Loading...
1#include <linux/ceph/ceph_debug.h>
2
3#include <linux/fs.h>
4#include <linux/kernel.h>
5#include <linux/sched.h>
6#include <linux/slab.h>
7#include <linux/vmalloc.h>
8#include <linux/wait.h>
9#include <linux/writeback.h>
10
11#include "super.h"
12#include "mds_client.h"
13#include "cache.h"
14#include <linux/ceph/decode.h>
15#include <linux/ceph/messenger.h>
16
17/*
18 * Capability management
19 *
20 * The Ceph metadata servers control client access to inode metadata
21 * and file data by issuing capabilities, granting clients permission
22 * to read and/or write both inode field and file data to OSDs
23 * (storage nodes). Each capability consists of a set of bits
24 * indicating which operations are allowed.
25 *
26 * If the client holds a *_SHARED cap, the client has a coherent value
27 * that can be safely read from the cached inode.
28 *
29 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
30 * client is allowed to change inode attributes (e.g., file size,
31 * mtime), note its dirty state in the ceph_cap, and asynchronously
32 * flush that metadata change to the MDS.
33 *
34 * In the event of a conflicting operation (perhaps by another
35 * client), the MDS will revoke the conflicting client capabilities.
36 *
37 * In order for a client to cache an inode, it must hold a capability
38 * with at least one MDS server. When inodes are released, release
39 * notifications are batched and periodically sent en masse to the MDS
40 * cluster to release server state.
41 */
42
43
44/*
45 * Generate readable cap strings for debugging output.
46 */
47#define MAX_CAP_STR 20
48static char cap_str[MAX_CAP_STR][40];
49static DEFINE_SPINLOCK(cap_str_lock);
50static int last_cap_str;
51
52static char *gcap_string(char *s, int c)
53{
54 if (c & CEPH_CAP_GSHARED)
55 *s++ = 's';
56 if (c & CEPH_CAP_GEXCL)
57 *s++ = 'x';
58 if (c & CEPH_CAP_GCACHE)
59 *s++ = 'c';
60 if (c & CEPH_CAP_GRD)
61 *s++ = 'r';
62 if (c & CEPH_CAP_GWR)
63 *s++ = 'w';
64 if (c & CEPH_CAP_GBUFFER)
65 *s++ = 'b';
66 if (c & CEPH_CAP_GLAZYIO)
67 *s++ = 'l';
68 return s;
69}
70
71const char *ceph_cap_string(int caps)
72{
73 int i;
74 char *s;
75 int c;
76
77 spin_lock(&cap_str_lock);
78 i = last_cap_str++;
79 if (last_cap_str == MAX_CAP_STR)
80 last_cap_str = 0;
81 spin_unlock(&cap_str_lock);
82
83 s = cap_str[i];
84
85 if (caps & CEPH_CAP_PIN)
86 *s++ = 'p';
87
88 c = (caps >> CEPH_CAP_SAUTH) & 3;
89 if (c) {
90 *s++ = 'A';
91 s = gcap_string(s, c);
92 }
93
94 c = (caps >> CEPH_CAP_SLINK) & 3;
95 if (c) {
96 *s++ = 'L';
97 s = gcap_string(s, c);
98 }
99
100 c = (caps >> CEPH_CAP_SXATTR) & 3;
101 if (c) {
102 *s++ = 'X';
103 s = gcap_string(s, c);
104 }
105
106 c = caps >> CEPH_CAP_SFILE;
107 if (c) {
108 *s++ = 'F';
109 s = gcap_string(s, c);
110 }
111
112 if (s == cap_str[i])
113 *s++ = '-';
114 *s = 0;
115 return cap_str[i];
116}
117
118void ceph_caps_init(struct ceph_mds_client *mdsc)
119{
120 INIT_LIST_HEAD(&mdsc->caps_list);
121 spin_lock_init(&mdsc->caps_list_lock);
122}
123
124void ceph_caps_finalize(struct ceph_mds_client *mdsc)
125{
126 struct ceph_cap *cap;
127
128 spin_lock(&mdsc->caps_list_lock);
129 while (!list_empty(&mdsc->caps_list)) {
130 cap = list_first_entry(&mdsc->caps_list,
131 struct ceph_cap, caps_item);
132 list_del(&cap->caps_item);
133 kmem_cache_free(ceph_cap_cachep, cap);
134 }
135 mdsc->caps_total_count = 0;
136 mdsc->caps_avail_count = 0;
137 mdsc->caps_use_count = 0;
138 mdsc->caps_reserve_count = 0;
139 mdsc->caps_min_count = 0;
140 spin_unlock(&mdsc->caps_list_lock);
141}
142
143void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
144{
145 spin_lock(&mdsc->caps_list_lock);
146 mdsc->caps_min_count += delta;
147 BUG_ON(mdsc->caps_min_count < 0);
148 spin_unlock(&mdsc->caps_list_lock);
149}
150
151void ceph_reserve_caps(struct ceph_mds_client *mdsc,
152 struct ceph_cap_reservation *ctx, int need)
153{
154 int i;
155 struct ceph_cap *cap;
156 int have;
157 int alloc = 0;
158 LIST_HEAD(newcaps);
159
160 dout("reserve caps ctx=%p need=%d\n", ctx, need);
161
162 /* first reserve any caps that are already allocated */
163 spin_lock(&mdsc->caps_list_lock);
164 if (mdsc->caps_avail_count >= need)
165 have = need;
166 else
167 have = mdsc->caps_avail_count;
168 mdsc->caps_avail_count -= have;
169 mdsc->caps_reserve_count += have;
170 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
171 mdsc->caps_reserve_count +
172 mdsc->caps_avail_count);
173 spin_unlock(&mdsc->caps_list_lock);
174
175 for (i = have; i < need; i++) {
176 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
177 if (!cap)
178 break;
179 list_add(&cap->caps_item, &newcaps);
180 alloc++;
181 }
182 /* we didn't manage to reserve as much as we needed */
183 if (have + alloc != need)
184 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
185 ctx, need, have + alloc);
186
187 spin_lock(&mdsc->caps_list_lock);
188 mdsc->caps_total_count += alloc;
189 mdsc->caps_reserve_count += alloc;
190 list_splice(&newcaps, &mdsc->caps_list);
191
192 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
193 mdsc->caps_reserve_count +
194 mdsc->caps_avail_count);
195 spin_unlock(&mdsc->caps_list_lock);
196
197 ctx->count = need;
198 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
199 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
200 mdsc->caps_reserve_count, mdsc->caps_avail_count);
201}
202
203int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
204 struct ceph_cap_reservation *ctx)
205{
206 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
207 if (ctx->count) {
208 spin_lock(&mdsc->caps_list_lock);
209 BUG_ON(mdsc->caps_reserve_count < ctx->count);
210 mdsc->caps_reserve_count -= ctx->count;
211 mdsc->caps_avail_count += ctx->count;
212 ctx->count = 0;
213 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
214 mdsc->caps_total_count, mdsc->caps_use_count,
215 mdsc->caps_reserve_count, mdsc->caps_avail_count);
216 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
217 mdsc->caps_reserve_count +
218 mdsc->caps_avail_count);
219 spin_unlock(&mdsc->caps_list_lock);
220 }
221 return 0;
222}
223
224struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
225 struct ceph_cap_reservation *ctx)
226{
227 struct ceph_cap *cap = NULL;
228
229 /* temporary, until we do something about cap import/export */
230 if (!ctx) {
231 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
232 if (cap) {
233 spin_lock(&mdsc->caps_list_lock);
234 mdsc->caps_use_count++;
235 mdsc->caps_total_count++;
236 spin_unlock(&mdsc->caps_list_lock);
237 }
238 return cap;
239 }
240
241 spin_lock(&mdsc->caps_list_lock);
242 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
243 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
244 mdsc->caps_reserve_count, mdsc->caps_avail_count);
245 BUG_ON(!ctx->count);
246 BUG_ON(ctx->count > mdsc->caps_reserve_count);
247 BUG_ON(list_empty(&mdsc->caps_list));
248
249 ctx->count--;
250 mdsc->caps_reserve_count--;
251 mdsc->caps_use_count++;
252
253 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
254 list_del(&cap->caps_item);
255
256 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
257 mdsc->caps_reserve_count + mdsc->caps_avail_count);
258 spin_unlock(&mdsc->caps_list_lock);
259 return cap;
260}
261
262void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
263{
264 spin_lock(&mdsc->caps_list_lock);
265 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
266 cap, mdsc->caps_total_count, mdsc->caps_use_count,
267 mdsc->caps_reserve_count, mdsc->caps_avail_count);
268 mdsc->caps_use_count--;
269 /*
270 * Keep some preallocated caps around (ceph_min_count), to
271 * avoid lots of free/alloc churn.
272 */
273 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
274 mdsc->caps_min_count) {
275 mdsc->caps_total_count--;
276 kmem_cache_free(ceph_cap_cachep, cap);
277 } else {
278 mdsc->caps_avail_count++;
279 list_add(&cap->caps_item, &mdsc->caps_list);
280 }
281
282 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
283 mdsc->caps_reserve_count + mdsc->caps_avail_count);
284 spin_unlock(&mdsc->caps_list_lock);
285}
286
287void ceph_reservation_status(struct ceph_fs_client *fsc,
288 int *total, int *avail, int *used, int *reserved,
289 int *min)
290{
291 struct ceph_mds_client *mdsc = fsc->mdsc;
292
293 if (total)
294 *total = mdsc->caps_total_count;
295 if (avail)
296 *avail = mdsc->caps_avail_count;
297 if (used)
298 *used = mdsc->caps_use_count;
299 if (reserved)
300 *reserved = mdsc->caps_reserve_count;
301 if (min)
302 *min = mdsc->caps_min_count;
303}
304
305/*
306 * Find ceph_cap for given mds, if any.
307 *
308 * Called with i_ceph_lock held.
309 */
310static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
311{
312 struct ceph_cap *cap;
313 struct rb_node *n = ci->i_caps.rb_node;
314
315 while (n) {
316 cap = rb_entry(n, struct ceph_cap, ci_node);
317 if (mds < cap->mds)
318 n = n->rb_left;
319 else if (mds > cap->mds)
320 n = n->rb_right;
321 else
322 return cap;
323 }
324 return NULL;
325}
326
327struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
328{
329 struct ceph_cap *cap;
330
331 spin_lock(&ci->i_ceph_lock);
332 cap = __get_cap_for_mds(ci, mds);
333 spin_unlock(&ci->i_ceph_lock);
334 return cap;
335}
336
337/*
338 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
339 */
340static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
341{
342 struct ceph_cap *cap;
343 int mds = -1;
344 struct rb_node *p;
345
346 /* prefer mds with WR|BUFFER|EXCL caps */
347 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
348 cap = rb_entry(p, struct ceph_cap, ci_node);
349 mds = cap->mds;
350 if (cap->issued & (CEPH_CAP_FILE_WR |
351 CEPH_CAP_FILE_BUFFER |
352 CEPH_CAP_FILE_EXCL))
353 break;
354 }
355 return mds;
356}
357
358int ceph_get_cap_mds(struct inode *inode)
359{
360 struct ceph_inode_info *ci = ceph_inode(inode);
361 int mds;
362 spin_lock(&ci->i_ceph_lock);
363 mds = __ceph_get_cap_mds(ceph_inode(inode));
364 spin_unlock(&ci->i_ceph_lock);
365 return mds;
366}
367
368/*
369 * Called under i_ceph_lock.
370 */
371static void __insert_cap_node(struct ceph_inode_info *ci,
372 struct ceph_cap *new)
373{
374 struct rb_node **p = &ci->i_caps.rb_node;
375 struct rb_node *parent = NULL;
376 struct ceph_cap *cap = NULL;
377
378 while (*p) {
379 parent = *p;
380 cap = rb_entry(parent, struct ceph_cap, ci_node);
381 if (new->mds < cap->mds)
382 p = &(*p)->rb_left;
383 else if (new->mds > cap->mds)
384 p = &(*p)->rb_right;
385 else
386 BUG();
387 }
388
389 rb_link_node(&new->ci_node, parent, p);
390 rb_insert_color(&new->ci_node, &ci->i_caps);
391}
392
393/*
394 * (re)set cap hold timeouts, which control the delayed release
395 * of unused caps back to the MDS. Should be called on cap use.
396 */
397static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
398 struct ceph_inode_info *ci)
399{
400 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
401
402 ci->i_hold_caps_min = round_jiffies(jiffies +
403 ma->caps_wanted_delay_min * HZ);
404 ci->i_hold_caps_max = round_jiffies(jiffies +
405 ma->caps_wanted_delay_max * HZ);
406 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
407 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
408}
409
410/*
411 * (Re)queue cap at the end of the delayed cap release list.
412 *
413 * If I_FLUSH is set, leave the inode at the front of the list.
414 *
415 * Caller holds i_ceph_lock
416 * -> we take mdsc->cap_delay_lock
417 */
418static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
419 struct ceph_inode_info *ci)
420{
421 __cap_set_timeouts(mdsc, ci);
422 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
423 ci->i_ceph_flags, ci->i_hold_caps_max);
424 if (!mdsc->stopping) {
425 spin_lock(&mdsc->cap_delay_lock);
426 if (!list_empty(&ci->i_cap_delay_list)) {
427 if (ci->i_ceph_flags & CEPH_I_FLUSH)
428 goto no_change;
429 list_del_init(&ci->i_cap_delay_list);
430 }
431 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
432no_change:
433 spin_unlock(&mdsc->cap_delay_lock);
434 }
435}
436
437/*
438 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
439 * indicating we should send a cap message to flush dirty metadata
440 * asap, and move to the front of the delayed cap list.
441 */
442static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
443 struct ceph_inode_info *ci)
444{
445 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
446 spin_lock(&mdsc->cap_delay_lock);
447 ci->i_ceph_flags |= CEPH_I_FLUSH;
448 if (!list_empty(&ci->i_cap_delay_list))
449 list_del_init(&ci->i_cap_delay_list);
450 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
451 spin_unlock(&mdsc->cap_delay_lock);
452}
453
454/*
455 * Cancel delayed work on cap.
456 *
457 * Caller must hold i_ceph_lock.
458 */
459static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
460 struct ceph_inode_info *ci)
461{
462 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
463 if (list_empty(&ci->i_cap_delay_list))
464 return;
465 spin_lock(&mdsc->cap_delay_lock);
466 list_del_init(&ci->i_cap_delay_list);
467 spin_unlock(&mdsc->cap_delay_lock);
468}
469
470/*
471 * Common issue checks for add_cap, handle_cap_grant.
472 */
473static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
474 unsigned issued)
475{
476 unsigned had = __ceph_caps_issued(ci, NULL);
477
478 /*
479 * Each time we receive FILE_CACHE anew, we increment
480 * i_rdcache_gen.
481 */
482 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
483 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
484 ci->i_rdcache_gen++;
485 }
486
487 /*
488 * if we are newly issued FILE_SHARED, mark dir not complete; we
489 * don't know what happened to this directory while we didn't
490 * have the cap.
491 */
492 if ((issued & CEPH_CAP_FILE_SHARED) &&
493 (had & CEPH_CAP_FILE_SHARED) == 0) {
494 ci->i_shared_gen++;
495 if (S_ISDIR(ci->vfs_inode.i_mode)) {
496 dout(" marking %p NOT complete\n", &ci->vfs_inode);
497 __ceph_dir_clear_complete(ci);
498 }
499 }
500}
501
502/*
503 * Add a capability under the given MDS session.
504 *
505 * Caller should hold session snap_rwsem (read) and s_mutex.
506 *
507 * @fmode is the open file mode, if we are opening a file, otherwise
508 * it is < 0. (This is so we can atomically add the cap and add an
509 * open file reference to it.)
510 */
511void ceph_add_cap(struct inode *inode,
512 struct ceph_mds_session *session, u64 cap_id,
513 int fmode, unsigned issued, unsigned wanted,
514 unsigned seq, unsigned mseq, u64 realmino, int flags,
515 struct ceph_cap **new_cap)
516{
517 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
518 struct ceph_inode_info *ci = ceph_inode(inode);
519 struct ceph_cap *cap;
520 int mds = session->s_mds;
521 int actual_wanted;
522
523 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
524 session->s_mds, cap_id, ceph_cap_string(issued), seq);
525
526 /*
527 * If we are opening the file, include file mode wanted bits
528 * in wanted.
529 */
530 if (fmode >= 0)
531 wanted |= ceph_caps_for_mode(fmode);
532
533 cap = __get_cap_for_mds(ci, mds);
534 if (!cap) {
535 cap = *new_cap;
536 *new_cap = NULL;
537
538 cap->issued = 0;
539 cap->implemented = 0;
540 cap->mds = mds;
541 cap->mds_wanted = 0;
542 cap->mseq = 0;
543
544 cap->ci = ci;
545 __insert_cap_node(ci, cap);
546
547 /* add to session cap list */
548 cap->session = session;
549 spin_lock(&session->s_cap_lock);
550 list_add_tail(&cap->session_caps, &session->s_caps);
551 session->s_nr_caps++;
552 spin_unlock(&session->s_cap_lock);
553 } else {
554 /*
555 * auth mds of the inode changed. we received the cap export
556 * message, but still haven't received the cap import message.
557 * handle_cap_export() updated the new auth MDS' cap.
558 *
559 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
560 * a message that was send before the cap import message. So
561 * don't remove caps.
562 */
563 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
564 WARN_ON(cap != ci->i_auth_cap);
565 WARN_ON(cap->cap_id != cap_id);
566 seq = cap->seq;
567 mseq = cap->mseq;
568 issued |= cap->issued;
569 flags |= CEPH_CAP_FLAG_AUTH;
570 }
571 }
572
573 if (!ci->i_snap_realm) {
574 /*
575 * add this inode to the appropriate snap realm
576 */
577 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
578 realmino);
579 if (realm) {
580 spin_lock(&realm->inodes_with_caps_lock);
581 ci->i_snap_realm = realm;
582 list_add(&ci->i_snap_realm_item,
583 &realm->inodes_with_caps);
584 spin_unlock(&realm->inodes_with_caps_lock);
585 } else {
586 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
587 realmino);
588 WARN_ON(!realm);
589 }
590 }
591
592 __check_cap_issue(ci, cap, issued);
593
594 /*
595 * If we are issued caps we don't want, or the mds' wanted
596 * value appears to be off, queue a check so we'll release
597 * later and/or update the mds wanted value.
598 */
599 actual_wanted = __ceph_caps_wanted(ci);
600 if ((wanted & ~actual_wanted) ||
601 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
602 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
603 ceph_cap_string(issued), ceph_cap_string(wanted),
604 ceph_cap_string(actual_wanted));
605 __cap_delay_requeue(mdsc, ci);
606 }
607
608 if (flags & CEPH_CAP_FLAG_AUTH) {
609 if (ci->i_auth_cap == NULL ||
610 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
611 ci->i_auth_cap = cap;
612 cap->mds_wanted = wanted;
613 }
614 } else {
615 WARN_ON(ci->i_auth_cap == cap);
616 }
617
618 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
619 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
620 ceph_cap_string(issued|cap->issued), seq, mds);
621 cap->cap_id = cap_id;
622 cap->issued = issued;
623 cap->implemented |= issued;
624 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
625 cap->mds_wanted = wanted;
626 else
627 cap->mds_wanted |= wanted;
628 cap->seq = seq;
629 cap->issue_seq = seq;
630 cap->mseq = mseq;
631 cap->cap_gen = session->s_cap_gen;
632
633 if (fmode >= 0)
634 __ceph_get_fmode(ci, fmode);
635}
636
637/*
638 * Return true if cap has not timed out and belongs to the current
639 * generation of the MDS session (i.e. has not gone 'stale' due to
640 * us losing touch with the mds).
641 */
642static int __cap_is_valid(struct ceph_cap *cap)
643{
644 unsigned long ttl;
645 u32 gen;
646
647 spin_lock(&cap->session->s_gen_ttl_lock);
648 gen = cap->session->s_cap_gen;
649 ttl = cap->session->s_cap_ttl;
650 spin_unlock(&cap->session->s_gen_ttl_lock);
651
652 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
653 dout("__cap_is_valid %p cap %p issued %s "
654 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
655 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
656 return 0;
657 }
658
659 return 1;
660}
661
662/*
663 * Return set of valid cap bits issued to us. Note that caps time
664 * out, and may be invalidated in bulk if the client session times out
665 * and session->s_cap_gen is bumped.
666 */
667int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
668{
669 int have = ci->i_snap_caps;
670 struct ceph_cap *cap;
671 struct rb_node *p;
672
673 if (implemented)
674 *implemented = 0;
675 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
676 cap = rb_entry(p, struct ceph_cap, ci_node);
677 if (!__cap_is_valid(cap))
678 continue;
679 dout("__ceph_caps_issued %p cap %p issued %s\n",
680 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
681 have |= cap->issued;
682 if (implemented)
683 *implemented |= cap->implemented;
684 }
685 /*
686 * exclude caps issued by non-auth MDS, but are been revoking
687 * by the auth MDS. The non-auth MDS should be revoking/exporting
688 * these caps, but the message is delayed.
689 */
690 if (ci->i_auth_cap) {
691 cap = ci->i_auth_cap;
692 have &= ~cap->implemented | cap->issued;
693 }
694 return have;
695}
696
697/*
698 * Get cap bits issued by caps other than @ocap
699 */
700int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
701{
702 int have = ci->i_snap_caps;
703 struct ceph_cap *cap;
704 struct rb_node *p;
705
706 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
707 cap = rb_entry(p, struct ceph_cap, ci_node);
708 if (cap == ocap)
709 continue;
710 if (!__cap_is_valid(cap))
711 continue;
712 have |= cap->issued;
713 }
714 return have;
715}
716
717/*
718 * Move a cap to the end of the LRU (oldest caps at list head, newest
719 * at list tail).
720 */
721static void __touch_cap(struct ceph_cap *cap)
722{
723 struct ceph_mds_session *s = cap->session;
724
725 spin_lock(&s->s_cap_lock);
726 if (s->s_cap_iterator == NULL) {
727 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
728 s->s_mds);
729 list_move_tail(&cap->session_caps, &s->s_caps);
730 } else {
731 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
732 &cap->ci->vfs_inode, cap, s->s_mds);
733 }
734 spin_unlock(&s->s_cap_lock);
735}
736
737/*
738 * Check if we hold the given mask. If so, move the cap(s) to the
739 * front of their respective LRUs. (This is the preferred way for
740 * callers to check for caps they want.)
741 */
742int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
743{
744 struct ceph_cap *cap;
745 struct rb_node *p;
746 int have = ci->i_snap_caps;
747
748 if ((have & mask) == mask) {
749 dout("__ceph_caps_issued_mask %p snap issued %s"
750 " (mask %s)\n", &ci->vfs_inode,
751 ceph_cap_string(have),
752 ceph_cap_string(mask));
753 return 1;
754 }
755
756 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
757 cap = rb_entry(p, struct ceph_cap, ci_node);
758 if (!__cap_is_valid(cap))
759 continue;
760 if ((cap->issued & mask) == mask) {
761 dout("__ceph_caps_issued_mask %p cap %p issued %s"
762 " (mask %s)\n", &ci->vfs_inode, cap,
763 ceph_cap_string(cap->issued),
764 ceph_cap_string(mask));
765 if (touch)
766 __touch_cap(cap);
767 return 1;
768 }
769
770 /* does a combination of caps satisfy mask? */
771 have |= cap->issued;
772 if ((have & mask) == mask) {
773 dout("__ceph_caps_issued_mask %p combo issued %s"
774 " (mask %s)\n", &ci->vfs_inode,
775 ceph_cap_string(cap->issued),
776 ceph_cap_string(mask));
777 if (touch) {
778 struct rb_node *q;
779
780 /* touch this + preceding caps */
781 __touch_cap(cap);
782 for (q = rb_first(&ci->i_caps); q != p;
783 q = rb_next(q)) {
784 cap = rb_entry(q, struct ceph_cap,
785 ci_node);
786 if (!__cap_is_valid(cap))
787 continue;
788 __touch_cap(cap);
789 }
790 }
791 return 1;
792 }
793 }
794
795 return 0;
796}
797
798/*
799 * Return true if mask caps are currently being revoked by an MDS.
800 */
801int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
802 struct ceph_cap *ocap, int mask)
803{
804 struct ceph_cap *cap;
805 struct rb_node *p;
806
807 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
808 cap = rb_entry(p, struct ceph_cap, ci_node);
809 if (cap != ocap &&
810 (cap->implemented & ~cap->issued & mask))
811 return 1;
812 }
813 return 0;
814}
815
816int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
817{
818 struct inode *inode = &ci->vfs_inode;
819 int ret;
820
821 spin_lock(&ci->i_ceph_lock);
822 ret = __ceph_caps_revoking_other(ci, NULL, mask);
823 spin_unlock(&ci->i_ceph_lock);
824 dout("ceph_caps_revoking %p %s = %d\n", inode,
825 ceph_cap_string(mask), ret);
826 return ret;
827}
828
829int __ceph_caps_used(struct ceph_inode_info *ci)
830{
831 int used = 0;
832 if (ci->i_pin_ref)
833 used |= CEPH_CAP_PIN;
834 if (ci->i_rd_ref)
835 used |= CEPH_CAP_FILE_RD;
836 if (ci->i_rdcache_ref ||
837 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
838 ci->vfs_inode.i_data.nrpages))
839 used |= CEPH_CAP_FILE_CACHE;
840 if (ci->i_wr_ref)
841 used |= CEPH_CAP_FILE_WR;
842 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
843 used |= CEPH_CAP_FILE_BUFFER;
844 return used;
845}
846
847/*
848 * wanted, by virtue of open file modes
849 */
850int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
851{
852 int want = 0;
853 int mode;
854 for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
855 if (ci->i_nr_by_mode[mode])
856 want |= ceph_caps_for_mode(mode);
857 return want;
858}
859
860/*
861 * Return caps we have registered with the MDS(s) as 'wanted'.
862 */
863int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
864{
865 struct ceph_cap *cap;
866 struct rb_node *p;
867 int mds_wanted = 0;
868
869 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
870 cap = rb_entry(p, struct ceph_cap, ci_node);
871 if (!__cap_is_valid(cap))
872 continue;
873 if (cap == ci->i_auth_cap)
874 mds_wanted |= cap->mds_wanted;
875 else
876 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
877 }
878 return mds_wanted;
879}
880
881/*
882 * called under i_ceph_lock
883 */
884static int __ceph_is_any_caps(struct ceph_inode_info *ci)
885{
886 return !RB_EMPTY_ROOT(&ci->i_caps);
887}
888
889int ceph_is_any_caps(struct inode *inode)
890{
891 struct ceph_inode_info *ci = ceph_inode(inode);
892 int ret;
893
894 spin_lock(&ci->i_ceph_lock);
895 ret = __ceph_is_any_caps(ci);
896 spin_unlock(&ci->i_ceph_lock);
897
898 return ret;
899}
900
901static void drop_inode_snap_realm(struct ceph_inode_info *ci)
902{
903 struct ceph_snap_realm *realm = ci->i_snap_realm;
904 spin_lock(&realm->inodes_with_caps_lock);
905 list_del_init(&ci->i_snap_realm_item);
906 ci->i_snap_realm_counter++;
907 ci->i_snap_realm = NULL;
908 spin_unlock(&realm->inodes_with_caps_lock);
909 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
910 realm);
911}
912
913/*
914 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
915 *
916 * caller should hold i_ceph_lock.
917 * caller will not hold session s_mutex if called from destroy_inode.
918 */
919void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
920{
921 struct ceph_mds_session *session = cap->session;
922 struct ceph_inode_info *ci = cap->ci;
923 struct ceph_mds_client *mdsc =
924 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
925 int removed = 0;
926
927 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
928
929 /* remove from session list */
930 spin_lock(&session->s_cap_lock);
931 if (session->s_cap_iterator == cap) {
932 /* not yet, we are iterating over this very cap */
933 dout("__ceph_remove_cap delaying %p removal from session %p\n",
934 cap, cap->session);
935 } else {
936 list_del_init(&cap->session_caps);
937 session->s_nr_caps--;
938 cap->session = NULL;
939 removed = 1;
940 }
941 /* protect backpointer with s_cap_lock: see iterate_session_caps */
942 cap->ci = NULL;
943
944 /*
945 * s_cap_reconnect is protected by s_cap_lock. no one changes
946 * s_cap_gen while session is in the reconnect state.
947 */
948 if (queue_release &&
949 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
950 cap->queue_release = 1;
951 if (removed) {
952 list_add_tail(&cap->session_caps,
953 &session->s_cap_releases);
954 session->s_num_cap_releases++;
955 removed = 0;
956 }
957 } else {
958 cap->queue_release = 0;
959 }
960 cap->cap_ino = ci->i_vino.ino;
961
962 spin_unlock(&session->s_cap_lock);
963
964 /* remove from inode list */
965 rb_erase(&cap->ci_node, &ci->i_caps);
966 if (ci->i_auth_cap == cap)
967 ci->i_auth_cap = NULL;
968
969 if (removed)
970 ceph_put_cap(mdsc, cap);
971
972 /* when reconnect denied, we remove session caps forcibly,
973 * i_wr_ref can be non-zero. If there are ongoing write,
974 * keep i_snap_realm.
975 */
976 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
977 drop_inode_snap_realm(ci);
978
979 if (!__ceph_is_any_real_caps(ci))
980 __cap_delay_cancel(mdsc, ci);
981}
982
983/*
984 * Build and send a cap message to the given MDS.
985 *
986 * Caller should be holding s_mutex.
987 */
988static int send_cap_msg(struct ceph_mds_session *session,
989 u64 ino, u64 cid, int op,
990 int caps, int wanted, int dirty,
991 u32 seq, u64 flush_tid, u64 oldest_flush_tid,
992 u32 issue_seq, u32 mseq, u64 size, u64 max_size,
993 struct timespec *mtime, struct timespec *atime,
994 struct timespec *ctime, u64 time_warp_seq,
995 kuid_t uid, kgid_t gid, umode_t mode,
996 u64 xattr_version,
997 struct ceph_buffer *xattrs_buf,
998 u64 follows, bool inline_data)
999{
1000 struct ceph_mds_caps *fc;
1001 struct ceph_msg *msg;
1002 void *p;
1003 size_t extra_len;
1004
1005 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1006 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1007 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
1008 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
1009 ceph_cap_string(dirty),
1010 seq, issue_seq, flush_tid, oldest_flush_tid,
1011 mseq, follows, size, max_size,
1012 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1013
1014 /* flock buffer size + inline version + inline data size +
1015 * osd_epoch_barrier + oldest_flush_tid */
1016 extra_len = 4 + 8 + 4 + 4 + 8;
1017 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1018 GFP_NOFS, false);
1019 if (!msg)
1020 return -ENOMEM;
1021
1022 msg->hdr.version = cpu_to_le16(6);
1023 msg->hdr.tid = cpu_to_le64(flush_tid);
1024
1025 fc = msg->front.iov_base;
1026 memset(fc, 0, sizeof(*fc));
1027
1028 fc->cap_id = cpu_to_le64(cid);
1029 fc->op = cpu_to_le32(op);
1030 fc->seq = cpu_to_le32(seq);
1031 fc->issue_seq = cpu_to_le32(issue_seq);
1032 fc->migrate_seq = cpu_to_le32(mseq);
1033 fc->caps = cpu_to_le32(caps);
1034 fc->wanted = cpu_to_le32(wanted);
1035 fc->dirty = cpu_to_le32(dirty);
1036 fc->ino = cpu_to_le64(ino);
1037 fc->snap_follows = cpu_to_le64(follows);
1038
1039 fc->size = cpu_to_le64(size);
1040 fc->max_size = cpu_to_le64(max_size);
1041 if (mtime)
1042 ceph_encode_timespec(&fc->mtime, mtime);
1043 if (atime)
1044 ceph_encode_timespec(&fc->atime, atime);
1045 if (ctime)
1046 ceph_encode_timespec(&fc->ctime, ctime);
1047 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1048
1049 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1050 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1051 fc->mode = cpu_to_le32(mode);
1052
1053 p = fc + 1;
1054 /* flock buffer size */
1055 ceph_encode_32(&p, 0);
1056 /* inline version */
1057 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1058 /* inline data size */
1059 ceph_encode_32(&p, 0);
1060 /* osd_epoch_barrier */
1061 ceph_encode_32(&p, 0);
1062 /* oldest_flush_tid */
1063 ceph_encode_64(&p, oldest_flush_tid);
1064
1065 fc->xattr_version = cpu_to_le64(xattr_version);
1066 if (xattrs_buf) {
1067 msg->middle = ceph_buffer_get(xattrs_buf);
1068 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1069 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1070 }
1071
1072 ceph_con_send(&session->s_con, msg);
1073 return 0;
1074}
1075
1076/*
1077 * Queue cap releases when an inode is dropped from our cache. Since
1078 * inode is about to be destroyed, there is no need for i_ceph_lock.
1079 */
1080void ceph_queue_caps_release(struct inode *inode)
1081{
1082 struct ceph_inode_info *ci = ceph_inode(inode);
1083 struct rb_node *p;
1084
1085 p = rb_first(&ci->i_caps);
1086 while (p) {
1087 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1088 p = rb_next(p);
1089 __ceph_remove_cap(cap, true);
1090 }
1091}
1092
1093/*
1094 * Send a cap msg on the given inode. Update our caps state, then
1095 * drop i_ceph_lock and send the message.
1096 *
1097 * Make note of max_size reported/requested from mds, revoked caps
1098 * that have now been implemented.
1099 *
1100 * Make half-hearted attempt ot to invalidate page cache if we are
1101 * dropping RDCACHE. Note that this will leave behind locked pages
1102 * that we'll then need to deal with elsewhere.
1103 *
1104 * Return non-zero if delayed release, or we experienced an error
1105 * such that the caller should requeue + retry later.
1106 *
1107 * called with i_ceph_lock, then drops it.
1108 * caller should hold snap_rwsem (read), s_mutex.
1109 */
1110static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1111 int op, int used, int want, int retain, int flushing,
1112 u64 flush_tid, u64 oldest_flush_tid)
1113 __releases(cap->ci->i_ceph_lock)
1114{
1115 struct ceph_inode_info *ci = cap->ci;
1116 struct inode *inode = &ci->vfs_inode;
1117 u64 cap_id = cap->cap_id;
1118 int held, revoking, dropping, keep;
1119 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1120 u64 size, max_size;
1121 struct timespec mtime, atime, ctime;
1122 int wake = 0;
1123 umode_t mode;
1124 kuid_t uid;
1125 kgid_t gid;
1126 struct ceph_mds_session *session;
1127 u64 xattr_version = 0;
1128 struct ceph_buffer *xattr_blob = NULL;
1129 int delayed = 0;
1130 int ret;
1131 bool inline_data;
1132
1133 held = cap->issued | cap->implemented;
1134 revoking = cap->implemented & ~cap->issued;
1135 retain &= ~revoking;
1136 dropping = cap->issued & ~retain;
1137
1138 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1139 inode, cap, cap->session,
1140 ceph_cap_string(held), ceph_cap_string(held & retain),
1141 ceph_cap_string(revoking));
1142 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1143
1144 session = cap->session;
1145
1146 /* don't release wanted unless we've waited a bit. */
1147 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1148 time_before(jiffies, ci->i_hold_caps_min)) {
1149 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1150 ceph_cap_string(cap->issued),
1151 ceph_cap_string(cap->issued & retain),
1152 ceph_cap_string(cap->mds_wanted),
1153 ceph_cap_string(want));
1154 want |= cap->mds_wanted;
1155 retain |= cap->issued;
1156 delayed = 1;
1157 }
1158 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1159
1160 cap->issued &= retain; /* drop bits we don't want */
1161 if (cap->implemented & ~cap->issued) {
1162 /*
1163 * Wake up any waiters on wanted -> needed transition.
1164 * This is due to the weird transition from buffered
1165 * to sync IO... we need to flush dirty pages _before_
1166 * allowing sync writes to avoid reordering.
1167 */
1168 wake = 1;
1169 }
1170 cap->implemented &= cap->issued | used;
1171 cap->mds_wanted = want;
1172
1173 follows = flushing ? ci->i_head_snapc->seq : 0;
1174
1175 keep = cap->implemented;
1176 seq = cap->seq;
1177 issue_seq = cap->issue_seq;
1178 mseq = cap->mseq;
1179 size = inode->i_size;
1180 ci->i_reported_size = size;
1181 max_size = ci->i_wanted_max_size;
1182 ci->i_requested_max_size = max_size;
1183 mtime = inode->i_mtime;
1184 atime = inode->i_atime;
1185 ctime = inode->i_ctime;
1186 time_warp_seq = ci->i_time_warp_seq;
1187 uid = inode->i_uid;
1188 gid = inode->i_gid;
1189 mode = inode->i_mode;
1190
1191 if (flushing & CEPH_CAP_XATTR_EXCL) {
1192 __ceph_build_xattrs_blob(ci);
1193 xattr_blob = ci->i_xattrs.blob;
1194 xattr_version = ci->i_xattrs.version;
1195 }
1196
1197 inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1198
1199 spin_unlock(&ci->i_ceph_lock);
1200
1201 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1202 op, keep, want, flushing, seq,
1203 flush_tid, oldest_flush_tid, issue_seq, mseq,
1204 size, max_size, &mtime, &atime, &ctime, time_warp_seq,
1205 uid, gid, mode, xattr_version, xattr_blob,
1206 follows, inline_data);
1207 if (ret < 0) {
1208 dout("error sending cap msg, must requeue %p\n", inode);
1209 delayed = 1;
1210 }
1211
1212 if (wake)
1213 wake_up_all(&ci->i_cap_wq);
1214
1215 return delayed;
1216}
1217
1218/*
1219 * When a snapshot is taken, clients accumulate dirty metadata on
1220 * inodes with capabilities in ceph_cap_snaps to describe the file
1221 * state at the time the snapshot was taken. This must be flushed
1222 * asynchronously back to the MDS once sync writes complete and dirty
1223 * data is written out.
1224 *
1225 * Unless @kick is true, skip cap_snaps that were already sent to
1226 * the MDS (i.e., during this session).
1227 *
1228 * Called under i_ceph_lock. Takes s_mutex as needed.
1229 */
1230void __ceph_flush_snaps(struct ceph_inode_info *ci,
1231 struct ceph_mds_session **psession,
1232 int kick)
1233 __releases(ci->i_ceph_lock)
1234 __acquires(ci->i_ceph_lock)
1235{
1236 struct inode *inode = &ci->vfs_inode;
1237 int mds;
1238 struct ceph_cap_snap *capsnap;
1239 u32 mseq;
1240 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1241 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1242 session->s_mutex */
1243 u64 next_follows = 0; /* keep track of how far we've gotten through the
1244 i_cap_snaps list, and skip these entries next time
1245 around to avoid an infinite loop */
1246
1247 if (psession)
1248 session = *psession;
1249
1250 dout("__flush_snaps %p\n", inode);
1251retry:
1252 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1253 /* avoid an infiniute loop after retry */
1254 if (capsnap->follows < next_follows)
1255 continue;
1256 /*
1257 * we need to wait for sync writes to complete and for dirty
1258 * pages to be written out.
1259 */
1260 if (capsnap->dirty_pages || capsnap->writing)
1261 break;
1262
1263 /* should be removed by ceph_try_drop_cap_snap() */
1264 BUG_ON(!capsnap->need_flush);
1265
1266 /* pick mds, take s_mutex */
1267 if (ci->i_auth_cap == NULL) {
1268 dout("no auth cap (migrating?), doing nothing\n");
1269 goto out;
1270 }
1271
1272 /* only flush each capsnap once */
1273 if (!kick && !list_empty(&capsnap->flushing_item)) {
1274 dout("already flushed %p, skipping\n", capsnap);
1275 continue;
1276 }
1277
1278 mds = ci->i_auth_cap->session->s_mds;
1279 mseq = ci->i_auth_cap->mseq;
1280
1281 if (session && session->s_mds != mds) {
1282 dout("oops, wrong session %p mutex\n", session);
1283 if (kick)
1284 goto out;
1285
1286 mutex_unlock(&session->s_mutex);
1287 ceph_put_mds_session(session);
1288 session = NULL;
1289 }
1290 if (!session) {
1291 spin_unlock(&ci->i_ceph_lock);
1292 mutex_lock(&mdsc->mutex);
1293 session = __ceph_lookup_mds_session(mdsc, mds);
1294 mutex_unlock(&mdsc->mutex);
1295 if (session) {
1296 dout("inverting session/ino locks on %p\n",
1297 session);
1298 mutex_lock(&session->s_mutex);
1299 }
1300 /*
1301 * if session == NULL, we raced against a cap
1302 * deletion or migration. retry, and we'll
1303 * get a better @mds value next time.
1304 */
1305 spin_lock(&ci->i_ceph_lock);
1306 goto retry;
1307 }
1308
1309 spin_lock(&mdsc->cap_dirty_lock);
1310 capsnap->flush_tid = ++mdsc->last_cap_flush_tid;
1311 spin_unlock(&mdsc->cap_dirty_lock);
1312
1313 atomic_inc(&capsnap->nref);
1314 if (list_empty(&capsnap->flushing_item))
1315 list_add_tail(&capsnap->flushing_item,
1316 &session->s_cap_snaps_flushing);
1317 spin_unlock(&ci->i_ceph_lock);
1318
1319 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1320 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1321 send_cap_msg(session, ceph_vino(inode).ino, 0,
1322 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1323 capsnap->dirty, 0, capsnap->flush_tid, 0,
1324 0, mseq, capsnap->size, 0,
1325 &capsnap->mtime, &capsnap->atime,
1326 &capsnap->ctime, capsnap->time_warp_seq,
1327 capsnap->uid, capsnap->gid, capsnap->mode,
1328 capsnap->xattr_version, capsnap->xattr_blob,
1329 capsnap->follows, capsnap->inline_data);
1330
1331 next_follows = capsnap->follows + 1;
1332 ceph_put_cap_snap(capsnap);
1333
1334 spin_lock(&ci->i_ceph_lock);
1335 goto retry;
1336 }
1337
1338 /* we flushed them all; remove this inode from the queue */
1339 spin_lock(&mdsc->snap_flush_lock);
1340 list_del_init(&ci->i_snap_flush_item);
1341 spin_unlock(&mdsc->snap_flush_lock);
1342
1343out:
1344 if (psession)
1345 *psession = session;
1346 else if (session) {
1347 mutex_unlock(&session->s_mutex);
1348 ceph_put_mds_session(session);
1349 }
1350}
1351
1352static void ceph_flush_snaps(struct ceph_inode_info *ci)
1353{
1354 spin_lock(&ci->i_ceph_lock);
1355 __ceph_flush_snaps(ci, NULL, 0);
1356 spin_unlock(&ci->i_ceph_lock);
1357}
1358
1359/*
1360 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1361 * Caller is then responsible for calling __mark_inode_dirty with the
1362 * returned flags value.
1363 */
1364int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1365 struct ceph_cap_flush **pcf)
1366{
1367 struct ceph_mds_client *mdsc =
1368 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1369 struct inode *inode = &ci->vfs_inode;
1370 int was = ci->i_dirty_caps;
1371 int dirty = 0;
1372
1373 if (!ci->i_auth_cap) {
1374 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1375 "but no auth cap (session was closed?)\n",
1376 inode, ceph_ino(inode), ceph_cap_string(mask));
1377 return 0;
1378 }
1379
1380 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1381 ceph_cap_string(mask), ceph_cap_string(was),
1382 ceph_cap_string(was | mask));
1383 ci->i_dirty_caps |= mask;
1384 if (was == 0) {
1385 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1386 swap(ci->i_prealloc_cap_flush, *pcf);
1387
1388 if (!ci->i_head_snapc) {
1389 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1390 ci->i_head_snapc = ceph_get_snap_context(
1391 ci->i_snap_realm->cached_context);
1392 }
1393 dout(" inode %p now dirty snapc %p auth cap %p\n",
1394 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1395 BUG_ON(!list_empty(&ci->i_dirty_item));
1396 spin_lock(&mdsc->cap_dirty_lock);
1397 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1398 spin_unlock(&mdsc->cap_dirty_lock);
1399 if (ci->i_flushing_caps == 0) {
1400 ihold(inode);
1401 dirty |= I_DIRTY_SYNC;
1402 }
1403 } else {
1404 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1405 }
1406 BUG_ON(list_empty(&ci->i_dirty_item));
1407 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1408 (mask & CEPH_CAP_FILE_BUFFER))
1409 dirty |= I_DIRTY_DATASYNC;
1410 __cap_delay_requeue(mdsc, ci);
1411 return dirty;
1412}
1413
1414static void __add_cap_flushing_to_inode(struct ceph_inode_info *ci,
1415 struct ceph_cap_flush *cf)
1416{
1417 struct rb_node **p = &ci->i_cap_flush_tree.rb_node;
1418 struct rb_node *parent = NULL;
1419 struct ceph_cap_flush *other = NULL;
1420
1421 while (*p) {
1422 parent = *p;
1423 other = rb_entry(parent, struct ceph_cap_flush, i_node);
1424
1425 if (cf->tid < other->tid)
1426 p = &(*p)->rb_left;
1427 else if (cf->tid > other->tid)
1428 p = &(*p)->rb_right;
1429 else
1430 BUG();
1431 }
1432
1433 rb_link_node(&cf->i_node, parent, p);
1434 rb_insert_color(&cf->i_node, &ci->i_cap_flush_tree);
1435}
1436
1437static void __add_cap_flushing_to_mdsc(struct ceph_mds_client *mdsc,
1438 struct ceph_cap_flush *cf)
1439{
1440 struct rb_node **p = &mdsc->cap_flush_tree.rb_node;
1441 struct rb_node *parent = NULL;
1442 struct ceph_cap_flush *other = NULL;
1443
1444 while (*p) {
1445 parent = *p;
1446 other = rb_entry(parent, struct ceph_cap_flush, g_node);
1447
1448 if (cf->tid < other->tid)
1449 p = &(*p)->rb_left;
1450 else if (cf->tid > other->tid)
1451 p = &(*p)->rb_right;
1452 else
1453 BUG();
1454 }
1455
1456 rb_link_node(&cf->g_node, parent, p);
1457 rb_insert_color(&cf->g_node, &mdsc->cap_flush_tree);
1458}
1459
1460struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1461{
1462 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1463}
1464
1465void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1466{
1467 if (cf)
1468 kmem_cache_free(ceph_cap_flush_cachep, cf);
1469}
1470
1471static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1472{
1473 struct rb_node *n = rb_first(&mdsc->cap_flush_tree);
1474 if (n) {
1475 struct ceph_cap_flush *cf =
1476 rb_entry(n, struct ceph_cap_flush, g_node);
1477 return cf->tid;
1478 }
1479 return 0;
1480}
1481
1482/*
1483 * Add dirty inode to the flushing list. Assigned a seq number so we
1484 * can wait for caps to flush without starving.
1485 *
1486 * Called under i_ceph_lock.
1487 */
1488static int __mark_caps_flushing(struct inode *inode,
1489 struct ceph_mds_session *session,
1490 u64 *flush_tid, u64 *oldest_flush_tid)
1491{
1492 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1493 struct ceph_inode_info *ci = ceph_inode(inode);
1494 struct ceph_cap_flush *cf = NULL;
1495 int flushing;
1496
1497 BUG_ON(ci->i_dirty_caps == 0);
1498 BUG_ON(list_empty(&ci->i_dirty_item));
1499 BUG_ON(!ci->i_prealloc_cap_flush);
1500
1501 flushing = ci->i_dirty_caps;
1502 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1503 ceph_cap_string(flushing),
1504 ceph_cap_string(ci->i_flushing_caps),
1505 ceph_cap_string(ci->i_flushing_caps | flushing));
1506 ci->i_flushing_caps |= flushing;
1507 ci->i_dirty_caps = 0;
1508 dout(" inode %p now !dirty\n", inode);
1509
1510 swap(cf, ci->i_prealloc_cap_flush);
1511 cf->caps = flushing;
1512
1513 spin_lock(&mdsc->cap_dirty_lock);
1514 list_del_init(&ci->i_dirty_item);
1515
1516 cf->tid = ++mdsc->last_cap_flush_tid;
1517 __add_cap_flushing_to_mdsc(mdsc, cf);
1518 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1519
1520 if (list_empty(&ci->i_flushing_item)) {
1521 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1522 mdsc->num_cap_flushing++;
1523 dout(" inode %p now flushing tid %llu\n", inode, cf->tid);
1524 } else {
1525 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1526 dout(" inode %p now flushing (more) tid %llu\n",
1527 inode, cf->tid);
1528 }
1529 spin_unlock(&mdsc->cap_dirty_lock);
1530
1531 __add_cap_flushing_to_inode(ci, cf);
1532
1533 *flush_tid = cf->tid;
1534 return flushing;
1535}
1536
1537/*
1538 * try to invalidate mapping pages without blocking.
1539 */
1540static int try_nonblocking_invalidate(struct inode *inode)
1541{
1542 struct ceph_inode_info *ci = ceph_inode(inode);
1543 u32 invalidating_gen = ci->i_rdcache_gen;
1544
1545 spin_unlock(&ci->i_ceph_lock);
1546 invalidate_mapping_pages(&inode->i_data, 0, -1);
1547 spin_lock(&ci->i_ceph_lock);
1548
1549 if (inode->i_data.nrpages == 0 &&
1550 invalidating_gen == ci->i_rdcache_gen) {
1551 /* success. */
1552 dout("try_nonblocking_invalidate %p success\n", inode);
1553 /* save any racing async invalidate some trouble */
1554 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1555 return 0;
1556 }
1557 dout("try_nonblocking_invalidate %p failed\n", inode);
1558 return -1;
1559}
1560
1561/*
1562 * Swiss army knife function to examine currently used and wanted
1563 * versus held caps. Release, flush, ack revoked caps to mds as
1564 * appropriate.
1565 *
1566 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1567 * cap release further.
1568 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1569 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1570 * further delay.
1571 */
1572void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1573 struct ceph_mds_session *session)
1574{
1575 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1576 struct ceph_mds_client *mdsc = fsc->mdsc;
1577 struct inode *inode = &ci->vfs_inode;
1578 struct ceph_cap *cap;
1579 u64 flush_tid, oldest_flush_tid;
1580 int file_wanted, used, cap_used;
1581 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1582 int issued, implemented, want, retain, revoking, flushing = 0;
1583 int mds = -1; /* keep track of how far we've gone through i_caps list
1584 to avoid an infinite loop on retry */
1585 struct rb_node *p;
1586 int tried_invalidate = 0;
1587 int delayed = 0, sent = 0, force_requeue = 0, num;
1588 int queue_invalidate = 0;
1589 int is_delayed = flags & CHECK_CAPS_NODELAY;
1590
1591 /* if we are unmounting, flush any unused caps immediately. */
1592 if (mdsc->stopping)
1593 is_delayed = 1;
1594
1595 spin_lock(&ci->i_ceph_lock);
1596
1597 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1598 flags |= CHECK_CAPS_FLUSH;
1599
1600 /* flush snaps first time around only */
1601 if (!list_empty(&ci->i_cap_snaps))
1602 __ceph_flush_snaps(ci, &session, 0);
1603 goto retry_locked;
1604retry:
1605 spin_lock(&ci->i_ceph_lock);
1606retry_locked:
1607 file_wanted = __ceph_caps_file_wanted(ci);
1608 used = __ceph_caps_used(ci);
1609 issued = __ceph_caps_issued(ci, &implemented);
1610 revoking = implemented & ~issued;
1611
1612 want = file_wanted;
1613 retain = file_wanted | used | CEPH_CAP_PIN;
1614 if (!mdsc->stopping && inode->i_nlink > 0) {
1615 if (file_wanted) {
1616 retain |= CEPH_CAP_ANY; /* be greedy */
1617 } else if (S_ISDIR(inode->i_mode) &&
1618 (issued & CEPH_CAP_FILE_SHARED) &&
1619 __ceph_dir_is_complete(ci)) {
1620 /*
1621 * If a directory is complete, we want to keep
1622 * the exclusive cap. So that MDS does not end up
1623 * revoking the shared cap on every create/unlink
1624 * operation.
1625 */
1626 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1627 retain |= want;
1628 } else {
1629
1630 retain |= CEPH_CAP_ANY_SHARED;
1631 /*
1632 * keep RD only if we didn't have the file open RW,
1633 * because then the mds would revoke it anyway to
1634 * journal max_size=0.
1635 */
1636 if (ci->i_max_size == 0)
1637 retain |= CEPH_CAP_ANY_RD;
1638 }
1639 }
1640
1641 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1642 " issued %s revoking %s retain %s %s%s%s\n", inode,
1643 ceph_cap_string(file_wanted),
1644 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1645 ceph_cap_string(ci->i_flushing_caps),
1646 ceph_cap_string(issued), ceph_cap_string(revoking),
1647 ceph_cap_string(retain),
1648 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1649 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1650 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1651
1652 /*
1653 * If we no longer need to hold onto old our caps, and we may
1654 * have cached pages, but don't want them, then try to invalidate.
1655 * If we fail, it's because pages are locked.... try again later.
1656 */
1657 if ((!is_delayed || mdsc->stopping) &&
1658 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1659 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1660 inode->i_data.nrpages && /* have cached pages */
1661 (revoking & (CEPH_CAP_FILE_CACHE|
1662 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1663 !tried_invalidate) {
1664 dout("check_caps trying to invalidate on %p\n", inode);
1665 if (try_nonblocking_invalidate(inode) < 0) {
1666 if (revoking & (CEPH_CAP_FILE_CACHE|
1667 CEPH_CAP_FILE_LAZYIO)) {
1668 dout("check_caps queuing invalidate\n");
1669 queue_invalidate = 1;
1670 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1671 } else {
1672 dout("check_caps failed to invalidate pages\n");
1673 /* we failed to invalidate pages. check these
1674 caps again later. */
1675 force_requeue = 1;
1676 __cap_set_timeouts(mdsc, ci);
1677 }
1678 }
1679 tried_invalidate = 1;
1680 goto retry_locked;
1681 }
1682
1683 num = 0;
1684 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1685 cap = rb_entry(p, struct ceph_cap, ci_node);
1686 num++;
1687
1688 /* avoid looping forever */
1689 if (mds >= cap->mds ||
1690 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1691 continue;
1692
1693 /* NOTE: no side-effects allowed, until we take s_mutex */
1694
1695 cap_used = used;
1696 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1697 cap_used &= ~ci->i_auth_cap->issued;
1698
1699 revoking = cap->implemented & ~cap->issued;
1700 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1701 cap->mds, cap, ceph_cap_string(cap->issued),
1702 ceph_cap_string(cap_used),
1703 ceph_cap_string(cap->implemented),
1704 ceph_cap_string(revoking));
1705
1706 if (cap == ci->i_auth_cap &&
1707 (cap->issued & CEPH_CAP_FILE_WR)) {
1708 /* request larger max_size from MDS? */
1709 if (ci->i_wanted_max_size > ci->i_max_size &&
1710 ci->i_wanted_max_size > ci->i_requested_max_size) {
1711 dout("requesting new max_size\n");
1712 goto ack;
1713 }
1714
1715 /* approaching file_max? */
1716 if ((inode->i_size << 1) >= ci->i_max_size &&
1717 (ci->i_reported_size << 1) < ci->i_max_size) {
1718 dout("i_size approaching max_size\n");
1719 goto ack;
1720 }
1721 }
1722 /* flush anything dirty? */
1723 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1724 ci->i_dirty_caps) {
1725 dout("flushing dirty caps\n");
1726 goto ack;
1727 }
1728
1729 /* completed revocation? going down and there are no caps? */
1730 if (revoking && (revoking & cap_used) == 0) {
1731 dout("completed revocation of %s\n",
1732 ceph_cap_string(cap->implemented & ~cap->issued));
1733 goto ack;
1734 }
1735
1736 /* want more caps from mds? */
1737 if (want & ~(cap->mds_wanted | cap->issued))
1738 goto ack;
1739
1740 /* things we might delay */
1741 if ((cap->issued & ~retain) == 0 &&
1742 cap->mds_wanted == want)
1743 continue; /* nope, all good */
1744
1745 if (is_delayed)
1746 goto ack;
1747
1748 /* delay? */
1749 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1750 time_before(jiffies, ci->i_hold_caps_max)) {
1751 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1752 ceph_cap_string(cap->issued),
1753 ceph_cap_string(cap->issued & retain),
1754 ceph_cap_string(cap->mds_wanted),
1755 ceph_cap_string(want));
1756 delayed++;
1757 continue;
1758 }
1759
1760ack:
1761 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1762 dout(" skipping %p I_NOFLUSH set\n", inode);
1763 continue;
1764 }
1765
1766 if (session && session != cap->session) {
1767 dout("oops, wrong session %p mutex\n", session);
1768 mutex_unlock(&session->s_mutex);
1769 session = NULL;
1770 }
1771 if (!session) {
1772 session = cap->session;
1773 if (mutex_trylock(&session->s_mutex) == 0) {
1774 dout("inverting session/ino locks on %p\n",
1775 session);
1776 spin_unlock(&ci->i_ceph_lock);
1777 if (took_snap_rwsem) {
1778 up_read(&mdsc->snap_rwsem);
1779 took_snap_rwsem = 0;
1780 }
1781 mutex_lock(&session->s_mutex);
1782 goto retry;
1783 }
1784 }
1785 /* take snap_rwsem after session mutex */
1786 if (!took_snap_rwsem) {
1787 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1788 dout("inverting snap/in locks on %p\n",
1789 inode);
1790 spin_unlock(&ci->i_ceph_lock);
1791 down_read(&mdsc->snap_rwsem);
1792 took_snap_rwsem = 1;
1793 goto retry;
1794 }
1795 took_snap_rwsem = 1;
1796 }
1797
1798 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1799 flushing = __mark_caps_flushing(inode, session,
1800 &flush_tid,
1801 &oldest_flush_tid);
1802 } else {
1803 flushing = 0;
1804 flush_tid = 0;
1805 spin_lock(&mdsc->cap_dirty_lock);
1806 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1807 spin_unlock(&mdsc->cap_dirty_lock);
1808 }
1809
1810 mds = cap->mds; /* remember mds, so we don't repeat */
1811 sent++;
1812
1813 /* __send_cap drops i_ceph_lock */
1814 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1815 want, retain, flushing,
1816 flush_tid, oldest_flush_tid);
1817 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1818 }
1819
1820 /*
1821 * Reschedule delayed caps release if we delayed anything,
1822 * otherwise cancel.
1823 */
1824 if (delayed && is_delayed)
1825 force_requeue = 1; /* __send_cap delayed release; requeue */
1826 if (!delayed && !is_delayed)
1827 __cap_delay_cancel(mdsc, ci);
1828 else if (!is_delayed || force_requeue)
1829 __cap_delay_requeue(mdsc, ci);
1830
1831 spin_unlock(&ci->i_ceph_lock);
1832
1833 if (queue_invalidate)
1834 ceph_queue_invalidate(inode);
1835
1836 if (session)
1837 mutex_unlock(&session->s_mutex);
1838 if (took_snap_rwsem)
1839 up_read(&mdsc->snap_rwsem);
1840}
1841
1842/*
1843 * Try to flush dirty caps back to the auth mds.
1844 */
1845static int try_flush_caps(struct inode *inode, u64 *ptid)
1846{
1847 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1848 struct ceph_inode_info *ci = ceph_inode(inode);
1849 struct ceph_mds_session *session = NULL;
1850 int flushing = 0;
1851 u64 flush_tid = 0, oldest_flush_tid = 0;
1852
1853retry:
1854 spin_lock(&ci->i_ceph_lock);
1855 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1856 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1857 goto out;
1858 }
1859 if (ci->i_dirty_caps && ci->i_auth_cap) {
1860 struct ceph_cap *cap = ci->i_auth_cap;
1861 int used = __ceph_caps_used(ci);
1862 int want = __ceph_caps_wanted(ci);
1863 int delayed;
1864
1865 if (!session || session != cap->session) {
1866 spin_unlock(&ci->i_ceph_lock);
1867 if (session)
1868 mutex_unlock(&session->s_mutex);
1869 session = cap->session;
1870 mutex_lock(&session->s_mutex);
1871 goto retry;
1872 }
1873 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1874 goto out;
1875
1876 flushing = __mark_caps_flushing(inode, session, &flush_tid,
1877 &oldest_flush_tid);
1878
1879 /* __send_cap drops i_ceph_lock */
1880 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1881 (cap->issued | cap->implemented),
1882 flushing, flush_tid, oldest_flush_tid);
1883
1884 if (delayed) {
1885 spin_lock(&ci->i_ceph_lock);
1886 __cap_delay_requeue(mdsc, ci);
1887 spin_unlock(&ci->i_ceph_lock);
1888 }
1889 } else {
1890 struct rb_node *n = rb_last(&ci->i_cap_flush_tree);
1891 if (n) {
1892 struct ceph_cap_flush *cf =
1893 rb_entry(n, struct ceph_cap_flush, i_node);
1894 flush_tid = cf->tid;
1895 }
1896 flushing = ci->i_flushing_caps;
1897 spin_unlock(&ci->i_ceph_lock);
1898 }
1899out:
1900 if (session)
1901 mutex_unlock(&session->s_mutex);
1902
1903 *ptid = flush_tid;
1904 return flushing;
1905}
1906
1907/*
1908 * Return true if we've flushed caps through the given flush_tid.
1909 */
1910static int caps_are_flushed(struct inode *inode, u64 flush_tid)
1911{
1912 struct ceph_inode_info *ci = ceph_inode(inode);
1913 struct ceph_cap_flush *cf;
1914 struct rb_node *n;
1915 int ret = 1;
1916
1917 spin_lock(&ci->i_ceph_lock);
1918 n = rb_first(&ci->i_cap_flush_tree);
1919 if (n) {
1920 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1921 if (cf->tid <= flush_tid)
1922 ret = 0;
1923 }
1924 spin_unlock(&ci->i_ceph_lock);
1925 return ret;
1926}
1927
1928/*
1929 * Wait on any unsafe replies for the given inode. First wait on the
1930 * newest request, and make that the upper bound. Then, if there are
1931 * more requests, keep waiting on the oldest as long as it is still older
1932 * than the original request.
1933 */
1934static void sync_write_wait(struct inode *inode)
1935{
1936 struct ceph_inode_info *ci = ceph_inode(inode);
1937 struct list_head *head = &ci->i_unsafe_writes;
1938 struct ceph_osd_request *req;
1939 u64 last_tid;
1940
1941 if (!S_ISREG(inode->i_mode))
1942 return;
1943
1944 spin_lock(&ci->i_unsafe_lock);
1945 if (list_empty(head))
1946 goto out;
1947
1948 /* set upper bound as _last_ entry in chain */
1949 req = list_last_entry(head, struct ceph_osd_request,
1950 r_unsafe_item);
1951 last_tid = req->r_tid;
1952
1953 do {
1954 ceph_osdc_get_request(req);
1955 spin_unlock(&ci->i_unsafe_lock);
1956 dout("sync_write_wait on tid %llu (until %llu)\n",
1957 req->r_tid, last_tid);
1958 wait_for_completion(&req->r_safe_completion);
1959 spin_lock(&ci->i_unsafe_lock);
1960 ceph_osdc_put_request(req);
1961
1962 /*
1963 * from here on look at first entry in chain, since we
1964 * only want to wait for anything older than last_tid
1965 */
1966 if (list_empty(head))
1967 break;
1968 req = list_first_entry(head, struct ceph_osd_request,
1969 r_unsafe_item);
1970 } while (req->r_tid < last_tid);
1971out:
1972 spin_unlock(&ci->i_unsafe_lock);
1973}
1974
1975/*
1976 * wait for any unsafe requests to complete.
1977 */
1978static int unsafe_request_wait(struct inode *inode)
1979{
1980 struct ceph_inode_info *ci = ceph_inode(inode);
1981 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
1982 int ret, err = 0;
1983
1984 spin_lock(&ci->i_unsafe_lock);
1985 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
1986 req1 = list_last_entry(&ci->i_unsafe_dirops,
1987 struct ceph_mds_request,
1988 r_unsafe_dir_item);
1989 ceph_mdsc_get_request(req1);
1990 }
1991 if (!list_empty(&ci->i_unsafe_iops)) {
1992 req2 = list_last_entry(&ci->i_unsafe_iops,
1993 struct ceph_mds_request,
1994 r_unsafe_target_item);
1995 ceph_mdsc_get_request(req2);
1996 }
1997 spin_unlock(&ci->i_unsafe_lock);
1998
1999 dout("unsafe_requeset_wait %p wait on tid %llu %llu\n",
2000 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2001 if (req1) {
2002 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2003 ceph_timeout_jiffies(req1->r_timeout));
2004 if (ret)
2005 err = -EIO;
2006 ceph_mdsc_put_request(req1);
2007 }
2008 if (req2) {
2009 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2010 ceph_timeout_jiffies(req2->r_timeout));
2011 if (ret)
2012 err = -EIO;
2013 ceph_mdsc_put_request(req2);
2014 }
2015 return err;
2016}
2017
2018int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2019{
2020 struct inode *inode = file->f_mapping->host;
2021 struct ceph_inode_info *ci = ceph_inode(inode);
2022 u64 flush_tid;
2023 int ret;
2024 int dirty;
2025
2026 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2027 sync_write_wait(inode);
2028
2029 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
2030 if (ret < 0)
2031 goto out;
2032
2033 if (datasync)
2034 goto out;
2035
2036 inode_lock(inode);
2037
2038 dirty = try_flush_caps(inode, &flush_tid);
2039 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2040
2041 ret = unsafe_request_wait(inode);
2042
2043 /*
2044 * only wait on non-file metadata writeback (the mds
2045 * can recover size and mtime, so we don't need to
2046 * wait for that)
2047 */
2048 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2049 ret = wait_event_interruptible(ci->i_cap_wq,
2050 caps_are_flushed(inode, flush_tid));
2051 }
2052 inode_unlock(inode);
2053out:
2054 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2055 return ret;
2056}
2057
2058/*
2059 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2060 * queue inode for flush but don't do so immediately, because we can
2061 * get by with fewer MDS messages if we wait for data writeback to
2062 * complete first.
2063 */
2064int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2065{
2066 struct ceph_inode_info *ci = ceph_inode(inode);
2067 u64 flush_tid;
2068 int err = 0;
2069 int dirty;
2070 int wait = wbc->sync_mode == WB_SYNC_ALL;
2071
2072 dout("write_inode %p wait=%d\n", inode, wait);
2073 if (wait) {
2074 dirty = try_flush_caps(inode, &flush_tid);
2075 if (dirty)
2076 err = wait_event_interruptible(ci->i_cap_wq,
2077 caps_are_flushed(inode, flush_tid));
2078 } else {
2079 struct ceph_mds_client *mdsc =
2080 ceph_sb_to_client(inode->i_sb)->mdsc;
2081
2082 spin_lock(&ci->i_ceph_lock);
2083 if (__ceph_caps_dirty(ci))
2084 __cap_delay_requeue_front(mdsc, ci);
2085 spin_unlock(&ci->i_ceph_lock);
2086 }
2087 return err;
2088}
2089
2090/*
2091 * After a recovering MDS goes active, we need to resend any caps
2092 * we were flushing.
2093 *
2094 * Caller holds session->s_mutex.
2095 */
2096static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
2097 struct ceph_mds_session *session)
2098{
2099 struct ceph_cap_snap *capsnap;
2100
2101 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
2102 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
2103 flushing_item) {
2104 struct ceph_inode_info *ci = capsnap->ci;
2105 struct inode *inode = &ci->vfs_inode;
2106 struct ceph_cap *cap;
2107
2108 spin_lock(&ci->i_ceph_lock);
2109 cap = ci->i_auth_cap;
2110 if (cap && cap->session == session) {
2111 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
2112 cap, capsnap);
2113 __ceph_flush_snaps(ci, &session, 1);
2114 } else {
2115 pr_err("%p auth cap %p not mds%d ???\n", inode,
2116 cap, session->s_mds);
2117 }
2118 spin_unlock(&ci->i_ceph_lock);
2119 }
2120}
2121
2122static int __kick_flushing_caps(struct ceph_mds_client *mdsc,
2123 struct ceph_mds_session *session,
2124 struct ceph_inode_info *ci)
2125{
2126 struct inode *inode = &ci->vfs_inode;
2127 struct ceph_cap *cap;
2128 struct ceph_cap_flush *cf;
2129 struct rb_node *n;
2130 int delayed = 0;
2131 u64 first_tid = 0;
2132 u64 oldest_flush_tid;
2133
2134 spin_lock(&mdsc->cap_dirty_lock);
2135 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2136 spin_unlock(&mdsc->cap_dirty_lock);
2137
2138 while (true) {
2139 spin_lock(&ci->i_ceph_lock);
2140 cap = ci->i_auth_cap;
2141 if (!(cap && cap->session == session)) {
2142 pr_err("%p auth cap %p not mds%d ???\n", inode,
2143 cap, session->s_mds);
2144 spin_unlock(&ci->i_ceph_lock);
2145 break;
2146 }
2147
2148 for (n = rb_first(&ci->i_cap_flush_tree); n; n = rb_next(n)) {
2149 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2150 if (cf->tid >= first_tid)
2151 break;
2152 }
2153 if (!n) {
2154 spin_unlock(&ci->i_ceph_lock);
2155 break;
2156 }
2157
2158 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2159
2160 first_tid = cf->tid + 1;
2161
2162 dout("kick_flushing_caps %p cap %p tid %llu %s\n", inode,
2163 cap, cf->tid, ceph_cap_string(cf->caps));
2164 delayed |= __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2165 __ceph_caps_used(ci),
2166 __ceph_caps_wanted(ci),
2167 cap->issued | cap->implemented,
2168 cf->caps, cf->tid, oldest_flush_tid);
2169 }
2170 return delayed;
2171}
2172
2173void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2174 struct ceph_mds_session *session)
2175{
2176 struct ceph_inode_info *ci;
2177 struct ceph_cap *cap;
2178
2179 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2180 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2181 spin_lock(&ci->i_ceph_lock);
2182 cap = ci->i_auth_cap;
2183 if (!(cap && cap->session == session)) {
2184 pr_err("%p auth cap %p not mds%d ???\n",
2185 &ci->vfs_inode, cap, session->s_mds);
2186 spin_unlock(&ci->i_ceph_lock);
2187 continue;
2188 }
2189
2190
2191 /*
2192 * if flushing caps were revoked, we re-send the cap flush
2193 * in client reconnect stage. This guarantees MDS * processes
2194 * the cap flush message before issuing the flushing caps to
2195 * other client.
2196 */
2197 if ((cap->issued & ci->i_flushing_caps) !=
2198 ci->i_flushing_caps) {
2199 spin_unlock(&ci->i_ceph_lock);
2200 if (!__kick_flushing_caps(mdsc, session, ci))
2201 continue;
2202 spin_lock(&ci->i_ceph_lock);
2203 }
2204
2205 spin_unlock(&ci->i_ceph_lock);
2206 }
2207}
2208
2209void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2210 struct ceph_mds_session *session)
2211{
2212 struct ceph_inode_info *ci;
2213
2214 kick_flushing_capsnaps(mdsc, session);
2215
2216 dout("kick_flushing_caps mds%d\n", session->s_mds);
2217 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2218 int delayed = __kick_flushing_caps(mdsc, session, ci);
2219 if (delayed) {
2220 spin_lock(&ci->i_ceph_lock);
2221 __cap_delay_requeue(mdsc, ci);
2222 spin_unlock(&ci->i_ceph_lock);
2223 }
2224 }
2225}
2226
2227static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2228 struct ceph_mds_session *session,
2229 struct inode *inode)
2230{
2231 struct ceph_inode_info *ci = ceph_inode(inode);
2232 struct ceph_cap *cap;
2233
2234 spin_lock(&ci->i_ceph_lock);
2235 cap = ci->i_auth_cap;
2236 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2237 ceph_cap_string(ci->i_flushing_caps));
2238
2239 __ceph_flush_snaps(ci, &session, 1);
2240
2241 if (ci->i_flushing_caps) {
2242 int delayed;
2243
2244 spin_lock(&mdsc->cap_dirty_lock);
2245 list_move_tail(&ci->i_flushing_item,
2246 &cap->session->s_cap_flushing);
2247 spin_unlock(&mdsc->cap_dirty_lock);
2248
2249 spin_unlock(&ci->i_ceph_lock);
2250
2251 delayed = __kick_flushing_caps(mdsc, session, ci);
2252 if (delayed) {
2253 spin_lock(&ci->i_ceph_lock);
2254 __cap_delay_requeue(mdsc, ci);
2255 spin_unlock(&ci->i_ceph_lock);
2256 }
2257 } else {
2258 spin_unlock(&ci->i_ceph_lock);
2259 }
2260}
2261
2262
2263/*
2264 * Take references to capabilities we hold, so that we don't release
2265 * them to the MDS prematurely.
2266 *
2267 * Protected by i_ceph_lock.
2268 */
2269static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2270 bool snap_rwsem_locked)
2271{
2272 if (got & CEPH_CAP_PIN)
2273 ci->i_pin_ref++;
2274 if (got & CEPH_CAP_FILE_RD)
2275 ci->i_rd_ref++;
2276 if (got & CEPH_CAP_FILE_CACHE)
2277 ci->i_rdcache_ref++;
2278 if (got & CEPH_CAP_FILE_WR) {
2279 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2280 BUG_ON(!snap_rwsem_locked);
2281 ci->i_head_snapc = ceph_get_snap_context(
2282 ci->i_snap_realm->cached_context);
2283 }
2284 ci->i_wr_ref++;
2285 }
2286 if (got & CEPH_CAP_FILE_BUFFER) {
2287 if (ci->i_wb_ref == 0)
2288 ihold(&ci->vfs_inode);
2289 ci->i_wb_ref++;
2290 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2291 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2292 }
2293}
2294
2295/*
2296 * Try to grab cap references. Specify those refs we @want, and the
2297 * minimal set we @need. Also include the larger offset we are writing
2298 * to (when applicable), and check against max_size here as well.
2299 * Note that caller is responsible for ensuring max_size increases are
2300 * requested from the MDS.
2301 */
2302static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2303 loff_t endoff, bool nonblock, int *got, int *err)
2304{
2305 struct inode *inode = &ci->vfs_inode;
2306 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2307 int ret = 0;
2308 int have, implemented;
2309 int file_wanted;
2310 bool snap_rwsem_locked = false;
2311
2312 dout("get_cap_refs %p need %s want %s\n", inode,
2313 ceph_cap_string(need), ceph_cap_string(want));
2314
2315again:
2316 spin_lock(&ci->i_ceph_lock);
2317
2318 /* make sure file is actually open */
2319 file_wanted = __ceph_caps_file_wanted(ci);
2320 if ((file_wanted & need) == 0) {
2321 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2322 ceph_cap_string(need), ceph_cap_string(file_wanted));
2323 *err = -EBADF;
2324 ret = 1;
2325 goto out_unlock;
2326 }
2327
2328 /* finish pending truncate */
2329 while (ci->i_truncate_pending) {
2330 spin_unlock(&ci->i_ceph_lock);
2331 if (snap_rwsem_locked) {
2332 up_read(&mdsc->snap_rwsem);
2333 snap_rwsem_locked = false;
2334 }
2335 __ceph_do_pending_vmtruncate(inode);
2336 spin_lock(&ci->i_ceph_lock);
2337 }
2338
2339 have = __ceph_caps_issued(ci, &implemented);
2340
2341 if (have & need & CEPH_CAP_FILE_WR) {
2342 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2343 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2344 inode, endoff, ci->i_max_size);
2345 if (endoff > ci->i_requested_max_size) {
2346 *err = -EAGAIN;
2347 ret = 1;
2348 }
2349 goto out_unlock;
2350 }
2351 /*
2352 * If a sync write is in progress, we must wait, so that we
2353 * can get a final snapshot value for size+mtime.
2354 */
2355 if (__ceph_have_pending_cap_snap(ci)) {
2356 dout("get_cap_refs %p cap_snap_pending\n", inode);
2357 goto out_unlock;
2358 }
2359 }
2360
2361 if ((have & need) == need) {
2362 /*
2363 * Look at (implemented & ~have & not) so that we keep waiting
2364 * on transition from wanted -> needed caps. This is needed
2365 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2366 * going before a prior buffered writeback happens.
2367 */
2368 int not = want & ~(have & need);
2369 int revoking = implemented & ~have;
2370 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2371 inode, ceph_cap_string(have), ceph_cap_string(not),
2372 ceph_cap_string(revoking));
2373 if ((revoking & not) == 0) {
2374 if (!snap_rwsem_locked &&
2375 !ci->i_head_snapc &&
2376 (need & CEPH_CAP_FILE_WR)) {
2377 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2378 /*
2379 * we can not call down_read() when
2380 * task isn't in TASK_RUNNING state
2381 */
2382 if (nonblock) {
2383 *err = -EAGAIN;
2384 ret = 1;
2385 goto out_unlock;
2386 }
2387
2388 spin_unlock(&ci->i_ceph_lock);
2389 down_read(&mdsc->snap_rwsem);
2390 snap_rwsem_locked = true;
2391 goto again;
2392 }
2393 snap_rwsem_locked = true;
2394 }
2395 *got = need | (have & want);
2396 __take_cap_refs(ci, *got, true);
2397 ret = 1;
2398 }
2399 } else {
2400 int session_readonly = false;
2401 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2402 struct ceph_mds_session *s = ci->i_auth_cap->session;
2403 spin_lock(&s->s_cap_lock);
2404 session_readonly = s->s_readonly;
2405 spin_unlock(&s->s_cap_lock);
2406 }
2407 if (session_readonly) {
2408 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2409 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2410 *err = -EROFS;
2411 ret = 1;
2412 goto out_unlock;
2413 }
2414
2415 if (!__ceph_is_any_caps(ci) &&
2416 ACCESS_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2417 dout("get_cap_refs %p forced umount\n", inode);
2418 *err = -EIO;
2419 ret = 1;
2420 goto out_unlock;
2421 }
2422
2423 dout("get_cap_refs %p have %s needed %s\n", inode,
2424 ceph_cap_string(have), ceph_cap_string(need));
2425 }
2426out_unlock:
2427 spin_unlock(&ci->i_ceph_lock);
2428 if (snap_rwsem_locked)
2429 up_read(&mdsc->snap_rwsem);
2430
2431 dout("get_cap_refs %p ret %d got %s\n", inode,
2432 ret, ceph_cap_string(*got));
2433 return ret;
2434}
2435
2436/*
2437 * Check the offset we are writing up to against our current
2438 * max_size. If necessary, tell the MDS we want to write to
2439 * a larger offset.
2440 */
2441static void check_max_size(struct inode *inode, loff_t endoff)
2442{
2443 struct ceph_inode_info *ci = ceph_inode(inode);
2444 int check = 0;
2445
2446 /* do we need to explicitly request a larger max_size? */
2447 spin_lock(&ci->i_ceph_lock);
2448 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2449 dout("write %p at large endoff %llu, req max_size\n",
2450 inode, endoff);
2451 ci->i_wanted_max_size = endoff;
2452 }
2453 /* duplicate ceph_check_caps()'s logic */
2454 if (ci->i_auth_cap &&
2455 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2456 ci->i_wanted_max_size > ci->i_max_size &&
2457 ci->i_wanted_max_size > ci->i_requested_max_size)
2458 check = 1;
2459 spin_unlock(&ci->i_ceph_lock);
2460 if (check)
2461 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2462}
2463
2464/*
2465 * Wait for caps, and take cap references. If we can't get a WR cap
2466 * due to a small max_size, make sure we check_max_size (and possibly
2467 * ask the mds) so we don't get hung up indefinitely.
2468 */
2469int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2470 loff_t endoff, int *got, struct page **pinned_page)
2471{
2472 int _got, ret, err = 0;
2473
2474 ret = ceph_pool_perm_check(ci, need);
2475 if (ret < 0)
2476 return ret;
2477
2478 while (true) {
2479 if (endoff > 0)
2480 check_max_size(&ci->vfs_inode, endoff);
2481
2482 err = 0;
2483 _got = 0;
2484 ret = try_get_cap_refs(ci, need, want, endoff,
2485 false, &_got, &err);
2486 if (ret) {
2487 if (err == -EAGAIN)
2488 continue;
2489 if (err < 0)
2490 return err;
2491 } else {
2492 ret = wait_event_interruptible(ci->i_cap_wq,
2493 try_get_cap_refs(ci, need, want, endoff,
2494 true, &_got, &err));
2495 if (err == -EAGAIN)
2496 continue;
2497 if (err < 0)
2498 ret = err;
2499 if (ret < 0)
2500 return ret;
2501 }
2502
2503 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2504 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2505 i_size_read(&ci->vfs_inode) > 0) {
2506 struct page *page =
2507 find_get_page(ci->vfs_inode.i_mapping, 0);
2508 if (page) {
2509 if (PageUptodate(page)) {
2510 *pinned_page = page;
2511 break;
2512 }
2513 put_page(page);
2514 }
2515 /*
2516 * drop cap refs first because getattr while
2517 * holding * caps refs can cause deadlock.
2518 */
2519 ceph_put_cap_refs(ci, _got);
2520 _got = 0;
2521
2522 /*
2523 * getattr request will bring inline data into
2524 * page cache
2525 */
2526 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2527 CEPH_STAT_CAP_INLINE_DATA,
2528 true);
2529 if (ret < 0)
2530 return ret;
2531 continue;
2532 }
2533 break;
2534 }
2535
2536 *got = _got;
2537 return 0;
2538}
2539
2540/*
2541 * Take cap refs. Caller must already know we hold at least one ref
2542 * on the caps in question or we don't know this is safe.
2543 */
2544void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2545{
2546 spin_lock(&ci->i_ceph_lock);
2547 __take_cap_refs(ci, caps, false);
2548 spin_unlock(&ci->i_ceph_lock);
2549}
2550
2551
2552/*
2553 * drop cap_snap that is not associated with any snapshot.
2554 * we don't need to send FLUSHSNAP message for it.
2555 */
2556static int ceph_try_drop_cap_snap(struct ceph_cap_snap *capsnap)
2557{
2558 if (!capsnap->need_flush &&
2559 !capsnap->writing && !capsnap->dirty_pages) {
2560
2561 dout("dropping cap_snap %p follows %llu\n",
2562 capsnap, capsnap->follows);
2563 ceph_put_snap_context(capsnap->context);
2564 list_del(&capsnap->ci_item);
2565 list_del(&capsnap->flushing_item);
2566 ceph_put_cap_snap(capsnap);
2567 return 1;
2568 }
2569 return 0;
2570}
2571
2572/*
2573 * Release cap refs.
2574 *
2575 * If we released the last ref on any given cap, call ceph_check_caps
2576 * to release (or schedule a release).
2577 *
2578 * If we are releasing a WR cap (from a sync write), finalize any affected
2579 * cap_snap, and wake up any waiters.
2580 */
2581void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2582{
2583 struct inode *inode = &ci->vfs_inode;
2584 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2585
2586 spin_lock(&ci->i_ceph_lock);
2587 if (had & CEPH_CAP_PIN)
2588 --ci->i_pin_ref;
2589 if (had & CEPH_CAP_FILE_RD)
2590 if (--ci->i_rd_ref == 0)
2591 last++;
2592 if (had & CEPH_CAP_FILE_CACHE)
2593 if (--ci->i_rdcache_ref == 0)
2594 last++;
2595 if (had & CEPH_CAP_FILE_BUFFER) {
2596 if (--ci->i_wb_ref == 0) {
2597 last++;
2598 put++;
2599 }
2600 dout("put_cap_refs %p wb %d -> %d (?)\n",
2601 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2602 }
2603 if (had & CEPH_CAP_FILE_WR)
2604 if (--ci->i_wr_ref == 0) {
2605 last++;
2606 if (__ceph_have_pending_cap_snap(ci)) {
2607 struct ceph_cap_snap *capsnap =
2608 list_last_entry(&ci->i_cap_snaps,
2609 struct ceph_cap_snap,
2610 ci_item);
2611 capsnap->writing = 0;
2612 if (ceph_try_drop_cap_snap(capsnap))
2613 put++;
2614 else if (__ceph_finish_cap_snap(ci, capsnap))
2615 flushsnaps = 1;
2616 wake = 1;
2617 }
2618 if (ci->i_wrbuffer_ref_head == 0 &&
2619 ci->i_dirty_caps == 0 &&
2620 ci->i_flushing_caps == 0) {
2621 BUG_ON(!ci->i_head_snapc);
2622 ceph_put_snap_context(ci->i_head_snapc);
2623 ci->i_head_snapc = NULL;
2624 }
2625 /* see comment in __ceph_remove_cap() */
2626 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2627 drop_inode_snap_realm(ci);
2628 }
2629 spin_unlock(&ci->i_ceph_lock);
2630
2631 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2632 last ? " last" : "", put ? " put" : "");
2633
2634 if (last && !flushsnaps)
2635 ceph_check_caps(ci, 0, NULL);
2636 else if (flushsnaps)
2637 ceph_flush_snaps(ci);
2638 if (wake)
2639 wake_up_all(&ci->i_cap_wq);
2640 while (put-- > 0)
2641 iput(inode);
2642}
2643
2644/*
2645 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2646 * context. Adjust per-snap dirty page accounting as appropriate.
2647 * Once all dirty data for a cap_snap is flushed, flush snapped file
2648 * metadata back to the MDS. If we dropped the last ref, call
2649 * ceph_check_caps.
2650 */
2651void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2652 struct ceph_snap_context *snapc)
2653{
2654 struct inode *inode = &ci->vfs_inode;
2655 int last = 0;
2656 int complete_capsnap = 0;
2657 int drop_capsnap = 0;
2658 int found = 0;
2659 struct ceph_cap_snap *capsnap = NULL;
2660
2661 spin_lock(&ci->i_ceph_lock);
2662 ci->i_wrbuffer_ref -= nr;
2663 last = !ci->i_wrbuffer_ref;
2664
2665 if (ci->i_head_snapc == snapc) {
2666 ci->i_wrbuffer_ref_head -= nr;
2667 if (ci->i_wrbuffer_ref_head == 0 &&
2668 ci->i_wr_ref == 0 &&
2669 ci->i_dirty_caps == 0 &&
2670 ci->i_flushing_caps == 0) {
2671 BUG_ON(!ci->i_head_snapc);
2672 ceph_put_snap_context(ci->i_head_snapc);
2673 ci->i_head_snapc = NULL;
2674 }
2675 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2676 inode,
2677 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2678 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2679 last ? " LAST" : "");
2680 } else {
2681 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2682 if (capsnap->context == snapc) {
2683 found = 1;
2684 break;
2685 }
2686 }
2687 BUG_ON(!found);
2688 capsnap->dirty_pages -= nr;
2689 if (capsnap->dirty_pages == 0) {
2690 complete_capsnap = 1;
2691 drop_capsnap = ceph_try_drop_cap_snap(capsnap);
2692 }
2693 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2694 " snap %lld %d/%d -> %d/%d %s%s\n",
2695 inode, capsnap, capsnap->context->seq,
2696 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2697 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2698 last ? " (wrbuffer last)" : "",
2699 complete_capsnap ? " (complete capsnap)" : "");
2700 }
2701
2702 spin_unlock(&ci->i_ceph_lock);
2703
2704 if (last) {
2705 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2706 iput(inode);
2707 } else if (complete_capsnap) {
2708 ceph_flush_snaps(ci);
2709 wake_up_all(&ci->i_cap_wq);
2710 }
2711 if (drop_capsnap)
2712 iput(inode);
2713}
2714
2715/*
2716 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2717 */
2718static void invalidate_aliases(struct inode *inode)
2719{
2720 struct dentry *dn, *prev = NULL;
2721
2722 dout("invalidate_aliases inode %p\n", inode);
2723 d_prune_aliases(inode);
2724 /*
2725 * For non-directory inode, d_find_alias() only returns
2726 * hashed dentry. After calling d_invalidate(), the
2727 * dentry becomes unhashed.
2728 *
2729 * For directory inode, d_find_alias() can return
2730 * unhashed dentry. But directory inode should have
2731 * one alias at most.
2732 */
2733 while ((dn = d_find_alias(inode))) {
2734 if (dn == prev) {
2735 dput(dn);
2736 break;
2737 }
2738 d_invalidate(dn);
2739 if (prev)
2740 dput(prev);
2741 prev = dn;
2742 }
2743 if (prev)
2744 dput(prev);
2745}
2746
2747/*
2748 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2749 * actually be a revocation if it specifies a smaller cap set.)
2750 *
2751 * caller holds s_mutex and i_ceph_lock, we drop both.
2752 */
2753static void handle_cap_grant(struct ceph_mds_client *mdsc,
2754 struct inode *inode, struct ceph_mds_caps *grant,
2755 u64 inline_version,
2756 void *inline_data, int inline_len,
2757 struct ceph_buffer *xattr_buf,
2758 struct ceph_mds_session *session,
2759 struct ceph_cap *cap, int issued,
2760 u32 pool_ns_len)
2761 __releases(ci->i_ceph_lock)
2762 __releases(mdsc->snap_rwsem)
2763{
2764 struct ceph_inode_info *ci = ceph_inode(inode);
2765 int mds = session->s_mds;
2766 int seq = le32_to_cpu(grant->seq);
2767 int newcaps = le32_to_cpu(grant->caps);
2768 int used, wanted, dirty;
2769 u64 size = le64_to_cpu(grant->size);
2770 u64 max_size = le64_to_cpu(grant->max_size);
2771 struct timespec mtime, atime, ctime;
2772 int check_caps = 0;
2773 bool wake = false;
2774 bool writeback = false;
2775 bool queue_trunc = false;
2776 bool queue_invalidate = false;
2777 bool queue_revalidate = false;
2778 bool deleted_inode = false;
2779 bool fill_inline = false;
2780
2781 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2782 inode, cap, mds, seq, ceph_cap_string(newcaps));
2783 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2784 inode->i_size);
2785
2786
2787 /*
2788 * auth mds of the inode changed. we received the cap export message,
2789 * but still haven't received the cap import message. handle_cap_export
2790 * updated the new auth MDS' cap.
2791 *
2792 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2793 * that was sent before the cap import message. So don't remove caps.
2794 */
2795 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2796 WARN_ON(cap != ci->i_auth_cap);
2797 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2798 seq = cap->seq;
2799 newcaps |= cap->issued;
2800 }
2801
2802 /*
2803 * If CACHE is being revoked, and we have no dirty buffers,
2804 * try to invalidate (once). (If there are dirty buffers, we
2805 * will invalidate _after_ writeback.)
2806 */
2807 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2808 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2809 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2810 !ci->i_wrbuffer_ref) {
2811 if (try_nonblocking_invalidate(inode)) {
2812 /* there were locked pages.. invalidate later
2813 in a separate thread. */
2814 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2815 queue_invalidate = true;
2816 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2817 }
2818 }
2819
2820 ceph_fscache_invalidate(inode);
2821 }
2822
2823 /* side effects now are allowed */
2824 cap->cap_gen = session->s_cap_gen;
2825 cap->seq = seq;
2826
2827 __check_cap_issue(ci, cap, newcaps);
2828
2829 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2830 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2831 inode->i_mode = le32_to_cpu(grant->mode);
2832 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2833 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2834 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2835 from_kuid(&init_user_ns, inode->i_uid),
2836 from_kgid(&init_user_ns, inode->i_gid));
2837 }
2838
2839 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2840 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2841 set_nlink(inode, le32_to_cpu(grant->nlink));
2842 if (inode->i_nlink == 0 &&
2843 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2844 deleted_inode = true;
2845 }
2846
2847 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2848 int len = le32_to_cpu(grant->xattr_len);
2849 u64 version = le64_to_cpu(grant->xattr_version);
2850
2851 if (version > ci->i_xattrs.version) {
2852 dout(" got new xattrs v%llu on %p len %d\n",
2853 version, inode, len);
2854 if (ci->i_xattrs.blob)
2855 ceph_buffer_put(ci->i_xattrs.blob);
2856 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2857 ci->i_xattrs.version = version;
2858 ceph_forget_all_cached_acls(inode);
2859 }
2860 }
2861
2862 /* Do we need to revalidate our fscache cookie. Don't bother on the
2863 * first cache cap as we already validate at cookie creation time. */
2864 if ((issued & CEPH_CAP_FILE_CACHE) && ci->i_rdcache_gen > 1)
2865 queue_revalidate = true;
2866
2867 if (newcaps & CEPH_CAP_ANY_RD) {
2868 /* ctime/mtime/atime? */
2869 ceph_decode_timespec(&mtime, &grant->mtime);
2870 ceph_decode_timespec(&atime, &grant->atime);
2871 ceph_decode_timespec(&ctime, &grant->ctime);
2872 ceph_fill_file_time(inode, issued,
2873 le32_to_cpu(grant->time_warp_seq),
2874 &ctime, &mtime, &atime);
2875 }
2876
2877 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2878 /* file layout may have changed */
2879 ci->i_layout = grant->layout;
2880 ci->i_pool_ns_len = pool_ns_len;
2881
2882 /* size/truncate_seq? */
2883 queue_trunc = ceph_fill_file_size(inode, issued,
2884 le32_to_cpu(grant->truncate_seq),
2885 le64_to_cpu(grant->truncate_size),
2886 size);
2887 /* max size increase? */
2888 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2889 dout("max_size %lld -> %llu\n",
2890 ci->i_max_size, max_size);
2891 ci->i_max_size = max_size;
2892 if (max_size >= ci->i_wanted_max_size) {
2893 ci->i_wanted_max_size = 0; /* reset */
2894 ci->i_requested_max_size = 0;
2895 }
2896 wake = true;
2897 }
2898 }
2899
2900 /* check cap bits */
2901 wanted = __ceph_caps_wanted(ci);
2902 used = __ceph_caps_used(ci);
2903 dirty = __ceph_caps_dirty(ci);
2904 dout(" my wanted = %s, used = %s, dirty %s\n",
2905 ceph_cap_string(wanted),
2906 ceph_cap_string(used),
2907 ceph_cap_string(dirty));
2908 if (wanted != le32_to_cpu(grant->wanted)) {
2909 dout("mds wanted %s -> %s\n",
2910 ceph_cap_string(le32_to_cpu(grant->wanted)),
2911 ceph_cap_string(wanted));
2912 /* imported cap may not have correct mds_wanted */
2913 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2914 check_caps = 1;
2915 }
2916
2917 /* revocation, grant, or no-op? */
2918 if (cap->issued & ~newcaps) {
2919 int revoking = cap->issued & ~newcaps;
2920
2921 dout("revocation: %s -> %s (revoking %s)\n",
2922 ceph_cap_string(cap->issued),
2923 ceph_cap_string(newcaps),
2924 ceph_cap_string(revoking));
2925 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2926 writeback = true; /* initiate writeback; will delay ack */
2927 else if (revoking == CEPH_CAP_FILE_CACHE &&
2928 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2929 queue_invalidate)
2930 ; /* do nothing yet, invalidation will be queued */
2931 else if (cap == ci->i_auth_cap)
2932 check_caps = 1; /* check auth cap only */
2933 else
2934 check_caps = 2; /* check all caps */
2935 cap->issued = newcaps;
2936 cap->implemented |= newcaps;
2937 } else if (cap->issued == newcaps) {
2938 dout("caps unchanged: %s -> %s\n",
2939 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2940 } else {
2941 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2942 ceph_cap_string(newcaps));
2943 /* non-auth MDS is revoking the newly grant caps ? */
2944 if (cap == ci->i_auth_cap &&
2945 __ceph_caps_revoking_other(ci, cap, newcaps))
2946 check_caps = 2;
2947
2948 cap->issued = newcaps;
2949 cap->implemented |= newcaps; /* add bits only, to
2950 * avoid stepping on a
2951 * pending revocation */
2952 wake = true;
2953 }
2954 BUG_ON(cap->issued & ~cap->implemented);
2955
2956 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
2957 ci->i_inline_version = inline_version;
2958 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2959 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
2960 fill_inline = true;
2961 }
2962
2963 spin_unlock(&ci->i_ceph_lock);
2964
2965 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
2966 kick_flushing_inode_caps(mdsc, session, inode);
2967 up_read(&mdsc->snap_rwsem);
2968 if (newcaps & ~issued)
2969 wake = true;
2970 }
2971
2972 if (fill_inline)
2973 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
2974
2975 if (queue_trunc) {
2976 ceph_queue_vmtruncate(inode);
2977 ceph_queue_revalidate(inode);
2978 } else if (queue_revalidate)
2979 ceph_queue_revalidate(inode);
2980
2981 if (writeback)
2982 /*
2983 * queue inode for writeback: we can't actually call
2984 * filemap_write_and_wait, etc. from message handler
2985 * context.
2986 */
2987 ceph_queue_writeback(inode);
2988 if (queue_invalidate)
2989 ceph_queue_invalidate(inode);
2990 if (deleted_inode)
2991 invalidate_aliases(inode);
2992 if (wake)
2993 wake_up_all(&ci->i_cap_wq);
2994
2995 if (check_caps == 1)
2996 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2997 session);
2998 else if (check_caps == 2)
2999 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3000 else
3001 mutex_unlock(&session->s_mutex);
3002}
3003
3004/*
3005 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3006 * MDS has been safely committed.
3007 */
3008static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3009 struct ceph_mds_caps *m,
3010 struct ceph_mds_session *session,
3011 struct ceph_cap *cap)
3012 __releases(ci->i_ceph_lock)
3013{
3014 struct ceph_inode_info *ci = ceph_inode(inode);
3015 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3016 struct ceph_cap_flush *cf;
3017 struct rb_node *n;
3018 LIST_HEAD(to_remove);
3019 unsigned seq = le32_to_cpu(m->seq);
3020 int dirty = le32_to_cpu(m->dirty);
3021 int cleaned = 0;
3022 int drop = 0;
3023
3024 n = rb_first(&ci->i_cap_flush_tree);
3025 while (n) {
3026 cf = rb_entry(n, struct ceph_cap_flush, i_node);
3027 n = rb_next(&cf->i_node);
3028 if (cf->tid == flush_tid)
3029 cleaned = cf->caps;
3030 if (cf->tid <= flush_tid) {
3031 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
3032 list_add_tail(&cf->list, &to_remove);
3033 } else {
3034 cleaned &= ~cf->caps;
3035 if (!cleaned)
3036 break;
3037 }
3038 }
3039
3040 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3041 " flushing %s -> %s\n",
3042 inode, session->s_mds, seq, ceph_cap_string(dirty),
3043 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3044 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3045
3046 if (list_empty(&to_remove) && !cleaned)
3047 goto out;
3048
3049 ci->i_flushing_caps &= ~cleaned;
3050
3051 spin_lock(&mdsc->cap_dirty_lock);
3052
3053 if (!list_empty(&to_remove)) {
3054 list_for_each_entry(cf, &to_remove, list)
3055 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
3056
3057 n = rb_first(&mdsc->cap_flush_tree);
3058 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
3059 if (!cf || cf->tid > flush_tid)
3060 wake_up_all(&mdsc->cap_flushing_wq);
3061 }
3062
3063 if (ci->i_flushing_caps == 0) {
3064 list_del_init(&ci->i_flushing_item);
3065 if (!list_empty(&session->s_cap_flushing))
3066 dout(" mds%d still flushing cap on %p\n",
3067 session->s_mds,
3068 &list_entry(session->s_cap_flushing.next,
3069 struct ceph_inode_info,
3070 i_flushing_item)->vfs_inode);
3071 mdsc->num_cap_flushing--;
3072 dout(" inode %p now !flushing\n", inode);
3073
3074 if (ci->i_dirty_caps == 0) {
3075 dout(" inode %p now clean\n", inode);
3076 BUG_ON(!list_empty(&ci->i_dirty_item));
3077 drop = 1;
3078 if (ci->i_wr_ref == 0 &&
3079 ci->i_wrbuffer_ref_head == 0) {
3080 BUG_ON(!ci->i_head_snapc);
3081 ceph_put_snap_context(ci->i_head_snapc);
3082 ci->i_head_snapc = NULL;
3083 }
3084 } else {
3085 BUG_ON(list_empty(&ci->i_dirty_item));
3086 }
3087 }
3088 spin_unlock(&mdsc->cap_dirty_lock);
3089 wake_up_all(&ci->i_cap_wq);
3090
3091out:
3092 spin_unlock(&ci->i_ceph_lock);
3093
3094 while (!list_empty(&to_remove)) {
3095 cf = list_first_entry(&to_remove,
3096 struct ceph_cap_flush, list);
3097 list_del(&cf->list);
3098 ceph_free_cap_flush(cf);
3099 }
3100 if (drop)
3101 iput(inode);
3102}
3103
3104/*
3105 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3106 * throw away our cap_snap.
3107 *
3108 * Caller hold s_mutex.
3109 */
3110static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3111 struct ceph_mds_caps *m,
3112 struct ceph_mds_session *session)
3113{
3114 struct ceph_inode_info *ci = ceph_inode(inode);
3115 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3116 u64 follows = le64_to_cpu(m->snap_follows);
3117 struct ceph_cap_snap *capsnap;
3118 int drop = 0;
3119
3120 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3121 inode, ci, session->s_mds, follows);
3122
3123 spin_lock(&ci->i_ceph_lock);
3124 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3125 if (capsnap->follows == follows) {
3126 if (capsnap->flush_tid != flush_tid) {
3127 dout(" cap_snap %p follows %lld tid %lld !="
3128 " %lld\n", capsnap, follows,
3129 flush_tid, capsnap->flush_tid);
3130 break;
3131 }
3132 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3133 dout(" removing %p cap_snap %p follows %lld\n",
3134 inode, capsnap, follows);
3135 ceph_put_snap_context(capsnap->context);
3136 list_del(&capsnap->ci_item);
3137 list_del(&capsnap->flushing_item);
3138 ceph_put_cap_snap(capsnap);
3139 wake_up_all(&mdsc->cap_flushing_wq);
3140 drop = 1;
3141 break;
3142 } else {
3143 dout(" skipping cap_snap %p follows %lld\n",
3144 capsnap, capsnap->follows);
3145 }
3146 }
3147 spin_unlock(&ci->i_ceph_lock);
3148 if (drop)
3149 iput(inode);
3150}
3151
3152/*
3153 * Handle TRUNC from MDS, indicating file truncation.
3154 *
3155 * caller hold s_mutex.
3156 */
3157static void handle_cap_trunc(struct inode *inode,
3158 struct ceph_mds_caps *trunc,
3159 struct ceph_mds_session *session)
3160 __releases(ci->i_ceph_lock)
3161{
3162 struct ceph_inode_info *ci = ceph_inode(inode);
3163 int mds = session->s_mds;
3164 int seq = le32_to_cpu(trunc->seq);
3165 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3166 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3167 u64 size = le64_to_cpu(trunc->size);
3168 int implemented = 0;
3169 int dirty = __ceph_caps_dirty(ci);
3170 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3171 int queue_trunc = 0;
3172
3173 issued |= implemented | dirty;
3174
3175 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3176 inode, mds, seq, truncate_size, truncate_seq);
3177 queue_trunc = ceph_fill_file_size(inode, issued,
3178 truncate_seq, truncate_size, size);
3179 spin_unlock(&ci->i_ceph_lock);
3180
3181 if (queue_trunc) {
3182 ceph_queue_vmtruncate(inode);
3183 ceph_fscache_invalidate(inode);
3184 }
3185}
3186
3187/*
3188 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3189 * different one. If we are the most recent migration we've seen (as
3190 * indicated by mseq), make note of the migrating cap bits for the
3191 * duration (until we see the corresponding IMPORT).
3192 *
3193 * caller holds s_mutex
3194 */
3195static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3196 struct ceph_mds_cap_peer *ph,
3197 struct ceph_mds_session *session)
3198{
3199 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3200 struct ceph_mds_session *tsession = NULL;
3201 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3202 struct ceph_inode_info *ci = ceph_inode(inode);
3203 u64 t_cap_id;
3204 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3205 unsigned t_seq, t_mseq;
3206 int target, issued;
3207 int mds = session->s_mds;
3208
3209 if (ph) {
3210 t_cap_id = le64_to_cpu(ph->cap_id);
3211 t_seq = le32_to_cpu(ph->seq);
3212 t_mseq = le32_to_cpu(ph->mseq);
3213 target = le32_to_cpu(ph->mds);
3214 } else {
3215 t_cap_id = t_seq = t_mseq = 0;
3216 target = -1;
3217 }
3218
3219 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3220 inode, ci, mds, mseq, target);
3221retry:
3222 spin_lock(&ci->i_ceph_lock);
3223 cap = __get_cap_for_mds(ci, mds);
3224 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3225 goto out_unlock;
3226
3227 if (target < 0) {
3228 __ceph_remove_cap(cap, false);
3229 goto out_unlock;
3230 }
3231
3232 /*
3233 * now we know we haven't received the cap import message yet
3234 * because the exported cap still exist.
3235 */
3236
3237 issued = cap->issued;
3238 WARN_ON(issued != cap->implemented);
3239
3240 tcap = __get_cap_for_mds(ci, target);
3241 if (tcap) {
3242 /* already have caps from the target */
3243 if (tcap->cap_id != t_cap_id ||
3244 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3245 dout(" updating import cap %p mds%d\n", tcap, target);
3246 tcap->cap_id = t_cap_id;
3247 tcap->seq = t_seq - 1;
3248 tcap->issue_seq = t_seq - 1;
3249 tcap->mseq = t_mseq;
3250 tcap->issued |= issued;
3251 tcap->implemented |= issued;
3252 if (cap == ci->i_auth_cap)
3253 ci->i_auth_cap = tcap;
3254 if (ci->i_flushing_caps && ci->i_auth_cap == tcap) {
3255 spin_lock(&mdsc->cap_dirty_lock);
3256 list_move_tail(&ci->i_flushing_item,
3257 &tcap->session->s_cap_flushing);
3258 spin_unlock(&mdsc->cap_dirty_lock);
3259 }
3260 }
3261 __ceph_remove_cap(cap, false);
3262 goto out_unlock;
3263 } else if (tsession) {
3264 /* add placeholder for the export tagert */
3265 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3266 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3267 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3268
3269 __ceph_remove_cap(cap, false);
3270 goto out_unlock;
3271 }
3272
3273 spin_unlock(&ci->i_ceph_lock);
3274 mutex_unlock(&session->s_mutex);
3275
3276 /* open target session */
3277 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3278 if (!IS_ERR(tsession)) {
3279 if (mds > target) {
3280 mutex_lock(&session->s_mutex);
3281 mutex_lock_nested(&tsession->s_mutex,
3282 SINGLE_DEPTH_NESTING);
3283 } else {
3284 mutex_lock(&tsession->s_mutex);
3285 mutex_lock_nested(&session->s_mutex,
3286 SINGLE_DEPTH_NESTING);
3287 }
3288 new_cap = ceph_get_cap(mdsc, NULL);
3289 } else {
3290 WARN_ON(1);
3291 tsession = NULL;
3292 target = -1;
3293 }
3294 goto retry;
3295
3296out_unlock:
3297 spin_unlock(&ci->i_ceph_lock);
3298 mutex_unlock(&session->s_mutex);
3299 if (tsession) {
3300 mutex_unlock(&tsession->s_mutex);
3301 ceph_put_mds_session(tsession);
3302 }
3303 if (new_cap)
3304 ceph_put_cap(mdsc, new_cap);
3305}
3306
3307/*
3308 * Handle cap IMPORT.
3309 *
3310 * caller holds s_mutex. acquires i_ceph_lock
3311 */
3312static void handle_cap_import(struct ceph_mds_client *mdsc,
3313 struct inode *inode, struct ceph_mds_caps *im,
3314 struct ceph_mds_cap_peer *ph,
3315 struct ceph_mds_session *session,
3316 struct ceph_cap **target_cap, int *old_issued)
3317 __acquires(ci->i_ceph_lock)
3318{
3319 struct ceph_inode_info *ci = ceph_inode(inode);
3320 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3321 int mds = session->s_mds;
3322 int issued;
3323 unsigned caps = le32_to_cpu(im->caps);
3324 unsigned wanted = le32_to_cpu(im->wanted);
3325 unsigned seq = le32_to_cpu(im->seq);
3326 unsigned mseq = le32_to_cpu(im->migrate_seq);
3327 u64 realmino = le64_to_cpu(im->realm);
3328 u64 cap_id = le64_to_cpu(im->cap_id);
3329 u64 p_cap_id;
3330 int peer;
3331
3332 if (ph) {
3333 p_cap_id = le64_to_cpu(ph->cap_id);
3334 peer = le32_to_cpu(ph->mds);
3335 } else {
3336 p_cap_id = 0;
3337 peer = -1;
3338 }
3339
3340 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3341 inode, ci, mds, mseq, peer);
3342
3343retry:
3344 spin_lock(&ci->i_ceph_lock);
3345 cap = __get_cap_for_mds(ci, mds);
3346 if (!cap) {
3347 if (!new_cap) {
3348 spin_unlock(&ci->i_ceph_lock);
3349 new_cap = ceph_get_cap(mdsc, NULL);
3350 goto retry;
3351 }
3352 cap = new_cap;
3353 } else {
3354 if (new_cap) {
3355 ceph_put_cap(mdsc, new_cap);
3356 new_cap = NULL;
3357 }
3358 }
3359
3360 __ceph_caps_issued(ci, &issued);
3361 issued |= __ceph_caps_dirty(ci);
3362
3363 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3364 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3365
3366 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3367 if (ocap && ocap->cap_id == p_cap_id) {
3368 dout(" remove export cap %p mds%d flags %d\n",
3369 ocap, peer, ph->flags);
3370 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3371 (ocap->seq != le32_to_cpu(ph->seq) ||
3372 ocap->mseq != le32_to_cpu(ph->mseq))) {
3373 pr_err("handle_cap_import: mismatched seq/mseq: "
3374 "ino (%llx.%llx) mds%d seq %d mseq %d "
3375 "importer mds%d has peer seq %d mseq %d\n",
3376 ceph_vinop(inode), peer, ocap->seq,
3377 ocap->mseq, mds, le32_to_cpu(ph->seq),
3378 le32_to_cpu(ph->mseq));
3379 }
3380 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3381 }
3382
3383 /* make sure we re-request max_size, if necessary */
3384 ci->i_wanted_max_size = 0;
3385 ci->i_requested_max_size = 0;
3386
3387 *old_issued = issued;
3388 *target_cap = cap;
3389}
3390
3391/*
3392 * Handle a caps message from the MDS.
3393 *
3394 * Identify the appropriate session, inode, and call the right handler
3395 * based on the cap op.
3396 */
3397void ceph_handle_caps(struct ceph_mds_session *session,
3398 struct ceph_msg *msg)
3399{
3400 struct ceph_mds_client *mdsc = session->s_mdsc;
3401 struct super_block *sb = mdsc->fsc->sb;
3402 struct inode *inode;
3403 struct ceph_inode_info *ci;
3404 struct ceph_cap *cap;
3405 struct ceph_mds_caps *h;
3406 struct ceph_mds_cap_peer *peer = NULL;
3407 struct ceph_snap_realm *realm;
3408 int mds = session->s_mds;
3409 int op, issued;
3410 u32 seq, mseq;
3411 struct ceph_vino vino;
3412 u64 cap_id;
3413 u64 size, max_size;
3414 u64 tid;
3415 u64 inline_version = 0;
3416 void *inline_data = NULL;
3417 u32 inline_len = 0;
3418 void *snaptrace;
3419 size_t snaptrace_len;
3420 u32 pool_ns_len = 0;
3421 void *p, *end;
3422
3423 dout("handle_caps from mds%d\n", mds);
3424
3425 /* decode */
3426 end = msg->front.iov_base + msg->front.iov_len;
3427 tid = le64_to_cpu(msg->hdr.tid);
3428 if (msg->front.iov_len < sizeof(*h))
3429 goto bad;
3430 h = msg->front.iov_base;
3431 op = le32_to_cpu(h->op);
3432 vino.ino = le64_to_cpu(h->ino);
3433 vino.snap = CEPH_NOSNAP;
3434 cap_id = le64_to_cpu(h->cap_id);
3435 seq = le32_to_cpu(h->seq);
3436 mseq = le32_to_cpu(h->migrate_seq);
3437 size = le64_to_cpu(h->size);
3438 max_size = le64_to_cpu(h->max_size);
3439
3440 snaptrace = h + 1;
3441 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3442 p = snaptrace + snaptrace_len;
3443
3444 if (le16_to_cpu(msg->hdr.version) >= 2) {
3445 u32 flock_len;
3446 ceph_decode_32_safe(&p, end, flock_len, bad);
3447 if (p + flock_len > end)
3448 goto bad;
3449 p += flock_len;
3450 }
3451
3452 if (le16_to_cpu(msg->hdr.version) >= 3) {
3453 if (op == CEPH_CAP_OP_IMPORT) {
3454 if (p + sizeof(*peer) > end)
3455 goto bad;
3456 peer = p;
3457 p += sizeof(*peer);
3458 } else if (op == CEPH_CAP_OP_EXPORT) {
3459 /* recorded in unused fields */
3460 peer = (void *)&h->size;
3461 }
3462 }
3463
3464 if (le16_to_cpu(msg->hdr.version) >= 4) {
3465 ceph_decode_64_safe(&p, end, inline_version, bad);
3466 ceph_decode_32_safe(&p, end, inline_len, bad);
3467 if (p + inline_len > end)
3468 goto bad;
3469 inline_data = p;
3470 p += inline_len;
3471 }
3472
3473 if (le16_to_cpu(msg->hdr.version) >= 8) {
3474 u64 flush_tid;
3475 u32 caller_uid, caller_gid;
3476 u32 osd_epoch_barrier;
3477 /* version >= 5 */
3478 ceph_decode_32_safe(&p, end, osd_epoch_barrier, bad);
3479 /* version >= 6 */
3480 ceph_decode_64_safe(&p, end, flush_tid, bad);
3481 /* version >= 7 */
3482 ceph_decode_32_safe(&p, end, caller_uid, bad);
3483 ceph_decode_32_safe(&p, end, caller_gid, bad);
3484 /* version >= 8 */
3485 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3486 }
3487
3488 /* lookup ino */
3489 inode = ceph_find_inode(sb, vino);
3490 ci = ceph_inode(inode);
3491 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3492 vino.snap, inode);
3493
3494 mutex_lock(&session->s_mutex);
3495 session->s_seq++;
3496 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3497 (unsigned)seq);
3498
3499 if (!inode) {
3500 dout(" i don't have ino %llx\n", vino.ino);
3501
3502 if (op == CEPH_CAP_OP_IMPORT) {
3503 cap = ceph_get_cap(mdsc, NULL);
3504 cap->cap_ino = vino.ino;
3505 cap->queue_release = 1;
3506 cap->cap_id = cap_id;
3507 cap->mseq = mseq;
3508 cap->seq = seq;
3509 spin_lock(&session->s_cap_lock);
3510 list_add_tail(&cap->session_caps,
3511 &session->s_cap_releases);
3512 session->s_num_cap_releases++;
3513 spin_unlock(&session->s_cap_lock);
3514 }
3515 goto flush_cap_releases;
3516 }
3517
3518 /* these will work even if we don't have a cap yet */
3519 switch (op) {
3520 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3521 handle_cap_flushsnap_ack(inode, tid, h, session);
3522 goto done;
3523
3524 case CEPH_CAP_OP_EXPORT:
3525 handle_cap_export(inode, h, peer, session);
3526 goto done_unlocked;
3527
3528 case CEPH_CAP_OP_IMPORT:
3529 realm = NULL;
3530 if (snaptrace_len) {
3531 down_write(&mdsc->snap_rwsem);
3532 ceph_update_snap_trace(mdsc, snaptrace,
3533 snaptrace + snaptrace_len,
3534 false, &realm);
3535 downgrade_write(&mdsc->snap_rwsem);
3536 } else {
3537 down_read(&mdsc->snap_rwsem);
3538 }
3539 handle_cap_import(mdsc, inode, h, peer, session,
3540 &cap, &issued);
3541 handle_cap_grant(mdsc, inode, h,
3542 inline_version, inline_data, inline_len,
3543 msg->middle, session, cap, issued,
3544 pool_ns_len);
3545 if (realm)
3546 ceph_put_snap_realm(mdsc, realm);
3547 goto done_unlocked;
3548 }
3549
3550 /* the rest require a cap */
3551 spin_lock(&ci->i_ceph_lock);
3552 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3553 if (!cap) {
3554 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3555 inode, ceph_ino(inode), ceph_snap(inode), mds);
3556 spin_unlock(&ci->i_ceph_lock);
3557 goto flush_cap_releases;
3558 }
3559
3560 /* note that each of these drops i_ceph_lock for us */
3561 switch (op) {
3562 case CEPH_CAP_OP_REVOKE:
3563 case CEPH_CAP_OP_GRANT:
3564 __ceph_caps_issued(ci, &issued);
3565 issued |= __ceph_caps_dirty(ci);
3566 handle_cap_grant(mdsc, inode, h,
3567 inline_version, inline_data, inline_len,
3568 msg->middle, session, cap, issued,
3569 pool_ns_len);
3570 goto done_unlocked;
3571
3572 case CEPH_CAP_OP_FLUSH_ACK:
3573 handle_cap_flush_ack(inode, tid, h, session, cap);
3574 break;
3575
3576 case CEPH_CAP_OP_TRUNC:
3577 handle_cap_trunc(inode, h, session);
3578 break;
3579
3580 default:
3581 spin_unlock(&ci->i_ceph_lock);
3582 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3583 ceph_cap_op_name(op));
3584 }
3585
3586 goto done;
3587
3588flush_cap_releases:
3589 /*
3590 * send any cap release message to try to move things
3591 * along for the mds (who clearly thinks we still have this
3592 * cap).
3593 */
3594 ceph_send_cap_releases(mdsc, session);
3595
3596done:
3597 mutex_unlock(&session->s_mutex);
3598done_unlocked:
3599 iput(inode);
3600 return;
3601
3602bad:
3603 pr_err("ceph_handle_caps: corrupt message\n");
3604 ceph_msg_dump(msg);
3605 return;
3606}
3607
3608/*
3609 * Delayed work handler to process end of delayed cap release LRU list.
3610 */
3611void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3612{
3613 struct ceph_inode_info *ci;
3614 int flags = CHECK_CAPS_NODELAY;
3615
3616 dout("check_delayed_caps\n");
3617 while (1) {
3618 spin_lock(&mdsc->cap_delay_lock);
3619 if (list_empty(&mdsc->cap_delay_list))
3620 break;
3621 ci = list_first_entry(&mdsc->cap_delay_list,
3622 struct ceph_inode_info,
3623 i_cap_delay_list);
3624 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3625 time_before(jiffies, ci->i_hold_caps_max))
3626 break;
3627 list_del_init(&ci->i_cap_delay_list);
3628 spin_unlock(&mdsc->cap_delay_lock);
3629 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3630 ceph_check_caps(ci, flags, NULL);
3631 }
3632 spin_unlock(&mdsc->cap_delay_lock);
3633}
3634
3635/*
3636 * Flush all dirty caps to the mds
3637 */
3638void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3639{
3640 struct ceph_inode_info *ci;
3641 struct inode *inode;
3642
3643 dout("flush_dirty_caps\n");
3644 spin_lock(&mdsc->cap_dirty_lock);
3645 while (!list_empty(&mdsc->cap_dirty)) {
3646 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3647 i_dirty_item);
3648 inode = &ci->vfs_inode;
3649 ihold(inode);
3650 dout("flush_dirty_caps %p\n", inode);
3651 spin_unlock(&mdsc->cap_dirty_lock);
3652 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3653 iput(inode);
3654 spin_lock(&mdsc->cap_dirty_lock);
3655 }
3656 spin_unlock(&mdsc->cap_dirty_lock);
3657 dout("flush_dirty_caps done\n");
3658}
3659
3660/*
3661 * Drop open file reference. If we were the last open file,
3662 * we may need to release capabilities to the MDS (or schedule
3663 * their delayed release).
3664 */
3665void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3666{
3667 struct inode *inode = &ci->vfs_inode;
3668 int last = 0;
3669
3670 spin_lock(&ci->i_ceph_lock);
3671 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
3672 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
3673 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
3674 if (--ci->i_nr_by_mode[fmode] == 0)
3675 last++;
3676 spin_unlock(&ci->i_ceph_lock);
3677
3678 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3679 ceph_check_caps(ci, 0, NULL);
3680}
3681
3682/*
3683 * Helpers for embedding cap and dentry lease releases into mds
3684 * requests.
3685 *
3686 * @force is used by dentry_release (below) to force inclusion of a
3687 * record for the directory inode, even when there aren't any caps to
3688 * drop.
3689 */
3690int ceph_encode_inode_release(void **p, struct inode *inode,
3691 int mds, int drop, int unless, int force)
3692{
3693 struct ceph_inode_info *ci = ceph_inode(inode);
3694 struct ceph_cap *cap;
3695 struct ceph_mds_request_release *rel = *p;
3696 int used, dirty;
3697 int ret = 0;
3698
3699 spin_lock(&ci->i_ceph_lock);
3700 used = __ceph_caps_used(ci);
3701 dirty = __ceph_caps_dirty(ci);
3702
3703 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3704 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3705 ceph_cap_string(unless));
3706
3707 /* only drop unused, clean caps */
3708 drop &= ~(used | dirty);
3709
3710 cap = __get_cap_for_mds(ci, mds);
3711 if (cap && __cap_is_valid(cap)) {
3712 if (force ||
3713 ((cap->issued & drop) &&
3714 (cap->issued & unless) == 0)) {
3715 if ((cap->issued & drop) &&
3716 (cap->issued & unless) == 0) {
3717 int wanted = __ceph_caps_wanted(ci);
3718 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3719 wanted |= cap->mds_wanted;
3720 dout("encode_inode_release %p cap %p "
3721 "%s -> %s, wanted %s -> %s\n", inode, cap,
3722 ceph_cap_string(cap->issued),
3723 ceph_cap_string(cap->issued & ~drop),
3724 ceph_cap_string(cap->mds_wanted),
3725 ceph_cap_string(wanted));
3726
3727 cap->issued &= ~drop;
3728 cap->implemented &= ~drop;
3729 cap->mds_wanted = wanted;
3730 } else {
3731 dout("encode_inode_release %p cap %p %s"
3732 " (force)\n", inode, cap,
3733 ceph_cap_string(cap->issued));
3734 }
3735
3736 rel->ino = cpu_to_le64(ceph_ino(inode));
3737 rel->cap_id = cpu_to_le64(cap->cap_id);
3738 rel->seq = cpu_to_le32(cap->seq);
3739 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3740 rel->mseq = cpu_to_le32(cap->mseq);
3741 rel->caps = cpu_to_le32(cap->implemented);
3742 rel->wanted = cpu_to_le32(cap->mds_wanted);
3743 rel->dname_len = 0;
3744 rel->dname_seq = 0;
3745 *p += sizeof(*rel);
3746 ret = 1;
3747 } else {
3748 dout("encode_inode_release %p cap %p %s\n",
3749 inode, cap, ceph_cap_string(cap->issued));
3750 }
3751 }
3752 spin_unlock(&ci->i_ceph_lock);
3753 return ret;
3754}
3755
3756int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3757 int mds, int drop, int unless)
3758{
3759 struct inode *dir = d_inode(dentry->d_parent);
3760 struct ceph_mds_request_release *rel = *p;
3761 struct ceph_dentry_info *di = ceph_dentry(dentry);
3762 int force = 0;
3763 int ret;
3764
3765 /*
3766 * force an record for the directory caps if we have a dentry lease.
3767 * this is racy (can't take i_ceph_lock and d_lock together), but it
3768 * doesn't have to be perfect; the mds will revoke anything we don't
3769 * release.
3770 */
3771 spin_lock(&dentry->d_lock);
3772 if (di->lease_session && di->lease_session->s_mds == mds)
3773 force = 1;
3774 spin_unlock(&dentry->d_lock);
3775
3776 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3777
3778 spin_lock(&dentry->d_lock);
3779 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3780 dout("encode_dentry_release %p mds%d seq %d\n",
3781 dentry, mds, (int)di->lease_seq);
3782 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3783 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3784 *p += dentry->d_name.len;
3785 rel->dname_seq = cpu_to_le32(di->lease_seq);
3786 __ceph_mdsc_drop_dentry_lease(dentry);
3787 }
3788 spin_unlock(&dentry->d_lock);
3789 return ret;
3790}
1#include <linux/ceph/ceph_debug.h>
2
3#include <linux/fs.h>
4#include <linux/kernel.h>
5#include <linux/sched.h>
6#include <linux/slab.h>
7#include <linux/vmalloc.h>
8#include <linux/wait.h>
9#include <linux/writeback.h>
10
11#include "super.h"
12#include "mds_client.h"
13#include "cache.h"
14#include <linux/ceph/decode.h>
15#include <linux/ceph/messenger.h>
16
17/*
18 * Capability management
19 *
20 * The Ceph metadata servers control client access to inode metadata
21 * and file data by issuing capabilities, granting clients permission
22 * to read and/or write both inode field and file data to OSDs
23 * (storage nodes). Each capability consists of a set of bits
24 * indicating which operations are allowed.
25 *
26 * If the client holds a *_SHARED cap, the client has a coherent value
27 * that can be safely read from the cached inode.
28 *
29 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
30 * client is allowed to change inode attributes (e.g., file size,
31 * mtime), note its dirty state in the ceph_cap, and asynchronously
32 * flush that metadata change to the MDS.
33 *
34 * In the event of a conflicting operation (perhaps by another
35 * client), the MDS will revoke the conflicting client capabilities.
36 *
37 * In order for a client to cache an inode, it must hold a capability
38 * with at least one MDS server. When inodes are released, release
39 * notifications are batched and periodically sent en masse to the MDS
40 * cluster to release server state.
41 */
42
43
44/*
45 * Generate readable cap strings for debugging output.
46 */
47#define MAX_CAP_STR 20
48static char cap_str[MAX_CAP_STR][40];
49static DEFINE_SPINLOCK(cap_str_lock);
50static int last_cap_str;
51
52static char *gcap_string(char *s, int c)
53{
54 if (c & CEPH_CAP_GSHARED)
55 *s++ = 's';
56 if (c & CEPH_CAP_GEXCL)
57 *s++ = 'x';
58 if (c & CEPH_CAP_GCACHE)
59 *s++ = 'c';
60 if (c & CEPH_CAP_GRD)
61 *s++ = 'r';
62 if (c & CEPH_CAP_GWR)
63 *s++ = 'w';
64 if (c & CEPH_CAP_GBUFFER)
65 *s++ = 'b';
66 if (c & CEPH_CAP_GLAZYIO)
67 *s++ = 'l';
68 return s;
69}
70
71const char *ceph_cap_string(int caps)
72{
73 int i;
74 char *s;
75 int c;
76
77 spin_lock(&cap_str_lock);
78 i = last_cap_str++;
79 if (last_cap_str == MAX_CAP_STR)
80 last_cap_str = 0;
81 spin_unlock(&cap_str_lock);
82
83 s = cap_str[i];
84
85 if (caps & CEPH_CAP_PIN)
86 *s++ = 'p';
87
88 c = (caps >> CEPH_CAP_SAUTH) & 3;
89 if (c) {
90 *s++ = 'A';
91 s = gcap_string(s, c);
92 }
93
94 c = (caps >> CEPH_CAP_SLINK) & 3;
95 if (c) {
96 *s++ = 'L';
97 s = gcap_string(s, c);
98 }
99
100 c = (caps >> CEPH_CAP_SXATTR) & 3;
101 if (c) {
102 *s++ = 'X';
103 s = gcap_string(s, c);
104 }
105
106 c = caps >> CEPH_CAP_SFILE;
107 if (c) {
108 *s++ = 'F';
109 s = gcap_string(s, c);
110 }
111
112 if (s == cap_str[i])
113 *s++ = '-';
114 *s = 0;
115 return cap_str[i];
116}
117
118void ceph_caps_init(struct ceph_mds_client *mdsc)
119{
120 INIT_LIST_HEAD(&mdsc->caps_list);
121 spin_lock_init(&mdsc->caps_list_lock);
122}
123
124void ceph_caps_finalize(struct ceph_mds_client *mdsc)
125{
126 struct ceph_cap *cap;
127
128 spin_lock(&mdsc->caps_list_lock);
129 while (!list_empty(&mdsc->caps_list)) {
130 cap = list_first_entry(&mdsc->caps_list,
131 struct ceph_cap, caps_item);
132 list_del(&cap->caps_item);
133 kmem_cache_free(ceph_cap_cachep, cap);
134 }
135 mdsc->caps_total_count = 0;
136 mdsc->caps_avail_count = 0;
137 mdsc->caps_use_count = 0;
138 mdsc->caps_reserve_count = 0;
139 mdsc->caps_min_count = 0;
140 spin_unlock(&mdsc->caps_list_lock);
141}
142
143void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
144{
145 spin_lock(&mdsc->caps_list_lock);
146 mdsc->caps_min_count += delta;
147 BUG_ON(mdsc->caps_min_count < 0);
148 spin_unlock(&mdsc->caps_list_lock);
149}
150
151void ceph_reserve_caps(struct ceph_mds_client *mdsc,
152 struct ceph_cap_reservation *ctx, int need)
153{
154 int i;
155 struct ceph_cap *cap;
156 int have;
157 int alloc = 0;
158 LIST_HEAD(newcaps);
159
160 dout("reserve caps ctx=%p need=%d\n", ctx, need);
161
162 /* first reserve any caps that are already allocated */
163 spin_lock(&mdsc->caps_list_lock);
164 if (mdsc->caps_avail_count >= need)
165 have = need;
166 else
167 have = mdsc->caps_avail_count;
168 mdsc->caps_avail_count -= have;
169 mdsc->caps_reserve_count += have;
170 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
171 mdsc->caps_reserve_count +
172 mdsc->caps_avail_count);
173 spin_unlock(&mdsc->caps_list_lock);
174
175 for (i = have; i < need; i++) {
176 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
177 if (!cap)
178 break;
179 list_add(&cap->caps_item, &newcaps);
180 alloc++;
181 }
182 /* we didn't manage to reserve as much as we needed */
183 if (have + alloc != need)
184 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
185 ctx, need, have + alloc);
186
187 spin_lock(&mdsc->caps_list_lock);
188 mdsc->caps_total_count += alloc;
189 mdsc->caps_reserve_count += alloc;
190 list_splice(&newcaps, &mdsc->caps_list);
191
192 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
193 mdsc->caps_reserve_count +
194 mdsc->caps_avail_count);
195 spin_unlock(&mdsc->caps_list_lock);
196
197 ctx->count = need;
198 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
199 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
200 mdsc->caps_reserve_count, mdsc->caps_avail_count);
201}
202
203int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
204 struct ceph_cap_reservation *ctx)
205{
206 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
207 if (ctx->count) {
208 spin_lock(&mdsc->caps_list_lock);
209 BUG_ON(mdsc->caps_reserve_count < ctx->count);
210 mdsc->caps_reserve_count -= ctx->count;
211 mdsc->caps_avail_count += ctx->count;
212 ctx->count = 0;
213 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
214 mdsc->caps_total_count, mdsc->caps_use_count,
215 mdsc->caps_reserve_count, mdsc->caps_avail_count);
216 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
217 mdsc->caps_reserve_count +
218 mdsc->caps_avail_count);
219 spin_unlock(&mdsc->caps_list_lock);
220 }
221 return 0;
222}
223
224static struct ceph_cap *get_cap(struct ceph_mds_client *mdsc,
225 struct ceph_cap_reservation *ctx)
226{
227 struct ceph_cap *cap = NULL;
228
229 /* temporary, until we do something about cap import/export */
230 if (!ctx) {
231 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
232 if (cap) {
233 spin_lock(&mdsc->caps_list_lock);
234 mdsc->caps_use_count++;
235 mdsc->caps_total_count++;
236 spin_unlock(&mdsc->caps_list_lock);
237 }
238 return cap;
239 }
240
241 spin_lock(&mdsc->caps_list_lock);
242 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
243 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
244 mdsc->caps_reserve_count, mdsc->caps_avail_count);
245 BUG_ON(!ctx->count);
246 BUG_ON(ctx->count > mdsc->caps_reserve_count);
247 BUG_ON(list_empty(&mdsc->caps_list));
248
249 ctx->count--;
250 mdsc->caps_reserve_count--;
251 mdsc->caps_use_count++;
252
253 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
254 list_del(&cap->caps_item);
255
256 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
257 mdsc->caps_reserve_count + mdsc->caps_avail_count);
258 spin_unlock(&mdsc->caps_list_lock);
259 return cap;
260}
261
262void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
263{
264 spin_lock(&mdsc->caps_list_lock);
265 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
266 cap, mdsc->caps_total_count, mdsc->caps_use_count,
267 mdsc->caps_reserve_count, mdsc->caps_avail_count);
268 mdsc->caps_use_count--;
269 /*
270 * Keep some preallocated caps around (ceph_min_count), to
271 * avoid lots of free/alloc churn.
272 */
273 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
274 mdsc->caps_min_count) {
275 mdsc->caps_total_count--;
276 kmem_cache_free(ceph_cap_cachep, cap);
277 } else {
278 mdsc->caps_avail_count++;
279 list_add(&cap->caps_item, &mdsc->caps_list);
280 }
281
282 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
283 mdsc->caps_reserve_count + mdsc->caps_avail_count);
284 spin_unlock(&mdsc->caps_list_lock);
285}
286
287void ceph_reservation_status(struct ceph_fs_client *fsc,
288 int *total, int *avail, int *used, int *reserved,
289 int *min)
290{
291 struct ceph_mds_client *mdsc = fsc->mdsc;
292
293 if (total)
294 *total = mdsc->caps_total_count;
295 if (avail)
296 *avail = mdsc->caps_avail_count;
297 if (used)
298 *used = mdsc->caps_use_count;
299 if (reserved)
300 *reserved = mdsc->caps_reserve_count;
301 if (min)
302 *min = mdsc->caps_min_count;
303}
304
305/*
306 * Find ceph_cap for given mds, if any.
307 *
308 * Called with i_ceph_lock held.
309 */
310static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
311{
312 struct ceph_cap *cap;
313 struct rb_node *n = ci->i_caps.rb_node;
314
315 while (n) {
316 cap = rb_entry(n, struct ceph_cap, ci_node);
317 if (mds < cap->mds)
318 n = n->rb_left;
319 else if (mds > cap->mds)
320 n = n->rb_right;
321 else
322 return cap;
323 }
324 return NULL;
325}
326
327struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
328{
329 struct ceph_cap *cap;
330
331 spin_lock(&ci->i_ceph_lock);
332 cap = __get_cap_for_mds(ci, mds);
333 spin_unlock(&ci->i_ceph_lock);
334 return cap;
335}
336
337/*
338 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
339 */
340static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
341{
342 struct ceph_cap *cap;
343 int mds = -1;
344 struct rb_node *p;
345
346 /* prefer mds with WR|BUFFER|EXCL caps */
347 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
348 cap = rb_entry(p, struct ceph_cap, ci_node);
349 mds = cap->mds;
350 if (cap->issued & (CEPH_CAP_FILE_WR |
351 CEPH_CAP_FILE_BUFFER |
352 CEPH_CAP_FILE_EXCL))
353 break;
354 }
355 return mds;
356}
357
358int ceph_get_cap_mds(struct inode *inode)
359{
360 struct ceph_inode_info *ci = ceph_inode(inode);
361 int mds;
362 spin_lock(&ci->i_ceph_lock);
363 mds = __ceph_get_cap_mds(ceph_inode(inode));
364 spin_unlock(&ci->i_ceph_lock);
365 return mds;
366}
367
368/*
369 * Called under i_ceph_lock.
370 */
371static void __insert_cap_node(struct ceph_inode_info *ci,
372 struct ceph_cap *new)
373{
374 struct rb_node **p = &ci->i_caps.rb_node;
375 struct rb_node *parent = NULL;
376 struct ceph_cap *cap = NULL;
377
378 while (*p) {
379 parent = *p;
380 cap = rb_entry(parent, struct ceph_cap, ci_node);
381 if (new->mds < cap->mds)
382 p = &(*p)->rb_left;
383 else if (new->mds > cap->mds)
384 p = &(*p)->rb_right;
385 else
386 BUG();
387 }
388
389 rb_link_node(&new->ci_node, parent, p);
390 rb_insert_color(&new->ci_node, &ci->i_caps);
391}
392
393/*
394 * (re)set cap hold timeouts, which control the delayed release
395 * of unused caps back to the MDS. Should be called on cap use.
396 */
397static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
398 struct ceph_inode_info *ci)
399{
400 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
401
402 ci->i_hold_caps_min = round_jiffies(jiffies +
403 ma->caps_wanted_delay_min * HZ);
404 ci->i_hold_caps_max = round_jiffies(jiffies +
405 ma->caps_wanted_delay_max * HZ);
406 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
407 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
408}
409
410/*
411 * (Re)queue cap at the end of the delayed cap release list.
412 *
413 * If I_FLUSH is set, leave the inode at the front of the list.
414 *
415 * Caller holds i_ceph_lock
416 * -> we take mdsc->cap_delay_lock
417 */
418static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
419 struct ceph_inode_info *ci)
420{
421 __cap_set_timeouts(mdsc, ci);
422 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
423 ci->i_ceph_flags, ci->i_hold_caps_max);
424 if (!mdsc->stopping) {
425 spin_lock(&mdsc->cap_delay_lock);
426 if (!list_empty(&ci->i_cap_delay_list)) {
427 if (ci->i_ceph_flags & CEPH_I_FLUSH)
428 goto no_change;
429 list_del_init(&ci->i_cap_delay_list);
430 }
431 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
432no_change:
433 spin_unlock(&mdsc->cap_delay_lock);
434 }
435}
436
437/*
438 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
439 * indicating we should send a cap message to flush dirty metadata
440 * asap, and move to the front of the delayed cap list.
441 */
442static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
443 struct ceph_inode_info *ci)
444{
445 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
446 spin_lock(&mdsc->cap_delay_lock);
447 ci->i_ceph_flags |= CEPH_I_FLUSH;
448 if (!list_empty(&ci->i_cap_delay_list))
449 list_del_init(&ci->i_cap_delay_list);
450 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
451 spin_unlock(&mdsc->cap_delay_lock);
452}
453
454/*
455 * Cancel delayed work on cap.
456 *
457 * Caller must hold i_ceph_lock.
458 */
459static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
460 struct ceph_inode_info *ci)
461{
462 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
463 if (list_empty(&ci->i_cap_delay_list))
464 return;
465 spin_lock(&mdsc->cap_delay_lock);
466 list_del_init(&ci->i_cap_delay_list);
467 spin_unlock(&mdsc->cap_delay_lock);
468}
469
470/*
471 * Common issue checks for add_cap, handle_cap_grant.
472 */
473static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
474 unsigned issued)
475{
476 unsigned had = __ceph_caps_issued(ci, NULL);
477
478 /*
479 * Each time we receive FILE_CACHE anew, we increment
480 * i_rdcache_gen.
481 */
482 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
483 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
484 ci->i_rdcache_gen++;
485 }
486
487 /*
488 * if we are newly issued FILE_SHARED, mark dir not complete; we
489 * don't know what happened to this directory while we didn't
490 * have the cap.
491 */
492 if ((issued & CEPH_CAP_FILE_SHARED) &&
493 (had & CEPH_CAP_FILE_SHARED) == 0) {
494 ci->i_shared_gen++;
495 if (S_ISDIR(ci->vfs_inode.i_mode)) {
496 dout(" marking %p NOT complete\n", &ci->vfs_inode);
497 __ceph_dir_clear_complete(ci);
498 }
499 }
500}
501
502/*
503 * Add a capability under the given MDS session.
504 *
505 * Caller should hold session snap_rwsem (read) and s_mutex.
506 *
507 * @fmode is the open file mode, if we are opening a file, otherwise
508 * it is < 0. (This is so we can atomically add the cap and add an
509 * open file reference to it.)
510 */
511int ceph_add_cap(struct inode *inode,
512 struct ceph_mds_session *session, u64 cap_id,
513 int fmode, unsigned issued, unsigned wanted,
514 unsigned seq, unsigned mseq, u64 realmino, int flags,
515 struct ceph_cap_reservation *caps_reservation)
516{
517 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
518 struct ceph_inode_info *ci = ceph_inode(inode);
519 struct ceph_cap *new_cap = NULL;
520 struct ceph_cap *cap;
521 int mds = session->s_mds;
522 int actual_wanted;
523
524 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
525 session->s_mds, cap_id, ceph_cap_string(issued), seq);
526
527 /*
528 * If we are opening the file, include file mode wanted bits
529 * in wanted.
530 */
531 if (fmode >= 0)
532 wanted |= ceph_caps_for_mode(fmode);
533
534retry:
535 spin_lock(&ci->i_ceph_lock);
536 cap = __get_cap_for_mds(ci, mds);
537 if (!cap) {
538 if (new_cap) {
539 cap = new_cap;
540 new_cap = NULL;
541 } else {
542 spin_unlock(&ci->i_ceph_lock);
543 new_cap = get_cap(mdsc, caps_reservation);
544 if (new_cap == NULL)
545 return -ENOMEM;
546 goto retry;
547 }
548
549 cap->issued = 0;
550 cap->implemented = 0;
551 cap->mds = mds;
552 cap->mds_wanted = 0;
553 cap->mseq = 0;
554
555 cap->ci = ci;
556 __insert_cap_node(ci, cap);
557
558 /* add to session cap list */
559 cap->session = session;
560 spin_lock(&session->s_cap_lock);
561 list_add_tail(&cap->session_caps, &session->s_caps);
562 session->s_nr_caps++;
563 spin_unlock(&session->s_cap_lock);
564 } else {
565 if (new_cap)
566 ceph_put_cap(mdsc, new_cap);
567
568 /*
569 * auth mds of the inode changed. we received the cap export
570 * message, but still haven't received the cap import message.
571 * handle_cap_export() updated the new auth MDS' cap.
572 *
573 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
574 * a message that was send before the cap import message. So
575 * don't remove caps.
576 */
577 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
578 WARN_ON(cap != ci->i_auth_cap);
579 WARN_ON(cap->cap_id != cap_id);
580 seq = cap->seq;
581 mseq = cap->mseq;
582 issued |= cap->issued;
583 flags |= CEPH_CAP_FLAG_AUTH;
584 }
585 }
586
587 if (!ci->i_snap_realm) {
588 /*
589 * add this inode to the appropriate snap realm
590 */
591 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
592 realmino);
593 if (realm) {
594 ceph_get_snap_realm(mdsc, realm);
595 spin_lock(&realm->inodes_with_caps_lock);
596 ci->i_snap_realm = realm;
597 list_add(&ci->i_snap_realm_item,
598 &realm->inodes_with_caps);
599 spin_unlock(&realm->inodes_with_caps_lock);
600 } else {
601 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
602 realmino);
603 WARN_ON(!realm);
604 }
605 }
606
607 __check_cap_issue(ci, cap, issued);
608
609 /*
610 * If we are issued caps we don't want, or the mds' wanted
611 * value appears to be off, queue a check so we'll release
612 * later and/or update the mds wanted value.
613 */
614 actual_wanted = __ceph_caps_wanted(ci);
615 if ((wanted & ~actual_wanted) ||
616 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
617 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
618 ceph_cap_string(issued), ceph_cap_string(wanted),
619 ceph_cap_string(actual_wanted));
620 __cap_delay_requeue(mdsc, ci);
621 }
622
623 if (flags & CEPH_CAP_FLAG_AUTH) {
624 if (ci->i_auth_cap == NULL ||
625 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
626 ci->i_auth_cap = cap;
627 cap->mds_wanted = wanted;
628 }
629 ci->i_cap_exporting_issued = 0;
630 } else {
631 WARN_ON(ci->i_auth_cap == cap);
632 }
633
634 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
635 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
636 ceph_cap_string(issued|cap->issued), seq, mds);
637 cap->cap_id = cap_id;
638 cap->issued = issued;
639 cap->implemented |= issued;
640 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
641 cap->mds_wanted = wanted;
642 else
643 cap->mds_wanted |= wanted;
644 cap->seq = seq;
645 cap->issue_seq = seq;
646 cap->mseq = mseq;
647 cap->cap_gen = session->s_cap_gen;
648
649 if (fmode >= 0)
650 __ceph_get_fmode(ci, fmode);
651 spin_unlock(&ci->i_ceph_lock);
652 wake_up_all(&ci->i_cap_wq);
653 return 0;
654}
655
656/*
657 * Return true if cap has not timed out and belongs to the current
658 * generation of the MDS session (i.e. has not gone 'stale' due to
659 * us losing touch with the mds).
660 */
661static int __cap_is_valid(struct ceph_cap *cap)
662{
663 unsigned long ttl;
664 u32 gen;
665
666 spin_lock(&cap->session->s_gen_ttl_lock);
667 gen = cap->session->s_cap_gen;
668 ttl = cap->session->s_cap_ttl;
669 spin_unlock(&cap->session->s_gen_ttl_lock);
670
671 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
672 dout("__cap_is_valid %p cap %p issued %s "
673 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
674 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
675 return 0;
676 }
677
678 return 1;
679}
680
681/*
682 * Return set of valid cap bits issued to us. Note that caps time
683 * out, and may be invalidated in bulk if the client session times out
684 * and session->s_cap_gen is bumped.
685 */
686int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
687{
688 int have = ci->i_snap_caps | ci->i_cap_exporting_issued;
689 struct ceph_cap *cap;
690 struct rb_node *p;
691
692 if (implemented)
693 *implemented = 0;
694 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
695 cap = rb_entry(p, struct ceph_cap, ci_node);
696 if (!__cap_is_valid(cap))
697 continue;
698 dout("__ceph_caps_issued %p cap %p issued %s\n",
699 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
700 have |= cap->issued;
701 if (implemented)
702 *implemented |= cap->implemented;
703 }
704 /*
705 * exclude caps issued by non-auth MDS, but are been revoking
706 * by the auth MDS. The non-auth MDS should be revoking/exporting
707 * these caps, but the message is delayed.
708 */
709 if (ci->i_auth_cap) {
710 cap = ci->i_auth_cap;
711 have &= ~cap->implemented | cap->issued;
712 }
713 return have;
714}
715
716/*
717 * Get cap bits issued by caps other than @ocap
718 */
719int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
720{
721 int have = ci->i_snap_caps;
722 struct ceph_cap *cap;
723 struct rb_node *p;
724
725 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
726 cap = rb_entry(p, struct ceph_cap, ci_node);
727 if (cap == ocap)
728 continue;
729 if (!__cap_is_valid(cap))
730 continue;
731 have |= cap->issued;
732 }
733 return have;
734}
735
736/*
737 * Move a cap to the end of the LRU (oldest caps at list head, newest
738 * at list tail).
739 */
740static void __touch_cap(struct ceph_cap *cap)
741{
742 struct ceph_mds_session *s = cap->session;
743
744 spin_lock(&s->s_cap_lock);
745 if (s->s_cap_iterator == NULL) {
746 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
747 s->s_mds);
748 list_move_tail(&cap->session_caps, &s->s_caps);
749 } else {
750 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
751 &cap->ci->vfs_inode, cap, s->s_mds);
752 }
753 spin_unlock(&s->s_cap_lock);
754}
755
756/*
757 * Check if we hold the given mask. If so, move the cap(s) to the
758 * front of their respective LRUs. (This is the preferred way for
759 * callers to check for caps they want.)
760 */
761int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
762{
763 struct ceph_cap *cap;
764 struct rb_node *p;
765 int have = ci->i_snap_caps;
766
767 if ((have & mask) == mask) {
768 dout("__ceph_caps_issued_mask %p snap issued %s"
769 " (mask %s)\n", &ci->vfs_inode,
770 ceph_cap_string(have),
771 ceph_cap_string(mask));
772 return 1;
773 }
774
775 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
776 cap = rb_entry(p, struct ceph_cap, ci_node);
777 if (!__cap_is_valid(cap))
778 continue;
779 if ((cap->issued & mask) == mask) {
780 dout("__ceph_caps_issued_mask %p cap %p issued %s"
781 " (mask %s)\n", &ci->vfs_inode, cap,
782 ceph_cap_string(cap->issued),
783 ceph_cap_string(mask));
784 if (touch)
785 __touch_cap(cap);
786 return 1;
787 }
788
789 /* does a combination of caps satisfy mask? */
790 have |= cap->issued;
791 if ((have & mask) == mask) {
792 dout("__ceph_caps_issued_mask %p combo issued %s"
793 " (mask %s)\n", &ci->vfs_inode,
794 ceph_cap_string(cap->issued),
795 ceph_cap_string(mask));
796 if (touch) {
797 struct rb_node *q;
798
799 /* touch this + preceding caps */
800 __touch_cap(cap);
801 for (q = rb_first(&ci->i_caps); q != p;
802 q = rb_next(q)) {
803 cap = rb_entry(q, struct ceph_cap,
804 ci_node);
805 if (!__cap_is_valid(cap))
806 continue;
807 __touch_cap(cap);
808 }
809 }
810 return 1;
811 }
812 }
813
814 return 0;
815}
816
817/*
818 * Return true if mask caps are currently being revoked by an MDS.
819 */
820int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
821 struct ceph_cap *ocap, int mask)
822{
823 struct ceph_cap *cap;
824 struct rb_node *p;
825
826 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
827 cap = rb_entry(p, struct ceph_cap, ci_node);
828 if (cap != ocap &&
829 (cap->implemented & ~cap->issued & mask))
830 return 1;
831 }
832 return 0;
833}
834
835int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
836{
837 struct inode *inode = &ci->vfs_inode;
838 int ret;
839
840 spin_lock(&ci->i_ceph_lock);
841 ret = __ceph_caps_revoking_other(ci, NULL, mask);
842 spin_unlock(&ci->i_ceph_lock);
843 dout("ceph_caps_revoking %p %s = %d\n", inode,
844 ceph_cap_string(mask), ret);
845 return ret;
846}
847
848int __ceph_caps_used(struct ceph_inode_info *ci)
849{
850 int used = 0;
851 if (ci->i_pin_ref)
852 used |= CEPH_CAP_PIN;
853 if (ci->i_rd_ref)
854 used |= CEPH_CAP_FILE_RD;
855 if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
856 used |= CEPH_CAP_FILE_CACHE;
857 if (ci->i_wr_ref)
858 used |= CEPH_CAP_FILE_WR;
859 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
860 used |= CEPH_CAP_FILE_BUFFER;
861 return used;
862}
863
864/*
865 * wanted, by virtue of open file modes
866 */
867int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
868{
869 int want = 0;
870 int mode;
871 for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
872 if (ci->i_nr_by_mode[mode])
873 want |= ceph_caps_for_mode(mode);
874 return want;
875}
876
877/*
878 * Return caps we have registered with the MDS(s) as 'wanted'.
879 */
880int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
881{
882 struct ceph_cap *cap;
883 struct rb_node *p;
884 int mds_wanted = 0;
885
886 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
887 cap = rb_entry(p, struct ceph_cap, ci_node);
888 if (!__cap_is_valid(cap))
889 continue;
890 if (cap == ci->i_auth_cap)
891 mds_wanted |= cap->mds_wanted;
892 else
893 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
894 }
895 return mds_wanted;
896}
897
898/*
899 * called under i_ceph_lock
900 */
901static int __ceph_is_any_caps(struct ceph_inode_info *ci)
902{
903 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_issued;
904}
905
906int ceph_is_any_caps(struct inode *inode)
907{
908 struct ceph_inode_info *ci = ceph_inode(inode);
909 int ret;
910
911 spin_lock(&ci->i_ceph_lock);
912 ret = __ceph_is_any_caps(ci);
913 spin_unlock(&ci->i_ceph_lock);
914
915 return ret;
916}
917
918/*
919 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
920 *
921 * caller should hold i_ceph_lock.
922 * caller will not hold session s_mutex if called from destroy_inode.
923 */
924void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
925{
926 struct ceph_mds_session *session = cap->session;
927 struct ceph_inode_info *ci = cap->ci;
928 struct ceph_mds_client *mdsc =
929 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
930 int removed = 0;
931
932 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
933
934 /* remove from session list */
935 spin_lock(&session->s_cap_lock);
936 /*
937 * s_cap_reconnect is protected by s_cap_lock. no one changes
938 * s_cap_gen while session is in the reconnect state.
939 */
940 if (queue_release &&
941 (!session->s_cap_reconnect ||
942 cap->cap_gen == session->s_cap_gen))
943 __queue_cap_release(session, ci->i_vino.ino, cap->cap_id,
944 cap->mseq, cap->issue_seq);
945
946 if (session->s_cap_iterator == cap) {
947 /* not yet, we are iterating over this very cap */
948 dout("__ceph_remove_cap delaying %p removal from session %p\n",
949 cap, cap->session);
950 } else {
951 list_del_init(&cap->session_caps);
952 session->s_nr_caps--;
953 cap->session = NULL;
954 removed = 1;
955 }
956 /* protect backpointer with s_cap_lock: see iterate_session_caps */
957 cap->ci = NULL;
958 spin_unlock(&session->s_cap_lock);
959
960 /* remove from inode list */
961 rb_erase(&cap->ci_node, &ci->i_caps);
962 if (ci->i_auth_cap == cap)
963 ci->i_auth_cap = NULL;
964
965 if (removed)
966 ceph_put_cap(mdsc, cap);
967
968 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
969 struct ceph_snap_realm *realm = ci->i_snap_realm;
970 spin_lock(&realm->inodes_with_caps_lock);
971 list_del_init(&ci->i_snap_realm_item);
972 ci->i_snap_realm_counter++;
973 ci->i_snap_realm = NULL;
974 spin_unlock(&realm->inodes_with_caps_lock);
975 ceph_put_snap_realm(mdsc, realm);
976 }
977 if (!__ceph_is_any_real_caps(ci))
978 __cap_delay_cancel(mdsc, ci);
979}
980
981/*
982 * Build and send a cap message to the given MDS.
983 *
984 * Caller should be holding s_mutex.
985 */
986static int send_cap_msg(struct ceph_mds_session *session,
987 u64 ino, u64 cid, int op,
988 int caps, int wanted, int dirty,
989 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
990 u64 size, u64 max_size,
991 struct timespec *mtime, struct timespec *atime,
992 u64 time_warp_seq,
993 kuid_t uid, kgid_t gid, umode_t mode,
994 u64 xattr_version,
995 struct ceph_buffer *xattrs_buf,
996 u64 follows)
997{
998 struct ceph_mds_caps *fc;
999 struct ceph_msg *msg;
1000
1001 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1002 " seq %u/%u mseq %u follows %lld size %llu/%llu"
1003 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
1004 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
1005 ceph_cap_string(dirty),
1006 seq, issue_seq, mseq, follows, size, max_size,
1007 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1008
1009 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), GFP_NOFS, false);
1010 if (!msg)
1011 return -ENOMEM;
1012
1013 msg->hdr.tid = cpu_to_le64(flush_tid);
1014
1015 fc = msg->front.iov_base;
1016 memset(fc, 0, sizeof(*fc));
1017
1018 fc->cap_id = cpu_to_le64(cid);
1019 fc->op = cpu_to_le32(op);
1020 fc->seq = cpu_to_le32(seq);
1021 fc->issue_seq = cpu_to_le32(issue_seq);
1022 fc->migrate_seq = cpu_to_le32(mseq);
1023 fc->caps = cpu_to_le32(caps);
1024 fc->wanted = cpu_to_le32(wanted);
1025 fc->dirty = cpu_to_le32(dirty);
1026 fc->ino = cpu_to_le64(ino);
1027 fc->snap_follows = cpu_to_le64(follows);
1028
1029 fc->size = cpu_to_le64(size);
1030 fc->max_size = cpu_to_le64(max_size);
1031 if (mtime)
1032 ceph_encode_timespec(&fc->mtime, mtime);
1033 if (atime)
1034 ceph_encode_timespec(&fc->atime, atime);
1035 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1036
1037 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1038 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1039 fc->mode = cpu_to_le32(mode);
1040
1041 fc->xattr_version = cpu_to_le64(xattr_version);
1042 if (xattrs_buf) {
1043 msg->middle = ceph_buffer_get(xattrs_buf);
1044 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1045 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1046 }
1047
1048 ceph_con_send(&session->s_con, msg);
1049 return 0;
1050}
1051
1052void __queue_cap_release(struct ceph_mds_session *session,
1053 u64 ino, u64 cap_id, u32 migrate_seq,
1054 u32 issue_seq)
1055{
1056 struct ceph_msg *msg;
1057 struct ceph_mds_cap_release *head;
1058 struct ceph_mds_cap_item *item;
1059
1060 BUG_ON(!session->s_num_cap_releases);
1061 msg = list_first_entry(&session->s_cap_releases,
1062 struct ceph_msg, list_head);
1063
1064 dout(" adding %llx release to mds%d msg %p (%d left)\n",
1065 ino, session->s_mds, msg, session->s_num_cap_releases);
1066
1067 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1068 head = msg->front.iov_base;
1069 le32_add_cpu(&head->num, 1);
1070 item = msg->front.iov_base + msg->front.iov_len;
1071 item->ino = cpu_to_le64(ino);
1072 item->cap_id = cpu_to_le64(cap_id);
1073 item->migrate_seq = cpu_to_le32(migrate_seq);
1074 item->seq = cpu_to_le32(issue_seq);
1075
1076 session->s_num_cap_releases--;
1077
1078 msg->front.iov_len += sizeof(*item);
1079 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1080 dout(" release msg %p full\n", msg);
1081 list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1082 } else {
1083 dout(" release msg %p at %d/%d (%d)\n", msg,
1084 (int)le32_to_cpu(head->num),
1085 (int)CEPH_CAPS_PER_RELEASE,
1086 (int)msg->front.iov_len);
1087 }
1088}
1089
1090/*
1091 * Queue cap releases when an inode is dropped from our cache. Since
1092 * inode is about to be destroyed, there is no need for i_ceph_lock.
1093 */
1094void ceph_queue_caps_release(struct inode *inode)
1095{
1096 struct ceph_inode_info *ci = ceph_inode(inode);
1097 struct rb_node *p;
1098
1099 p = rb_first(&ci->i_caps);
1100 while (p) {
1101 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1102 p = rb_next(p);
1103 __ceph_remove_cap(cap, true);
1104 }
1105}
1106
1107/*
1108 * Send a cap msg on the given inode. Update our caps state, then
1109 * drop i_ceph_lock and send the message.
1110 *
1111 * Make note of max_size reported/requested from mds, revoked caps
1112 * that have now been implemented.
1113 *
1114 * Make half-hearted attempt ot to invalidate page cache if we are
1115 * dropping RDCACHE. Note that this will leave behind locked pages
1116 * that we'll then need to deal with elsewhere.
1117 *
1118 * Return non-zero if delayed release, or we experienced an error
1119 * such that the caller should requeue + retry later.
1120 *
1121 * called with i_ceph_lock, then drops it.
1122 * caller should hold snap_rwsem (read), s_mutex.
1123 */
1124static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1125 int op, int used, int want, int retain, int flushing,
1126 unsigned *pflush_tid)
1127 __releases(cap->ci->i_ceph_lock)
1128{
1129 struct ceph_inode_info *ci = cap->ci;
1130 struct inode *inode = &ci->vfs_inode;
1131 u64 cap_id = cap->cap_id;
1132 int held, revoking, dropping, keep;
1133 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1134 u64 size, max_size;
1135 struct timespec mtime, atime;
1136 int wake = 0;
1137 umode_t mode;
1138 kuid_t uid;
1139 kgid_t gid;
1140 struct ceph_mds_session *session;
1141 u64 xattr_version = 0;
1142 struct ceph_buffer *xattr_blob = NULL;
1143 int delayed = 0;
1144 u64 flush_tid = 0;
1145 int i;
1146 int ret;
1147
1148 held = cap->issued | cap->implemented;
1149 revoking = cap->implemented & ~cap->issued;
1150 retain &= ~revoking;
1151 dropping = cap->issued & ~retain;
1152
1153 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1154 inode, cap, cap->session,
1155 ceph_cap_string(held), ceph_cap_string(held & retain),
1156 ceph_cap_string(revoking));
1157 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1158
1159 session = cap->session;
1160
1161 /* don't release wanted unless we've waited a bit. */
1162 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1163 time_before(jiffies, ci->i_hold_caps_min)) {
1164 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1165 ceph_cap_string(cap->issued),
1166 ceph_cap_string(cap->issued & retain),
1167 ceph_cap_string(cap->mds_wanted),
1168 ceph_cap_string(want));
1169 want |= cap->mds_wanted;
1170 retain |= cap->issued;
1171 delayed = 1;
1172 }
1173 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1174
1175 cap->issued &= retain; /* drop bits we don't want */
1176 if (cap->implemented & ~cap->issued) {
1177 /*
1178 * Wake up any waiters on wanted -> needed transition.
1179 * This is due to the weird transition from buffered
1180 * to sync IO... we need to flush dirty pages _before_
1181 * allowing sync writes to avoid reordering.
1182 */
1183 wake = 1;
1184 }
1185 cap->implemented &= cap->issued | used;
1186 cap->mds_wanted = want;
1187
1188 if (flushing) {
1189 /*
1190 * assign a tid for flush operations so we can avoid
1191 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1192 * clean type races. track latest tid for every bit
1193 * so we can handle flush AxFw, flush Fw, and have the
1194 * first ack clean Ax.
1195 */
1196 flush_tid = ++ci->i_cap_flush_last_tid;
1197 if (pflush_tid)
1198 *pflush_tid = flush_tid;
1199 dout(" cap_flush_tid %d\n", (int)flush_tid);
1200 for (i = 0; i < CEPH_CAP_BITS; i++)
1201 if (flushing & (1 << i))
1202 ci->i_cap_flush_tid[i] = flush_tid;
1203
1204 follows = ci->i_head_snapc->seq;
1205 } else {
1206 follows = 0;
1207 }
1208
1209 keep = cap->implemented;
1210 seq = cap->seq;
1211 issue_seq = cap->issue_seq;
1212 mseq = cap->mseq;
1213 size = inode->i_size;
1214 ci->i_reported_size = size;
1215 max_size = ci->i_wanted_max_size;
1216 ci->i_requested_max_size = max_size;
1217 mtime = inode->i_mtime;
1218 atime = inode->i_atime;
1219 time_warp_seq = ci->i_time_warp_seq;
1220 uid = inode->i_uid;
1221 gid = inode->i_gid;
1222 mode = inode->i_mode;
1223
1224 if (flushing & CEPH_CAP_XATTR_EXCL) {
1225 __ceph_build_xattrs_blob(ci);
1226 xattr_blob = ci->i_xattrs.blob;
1227 xattr_version = ci->i_xattrs.version;
1228 }
1229
1230 spin_unlock(&ci->i_ceph_lock);
1231
1232 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1233 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1234 size, max_size, &mtime, &atime, time_warp_seq,
1235 uid, gid, mode, xattr_version, xattr_blob,
1236 follows);
1237 if (ret < 0) {
1238 dout("error sending cap msg, must requeue %p\n", inode);
1239 delayed = 1;
1240 }
1241
1242 if (wake)
1243 wake_up_all(&ci->i_cap_wq);
1244
1245 return delayed;
1246}
1247
1248/*
1249 * When a snapshot is taken, clients accumulate dirty metadata on
1250 * inodes with capabilities in ceph_cap_snaps to describe the file
1251 * state at the time the snapshot was taken. This must be flushed
1252 * asynchronously back to the MDS once sync writes complete and dirty
1253 * data is written out.
1254 *
1255 * Unless @again is true, skip cap_snaps that were already sent to
1256 * the MDS (i.e., during this session).
1257 *
1258 * Called under i_ceph_lock. Takes s_mutex as needed.
1259 */
1260void __ceph_flush_snaps(struct ceph_inode_info *ci,
1261 struct ceph_mds_session **psession,
1262 int again)
1263 __releases(ci->i_ceph_lock)
1264 __acquires(ci->i_ceph_lock)
1265{
1266 struct inode *inode = &ci->vfs_inode;
1267 int mds;
1268 struct ceph_cap_snap *capsnap;
1269 u32 mseq;
1270 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1271 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1272 session->s_mutex */
1273 u64 next_follows = 0; /* keep track of how far we've gotten through the
1274 i_cap_snaps list, and skip these entries next time
1275 around to avoid an infinite loop */
1276
1277 if (psession)
1278 session = *psession;
1279
1280 dout("__flush_snaps %p\n", inode);
1281retry:
1282 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1283 /* avoid an infiniute loop after retry */
1284 if (capsnap->follows < next_follows)
1285 continue;
1286 /*
1287 * we need to wait for sync writes to complete and for dirty
1288 * pages to be written out.
1289 */
1290 if (capsnap->dirty_pages || capsnap->writing)
1291 break;
1292
1293 /*
1294 * if cap writeback already occurred, we should have dropped
1295 * the capsnap in ceph_put_wrbuffer_cap_refs.
1296 */
1297 BUG_ON(capsnap->dirty == 0);
1298
1299 /* pick mds, take s_mutex */
1300 if (ci->i_auth_cap == NULL) {
1301 dout("no auth cap (migrating?), doing nothing\n");
1302 goto out;
1303 }
1304
1305 /* only flush each capsnap once */
1306 if (!again && !list_empty(&capsnap->flushing_item)) {
1307 dout("already flushed %p, skipping\n", capsnap);
1308 continue;
1309 }
1310
1311 mds = ci->i_auth_cap->session->s_mds;
1312 mseq = ci->i_auth_cap->mseq;
1313
1314 if (session && session->s_mds != mds) {
1315 dout("oops, wrong session %p mutex\n", session);
1316 mutex_unlock(&session->s_mutex);
1317 ceph_put_mds_session(session);
1318 session = NULL;
1319 }
1320 if (!session) {
1321 spin_unlock(&ci->i_ceph_lock);
1322 mutex_lock(&mdsc->mutex);
1323 session = __ceph_lookup_mds_session(mdsc, mds);
1324 mutex_unlock(&mdsc->mutex);
1325 if (session) {
1326 dout("inverting session/ino locks on %p\n",
1327 session);
1328 mutex_lock(&session->s_mutex);
1329 }
1330 /*
1331 * if session == NULL, we raced against a cap
1332 * deletion or migration. retry, and we'll
1333 * get a better @mds value next time.
1334 */
1335 spin_lock(&ci->i_ceph_lock);
1336 goto retry;
1337 }
1338
1339 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1340 atomic_inc(&capsnap->nref);
1341 if (!list_empty(&capsnap->flushing_item))
1342 list_del_init(&capsnap->flushing_item);
1343 list_add_tail(&capsnap->flushing_item,
1344 &session->s_cap_snaps_flushing);
1345 spin_unlock(&ci->i_ceph_lock);
1346
1347 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1348 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1349 send_cap_msg(session, ceph_vino(inode).ino, 0,
1350 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1351 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1352 capsnap->size, 0,
1353 &capsnap->mtime, &capsnap->atime,
1354 capsnap->time_warp_seq,
1355 capsnap->uid, capsnap->gid, capsnap->mode,
1356 capsnap->xattr_version, capsnap->xattr_blob,
1357 capsnap->follows);
1358
1359 next_follows = capsnap->follows + 1;
1360 ceph_put_cap_snap(capsnap);
1361
1362 spin_lock(&ci->i_ceph_lock);
1363 goto retry;
1364 }
1365
1366 /* we flushed them all; remove this inode from the queue */
1367 spin_lock(&mdsc->snap_flush_lock);
1368 list_del_init(&ci->i_snap_flush_item);
1369 spin_unlock(&mdsc->snap_flush_lock);
1370
1371out:
1372 if (psession)
1373 *psession = session;
1374 else if (session) {
1375 mutex_unlock(&session->s_mutex);
1376 ceph_put_mds_session(session);
1377 }
1378}
1379
1380static void ceph_flush_snaps(struct ceph_inode_info *ci)
1381{
1382 spin_lock(&ci->i_ceph_lock);
1383 __ceph_flush_snaps(ci, NULL, 0);
1384 spin_unlock(&ci->i_ceph_lock);
1385}
1386
1387/*
1388 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1389 * Caller is then responsible for calling __mark_inode_dirty with the
1390 * returned flags value.
1391 */
1392int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1393{
1394 struct ceph_mds_client *mdsc =
1395 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1396 struct inode *inode = &ci->vfs_inode;
1397 int was = ci->i_dirty_caps;
1398 int dirty = 0;
1399
1400 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1401 ceph_cap_string(mask), ceph_cap_string(was),
1402 ceph_cap_string(was | mask));
1403 ci->i_dirty_caps |= mask;
1404 if (was == 0) {
1405 if (!ci->i_head_snapc)
1406 ci->i_head_snapc = ceph_get_snap_context(
1407 ci->i_snap_realm->cached_context);
1408 dout(" inode %p now dirty snapc %p auth cap %p\n",
1409 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1410 WARN_ON(!ci->i_auth_cap);
1411 BUG_ON(!list_empty(&ci->i_dirty_item));
1412 spin_lock(&mdsc->cap_dirty_lock);
1413 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1414 spin_unlock(&mdsc->cap_dirty_lock);
1415 if (ci->i_flushing_caps == 0) {
1416 ihold(inode);
1417 dirty |= I_DIRTY_SYNC;
1418 }
1419 }
1420 BUG_ON(list_empty(&ci->i_dirty_item));
1421 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1422 (mask & CEPH_CAP_FILE_BUFFER))
1423 dirty |= I_DIRTY_DATASYNC;
1424 __cap_delay_requeue(mdsc, ci);
1425 return dirty;
1426}
1427
1428/*
1429 * Add dirty inode to the flushing list. Assigned a seq number so we
1430 * can wait for caps to flush without starving.
1431 *
1432 * Called under i_ceph_lock.
1433 */
1434static int __mark_caps_flushing(struct inode *inode,
1435 struct ceph_mds_session *session)
1436{
1437 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1438 struct ceph_inode_info *ci = ceph_inode(inode);
1439 int flushing;
1440
1441 BUG_ON(ci->i_dirty_caps == 0);
1442 BUG_ON(list_empty(&ci->i_dirty_item));
1443
1444 flushing = ci->i_dirty_caps;
1445 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1446 ceph_cap_string(flushing),
1447 ceph_cap_string(ci->i_flushing_caps),
1448 ceph_cap_string(ci->i_flushing_caps | flushing));
1449 ci->i_flushing_caps |= flushing;
1450 ci->i_dirty_caps = 0;
1451 dout(" inode %p now !dirty\n", inode);
1452
1453 spin_lock(&mdsc->cap_dirty_lock);
1454 list_del_init(&ci->i_dirty_item);
1455
1456 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1457 if (list_empty(&ci->i_flushing_item)) {
1458 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1459 mdsc->num_cap_flushing++;
1460 dout(" inode %p now flushing seq %lld\n", inode,
1461 ci->i_cap_flush_seq);
1462 } else {
1463 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1464 dout(" inode %p now flushing (more) seq %lld\n", inode,
1465 ci->i_cap_flush_seq);
1466 }
1467 spin_unlock(&mdsc->cap_dirty_lock);
1468
1469 return flushing;
1470}
1471
1472/*
1473 * try to invalidate mapping pages without blocking.
1474 */
1475static int try_nonblocking_invalidate(struct inode *inode)
1476{
1477 struct ceph_inode_info *ci = ceph_inode(inode);
1478 u32 invalidating_gen = ci->i_rdcache_gen;
1479
1480 spin_unlock(&ci->i_ceph_lock);
1481 invalidate_mapping_pages(&inode->i_data, 0, -1);
1482 spin_lock(&ci->i_ceph_lock);
1483
1484 if (inode->i_data.nrpages == 0 &&
1485 invalidating_gen == ci->i_rdcache_gen) {
1486 /* success. */
1487 dout("try_nonblocking_invalidate %p success\n", inode);
1488 /* save any racing async invalidate some trouble */
1489 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1490 return 0;
1491 }
1492 dout("try_nonblocking_invalidate %p failed\n", inode);
1493 return -1;
1494}
1495
1496/*
1497 * Swiss army knife function to examine currently used and wanted
1498 * versus held caps. Release, flush, ack revoked caps to mds as
1499 * appropriate.
1500 *
1501 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1502 * cap release further.
1503 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1504 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1505 * further delay.
1506 */
1507void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1508 struct ceph_mds_session *session)
1509{
1510 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1511 struct ceph_mds_client *mdsc = fsc->mdsc;
1512 struct inode *inode = &ci->vfs_inode;
1513 struct ceph_cap *cap;
1514 int file_wanted, used, cap_used;
1515 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1516 int issued, implemented, want, retain, revoking, flushing = 0;
1517 int mds = -1; /* keep track of how far we've gone through i_caps list
1518 to avoid an infinite loop on retry */
1519 struct rb_node *p;
1520 int tried_invalidate = 0;
1521 int delayed = 0, sent = 0, force_requeue = 0, num;
1522 int queue_invalidate = 0;
1523 int is_delayed = flags & CHECK_CAPS_NODELAY;
1524
1525 /* if we are unmounting, flush any unused caps immediately. */
1526 if (mdsc->stopping)
1527 is_delayed = 1;
1528
1529 spin_lock(&ci->i_ceph_lock);
1530
1531 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1532 flags |= CHECK_CAPS_FLUSH;
1533
1534 /* flush snaps first time around only */
1535 if (!list_empty(&ci->i_cap_snaps))
1536 __ceph_flush_snaps(ci, &session, 0);
1537 goto retry_locked;
1538retry:
1539 spin_lock(&ci->i_ceph_lock);
1540retry_locked:
1541 file_wanted = __ceph_caps_file_wanted(ci);
1542 used = __ceph_caps_used(ci);
1543 want = file_wanted | used;
1544 issued = __ceph_caps_issued(ci, &implemented);
1545 revoking = implemented & ~issued;
1546
1547 retain = want | CEPH_CAP_PIN;
1548 if (!mdsc->stopping && inode->i_nlink > 0) {
1549 if (want) {
1550 retain |= CEPH_CAP_ANY; /* be greedy */
1551 } else {
1552 retain |= CEPH_CAP_ANY_SHARED;
1553 /*
1554 * keep RD only if we didn't have the file open RW,
1555 * because then the mds would revoke it anyway to
1556 * journal max_size=0.
1557 */
1558 if (ci->i_max_size == 0)
1559 retain |= CEPH_CAP_ANY_RD;
1560 }
1561 }
1562
1563 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1564 " issued %s revoking %s retain %s %s%s%s\n", inode,
1565 ceph_cap_string(file_wanted),
1566 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1567 ceph_cap_string(ci->i_flushing_caps),
1568 ceph_cap_string(issued), ceph_cap_string(revoking),
1569 ceph_cap_string(retain),
1570 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1571 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1572 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1573
1574 /*
1575 * If we no longer need to hold onto old our caps, and we may
1576 * have cached pages, but don't want them, then try to invalidate.
1577 * If we fail, it's because pages are locked.... try again later.
1578 */
1579 if ((!is_delayed || mdsc->stopping) &&
1580 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1581 inode->i_data.nrpages && /* have cached pages */
1582 (file_wanted == 0 || /* no open files */
1583 (revoking & (CEPH_CAP_FILE_CACHE|
1584 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */
1585 !tried_invalidate) {
1586 dout("check_caps trying to invalidate on %p\n", inode);
1587 if (try_nonblocking_invalidate(inode) < 0) {
1588 if (revoking & (CEPH_CAP_FILE_CACHE|
1589 CEPH_CAP_FILE_LAZYIO)) {
1590 dout("check_caps queuing invalidate\n");
1591 queue_invalidate = 1;
1592 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1593 } else {
1594 dout("check_caps failed to invalidate pages\n");
1595 /* we failed to invalidate pages. check these
1596 caps again later. */
1597 force_requeue = 1;
1598 __cap_set_timeouts(mdsc, ci);
1599 }
1600 }
1601 tried_invalidate = 1;
1602 goto retry_locked;
1603 }
1604
1605 num = 0;
1606 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1607 cap = rb_entry(p, struct ceph_cap, ci_node);
1608 num++;
1609
1610 /* avoid looping forever */
1611 if (mds >= cap->mds ||
1612 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1613 continue;
1614
1615 /* NOTE: no side-effects allowed, until we take s_mutex */
1616
1617 cap_used = used;
1618 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1619 cap_used &= ~ci->i_auth_cap->issued;
1620
1621 revoking = cap->implemented & ~cap->issued;
1622 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1623 cap->mds, cap, ceph_cap_string(cap->issued),
1624 ceph_cap_string(cap_used),
1625 ceph_cap_string(cap->implemented),
1626 ceph_cap_string(revoking));
1627
1628 if (cap == ci->i_auth_cap &&
1629 (cap->issued & CEPH_CAP_FILE_WR)) {
1630 /* request larger max_size from MDS? */
1631 if (ci->i_wanted_max_size > ci->i_max_size &&
1632 ci->i_wanted_max_size > ci->i_requested_max_size) {
1633 dout("requesting new max_size\n");
1634 goto ack;
1635 }
1636
1637 /* approaching file_max? */
1638 if ((inode->i_size << 1) >= ci->i_max_size &&
1639 (ci->i_reported_size << 1) < ci->i_max_size) {
1640 dout("i_size approaching max_size\n");
1641 goto ack;
1642 }
1643 }
1644 /* flush anything dirty? */
1645 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1646 ci->i_dirty_caps) {
1647 dout("flushing dirty caps\n");
1648 goto ack;
1649 }
1650
1651 /* completed revocation? going down and there are no caps? */
1652 if (revoking && (revoking & cap_used) == 0) {
1653 dout("completed revocation of %s\n",
1654 ceph_cap_string(cap->implemented & ~cap->issued));
1655 goto ack;
1656 }
1657
1658 /* want more caps from mds? */
1659 if (want & ~(cap->mds_wanted | cap->issued))
1660 goto ack;
1661
1662 /* things we might delay */
1663 if ((cap->issued & ~retain) == 0 &&
1664 cap->mds_wanted == want)
1665 continue; /* nope, all good */
1666
1667 if (is_delayed)
1668 goto ack;
1669
1670 /* delay? */
1671 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1672 time_before(jiffies, ci->i_hold_caps_max)) {
1673 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1674 ceph_cap_string(cap->issued),
1675 ceph_cap_string(cap->issued & retain),
1676 ceph_cap_string(cap->mds_wanted),
1677 ceph_cap_string(want));
1678 delayed++;
1679 continue;
1680 }
1681
1682ack:
1683 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1684 dout(" skipping %p I_NOFLUSH set\n", inode);
1685 continue;
1686 }
1687
1688 if (session && session != cap->session) {
1689 dout("oops, wrong session %p mutex\n", session);
1690 mutex_unlock(&session->s_mutex);
1691 session = NULL;
1692 }
1693 if (!session) {
1694 session = cap->session;
1695 if (mutex_trylock(&session->s_mutex) == 0) {
1696 dout("inverting session/ino locks on %p\n",
1697 session);
1698 spin_unlock(&ci->i_ceph_lock);
1699 if (took_snap_rwsem) {
1700 up_read(&mdsc->snap_rwsem);
1701 took_snap_rwsem = 0;
1702 }
1703 mutex_lock(&session->s_mutex);
1704 goto retry;
1705 }
1706 }
1707 /* take snap_rwsem after session mutex */
1708 if (!took_snap_rwsem) {
1709 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1710 dout("inverting snap/in locks on %p\n",
1711 inode);
1712 spin_unlock(&ci->i_ceph_lock);
1713 down_read(&mdsc->snap_rwsem);
1714 took_snap_rwsem = 1;
1715 goto retry;
1716 }
1717 took_snap_rwsem = 1;
1718 }
1719
1720 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1721 flushing = __mark_caps_flushing(inode, session);
1722 else
1723 flushing = 0;
1724
1725 mds = cap->mds; /* remember mds, so we don't repeat */
1726 sent++;
1727
1728 /* __send_cap drops i_ceph_lock */
1729 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1730 want, retain, flushing, NULL);
1731 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1732 }
1733
1734 /*
1735 * Reschedule delayed caps release if we delayed anything,
1736 * otherwise cancel.
1737 */
1738 if (delayed && is_delayed)
1739 force_requeue = 1; /* __send_cap delayed release; requeue */
1740 if (!delayed && !is_delayed)
1741 __cap_delay_cancel(mdsc, ci);
1742 else if (!is_delayed || force_requeue)
1743 __cap_delay_requeue(mdsc, ci);
1744
1745 spin_unlock(&ci->i_ceph_lock);
1746
1747 if (queue_invalidate)
1748 ceph_queue_invalidate(inode);
1749
1750 if (session)
1751 mutex_unlock(&session->s_mutex);
1752 if (took_snap_rwsem)
1753 up_read(&mdsc->snap_rwsem);
1754}
1755
1756/*
1757 * Try to flush dirty caps back to the auth mds.
1758 */
1759static int try_flush_caps(struct inode *inode, unsigned *flush_tid)
1760{
1761 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1762 struct ceph_inode_info *ci = ceph_inode(inode);
1763 int flushing = 0;
1764 struct ceph_mds_session *session = NULL;
1765
1766retry:
1767 spin_lock(&ci->i_ceph_lock);
1768 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1769 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1770 goto out;
1771 }
1772 if (ci->i_dirty_caps && ci->i_auth_cap) {
1773 struct ceph_cap *cap = ci->i_auth_cap;
1774 int used = __ceph_caps_used(ci);
1775 int want = __ceph_caps_wanted(ci);
1776 int delayed;
1777
1778 if (!session || session != cap->session) {
1779 spin_unlock(&ci->i_ceph_lock);
1780 if (session)
1781 mutex_unlock(&session->s_mutex);
1782 session = cap->session;
1783 mutex_lock(&session->s_mutex);
1784 goto retry;
1785 }
1786 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1787 goto out;
1788
1789 flushing = __mark_caps_flushing(inode, session);
1790
1791 /* __send_cap drops i_ceph_lock */
1792 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1793 cap->issued | cap->implemented, flushing,
1794 flush_tid);
1795 if (!delayed)
1796 goto out_unlocked;
1797
1798 spin_lock(&ci->i_ceph_lock);
1799 __cap_delay_requeue(mdsc, ci);
1800 }
1801out:
1802 spin_unlock(&ci->i_ceph_lock);
1803out_unlocked:
1804 if (session)
1805 mutex_unlock(&session->s_mutex);
1806 return flushing;
1807}
1808
1809/*
1810 * Return true if we've flushed caps through the given flush_tid.
1811 */
1812static int caps_are_flushed(struct inode *inode, unsigned tid)
1813{
1814 struct ceph_inode_info *ci = ceph_inode(inode);
1815 int i, ret = 1;
1816
1817 spin_lock(&ci->i_ceph_lock);
1818 for (i = 0; i < CEPH_CAP_BITS; i++)
1819 if ((ci->i_flushing_caps & (1 << i)) &&
1820 ci->i_cap_flush_tid[i] <= tid) {
1821 /* still flushing this bit */
1822 ret = 0;
1823 break;
1824 }
1825 spin_unlock(&ci->i_ceph_lock);
1826 return ret;
1827}
1828
1829/*
1830 * Wait on any unsafe replies for the given inode. First wait on the
1831 * newest request, and make that the upper bound. Then, if there are
1832 * more requests, keep waiting on the oldest as long as it is still older
1833 * than the original request.
1834 */
1835static void sync_write_wait(struct inode *inode)
1836{
1837 struct ceph_inode_info *ci = ceph_inode(inode);
1838 struct list_head *head = &ci->i_unsafe_writes;
1839 struct ceph_osd_request *req;
1840 u64 last_tid;
1841
1842 spin_lock(&ci->i_unsafe_lock);
1843 if (list_empty(head))
1844 goto out;
1845
1846 /* set upper bound as _last_ entry in chain */
1847 req = list_entry(head->prev, struct ceph_osd_request,
1848 r_unsafe_item);
1849 last_tid = req->r_tid;
1850
1851 do {
1852 ceph_osdc_get_request(req);
1853 spin_unlock(&ci->i_unsafe_lock);
1854 dout("sync_write_wait on tid %llu (until %llu)\n",
1855 req->r_tid, last_tid);
1856 wait_for_completion(&req->r_safe_completion);
1857 spin_lock(&ci->i_unsafe_lock);
1858 ceph_osdc_put_request(req);
1859
1860 /*
1861 * from here on look at first entry in chain, since we
1862 * only want to wait for anything older than last_tid
1863 */
1864 if (list_empty(head))
1865 break;
1866 req = list_entry(head->next, struct ceph_osd_request,
1867 r_unsafe_item);
1868 } while (req->r_tid < last_tid);
1869out:
1870 spin_unlock(&ci->i_unsafe_lock);
1871}
1872
1873int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1874{
1875 struct inode *inode = file->f_mapping->host;
1876 struct ceph_inode_info *ci = ceph_inode(inode);
1877 unsigned flush_tid;
1878 int ret;
1879 int dirty;
1880
1881 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1882 sync_write_wait(inode);
1883
1884 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1885 if (ret < 0)
1886 return ret;
1887 mutex_lock(&inode->i_mutex);
1888
1889 dirty = try_flush_caps(inode, &flush_tid);
1890 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1891
1892 /*
1893 * only wait on non-file metadata writeback (the mds
1894 * can recover size and mtime, so we don't need to
1895 * wait for that)
1896 */
1897 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1898 dout("fsync waiting for flush_tid %u\n", flush_tid);
1899 ret = wait_event_interruptible(ci->i_cap_wq,
1900 caps_are_flushed(inode, flush_tid));
1901 }
1902
1903 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1904 mutex_unlock(&inode->i_mutex);
1905 return ret;
1906}
1907
1908/*
1909 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1910 * queue inode for flush but don't do so immediately, because we can
1911 * get by with fewer MDS messages if we wait for data writeback to
1912 * complete first.
1913 */
1914int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1915{
1916 struct ceph_inode_info *ci = ceph_inode(inode);
1917 unsigned flush_tid;
1918 int err = 0;
1919 int dirty;
1920 int wait = wbc->sync_mode == WB_SYNC_ALL;
1921
1922 dout("write_inode %p wait=%d\n", inode, wait);
1923 if (wait) {
1924 dirty = try_flush_caps(inode, &flush_tid);
1925 if (dirty)
1926 err = wait_event_interruptible(ci->i_cap_wq,
1927 caps_are_flushed(inode, flush_tid));
1928 } else {
1929 struct ceph_mds_client *mdsc =
1930 ceph_sb_to_client(inode->i_sb)->mdsc;
1931
1932 spin_lock(&ci->i_ceph_lock);
1933 if (__ceph_caps_dirty(ci))
1934 __cap_delay_requeue_front(mdsc, ci);
1935 spin_unlock(&ci->i_ceph_lock);
1936 }
1937 return err;
1938}
1939
1940/*
1941 * After a recovering MDS goes active, we need to resend any caps
1942 * we were flushing.
1943 *
1944 * Caller holds session->s_mutex.
1945 */
1946static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1947 struct ceph_mds_session *session)
1948{
1949 struct ceph_cap_snap *capsnap;
1950
1951 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1952 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1953 flushing_item) {
1954 struct ceph_inode_info *ci = capsnap->ci;
1955 struct inode *inode = &ci->vfs_inode;
1956 struct ceph_cap *cap;
1957
1958 spin_lock(&ci->i_ceph_lock);
1959 cap = ci->i_auth_cap;
1960 if (cap && cap->session == session) {
1961 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1962 cap, capsnap);
1963 __ceph_flush_snaps(ci, &session, 1);
1964 } else {
1965 pr_err("%p auth cap %p not mds%d ???\n", inode,
1966 cap, session->s_mds);
1967 }
1968 spin_unlock(&ci->i_ceph_lock);
1969 }
1970}
1971
1972void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1973 struct ceph_mds_session *session)
1974{
1975 struct ceph_inode_info *ci;
1976
1977 kick_flushing_capsnaps(mdsc, session);
1978
1979 dout("kick_flushing_caps mds%d\n", session->s_mds);
1980 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1981 struct inode *inode = &ci->vfs_inode;
1982 struct ceph_cap *cap;
1983 int delayed = 0;
1984
1985 spin_lock(&ci->i_ceph_lock);
1986 cap = ci->i_auth_cap;
1987 if (cap && cap->session == session) {
1988 dout("kick_flushing_caps %p cap %p %s\n", inode,
1989 cap, ceph_cap_string(ci->i_flushing_caps));
1990 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1991 __ceph_caps_used(ci),
1992 __ceph_caps_wanted(ci),
1993 cap->issued | cap->implemented,
1994 ci->i_flushing_caps, NULL);
1995 if (delayed) {
1996 spin_lock(&ci->i_ceph_lock);
1997 __cap_delay_requeue(mdsc, ci);
1998 spin_unlock(&ci->i_ceph_lock);
1999 }
2000 } else {
2001 pr_err("%p auth cap %p not mds%d ???\n", inode,
2002 cap, session->s_mds);
2003 spin_unlock(&ci->i_ceph_lock);
2004 }
2005 }
2006}
2007
2008static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2009 struct ceph_mds_session *session,
2010 struct inode *inode)
2011{
2012 struct ceph_inode_info *ci = ceph_inode(inode);
2013 struct ceph_cap *cap;
2014 int delayed = 0;
2015
2016 spin_lock(&ci->i_ceph_lock);
2017 cap = ci->i_auth_cap;
2018 dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode,
2019 ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq);
2020
2021 __ceph_flush_snaps(ci, &session, 1);
2022
2023 if (ci->i_flushing_caps) {
2024 spin_lock(&mdsc->cap_dirty_lock);
2025 list_move_tail(&ci->i_flushing_item,
2026 &cap->session->s_cap_flushing);
2027 spin_unlock(&mdsc->cap_dirty_lock);
2028
2029 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2030 __ceph_caps_used(ci),
2031 __ceph_caps_wanted(ci),
2032 cap->issued | cap->implemented,
2033 ci->i_flushing_caps, NULL);
2034 if (delayed) {
2035 spin_lock(&ci->i_ceph_lock);
2036 __cap_delay_requeue(mdsc, ci);
2037 spin_unlock(&ci->i_ceph_lock);
2038 }
2039 } else {
2040 spin_unlock(&ci->i_ceph_lock);
2041 }
2042}
2043
2044
2045/*
2046 * Take references to capabilities we hold, so that we don't release
2047 * them to the MDS prematurely.
2048 *
2049 * Protected by i_ceph_lock.
2050 */
2051static void __take_cap_refs(struct ceph_inode_info *ci, int got)
2052{
2053 if (got & CEPH_CAP_PIN)
2054 ci->i_pin_ref++;
2055 if (got & CEPH_CAP_FILE_RD)
2056 ci->i_rd_ref++;
2057 if (got & CEPH_CAP_FILE_CACHE)
2058 ci->i_rdcache_ref++;
2059 if (got & CEPH_CAP_FILE_WR)
2060 ci->i_wr_ref++;
2061 if (got & CEPH_CAP_FILE_BUFFER) {
2062 if (ci->i_wb_ref == 0)
2063 ihold(&ci->vfs_inode);
2064 ci->i_wb_ref++;
2065 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2066 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2067 }
2068}
2069
2070/*
2071 * Try to grab cap references. Specify those refs we @want, and the
2072 * minimal set we @need. Also include the larger offset we are writing
2073 * to (when applicable), and check against max_size here as well.
2074 * Note that caller is responsible for ensuring max_size increases are
2075 * requested from the MDS.
2076 */
2077static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2078 int *got, loff_t endoff, int *check_max, int *err)
2079{
2080 struct inode *inode = &ci->vfs_inode;
2081 int ret = 0;
2082 int have, implemented;
2083 int file_wanted;
2084
2085 dout("get_cap_refs %p need %s want %s\n", inode,
2086 ceph_cap_string(need), ceph_cap_string(want));
2087 spin_lock(&ci->i_ceph_lock);
2088
2089 /* make sure file is actually open */
2090 file_wanted = __ceph_caps_file_wanted(ci);
2091 if ((file_wanted & need) == 0) {
2092 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2093 ceph_cap_string(need), ceph_cap_string(file_wanted));
2094 *err = -EBADF;
2095 ret = 1;
2096 goto out;
2097 }
2098
2099 /* finish pending truncate */
2100 while (ci->i_truncate_pending) {
2101 spin_unlock(&ci->i_ceph_lock);
2102 __ceph_do_pending_vmtruncate(inode);
2103 spin_lock(&ci->i_ceph_lock);
2104 }
2105
2106 have = __ceph_caps_issued(ci, &implemented);
2107
2108 if (have & need & CEPH_CAP_FILE_WR) {
2109 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2110 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2111 inode, endoff, ci->i_max_size);
2112 if (endoff > ci->i_requested_max_size) {
2113 *check_max = 1;
2114 ret = 1;
2115 }
2116 goto out;
2117 }
2118 /*
2119 * If a sync write is in progress, we must wait, so that we
2120 * can get a final snapshot value for size+mtime.
2121 */
2122 if (__ceph_have_pending_cap_snap(ci)) {
2123 dout("get_cap_refs %p cap_snap_pending\n", inode);
2124 goto out;
2125 }
2126 }
2127
2128 if ((have & need) == need) {
2129 /*
2130 * Look at (implemented & ~have & not) so that we keep waiting
2131 * on transition from wanted -> needed caps. This is needed
2132 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2133 * going before a prior buffered writeback happens.
2134 */
2135 int not = want & ~(have & need);
2136 int revoking = implemented & ~have;
2137 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2138 inode, ceph_cap_string(have), ceph_cap_string(not),
2139 ceph_cap_string(revoking));
2140 if ((revoking & not) == 0) {
2141 *got = need | (have & want);
2142 __take_cap_refs(ci, *got);
2143 ret = 1;
2144 }
2145 } else {
2146 dout("get_cap_refs %p have %s needed %s\n", inode,
2147 ceph_cap_string(have), ceph_cap_string(need));
2148 }
2149out:
2150 spin_unlock(&ci->i_ceph_lock);
2151 dout("get_cap_refs %p ret %d got %s\n", inode,
2152 ret, ceph_cap_string(*got));
2153 return ret;
2154}
2155
2156/*
2157 * Check the offset we are writing up to against our current
2158 * max_size. If necessary, tell the MDS we want to write to
2159 * a larger offset.
2160 */
2161static void check_max_size(struct inode *inode, loff_t endoff)
2162{
2163 struct ceph_inode_info *ci = ceph_inode(inode);
2164 int check = 0;
2165
2166 /* do we need to explicitly request a larger max_size? */
2167 spin_lock(&ci->i_ceph_lock);
2168 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2169 dout("write %p at large endoff %llu, req max_size\n",
2170 inode, endoff);
2171 ci->i_wanted_max_size = endoff;
2172 }
2173 /* duplicate ceph_check_caps()'s logic */
2174 if (ci->i_auth_cap &&
2175 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2176 ci->i_wanted_max_size > ci->i_max_size &&
2177 ci->i_wanted_max_size > ci->i_requested_max_size)
2178 check = 1;
2179 spin_unlock(&ci->i_ceph_lock);
2180 if (check)
2181 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2182}
2183
2184/*
2185 * Wait for caps, and take cap references. If we can't get a WR cap
2186 * due to a small max_size, make sure we check_max_size (and possibly
2187 * ask the mds) so we don't get hung up indefinitely.
2188 */
2189int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
2190 loff_t endoff)
2191{
2192 int check_max, ret, err;
2193
2194retry:
2195 if (endoff > 0)
2196 check_max_size(&ci->vfs_inode, endoff);
2197 check_max = 0;
2198 err = 0;
2199 ret = wait_event_interruptible(ci->i_cap_wq,
2200 try_get_cap_refs(ci, need, want,
2201 got, endoff,
2202 &check_max, &err));
2203 if (err)
2204 ret = err;
2205 if (check_max)
2206 goto retry;
2207 return ret;
2208}
2209
2210/*
2211 * Take cap refs. Caller must already know we hold at least one ref
2212 * on the caps in question or we don't know this is safe.
2213 */
2214void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2215{
2216 spin_lock(&ci->i_ceph_lock);
2217 __take_cap_refs(ci, caps);
2218 spin_unlock(&ci->i_ceph_lock);
2219}
2220
2221/*
2222 * Release cap refs.
2223 *
2224 * If we released the last ref on any given cap, call ceph_check_caps
2225 * to release (or schedule a release).
2226 *
2227 * If we are releasing a WR cap (from a sync write), finalize any affected
2228 * cap_snap, and wake up any waiters.
2229 */
2230void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2231{
2232 struct inode *inode = &ci->vfs_inode;
2233 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2234 struct ceph_cap_snap *capsnap;
2235
2236 spin_lock(&ci->i_ceph_lock);
2237 if (had & CEPH_CAP_PIN)
2238 --ci->i_pin_ref;
2239 if (had & CEPH_CAP_FILE_RD)
2240 if (--ci->i_rd_ref == 0)
2241 last++;
2242 if (had & CEPH_CAP_FILE_CACHE)
2243 if (--ci->i_rdcache_ref == 0)
2244 last++;
2245 if (had & CEPH_CAP_FILE_BUFFER) {
2246 if (--ci->i_wb_ref == 0) {
2247 last++;
2248 put++;
2249 }
2250 dout("put_cap_refs %p wb %d -> %d (?)\n",
2251 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2252 }
2253 if (had & CEPH_CAP_FILE_WR)
2254 if (--ci->i_wr_ref == 0) {
2255 last++;
2256 if (!list_empty(&ci->i_cap_snaps)) {
2257 capsnap = list_first_entry(&ci->i_cap_snaps,
2258 struct ceph_cap_snap,
2259 ci_item);
2260 if (capsnap->writing) {
2261 capsnap->writing = 0;
2262 flushsnaps =
2263 __ceph_finish_cap_snap(ci,
2264 capsnap);
2265 wake = 1;
2266 }
2267 }
2268 }
2269 spin_unlock(&ci->i_ceph_lock);
2270
2271 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2272 last ? " last" : "", put ? " put" : "");
2273
2274 if (last && !flushsnaps)
2275 ceph_check_caps(ci, 0, NULL);
2276 else if (flushsnaps)
2277 ceph_flush_snaps(ci);
2278 if (wake)
2279 wake_up_all(&ci->i_cap_wq);
2280 if (put)
2281 iput(inode);
2282}
2283
2284/*
2285 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2286 * context. Adjust per-snap dirty page accounting as appropriate.
2287 * Once all dirty data for a cap_snap is flushed, flush snapped file
2288 * metadata back to the MDS. If we dropped the last ref, call
2289 * ceph_check_caps.
2290 */
2291void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2292 struct ceph_snap_context *snapc)
2293{
2294 struct inode *inode = &ci->vfs_inode;
2295 int last = 0;
2296 int complete_capsnap = 0;
2297 int drop_capsnap = 0;
2298 int found = 0;
2299 struct ceph_cap_snap *capsnap = NULL;
2300
2301 spin_lock(&ci->i_ceph_lock);
2302 ci->i_wrbuffer_ref -= nr;
2303 last = !ci->i_wrbuffer_ref;
2304
2305 if (ci->i_head_snapc == snapc) {
2306 ci->i_wrbuffer_ref_head -= nr;
2307 if (ci->i_wrbuffer_ref_head == 0 &&
2308 ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) {
2309 BUG_ON(!ci->i_head_snapc);
2310 ceph_put_snap_context(ci->i_head_snapc);
2311 ci->i_head_snapc = NULL;
2312 }
2313 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2314 inode,
2315 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2316 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2317 last ? " LAST" : "");
2318 } else {
2319 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2320 if (capsnap->context == snapc) {
2321 found = 1;
2322 break;
2323 }
2324 }
2325 BUG_ON(!found);
2326 capsnap->dirty_pages -= nr;
2327 if (capsnap->dirty_pages == 0) {
2328 complete_capsnap = 1;
2329 if (capsnap->dirty == 0)
2330 /* cap writeback completed before we created
2331 * the cap_snap; no FLUSHSNAP is needed */
2332 drop_capsnap = 1;
2333 }
2334 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2335 " snap %lld %d/%d -> %d/%d %s%s%s\n",
2336 inode, capsnap, capsnap->context->seq,
2337 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2338 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2339 last ? " (wrbuffer last)" : "",
2340 complete_capsnap ? " (complete capsnap)" : "",
2341 drop_capsnap ? " (drop capsnap)" : "");
2342 if (drop_capsnap) {
2343 ceph_put_snap_context(capsnap->context);
2344 list_del(&capsnap->ci_item);
2345 list_del(&capsnap->flushing_item);
2346 ceph_put_cap_snap(capsnap);
2347 }
2348 }
2349
2350 spin_unlock(&ci->i_ceph_lock);
2351
2352 if (last) {
2353 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2354 iput(inode);
2355 } else if (complete_capsnap) {
2356 ceph_flush_snaps(ci);
2357 wake_up_all(&ci->i_cap_wq);
2358 }
2359 if (drop_capsnap)
2360 iput(inode);
2361}
2362
2363/*
2364 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2365 */
2366static void invalidate_aliases(struct inode *inode)
2367{
2368 struct dentry *dn, *prev = NULL;
2369
2370 dout("invalidate_aliases inode %p\n", inode);
2371 d_prune_aliases(inode);
2372 /*
2373 * For non-directory inode, d_find_alias() only returns
2374 * hashed dentry. After calling d_invalidate(), the
2375 * dentry becomes unhashed.
2376 *
2377 * For directory inode, d_find_alias() can return
2378 * unhashed dentry. But directory inode should have
2379 * one alias at most.
2380 */
2381 while ((dn = d_find_alias(inode))) {
2382 if (dn == prev) {
2383 dput(dn);
2384 break;
2385 }
2386 d_invalidate(dn);
2387 if (prev)
2388 dput(prev);
2389 prev = dn;
2390 }
2391 if (prev)
2392 dput(prev);
2393}
2394
2395/*
2396 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2397 * actually be a revocation if it specifies a smaller cap set.)
2398 *
2399 * caller holds s_mutex and i_ceph_lock, we drop both.
2400 *
2401 * return value:
2402 * 0 - ok
2403 * 1 - check_caps on auth cap only (writeback)
2404 * 2 - check_caps (ack revoke)
2405 */
2406static void handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2407 struct ceph_mds_session *session,
2408 struct ceph_cap *cap,
2409 struct ceph_buffer *xattr_buf)
2410 __releases(ci->i_ceph_lock)
2411{
2412 struct ceph_inode_info *ci = ceph_inode(inode);
2413 int mds = session->s_mds;
2414 int seq = le32_to_cpu(grant->seq);
2415 int newcaps = le32_to_cpu(grant->caps);
2416 int issued, implemented, used, wanted, dirty;
2417 u64 size = le64_to_cpu(grant->size);
2418 u64 max_size = le64_to_cpu(grant->max_size);
2419 struct timespec mtime, atime, ctime;
2420 int check_caps = 0;
2421 int wake = 0;
2422 int writeback = 0;
2423 int queue_invalidate = 0;
2424 int deleted_inode = 0;
2425 int queue_revalidate = 0;
2426
2427 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2428 inode, cap, mds, seq, ceph_cap_string(newcaps));
2429 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2430 inode->i_size);
2431
2432
2433 /*
2434 * auth mds of the inode changed. we received the cap export message,
2435 * but still haven't received the cap import message. handle_cap_export
2436 * updated the new auth MDS' cap.
2437 *
2438 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2439 * that was sent before the cap import message. So don't remove caps.
2440 */
2441 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2442 WARN_ON(cap != ci->i_auth_cap);
2443 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2444 seq = cap->seq;
2445 newcaps |= cap->issued;
2446 }
2447
2448 /*
2449 * If CACHE is being revoked, and we have no dirty buffers,
2450 * try to invalidate (once). (If there are dirty buffers, we
2451 * will invalidate _after_ writeback.)
2452 */
2453 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2454 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2455 !ci->i_wrbuffer_ref) {
2456 if (try_nonblocking_invalidate(inode)) {
2457 /* there were locked pages.. invalidate later
2458 in a separate thread. */
2459 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2460 queue_invalidate = 1;
2461 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2462 }
2463 }
2464
2465 ceph_fscache_invalidate(inode);
2466 }
2467
2468 /* side effects now are allowed */
2469
2470 issued = __ceph_caps_issued(ci, &implemented);
2471 issued |= implemented | __ceph_caps_dirty(ci);
2472
2473 cap->cap_gen = session->s_cap_gen;
2474 cap->seq = seq;
2475
2476 __check_cap_issue(ci, cap, newcaps);
2477
2478 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2479 inode->i_mode = le32_to_cpu(grant->mode);
2480 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2481 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2482 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2483 from_kuid(&init_user_ns, inode->i_uid),
2484 from_kgid(&init_user_ns, inode->i_gid));
2485 }
2486
2487 if ((issued & CEPH_CAP_LINK_EXCL) == 0) {
2488 set_nlink(inode, le32_to_cpu(grant->nlink));
2489 if (inode->i_nlink == 0 &&
2490 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2491 deleted_inode = 1;
2492 }
2493
2494 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2495 int len = le32_to_cpu(grant->xattr_len);
2496 u64 version = le64_to_cpu(grant->xattr_version);
2497
2498 if (version > ci->i_xattrs.version) {
2499 dout(" got new xattrs v%llu on %p len %d\n",
2500 version, inode, len);
2501 if (ci->i_xattrs.blob)
2502 ceph_buffer_put(ci->i_xattrs.blob);
2503 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2504 ci->i_xattrs.version = version;
2505 ceph_forget_all_cached_acls(inode);
2506 }
2507 }
2508
2509 /* Do we need to revalidate our fscache cookie. Don't bother on the
2510 * first cache cap as we already validate at cookie creation time. */
2511 if ((issued & CEPH_CAP_FILE_CACHE) && ci->i_rdcache_gen > 1)
2512 queue_revalidate = 1;
2513
2514 /* size/ctime/mtime/atime? */
2515 ceph_fill_file_size(inode, issued,
2516 le32_to_cpu(grant->truncate_seq),
2517 le64_to_cpu(grant->truncate_size), size);
2518 ceph_decode_timespec(&mtime, &grant->mtime);
2519 ceph_decode_timespec(&atime, &grant->atime);
2520 ceph_decode_timespec(&ctime, &grant->ctime);
2521 ceph_fill_file_time(inode, issued,
2522 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2523 &atime);
2524
2525
2526 /* file layout may have changed */
2527 ci->i_layout = grant->layout;
2528
2529 /* max size increase? */
2530 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2531 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2532 ci->i_max_size = max_size;
2533 if (max_size >= ci->i_wanted_max_size) {
2534 ci->i_wanted_max_size = 0; /* reset */
2535 ci->i_requested_max_size = 0;
2536 }
2537 wake = 1;
2538 }
2539
2540 /* check cap bits */
2541 wanted = __ceph_caps_wanted(ci);
2542 used = __ceph_caps_used(ci);
2543 dirty = __ceph_caps_dirty(ci);
2544 dout(" my wanted = %s, used = %s, dirty %s\n",
2545 ceph_cap_string(wanted),
2546 ceph_cap_string(used),
2547 ceph_cap_string(dirty));
2548 if (wanted != le32_to_cpu(grant->wanted)) {
2549 dout("mds wanted %s -> %s\n",
2550 ceph_cap_string(le32_to_cpu(grant->wanted)),
2551 ceph_cap_string(wanted));
2552 /* imported cap may not have correct mds_wanted */
2553 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2554 check_caps = 1;
2555 }
2556
2557 /* revocation, grant, or no-op? */
2558 if (cap->issued & ~newcaps) {
2559 int revoking = cap->issued & ~newcaps;
2560
2561 dout("revocation: %s -> %s (revoking %s)\n",
2562 ceph_cap_string(cap->issued),
2563 ceph_cap_string(newcaps),
2564 ceph_cap_string(revoking));
2565 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2566 writeback = 1; /* initiate writeback; will delay ack */
2567 else if (revoking == CEPH_CAP_FILE_CACHE &&
2568 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2569 queue_invalidate)
2570 ; /* do nothing yet, invalidation will be queued */
2571 else if (cap == ci->i_auth_cap)
2572 check_caps = 1; /* check auth cap only */
2573 else
2574 check_caps = 2; /* check all caps */
2575 cap->issued = newcaps;
2576 cap->implemented |= newcaps;
2577 } else if (cap->issued == newcaps) {
2578 dout("caps unchanged: %s -> %s\n",
2579 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2580 } else {
2581 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2582 ceph_cap_string(newcaps));
2583 /* non-auth MDS is revoking the newly grant caps ? */
2584 if (cap == ci->i_auth_cap &&
2585 __ceph_caps_revoking_other(ci, cap, newcaps))
2586 check_caps = 2;
2587
2588 cap->issued = newcaps;
2589 cap->implemented |= newcaps; /* add bits only, to
2590 * avoid stepping on a
2591 * pending revocation */
2592 wake = 1;
2593 }
2594 BUG_ON(cap->issued & ~cap->implemented);
2595
2596 spin_unlock(&ci->i_ceph_lock);
2597
2598 if (writeback)
2599 /*
2600 * queue inode for writeback: we can't actually call
2601 * filemap_write_and_wait, etc. from message handler
2602 * context.
2603 */
2604 ceph_queue_writeback(inode);
2605 if (queue_invalidate)
2606 ceph_queue_invalidate(inode);
2607 if (deleted_inode)
2608 invalidate_aliases(inode);
2609 if (queue_revalidate)
2610 ceph_queue_revalidate(inode);
2611 if (wake)
2612 wake_up_all(&ci->i_cap_wq);
2613
2614 if (check_caps == 1)
2615 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2616 session);
2617 else if (check_caps == 2)
2618 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2619 else
2620 mutex_unlock(&session->s_mutex);
2621}
2622
2623/*
2624 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2625 * MDS has been safely committed.
2626 */
2627static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2628 struct ceph_mds_caps *m,
2629 struct ceph_mds_session *session,
2630 struct ceph_cap *cap)
2631 __releases(ci->i_ceph_lock)
2632{
2633 struct ceph_inode_info *ci = ceph_inode(inode);
2634 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2635 unsigned seq = le32_to_cpu(m->seq);
2636 int dirty = le32_to_cpu(m->dirty);
2637 int cleaned = 0;
2638 int drop = 0;
2639 int i;
2640
2641 for (i = 0; i < CEPH_CAP_BITS; i++)
2642 if ((dirty & (1 << i)) &&
2643 flush_tid == ci->i_cap_flush_tid[i])
2644 cleaned |= 1 << i;
2645
2646 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2647 " flushing %s -> %s\n",
2648 inode, session->s_mds, seq, ceph_cap_string(dirty),
2649 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2650 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2651
2652 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2653 goto out;
2654
2655 ci->i_flushing_caps &= ~cleaned;
2656
2657 spin_lock(&mdsc->cap_dirty_lock);
2658 if (ci->i_flushing_caps == 0) {
2659 list_del_init(&ci->i_flushing_item);
2660 if (!list_empty(&session->s_cap_flushing))
2661 dout(" mds%d still flushing cap on %p\n",
2662 session->s_mds,
2663 &list_entry(session->s_cap_flushing.next,
2664 struct ceph_inode_info,
2665 i_flushing_item)->vfs_inode);
2666 mdsc->num_cap_flushing--;
2667 wake_up_all(&mdsc->cap_flushing_wq);
2668 dout(" inode %p now !flushing\n", inode);
2669
2670 if (ci->i_dirty_caps == 0) {
2671 dout(" inode %p now clean\n", inode);
2672 BUG_ON(!list_empty(&ci->i_dirty_item));
2673 drop = 1;
2674 if (ci->i_wrbuffer_ref_head == 0) {
2675 BUG_ON(!ci->i_head_snapc);
2676 ceph_put_snap_context(ci->i_head_snapc);
2677 ci->i_head_snapc = NULL;
2678 }
2679 } else {
2680 BUG_ON(list_empty(&ci->i_dirty_item));
2681 }
2682 }
2683 spin_unlock(&mdsc->cap_dirty_lock);
2684 wake_up_all(&ci->i_cap_wq);
2685
2686out:
2687 spin_unlock(&ci->i_ceph_lock);
2688 if (drop)
2689 iput(inode);
2690}
2691
2692/*
2693 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2694 * throw away our cap_snap.
2695 *
2696 * Caller hold s_mutex.
2697 */
2698static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2699 struct ceph_mds_caps *m,
2700 struct ceph_mds_session *session)
2701{
2702 struct ceph_inode_info *ci = ceph_inode(inode);
2703 u64 follows = le64_to_cpu(m->snap_follows);
2704 struct ceph_cap_snap *capsnap;
2705 int drop = 0;
2706
2707 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2708 inode, ci, session->s_mds, follows);
2709
2710 spin_lock(&ci->i_ceph_lock);
2711 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2712 if (capsnap->follows == follows) {
2713 if (capsnap->flush_tid != flush_tid) {
2714 dout(" cap_snap %p follows %lld tid %lld !="
2715 " %lld\n", capsnap, follows,
2716 flush_tid, capsnap->flush_tid);
2717 break;
2718 }
2719 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2720 dout(" removing %p cap_snap %p follows %lld\n",
2721 inode, capsnap, follows);
2722 ceph_put_snap_context(capsnap->context);
2723 list_del(&capsnap->ci_item);
2724 list_del(&capsnap->flushing_item);
2725 ceph_put_cap_snap(capsnap);
2726 drop = 1;
2727 break;
2728 } else {
2729 dout(" skipping cap_snap %p follows %lld\n",
2730 capsnap, capsnap->follows);
2731 }
2732 }
2733 spin_unlock(&ci->i_ceph_lock);
2734 if (drop)
2735 iput(inode);
2736}
2737
2738/*
2739 * Handle TRUNC from MDS, indicating file truncation.
2740 *
2741 * caller hold s_mutex.
2742 */
2743static void handle_cap_trunc(struct inode *inode,
2744 struct ceph_mds_caps *trunc,
2745 struct ceph_mds_session *session)
2746 __releases(ci->i_ceph_lock)
2747{
2748 struct ceph_inode_info *ci = ceph_inode(inode);
2749 int mds = session->s_mds;
2750 int seq = le32_to_cpu(trunc->seq);
2751 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2752 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2753 u64 size = le64_to_cpu(trunc->size);
2754 int implemented = 0;
2755 int dirty = __ceph_caps_dirty(ci);
2756 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2757 int queue_trunc = 0;
2758
2759 issued |= implemented | dirty;
2760
2761 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2762 inode, mds, seq, truncate_size, truncate_seq);
2763 queue_trunc = ceph_fill_file_size(inode, issued,
2764 truncate_seq, truncate_size, size);
2765 spin_unlock(&ci->i_ceph_lock);
2766
2767 if (queue_trunc) {
2768 ceph_queue_vmtruncate(inode);
2769 ceph_fscache_invalidate(inode);
2770 }
2771}
2772
2773/*
2774 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2775 * different one. If we are the most recent migration we've seen (as
2776 * indicated by mseq), make note of the migrating cap bits for the
2777 * duration (until we see the corresponding IMPORT).
2778 *
2779 * caller holds s_mutex
2780 */
2781static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2782 struct ceph_mds_cap_peer *ph,
2783 struct ceph_mds_session *session)
2784{
2785 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2786 struct ceph_mds_session *tsession = NULL;
2787 struct ceph_cap *cap, *tcap;
2788 struct ceph_inode_info *ci = ceph_inode(inode);
2789 u64 t_cap_id;
2790 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2791 unsigned t_seq, t_mseq;
2792 int target, issued;
2793 int mds = session->s_mds;
2794
2795 if (ph) {
2796 t_cap_id = le64_to_cpu(ph->cap_id);
2797 t_seq = le32_to_cpu(ph->seq);
2798 t_mseq = le32_to_cpu(ph->mseq);
2799 target = le32_to_cpu(ph->mds);
2800 } else {
2801 t_cap_id = t_seq = t_mseq = 0;
2802 target = -1;
2803 }
2804
2805 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
2806 inode, ci, mds, mseq, target);
2807retry:
2808 spin_lock(&ci->i_ceph_lock);
2809 cap = __get_cap_for_mds(ci, mds);
2810 if (!cap)
2811 goto out_unlock;
2812
2813 if (target < 0) {
2814 __ceph_remove_cap(cap, false);
2815 goto out_unlock;
2816 }
2817
2818 /*
2819 * now we know we haven't received the cap import message yet
2820 * because the exported cap still exist.
2821 */
2822
2823 issued = cap->issued;
2824 WARN_ON(issued != cap->implemented);
2825
2826 tcap = __get_cap_for_mds(ci, target);
2827 if (tcap) {
2828 /* already have caps from the target */
2829 if (tcap->cap_id != t_cap_id ||
2830 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
2831 dout(" updating import cap %p mds%d\n", tcap, target);
2832 tcap->cap_id = t_cap_id;
2833 tcap->seq = t_seq - 1;
2834 tcap->issue_seq = t_seq - 1;
2835 tcap->mseq = t_mseq;
2836 tcap->issued |= issued;
2837 tcap->implemented |= issued;
2838 if (cap == ci->i_auth_cap)
2839 ci->i_auth_cap = tcap;
2840 if (ci->i_flushing_caps && ci->i_auth_cap == tcap) {
2841 spin_lock(&mdsc->cap_dirty_lock);
2842 list_move_tail(&ci->i_flushing_item,
2843 &tcap->session->s_cap_flushing);
2844 spin_unlock(&mdsc->cap_dirty_lock);
2845 }
2846 }
2847 __ceph_remove_cap(cap, false);
2848 goto out_unlock;
2849 }
2850
2851 if (tsession) {
2852 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
2853 spin_unlock(&ci->i_ceph_lock);
2854 /* add placeholder for the export tagert */
2855 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
2856 t_seq - 1, t_mseq, (u64)-1, flag, NULL);
2857 goto retry;
2858 }
2859
2860 spin_unlock(&ci->i_ceph_lock);
2861 mutex_unlock(&session->s_mutex);
2862
2863 /* open target session */
2864 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
2865 if (!IS_ERR(tsession)) {
2866 if (mds > target) {
2867 mutex_lock(&session->s_mutex);
2868 mutex_lock_nested(&tsession->s_mutex,
2869 SINGLE_DEPTH_NESTING);
2870 } else {
2871 mutex_lock(&tsession->s_mutex);
2872 mutex_lock_nested(&session->s_mutex,
2873 SINGLE_DEPTH_NESTING);
2874 }
2875 ceph_add_cap_releases(mdsc, tsession);
2876 } else {
2877 WARN_ON(1);
2878 tsession = NULL;
2879 target = -1;
2880 }
2881 goto retry;
2882
2883out_unlock:
2884 spin_unlock(&ci->i_ceph_lock);
2885 mutex_unlock(&session->s_mutex);
2886 if (tsession) {
2887 mutex_unlock(&tsession->s_mutex);
2888 ceph_put_mds_session(tsession);
2889 }
2890}
2891
2892/*
2893 * Handle cap IMPORT. If there are temp bits from an older EXPORT,
2894 * clean them up.
2895 *
2896 * caller holds s_mutex.
2897 */
2898static void handle_cap_import(struct ceph_mds_client *mdsc,
2899 struct inode *inode, struct ceph_mds_caps *im,
2900 struct ceph_mds_cap_peer *ph,
2901 struct ceph_mds_session *session,
2902 void *snaptrace, int snaptrace_len)
2903{
2904 struct ceph_inode_info *ci = ceph_inode(inode);
2905 struct ceph_cap *cap;
2906 int mds = session->s_mds;
2907 unsigned issued = le32_to_cpu(im->caps);
2908 unsigned wanted = le32_to_cpu(im->wanted);
2909 unsigned seq = le32_to_cpu(im->seq);
2910 unsigned mseq = le32_to_cpu(im->migrate_seq);
2911 u64 realmino = le64_to_cpu(im->realm);
2912 u64 cap_id = le64_to_cpu(im->cap_id);
2913 u64 p_cap_id;
2914 int peer;
2915
2916 if (ph) {
2917 p_cap_id = le64_to_cpu(ph->cap_id);
2918 peer = le32_to_cpu(ph->mds);
2919 } else {
2920 p_cap_id = 0;
2921 peer = -1;
2922 }
2923
2924 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
2925 inode, ci, mds, mseq, peer);
2926
2927 spin_lock(&ci->i_ceph_lock);
2928 cap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
2929 if (cap && cap->cap_id == p_cap_id) {
2930 dout(" remove export cap %p mds%d flags %d\n",
2931 cap, peer, ph->flags);
2932 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
2933 (cap->seq != le32_to_cpu(ph->seq) ||
2934 cap->mseq != le32_to_cpu(ph->mseq))) {
2935 pr_err("handle_cap_import: mismatched seq/mseq: "
2936 "ino (%llx.%llx) mds%d seq %d mseq %d "
2937 "importer mds%d has peer seq %d mseq %d\n",
2938 ceph_vinop(inode), peer, cap->seq,
2939 cap->mseq, mds, le32_to_cpu(ph->seq),
2940 le32_to_cpu(ph->mseq));
2941 }
2942 ci->i_cap_exporting_issued = cap->issued;
2943 __ceph_remove_cap(cap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
2944 }
2945
2946 /* make sure we re-request max_size, if necessary */
2947 ci->i_wanted_max_size = 0;
2948 ci->i_requested_max_size = 0;
2949 spin_unlock(&ci->i_ceph_lock);
2950
2951 down_write(&mdsc->snap_rwsem);
2952 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2953 false);
2954 downgrade_write(&mdsc->snap_rwsem);
2955 ceph_add_cap(inode, session, cap_id, -1,
2956 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2957 NULL /* no caps context */);
2958 kick_flushing_inode_caps(mdsc, session, inode);
2959 up_read(&mdsc->snap_rwsem);
2960
2961}
2962
2963/*
2964 * Handle a caps message from the MDS.
2965 *
2966 * Identify the appropriate session, inode, and call the right handler
2967 * based on the cap op.
2968 */
2969void ceph_handle_caps(struct ceph_mds_session *session,
2970 struct ceph_msg *msg)
2971{
2972 struct ceph_mds_client *mdsc = session->s_mdsc;
2973 struct super_block *sb = mdsc->fsc->sb;
2974 struct inode *inode;
2975 struct ceph_inode_info *ci;
2976 struct ceph_cap *cap;
2977 struct ceph_mds_caps *h;
2978 struct ceph_mds_cap_peer *peer = NULL;
2979 int mds = session->s_mds;
2980 int op;
2981 u32 seq, mseq;
2982 struct ceph_vino vino;
2983 u64 cap_id;
2984 u64 size, max_size;
2985 u64 tid;
2986 void *snaptrace;
2987 size_t snaptrace_len;
2988 void *flock;
2989 void *end;
2990 u32 flock_len;
2991
2992 dout("handle_caps from mds%d\n", mds);
2993
2994 /* decode */
2995 end = msg->front.iov_base + msg->front.iov_len;
2996 tid = le64_to_cpu(msg->hdr.tid);
2997 if (msg->front.iov_len < sizeof(*h))
2998 goto bad;
2999 h = msg->front.iov_base;
3000 op = le32_to_cpu(h->op);
3001 vino.ino = le64_to_cpu(h->ino);
3002 vino.snap = CEPH_NOSNAP;
3003 cap_id = le64_to_cpu(h->cap_id);
3004 seq = le32_to_cpu(h->seq);
3005 mseq = le32_to_cpu(h->migrate_seq);
3006 size = le64_to_cpu(h->size);
3007 max_size = le64_to_cpu(h->max_size);
3008
3009 snaptrace = h + 1;
3010 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3011
3012 if (le16_to_cpu(msg->hdr.version) >= 2) {
3013 void *p = snaptrace + snaptrace_len;
3014 ceph_decode_32_safe(&p, end, flock_len, bad);
3015 if (p + flock_len > end)
3016 goto bad;
3017 flock = p;
3018 } else {
3019 flock = NULL;
3020 flock_len = 0;
3021 }
3022
3023 if (le16_to_cpu(msg->hdr.version) >= 3) {
3024 if (op == CEPH_CAP_OP_IMPORT) {
3025 void *p = flock + flock_len;
3026 if (p + sizeof(*peer) > end)
3027 goto bad;
3028 peer = p;
3029 } else if (op == CEPH_CAP_OP_EXPORT) {
3030 /* recorded in unused fields */
3031 peer = (void *)&h->size;
3032 }
3033 }
3034
3035 mutex_lock(&session->s_mutex);
3036 session->s_seq++;
3037 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3038 (unsigned)seq);
3039
3040 if (op == CEPH_CAP_OP_IMPORT)
3041 ceph_add_cap_releases(mdsc, session);
3042
3043 /* lookup ino */
3044 inode = ceph_find_inode(sb, vino);
3045 ci = ceph_inode(inode);
3046 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3047 vino.snap, inode);
3048 if (!inode) {
3049 dout(" i don't have ino %llx\n", vino.ino);
3050
3051 if (op == CEPH_CAP_OP_IMPORT) {
3052 spin_lock(&session->s_cap_lock);
3053 __queue_cap_release(session, vino.ino, cap_id,
3054 mseq, seq);
3055 spin_unlock(&session->s_cap_lock);
3056 }
3057 goto flush_cap_releases;
3058 }
3059
3060 /* these will work even if we don't have a cap yet */
3061 switch (op) {
3062 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3063 handle_cap_flushsnap_ack(inode, tid, h, session);
3064 goto done;
3065
3066 case CEPH_CAP_OP_EXPORT:
3067 handle_cap_export(inode, h, peer, session);
3068 goto done_unlocked;
3069
3070 case CEPH_CAP_OP_IMPORT:
3071 handle_cap_import(mdsc, inode, h, peer, session,
3072 snaptrace, snaptrace_len);
3073 }
3074
3075 /* the rest require a cap */
3076 spin_lock(&ci->i_ceph_lock);
3077 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3078 if (!cap) {
3079 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3080 inode, ceph_ino(inode), ceph_snap(inode), mds);
3081 spin_unlock(&ci->i_ceph_lock);
3082 goto flush_cap_releases;
3083 }
3084
3085 /* note that each of these drops i_ceph_lock for us */
3086 switch (op) {
3087 case CEPH_CAP_OP_REVOKE:
3088 case CEPH_CAP_OP_GRANT:
3089 case CEPH_CAP_OP_IMPORT:
3090 handle_cap_grant(inode, h, session, cap, msg->middle);
3091 goto done_unlocked;
3092
3093 case CEPH_CAP_OP_FLUSH_ACK:
3094 handle_cap_flush_ack(inode, tid, h, session, cap);
3095 break;
3096
3097 case CEPH_CAP_OP_TRUNC:
3098 handle_cap_trunc(inode, h, session);
3099 break;
3100
3101 default:
3102 spin_unlock(&ci->i_ceph_lock);
3103 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3104 ceph_cap_op_name(op));
3105 }
3106
3107 goto done;
3108
3109flush_cap_releases:
3110 /*
3111 * send any full release message to try to move things
3112 * along for the mds (who clearly thinks we still have this
3113 * cap).
3114 */
3115 ceph_add_cap_releases(mdsc, session);
3116 ceph_send_cap_releases(mdsc, session);
3117
3118done:
3119 mutex_unlock(&session->s_mutex);
3120done_unlocked:
3121 if (inode)
3122 iput(inode);
3123 return;
3124
3125bad:
3126 pr_err("ceph_handle_caps: corrupt message\n");
3127 ceph_msg_dump(msg);
3128 return;
3129}
3130
3131/*
3132 * Delayed work handler to process end of delayed cap release LRU list.
3133 */
3134void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3135{
3136 struct ceph_inode_info *ci;
3137 int flags = CHECK_CAPS_NODELAY;
3138
3139 dout("check_delayed_caps\n");
3140 while (1) {
3141 spin_lock(&mdsc->cap_delay_lock);
3142 if (list_empty(&mdsc->cap_delay_list))
3143 break;
3144 ci = list_first_entry(&mdsc->cap_delay_list,
3145 struct ceph_inode_info,
3146 i_cap_delay_list);
3147 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3148 time_before(jiffies, ci->i_hold_caps_max))
3149 break;
3150 list_del_init(&ci->i_cap_delay_list);
3151 spin_unlock(&mdsc->cap_delay_lock);
3152 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3153 ceph_check_caps(ci, flags, NULL);
3154 }
3155 spin_unlock(&mdsc->cap_delay_lock);
3156}
3157
3158/*
3159 * Flush all dirty caps to the mds
3160 */
3161void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3162{
3163 struct ceph_inode_info *ci;
3164 struct inode *inode;
3165
3166 dout("flush_dirty_caps\n");
3167 spin_lock(&mdsc->cap_dirty_lock);
3168 while (!list_empty(&mdsc->cap_dirty)) {
3169 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3170 i_dirty_item);
3171 inode = &ci->vfs_inode;
3172 ihold(inode);
3173 dout("flush_dirty_caps %p\n", inode);
3174 spin_unlock(&mdsc->cap_dirty_lock);
3175 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3176 iput(inode);
3177 spin_lock(&mdsc->cap_dirty_lock);
3178 }
3179 spin_unlock(&mdsc->cap_dirty_lock);
3180 dout("flush_dirty_caps done\n");
3181}
3182
3183/*
3184 * Drop open file reference. If we were the last open file,
3185 * we may need to release capabilities to the MDS (or schedule
3186 * their delayed release).
3187 */
3188void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3189{
3190 struct inode *inode = &ci->vfs_inode;
3191 int last = 0;
3192
3193 spin_lock(&ci->i_ceph_lock);
3194 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
3195 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
3196 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
3197 if (--ci->i_nr_by_mode[fmode] == 0)
3198 last++;
3199 spin_unlock(&ci->i_ceph_lock);
3200
3201 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3202 ceph_check_caps(ci, 0, NULL);
3203}
3204
3205/*
3206 * Helpers for embedding cap and dentry lease releases into mds
3207 * requests.
3208 *
3209 * @force is used by dentry_release (below) to force inclusion of a
3210 * record for the directory inode, even when there aren't any caps to
3211 * drop.
3212 */
3213int ceph_encode_inode_release(void **p, struct inode *inode,
3214 int mds, int drop, int unless, int force)
3215{
3216 struct ceph_inode_info *ci = ceph_inode(inode);
3217 struct ceph_cap *cap;
3218 struct ceph_mds_request_release *rel = *p;
3219 int used, dirty;
3220 int ret = 0;
3221
3222 spin_lock(&ci->i_ceph_lock);
3223 used = __ceph_caps_used(ci);
3224 dirty = __ceph_caps_dirty(ci);
3225
3226 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3227 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3228 ceph_cap_string(unless));
3229
3230 /* only drop unused, clean caps */
3231 drop &= ~(used | dirty);
3232
3233 cap = __get_cap_for_mds(ci, mds);
3234 if (cap && __cap_is_valid(cap)) {
3235 if (force ||
3236 ((cap->issued & drop) &&
3237 (cap->issued & unless) == 0)) {
3238 if ((cap->issued & drop) &&
3239 (cap->issued & unless) == 0) {
3240 int wanted = __ceph_caps_wanted(ci);
3241 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3242 wanted |= cap->mds_wanted;
3243 dout("encode_inode_release %p cap %p "
3244 "%s -> %s, wanted %s -> %s\n", inode, cap,
3245 ceph_cap_string(cap->issued),
3246 ceph_cap_string(cap->issued & ~drop),
3247 ceph_cap_string(cap->mds_wanted),
3248 ceph_cap_string(wanted));
3249
3250 cap->issued &= ~drop;
3251 cap->implemented &= ~drop;
3252 cap->mds_wanted = wanted;
3253 } else {
3254 dout("encode_inode_release %p cap %p %s"
3255 " (force)\n", inode, cap,
3256 ceph_cap_string(cap->issued));
3257 }
3258
3259 rel->ino = cpu_to_le64(ceph_ino(inode));
3260 rel->cap_id = cpu_to_le64(cap->cap_id);
3261 rel->seq = cpu_to_le32(cap->seq);
3262 rel->issue_seq = cpu_to_le32(cap->issue_seq),
3263 rel->mseq = cpu_to_le32(cap->mseq);
3264 rel->caps = cpu_to_le32(cap->implemented);
3265 rel->wanted = cpu_to_le32(cap->mds_wanted);
3266 rel->dname_len = 0;
3267 rel->dname_seq = 0;
3268 *p += sizeof(*rel);
3269 ret = 1;
3270 } else {
3271 dout("encode_inode_release %p cap %p %s\n",
3272 inode, cap, ceph_cap_string(cap->issued));
3273 }
3274 }
3275 spin_unlock(&ci->i_ceph_lock);
3276 return ret;
3277}
3278
3279int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3280 int mds, int drop, int unless)
3281{
3282 struct inode *dir = dentry->d_parent->d_inode;
3283 struct ceph_mds_request_release *rel = *p;
3284 struct ceph_dentry_info *di = ceph_dentry(dentry);
3285 int force = 0;
3286 int ret;
3287
3288 /*
3289 * force an record for the directory caps if we have a dentry lease.
3290 * this is racy (can't take i_ceph_lock and d_lock together), but it
3291 * doesn't have to be perfect; the mds will revoke anything we don't
3292 * release.
3293 */
3294 spin_lock(&dentry->d_lock);
3295 if (di->lease_session && di->lease_session->s_mds == mds)
3296 force = 1;
3297 spin_unlock(&dentry->d_lock);
3298
3299 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3300
3301 spin_lock(&dentry->d_lock);
3302 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3303 dout("encode_dentry_release %p mds%d seq %d\n",
3304 dentry, mds, (int)di->lease_seq);
3305 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3306 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3307 *p += dentry->d_name.len;
3308 rel->dname_seq = cpu_to_le32(di->lease_seq);
3309 __ceph_mdsc_drop_dentry_lease(dentry);
3310 }
3311 spin_unlock(&dentry->d_lock);
3312 return ret;
3313}