mds = __choose_mds(mdsc, req);
2126 if (mds < 0 ||
2127 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2128 dout("do_request no mds or not active, waiting for map\n");
2129 list_add(&req->r_wait, &mdsc->waiting_for_map);
2130 goto out;
2131 }
2132
2133 /* get, open session */
2134 session = __ceph_lookup_mds_session(mdsc, mds);
9c423956 2135 if (!session) {
2f2dc053 2136 session = register_session(mdsc, mds);
9c423956
SW		 2137 if (IS_ERR(session)) {
2138 err = PTR_ERR(session);
2139 goto finish;
2140 }
2141 }
dc69e2e9
SW		 2142 req->r_session = get_session(session);
2143
2f2dc053 2144 dout("do_request mds%d session %p state %s\n", mds, session,
a687ecaf 2145 ceph_session_state_name(session->s_state));
2f2dc053
SW		 2146 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2147 session->s_state != CEPH_MDS_SESSION_HUNG) {
2148 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2149 session->s_state == CEPH_MDS_SESSION_CLOSING)
2150 __open_session(mdsc, session);
2151 list_add(&req->r_wait, &session->s_waiting);
2152 goto out_session;
2153 }
2154
2155 /* send request */
2f2dc053
SW		 2156 req->r_resend_mds = -1; /* forget any previous mds hint */
2157
2158 if (req->r_request_started == 0) /* note request start time */
2159 req->r_request_started = jiffies;
2160
6e6f0923 2161 err = __prepare_send_request(mdsc, req, mds, false);
2f2dc053
SW		 2162 if (!err) {
2163 ceph_msg_get(req->r_request);
2164 ceph_con_send(&session->s_con, req->r_request);
2165 }
2166
2167out_session:
2168 ceph_put_mds_session(session);
2169out:
2170 return err;
2171
2172finish:
e1518c7c 2173 req->r_err = err;
2f2dc053
SW		 2174 complete_request(mdsc, req);
2175 goto out;
2176}
2177
2178/*
2179 * called under mdsc->mutex
2180 */
2181static void __wake_requests(struct ceph_mds_client *mdsc,
2182 struct list_head *head)
2183{
ed75ec2c
YZ		 2184 struct ceph_mds_request *req;
2185 LIST_HEAD(tmp_list);
2186
2187 list_splice_init(head, &tmp_list);
2f2dc053 2188
ed75ec2c
YZ		 2189 while (!list_empty(&tmp_list)) {
2190 req = list_entry(tmp_list.next,
2191 struct ceph_mds_request, r_wait);
2f2dc053 2192 list_del_init(&req->r_wait);
7971bd92 2193 dout(" wake request %p tid %llu\n", req, req->r_tid);
2f2dc053
SW		 2194 __do_request(mdsc, req);
2195 }
2196}
2197
2198/*
2199 * Wake up threads with requests pending for @mds, so that they can
29790f26 2200 * resubmit their requests to a possibly different mds.
2f2dc053 2201 */
29790f26 2202static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2f2dc053 2203{
44ca18f2 2204 struct ceph_mds_request *req;
282c1052 2205 struct rb_node *p = rb_first(&mdsc->request_tree);
2f2dc053
SW		 2206
2207 dout("kick_requests mds%d\n", mds);
282c1052 2208 while (p) {
44ca18f2 2209 req = rb_entry(p, struct ceph_mds_request, r_node);
282c1052 2210 p = rb_next(p);
44ca18f2
SW		 2211 if (req->r_got_unsafe)
2212 continue;
3de22be6
YZ		 2213 if (req->r_attempts > 0)
2214 continue; /* only new requests */
44ca18f2
SW		 2215 if (req->r_session &&
2216 req->r_session->s_mds == mds) {
2217 dout(" kicking tid %llu\n", req->r_tid);
03974e81 2218 list_del_init(&req->r_wait);
44ca18f2 2219 __do_request(mdsc, req);
2f2dc053
SW		 2220 }
2221 }
2222}
2223
2224void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2225 struct ceph_mds_request *req)
2226{
2227 dout("submit_request on %p\n", req);
2228 mutex_lock(&mdsc->mutex);
2229 __register_request(mdsc, req, NULL);
2230 __do_request(mdsc, req);
2231 mutex_unlock(&mdsc->mutex);
2232}
2233
2234/*
2235 * Synchrously perform an mds request. Take care of all of the
2236 * session setup, forwarding, retry details.
2237 */
2238int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2239 struct inode *dir,
2240 struct ceph_mds_request *req)
2241{
2242 int err;
2243
2244 dout("do_request on %p\n", req);
2245
2246 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2247 if (req->r_inode)
2248 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2249 if (req->r_locked_dir)
2250 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
844d87c3 2251 if (req->r_old_dentry_dir)
41b02e1f
SW		 2252 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2253 CEPH_CAP_PIN);
2f2dc053
SW		 2254
2255 /* issue */
2256 mutex_lock(&mdsc->mutex);
2257 __register_request(mdsc, req, dir);
2258 __do_request(mdsc, req);
2259
e1518c7c
SW		 2260 if (req->r_err) {
2261 err = req->r_err;
2262 __unregister_request(mdsc, req);
2263 dout("do_request early error %d\n", err);
2264 goto out;
2f2dc053
SW		 2265 }
2266
e1518c7c
SW		 2267 /* wait */
2268 mutex_unlock(&mdsc->mutex);
2269 dout("do_request waiting\n");
2270 if (req->r_timeout) {
aa91647c 2271 err = (long)wait_for_completion_killable_timeout(
e1518c7c
SW		 2272 &req->r_completion, req->r_timeout);
2273 if (err == 0)
2274 err = -EIO;
9280be24
YZ		 2275 } else if (req->r_wait_for_completion) {
2276 err = req->r_wait_for_completion(mdsc, req);
e1518c7c 2277 } else {
aa91647c 2278 err = wait_for_completion_killable(&req->r_completion);
e1518c7c
SW		 2279 }
2280 dout("do_request waited, got %d\n", err);
2281 mutex_lock(&mdsc->mutex);
5b1daecd 2282
e1518c7c
SW		 2283 /* only abort if we didn't race with a real reply */
2284 if (req->r_got_result) {
2285 err = le32_to_cpu(req->r_reply_info.head->result);
2286 } else if (err < 0) {
2287 dout("aborted request %lld with %d\n", req->r_tid, err);
b4556396
SW		 2288
2289 /*
2290 * ensure we aren't running concurrently with
2291 * ceph_fill_trace or ceph_readdir_prepopulate, which
2292 * rely on locks (dir mutex) held by our caller.
2293 */
2294 mutex_lock(&req->r_fill_mutex);
e1518c7c
SW		 2295 req->r_err = err;
2296 req->r_aborted = true;
b4556396 2297 mutex_unlock(&req->r_fill_mutex);
5b1daecd 2298
e1518c7c 2299 if (req->r_locked_dir &&
167c9e35
SW		 2300 (req->r_op & CEPH_MDS_OP_WRITE))
2301 ceph_invalidate_dir_request(req);
2f2dc053 2302 } else {
e1518c7c 2303 err = req->r_err;
2f2dc053 2304 }
2f2dc053 2305
e1518c7c
SW		 2306out:
2307 mutex_unlock(&mdsc->mutex);
2f2dc053
SW		 2308 dout("do_request %p done, result %d\n", req, err);
2309 return err;
2310}
2311
167c9e35 2312/*
2f276c51 2313 * Invalidate dir's completeness, dentry lease state on an aborted MDS
167c9e35
SW		 2314 * namespace request.
2315 */
2316void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2317{
2318 struct inode *inode = req->r_locked_dir;
167c9e35 2319
2f276c51 2320 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
167c9e35 2321
2f276c51 2322 ceph_dir_clear_complete(inode);
167c9e35
SW		 2323 if (req->r_dentry)
2324 ceph_invalidate_dentry_lease(req->r_dentry);
2325 if (req->r_old_dentry)
2326 ceph_invalidate_dentry_lease(req->r_old_dentry);
2327}
2328
2f2dc053
SW		 2329/*
2330 * Handle mds reply.
2331 *
2332 * We take the session mutex and parse and process the reply immediately.
2333 * This preserves the logical ordering of replies, capabilities, etc., sent
2334 * by the MDS as they are applied to our local cache.
2335 */
2336static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2337{
2338 struct ceph_mds_client *mdsc = session->s_mdsc;
2339 struct ceph_mds_request *req;
2340 struct ceph_mds_reply_head *head = msg->front.iov_base;
2341 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
982d6011 2342 struct ceph_snap_realm *realm;
2f2dc053
SW		 2343 u64 tid;
2344 int err, result;
2600d2dd 2345 int mds = session->s_mds;
2f2dc053 2346
2f2dc053
SW		 2347 if (msg->front.iov_len < sizeof(*head)) {
2348 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
9ec7cab1 2349 ceph_msg_dump(msg);
2f2dc053
SW		 2350 return;
2351 }
2352
2353 /* get request, session */
6df058c0 2354 tid = le64_to_cpu(msg->hdr.tid);
2f2dc053
SW		 2355 mutex_lock(&mdsc->mutex);
2356 req = __lookup_request(mdsc, tid);
2357 if (!req) {
2358 dout("handle_reply on unknown tid %llu\n", tid);
2359 mutex_unlock(&mdsc->mutex);
2360 return;
2361 }
2362 dout("handle_reply %p\n", req);
2f2dc053
SW		 2363
2364 /* correct session? */
d96d6049 2365 if (req->r_session != session) {
2f2dc053
SW		 2366 pr_err("mdsc_handle_reply got %llu on session mds%d"
2367 " not mds%d\n", tid, session->s_mds,
2368 req->r_session ? req->r_session->s_mds : -1);
2369 mutex_unlock(&mdsc->mutex);
2370 goto out;
2371 }
2372
2373 /* dup? */
2374 if ((req->r_got_unsafe && !head->safe) ||
2375 (req->r_got_safe && head->safe)) {
f3ae1b97 2376 pr_warn("got a dup %s reply on %llu from mds%d\n",
2f2dc053
SW		 2377 head->safe ? "safe" : "unsafe", tid, mds);
2378 mutex_unlock(&mdsc->mutex);
2379 goto out;
2380 }
85792d0d 2381 if (req->r_got_safe && !head->safe) {
f3ae1b97 2382 pr_warn("got unsafe after safe on %llu from mds%d\n",
85792d0d
SW		 2383 tid, mds);
2384 mutex_unlock(&mdsc->mutex);
2385 goto out;
2386 }
2f2dc053
SW		 2387
2388 result = le32_to_cpu(head->result);
2389
2390 /*
e55b71f8
GF		 2391 * Handle an ESTALE
2392 * if we're not talking to the authority, send to them
2393 * if the authority has changed while we weren't looking,
2394 * send to new authority
2395 * Otherwise we just have to return an ESTALE
2f2dc053
SW		 2396 */
2397 if (result == -ESTALE) {
e55b71f8 2398 dout("got ESTALE on request %llu", req->r_tid);
51da8e8c 2399 req->r_resend_mds = -1;
ca18bede 2400 if (req->r_direct_mode != USE_AUTH_MDS) {
e55b71f8
GF		 2401 dout("not using auth, setting for that now");
2402 req->r_direct_mode = USE_AUTH_MDS;
2f2dc053
SW		 2403 __do_request(mdsc, req);
2404 mutex_unlock(&mdsc->mutex);
2405 goto out;
e55b71f8 2406 } else {
ca18bede
YZ		 2407 int mds = __choose_mds(mdsc, req);
2408 if (mds >= 0 && mds != req->r_session->s_mds) {
2409 dout("but auth changed, so resending");
e55b71f8
GF		 2410 __do_request(mdsc, req);
2411 mutex_unlock(&mdsc->mutex);
2412 goto out;
2413 }
2f2dc053 2414 }
e55b71f8 2415 dout("have to return ESTALE on request %llu", req->r_tid);
2f2dc053
SW		 2416 }
2417
e55b71f8 2418
2f2dc053
SW		 2419 if (head->safe) {
2420 req->r_got_safe = true;
2421 __unregister_request(mdsc, req);
2f2dc053
SW		 2422
2423 if (req->r_got_unsafe) {
2424 /*
2425 * We already handled the unsafe response, now do the
2426 * cleanup. No need to examine the response; the MDS
2427 * doesn't include any result info in the safe
2428 * response. And even if it did, there is nothing
2429 * useful we could do with a revised return value.
2430 */
2431 dout("got safe reply %llu, mds%d\n", tid, mds);
2432 list_del_init(&req->r_unsafe_item);
2433
2434 /* last unsafe request during umount? */
44ca18f2 2435 if (mdsc->stopping && !__get_oldest_req(mdsc))
03066f23 2436 complete_all(&mdsc->safe_umount_waiters);
2f2dc053
SW		 2437 mutex_unlock(&mdsc->mutex);
2438 goto out;
2439 }
e1518c7c 2440 } else {
2f2dc053
SW		 2441 req->r_got_unsafe = true;
2442 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2443 }
2444
2445 dout("handle_reply tid %lld result %d\n", tid, result);
2446 rinfo = &req->r_reply_info;
14303d20 2447 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2f2dc053
SW		 2448 mutex_unlock(&mdsc->mutex);
2449
2450 mutex_lock(&session->s_mutex);
2451 if (err < 0) {
25933abd 2452 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
9ec7cab1 2453 ceph_msg_dump(msg);
2f2dc053
SW		 2454 goto out_err;
2455 }
2456
2457 /* snap trace */
982d6011 2458 realm = NULL;
2f2dc053
SW		 2459 if (rinfo->snapblob_len) {
2460 down_write(&mdsc->snap_rwsem);
2461 ceph_update_snap_trace(mdsc, rinfo->snapblob,
982d6011
YZ		 2462 rinfo->snapblob + rinfo->snapblob_len,
2463 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2464 &realm);
2f2dc053
SW		 2465 downgrade_write(&mdsc->snap_rwsem);
2466 } else {
2467 down_read(&mdsc->snap_rwsem);
2468 }
2469
2470 /* insert trace into our cache */
b4556396 2471 mutex_lock(&req->r_fill_mutex);
3d14c5d2 2472 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2f2dc053 2473 if (err == 0) {
6e8575fa 2474 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
81c6aea5 2475 req->r_op == CEPH_MDS_OP_LSSNAP))
2f2dc053 2476 ceph_readdir_prepopulate(req, req->r_session);
37151668 2477 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2f2dc053 2478 }
b4556396 2479 mutex_unlock(&req->r_fill_mutex);
2f2dc053
SW		 2480
2481 up_read(&mdsc->snap_rwsem);
982d6011
YZ		 2482 if (realm)
2483 ceph_put_snap_realm(mdsc, realm);
2f2dc053 2484out_err:
e1518c7c
SW		 2485 mutex_lock(&mdsc->mutex);
2486 if (!req->r_aborted) {
2487 if (err) {
2488 req->r_err = err;
2489 } else {
2490 req->r_reply = msg;
2491 ceph_msg_get(msg);
2492 req->r_got_result = true;
2493 }
2f2dc053 2494 } else {
e1518c7c 2495 dout("reply arrived after request %lld was aborted\n", tid);
2f2dc053 2496 }
e1518c7c 2497 mutex_unlock(&mdsc->mutex);
2f2dc053 2498
ee6b272b 2499 ceph_add_cap_releases(mdsc, req->r_session);
2f2dc053
SW		 2500 mutex_unlock(&session->s_mutex);
2501
2502 /* kick calling process */
2503 complete_request(mdsc, req);
2504out:
2505 ceph_mdsc_put_request(req);
2506 return;
2507}
2508
2509
2510
2511/*
2512 * handle mds notification that our request has been forwarded.
2513 */
2600d2dd
SW		 2514static void handle_forward(struct ceph_mds_client *mdsc,
2515 struct ceph_mds_session *session,
2516 struct ceph_msg *msg)
2f2dc053
SW		 2517{
2518 struct ceph_mds_request *req;
a1ea787c 2519 u64 tid = le64_to_cpu(msg->hdr.tid);
2f2dc053
SW		 2520 u32 next_mds;
2521 u32 fwd_seq;
2f2dc053
SW		 2522 int err = -EINVAL;
2523 void *p = msg->front.iov_base;
2524 void *end = p + msg->front.iov_len;
2f2dc053 2525
a1ea787c 2526 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
c89136ea
SW		 2527 next_mds = ceph_decode_32(&p);
2528 fwd_seq = ceph_decode_32(&p);
2f2dc053
SW		 2529
2530 mutex_lock(&mdsc->mutex);
2531 req = __lookup_request(mdsc, tid);
2532 if (!req) {
2a8e5e36 2533 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2f2dc053
SW		 2534 goto out; /* dup reply? */
2535 }
2536
2a8e5e36
SW		 2537 if (req->r_aborted) {
2538 dout("forward tid %llu aborted, unregistering\n", tid);
2539 __unregister_request(mdsc, req);
2540 } else if (fwd_seq <= req->r_num_fwd) {
2541 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2f2dc053
SW		 2542 tid, next_mds, req->r_num_fwd, fwd_seq);
2543 } else {
2544 /* resend. forward race not possible; mds would drop */
2a8e5e36
SW		 2545 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2546 BUG_ON(req->r_err);
2547 BUG_ON(req->r_got_result);
3de22be6 2548 req->r_attempts = 0;
2f2dc053
SW		 2549 req->r_num_fwd = fwd_seq;
2550 req->r_resend_mds = next_mds;
2551 put_request_session(req);
2552 __do_request(mdsc, req);
2553 }
2554 ceph_mdsc_put_request(req);
2555out:
2556 mutex_unlock(&mdsc->mutex);
2557 return;
2558
2559bad:
2560 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2561}
2562
2563/*
2564 * handle a mds session control message
2565 */
2566static void handle_session(struct ceph_mds_session *session,
2567 struct ceph_msg *msg)
2568{
2569 struct ceph_mds_client *mdsc = session->s_mdsc;
2570 u32 op;
2571 u64 seq;
2600d2dd 2572 int mds = session->s_mds;
2f2dc053
SW		 2573 struct ceph_mds_session_head *h = msg->front.iov_base;
2574 int wake = 0;
2575
2f2dc053
SW		 2576 /* decode */
2577 if (msg->front.iov_len != sizeof(*h))
2578 goto bad;
2579 op = le32_to_cpu(h->op);
2580 seq = le64_to_cpu(h->seq);
2581
2582 mutex_lock(&mdsc->mutex);
2600d2dd
SW		 2583 if (op == CEPH_SESSION_CLOSE)
2584 __unregister_session(mdsc, session);
2f2dc053
SW		 2585 /* FIXME: this ttl calculation is generous */
2586 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2587 mutex_unlock(&mdsc->mutex);
2588
2589 mutex_lock(&session->s_mutex);
2590
2591 dout("handle_session mds%d %s %p state %s seq %llu\n",
2592 mds, ceph_session_op_name(op), session,
a687ecaf 2593 ceph_session_state_name(session->s_state), seq);
2f2dc053
SW		 2594
2595 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2596 session->s_state = CEPH_MDS_SESSION_OPEN;
2597 pr_info("mds%d came back\n", session->s_mds);
2598 }
2599
2600 switch (op) {
2601 case CEPH_SESSION_OPEN:
29790f26
SW		 2602 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2603 pr_info("mds%d reconnect success\n", session->s_mds);
2f2dc053
SW		 2604 session->s_state = CEPH_MDS_SESSION_OPEN;
2605 renewed_caps(mdsc, session, 0);
2606 wake = 1;
2607 if (mdsc->stopping)
2608 __close_session(mdsc, session);
2609 break;
2610
2611 case CEPH_SESSION_RENEWCAPS:
2612 if (session->s_renew_seq == seq)
2613 renewed_caps(mdsc, session, 1);
2614 break;
2615
2616 case CEPH_SESSION_CLOSE:
29790f26
SW		 2617 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2618 pr_info("mds%d reconnect denied\n", session->s_mds);
1c841a96 2619 cleanup_session_requests(mdsc, session);
2f2dc053 2620 remove_session_caps(session);
656e4382 2621 wake = 2; /* for good measure */
f3c60c59 2622 wake_up_all(&mdsc->session_close_wq);
2f2dc053
SW		 2623 break;
2624
2625 case CEPH_SESSION_STALE:
2626 pr_info("mds%d caps went stale, renewing\n",
2627 session->s_mds);
d8fb02ab 2628 spin_lock(&session->s_gen_ttl_lock);
2f2dc053 2629 session->s_cap_gen++;
1ce208a6 2630 session->s_cap_ttl = jiffies - 1;
d8fb02ab 2631 spin_unlock(&session->s_gen_ttl_lock);
2f2dc053
SW		 2632 send_renew_caps(mdsc, session);
2633 break;
2634
2635 case CEPH_SESSION_RECALL_STATE:
2636 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2637 break;
2638
186e4f7a
YZ		 2639 case CEPH_SESSION_FLUSHMSG:
2640 send_flushmsg_ack(mdsc, session, seq);
2641 break;
2642
03f4fcb0
YZ		 2643 case CEPH_SESSION_FORCE_RO:
2644 dout("force_session_readonly %p\n", session);
2645 spin_lock(&session->s_cap_lock);
2646 session->s_readonly = true;
2647 spin_unlock(&session->s_cap_lock);
2648 wake_up_session_caps(session, 0);
2649 break;
2650
2f2dc053
SW		 2651 default:
2652 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2653 WARN_ON(1);
2654 }
2655
2656 mutex_unlock(&session->s_mutex);
2657 if (wake) {
2658 mutex_lock(&mdsc->mutex);
2659 __wake_requests(mdsc, &session->s_waiting);
656e4382
YZ		 2660 if (wake == 2)
2661 kick_requests(mdsc, mds);
2f2dc053
SW		 2662 mutex_unlock(&mdsc->mutex);
2663 }
2664 return;
2665
2666bad:
2667 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2668 (int)msg->front.iov_len);
9ec7cab1 2669 ceph_msg_dump(msg);
2f2dc053
SW		 2670 return;
2671}
2672
2673
2674/*
2675 * called under session->mutex.
2676 */
2677static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2678 struct ceph_mds_session *session)
2679{
2680 struct ceph_mds_request *req, *nreq;
3de22be6 2681 struct rb_node *p;
2f2dc053
SW		 2682 int err;
2683
2684 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2685
2686 mutex_lock(&mdsc->mutex);
2687 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
6e6f0923 2688 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2f2dc053
SW		 2689 if (!err) {
2690 ceph_msg_get(req->r_request);
2691 ceph_con_send(&session->s_con, req->r_request);
2692 }
2693 }
3de22be6
YZ		 2694
2695 /*
2696 * also re-send old requests when MDS enters reconnect stage. So that MDS
2697 * can process completed request in clientreplay stage.
2698 */
2699 p = rb_first(&mdsc->request_tree);
2700 while (p) {
2701 req = rb_entry(p, struct ceph_mds_request, r_node);
2702 p = rb_next(p);
2703 if (req->r_got_unsafe)
2704 continue;
2705 if (req->r_attempts == 0)
2706 continue; /* only old requests */
2707 if (req->r_session &&
2708 req->r_session->s_mds == session->s_mds) {
6e6f0923
YZ		 2709 err = __prepare_send_request(mdsc, req,
2710 session->s_mds, true);
3de22be6
YZ		 2711 if (!err) {
2712 ceph_msg_get(req->r_request);
2713 ceph_con_send(&session->s_con, req->r_request);
2714 }
2715 }
2716 }
2f2dc053
SW		 2717 mutex_unlock(&mdsc->mutex);
2718}
2719
2720/*
2721 * Encode information about a cap for a reconnect with the MDS.
2722 */
2f2dc053
SW		 2723static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2724 void *arg)
2725{
20cb34ae
SW		 2726 union {
2727 struct ceph_mds_cap_reconnect v2;
2728 struct ceph_mds_cap_reconnect_v1 v1;
2729 } rec;
2730 size_t reclen;
2f2dc053 2731 struct ceph_inode_info *ci;
20cb34ae
SW		 2732 struct ceph_reconnect_state *recon_state = arg;
2733 struct ceph_pagelist *pagelist = recon_state->pagelist;
2f2dc053
SW		 2734 char *path;
2735 int pathlen, err;
2736 u64 pathbase;
2737 struct dentry *dentry;
2738
2739 ci = cap->ci;
2740
2741 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2742 inode, ceph_vinop(inode), cap, cap->cap_id,
2743 ceph_cap_string(cap->issued));
93cea5be
SW		 2744 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2745 if (err)
2746 return err;
2f2dc053
SW		 2747
2748 dentry = d_find_alias(inode);
2749 if (dentry) {
2750 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2751 if (IS_ERR(path)) {
2752 err = PTR_ERR(path);
e072f8aa 2753 goto out_dput;
2f2dc053
SW		 2754 }
2755 } else {
2756 path = NULL;
2757 pathlen = 0;
2758 }
93cea5be
SW		 2759 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2760 if (err)
e072f8aa 2761 goto out_free;
2f2dc053 2762
be655596 2763 spin_lock(&ci->i_ceph_lock);
2f2dc053
SW		 2764 cap->seq = 0; /* reset cap seq */
2765 cap->issue_seq = 0; /* and issue_seq */
667ca05c 2766 cap->mseq = 0; /* and migrate_seq */
99a9c273 2767 cap->cap_gen = cap->session->s_cap_gen;
20cb34ae
SW		 2768
2769 if (recon_state->flock) {
2770 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2771 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2772 rec.v2.issued = cpu_to_le32(cap->issued);
2773 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2774 rec.v2.pathbase = cpu_to_le64(pathbase);
2775 rec.v2.flock_len = 0;
2776 reclen = sizeof(rec.v2);
2777 } else {
2778 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2779 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2780 rec.v1.issued = cpu_to_le32(cap->issued);
2781 rec.v1.size = cpu_to_le64(inode->i_size);
2782 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2783 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2784 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2785 rec.v1.pathbase = cpu_to_le64(pathbase);
2786 reclen = sizeof(rec.v1);
2787 }
be655596 2788 spin_unlock(&ci->i_ceph_lock);
2f2dc053 2789
40819f6f
GF		 2790 if (recon_state->flock) {
2791 int num_fcntl_locks, num_flock_locks;
39be95e9
JS		 2792 struct ceph_filelock *flocks;
2793
2794encode_again:
39be95e9 2795 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
39be95e9
JS		 2796 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2797 sizeof(struct ceph_filelock), GFP_NOFS);
2798 if (!flocks) {
2799 err = -ENOMEM;
2800 goto out_free;
2801 }
39be95e9
JS		 2802 err = ceph_encode_locks_to_buffer(inode, flocks,
2803 num_fcntl_locks,
2804 num_flock_locks);
39be95e9
JS		 2805 if (err) {
2806 kfree(flocks);
2807 if (err == -ENOSPC)
2808 goto encode_again;
2809 goto out_free;
2810 }
2811 /*
2812 * number of encoded locks is stable, so copy to pagelist
2813 */
2814 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2815 (num_fcntl_locks+num_flock_locks) *
2816 sizeof(struct ceph_filelock));
2817 err = ceph_pagelist_append(pagelist, &rec, reclen);
2818 if (!err)
2819 err = ceph_locks_to_pagelist(flocks, pagelist,
2820 num_fcntl_locks,
2821 num_flock_locks);
2822 kfree(flocks);
3612abbd
SW		 2823 } else {
2824 err = ceph_pagelist_append(pagelist, &rec, reclen);
40819f6f 2825 }
44c99757
YZ		 2826
2827 recon_state->nr_caps++;
e072f8aa 2828out_free:
2f2dc053 2829 kfree(path);
e072f8aa 2830out_dput:
2f2dc053 2831 dput(dentry);
93cea5be 2832 return err;
2f2dc053
SW		 2833}
2834
2835
2836/*
2837 * If an MDS fails and recovers, clients need to reconnect in order to
2838 * reestablish shared state. This includes all caps issued through
2839 * this session _and_ the snap_realm hierarchy. Because it's not
2840 * clear which snap realms the mds cares about, we send everything we
2841 * know about.. that ensures we'll then get any new info the
2842 * recovering MDS might have.
2843 *
2844 * This is a relatively heavyweight operation, but it's rare.
2845 *
2846 * called with mdsc->mutex held.
2847 */
34b6c855
SW		 2848static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2849 struct ceph_mds_session *session)
2f2dc053 2850{
2f2dc053 2851 struct ceph_msg *reply;
a105f00c 2852 struct rb_node *p;
34b6c855 2853 int mds = session->s_mds;
9abf82b8 2854 int err = -ENOMEM;
44c99757 2855 int s_nr_caps;
93cea5be 2856 struct ceph_pagelist *pagelist;
20cb34ae 2857 struct ceph_reconnect_state recon_state;
2f2dc053 2858
34b6c855 2859 pr_info("mds%d reconnect start\n", mds);
2f2dc053 2860
93cea5be
SW		 2861 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2862 if (!pagelist)
2863 goto fail_nopagelist;
2864 ceph_pagelist_init(pagelist);
2865
b61c2763 2866 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
a79832f2 2867 if (!reply)
93cea5be 2868 goto fail_nomsg;
93cea5be 2869
34b6c855
SW		 2870 mutex_lock(&session->s_mutex);
2871 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2872 session->s_seq = 0;
2f2dc053 2873
2f2dc053 2874 dout("session %p state %s\n", session,
a687ecaf 2875 ceph_session_state_name(session->s_state));
2f2dc053 2876
99a9c273
YZ		 2877 spin_lock(&session->s_gen_ttl_lock);
2878 session->s_cap_gen++;
2879 spin_unlock(&session->s_gen_ttl_lock);
2880
2881 spin_lock(&session->s_cap_lock);
03f4fcb0
YZ		 2882 /* don't know if session is readonly */
2883 session->s_readonly = 0;
99a9c273
YZ		 2884 /*
2885 * notify __ceph_remove_cap() that we are composing cap reconnect.
2886 * If a cap get released before being added to the cap reconnect,
2887 * __ceph_remove_cap() should skip queuing cap release.
2888 */
2889 session->s_cap_reconnect = 1;
e01a5946
SW		 2890 /* drop old cap expires; we're about to reestablish that state */
2891 discard_cap_releases(mdsc, session);
99a9c273 2892 spin_unlock(&session->s_cap_lock);
e01a5946 2893
5d23371f
YZ		 2894 /* trim unused caps to reduce MDS's cache rejoin time */
2895 shrink_dcache_parent(mdsc->fsc->sb->s_root);
2896
2897 ceph_con_close(&session->s_con);
2898 ceph_con_open(&session->s_con,
2899 CEPH_ENTITY_TYPE_MDS, mds,
2900 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2901
2902 /* replay unsafe requests */
2903 replay_unsafe_requests(mdsc, session);
2904
2905 down_read(&mdsc->snap_rwsem);
2906
2f2dc053 2907 /* traverse this session's caps */
44c99757
YZ		 2908 s_nr_caps = session->s_nr_caps;
2909 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
93cea5be
SW		 2910 if (err)
2911 goto fail;
20cb34ae 2912
44c99757 2913 recon_state.nr_caps = 0;
20cb34ae
SW		 2914 recon_state.pagelist = pagelist;
2915 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2916 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2f2dc053 2917 if (err < 0)
9abf82b8 2918 goto fail;
2f2dc053 2919
99a9c273
YZ		 2920 spin_lock(&session->s_cap_lock);
2921 session->s_cap_reconnect = 0;
2922 spin_unlock(&session->s_cap_lock);
2923
2f2dc053
SW		 2924 /*
2925 * snaprealms. we provide mds with the ino, seq (version), and
2926 * parent for all of our realms. If the mds has any newer info,
2927 * it will tell us.
2928 */
a105f00c
SW		 2929 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2930 struct ceph_snap_realm *realm =
2931 rb_entry(p, struct ceph_snap_realm, node);
93cea5be 2932 struct ceph_mds_snaprealm_reconnect sr_rec;
2f2dc053
SW		 2933
2934 dout(" adding snap realm %llx seq %lld parent %llx\n",
2935 realm->ino, realm->seq, realm->parent_ino);
93cea5be
SW		 2936 sr_rec.ino = cpu_to_le64(realm->ino);
2937 sr_rec.seq = cpu_to_le64(realm->seq);
2938 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2939 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2940 if (err)
2941 goto fail;
2f2dc053 2942 }
2f2dc053 2943
20cb34ae
SW		 2944 if (recon_state.flock)
2945 reply->hdr.version = cpu_to_le16(2);
44c99757
YZ		 2946
2947 /* raced with cap release? */
2948 if (s_nr_caps != recon_state.nr_caps) {
2949 struct page *page = list_first_entry(&pagelist->head,
2950 struct page, lru);
2951 __le32 *addr = kmap_atomic(page);
2952 *addr = cpu_to_le32(recon_state.nr_caps);
2953 kunmap_atomic(addr);
ebf18f47 2954 }
44c99757
YZ		 2955
2956 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2957 ceph_msg_data_add_pagelist(reply, pagelist);
2f2dc053
SW		 2958 ceph_con_send(&session->s_con, reply);
2959
9abf82b8
SW		 2960 mutex_unlock(&session->s_mutex);
2961
2962 mutex_lock(&mdsc->mutex);
2963 __wake_requests(mdsc, &session->s_waiting);
2964 mutex_unlock(&mdsc->mutex);
2965
2f2dc053 2966 up_read(&mdsc->snap_rwsem);
2f2dc053
SW		 2967 return;
2968
93cea5be 2969fail:
2f2dc053 2970 ceph_msg_put(reply);
9abf82b8
SW		 2971 up_read(&mdsc->snap_rwsem);
2972 mutex_unlock(&session->s_mutex);
93cea5be
SW		 2973fail_nomsg:
2974 ceph_pagelist_release(pagelist);
93cea5be 2975fail_nopagelist:
9abf82b8 2976 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
9abf82b8 2977 return;
2f2dc053
SW		 2978}
2979
2980
2981/*
2982 * compare old and new mdsmaps, kicking requests
2983 * and closing out old connections as necessary
2984 *
2985 * called under mdsc->mutex.
2986 */
2987static void check_new_map(struct ceph_mds_client *mdsc,
2988 struct ceph_mdsmap *newmap,
2989 struct ceph_mdsmap *oldmap)
2990{
2991 int i;
2992 int oldstate, newstate;
2993 struct ceph_mds_session *s;
2994
2995 dout("check_new_map new %u old %u\n",
2996 newmap->m_epoch, oldmap->m_epoch);
2997
2998 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2999 if (mdsc->sessions[i] == NULL)
3000 continue;
3001 s = mdsc->sessions[i];
3002 oldstate = ceph_mdsmap_get_state(oldmap, i);
3003 newstate = ceph_mdsmap_get_state(newmap, i);
3004
0deb01c9 3005 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2f2dc053 3006 i, ceph_mds_state_name(oldstate),
0deb01c9 3007 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2f2dc053 3008 ceph_mds_state_name(newstate),
0deb01c9 3009 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
a687ecaf 3010 ceph_session_state_name(s->s_state));
2f2dc053 3011
3e8f43a0
YZ		 3012 if (i >= newmap->m_max_mds ||
3013 memcmp(ceph_mdsmap_get_addr(oldmap, i),
2f2dc053
SW		 3014 ceph_mdsmap_get_addr(newmap, i),
3015 sizeof(struct ceph_entity_addr))) {
3016 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3017 /* the session never opened, just close it
3018 * out now */
3019 __wake_requests(mdsc, &s->s_waiting);
2600d2dd 3020 __unregister_session(mdsc, s);
2f2dc053
SW		 3021 } else {
3022 /* just close it */
3023 mutex_unlock(&mdsc->mutex);
3024 mutex_lock(&s->s_mutex);
3025 mutex_lock(&mdsc->mutex);
3026 ceph_con_close(&s->s_con);
3027 mutex_unlock(&s->s_mutex);
3028 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3029 }
2f2dc053
SW		 3030 } else if (oldstate == newstate) {
3031 continue; /* nothing new with this mds */
3032 }
3033
3034 /*
3035 * send reconnect?
3036 */
3037 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
34b6c855
SW		 3038 newstate >= CEPH_MDS_STATE_RECONNECT) {
3039 mutex_unlock(&mdsc->mutex);
3040 send_mds_reconnect(mdsc, s);
3041 mutex_lock(&mdsc->mutex);
3042 }
2f2dc053
SW		 3043
3044 /*
29790f26 3045 * kick request on any mds that has gone active.
2f2dc053
SW		 3046 */
3047 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3048 newstate >= CEPH_MDS_STATE_ACTIVE) {
29790f26
SW		 3049 if (oldstate != CEPH_MDS_STATE_CREATING &&
3050 oldstate != CEPH_MDS_STATE_STARTING)
3051 pr_info("mds%d recovery completed\n", s->s_mds);
3052 kick_requests(mdsc, i);
2f2dc053 3053 ceph_kick_flushing_caps(mdsc, s);
0dc2570f 3054 wake_up_session_caps(s, 1);
2f2dc053
SW		 3055 }
3056 }
cb170a22
SW		 3057
3058 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
3059 s = mdsc->sessions[i];
3060 if (!s)
3061 continue;
3062 if (!ceph_mdsmap_is_laggy(newmap, i))
3063 continue;
3064 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3065 s->s_state == CEPH_MDS_SESSION_HUNG ||
3066 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3067 dout(" connecting to export targets of laggy mds%d\n",
3068 i);
3069 __open_export_target_sessions(mdsc, s);
3070 }
3071 }
2f2dc053
SW		 3072}
3073
3074
3075
3076/*
3077 * leases
3078 */
3079
3080/*
3081 * caller must hold session s_mutex, dentry->d_lock
3082 */
3083void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3084{
3085 struct ceph_dentry_info *di = ceph_dentry(dentry);
3086
3087 ceph_put_mds_session(di->lease_session);
3088 di->lease_session = NULL;
3089}
3090
2600d2dd
SW		 3091static void handle_lease(struct ceph_mds_client *mdsc,
3092 struct ceph_mds_session *session,
3093 struct ceph_msg *msg)
2f2dc053 3094{
3d14c5d2 3095 struct super_block *sb = mdsc->fsc->sb;
2f2dc053 3096 struct inode *inode;
2f2dc053
SW		 3097 struct dentry *parent, *dentry;
3098 struct ceph_dentry_info *di;
2600d2dd 3099 int mds = session->s_mds;
2f2dc053 3100 struct ceph_mds_lease *h = msg->front.iov_base;
1e5ea23d 3101 u32 seq;
2f2dc053 3102 struct ceph_vino vino;
2f2dc053
SW		 3103 struct qstr dname;
3104 int release = 0;
3105
2f2dc053
SW		 3106 dout("handle_lease from mds%d\n", mds);
3107
3108 /* decode */
3109 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3110 goto bad;
3111 vino.ino = le64_to_cpu(h->ino);
3112 vino.snap = CEPH_NOSNAP;
1e5ea23d 3113 seq = le32_to_cpu(h->seq);
2f2dc053
SW		 3114 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3115 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3116 if (dname.len != get_unaligned_le32(h+1))
3117 goto bad;
3118
2f2dc053
SW		 3119 /* lookup inode */
3120 inode = ceph_find_inode(sb, vino);
2f90b852
SW		 3121 dout("handle_lease %s, ino %llx %p %.*s\n",
3122 ceph_lease_op_name(h->action), vino.ino, inode,
1e5ea23d 3123 dname.len, dname.name);
6cd3bcad
YZ		 3124
3125 mutex_lock(&session->s_mutex);
3126 session->s_seq++;
3127
2f2dc053
SW		 3128 if (inode == NULL) {
3129 dout("handle_lease no inode %llx\n", vino.ino);
3130 goto release;
3131 }
2f2dc053
SW		 3132
3133 /* dentry */
3134 parent = d_find_alias(inode);
3135 if (!parent) {
3136 dout("no parent dentry on inode %p\n", inode);
3137 WARN_ON(1);
3138 goto release; /* hrm... */
3139 }
3140 dname.hash = full_name_hash(dname.name, dname.len);
3141 dentry = d_lookup(parent, &dname);
3142 dput(parent);
3143 if (!dentry)
3144 goto release;
3145
3146 spin_lock(&dentry->d_lock);
3147 di = ceph_dentry(dentry);
3148 switch (h->action) {
3149 case CEPH_MDS_LEASE_REVOKE:
3d8eb7a9 3150 if (di->lease_session == session) {
1e5ea23d
SW		 3151 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3152 h->seq = cpu_to_le32(di->lease_seq);
2f2dc053
SW		 3153 __ceph_mdsc_drop_dentry_lease(dentry);
3154 }
3155 release = 1;
3156 break;
3157
3158 case CEPH_MDS_LEASE_RENEW:
3d8eb7a9 3159 if (di->lease_session == session &&
2f2dc053
SW		 3160 di->lease_gen == session->s_cap_gen &&
3161 di->lease_renew_from &&
3162 di->lease_renew_after == 0) {
3163 unsigned long duration =
3563dbdd 3164 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
2f2dc053 3165
1e5ea23d 3166 di->lease_seq = seq;
2f2dc053
SW		 3167 dentry->d_time = di->lease_renew_from + duration;
3168 di->lease_renew_after = di->lease_renew_from +
3169 (duration >> 1);
3170 di->lease_renew_from = 0;
3171 }
3172 break;
3173 }
3174 spin_unlock(&dentry->d_lock);
3175 dput(dentry);
3176
3177 if (!release)
3178 goto out;
3179
3180release:
3181 /* let's just reuse the same message */
3182 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3183 ceph_msg_get(msg);
3184 ceph_con_send(&session->s_con, msg);
3185
3186out:
3187 iput(inode);
3188 mutex_unlock(&session->s_mutex);
2f2dc053
SW		 3189 return;
3190
3191bad:
3192 pr_err("corrupt lease message\n");
9ec7cab1 3193 ceph_msg_dump(msg);
2f2dc053
SW		 3194}
3195
3196void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3197 struct inode *inode,
3198 struct dentry *dentry, char action,
3199 u32 seq)
3200{
3201 struct ceph_msg *msg;
3202 struct ceph_mds_lease *lease;
3203 int len = sizeof(*lease) + sizeof(u32);
3204 int dnamelen = 0;
3205
3206 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3207 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3208 dnamelen = dentry->d_name.len;
3209 len += dnamelen;
3210
b61c2763 3211 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
a79832f2 3212 if (!msg)
2f2dc053
SW		 3213 return;
3214 lease = msg->front.iov_base;
3215 lease->action = action;
2f2dc053
SW		 3216 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3217 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3218 lease->seq = cpu_to_le32(seq);
3219 put_unaligned_le32(dnamelen, lease + 1);
3220 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3221
3222 /*
3223 * if this is a preemptive lease RELEASE, no need to
3224 * flush request stream, since the actual request will
3225 * soon follow.
3226 */
3227 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3228
3229 ceph_con_send(&session->s_con, msg);
3230}
3231
3232/*
3233 * Preemptively release a lease we expect to invalidate anyway.
3234 * Pass @inode always, @dentry is optional.
3235 */
3236void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2f90b852 3237 struct dentry *dentry)
2f2dc053
SW		 3238{
3239 struct ceph_dentry_info *di;
3240 struct ceph_mds_session *session;
3241 u32 seq;
3242
3243 BUG_ON(inode == NULL);
3244 BUG_ON(dentry == NULL);
2f2dc053
SW		 3245
3246 /* is dentry lease valid? */
3247 spin_lock(&dentry->d_lock);
3248 di = ceph_dentry(dentry);
3249 if (!di || !di->lease_session ||
3250 di->lease_session->s_mds < 0 ||
3251 di->lease_gen != di->lease_session->s_cap_gen ||
3252 !time_before(jiffies, dentry->d_time)) {
3253 dout("lease_release inode %p dentry %p -- "
2f90b852
SW		 3254 "no lease\n",
3255 inode, dentry);
2f2dc053
SW		 3256 spin_unlock(&dentry->d_lock);
3257 return;
3258 }
3259
3260 /* we do have a lease on this dentry; note mds and seq */
3261 session = ceph_get_mds_session(di->lease_session);
3262 seq = di->lease_seq;
3263 __ceph_mdsc_drop_dentry_lease(dentry);
3264 spin_unlock(&dentry->d_lock);
3265
2f90b852
SW		 3266 dout("lease_release inode %p dentry %p to mds%d\n",
3267 inode, dentry, session->s_mds);
2f2dc053
SW		 3268 ceph_mdsc_lease_send_msg(session, inode, dentry,
3269 CEPH_MDS_LEASE_RELEASE, seq);
3270 ceph_put_mds_session(session);
3271}
3272
3273/*
3274 * drop all leases (and dentry refs) in preparation for umount
3275 */
3276static void drop_leases(struct ceph_mds_client *mdsc)
3277{
3278 int i;
3279
3280 dout("drop_leases\n");
3281 mutex_lock(&mdsc->mutex);
3282 for (i = 0; i < mdsc->max_sessions; i++) {
3283 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3284 if (!s)
3285 continue;
3286 mutex_unlock(&mdsc->mutex);
3287 mutex_lock(&s->s_mutex);
3288 mutex_unlock(&s->s_mutex);
3289 ceph_put_mds_session(s);
3290 mutex_lock(&mdsc->mutex);
3291 }
3292 mutex_unlock(&mdsc->mutex);
3293}
3294
3295
3296
3297/*
3298 * delayed work -- periodically trim expired leases, renew caps with mds
3299 */
3300static void schedule_delayed(struct ceph_mds_client *mdsc)
3301{
3302 int delay = 5;
3303 unsigned hz = round_jiffies_relative(HZ * delay);
3304 schedule_delayed_work(&mdsc->delayed_work, hz);
3305}
3306
3307static void delayed_work(struct work_struct *work)
3308{
3309 int i;
3310 struct ceph_mds_client *mdsc =
3311 container_of(work, struct ceph_mds_client, delayed_work.work);
3312 int renew_interval;
3313 int renew_caps;
3314
3315 dout("mdsc delayed_work\n");
afcdaea3 3316 ceph_check_delayed_caps(mdsc);
2f2dc053
SW		 3317
3318 mutex_lock(&mdsc->mutex);
3319 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3320 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3321 mdsc->last_renew_caps);
3322 if (renew_caps)
3323 mdsc->last_renew_caps = jiffies;
3324
3325 for (i = 0; i < mdsc->max_sessions; i++) {
3326 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3327 if (s == NULL)
3328 continue;
3329 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3330 dout("resending session close request for mds%d\n",
3331 s->s_mds);
3332 request_close_session(mdsc, s);
3333 ceph_put_mds_session(s);
3334 continue;
3335 }
3336 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3337 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3338 s->s_state = CEPH_MDS_SESSION_HUNG;
3339 pr_info("mds%d hung\n", s->s_mds);
3340 }
3341 }
3342 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3343 /* this mds is failed or recovering, just wait */
3344 ceph_put_mds_session(s);
3345 continue;
3346 }
3347 mutex_unlock(&mdsc->mutex);
3348
3349 mutex_lock(&s->s_mutex);
3350 if (renew_caps)
3351 send_renew_caps(mdsc, s);
3352 else
3353 ceph_con_keepalive(&s->s_con);
ee6b272b 3354 ceph_add_cap_releases(mdsc, s);
aab53dd9
SW		 3355 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3356 s->s_state == CEPH_MDS_SESSION_HUNG)
3d7ded4d 3357 ceph_send_cap_releases(mdsc, s);
2f2dc053
SW		 3358 mutex_unlock(&s->s_mutex);
3359 ceph_put_mds_session(s);
3360
3361 mutex_lock(&mdsc->mutex);
3362 }
3363 mutex_unlock(&mdsc->mutex);
3364
3365 schedule_delayed(mdsc);
3366}
3367
3d14c5d2 3368int ceph_mdsc_init(struct ceph_fs_client *fsc)
2f2dc053 3369
2f2dc053 3370{
3d14c5d2
YS		 3371 struct ceph_mds_client *mdsc;
3372
3373 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3374 if (!mdsc)
3375 return -ENOMEM;
3376 mdsc->fsc = fsc;
3377 fsc->mdsc = mdsc;
2f2dc053
SW		 3378 mutex_init(&mdsc->mutex);
3379 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
fb3101b6 3380 if (mdsc->mdsmap == NULL) {
3381 kfree(mdsc);
2d06eeb8 3382 return -ENOMEM;
fb3101b6 3383 }
2d06eeb8 3384
2f2dc053 3385 init_completion(&mdsc->safe_umount_waiters);
f3c60c59 3386 init_waitqueue_head(&mdsc->session_close_wq);
2f2dc053
SW		 3387 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3388 mdsc->sessions = NULL;
86d8f67b 3389 atomic_set(&mdsc->num_sessions, 0);
2f2dc053
SW		 3390 mdsc->max_sessions = 0;
3391 mdsc->stopping = 0;
3392 init_rwsem(&mdsc->snap_rwsem);
a105f00c 3393 mdsc->snap_realms = RB_ROOT;
2f2dc053
SW		 3394 INIT_LIST_HEAD(&mdsc->snap_empty);
3395 spin_lock_init(&mdsc->snap_empty_lock);
3396 mdsc->last_tid = 0;
44ca18f2 3397 mdsc->request_tree = RB_ROOT;
2f2dc053
SW		 3398 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3399 mdsc->last_renew_caps = jiffies;
3400 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3401 spin_lock_init(&mdsc->cap_delay_lock);
3402 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3403 spin_lock_init(&mdsc->snap_flush_lock);
3404 mdsc->cap_flush_seq = 0;
3405 INIT_LIST_HEAD(&mdsc->cap_dirty);
db354052 3406 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
2f2dc053
SW		 3407 mdsc->num_cap_flushing = 0;
3408 spin_lock_init(&mdsc->cap_dirty_lock);
3409 init_waitqueue_head(&mdsc->cap_flushing_wq);
3410 spin_lock_init(&mdsc->dentry_lru_lock);
3411 INIT_LIST_HEAD(&mdsc->dentry_lru);
2d06eeb8 3412
37151668 3413 ceph_caps_init(mdsc);
3d14c5d2 3414 ceph_adjust_min_caps(mdsc, fsc->min_caps);
37151668 3415
5f44f142 3416 return 0;
2f2dc053
SW		 3417}
3418
3419/*
3420 * Wait for safe replies on open mds requests. If we time out, drop
3421 * all requests from the tree to avoid dangling dentry refs.
3422 */
3423static void wait_requests(struct ceph_mds_client *mdsc)
3424{
3425 struct ceph_mds_request *req;
3d14c5d2 3426 struct ceph_fs_client *fsc = mdsc->fsc;
2f2dc053
SW		 3427
3428 mutex_lock(&mdsc->mutex);
44ca18f2 3429 if (__get_oldest_req(mdsc)) {
2f2dc053 3430 mutex_unlock(&mdsc->mutex);
44ca18f2 3431
2f2dc053
SW		 3432 dout("wait_requests waiting for requests\n");
3433 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3d14c5d2 3434 fsc->client->options->mount_timeout * HZ);
2f2dc053
SW		 3435
3436 /* tear down remaining requests */
44ca18f2
SW		 3437 mutex_lock(&mdsc->mutex);
3438 while ((req = __get_oldest_req(mdsc))) {
2f2dc053
SW		 3439 dout("wait_requests timed out on tid %llu\n",
3440 req->r_tid);
44ca18f2 3441 __unregister_request(mdsc, req);
2f2dc053
SW		 3442 }
3443 }
3444 mutex_unlock(&mdsc->mutex);
3445 dout("wait_requests done\n");
3446}
3447
3448/*
3449 * called before mount is ro, and before dentries are torn down.
3450 * (hmm, does this still race with new lookups?)
3451 */
3452void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3453{
3454 dout("pre_umount\n");
3455 mdsc->stopping = 1;
3456
3457 drop_leases(mdsc);
afcdaea3 3458 ceph_flush_dirty_caps(mdsc);
2f2dc053 3459 wait_requests(mdsc);
17c688c3
SW		 3460
3461 /*
3462 * wait for reply handlers to drop their request refs and
3463 * their inode/dcache refs
3464 */
3465 ceph_msgr_flush();
2f2dc053
SW		 3466}
3467
3468/*
3469 * wait for all write mds requests to flush.
3470 */
3471static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3472{
80fc7314 3473 struct ceph_mds_request *req = NULL, *nextreq;
44ca18f2 3474 struct rb_node *n;
2f2dc053
SW		 3475
3476 mutex_lock(&mdsc->mutex);
3477 dout("wait_unsafe_requests want %lld\n", want_tid);
80fc7314 3478restart:
44ca18f2
SW		 3479 req = __get_oldest_req(mdsc);
3480 while (req && req->r_tid <= want_tid) {
80fc7314
SW		 3481 /* find next request */
3482 n = rb_next(&req->r_node);
3483 if (n)
3484 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3485 else
3486 nextreq = NULL;
44ca18f2
SW		 3487 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3488 /* write op */
3489 ceph_mdsc_get_request(req);
80fc7314
SW		 3490 if (nextreq)
3491 ceph_mdsc_get_request(nextreq);
44ca18f2
SW		 3492 mutex_unlock(&mdsc->mutex);
3493 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3494 req->r_tid, want_tid);
3495 wait_for_completion(&req->r_safe_completion);
3496 mutex_lock(&mdsc->mutex);
44ca18f2 3497 ceph_mdsc_put_request(req);
80fc7314
SW		 3498 if (!nextreq)
3499 break; /* next dne before, so we're done! */
3500 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3501 /* next request was removed from tree */
3502 ceph_mdsc_put_request(nextreq);
3503 goto restart;
3504 }
3505 ceph_mdsc_put_request(nextreq); /* won't go away */
44ca18f2 3506 }
80fc7314 3507 req = nextreq;
2f2dc053
SW		 3508 }
3509 mutex_unlock(&mdsc->mutex);
3510 dout("wait_unsafe_requests done\n");
3511}
3512
3513void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3514{
3515 u64 want_tid, want_flush;
3516
3d14c5d2 3517 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
56b7cf95
SW		 3518 return;
3519
2f2dc053
SW		 3520 dout("sync\n");
3521 mutex_lock(&mdsc->mutex);
3522 want_tid = mdsc->last_tid;
2f2dc053 3523 mutex_unlock(&mdsc->mutex);
2f2dc053 3524
afcdaea3 3525 ceph_flush_dirty_caps(mdsc);
d3383a8e
YZ		 3526 spin_lock(&mdsc->cap_dirty_lock);
3527 want_flush = mdsc->cap_flush_seq;
3528 spin_unlock(&mdsc->cap_dirty_lock);
3529
3530 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2f2dc053
SW		 3531
3532 wait_unsafe_requests(mdsc, want_tid);
d3383a8e 3533 wait_caps_flush(mdsc, want_flush);
2f2dc053
SW		 3534}
3535
f3c60c59
SW		 3536/*
3537 * true if all sessions are closed, or we force unmount
3538 */
7fd7d101 3539static bool done_closing_sessions(struct ceph_mds_client *mdsc)
f3c60c59 3540{
3d14c5d2 3541 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
f3c60c59 3542 return true;
86d8f67b 3543 return atomic_read(&mdsc->num_sessions) == 0;
f3c60c59 3544}
2f2dc053
SW		 3545
3546/*
3547 * called after sb is ro.
3548 */
3549void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3550{
3551 struct ceph_mds_session *session;
3552 int i;
3d14c5d2
YS		 3553 struct ceph_fs_client *fsc = mdsc->fsc;
3554 unsigned long timeout = fsc->client->options->mount_timeout * HZ;
2f2dc053
SW		 3555
3556 dout("close_sessions\n");
3557
2f2dc053 3558 /* close sessions */
f3c60c59
SW		 3559 mutex_lock(&mdsc->mutex);
3560 for (i = 0; i < mdsc->max_sessions; i++) {
3561 session = __ceph_lookup_mds_session(mdsc, i);
3562 if (!session)
3563 continue;
2f2dc053 3564 mutex_unlock(&mdsc->mutex);
f3c60c59
SW		 3565 mutex_lock(&session->s_mutex);
3566 __close_session(mdsc, session);
3567 mutex_unlock(&session->s_mutex);
3568 ceph_put_mds_session(session);
2f2dc053
SW		 3569 mutex_lock(&mdsc->mutex);
3570 }
f3c60c59
SW		 3571 mutex_unlock(&mdsc->mutex);
3572
3573 dout("waiting for sessions to close\n");
3574 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3575 timeout);
2f2dc053
SW		 3576
3577 /* tear down remaining sessions */
f3c60c59 3578 mutex_lock(&mdsc->mutex);
2f2dc053
SW		 3579 for (i = 0; i < mdsc->max_sessions; i++) {
3580 if (mdsc->sessions[i]) {
3581 session = get_session(mdsc->sessions[i]);
2600d2dd 3582 __unregister_session(mdsc, session);
2f2dc053
SW		 3583 mutex_unlock(&mdsc->mutex);
3584 mutex_lock(&session->s_mutex);
3585 remove_session_caps(session);
3586 mutex_unlock(&session->s_mutex);
3587 ceph_put_mds_session(session);
3588 mutex_lock(&mdsc->mutex);
3589 }
3590 }
2f2dc053 3591 WARN_ON(!list_empty(&mdsc->cap_delay_list));
2f2dc053
SW		 3592 mutex_unlock(&mdsc->mutex);
3593
3594 ceph_cleanup_empty_realms(mdsc);
3595
3596 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3597
3598 dout("stopped\n");
3599}
3600
3d14c5d2 3601static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2f2dc053
SW		 3602{
3603 dout("stop\n");
3604 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3605 if (mdsc->mdsmap)
3606 ceph_mdsmap_destroy(mdsc->mdsmap);
3607 kfree(mdsc->sessions);
37151668 3608 ceph_caps_finalize(mdsc);
2f2dc053
SW		 3609}
3610
3d14c5d2
YS		 3611void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3612{
3613 struct ceph_mds_client *mdsc = fsc->mdsc;
3614
ef550f6f 3615 dout("mdsc_destroy %p\n", mdsc);
3d14c5d2 3616 ceph_mdsc_stop(mdsc);
ef550f6f
SW		 3617
3618 /* flush out any connection work with references to us */
3619 ceph_msgr_flush();
3620
3d14c5d2
YS		 3621 fsc->mdsc = NULL;
3622 kfree(mdsc);
ef550f6f 3623 dout("mdsc_destroy %p done\n", mdsc);
3d14c5d2
YS		 3624}
3625
2f2dc053
SW		 3626
3627/*
3628 * handle mds map update.
3629 */
3630void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3631{
3632 u32 epoch;
3633 u32 maplen;
3634 void *p = msg->front.iov_base;
3635 void *end = p + msg->front.iov_len;
3636 struct ceph_mdsmap *newmap, *oldmap;
3637 struct ceph_fsid fsid;
3638 int err = -EINVAL;
3639
3640 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3641 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3d14c5d2 3642 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
0743304d 3643 return;
c89136ea
SW		 3644 epoch = ceph_decode_32(&p);
3645 maplen = ceph_decode_32(&p);
2f2dc053
SW		 3646 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3647
3648 /* do we need it? */
3d14c5d2 3649 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
2f2dc053
SW		 3650 mutex_lock(&mdsc->mutex);
3651 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3652 dout("handle_map epoch %u <= our %u\n",
3653 epoch, mdsc->mdsmap->m_epoch);
3654 mutex_unlock(&mdsc->mutex);
3655 return;
3656 }
3657
3658 newmap = ceph_mdsmap_decode(&p, end);
3659 if (IS_ERR(newmap)) {
3660 err = PTR_ERR(newmap);
3661 goto bad_unlock;
3662 }
3663
3664 /* swap into place */
3665 if (mdsc->mdsmap) {
3666 oldmap = mdsc->mdsmap;
3667 mdsc->mdsmap = newmap;
3668 check_new_map(mdsc, newmap, oldmap);
3669 ceph_mdsmap_destroy(oldmap);
3670 } else {
3671 mdsc->mdsmap = newmap; /* first mds map */
3672 }
3d14c5d2 3673 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2f2dc053
SW		 3674
3675 __wake_requests(mdsc, &mdsc->waiting_for_map);
3676
3677 mutex_unlock(&mdsc->mutex);
3678 schedule_delayed(mdsc);
3679 return;
3680
3681bad_unlock:
3682 mutex_unlock(&mdsc->mutex);
3683bad:
3684 pr_err("error decoding mdsmap %d\n", err);
3685 return;
3686}
3687
3688static struct ceph_connection *con_get(struct ceph_connection *con)
3689{
3690 struct ceph_mds_session *s = con->private;
3691
3692 if (get_session(s)) {
2600d2dd 3693 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2f2dc053
SW		 3694 return con;
3695 }
3696 dout("mdsc con_get %p FAIL\n", s);
3697 return NULL;
3698}
3699
3700static void con_put(struct ceph_connection *con)
3701{
3702 struct ceph_mds_session *s = con->private;
3703
7d8e18a6 3704 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
2f2dc053
SW		 3705 ceph_put_mds_session(s);
3706}
3707
3708/*
3709 * if the client is unresponsive for long enough, the mds will kill
3710 * the session entirely.
3711 */
3712static void peer_reset(struct ceph_connection *con)
3713{
3714 struct ceph_mds_session *s = con->private;
7e70f0ed 3715 struct ceph_mds_client *mdsc = s->s_mdsc;
2f2dc053 3716
f3ae1b97 3717 pr_warn("mds%d closed our session\n", s->s_mds);
7e70f0ed 3718 send_mds_reconnect(mdsc, s);
2f2dc053
SW		 3719}
3720
3721static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3722{
3723 struct ceph_mds_session *s = con->private;
3724 struct ceph_mds_client *mdsc = s->s_mdsc;
3725 int type = le16_to_cpu(msg->hdr.type);
3726
2600d2dd
SW		 3727 mutex_lock(&mdsc->mutex);
3728 if (__verify_registered_session(mdsc, s) < 0) {
3729 mutex_unlock(&mdsc->mutex);
3730 goto out;
3731 }
3732 mutex_unlock(&mdsc->mutex);
3733
2f2dc053
SW		 3734 switch (type) {
3735 case CEPH_MSG_MDS_MAP:
3736 ceph_mdsc_handle_map(mdsc, msg);
3737 break;
3738 case CEPH_MSG_CLIENT_SESSION:
3739 handle_session(s, msg);
3740 break;
3741 case CEPH_MSG_CLIENT_REPLY:
3742 handle_reply(s, msg);
3743 break;
3744 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2600d2dd 3745 handle_forward(mdsc, s, msg);
2f2dc053
SW		 3746 break;
3747 case CEPH_MSG_CLIENT_CAPS:
3748 ceph_handle_caps(s, msg);
3749 break;
3750 case CEPH_MSG_CLIENT_SNAP:
2600d2dd 3751 ceph_handle_snap(mdsc, s, msg);
2f2dc053
SW		 3752 break;
3753 case CEPH_MSG_CLIENT_LEASE:
2600d2dd 3754 handle_lease(mdsc, s, msg);
2f2dc053
SW		 3755 break;
3756
3757 default:
3758 pr_err("received unknown message type %d %s\n", type,
3759 ceph_msg_type_name(type));
3760 }
2600d2dd 3761out:
2f2dc053
SW		 3762 ceph_msg_put(msg);
3763}
3764
4e7a5dcd
SW		 3765/*
3766 * authentication
3767 */
a3530df3
AE		 3768
3769/*
3770 * Note: returned pointer is the address of a structure that's
3771 * managed separately. Caller must *not* attempt to free it.
3772 */
3773static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
8f43fb53 3774 int *proto, int force_new)
4e7a5dcd
SW		 3775{
3776 struct ceph_mds_session *s = con->private;
3777 struct ceph_mds_client *mdsc = s->s_mdsc;
3d14c5d2 3778 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
74f1869f 3779 struct ceph_auth_handshake *auth = &s->s_auth;
4e7a5dcd 3780
74f1869f 3781 if (force_new && auth->authorizer) {
27859f97 3782 ceph_auth_destroy_authorizer(ac, auth->authorizer);
74f1869f 3783 auth->authorizer = NULL;
4e7a5dcd 3784 }
27859f97
SW		 3785 if (!auth->authorizer) {
3786 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3787 auth);
0bed9b5c
SW		 3788 if (ret)
3789 return ERR_PTR(ret);
27859f97
SW		 3790 } else {
3791 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3792 auth);
a255651d 3793 if (ret)
a3530df3 3794 return ERR_PTR(ret);
4e7a5dcd 3795 }
4e7a5dcd 3796 *proto = ac->protocol;
74f1869f 3797
a3530df3 3798 return auth;
4e7a5dcd
SW		 3799}
3800
3801
3802static int verify_authorizer_reply(struct ceph_connection *con, int len)
3803{
3804 struct ceph_mds_session *s = con->private;
3805 struct ceph_mds_client *mdsc = s->s_mdsc;
3d14c5d2 3806 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4e7a5dcd 3807
27859f97 3808 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
4e7a5dcd
SW		 3809}
3810
9bd2e6f8
SW		 3811static int invalidate_authorizer(struct ceph_connection *con)
3812{
3813 struct ceph_mds_session *s = con->private;
3814 struct ceph_mds_client *mdsc = s->s_mdsc;
3d14c5d2 3815 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
9bd2e6f8 3816
27859f97 3817 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
9bd2e6f8 3818
3d14c5d2 3819 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
9bd2e6f8
SW		 3820}
3821
53ded495
AE		 3822static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3823 struct ceph_msg_header *hdr, int *skip)
3824{
3825 struct ceph_msg *msg;
3826 int type = (int) le16_to_cpu(hdr->type);
3827 int front_len = (int) le32_to_cpu(hdr->front_len);
3828
3829 if (con->in_msg)
3830 return con->in_msg;
3831
3832 *skip = 0;
3833 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3834 if (!msg) {
3835 pr_err("unable to allocate msg type %d len %d\n",
3836 type, front_len);
3837 return NULL;
3838 }
53ded495
AE		 3839
3840 return msg;
3841}
3842
33d07337
YZ		 3843static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
3844{
3845 struct ceph_mds_session *s = con->private;
3846 struct ceph_auth_handshake *auth = &s->s_auth;
3847 return ceph_auth_sign_message(auth, msg);
3848}
3849
3850static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3851{
3852 struct ceph_mds_session *s = con->private;
3853 struct ceph_auth_handshake *auth = &s->s_auth;
3854 return ceph_auth_check_message_signature(auth, msg);
3855}
3856
9e32789f 3857static const struct ceph_connection_operations mds_con_ops = {
2f2dc053
SW		 3858 .get = con_get,
3859 .put = con_put,
3860 .dispatch = dispatch,
4e7a5dcd
SW		 3861 .get_authorizer = get_authorizer,
3862 .verify_authorizer_reply = verify_authorizer_reply,
9bd2e6f8 3863 .invalidate_authorizer = invalidate_authorizer,
2f2dc053 3864 .peer_reset = peer_reset,
53ded495 3865 .alloc_msg = mds_alloc_msg,
33d07337
YZ		 3866 .sign_message = sign_message,
3867 .check_message_signature = check_message_signature,
2f2dc053
SW		 3868};
3869
2f2dc053 3870/* eof */
ubuntu-bionic-kernel mirror