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