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