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// SPDX-License-Identifier: GPL-2.0
2#include <linux/ceph/ceph_debug.h>
3
4#include <linux/fs.h>
5#include <linux/kernel.h>
6#include <linux/sched/signal.h>
7#include <linux/slab.h>
8#include <linux/vmalloc.h>
9#include <linux/wait.h>
10#include <linux/writeback.h>
11#include <linux/iversion.h>
12
13#include "super.h"
14#include "mds_client.h"
15#include "cache.h"
16#include <linux/ceph/decode.h>
17#include <linux/ceph/messenger.h>
18
19/*
20 * Capability management
21 *
22 * The Ceph metadata servers control client access to inode metadata
23 * and file data by issuing capabilities, granting clients permission
24 * to read and/or write both inode field and file data to OSDs
25 * (storage nodes). Each capability consists of a set of bits
26 * indicating which operations are allowed.
27 *
28 * If the client holds a *_SHARED cap, the client has a coherent value
29 * that can be safely read from the cached inode.
30 *
31 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
32 * client is allowed to change inode attributes (e.g., file size,
33 * mtime), note its dirty state in the ceph_cap, and asynchronously
34 * flush that metadata change to the MDS.
35 *
36 * In the event of a conflicting operation (perhaps by another
37 * client), the MDS will revoke the conflicting client capabilities.
38 *
39 * In order for a client to cache an inode, it must hold a capability
40 * with at least one MDS server. When inodes are released, release
41 * notifications are batched and periodically sent en masse to the MDS
42 * cluster to release server state.
43 */
44
45static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
46static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
47 struct ceph_mds_session *session,
48 struct ceph_inode_info *ci,
49 u64 oldest_flush_tid);
50
51/*
52 * Generate readable cap strings for debugging output.
53 */
54#define MAX_CAP_STR 20
55static char cap_str[MAX_CAP_STR][40];
56static DEFINE_SPINLOCK(cap_str_lock);
57static int last_cap_str;
58
59static char *gcap_string(char *s, int c)
60{
61 if (c & CEPH_CAP_GSHARED)
62 *s++ = 's';
63 if (c & CEPH_CAP_GEXCL)
64 *s++ = 'x';
65 if (c & CEPH_CAP_GCACHE)
66 *s++ = 'c';
67 if (c & CEPH_CAP_GRD)
68 *s++ = 'r';
69 if (c & CEPH_CAP_GWR)
70 *s++ = 'w';
71 if (c & CEPH_CAP_GBUFFER)
72 *s++ = 'b';
73 if (c & CEPH_CAP_GWREXTEND)
74 *s++ = 'a';
75 if (c & CEPH_CAP_GLAZYIO)
76 *s++ = 'l';
77 return s;
78}
79
80const char *ceph_cap_string(int caps)
81{
82 int i;
83 char *s;
84 int c;
85
86 spin_lock(&cap_str_lock);
87 i = last_cap_str++;
88 if (last_cap_str == MAX_CAP_STR)
89 last_cap_str = 0;
90 spin_unlock(&cap_str_lock);
91
92 s = cap_str[i];
93
94 if (caps & CEPH_CAP_PIN)
95 *s++ = 'p';
96
97 c = (caps >> CEPH_CAP_SAUTH) & 3;
98 if (c) {
99 *s++ = 'A';
100 s = gcap_string(s, c);
101 }
102
103 c = (caps >> CEPH_CAP_SLINK) & 3;
104 if (c) {
105 *s++ = 'L';
106 s = gcap_string(s, c);
107 }
108
109 c = (caps >> CEPH_CAP_SXATTR) & 3;
110 if (c) {
111 *s++ = 'X';
112 s = gcap_string(s, c);
113 }
114
115 c = caps >> CEPH_CAP_SFILE;
116 if (c) {
117 *s++ = 'F';
118 s = gcap_string(s, c);
119 }
120
121 if (s == cap_str[i])
122 *s++ = '-';
123 *s = 0;
124 return cap_str[i];
125}
126
127void ceph_caps_init(struct ceph_mds_client *mdsc)
128{
129 INIT_LIST_HEAD(&mdsc->caps_list);
130 spin_lock_init(&mdsc->caps_list_lock);
131}
132
133void ceph_caps_finalize(struct ceph_mds_client *mdsc)
134{
135 struct ceph_cap *cap;
136
137 spin_lock(&mdsc->caps_list_lock);
138 while (!list_empty(&mdsc->caps_list)) {
139 cap = list_first_entry(&mdsc->caps_list,
140 struct ceph_cap, caps_item);
141 list_del(&cap->caps_item);
142 kmem_cache_free(ceph_cap_cachep, cap);
143 }
144 mdsc->caps_total_count = 0;
145 mdsc->caps_avail_count = 0;
146 mdsc->caps_use_count = 0;
147 mdsc->caps_reserve_count = 0;
148 mdsc->caps_min_count = 0;
149 spin_unlock(&mdsc->caps_list_lock);
150}
151
152void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc,
153 struct ceph_mount_options *fsopt)
154{
155 spin_lock(&mdsc->caps_list_lock);
156 mdsc->caps_min_count = fsopt->max_readdir;
157 if (mdsc->caps_min_count < 1024)
158 mdsc->caps_min_count = 1024;
159 mdsc->caps_use_max = fsopt->caps_max;
160 if (mdsc->caps_use_max > 0 &&
161 mdsc->caps_use_max < mdsc->caps_min_count)
162 mdsc->caps_use_max = mdsc->caps_min_count;
163 spin_unlock(&mdsc->caps_list_lock);
164}
165
166static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
167{
168 struct ceph_cap *cap;
169 int i;
170
171 if (nr_caps) {
172 BUG_ON(mdsc->caps_reserve_count < nr_caps);
173 mdsc->caps_reserve_count -= nr_caps;
174 if (mdsc->caps_avail_count >=
175 mdsc->caps_reserve_count + mdsc->caps_min_count) {
176 mdsc->caps_total_count -= nr_caps;
177 for (i = 0; i < nr_caps; i++) {
178 cap = list_first_entry(&mdsc->caps_list,
179 struct ceph_cap, caps_item);
180 list_del(&cap->caps_item);
181 kmem_cache_free(ceph_cap_cachep, cap);
182 }
183 } else {
184 mdsc->caps_avail_count += nr_caps;
185 }
186
187 dout("%s: caps %d = %d used + %d resv + %d avail\n",
188 __func__,
189 mdsc->caps_total_count, mdsc->caps_use_count,
190 mdsc->caps_reserve_count, mdsc->caps_avail_count);
191 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 mdsc->caps_reserve_count +
193 mdsc->caps_avail_count);
194 }
195}
196
197/*
198 * Called under mdsc->mutex.
199 */
200int ceph_reserve_caps(struct ceph_mds_client *mdsc,
201 struct ceph_cap_reservation *ctx, int need)
202{
203 int i, j;
204 struct ceph_cap *cap;
205 int have;
206 int alloc = 0;
207 int max_caps;
208 int err = 0;
209 bool trimmed = false;
210 struct ceph_mds_session *s;
211 LIST_HEAD(newcaps);
212
213 dout("reserve caps ctx=%p need=%d\n", ctx, need);
214
215 /* first reserve any caps that are already allocated */
216 spin_lock(&mdsc->caps_list_lock);
217 if (mdsc->caps_avail_count >= need)
218 have = need;
219 else
220 have = mdsc->caps_avail_count;
221 mdsc->caps_avail_count -= have;
222 mdsc->caps_reserve_count += have;
223 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
224 mdsc->caps_reserve_count +
225 mdsc->caps_avail_count);
226 spin_unlock(&mdsc->caps_list_lock);
227
228 for (i = have; i < need; ) {
229 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
230 if (cap) {
231 list_add(&cap->caps_item, &newcaps);
232 alloc++;
233 i++;
234 continue;
235 }
236
237 if (!trimmed) {
238 for (j = 0; j < mdsc->max_sessions; j++) {
239 s = __ceph_lookup_mds_session(mdsc, j);
240 if (!s)
241 continue;
242 mutex_unlock(&mdsc->mutex);
243
244 mutex_lock(&s->s_mutex);
245 max_caps = s->s_nr_caps - (need - i);
246 ceph_trim_caps(mdsc, s, max_caps);
247 mutex_unlock(&s->s_mutex);
248
249 ceph_put_mds_session(s);
250 mutex_lock(&mdsc->mutex);
251 }
252 trimmed = true;
253
254 spin_lock(&mdsc->caps_list_lock);
255 if (mdsc->caps_avail_count) {
256 int more_have;
257 if (mdsc->caps_avail_count >= need - i)
258 more_have = need - i;
259 else
260 more_have = mdsc->caps_avail_count;
261
262 i += more_have;
263 have += more_have;
264 mdsc->caps_avail_count -= more_have;
265 mdsc->caps_reserve_count += more_have;
266
267 }
268 spin_unlock(&mdsc->caps_list_lock);
269
270 continue;
271 }
272
273 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
274 ctx, need, have + alloc);
275 err = -ENOMEM;
276 break;
277 }
278
279 if (!err) {
280 BUG_ON(have + alloc != need);
281 ctx->count = need;
282 ctx->used = 0;
283 }
284
285 spin_lock(&mdsc->caps_list_lock);
286 mdsc->caps_total_count += alloc;
287 mdsc->caps_reserve_count += alloc;
288 list_splice(&newcaps, &mdsc->caps_list);
289
290 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
291 mdsc->caps_reserve_count +
292 mdsc->caps_avail_count);
293
294 if (err)
295 __ceph_unreserve_caps(mdsc, have + alloc);
296
297 spin_unlock(&mdsc->caps_list_lock);
298
299 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
300 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
301 mdsc->caps_reserve_count, mdsc->caps_avail_count);
302 return err;
303}
304
305void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
306 struct ceph_cap_reservation *ctx)
307{
308 bool reclaim = false;
309 if (!ctx->count)
310 return;
311
312 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
313 spin_lock(&mdsc->caps_list_lock);
314 __ceph_unreserve_caps(mdsc, ctx->count);
315 ctx->count = 0;
316
317 if (mdsc->caps_use_max > 0 &&
318 mdsc->caps_use_count > mdsc->caps_use_max)
319 reclaim = true;
320 spin_unlock(&mdsc->caps_list_lock);
321
322 if (reclaim)
323 ceph_reclaim_caps_nr(mdsc, ctx->used);
324}
325
326struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
327 struct ceph_cap_reservation *ctx)
328{
329 struct ceph_cap *cap = NULL;
330
331 /* temporary, until we do something about cap import/export */
332 if (!ctx) {
333 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
334 if (cap) {
335 spin_lock(&mdsc->caps_list_lock);
336 mdsc->caps_use_count++;
337 mdsc->caps_total_count++;
338 spin_unlock(&mdsc->caps_list_lock);
339 } else {
340 spin_lock(&mdsc->caps_list_lock);
341 if (mdsc->caps_avail_count) {
342 BUG_ON(list_empty(&mdsc->caps_list));
343
344 mdsc->caps_avail_count--;
345 mdsc->caps_use_count++;
346 cap = list_first_entry(&mdsc->caps_list,
347 struct ceph_cap, caps_item);
348 list_del(&cap->caps_item);
349
350 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
351 mdsc->caps_reserve_count + mdsc->caps_avail_count);
352 }
353 spin_unlock(&mdsc->caps_list_lock);
354 }
355
356 return cap;
357 }
358
359 spin_lock(&mdsc->caps_list_lock);
360 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
361 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
362 mdsc->caps_reserve_count, mdsc->caps_avail_count);
363 BUG_ON(!ctx->count);
364 BUG_ON(ctx->count > mdsc->caps_reserve_count);
365 BUG_ON(list_empty(&mdsc->caps_list));
366
367 ctx->count--;
368 ctx->used++;
369 mdsc->caps_reserve_count--;
370 mdsc->caps_use_count++;
371
372 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
373 list_del(&cap->caps_item);
374
375 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
376 mdsc->caps_reserve_count + mdsc->caps_avail_count);
377 spin_unlock(&mdsc->caps_list_lock);
378 return cap;
379}
380
381void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
382{
383 spin_lock(&mdsc->caps_list_lock);
384 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
385 cap, mdsc->caps_total_count, mdsc->caps_use_count,
386 mdsc->caps_reserve_count, mdsc->caps_avail_count);
387 mdsc->caps_use_count--;
388 /*
389 * Keep some preallocated caps around (ceph_min_count), to
390 * avoid lots of free/alloc churn.
391 */
392 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
393 mdsc->caps_min_count) {
394 mdsc->caps_total_count--;
395 kmem_cache_free(ceph_cap_cachep, cap);
396 } else {
397 mdsc->caps_avail_count++;
398 list_add(&cap->caps_item, &mdsc->caps_list);
399 }
400
401 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
402 mdsc->caps_reserve_count + mdsc->caps_avail_count);
403 spin_unlock(&mdsc->caps_list_lock);
404}
405
406void ceph_reservation_status(struct ceph_fs_client *fsc,
407 int *total, int *avail, int *used, int *reserved,
408 int *min)
409{
410 struct ceph_mds_client *mdsc = fsc->mdsc;
411
412 spin_lock(&mdsc->caps_list_lock);
413
414 if (total)
415 *total = mdsc->caps_total_count;
416 if (avail)
417 *avail = mdsc->caps_avail_count;
418 if (used)
419 *used = mdsc->caps_use_count;
420 if (reserved)
421 *reserved = mdsc->caps_reserve_count;
422 if (min)
423 *min = mdsc->caps_min_count;
424
425 spin_unlock(&mdsc->caps_list_lock);
426}
427
428/*
429 * Find ceph_cap for given mds, if any.
430 *
431 * Called with i_ceph_lock held.
432 */
433static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
434{
435 struct ceph_cap *cap;
436 struct rb_node *n = ci->i_caps.rb_node;
437
438 while (n) {
439 cap = rb_entry(n, struct ceph_cap, ci_node);
440 if (mds < cap->mds)
441 n = n->rb_left;
442 else if (mds > cap->mds)
443 n = n->rb_right;
444 else
445 return cap;
446 }
447 return NULL;
448}
449
450struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
451{
452 struct ceph_cap *cap;
453
454 spin_lock(&ci->i_ceph_lock);
455 cap = __get_cap_for_mds(ci, mds);
456 spin_unlock(&ci->i_ceph_lock);
457 return cap;
458}
459
460/*
461 * Called under i_ceph_lock.
462 */
463static void __insert_cap_node(struct ceph_inode_info *ci,
464 struct ceph_cap *new)
465{
466 struct rb_node **p = &ci->i_caps.rb_node;
467 struct rb_node *parent = NULL;
468 struct ceph_cap *cap = NULL;
469
470 while (*p) {
471 parent = *p;
472 cap = rb_entry(parent, struct ceph_cap, ci_node);
473 if (new->mds < cap->mds)
474 p = &(*p)->rb_left;
475 else if (new->mds > cap->mds)
476 p = &(*p)->rb_right;
477 else
478 BUG();
479 }
480
481 rb_link_node(&new->ci_node, parent, p);
482 rb_insert_color(&new->ci_node, &ci->i_caps);
483}
484
485/*
486 * (re)set cap hold timeouts, which control the delayed release
487 * of unused caps back to the MDS. Should be called on cap use.
488 */
489static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
490 struct ceph_inode_info *ci)
491{
492 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
493 ci->i_hold_caps_max = round_jiffies(jiffies +
494 opt->caps_wanted_delay_max * HZ);
495 dout("__cap_set_timeouts %p %lu\n", &ci->vfs_inode,
496 ci->i_hold_caps_max - jiffies);
497}
498
499/*
500 * (Re)queue cap at the end of the delayed cap release list.
501 *
502 * If I_FLUSH is set, leave the inode at the front of the list.
503 *
504 * Caller holds i_ceph_lock
505 * -> we take mdsc->cap_delay_lock
506 */
507static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
508 struct ceph_inode_info *ci)
509{
510 dout("__cap_delay_requeue %p flags 0x%lx at %lu\n", &ci->vfs_inode,
511 ci->i_ceph_flags, ci->i_hold_caps_max);
512 if (!mdsc->stopping) {
513 spin_lock(&mdsc->cap_delay_lock);
514 if (!list_empty(&ci->i_cap_delay_list)) {
515 if (ci->i_ceph_flags & CEPH_I_FLUSH)
516 goto no_change;
517 list_del_init(&ci->i_cap_delay_list);
518 }
519 __cap_set_timeouts(mdsc, ci);
520 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
521no_change:
522 spin_unlock(&mdsc->cap_delay_lock);
523 }
524}
525
526/*
527 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
528 * indicating we should send a cap message to flush dirty metadata
529 * asap, and move to the front of the delayed cap list.
530 */
531static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
532 struct ceph_inode_info *ci)
533{
534 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
535 spin_lock(&mdsc->cap_delay_lock);
536 ci->i_ceph_flags |= CEPH_I_FLUSH;
537 if (!list_empty(&ci->i_cap_delay_list))
538 list_del_init(&ci->i_cap_delay_list);
539 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
540 spin_unlock(&mdsc->cap_delay_lock);
541}
542
543/*
544 * Cancel delayed work on cap.
545 *
546 * Caller must hold i_ceph_lock.
547 */
548static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
549 struct ceph_inode_info *ci)
550{
551 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
552 if (list_empty(&ci->i_cap_delay_list))
553 return;
554 spin_lock(&mdsc->cap_delay_lock);
555 list_del_init(&ci->i_cap_delay_list);
556 spin_unlock(&mdsc->cap_delay_lock);
557}
558
559/* Common issue checks for add_cap, handle_cap_grant. */
560static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
561 unsigned issued)
562{
563 unsigned had = __ceph_caps_issued(ci, NULL);
564
565 lockdep_assert_held(&ci->i_ceph_lock);
566
567 /*
568 * Each time we receive FILE_CACHE anew, we increment
569 * i_rdcache_gen.
570 */
571 if (S_ISREG(ci->vfs_inode.i_mode) &&
572 (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
573 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
574 ci->i_rdcache_gen++;
575 }
576
577 /*
578 * If FILE_SHARED is newly issued, mark dir not complete. We don't
579 * know what happened to this directory while we didn't have the cap.
580 * If FILE_SHARED is being revoked, also mark dir not complete. It
581 * stops on-going cached readdir.
582 */
583 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
584 if (issued & CEPH_CAP_FILE_SHARED)
585 atomic_inc(&ci->i_shared_gen);
586 if (S_ISDIR(ci->vfs_inode.i_mode)) {
587 dout(" marking %p NOT complete\n", &ci->vfs_inode);
588 __ceph_dir_clear_complete(ci);
589 }
590 }
591
592 /* Wipe saved layout if we're losing DIR_CREATE caps */
593 if (S_ISDIR(ci->vfs_inode.i_mode) && (had & CEPH_CAP_DIR_CREATE) &&
594 !(issued & CEPH_CAP_DIR_CREATE)) {
595 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
596 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
597 }
598}
599
600/**
601 * change_auth_cap_ses - move inode to appropriate lists when auth caps change
602 * @ci: inode to be moved
603 * @session: new auth caps session
604 */
605static void change_auth_cap_ses(struct ceph_inode_info *ci,
606 struct ceph_mds_session *session)
607{
608 lockdep_assert_held(&ci->i_ceph_lock);
609
610 if (list_empty(&ci->i_dirty_item) && list_empty(&ci->i_flushing_item))
611 return;
612
613 spin_lock(&session->s_mdsc->cap_dirty_lock);
614 if (!list_empty(&ci->i_dirty_item))
615 list_move(&ci->i_dirty_item, &session->s_cap_dirty);
616 if (!list_empty(&ci->i_flushing_item))
617 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
618 spin_unlock(&session->s_mdsc->cap_dirty_lock);
619}
620
621/*
622 * Add a capability under the given MDS session.
623 *
624 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
625 *
626 * @fmode is the open file mode, if we are opening a file, otherwise
627 * it is < 0. (This is so we can atomically add the cap and add an
628 * open file reference to it.)
629 */
630void ceph_add_cap(struct inode *inode,
631 struct ceph_mds_session *session, u64 cap_id,
632 unsigned issued, unsigned wanted,
633 unsigned seq, unsigned mseq, u64 realmino, int flags,
634 struct ceph_cap **new_cap)
635{
636 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
637 struct ceph_inode_info *ci = ceph_inode(inode);
638 struct ceph_cap *cap;
639 int mds = session->s_mds;
640 int actual_wanted;
641 u32 gen;
642
643 lockdep_assert_held(&ci->i_ceph_lock);
644
645 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
646 session->s_mds, cap_id, ceph_cap_string(issued), seq);
647
648 gen = atomic_read(&session->s_cap_gen);
649
650 cap = __get_cap_for_mds(ci, mds);
651 if (!cap) {
652 cap = *new_cap;
653 *new_cap = NULL;
654
655 cap->issued = 0;
656 cap->implemented = 0;
657 cap->mds = mds;
658 cap->mds_wanted = 0;
659 cap->mseq = 0;
660
661 cap->ci = ci;
662 __insert_cap_node(ci, cap);
663
664 /* add to session cap list */
665 cap->session = session;
666 spin_lock(&session->s_cap_lock);
667 list_add_tail(&cap->session_caps, &session->s_caps);
668 session->s_nr_caps++;
669 atomic64_inc(&mdsc->metric.total_caps);
670 spin_unlock(&session->s_cap_lock);
671 } else {
672 spin_lock(&session->s_cap_lock);
673 list_move_tail(&cap->session_caps, &session->s_caps);
674 spin_unlock(&session->s_cap_lock);
675
676 if (cap->cap_gen < gen)
677 cap->issued = cap->implemented = CEPH_CAP_PIN;
678
679 /*
680 * auth mds of the inode changed. we received the cap export
681 * message, but still haven't received the cap import message.
682 * handle_cap_export() updated the new auth MDS' cap.
683 *
684 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
685 * a message that was send before the cap import message. So
686 * don't remove caps.
687 */
688 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
689 WARN_ON(cap != ci->i_auth_cap);
690 WARN_ON(cap->cap_id != cap_id);
691 seq = cap->seq;
692 mseq = cap->mseq;
693 issued |= cap->issued;
694 flags |= CEPH_CAP_FLAG_AUTH;
695 }
696 }
697
698 if (!ci->i_snap_realm ||
699 ((flags & CEPH_CAP_FLAG_AUTH) &&
700 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
701 /*
702 * add this inode to the appropriate snap realm
703 */
704 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
705 realmino);
706 if (realm) {
707 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
708 if (oldrealm) {
709 spin_lock(&oldrealm->inodes_with_caps_lock);
710 list_del_init(&ci->i_snap_realm_item);
711 spin_unlock(&oldrealm->inodes_with_caps_lock);
712 }
713
714 spin_lock(&realm->inodes_with_caps_lock);
715 list_add(&ci->i_snap_realm_item,
716 &realm->inodes_with_caps);
717 ci->i_snap_realm = realm;
718 if (realm->ino == ci->i_vino.ino)
719 realm->inode = inode;
720 spin_unlock(&realm->inodes_with_caps_lock);
721
722 if (oldrealm)
723 ceph_put_snap_realm(mdsc, oldrealm);
724 } else {
725 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
726 realmino);
727 WARN_ON(!realm);
728 }
729 }
730
731 __check_cap_issue(ci, cap, issued);
732
733 /*
734 * If we are issued caps we don't want, or the mds' wanted
735 * value appears to be off, queue a check so we'll release
736 * later and/or update the mds wanted value.
737 */
738 actual_wanted = __ceph_caps_wanted(ci);
739 if ((wanted & ~actual_wanted) ||
740 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
741 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
742 ceph_cap_string(issued), ceph_cap_string(wanted),
743 ceph_cap_string(actual_wanted));
744 __cap_delay_requeue(mdsc, ci);
745 }
746
747 if (flags & CEPH_CAP_FLAG_AUTH) {
748 if (!ci->i_auth_cap ||
749 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
750 if (ci->i_auth_cap &&
751 ci->i_auth_cap->session != cap->session)
752 change_auth_cap_ses(ci, cap->session);
753 ci->i_auth_cap = cap;
754 cap->mds_wanted = wanted;
755 }
756 } else {
757 WARN_ON(ci->i_auth_cap == cap);
758 }
759
760 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
761 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
762 ceph_cap_string(issued|cap->issued), seq, mds);
763 cap->cap_id = cap_id;
764 cap->issued = issued;
765 cap->implemented |= issued;
766 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
767 cap->mds_wanted = wanted;
768 else
769 cap->mds_wanted |= wanted;
770 cap->seq = seq;
771 cap->issue_seq = seq;
772 cap->mseq = mseq;
773 cap->cap_gen = gen;
774}
775
776/*
777 * Return true if cap has not timed out and belongs to the current
778 * generation of the MDS session (i.e. has not gone 'stale' due to
779 * us losing touch with the mds).
780 */
781static int __cap_is_valid(struct ceph_cap *cap)
782{
783 unsigned long ttl;
784 u32 gen;
785
786 gen = atomic_read(&cap->session->s_cap_gen);
787 ttl = cap->session->s_cap_ttl;
788
789 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
790 dout("__cap_is_valid %p cap %p issued %s "
791 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
792 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
793 return 0;
794 }
795
796 return 1;
797}
798
799/*
800 * Return set of valid cap bits issued to us. Note that caps time
801 * out, and may be invalidated in bulk if the client session times out
802 * and session->s_cap_gen is bumped.
803 */
804int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
805{
806 int have = ci->i_snap_caps;
807 struct ceph_cap *cap;
808 struct rb_node *p;
809
810 if (implemented)
811 *implemented = 0;
812 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
813 cap = rb_entry(p, struct ceph_cap, ci_node);
814 if (!__cap_is_valid(cap))
815 continue;
816 dout("__ceph_caps_issued %p cap %p issued %s\n",
817 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
818 have |= cap->issued;
819 if (implemented)
820 *implemented |= cap->implemented;
821 }
822 /*
823 * exclude caps issued by non-auth MDS, but are been revoking
824 * by the auth MDS. The non-auth MDS should be revoking/exporting
825 * these caps, but the message is delayed.
826 */
827 if (ci->i_auth_cap) {
828 cap = ci->i_auth_cap;
829 have &= ~cap->implemented | cap->issued;
830 }
831 return have;
832}
833
834/*
835 * Get cap bits issued by caps other than @ocap
836 */
837int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
838{
839 int have = ci->i_snap_caps;
840 struct ceph_cap *cap;
841 struct rb_node *p;
842
843 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
844 cap = rb_entry(p, struct ceph_cap, ci_node);
845 if (cap == ocap)
846 continue;
847 if (!__cap_is_valid(cap))
848 continue;
849 have |= cap->issued;
850 }
851 return have;
852}
853
854/*
855 * Move a cap to the end of the LRU (oldest caps at list head, newest
856 * at list tail).
857 */
858static void __touch_cap(struct ceph_cap *cap)
859{
860 struct ceph_mds_session *s = cap->session;
861
862 spin_lock(&s->s_cap_lock);
863 if (!s->s_cap_iterator) {
864 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
865 s->s_mds);
866 list_move_tail(&cap->session_caps, &s->s_caps);
867 } else {
868 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
869 &cap->ci->vfs_inode, cap, s->s_mds);
870 }
871 spin_unlock(&s->s_cap_lock);
872}
873
874/*
875 * Check if we hold the given mask. If so, move the cap(s) to the
876 * front of their respective LRUs. (This is the preferred way for
877 * callers to check for caps they want.)
878 */
879int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
880{
881 struct ceph_cap *cap;
882 struct rb_node *p;
883 int have = ci->i_snap_caps;
884
885 if ((have & mask) == mask) {
886 dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
887 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
888 ceph_cap_string(have),
889 ceph_cap_string(mask));
890 return 1;
891 }
892
893 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
894 cap = rb_entry(p, struct ceph_cap, ci_node);
895 if (!__cap_is_valid(cap))
896 continue;
897 if ((cap->issued & mask) == mask) {
898 dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
899 " (mask %s)\n", ceph_ino(&ci->vfs_inode), cap,
900 ceph_cap_string(cap->issued),
901 ceph_cap_string(mask));
902 if (touch)
903 __touch_cap(cap);
904 return 1;
905 }
906
907 /* does a combination of caps satisfy mask? */
908 have |= cap->issued;
909 if ((have & mask) == mask) {
910 dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
911 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
912 ceph_cap_string(cap->issued),
913 ceph_cap_string(mask));
914 if (touch) {
915 struct rb_node *q;
916
917 /* touch this + preceding caps */
918 __touch_cap(cap);
919 for (q = rb_first(&ci->i_caps); q != p;
920 q = rb_next(q)) {
921 cap = rb_entry(q, struct ceph_cap,
922 ci_node);
923 if (!__cap_is_valid(cap))
924 continue;
925 if (cap->issued & mask)
926 __touch_cap(cap);
927 }
928 }
929 return 1;
930 }
931 }
932
933 return 0;
934}
935
936int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
937 int touch)
938{
939 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
940 int r;
941
942 r = __ceph_caps_issued_mask(ci, mask, touch);
943 if (r)
944 ceph_update_cap_hit(&fsc->mdsc->metric);
945 else
946 ceph_update_cap_mis(&fsc->mdsc->metric);
947 return r;
948}
949
950/*
951 * Return true if mask caps are currently being revoked by an MDS.
952 */
953int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
954 struct ceph_cap *ocap, int mask)
955{
956 struct ceph_cap *cap;
957 struct rb_node *p;
958
959 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
960 cap = rb_entry(p, struct ceph_cap, ci_node);
961 if (cap != ocap &&
962 (cap->implemented & ~cap->issued & mask))
963 return 1;
964 }
965 return 0;
966}
967
968int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
969{
970 struct inode *inode = &ci->vfs_inode;
971 int ret;
972
973 spin_lock(&ci->i_ceph_lock);
974 ret = __ceph_caps_revoking_other(ci, NULL, mask);
975 spin_unlock(&ci->i_ceph_lock);
976 dout("ceph_caps_revoking %p %s = %d\n", inode,
977 ceph_cap_string(mask), ret);
978 return ret;
979}
980
981int __ceph_caps_used(struct ceph_inode_info *ci)
982{
983 int used = 0;
984 if (ci->i_pin_ref)
985 used |= CEPH_CAP_PIN;
986 if (ci->i_rd_ref)
987 used |= CEPH_CAP_FILE_RD;
988 if (ci->i_rdcache_ref ||
989 (S_ISREG(ci->vfs_inode.i_mode) &&
990 ci->vfs_inode.i_data.nrpages))
991 used |= CEPH_CAP_FILE_CACHE;
992 if (ci->i_wr_ref)
993 used |= CEPH_CAP_FILE_WR;
994 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
995 used |= CEPH_CAP_FILE_BUFFER;
996 if (ci->i_fx_ref)
997 used |= CEPH_CAP_FILE_EXCL;
998 return used;
999}
1000
1001#define FMODE_WAIT_BIAS 1000
1002
1003/*
1004 * wanted, by virtue of open file modes
1005 */
1006int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
1007{
1008 const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
1009 const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
1010 const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
1011 const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
1012 struct ceph_mount_options *opt =
1013 ceph_inode_to_client(&ci->vfs_inode)->mount_options;
1014 unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
1015 unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1016
1017 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1018 int want = 0;
1019
1020 /* use used_cutoff here, to keep dir's wanted caps longer */
1021 if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1022 time_after(ci->i_last_rd, used_cutoff))
1023 want |= CEPH_CAP_ANY_SHARED;
1024
1025 if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1026 time_after(ci->i_last_wr, used_cutoff)) {
1027 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1028 if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1029 want |= CEPH_CAP_ANY_DIR_OPS;
1030 }
1031
1032 if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1033 want |= CEPH_CAP_PIN;
1034
1035 return want;
1036 } else {
1037 int bits = 0;
1038
1039 if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1040 if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1041 time_after(ci->i_last_rd, used_cutoff))
1042 bits |= 1 << RD_SHIFT;
1043 } else if (time_after(ci->i_last_rd, idle_cutoff)) {
1044 bits |= 1 << RD_SHIFT;
1045 }
1046
1047 if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1048 if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1049 time_after(ci->i_last_wr, used_cutoff))
1050 bits |= 1 << WR_SHIFT;
1051 } else if (time_after(ci->i_last_wr, idle_cutoff)) {
1052 bits |= 1 << WR_SHIFT;
1053 }
1054
1055 /* check lazyio only when read/write is wanted */
1056 if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1057 ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1058 bits |= 1 << LAZY_SHIFT;
1059
1060 return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1061 }
1062}
1063
1064/*
1065 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1066 */
1067int __ceph_caps_wanted(struct ceph_inode_info *ci)
1068{
1069 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1070 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1071 /* we want EXCL if holding caps of dir ops */
1072 if (w & CEPH_CAP_ANY_DIR_OPS)
1073 w |= CEPH_CAP_FILE_EXCL;
1074 } else {
1075 /* we want EXCL if dirty data */
1076 if (w & CEPH_CAP_FILE_BUFFER)
1077 w |= CEPH_CAP_FILE_EXCL;
1078 }
1079 return w;
1080}
1081
1082/*
1083 * Return caps we have registered with the MDS(s) as 'wanted'.
1084 */
1085int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1086{
1087 struct ceph_cap *cap;
1088 struct rb_node *p;
1089 int mds_wanted = 0;
1090
1091 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1092 cap = rb_entry(p, struct ceph_cap, ci_node);
1093 if (check && !__cap_is_valid(cap))
1094 continue;
1095 if (cap == ci->i_auth_cap)
1096 mds_wanted |= cap->mds_wanted;
1097 else
1098 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1099 }
1100 return mds_wanted;
1101}
1102
1103int ceph_is_any_caps(struct inode *inode)
1104{
1105 struct ceph_inode_info *ci = ceph_inode(inode);
1106 int ret;
1107
1108 spin_lock(&ci->i_ceph_lock);
1109 ret = __ceph_is_any_real_caps(ci);
1110 spin_unlock(&ci->i_ceph_lock);
1111
1112 return ret;
1113}
1114
1115static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1116{
1117 struct ceph_snap_realm *realm = ci->i_snap_realm;
1118 spin_lock(&realm->inodes_with_caps_lock);
1119 list_del_init(&ci->i_snap_realm_item);
1120 ci->i_snap_realm_counter++;
1121 ci->i_snap_realm = NULL;
1122 if (realm->ino == ci->i_vino.ino)
1123 realm->inode = NULL;
1124 spin_unlock(&realm->inodes_with_caps_lock);
1125 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1126 realm);
1127}
1128
1129/*
1130 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1131 *
1132 * caller should hold i_ceph_lock.
1133 * caller will not hold session s_mutex if called from destroy_inode.
1134 */
1135void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1136{
1137 struct ceph_mds_session *session = cap->session;
1138 struct ceph_inode_info *ci = cap->ci;
1139 struct ceph_mds_client *mdsc;
1140 int removed = 0;
1141
1142 /* 'ci' being NULL means the remove have already occurred */
1143 if (!ci) {
1144 dout("%s: cap inode is NULL\n", __func__);
1145 return;
1146 }
1147
1148 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1149
1150 mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1151
1152 /* remove from inode's cap rbtree, and clear auth cap */
1153 rb_erase(&cap->ci_node, &ci->i_caps);
1154 if (ci->i_auth_cap == cap) {
1155 WARN_ON_ONCE(!list_empty(&ci->i_dirty_item) &&
1156 !mdsc->fsc->blocklisted);
1157 ci->i_auth_cap = NULL;
1158 }
1159
1160 /* remove from session list */
1161 spin_lock(&session->s_cap_lock);
1162 if (session->s_cap_iterator == cap) {
1163 /* not yet, we are iterating over this very cap */
1164 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1165 cap, cap->session);
1166 } else {
1167 list_del_init(&cap->session_caps);
1168 session->s_nr_caps--;
1169 atomic64_dec(&mdsc->metric.total_caps);
1170 cap->session = NULL;
1171 removed = 1;
1172 }
1173 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1174 cap->ci = NULL;
1175
1176 /*
1177 * s_cap_reconnect is protected by s_cap_lock. no one changes
1178 * s_cap_gen while session is in the reconnect state.
1179 */
1180 if (queue_release &&
1181 (!session->s_cap_reconnect ||
1182 cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1183 cap->queue_release = 1;
1184 if (removed) {
1185 __ceph_queue_cap_release(session, cap);
1186 removed = 0;
1187 }
1188 } else {
1189 cap->queue_release = 0;
1190 }
1191 cap->cap_ino = ci->i_vino.ino;
1192
1193 spin_unlock(&session->s_cap_lock);
1194
1195 if (removed)
1196 ceph_put_cap(mdsc, cap);
1197
1198 if (!__ceph_is_any_real_caps(ci)) {
1199 /* when reconnect denied, we remove session caps forcibly,
1200 * i_wr_ref can be non-zero. If there are ongoing write,
1201 * keep i_snap_realm.
1202 */
1203 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1204 drop_inode_snap_realm(ci);
1205
1206 __cap_delay_cancel(mdsc, ci);
1207 }
1208}
1209
1210struct cap_msg_args {
1211 struct ceph_mds_session *session;
1212 u64 ino, cid, follows;
1213 u64 flush_tid, oldest_flush_tid, size, max_size;
1214 u64 xattr_version;
1215 u64 change_attr;
1216 struct ceph_buffer *xattr_buf;
1217 struct ceph_buffer *old_xattr_buf;
1218 struct timespec64 atime, mtime, ctime, btime;
1219 int op, caps, wanted, dirty;
1220 u32 seq, issue_seq, mseq, time_warp_seq;
1221 u32 flags;
1222 kuid_t uid;
1223 kgid_t gid;
1224 umode_t mode;
1225 bool inline_data;
1226 bool wake;
1227};
1228
1229/*
1230 * cap struct size + flock buffer size + inline version + inline data size +
1231 * osd_epoch_barrier + oldest_flush_tid
1232 */
1233#define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1234 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1235
1236/* Marshal up the cap msg to the MDS */
1237static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1238{
1239 struct ceph_mds_caps *fc;
1240 void *p;
1241 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1242
1243 dout("%s %s %llx %llx caps %s wanted %s dirty %s seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu xattr_ver %llu xattr_len %d\n",
1244 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1245 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1246 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1247 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1248 arg->size, arg->max_size, arg->xattr_version,
1249 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1250
1251 msg->hdr.version = cpu_to_le16(10);
1252 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1253
1254 fc = msg->front.iov_base;
1255 memset(fc, 0, sizeof(*fc));
1256
1257 fc->cap_id = cpu_to_le64(arg->cid);
1258 fc->op = cpu_to_le32(arg->op);
1259 fc->seq = cpu_to_le32(arg->seq);
1260 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1261 fc->migrate_seq = cpu_to_le32(arg->mseq);
1262 fc->caps = cpu_to_le32(arg->caps);
1263 fc->wanted = cpu_to_le32(arg->wanted);
1264 fc->dirty = cpu_to_le32(arg->dirty);
1265 fc->ino = cpu_to_le64(arg->ino);
1266 fc->snap_follows = cpu_to_le64(arg->follows);
1267
1268 fc->size = cpu_to_le64(arg->size);
1269 fc->max_size = cpu_to_le64(arg->max_size);
1270 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1271 ceph_encode_timespec64(&fc->atime, &arg->atime);
1272 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1273 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1274
1275 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1276 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1277 fc->mode = cpu_to_le32(arg->mode);
1278
1279 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1280 if (arg->xattr_buf) {
1281 msg->middle = ceph_buffer_get(arg->xattr_buf);
1282 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1283 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1284 }
1285
1286 p = fc + 1;
1287 /* flock buffer size (version 2) */
1288 ceph_encode_32(&p, 0);
1289 /* inline version (version 4) */
1290 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1291 /* inline data size */
1292 ceph_encode_32(&p, 0);
1293 /*
1294 * osd_epoch_barrier (version 5)
1295 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1296 * case it was recently changed
1297 */
1298 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1299 /* oldest_flush_tid (version 6) */
1300 ceph_encode_64(&p, arg->oldest_flush_tid);
1301
1302 /*
1303 * caller_uid/caller_gid (version 7)
1304 *
1305 * Currently, we don't properly track which caller dirtied the caps
1306 * last, and force a flush of them when there is a conflict. For now,
1307 * just set this to 0:0, to emulate how the MDS has worked up to now.
1308 */
1309 ceph_encode_32(&p, 0);
1310 ceph_encode_32(&p, 0);
1311
1312 /* pool namespace (version 8) (mds always ignores this) */
1313 ceph_encode_32(&p, 0);
1314
1315 /* btime and change_attr (version 9) */
1316 ceph_encode_timespec64(p, &arg->btime);
1317 p += sizeof(struct ceph_timespec);
1318 ceph_encode_64(&p, arg->change_attr);
1319
1320 /* Advisory flags (version 10) */
1321 ceph_encode_32(&p, arg->flags);
1322}
1323
1324/*
1325 * Queue cap releases when an inode is dropped from our cache.
1326 */
1327void __ceph_remove_caps(struct ceph_inode_info *ci)
1328{
1329 struct rb_node *p;
1330
1331 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1332 * may call __ceph_caps_issued_mask() on a freeing inode. */
1333 spin_lock(&ci->i_ceph_lock);
1334 p = rb_first(&ci->i_caps);
1335 while (p) {
1336 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1337 p = rb_next(p);
1338 __ceph_remove_cap(cap, true);
1339 }
1340 spin_unlock(&ci->i_ceph_lock);
1341}
1342
1343/*
1344 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1345 * the arg struct with the parameters that will need to be sent. This should
1346 * be done under the i_ceph_lock to guard against changes to cap state.
1347 *
1348 * Make note of max_size reported/requested from mds, revoked caps
1349 * that have now been implemented.
1350 */
1351static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1352 int op, int flags, int used, int want, int retain,
1353 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1354{
1355 struct ceph_inode_info *ci = cap->ci;
1356 struct inode *inode = &ci->vfs_inode;
1357 int held, revoking;
1358
1359 lockdep_assert_held(&ci->i_ceph_lock);
1360
1361 held = cap->issued | cap->implemented;
1362 revoking = cap->implemented & ~cap->issued;
1363 retain &= ~revoking;
1364
1365 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1366 __func__, inode, cap, cap->session,
1367 ceph_cap_string(held), ceph_cap_string(held & retain),
1368 ceph_cap_string(revoking));
1369 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1370
1371 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1372
1373 cap->issued &= retain; /* drop bits we don't want */
1374 /*
1375 * Wake up any waiters on wanted -> needed transition. This is due to
1376 * the weird transition from buffered to sync IO... we need to flush
1377 * dirty pages _before_ allowing sync writes to avoid reordering.
1378 */
1379 arg->wake = cap->implemented & ~cap->issued;
1380 cap->implemented &= cap->issued | used;
1381 cap->mds_wanted = want;
1382
1383 arg->session = cap->session;
1384 arg->ino = ceph_vino(inode).ino;
1385 arg->cid = cap->cap_id;
1386 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1387 arg->flush_tid = flush_tid;
1388 arg->oldest_flush_tid = oldest_flush_tid;
1389
1390 arg->size = i_size_read(inode);
1391 ci->i_reported_size = arg->size;
1392 arg->max_size = ci->i_wanted_max_size;
1393 if (cap == ci->i_auth_cap) {
1394 if (want & CEPH_CAP_ANY_FILE_WR)
1395 ci->i_requested_max_size = arg->max_size;
1396 else
1397 ci->i_requested_max_size = 0;
1398 }
1399
1400 if (flushing & CEPH_CAP_XATTR_EXCL) {
1401 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1402 arg->xattr_version = ci->i_xattrs.version;
1403 arg->xattr_buf = ci->i_xattrs.blob;
1404 } else {
1405 arg->xattr_buf = NULL;
1406 arg->old_xattr_buf = NULL;
1407 }
1408
1409 arg->mtime = inode->i_mtime;
1410 arg->atime = inode->i_atime;
1411 arg->ctime = inode->i_ctime;
1412 arg->btime = ci->i_btime;
1413 arg->change_attr = inode_peek_iversion_raw(inode);
1414
1415 arg->op = op;
1416 arg->caps = cap->implemented;
1417 arg->wanted = want;
1418 arg->dirty = flushing;
1419
1420 arg->seq = cap->seq;
1421 arg->issue_seq = cap->issue_seq;
1422 arg->mseq = cap->mseq;
1423 arg->time_warp_seq = ci->i_time_warp_seq;
1424
1425 arg->uid = inode->i_uid;
1426 arg->gid = inode->i_gid;
1427 arg->mode = inode->i_mode;
1428
1429 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1430 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1431 !list_empty(&ci->i_cap_snaps)) {
1432 struct ceph_cap_snap *capsnap;
1433 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1434 if (capsnap->cap_flush.tid)
1435 break;
1436 if (capsnap->need_flush) {
1437 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1438 break;
1439 }
1440 }
1441 }
1442 arg->flags = flags;
1443}
1444
1445/*
1446 * Send a cap msg on the given inode.
1447 *
1448 * Caller should hold snap_rwsem (read), s_mutex.
1449 */
1450static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1451{
1452 struct ceph_msg *msg;
1453 struct inode *inode = &ci->vfs_inode;
1454
1455 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1456 if (!msg) {
1457 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1458 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1459 arg->flush_tid);
1460 spin_lock(&ci->i_ceph_lock);
1461 __cap_delay_requeue(arg->session->s_mdsc, ci);
1462 spin_unlock(&ci->i_ceph_lock);
1463 return;
1464 }
1465
1466 encode_cap_msg(msg, arg);
1467 ceph_con_send(&arg->session->s_con, msg);
1468 ceph_buffer_put(arg->old_xattr_buf);
1469 if (arg->wake)
1470 wake_up_all(&ci->i_cap_wq);
1471}
1472
1473static inline int __send_flush_snap(struct inode *inode,
1474 struct ceph_mds_session *session,
1475 struct ceph_cap_snap *capsnap,
1476 u32 mseq, u64 oldest_flush_tid)
1477{
1478 struct cap_msg_args arg;
1479 struct ceph_msg *msg;
1480
1481 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1482 if (!msg)
1483 return -ENOMEM;
1484
1485 arg.session = session;
1486 arg.ino = ceph_vino(inode).ino;
1487 arg.cid = 0;
1488 arg.follows = capsnap->follows;
1489 arg.flush_tid = capsnap->cap_flush.tid;
1490 arg.oldest_flush_tid = oldest_flush_tid;
1491
1492 arg.size = capsnap->size;
1493 arg.max_size = 0;
1494 arg.xattr_version = capsnap->xattr_version;
1495 arg.xattr_buf = capsnap->xattr_blob;
1496 arg.old_xattr_buf = NULL;
1497
1498 arg.atime = capsnap->atime;
1499 arg.mtime = capsnap->mtime;
1500 arg.ctime = capsnap->ctime;
1501 arg.btime = capsnap->btime;
1502 arg.change_attr = capsnap->change_attr;
1503
1504 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1505 arg.caps = capsnap->issued;
1506 arg.wanted = 0;
1507 arg.dirty = capsnap->dirty;
1508
1509 arg.seq = 0;
1510 arg.issue_seq = 0;
1511 arg.mseq = mseq;
1512 arg.time_warp_seq = capsnap->time_warp_seq;
1513
1514 arg.uid = capsnap->uid;
1515 arg.gid = capsnap->gid;
1516 arg.mode = capsnap->mode;
1517
1518 arg.inline_data = capsnap->inline_data;
1519 arg.flags = 0;
1520 arg.wake = false;
1521
1522 encode_cap_msg(msg, &arg);
1523 ceph_con_send(&arg.session->s_con, msg);
1524 return 0;
1525}
1526
1527/*
1528 * When a snapshot is taken, clients accumulate dirty metadata on
1529 * inodes with capabilities in ceph_cap_snaps to describe the file
1530 * state at the time the snapshot was taken. This must be flushed
1531 * asynchronously back to the MDS once sync writes complete and dirty
1532 * data is written out.
1533 *
1534 * Called under i_ceph_lock.
1535 */
1536static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1537 struct ceph_mds_session *session)
1538 __releases(ci->i_ceph_lock)
1539 __acquires(ci->i_ceph_lock)
1540{
1541 struct inode *inode = &ci->vfs_inode;
1542 struct ceph_mds_client *mdsc = session->s_mdsc;
1543 struct ceph_cap_snap *capsnap;
1544 u64 oldest_flush_tid = 0;
1545 u64 first_tid = 1, last_tid = 0;
1546
1547 dout("__flush_snaps %p session %p\n", inode, session);
1548
1549 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1550 /*
1551 * we need to wait for sync writes to complete and for dirty
1552 * pages to be written out.
1553 */
1554 if (capsnap->dirty_pages || capsnap->writing)
1555 break;
1556
1557 /* should be removed by ceph_try_drop_cap_snap() */
1558 BUG_ON(!capsnap->need_flush);
1559
1560 /* only flush each capsnap once */
1561 if (capsnap->cap_flush.tid > 0) {
1562 dout(" already flushed %p, skipping\n", capsnap);
1563 continue;
1564 }
1565
1566 spin_lock(&mdsc->cap_dirty_lock);
1567 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1568 list_add_tail(&capsnap->cap_flush.g_list,
1569 &mdsc->cap_flush_list);
1570 if (oldest_flush_tid == 0)
1571 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1572 if (list_empty(&ci->i_flushing_item)) {
1573 list_add_tail(&ci->i_flushing_item,
1574 &session->s_cap_flushing);
1575 }
1576 spin_unlock(&mdsc->cap_dirty_lock);
1577
1578 list_add_tail(&capsnap->cap_flush.i_list,
1579 &ci->i_cap_flush_list);
1580
1581 if (first_tid == 1)
1582 first_tid = capsnap->cap_flush.tid;
1583 last_tid = capsnap->cap_flush.tid;
1584 }
1585
1586 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1587
1588 while (first_tid <= last_tid) {
1589 struct ceph_cap *cap = ci->i_auth_cap;
1590 struct ceph_cap_flush *cf;
1591 int ret;
1592
1593 if (!(cap && cap->session == session)) {
1594 dout("__flush_snaps %p auth cap %p not mds%d, "
1595 "stop\n", inode, cap, session->s_mds);
1596 break;
1597 }
1598
1599 ret = -ENOENT;
1600 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1601 if (cf->tid >= first_tid) {
1602 ret = 0;
1603 break;
1604 }
1605 }
1606 if (ret < 0)
1607 break;
1608
1609 first_tid = cf->tid + 1;
1610
1611 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1612 refcount_inc(&capsnap->nref);
1613 spin_unlock(&ci->i_ceph_lock);
1614
1615 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1616 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1617
1618 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1619 oldest_flush_tid);
1620 if (ret < 0) {
1621 pr_err("__flush_snaps: error sending cap flushsnap, "
1622 "ino (%llx.%llx) tid %llu follows %llu\n",
1623 ceph_vinop(inode), cf->tid, capsnap->follows);
1624 }
1625
1626 ceph_put_cap_snap(capsnap);
1627 spin_lock(&ci->i_ceph_lock);
1628 }
1629}
1630
1631void ceph_flush_snaps(struct ceph_inode_info *ci,
1632 struct ceph_mds_session **psession)
1633{
1634 struct inode *inode = &ci->vfs_inode;
1635 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1636 struct ceph_mds_session *session = NULL;
1637 int mds;
1638
1639 dout("ceph_flush_snaps %p\n", inode);
1640 if (psession)
1641 session = *psession;
1642retry:
1643 spin_lock(&ci->i_ceph_lock);
1644 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1645 dout(" no capsnap needs flush, doing nothing\n");
1646 goto out;
1647 }
1648 if (!ci->i_auth_cap) {
1649 dout(" no auth cap (migrating?), doing nothing\n");
1650 goto out;
1651 }
1652
1653 mds = ci->i_auth_cap->session->s_mds;
1654 if (session && session->s_mds != mds) {
1655 dout(" oops, wrong session %p mutex\n", session);
1656 ceph_put_mds_session(session);
1657 session = NULL;
1658 }
1659 if (!session) {
1660 spin_unlock(&ci->i_ceph_lock);
1661 mutex_lock(&mdsc->mutex);
1662 session = __ceph_lookup_mds_session(mdsc, mds);
1663 mutex_unlock(&mdsc->mutex);
1664 goto retry;
1665 }
1666
1667 // make sure flushsnap messages are sent in proper order.
1668 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1669 __kick_flushing_caps(mdsc, session, ci, 0);
1670
1671 __ceph_flush_snaps(ci, session);
1672out:
1673 spin_unlock(&ci->i_ceph_lock);
1674
1675 if (psession)
1676 *psession = session;
1677 else
1678 ceph_put_mds_session(session);
1679 /* we flushed them all; remove this inode from the queue */
1680 spin_lock(&mdsc->snap_flush_lock);
1681 list_del_init(&ci->i_snap_flush_item);
1682 spin_unlock(&mdsc->snap_flush_lock);
1683}
1684
1685/*
1686 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1687 * Caller is then responsible for calling __mark_inode_dirty with the
1688 * returned flags value.
1689 */
1690int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1691 struct ceph_cap_flush **pcf)
1692{
1693 struct ceph_mds_client *mdsc =
1694 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1695 struct inode *inode = &ci->vfs_inode;
1696 int was = ci->i_dirty_caps;
1697 int dirty = 0;
1698
1699 lockdep_assert_held(&ci->i_ceph_lock);
1700
1701 if (!ci->i_auth_cap) {
1702 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1703 "but no auth cap (session was closed?)\n",
1704 inode, ceph_ino(inode), ceph_cap_string(mask));
1705 return 0;
1706 }
1707
1708 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1709 ceph_cap_string(mask), ceph_cap_string(was),
1710 ceph_cap_string(was | mask));
1711 ci->i_dirty_caps |= mask;
1712 if (was == 0) {
1713 struct ceph_mds_session *session = ci->i_auth_cap->session;
1714
1715 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1716 swap(ci->i_prealloc_cap_flush, *pcf);
1717
1718 if (!ci->i_head_snapc) {
1719 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1720 ci->i_head_snapc = ceph_get_snap_context(
1721 ci->i_snap_realm->cached_context);
1722 }
1723 dout(" inode %p now dirty snapc %p auth cap %p\n",
1724 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1725 BUG_ON(!list_empty(&ci->i_dirty_item));
1726 spin_lock(&mdsc->cap_dirty_lock);
1727 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1728 spin_unlock(&mdsc->cap_dirty_lock);
1729 if (ci->i_flushing_caps == 0) {
1730 ihold(inode);
1731 dirty |= I_DIRTY_SYNC;
1732 }
1733 } else {
1734 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1735 }
1736 BUG_ON(list_empty(&ci->i_dirty_item));
1737 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1738 (mask & CEPH_CAP_FILE_BUFFER))
1739 dirty |= I_DIRTY_DATASYNC;
1740 __cap_delay_requeue(mdsc, ci);
1741 return dirty;
1742}
1743
1744struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1745{
1746 struct ceph_cap_flush *cf;
1747
1748 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1749 if (!cf)
1750 return NULL;
1751
1752 cf->is_capsnap = false;
1753 return cf;
1754}
1755
1756void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1757{
1758 if (cf)
1759 kmem_cache_free(ceph_cap_flush_cachep, cf);
1760}
1761
1762static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1763{
1764 if (!list_empty(&mdsc->cap_flush_list)) {
1765 struct ceph_cap_flush *cf =
1766 list_first_entry(&mdsc->cap_flush_list,
1767 struct ceph_cap_flush, g_list);
1768 return cf->tid;
1769 }
1770 return 0;
1771}
1772
1773/*
1774 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1775 * Return true if caller needs to wake up flush waiters.
1776 */
1777static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1778 struct ceph_cap_flush *cf)
1779{
1780 struct ceph_cap_flush *prev;
1781 bool wake = cf->wake;
1782
1783 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1784 prev = list_prev_entry(cf, g_list);
1785 prev->wake = true;
1786 wake = false;
1787 }
1788 list_del_init(&cf->g_list);
1789 return wake;
1790}
1791
1792static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1793 struct ceph_cap_flush *cf)
1794{
1795 struct ceph_cap_flush *prev;
1796 bool wake = cf->wake;
1797
1798 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1799 prev = list_prev_entry(cf, i_list);
1800 prev->wake = true;
1801 wake = false;
1802 }
1803 list_del_init(&cf->i_list);
1804 return wake;
1805}
1806
1807/*
1808 * Add dirty inode to the flushing list. Assigned a seq number so we
1809 * can wait for caps to flush without starving.
1810 *
1811 * Called under i_ceph_lock. Returns the flush tid.
1812 */
1813static u64 __mark_caps_flushing(struct inode *inode,
1814 struct ceph_mds_session *session, bool wake,
1815 u64 *oldest_flush_tid)
1816{
1817 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1818 struct ceph_inode_info *ci = ceph_inode(inode);
1819 struct ceph_cap_flush *cf = NULL;
1820 int flushing;
1821
1822 lockdep_assert_held(&ci->i_ceph_lock);
1823 BUG_ON(ci->i_dirty_caps == 0);
1824 BUG_ON(list_empty(&ci->i_dirty_item));
1825 BUG_ON(!ci->i_prealloc_cap_flush);
1826
1827 flushing = ci->i_dirty_caps;
1828 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1829 ceph_cap_string(flushing),
1830 ceph_cap_string(ci->i_flushing_caps),
1831 ceph_cap_string(ci->i_flushing_caps | flushing));
1832 ci->i_flushing_caps |= flushing;
1833 ci->i_dirty_caps = 0;
1834 dout(" inode %p now !dirty\n", inode);
1835
1836 swap(cf, ci->i_prealloc_cap_flush);
1837 cf->caps = flushing;
1838 cf->wake = wake;
1839
1840 spin_lock(&mdsc->cap_dirty_lock);
1841 list_del_init(&ci->i_dirty_item);
1842
1843 cf->tid = ++mdsc->last_cap_flush_tid;
1844 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1845 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1846
1847 if (list_empty(&ci->i_flushing_item)) {
1848 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1849 mdsc->num_cap_flushing++;
1850 }
1851 spin_unlock(&mdsc->cap_dirty_lock);
1852
1853 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1854
1855 return cf->tid;
1856}
1857
1858/*
1859 * try to invalidate mapping pages without blocking.
1860 */
1861static int try_nonblocking_invalidate(struct inode *inode)
1862 __releases(ci->i_ceph_lock)
1863 __acquires(ci->i_ceph_lock)
1864{
1865 struct ceph_inode_info *ci = ceph_inode(inode);
1866 u32 invalidating_gen = ci->i_rdcache_gen;
1867
1868 spin_unlock(&ci->i_ceph_lock);
1869 ceph_fscache_invalidate(inode);
1870 invalidate_mapping_pages(&inode->i_data, 0, -1);
1871 spin_lock(&ci->i_ceph_lock);
1872
1873 if (inode->i_data.nrpages == 0 &&
1874 invalidating_gen == ci->i_rdcache_gen) {
1875 /* success. */
1876 dout("try_nonblocking_invalidate %p success\n", inode);
1877 /* save any racing async invalidate some trouble */
1878 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1879 return 0;
1880 }
1881 dout("try_nonblocking_invalidate %p failed\n", inode);
1882 return -1;
1883}
1884
1885bool __ceph_should_report_size(struct ceph_inode_info *ci)
1886{
1887 loff_t size = i_size_read(&ci->vfs_inode);
1888 /* mds will adjust max size according to the reported size */
1889 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1890 return false;
1891 if (size >= ci->i_max_size)
1892 return true;
1893 /* half of previous max_size increment has been used */
1894 if (ci->i_max_size > ci->i_reported_size &&
1895 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1896 return true;
1897 return false;
1898}
1899
1900/*
1901 * Swiss army knife function to examine currently used and wanted
1902 * versus held caps. Release, flush, ack revoked caps to mds as
1903 * appropriate.
1904 *
1905 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1906 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1907 * further delay.
1908 */
1909void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1910 struct ceph_mds_session *session)
1911{
1912 struct inode *inode = &ci->vfs_inode;
1913 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1914 struct ceph_cap *cap;
1915 u64 flush_tid, oldest_flush_tid;
1916 int file_wanted, used, cap_used;
1917 int issued, implemented, want, retain, revoking, flushing = 0;
1918 int mds = -1; /* keep track of how far we've gone through i_caps list
1919 to avoid an infinite loop on retry */
1920 struct rb_node *p;
1921 bool queue_invalidate = false;
1922 bool tried_invalidate = false;
1923
1924 if (session)
1925 ceph_get_mds_session(session);
1926
1927 spin_lock(&ci->i_ceph_lock);
1928 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1929 flags |= CHECK_CAPS_FLUSH;
1930retry:
1931 /* Caps wanted by virtue of active open files. */
1932 file_wanted = __ceph_caps_file_wanted(ci);
1933
1934 /* Caps which have active references against them */
1935 used = __ceph_caps_used(ci);
1936
1937 /*
1938 * "issued" represents the current caps that the MDS wants us to have.
1939 * "implemented" is the set that we have been granted, and includes the
1940 * ones that have not yet been returned to the MDS (the "revoking" set,
1941 * usually because they have outstanding references).
1942 */
1943 issued = __ceph_caps_issued(ci, &implemented);
1944 revoking = implemented & ~issued;
1945
1946 want = file_wanted;
1947
1948 /* The ones we currently want to retain (may be adjusted below) */
1949 retain = file_wanted | used | CEPH_CAP_PIN;
1950 if (!mdsc->stopping && inode->i_nlink > 0) {
1951 if (file_wanted) {
1952 retain |= CEPH_CAP_ANY; /* be greedy */
1953 } else if (S_ISDIR(inode->i_mode) &&
1954 (issued & CEPH_CAP_FILE_SHARED) &&
1955 __ceph_dir_is_complete(ci)) {
1956 /*
1957 * If a directory is complete, we want to keep
1958 * the exclusive cap. So that MDS does not end up
1959 * revoking the shared cap on every create/unlink
1960 * operation.
1961 */
1962 if (IS_RDONLY(inode)) {
1963 want = CEPH_CAP_ANY_SHARED;
1964 } else {
1965 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1966 }
1967 retain |= want;
1968 } else {
1969
1970 retain |= CEPH_CAP_ANY_SHARED;
1971 /*
1972 * keep RD only if we didn't have the file open RW,
1973 * because then the mds would revoke it anyway to
1974 * journal max_size=0.
1975 */
1976 if (ci->i_max_size == 0)
1977 retain |= CEPH_CAP_ANY_RD;
1978 }
1979 }
1980
1981 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1982 " issued %s revoking %s retain %s %s%s\n", inode,
1983 ceph_cap_string(file_wanted),
1984 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1985 ceph_cap_string(ci->i_flushing_caps),
1986 ceph_cap_string(issued), ceph_cap_string(revoking),
1987 ceph_cap_string(retain),
1988 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1989 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1990
1991 /*
1992 * If we no longer need to hold onto old our caps, and we may
1993 * have cached pages, but don't want them, then try to invalidate.
1994 * If we fail, it's because pages are locked.... try again later.
1995 */
1996 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1997 S_ISREG(inode->i_mode) &&
1998 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1999 inode->i_data.nrpages && /* have cached pages */
2000 (revoking & (CEPH_CAP_FILE_CACHE|
2001 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2002 !tried_invalidate) {
2003 dout("check_caps trying to invalidate on %p\n", inode);
2004 if (try_nonblocking_invalidate(inode) < 0) {
2005 dout("check_caps queuing invalidate\n");
2006 queue_invalidate = true;
2007 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2008 }
2009 tried_invalidate = true;
2010 goto retry;
2011 }
2012
2013 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2014 int mflags = 0;
2015 struct cap_msg_args arg;
2016
2017 cap = rb_entry(p, struct ceph_cap, ci_node);
2018
2019 /* avoid looping forever */
2020 if (mds >= cap->mds ||
2021 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2022 continue;
2023
2024 /*
2025 * If we have an auth cap, we don't need to consider any
2026 * overlapping caps as used.
2027 */
2028 cap_used = used;
2029 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2030 cap_used &= ~ci->i_auth_cap->issued;
2031
2032 revoking = cap->implemented & ~cap->issued;
2033 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2034 cap->mds, cap, ceph_cap_string(cap_used),
2035 ceph_cap_string(cap->issued),
2036 ceph_cap_string(cap->implemented),
2037 ceph_cap_string(revoking));
2038
2039 if (cap == ci->i_auth_cap &&
2040 (cap->issued & CEPH_CAP_FILE_WR)) {
2041 /* request larger max_size from MDS? */
2042 if (ci->i_wanted_max_size > ci->i_max_size &&
2043 ci->i_wanted_max_size > ci->i_requested_max_size) {
2044 dout("requesting new max_size\n");
2045 goto ack;
2046 }
2047
2048 /* approaching file_max? */
2049 if (__ceph_should_report_size(ci)) {
2050 dout("i_size approaching max_size\n");
2051 goto ack;
2052 }
2053 }
2054 /* flush anything dirty? */
2055 if (cap == ci->i_auth_cap) {
2056 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2057 dout("flushing dirty caps\n");
2058 goto ack;
2059 }
2060 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2061 dout("flushing snap caps\n");
2062 goto ack;
2063 }
2064 }
2065
2066 /* completed revocation? going down and there are no caps? */
2067 if (revoking && (revoking & cap_used) == 0) {
2068 dout("completed revocation of %s\n",
2069 ceph_cap_string(cap->implemented & ~cap->issued));
2070 goto ack;
2071 }
2072
2073 /* want more caps from mds? */
2074 if (want & ~cap->mds_wanted) {
2075 if (want & ~(cap->mds_wanted | cap->issued))
2076 goto ack;
2077 if (!__cap_is_valid(cap))
2078 goto ack;
2079 }
2080
2081 /* things we might delay */
2082 if ((cap->issued & ~retain) == 0)
2083 continue; /* nope, all good */
2084
2085ack:
2086 ceph_put_mds_session(session);
2087 session = ceph_get_mds_session(cap->session);
2088
2089 /* kick flushing and flush snaps before sending normal
2090 * cap message */
2091 if (cap == ci->i_auth_cap &&
2092 (ci->i_ceph_flags &
2093 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2094 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2095 __kick_flushing_caps(mdsc, session, ci, 0);
2096 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2097 __ceph_flush_snaps(ci, session);
2098
2099 goto retry;
2100 }
2101
2102 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2103 flushing = ci->i_dirty_caps;
2104 flush_tid = __mark_caps_flushing(inode, session, false,
2105 &oldest_flush_tid);
2106 if (flags & CHECK_CAPS_FLUSH &&
2107 list_empty(&session->s_cap_dirty))
2108 mflags |= CEPH_CLIENT_CAPS_SYNC;
2109 } else {
2110 flushing = 0;
2111 flush_tid = 0;
2112 spin_lock(&mdsc->cap_dirty_lock);
2113 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2114 spin_unlock(&mdsc->cap_dirty_lock);
2115 }
2116
2117 mds = cap->mds; /* remember mds, so we don't repeat */
2118
2119 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2120 want, retain, flushing, flush_tid, oldest_flush_tid);
2121
2122 spin_unlock(&ci->i_ceph_lock);
2123 __send_cap(&arg, ci);
2124 spin_lock(&ci->i_ceph_lock);
2125
2126 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2127 }
2128
2129 /* periodically re-calculate caps wanted by open files */
2130 if (__ceph_is_any_real_caps(ci) &&
2131 list_empty(&ci->i_cap_delay_list) &&
2132 (file_wanted & ~CEPH_CAP_PIN) &&
2133 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2134 __cap_delay_requeue(mdsc, ci);
2135 }
2136
2137 spin_unlock(&ci->i_ceph_lock);
2138
2139 ceph_put_mds_session(session);
2140 if (queue_invalidate)
2141 ceph_queue_invalidate(inode);
2142}
2143
2144/*
2145 * Try to flush dirty caps back to the auth mds.
2146 */
2147static int try_flush_caps(struct inode *inode, u64 *ptid)
2148{
2149 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2150 struct ceph_inode_info *ci = ceph_inode(inode);
2151 int flushing = 0;
2152 u64 flush_tid = 0, oldest_flush_tid = 0;
2153
2154 spin_lock(&ci->i_ceph_lock);
2155retry_locked:
2156 if (ci->i_dirty_caps && ci->i_auth_cap) {
2157 struct ceph_cap *cap = ci->i_auth_cap;
2158 struct cap_msg_args arg;
2159 struct ceph_mds_session *session = cap->session;
2160
2161 if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2162 spin_unlock(&ci->i_ceph_lock);
2163 goto out;
2164 }
2165
2166 if (ci->i_ceph_flags &
2167 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2168 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2169 __kick_flushing_caps(mdsc, session, ci, 0);
2170 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2171 __ceph_flush_snaps(ci, session);
2172 goto retry_locked;
2173 }
2174
2175 flushing = ci->i_dirty_caps;
2176 flush_tid = __mark_caps_flushing(inode, session, true,
2177 &oldest_flush_tid);
2178
2179 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2180 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2181 (cap->issued | cap->implemented),
2182 flushing, flush_tid, oldest_flush_tid);
2183 spin_unlock(&ci->i_ceph_lock);
2184
2185 __send_cap(&arg, ci);
2186 } else {
2187 if (!list_empty(&ci->i_cap_flush_list)) {
2188 struct ceph_cap_flush *cf =
2189 list_last_entry(&ci->i_cap_flush_list,
2190 struct ceph_cap_flush, i_list);
2191 cf->wake = true;
2192 flush_tid = cf->tid;
2193 }
2194 flushing = ci->i_flushing_caps;
2195 spin_unlock(&ci->i_ceph_lock);
2196 }
2197out:
2198 *ptid = flush_tid;
2199 return flushing;
2200}
2201
2202/*
2203 * Return true if we've flushed caps through the given flush_tid.
2204 */
2205static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2206{
2207 struct ceph_inode_info *ci = ceph_inode(inode);
2208 int ret = 1;
2209
2210 spin_lock(&ci->i_ceph_lock);
2211 if (!list_empty(&ci->i_cap_flush_list)) {
2212 struct ceph_cap_flush * cf =
2213 list_first_entry(&ci->i_cap_flush_list,
2214 struct ceph_cap_flush, i_list);
2215 if (cf->tid <= flush_tid)
2216 ret = 0;
2217 }
2218 spin_unlock(&ci->i_ceph_lock);
2219 return ret;
2220}
2221
2222/*
2223 * wait for any unsafe requests to complete.
2224 */
2225static int unsafe_request_wait(struct inode *inode)
2226{
2227 struct ceph_inode_info *ci = ceph_inode(inode);
2228 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2229 int ret, err = 0;
2230
2231 spin_lock(&ci->i_unsafe_lock);
2232 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2233 req1 = list_last_entry(&ci->i_unsafe_dirops,
2234 struct ceph_mds_request,
2235 r_unsafe_dir_item);
2236 ceph_mdsc_get_request(req1);
2237 }
2238 if (!list_empty(&ci->i_unsafe_iops)) {
2239 req2 = list_last_entry(&ci->i_unsafe_iops,
2240 struct ceph_mds_request,
2241 r_unsafe_target_item);
2242 ceph_mdsc_get_request(req2);
2243 }
2244 spin_unlock(&ci->i_unsafe_lock);
2245
2246 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2247 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2248 if (req1) {
2249 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2250 ceph_timeout_jiffies(req1->r_timeout));
2251 if (ret)
2252 err = -EIO;
2253 ceph_mdsc_put_request(req1);
2254 }
2255 if (req2) {
2256 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2257 ceph_timeout_jiffies(req2->r_timeout));
2258 if (ret)
2259 err = -EIO;
2260 ceph_mdsc_put_request(req2);
2261 }
2262 return err;
2263}
2264
2265int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2266{
2267 struct inode *inode = file->f_mapping->host;
2268 struct ceph_inode_info *ci = ceph_inode(inode);
2269 u64 flush_tid;
2270 int ret, err;
2271 int dirty;
2272
2273 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2274
2275 ret = file_write_and_wait_range(file, start, end);
2276 if (datasync)
2277 goto out;
2278
2279 ret = ceph_wait_on_async_create(inode);
2280 if (ret)
2281 goto out;
2282
2283 dirty = try_flush_caps(inode, &flush_tid);
2284 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2285
2286 err = unsafe_request_wait(inode);
2287
2288 /*
2289 * only wait on non-file metadata writeback (the mds
2290 * can recover size and mtime, so we don't need to
2291 * wait for that)
2292 */
2293 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2294 err = wait_event_interruptible(ci->i_cap_wq,
2295 caps_are_flushed(inode, flush_tid));
2296 }
2297
2298 if (err < 0)
2299 ret = err;
2300
2301 err = file_check_and_advance_wb_err(file);
2302 if (err < 0)
2303 ret = err;
2304out:
2305 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2306 return ret;
2307}
2308
2309/*
2310 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2311 * queue inode for flush but don't do so immediately, because we can
2312 * get by with fewer MDS messages if we wait for data writeback to
2313 * complete first.
2314 */
2315int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2316{
2317 struct ceph_inode_info *ci = ceph_inode(inode);
2318 u64 flush_tid;
2319 int err = 0;
2320 int dirty;
2321 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2322
2323 dout("write_inode %p wait=%d\n", inode, wait);
2324 if (wait) {
2325 dirty = try_flush_caps(inode, &flush_tid);
2326 if (dirty)
2327 err = wait_event_interruptible(ci->i_cap_wq,
2328 caps_are_flushed(inode, flush_tid));
2329 } else {
2330 struct ceph_mds_client *mdsc =
2331 ceph_sb_to_client(inode->i_sb)->mdsc;
2332
2333 spin_lock(&ci->i_ceph_lock);
2334 if (__ceph_caps_dirty(ci))
2335 __cap_delay_requeue_front(mdsc, ci);
2336 spin_unlock(&ci->i_ceph_lock);
2337 }
2338 return err;
2339}
2340
2341static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2342 struct ceph_mds_session *session,
2343 struct ceph_inode_info *ci,
2344 u64 oldest_flush_tid)
2345 __releases(ci->i_ceph_lock)
2346 __acquires(ci->i_ceph_lock)
2347{
2348 struct inode *inode = &ci->vfs_inode;
2349 struct ceph_cap *cap;
2350 struct ceph_cap_flush *cf;
2351 int ret;
2352 u64 first_tid = 0;
2353 u64 last_snap_flush = 0;
2354
2355 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2356
2357 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2358 if (cf->is_capsnap) {
2359 last_snap_flush = cf->tid;
2360 break;
2361 }
2362 }
2363
2364 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2365 if (cf->tid < first_tid)
2366 continue;
2367
2368 cap = ci->i_auth_cap;
2369 if (!(cap && cap->session == session)) {
2370 pr_err("%p auth cap %p not mds%d ???\n",
2371 inode, cap, session->s_mds);
2372 break;
2373 }
2374
2375 first_tid = cf->tid + 1;
2376
2377 if (!cf->is_capsnap) {
2378 struct cap_msg_args arg;
2379
2380 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2381 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2382 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2383 (cf->tid < last_snap_flush ?
2384 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2385 __ceph_caps_used(ci),
2386 __ceph_caps_wanted(ci),
2387 (cap->issued | cap->implemented),
2388 cf->caps, cf->tid, oldest_flush_tid);
2389 spin_unlock(&ci->i_ceph_lock);
2390 __send_cap(&arg, ci);
2391 } else {
2392 struct ceph_cap_snap *capsnap =
2393 container_of(cf, struct ceph_cap_snap,
2394 cap_flush);
2395 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2396 inode, capsnap, cf->tid,
2397 ceph_cap_string(capsnap->dirty));
2398
2399 refcount_inc(&capsnap->nref);
2400 spin_unlock(&ci->i_ceph_lock);
2401
2402 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2403 oldest_flush_tid);
2404 if (ret < 0) {
2405 pr_err("kick_flushing_caps: error sending "
2406 "cap flushsnap, ino (%llx.%llx) "
2407 "tid %llu follows %llu\n",
2408 ceph_vinop(inode), cf->tid,
2409 capsnap->follows);
2410 }
2411
2412 ceph_put_cap_snap(capsnap);
2413 }
2414
2415 spin_lock(&ci->i_ceph_lock);
2416 }
2417}
2418
2419void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2420 struct ceph_mds_session *session)
2421{
2422 struct ceph_inode_info *ci;
2423 struct ceph_cap *cap;
2424 u64 oldest_flush_tid;
2425
2426 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2427
2428 spin_lock(&mdsc->cap_dirty_lock);
2429 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2430 spin_unlock(&mdsc->cap_dirty_lock);
2431
2432 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2433 spin_lock(&ci->i_ceph_lock);
2434 cap = ci->i_auth_cap;
2435 if (!(cap && cap->session == session)) {
2436 pr_err("%p auth cap %p not mds%d ???\n",
2437 &ci->vfs_inode, cap, session->s_mds);
2438 spin_unlock(&ci->i_ceph_lock);
2439 continue;
2440 }
2441
2442
2443 /*
2444 * if flushing caps were revoked, we re-send the cap flush
2445 * in client reconnect stage. This guarantees MDS * processes
2446 * the cap flush message before issuing the flushing caps to
2447 * other client.
2448 */
2449 if ((cap->issued & ci->i_flushing_caps) !=
2450 ci->i_flushing_caps) {
2451 /* encode_caps_cb() also will reset these sequence
2452 * numbers. make sure sequence numbers in cap flush
2453 * message match later reconnect message */
2454 cap->seq = 0;
2455 cap->issue_seq = 0;
2456 cap->mseq = 0;
2457 __kick_flushing_caps(mdsc, session, ci,
2458 oldest_flush_tid);
2459 } else {
2460 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2461 }
2462
2463 spin_unlock(&ci->i_ceph_lock);
2464 }
2465}
2466
2467void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2468 struct ceph_mds_session *session)
2469{
2470 struct ceph_inode_info *ci;
2471 struct ceph_cap *cap;
2472 u64 oldest_flush_tid;
2473
2474 lockdep_assert_held(&session->s_mutex);
2475
2476 dout("kick_flushing_caps mds%d\n", session->s_mds);
2477
2478 spin_lock(&mdsc->cap_dirty_lock);
2479 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2480 spin_unlock(&mdsc->cap_dirty_lock);
2481
2482 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2483 spin_lock(&ci->i_ceph_lock);
2484 cap = ci->i_auth_cap;
2485 if (!(cap && cap->session == session)) {
2486 pr_err("%p auth cap %p not mds%d ???\n",
2487 &ci->vfs_inode, cap, session->s_mds);
2488 spin_unlock(&ci->i_ceph_lock);
2489 continue;
2490 }
2491 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2492 __kick_flushing_caps(mdsc, session, ci,
2493 oldest_flush_tid);
2494 }
2495 spin_unlock(&ci->i_ceph_lock);
2496 }
2497}
2498
2499void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2500 struct ceph_inode_info *ci)
2501{
2502 struct ceph_mds_client *mdsc = session->s_mdsc;
2503 struct ceph_cap *cap = ci->i_auth_cap;
2504
2505 lockdep_assert_held(&ci->i_ceph_lock);
2506
2507 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2508 ceph_cap_string(ci->i_flushing_caps));
2509
2510 if (!list_empty(&ci->i_cap_flush_list)) {
2511 u64 oldest_flush_tid;
2512 spin_lock(&mdsc->cap_dirty_lock);
2513 list_move_tail(&ci->i_flushing_item,
2514 &cap->session->s_cap_flushing);
2515 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2516 spin_unlock(&mdsc->cap_dirty_lock);
2517
2518 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2519 }
2520}
2521
2522
2523/*
2524 * Take references to capabilities we hold, so that we don't release
2525 * them to the MDS prematurely.
2526 */
2527void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2528 bool snap_rwsem_locked)
2529{
2530 lockdep_assert_held(&ci->i_ceph_lock);
2531
2532 if (got & CEPH_CAP_PIN)
2533 ci->i_pin_ref++;
2534 if (got & CEPH_CAP_FILE_RD)
2535 ci->i_rd_ref++;
2536 if (got & CEPH_CAP_FILE_CACHE)
2537 ci->i_rdcache_ref++;
2538 if (got & CEPH_CAP_FILE_EXCL)
2539 ci->i_fx_ref++;
2540 if (got & CEPH_CAP_FILE_WR) {
2541 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2542 BUG_ON(!snap_rwsem_locked);
2543 ci->i_head_snapc = ceph_get_snap_context(
2544 ci->i_snap_realm->cached_context);
2545 }
2546 ci->i_wr_ref++;
2547 }
2548 if (got & CEPH_CAP_FILE_BUFFER) {
2549 if (ci->i_wb_ref == 0)
2550 ihold(&ci->vfs_inode);
2551 ci->i_wb_ref++;
2552 dout("%s %p wb %d -> %d (?)\n", __func__,
2553 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2554 }
2555}
2556
2557/*
2558 * Try to grab cap references. Specify those refs we @want, and the
2559 * minimal set we @need. Also include the larger offset we are writing
2560 * to (when applicable), and check against max_size here as well.
2561 * Note that caller is responsible for ensuring max_size increases are
2562 * requested from the MDS.
2563 *
2564 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2565 * or a negative error code. There are 3 speical error codes:
2566 * -EAGAIN: need to sleep but non-blocking is specified
2567 * -EFBIG: ask caller to call check_max_size() and try again.
2568 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2569 */
2570enum {
2571 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2572 NON_BLOCKING = (1 << 8),
2573 CHECK_FILELOCK = (1 << 9),
2574};
2575
2576static int try_get_cap_refs(struct inode *inode, int need, int want,
2577 loff_t endoff, int flags, int *got)
2578{
2579 struct ceph_inode_info *ci = ceph_inode(inode);
2580 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2581 int ret = 0;
2582 int have, implemented;
2583 bool snap_rwsem_locked = false;
2584
2585 dout("get_cap_refs %p need %s want %s\n", inode,
2586 ceph_cap_string(need), ceph_cap_string(want));
2587
2588again:
2589 spin_lock(&ci->i_ceph_lock);
2590
2591 if ((flags & CHECK_FILELOCK) &&
2592 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2593 dout("try_get_cap_refs %p error filelock\n", inode);
2594 ret = -EIO;
2595 goto out_unlock;
2596 }
2597
2598 /* finish pending truncate */
2599 while (ci->i_truncate_pending) {
2600 spin_unlock(&ci->i_ceph_lock);
2601 if (snap_rwsem_locked) {
2602 up_read(&mdsc->snap_rwsem);
2603 snap_rwsem_locked = false;
2604 }
2605 __ceph_do_pending_vmtruncate(inode);
2606 spin_lock(&ci->i_ceph_lock);
2607 }
2608
2609 have = __ceph_caps_issued(ci, &implemented);
2610
2611 if (have & need & CEPH_CAP_FILE_WR) {
2612 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2613 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2614 inode, endoff, ci->i_max_size);
2615 if (endoff > ci->i_requested_max_size)
2616 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2617 goto out_unlock;
2618 }
2619 /*
2620 * If a sync write is in progress, we must wait, so that we
2621 * can get a final snapshot value for size+mtime.
2622 */
2623 if (__ceph_have_pending_cap_snap(ci)) {
2624 dout("get_cap_refs %p cap_snap_pending\n", inode);
2625 goto out_unlock;
2626 }
2627 }
2628
2629 if ((have & need) == need) {
2630 /*
2631 * Look at (implemented & ~have & not) so that we keep waiting
2632 * on transition from wanted -> needed caps. This is needed
2633 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2634 * going before a prior buffered writeback happens.
2635 */
2636 int not = want & ~(have & need);
2637 int revoking = implemented & ~have;
2638 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2639 inode, ceph_cap_string(have), ceph_cap_string(not),
2640 ceph_cap_string(revoking));
2641 if ((revoking & not) == 0) {
2642 if (!snap_rwsem_locked &&
2643 !ci->i_head_snapc &&
2644 (need & CEPH_CAP_FILE_WR)) {
2645 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2646 /*
2647 * we can not call down_read() when
2648 * task isn't in TASK_RUNNING state
2649 */
2650 if (flags & NON_BLOCKING) {
2651 ret = -EAGAIN;
2652 goto out_unlock;
2653 }
2654
2655 spin_unlock(&ci->i_ceph_lock);
2656 down_read(&mdsc->snap_rwsem);
2657 snap_rwsem_locked = true;
2658 goto again;
2659 }
2660 snap_rwsem_locked = true;
2661 }
2662 if ((have & want) == want)
2663 *got = need | want;
2664 else
2665 *got = need;
2666 ceph_take_cap_refs(ci, *got, true);
2667 ret = 1;
2668 }
2669 } else {
2670 int session_readonly = false;
2671 int mds_wanted;
2672 if (ci->i_auth_cap &&
2673 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2674 struct ceph_mds_session *s = ci->i_auth_cap->session;
2675 spin_lock(&s->s_cap_lock);
2676 session_readonly = s->s_readonly;
2677 spin_unlock(&s->s_cap_lock);
2678 }
2679 if (session_readonly) {
2680 dout("get_cap_refs %p need %s but mds%d readonly\n",
2681 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2682 ret = -EROFS;
2683 goto out_unlock;
2684 }
2685
2686 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
2687 dout("get_cap_refs %p forced umount\n", inode);
2688 ret = -EIO;
2689 goto out_unlock;
2690 }
2691 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2692 if (need & ~mds_wanted) {
2693 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2694 inode, ceph_cap_string(need),
2695 ceph_cap_string(mds_wanted));
2696 ret = -ESTALE;
2697 goto out_unlock;
2698 }
2699
2700 dout("get_cap_refs %p have %s need %s\n", inode,
2701 ceph_cap_string(have), ceph_cap_string(need));
2702 }
2703out_unlock:
2704
2705 __ceph_touch_fmode(ci, mdsc, flags);
2706
2707 spin_unlock(&ci->i_ceph_lock);
2708 if (snap_rwsem_locked)
2709 up_read(&mdsc->snap_rwsem);
2710
2711 if (!ret)
2712 ceph_update_cap_mis(&mdsc->metric);
2713 else if (ret == 1)
2714 ceph_update_cap_hit(&mdsc->metric);
2715
2716 dout("get_cap_refs %p ret %d got %s\n", inode,
2717 ret, ceph_cap_string(*got));
2718 return ret;
2719}
2720
2721/*
2722 * Check the offset we are writing up to against our current
2723 * max_size. If necessary, tell the MDS we want to write to
2724 * a larger offset.
2725 */
2726static void check_max_size(struct inode *inode, loff_t endoff)
2727{
2728 struct ceph_inode_info *ci = ceph_inode(inode);
2729 int check = 0;
2730
2731 /* do we need to explicitly request a larger max_size? */
2732 spin_lock(&ci->i_ceph_lock);
2733 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2734 dout("write %p at large endoff %llu, req max_size\n",
2735 inode, endoff);
2736 ci->i_wanted_max_size = endoff;
2737 }
2738 /* duplicate ceph_check_caps()'s logic */
2739 if (ci->i_auth_cap &&
2740 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2741 ci->i_wanted_max_size > ci->i_max_size &&
2742 ci->i_wanted_max_size > ci->i_requested_max_size)
2743 check = 1;
2744 spin_unlock(&ci->i_ceph_lock);
2745 if (check)
2746 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2747}
2748
2749static inline int get_used_fmode(int caps)
2750{
2751 int fmode = 0;
2752 if (caps & CEPH_CAP_FILE_RD)
2753 fmode |= CEPH_FILE_MODE_RD;
2754 if (caps & CEPH_CAP_FILE_WR)
2755 fmode |= CEPH_FILE_MODE_WR;
2756 return fmode;
2757}
2758
2759int ceph_try_get_caps(struct inode *inode, int need, int want,
2760 bool nonblock, int *got)
2761{
2762 int ret, flags;
2763
2764 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2765 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2766 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2767 CEPH_CAP_ANY_DIR_OPS));
2768 if (need) {
2769 ret = ceph_pool_perm_check(inode, need);
2770 if (ret < 0)
2771 return ret;
2772 }
2773
2774 flags = get_used_fmode(need | want);
2775 if (nonblock)
2776 flags |= NON_BLOCKING;
2777
2778 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2779 /* three special error codes */
2780 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2781 ret = 0;
2782 return ret;
2783}
2784
2785/*
2786 * Wait for caps, and take cap references. If we can't get a WR cap
2787 * due to a small max_size, make sure we check_max_size (and possibly
2788 * ask the mds) so we don't get hung up indefinitely.
2789 */
2790int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2791{
2792 struct ceph_file_info *fi = filp->private_data;
2793 struct inode *inode = file_inode(filp);
2794 struct ceph_inode_info *ci = ceph_inode(inode);
2795 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2796 int ret, _got, flags;
2797
2798 ret = ceph_pool_perm_check(inode, need);
2799 if (ret < 0)
2800 return ret;
2801
2802 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2803 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2804 return -EBADF;
2805
2806 flags = get_used_fmode(need | want);
2807
2808 while (true) {
2809 flags &= CEPH_FILE_MODE_MASK;
2810 if (atomic_read(&fi->num_locks))
2811 flags |= CHECK_FILELOCK;
2812 _got = 0;
2813 ret = try_get_cap_refs(inode, need, want, endoff,
2814 flags, &_got);
2815 WARN_ON_ONCE(ret == -EAGAIN);
2816 if (!ret) {
2817 struct ceph_mds_client *mdsc = fsc->mdsc;
2818 struct cap_wait cw;
2819 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2820
2821 cw.ino = ceph_ino(inode);
2822 cw.tgid = current->tgid;
2823 cw.need = need;
2824 cw.want = want;
2825
2826 spin_lock(&mdsc->caps_list_lock);
2827 list_add(&cw.list, &mdsc->cap_wait_list);
2828 spin_unlock(&mdsc->caps_list_lock);
2829
2830 /* make sure used fmode not timeout */
2831 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2832 add_wait_queue(&ci->i_cap_wq, &wait);
2833
2834 flags |= NON_BLOCKING;
2835 while (!(ret = try_get_cap_refs(inode, need, want,
2836 endoff, flags, &_got))) {
2837 if (signal_pending(current)) {
2838 ret = -ERESTARTSYS;
2839 break;
2840 }
2841 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2842 }
2843
2844 remove_wait_queue(&ci->i_cap_wq, &wait);
2845 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2846
2847 spin_lock(&mdsc->caps_list_lock);
2848 list_del(&cw.list);
2849 spin_unlock(&mdsc->caps_list_lock);
2850
2851 if (ret == -EAGAIN)
2852 continue;
2853 }
2854
2855 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2856 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2857 if (ret >= 0 && _got)
2858 ceph_put_cap_refs(ci, _got);
2859 return -EBADF;
2860 }
2861
2862 if (ret < 0) {
2863 if (ret == -EFBIG || ret == -ESTALE) {
2864 int ret2 = ceph_wait_on_async_create(inode);
2865 if (ret2 < 0)
2866 return ret2;
2867 }
2868 if (ret == -EFBIG) {
2869 check_max_size(inode, endoff);
2870 continue;
2871 }
2872 if (ret == -ESTALE) {
2873 /* session was killed, try renew caps */
2874 ret = ceph_renew_caps(inode, flags);
2875 if (ret == 0)
2876 continue;
2877 }
2878 return ret;
2879 }
2880
2881 if (S_ISREG(ci->vfs_inode.i_mode) &&
2882 ci->i_inline_version != CEPH_INLINE_NONE &&
2883 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2884 i_size_read(inode) > 0) {
2885 struct page *page =
2886 find_get_page(inode->i_mapping, 0);
2887 if (page) {
2888 bool uptodate = PageUptodate(page);
2889
2890 put_page(page);
2891 if (uptodate)
2892 break;
2893 }
2894 /*
2895 * drop cap refs first because getattr while
2896 * holding * caps refs can cause deadlock.
2897 */
2898 ceph_put_cap_refs(ci, _got);
2899 _got = 0;
2900
2901 /*
2902 * getattr request will bring inline data into
2903 * page cache
2904 */
2905 ret = __ceph_do_getattr(inode, NULL,
2906 CEPH_STAT_CAP_INLINE_DATA,
2907 true);
2908 if (ret < 0)
2909 return ret;
2910 continue;
2911 }
2912 break;
2913 }
2914 *got = _got;
2915 return 0;
2916}
2917
2918/*
2919 * Take cap refs. Caller must already know we hold at least one ref
2920 * on the caps in question or we don't know this is safe.
2921 */
2922void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2923{
2924 spin_lock(&ci->i_ceph_lock);
2925 ceph_take_cap_refs(ci, caps, false);
2926 spin_unlock(&ci->i_ceph_lock);
2927}
2928
2929
2930/*
2931 * drop cap_snap that is not associated with any snapshot.
2932 * we don't need to send FLUSHSNAP message for it.
2933 */
2934static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2935 struct ceph_cap_snap *capsnap)
2936{
2937 if (!capsnap->need_flush &&
2938 !capsnap->writing && !capsnap->dirty_pages) {
2939 dout("dropping cap_snap %p follows %llu\n",
2940 capsnap, capsnap->follows);
2941 BUG_ON(capsnap->cap_flush.tid > 0);
2942 ceph_put_snap_context(capsnap->context);
2943 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2944 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2945
2946 list_del(&capsnap->ci_item);
2947 ceph_put_cap_snap(capsnap);
2948 return 1;
2949 }
2950 return 0;
2951}
2952
2953enum put_cap_refs_mode {
2954 PUT_CAP_REFS_SYNC = 0,
2955 PUT_CAP_REFS_NO_CHECK,
2956 PUT_CAP_REFS_ASYNC,
2957};
2958
2959/*
2960 * Release cap refs.
2961 *
2962 * If we released the last ref on any given cap, call ceph_check_caps
2963 * to release (or schedule a release).
2964 *
2965 * If we are releasing a WR cap (from a sync write), finalize any affected
2966 * cap_snap, and wake up any waiters.
2967 */
2968static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
2969 enum put_cap_refs_mode mode)
2970{
2971 struct inode *inode = &ci->vfs_inode;
2972 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2973 bool check_flushsnaps = false;
2974
2975 spin_lock(&ci->i_ceph_lock);
2976 if (had & CEPH_CAP_PIN)
2977 --ci->i_pin_ref;
2978 if (had & CEPH_CAP_FILE_RD)
2979 if (--ci->i_rd_ref == 0)
2980 last++;
2981 if (had & CEPH_CAP_FILE_CACHE)
2982 if (--ci->i_rdcache_ref == 0)
2983 last++;
2984 if (had & CEPH_CAP_FILE_EXCL)
2985 if (--ci->i_fx_ref == 0)
2986 last++;
2987 if (had & CEPH_CAP_FILE_BUFFER) {
2988 if (--ci->i_wb_ref == 0) {
2989 last++;
2990 /* put the ref held by ceph_take_cap_refs() */
2991 put++;
2992 check_flushsnaps = true;
2993 }
2994 dout("put_cap_refs %p wb %d -> %d (?)\n",
2995 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2996 }
2997 if (had & CEPH_CAP_FILE_WR) {
2998 if (--ci->i_wr_ref == 0) {
2999 last++;
3000 check_flushsnaps = true;
3001 if (ci->i_wrbuffer_ref_head == 0 &&
3002 ci->i_dirty_caps == 0 &&
3003 ci->i_flushing_caps == 0) {
3004 BUG_ON(!ci->i_head_snapc);
3005 ceph_put_snap_context(ci->i_head_snapc);
3006 ci->i_head_snapc = NULL;
3007 }
3008 /* see comment in __ceph_remove_cap() */
3009 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3010 drop_inode_snap_realm(ci);
3011 }
3012 }
3013 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3014 struct ceph_cap_snap *capsnap =
3015 list_last_entry(&ci->i_cap_snaps,
3016 struct ceph_cap_snap,
3017 ci_item);
3018
3019 capsnap->writing = 0;
3020 if (ceph_try_drop_cap_snap(ci, capsnap))
3021 /* put the ref held by ceph_queue_cap_snap() */
3022 put++;
3023 else if (__ceph_finish_cap_snap(ci, capsnap))
3024 flushsnaps = 1;
3025 wake = 1;
3026 }
3027 spin_unlock(&ci->i_ceph_lock);
3028
3029 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3030 last ? " last" : "", put ? " put" : "");
3031
3032 switch (mode) {
3033 case PUT_CAP_REFS_SYNC:
3034 if (last)
3035 ceph_check_caps(ci, 0, NULL);
3036 else if (flushsnaps)
3037 ceph_flush_snaps(ci, NULL);
3038 break;
3039 case PUT_CAP_REFS_ASYNC:
3040 if (last)
3041 ceph_queue_check_caps(inode);
3042 else if (flushsnaps)
3043 ceph_queue_flush_snaps(inode);
3044 break;
3045 default:
3046 break;
3047 }
3048 if (wake)
3049 wake_up_all(&ci->i_cap_wq);
3050 while (put-- > 0)
3051 iput(inode);
3052}
3053
3054void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3055{
3056 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3057}
3058
3059void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3060{
3061 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3062}
3063
3064void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3065{
3066 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3067}
3068
3069/*
3070 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3071 * context. Adjust per-snap dirty page accounting as appropriate.
3072 * Once all dirty data for a cap_snap is flushed, flush snapped file
3073 * metadata back to the MDS. If we dropped the last ref, call
3074 * ceph_check_caps.
3075 */
3076void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3077 struct ceph_snap_context *snapc)
3078{
3079 struct inode *inode = &ci->vfs_inode;
3080 struct ceph_cap_snap *capsnap = NULL;
3081 int put = 0;
3082 bool last = false;
3083 bool found = false;
3084 bool flush_snaps = false;
3085 bool complete_capsnap = false;
3086
3087 spin_lock(&ci->i_ceph_lock);
3088 ci->i_wrbuffer_ref -= nr;
3089 if (ci->i_wrbuffer_ref == 0) {
3090 last = true;
3091 put++;
3092 }
3093
3094 if (ci->i_head_snapc == snapc) {
3095 ci->i_wrbuffer_ref_head -= nr;
3096 if (ci->i_wrbuffer_ref_head == 0 &&
3097 ci->i_wr_ref == 0 &&
3098 ci->i_dirty_caps == 0 &&
3099 ci->i_flushing_caps == 0) {
3100 BUG_ON(!ci->i_head_snapc);
3101 ceph_put_snap_context(ci->i_head_snapc);
3102 ci->i_head_snapc = NULL;
3103 }
3104 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3105 inode,
3106 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3107 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3108 last ? " LAST" : "");
3109 } else {
3110 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3111 if (capsnap->context == snapc) {
3112 found = true;
3113 break;
3114 }
3115 }
3116
3117 if (!found) {
3118 /*
3119 * The capsnap should already be removed when removing
3120 * auth cap in the case of a forced unmount.
3121 */
3122 WARN_ON_ONCE(ci->i_auth_cap);
3123 goto unlock;
3124 }
3125
3126 capsnap->dirty_pages -= nr;
3127 if (capsnap->dirty_pages == 0) {
3128 complete_capsnap = true;
3129 if (!capsnap->writing) {
3130 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3131 put++;
3132 } else {
3133 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3134 flush_snaps = true;
3135 }
3136 }
3137 }
3138 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3139 " snap %lld %d/%d -> %d/%d %s%s\n",
3140 inode, capsnap, capsnap->context->seq,
3141 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3142 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3143 last ? " (wrbuffer last)" : "",
3144 complete_capsnap ? " (complete capsnap)" : "");
3145 }
3146
3147unlock:
3148 spin_unlock(&ci->i_ceph_lock);
3149
3150 if (last) {
3151 ceph_check_caps(ci, 0, NULL);
3152 } else if (flush_snaps) {
3153 ceph_flush_snaps(ci, NULL);
3154 }
3155 if (complete_capsnap)
3156 wake_up_all(&ci->i_cap_wq);
3157 while (put-- > 0) {
3158 iput(inode);
3159 }
3160}
3161
3162/*
3163 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3164 */
3165static void invalidate_aliases(struct inode *inode)
3166{
3167 struct dentry *dn, *prev = NULL;
3168
3169 dout("invalidate_aliases inode %p\n", inode);
3170 d_prune_aliases(inode);
3171 /*
3172 * For non-directory inode, d_find_alias() only returns
3173 * hashed dentry. After calling d_invalidate(), the
3174 * dentry becomes unhashed.
3175 *
3176 * For directory inode, d_find_alias() can return
3177 * unhashed dentry. But directory inode should have
3178 * one alias at most.
3179 */
3180 while ((dn = d_find_alias(inode))) {
3181 if (dn == prev) {
3182 dput(dn);
3183 break;
3184 }
3185 d_invalidate(dn);
3186 if (prev)
3187 dput(prev);
3188 prev = dn;
3189 }
3190 if (prev)
3191 dput(prev);
3192}
3193
3194struct cap_extra_info {
3195 struct ceph_string *pool_ns;
3196 /* inline data */
3197 u64 inline_version;
3198 void *inline_data;
3199 u32 inline_len;
3200 /* dirstat */
3201 bool dirstat_valid;
3202 u64 nfiles;
3203 u64 nsubdirs;
3204 u64 change_attr;
3205 /* currently issued */
3206 int issued;
3207 struct timespec64 btime;
3208};
3209
3210/*
3211 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3212 * actually be a revocation if it specifies a smaller cap set.)
3213 *
3214 * caller holds s_mutex and i_ceph_lock, we drop both.
3215 */
3216static void handle_cap_grant(struct inode *inode,
3217 struct ceph_mds_session *session,
3218 struct ceph_cap *cap,
3219 struct ceph_mds_caps *grant,
3220 struct ceph_buffer *xattr_buf,
3221 struct cap_extra_info *extra_info)
3222 __releases(ci->i_ceph_lock)
3223 __releases(session->s_mdsc->snap_rwsem)
3224{
3225 struct ceph_inode_info *ci = ceph_inode(inode);
3226 int seq = le32_to_cpu(grant->seq);
3227 int newcaps = le32_to_cpu(grant->caps);
3228 int used, wanted, dirty;
3229 u64 size = le64_to_cpu(grant->size);
3230 u64 max_size = le64_to_cpu(grant->max_size);
3231 unsigned char check_caps = 0;
3232 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3233 bool wake = false;
3234 bool writeback = false;
3235 bool queue_trunc = false;
3236 bool queue_invalidate = false;
3237 bool deleted_inode = false;
3238 bool fill_inline = false;
3239
3240 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3241 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3242 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3243 i_size_read(inode));
3244
3245
3246 /*
3247 * If CACHE is being revoked, and we have no dirty buffers,
3248 * try to invalidate (once). (If there are dirty buffers, we
3249 * will invalidate _after_ writeback.)
3250 */
3251 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3252 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3253 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3254 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3255 if (try_nonblocking_invalidate(inode)) {
3256 /* there were locked pages.. invalidate later
3257 in a separate thread. */
3258 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3259 queue_invalidate = true;
3260 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3261 }
3262 }
3263 }
3264
3265 if (was_stale)
3266 cap->issued = cap->implemented = CEPH_CAP_PIN;
3267
3268 /*
3269 * auth mds of the inode changed. we received the cap export message,
3270 * but still haven't received the cap import message. handle_cap_export
3271 * updated the new auth MDS' cap.
3272 *
3273 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3274 * that was sent before the cap import message. So don't remove caps.
3275 */
3276 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3277 WARN_ON(cap != ci->i_auth_cap);
3278 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3279 seq = cap->seq;
3280 newcaps |= cap->issued;
3281 }
3282
3283 /* side effects now are allowed */
3284 cap->cap_gen = atomic_read(&session->s_cap_gen);
3285 cap->seq = seq;
3286
3287 __check_cap_issue(ci, cap, newcaps);
3288
3289 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3290
3291 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3292 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3293 umode_t mode = le32_to_cpu(grant->mode);
3294
3295 if (inode_wrong_type(inode, mode))
3296 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3297 ceph_vinop(inode), inode->i_mode, mode);
3298 else
3299 inode->i_mode = mode;
3300 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3301 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3302 ci->i_btime = extra_info->btime;
3303 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3304 from_kuid(&init_user_ns, inode->i_uid),
3305 from_kgid(&init_user_ns, inode->i_gid));
3306 }
3307
3308 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3309 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3310 set_nlink(inode, le32_to_cpu(grant->nlink));
3311 if (inode->i_nlink == 0 &&
3312 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3313 deleted_inode = true;
3314 }
3315
3316 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3317 grant->xattr_len) {
3318 int len = le32_to_cpu(grant->xattr_len);
3319 u64 version = le64_to_cpu(grant->xattr_version);
3320
3321 if (version > ci->i_xattrs.version) {
3322 dout(" got new xattrs v%llu on %p len %d\n",
3323 version, inode, len);
3324 if (ci->i_xattrs.blob)
3325 ceph_buffer_put(ci->i_xattrs.blob);
3326 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3327 ci->i_xattrs.version = version;
3328 ceph_forget_all_cached_acls(inode);
3329 ceph_security_invalidate_secctx(inode);
3330 }
3331 }
3332
3333 if (newcaps & CEPH_CAP_ANY_RD) {
3334 struct timespec64 mtime, atime, ctime;
3335 /* ctime/mtime/atime? */
3336 ceph_decode_timespec64(&mtime, &grant->mtime);
3337 ceph_decode_timespec64(&atime, &grant->atime);
3338 ceph_decode_timespec64(&ctime, &grant->ctime);
3339 ceph_fill_file_time(inode, extra_info->issued,
3340 le32_to_cpu(grant->time_warp_seq),
3341 &ctime, &mtime, &atime);
3342 }
3343
3344 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3345 ci->i_files = extra_info->nfiles;
3346 ci->i_subdirs = extra_info->nsubdirs;
3347 }
3348
3349 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3350 /* file layout may have changed */
3351 s64 old_pool = ci->i_layout.pool_id;
3352 struct ceph_string *old_ns;
3353
3354 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3355 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3356 lockdep_is_held(&ci->i_ceph_lock));
3357 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3358
3359 if (ci->i_layout.pool_id != old_pool ||
3360 extra_info->pool_ns != old_ns)
3361 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3362
3363 extra_info->pool_ns = old_ns;
3364
3365 /* size/truncate_seq? */
3366 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3367 le32_to_cpu(grant->truncate_seq),
3368 le64_to_cpu(grant->truncate_size),
3369 size);
3370 }
3371
3372 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3373 if (max_size != ci->i_max_size) {
3374 dout("max_size %lld -> %llu\n",
3375 ci->i_max_size, max_size);
3376 ci->i_max_size = max_size;
3377 if (max_size >= ci->i_wanted_max_size) {
3378 ci->i_wanted_max_size = 0; /* reset */
3379 ci->i_requested_max_size = 0;
3380 }
3381 wake = true;
3382 }
3383 }
3384
3385 /* check cap bits */
3386 wanted = __ceph_caps_wanted(ci);
3387 used = __ceph_caps_used(ci);
3388 dirty = __ceph_caps_dirty(ci);
3389 dout(" my wanted = %s, used = %s, dirty %s\n",
3390 ceph_cap_string(wanted),
3391 ceph_cap_string(used),
3392 ceph_cap_string(dirty));
3393
3394 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3395 (wanted & ~(cap->mds_wanted | newcaps))) {
3396 /*
3397 * If mds is importing cap, prior cap messages that update
3398 * 'wanted' may get dropped by mds (migrate seq mismatch).
3399 *
3400 * We don't send cap message to update 'wanted' if what we
3401 * want are already issued. If mds revokes caps, cap message
3402 * that releases caps also tells mds what we want. But if
3403 * caps got revoked by mds forcedly (session stale). We may
3404 * haven't told mds what we want.
3405 */
3406 check_caps = 1;
3407 }
3408
3409 /* revocation, grant, or no-op? */
3410 if (cap->issued & ~newcaps) {
3411 int revoking = cap->issued & ~newcaps;
3412
3413 dout("revocation: %s -> %s (revoking %s)\n",
3414 ceph_cap_string(cap->issued),
3415 ceph_cap_string(newcaps),
3416 ceph_cap_string(revoking));
3417 if (S_ISREG(inode->i_mode) &&
3418 (revoking & used & CEPH_CAP_FILE_BUFFER))
3419 writeback = true; /* initiate writeback; will delay ack */
3420 else if (queue_invalidate &&
3421 revoking == CEPH_CAP_FILE_CACHE &&
3422 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3423 ; /* do nothing yet, invalidation will be queued */
3424 else if (cap == ci->i_auth_cap)
3425 check_caps = 1; /* check auth cap only */
3426 else
3427 check_caps = 2; /* check all caps */
3428 cap->issued = newcaps;
3429 cap->implemented |= newcaps;
3430 } else if (cap->issued == newcaps) {
3431 dout("caps unchanged: %s -> %s\n",
3432 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3433 } else {
3434 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3435 ceph_cap_string(newcaps));
3436 /* non-auth MDS is revoking the newly grant caps ? */
3437 if (cap == ci->i_auth_cap &&
3438 __ceph_caps_revoking_other(ci, cap, newcaps))
3439 check_caps = 2;
3440
3441 cap->issued = newcaps;
3442 cap->implemented |= newcaps; /* add bits only, to
3443 * avoid stepping on a
3444 * pending revocation */
3445 wake = true;
3446 }
3447 BUG_ON(cap->issued & ~cap->implemented);
3448
3449 if (extra_info->inline_version > 0 &&
3450 extra_info->inline_version >= ci->i_inline_version) {
3451 ci->i_inline_version = extra_info->inline_version;
3452 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3453 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3454 fill_inline = true;
3455 }
3456
3457 if (ci->i_auth_cap == cap &&
3458 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3459 if (newcaps & ~extra_info->issued)
3460 wake = true;
3461
3462 if (ci->i_requested_max_size > max_size ||
3463 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3464 /* re-request max_size if necessary */
3465 ci->i_requested_max_size = 0;
3466 wake = true;
3467 }
3468
3469 ceph_kick_flushing_inode_caps(session, ci);
3470 spin_unlock(&ci->i_ceph_lock);
3471 up_read(&session->s_mdsc->snap_rwsem);
3472 } else {
3473 spin_unlock(&ci->i_ceph_lock);
3474 }
3475
3476 if (fill_inline)
3477 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3478 extra_info->inline_len);
3479
3480 if (queue_trunc)
3481 ceph_queue_vmtruncate(inode);
3482
3483 if (writeback)
3484 /*
3485 * queue inode for writeback: we can't actually call
3486 * filemap_write_and_wait, etc. from message handler
3487 * context.
3488 */
3489 ceph_queue_writeback(inode);
3490 if (queue_invalidate)
3491 ceph_queue_invalidate(inode);
3492 if (deleted_inode)
3493 invalidate_aliases(inode);
3494 if (wake)
3495 wake_up_all(&ci->i_cap_wq);
3496
3497 mutex_unlock(&session->s_mutex);
3498 if (check_caps == 1)
3499 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3500 session);
3501 else if (check_caps == 2)
3502 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3503}
3504
3505/*
3506 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3507 * MDS has been safely committed.
3508 */
3509static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3510 struct ceph_mds_caps *m,
3511 struct ceph_mds_session *session,
3512 struct ceph_cap *cap)
3513 __releases(ci->i_ceph_lock)
3514{
3515 struct ceph_inode_info *ci = ceph_inode(inode);
3516 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3517 struct ceph_cap_flush *cf, *tmp_cf;
3518 LIST_HEAD(to_remove);
3519 unsigned seq = le32_to_cpu(m->seq);
3520 int dirty = le32_to_cpu(m->dirty);
3521 int cleaned = 0;
3522 bool drop = false;
3523 bool wake_ci = false;
3524 bool wake_mdsc = false;
3525
3526 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3527 /* Is this the one that was flushed? */
3528 if (cf->tid == flush_tid)
3529 cleaned = cf->caps;
3530
3531 /* Is this a capsnap? */
3532 if (cf->is_capsnap)
3533 continue;
3534
3535 if (cf->tid <= flush_tid) {
3536 /*
3537 * An earlier or current tid. The FLUSH_ACK should
3538 * represent a superset of this flush's caps.
3539 */
3540 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3541 list_add_tail(&cf->i_list, &to_remove);
3542 } else {
3543 /*
3544 * This is a later one. Any caps in it are still dirty
3545 * so don't count them as cleaned.
3546 */
3547 cleaned &= ~cf->caps;
3548 if (!cleaned)
3549 break;
3550 }
3551 }
3552
3553 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3554 " flushing %s -> %s\n",
3555 inode, session->s_mds, seq, ceph_cap_string(dirty),
3556 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3557 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3558
3559 if (list_empty(&to_remove) && !cleaned)
3560 goto out;
3561
3562 ci->i_flushing_caps &= ~cleaned;
3563
3564 spin_lock(&mdsc->cap_dirty_lock);
3565
3566 list_for_each_entry(cf, &to_remove, i_list)
3567 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3568
3569 if (ci->i_flushing_caps == 0) {
3570 if (list_empty(&ci->i_cap_flush_list)) {
3571 list_del_init(&ci->i_flushing_item);
3572 if (!list_empty(&session->s_cap_flushing)) {
3573 dout(" mds%d still flushing cap on %p\n",
3574 session->s_mds,
3575 &list_first_entry(&session->s_cap_flushing,
3576 struct ceph_inode_info,
3577 i_flushing_item)->vfs_inode);
3578 }
3579 }
3580 mdsc->num_cap_flushing--;
3581 dout(" inode %p now !flushing\n", inode);
3582
3583 if (ci->i_dirty_caps == 0) {
3584 dout(" inode %p now clean\n", inode);
3585 BUG_ON(!list_empty(&ci->i_dirty_item));
3586 drop = true;
3587 if (ci->i_wr_ref == 0 &&
3588 ci->i_wrbuffer_ref_head == 0) {
3589 BUG_ON(!ci->i_head_snapc);
3590 ceph_put_snap_context(ci->i_head_snapc);
3591 ci->i_head_snapc = NULL;
3592 }
3593 } else {
3594 BUG_ON(list_empty(&ci->i_dirty_item));
3595 }
3596 }
3597 spin_unlock(&mdsc->cap_dirty_lock);
3598
3599out:
3600 spin_unlock(&ci->i_ceph_lock);
3601
3602 while (!list_empty(&to_remove)) {
3603 cf = list_first_entry(&to_remove,
3604 struct ceph_cap_flush, i_list);
3605 list_del_init(&cf->i_list);
3606 if (!cf->is_capsnap)
3607 ceph_free_cap_flush(cf);
3608 }
3609
3610 if (wake_ci)
3611 wake_up_all(&ci->i_cap_wq);
3612 if (wake_mdsc)
3613 wake_up_all(&mdsc->cap_flushing_wq);
3614 if (drop)
3615 iput(inode);
3616}
3617
3618void __ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3619 bool *wake_ci, bool *wake_mdsc)
3620{
3621 struct ceph_inode_info *ci = ceph_inode(inode);
3622 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3623 bool ret;
3624
3625 lockdep_assert_held(&ci->i_ceph_lock);
3626
3627 dout("removing capsnap %p, inode %p ci %p\n", capsnap, inode, ci);
3628
3629 list_del_init(&capsnap->ci_item);
3630 ret = __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3631 if (wake_ci)
3632 *wake_ci = ret;
3633
3634 spin_lock(&mdsc->cap_dirty_lock);
3635 if (list_empty(&ci->i_cap_flush_list))
3636 list_del_init(&ci->i_flushing_item);
3637
3638 ret = __detach_cap_flush_from_mdsc(mdsc, &capsnap->cap_flush);
3639 if (wake_mdsc)
3640 *wake_mdsc = ret;
3641 spin_unlock(&mdsc->cap_dirty_lock);
3642}
3643
3644void ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3645 bool *wake_ci, bool *wake_mdsc)
3646{
3647 struct ceph_inode_info *ci = ceph_inode(inode);
3648
3649 lockdep_assert_held(&ci->i_ceph_lock);
3650
3651 WARN_ON_ONCE(capsnap->dirty_pages || capsnap->writing);
3652 __ceph_remove_capsnap(inode, capsnap, wake_ci, wake_mdsc);
3653}
3654
3655/*
3656 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3657 * throw away our cap_snap.
3658 *
3659 * Caller hold s_mutex.
3660 */
3661static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3662 struct ceph_mds_caps *m,
3663 struct ceph_mds_session *session)
3664{
3665 struct ceph_inode_info *ci = ceph_inode(inode);
3666 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3667 u64 follows = le64_to_cpu(m->snap_follows);
3668 struct ceph_cap_snap *capsnap;
3669 bool flushed = false;
3670 bool wake_ci = false;
3671 bool wake_mdsc = false;
3672
3673 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3674 inode, ci, session->s_mds, follows);
3675
3676 spin_lock(&ci->i_ceph_lock);
3677 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3678 if (capsnap->follows == follows) {
3679 if (capsnap->cap_flush.tid != flush_tid) {
3680 dout(" cap_snap %p follows %lld tid %lld !="
3681 " %lld\n", capsnap, follows,
3682 flush_tid, capsnap->cap_flush.tid);
3683 break;
3684 }
3685 flushed = true;
3686 break;
3687 } else {
3688 dout(" skipping cap_snap %p follows %lld\n",
3689 capsnap, capsnap->follows);
3690 }
3691 }
3692 if (flushed)
3693 ceph_remove_capsnap(inode, capsnap, &wake_ci, &wake_mdsc);
3694 spin_unlock(&ci->i_ceph_lock);
3695
3696 if (flushed) {
3697 ceph_put_snap_context(capsnap->context);
3698 ceph_put_cap_snap(capsnap);
3699 if (wake_ci)
3700 wake_up_all(&ci->i_cap_wq);
3701 if (wake_mdsc)
3702 wake_up_all(&mdsc->cap_flushing_wq);
3703 iput(inode);
3704 }
3705}
3706
3707/*
3708 * Handle TRUNC from MDS, indicating file truncation.
3709 *
3710 * caller hold s_mutex.
3711 */
3712static bool handle_cap_trunc(struct inode *inode,
3713 struct ceph_mds_caps *trunc,
3714 struct ceph_mds_session *session)
3715{
3716 struct ceph_inode_info *ci = ceph_inode(inode);
3717 int mds = session->s_mds;
3718 int seq = le32_to_cpu(trunc->seq);
3719 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3720 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3721 u64 size = le64_to_cpu(trunc->size);
3722 int implemented = 0;
3723 int dirty = __ceph_caps_dirty(ci);
3724 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3725 bool queue_trunc = false;
3726
3727 lockdep_assert_held(&ci->i_ceph_lock);
3728
3729 issued |= implemented | dirty;
3730
3731 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3732 inode, mds, seq, truncate_size, truncate_seq);
3733 queue_trunc = ceph_fill_file_size(inode, issued,
3734 truncate_seq, truncate_size, size);
3735 return queue_trunc;
3736}
3737
3738/*
3739 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3740 * different one. If we are the most recent migration we've seen (as
3741 * indicated by mseq), make note of the migrating cap bits for the
3742 * duration (until we see the corresponding IMPORT).
3743 *
3744 * caller holds s_mutex
3745 */
3746static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3747 struct ceph_mds_cap_peer *ph,
3748 struct ceph_mds_session *session)
3749{
3750 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3751 struct ceph_mds_session *tsession = NULL;
3752 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3753 struct ceph_inode_info *ci = ceph_inode(inode);
3754 u64 t_cap_id;
3755 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3756 unsigned t_seq, t_mseq;
3757 int target, issued;
3758 int mds = session->s_mds;
3759
3760 if (ph) {
3761 t_cap_id = le64_to_cpu(ph->cap_id);
3762 t_seq = le32_to_cpu(ph->seq);
3763 t_mseq = le32_to_cpu(ph->mseq);
3764 target = le32_to_cpu(ph->mds);
3765 } else {
3766 t_cap_id = t_seq = t_mseq = 0;
3767 target = -1;
3768 }
3769
3770 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3771 inode, ci, mds, mseq, target);
3772retry:
3773 spin_lock(&ci->i_ceph_lock);
3774 cap = __get_cap_for_mds(ci, mds);
3775 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3776 goto out_unlock;
3777
3778 if (target < 0) {
3779 __ceph_remove_cap(cap, false);
3780 goto out_unlock;
3781 }
3782
3783 /*
3784 * now we know we haven't received the cap import message yet
3785 * because the exported cap still exist.
3786 */
3787
3788 issued = cap->issued;
3789 if (issued != cap->implemented)
3790 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3791 "ino (%llx.%llx) mds%d seq %d mseq %d "
3792 "issued %s implemented %s\n",
3793 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3794 ceph_cap_string(issued),
3795 ceph_cap_string(cap->implemented));
3796
3797
3798 tcap = __get_cap_for_mds(ci, target);
3799 if (tcap) {
3800 /* already have caps from the target */
3801 if (tcap->cap_id == t_cap_id &&
3802 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3803 dout(" updating import cap %p mds%d\n", tcap, target);
3804 tcap->cap_id = t_cap_id;
3805 tcap->seq = t_seq - 1;
3806 tcap->issue_seq = t_seq - 1;
3807 tcap->issued |= issued;
3808 tcap->implemented |= issued;
3809 if (cap == ci->i_auth_cap) {
3810 ci->i_auth_cap = tcap;
3811 change_auth_cap_ses(ci, tcap->session);
3812 }
3813 }
3814 __ceph_remove_cap(cap, false);
3815 goto out_unlock;
3816 } else if (tsession) {
3817 /* add placeholder for the export tagert */
3818 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3819 tcap = new_cap;
3820 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3821 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3822
3823 if (!list_empty(&ci->i_cap_flush_list) &&
3824 ci->i_auth_cap == tcap) {
3825 spin_lock(&mdsc->cap_dirty_lock);
3826 list_move_tail(&ci->i_flushing_item,
3827 &tcap->session->s_cap_flushing);
3828 spin_unlock(&mdsc->cap_dirty_lock);
3829 }
3830
3831 __ceph_remove_cap(cap, false);
3832 goto out_unlock;
3833 }
3834
3835 spin_unlock(&ci->i_ceph_lock);
3836 mutex_unlock(&session->s_mutex);
3837
3838 /* open target session */
3839 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3840 if (!IS_ERR(tsession)) {
3841 if (mds > target) {
3842 mutex_lock(&session->s_mutex);
3843 mutex_lock_nested(&tsession->s_mutex,
3844 SINGLE_DEPTH_NESTING);
3845 } else {
3846 mutex_lock(&tsession->s_mutex);
3847 mutex_lock_nested(&session->s_mutex,
3848 SINGLE_DEPTH_NESTING);
3849 }
3850 new_cap = ceph_get_cap(mdsc, NULL);
3851 } else {
3852 WARN_ON(1);
3853 tsession = NULL;
3854 target = -1;
3855 mutex_lock(&session->s_mutex);
3856 }
3857 goto retry;
3858
3859out_unlock:
3860 spin_unlock(&ci->i_ceph_lock);
3861 mutex_unlock(&session->s_mutex);
3862 if (tsession) {
3863 mutex_unlock(&tsession->s_mutex);
3864 ceph_put_mds_session(tsession);
3865 }
3866 if (new_cap)
3867 ceph_put_cap(mdsc, new_cap);
3868}
3869
3870/*
3871 * Handle cap IMPORT.
3872 *
3873 * caller holds s_mutex. acquires i_ceph_lock
3874 */
3875static void handle_cap_import(struct ceph_mds_client *mdsc,
3876 struct inode *inode, struct ceph_mds_caps *im,
3877 struct ceph_mds_cap_peer *ph,
3878 struct ceph_mds_session *session,
3879 struct ceph_cap **target_cap, int *old_issued)
3880{
3881 struct ceph_inode_info *ci = ceph_inode(inode);
3882 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3883 int mds = session->s_mds;
3884 int issued;
3885 unsigned caps = le32_to_cpu(im->caps);
3886 unsigned wanted = le32_to_cpu(im->wanted);
3887 unsigned seq = le32_to_cpu(im->seq);
3888 unsigned mseq = le32_to_cpu(im->migrate_seq);
3889 u64 realmino = le64_to_cpu(im->realm);
3890 u64 cap_id = le64_to_cpu(im->cap_id);
3891 u64 p_cap_id;
3892 int peer;
3893
3894 if (ph) {
3895 p_cap_id = le64_to_cpu(ph->cap_id);
3896 peer = le32_to_cpu(ph->mds);
3897 } else {
3898 p_cap_id = 0;
3899 peer = -1;
3900 }
3901
3902 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3903 inode, ci, mds, mseq, peer);
3904retry:
3905 cap = __get_cap_for_mds(ci, mds);
3906 if (!cap) {
3907 if (!new_cap) {
3908 spin_unlock(&ci->i_ceph_lock);
3909 new_cap = ceph_get_cap(mdsc, NULL);
3910 spin_lock(&ci->i_ceph_lock);
3911 goto retry;
3912 }
3913 cap = new_cap;
3914 } else {
3915 if (new_cap) {
3916 ceph_put_cap(mdsc, new_cap);
3917 new_cap = NULL;
3918 }
3919 }
3920
3921 __ceph_caps_issued(ci, &issued);
3922 issued |= __ceph_caps_dirty(ci);
3923
3924 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3925 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3926
3927 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3928 if (ocap && ocap->cap_id == p_cap_id) {
3929 dout(" remove export cap %p mds%d flags %d\n",
3930 ocap, peer, ph->flags);
3931 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3932 (ocap->seq != le32_to_cpu(ph->seq) ||
3933 ocap->mseq != le32_to_cpu(ph->mseq))) {
3934 pr_err_ratelimited("handle_cap_import: "
3935 "mismatched seq/mseq: ino (%llx.%llx) "
3936 "mds%d seq %d mseq %d importer mds%d "
3937 "has peer seq %d mseq %d\n",
3938 ceph_vinop(inode), peer, ocap->seq,
3939 ocap->mseq, mds, le32_to_cpu(ph->seq),
3940 le32_to_cpu(ph->mseq));
3941 }
3942 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3943 }
3944
3945 *old_issued = issued;
3946 *target_cap = cap;
3947}
3948
3949/*
3950 * Handle a caps message from the MDS.
3951 *
3952 * Identify the appropriate session, inode, and call the right handler
3953 * based on the cap op.
3954 */
3955void ceph_handle_caps(struct ceph_mds_session *session,
3956 struct ceph_msg *msg)
3957{
3958 struct ceph_mds_client *mdsc = session->s_mdsc;
3959 struct inode *inode;
3960 struct ceph_inode_info *ci;
3961 struct ceph_cap *cap;
3962 struct ceph_mds_caps *h;
3963 struct ceph_mds_cap_peer *peer = NULL;
3964 struct ceph_snap_realm *realm = NULL;
3965 int op;
3966 int msg_version = le16_to_cpu(msg->hdr.version);
3967 u32 seq, mseq;
3968 struct ceph_vino vino;
3969 void *snaptrace;
3970 size_t snaptrace_len;
3971 void *p, *end;
3972 struct cap_extra_info extra_info = {};
3973 bool queue_trunc;
3974
3975 dout("handle_caps from mds%d\n", session->s_mds);
3976
3977 /* decode */
3978 end = msg->front.iov_base + msg->front.iov_len;
3979 if (msg->front.iov_len < sizeof(*h))
3980 goto bad;
3981 h = msg->front.iov_base;
3982 op = le32_to_cpu(h->op);
3983 vino.ino = le64_to_cpu(h->ino);
3984 vino.snap = CEPH_NOSNAP;
3985 seq = le32_to_cpu(h->seq);
3986 mseq = le32_to_cpu(h->migrate_seq);
3987
3988 snaptrace = h + 1;
3989 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3990 p = snaptrace + snaptrace_len;
3991
3992 if (msg_version >= 2) {
3993 u32 flock_len;
3994 ceph_decode_32_safe(&p, end, flock_len, bad);
3995 if (p + flock_len > end)
3996 goto bad;
3997 p += flock_len;
3998 }
3999
4000 if (msg_version >= 3) {
4001 if (op == CEPH_CAP_OP_IMPORT) {
4002 if (p + sizeof(*peer) > end)
4003 goto bad;
4004 peer = p;
4005 p += sizeof(*peer);
4006 } else if (op == CEPH_CAP_OP_EXPORT) {
4007 /* recorded in unused fields */
4008 peer = (void *)&h->size;
4009 }
4010 }
4011
4012 if (msg_version >= 4) {
4013 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4014 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4015 if (p + extra_info.inline_len > end)
4016 goto bad;
4017 extra_info.inline_data = p;
4018 p += extra_info.inline_len;
4019 }
4020
4021 if (msg_version >= 5) {
4022 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4023 u32 epoch_barrier;
4024
4025 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4026 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4027 }
4028
4029 if (msg_version >= 8) {
4030 u32 pool_ns_len;
4031
4032 /* version >= 6 */
4033 ceph_decode_skip_64(&p, end, bad); // flush_tid
4034 /* version >= 7 */
4035 ceph_decode_skip_32(&p, end, bad); // caller_uid
4036 ceph_decode_skip_32(&p, end, bad); // caller_gid
4037 /* version >= 8 */
4038 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4039 if (pool_ns_len > 0) {
4040 ceph_decode_need(&p, end, pool_ns_len, bad);
4041 extra_info.pool_ns =
4042 ceph_find_or_create_string(p, pool_ns_len);
4043 p += pool_ns_len;
4044 }
4045 }
4046
4047 if (msg_version >= 9) {
4048 struct ceph_timespec *btime;
4049
4050 if (p + sizeof(*btime) > end)
4051 goto bad;
4052 btime = p;
4053 ceph_decode_timespec64(&extra_info.btime, btime);
4054 p += sizeof(*btime);
4055 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4056 }
4057
4058 if (msg_version >= 11) {
4059 /* version >= 10 */
4060 ceph_decode_skip_32(&p, end, bad); // flags
4061 /* version >= 11 */
4062 extra_info.dirstat_valid = true;
4063 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4064 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4065 }
4066
4067 /* lookup ino */
4068 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4069 ci = ceph_inode(inode);
4070 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4071 vino.snap, inode);
4072
4073 mutex_lock(&session->s_mutex);
4074 inc_session_sequence(session);
4075 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4076 (unsigned)seq);
4077
4078 if (!inode) {
4079 dout(" i don't have ino %llx\n", vino.ino);
4080
4081 if (op == CEPH_CAP_OP_IMPORT) {
4082 cap = ceph_get_cap(mdsc, NULL);
4083 cap->cap_ino = vino.ino;
4084 cap->queue_release = 1;
4085 cap->cap_id = le64_to_cpu(h->cap_id);
4086 cap->mseq = mseq;
4087 cap->seq = seq;
4088 cap->issue_seq = seq;
4089 spin_lock(&session->s_cap_lock);
4090 __ceph_queue_cap_release(session, cap);
4091 spin_unlock(&session->s_cap_lock);
4092 }
4093 goto flush_cap_releases;
4094 }
4095
4096 /* these will work even if we don't have a cap yet */
4097 switch (op) {
4098 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4099 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4100 h, session);
4101 goto done;
4102
4103 case CEPH_CAP_OP_EXPORT:
4104 handle_cap_export(inode, h, peer, session);
4105 goto done_unlocked;
4106
4107 case CEPH_CAP_OP_IMPORT:
4108 realm = NULL;
4109 if (snaptrace_len) {
4110 down_write(&mdsc->snap_rwsem);
4111 ceph_update_snap_trace(mdsc, snaptrace,
4112 snaptrace + snaptrace_len,
4113 false, &realm);
4114 downgrade_write(&mdsc->snap_rwsem);
4115 } else {
4116 down_read(&mdsc->snap_rwsem);
4117 }
4118 spin_lock(&ci->i_ceph_lock);
4119 handle_cap_import(mdsc, inode, h, peer, session,
4120 &cap, &extra_info.issued);
4121 handle_cap_grant(inode, session, cap,
4122 h, msg->middle, &extra_info);
4123 if (realm)
4124 ceph_put_snap_realm(mdsc, realm);
4125 goto done_unlocked;
4126 }
4127
4128 /* the rest require a cap */
4129 spin_lock(&ci->i_ceph_lock);
4130 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4131 if (!cap) {
4132 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4133 inode, ceph_ino(inode), ceph_snap(inode),
4134 session->s_mds);
4135 spin_unlock(&ci->i_ceph_lock);
4136 goto flush_cap_releases;
4137 }
4138
4139 /* note that each of these drops i_ceph_lock for us */
4140 switch (op) {
4141 case CEPH_CAP_OP_REVOKE:
4142 case CEPH_CAP_OP_GRANT:
4143 __ceph_caps_issued(ci, &extra_info.issued);
4144 extra_info.issued |= __ceph_caps_dirty(ci);
4145 handle_cap_grant(inode, session, cap,
4146 h, msg->middle, &extra_info);
4147 goto done_unlocked;
4148
4149 case CEPH_CAP_OP_FLUSH_ACK:
4150 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4151 h, session, cap);
4152 break;
4153
4154 case CEPH_CAP_OP_TRUNC:
4155 queue_trunc = handle_cap_trunc(inode, h, session);
4156 spin_unlock(&ci->i_ceph_lock);
4157 if (queue_trunc)
4158 ceph_queue_vmtruncate(inode);
4159 break;
4160
4161 default:
4162 spin_unlock(&ci->i_ceph_lock);
4163 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4164 ceph_cap_op_name(op));
4165 }
4166
4167done:
4168 mutex_unlock(&session->s_mutex);
4169done_unlocked:
4170 iput(inode);
4171out:
4172 ceph_put_string(extra_info.pool_ns);
4173 return;
4174
4175flush_cap_releases:
4176 /*
4177 * send any cap release message to try to move things
4178 * along for the mds (who clearly thinks we still have this
4179 * cap).
4180 */
4181 ceph_flush_cap_releases(mdsc, session);
4182 goto done;
4183
4184bad:
4185 pr_err("ceph_handle_caps: corrupt message\n");
4186 ceph_msg_dump(msg);
4187 goto out;
4188}
4189
4190/*
4191 * Delayed work handler to process end of delayed cap release LRU list.
4192 *
4193 * If new caps are added to the list while processing it, these won't get
4194 * processed in this run. In this case, the ci->i_hold_caps_max will be
4195 * returned so that the work can be scheduled accordingly.
4196 */
4197unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4198{
4199 struct inode *inode;
4200 struct ceph_inode_info *ci;
4201 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4202 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4203 unsigned long loop_start = jiffies;
4204 unsigned long delay = 0;
4205
4206 dout("check_delayed_caps\n");
4207 spin_lock(&mdsc->cap_delay_lock);
4208 while (!list_empty(&mdsc->cap_delay_list)) {
4209 ci = list_first_entry(&mdsc->cap_delay_list,
4210 struct ceph_inode_info,
4211 i_cap_delay_list);
4212 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4213 dout("%s caps added recently. Exiting loop", __func__);
4214 delay = ci->i_hold_caps_max;
4215 break;
4216 }
4217 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4218 time_before(jiffies, ci->i_hold_caps_max))
4219 break;
4220 list_del_init(&ci->i_cap_delay_list);
4221
4222 inode = igrab(&ci->vfs_inode);
4223 if (inode) {
4224 spin_unlock(&mdsc->cap_delay_lock);
4225 dout("check_delayed_caps on %p\n", inode);
4226 ceph_check_caps(ci, 0, NULL);
4227 iput(inode);
4228 spin_lock(&mdsc->cap_delay_lock);
4229 }
4230 }
4231 spin_unlock(&mdsc->cap_delay_lock);
4232
4233 return delay;
4234}
4235
4236/*
4237 * Flush all dirty caps to the mds
4238 */
4239static void flush_dirty_session_caps(struct ceph_mds_session *s)
4240{
4241 struct ceph_mds_client *mdsc = s->s_mdsc;
4242 struct ceph_inode_info *ci;
4243 struct inode *inode;
4244
4245 dout("flush_dirty_caps\n");
4246 spin_lock(&mdsc->cap_dirty_lock);
4247 while (!list_empty(&s->s_cap_dirty)) {
4248 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4249 i_dirty_item);
4250 inode = &ci->vfs_inode;
4251 ihold(inode);
4252 dout("flush_dirty_caps %p\n", inode);
4253 spin_unlock(&mdsc->cap_dirty_lock);
4254 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4255 iput(inode);
4256 spin_lock(&mdsc->cap_dirty_lock);
4257 }
4258 spin_unlock(&mdsc->cap_dirty_lock);
4259 dout("flush_dirty_caps done\n");
4260}
4261
4262static void iterate_sessions(struct ceph_mds_client *mdsc,
4263 void (*cb)(struct ceph_mds_session *))
4264{
4265 int mds;
4266
4267 mutex_lock(&mdsc->mutex);
4268 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
4269 struct ceph_mds_session *s;
4270
4271 if (!mdsc->sessions[mds])
4272 continue;
4273
4274 s = ceph_get_mds_session(mdsc->sessions[mds]);
4275 if (!s)
4276 continue;
4277
4278 mutex_unlock(&mdsc->mutex);
4279 cb(s);
4280 ceph_put_mds_session(s);
4281 mutex_lock(&mdsc->mutex);
4282 }
4283 mutex_unlock(&mdsc->mutex);
4284}
4285
4286void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4287{
4288 iterate_sessions(mdsc, flush_dirty_session_caps);
4289}
4290
4291void __ceph_touch_fmode(struct ceph_inode_info *ci,
4292 struct ceph_mds_client *mdsc, int fmode)
4293{
4294 unsigned long now = jiffies;
4295 if (fmode & CEPH_FILE_MODE_RD)
4296 ci->i_last_rd = now;
4297 if (fmode & CEPH_FILE_MODE_WR)
4298 ci->i_last_wr = now;
4299 /* queue periodic check */
4300 if (fmode &&
4301 __ceph_is_any_real_caps(ci) &&
4302 list_empty(&ci->i_cap_delay_list))
4303 __cap_delay_requeue(mdsc, ci);
4304}
4305
4306void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4307{
4308 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4309 int bits = (fmode << 1) | 1;
4310 bool is_opened = false;
4311 int i;
4312
4313 if (count == 1)
4314 atomic64_inc(&mdsc->metric.opened_files);
4315
4316 spin_lock(&ci->i_ceph_lock);
4317 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4318 if (bits & (1 << i))
4319 ci->i_nr_by_mode[i] += count;
4320
4321 /*
4322 * If any of the mode ref is larger than 1,
4323 * that means it has been already opened by
4324 * others. Just skip checking the PIN ref.
4325 */
4326 if (i && ci->i_nr_by_mode[i] > 1)
4327 is_opened = true;
4328 }
4329
4330 if (!is_opened)
4331 percpu_counter_inc(&mdsc->metric.opened_inodes);
4332 spin_unlock(&ci->i_ceph_lock);
4333}
4334
4335/*
4336 * Drop open file reference. If we were the last open file,
4337 * we may need to release capabilities to the MDS (or schedule
4338 * their delayed release).
4339 */
4340void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4341{
4342 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4343 int bits = (fmode << 1) | 1;
4344 bool is_closed = true;
4345 int i;
4346
4347 if (count == 1)
4348 atomic64_dec(&mdsc->metric.opened_files);
4349
4350 spin_lock(&ci->i_ceph_lock);
4351 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4352 if (bits & (1 << i)) {
4353 BUG_ON(ci->i_nr_by_mode[i] < count);
4354 ci->i_nr_by_mode[i] -= count;
4355 }
4356
4357 /*
4358 * If any of the mode ref is not 0 after
4359 * decreased, that means it is still opened
4360 * by others. Just skip checking the PIN ref.
4361 */
4362 if (i && ci->i_nr_by_mode[i])
4363 is_closed = false;
4364 }
4365
4366 if (is_closed)
4367 percpu_counter_dec(&mdsc->metric.opened_inodes);
4368 spin_unlock(&ci->i_ceph_lock);
4369}
4370
4371/*
4372 * For a soon-to-be unlinked file, drop the LINK caps. If it
4373 * looks like the link count will hit 0, drop any other caps (other
4374 * than PIN) we don't specifically want (due to the file still being
4375 * open).
4376 */
4377int ceph_drop_caps_for_unlink(struct inode *inode)
4378{
4379 struct ceph_inode_info *ci = ceph_inode(inode);
4380 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4381
4382 spin_lock(&ci->i_ceph_lock);
4383 if (inode->i_nlink == 1) {
4384 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4385
4386 if (__ceph_caps_dirty(ci)) {
4387 struct ceph_mds_client *mdsc =
4388 ceph_inode_to_client(inode)->mdsc;
4389 __cap_delay_requeue_front(mdsc, ci);
4390 }
4391 }
4392 spin_unlock(&ci->i_ceph_lock);
4393 return drop;
4394}
4395
4396/*
4397 * Helpers for embedding cap and dentry lease releases into mds
4398 * requests.
4399 *
4400 * @force is used by dentry_release (below) to force inclusion of a
4401 * record for the directory inode, even when there aren't any caps to
4402 * drop.
4403 */
4404int ceph_encode_inode_release(void **p, struct inode *inode,
4405 int mds, int drop, int unless, int force)
4406{
4407 struct ceph_inode_info *ci = ceph_inode(inode);
4408 struct ceph_cap *cap;
4409 struct ceph_mds_request_release *rel = *p;
4410 int used, dirty;
4411 int ret = 0;
4412
4413 spin_lock(&ci->i_ceph_lock);
4414 used = __ceph_caps_used(ci);
4415 dirty = __ceph_caps_dirty(ci);
4416
4417 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4418 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4419 ceph_cap_string(unless));
4420
4421 /* only drop unused, clean caps */
4422 drop &= ~(used | dirty);
4423
4424 cap = __get_cap_for_mds(ci, mds);
4425 if (cap && __cap_is_valid(cap)) {
4426 unless &= cap->issued;
4427 if (unless) {
4428 if (unless & CEPH_CAP_AUTH_EXCL)
4429 drop &= ~CEPH_CAP_AUTH_SHARED;
4430 if (unless & CEPH_CAP_LINK_EXCL)
4431 drop &= ~CEPH_CAP_LINK_SHARED;
4432 if (unless & CEPH_CAP_XATTR_EXCL)
4433 drop &= ~CEPH_CAP_XATTR_SHARED;
4434 if (unless & CEPH_CAP_FILE_EXCL)
4435 drop &= ~CEPH_CAP_FILE_SHARED;
4436 }
4437
4438 if (force || (cap->issued & drop)) {
4439 if (cap->issued & drop) {
4440 int wanted = __ceph_caps_wanted(ci);
4441 dout("encode_inode_release %p cap %p "
4442 "%s -> %s, wanted %s -> %s\n", inode, cap,
4443 ceph_cap_string(cap->issued),
4444 ceph_cap_string(cap->issued & ~drop),
4445 ceph_cap_string(cap->mds_wanted),
4446 ceph_cap_string(wanted));
4447
4448 cap->issued &= ~drop;
4449 cap->implemented &= ~drop;
4450 cap->mds_wanted = wanted;
4451 if (cap == ci->i_auth_cap &&
4452 !(wanted & CEPH_CAP_ANY_FILE_WR))
4453 ci->i_requested_max_size = 0;
4454 } else {
4455 dout("encode_inode_release %p cap %p %s"
4456 " (force)\n", inode, cap,
4457 ceph_cap_string(cap->issued));
4458 }
4459
4460 rel->ino = cpu_to_le64(ceph_ino(inode));
4461 rel->cap_id = cpu_to_le64(cap->cap_id);
4462 rel->seq = cpu_to_le32(cap->seq);
4463 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4464 rel->mseq = cpu_to_le32(cap->mseq);
4465 rel->caps = cpu_to_le32(cap->implemented);
4466 rel->wanted = cpu_to_le32(cap->mds_wanted);
4467 rel->dname_len = 0;
4468 rel->dname_seq = 0;
4469 *p += sizeof(*rel);
4470 ret = 1;
4471 } else {
4472 dout("encode_inode_release %p cap %p %s (noop)\n",
4473 inode, cap, ceph_cap_string(cap->issued));
4474 }
4475 }
4476 spin_unlock(&ci->i_ceph_lock);
4477 return ret;
4478}
4479
4480int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4481 struct inode *dir,
4482 int mds, int drop, int unless)
4483{
4484 struct dentry *parent = NULL;
4485 struct ceph_mds_request_release *rel = *p;
4486 struct ceph_dentry_info *di = ceph_dentry(dentry);
4487 int force = 0;
4488 int ret;
4489
4490 /*
4491 * force an record for the directory caps if we have a dentry lease.
4492 * this is racy (can't take i_ceph_lock and d_lock together), but it
4493 * doesn't have to be perfect; the mds will revoke anything we don't
4494 * release.
4495 */
4496 spin_lock(&dentry->d_lock);
4497 if (di->lease_session && di->lease_session->s_mds == mds)
4498 force = 1;
4499 if (!dir) {
4500 parent = dget(dentry->d_parent);
4501 dir = d_inode(parent);
4502 }
4503 spin_unlock(&dentry->d_lock);
4504
4505 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4506 dput(parent);
4507
4508 spin_lock(&dentry->d_lock);
4509 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4510 dout("encode_dentry_release %p mds%d seq %d\n",
4511 dentry, mds, (int)di->lease_seq);
4512 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4513 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4514 *p += dentry->d_name.len;
4515 rel->dname_seq = cpu_to_le32(di->lease_seq);
4516 __ceph_mdsc_drop_dentry_lease(dentry);
4517 }
4518 spin_unlock(&dentry->d_lock);
4519 return ret;
4520}