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 <linux/ceph/decode.h>
14#include <linux/ceph/messenger.h>
15
16/*
17 * Capability management
18 *
19 * The Ceph metadata servers control client access to inode metadata
20 * and file data by issuing capabilities, granting clients permission
21 * to read and/or write both inode field and file data to OSDs
22 * (storage nodes). Each capability consists of a set of bits
23 * indicating which operations are allowed.
24 *
25 * If the client holds a *_SHARED cap, the client has a coherent value
26 * that can be safely read from the cached inode.
27 *
28 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
29 * client is allowed to change inode attributes (e.g., file size,
30 * mtime), note its dirty state in the ceph_cap, and asynchronously
31 * flush that metadata change to the MDS.
32 *
33 * In the event of a conflicting operation (perhaps by another
34 * client), the MDS will revoke the conflicting client capabilities.
35 *
36 * In order for a client to cache an inode, it must hold a capability
37 * with at least one MDS server. When inodes are released, release
38 * notifications are batched and periodically sent en masse to the MDS
39 * cluster to release server state.
40 */
41
42
43/*
44 * Generate readable cap strings for debugging output.
45 */
46#define MAX_CAP_STR 20
47static char cap_str[MAX_CAP_STR][40];
48static DEFINE_SPINLOCK(cap_str_lock);
49static int last_cap_str;
50
51static char *gcap_string(char *s, int c)
52{
53 if (c & CEPH_CAP_GSHARED)
54 *s++ = 's';
55 if (c & CEPH_CAP_GEXCL)
56 *s++ = 'x';
57 if (c & CEPH_CAP_GCACHE)
58 *s++ = 'c';
59 if (c & CEPH_CAP_GRD)
60 *s++ = 'r';
61 if (c & CEPH_CAP_GWR)
62 *s++ = 'w';
63 if (c & CEPH_CAP_GBUFFER)
64 *s++ = 'b';
65 if (c & CEPH_CAP_GLAZYIO)
66 *s++ = 'l';
67 return s;
68}
69
70const char *ceph_cap_string(int caps)
71{
72 int i;
73 char *s;
74 int c;
75
76 spin_lock(&cap_str_lock);
77 i = last_cap_str++;
78 if (last_cap_str == MAX_CAP_STR)
79 last_cap_str = 0;
80 spin_unlock(&cap_str_lock);
81
82 s = cap_str[i];
83
84 if (caps & CEPH_CAP_PIN)
85 *s++ = 'p';
86
87 c = (caps >> CEPH_CAP_SAUTH) & 3;
88 if (c) {
89 *s++ = 'A';
90 s = gcap_string(s, c);
91 }
92
93 c = (caps >> CEPH_CAP_SLINK) & 3;
94 if (c) {
95 *s++ = 'L';
96 s = gcap_string(s, c);
97 }
98
99 c = (caps >> CEPH_CAP_SXATTR) & 3;
100 if (c) {
101 *s++ = 'X';
102 s = gcap_string(s, c);
103 }
104
105 c = caps >> CEPH_CAP_SFILE;
106 if (c) {
107 *s++ = 'F';
108 s = gcap_string(s, c);
109 }
110
111 if (s == cap_str[i])
112 *s++ = '-';
113 *s = 0;
114 return cap_str[i];
115}
116
117void ceph_caps_init(struct ceph_mds_client *mdsc)
118{
119 INIT_LIST_HEAD(&mdsc->caps_list);
120 spin_lock_init(&mdsc->caps_list_lock);
121}
122
123void ceph_caps_finalize(struct ceph_mds_client *mdsc)
124{
125 struct ceph_cap *cap;
126
127 spin_lock(&mdsc->caps_list_lock);
128 while (!list_empty(&mdsc->caps_list)) {
129 cap = list_first_entry(&mdsc->caps_list,
130 struct ceph_cap, caps_item);
131 list_del(&cap->caps_item);
132 kmem_cache_free(ceph_cap_cachep, cap);
133 }
134 mdsc->caps_total_count = 0;
135 mdsc->caps_avail_count = 0;
136 mdsc->caps_use_count = 0;
137 mdsc->caps_reserve_count = 0;
138 mdsc->caps_min_count = 0;
139 spin_unlock(&mdsc->caps_list_lock);
140}
141
142void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
143{
144 spin_lock(&mdsc->caps_list_lock);
145 mdsc->caps_min_count += delta;
146 BUG_ON(mdsc->caps_min_count < 0);
147 spin_unlock(&mdsc->caps_list_lock);
148}
149
150int ceph_reserve_caps(struct ceph_mds_client *mdsc,
151 struct ceph_cap_reservation *ctx, int need)
152{
153 int i;
154 struct ceph_cap *cap;
155 int have;
156 int alloc = 0;
157 LIST_HEAD(newcaps);
158 int ret = 0;
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 ret = -ENOMEM;
179 goto out_alloc_count;
180 }
181 list_add(&cap->caps_item, &newcaps);
182 alloc++;
183 }
184 BUG_ON(have + alloc != need);
185
186 spin_lock(&mdsc->caps_list_lock);
187 mdsc->caps_total_count += alloc;
188 mdsc->caps_reserve_count += alloc;
189 list_splice(&newcaps, &mdsc->caps_list);
190
191 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 mdsc->caps_reserve_count +
193 mdsc->caps_avail_count);
194 spin_unlock(&mdsc->caps_list_lock);
195
196 ctx->count = need;
197 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
198 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
199 mdsc->caps_reserve_count, mdsc->caps_avail_count);
200 return 0;
201
202out_alloc_count:
203 /* we didn't manage to reserve as much as we needed */
204 pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
205 ctx, need, have);
206 return ret;
207}
208
209int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
210 struct ceph_cap_reservation *ctx)
211{
212 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
213 if (ctx->count) {
214 spin_lock(&mdsc->caps_list_lock);
215 BUG_ON(mdsc->caps_reserve_count < ctx->count);
216 mdsc->caps_reserve_count -= ctx->count;
217 mdsc->caps_avail_count += ctx->count;
218 ctx->count = 0;
219 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
220 mdsc->caps_total_count, mdsc->caps_use_count,
221 mdsc->caps_reserve_count, mdsc->caps_avail_count);
222 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
223 mdsc->caps_reserve_count +
224 mdsc->caps_avail_count);
225 spin_unlock(&mdsc->caps_list_lock);
226 }
227 return 0;
228}
229
230static struct ceph_cap *get_cap(struct ceph_mds_client *mdsc,
231 struct ceph_cap_reservation *ctx)
232{
233 struct ceph_cap *cap = NULL;
234
235 /* temporary, until we do something about cap import/export */
236 if (!ctx) {
237 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
238 if (cap) {
239 mdsc->caps_use_count++;
240 mdsc->caps_total_count++;
241 }
242 return cap;
243 }
244
245 spin_lock(&mdsc->caps_list_lock);
246 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
247 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
248 mdsc->caps_reserve_count, mdsc->caps_avail_count);
249 BUG_ON(!ctx->count);
250 BUG_ON(ctx->count > mdsc->caps_reserve_count);
251 BUG_ON(list_empty(&mdsc->caps_list));
252
253 ctx->count--;
254 mdsc->caps_reserve_count--;
255 mdsc->caps_use_count++;
256
257 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
258 list_del(&cap->caps_item);
259
260 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
261 mdsc->caps_reserve_count + mdsc->caps_avail_count);
262 spin_unlock(&mdsc->caps_list_lock);
263 return cap;
264}
265
266void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
267{
268 spin_lock(&mdsc->caps_list_lock);
269 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
270 cap, mdsc->caps_total_count, mdsc->caps_use_count,
271 mdsc->caps_reserve_count, mdsc->caps_avail_count);
272 mdsc->caps_use_count--;
273 /*
274 * Keep some preallocated caps around (ceph_min_count), to
275 * avoid lots of free/alloc churn.
276 */
277 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
278 mdsc->caps_min_count) {
279 mdsc->caps_total_count--;
280 kmem_cache_free(ceph_cap_cachep, cap);
281 } else {
282 mdsc->caps_avail_count++;
283 list_add(&cap->caps_item, &mdsc->caps_list);
284 }
285
286 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
287 mdsc->caps_reserve_count + mdsc->caps_avail_count);
288 spin_unlock(&mdsc->caps_list_lock);
289}
290
291void ceph_reservation_status(struct ceph_fs_client *fsc,
292 int *total, int *avail, int *used, int *reserved,
293 int *min)
294{
295 struct ceph_mds_client *mdsc = fsc->mdsc;
296
297 if (total)
298 *total = mdsc->caps_total_count;
299 if (avail)
300 *avail = mdsc->caps_avail_count;
301 if (used)
302 *used = mdsc->caps_use_count;
303 if (reserved)
304 *reserved = mdsc->caps_reserve_count;
305 if (min)
306 *min = mdsc->caps_min_count;
307}
308
309/*
310 * Find ceph_cap for given mds, if any.
311 *
312 * Called with i_ceph_lock held.
313 */
314static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
315{
316 struct ceph_cap *cap;
317 struct rb_node *n = ci->i_caps.rb_node;
318
319 while (n) {
320 cap = rb_entry(n, struct ceph_cap, ci_node);
321 if (mds < cap->mds)
322 n = n->rb_left;
323 else if (mds > cap->mds)
324 n = n->rb_right;
325 else
326 return cap;
327 }
328 return NULL;
329}
330
331struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
332{
333 struct ceph_cap *cap;
334
335 spin_lock(&ci->i_ceph_lock);
336 cap = __get_cap_for_mds(ci, mds);
337 spin_unlock(&ci->i_ceph_lock);
338 return cap;
339}
340
341/*
342 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
343 */
344static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
345{
346 struct ceph_cap *cap;
347 int mds = -1;
348 struct rb_node *p;
349
350 /* prefer mds with WR|BUFFER|EXCL caps */
351 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
352 cap = rb_entry(p, struct ceph_cap, ci_node);
353 mds = cap->mds;
354 if (cap->issued & (CEPH_CAP_FILE_WR |
355 CEPH_CAP_FILE_BUFFER |
356 CEPH_CAP_FILE_EXCL))
357 break;
358 }
359 return mds;
360}
361
362int ceph_get_cap_mds(struct inode *inode)
363{
364 struct ceph_inode_info *ci = ceph_inode(inode);
365 int mds;
366 spin_lock(&ci->i_ceph_lock);
367 mds = __ceph_get_cap_mds(ceph_inode(inode));
368 spin_unlock(&ci->i_ceph_lock);
369 return mds;
370}
371
372/*
373 * Called under i_ceph_lock.
374 */
375static void __insert_cap_node(struct ceph_inode_info *ci,
376 struct ceph_cap *new)
377{
378 struct rb_node **p = &ci->i_caps.rb_node;
379 struct rb_node *parent = NULL;
380 struct ceph_cap *cap = NULL;
381
382 while (*p) {
383 parent = *p;
384 cap = rb_entry(parent, struct ceph_cap, ci_node);
385 if (new->mds < cap->mds)
386 p = &(*p)->rb_left;
387 else if (new->mds > cap->mds)
388 p = &(*p)->rb_right;
389 else
390 BUG();
391 }
392
393 rb_link_node(&new->ci_node, parent, p);
394 rb_insert_color(&new->ci_node, &ci->i_caps);
395}
396
397/*
398 * (re)set cap hold timeouts, which control the delayed release
399 * of unused caps back to the MDS. Should be called on cap use.
400 */
401static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
402 struct ceph_inode_info *ci)
403{
404 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
405
406 ci->i_hold_caps_min = round_jiffies(jiffies +
407 ma->caps_wanted_delay_min * HZ);
408 ci->i_hold_caps_max = round_jiffies(jiffies +
409 ma->caps_wanted_delay_max * HZ);
410 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
411 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
412}
413
414/*
415 * (Re)queue cap at the end of the delayed cap release list.
416 *
417 * If I_FLUSH is set, leave the inode at the front of the list.
418 *
419 * Caller holds i_ceph_lock
420 * -> we take mdsc->cap_delay_lock
421 */
422static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
423 struct ceph_inode_info *ci)
424{
425 __cap_set_timeouts(mdsc, ci);
426 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
427 ci->i_ceph_flags, ci->i_hold_caps_max);
428 if (!mdsc->stopping) {
429 spin_lock(&mdsc->cap_delay_lock);
430 if (!list_empty(&ci->i_cap_delay_list)) {
431 if (ci->i_ceph_flags & CEPH_I_FLUSH)
432 goto no_change;
433 list_del_init(&ci->i_cap_delay_list);
434 }
435 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
436no_change:
437 spin_unlock(&mdsc->cap_delay_lock);
438 }
439}
440
441/*
442 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
443 * indicating we should send a cap message to flush dirty metadata
444 * asap, and move to the front of the delayed cap list.
445 */
446static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
447 struct ceph_inode_info *ci)
448{
449 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
450 spin_lock(&mdsc->cap_delay_lock);
451 ci->i_ceph_flags |= CEPH_I_FLUSH;
452 if (!list_empty(&ci->i_cap_delay_list))
453 list_del_init(&ci->i_cap_delay_list);
454 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
455 spin_unlock(&mdsc->cap_delay_lock);
456}
457
458/*
459 * Cancel delayed work on cap.
460 *
461 * Caller must hold i_ceph_lock.
462 */
463static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
464 struct ceph_inode_info *ci)
465{
466 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
467 if (list_empty(&ci->i_cap_delay_list))
468 return;
469 spin_lock(&mdsc->cap_delay_lock);
470 list_del_init(&ci->i_cap_delay_list);
471 spin_unlock(&mdsc->cap_delay_lock);
472}
473
474/*
475 * Common issue checks for add_cap, handle_cap_grant.
476 */
477static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
478 unsigned issued)
479{
480 unsigned had = __ceph_caps_issued(ci, NULL);
481
482 /*
483 * Each time we receive FILE_CACHE anew, we increment
484 * i_rdcache_gen.
485 */
486 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
487 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
488 ci->i_rdcache_gen++;
489
490 /*
491 * if we are newly issued FILE_SHARED, clear D_COMPLETE; we
492 * don't know what happened to this directory while we didn't
493 * have the cap.
494 */
495 if ((issued & CEPH_CAP_FILE_SHARED) &&
496 (had & CEPH_CAP_FILE_SHARED) == 0) {
497 ci->i_shared_gen++;
498 if (S_ISDIR(ci->vfs_inode.i_mode))
499 ceph_dir_clear_complete(&ci->vfs_inode);
500 }
501}
502
503/*
504 * Add a capability under the given MDS session.
505 *
506 * Caller should hold session snap_rwsem (read) and s_mutex.
507 *
508 * @fmode is the open file mode, if we are opening a file, otherwise
509 * it is < 0. (This is so we can atomically add the cap and add an
510 * open file reference to it.)
511 */
512int ceph_add_cap(struct inode *inode,
513 struct ceph_mds_session *session, u64 cap_id,
514 int fmode, unsigned issued, unsigned wanted,
515 unsigned seq, unsigned mseq, u64 realmino, int flags,
516 struct ceph_cap_reservation *caps_reservation)
517{
518 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
519 struct ceph_inode_info *ci = ceph_inode(inode);
520 struct ceph_cap *new_cap = NULL;
521 struct ceph_cap *cap;
522 int mds = session->s_mds;
523 int actual_wanted;
524
525 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
526 session->s_mds, cap_id, ceph_cap_string(issued), seq);
527
528 /*
529 * If we are opening the file, include file mode wanted bits
530 * in wanted.
531 */
532 if (fmode >= 0)
533 wanted |= ceph_caps_for_mode(fmode);
534
535retry:
536 spin_lock(&ci->i_ceph_lock);
537 cap = __get_cap_for_mds(ci, mds);
538 if (!cap) {
539 if (new_cap) {
540 cap = new_cap;
541 new_cap = NULL;
542 } else {
543 spin_unlock(&ci->i_ceph_lock);
544 new_cap = get_cap(mdsc, caps_reservation);
545 if (new_cap == NULL)
546 return -ENOMEM;
547 goto retry;
548 }
549
550 cap->issued = 0;
551 cap->implemented = 0;
552 cap->mds = mds;
553 cap->mds_wanted = 0;
554
555 cap->ci = ci;
556 __insert_cap_node(ci, cap);
557
558 /* clear out old exporting info? (i.e. on cap import) */
559 if (ci->i_cap_exporting_mds == mds) {
560 ci->i_cap_exporting_issued = 0;
561 ci->i_cap_exporting_mseq = 0;
562 ci->i_cap_exporting_mds = -1;
563 }
564
565 /* add to session cap list */
566 cap->session = session;
567 spin_lock(&session->s_cap_lock);
568 list_add_tail(&cap->session_caps, &session->s_caps);
569 session->s_nr_caps++;
570 spin_unlock(&session->s_cap_lock);
571 } else if (new_cap)
572 ceph_put_cap(mdsc, new_cap);
573
574 if (!ci->i_snap_realm) {
575 /*
576 * add this inode to the appropriate snap realm
577 */
578 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
579 realmino);
580 if (realm) {
581 ceph_get_snap_realm(mdsc, realm);
582 spin_lock(&realm->inodes_with_caps_lock);
583 ci->i_snap_realm = realm;
584 list_add(&ci->i_snap_realm_item,
585 &realm->inodes_with_caps);
586 spin_unlock(&realm->inodes_with_caps_lock);
587 } else {
588 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
589 realmino);
590 WARN_ON(!realm);
591 }
592 }
593
594 __check_cap_issue(ci, cap, issued);
595
596 /*
597 * If we are issued caps we don't want, or the mds' wanted
598 * value appears to be off, queue a check so we'll release
599 * later and/or update the mds wanted value.
600 */
601 actual_wanted = __ceph_caps_wanted(ci);
602 if ((wanted & ~actual_wanted) ||
603 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
604 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
605 ceph_cap_string(issued), ceph_cap_string(wanted),
606 ceph_cap_string(actual_wanted));
607 __cap_delay_requeue(mdsc, ci);
608 }
609
610 if (flags & CEPH_CAP_FLAG_AUTH)
611 ci->i_auth_cap = cap;
612 else if (ci->i_auth_cap == cap)
613 ci->i_auth_cap = NULL;
614
615 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
616 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
617 ceph_cap_string(issued|cap->issued), seq, mds);
618 cap->cap_id = cap_id;
619 cap->issued = issued;
620 cap->implemented |= issued;
621 cap->mds_wanted |= wanted;
622 cap->seq = seq;
623 cap->issue_seq = seq;
624 cap->mseq = mseq;
625 cap->cap_gen = session->s_cap_gen;
626
627 if (fmode >= 0)
628 __ceph_get_fmode(ci, fmode);
629 spin_unlock(&ci->i_ceph_lock);
630 wake_up_all(&ci->i_cap_wq);
631 return 0;
632}
633
634/*
635 * Return true if cap has not timed out and belongs to the current
636 * generation of the MDS session (i.e. has not gone 'stale' due to
637 * us losing touch with the mds).
638 */
639static int __cap_is_valid(struct ceph_cap *cap)
640{
641 unsigned long ttl;
642 u32 gen;
643
644 spin_lock(&cap->session->s_gen_ttl_lock);
645 gen = cap->session->s_cap_gen;
646 ttl = cap->session->s_cap_ttl;
647 spin_unlock(&cap->session->s_gen_ttl_lock);
648
649 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
650 dout("__cap_is_valid %p cap %p issued %s "
651 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
652 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
653 return 0;
654 }
655
656 return 1;
657}
658
659/*
660 * Return set of valid cap bits issued to us. Note that caps time
661 * out, and may be invalidated in bulk if the client session times out
662 * and session->s_cap_gen is bumped.
663 */
664int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
665{
666 int have = ci->i_snap_caps | ci->i_cap_exporting_issued;
667 struct ceph_cap *cap;
668 struct rb_node *p;
669
670 if (implemented)
671 *implemented = 0;
672 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
673 cap = rb_entry(p, struct ceph_cap, ci_node);
674 if (!__cap_is_valid(cap))
675 continue;
676 dout("__ceph_caps_issued %p cap %p issued %s\n",
677 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
678 have |= cap->issued;
679 if (implemented)
680 *implemented |= cap->implemented;
681 }
682 return have;
683}
684
685/*
686 * Get cap bits issued by caps other than @ocap
687 */
688int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
689{
690 int have = ci->i_snap_caps;
691 struct ceph_cap *cap;
692 struct rb_node *p;
693
694 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
695 cap = rb_entry(p, struct ceph_cap, ci_node);
696 if (cap == ocap)
697 continue;
698 if (!__cap_is_valid(cap))
699 continue;
700 have |= cap->issued;
701 }
702 return have;
703}
704
705/*
706 * Move a cap to the end of the LRU (oldest caps at list head, newest
707 * at list tail).
708 */
709static void __touch_cap(struct ceph_cap *cap)
710{
711 struct ceph_mds_session *s = cap->session;
712
713 spin_lock(&s->s_cap_lock);
714 if (s->s_cap_iterator == NULL) {
715 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
716 s->s_mds);
717 list_move_tail(&cap->session_caps, &s->s_caps);
718 } else {
719 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
720 &cap->ci->vfs_inode, cap, s->s_mds);
721 }
722 spin_unlock(&s->s_cap_lock);
723}
724
725/*
726 * Check if we hold the given mask. If so, move the cap(s) to the
727 * front of their respective LRUs. (This is the preferred way for
728 * callers to check for caps they want.)
729 */
730int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
731{
732 struct ceph_cap *cap;
733 struct rb_node *p;
734 int have = ci->i_snap_caps;
735
736 if ((have & mask) == mask) {
737 dout("__ceph_caps_issued_mask %p snap issued %s"
738 " (mask %s)\n", &ci->vfs_inode,
739 ceph_cap_string(have),
740 ceph_cap_string(mask));
741 return 1;
742 }
743
744 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
745 cap = rb_entry(p, struct ceph_cap, ci_node);
746 if (!__cap_is_valid(cap))
747 continue;
748 if ((cap->issued & mask) == mask) {
749 dout("__ceph_caps_issued_mask %p cap %p issued %s"
750 " (mask %s)\n", &ci->vfs_inode, cap,
751 ceph_cap_string(cap->issued),
752 ceph_cap_string(mask));
753 if (touch)
754 __touch_cap(cap);
755 return 1;
756 }
757
758 /* does a combination of caps satisfy mask? */
759 have |= cap->issued;
760 if ((have & mask) == mask) {
761 dout("__ceph_caps_issued_mask %p combo issued %s"
762 " (mask %s)\n", &ci->vfs_inode,
763 ceph_cap_string(cap->issued),
764 ceph_cap_string(mask));
765 if (touch) {
766 struct rb_node *q;
767
768 /* touch this + preceding caps */
769 __touch_cap(cap);
770 for (q = rb_first(&ci->i_caps); q != p;
771 q = rb_next(q)) {
772 cap = rb_entry(q, struct ceph_cap,
773 ci_node);
774 if (!__cap_is_valid(cap))
775 continue;
776 __touch_cap(cap);
777 }
778 }
779 return 1;
780 }
781 }
782
783 return 0;
784}
785
786/*
787 * Return true if mask caps are currently being revoked by an MDS.
788 */
789int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
790{
791 struct inode *inode = &ci->vfs_inode;
792 struct ceph_cap *cap;
793 struct rb_node *p;
794 int ret = 0;
795
796 spin_lock(&ci->i_ceph_lock);
797 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
798 cap = rb_entry(p, struct ceph_cap, ci_node);
799 if (__cap_is_valid(cap) &&
800 (cap->implemented & ~cap->issued & mask)) {
801 ret = 1;
802 break;
803 }
804 }
805 spin_unlock(&ci->i_ceph_lock);
806 dout("ceph_caps_revoking %p %s = %d\n", inode,
807 ceph_cap_string(mask), ret);
808 return ret;
809}
810
811int __ceph_caps_used(struct ceph_inode_info *ci)
812{
813 int used = 0;
814 if (ci->i_pin_ref)
815 used |= CEPH_CAP_PIN;
816 if (ci->i_rd_ref)
817 used |= CEPH_CAP_FILE_RD;
818 if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
819 used |= CEPH_CAP_FILE_CACHE;
820 if (ci->i_wr_ref)
821 used |= CEPH_CAP_FILE_WR;
822 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
823 used |= CEPH_CAP_FILE_BUFFER;
824 return used;
825}
826
827/*
828 * wanted, by virtue of open file modes
829 */
830int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
831{
832 int want = 0;
833 int mode;
834 for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
835 if (ci->i_nr_by_mode[mode])
836 want |= ceph_caps_for_mode(mode);
837 return want;
838}
839
840/*
841 * Return caps we have registered with the MDS(s) as 'wanted'.
842 */
843int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
844{
845 struct ceph_cap *cap;
846 struct rb_node *p;
847 int mds_wanted = 0;
848
849 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
850 cap = rb_entry(p, struct ceph_cap, ci_node);
851 if (!__cap_is_valid(cap))
852 continue;
853 mds_wanted |= cap->mds_wanted;
854 }
855 return mds_wanted;
856}
857
858/*
859 * called under i_ceph_lock
860 */
861static int __ceph_is_any_caps(struct ceph_inode_info *ci)
862{
863 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0;
864}
865
866/*
867 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
868 *
869 * caller should hold i_ceph_lock.
870 * caller will not hold session s_mutex if called from destroy_inode.
871 */
872void __ceph_remove_cap(struct ceph_cap *cap)
873{
874 struct ceph_mds_session *session = cap->session;
875 struct ceph_inode_info *ci = cap->ci;
876 struct ceph_mds_client *mdsc =
877 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
878 int removed = 0;
879
880 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
881
882 /* remove from session list */
883 spin_lock(&session->s_cap_lock);
884 if (session->s_cap_iterator == cap) {
885 /* not yet, we are iterating over this very cap */
886 dout("__ceph_remove_cap delaying %p removal from session %p\n",
887 cap, cap->session);
888 } else {
889 list_del_init(&cap->session_caps);
890 session->s_nr_caps--;
891 cap->session = NULL;
892 removed = 1;
893 }
894 /* protect backpointer with s_cap_lock: see iterate_session_caps */
895 cap->ci = NULL;
896 spin_unlock(&session->s_cap_lock);
897
898 /* remove from inode list */
899 rb_erase(&cap->ci_node, &ci->i_caps);
900 if (ci->i_auth_cap == cap)
901 ci->i_auth_cap = NULL;
902
903 if (removed)
904 ceph_put_cap(mdsc, cap);
905
906 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
907 struct ceph_snap_realm *realm = ci->i_snap_realm;
908 spin_lock(&realm->inodes_with_caps_lock);
909 list_del_init(&ci->i_snap_realm_item);
910 ci->i_snap_realm_counter++;
911 ci->i_snap_realm = NULL;
912 spin_unlock(&realm->inodes_with_caps_lock);
913 ceph_put_snap_realm(mdsc, realm);
914 }
915 if (!__ceph_is_any_real_caps(ci))
916 __cap_delay_cancel(mdsc, ci);
917}
918
919/*
920 * Build and send a cap message to the given MDS.
921 *
922 * Caller should be holding s_mutex.
923 */
924static int send_cap_msg(struct ceph_mds_session *session,
925 u64 ino, u64 cid, int op,
926 int caps, int wanted, int dirty,
927 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
928 u64 size, u64 max_size,
929 struct timespec *mtime, struct timespec *atime,
930 u64 time_warp_seq,
931 uid_t uid, gid_t gid, umode_t mode,
932 u64 xattr_version,
933 struct ceph_buffer *xattrs_buf,
934 u64 follows)
935{
936 struct ceph_mds_caps *fc;
937 struct ceph_msg *msg;
938
939 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
940 " seq %u/%u mseq %u follows %lld size %llu/%llu"
941 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
942 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
943 ceph_cap_string(dirty),
944 seq, issue_seq, mseq, follows, size, max_size,
945 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
946
947 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), GFP_NOFS, false);
948 if (!msg)
949 return -ENOMEM;
950
951 msg->hdr.tid = cpu_to_le64(flush_tid);
952
953 fc = msg->front.iov_base;
954 memset(fc, 0, sizeof(*fc));
955
956 fc->cap_id = cpu_to_le64(cid);
957 fc->op = cpu_to_le32(op);
958 fc->seq = cpu_to_le32(seq);
959 fc->issue_seq = cpu_to_le32(issue_seq);
960 fc->migrate_seq = cpu_to_le32(mseq);
961 fc->caps = cpu_to_le32(caps);
962 fc->wanted = cpu_to_le32(wanted);
963 fc->dirty = cpu_to_le32(dirty);
964 fc->ino = cpu_to_le64(ino);
965 fc->snap_follows = cpu_to_le64(follows);
966
967 fc->size = cpu_to_le64(size);
968 fc->max_size = cpu_to_le64(max_size);
969 if (mtime)
970 ceph_encode_timespec(&fc->mtime, mtime);
971 if (atime)
972 ceph_encode_timespec(&fc->atime, atime);
973 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
974
975 fc->uid = cpu_to_le32(uid);
976 fc->gid = cpu_to_le32(gid);
977 fc->mode = cpu_to_le32(mode);
978
979 fc->xattr_version = cpu_to_le64(xattr_version);
980 if (xattrs_buf) {
981 msg->middle = ceph_buffer_get(xattrs_buf);
982 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
983 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
984 }
985
986 ceph_con_send(&session->s_con, msg);
987 return 0;
988}
989
990static void __queue_cap_release(struct ceph_mds_session *session,
991 u64 ino, u64 cap_id, u32 migrate_seq,
992 u32 issue_seq)
993{
994 struct ceph_msg *msg;
995 struct ceph_mds_cap_release *head;
996 struct ceph_mds_cap_item *item;
997
998 spin_lock(&session->s_cap_lock);
999 BUG_ON(!session->s_num_cap_releases);
1000 msg = list_first_entry(&session->s_cap_releases,
1001 struct ceph_msg, list_head);
1002
1003 dout(" adding %llx release to mds%d msg %p (%d left)\n",
1004 ino, session->s_mds, msg, session->s_num_cap_releases);
1005
1006 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1007 head = msg->front.iov_base;
1008 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1);
1009 item = msg->front.iov_base + msg->front.iov_len;
1010 item->ino = cpu_to_le64(ino);
1011 item->cap_id = cpu_to_le64(cap_id);
1012 item->migrate_seq = cpu_to_le32(migrate_seq);
1013 item->seq = cpu_to_le32(issue_seq);
1014
1015 session->s_num_cap_releases--;
1016
1017 msg->front.iov_len += sizeof(*item);
1018 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1019 dout(" release msg %p full\n", msg);
1020 list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1021 } else {
1022 dout(" release msg %p at %d/%d (%d)\n", msg,
1023 (int)le32_to_cpu(head->num),
1024 (int)CEPH_CAPS_PER_RELEASE,
1025 (int)msg->front.iov_len);
1026 }
1027 spin_unlock(&session->s_cap_lock);
1028}
1029
1030/*
1031 * Queue cap releases when an inode is dropped from our cache. Since
1032 * inode is about to be destroyed, there is no need for i_ceph_lock.
1033 */
1034void ceph_queue_caps_release(struct inode *inode)
1035{
1036 struct ceph_inode_info *ci = ceph_inode(inode);
1037 struct rb_node *p;
1038
1039 p = rb_first(&ci->i_caps);
1040 while (p) {
1041 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1042 struct ceph_mds_session *session = cap->session;
1043
1044 __queue_cap_release(session, ceph_ino(inode), cap->cap_id,
1045 cap->mseq, cap->issue_seq);
1046 p = rb_next(p);
1047 __ceph_remove_cap(cap);
1048 }
1049}
1050
1051/*
1052 * Send a cap msg on the given inode. Update our caps state, then
1053 * drop i_ceph_lock and send the message.
1054 *
1055 * Make note of max_size reported/requested from mds, revoked caps
1056 * that have now been implemented.
1057 *
1058 * Make half-hearted attempt ot to invalidate page cache if we are
1059 * dropping RDCACHE. Note that this will leave behind locked pages
1060 * that we'll then need to deal with elsewhere.
1061 *
1062 * Return non-zero if delayed release, or we experienced an error
1063 * such that the caller should requeue + retry later.
1064 *
1065 * called with i_ceph_lock, then drops it.
1066 * caller should hold snap_rwsem (read), s_mutex.
1067 */
1068static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1069 int op, int used, int want, int retain, int flushing,
1070 unsigned *pflush_tid)
1071 __releases(cap->ci->i_ceph_lock)
1072{
1073 struct ceph_inode_info *ci = cap->ci;
1074 struct inode *inode = &ci->vfs_inode;
1075 u64 cap_id = cap->cap_id;
1076 int held, revoking, dropping, keep;
1077 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1078 u64 size, max_size;
1079 struct timespec mtime, atime;
1080 int wake = 0;
1081 umode_t mode;
1082 uid_t uid;
1083 gid_t gid;
1084 struct ceph_mds_session *session;
1085 u64 xattr_version = 0;
1086 struct ceph_buffer *xattr_blob = NULL;
1087 int delayed = 0;
1088 u64 flush_tid = 0;
1089 int i;
1090 int ret;
1091
1092 held = cap->issued | cap->implemented;
1093 revoking = cap->implemented & ~cap->issued;
1094 retain &= ~revoking;
1095 dropping = cap->issued & ~retain;
1096
1097 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1098 inode, cap, cap->session,
1099 ceph_cap_string(held), ceph_cap_string(held & retain),
1100 ceph_cap_string(revoking));
1101 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1102
1103 session = cap->session;
1104
1105 /* don't release wanted unless we've waited a bit. */
1106 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1107 time_before(jiffies, ci->i_hold_caps_min)) {
1108 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1109 ceph_cap_string(cap->issued),
1110 ceph_cap_string(cap->issued & retain),
1111 ceph_cap_string(cap->mds_wanted),
1112 ceph_cap_string(want));
1113 want |= cap->mds_wanted;
1114 retain |= cap->issued;
1115 delayed = 1;
1116 }
1117 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1118
1119 cap->issued &= retain; /* drop bits we don't want */
1120 if (cap->implemented & ~cap->issued) {
1121 /*
1122 * Wake up any waiters on wanted -> needed transition.
1123 * This is due to the weird transition from buffered
1124 * to sync IO... we need to flush dirty pages _before_
1125 * allowing sync writes to avoid reordering.
1126 */
1127 wake = 1;
1128 }
1129 cap->implemented &= cap->issued | used;
1130 cap->mds_wanted = want;
1131
1132 if (flushing) {
1133 /*
1134 * assign a tid for flush operations so we can avoid
1135 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1136 * clean type races. track latest tid for every bit
1137 * so we can handle flush AxFw, flush Fw, and have the
1138 * first ack clean Ax.
1139 */
1140 flush_tid = ++ci->i_cap_flush_last_tid;
1141 if (pflush_tid)
1142 *pflush_tid = flush_tid;
1143 dout(" cap_flush_tid %d\n", (int)flush_tid);
1144 for (i = 0; i < CEPH_CAP_BITS; i++)
1145 if (flushing & (1 << i))
1146 ci->i_cap_flush_tid[i] = flush_tid;
1147
1148 follows = ci->i_head_snapc->seq;
1149 } else {
1150 follows = 0;
1151 }
1152
1153 keep = cap->implemented;
1154 seq = cap->seq;
1155 issue_seq = cap->issue_seq;
1156 mseq = cap->mseq;
1157 size = inode->i_size;
1158 ci->i_reported_size = size;
1159 max_size = ci->i_wanted_max_size;
1160 ci->i_requested_max_size = max_size;
1161 mtime = inode->i_mtime;
1162 atime = inode->i_atime;
1163 time_warp_seq = ci->i_time_warp_seq;
1164 uid = inode->i_uid;
1165 gid = inode->i_gid;
1166 mode = inode->i_mode;
1167
1168 if (flushing & CEPH_CAP_XATTR_EXCL) {
1169 __ceph_build_xattrs_blob(ci);
1170 xattr_blob = ci->i_xattrs.blob;
1171 xattr_version = ci->i_xattrs.version;
1172 }
1173
1174 spin_unlock(&ci->i_ceph_lock);
1175
1176 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1177 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1178 size, max_size, &mtime, &atime, time_warp_seq,
1179 uid, gid, mode, xattr_version, xattr_blob,
1180 follows);
1181 if (ret < 0) {
1182 dout("error sending cap msg, must requeue %p\n", inode);
1183 delayed = 1;
1184 }
1185
1186 if (wake)
1187 wake_up_all(&ci->i_cap_wq);
1188
1189 return delayed;
1190}
1191
1192/*
1193 * When a snapshot is taken, clients accumulate dirty metadata on
1194 * inodes with capabilities in ceph_cap_snaps to describe the file
1195 * state at the time the snapshot was taken. This must be flushed
1196 * asynchronously back to the MDS once sync writes complete and dirty
1197 * data is written out.
1198 *
1199 * Unless @again is true, skip cap_snaps that were already sent to
1200 * the MDS (i.e., during this session).
1201 *
1202 * Called under i_ceph_lock. Takes s_mutex as needed.
1203 */
1204void __ceph_flush_snaps(struct ceph_inode_info *ci,
1205 struct ceph_mds_session **psession,
1206 int again)
1207 __releases(ci->i_ceph_lock)
1208 __acquires(ci->i_ceph_lock)
1209{
1210 struct inode *inode = &ci->vfs_inode;
1211 int mds;
1212 struct ceph_cap_snap *capsnap;
1213 u32 mseq;
1214 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1215 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1216 session->s_mutex */
1217 u64 next_follows = 0; /* keep track of how far we've gotten through the
1218 i_cap_snaps list, and skip these entries next time
1219 around to avoid an infinite loop */
1220
1221 if (psession)
1222 session = *psession;
1223
1224 dout("__flush_snaps %p\n", inode);
1225retry:
1226 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1227 /* avoid an infiniute loop after retry */
1228 if (capsnap->follows < next_follows)
1229 continue;
1230 /*
1231 * we need to wait for sync writes to complete and for dirty
1232 * pages to be written out.
1233 */
1234 if (capsnap->dirty_pages || capsnap->writing)
1235 break;
1236
1237 /*
1238 * if cap writeback already occurred, we should have dropped
1239 * the capsnap in ceph_put_wrbuffer_cap_refs.
1240 */
1241 BUG_ON(capsnap->dirty == 0);
1242
1243 /* pick mds, take s_mutex */
1244 if (ci->i_auth_cap == NULL) {
1245 dout("no auth cap (migrating?), doing nothing\n");
1246 goto out;
1247 }
1248
1249 /* only flush each capsnap once */
1250 if (!again && !list_empty(&capsnap->flushing_item)) {
1251 dout("already flushed %p, skipping\n", capsnap);
1252 continue;
1253 }
1254
1255 mds = ci->i_auth_cap->session->s_mds;
1256 mseq = ci->i_auth_cap->mseq;
1257
1258 if (session && session->s_mds != mds) {
1259 dout("oops, wrong session %p mutex\n", session);
1260 mutex_unlock(&session->s_mutex);
1261 ceph_put_mds_session(session);
1262 session = NULL;
1263 }
1264 if (!session) {
1265 spin_unlock(&ci->i_ceph_lock);
1266 mutex_lock(&mdsc->mutex);
1267 session = __ceph_lookup_mds_session(mdsc, mds);
1268 mutex_unlock(&mdsc->mutex);
1269 if (session) {
1270 dout("inverting session/ino locks on %p\n",
1271 session);
1272 mutex_lock(&session->s_mutex);
1273 }
1274 /*
1275 * if session == NULL, we raced against a cap
1276 * deletion or migration. retry, and we'll
1277 * get a better @mds value next time.
1278 */
1279 spin_lock(&ci->i_ceph_lock);
1280 goto retry;
1281 }
1282
1283 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1284 atomic_inc(&capsnap->nref);
1285 if (!list_empty(&capsnap->flushing_item))
1286 list_del_init(&capsnap->flushing_item);
1287 list_add_tail(&capsnap->flushing_item,
1288 &session->s_cap_snaps_flushing);
1289 spin_unlock(&ci->i_ceph_lock);
1290
1291 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1292 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1293 send_cap_msg(session, ceph_vino(inode).ino, 0,
1294 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1295 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1296 capsnap->size, 0,
1297 &capsnap->mtime, &capsnap->atime,
1298 capsnap->time_warp_seq,
1299 capsnap->uid, capsnap->gid, capsnap->mode,
1300 capsnap->xattr_version, capsnap->xattr_blob,
1301 capsnap->follows);
1302
1303 next_follows = capsnap->follows + 1;
1304 ceph_put_cap_snap(capsnap);
1305
1306 spin_lock(&ci->i_ceph_lock);
1307 goto retry;
1308 }
1309
1310 /* we flushed them all; remove this inode from the queue */
1311 spin_lock(&mdsc->snap_flush_lock);
1312 list_del_init(&ci->i_snap_flush_item);
1313 spin_unlock(&mdsc->snap_flush_lock);
1314
1315out:
1316 if (psession)
1317 *psession = session;
1318 else if (session) {
1319 mutex_unlock(&session->s_mutex);
1320 ceph_put_mds_session(session);
1321 }
1322}
1323
1324static void ceph_flush_snaps(struct ceph_inode_info *ci)
1325{
1326 spin_lock(&ci->i_ceph_lock);
1327 __ceph_flush_snaps(ci, NULL, 0);
1328 spin_unlock(&ci->i_ceph_lock);
1329}
1330
1331/*
1332 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1333 * Caller is then responsible for calling __mark_inode_dirty with the
1334 * returned flags value.
1335 */
1336int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1337{
1338 struct ceph_mds_client *mdsc =
1339 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1340 struct inode *inode = &ci->vfs_inode;
1341 int was = ci->i_dirty_caps;
1342 int dirty = 0;
1343
1344 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1345 ceph_cap_string(mask), ceph_cap_string(was),
1346 ceph_cap_string(was | mask));
1347 ci->i_dirty_caps |= mask;
1348 if (was == 0) {
1349 if (!ci->i_head_snapc)
1350 ci->i_head_snapc = ceph_get_snap_context(
1351 ci->i_snap_realm->cached_context);
1352 dout(" inode %p now dirty snapc %p\n", &ci->vfs_inode,
1353 ci->i_head_snapc);
1354 BUG_ON(!list_empty(&ci->i_dirty_item));
1355 spin_lock(&mdsc->cap_dirty_lock);
1356 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1357 spin_unlock(&mdsc->cap_dirty_lock);
1358 if (ci->i_flushing_caps == 0) {
1359 ihold(inode);
1360 dirty |= I_DIRTY_SYNC;
1361 }
1362 }
1363 BUG_ON(list_empty(&ci->i_dirty_item));
1364 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1365 (mask & CEPH_CAP_FILE_BUFFER))
1366 dirty |= I_DIRTY_DATASYNC;
1367 __cap_delay_requeue(mdsc, ci);
1368 return dirty;
1369}
1370
1371/*
1372 * Add dirty inode to the flushing list. Assigned a seq number so we
1373 * can wait for caps to flush without starving.
1374 *
1375 * Called under i_ceph_lock.
1376 */
1377static int __mark_caps_flushing(struct inode *inode,
1378 struct ceph_mds_session *session)
1379{
1380 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1381 struct ceph_inode_info *ci = ceph_inode(inode);
1382 int flushing;
1383
1384 BUG_ON(ci->i_dirty_caps == 0);
1385 BUG_ON(list_empty(&ci->i_dirty_item));
1386
1387 flushing = ci->i_dirty_caps;
1388 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1389 ceph_cap_string(flushing),
1390 ceph_cap_string(ci->i_flushing_caps),
1391 ceph_cap_string(ci->i_flushing_caps | flushing));
1392 ci->i_flushing_caps |= flushing;
1393 ci->i_dirty_caps = 0;
1394 dout(" inode %p now !dirty\n", inode);
1395
1396 spin_lock(&mdsc->cap_dirty_lock);
1397 list_del_init(&ci->i_dirty_item);
1398
1399 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1400 if (list_empty(&ci->i_flushing_item)) {
1401 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1402 mdsc->num_cap_flushing++;
1403 dout(" inode %p now flushing seq %lld\n", inode,
1404 ci->i_cap_flush_seq);
1405 } else {
1406 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1407 dout(" inode %p now flushing (more) seq %lld\n", inode,
1408 ci->i_cap_flush_seq);
1409 }
1410 spin_unlock(&mdsc->cap_dirty_lock);
1411
1412 return flushing;
1413}
1414
1415/*
1416 * try to invalidate mapping pages without blocking.
1417 */
1418static int try_nonblocking_invalidate(struct inode *inode)
1419{
1420 struct ceph_inode_info *ci = ceph_inode(inode);
1421 u32 invalidating_gen = ci->i_rdcache_gen;
1422
1423 spin_unlock(&ci->i_ceph_lock);
1424 invalidate_mapping_pages(&inode->i_data, 0, -1);
1425 spin_lock(&ci->i_ceph_lock);
1426
1427 if (inode->i_data.nrpages == 0 &&
1428 invalidating_gen == ci->i_rdcache_gen) {
1429 /* success. */
1430 dout("try_nonblocking_invalidate %p success\n", inode);
1431 /* save any racing async invalidate some trouble */
1432 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1433 return 0;
1434 }
1435 dout("try_nonblocking_invalidate %p failed\n", inode);
1436 return -1;
1437}
1438
1439/*
1440 * Swiss army knife function to examine currently used and wanted
1441 * versus held caps. Release, flush, ack revoked caps to mds as
1442 * appropriate.
1443 *
1444 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1445 * cap release further.
1446 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1447 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1448 * further delay.
1449 */
1450void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1451 struct ceph_mds_session *session)
1452{
1453 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1454 struct ceph_mds_client *mdsc = fsc->mdsc;
1455 struct inode *inode = &ci->vfs_inode;
1456 struct ceph_cap *cap;
1457 int file_wanted, used;
1458 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1459 int issued, implemented, want, retain, revoking, flushing = 0;
1460 int mds = -1; /* keep track of how far we've gone through i_caps list
1461 to avoid an infinite loop on retry */
1462 struct rb_node *p;
1463 int tried_invalidate = 0;
1464 int delayed = 0, sent = 0, force_requeue = 0, num;
1465 int queue_invalidate = 0;
1466 int is_delayed = flags & CHECK_CAPS_NODELAY;
1467
1468 /* if we are unmounting, flush any unused caps immediately. */
1469 if (mdsc->stopping)
1470 is_delayed = 1;
1471
1472 spin_lock(&ci->i_ceph_lock);
1473
1474 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1475 flags |= CHECK_CAPS_FLUSH;
1476
1477 /* flush snaps first time around only */
1478 if (!list_empty(&ci->i_cap_snaps))
1479 __ceph_flush_snaps(ci, &session, 0);
1480 goto retry_locked;
1481retry:
1482 spin_lock(&ci->i_ceph_lock);
1483retry_locked:
1484 file_wanted = __ceph_caps_file_wanted(ci);
1485 used = __ceph_caps_used(ci);
1486 want = file_wanted | used;
1487 issued = __ceph_caps_issued(ci, &implemented);
1488 revoking = implemented & ~issued;
1489
1490 retain = want | CEPH_CAP_PIN;
1491 if (!mdsc->stopping && inode->i_nlink > 0) {
1492 if (want) {
1493 retain |= CEPH_CAP_ANY; /* be greedy */
1494 } else {
1495 retain |= CEPH_CAP_ANY_SHARED;
1496 /*
1497 * keep RD only if we didn't have the file open RW,
1498 * because then the mds would revoke it anyway to
1499 * journal max_size=0.
1500 */
1501 if (ci->i_max_size == 0)
1502 retain |= CEPH_CAP_ANY_RD;
1503 }
1504 }
1505
1506 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1507 " issued %s revoking %s retain %s %s%s%s\n", inode,
1508 ceph_cap_string(file_wanted),
1509 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1510 ceph_cap_string(ci->i_flushing_caps),
1511 ceph_cap_string(issued), ceph_cap_string(revoking),
1512 ceph_cap_string(retain),
1513 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1514 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1515 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1516
1517 /*
1518 * If we no longer need to hold onto old our caps, and we may
1519 * have cached pages, but don't want them, then try to invalidate.
1520 * If we fail, it's because pages are locked.... try again later.
1521 */
1522 if ((!is_delayed || mdsc->stopping) &&
1523 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1524 inode->i_data.nrpages && /* have cached pages */
1525 (file_wanted == 0 || /* no open files */
1526 (revoking & (CEPH_CAP_FILE_CACHE|
1527 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */
1528 !tried_invalidate) {
1529 dout("check_caps trying to invalidate on %p\n", inode);
1530 if (try_nonblocking_invalidate(inode) < 0) {
1531 if (revoking & (CEPH_CAP_FILE_CACHE|
1532 CEPH_CAP_FILE_LAZYIO)) {
1533 dout("check_caps queuing invalidate\n");
1534 queue_invalidate = 1;
1535 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1536 } else {
1537 dout("check_caps failed to invalidate pages\n");
1538 /* we failed to invalidate pages. check these
1539 caps again later. */
1540 force_requeue = 1;
1541 __cap_set_timeouts(mdsc, ci);
1542 }
1543 }
1544 tried_invalidate = 1;
1545 goto retry_locked;
1546 }
1547
1548 num = 0;
1549 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1550 cap = rb_entry(p, struct ceph_cap, ci_node);
1551 num++;
1552
1553 /* avoid looping forever */
1554 if (mds >= cap->mds ||
1555 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1556 continue;
1557
1558 /* NOTE: no side-effects allowed, until we take s_mutex */
1559
1560 revoking = cap->implemented & ~cap->issued;
1561 dout(" mds%d cap %p issued %s implemented %s revoking %s\n",
1562 cap->mds, cap, ceph_cap_string(cap->issued),
1563 ceph_cap_string(cap->implemented),
1564 ceph_cap_string(revoking));
1565
1566 if (cap == ci->i_auth_cap &&
1567 (cap->issued & CEPH_CAP_FILE_WR)) {
1568 /* request larger max_size from MDS? */
1569 if (ci->i_wanted_max_size > ci->i_max_size &&
1570 ci->i_wanted_max_size > ci->i_requested_max_size) {
1571 dout("requesting new max_size\n");
1572 goto ack;
1573 }
1574
1575 /* approaching file_max? */
1576 if ((inode->i_size << 1) >= ci->i_max_size &&
1577 (ci->i_reported_size << 1) < ci->i_max_size) {
1578 dout("i_size approaching max_size\n");
1579 goto ack;
1580 }
1581 }
1582 /* flush anything dirty? */
1583 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1584 ci->i_dirty_caps) {
1585 dout("flushing dirty caps\n");
1586 goto ack;
1587 }
1588
1589 /* completed revocation? going down and there are no caps? */
1590 if (revoking && (revoking & used) == 0) {
1591 dout("completed revocation of %s\n",
1592 ceph_cap_string(cap->implemented & ~cap->issued));
1593 goto ack;
1594 }
1595
1596 /* want more caps from mds? */
1597 if (want & ~(cap->mds_wanted | cap->issued))
1598 goto ack;
1599
1600 /* things we might delay */
1601 if ((cap->issued & ~retain) == 0 &&
1602 cap->mds_wanted == want)
1603 continue; /* nope, all good */
1604
1605 if (is_delayed)
1606 goto ack;
1607
1608 /* delay? */
1609 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1610 time_before(jiffies, ci->i_hold_caps_max)) {
1611 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1612 ceph_cap_string(cap->issued),
1613 ceph_cap_string(cap->issued & retain),
1614 ceph_cap_string(cap->mds_wanted),
1615 ceph_cap_string(want));
1616 delayed++;
1617 continue;
1618 }
1619
1620ack:
1621 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1622 dout(" skipping %p I_NOFLUSH set\n", inode);
1623 continue;
1624 }
1625
1626 if (session && session != cap->session) {
1627 dout("oops, wrong session %p mutex\n", session);
1628 mutex_unlock(&session->s_mutex);
1629 session = NULL;
1630 }
1631 if (!session) {
1632 session = cap->session;
1633 if (mutex_trylock(&session->s_mutex) == 0) {
1634 dout("inverting session/ino locks on %p\n",
1635 session);
1636 spin_unlock(&ci->i_ceph_lock);
1637 if (took_snap_rwsem) {
1638 up_read(&mdsc->snap_rwsem);
1639 took_snap_rwsem = 0;
1640 }
1641 mutex_lock(&session->s_mutex);
1642 goto retry;
1643 }
1644 }
1645 /* take snap_rwsem after session mutex */
1646 if (!took_snap_rwsem) {
1647 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1648 dout("inverting snap/in locks on %p\n",
1649 inode);
1650 spin_unlock(&ci->i_ceph_lock);
1651 down_read(&mdsc->snap_rwsem);
1652 took_snap_rwsem = 1;
1653 goto retry;
1654 }
1655 took_snap_rwsem = 1;
1656 }
1657
1658 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1659 flushing = __mark_caps_flushing(inode, session);
1660 else
1661 flushing = 0;
1662
1663 mds = cap->mds; /* remember mds, so we don't repeat */
1664 sent++;
1665
1666 /* __send_cap drops i_ceph_lock */
1667 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want,
1668 retain, flushing, NULL);
1669 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1670 }
1671
1672 /*
1673 * Reschedule delayed caps release if we delayed anything,
1674 * otherwise cancel.
1675 */
1676 if (delayed && is_delayed)
1677 force_requeue = 1; /* __send_cap delayed release; requeue */
1678 if (!delayed && !is_delayed)
1679 __cap_delay_cancel(mdsc, ci);
1680 else if (!is_delayed || force_requeue)
1681 __cap_delay_requeue(mdsc, ci);
1682
1683 spin_unlock(&ci->i_ceph_lock);
1684
1685 if (queue_invalidate)
1686 ceph_queue_invalidate(inode);
1687
1688 if (session)
1689 mutex_unlock(&session->s_mutex);
1690 if (took_snap_rwsem)
1691 up_read(&mdsc->snap_rwsem);
1692}
1693
1694/*
1695 * Try to flush dirty caps back to the auth mds.
1696 */
1697static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session,
1698 unsigned *flush_tid)
1699{
1700 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1701 struct ceph_inode_info *ci = ceph_inode(inode);
1702 int unlock_session = session ? 0 : 1;
1703 int flushing = 0;
1704
1705retry:
1706 spin_lock(&ci->i_ceph_lock);
1707 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1708 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1709 goto out;
1710 }
1711 if (ci->i_dirty_caps && ci->i_auth_cap) {
1712 struct ceph_cap *cap = ci->i_auth_cap;
1713 int used = __ceph_caps_used(ci);
1714 int want = __ceph_caps_wanted(ci);
1715 int delayed;
1716
1717 if (!session) {
1718 spin_unlock(&ci->i_ceph_lock);
1719 session = cap->session;
1720 mutex_lock(&session->s_mutex);
1721 goto retry;
1722 }
1723 BUG_ON(session != cap->session);
1724 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1725 goto out;
1726
1727 flushing = __mark_caps_flushing(inode, session);
1728
1729 /* __send_cap drops i_ceph_lock */
1730 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1731 cap->issued | cap->implemented, flushing,
1732 flush_tid);
1733 if (!delayed)
1734 goto out_unlocked;
1735
1736 spin_lock(&ci->i_ceph_lock);
1737 __cap_delay_requeue(mdsc, ci);
1738 }
1739out:
1740 spin_unlock(&ci->i_ceph_lock);
1741out_unlocked:
1742 if (session && unlock_session)
1743 mutex_unlock(&session->s_mutex);
1744 return flushing;
1745}
1746
1747/*
1748 * Return true if we've flushed caps through the given flush_tid.
1749 */
1750static int caps_are_flushed(struct inode *inode, unsigned tid)
1751{
1752 struct ceph_inode_info *ci = ceph_inode(inode);
1753 int i, ret = 1;
1754
1755 spin_lock(&ci->i_ceph_lock);
1756 for (i = 0; i < CEPH_CAP_BITS; i++)
1757 if ((ci->i_flushing_caps & (1 << i)) &&
1758 ci->i_cap_flush_tid[i] <= tid) {
1759 /* still flushing this bit */
1760 ret = 0;
1761 break;
1762 }
1763 spin_unlock(&ci->i_ceph_lock);
1764 return ret;
1765}
1766
1767/*
1768 * Wait on any unsafe replies for the given inode. First wait on the
1769 * newest request, and make that the upper bound. Then, if there are
1770 * more requests, keep waiting on the oldest as long as it is still older
1771 * than the original request.
1772 */
1773static void sync_write_wait(struct inode *inode)
1774{
1775 struct ceph_inode_info *ci = ceph_inode(inode);
1776 struct list_head *head = &ci->i_unsafe_writes;
1777 struct ceph_osd_request *req;
1778 u64 last_tid;
1779
1780 spin_lock(&ci->i_unsafe_lock);
1781 if (list_empty(head))
1782 goto out;
1783
1784 /* set upper bound as _last_ entry in chain */
1785 req = list_entry(head->prev, struct ceph_osd_request,
1786 r_unsafe_item);
1787 last_tid = req->r_tid;
1788
1789 do {
1790 ceph_osdc_get_request(req);
1791 spin_unlock(&ci->i_unsafe_lock);
1792 dout("sync_write_wait on tid %llu (until %llu)\n",
1793 req->r_tid, last_tid);
1794 wait_for_completion(&req->r_safe_completion);
1795 spin_lock(&ci->i_unsafe_lock);
1796 ceph_osdc_put_request(req);
1797
1798 /*
1799 * from here on look at first entry in chain, since we
1800 * only want to wait for anything older than last_tid
1801 */
1802 if (list_empty(head))
1803 break;
1804 req = list_entry(head->next, struct ceph_osd_request,
1805 r_unsafe_item);
1806 } while (req->r_tid < last_tid);
1807out:
1808 spin_unlock(&ci->i_unsafe_lock);
1809}
1810
1811int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1812{
1813 struct inode *inode = file->f_mapping->host;
1814 struct ceph_inode_info *ci = ceph_inode(inode);
1815 unsigned flush_tid;
1816 int ret;
1817 int dirty;
1818
1819 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1820 sync_write_wait(inode);
1821
1822 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1823 if (ret < 0)
1824 return ret;
1825 mutex_lock(&inode->i_mutex);
1826
1827 dirty = try_flush_caps(inode, NULL, &flush_tid);
1828 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1829
1830 /*
1831 * only wait on non-file metadata writeback (the mds
1832 * can recover size and mtime, so we don't need to
1833 * wait for that)
1834 */
1835 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1836 dout("fsync waiting for flush_tid %u\n", flush_tid);
1837 ret = wait_event_interruptible(ci->i_cap_wq,
1838 caps_are_flushed(inode, flush_tid));
1839 }
1840
1841 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1842 mutex_unlock(&inode->i_mutex);
1843 return ret;
1844}
1845
1846/*
1847 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1848 * queue inode for flush but don't do so immediately, because we can
1849 * get by with fewer MDS messages if we wait for data writeback to
1850 * complete first.
1851 */
1852int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1853{
1854 struct ceph_inode_info *ci = ceph_inode(inode);
1855 unsigned flush_tid;
1856 int err = 0;
1857 int dirty;
1858 int wait = wbc->sync_mode == WB_SYNC_ALL;
1859
1860 dout("write_inode %p wait=%d\n", inode, wait);
1861 if (wait) {
1862 dirty = try_flush_caps(inode, NULL, &flush_tid);
1863 if (dirty)
1864 err = wait_event_interruptible(ci->i_cap_wq,
1865 caps_are_flushed(inode, flush_tid));
1866 } else {
1867 struct ceph_mds_client *mdsc =
1868 ceph_sb_to_client(inode->i_sb)->mdsc;
1869
1870 spin_lock(&ci->i_ceph_lock);
1871 if (__ceph_caps_dirty(ci))
1872 __cap_delay_requeue_front(mdsc, ci);
1873 spin_unlock(&ci->i_ceph_lock);
1874 }
1875 return err;
1876}
1877
1878/*
1879 * After a recovering MDS goes active, we need to resend any caps
1880 * we were flushing.
1881 *
1882 * Caller holds session->s_mutex.
1883 */
1884static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1885 struct ceph_mds_session *session)
1886{
1887 struct ceph_cap_snap *capsnap;
1888
1889 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1890 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1891 flushing_item) {
1892 struct ceph_inode_info *ci = capsnap->ci;
1893 struct inode *inode = &ci->vfs_inode;
1894 struct ceph_cap *cap;
1895
1896 spin_lock(&ci->i_ceph_lock);
1897 cap = ci->i_auth_cap;
1898 if (cap && cap->session == session) {
1899 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1900 cap, capsnap);
1901 __ceph_flush_snaps(ci, &session, 1);
1902 } else {
1903 pr_err("%p auth cap %p not mds%d ???\n", inode,
1904 cap, session->s_mds);
1905 }
1906 spin_unlock(&ci->i_ceph_lock);
1907 }
1908}
1909
1910void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1911 struct ceph_mds_session *session)
1912{
1913 struct ceph_inode_info *ci;
1914
1915 kick_flushing_capsnaps(mdsc, session);
1916
1917 dout("kick_flushing_caps mds%d\n", session->s_mds);
1918 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1919 struct inode *inode = &ci->vfs_inode;
1920 struct ceph_cap *cap;
1921 int delayed = 0;
1922
1923 spin_lock(&ci->i_ceph_lock);
1924 cap = ci->i_auth_cap;
1925 if (cap && cap->session == session) {
1926 dout("kick_flushing_caps %p cap %p %s\n", inode,
1927 cap, ceph_cap_string(ci->i_flushing_caps));
1928 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1929 __ceph_caps_used(ci),
1930 __ceph_caps_wanted(ci),
1931 cap->issued | cap->implemented,
1932 ci->i_flushing_caps, NULL);
1933 if (delayed) {
1934 spin_lock(&ci->i_ceph_lock);
1935 __cap_delay_requeue(mdsc, ci);
1936 spin_unlock(&ci->i_ceph_lock);
1937 }
1938 } else {
1939 pr_err("%p auth cap %p not mds%d ???\n", inode,
1940 cap, session->s_mds);
1941 spin_unlock(&ci->i_ceph_lock);
1942 }
1943 }
1944}
1945
1946static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
1947 struct ceph_mds_session *session,
1948 struct inode *inode)
1949{
1950 struct ceph_inode_info *ci = ceph_inode(inode);
1951 struct ceph_cap *cap;
1952 int delayed = 0;
1953
1954 spin_lock(&ci->i_ceph_lock);
1955 cap = ci->i_auth_cap;
1956 dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode,
1957 ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq);
1958 __ceph_flush_snaps(ci, &session, 1);
1959 if (ci->i_flushing_caps) {
1960 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1961 __ceph_caps_used(ci),
1962 __ceph_caps_wanted(ci),
1963 cap->issued | cap->implemented,
1964 ci->i_flushing_caps, NULL);
1965 if (delayed) {
1966 spin_lock(&ci->i_ceph_lock);
1967 __cap_delay_requeue(mdsc, ci);
1968 spin_unlock(&ci->i_ceph_lock);
1969 }
1970 } else {
1971 spin_unlock(&ci->i_ceph_lock);
1972 }
1973}
1974
1975
1976/*
1977 * Take references to capabilities we hold, so that we don't release
1978 * them to the MDS prematurely.
1979 *
1980 * Protected by i_ceph_lock.
1981 */
1982static void __take_cap_refs(struct ceph_inode_info *ci, int got)
1983{
1984 if (got & CEPH_CAP_PIN)
1985 ci->i_pin_ref++;
1986 if (got & CEPH_CAP_FILE_RD)
1987 ci->i_rd_ref++;
1988 if (got & CEPH_CAP_FILE_CACHE)
1989 ci->i_rdcache_ref++;
1990 if (got & CEPH_CAP_FILE_WR)
1991 ci->i_wr_ref++;
1992 if (got & CEPH_CAP_FILE_BUFFER) {
1993 if (ci->i_wb_ref == 0)
1994 ihold(&ci->vfs_inode);
1995 ci->i_wb_ref++;
1996 dout("__take_cap_refs %p wb %d -> %d (?)\n",
1997 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
1998 }
1999}
2000
2001/*
2002 * Try to grab cap references. Specify those refs we @want, and the
2003 * minimal set we @need. Also include the larger offset we are writing
2004 * to (when applicable), and check against max_size here as well.
2005 * Note that caller is responsible for ensuring max_size increases are
2006 * requested from the MDS.
2007 */
2008static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2009 int *got, loff_t endoff, int *check_max, int *err)
2010{
2011 struct inode *inode = &ci->vfs_inode;
2012 int ret = 0;
2013 int have, implemented;
2014 int file_wanted;
2015
2016 dout("get_cap_refs %p need %s want %s\n", inode,
2017 ceph_cap_string(need), ceph_cap_string(want));
2018 spin_lock(&ci->i_ceph_lock);
2019
2020 /* make sure file is actually open */
2021 file_wanted = __ceph_caps_file_wanted(ci);
2022 if ((file_wanted & need) == 0) {
2023 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2024 ceph_cap_string(need), ceph_cap_string(file_wanted));
2025 *err = -EBADF;
2026 ret = 1;
2027 goto out;
2028 }
2029
2030 if (need & CEPH_CAP_FILE_WR) {
2031 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2032 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2033 inode, endoff, ci->i_max_size);
2034 if (endoff > ci->i_wanted_max_size) {
2035 *check_max = 1;
2036 ret = 1;
2037 }
2038 goto out;
2039 }
2040 /*
2041 * If a sync write is in progress, we must wait, so that we
2042 * can get a final snapshot value for size+mtime.
2043 */
2044 if (__ceph_have_pending_cap_snap(ci)) {
2045 dout("get_cap_refs %p cap_snap_pending\n", inode);
2046 goto out;
2047 }
2048 }
2049 have = __ceph_caps_issued(ci, &implemented);
2050
2051 /*
2052 * disallow writes while a truncate is pending
2053 */
2054 if (ci->i_truncate_pending)
2055 have &= ~CEPH_CAP_FILE_WR;
2056
2057 if ((have & need) == need) {
2058 /*
2059 * Look at (implemented & ~have & not) so that we keep waiting
2060 * on transition from wanted -> needed caps. This is needed
2061 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2062 * going before a prior buffered writeback happens.
2063 */
2064 int not = want & ~(have & need);
2065 int revoking = implemented & ~have;
2066 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2067 inode, ceph_cap_string(have), ceph_cap_string(not),
2068 ceph_cap_string(revoking));
2069 if ((revoking & not) == 0) {
2070 *got = need | (have & want);
2071 __take_cap_refs(ci, *got);
2072 ret = 1;
2073 }
2074 } else {
2075 dout("get_cap_refs %p have %s needed %s\n", inode,
2076 ceph_cap_string(have), ceph_cap_string(need));
2077 }
2078out:
2079 spin_unlock(&ci->i_ceph_lock);
2080 dout("get_cap_refs %p ret %d got %s\n", inode,
2081 ret, ceph_cap_string(*got));
2082 return ret;
2083}
2084
2085/*
2086 * Check the offset we are writing up to against our current
2087 * max_size. If necessary, tell the MDS we want to write to
2088 * a larger offset.
2089 */
2090static void check_max_size(struct inode *inode, loff_t endoff)
2091{
2092 struct ceph_inode_info *ci = ceph_inode(inode);
2093 int check = 0;
2094
2095 /* do we need to explicitly request a larger max_size? */
2096 spin_lock(&ci->i_ceph_lock);
2097 if ((endoff >= ci->i_max_size ||
2098 endoff > (inode->i_size << 1)) &&
2099 endoff > ci->i_wanted_max_size) {
2100 dout("write %p at large endoff %llu, req max_size\n",
2101 inode, endoff);
2102 ci->i_wanted_max_size = endoff;
2103 check = 1;
2104 }
2105 spin_unlock(&ci->i_ceph_lock);
2106 if (check)
2107 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2108}
2109
2110/*
2111 * Wait for caps, and take cap references. If we can't get a WR cap
2112 * due to a small max_size, make sure we check_max_size (and possibly
2113 * ask the mds) so we don't get hung up indefinitely.
2114 */
2115int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
2116 loff_t endoff)
2117{
2118 int check_max, ret, err;
2119
2120retry:
2121 if (endoff > 0)
2122 check_max_size(&ci->vfs_inode, endoff);
2123 check_max = 0;
2124 err = 0;
2125 ret = wait_event_interruptible(ci->i_cap_wq,
2126 try_get_cap_refs(ci, need, want,
2127 got, endoff,
2128 &check_max, &err));
2129 if (err)
2130 ret = err;
2131 if (check_max)
2132 goto retry;
2133 return ret;
2134}
2135
2136/*
2137 * Take cap refs. Caller must already know we hold at least one ref
2138 * on the caps in question or we don't know this is safe.
2139 */
2140void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2141{
2142 spin_lock(&ci->i_ceph_lock);
2143 __take_cap_refs(ci, caps);
2144 spin_unlock(&ci->i_ceph_lock);
2145}
2146
2147/*
2148 * Release cap refs.
2149 *
2150 * If we released the last ref on any given cap, call ceph_check_caps
2151 * to release (or schedule a release).
2152 *
2153 * If we are releasing a WR cap (from a sync write), finalize any affected
2154 * cap_snap, and wake up any waiters.
2155 */
2156void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2157{
2158 struct inode *inode = &ci->vfs_inode;
2159 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2160 struct ceph_cap_snap *capsnap;
2161
2162 spin_lock(&ci->i_ceph_lock);
2163 if (had & CEPH_CAP_PIN)
2164 --ci->i_pin_ref;
2165 if (had & CEPH_CAP_FILE_RD)
2166 if (--ci->i_rd_ref == 0)
2167 last++;
2168 if (had & CEPH_CAP_FILE_CACHE)
2169 if (--ci->i_rdcache_ref == 0)
2170 last++;
2171 if (had & CEPH_CAP_FILE_BUFFER) {
2172 if (--ci->i_wb_ref == 0) {
2173 last++;
2174 put++;
2175 }
2176 dout("put_cap_refs %p wb %d -> %d (?)\n",
2177 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2178 }
2179 if (had & CEPH_CAP_FILE_WR)
2180 if (--ci->i_wr_ref == 0) {
2181 last++;
2182 if (!list_empty(&ci->i_cap_snaps)) {
2183 capsnap = list_first_entry(&ci->i_cap_snaps,
2184 struct ceph_cap_snap,
2185 ci_item);
2186 if (capsnap->writing) {
2187 capsnap->writing = 0;
2188 flushsnaps =
2189 __ceph_finish_cap_snap(ci,
2190 capsnap);
2191 wake = 1;
2192 }
2193 }
2194 }
2195 spin_unlock(&ci->i_ceph_lock);
2196
2197 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2198 last ? " last" : "", put ? " put" : "");
2199
2200 if (last && !flushsnaps)
2201 ceph_check_caps(ci, 0, NULL);
2202 else if (flushsnaps)
2203 ceph_flush_snaps(ci);
2204 if (wake)
2205 wake_up_all(&ci->i_cap_wq);
2206 if (put)
2207 iput(inode);
2208}
2209
2210/*
2211 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2212 * context. Adjust per-snap dirty page accounting as appropriate.
2213 * Once all dirty data for a cap_snap is flushed, flush snapped file
2214 * metadata back to the MDS. If we dropped the last ref, call
2215 * ceph_check_caps.
2216 */
2217void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2218 struct ceph_snap_context *snapc)
2219{
2220 struct inode *inode = &ci->vfs_inode;
2221 int last = 0;
2222 int complete_capsnap = 0;
2223 int drop_capsnap = 0;
2224 int found = 0;
2225 struct ceph_cap_snap *capsnap = NULL;
2226
2227 spin_lock(&ci->i_ceph_lock);
2228 ci->i_wrbuffer_ref -= nr;
2229 last = !ci->i_wrbuffer_ref;
2230
2231 if (ci->i_head_snapc == snapc) {
2232 ci->i_wrbuffer_ref_head -= nr;
2233 if (ci->i_wrbuffer_ref_head == 0 &&
2234 ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) {
2235 BUG_ON(!ci->i_head_snapc);
2236 ceph_put_snap_context(ci->i_head_snapc);
2237 ci->i_head_snapc = NULL;
2238 }
2239 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2240 inode,
2241 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2242 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2243 last ? " LAST" : "");
2244 } else {
2245 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2246 if (capsnap->context == snapc) {
2247 found = 1;
2248 break;
2249 }
2250 }
2251 BUG_ON(!found);
2252 capsnap->dirty_pages -= nr;
2253 if (capsnap->dirty_pages == 0) {
2254 complete_capsnap = 1;
2255 if (capsnap->dirty == 0)
2256 /* cap writeback completed before we created
2257 * the cap_snap; no FLUSHSNAP is needed */
2258 drop_capsnap = 1;
2259 }
2260 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2261 " snap %lld %d/%d -> %d/%d %s%s%s\n",
2262 inode, capsnap, capsnap->context->seq,
2263 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2264 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2265 last ? " (wrbuffer last)" : "",
2266 complete_capsnap ? " (complete capsnap)" : "",
2267 drop_capsnap ? " (drop capsnap)" : "");
2268 if (drop_capsnap) {
2269 ceph_put_snap_context(capsnap->context);
2270 list_del(&capsnap->ci_item);
2271 list_del(&capsnap->flushing_item);
2272 ceph_put_cap_snap(capsnap);
2273 }
2274 }
2275
2276 spin_unlock(&ci->i_ceph_lock);
2277
2278 if (last) {
2279 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2280 iput(inode);
2281 } else if (complete_capsnap) {
2282 ceph_flush_snaps(ci);
2283 wake_up_all(&ci->i_cap_wq);
2284 }
2285 if (drop_capsnap)
2286 iput(inode);
2287}
2288
2289/*
2290 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2291 * actually be a revocation if it specifies a smaller cap set.)
2292 *
2293 * caller holds s_mutex and i_ceph_lock, we drop both.
2294 *
2295 * return value:
2296 * 0 - ok
2297 * 1 - check_caps on auth cap only (writeback)
2298 * 2 - check_caps (ack revoke)
2299 */
2300static void handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2301 struct ceph_mds_session *session,
2302 struct ceph_cap *cap,
2303 struct ceph_buffer *xattr_buf)
2304 __releases(ci->i_ceph_lock)
2305{
2306 struct ceph_inode_info *ci = ceph_inode(inode);
2307 int mds = session->s_mds;
2308 int seq = le32_to_cpu(grant->seq);
2309 int newcaps = le32_to_cpu(grant->caps);
2310 int issued, implemented, used, wanted, dirty;
2311 u64 size = le64_to_cpu(grant->size);
2312 u64 max_size = le64_to_cpu(grant->max_size);
2313 struct timespec mtime, atime, ctime;
2314 int check_caps = 0;
2315 int wake = 0;
2316 int writeback = 0;
2317 int revoked_rdcache = 0;
2318 int queue_invalidate = 0;
2319
2320 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2321 inode, cap, mds, seq, ceph_cap_string(newcaps));
2322 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2323 inode->i_size);
2324
2325 /*
2326 * If CACHE is being revoked, and we have no dirty buffers,
2327 * try to invalidate (once). (If there are dirty buffers, we
2328 * will invalidate _after_ writeback.)
2329 */
2330 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2331 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2332 !ci->i_wrbuffer_ref) {
2333 if (try_nonblocking_invalidate(inode) == 0) {
2334 revoked_rdcache = 1;
2335 } else {
2336 /* there were locked pages.. invalidate later
2337 in a separate thread. */
2338 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2339 queue_invalidate = 1;
2340 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2341 }
2342 }
2343 }
2344
2345 /* side effects now are allowed */
2346
2347 issued = __ceph_caps_issued(ci, &implemented);
2348 issued |= implemented | __ceph_caps_dirty(ci);
2349
2350 cap->cap_gen = session->s_cap_gen;
2351
2352 __check_cap_issue(ci, cap, newcaps);
2353
2354 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2355 inode->i_mode = le32_to_cpu(grant->mode);
2356 inode->i_uid = le32_to_cpu(grant->uid);
2357 inode->i_gid = le32_to_cpu(grant->gid);
2358 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2359 inode->i_uid, inode->i_gid);
2360 }
2361
2362 if ((issued & CEPH_CAP_LINK_EXCL) == 0)
2363 set_nlink(inode, le32_to_cpu(grant->nlink));
2364
2365 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2366 int len = le32_to_cpu(grant->xattr_len);
2367 u64 version = le64_to_cpu(grant->xattr_version);
2368
2369 if (version > ci->i_xattrs.version) {
2370 dout(" got new xattrs v%llu on %p len %d\n",
2371 version, inode, len);
2372 if (ci->i_xattrs.blob)
2373 ceph_buffer_put(ci->i_xattrs.blob);
2374 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2375 ci->i_xattrs.version = version;
2376 }
2377 }
2378
2379 /* size/ctime/mtime/atime? */
2380 ceph_fill_file_size(inode, issued,
2381 le32_to_cpu(grant->truncate_seq),
2382 le64_to_cpu(grant->truncate_size), size);
2383 ceph_decode_timespec(&mtime, &grant->mtime);
2384 ceph_decode_timespec(&atime, &grant->atime);
2385 ceph_decode_timespec(&ctime, &grant->ctime);
2386 ceph_fill_file_time(inode, issued,
2387 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2388 &atime);
2389
2390 /* max size increase? */
2391 if (max_size != ci->i_max_size) {
2392 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2393 ci->i_max_size = max_size;
2394 if (max_size >= ci->i_wanted_max_size) {
2395 ci->i_wanted_max_size = 0; /* reset */
2396 ci->i_requested_max_size = 0;
2397 }
2398 wake = 1;
2399 }
2400
2401 /* check cap bits */
2402 wanted = __ceph_caps_wanted(ci);
2403 used = __ceph_caps_used(ci);
2404 dirty = __ceph_caps_dirty(ci);
2405 dout(" my wanted = %s, used = %s, dirty %s\n",
2406 ceph_cap_string(wanted),
2407 ceph_cap_string(used),
2408 ceph_cap_string(dirty));
2409 if (wanted != le32_to_cpu(grant->wanted)) {
2410 dout("mds wanted %s -> %s\n",
2411 ceph_cap_string(le32_to_cpu(grant->wanted)),
2412 ceph_cap_string(wanted));
2413 grant->wanted = cpu_to_le32(wanted);
2414 }
2415
2416 cap->seq = seq;
2417
2418 /* file layout may have changed */
2419 ci->i_layout = grant->layout;
2420
2421 /* revocation, grant, or no-op? */
2422 if (cap->issued & ~newcaps) {
2423 int revoking = cap->issued & ~newcaps;
2424
2425 dout("revocation: %s -> %s (revoking %s)\n",
2426 ceph_cap_string(cap->issued),
2427 ceph_cap_string(newcaps),
2428 ceph_cap_string(revoking));
2429 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2430 writeback = 1; /* initiate writeback; will delay ack */
2431 else if (revoking == CEPH_CAP_FILE_CACHE &&
2432 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2433 queue_invalidate)
2434 ; /* do nothing yet, invalidation will be queued */
2435 else if (cap == ci->i_auth_cap)
2436 check_caps = 1; /* check auth cap only */
2437 else
2438 check_caps = 2; /* check all caps */
2439 cap->issued = newcaps;
2440 cap->implemented |= newcaps;
2441 } else if (cap->issued == newcaps) {
2442 dout("caps unchanged: %s -> %s\n",
2443 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2444 } else {
2445 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2446 ceph_cap_string(newcaps));
2447 cap->issued = newcaps;
2448 cap->implemented |= newcaps; /* add bits only, to
2449 * avoid stepping on a
2450 * pending revocation */
2451 wake = 1;
2452 }
2453 BUG_ON(cap->issued & ~cap->implemented);
2454
2455 spin_unlock(&ci->i_ceph_lock);
2456 if (writeback)
2457 /*
2458 * queue inode for writeback: we can't actually call
2459 * filemap_write_and_wait, etc. from message handler
2460 * context.
2461 */
2462 ceph_queue_writeback(inode);
2463 if (queue_invalidate)
2464 ceph_queue_invalidate(inode);
2465 if (wake)
2466 wake_up_all(&ci->i_cap_wq);
2467
2468 if (check_caps == 1)
2469 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2470 session);
2471 else if (check_caps == 2)
2472 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2473 else
2474 mutex_unlock(&session->s_mutex);
2475}
2476
2477/*
2478 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2479 * MDS has been safely committed.
2480 */
2481static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2482 struct ceph_mds_caps *m,
2483 struct ceph_mds_session *session,
2484 struct ceph_cap *cap)
2485 __releases(ci->i_ceph_lock)
2486{
2487 struct ceph_inode_info *ci = ceph_inode(inode);
2488 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2489 unsigned seq = le32_to_cpu(m->seq);
2490 int dirty = le32_to_cpu(m->dirty);
2491 int cleaned = 0;
2492 int drop = 0;
2493 int i;
2494
2495 for (i = 0; i < CEPH_CAP_BITS; i++)
2496 if ((dirty & (1 << i)) &&
2497 flush_tid == ci->i_cap_flush_tid[i])
2498 cleaned |= 1 << i;
2499
2500 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2501 " flushing %s -> %s\n",
2502 inode, session->s_mds, seq, ceph_cap_string(dirty),
2503 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2504 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2505
2506 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2507 goto out;
2508
2509 ci->i_flushing_caps &= ~cleaned;
2510
2511 spin_lock(&mdsc->cap_dirty_lock);
2512 if (ci->i_flushing_caps == 0) {
2513 list_del_init(&ci->i_flushing_item);
2514 if (!list_empty(&session->s_cap_flushing))
2515 dout(" mds%d still flushing cap on %p\n",
2516 session->s_mds,
2517 &list_entry(session->s_cap_flushing.next,
2518 struct ceph_inode_info,
2519 i_flushing_item)->vfs_inode);
2520 mdsc->num_cap_flushing--;
2521 wake_up_all(&mdsc->cap_flushing_wq);
2522 dout(" inode %p now !flushing\n", inode);
2523
2524 if (ci->i_dirty_caps == 0) {
2525 dout(" inode %p now clean\n", inode);
2526 BUG_ON(!list_empty(&ci->i_dirty_item));
2527 drop = 1;
2528 if (ci->i_wrbuffer_ref_head == 0) {
2529 BUG_ON(!ci->i_head_snapc);
2530 ceph_put_snap_context(ci->i_head_snapc);
2531 ci->i_head_snapc = NULL;
2532 }
2533 } else {
2534 BUG_ON(list_empty(&ci->i_dirty_item));
2535 }
2536 }
2537 spin_unlock(&mdsc->cap_dirty_lock);
2538 wake_up_all(&ci->i_cap_wq);
2539
2540out:
2541 spin_unlock(&ci->i_ceph_lock);
2542 if (drop)
2543 iput(inode);
2544}
2545
2546/*
2547 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2548 * throw away our cap_snap.
2549 *
2550 * Caller hold s_mutex.
2551 */
2552static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2553 struct ceph_mds_caps *m,
2554 struct ceph_mds_session *session)
2555{
2556 struct ceph_inode_info *ci = ceph_inode(inode);
2557 u64 follows = le64_to_cpu(m->snap_follows);
2558 struct ceph_cap_snap *capsnap;
2559 int drop = 0;
2560
2561 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2562 inode, ci, session->s_mds, follows);
2563
2564 spin_lock(&ci->i_ceph_lock);
2565 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2566 if (capsnap->follows == follows) {
2567 if (capsnap->flush_tid != flush_tid) {
2568 dout(" cap_snap %p follows %lld tid %lld !="
2569 " %lld\n", capsnap, follows,
2570 flush_tid, capsnap->flush_tid);
2571 break;
2572 }
2573 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2574 dout(" removing %p cap_snap %p follows %lld\n",
2575 inode, capsnap, follows);
2576 ceph_put_snap_context(capsnap->context);
2577 list_del(&capsnap->ci_item);
2578 list_del(&capsnap->flushing_item);
2579 ceph_put_cap_snap(capsnap);
2580 drop = 1;
2581 break;
2582 } else {
2583 dout(" skipping cap_snap %p follows %lld\n",
2584 capsnap, capsnap->follows);
2585 }
2586 }
2587 spin_unlock(&ci->i_ceph_lock);
2588 if (drop)
2589 iput(inode);
2590}
2591
2592/*
2593 * Handle TRUNC from MDS, indicating file truncation.
2594 *
2595 * caller hold s_mutex.
2596 */
2597static void handle_cap_trunc(struct inode *inode,
2598 struct ceph_mds_caps *trunc,
2599 struct ceph_mds_session *session)
2600 __releases(ci->i_ceph_lock)
2601{
2602 struct ceph_inode_info *ci = ceph_inode(inode);
2603 int mds = session->s_mds;
2604 int seq = le32_to_cpu(trunc->seq);
2605 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2606 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2607 u64 size = le64_to_cpu(trunc->size);
2608 int implemented = 0;
2609 int dirty = __ceph_caps_dirty(ci);
2610 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2611 int queue_trunc = 0;
2612
2613 issued |= implemented | dirty;
2614
2615 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2616 inode, mds, seq, truncate_size, truncate_seq);
2617 queue_trunc = ceph_fill_file_size(inode, issued,
2618 truncate_seq, truncate_size, size);
2619 spin_unlock(&ci->i_ceph_lock);
2620
2621 if (queue_trunc)
2622 ceph_queue_vmtruncate(inode);
2623}
2624
2625/*
2626 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2627 * different one. If we are the most recent migration we've seen (as
2628 * indicated by mseq), make note of the migrating cap bits for the
2629 * duration (until we see the corresponding IMPORT).
2630 *
2631 * caller holds s_mutex
2632 */
2633static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2634 struct ceph_mds_session *session,
2635 int *open_target_sessions)
2636{
2637 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2638 struct ceph_inode_info *ci = ceph_inode(inode);
2639 int mds = session->s_mds;
2640 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2641 struct ceph_cap *cap = NULL, *t;
2642 struct rb_node *p;
2643 int remember = 1;
2644
2645 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n",
2646 inode, ci, mds, mseq);
2647
2648 spin_lock(&ci->i_ceph_lock);
2649
2650 /* make sure we haven't seen a higher mseq */
2651 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2652 t = rb_entry(p, struct ceph_cap, ci_node);
2653 if (ceph_seq_cmp(t->mseq, mseq) > 0) {
2654 dout(" higher mseq on cap from mds%d\n",
2655 t->session->s_mds);
2656 remember = 0;
2657 }
2658 if (t->session->s_mds == mds)
2659 cap = t;
2660 }
2661
2662 if (cap) {
2663 if (remember) {
2664 /* make note */
2665 ci->i_cap_exporting_mds = mds;
2666 ci->i_cap_exporting_mseq = mseq;
2667 ci->i_cap_exporting_issued = cap->issued;
2668
2669 /*
2670 * make sure we have open sessions with all possible
2671 * export targets, so that we get the matching IMPORT
2672 */
2673 *open_target_sessions = 1;
2674
2675 /*
2676 * we can't flush dirty caps that we've seen the
2677 * EXPORT but no IMPORT for
2678 */
2679 spin_lock(&mdsc->cap_dirty_lock);
2680 if (!list_empty(&ci->i_dirty_item)) {
2681 dout(" moving %p to cap_dirty_migrating\n",
2682 inode);
2683 list_move(&ci->i_dirty_item,
2684 &mdsc->cap_dirty_migrating);
2685 }
2686 spin_unlock(&mdsc->cap_dirty_lock);
2687 }
2688 __ceph_remove_cap(cap);
2689 }
2690 /* else, we already released it */
2691
2692 spin_unlock(&ci->i_ceph_lock);
2693}
2694
2695/*
2696 * Handle cap IMPORT. If there are temp bits from an older EXPORT,
2697 * clean them up.
2698 *
2699 * caller holds s_mutex.
2700 */
2701static void handle_cap_import(struct ceph_mds_client *mdsc,
2702 struct inode *inode, struct ceph_mds_caps *im,
2703 struct ceph_mds_session *session,
2704 void *snaptrace, int snaptrace_len)
2705{
2706 struct ceph_inode_info *ci = ceph_inode(inode);
2707 int mds = session->s_mds;
2708 unsigned issued = le32_to_cpu(im->caps);
2709 unsigned wanted = le32_to_cpu(im->wanted);
2710 unsigned seq = le32_to_cpu(im->seq);
2711 unsigned mseq = le32_to_cpu(im->migrate_seq);
2712 u64 realmino = le64_to_cpu(im->realm);
2713 u64 cap_id = le64_to_cpu(im->cap_id);
2714
2715 if (ci->i_cap_exporting_mds >= 0 &&
2716 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) {
2717 dout("handle_cap_import inode %p ci %p mds%d mseq %d"
2718 " - cleared exporting from mds%d\n",
2719 inode, ci, mds, mseq,
2720 ci->i_cap_exporting_mds);
2721 ci->i_cap_exporting_issued = 0;
2722 ci->i_cap_exporting_mseq = 0;
2723 ci->i_cap_exporting_mds = -1;
2724
2725 spin_lock(&mdsc->cap_dirty_lock);
2726 if (!list_empty(&ci->i_dirty_item)) {
2727 dout(" moving %p back to cap_dirty\n", inode);
2728 list_move(&ci->i_dirty_item, &mdsc->cap_dirty);
2729 }
2730 spin_unlock(&mdsc->cap_dirty_lock);
2731 } else {
2732 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
2733 inode, ci, mds, mseq);
2734 }
2735
2736 down_write(&mdsc->snap_rwsem);
2737 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2738 false);
2739 downgrade_write(&mdsc->snap_rwsem);
2740 ceph_add_cap(inode, session, cap_id, -1,
2741 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2742 NULL /* no caps context */);
2743 kick_flushing_inode_caps(mdsc, session, inode);
2744 up_read(&mdsc->snap_rwsem);
2745
2746 /* make sure we re-request max_size, if necessary */
2747 spin_lock(&ci->i_ceph_lock);
2748 ci->i_requested_max_size = 0;
2749 spin_unlock(&ci->i_ceph_lock);
2750}
2751
2752/*
2753 * Handle a caps message from the MDS.
2754 *
2755 * Identify the appropriate session, inode, and call the right handler
2756 * based on the cap op.
2757 */
2758void ceph_handle_caps(struct ceph_mds_session *session,
2759 struct ceph_msg *msg)
2760{
2761 struct ceph_mds_client *mdsc = session->s_mdsc;
2762 struct super_block *sb = mdsc->fsc->sb;
2763 struct inode *inode;
2764 struct ceph_inode_info *ci;
2765 struct ceph_cap *cap;
2766 struct ceph_mds_caps *h;
2767 int mds = session->s_mds;
2768 int op;
2769 u32 seq, mseq;
2770 struct ceph_vino vino;
2771 u64 cap_id;
2772 u64 size, max_size;
2773 u64 tid;
2774 void *snaptrace;
2775 size_t snaptrace_len;
2776 void *flock;
2777 u32 flock_len;
2778 int open_target_sessions = 0;
2779
2780 dout("handle_caps from mds%d\n", mds);
2781
2782 /* decode */
2783 tid = le64_to_cpu(msg->hdr.tid);
2784 if (msg->front.iov_len < sizeof(*h))
2785 goto bad;
2786 h = msg->front.iov_base;
2787 op = le32_to_cpu(h->op);
2788 vino.ino = le64_to_cpu(h->ino);
2789 vino.snap = CEPH_NOSNAP;
2790 cap_id = le64_to_cpu(h->cap_id);
2791 seq = le32_to_cpu(h->seq);
2792 mseq = le32_to_cpu(h->migrate_seq);
2793 size = le64_to_cpu(h->size);
2794 max_size = le64_to_cpu(h->max_size);
2795
2796 snaptrace = h + 1;
2797 snaptrace_len = le32_to_cpu(h->snap_trace_len);
2798
2799 if (le16_to_cpu(msg->hdr.version) >= 2) {
2800 void *p, *end;
2801
2802 p = snaptrace + snaptrace_len;
2803 end = msg->front.iov_base + msg->front.iov_len;
2804 ceph_decode_32_safe(&p, end, flock_len, bad);
2805 flock = p;
2806 } else {
2807 flock = NULL;
2808 flock_len = 0;
2809 }
2810
2811 mutex_lock(&session->s_mutex);
2812 session->s_seq++;
2813 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
2814 (unsigned)seq);
2815
2816 /* lookup ino */
2817 inode = ceph_find_inode(sb, vino);
2818 ci = ceph_inode(inode);
2819 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
2820 vino.snap, inode);
2821 if (!inode) {
2822 dout(" i don't have ino %llx\n", vino.ino);
2823
2824 if (op == CEPH_CAP_OP_IMPORT)
2825 __queue_cap_release(session, vino.ino, cap_id,
2826 mseq, seq);
2827 goto flush_cap_releases;
2828 }
2829
2830 /* these will work even if we don't have a cap yet */
2831 switch (op) {
2832 case CEPH_CAP_OP_FLUSHSNAP_ACK:
2833 handle_cap_flushsnap_ack(inode, tid, h, session);
2834 goto done;
2835
2836 case CEPH_CAP_OP_EXPORT:
2837 handle_cap_export(inode, h, session, &open_target_sessions);
2838 goto done;
2839
2840 case CEPH_CAP_OP_IMPORT:
2841 handle_cap_import(mdsc, inode, h, session,
2842 snaptrace, snaptrace_len);
2843 ceph_check_caps(ceph_inode(inode), 0, session);
2844 goto done_unlocked;
2845 }
2846
2847 /* the rest require a cap */
2848 spin_lock(&ci->i_ceph_lock);
2849 cap = __get_cap_for_mds(ceph_inode(inode), mds);
2850 if (!cap) {
2851 dout(" no cap on %p ino %llx.%llx from mds%d\n",
2852 inode, ceph_ino(inode), ceph_snap(inode), mds);
2853 spin_unlock(&ci->i_ceph_lock);
2854 goto flush_cap_releases;
2855 }
2856
2857 /* note that each of these drops i_ceph_lock for us */
2858 switch (op) {
2859 case CEPH_CAP_OP_REVOKE:
2860 case CEPH_CAP_OP_GRANT:
2861 handle_cap_grant(inode, h, session, cap, msg->middle);
2862 goto done_unlocked;
2863
2864 case CEPH_CAP_OP_FLUSH_ACK:
2865 handle_cap_flush_ack(inode, tid, h, session, cap);
2866 break;
2867
2868 case CEPH_CAP_OP_TRUNC:
2869 handle_cap_trunc(inode, h, session);
2870 break;
2871
2872 default:
2873 spin_unlock(&ci->i_ceph_lock);
2874 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
2875 ceph_cap_op_name(op));
2876 }
2877
2878 goto done;
2879
2880flush_cap_releases:
2881 /*
2882 * send any full release message to try to move things
2883 * along for the mds (who clearly thinks we still have this
2884 * cap).
2885 */
2886 ceph_add_cap_releases(mdsc, session);
2887 ceph_send_cap_releases(mdsc, session);
2888
2889done:
2890 mutex_unlock(&session->s_mutex);
2891done_unlocked:
2892 if (inode)
2893 iput(inode);
2894 if (open_target_sessions)
2895 ceph_mdsc_open_export_target_sessions(mdsc, session);
2896 return;
2897
2898bad:
2899 pr_err("ceph_handle_caps: corrupt message\n");
2900 ceph_msg_dump(msg);
2901 return;
2902}
2903
2904/*
2905 * Delayed work handler to process end of delayed cap release LRU list.
2906 */
2907void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
2908{
2909 struct ceph_inode_info *ci;
2910 int flags = CHECK_CAPS_NODELAY;
2911
2912 dout("check_delayed_caps\n");
2913 while (1) {
2914 spin_lock(&mdsc->cap_delay_lock);
2915 if (list_empty(&mdsc->cap_delay_list))
2916 break;
2917 ci = list_first_entry(&mdsc->cap_delay_list,
2918 struct ceph_inode_info,
2919 i_cap_delay_list);
2920 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
2921 time_before(jiffies, ci->i_hold_caps_max))
2922 break;
2923 list_del_init(&ci->i_cap_delay_list);
2924 spin_unlock(&mdsc->cap_delay_lock);
2925 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
2926 ceph_check_caps(ci, flags, NULL);
2927 }
2928 spin_unlock(&mdsc->cap_delay_lock);
2929}
2930
2931/*
2932 * Flush all dirty caps to the mds
2933 */
2934void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
2935{
2936 struct ceph_inode_info *ci;
2937 struct inode *inode;
2938
2939 dout("flush_dirty_caps\n");
2940 spin_lock(&mdsc->cap_dirty_lock);
2941 while (!list_empty(&mdsc->cap_dirty)) {
2942 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
2943 i_dirty_item);
2944 inode = &ci->vfs_inode;
2945 ihold(inode);
2946 dout("flush_dirty_caps %p\n", inode);
2947 spin_unlock(&mdsc->cap_dirty_lock);
2948 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
2949 iput(inode);
2950 spin_lock(&mdsc->cap_dirty_lock);
2951 }
2952 spin_unlock(&mdsc->cap_dirty_lock);
2953 dout("flush_dirty_caps done\n");
2954}
2955
2956/*
2957 * Drop open file reference. If we were the last open file,
2958 * we may need to release capabilities to the MDS (or schedule
2959 * their delayed release).
2960 */
2961void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
2962{
2963 struct inode *inode = &ci->vfs_inode;
2964 int last = 0;
2965
2966 spin_lock(&ci->i_ceph_lock);
2967 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
2968 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
2969 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
2970 if (--ci->i_nr_by_mode[fmode] == 0)
2971 last++;
2972 spin_unlock(&ci->i_ceph_lock);
2973
2974 if (last && ci->i_vino.snap == CEPH_NOSNAP)
2975 ceph_check_caps(ci, 0, NULL);
2976}
2977
2978/*
2979 * Helpers for embedding cap and dentry lease releases into mds
2980 * requests.
2981 *
2982 * @force is used by dentry_release (below) to force inclusion of a
2983 * record for the directory inode, even when there aren't any caps to
2984 * drop.
2985 */
2986int ceph_encode_inode_release(void **p, struct inode *inode,
2987 int mds, int drop, int unless, int force)
2988{
2989 struct ceph_inode_info *ci = ceph_inode(inode);
2990 struct ceph_cap *cap;
2991 struct ceph_mds_request_release *rel = *p;
2992 int used, dirty;
2993 int ret = 0;
2994
2995 spin_lock(&ci->i_ceph_lock);
2996 used = __ceph_caps_used(ci);
2997 dirty = __ceph_caps_dirty(ci);
2998
2999 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3000 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3001 ceph_cap_string(unless));
3002
3003 /* only drop unused, clean caps */
3004 drop &= ~(used | dirty);
3005
3006 cap = __get_cap_for_mds(ci, mds);
3007 if (cap && __cap_is_valid(cap)) {
3008 if (force ||
3009 ((cap->issued & drop) &&
3010 (cap->issued & unless) == 0)) {
3011 if ((cap->issued & drop) &&
3012 (cap->issued & unless) == 0) {
3013 dout("encode_inode_release %p cap %p %s -> "
3014 "%s\n", inode, cap,
3015 ceph_cap_string(cap->issued),
3016 ceph_cap_string(cap->issued & ~drop));
3017 cap->issued &= ~drop;
3018 cap->implemented &= ~drop;
3019 if (ci->i_ceph_flags & CEPH_I_NODELAY) {
3020 int wanted = __ceph_caps_wanted(ci);
3021 dout(" wanted %s -> %s (act %s)\n",
3022 ceph_cap_string(cap->mds_wanted),
3023 ceph_cap_string(cap->mds_wanted &
3024 ~wanted),
3025 ceph_cap_string(wanted));
3026 cap->mds_wanted &= wanted;
3027 }
3028 } else {
3029 dout("encode_inode_release %p cap %p %s"
3030 " (force)\n", inode, cap,
3031 ceph_cap_string(cap->issued));
3032 }
3033
3034 rel->ino = cpu_to_le64(ceph_ino(inode));
3035 rel->cap_id = cpu_to_le64(cap->cap_id);
3036 rel->seq = cpu_to_le32(cap->seq);
3037 rel->issue_seq = cpu_to_le32(cap->issue_seq),
3038 rel->mseq = cpu_to_le32(cap->mseq);
3039 rel->caps = cpu_to_le32(cap->issued);
3040 rel->wanted = cpu_to_le32(cap->mds_wanted);
3041 rel->dname_len = 0;
3042 rel->dname_seq = 0;
3043 *p += sizeof(*rel);
3044 ret = 1;
3045 } else {
3046 dout("encode_inode_release %p cap %p %s\n",
3047 inode, cap, ceph_cap_string(cap->issued));
3048 }
3049 }
3050 spin_unlock(&ci->i_ceph_lock);
3051 return ret;
3052}
3053
3054int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3055 int mds, int drop, int unless)
3056{
3057 struct inode *dir = dentry->d_parent->d_inode;
3058 struct ceph_mds_request_release *rel = *p;
3059 struct ceph_dentry_info *di = ceph_dentry(dentry);
3060 int force = 0;
3061 int ret;
3062
3063 /*
3064 * force an record for the directory caps if we have a dentry lease.
3065 * this is racy (can't take i_ceph_lock and d_lock together), but it
3066 * doesn't have to be perfect; the mds will revoke anything we don't
3067 * release.
3068 */
3069 spin_lock(&dentry->d_lock);
3070 if (di->lease_session && di->lease_session->s_mds == mds)
3071 force = 1;
3072 spin_unlock(&dentry->d_lock);
3073
3074 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3075
3076 spin_lock(&dentry->d_lock);
3077 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3078 dout("encode_dentry_release %p mds%d seq %d\n",
3079 dentry, mds, (int)di->lease_seq);
3080 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3081 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3082 *p += dentry->d_name.len;
3083 rel->dname_seq = cpu_to_le32(di->lease_seq);
3084 __ceph_mdsc_drop_dentry_lease(dentry);
3085 }
3086 spin_unlock(&dentry->d_lock);
3087 return ret;
3088}
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
43static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
44static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
45 struct ceph_mds_session *session,
46 struct ceph_inode_info *ci,
47 u64 oldest_flush_tid);
48
49/*
50 * Generate readable cap strings for debugging output.
51 */
52#define MAX_CAP_STR 20
53static char cap_str[MAX_CAP_STR][40];
54static DEFINE_SPINLOCK(cap_str_lock);
55static int last_cap_str;
56
57static char *gcap_string(char *s, int c)
58{
59 if (c & CEPH_CAP_GSHARED)
60 *s++ = 's';
61 if (c & CEPH_CAP_GEXCL)
62 *s++ = 'x';
63 if (c & CEPH_CAP_GCACHE)
64 *s++ = 'c';
65 if (c & CEPH_CAP_GRD)
66 *s++ = 'r';
67 if (c & CEPH_CAP_GWR)
68 *s++ = 'w';
69 if (c & CEPH_CAP_GBUFFER)
70 *s++ = 'b';
71 if (c & CEPH_CAP_GLAZYIO)
72 *s++ = 'l';
73 return s;
74}
75
76const char *ceph_cap_string(int caps)
77{
78 int i;
79 char *s;
80 int c;
81
82 spin_lock(&cap_str_lock);
83 i = last_cap_str++;
84 if (last_cap_str == MAX_CAP_STR)
85 last_cap_str = 0;
86 spin_unlock(&cap_str_lock);
87
88 s = cap_str[i];
89
90 if (caps & CEPH_CAP_PIN)
91 *s++ = 'p';
92
93 c = (caps >> CEPH_CAP_SAUTH) & 3;
94 if (c) {
95 *s++ = 'A';
96 s = gcap_string(s, c);
97 }
98
99 c = (caps >> CEPH_CAP_SLINK) & 3;
100 if (c) {
101 *s++ = 'L';
102 s = gcap_string(s, c);
103 }
104
105 c = (caps >> CEPH_CAP_SXATTR) & 3;
106 if (c) {
107 *s++ = 'X';
108 s = gcap_string(s, c);
109 }
110
111 c = caps >> CEPH_CAP_SFILE;
112 if (c) {
113 *s++ = 'F';
114 s = gcap_string(s, c);
115 }
116
117 if (s == cap_str[i])
118 *s++ = '-';
119 *s = 0;
120 return cap_str[i];
121}
122
123void ceph_caps_init(struct ceph_mds_client *mdsc)
124{
125 INIT_LIST_HEAD(&mdsc->caps_list);
126 spin_lock_init(&mdsc->caps_list_lock);
127}
128
129void ceph_caps_finalize(struct ceph_mds_client *mdsc)
130{
131 struct ceph_cap *cap;
132
133 spin_lock(&mdsc->caps_list_lock);
134 while (!list_empty(&mdsc->caps_list)) {
135 cap = list_first_entry(&mdsc->caps_list,
136 struct ceph_cap, caps_item);
137 list_del(&cap->caps_item);
138 kmem_cache_free(ceph_cap_cachep, cap);
139 }
140 mdsc->caps_total_count = 0;
141 mdsc->caps_avail_count = 0;
142 mdsc->caps_use_count = 0;
143 mdsc->caps_reserve_count = 0;
144 mdsc->caps_min_count = 0;
145 spin_unlock(&mdsc->caps_list_lock);
146}
147
148void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
149{
150 spin_lock(&mdsc->caps_list_lock);
151 mdsc->caps_min_count += delta;
152 BUG_ON(mdsc->caps_min_count < 0);
153 spin_unlock(&mdsc->caps_list_lock);
154}
155
156void ceph_reserve_caps(struct ceph_mds_client *mdsc,
157 struct ceph_cap_reservation *ctx, int need)
158{
159 int i;
160 struct ceph_cap *cap;
161 int have;
162 int alloc = 0;
163 LIST_HEAD(newcaps);
164
165 dout("reserve caps ctx=%p need=%d\n", ctx, need);
166
167 /* first reserve any caps that are already allocated */
168 spin_lock(&mdsc->caps_list_lock);
169 if (mdsc->caps_avail_count >= need)
170 have = need;
171 else
172 have = mdsc->caps_avail_count;
173 mdsc->caps_avail_count -= have;
174 mdsc->caps_reserve_count += have;
175 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
176 mdsc->caps_reserve_count +
177 mdsc->caps_avail_count);
178 spin_unlock(&mdsc->caps_list_lock);
179
180 for (i = have; i < need; i++) {
181 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
182 if (!cap)
183 break;
184 list_add(&cap->caps_item, &newcaps);
185 alloc++;
186 }
187 /* we didn't manage to reserve as much as we needed */
188 if (have + alloc != need)
189 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
190 ctx, need, have + alloc);
191
192 spin_lock(&mdsc->caps_list_lock);
193 mdsc->caps_total_count += alloc;
194 mdsc->caps_reserve_count += alloc;
195 list_splice(&newcaps, &mdsc->caps_list);
196
197 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
198 mdsc->caps_reserve_count +
199 mdsc->caps_avail_count);
200 spin_unlock(&mdsc->caps_list_lock);
201
202 ctx->count = need;
203 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
204 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
205 mdsc->caps_reserve_count, mdsc->caps_avail_count);
206}
207
208int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
209 struct ceph_cap_reservation *ctx)
210{
211 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
212 if (ctx->count) {
213 spin_lock(&mdsc->caps_list_lock);
214 BUG_ON(mdsc->caps_reserve_count < ctx->count);
215 mdsc->caps_reserve_count -= ctx->count;
216 mdsc->caps_avail_count += ctx->count;
217 ctx->count = 0;
218 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
219 mdsc->caps_total_count, mdsc->caps_use_count,
220 mdsc->caps_reserve_count, mdsc->caps_avail_count);
221 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
222 mdsc->caps_reserve_count +
223 mdsc->caps_avail_count);
224 spin_unlock(&mdsc->caps_list_lock);
225 }
226 return 0;
227}
228
229struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
230 struct ceph_cap_reservation *ctx)
231{
232 struct ceph_cap *cap = NULL;
233
234 /* temporary, until we do something about cap import/export */
235 if (!ctx) {
236 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
237 if (cap) {
238 spin_lock(&mdsc->caps_list_lock);
239 mdsc->caps_use_count++;
240 mdsc->caps_total_count++;
241 spin_unlock(&mdsc->caps_list_lock);
242 }
243 return cap;
244 }
245
246 spin_lock(&mdsc->caps_list_lock);
247 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
248 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
249 mdsc->caps_reserve_count, mdsc->caps_avail_count);
250 BUG_ON(!ctx->count);
251 BUG_ON(ctx->count > mdsc->caps_reserve_count);
252 BUG_ON(list_empty(&mdsc->caps_list));
253
254 ctx->count--;
255 mdsc->caps_reserve_count--;
256 mdsc->caps_use_count++;
257
258 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
259 list_del(&cap->caps_item);
260
261 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
262 mdsc->caps_reserve_count + mdsc->caps_avail_count);
263 spin_unlock(&mdsc->caps_list_lock);
264 return cap;
265}
266
267void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
268{
269 spin_lock(&mdsc->caps_list_lock);
270 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
271 cap, mdsc->caps_total_count, mdsc->caps_use_count,
272 mdsc->caps_reserve_count, mdsc->caps_avail_count);
273 mdsc->caps_use_count--;
274 /*
275 * Keep some preallocated caps around (ceph_min_count), to
276 * avoid lots of free/alloc churn.
277 */
278 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
279 mdsc->caps_min_count) {
280 mdsc->caps_total_count--;
281 kmem_cache_free(ceph_cap_cachep, cap);
282 } else {
283 mdsc->caps_avail_count++;
284 list_add(&cap->caps_item, &mdsc->caps_list);
285 }
286
287 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
288 mdsc->caps_reserve_count + mdsc->caps_avail_count);
289 spin_unlock(&mdsc->caps_list_lock);
290}
291
292void ceph_reservation_status(struct ceph_fs_client *fsc,
293 int *total, int *avail, int *used, int *reserved,
294 int *min)
295{
296 struct ceph_mds_client *mdsc = fsc->mdsc;
297
298 if (total)
299 *total = mdsc->caps_total_count;
300 if (avail)
301 *avail = mdsc->caps_avail_count;
302 if (used)
303 *used = mdsc->caps_use_count;
304 if (reserved)
305 *reserved = mdsc->caps_reserve_count;
306 if (min)
307 *min = mdsc->caps_min_count;
308}
309
310/*
311 * Find ceph_cap for given mds, if any.
312 *
313 * Called with i_ceph_lock held.
314 */
315static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
316{
317 struct ceph_cap *cap;
318 struct rb_node *n = ci->i_caps.rb_node;
319
320 while (n) {
321 cap = rb_entry(n, struct ceph_cap, ci_node);
322 if (mds < cap->mds)
323 n = n->rb_left;
324 else if (mds > cap->mds)
325 n = n->rb_right;
326 else
327 return cap;
328 }
329 return NULL;
330}
331
332struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
333{
334 struct ceph_cap *cap;
335
336 spin_lock(&ci->i_ceph_lock);
337 cap = __get_cap_for_mds(ci, mds);
338 spin_unlock(&ci->i_ceph_lock);
339 return cap;
340}
341
342/*
343 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
344 */
345static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
346{
347 struct ceph_cap *cap;
348 int mds = -1;
349 struct rb_node *p;
350
351 /* prefer mds with WR|BUFFER|EXCL caps */
352 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
353 cap = rb_entry(p, struct ceph_cap, ci_node);
354 mds = cap->mds;
355 if (cap->issued & (CEPH_CAP_FILE_WR |
356 CEPH_CAP_FILE_BUFFER |
357 CEPH_CAP_FILE_EXCL))
358 break;
359 }
360 return mds;
361}
362
363int ceph_get_cap_mds(struct inode *inode)
364{
365 struct ceph_inode_info *ci = ceph_inode(inode);
366 int mds;
367 spin_lock(&ci->i_ceph_lock);
368 mds = __ceph_get_cap_mds(ceph_inode(inode));
369 spin_unlock(&ci->i_ceph_lock);
370 return mds;
371}
372
373/*
374 * Called under i_ceph_lock.
375 */
376static void __insert_cap_node(struct ceph_inode_info *ci,
377 struct ceph_cap *new)
378{
379 struct rb_node **p = &ci->i_caps.rb_node;
380 struct rb_node *parent = NULL;
381 struct ceph_cap *cap = NULL;
382
383 while (*p) {
384 parent = *p;
385 cap = rb_entry(parent, struct ceph_cap, ci_node);
386 if (new->mds < cap->mds)
387 p = &(*p)->rb_left;
388 else if (new->mds > cap->mds)
389 p = &(*p)->rb_right;
390 else
391 BUG();
392 }
393
394 rb_link_node(&new->ci_node, parent, p);
395 rb_insert_color(&new->ci_node, &ci->i_caps);
396}
397
398/*
399 * (re)set cap hold timeouts, which control the delayed release
400 * of unused caps back to the MDS. Should be called on cap use.
401 */
402static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
403 struct ceph_inode_info *ci)
404{
405 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
406
407 ci->i_hold_caps_min = round_jiffies(jiffies +
408 ma->caps_wanted_delay_min * HZ);
409 ci->i_hold_caps_max = round_jiffies(jiffies +
410 ma->caps_wanted_delay_max * HZ);
411 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
412 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
413}
414
415/*
416 * (Re)queue cap at the end of the delayed cap release list.
417 *
418 * If I_FLUSH is set, leave the inode at the front of the list.
419 *
420 * Caller holds i_ceph_lock
421 * -> we take mdsc->cap_delay_lock
422 */
423static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
424 struct ceph_inode_info *ci)
425{
426 __cap_set_timeouts(mdsc, ci);
427 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
428 ci->i_ceph_flags, ci->i_hold_caps_max);
429 if (!mdsc->stopping) {
430 spin_lock(&mdsc->cap_delay_lock);
431 if (!list_empty(&ci->i_cap_delay_list)) {
432 if (ci->i_ceph_flags & CEPH_I_FLUSH)
433 goto no_change;
434 list_del_init(&ci->i_cap_delay_list);
435 }
436 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
437no_change:
438 spin_unlock(&mdsc->cap_delay_lock);
439 }
440}
441
442/*
443 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
444 * indicating we should send a cap message to flush dirty metadata
445 * asap, and move to the front of the delayed cap list.
446 */
447static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
448 struct ceph_inode_info *ci)
449{
450 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
451 spin_lock(&mdsc->cap_delay_lock);
452 ci->i_ceph_flags |= CEPH_I_FLUSH;
453 if (!list_empty(&ci->i_cap_delay_list))
454 list_del_init(&ci->i_cap_delay_list);
455 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
456 spin_unlock(&mdsc->cap_delay_lock);
457}
458
459/*
460 * Cancel delayed work on cap.
461 *
462 * Caller must hold i_ceph_lock.
463 */
464static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
465 struct ceph_inode_info *ci)
466{
467 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
468 if (list_empty(&ci->i_cap_delay_list))
469 return;
470 spin_lock(&mdsc->cap_delay_lock);
471 list_del_init(&ci->i_cap_delay_list);
472 spin_unlock(&mdsc->cap_delay_lock);
473}
474
475/*
476 * Common issue checks for add_cap, handle_cap_grant.
477 */
478static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
479 unsigned issued)
480{
481 unsigned had = __ceph_caps_issued(ci, NULL);
482
483 /*
484 * Each time we receive FILE_CACHE anew, we increment
485 * i_rdcache_gen.
486 */
487 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
488 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
489 ci->i_rdcache_gen++;
490 }
491
492 /*
493 * if we are newly issued FILE_SHARED, mark dir not complete; we
494 * don't know what happened to this directory while we didn't
495 * have the cap.
496 */
497 if ((issued & CEPH_CAP_FILE_SHARED) &&
498 (had & CEPH_CAP_FILE_SHARED) == 0) {
499 ci->i_shared_gen++;
500 if (S_ISDIR(ci->vfs_inode.i_mode)) {
501 dout(" marking %p NOT complete\n", &ci->vfs_inode);
502 __ceph_dir_clear_complete(ci);
503 }
504 }
505}
506
507/*
508 * Add a capability under the given MDS session.
509 *
510 * Caller should hold session snap_rwsem (read) and s_mutex.
511 *
512 * @fmode is the open file mode, if we are opening a file, otherwise
513 * it is < 0. (This is so we can atomically add the cap and add an
514 * open file reference to it.)
515 */
516void ceph_add_cap(struct inode *inode,
517 struct ceph_mds_session *session, u64 cap_id,
518 int fmode, unsigned issued, unsigned wanted,
519 unsigned seq, unsigned mseq, u64 realmino, int flags,
520 struct ceph_cap **new_cap)
521{
522 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
523 struct ceph_inode_info *ci = ceph_inode(inode);
524 struct ceph_cap *cap;
525 int mds = session->s_mds;
526 int actual_wanted;
527
528 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
529 session->s_mds, cap_id, ceph_cap_string(issued), seq);
530
531 /*
532 * If we are opening the file, include file mode wanted bits
533 * in wanted.
534 */
535 if (fmode >= 0)
536 wanted |= ceph_caps_for_mode(fmode);
537
538 cap = __get_cap_for_mds(ci, mds);
539 if (!cap) {
540 cap = *new_cap;
541 *new_cap = NULL;
542
543 cap->issued = 0;
544 cap->implemented = 0;
545 cap->mds = mds;
546 cap->mds_wanted = 0;
547 cap->mseq = 0;
548
549 cap->ci = ci;
550 __insert_cap_node(ci, cap);
551
552 /* add to session cap list */
553 cap->session = session;
554 spin_lock(&session->s_cap_lock);
555 list_add_tail(&cap->session_caps, &session->s_caps);
556 session->s_nr_caps++;
557 spin_unlock(&session->s_cap_lock);
558 } else {
559 /*
560 * auth mds of the inode changed. we received the cap export
561 * message, but still haven't received the cap import message.
562 * handle_cap_export() updated the new auth MDS' cap.
563 *
564 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
565 * a message that was send before the cap import message. So
566 * don't remove caps.
567 */
568 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
569 WARN_ON(cap != ci->i_auth_cap);
570 WARN_ON(cap->cap_id != cap_id);
571 seq = cap->seq;
572 mseq = cap->mseq;
573 issued |= cap->issued;
574 flags |= CEPH_CAP_FLAG_AUTH;
575 }
576 }
577
578 if (!ci->i_snap_realm) {
579 /*
580 * add this inode to the appropriate snap realm
581 */
582 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
583 realmino);
584 if (realm) {
585 spin_lock(&realm->inodes_with_caps_lock);
586 ci->i_snap_realm = realm;
587 list_add(&ci->i_snap_realm_item,
588 &realm->inodes_with_caps);
589 spin_unlock(&realm->inodes_with_caps_lock);
590 } else {
591 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
592 realmino);
593 WARN_ON(!realm);
594 }
595 }
596
597 __check_cap_issue(ci, cap, issued);
598
599 /*
600 * If we are issued caps we don't want, or the mds' wanted
601 * value appears to be off, queue a check so we'll release
602 * later and/or update the mds wanted value.
603 */
604 actual_wanted = __ceph_caps_wanted(ci);
605 if ((wanted & ~actual_wanted) ||
606 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
607 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
608 ceph_cap_string(issued), ceph_cap_string(wanted),
609 ceph_cap_string(actual_wanted));
610 __cap_delay_requeue(mdsc, ci);
611 }
612
613 if (flags & CEPH_CAP_FLAG_AUTH) {
614 if (ci->i_auth_cap == NULL ||
615 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
616 ci->i_auth_cap = cap;
617 cap->mds_wanted = wanted;
618 }
619 } else {
620 WARN_ON(ci->i_auth_cap == cap);
621 }
622
623 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
624 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
625 ceph_cap_string(issued|cap->issued), seq, mds);
626 cap->cap_id = cap_id;
627 cap->issued = issued;
628 cap->implemented |= issued;
629 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
630 cap->mds_wanted = wanted;
631 else
632 cap->mds_wanted |= wanted;
633 cap->seq = seq;
634 cap->issue_seq = seq;
635 cap->mseq = mseq;
636 cap->cap_gen = session->s_cap_gen;
637
638 if (fmode >= 0)
639 __ceph_get_fmode(ci, fmode);
640}
641
642/*
643 * Return true if cap has not timed out and belongs to the current
644 * generation of the MDS session (i.e. has not gone 'stale' due to
645 * us losing touch with the mds).
646 */
647static int __cap_is_valid(struct ceph_cap *cap)
648{
649 unsigned long ttl;
650 u32 gen;
651
652 spin_lock(&cap->session->s_gen_ttl_lock);
653 gen = cap->session->s_cap_gen;
654 ttl = cap->session->s_cap_ttl;
655 spin_unlock(&cap->session->s_gen_ttl_lock);
656
657 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
658 dout("__cap_is_valid %p cap %p issued %s "
659 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
660 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
661 return 0;
662 }
663
664 return 1;
665}
666
667/*
668 * Return set of valid cap bits issued to us. Note that caps time
669 * out, and may be invalidated in bulk if the client session times out
670 * and session->s_cap_gen is bumped.
671 */
672int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
673{
674 int have = ci->i_snap_caps;
675 struct ceph_cap *cap;
676 struct rb_node *p;
677
678 if (implemented)
679 *implemented = 0;
680 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
681 cap = rb_entry(p, struct ceph_cap, ci_node);
682 if (!__cap_is_valid(cap))
683 continue;
684 dout("__ceph_caps_issued %p cap %p issued %s\n",
685 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
686 have |= cap->issued;
687 if (implemented)
688 *implemented |= cap->implemented;
689 }
690 /*
691 * exclude caps issued by non-auth MDS, but are been revoking
692 * by the auth MDS. The non-auth MDS should be revoking/exporting
693 * these caps, but the message is delayed.
694 */
695 if (ci->i_auth_cap) {
696 cap = ci->i_auth_cap;
697 have &= ~cap->implemented | cap->issued;
698 }
699 return have;
700}
701
702/*
703 * Get cap bits issued by caps other than @ocap
704 */
705int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
706{
707 int have = ci->i_snap_caps;
708 struct ceph_cap *cap;
709 struct rb_node *p;
710
711 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
712 cap = rb_entry(p, struct ceph_cap, ci_node);
713 if (cap == ocap)
714 continue;
715 if (!__cap_is_valid(cap))
716 continue;
717 have |= cap->issued;
718 }
719 return have;
720}
721
722/*
723 * Move a cap to the end of the LRU (oldest caps at list head, newest
724 * at list tail).
725 */
726static void __touch_cap(struct ceph_cap *cap)
727{
728 struct ceph_mds_session *s = cap->session;
729
730 spin_lock(&s->s_cap_lock);
731 if (s->s_cap_iterator == NULL) {
732 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
733 s->s_mds);
734 list_move_tail(&cap->session_caps, &s->s_caps);
735 } else {
736 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
737 &cap->ci->vfs_inode, cap, s->s_mds);
738 }
739 spin_unlock(&s->s_cap_lock);
740}
741
742/*
743 * Check if we hold the given mask. If so, move the cap(s) to the
744 * front of their respective LRUs. (This is the preferred way for
745 * callers to check for caps they want.)
746 */
747int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
748{
749 struct ceph_cap *cap;
750 struct rb_node *p;
751 int have = ci->i_snap_caps;
752
753 if ((have & mask) == mask) {
754 dout("__ceph_caps_issued_mask %p snap issued %s"
755 " (mask %s)\n", &ci->vfs_inode,
756 ceph_cap_string(have),
757 ceph_cap_string(mask));
758 return 1;
759 }
760
761 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
762 cap = rb_entry(p, struct ceph_cap, ci_node);
763 if (!__cap_is_valid(cap))
764 continue;
765 if ((cap->issued & mask) == mask) {
766 dout("__ceph_caps_issued_mask %p cap %p issued %s"
767 " (mask %s)\n", &ci->vfs_inode, cap,
768 ceph_cap_string(cap->issued),
769 ceph_cap_string(mask));
770 if (touch)
771 __touch_cap(cap);
772 return 1;
773 }
774
775 /* does a combination of caps satisfy mask? */
776 have |= cap->issued;
777 if ((have & mask) == mask) {
778 dout("__ceph_caps_issued_mask %p combo issued %s"
779 " (mask %s)\n", &ci->vfs_inode,
780 ceph_cap_string(cap->issued),
781 ceph_cap_string(mask));
782 if (touch) {
783 struct rb_node *q;
784
785 /* touch this + preceding caps */
786 __touch_cap(cap);
787 for (q = rb_first(&ci->i_caps); q != p;
788 q = rb_next(q)) {
789 cap = rb_entry(q, struct ceph_cap,
790 ci_node);
791 if (!__cap_is_valid(cap))
792 continue;
793 __touch_cap(cap);
794 }
795 }
796 return 1;
797 }
798 }
799
800 return 0;
801}
802
803/*
804 * Return true if mask caps are currently being revoked by an MDS.
805 */
806int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
807 struct ceph_cap *ocap, int mask)
808{
809 struct ceph_cap *cap;
810 struct rb_node *p;
811
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 != ocap &&
815 (cap->implemented & ~cap->issued & mask))
816 return 1;
817 }
818 return 0;
819}
820
821int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
822{
823 struct inode *inode = &ci->vfs_inode;
824 int ret;
825
826 spin_lock(&ci->i_ceph_lock);
827 ret = __ceph_caps_revoking_other(ci, NULL, mask);
828 spin_unlock(&ci->i_ceph_lock);
829 dout("ceph_caps_revoking %p %s = %d\n", inode,
830 ceph_cap_string(mask), ret);
831 return ret;
832}
833
834int __ceph_caps_used(struct ceph_inode_info *ci)
835{
836 int used = 0;
837 if (ci->i_pin_ref)
838 used |= CEPH_CAP_PIN;
839 if (ci->i_rd_ref)
840 used |= CEPH_CAP_FILE_RD;
841 if (ci->i_rdcache_ref ||
842 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
843 ci->vfs_inode.i_data.nrpages))
844 used |= CEPH_CAP_FILE_CACHE;
845 if (ci->i_wr_ref)
846 used |= CEPH_CAP_FILE_WR;
847 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
848 used |= CEPH_CAP_FILE_BUFFER;
849 return used;
850}
851
852/*
853 * wanted, by virtue of open file modes
854 */
855int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
856{
857 int i, bits = 0;
858 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
859 if (ci->i_nr_by_mode[i])
860 bits |= 1 << i;
861 }
862 if (bits == 0)
863 return 0;
864 return ceph_caps_for_mode(bits >> 1);
865}
866
867/*
868 * Return caps we have registered with the MDS(s) as 'wanted'.
869 */
870int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
871{
872 struct ceph_cap *cap;
873 struct rb_node *p;
874 int mds_wanted = 0;
875
876 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
877 cap = rb_entry(p, struct ceph_cap, ci_node);
878 if (!__cap_is_valid(cap))
879 continue;
880 if (cap == ci->i_auth_cap)
881 mds_wanted |= cap->mds_wanted;
882 else
883 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
884 }
885 return mds_wanted;
886}
887
888/*
889 * called under i_ceph_lock
890 */
891static int __ceph_is_any_caps(struct ceph_inode_info *ci)
892{
893 return !RB_EMPTY_ROOT(&ci->i_caps);
894}
895
896int ceph_is_any_caps(struct inode *inode)
897{
898 struct ceph_inode_info *ci = ceph_inode(inode);
899 int ret;
900
901 spin_lock(&ci->i_ceph_lock);
902 ret = __ceph_is_any_caps(ci);
903 spin_unlock(&ci->i_ceph_lock);
904
905 return ret;
906}
907
908static void drop_inode_snap_realm(struct ceph_inode_info *ci)
909{
910 struct ceph_snap_realm *realm = ci->i_snap_realm;
911 spin_lock(&realm->inodes_with_caps_lock);
912 list_del_init(&ci->i_snap_realm_item);
913 ci->i_snap_realm_counter++;
914 ci->i_snap_realm = NULL;
915 spin_unlock(&realm->inodes_with_caps_lock);
916 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
917 realm);
918}
919
920/*
921 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
922 *
923 * caller should hold i_ceph_lock.
924 * caller will not hold session s_mutex if called from destroy_inode.
925 */
926void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
927{
928 struct ceph_mds_session *session = cap->session;
929 struct ceph_inode_info *ci = cap->ci;
930 struct ceph_mds_client *mdsc =
931 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
932 int removed = 0;
933
934 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
935
936 /* remove from session list */
937 spin_lock(&session->s_cap_lock);
938 if (session->s_cap_iterator == cap) {
939 /* not yet, we are iterating over this very cap */
940 dout("__ceph_remove_cap delaying %p removal from session %p\n",
941 cap, cap->session);
942 } else {
943 list_del_init(&cap->session_caps);
944 session->s_nr_caps--;
945 cap->session = NULL;
946 removed = 1;
947 }
948 /* protect backpointer with s_cap_lock: see iterate_session_caps */
949 cap->ci = NULL;
950
951 /*
952 * s_cap_reconnect is protected by s_cap_lock. no one changes
953 * s_cap_gen while session is in the reconnect state.
954 */
955 if (queue_release &&
956 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
957 cap->queue_release = 1;
958 if (removed) {
959 list_add_tail(&cap->session_caps,
960 &session->s_cap_releases);
961 session->s_num_cap_releases++;
962 removed = 0;
963 }
964 } else {
965 cap->queue_release = 0;
966 }
967 cap->cap_ino = ci->i_vino.ino;
968
969 spin_unlock(&session->s_cap_lock);
970
971 /* remove from inode list */
972 rb_erase(&cap->ci_node, &ci->i_caps);
973 if (ci->i_auth_cap == cap)
974 ci->i_auth_cap = NULL;
975
976 if (removed)
977 ceph_put_cap(mdsc, cap);
978
979 /* when reconnect denied, we remove session caps forcibly,
980 * i_wr_ref can be non-zero. If there are ongoing write,
981 * keep i_snap_realm.
982 */
983 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
984 drop_inode_snap_realm(ci);
985
986 if (!__ceph_is_any_real_caps(ci))
987 __cap_delay_cancel(mdsc, ci);
988}
989
990struct cap_msg_args {
991 struct ceph_mds_session *session;
992 u64 ino, cid, follows;
993 u64 flush_tid, oldest_flush_tid, size, max_size;
994 u64 xattr_version;
995 struct ceph_buffer *xattr_buf;
996 struct timespec atime, mtime, ctime;
997 int op, caps, wanted, dirty;
998 u32 seq, issue_seq, mseq, time_warp_seq;
999 u32 flags;
1000 kuid_t uid;
1001 kgid_t gid;
1002 umode_t mode;
1003 bool inline_data;
1004};
1005
1006/*
1007 * Build and send a cap message to the given MDS.
1008 *
1009 * Caller should be holding s_mutex.
1010 */
1011static int send_cap_msg(struct cap_msg_args *arg)
1012{
1013 struct ceph_mds_caps *fc;
1014 struct ceph_msg *msg;
1015 void *p;
1016 size_t extra_len;
1017 struct timespec zerotime = {0};
1018
1019 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1020 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1021 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1022 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1023 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1024 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1025 arg->mseq, arg->follows, arg->size, arg->max_size,
1026 arg->xattr_version,
1027 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1028
1029 /* flock buffer size + inline version + inline data size +
1030 * osd_epoch_barrier + oldest_flush_tid */
1031 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1032 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1033 GFP_NOFS, false);
1034 if (!msg)
1035 return -ENOMEM;
1036
1037 msg->hdr.version = cpu_to_le16(10);
1038 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1039
1040 fc = msg->front.iov_base;
1041 memset(fc, 0, sizeof(*fc));
1042
1043 fc->cap_id = cpu_to_le64(arg->cid);
1044 fc->op = cpu_to_le32(arg->op);
1045 fc->seq = cpu_to_le32(arg->seq);
1046 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1047 fc->migrate_seq = cpu_to_le32(arg->mseq);
1048 fc->caps = cpu_to_le32(arg->caps);
1049 fc->wanted = cpu_to_le32(arg->wanted);
1050 fc->dirty = cpu_to_le32(arg->dirty);
1051 fc->ino = cpu_to_le64(arg->ino);
1052 fc->snap_follows = cpu_to_le64(arg->follows);
1053
1054 fc->size = cpu_to_le64(arg->size);
1055 fc->max_size = cpu_to_le64(arg->max_size);
1056 ceph_encode_timespec(&fc->mtime, &arg->mtime);
1057 ceph_encode_timespec(&fc->atime, &arg->atime);
1058 ceph_encode_timespec(&fc->ctime, &arg->ctime);
1059 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1060
1061 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1062 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1063 fc->mode = cpu_to_le32(arg->mode);
1064
1065 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1066 if (arg->xattr_buf) {
1067 msg->middle = ceph_buffer_get(arg->xattr_buf);
1068 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1069 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1070 }
1071
1072 p = fc + 1;
1073 /* flock buffer size (version 2) */
1074 ceph_encode_32(&p, 0);
1075 /* inline version (version 4) */
1076 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1077 /* inline data size */
1078 ceph_encode_32(&p, 0);
1079 /* osd_epoch_barrier (version 5) */
1080 ceph_encode_32(&p, 0);
1081 /* oldest_flush_tid (version 6) */
1082 ceph_encode_64(&p, arg->oldest_flush_tid);
1083
1084 /*
1085 * caller_uid/caller_gid (version 7)
1086 *
1087 * Currently, we don't properly track which caller dirtied the caps
1088 * last, and force a flush of them when there is a conflict. For now,
1089 * just set this to 0:0, to emulate how the MDS has worked up to now.
1090 */
1091 ceph_encode_32(&p, 0);
1092 ceph_encode_32(&p, 0);
1093
1094 /* pool namespace (version 8) (mds always ignores this) */
1095 ceph_encode_32(&p, 0);
1096
1097 /*
1098 * btime and change_attr (version 9)
1099 *
1100 * We just zero these out for now, as the MDS ignores them unless
1101 * the requisite feature flags are set (which we don't do yet).
1102 */
1103 ceph_encode_timespec(p, &zerotime);
1104 p += sizeof(struct ceph_timespec);
1105 ceph_encode_64(&p, 0);
1106
1107 /* Advisory flags (version 10) */
1108 ceph_encode_32(&p, arg->flags);
1109
1110 ceph_con_send(&arg->session->s_con, msg);
1111 return 0;
1112}
1113
1114/*
1115 * Queue cap releases when an inode is dropped from our cache. Since
1116 * inode is about to be destroyed, there is no need for i_ceph_lock.
1117 */
1118void ceph_queue_caps_release(struct inode *inode)
1119{
1120 struct ceph_inode_info *ci = ceph_inode(inode);
1121 struct rb_node *p;
1122
1123 p = rb_first(&ci->i_caps);
1124 while (p) {
1125 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1126 p = rb_next(p);
1127 __ceph_remove_cap(cap, true);
1128 }
1129}
1130
1131/*
1132 * Send a cap msg on the given inode. Update our caps state, then
1133 * drop i_ceph_lock and send the message.
1134 *
1135 * Make note of max_size reported/requested from mds, revoked caps
1136 * that have now been implemented.
1137 *
1138 * Make half-hearted attempt ot to invalidate page cache if we are
1139 * dropping RDCACHE. Note that this will leave behind locked pages
1140 * that we'll then need to deal with elsewhere.
1141 *
1142 * Return non-zero if delayed release, or we experienced an error
1143 * such that the caller should requeue + retry later.
1144 *
1145 * called with i_ceph_lock, then drops it.
1146 * caller should hold snap_rwsem (read), s_mutex.
1147 */
1148static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1149 int op, bool sync, int used, int want, int retain,
1150 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1151 __releases(cap->ci->i_ceph_lock)
1152{
1153 struct ceph_inode_info *ci = cap->ci;
1154 struct inode *inode = &ci->vfs_inode;
1155 struct cap_msg_args arg;
1156 int held, revoking, dropping;
1157 int wake = 0;
1158 int delayed = 0;
1159 int ret;
1160
1161 held = cap->issued | cap->implemented;
1162 revoking = cap->implemented & ~cap->issued;
1163 retain &= ~revoking;
1164 dropping = cap->issued & ~retain;
1165
1166 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1167 inode, cap, cap->session,
1168 ceph_cap_string(held), ceph_cap_string(held & retain),
1169 ceph_cap_string(revoking));
1170 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1171
1172 arg.session = cap->session;
1173
1174 /* don't release wanted unless we've waited a bit. */
1175 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1176 time_before(jiffies, ci->i_hold_caps_min)) {
1177 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1178 ceph_cap_string(cap->issued),
1179 ceph_cap_string(cap->issued & retain),
1180 ceph_cap_string(cap->mds_wanted),
1181 ceph_cap_string(want));
1182 want |= cap->mds_wanted;
1183 retain |= cap->issued;
1184 delayed = 1;
1185 }
1186 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1187
1188 cap->issued &= retain; /* drop bits we don't want */
1189 if (cap->implemented & ~cap->issued) {
1190 /*
1191 * Wake up any waiters on wanted -> needed transition.
1192 * This is due to the weird transition from buffered
1193 * to sync IO... we need to flush dirty pages _before_
1194 * allowing sync writes to avoid reordering.
1195 */
1196 wake = 1;
1197 }
1198 cap->implemented &= cap->issued | used;
1199 cap->mds_wanted = want;
1200
1201 arg.ino = ceph_vino(inode).ino;
1202 arg.cid = cap->cap_id;
1203 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1204 arg.flush_tid = flush_tid;
1205 arg.oldest_flush_tid = oldest_flush_tid;
1206
1207 arg.size = inode->i_size;
1208 ci->i_reported_size = arg.size;
1209 arg.max_size = ci->i_wanted_max_size;
1210 ci->i_requested_max_size = arg.max_size;
1211
1212 if (flushing & CEPH_CAP_XATTR_EXCL) {
1213 __ceph_build_xattrs_blob(ci);
1214 arg.xattr_version = ci->i_xattrs.version;
1215 arg.xattr_buf = ci->i_xattrs.blob;
1216 } else {
1217 arg.xattr_buf = NULL;
1218 }
1219
1220 arg.mtime = inode->i_mtime;
1221 arg.atime = inode->i_atime;
1222 arg.ctime = inode->i_ctime;
1223
1224 arg.op = op;
1225 arg.caps = cap->implemented;
1226 arg.wanted = want;
1227 arg.dirty = flushing;
1228
1229 arg.seq = cap->seq;
1230 arg.issue_seq = cap->issue_seq;
1231 arg.mseq = cap->mseq;
1232 arg.time_warp_seq = ci->i_time_warp_seq;
1233
1234 arg.uid = inode->i_uid;
1235 arg.gid = inode->i_gid;
1236 arg.mode = inode->i_mode;
1237
1238 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1239 arg.flags = 0;
1240 if (sync)
1241 arg.flags |= CEPH_CLIENT_CAPS_SYNC;
1242
1243 spin_unlock(&ci->i_ceph_lock);
1244
1245 ret = send_cap_msg(&arg);
1246 if (ret < 0) {
1247 dout("error sending cap msg, must requeue %p\n", inode);
1248 delayed = 1;
1249 }
1250
1251 if (wake)
1252 wake_up_all(&ci->i_cap_wq);
1253
1254 return delayed;
1255}
1256
1257static inline int __send_flush_snap(struct inode *inode,
1258 struct ceph_mds_session *session,
1259 struct ceph_cap_snap *capsnap,
1260 u32 mseq, u64 oldest_flush_tid)
1261{
1262 struct cap_msg_args arg;
1263
1264 arg.session = session;
1265 arg.ino = ceph_vino(inode).ino;
1266 arg.cid = 0;
1267 arg.follows = capsnap->follows;
1268 arg.flush_tid = capsnap->cap_flush.tid;
1269 arg.oldest_flush_tid = oldest_flush_tid;
1270
1271 arg.size = capsnap->size;
1272 arg.max_size = 0;
1273 arg.xattr_version = capsnap->xattr_version;
1274 arg.xattr_buf = capsnap->xattr_blob;
1275
1276 arg.atime = capsnap->atime;
1277 arg.mtime = capsnap->mtime;
1278 arg.ctime = capsnap->ctime;
1279
1280 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1281 arg.caps = capsnap->issued;
1282 arg.wanted = 0;
1283 arg.dirty = capsnap->dirty;
1284
1285 arg.seq = 0;
1286 arg.issue_seq = 0;
1287 arg.mseq = mseq;
1288 arg.time_warp_seq = capsnap->time_warp_seq;
1289
1290 arg.uid = capsnap->uid;
1291 arg.gid = capsnap->gid;
1292 arg.mode = capsnap->mode;
1293
1294 arg.inline_data = capsnap->inline_data;
1295 arg.flags = 0;
1296
1297 return send_cap_msg(&arg);
1298}
1299
1300/*
1301 * When a snapshot is taken, clients accumulate dirty metadata on
1302 * inodes with capabilities in ceph_cap_snaps to describe the file
1303 * state at the time the snapshot was taken. This must be flushed
1304 * asynchronously back to the MDS once sync writes complete and dirty
1305 * data is written out.
1306 *
1307 * Called under i_ceph_lock. Takes s_mutex as needed.
1308 */
1309static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1310 struct ceph_mds_session *session)
1311 __releases(ci->i_ceph_lock)
1312 __acquires(ci->i_ceph_lock)
1313{
1314 struct inode *inode = &ci->vfs_inode;
1315 struct ceph_mds_client *mdsc = session->s_mdsc;
1316 struct ceph_cap_snap *capsnap;
1317 u64 oldest_flush_tid = 0;
1318 u64 first_tid = 1, last_tid = 0;
1319
1320 dout("__flush_snaps %p session %p\n", inode, session);
1321
1322 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1323 /*
1324 * we need to wait for sync writes to complete and for dirty
1325 * pages to be written out.
1326 */
1327 if (capsnap->dirty_pages || capsnap->writing)
1328 break;
1329
1330 /* should be removed by ceph_try_drop_cap_snap() */
1331 BUG_ON(!capsnap->need_flush);
1332
1333 /* only flush each capsnap once */
1334 if (capsnap->cap_flush.tid > 0) {
1335 dout(" already flushed %p, skipping\n", capsnap);
1336 continue;
1337 }
1338
1339 spin_lock(&mdsc->cap_dirty_lock);
1340 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1341 list_add_tail(&capsnap->cap_flush.g_list,
1342 &mdsc->cap_flush_list);
1343 if (oldest_flush_tid == 0)
1344 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1345 if (list_empty(&ci->i_flushing_item)) {
1346 list_add_tail(&ci->i_flushing_item,
1347 &session->s_cap_flushing);
1348 }
1349 spin_unlock(&mdsc->cap_dirty_lock);
1350
1351 list_add_tail(&capsnap->cap_flush.i_list,
1352 &ci->i_cap_flush_list);
1353
1354 if (first_tid == 1)
1355 first_tid = capsnap->cap_flush.tid;
1356 last_tid = capsnap->cap_flush.tid;
1357 }
1358
1359 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1360
1361 while (first_tid <= last_tid) {
1362 struct ceph_cap *cap = ci->i_auth_cap;
1363 struct ceph_cap_flush *cf;
1364 int ret;
1365
1366 if (!(cap && cap->session == session)) {
1367 dout("__flush_snaps %p auth cap %p not mds%d, "
1368 "stop\n", inode, cap, session->s_mds);
1369 break;
1370 }
1371
1372 ret = -ENOENT;
1373 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1374 if (cf->tid >= first_tid) {
1375 ret = 0;
1376 break;
1377 }
1378 }
1379 if (ret < 0)
1380 break;
1381
1382 first_tid = cf->tid + 1;
1383
1384 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1385 atomic_inc(&capsnap->nref);
1386 spin_unlock(&ci->i_ceph_lock);
1387
1388 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1389 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1390
1391 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1392 oldest_flush_tid);
1393 if (ret < 0) {
1394 pr_err("__flush_snaps: error sending cap flushsnap, "
1395 "ino (%llx.%llx) tid %llu follows %llu\n",
1396 ceph_vinop(inode), cf->tid, capsnap->follows);
1397 }
1398
1399 ceph_put_cap_snap(capsnap);
1400 spin_lock(&ci->i_ceph_lock);
1401 }
1402}
1403
1404void ceph_flush_snaps(struct ceph_inode_info *ci,
1405 struct ceph_mds_session **psession)
1406{
1407 struct inode *inode = &ci->vfs_inode;
1408 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1409 struct ceph_mds_session *session = NULL;
1410 int mds;
1411
1412 dout("ceph_flush_snaps %p\n", inode);
1413 if (psession)
1414 session = *psession;
1415retry:
1416 spin_lock(&ci->i_ceph_lock);
1417 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1418 dout(" no capsnap needs flush, doing nothing\n");
1419 goto out;
1420 }
1421 if (!ci->i_auth_cap) {
1422 dout(" no auth cap (migrating?), doing nothing\n");
1423 goto out;
1424 }
1425
1426 mds = ci->i_auth_cap->session->s_mds;
1427 if (session && session->s_mds != mds) {
1428 dout(" oops, wrong session %p mutex\n", session);
1429 mutex_unlock(&session->s_mutex);
1430 ceph_put_mds_session(session);
1431 session = NULL;
1432 }
1433 if (!session) {
1434 spin_unlock(&ci->i_ceph_lock);
1435 mutex_lock(&mdsc->mutex);
1436 session = __ceph_lookup_mds_session(mdsc, mds);
1437 mutex_unlock(&mdsc->mutex);
1438 if (session) {
1439 dout(" inverting session/ino locks on %p\n", session);
1440 mutex_lock(&session->s_mutex);
1441 }
1442 goto retry;
1443 }
1444
1445 __ceph_flush_snaps(ci, session);
1446out:
1447 spin_unlock(&ci->i_ceph_lock);
1448
1449 if (psession) {
1450 *psession = session;
1451 } else {
1452 mutex_unlock(&session->s_mutex);
1453 ceph_put_mds_session(session);
1454 }
1455 /* we flushed them all; remove this inode from the queue */
1456 spin_lock(&mdsc->snap_flush_lock);
1457 list_del_init(&ci->i_snap_flush_item);
1458 spin_unlock(&mdsc->snap_flush_lock);
1459}
1460
1461/*
1462 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1463 * Caller is then responsible for calling __mark_inode_dirty with the
1464 * returned flags value.
1465 */
1466int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1467 struct ceph_cap_flush **pcf)
1468{
1469 struct ceph_mds_client *mdsc =
1470 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1471 struct inode *inode = &ci->vfs_inode;
1472 int was = ci->i_dirty_caps;
1473 int dirty = 0;
1474
1475 if (!ci->i_auth_cap) {
1476 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1477 "but no auth cap (session was closed?)\n",
1478 inode, ceph_ino(inode), ceph_cap_string(mask));
1479 return 0;
1480 }
1481
1482 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1483 ceph_cap_string(mask), ceph_cap_string(was),
1484 ceph_cap_string(was | mask));
1485 ci->i_dirty_caps |= mask;
1486 if (was == 0) {
1487 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1488 swap(ci->i_prealloc_cap_flush, *pcf);
1489
1490 if (!ci->i_head_snapc) {
1491 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1492 ci->i_head_snapc = ceph_get_snap_context(
1493 ci->i_snap_realm->cached_context);
1494 }
1495 dout(" inode %p now dirty snapc %p auth cap %p\n",
1496 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1497 BUG_ON(!list_empty(&ci->i_dirty_item));
1498 spin_lock(&mdsc->cap_dirty_lock);
1499 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1500 spin_unlock(&mdsc->cap_dirty_lock);
1501 if (ci->i_flushing_caps == 0) {
1502 ihold(inode);
1503 dirty |= I_DIRTY_SYNC;
1504 }
1505 } else {
1506 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1507 }
1508 BUG_ON(list_empty(&ci->i_dirty_item));
1509 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1510 (mask & CEPH_CAP_FILE_BUFFER))
1511 dirty |= I_DIRTY_DATASYNC;
1512 __cap_delay_requeue(mdsc, ci);
1513 return dirty;
1514}
1515
1516struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1517{
1518 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1519}
1520
1521void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1522{
1523 if (cf)
1524 kmem_cache_free(ceph_cap_flush_cachep, cf);
1525}
1526
1527static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1528{
1529 if (!list_empty(&mdsc->cap_flush_list)) {
1530 struct ceph_cap_flush *cf =
1531 list_first_entry(&mdsc->cap_flush_list,
1532 struct ceph_cap_flush, g_list);
1533 return cf->tid;
1534 }
1535 return 0;
1536}
1537
1538/*
1539 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1540 * Return true if caller needs to wake up flush waiters.
1541 */
1542static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1543 struct ceph_inode_info *ci,
1544 struct ceph_cap_flush *cf)
1545{
1546 struct ceph_cap_flush *prev;
1547 bool wake = cf->wake;
1548 if (mdsc) {
1549 /* are there older pending cap flushes? */
1550 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1551 prev = list_prev_entry(cf, g_list);
1552 prev->wake = true;
1553 wake = false;
1554 }
1555 list_del(&cf->g_list);
1556 } else if (ci) {
1557 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1558 prev = list_prev_entry(cf, i_list);
1559 prev->wake = true;
1560 wake = false;
1561 }
1562 list_del(&cf->i_list);
1563 } else {
1564 BUG_ON(1);
1565 }
1566 return wake;
1567}
1568
1569/*
1570 * Add dirty inode to the flushing list. Assigned a seq number so we
1571 * can wait for caps to flush without starving.
1572 *
1573 * Called under i_ceph_lock.
1574 */
1575static int __mark_caps_flushing(struct inode *inode,
1576 struct ceph_mds_session *session, bool wake,
1577 u64 *flush_tid, u64 *oldest_flush_tid)
1578{
1579 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1580 struct ceph_inode_info *ci = ceph_inode(inode);
1581 struct ceph_cap_flush *cf = NULL;
1582 int flushing;
1583
1584 BUG_ON(ci->i_dirty_caps == 0);
1585 BUG_ON(list_empty(&ci->i_dirty_item));
1586 BUG_ON(!ci->i_prealloc_cap_flush);
1587
1588 flushing = ci->i_dirty_caps;
1589 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1590 ceph_cap_string(flushing),
1591 ceph_cap_string(ci->i_flushing_caps),
1592 ceph_cap_string(ci->i_flushing_caps | flushing));
1593 ci->i_flushing_caps |= flushing;
1594 ci->i_dirty_caps = 0;
1595 dout(" inode %p now !dirty\n", inode);
1596
1597 swap(cf, ci->i_prealloc_cap_flush);
1598 cf->caps = flushing;
1599 cf->wake = wake;
1600
1601 spin_lock(&mdsc->cap_dirty_lock);
1602 list_del_init(&ci->i_dirty_item);
1603
1604 cf->tid = ++mdsc->last_cap_flush_tid;
1605 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1606 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1607
1608 if (list_empty(&ci->i_flushing_item)) {
1609 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1610 mdsc->num_cap_flushing++;
1611 }
1612 spin_unlock(&mdsc->cap_dirty_lock);
1613
1614 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1615
1616 *flush_tid = cf->tid;
1617 return flushing;
1618}
1619
1620/*
1621 * try to invalidate mapping pages without blocking.
1622 */
1623static int try_nonblocking_invalidate(struct inode *inode)
1624{
1625 struct ceph_inode_info *ci = ceph_inode(inode);
1626 u32 invalidating_gen = ci->i_rdcache_gen;
1627
1628 spin_unlock(&ci->i_ceph_lock);
1629 invalidate_mapping_pages(&inode->i_data, 0, -1);
1630 spin_lock(&ci->i_ceph_lock);
1631
1632 if (inode->i_data.nrpages == 0 &&
1633 invalidating_gen == ci->i_rdcache_gen) {
1634 /* success. */
1635 dout("try_nonblocking_invalidate %p success\n", inode);
1636 /* save any racing async invalidate some trouble */
1637 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1638 return 0;
1639 }
1640 dout("try_nonblocking_invalidate %p failed\n", inode);
1641 return -1;
1642}
1643
1644/*
1645 * Swiss army knife function to examine currently used and wanted
1646 * versus held caps. Release, flush, ack revoked caps to mds as
1647 * appropriate.
1648 *
1649 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1650 * cap release further.
1651 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1652 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1653 * further delay.
1654 */
1655void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1656 struct ceph_mds_session *session)
1657{
1658 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1659 struct ceph_mds_client *mdsc = fsc->mdsc;
1660 struct inode *inode = &ci->vfs_inode;
1661 struct ceph_cap *cap;
1662 u64 flush_tid, oldest_flush_tid;
1663 int file_wanted, used, cap_used;
1664 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1665 int issued, implemented, want, retain, revoking, flushing = 0;
1666 int mds = -1; /* keep track of how far we've gone through i_caps list
1667 to avoid an infinite loop on retry */
1668 struct rb_node *p;
1669 int delayed = 0, sent = 0, num;
1670 bool is_delayed = flags & CHECK_CAPS_NODELAY;
1671 bool queue_invalidate = false;
1672 bool force_requeue = false;
1673 bool tried_invalidate = false;
1674
1675 /* if we are unmounting, flush any unused caps immediately. */
1676 if (mdsc->stopping)
1677 is_delayed = 1;
1678
1679 spin_lock(&ci->i_ceph_lock);
1680
1681 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1682 flags |= CHECK_CAPS_FLUSH;
1683
1684 goto retry_locked;
1685retry:
1686 spin_lock(&ci->i_ceph_lock);
1687retry_locked:
1688 file_wanted = __ceph_caps_file_wanted(ci);
1689 used = __ceph_caps_used(ci);
1690 issued = __ceph_caps_issued(ci, &implemented);
1691 revoking = implemented & ~issued;
1692
1693 want = file_wanted;
1694 retain = file_wanted | used | CEPH_CAP_PIN;
1695 if (!mdsc->stopping && inode->i_nlink > 0) {
1696 if (file_wanted) {
1697 retain |= CEPH_CAP_ANY; /* be greedy */
1698 } else if (S_ISDIR(inode->i_mode) &&
1699 (issued & CEPH_CAP_FILE_SHARED) &&
1700 __ceph_dir_is_complete(ci)) {
1701 /*
1702 * If a directory is complete, we want to keep
1703 * the exclusive cap. So that MDS does not end up
1704 * revoking the shared cap on every create/unlink
1705 * operation.
1706 */
1707 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1708 retain |= want;
1709 } else {
1710
1711 retain |= CEPH_CAP_ANY_SHARED;
1712 /*
1713 * keep RD only if we didn't have the file open RW,
1714 * because then the mds would revoke it anyway to
1715 * journal max_size=0.
1716 */
1717 if (ci->i_max_size == 0)
1718 retain |= CEPH_CAP_ANY_RD;
1719 }
1720 }
1721
1722 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1723 " issued %s revoking %s retain %s %s%s%s\n", inode,
1724 ceph_cap_string(file_wanted),
1725 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1726 ceph_cap_string(ci->i_flushing_caps),
1727 ceph_cap_string(issued), ceph_cap_string(revoking),
1728 ceph_cap_string(retain),
1729 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1730 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1731 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1732
1733 /*
1734 * If we no longer need to hold onto old our caps, and we may
1735 * have cached pages, but don't want them, then try to invalidate.
1736 * If we fail, it's because pages are locked.... try again later.
1737 */
1738 if ((!is_delayed || mdsc->stopping) &&
1739 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1740 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1741 inode->i_data.nrpages && /* have cached pages */
1742 (revoking & (CEPH_CAP_FILE_CACHE|
1743 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1744 !tried_invalidate) {
1745 dout("check_caps trying to invalidate on %p\n", inode);
1746 if (try_nonblocking_invalidate(inode) < 0) {
1747 if (revoking & (CEPH_CAP_FILE_CACHE|
1748 CEPH_CAP_FILE_LAZYIO)) {
1749 dout("check_caps queuing invalidate\n");
1750 queue_invalidate = true;
1751 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1752 } else {
1753 dout("check_caps failed to invalidate pages\n");
1754 /* we failed to invalidate pages. check these
1755 caps again later. */
1756 force_requeue = true;
1757 __cap_set_timeouts(mdsc, ci);
1758 }
1759 }
1760 tried_invalidate = true;
1761 goto retry_locked;
1762 }
1763
1764 num = 0;
1765 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1766 cap = rb_entry(p, struct ceph_cap, ci_node);
1767 num++;
1768
1769 /* avoid looping forever */
1770 if (mds >= cap->mds ||
1771 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1772 continue;
1773
1774 /* NOTE: no side-effects allowed, until we take s_mutex */
1775
1776 cap_used = used;
1777 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1778 cap_used &= ~ci->i_auth_cap->issued;
1779
1780 revoking = cap->implemented & ~cap->issued;
1781 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1782 cap->mds, cap, ceph_cap_string(cap_used),
1783 ceph_cap_string(cap->issued),
1784 ceph_cap_string(cap->implemented),
1785 ceph_cap_string(revoking));
1786
1787 if (cap == ci->i_auth_cap &&
1788 (cap->issued & CEPH_CAP_FILE_WR)) {
1789 /* request larger max_size from MDS? */
1790 if (ci->i_wanted_max_size > ci->i_max_size &&
1791 ci->i_wanted_max_size > ci->i_requested_max_size) {
1792 dout("requesting new max_size\n");
1793 goto ack;
1794 }
1795
1796 /* approaching file_max? */
1797 if ((inode->i_size << 1) >= ci->i_max_size &&
1798 (ci->i_reported_size << 1) < ci->i_max_size) {
1799 dout("i_size approaching max_size\n");
1800 goto ack;
1801 }
1802 }
1803 /* flush anything dirty? */
1804 if (cap == ci->i_auth_cap) {
1805 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1806 dout("flushing dirty caps\n");
1807 goto ack;
1808 }
1809 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1810 dout("flushing snap caps\n");
1811 goto ack;
1812 }
1813 }
1814
1815 /* completed revocation? going down and there are no caps? */
1816 if (revoking && (revoking & cap_used) == 0) {
1817 dout("completed revocation of %s\n",
1818 ceph_cap_string(cap->implemented & ~cap->issued));
1819 goto ack;
1820 }
1821
1822 /* want more caps from mds? */
1823 if (want & ~(cap->mds_wanted | cap->issued))
1824 goto ack;
1825
1826 /* things we might delay */
1827 if ((cap->issued & ~retain) == 0 &&
1828 cap->mds_wanted == want)
1829 continue; /* nope, all good */
1830
1831 if (is_delayed)
1832 goto ack;
1833
1834 /* delay? */
1835 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1836 time_before(jiffies, ci->i_hold_caps_max)) {
1837 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1838 ceph_cap_string(cap->issued),
1839 ceph_cap_string(cap->issued & retain),
1840 ceph_cap_string(cap->mds_wanted),
1841 ceph_cap_string(want));
1842 delayed++;
1843 continue;
1844 }
1845
1846ack:
1847 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1848 dout(" skipping %p I_NOFLUSH set\n", inode);
1849 continue;
1850 }
1851
1852 if (session && session != cap->session) {
1853 dout("oops, wrong session %p mutex\n", session);
1854 mutex_unlock(&session->s_mutex);
1855 session = NULL;
1856 }
1857 if (!session) {
1858 session = cap->session;
1859 if (mutex_trylock(&session->s_mutex) == 0) {
1860 dout("inverting session/ino locks on %p\n",
1861 session);
1862 spin_unlock(&ci->i_ceph_lock);
1863 if (took_snap_rwsem) {
1864 up_read(&mdsc->snap_rwsem);
1865 took_snap_rwsem = 0;
1866 }
1867 mutex_lock(&session->s_mutex);
1868 goto retry;
1869 }
1870 }
1871
1872 /* kick flushing and flush snaps before sending normal
1873 * cap message */
1874 if (cap == ci->i_auth_cap &&
1875 (ci->i_ceph_flags &
1876 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
1877 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1878 spin_lock(&mdsc->cap_dirty_lock);
1879 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1880 spin_unlock(&mdsc->cap_dirty_lock);
1881 __kick_flushing_caps(mdsc, session, ci,
1882 oldest_flush_tid);
1883 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1884 }
1885 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
1886 __ceph_flush_snaps(ci, session);
1887
1888 goto retry_locked;
1889 }
1890
1891 /* take snap_rwsem after session mutex */
1892 if (!took_snap_rwsem) {
1893 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1894 dout("inverting snap/in locks on %p\n",
1895 inode);
1896 spin_unlock(&ci->i_ceph_lock);
1897 down_read(&mdsc->snap_rwsem);
1898 took_snap_rwsem = 1;
1899 goto retry;
1900 }
1901 took_snap_rwsem = 1;
1902 }
1903
1904 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1905 flushing = __mark_caps_flushing(inode, session, false,
1906 &flush_tid,
1907 &oldest_flush_tid);
1908 } else {
1909 flushing = 0;
1910 flush_tid = 0;
1911 spin_lock(&mdsc->cap_dirty_lock);
1912 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1913 spin_unlock(&mdsc->cap_dirty_lock);
1914 }
1915
1916 mds = cap->mds; /* remember mds, so we don't repeat */
1917 sent++;
1918
1919 /* __send_cap drops i_ceph_lock */
1920 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, false,
1921 cap_used, want, retain, flushing,
1922 flush_tid, oldest_flush_tid);
1923 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1924 }
1925
1926 /*
1927 * Reschedule delayed caps release if we delayed anything,
1928 * otherwise cancel.
1929 */
1930 if (delayed && is_delayed)
1931 force_requeue = true; /* __send_cap delayed release; requeue */
1932 if (!delayed && !is_delayed)
1933 __cap_delay_cancel(mdsc, ci);
1934 else if (!is_delayed || force_requeue)
1935 __cap_delay_requeue(mdsc, ci);
1936
1937 spin_unlock(&ci->i_ceph_lock);
1938
1939 if (queue_invalidate)
1940 ceph_queue_invalidate(inode);
1941
1942 if (session)
1943 mutex_unlock(&session->s_mutex);
1944 if (took_snap_rwsem)
1945 up_read(&mdsc->snap_rwsem);
1946}
1947
1948/*
1949 * Try to flush dirty caps back to the auth mds.
1950 */
1951static int try_flush_caps(struct inode *inode, u64 *ptid)
1952{
1953 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1954 struct ceph_inode_info *ci = ceph_inode(inode);
1955 struct ceph_mds_session *session = NULL;
1956 int flushing = 0;
1957 u64 flush_tid = 0, oldest_flush_tid = 0;
1958
1959retry:
1960 spin_lock(&ci->i_ceph_lock);
1961 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1962 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1963 goto out;
1964 }
1965 if (ci->i_dirty_caps && ci->i_auth_cap) {
1966 struct ceph_cap *cap = ci->i_auth_cap;
1967 int used = __ceph_caps_used(ci);
1968 int want = __ceph_caps_wanted(ci);
1969 int delayed;
1970
1971 if (!session || session != cap->session) {
1972 spin_unlock(&ci->i_ceph_lock);
1973 if (session)
1974 mutex_unlock(&session->s_mutex);
1975 session = cap->session;
1976 mutex_lock(&session->s_mutex);
1977 goto retry;
1978 }
1979 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1980 goto out;
1981
1982 flushing = __mark_caps_flushing(inode, session, true,
1983 &flush_tid, &oldest_flush_tid);
1984
1985 /* __send_cap drops i_ceph_lock */
1986 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, true,
1987 used, want, (cap->issued | cap->implemented),
1988 flushing, flush_tid, oldest_flush_tid);
1989
1990 if (delayed) {
1991 spin_lock(&ci->i_ceph_lock);
1992 __cap_delay_requeue(mdsc, ci);
1993 spin_unlock(&ci->i_ceph_lock);
1994 }
1995 } else {
1996 if (!list_empty(&ci->i_cap_flush_list)) {
1997 struct ceph_cap_flush *cf =
1998 list_last_entry(&ci->i_cap_flush_list,
1999 struct ceph_cap_flush, i_list);
2000 cf->wake = true;
2001 flush_tid = cf->tid;
2002 }
2003 flushing = ci->i_flushing_caps;
2004 spin_unlock(&ci->i_ceph_lock);
2005 }
2006out:
2007 if (session)
2008 mutex_unlock(&session->s_mutex);
2009
2010 *ptid = flush_tid;
2011 return flushing;
2012}
2013
2014/*
2015 * Return true if we've flushed caps through the given flush_tid.
2016 */
2017static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2018{
2019 struct ceph_inode_info *ci = ceph_inode(inode);
2020 int ret = 1;
2021
2022 spin_lock(&ci->i_ceph_lock);
2023 if (!list_empty(&ci->i_cap_flush_list)) {
2024 struct ceph_cap_flush * cf =
2025 list_first_entry(&ci->i_cap_flush_list,
2026 struct ceph_cap_flush, i_list);
2027 if (cf->tid <= flush_tid)
2028 ret = 0;
2029 }
2030 spin_unlock(&ci->i_ceph_lock);
2031 return ret;
2032}
2033
2034/*
2035 * wait for any unsafe requests to complete.
2036 */
2037static int unsafe_request_wait(struct inode *inode)
2038{
2039 struct ceph_inode_info *ci = ceph_inode(inode);
2040 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2041 int ret, err = 0;
2042
2043 spin_lock(&ci->i_unsafe_lock);
2044 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2045 req1 = list_last_entry(&ci->i_unsafe_dirops,
2046 struct ceph_mds_request,
2047 r_unsafe_dir_item);
2048 ceph_mdsc_get_request(req1);
2049 }
2050 if (!list_empty(&ci->i_unsafe_iops)) {
2051 req2 = list_last_entry(&ci->i_unsafe_iops,
2052 struct ceph_mds_request,
2053 r_unsafe_target_item);
2054 ceph_mdsc_get_request(req2);
2055 }
2056 spin_unlock(&ci->i_unsafe_lock);
2057
2058 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2059 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2060 if (req1) {
2061 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2062 ceph_timeout_jiffies(req1->r_timeout));
2063 if (ret)
2064 err = -EIO;
2065 ceph_mdsc_put_request(req1);
2066 }
2067 if (req2) {
2068 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2069 ceph_timeout_jiffies(req2->r_timeout));
2070 if (ret)
2071 err = -EIO;
2072 ceph_mdsc_put_request(req2);
2073 }
2074 return err;
2075}
2076
2077int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2078{
2079 struct inode *inode = file->f_mapping->host;
2080 struct ceph_inode_info *ci = ceph_inode(inode);
2081 u64 flush_tid;
2082 int ret;
2083 int dirty;
2084
2085 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2086
2087 ceph_sync_write_wait(inode);
2088
2089 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
2090 if (ret < 0)
2091 goto out;
2092
2093 if (datasync)
2094 goto out;
2095
2096 inode_lock(inode);
2097
2098 dirty = try_flush_caps(inode, &flush_tid);
2099 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2100
2101 ret = unsafe_request_wait(inode);
2102
2103 /*
2104 * only wait on non-file metadata writeback (the mds
2105 * can recover size and mtime, so we don't need to
2106 * wait for that)
2107 */
2108 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2109 ret = wait_event_interruptible(ci->i_cap_wq,
2110 caps_are_flushed(inode, flush_tid));
2111 }
2112 inode_unlock(inode);
2113out:
2114 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2115 return ret;
2116}
2117
2118/*
2119 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2120 * queue inode for flush but don't do so immediately, because we can
2121 * get by with fewer MDS messages if we wait for data writeback to
2122 * complete first.
2123 */
2124int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2125{
2126 struct ceph_inode_info *ci = ceph_inode(inode);
2127 u64 flush_tid;
2128 int err = 0;
2129 int dirty;
2130 int wait = wbc->sync_mode == WB_SYNC_ALL;
2131
2132 dout("write_inode %p wait=%d\n", inode, wait);
2133 if (wait) {
2134 dirty = try_flush_caps(inode, &flush_tid);
2135 if (dirty)
2136 err = wait_event_interruptible(ci->i_cap_wq,
2137 caps_are_flushed(inode, flush_tid));
2138 } else {
2139 struct ceph_mds_client *mdsc =
2140 ceph_sb_to_client(inode->i_sb)->mdsc;
2141
2142 spin_lock(&ci->i_ceph_lock);
2143 if (__ceph_caps_dirty(ci))
2144 __cap_delay_requeue_front(mdsc, ci);
2145 spin_unlock(&ci->i_ceph_lock);
2146 }
2147 return err;
2148}
2149
2150static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2151 struct ceph_mds_session *session,
2152 struct ceph_inode_info *ci,
2153 u64 oldest_flush_tid)
2154 __releases(ci->i_ceph_lock)
2155 __acquires(ci->i_ceph_lock)
2156{
2157 struct inode *inode = &ci->vfs_inode;
2158 struct ceph_cap *cap;
2159 struct ceph_cap_flush *cf;
2160 int ret;
2161 u64 first_tid = 0;
2162
2163 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2164 if (cf->tid < first_tid)
2165 continue;
2166
2167 cap = ci->i_auth_cap;
2168 if (!(cap && cap->session == session)) {
2169 pr_err("%p auth cap %p not mds%d ???\n",
2170 inode, cap, session->s_mds);
2171 break;
2172 }
2173
2174 first_tid = cf->tid + 1;
2175
2176 if (cf->caps) {
2177 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2178 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2179 ci->i_ceph_flags |= CEPH_I_NODELAY;
2180 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2181 false, __ceph_caps_used(ci),
2182 __ceph_caps_wanted(ci),
2183 cap->issued | cap->implemented,
2184 cf->caps, cf->tid, oldest_flush_tid);
2185 if (ret) {
2186 pr_err("kick_flushing_caps: error sending "
2187 "cap flush, ino (%llx.%llx) "
2188 "tid %llu flushing %s\n",
2189 ceph_vinop(inode), cf->tid,
2190 ceph_cap_string(cf->caps));
2191 }
2192 } else {
2193 struct ceph_cap_snap *capsnap =
2194 container_of(cf, struct ceph_cap_snap,
2195 cap_flush);
2196 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2197 inode, capsnap, cf->tid,
2198 ceph_cap_string(capsnap->dirty));
2199
2200 atomic_inc(&capsnap->nref);
2201 spin_unlock(&ci->i_ceph_lock);
2202
2203 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2204 oldest_flush_tid);
2205 if (ret < 0) {
2206 pr_err("kick_flushing_caps: error sending "
2207 "cap flushsnap, ino (%llx.%llx) "
2208 "tid %llu follows %llu\n",
2209 ceph_vinop(inode), cf->tid,
2210 capsnap->follows);
2211 }
2212
2213 ceph_put_cap_snap(capsnap);
2214 }
2215
2216 spin_lock(&ci->i_ceph_lock);
2217 }
2218}
2219
2220void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2221 struct ceph_mds_session *session)
2222{
2223 struct ceph_inode_info *ci;
2224 struct ceph_cap *cap;
2225 u64 oldest_flush_tid;
2226
2227 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2228
2229 spin_lock(&mdsc->cap_dirty_lock);
2230 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2231 spin_unlock(&mdsc->cap_dirty_lock);
2232
2233 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2234 spin_lock(&ci->i_ceph_lock);
2235 cap = ci->i_auth_cap;
2236 if (!(cap && cap->session == session)) {
2237 pr_err("%p auth cap %p not mds%d ???\n",
2238 &ci->vfs_inode, cap, session->s_mds);
2239 spin_unlock(&ci->i_ceph_lock);
2240 continue;
2241 }
2242
2243
2244 /*
2245 * if flushing caps were revoked, we re-send the cap flush
2246 * in client reconnect stage. This guarantees MDS * processes
2247 * the cap flush message before issuing the flushing caps to
2248 * other client.
2249 */
2250 if ((cap->issued & ci->i_flushing_caps) !=
2251 ci->i_flushing_caps) {
2252 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2253 __kick_flushing_caps(mdsc, session, ci,
2254 oldest_flush_tid);
2255 } else {
2256 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2257 }
2258
2259 spin_unlock(&ci->i_ceph_lock);
2260 }
2261}
2262
2263void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2264 struct ceph_mds_session *session)
2265{
2266 struct ceph_inode_info *ci;
2267 struct ceph_cap *cap;
2268 u64 oldest_flush_tid;
2269
2270 dout("kick_flushing_caps mds%d\n", session->s_mds);
2271
2272 spin_lock(&mdsc->cap_dirty_lock);
2273 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2274 spin_unlock(&mdsc->cap_dirty_lock);
2275
2276 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2277 spin_lock(&ci->i_ceph_lock);
2278 cap = ci->i_auth_cap;
2279 if (!(cap && cap->session == session)) {
2280 pr_err("%p auth cap %p not mds%d ???\n",
2281 &ci->vfs_inode, cap, session->s_mds);
2282 spin_unlock(&ci->i_ceph_lock);
2283 continue;
2284 }
2285 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2286 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2287 __kick_flushing_caps(mdsc, session, ci,
2288 oldest_flush_tid);
2289 }
2290 spin_unlock(&ci->i_ceph_lock);
2291 }
2292}
2293
2294static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2295 struct ceph_mds_session *session,
2296 struct inode *inode)
2297 __releases(ci->i_ceph_lock)
2298{
2299 struct ceph_inode_info *ci = ceph_inode(inode);
2300 struct ceph_cap *cap;
2301
2302 cap = ci->i_auth_cap;
2303 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2304 ceph_cap_string(ci->i_flushing_caps));
2305
2306 if (!list_empty(&ci->i_cap_flush_list)) {
2307 u64 oldest_flush_tid;
2308 spin_lock(&mdsc->cap_dirty_lock);
2309 list_move_tail(&ci->i_flushing_item,
2310 &cap->session->s_cap_flushing);
2311 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2312 spin_unlock(&mdsc->cap_dirty_lock);
2313
2314 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2315 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2316 spin_unlock(&ci->i_ceph_lock);
2317 } else {
2318 spin_unlock(&ci->i_ceph_lock);
2319 }
2320}
2321
2322
2323/*
2324 * Take references to capabilities we hold, so that we don't release
2325 * them to the MDS prematurely.
2326 *
2327 * Protected by i_ceph_lock.
2328 */
2329static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2330 bool snap_rwsem_locked)
2331{
2332 if (got & CEPH_CAP_PIN)
2333 ci->i_pin_ref++;
2334 if (got & CEPH_CAP_FILE_RD)
2335 ci->i_rd_ref++;
2336 if (got & CEPH_CAP_FILE_CACHE)
2337 ci->i_rdcache_ref++;
2338 if (got & CEPH_CAP_FILE_WR) {
2339 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2340 BUG_ON(!snap_rwsem_locked);
2341 ci->i_head_snapc = ceph_get_snap_context(
2342 ci->i_snap_realm->cached_context);
2343 }
2344 ci->i_wr_ref++;
2345 }
2346 if (got & CEPH_CAP_FILE_BUFFER) {
2347 if (ci->i_wb_ref == 0)
2348 ihold(&ci->vfs_inode);
2349 ci->i_wb_ref++;
2350 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2351 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2352 }
2353}
2354
2355/*
2356 * Try to grab cap references. Specify those refs we @want, and the
2357 * minimal set we @need. Also include the larger offset we are writing
2358 * to (when applicable), and check against max_size here as well.
2359 * Note that caller is responsible for ensuring max_size increases are
2360 * requested from the MDS.
2361 */
2362static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2363 loff_t endoff, bool nonblock, int *got, int *err)
2364{
2365 struct inode *inode = &ci->vfs_inode;
2366 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2367 int ret = 0;
2368 int have, implemented;
2369 int file_wanted;
2370 bool snap_rwsem_locked = false;
2371
2372 dout("get_cap_refs %p need %s want %s\n", inode,
2373 ceph_cap_string(need), ceph_cap_string(want));
2374
2375again:
2376 spin_lock(&ci->i_ceph_lock);
2377
2378 /* make sure file is actually open */
2379 file_wanted = __ceph_caps_file_wanted(ci);
2380 if ((file_wanted & need) != need) {
2381 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2382 ceph_cap_string(need), ceph_cap_string(file_wanted));
2383 *err = -EBADF;
2384 ret = 1;
2385 goto out_unlock;
2386 }
2387
2388 /* finish pending truncate */
2389 while (ci->i_truncate_pending) {
2390 spin_unlock(&ci->i_ceph_lock);
2391 if (snap_rwsem_locked) {
2392 up_read(&mdsc->snap_rwsem);
2393 snap_rwsem_locked = false;
2394 }
2395 __ceph_do_pending_vmtruncate(inode);
2396 spin_lock(&ci->i_ceph_lock);
2397 }
2398
2399 have = __ceph_caps_issued(ci, &implemented);
2400
2401 if (have & need & CEPH_CAP_FILE_WR) {
2402 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2403 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2404 inode, endoff, ci->i_max_size);
2405 if (endoff > ci->i_requested_max_size) {
2406 *err = -EAGAIN;
2407 ret = 1;
2408 }
2409 goto out_unlock;
2410 }
2411 /*
2412 * If a sync write is in progress, we must wait, so that we
2413 * can get a final snapshot value for size+mtime.
2414 */
2415 if (__ceph_have_pending_cap_snap(ci)) {
2416 dout("get_cap_refs %p cap_snap_pending\n", inode);
2417 goto out_unlock;
2418 }
2419 }
2420
2421 if ((have & need) == need) {
2422 /*
2423 * Look at (implemented & ~have & not) so that we keep waiting
2424 * on transition from wanted -> needed caps. This is needed
2425 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2426 * going before a prior buffered writeback happens.
2427 */
2428 int not = want & ~(have & need);
2429 int revoking = implemented & ~have;
2430 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2431 inode, ceph_cap_string(have), ceph_cap_string(not),
2432 ceph_cap_string(revoking));
2433 if ((revoking & not) == 0) {
2434 if (!snap_rwsem_locked &&
2435 !ci->i_head_snapc &&
2436 (need & CEPH_CAP_FILE_WR)) {
2437 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2438 /*
2439 * we can not call down_read() when
2440 * task isn't in TASK_RUNNING state
2441 */
2442 if (nonblock) {
2443 *err = -EAGAIN;
2444 ret = 1;
2445 goto out_unlock;
2446 }
2447
2448 spin_unlock(&ci->i_ceph_lock);
2449 down_read(&mdsc->snap_rwsem);
2450 snap_rwsem_locked = true;
2451 goto again;
2452 }
2453 snap_rwsem_locked = true;
2454 }
2455 *got = need | (have & want);
2456 if ((need & CEPH_CAP_FILE_RD) &&
2457 !(*got & CEPH_CAP_FILE_CACHE))
2458 ceph_disable_fscache_readpage(ci);
2459 __take_cap_refs(ci, *got, true);
2460 ret = 1;
2461 }
2462 } else {
2463 int session_readonly = false;
2464 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2465 struct ceph_mds_session *s = ci->i_auth_cap->session;
2466 spin_lock(&s->s_cap_lock);
2467 session_readonly = s->s_readonly;
2468 spin_unlock(&s->s_cap_lock);
2469 }
2470 if (session_readonly) {
2471 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2472 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2473 *err = -EROFS;
2474 ret = 1;
2475 goto out_unlock;
2476 }
2477
2478 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2479 int mds_wanted;
2480 if (ACCESS_ONCE(mdsc->fsc->mount_state) ==
2481 CEPH_MOUNT_SHUTDOWN) {
2482 dout("get_cap_refs %p forced umount\n", inode);
2483 *err = -EIO;
2484 ret = 1;
2485 goto out_unlock;
2486 }
2487 mds_wanted = __ceph_caps_mds_wanted(ci);
2488 if ((mds_wanted & need) != need) {
2489 dout("get_cap_refs %p caps were dropped"
2490 " (session killed?)\n", inode);
2491 *err = -ESTALE;
2492 ret = 1;
2493 goto out_unlock;
2494 }
2495 if ((mds_wanted & file_wanted) ==
2496 (file_wanted & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR)))
2497 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2498 }
2499
2500 dout("get_cap_refs %p have %s needed %s\n", inode,
2501 ceph_cap_string(have), ceph_cap_string(need));
2502 }
2503out_unlock:
2504 spin_unlock(&ci->i_ceph_lock);
2505 if (snap_rwsem_locked)
2506 up_read(&mdsc->snap_rwsem);
2507
2508 dout("get_cap_refs %p ret %d got %s\n", inode,
2509 ret, ceph_cap_string(*got));
2510 return ret;
2511}
2512
2513/*
2514 * Check the offset we are writing up to against our current
2515 * max_size. If necessary, tell the MDS we want to write to
2516 * a larger offset.
2517 */
2518static void check_max_size(struct inode *inode, loff_t endoff)
2519{
2520 struct ceph_inode_info *ci = ceph_inode(inode);
2521 int check = 0;
2522
2523 /* do we need to explicitly request a larger max_size? */
2524 spin_lock(&ci->i_ceph_lock);
2525 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2526 dout("write %p at large endoff %llu, req max_size\n",
2527 inode, endoff);
2528 ci->i_wanted_max_size = endoff;
2529 }
2530 /* duplicate ceph_check_caps()'s logic */
2531 if (ci->i_auth_cap &&
2532 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2533 ci->i_wanted_max_size > ci->i_max_size &&
2534 ci->i_wanted_max_size > ci->i_requested_max_size)
2535 check = 1;
2536 spin_unlock(&ci->i_ceph_lock);
2537 if (check)
2538 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2539}
2540
2541int ceph_try_get_caps(struct ceph_inode_info *ci, int need, int want, int *got)
2542{
2543 int ret, err = 0;
2544
2545 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2546 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
2547 ret = ceph_pool_perm_check(ci, need);
2548 if (ret < 0)
2549 return ret;
2550
2551 ret = try_get_cap_refs(ci, need, want, 0, true, got, &err);
2552 if (ret) {
2553 if (err == -EAGAIN) {
2554 ret = 0;
2555 } else if (err < 0) {
2556 ret = err;
2557 }
2558 }
2559 return ret;
2560}
2561
2562/*
2563 * Wait for caps, and take cap references. If we can't get a WR cap
2564 * due to a small max_size, make sure we check_max_size (and possibly
2565 * ask the mds) so we don't get hung up indefinitely.
2566 */
2567int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2568 loff_t endoff, int *got, struct page **pinned_page)
2569{
2570 int _got, ret, err = 0;
2571
2572 ret = ceph_pool_perm_check(ci, need);
2573 if (ret < 0)
2574 return ret;
2575
2576 while (true) {
2577 if (endoff > 0)
2578 check_max_size(&ci->vfs_inode, endoff);
2579
2580 err = 0;
2581 _got = 0;
2582 ret = try_get_cap_refs(ci, need, want, endoff,
2583 false, &_got, &err);
2584 if (ret) {
2585 if (err == -EAGAIN)
2586 continue;
2587 if (err < 0)
2588 ret = err;
2589 } else {
2590 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2591 add_wait_queue(&ci->i_cap_wq, &wait);
2592
2593 while (!try_get_cap_refs(ci, need, want, endoff,
2594 true, &_got, &err)) {
2595 if (signal_pending(current)) {
2596 ret = -ERESTARTSYS;
2597 break;
2598 }
2599 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2600 }
2601
2602 remove_wait_queue(&ci->i_cap_wq, &wait);
2603
2604 if (err == -EAGAIN)
2605 continue;
2606 if (err < 0)
2607 ret = err;
2608 }
2609 if (ret < 0) {
2610 if (err == -ESTALE) {
2611 /* session was killed, try renew caps */
2612 ret = ceph_renew_caps(&ci->vfs_inode);
2613 if (ret == 0)
2614 continue;
2615 }
2616 return ret;
2617 }
2618
2619 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2620 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2621 i_size_read(&ci->vfs_inode) > 0) {
2622 struct page *page =
2623 find_get_page(ci->vfs_inode.i_mapping, 0);
2624 if (page) {
2625 if (PageUptodate(page)) {
2626 *pinned_page = page;
2627 break;
2628 }
2629 put_page(page);
2630 }
2631 /*
2632 * drop cap refs first because getattr while
2633 * holding * caps refs can cause deadlock.
2634 */
2635 ceph_put_cap_refs(ci, _got);
2636 _got = 0;
2637
2638 /*
2639 * getattr request will bring inline data into
2640 * page cache
2641 */
2642 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2643 CEPH_STAT_CAP_INLINE_DATA,
2644 true);
2645 if (ret < 0)
2646 return ret;
2647 continue;
2648 }
2649 break;
2650 }
2651
2652 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2653 ceph_fscache_revalidate_cookie(ci);
2654
2655 *got = _got;
2656 return 0;
2657}
2658
2659/*
2660 * Take cap refs. Caller must already know we hold at least one ref
2661 * on the caps in question or we don't know this is safe.
2662 */
2663void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2664{
2665 spin_lock(&ci->i_ceph_lock);
2666 __take_cap_refs(ci, caps, false);
2667 spin_unlock(&ci->i_ceph_lock);
2668}
2669
2670
2671/*
2672 * drop cap_snap that is not associated with any snapshot.
2673 * we don't need to send FLUSHSNAP message for it.
2674 */
2675static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2676 struct ceph_cap_snap *capsnap)
2677{
2678 if (!capsnap->need_flush &&
2679 !capsnap->writing && !capsnap->dirty_pages) {
2680 dout("dropping cap_snap %p follows %llu\n",
2681 capsnap, capsnap->follows);
2682 BUG_ON(capsnap->cap_flush.tid > 0);
2683 ceph_put_snap_context(capsnap->context);
2684 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2685 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2686
2687 list_del(&capsnap->ci_item);
2688 ceph_put_cap_snap(capsnap);
2689 return 1;
2690 }
2691 return 0;
2692}
2693
2694/*
2695 * Release cap refs.
2696 *
2697 * If we released the last ref on any given cap, call ceph_check_caps
2698 * to release (or schedule a release).
2699 *
2700 * If we are releasing a WR cap (from a sync write), finalize any affected
2701 * cap_snap, and wake up any waiters.
2702 */
2703void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2704{
2705 struct inode *inode = &ci->vfs_inode;
2706 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2707
2708 spin_lock(&ci->i_ceph_lock);
2709 if (had & CEPH_CAP_PIN)
2710 --ci->i_pin_ref;
2711 if (had & CEPH_CAP_FILE_RD)
2712 if (--ci->i_rd_ref == 0)
2713 last++;
2714 if (had & CEPH_CAP_FILE_CACHE)
2715 if (--ci->i_rdcache_ref == 0)
2716 last++;
2717 if (had & CEPH_CAP_FILE_BUFFER) {
2718 if (--ci->i_wb_ref == 0) {
2719 last++;
2720 put++;
2721 }
2722 dout("put_cap_refs %p wb %d -> %d (?)\n",
2723 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2724 }
2725 if (had & CEPH_CAP_FILE_WR)
2726 if (--ci->i_wr_ref == 0) {
2727 last++;
2728 if (__ceph_have_pending_cap_snap(ci)) {
2729 struct ceph_cap_snap *capsnap =
2730 list_last_entry(&ci->i_cap_snaps,
2731 struct ceph_cap_snap,
2732 ci_item);
2733 capsnap->writing = 0;
2734 if (ceph_try_drop_cap_snap(ci, capsnap))
2735 put++;
2736 else if (__ceph_finish_cap_snap(ci, capsnap))
2737 flushsnaps = 1;
2738 wake = 1;
2739 }
2740 if (ci->i_wrbuffer_ref_head == 0 &&
2741 ci->i_dirty_caps == 0 &&
2742 ci->i_flushing_caps == 0) {
2743 BUG_ON(!ci->i_head_snapc);
2744 ceph_put_snap_context(ci->i_head_snapc);
2745 ci->i_head_snapc = NULL;
2746 }
2747 /* see comment in __ceph_remove_cap() */
2748 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2749 drop_inode_snap_realm(ci);
2750 }
2751 spin_unlock(&ci->i_ceph_lock);
2752
2753 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2754 last ? " last" : "", put ? " put" : "");
2755
2756 if (last && !flushsnaps)
2757 ceph_check_caps(ci, 0, NULL);
2758 else if (flushsnaps)
2759 ceph_flush_snaps(ci, NULL);
2760 if (wake)
2761 wake_up_all(&ci->i_cap_wq);
2762 while (put-- > 0)
2763 iput(inode);
2764}
2765
2766/*
2767 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2768 * context. Adjust per-snap dirty page accounting as appropriate.
2769 * Once all dirty data for a cap_snap is flushed, flush snapped file
2770 * metadata back to the MDS. If we dropped the last ref, call
2771 * ceph_check_caps.
2772 */
2773void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2774 struct ceph_snap_context *snapc)
2775{
2776 struct inode *inode = &ci->vfs_inode;
2777 struct ceph_cap_snap *capsnap = NULL;
2778 int put = 0;
2779 bool last = false;
2780 bool found = false;
2781 bool flush_snaps = false;
2782 bool complete_capsnap = false;
2783
2784 spin_lock(&ci->i_ceph_lock);
2785 ci->i_wrbuffer_ref -= nr;
2786 if (ci->i_wrbuffer_ref == 0) {
2787 last = true;
2788 put++;
2789 }
2790
2791 if (ci->i_head_snapc == snapc) {
2792 ci->i_wrbuffer_ref_head -= nr;
2793 if (ci->i_wrbuffer_ref_head == 0 &&
2794 ci->i_wr_ref == 0 &&
2795 ci->i_dirty_caps == 0 &&
2796 ci->i_flushing_caps == 0) {
2797 BUG_ON(!ci->i_head_snapc);
2798 ceph_put_snap_context(ci->i_head_snapc);
2799 ci->i_head_snapc = NULL;
2800 }
2801 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2802 inode,
2803 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2804 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2805 last ? " LAST" : "");
2806 } else {
2807 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2808 if (capsnap->context == snapc) {
2809 found = true;
2810 break;
2811 }
2812 }
2813 BUG_ON(!found);
2814 capsnap->dirty_pages -= nr;
2815 if (capsnap->dirty_pages == 0) {
2816 complete_capsnap = true;
2817 if (!capsnap->writing) {
2818 if (ceph_try_drop_cap_snap(ci, capsnap)) {
2819 put++;
2820 } else {
2821 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2822 flush_snaps = true;
2823 }
2824 }
2825 }
2826 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2827 " snap %lld %d/%d -> %d/%d %s%s\n",
2828 inode, capsnap, capsnap->context->seq,
2829 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2830 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2831 last ? " (wrbuffer last)" : "",
2832 complete_capsnap ? " (complete capsnap)" : "");
2833 }
2834
2835 spin_unlock(&ci->i_ceph_lock);
2836
2837 if (last) {
2838 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2839 } else if (flush_snaps) {
2840 ceph_flush_snaps(ci, NULL);
2841 }
2842 if (complete_capsnap)
2843 wake_up_all(&ci->i_cap_wq);
2844 while (put-- > 0)
2845 iput(inode);
2846}
2847
2848/*
2849 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2850 */
2851static void invalidate_aliases(struct inode *inode)
2852{
2853 struct dentry *dn, *prev = NULL;
2854
2855 dout("invalidate_aliases inode %p\n", inode);
2856 d_prune_aliases(inode);
2857 /*
2858 * For non-directory inode, d_find_alias() only returns
2859 * hashed dentry. After calling d_invalidate(), the
2860 * dentry becomes unhashed.
2861 *
2862 * For directory inode, d_find_alias() can return
2863 * unhashed dentry. But directory inode should have
2864 * one alias at most.
2865 */
2866 while ((dn = d_find_alias(inode))) {
2867 if (dn == prev) {
2868 dput(dn);
2869 break;
2870 }
2871 d_invalidate(dn);
2872 if (prev)
2873 dput(prev);
2874 prev = dn;
2875 }
2876 if (prev)
2877 dput(prev);
2878}
2879
2880/*
2881 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2882 * actually be a revocation if it specifies a smaller cap set.)
2883 *
2884 * caller holds s_mutex and i_ceph_lock, we drop both.
2885 */
2886static void handle_cap_grant(struct ceph_mds_client *mdsc,
2887 struct inode *inode, struct ceph_mds_caps *grant,
2888 struct ceph_string **pns, u64 inline_version,
2889 void *inline_data, u32 inline_len,
2890 struct ceph_buffer *xattr_buf,
2891 struct ceph_mds_session *session,
2892 struct ceph_cap *cap, int issued)
2893 __releases(ci->i_ceph_lock)
2894 __releases(mdsc->snap_rwsem)
2895{
2896 struct ceph_inode_info *ci = ceph_inode(inode);
2897 int mds = session->s_mds;
2898 int seq = le32_to_cpu(grant->seq);
2899 int newcaps = le32_to_cpu(grant->caps);
2900 int used, wanted, dirty;
2901 u64 size = le64_to_cpu(grant->size);
2902 u64 max_size = le64_to_cpu(grant->max_size);
2903 struct timespec mtime, atime, ctime;
2904 int check_caps = 0;
2905 bool wake = false;
2906 bool writeback = false;
2907 bool queue_trunc = false;
2908 bool queue_invalidate = false;
2909 bool deleted_inode = false;
2910 bool fill_inline = false;
2911
2912 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2913 inode, cap, mds, seq, ceph_cap_string(newcaps));
2914 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2915 inode->i_size);
2916
2917
2918 /*
2919 * auth mds of the inode changed. we received the cap export message,
2920 * but still haven't received the cap import message. handle_cap_export
2921 * updated the new auth MDS' cap.
2922 *
2923 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2924 * that was sent before the cap import message. So don't remove caps.
2925 */
2926 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2927 WARN_ON(cap != ci->i_auth_cap);
2928 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2929 seq = cap->seq;
2930 newcaps |= cap->issued;
2931 }
2932
2933 /*
2934 * If CACHE is being revoked, and we have no dirty buffers,
2935 * try to invalidate (once). (If there are dirty buffers, we
2936 * will invalidate _after_ writeback.)
2937 */
2938 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2939 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2940 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2941 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
2942 if (try_nonblocking_invalidate(inode)) {
2943 /* there were locked pages.. invalidate later
2944 in a separate thread. */
2945 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2946 queue_invalidate = true;
2947 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2948 }
2949 }
2950 }
2951
2952 /* side effects now are allowed */
2953 cap->cap_gen = session->s_cap_gen;
2954 cap->seq = seq;
2955
2956 __check_cap_issue(ci, cap, newcaps);
2957
2958 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2959 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2960 inode->i_mode = le32_to_cpu(grant->mode);
2961 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2962 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2963 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2964 from_kuid(&init_user_ns, inode->i_uid),
2965 from_kgid(&init_user_ns, inode->i_gid));
2966 }
2967
2968 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2969 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2970 set_nlink(inode, le32_to_cpu(grant->nlink));
2971 if (inode->i_nlink == 0 &&
2972 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2973 deleted_inode = true;
2974 }
2975
2976 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2977 int len = le32_to_cpu(grant->xattr_len);
2978 u64 version = le64_to_cpu(grant->xattr_version);
2979
2980 if (version > ci->i_xattrs.version) {
2981 dout(" got new xattrs v%llu on %p len %d\n",
2982 version, inode, len);
2983 if (ci->i_xattrs.blob)
2984 ceph_buffer_put(ci->i_xattrs.blob);
2985 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2986 ci->i_xattrs.version = version;
2987 ceph_forget_all_cached_acls(inode);
2988 }
2989 }
2990
2991 if (newcaps & CEPH_CAP_ANY_RD) {
2992 /* ctime/mtime/atime? */
2993 ceph_decode_timespec(&mtime, &grant->mtime);
2994 ceph_decode_timespec(&atime, &grant->atime);
2995 ceph_decode_timespec(&ctime, &grant->ctime);
2996 ceph_fill_file_time(inode, issued,
2997 le32_to_cpu(grant->time_warp_seq),
2998 &ctime, &mtime, &atime);
2999 }
3000
3001 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3002 /* file layout may have changed */
3003 s64 old_pool = ci->i_layout.pool_id;
3004 struct ceph_string *old_ns;
3005
3006 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3007 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3008 lockdep_is_held(&ci->i_ceph_lock));
3009 rcu_assign_pointer(ci->i_layout.pool_ns, *pns);
3010
3011 if (ci->i_layout.pool_id != old_pool || *pns != old_ns)
3012 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3013
3014 *pns = old_ns;
3015
3016 /* size/truncate_seq? */
3017 queue_trunc = ceph_fill_file_size(inode, issued,
3018 le32_to_cpu(grant->truncate_seq),
3019 le64_to_cpu(grant->truncate_size),
3020 size);
3021 /* max size increase? */
3022 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
3023 dout("max_size %lld -> %llu\n",
3024 ci->i_max_size, max_size);
3025 ci->i_max_size = max_size;
3026 if (max_size >= ci->i_wanted_max_size) {
3027 ci->i_wanted_max_size = 0; /* reset */
3028 ci->i_requested_max_size = 0;
3029 }
3030 wake = true;
3031 }
3032 }
3033
3034 /* check cap bits */
3035 wanted = __ceph_caps_wanted(ci);
3036 used = __ceph_caps_used(ci);
3037 dirty = __ceph_caps_dirty(ci);
3038 dout(" my wanted = %s, used = %s, dirty %s\n",
3039 ceph_cap_string(wanted),
3040 ceph_cap_string(used),
3041 ceph_cap_string(dirty));
3042 if (wanted != le32_to_cpu(grant->wanted)) {
3043 dout("mds wanted %s -> %s\n",
3044 ceph_cap_string(le32_to_cpu(grant->wanted)),
3045 ceph_cap_string(wanted));
3046 /* imported cap may not have correct mds_wanted */
3047 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
3048 check_caps = 1;
3049 }
3050
3051 /* revocation, grant, or no-op? */
3052 if (cap->issued & ~newcaps) {
3053 int revoking = cap->issued & ~newcaps;
3054
3055 dout("revocation: %s -> %s (revoking %s)\n",
3056 ceph_cap_string(cap->issued),
3057 ceph_cap_string(newcaps),
3058 ceph_cap_string(revoking));
3059 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3060 writeback = true; /* initiate writeback; will delay ack */
3061 else if (revoking == CEPH_CAP_FILE_CACHE &&
3062 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3063 queue_invalidate)
3064 ; /* do nothing yet, invalidation will be queued */
3065 else if (cap == ci->i_auth_cap)
3066 check_caps = 1; /* check auth cap only */
3067 else
3068 check_caps = 2; /* check all caps */
3069 cap->issued = newcaps;
3070 cap->implemented |= newcaps;
3071 } else if (cap->issued == newcaps) {
3072 dout("caps unchanged: %s -> %s\n",
3073 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3074 } else {
3075 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3076 ceph_cap_string(newcaps));
3077 /* non-auth MDS is revoking the newly grant caps ? */
3078 if (cap == ci->i_auth_cap &&
3079 __ceph_caps_revoking_other(ci, cap, newcaps))
3080 check_caps = 2;
3081
3082 cap->issued = newcaps;
3083 cap->implemented |= newcaps; /* add bits only, to
3084 * avoid stepping on a
3085 * pending revocation */
3086 wake = true;
3087 }
3088 BUG_ON(cap->issued & ~cap->implemented);
3089
3090 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
3091 ci->i_inline_version = inline_version;
3092 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3093 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3094 fill_inline = true;
3095 }
3096
3097 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3098 if (newcaps & ~issued)
3099 wake = true;
3100 kick_flushing_inode_caps(mdsc, session, inode);
3101 up_read(&mdsc->snap_rwsem);
3102 } else {
3103 spin_unlock(&ci->i_ceph_lock);
3104 }
3105
3106 if (fill_inline)
3107 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
3108
3109 if (queue_trunc)
3110 ceph_queue_vmtruncate(inode);
3111
3112 if (writeback)
3113 /*
3114 * queue inode for writeback: we can't actually call
3115 * filemap_write_and_wait, etc. from message handler
3116 * context.
3117 */
3118 ceph_queue_writeback(inode);
3119 if (queue_invalidate)
3120 ceph_queue_invalidate(inode);
3121 if (deleted_inode)
3122 invalidate_aliases(inode);
3123 if (wake)
3124 wake_up_all(&ci->i_cap_wq);
3125
3126 if (check_caps == 1)
3127 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3128 session);
3129 else if (check_caps == 2)
3130 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3131 else
3132 mutex_unlock(&session->s_mutex);
3133}
3134
3135/*
3136 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3137 * MDS has been safely committed.
3138 */
3139static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3140 struct ceph_mds_caps *m,
3141 struct ceph_mds_session *session,
3142 struct ceph_cap *cap)
3143 __releases(ci->i_ceph_lock)
3144{
3145 struct ceph_inode_info *ci = ceph_inode(inode);
3146 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3147 struct ceph_cap_flush *cf, *tmp_cf;
3148 LIST_HEAD(to_remove);
3149 unsigned seq = le32_to_cpu(m->seq);
3150 int dirty = le32_to_cpu(m->dirty);
3151 int cleaned = 0;
3152 bool drop = false;
3153 bool wake_ci = 0;
3154 bool wake_mdsc = 0;
3155
3156 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3157 if (cf->tid == flush_tid)
3158 cleaned = cf->caps;
3159 if (cf->caps == 0) /* capsnap */
3160 continue;
3161 if (cf->tid <= flush_tid) {
3162 if (__finish_cap_flush(NULL, ci, cf))
3163 wake_ci = true;
3164 list_add_tail(&cf->i_list, &to_remove);
3165 } else {
3166 cleaned &= ~cf->caps;
3167 if (!cleaned)
3168 break;
3169 }
3170 }
3171
3172 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3173 " flushing %s -> %s\n",
3174 inode, session->s_mds, seq, ceph_cap_string(dirty),
3175 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3176 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3177
3178 if (list_empty(&to_remove) && !cleaned)
3179 goto out;
3180
3181 ci->i_flushing_caps &= ~cleaned;
3182
3183 spin_lock(&mdsc->cap_dirty_lock);
3184
3185 list_for_each_entry(cf, &to_remove, i_list) {
3186 if (__finish_cap_flush(mdsc, NULL, cf))
3187 wake_mdsc = true;
3188 }
3189
3190 if (ci->i_flushing_caps == 0) {
3191 if (list_empty(&ci->i_cap_flush_list)) {
3192 list_del_init(&ci->i_flushing_item);
3193 if (!list_empty(&session->s_cap_flushing)) {
3194 dout(" mds%d still flushing cap on %p\n",
3195 session->s_mds,
3196 &list_first_entry(&session->s_cap_flushing,
3197 struct ceph_inode_info,
3198 i_flushing_item)->vfs_inode);
3199 }
3200 }
3201 mdsc->num_cap_flushing--;
3202 dout(" inode %p now !flushing\n", inode);
3203
3204 if (ci->i_dirty_caps == 0) {
3205 dout(" inode %p now clean\n", inode);
3206 BUG_ON(!list_empty(&ci->i_dirty_item));
3207 drop = true;
3208 if (ci->i_wr_ref == 0 &&
3209 ci->i_wrbuffer_ref_head == 0) {
3210 BUG_ON(!ci->i_head_snapc);
3211 ceph_put_snap_context(ci->i_head_snapc);
3212 ci->i_head_snapc = NULL;
3213 }
3214 } else {
3215 BUG_ON(list_empty(&ci->i_dirty_item));
3216 }
3217 }
3218 spin_unlock(&mdsc->cap_dirty_lock);
3219
3220out:
3221 spin_unlock(&ci->i_ceph_lock);
3222
3223 while (!list_empty(&to_remove)) {
3224 cf = list_first_entry(&to_remove,
3225 struct ceph_cap_flush, i_list);
3226 list_del(&cf->i_list);
3227 ceph_free_cap_flush(cf);
3228 }
3229
3230 if (wake_ci)
3231 wake_up_all(&ci->i_cap_wq);
3232 if (wake_mdsc)
3233 wake_up_all(&mdsc->cap_flushing_wq);
3234 if (drop)
3235 iput(inode);
3236}
3237
3238/*
3239 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3240 * throw away our cap_snap.
3241 *
3242 * Caller hold s_mutex.
3243 */
3244static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3245 struct ceph_mds_caps *m,
3246 struct ceph_mds_session *session)
3247{
3248 struct ceph_inode_info *ci = ceph_inode(inode);
3249 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3250 u64 follows = le64_to_cpu(m->snap_follows);
3251 struct ceph_cap_snap *capsnap;
3252 bool flushed = false;
3253 bool wake_ci = false;
3254 bool wake_mdsc = false;
3255
3256 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3257 inode, ci, session->s_mds, follows);
3258
3259 spin_lock(&ci->i_ceph_lock);
3260 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3261 if (capsnap->follows == follows) {
3262 if (capsnap->cap_flush.tid != flush_tid) {
3263 dout(" cap_snap %p follows %lld tid %lld !="
3264 " %lld\n", capsnap, follows,
3265 flush_tid, capsnap->cap_flush.tid);
3266 break;
3267 }
3268 flushed = true;
3269 break;
3270 } else {
3271 dout(" skipping cap_snap %p follows %lld\n",
3272 capsnap, capsnap->follows);
3273 }
3274 }
3275 if (flushed) {
3276 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3277 dout(" removing %p cap_snap %p follows %lld\n",
3278 inode, capsnap, follows);
3279 list_del(&capsnap->ci_item);
3280 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3281 wake_ci = true;
3282
3283 spin_lock(&mdsc->cap_dirty_lock);
3284
3285 if (list_empty(&ci->i_cap_flush_list))
3286 list_del_init(&ci->i_flushing_item);
3287
3288 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3289 wake_mdsc = true;
3290
3291 spin_unlock(&mdsc->cap_dirty_lock);
3292 }
3293 spin_unlock(&ci->i_ceph_lock);
3294 if (flushed) {
3295 ceph_put_snap_context(capsnap->context);
3296 ceph_put_cap_snap(capsnap);
3297 if (wake_ci)
3298 wake_up_all(&ci->i_cap_wq);
3299 if (wake_mdsc)
3300 wake_up_all(&mdsc->cap_flushing_wq);
3301 iput(inode);
3302 }
3303}
3304
3305/*
3306 * Handle TRUNC from MDS, indicating file truncation.
3307 *
3308 * caller hold s_mutex.
3309 */
3310static void handle_cap_trunc(struct inode *inode,
3311 struct ceph_mds_caps *trunc,
3312 struct ceph_mds_session *session)
3313 __releases(ci->i_ceph_lock)
3314{
3315 struct ceph_inode_info *ci = ceph_inode(inode);
3316 int mds = session->s_mds;
3317 int seq = le32_to_cpu(trunc->seq);
3318 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3319 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3320 u64 size = le64_to_cpu(trunc->size);
3321 int implemented = 0;
3322 int dirty = __ceph_caps_dirty(ci);
3323 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3324 int queue_trunc = 0;
3325
3326 issued |= implemented | dirty;
3327
3328 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3329 inode, mds, seq, truncate_size, truncate_seq);
3330 queue_trunc = ceph_fill_file_size(inode, issued,
3331 truncate_seq, truncate_size, size);
3332 spin_unlock(&ci->i_ceph_lock);
3333
3334 if (queue_trunc)
3335 ceph_queue_vmtruncate(inode);
3336}
3337
3338/*
3339 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3340 * different one. If we are the most recent migration we've seen (as
3341 * indicated by mseq), make note of the migrating cap bits for the
3342 * duration (until we see the corresponding IMPORT).
3343 *
3344 * caller holds s_mutex
3345 */
3346static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3347 struct ceph_mds_cap_peer *ph,
3348 struct ceph_mds_session *session)
3349{
3350 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3351 struct ceph_mds_session *tsession = NULL;
3352 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3353 struct ceph_inode_info *ci = ceph_inode(inode);
3354 u64 t_cap_id;
3355 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3356 unsigned t_seq, t_mseq;
3357 int target, issued;
3358 int mds = session->s_mds;
3359
3360 if (ph) {
3361 t_cap_id = le64_to_cpu(ph->cap_id);
3362 t_seq = le32_to_cpu(ph->seq);
3363 t_mseq = le32_to_cpu(ph->mseq);
3364 target = le32_to_cpu(ph->mds);
3365 } else {
3366 t_cap_id = t_seq = t_mseq = 0;
3367 target = -1;
3368 }
3369
3370 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3371 inode, ci, mds, mseq, target);
3372retry:
3373 spin_lock(&ci->i_ceph_lock);
3374 cap = __get_cap_for_mds(ci, mds);
3375 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3376 goto out_unlock;
3377
3378 if (target < 0) {
3379 __ceph_remove_cap(cap, false);
3380 if (!ci->i_auth_cap)
3381 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3382 goto out_unlock;
3383 }
3384
3385 /*
3386 * now we know we haven't received the cap import message yet
3387 * because the exported cap still exist.
3388 */
3389
3390 issued = cap->issued;
3391 WARN_ON(issued != cap->implemented);
3392
3393 tcap = __get_cap_for_mds(ci, target);
3394 if (tcap) {
3395 /* already have caps from the target */
3396 if (tcap->cap_id != t_cap_id ||
3397 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3398 dout(" updating import cap %p mds%d\n", tcap, target);
3399 tcap->cap_id = t_cap_id;
3400 tcap->seq = t_seq - 1;
3401 tcap->issue_seq = t_seq - 1;
3402 tcap->mseq = t_mseq;
3403 tcap->issued |= issued;
3404 tcap->implemented |= issued;
3405 if (cap == ci->i_auth_cap)
3406 ci->i_auth_cap = tcap;
3407 if (!list_empty(&ci->i_cap_flush_list) &&
3408 ci->i_auth_cap == tcap) {
3409 spin_lock(&mdsc->cap_dirty_lock);
3410 list_move_tail(&ci->i_flushing_item,
3411 &tcap->session->s_cap_flushing);
3412 spin_unlock(&mdsc->cap_dirty_lock);
3413 }
3414 }
3415 __ceph_remove_cap(cap, false);
3416 goto out_unlock;
3417 } else if (tsession) {
3418 /* add placeholder for the export tagert */
3419 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3420 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3421 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3422
3423 __ceph_remove_cap(cap, false);
3424 goto out_unlock;
3425 }
3426
3427 spin_unlock(&ci->i_ceph_lock);
3428 mutex_unlock(&session->s_mutex);
3429
3430 /* open target session */
3431 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3432 if (!IS_ERR(tsession)) {
3433 if (mds > target) {
3434 mutex_lock(&session->s_mutex);
3435 mutex_lock_nested(&tsession->s_mutex,
3436 SINGLE_DEPTH_NESTING);
3437 } else {
3438 mutex_lock(&tsession->s_mutex);
3439 mutex_lock_nested(&session->s_mutex,
3440 SINGLE_DEPTH_NESTING);
3441 }
3442 new_cap = ceph_get_cap(mdsc, NULL);
3443 } else {
3444 WARN_ON(1);
3445 tsession = NULL;
3446 target = -1;
3447 }
3448 goto retry;
3449
3450out_unlock:
3451 spin_unlock(&ci->i_ceph_lock);
3452 mutex_unlock(&session->s_mutex);
3453 if (tsession) {
3454 mutex_unlock(&tsession->s_mutex);
3455 ceph_put_mds_session(tsession);
3456 }
3457 if (new_cap)
3458 ceph_put_cap(mdsc, new_cap);
3459}
3460
3461/*
3462 * Handle cap IMPORT.
3463 *
3464 * caller holds s_mutex. acquires i_ceph_lock
3465 */
3466static void handle_cap_import(struct ceph_mds_client *mdsc,
3467 struct inode *inode, struct ceph_mds_caps *im,
3468 struct ceph_mds_cap_peer *ph,
3469 struct ceph_mds_session *session,
3470 struct ceph_cap **target_cap, int *old_issued)
3471 __acquires(ci->i_ceph_lock)
3472{
3473 struct ceph_inode_info *ci = ceph_inode(inode);
3474 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3475 int mds = session->s_mds;
3476 int issued;
3477 unsigned caps = le32_to_cpu(im->caps);
3478 unsigned wanted = le32_to_cpu(im->wanted);
3479 unsigned seq = le32_to_cpu(im->seq);
3480 unsigned mseq = le32_to_cpu(im->migrate_seq);
3481 u64 realmino = le64_to_cpu(im->realm);
3482 u64 cap_id = le64_to_cpu(im->cap_id);
3483 u64 p_cap_id;
3484 int peer;
3485
3486 if (ph) {
3487 p_cap_id = le64_to_cpu(ph->cap_id);
3488 peer = le32_to_cpu(ph->mds);
3489 } else {
3490 p_cap_id = 0;
3491 peer = -1;
3492 }
3493
3494 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3495 inode, ci, mds, mseq, peer);
3496
3497retry:
3498 spin_lock(&ci->i_ceph_lock);
3499 cap = __get_cap_for_mds(ci, mds);
3500 if (!cap) {
3501 if (!new_cap) {
3502 spin_unlock(&ci->i_ceph_lock);
3503 new_cap = ceph_get_cap(mdsc, NULL);
3504 goto retry;
3505 }
3506 cap = new_cap;
3507 } else {
3508 if (new_cap) {
3509 ceph_put_cap(mdsc, new_cap);
3510 new_cap = NULL;
3511 }
3512 }
3513
3514 __ceph_caps_issued(ci, &issued);
3515 issued |= __ceph_caps_dirty(ci);
3516
3517 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3518 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3519
3520 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3521 if (ocap && ocap->cap_id == p_cap_id) {
3522 dout(" remove export cap %p mds%d flags %d\n",
3523 ocap, peer, ph->flags);
3524 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3525 (ocap->seq != le32_to_cpu(ph->seq) ||
3526 ocap->mseq != le32_to_cpu(ph->mseq))) {
3527 pr_err("handle_cap_import: mismatched seq/mseq: "
3528 "ino (%llx.%llx) mds%d seq %d mseq %d "
3529 "importer mds%d has peer seq %d mseq %d\n",
3530 ceph_vinop(inode), peer, ocap->seq,
3531 ocap->mseq, mds, le32_to_cpu(ph->seq),
3532 le32_to_cpu(ph->mseq));
3533 }
3534 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3535 }
3536
3537 /* make sure we re-request max_size, if necessary */
3538 ci->i_wanted_max_size = 0;
3539 ci->i_requested_max_size = 0;
3540
3541 *old_issued = issued;
3542 *target_cap = cap;
3543}
3544
3545/*
3546 * Handle a caps message from the MDS.
3547 *
3548 * Identify the appropriate session, inode, and call the right handler
3549 * based on the cap op.
3550 */
3551void ceph_handle_caps(struct ceph_mds_session *session,
3552 struct ceph_msg *msg)
3553{
3554 struct ceph_mds_client *mdsc = session->s_mdsc;
3555 struct super_block *sb = mdsc->fsc->sb;
3556 struct inode *inode;
3557 struct ceph_inode_info *ci;
3558 struct ceph_cap *cap;
3559 struct ceph_mds_caps *h;
3560 struct ceph_mds_cap_peer *peer = NULL;
3561 struct ceph_snap_realm *realm = NULL;
3562 struct ceph_string *pool_ns = NULL;
3563 int mds = session->s_mds;
3564 int op, issued;
3565 u32 seq, mseq;
3566 struct ceph_vino vino;
3567 u64 tid;
3568 u64 inline_version = 0;
3569 void *inline_data = NULL;
3570 u32 inline_len = 0;
3571 void *snaptrace;
3572 size_t snaptrace_len;
3573 void *p, *end;
3574
3575 dout("handle_caps from mds%d\n", mds);
3576
3577 /* decode */
3578 end = msg->front.iov_base + msg->front.iov_len;
3579 tid = le64_to_cpu(msg->hdr.tid);
3580 if (msg->front.iov_len < sizeof(*h))
3581 goto bad;
3582 h = msg->front.iov_base;
3583 op = le32_to_cpu(h->op);
3584 vino.ino = le64_to_cpu(h->ino);
3585 vino.snap = CEPH_NOSNAP;
3586 seq = le32_to_cpu(h->seq);
3587 mseq = le32_to_cpu(h->migrate_seq);
3588
3589 snaptrace = h + 1;
3590 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3591 p = snaptrace + snaptrace_len;
3592
3593 if (le16_to_cpu(msg->hdr.version) >= 2) {
3594 u32 flock_len;
3595 ceph_decode_32_safe(&p, end, flock_len, bad);
3596 if (p + flock_len > end)
3597 goto bad;
3598 p += flock_len;
3599 }
3600
3601 if (le16_to_cpu(msg->hdr.version) >= 3) {
3602 if (op == CEPH_CAP_OP_IMPORT) {
3603 if (p + sizeof(*peer) > end)
3604 goto bad;
3605 peer = p;
3606 p += sizeof(*peer);
3607 } else if (op == CEPH_CAP_OP_EXPORT) {
3608 /* recorded in unused fields */
3609 peer = (void *)&h->size;
3610 }
3611 }
3612
3613 if (le16_to_cpu(msg->hdr.version) >= 4) {
3614 ceph_decode_64_safe(&p, end, inline_version, bad);
3615 ceph_decode_32_safe(&p, end, inline_len, bad);
3616 if (p + inline_len > end)
3617 goto bad;
3618 inline_data = p;
3619 p += inline_len;
3620 }
3621
3622 if (le16_to_cpu(msg->hdr.version) >= 8) {
3623 u64 flush_tid;
3624 u32 caller_uid, caller_gid;
3625 u32 osd_epoch_barrier;
3626 u32 pool_ns_len;
3627 /* version >= 5 */
3628 ceph_decode_32_safe(&p, end, osd_epoch_barrier, bad);
3629 /* version >= 6 */
3630 ceph_decode_64_safe(&p, end, flush_tid, bad);
3631 /* version >= 7 */
3632 ceph_decode_32_safe(&p, end, caller_uid, bad);
3633 ceph_decode_32_safe(&p, end, caller_gid, bad);
3634 /* version >= 8 */
3635 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3636 if (pool_ns_len > 0) {
3637 ceph_decode_need(&p, end, pool_ns_len, bad);
3638 pool_ns = ceph_find_or_create_string(p, pool_ns_len);
3639 p += pool_ns_len;
3640 }
3641 }
3642
3643 /* lookup ino */
3644 inode = ceph_find_inode(sb, vino);
3645 ci = ceph_inode(inode);
3646 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3647 vino.snap, inode);
3648
3649 mutex_lock(&session->s_mutex);
3650 session->s_seq++;
3651 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3652 (unsigned)seq);
3653
3654 if (!inode) {
3655 dout(" i don't have ino %llx\n", vino.ino);
3656
3657 if (op == CEPH_CAP_OP_IMPORT) {
3658 cap = ceph_get_cap(mdsc, NULL);
3659 cap->cap_ino = vino.ino;
3660 cap->queue_release = 1;
3661 cap->cap_id = le64_to_cpu(h->cap_id);
3662 cap->mseq = mseq;
3663 cap->seq = seq;
3664 cap->issue_seq = seq;
3665 spin_lock(&session->s_cap_lock);
3666 list_add_tail(&cap->session_caps,
3667 &session->s_cap_releases);
3668 session->s_num_cap_releases++;
3669 spin_unlock(&session->s_cap_lock);
3670 }
3671 goto flush_cap_releases;
3672 }
3673
3674 /* these will work even if we don't have a cap yet */
3675 switch (op) {
3676 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3677 handle_cap_flushsnap_ack(inode, tid, h, session);
3678 goto done;
3679
3680 case CEPH_CAP_OP_EXPORT:
3681 handle_cap_export(inode, h, peer, session);
3682 goto done_unlocked;
3683
3684 case CEPH_CAP_OP_IMPORT:
3685 realm = NULL;
3686 if (snaptrace_len) {
3687 down_write(&mdsc->snap_rwsem);
3688 ceph_update_snap_trace(mdsc, snaptrace,
3689 snaptrace + snaptrace_len,
3690 false, &realm);
3691 downgrade_write(&mdsc->snap_rwsem);
3692 } else {
3693 down_read(&mdsc->snap_rwsem);
3694 }
3695 handle_cap_import(mdsc, inode, h, peer, session,
3696 &cap, &issued);
3697 handle_cap_grant(mdsc, inode, h, &pool_ns,
3698 inline_version, inline_data, inline_len,
3699 msg->middle, session, cap, issued);
3700 if (realm)
3701 ceph_put_snap_realm(mdsc, realm);
3702 goto done_unlocked;
3703 }
3704
3705 /* the rest require a cap */
3706 spin_lock(&ci->i_ceph_lock);
3707 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3708 if (!cap) {
3709 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3710 inode, ceph_ino(inode), ceph_snap(inode), mds);
3711 spin_unlock(&ci->i_ceph_lock);
3712 goto flush_cap_releases;
3713 }
3714
3715 /* note that each of these drops i_ceph_lock for us */
3716 switch (op) {
3717 case CEPH_CAP_OP_REVOKE:
3718 case CEPH_CAP_OP_GRANT:
3719 __ceph_caps_issued(ci, &issued);
3720 issued |= __ceph_caps_dirty(ci);
3721 handle_cap_grant(mdsc, inode, h, &pool_ns,
3722 inline_version, inline_data, inline_len,
3723 msg->middle, session, cap, issued);
3724 goto done_unlocked;
3725
3726 case CEPH_CAP_OP_FLUSH_ACK:
3727 handle_cap_flush_ack(inode, tid, h, session, cap);
3728 break;
3729
3730 case CEPH_CAP_OP_TRUNC:
3731 handle_cap_trunc(inode, h, session);
3732 break;
3733
3734 default:
3735 spin_unlock(&ci->i_ceph_lock);
3736 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3737 ceph_cap_op_name(op));
3738 }
3739
3740 goto done;
3741
3742flush_cap_releases:
3743 /*
3744 * send any cap release message to try to move things
3745 * along for the mds (who clearly thinks we still have this
3746 * cap).
3747 */
3748 ceph_send_cap_releases(mdsc, session);
3749
3750done:
3751 mutex_unlock(&session->s_mutex);
3752done_unlocked:
3753 iput(inode);
3754 ceph_put_string(pool_ns);
3755 return;
3756
3757bad:
3758 pr_err("ceph_handle_caps: corrupt message\n");
3759 ceph_msg_dump(msg);
3760 return;
3761}
3762
3763/*
3764 * Delayed work handler to process end of delayed cap release LRU list.
3765 */
3766void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3767{
3768 struct ceph_inode_info *ci;
3769 int flags = CHECK_CAPS_NODELAY;
3770
3771 dout("check_delayed_caps\n");
3772 while (1) {
3773 spin_lock(&mdsc->cap_delay_lock);
3774 if (list_empty(&mdsc->cap_delay_list))
3775 break;
3776 ci = list_first_entry(&mdsc->cap_delay_list,
3777 struct ceph_inode_info,
3778 i_cap_delay_list);
3779 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3780 time_before(jiffies, ci->i_hold_caps_max))
3781 break;
3782 list_del_init(&ci->i_cap_delay_list);
3783 spin_unlock(&mdsc->cap_delay_lock);
3784 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3785 ceph_check_caps(ci, flags, NULL);
3786 }
3787 spin_unlock(&mdsc->cap_delay_lock);
3788}
3789
3790/*
3791 * Flush all dirty caps to the mds
3792 */
3793void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3794{
3795 struct ceph_inode_info *ci;
3796 struct inode *inode;
3797
3798 dout("flush_dirty_caps\n");
3799 spin_lock(&mdsc->cap_dirty_lock);
3800 while (!list_empty(&mdsc->cap_dirty)) {
3801 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3802 i_dirty_item);
3803 inode = &ci->vfs_inode;
3804 ihold(inode);
3805 dout("flush_dirty_caps %p\n", inode);
3806 spin_unlock(&mdsc->cap_dirty_lock);
3807 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3808 iput(inode);
3809 spin_lock(&mdsc->cap_dirty_lock);
3810 }
3811 spin_unlock(&mdsc->cap_dirty_lock);
3812 dout("flush_dirty_caps done\n");
3813}
3814
3815void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
3816{
3817 int i;
3818 int bits = (fmode << 1) | 1;
3819 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3820 if (bits & (1 << i))
3821 ci->i_nr_by_mode[i]++;
3822 }
3823}
3824
3825/*
3826 * Drop open file reference. If we were the last open file,
3827 * we may need to release capabilities to the MDS (or schedule
3828 * their delayed release).
3829 */
3830void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3831{
3832 int i, last = 0;
3833 int bits = (fmode << 1) | 1;
3834 spin_lock(&ci->i_ceph_lock);
3835 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3836 if (bits & (1 << i)) {
3837 BUG_ON(ci->i_nr_by_mode[i] == 0);
3838 if (--ci->i_nr_by_mode[i] == 0)
3839 last++;
3840 }
3841 }
3842 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3843 &ci->vfs_inode, fmode,
3844 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
3845 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
3846 spin_unlock(&ci->i_ceph_lock);
3847
3848 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3849 ceph_check_caps(ci, 0, NULL);
3850}
3851
3852/*
3853 * Helpers for embedding cap and dentry lease releases into mds
3854 * requests.
3855 *
3856 * @force is used by dentry_release (below) to force inclusion of a
3857 * record for the directory inode, even when there aren't any caps to
3858 * drop.
3859 */
3860int ceph_encode_inode_release(void **p, struct inode *inode,
3861 int mds, int drop, int unless, int force)
3862{
3863 struct ceph_inode_info *ci = ceph_inode(inode);
3864 struct ceph_cap *cap;
3865 struct ceph_mds_request_release *rel = *p;
3866 int used, dirty;
3867 int ret = 0;
3868
3869 spin_lock(&ci->i_ceph_lock);
3870 used = __ceph_caps_used(ci);
3871 dirty = __ceph_caps_dirty(ci);
3872
3873 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3874 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3875 ceph_cap_string(unless));
3876
3877 /* only drop unused, clean caps */
3878 drop &= ~(used | dirty);
3879
3880 cap = __get_cap_for_mds(ci, mds);
3881 if (cap && __cap_is_valid(cap)) {
3882 if (force ||
3883 ((cap->issued & drop) &&
3884 (cap->issued & unless) == 0)) {
3885 if ((cap->issued & drop) &&
3886 (cap->issued & unless) == 0) {
3887 int wanted = __ceph_caps_wanted(ci);
3888 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3889 wanted |= cap->mds_wanted;
3890 dout("encode_inode_release %p cap %p "
3891 "%s -> %s, wanted %s -> %s\n", inode, cap,
3892 ceph_cap_string(cap->issued),
3893 ceph_cap_string(cap->issued & ~drop),
3894 ceph_cap_string(cap->mds_wanted),
3895 ceph_cap_string(wanted));
3896
3897 cap->issued &= ~drop;
3898 cap->implemented &= ~drop;
3899 cap->mds_wanted = wanted;
3900 } else {
3901 dout("encode_inode_release %p cap %p %s"
3902 " (force)\n", inode, cap,
3903 ceph_cap_string(cap->issued));
3904 }
3905
3906 rel->ino = cpu_to_le64(ceph_ino(inode));
3907 rel->cap_id = cpu_to_le64(cap->cap_id);
3908 rel->seq = cpu_to_le32(cap->seq);
3909 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3910 rel->mseq = cpu_to_le32(cap->mseq);
3911 rel->caps = cpu_to_le32(cap->implemented);
3912 rel->wanted = cpu_to_le32(cap->mds_wanted);
3913 rel->dname_len = 0;
3914 rel->dname_seq = 0;
3915 *p += sizeof(*rel);
3916 ret = 1;
3917 } else {
3918 dout("encode_inode_release %p cap %p %s\n",
3919 inode, cap, ceph_cap_string(cap->issued));
3920 }
3921 }
3922 spin_unlock(&ci->i_ceph_lock);
3923 return ret;
3924}
3925
3926int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3927 int mds, int drop, int unless)
3928{
3929 struct inode *dir = d_inode(dentry->d_parent);
3930 struct ceph_mds_request_release *rel = *p;
3931 struct ceph_dentry_info *di = ceph_dentry(dentry);
3932 int force = 0;
3933 int ret;
3934
3935 /*
3936 * force an record for the directory caps if we have a dentry lease.
3937 * this is racy (can't take i_ceph_lock and d_lock together), but it
3938 * doesn't have to be perfect; the mds will revoke anything we don't
3939 * release.
3940 */
3941 spin_lock(&dentry->d_lock);
3942 if (di->lease_session && di->lease_session->s_mds == mds)
3943 force = 1;
3944 spin_unlock(&dentry->d_lock);
3945
3946 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3947
3948 spin_lock(&dentry->d_lock);
3949 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3950 dout("encode_dentry_release %p mds%d seq %d\n",
3951 dentry, mds, (int)di->lease_seq);
3952 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3953 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3954 *p += dentry->d_name.len;
3955 rel->dname_seq = cpu_to_le32(di->lease_seq);
3956 __ceph_mdsc_drop_dentry_lease(dentry);
3957 }
3958 spin_unlock(&dentry->d_lock);
3959 return ret;
3960}