aboutsummaryrefslogtreecommitdiffstats
path: root/fs/ceph/caps.c
diff options
context:
space:
mode:
Diffstat (limited to 'fs/ceph/caps.c')
-rw-r--r--fs/ceph/caps.c2927
1 files changed, 2927 insertions, 0 deletions
diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
new file mode 100644
index 000000000000..db122bb357b8
--- /dev/null
+++ b/fs/ceph/caps.c
@@ -0,0 +1,2927 @@
1#include "ceph_debug.h"
2
3#include <linux/fs.h>
4#include <linux/kernel.h>
5#include <linux/sched.h>
6#include <linux/vmalloc.h>
7#include <linux/wait.h>
8#include <linux/writeback.h>
9
10#include "super.h"
11#include "decode.h"
12#include "messenger.h"
13
14/*
15 * Capability management
16 *
17 * The Ceph metadata servers control client access to inode metadata
18 * and file data by issuing capabilities, granting clients permission
19 * to read and/or write both inode field and file data to OSDs
20 * (storage nodes). Each capability consists of a set of bits
21 * indicating which operations are allowed.
22 *
23 * If the client holds a *_SHARED cap, the client has a coherent value
24 * that can be safely read from the cached inode.
25 *
26 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
27 * client is allowed to change inode attributes (e.g., file size,
28 * mtime), note its dirty state in the ceph_cap, and asynchronously
29 * flush that metadata change to the MDS.
30 *
31 * In the event of a conflicting operation (perhaps by another
32 * client), the MDS will revoke the conflicting client capabilities.
33 *
34 * In order for a client to cache an inode, it must hold a capability
35 * with at least one MDS server. When inodes are released, release
36 * notifications are batched and periodically sent en masse to the MDS
37 * cluster to release server state.
38 */
39
40
41/*
42 * Generate readable cap strings for debugging output.
43 */
44#define MAX_CAP_STR 20
45static char cap_str[MAX_CAP_STR][40];
46static DEFINE_SPINLOCK(cap_str_lock);
47static int last_cap_str;
48
49static char *gcap_string(char *s, int c)
50{
51 if (c & CEPH_CAP_GSHARED)
52 *s++ = 's';
53 if (c & CEPH_CAP_GEXCL)
54 *s++ = 'x';
55 if (c & CEPH_CAP_GCACHE)
56 *s++ = 'c';
57 if (c & CEPH_CAP_GRD)
58 *s++ = 'r';
59 if (c & CEPH_CAP_GWR)
60 *s++ = 'w';
61 if (c & CEPH_CAP_GBUFFER)
62 *s++ = 'b';
63 if (c & CEPH_CAP_GLAZYIO)
64 *s++ = 'l';
65 return s;
66}
67
68const char *ceph_cap_string(int caps)
69{
70 int i;
71 char *s;
72 int c;
73
74 spin_lock(&cap_str_lock);
75 i = last_cap_str++;
76 if (last_cap_str == MAX_CAP_STR)
77 last_cap_str = 0;
78 spin_unlock(&cap_str_lock);
79
80 s = cap_str[i];
81
82 if (caps & CEPH_CAP_PIN)
83 *s++ = 'p';
84
85 c = (caps >> CEPH_CAP_SAUTH) & 3;
86 if (c) {
87 *s++ = 'A';
88 s = gcap_string(s, c);
89 }
90
91 c = (caps >> CEPH_CAP_SLINK) & 3;
92 if (c) {
93 *s++ = 'L';
94 s = gcap_string(s, c);
95 }
96
97 c = (caps >> CEPH_CAP_SXATTR) & 3;
98 if (c) {
99 *s++ = 'X';
100 s = gcap_string(s, c);
101 }
102
103 c = caps >> CEPH_CAP_SFILE;
104 if (c) {
105 *s++ = 'F';
106 s = gcap_string(s, c);
107 }
108
109 if (s == cap_str[i])
110 *s++ = '-';
111 *s = 0;
112 return cap_str[i];
113}
114
115/*
116 * Cap reservations
117 *
118 * Maintain a global pool of preallocated struct ceph_caps, referenced
119 * by struct ceph_caps_reservations. This ensures that we preallocate
120 * memory needed to successfully process an MDS response. (If an MDS
121 * sends us cap information and we fail to process it, we will have
122 * problems due to the client and MDS being out of sync.)
123 *
124 * Reservations are 'owned' by a ceph_cap_reservation context.
125 */
126static spinlock_t caps_list_lock;
127static struct list_head caps_list; /* unused (reserved or unreserved) */
128static int caps_total_count; /* total caps allocated */
129static int caps_use_count; /* in use */
130static int caps_reserve_count; /* unused, reserved */
131static int caps_avail_count; /* unused, unreserved */
132static int caps_min_count; /* keep at least this many (unreserved) */
133
134void __init ceph_caps_init(void)
135{
136 INIT_LIST_HEAD(&caps_list);
137 spin_lock_init(&caps_list_lock);
138}
139
140void ceph_caps_finalize(void)
141{
142 struct ceph_cap *cap;
143
144 spin_lock(&caps_list_lock);
145 while (!list_empty(&caps_list)) {
146 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item);
147 list_del(&cap->caps_item);
148 kmem_cache_free(ceph_cap_cachep, cap);
149 }
150 caps_total_count = 0;
151 caps_avail_count = 0;
152 caps_use_count = 0;
153 caps_reserve_count = 0;
154 caps_min_count = 0;
155 spin_unlock(&caps_list_lock);
156}
157
158void ceph_adjust_min_caps(int delta)
159{
160 spin_lock(&caps_list_lock);
161 caps_min_count += delta;
162 BUG_ON(caps_min_count < 0);
163 spin_unlock(&caps_list_lock);
164}
165
166int ceph_reserve_caps(struct ceph_cap_reservation *ctx, int need)
167{
168 int i;
169 struct ceph_cap *cap;
170 int have;
171 int alloc = 0;
172 LIST_HEAD(newcaps);
173 int ret = 0;
174
175 dout("reserve caps ctx=%p need=%d\n", ctx, need);
176
177 /* first reserve any caps that are already allocated */
178 spin_lock(&caps_list_lock);
179 if (caps_avail_count >= need)
180 have = need;
181 else
182 have = caps_avail_count;
183 caps_avail_count -= have;
184 caps_reserve_count += have;
185 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
186 caps_avail_count);
187 spin_unlock(&caps_list_lock);
188
189 for (i = have; i < need; i++) {
190 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
191 if (!cap) {
192 ret = -ENOMEM;
193 goto out_alloc_count;
194 }
195 list_add(&cap->caps_item, &newcaps);
196 alloc++;
197 }
198 BUG_ON(have + alloc != need);
199
200 spin_lock(&caps_list_lock);
201 caps_total_count += alloc;
202 caps_reserve_count += alloc;
203 list_splice(&newcaps, &caps_list);
204
205 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
206 caps_avail_count);
207 spin_unlock(&caps_list_lock);
208
209 ctx->count = need;
210 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
211 ctx, caps_total_count, caps_use_count, caps_reserve_count,
212 caps_avail_count);
213 return 0;
214
215out_alloc_count:
216 /* we didn't manage to reserve as much as we needed */
217 pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
218 ctx, need, have);
219 return ret;
220}
221
222int ceph_unreserve_caps(struct ceph_cap_reservation *ctx)
223{
224 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
225 if (ctx->count) {
226 spin_lock(&caps_list_lock);
227 BUG_ON(caps_reserve_count < ctx->count);
228 caps_reserve_count -= ctx->count;
229 caps_avail_count += ctx->count;
230 ctx->count = 0;
231 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
232 caps_total_count, caps_use_count, caps_reserve_count,
233 caps_avail_count);
234 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
235 caps_avail_count);
236 spin_unlock(&caps_list_lock);
237 }
238 return 0;
239}
240
241static struct ceph_cap *get_cap(struct ceph_cap_reservation *ctx)
242{
243 struct ceph_cap *cap = NULL;
244
245 /* temporary, until we do something about cap import/export */
246 if (!ctx)
247 return kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
248
249 spin_lock(&caps_list_lock);
250 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
251 ctx, ctx->count, caps_total_count, caps_use_count,
252 caps_reserve_count, caps_avail_count);
253 BUG_ON(!ctx->count);
254 BUG_ON(ctx->count > caps_reserve_count);
255 BUG_ON(list_empty(&caps_list));
256
257 ctx->count--;
258 caps_reserve_count--;
259 caps_use_count++;
260
261 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item);
262 list_del(&cap->caps_item);
263
264 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
265 caps_avail_count);
266 spin_unlock(&caps_list_lock);
267 return cap;
268}
269
270void ceph_put_cap(struct ceph_cap *cap)
271{
272 spin_lock(&caps_list_lock);
273 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
274 cap, caps_total_count, caps_use_count,
275 caps_reserve_count, caps_avail_count);
276 caps_use_count--;
277 /*
278 * Keep some preallocated caps around (ceph_min_count), to
279 * avoid lots of free/alloc churn.
280 */
281 if (caps_avail_count >= caps_reserve_count + caps_min_count) {
282 caps_total_count--;
283 kmem_cache_free(ceph_cap_cachep, cap);
284 } else {
285 caps_avail_count++;
286 list_add(&cap->caps_item, &caps_list);
287 }
288
289 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
290 caps_avail_count);
291 spin_unlock(&caps_list_lock);
292}
293
294void ceph_reservation_status(struct ceph_client *client,
295 int *total, int *avail, int *used, int *reserved,
296 int *min)
297{
298 if (total)
299 *total = caps_total_count;
300 if (avail)
301 *avail = caps_avail_count;
302 if (used)
303 *used = caps_use_count;
304 if (reserved)
305 *reserved = caps_reserve_count;
306 if (min)
307 *min = caps_min_count;
308}
309
310/*
311 * Find ceph_cap for given mds, if any.
312 *
313 * Called with i_lock held.
314 */
315static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
316{
317 struct ceph_cap *cap;
318 struct rb_node *n = ci->i_caps.rb_node;
319
320 while (n) {
321 cap = rb_entry(n, struct ceph_cap, ci_node);
322 if (mds < cap->mds)
323 n = n->rb_left;
324 else if (mds > cap->mds)
325 n = n->rb_right;
326 else
327 return cap;
328 }
329 return NULL;
330}
331
332/*
333 * Return id of any MDS with a cap, preferably FILE_WR|WRBUFFER|EXCL, else
334 * -1.
335 */
336static int __ceph_get_cap_mds(struct ceph_inode_info *ci, u32 *mseq)
337{
338 struct ceph_cap *cap;
339 int mds = -1;
340 struct rb_node *p;
341
342 /* prefer mds with WR|WRBUFFER|EXCL caps */
343 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
344 cap = rb_entry(p, struct ceph_cap, ci_node);
345 mds = cap->mds;
346 if (mseq)
347 *mseq = cap->mseq;
348 if (cap->issued & (CEPH_CAP_FILE_WR |
349 CEPH_CAP_FILE_BUFFER |
350 CEPH_CAP_FILE_EXCL))
351 break;
352 }
353 return mds;
354}
355
356int ceph_get_cap_mds(struct inode *inode)
357{
358 int mds;
359 spin_lock(&inode->i_lock);
360 mds = __ceph_get_cap_mds(ceph_inode(inode), NULL);
361 spin_unlock(&inode->i_lock);
362 return mds;
363}
364
365/*
366 * Called under i_lock.
367 */
368static void __insert_cap_node(struct ceph_inode_info *ci,
369 struct ceph_cap *new)
370{
371 struct rb_node **p = &ci->i_caps.rb_node;
372 struct rb_node *parent = NULL;
373 struct ceph_cap *cap = NULL;
374
375 while (*p) {
376 parent = *p;
377 cap = rb_entry(parent, struct ceph_cap, ci_node);
378 if (new->mds < cap->mds)
379 p = &(*p)->rb_left;
380 else if (new->mds > cap->mds)
381 p = &(*p)->rb_right;
382 else
383 BUG();
384 }
385
386 rb_link_node(&new->ci_node, parent, p);
387 rb_insert_color(&new->ci_node, &ci->i_caps);
388}
389
390/*
391 * (re)set cap hold timeouts, which control the delayed release
392 * of unused caps back to the MDS. Should be called on cap use.
393 */
394static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
395 struct ceph_inode_info *ci)
396{
397 struct ceph_mount_args *ma = mdsc->client->mount_args;
398
399 ci->i_hold_caps_min = round_jiffies(jiffies +
400 ma->caps_wanted_delay_min * HZ);
401 ci->i_hold_caps_max = round_jiffies(jiffies +
402 ma->caps_wanted_delay_max * HZ);
403 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
404 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
405}
406
407/*
408 * (Re)queue cap at the end of the delayed cap release list.
409 *
410 * If I_FLUSH is set, leave the inode at the front of the list.
411 *
412 * Caller holds i_lock
413 * -> we take mdsc->cap_delay_lock
414 */
415static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
416 struct ceph_inode_info *ci)
417{
418 __cap_set_timeouts(mdsc, ci);
419 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
420 ci->i_ceph_flags, ci->i_hold_caps_max);
421 if (!mdsc->stopping) {
422 spin_lock(&mdsc->cap_delay_lock);
423 if (!list_empty(&ci->i_cap_delay_list)) {
424 if (ci->i_ceph_flags & CEPH_I_FLUSH)
425 goto no_change;
426 list_del_init(&ci->i_cap_delay_list);
427 }
428 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
429no_change:
430 spin_unlock(&mdsc->cap_delay_lock);
431 }
432}
433
434/*
435 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
436 * indicating we should send a cap message to flush dirty metadata
437 * asap, and move to the front of the delayed cap list.
438 */
439static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
440 struct ceph_inode_info *ci)
441{
442 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
443 spin_lock(&mdsc->cap_delay_lock);
444 ci->i_ceph_flags |= CEPH_I_FLUSH;
445 if (!list_empty(&ci->i_cap_delay_list))
446 list_del_init(&ci->i_cap_delay_list);
447 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
448 spin_unlock(&mdsc->cap_delay_lock);
449}
450
451/*
452 * Cancel delayed work on cap.
453 *
454 * Caller must hold i_lock.
455 */
456static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
457 struct ceph_inode_info *ci)
458{
459 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
460 if (list_empty(&ci->i_cap_delay_list))
461 return;
462 spin_lock(&mdsc->cap_delay_lock);
463 list_del_init(&ci->i_cap_delay_list);
464 spin_unlock(&mdsc->cap_delay_lock);
465}
466
467/*
468 * Common issue checks for add_cap, handle_cap_grant.
469 */
470static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
471 unsigned issued)
472{
473 unsigned had = __ceph_caps_issued(ci, NULL);
474
475 /*
476 * Each time we receive FILE_CACHE anew, we increment
477 * i_rdcache_gen.
478 */
479 if ((issued & CEPH_CAP_FILE_CACHE) &&
480 (had & CEPH_CAP_FILE_CACHE) == 0)
481 ci->i_rdcache_gen++;
482
483 /*
484 * if we are newly issued FILE_SHARED, clear I_COMPLETE; we
485 * don't know what happened to this directory while we didn't
486 * have the cap.
487 */
488 if ((issued & CEPH_CAP_FILE_SHARED) &&
489 (had & CEPH_CAP_FILE_SHARED) == 0) {
490 ci->i_shared_gen++;
491 if (S_ISDIR(ci->vfs_inode.i_mode)) {
492 dout(" marking %p NOT complete\n", &ci->vfs_inode);
493 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
494 }
495 }
496}
497
498/*
499 * Add a capability under the given MDS session.
500 *
501 * Caller should hold session snap_rwsem (read) and s_mutex.
502 *
503 * @fmode is the open file mode, if we are opening a file, otherwise
504 * it is < 0. (This is so we can atomically add the cap and add an
505 * open file reference to it.)
506 */
507int ceph_add_cap(struct inode *inode,
508 struct ceph_mds_session *session, u64 cap_id,
509 int fmode, unsigned issued, unsigned wanted,
510 unsigned seq, unsigned mseq, u64 realmino, int flags,
511 struct ceph_cap_reservation *caps_reservation)
512{
513 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
514 struct ceph_inode_info *ci = ceph_inode(inode);
515 struct ceph_cap *new_cap = NULL;
516 struct ceph_cap *cap;
517 int mds = session->s_mds;
518 int actual_wanted;
519
520 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
521 session->s_mds, cap_id, ceph_cap_string(issued), seq);
522
523 /*
524 * If we are opening the file, include file mode wanted bits
525 * in wanted.
526 */
527 if (fmode >= 0)
528 wanted |= ceph_caps_for_mode(fmode);
529
530retry:
531 spin_lock(&inode->i_lock);
532 cap = __get_cap_for_mds(ci, mds);
533 if (!cap) {
534 if (new_cap) {
535 cap = new_cap;
536 new_cap = NULL;
537 } else {
538 spin_unlock(&inode->i_lock);
539 new_cap = get_cap(caps_reservation);
540 if (new_cap == NULL)
541 return -ENOMEM;
542 goto retry;
543 }
544
545 cap->issued = 0;
546 cap->implemented = 0;
547 cap->mds = mds;
548 cap->mds_wanted = 0;
549
550 cap->ci = ci;
551 __insert_cap_node(ci, cap);
552
553 /* clear out old exporting info? (i.e. on cap import) */
554 if (ci->i_cap_exporting_mds == mds) {
555 ci->i_cap_exporting_issued = 0;
556 ci->i_cap_exporting_mseq = 0;
557 ci->i_cap_exporting_mds = -1;
558 }
559
560 /* add to session cap list */
561 cap->session = session;
562 spin_lock(&session->s_cap_lock);
563 list_add_tail(&cap->session_caps, &session->s_caps);
564 session->s_nr_caps++;
565 spin_unlock(&session->s_cap_lock);
566 }
567
568 if (!ci->i_snap_realm) {
569 /*
570 * add this inode to the appropriate snap realm
571 */
572 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
573 realmino);
574 if (realm) {
575 ceph_get_snap_realm(mdsc, realm);
576 spin_lock(&realm->inodes_with_caps_lock);
577 ci->i_snap_realm = realm;
578 list_add(&ci->i_snap_realm_item,
579 &realm->inodes_with_caps);
580 spin_unlock(&realm->inodes_with_caps_lock);
581 } else {
582 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
583 realmino);
584 }
585 }
586
587 __check_cap_issue(ci, cap, issued);
588
589 /*
590 * If we are issued caps we don't want, or the mds' wanted
591 * value appears to be off, queue a check so we'll release
592 * later and/or update the mds wanted value.
593 */
594 actual_wanted = __ceph_caps_wanted(ci);
595 if ((wanted & ~actual_wanted) ||
596 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
597 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
598 ceph_cap_string(issued), ceph_cap_string(wanted),
599 ceph_cap_string(actual_wanted));
600 __cap_delay_requeue(mdsc, ci);
601 }
602
603 if (flags & CEPH_CAP_FLAG_AUTH)
604 ci->i_auth_cap = cap;
605 else if (ci->i_auth_cap == cap)
606 ci->i_auth_cap = NULL;
607
608 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
609 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
610 ceph_cap_string(issued|cap->issued), seq, mds);
611 cap->cap_id = cap_id;
612 cap->issued = issued;
613 cap->implemented |= issued;
614 cap->mds_wanted |= wanted;
615 cap->seq = seq;
616 cap->issue_seq = seq;
617 cap->mseq = mseq;
618 cap->cap_gen = session->s_cap_gen;
619
620 if (fmode >= 0)
621 __ceph_get_fmode(ci, fmode);
622 spin_unlock(&inode->i_lock);
623 wake_up(&ci->i_cap_wq);
624 return 0;
625}
626
627/*
628 * Return true if cap has not timed out and belongs to the current
629 * generation of the MDS session (i.e. has not gone 'stale' due to
630 * us losing touch with the mds).
631 */
632static int __cap_is_valid(struct ceph_cap *cap)
633{
634 unsigned long ttl;
635 u32 gen;
636
637 spin_lock(&cap->session->s_cap_lock);
638 gen = cap->session->s_cap_gen;
639 ttl = cap->session->s_cap_ttl;
640 spin_unlock(&cap->session->s_cap_lock);
641
642 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
643 dout("__cap_is_valid %p cap %p issued %s "
644 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
645 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
646 return 0;
647 }
648
649 return 1;
650}
651
652/*
653 * Return set of valid cap bits issued to us. Note that caps time
654 * out, and may be invalidated in bulk if the client session times out
655 * and session->s_cap_gen is bumped.
656 */
657int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
658{
659 int have = ci->i_snap_caps | ci->i_cap_exporting_issued;
660 struct ceph_cap *cap;
661 struct rb_node *p;
662
663 if (implemented)
664 *implemented = 0;
665 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
666 cap = rb_entry(p, struct ceph_cap, ci_node);
667 if (!__cap_is_valid(cap))
668 continue;
669 dout("__ceph_caps_issued %p cap %p issued %s\n",
670 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
671 have |= cap->issued;
672 if (implemented)
673 *implemented |= cap->implemented;
674 }
675 return have;
676}
677
678/*
679 * Get cap bits issued by caps other than @ocap
680 */
681int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
682{
683 int have = ci->i_snap_caps;
684 struct ceph_cap *cap;
685 struct rb_node *p;
686
687 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
688 cap = rb_entry(p, struct ceph_cap, ci_node);
689 if (cap == ocap)
690 continue;
691 if (!__cap_is_valid(cap))
692 continue;
693 have |= cap->issued;
694 }
695 return have;
696}
697
698/*
699 * Move a cap to the end of the LRU (oldest caps at list head, newest
700 * at list tail).
701 */
702static void __touch_cap(struct ceph_cap *cap)
703{
704 struct ceph_mds_session *s = cap->session;
705
706 spin_lock(&s->s_cap_lock);
707 if (s->s_cap_iterator == NULL) {
708 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
709 s->s_mds);
710 list_move_tail(&cap->session_caps, &s->s_caps);
711 } else {
712 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
713 &cap->ci->vfs_inode, cap, s->s_mds);
714 }
715 spin_unlock(&s->s_cap_lock);
716}
717
718/*
719 * Check if we hold the given mask. If so, move the cap(s) to the
720 * front of their respective LRUs. (This is the preferred way for
721 * callers to check for caps they want.)
722 */
723int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
724{
725 struct ceph_cap *cap;
726 struct rb_node *p;
727 int have = ci->i_snap_caps;
728
729 if ((have & mask) == mask) {
730 dout("__ceph_caps_issued_mask %p snap issued %s"
731 " (mask %s)\n", &ci->vfs_inode,
732 ceph_cap_string(have),
733 ceph_cap_string(mask));
734 return 1;
735 }
736
737 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
738 cap = rb_entry(p, struct ceph_cap, ci_node);
739 if (!__cap_is_valid(cap))
740 continue;
741 if ((cap->issued & mask) == mask) {
742 dout("__ceph_caps_issued_mask %p cap %p issued %s"
743 " (mask %s)\n", &ci->vfs_inode, cap,
744 ceph_cap_string(cap->issued),
745 ceph_cap_string(mask));
746 if (touch)
747 __touch_cap(cap);
748 return 1;
749 }
750
751 /* does a combination of caps satisfy mask? */
752 have |= cap->issued;
753 if ((have & mask) == mask) {
754 dout("__ceph_caps_issued_mask %p combo issued %s"
755 " (mask %s)\n", &ci->vfs_inode,
756 ceph_cap_string(cap->issued),
757 ceph_cap_string(mask));
758 if (touch) {
759 struct rb_node *q;
760
761 /* touch this + preceeding caps */
762 __touch_cap(cap);
763 for (q = rb_first(&ci->i_caps); q != p;
764 q = rb_next(q)) {
765 cap = rb_entry(q, struct ceph_cap,
766 ci_node);
767 if (!__cap_is_valid(cap))
768 continue;
769 __touch_cap(cap);
770 }
771 }
772 return 1;
773 }
774 }
775
776 return 0;
777}
778
779/*
780 * Return true if mask caps are currently being revoked by an MDS.
781 */
782int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
783{
784 struct inode *inode = &ci->vfs_inode;
785 struct ceph_cap *cap;
786 struct rb_node *p;
787 int ret = 0;
788
789 spin_lock(&inode->i_lock);
790 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
791 cap = rb_entry(p, struct ceph_cap, ci_node);
792 if (__cap_is_valid(cap) &&
793 (cap->implemented & ~cap->issued & mask)) {
794 ret = 1;
795 break;
796 }
797 }
798 spin_unlock(&inode->i_lock);
799 dout("ceph_caps_revoking %p %s = %d\n", inode,
800 ceph_cap_string(mask), ret);
801 return ret;
802}
803
804int __ceph_caps_used(struct ceph_inode_info *ci)
805{
806 int used = 0;
807 if (ci->i_pin_ref)
808 used |= CEPH_CAP_PIN;
809 if (ci->i_rd_ref)
810 used |= CEPH_CAP_FILE_RD;
811 if (ci->i_rdcache_ref || ci->i_rdcache_gen)
812 used |= CEPH_CAP_FILE_CACHE;
813 if (ci->i_wr_ref)
814 used |= CEPH_CAP_FILE_WR;
815 if (ci->i_wrbuffer_ref)
816 used |= CEPH_CAP_FILE_BUFFER;
817 return used;
818}
819
820/*
821 * wanted, by virtue of open file modes
822 */
823int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
824{
825 int want = 0;
826 int mode;
827 for (mode = 0; mode < 4; mode++)
828 if (ci->i_nr_by_mode[mode])
829 want |= ceph_caps_for_mode(mode);
830 return want;
831}
832
833/*
834 * Return caps we have registered with the MDS(s) as 'wanted'.
835 */
836int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
837{
838 struct ceph_cap *cap;
839 struct rb_node *p;
840 int mds_wanted = 0;
841
842 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
843 cap = rb_entry(p, struct ceph_cap, ci_node);
844 if (!__cap_is_valid(cap))
845 continue;
846 mds_wanted |= cap->mds_wanted;
847 }
848 return mds_wanted;
849}
850
851/*
852 * called under i_lock
853 */
854static int __ceph_is_any_caps(struct ceph_inode_info *ci)
855{
856 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0;
857}
858
859/*
860 * caller should hold i_lock.
861 * caller will not hold session s_mutex if called from destroy_inode.
862 */
863void __ceph_remove_cap(struct ceph_cap *cap)
864{
865 struct ceph_mds_session *session = cap->session;
866 struct ceph_inode_info *ci = cap->ci;
867 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
868
869 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
870
871 /* remove from inode list */
872 rb_erase(&cap->ci_node, &ci->i_caps);
873 cap->ci = NULL;
874 if (ci->i_auth_cap == cap)
875 ci->i_auth_cap = NULL;
876
877 /* remove from session list */
878 spin_lock(&session->s_cap_lock);
879 if (session->s_cap_iterator == cap) {
880 /* not yet, we are iterating over this very cap */
881 dout("__ceph_remove_cap delaying %p removal from session %p\n",
882 cap, cap->session);
883 } else {
884 list_del_init(&cap->session_caps);
885 session->s_nr_caps--;
886 cap->session = NULL;
887 }
888 spin_unlock(&session->s_cap_lock);
889
890 if (cap->session == NULL)
891 ceph_put_cap(cap);
892
893 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
894 struct ceph_snap_realm *realm = ci->i_snap_realm;
895 spin_lock(&realm->inodes_with_caps_lock);
896 list_del_init(&ci->i_snap_realm_item);
897 ci->i_snap_realm_counter++;
898 ci->i_snap_realm = NULL;
899 spin_unlock(&realm->inodes_with_caps_lock);
900 ceph_put_snap_realm(mdsc, realm);
901 }
902 if (!__ceph_is_any_real_caps(ci))
903 __cap_delay_cancel(mdsc, ci);
904}
905
906/*
907 * Build and send a cap message to the given MDS.
908 *
909 * Caller should be holding s_mutex.
910 */
911static int send_cap_msg(struct ceph_mds_session *session,
912 u64 ino, u64 cid, int op,
913 int caps, int wanted, int dirty,
914 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
915 u64 size, u64 max_size,
916 struct timespec *mtime, struct timespec *atime,
917 u64 time_warp_seq,
918 uid_t uid, gid_t gid, mode_t mode,
919 u64 xattr_version,
920 struct ceph_buffer *xattrs_buf,
921 u64 follows)
922{
923 struct ceph_mds_caps *fc;
924 struct ceph_msg *msg;
925
926 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
927 " seq %u/%u mseq %u follows %lld size %llu/%llu"
928 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
929 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
930 ceph_cap_string(dirty),
931 seq, issue_seq, mseq, follows, size, max_size,
932 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
933
934 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), 0, 0, NULL);
935 if (IS_ERR(msg))
936 return PTR_ERR(msg);
937
938 msg->hdr.tid = cpu_to_le64(flush_tid);
939
940 fc = msg->front.iov_base;
941 memset(fc, 0, sizeof(*fc));
942
943 fc->cap_id = cpu_to_le64(cid);
944 fc->op = cpu_to_le32(op);
945 fc->seq = cpu_to_le32(seq);
946 fc->issue_seq = cpu_to_le32(issue_seq);
947 fc->migrate_seq = cpu_to_le32(mseq);
948 fc->caps = cpu_to_le32(caps);
949 fc->wanted = cpu_to_le32(wanted);
950 fc->dirty = cpu_to_le32(dirty);
951 fc->ino = cpu_to_le64(ino);
952 fc->snap_follows = cpu_to_le64(follows);
953
954 fc->size = cpu_to_le64(size);
955 fc->max_size = cpu_to_le64(max_size);
956 if (mtime)
957 ceph_encode_timespec(&fc->mtime, mtime);
958 if (atime)
959 ceph_encode_timespec(&fc->atime, atime);
960 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
961
962 fc->uid = cpu_to_le32(uid);
963 fc->gid = cpu_to_le32(gid);
964 fc->mode = cpu_to_le32(mode);
965
966 fc->xattr_version = cpu_to_le64(xattr_version);
967 if (xattrs_buf) {
968 msg->middle = ceph_buffer_get(xattrs_buf);
969 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
970 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
971 }
972
973 ceph_con_send(&session->s_con, msg);
974 return 0;
975}
976
977/*
978 * Queue cap releases when an inode is dropped from our cache. Since
979 * inode is about to be destroyed, there is no need for i_lock.
980 */
981void ceph_queue_caps_release(struct inode *inode)
982{
983 struct ceph_inode_info *ci = ceph_inode(inode);
984 struct rb_node *p;
985
986 p = rb_first(&ci->i_caps);
987 while (p) {
988 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
989 struct ceph_mds_session *session = cap->session;
990 struct ceph_msg *msg;
991 struct ceph_mds_cap_release *head;
992 struct ceph_mds_cap_item *item;
993
994 spin_lock(&session->s_cap_lock);
995 BUG_ON(!session->s_num_cap_releases);
996 msg = list_first_entry(&session->s_cap_releases,
997 struct ceph_msg, list_head);
998
999 dout(" adding %p release to mds%d msg %p (%d left)\n",
1000 inode, session->s_mds, msg, session->s_num_cap_releases);
1001
1002 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1003 head = msg->front.iov_base;
1004 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1);
1005 item = msg->front.iov_base + msg->front.iov_len;
1006 item->ino = cpu_to_le64(ceph_ino(inode));
1007 item->cap_id = cpu_to_le64(cap->cap_id);
1008 item->migrate_seq = cpu_to_le32(cap->mseq);
1009 item->seq = cpu_to_le32(cap->issue_seq);
1010
1011 session->s_num_cap_releases--;
1012
1013 msg->front.iov_len += sizeof(*item);
1014 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1015 dout(" release msg %p full\n", msg);
1016 list_move_tail(&msg->list_head,
1017 &session->s_cap_releases_done);
1018 } else {
1019 dout(" release msg %p at %d/%d (%d)\n", msg,
1020 (int)le32_to_cpu(head->num),
1021 (int)CEPH_CAPS_PER_RELEASE,
1022 (int)msg->front.iov_len);
1023 }
1024 spin_unlock(&session->s_cap_lock);
1025 p = rb_next(p);
1026 __ceph_remove_cap(cap);
1027 }
1028}
1029
1030/*
1031 * Send a cap msg on the given inode. Update our caps state, then
1032 * drop i_lock and send the message.
1033 *
1034 * Make note of max_size reported/requested from mds, revoked caps
1035 * that have now been implemented.
1036 *
1037 * Make half-hearted attempt ot to invalidate page cache if we are
1038 * dropping RDCACHE. Note that this will leave behind locked pages
1039 * that we'll then need to deal with elsewhere.
1040 *
1041 * Return non-zero if delayed release, or we experienced an error
1042 * such that the caller should requeue + retry later.
1043 *
1044 * called with i_lock, then drops it.
1045 * caller should hold snap_rwsem (read), s_mutex.
1046 */
1047static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1048 int op, int used, int want, int retain, int flushing,
1049 unsigned *pflush_tid)
1050 __releases(cap->ci->vfs_inode->i_lock)
1051{
1052 struct ceph_inode_info *ci = cap->ci;
1053 struct inode *inode = &ci->vfs_inode;
1054 u64 cap_id = cap->cap_id;
1055 int held, revoking, dropping, keep;
1056 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1057 u64 size, max_size;
1058 struct timespec mtime, atime;
1059 int wake = 0;
1060 mode_t mode;
1061 uid_t uid;
1062 gid_t gid;
1063 struct ceph_mds_session *session;
1064 u64 xattr_version = 0;
1065 int delayed = 0;
1066 u64 flush_tid = 0;
1067 int i;
1068 int ret;
1069
1070 held = cap->issued | cap->implemented;
1071 revoking = cap->implemented & ~cap->issued;
1072 retain &= ~revoking;
1073 dropping = cap->issued & ~retain;
1074
1075 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1076 inode, cap, cap->session,
1077 ceph_cap_string(held), ceph_cap_string(held & retain),
1078 ceph_cap_string(revoking));
1079 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1080
1081 session = cap->session;
1082
1083 /* don't release wanted unless we've waited a bit. */
1084 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1085 time_before(jiffies, ci->i_hold_caps_min)) {
1086 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1087 ceph_cap_string(cap->issued),
1088 ceph_cap_string(cap->issued & retain),
1089 ceph_cap_string(cap->mds_wanted),
1090 ceph_cap_string(want));
1091 want |= cap->mds_wanted;
1092 retain |= cap->issued;
1093 delayed = 1;
1094 }
1095 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1096
1097 cap->issued &= retain; /* drop bits we don't want */
1098 if (cap->implemented & ~cap->issued) {
1099 /*
1100 * Wake up any waiters on wanted -> needed transition.
1101 * This is due to the weird transition from buffered
1102 * to sync IO... we need to flush dirty pages _before_
1103 * allowing sync writes to avoid reordering.
1104 */
1105 wake = 1;
1106 }
1107 cap->implemented &= cap->issued | used;
1108 cap->mds_wanted = want;
1109
1110 if (flushing) {
1111 /*
1112 * assign a tid for flush operations so we can avoid
1113 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1114 * clean type races. track latest tid for every bit
1115 * so we can handle flush AxFw, flush Fw, and have the
1116 * first ack clean Ax.
1117 */
1118 flush_tid = ++ci->i_cap_flush_last_tid;
1119 if (pflush_tid)
1120 *pflush_tid = flush_tid;
1121 dout(" cap_flush_tid %d\n", (int)flush_tid);
1122 for (i = 0; i < CEPH_CAP_BITS; i++)
1123 if (flushing & (1 << i))
1124 ci->i_cap_flush_tid[i] = flush_tid;
1125 }
1126
1127 keep = cap->implemented;
1128 seq = cap->seq;
1129 issue_seq = cap->issue_seq;
1130 mseq = cap->mseq;
1131 size = inode->i_size;
1132 ci->i_reported_size = size;
1133 max_size = ci->i_wanted_max_size;
1134 ci->i_requested_max_size = max_size;
1135 mtime = inode->i_mtime;
1136 atime = inode->i_atime;
1137 time_warp_seq = ci->i_time_warp_seq;
1138 follows = ci->i_snap_realm->cached_context->seq;
1139 uid = inode->i_uid;
1140 gid = inode->i_gid;
1141 mode = inode->i_mode;
1142
1143 if (dropping & CEPH_CAP_XATTR_EXCL) {
1144 __ceph_build_xattrs_blob(ci);
1145 xattr_version = ci->i_xattrs.version + 1;
1146 }
1147
1148 spin_unlock(&inode->i_lock);
1149
1150 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1151 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1152 size, max_size, &mtime, &atime, time_warp_seq,
1153 uid, gid, mode,
1154 xattr_version,
1155 (flushing & CEPH_CAP_XATTR_EXCL) ? ci->i_xattrs.blob : NULL,
1156 follows);
1157 if (ret < 0) {
1158 dout("error sending cap msg, must requeue %p\n", inode);
1159 delayed = 1;
1160 }
1161
1162 if (wake)
1163 wake_up(&ci->i_cap_wq);
1164
1165 return delayed;
1166}
1167
1168/*
1169 * When a snapshot is taken, clients accumulate dirty metadata on
1170 * inodes with capabilities in ceph_cap_snaps to describe the file
1171 * state at the time the snapshot was taken. This must be flushed
1172 * asynchronously back to the MDS once sync writes complete and dirty
1173 * data is written out.
1174 *
1175 * Called under i_lock. Takes s_mutex as needed.
1176 */
1177void __ceph_flush_snaps(struct ceph_inode_info *ci,
1178 struct ceph_mds_session **psession)
1179{
1180 struct inode *inode = &ci->vfs_inode;
1181 int mds;
1182 struct ceph_cap_snap *capsnap;
1183 u32 mseq;
1184 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1185 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1186 session->s_mutex */
1187 u64 next_follows = 0; /* keep track of how far we've gotten through the
1188 i_cap_snaps list, and skip these entries next time
1189 around to avoid an infinite loop */
1190
1191 if (psession)
1192 session = *psession;
1193
1194 dout("__flush_snaps %p\n", inode);
1195retry:
1196 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1197 /* avoid an infiniute loop after retry */
1198 if (capsnap->follows < next_follows)
1199 continue;
1200 /*
1201 * we need to wait for sync writes to complete and for dirty
1202 * pages to be written out.
1203 */
1204 if (capsnap->dirty_pages || capsnap->writing)
1205 continue;
1206
1207 /* pick mds, take s_mutex */
1208 mds = __ceph_get_cap_mds(ci, &mseq);
1209 if (session && session->s_mds != mds) {
1210 dout("oops, wrong session %p mutex\n", session);
1211 mutex_unlock(&session->s_mutex);
1212 ceph_put_mds_session(session);
1213 session = NULL;
1214 }
1215 if (!session) {
1216 spin_unlock(&inode->i_lock);
1217 mutex_lock(&mdsc->mutex);
1218 session = __ceph_lookup_mds_session(mdsc, mds);
1219 mutex_unlock(&mdsc->mutex);
1220 if (session) {
1221 dout("inverting session/ino locks on %p\n",
1222 session);
1223 mutex_lock(&session->s_mutex);
1224 }
1225 /*
1226 * if session == NULL, we raced against a cap
1227 * deletion. retry, and we'll get a better
1228 * @mds value next time.
1229 */
1230 spin_lock(&inode->i_lock);
1231 goto retry;
1232 }
1233
1234 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1235 atomic_inc(&capsnap->nref);
1236 if (!list_empty(&capsnap->flushing_item))
1237 list_del_init(&capsnap->flushing_item);
1238 list_add_tail(&capsnap->flushing_item,
1239 &session->s_cap_snaps_flushing);
1240 spin_unlock(&inode->i_lock);
1241
1242 dout("flush_snaps %p cap_snap %p follows %lld size %llu\n",
1243 inode, capsnap, next_follows, capsnap->size);
1244 send_cap_msg(session, ceph_vino(inode).ino, 0,
1245 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1246 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1247 capsnap->size, 0,
1248 &capsnap->mtime, &capsnap->atime,
1249 capsnap->time_warp_seq,
1250 capsnap->uid, capsnap->gid, capsnap->mode,
1251 0, NULL,
1252 capsnap->follows);
1253
1254 next_follows = capsnap->follows + 1;
1255 ceph_put_cap_snap(capsnap);
1256
1257 spin_lock(&inode->i_lock);
1258 goto retry;
1259 }
1260
1261 /* we flushed them all; remove this inode from the queue */
1262 spin_lock(&mdsc->snap_flush_lock);
1263 list_del_init(&ci->i_snap_flush_item);
1264 spin_unlock(&mdsc->snap_flush_lock);
1265
1266 if (psession)
1267 *psession = session;
1268 else if (session) {
1269 mutex_unlock(&session->s_mutex);
1270 ceph_put_mds_session(session);
1271 }
1272}
1273
1274static void ceph_flush_snaps(struct ceph_inode_info *ci)
1275{
1276 struct inode *inode = &ci->vfs_inode;
1277
1278 spin_lock(&inode->i_lock);
1279 __ceph_flush_snaps(ci, NULL);
1280 spin_unlock(&inode->i_lock);
1281}
1282
1283/*
1284 * Mark caps dirty. If inode is newly dirty, add to the global dirty
1285 * list.
1286 */
1287void __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1288{
1289 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
1290 struct inode *inode = &ci->vfs_inode;
1291 int was = ci->i_dirty_caps;
1292 int dirty = 0;
1293
1294 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1295 ceph_cap_string(mask), ceph_cap_string(was),
1296 ceph_cap_string(was | mask));
1297 ci->i_dirty_caps |= mask;
1298 if (was == 0) {
1299 dout(" inode %p now dirty\n", &ci->vfs_inode);
1300 BUG_ON(!list_empty(&ci->i_dirty_item));
1301 spin_lock(&mdsc->cap_dirty_lock);
1302 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1303 spin_unlock(&mdsc->cap_dirty_lock);
1304 if (ci->i_flushing_caps == 0) {
1305 igrab(inode);
1306 dirty |= I_DIRTY_SYNC;
1307 }
1308 }
1309 BUG_ON(list_empty(&ci->i_dirty_item));
1310 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1311 (mask & CEPH_CAP_FILE_BUFFER))
1312 dirty |= I_DIRTY_DATASYNC;
1313 if (dirty)
1314 __mark_inode_dirty(inode, dirty);
1315 __cap_delay_requeue(mdsc, ci);
1316}
1317
1318/*
1319 * Add dirty inode to the flushing list. Assigned a seq number so we
1320 * can wait for caps to flush without starving.
1321 *
1322 * Called under i_lock.
1323 */
1324static int __mark_caps_flushing(struct inode *inode,
1325 struct ceph_mds_session *session)
1326{
1327 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1328 struct ceph_inode_info *ci = ceph_inode(inode);
1329 int flushing;
1330
1331 BUG_ON(ci->i_dirty_caps == 0);
1332 BUG_ON(list_empty(&ci->i_dirty_item));
1333
1334 flushing = ci->i_dirty_caps;
1335 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1336 ceph_cap_string(flushing),
1337 ceph_cap_string(ci->i_flushing_caps),
1338 ceph_cap_string(ci->i_flushing_caps | flushing));
1339 ci->i_flushing_caps |= flushing;
1340 ci->i_dirty_caps = 0;
1341 dout(" inode %p now !dirty\n", inode);
1342
1343 spin_lock(&mdsc->cap_dirty_lock);
1344 list_del_init(&ci->i_dirty_item);
1345
1346 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1347 if (list_empty(&ci->i_flushing_item)) {
1348 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1349 mdsc->num_cap_flushing++;
1350 dout(" inode %p now flushing seq %lld\n", inode,
1351 ci->i_cap_flush_seq);
1352 } else {
1353 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1354 dout(" inode %p now flushing (more) seq %lld\n", inode,
1355 ci->i_cap_flush_seq);
1356 }
1357 spin_unlock(&mdsc->cap_dirty_lock);
1358
1359 return flushing;
1360}
1361
1362/*
1363 * try to invalidate mapping pages without blocking.
1364 */
1365static int mapping_is_empty(struct address_space *mapping)
1366{
1367 struct page *page = find_get_page(mapping, 0);
1368
1369 if (!page)
1370 return 1;
1371
1372 put_page(page);
1373 return 0;
1374}
1375
1376static int try_nonblocking_invalidate(struct inode *inode)
1377{
1378 struct ceph_inode_info *ci = ceph_inode(inode);
1379 u32 invalidating_gen = ci->i_rdcache_gen;
1380
1381 spin_unlock(&inode->i_lock);
1382 invalidate_mapping_pages(&inode->i_data, 0, -1);
1383 spin_lock(&inode->i_lock);
1384
1385 if (mapping_is_empty(&inode->i_data) &&
1386 invalidating_gen == ci->i_rdcache_gen) {
1387 /* success. */
1388 dout("try_nonblocking_invalidate %p success\n", inode);
1389 ci->i_rdcache_gen = 0;
1390 ci->i_rdcache_revoking = 0;
1391 return 0;
1392 }
1393 dout("try_nonblocking_invalidate %p failed\n", inode);
1394 return -1;
1395}
1396
1397/*
1398 * Swiss army knife function to examine currently used and wanted
1399 * versus held caps. Release, flush, ack revoked caps to mds as
1400 * appropriate.
1401 *
1402 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1403 * cap release further.
1404 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1405 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1406 * further delay.
1407 */
1408void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1409 struct ceph_mds_session *session)
1410{
1411 struct ceph_client *client = ceph_inode_to_client(&ci->vfs_inode);
1412 struct ceph_mds_client *mdsc = &client->mdsc;
1413 struct inode *inode = &ci->vfs_inode;
1414 struct ceph_cap *cap;
1415 int file_wanted, used;
1416 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1417 int drop_session_lock = session ? 0 : 1;
1418 int issued, implemented, want, retain, revoking, flushing = 0;
1419 int mds = -1; /* keep track of how far we've gone through i_caps list
1420 to avoid an infinite loop on retry */
1421 struct rb_node *p;
1422 int tried_invalidate = 0;
1423 int delayed = 0, sent = 0, force_requeue = 0, num;
1424 int queue_invalidate = 0;
1425 int is_delayed = flags & CHECK_CAPS_NODELAY;
1426
1427 /* if we are unmounting, flush any unused caps immediately. */
1428 if (mdsc->stopping)
1429 is_delayed = 1;
1430
1431 spin_lock(&inode->i_lock);
1432
1433 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1434 flags |= CHECK_CAPS_FLUSH;
1435
1436 /* flush snaps first time around only */
1437 if (!list_empty(&ci->i_cap_snaps))
1438 __ceph_flush_snaps(ci, &session);
1439 goto retry_locked;
1440retry:
1441 spin_lock(&inode->i_lock);
1442retry_locked:
1443 file_wanted = __ceph_caps_file_wanted(ci);
1444 used = __ceph_caps_used(ci);
1445 want = file_wanted | used;
1446 issued = __ceph_caps_issued(ci, &implemented);
1447 revoking = implemented & ~issued;
1448
1449 retain = want | CEPH_CAP_PIN;
1450 if (!mdsc->stopping && inode->i_nlink > 0) {
1451 if (want) {
1452 retain |= CEPH_CAP_ANY; /* be greedy */
1453 } else {
1454 retain |= CEPH_CAP_ANY_SHARED;
1455 /*
1456 * keep RD only if we didn't have the file open RW,
1457 * because then the mds would revoke it anyway to
1458 * journal max_size=0.
1459 */
1460 if (ci->i_max_size == 0)
1461 retain |= CEPH_CAP_ANY_RD;
1462 }
1463 }
1464
1465 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1466 " issued %s revoking %s retain %s %s%s%s\n", inode,
1467 ceph_cap_string(file_wanted),
1468 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1469 ceph_cap_string(ci->i_flushing_caps),
1470 ceph_cap_string(issued), ceph_cap_string(revoking),
1471 ceph_cap_string(retain),
1472 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1473 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1474 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1475
1476 /*
1477 * If we no longer need to hold onto old our caps, and we may
1478 * have cached pages, but don't want them, then try to invalidate.
1479 * If we fail, it's because pages are locked.... try again later.
1480 */
1481 if ((!is_delayed || mdsc->stopping) &&
1482 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1483 ci->i_rdcache_gen && /* may have cached pages */
1484 (file_wanted == 0 || /* no open files */
1485 (revoking & CEPH_CAP_FILE_CACHE)) && /* or revoking cache */
1486 !tried_invalidate) {
1487 dout("check_caps trying to invalidate on %p\n", inode);
1488 if (try_nonblocking_invalidate(inode) < 0) {
1489 if (revoking & CEPH_CAP_FILE_CACHE) {
1490 dout("check_caps queuing invalidate\n");
1491 queue_invalidate = 1;
1492 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1493 } else {
1494 dout("check_caps failed to invalidate pages\n");
1495 /* we failed to invalidate pages. check these
1496 caps again later. */
1497 force_requeue = 1;
1498 __cap_set_timeouts(mdsc, ci);
1499 }
1500 }
1501 tried_invalidate = 1;
1502 goto retry_locked;
1503 }
1504
1505 num = 0;
1506 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1507 cap = rb_entry(p, struct ceph_cap, ci_node);
1508 num++;
1509
1510 /* avoid looping forever */
1511 if (mds >= cap->mds ||
1512 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1513 continue;
1514
1515 /* NOTE: no side-effects allowed, until we take s_mutex */
1516
1517 revoking = cap->implemented & ~cap->issued;
1518 if (revoking)
1519 dout(" mds%d revoking %s\n", cap->mds,
1520 ceph_cap_string(revoking));
1521
1522 if (cap == ci->i_auth_cap &&
1523 (cap->issued & CEPH_CAP_FILE_WR)) {
1524 /* request larger max_size from MDS? */
1525 if (ci->i_wanted_max_size > ci->i_max_size &&
1526 ci->i_wanted_max_size > ci->i_requested_max_size) {
1527 dout("requesting new max_size\n");
1528 goto ack;
1529 }
1530
1531 /* approaching file_max? */
1532 if ((inode->i_size << 1) >= ci->i_max_size &&
1533 (ci->i_reported_size << 1) < ci->i_max_size) {
1534 dout("i_size approaching max_size\n");
1535 goto ack;
1536 }
1537 }
1538 /* flush anything dirty? */
1539 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1540 ci->i_dirty_caps) {
1541 dout("flushing dirty caps\n");
1542 goto ack;
1543 }
1544
1545 /* completed revocation? going down and there are no caps? */
1546 if (revoking && (revoking & used) == 0) {
1547 dout("completed revocation of %s\n",
1548 ceph_cap_string(cap->implemented & ~cap->issued));
1549 goto ack;
1550 }
1551
1552 /* want more caps from mds? */
1553 if (want & ~(cap->mds_wanted | cap->issued))
1554 goto ack;
1555
1556 /* things we might delay */
1557 if ((cap->issued & ~retain) == 0 &&
1558 cap->mds_wanted == want)
1559 continue; /* nope, all good */
1560
1561 if (is_delayed)
1562 goto ack;
1563
1564 /* delay? */
1565 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1566 time_before(jiffies, ci->i_hold_caps_max)) {
1567 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1568 ceph_cap_string(cap->issued),
1569 ceph_cap_string(cap->issued & retain),
1570 ceph_cap_string(cap->mds_wanted),
1571 ceph_cap_string(want));
1572 delayed++;
1573 continue;
1574 }
1575
1576ack:
1577 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1578 dout(" skipping %p I_NOFLUSH set\n", inode);
1579 continue;
1580 }
1581
1582 if (session && session != cap->session) {
1583 dout("oops, wrong session %p mutex\n", session);
1584 mutex_unlock(&session->s_mutex);
1585 session = NULL;
1586 }
1587 if (!session) {
1588 session = cap->session;
1589 if (mutex_trylock(&session->s_mutex) == 0) {
1590 dout("inverting session/ino locks on %p\n",
1591 session);
1592 spin_unlock(&inode->i_lock);
1593 if (took_snap_rwsem) {
1594 up_read(&mdsc->snap_rwsem);
1595 took_snap_rwsem = 0;
1596 }
1597 mutex_lock(&session->s_mutex);
1598 goto retry;
1599 }
1600 }
1601 /* take snap_rwsem after session mutex */
1602 if (!took_snap_rwsem) {
1603 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1604 dout("inverting snap/in locks on %p\n",
1605 inode);
1606 spin_unlock(&inode->i_lock);
1607 down_read(&mdsc->snap_rwsem);
1608 took_snap_rwsem = 1;
1609 goto retry;
1610 }
1611 took_snap_rwsem = 1;
1612 }
1613
1614 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1615 flushing = __mark_caps_flushing(inode, session);
1616
1617 mds = cap->mds; /* remember mds, so we don't repeat */
1618 sent++;
1619
1620 /* __send_cap drops i_lock */
1621 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want,
1622 retain, flushing, NULL);
1623 goto retry; /* retake i_lock and restart our cap scan. */
1624 }
1625
1626 /*
1627 * Reschedule delayed caps release if we delayed anything,
1628 * otherwise cancel.
1629 */
1630 if (delayed && is_delayed)
1631 force_requeue = 1; /* __send_cap delayed release; requeue */
1632 if (!delayed && !is_delayed)
1633 __cap_delay_cancel(mdsc, ci);
1634 else if (!is_delayed || force_requeue)
1635 __cap_delay_requeue(mdsc, ci);
1636
1637 spin_unlock(&inode->i_lock);
1638
1639 if (queue_invalidate)
1640 ceph_queue_invalidate(inode);
1641
1642 if (session && drop_session_lock)
1643 mutex_unlock(&session->s_mutex);
1644 if (took_snap_rwsem)
1645 up_read(&mdsc->snap_rwsem);
1646}
1647
1648/*
1649 * Try to flush dirty caps back to the auth mds.
1650 */
1651static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session,
1652 unsigned *flush_tid)
1653{
1654 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1655 struct ceph_inode_info *ci = ceph_inode(inode);
1656 int unlock_session = session ? 0 : 1;
1657 int flushing = 0;
1658
1659retry:
1660 spin_lock(&inode->i_lock);
1661 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1662 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1663 goto out;
1664 }
1665 if (ci->i_dirty_caps && ci->i_auth_cap) {
1666 struct ceph_cap *cap = ci->i_auth_cap;
1667 int used = __ceph_caps_used(ci);
1668 int want = __ceph_caps_wanted(ci);
1669 int delayed;
1670
1671 if (!session) {
1672 spin_unlock(&inode->i_lock);
1673 session = cap->session;
1674 mutex_lock(&session->s_mutex);
1675 goto retry;
1676 }
1677 BUG_ON(session != cap->session);
1678 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1679 goto out;
1680
1681 flushing = __mark_caps_flushing(inode, session);
1682
1683 /* __send_cap drops i_lock */
1684 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1685 cap->issued | cap->implemented, flushing,
1686 flush_tid);
1687 if (!delayed)
1688 goto out_unlocked;
1689
1690 spin_lock(&inode->i_lock);
1691 __cap_delay_requeue(mdsc, ci);
1692 }
1693out:
1694 spin_unlock(&inode->i_lock);
1695out_unlocked:
1696 if (session && unlock_session)
1697 mutex_unlock(&session->s_mutex);
1698 return flushing;
1699}
1700
1701/*
1702 * Return true if we've flushed caps through the given flush_tid.
1703 */
1704static int caps_are_flushed(struct inode *inode, unsigned tid)
1705{
1706 struct ceph_inode_info *ci = ceph_inode(inode);
1707 int dirty, i, ret = 1;
1708
1709 spin_lock(&inode->i_lock);
1710 dirty = __ceph_caps_dirty(ci);
1711 for (i = 0; i < CEPH_CAP_BITS; i++)
1712 if ((ci->i_flushing_caps & (1 << i)) &&
1713 ci->i_cap_flush_tid[i] <= tid) {
1714 /* still flushing this bit */
1715 ret = 0;
1716 break;
1717 }
1718 spin_unlock(&inode->i_lock);
1719 return ret;
1720}
1721
1722/*
1723 * Wait on any unsafe replies for the given inode. First wait on the
1724 * newest request, and make that the upper bound. Then, if there are
1725 * more requests, keep waiting on the oldest as long as it is still older
1726 * than the original request.
1727 */
1728static void sync_write_wait(struct inode *inode)
1729{
1730 struct ceph_inode_info *ci = ceph_inode(inode);
1731 struct list_head *head = &ci->i_unsafe_writes;
1732 struct ceph_osd_request *req;
1733 u64 last_tid;
1734
1735 spin_lock(&ci->i_unsafe_lock);
1736 if (list_empty(head))
1737 goto out;
1738
1739 /* set upper bound as _last_ entry in chain */
1740 req = list_entry(head->prev, struct ceph_osd_request,
1741 r_unsafe_item);
1742 last_tid = req->r_tid;
1743
1744 do {
1745 ceph_osdc_get_request(req);
1746 spin_unlock(&ci->i_unsafe_lock);
1747 dout("sync_write_wait on tid %llu (until %llu)\n",
1748 req->r_tid, last_tid);
1749 wait_for_completion(&req->r_safe_completion);
1750 spin_lock(&ci->i_unsafe_lock);
1751 ceph_osdc_put_request(req);
1752
1753 /*
1754 * from here on look at first entry in chain, since we
1755 * only want to wait for anything older than last_tid
1756 */
1757 if (list_empty(head))
1758 break;
1759 req = list_entry(head->next, struct ceph_osd_request,
1760 r_unsafe_item);
1761 } while (req->r_tid < last_tid);
1762out:
1763 spin_unlock(&ci->i_unsafe_lock);
1764}
1765
1766int ceph_fsync(struct file *file, struct dentry *dentry, int datasync)
1767{
1768 struct inode *inode = dentry->d_inode;
1769 struct ceph_inode_info *ci = ceph_inode(inode);
1770 unsigned flush_tid;
1771 int ret;
1772 int dirty;
1773
1774 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1775 sync_write_wait(inode);
1776
1777 ret = filemap_write_and_wait(inode->i_mapping);
1778 if (ret < 0)
1779 return ret;
1780
1781 dirty = try_flush_caps(inode, NULL, &flush_tid);
1782 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1783
1784 /*
1785 * only wait on non-file metadata writeback (the mds
1786 * can recover size and mtime, so we don't need to
1787 * wait for that)
1788 */
1789 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1790 dout("fsync waiting for flush_tid %u\n", flush_tid);
1791 ret = wait_event_interruptible(ci->i_cap_wq,
1792 caps_are_flushed(inode, flush_tid));
1793 }
1794
1795 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1796 return ret;
1797}
1798
1799/*
1800 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1801 * queue inode for flush but don't do so immediately, because we can
1802 * get by with fewer MDS messages if we wait for data writeback to
1803 * complete first.
1804 */
1805int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1806{
1807 struct ceph_inode_info *ci = ceph_inode(inode);
1808 unsigned flush_tid;
1809 int err = 0;
1810 int dirty;
1811 int wait = wbc->sync_mode == WB_SYNC_ALL;
1812
1813 dout("write_inode %p wait=%d\n", inode, wait);
1814 if (wait) {
1815 dirty = try_flush_caps(inode, NULL, &flush_tid);
1816 if (dirty)
1817 err = wait_event_interruptible(ci->i_cap_wq,
1818 caps_are_flushed(inode, flush_tid));
1819 } else {
1820 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1821
1822 spin_lock(&inode->i_lock);
1823 if (__ceph_caps_dirty(ci))
1824 __cap_delay_requeue_front(mdsc, ci);
1825 spin_unlock(&inode->i_lock);
1826 }
1827 return err;
1828}
1829
1830/*
1831 * After a recovering MDS goes active, we need to resend any caps
1832 * we were flushing.
1833 *
1834 * Caller holds session->s_mutex.
1835 */
1836static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1837 struct ceph_mds_session *session)
1838{
1839 struct ceph_cap_snap *capsnap;
1840
1841 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1842 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1843 flushing_item) {
1844 struct ceph_inode_info *ci = capsnap->ci;
1845 struct inode *inode = &ci->vfs_inode;
1846 struct ceph_cap *cap;
1847
1848 spin_lock(&inode->i_lock);
1849 cap = ci->i_auth_cap;
1850 if (cap && cap->session == session) {
1851 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1852 cap, capsnap);
1853 __ceph_flush_snaps(ci, &session);
1854 } else {
1855 pr_err("%p auth cap %p not mds%d ???\n", inode,
1856 cap, session->s_mds);
1857 spin_unlock(&inode->i_lock);
1858 }
1859 }
1860}
1861
1862void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1863 struct ceph_mds_session *session)
1864{
1865 struct ceph_inode_info *ci;
1866
1867 kick_flushing_capsnaps(mdsc, session);
1868
1869 dout("kick_flushing_caps mds%d\n", session->s_mds);
1870 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1871 struct inode *inode = &ci->vfs_inode;
1872 struct ceph_cap *cap;
1873 int delayed = 0;
1874
1875 spin_lock(&inode->i_lock);
1876 cap = ci->i_auth_cap;
1877 if (cap && cap->session == session) {
1878 dout("kick_flushing_caps %p cap %p %s\n", inode,
1879 cap, ceph_cap_string(ci->i_flushing_caps));
1880 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1881 __ceph_caps_used(ci),
1882 __ceph_caps_wanted(ci),
1883 cap->issued | cap->implemented,
1884 ci->i_flushing_caps, NULL);
1885 if (delayed) {
1886 spin_lock(&inode->i_lock);
1887 __cap_delay_requeue(mdsc, ci);
1888 spin_unlock(&inode->i_lock);
1889 }
1890 } else {
1891 pr_err("%p auth cap %p not mds%d ???\n", inode,
1892 cap, session->s_mds);
1893 spin_unlock(&inode->i_lock);
1894 }
1895 }
1896}
1897
1898
1899/*
1900 * Take references to capabilities we hold, so that we don't release
1901 * them to the MDS prematurely.
1902 *
1903 * Protected by i_lock.
1904 */
1905static void __take_cap_refs(struct ceph_inode_info *ci, int got)
1906{
1907 if (got & CEPH_CAP_PIN)
1908 ci->i_pin_ref++;
1909 if (got & CEPH_CAP_FILE_RD)
1910 ci->i_rd_ref++;
1911 if (got & CEPH_CAP_FILE_CACHE)
1912 ci->i_rdcache_ref++;
1913 if (got & CEPH_CAP_FILE_WR)
1914 ci->i_wr_ref++;
1915 if (got & CEPH_CAP_FILE_BUFFER) {
1916 if (ci->i_wrbuffer_ref == 0)
1917 igrab(&ci->vfs_inode);
1918 ci->i_wrbuffer_ref++;
1919 dout("__take_cap_refs %p wrbuffer %d -> %d (?)\n",
1920 &ci->vfs_inode, ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref);
1921 }
1922}
1923
1924/*
1925 * Try to grab cap references. Specify those refs we @want, and the
1926 * minimal set we @need. Also include the larger offset we are writing
1927 * to (when applicable), and check against max_size here as well.
1928 * Note that caller is responsible for ensuring max_size increases are
1929 * requested from the MDS.
1930 */
1931static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
1932 int *got, loff_t endoff, int *check_max, int *err)
1933{
1934 struct inode *inode = &ci->vfs_inode;
1935 int ret = 0;
1936 int have, implemented;
1937 int file_wanted;
1938
1939 dout("get_cap_refs %p need %s want %s\n", inode,
1940 ceph_cap_string(need), ceph_cap_string(want));
1941 spin_lock(&inode->i_lock);
1942
1943 /* make sure file is actually open */
1944 file_wanted = __ceph_caps_file_wanted(ci);
1945 if ((file_wanted & need) == 0) {
1946 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
1947 ceph_cap_string(need), ceph_cap_string(file_wanted));
1948 *err = -EBADF;
1949 ret = 1;
1950 goto out;
1951 }
1952
1953 if (need & CEPH_CAP_FILE_WR) {
1954 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
1955 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
1956 inode, endoff, ci->i_max_size);
1957 if (endoff > ci->i_wanted_max_size) {
1958 *check_max = 1;
1959 ret = 1;
1960 }
1961 goto out;
1962 }
1963 /*
1964 * If a sync write is in progress, we must wait, so that we
1965 * can get a final snapshot value for size+mtime.
1966 */
1967 if (__ceph_have_pending_cap_snap(ci)) {
1968 dout("get_cap_refs %p cap_snap_pending\n", inode);
1969 goto out;
1970 }
1971 }
1972 have = __ceph_caps_issued(ci, &implemented);
1973
1974 /*
1975 * disallow writes while a truncate is pending
1976 */
1977 if (ci->i_truncate_pending)
1978 have &= ~CEPH_CAP_FILE_WR;
1979
1980 if ((have & need) == need) {
1981 /*
1982 * Look at (implemented & ~have & not) so that we keep waiting
1983 * on transition from wanted -> needed caps. This is needed
1984 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
1985 * going before a prior buffered writeback happens.
1986 */
1987 int not = want & ~(have & need);
1988 int revoking = implemented & ~have;
1989 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
1990 inode, ceph_cap_string(have), ceph_cap_string(not),
1991 ceph_cap_string(revoking));
1992 if ((revoking & not) == 0) {
1993 *got = need | (have & want);
1994 __take_cap_refs(ci, *got);
1995 ret = 1;
1996 }
1997 } else {
1998 dout("get_cap_refs %p have %s needed %s\n", inode,
1999 ceph_cap_string(have), ceph_cap_string(need));
2000 }
2001out:
2002 spin_unlock(&inode->i_lock);
2003 dout("get_cap_refs %p ret %d got %s\n", inode,
2004 ret, ceph_cap_string(*got));
2005 return ret;
2006}
2007
2008/*
2009 * Check the offset we are writing up to against our current
2010 * max_size. If necessary, tell the MDS we want to write to
2011 * a larger offset.
2012 */
2013static void check_max_size(struct inode *inode, loff_t endoff)
2014{
2015 struct ceph_inode_info *ci = ceph_inode(inode);
2016 int check = 0;
2017
2018 /* do we need to explicitly request a larger max_size? */
2019 spin_lock(&inode->i_lock);
2020 if ((endoff >= ci->i_max_size ||
2021 endoff > (inode->i_size << 1)) &&
2022 endoff > ci->i_wanted_max_size) {
2023 dout("write %p at large endoff %llu, req max_size\n",
2024 inode, endoff);
2025 ci->i_wanted_max_size = endoff;
2026 check = 1;
2027 }
2028 spin_unlock(&inode->i_lock);
2029 if (check)
2030 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2031}
2032
2033/*
2034 * Wait for caps, and take cap references. If we can't get a WR cap
2035 * due to a small max_size, make sure we check_max_size (and possibly
2036 * ask the mds) so we don't get hung up indefinitely.
2037 */
2038int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
2039 loff_t endoff)
2040{
2041 int check_max, ret, err;
2042
2043retry:
2044 if (endoff > 0)
2045 check_max_size(&ci->vfs_inode, endoff);
2046 check_max = 0;
2047 err = 0;
2048 ret = wait_event_interruptible(ci->i_cap_wq,
2049 try_get_cap_refs(ci, need, want,
2050 got, endoff,
2051 &check_max, &err));
2052 if (err)
2053 ret = err;
2054 if (check_max)
2055 goto retry;
2056 return ret;
2057}
2058
2059/*
2060 * Take cap refs. Caller must already know we hold at least one ref
2061 * on the caps in question or we don't know this is safe.
2062 */
2063void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2064{
2065 spin_lock(&ci->vfs_inode.i_lock);
2066 __take_cap_refs(ci, caps);
2067 spin_unlock(&ci->vfs_inode.i_lock);
2068}
2069
2070/*
2071 * Release cap refs.
2072 *
2073 * If we released the last ref on any given cap, call ceph_check_caps
2074 * to release (or schedule a release).
2075 *
2076 * If we are releasing a WR cap (from a sync write), finalize any affected
2077 * cap_snap, and wake up any waiters.
2078 */
2079void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2080{
2081 struct inode *inode = &ci->vfs_inode;
2082 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2083 struct ceph_cap_snap *capsnap;
2084
2085 spin_lock(&inode->i_lock);
2086 if (had & CEPH_CAP_PIN)
2087 --ci->i_pin_ref;
2088 if (had & CEPH_CAP_FILE_RD)
2089 if (--ci->i_rd_ref == 0)
2090 last++;
2091 if (had & CEPH_CAP_FILE_CACHE)
2092 if (--ci->i_rdcache_ref == 0)
2093 last++;
2094 if (had & CEPH_CAP_FILE_BUFFER) {
2095 if (--ci->i_wrbuffer_ref == 0) {
2096 last++;
2097 put++;
2098 }
2099 dout("put_cap_refs %p wrbuffer %d -> %d (?)\n",
2100 inode, ci->i_wrbuffer_ref+1, ci->i_wrbuffer_ref);
2101 }
2102 if (had & CEPH_CAP_FILE_WR)
2103 if (--ci->i_wr_ref == 0) {
2104 last++;
2105 if (!list_empty(&ci->i_cap_snaps)) {
2106 capsnap = list_first_entry(&ci->i_cap_snaps,
2107 struct ceph_cap_snap,
2108 ci_item);
2109 if (capsnap->writing) {
2110 capsnap->writing = 0;
2111 flushsnaps =
2112 __ceph_finish_cap_snap(ci,
2113 capsnap);
2114 wake = 1;
2115 }
2116 }
2117 }
2118 spin_unlock(&inode->i_lock);
2119
2120 dout("put_cap_refs %p had %s %s\n", inode, ceph_cap_string(had),
2121 last ? "last" : "");
2122
2123 if (last && !flushsnaps)
2124 ceph_check_caps(ci, 0, NULL);
2125 else if (flushsnaps)
2126 ceph_flush_snaps(ci);
2127 if (wake)
2128 wake_up(&ci->i_cap_wq);
2129 if (put)
2130 iput(inode);
2131}
2132
2133/*
2134 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2135 * context. Adjust per-snap dirty page accounting as appropriate.
2136 * Once all dirty data for a cap_snap is flushed, flush snapped file
2137 * metadata back to the MDS. If we dropped the last ref, call
2138 * ceph_check_caps.
2139 */
2140void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2141 struct ceph_snap_context *snapc)
2142{
2143 struct inode *inode = &ci->vfs_inode;
2144 int last = 0;
2145 int last_snap = 0;
2146 int found = 0;
2147 struct ceph_cap_snap *capsnap = NULL;
2148
2149 spin_lock(&inode->i_lock);
2150 ci->i_wrbuffer_ref -= nr;
2151 last = !ci->i_wrbuffer_ref;
2152
2153 if (ci->i_head_snapc == snapc) {
2154 ci->i_wrbuffer_ref_head -= nr;
2155 if (!ci->i_wrbuffer_ref_head) {
2156 ceph_put_snap_context(ci->i_head_snapc);
2157 ci->i_head_snapc = NULL;
2158 }
2159 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2160 inode,
2161 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2162 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2163 last ? " LAST" : "");
2164 } else {
2165 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2166 if (capsnap->context == snapc) {
2167 found = 1;
2168 capsnap->dirty_pages -= nr;
2169 last_snap = !capsnap->dirty_pages;
2170 break;
2171 }
2172 }
2173 BUG_ON(!found);
2174 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2175 " snap %lld %d/%d -> %d/%d %s%s\n",
2176 inode, capsnap, capsnap->context->seq,
2177 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2178 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2179 last ? " (wrbuffer last)" : "",
2180 last_snap ? " (capsnap last)" : "");
2181 }
2182
2183 spin_unlock(&inode->i_lock);
2184
2185 if (last) {
2186 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2187 iput(inode);
2188 } else if (last_snap) {
2189 ceph_flush_snaps(ci);
2190 wake_up(&ci->i_cap_wq);
2191 }
2192}
2193
2194/*
2195 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2196 * actually be a revocation if it specifies a smaller cap set.)
2197 *
2198 * caller holds s_mutex.
2199 * return value:
2200 * 0 - ok
2201 * 1 - check_caps on auth cap only (writeback)
2202 * 2 - check_caps (ack revoke)
2203 */
2204static int handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2205 struct ceph_mds_session *session,
2206 struct ceph_cap *cap,
2207 struct ceph_buffer *xattr_buf)
2208 __releases(inode->i_lock)
2209
2210{
2211 struct ceph_inode_info *ci = ceph_inode(inode);
2212 int mds = session->s_mds;
2213 int seq = le32_to_cpu(grant->seq);
2214 int newcaps = le32_to_cpu(grant->caps);
2215 int issued, implemented, used, wanted, dirty;
2216 u64 size = le64_to_cpu(grant->size);
2217 u64 max_size = le64_to_cpu(grant->max_size);
2218 struct timespec mtime, atime, ctime;
2219 int reply = 0;
2220 int wake = 0;
2221 int writeback = 0;
2222 int revoked_rdcache = 0;
2223 int queue_invalidate = 0;
2224
2225 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2226 inode, cap, mds, seq, ceph_cap_string(newcaps));
2227 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2228 inode->i_size);
2229
2230 /*
2231 * If CACHE is being revoked, and we have no dirty buffers,
2232 * try to invalidate (once). (If there are dirty buffers, we
2233 * will invalidate _after_ writeback.)
2234 */
2235 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2236 !ci->i_wrbuffer_ref) {
2237 if (try_nonblocking_invalidate(inode) == 0) {
2238 revoked_rdcache = 1;
2239 } else {
2240 /* there were locked pages.. invalidate later
2241 in a separate thread. */
2242 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2243 queue_invalidate = 1;
2244 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2245 }
2246 }
2247 }
2248
2249 /* side effects now are allowed */
2250
2251 issued = __ceph_caps_issued(ci, &implemented);
2252 issued |= implemented | __ceph_caps_dirty(ci);
2253
2254 cap->cap_gen = session->s_cap_gen;
2255
2256 __check_cap_issue(ci, cap, newcaps);
2257
2258 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2259 inode->i_mode = le32_to_cpu(grant->mode);
2260 inode->i_uid = le32_to_cpu(grant->uid);
2261 inode->i_gid = le32_to_cpu(grant->gid);
2262 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2263 inode->i_uid, inode->i_gid);
2264 }
2265
2266 if ((issued & CEPH_CAP_LINK_EXCL) == 0)
2267 inode->i_nlink = le32_to_cpu(grant->nlink);
2268
2269 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2270 int len = le32_to_cpu(grant->xattr_len);
2271 u64 version = le64_to_cpu(grant->xattr_version);
2272
2273 if (version > ci->i_xattrs.version) {
2274 dout(" got new xattrs v%llu on %p len %d\n",
2275 version, inode, len);
2276 if (ci->i_xattrs.blob)
2277 ceph_buffer_put(ci->i_xattrs.blob);
2278 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2279 ci->i_xattrs.version = version;
2280 }
2281 }
2282
2283 /* size/ctime/mtime/atime? */
2284 ceph_fill_file_size(inode, issued,
2285 le32_to_cpu(grant->truncate_seq),
2286 le64_to_cpu(grant->truncate_size), size);
2287 ceph_decode_timespec(&mtime, &grant->mtime);
2288 ceph_decode_timespec(&atime, &grant->atime);
2289 ceph_decode_timespec(&ctime, &grant->ctime);
2290 ceph_fill_file_time(inode, issued,
2291 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2292 &atime);
2293
2294 /* max size increase? */
2295 if (max_size != ci->i_max_size) {
2296 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2297 ci->i_max_size = max_size;
2298 if (max_size >= ci->i_wanted_max_size) {
2299 ci->i_wanted_max_size = 0; /* reset */
2300 ci->i_requested_max_size = 0;
2301 }
2302 wake = 1;
2303 }
2304
2305 /* check cap bits */
2306 wanted = __ceph_caps_wanted(ci);
2307 used = __ceph_caps_used(ci);
2308 dirty = __ceph_caps_dirty(ci);
2309 dout(" my wanted = %s, used = %s, dirty %s\n",
2310 ceph_cap_string(wanted),
2311 ceph_cap_string(used),
2312 ceph_cap_string(dirty));
2313 if (wanted != le32_to_cpu(grant->wanted)) {
2314 dout("mds wanted %s -> %s\n",
2315 ceph_cap_string(le32_to_cpu(grant->wanted)),
2316 ceph_cap_string(wanted));
2317 grant->wanted = cpu_to_le32(wanted);
2318 }
2319
2320 cap->seq = seq;
2321
2322 /* file layout may have changed */
2323 ci->i_layout = grant->layout;
2324
2325 /* revocation, grant, or no-op? */
2326 if (cap->issued & ~newcaps) {
2327 dout("revocation: %s -> %s\n", ceph_cap_string(cap->issued),
2328 ceph_cap_string(newcaps));
2329 if ((used & ~newcaps) & CEPH_CAP_FILE_BUFFER)
2330 writeback = 1; /* will delay ack */
2331 else if (dirty & ~newcaps)
2332 reply = 1; /* initiate writeback in check_caps */
2333 else if (((used & ~newcaps) & CEPH_CAP_FILE_CACHE) == 0 ||
2334 revoked_rdcache)
2335 reply = 2; /* send revoke ack in check_caps */
2336 cap->issued = newcaps;
2337 } else if (cap->issued == newcaps) {
2338 dout("caps unchanged: %s -> %s\n",
2339 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2340 } else {
2341 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2342 ceph_cap_string(newcaps));
2343 cap->issued = newcaps;
2344 cap->implemented |= newcaps; /* add bits only, to
2345 * avoid stepping on a
2346 * pending revocation */
2347 wake = 1;
2348 }
2349
2350 spin_unlock(&inode->i_lock);
2351 if (writeback)
2352 /*
2353 * queue inode for writeback: we can't actually call
2354 * filemap_write_and_wait, etc. from message handler
2355 * context.
2356 */
2357 ceph_queue_writeback(inode);
2358 if (queue_invalidate)
2359 ceph_queue_invalidate(inode);
2360 if (wake)
2361 wake_up(&ci->i_cap_wq);
2362 return reply;
2363}
2364
2365/*
2366 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2367 * MDS has been safely committed.
2368 */
2369static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2370 struct ceph_mds_caps *m,
2371 struct ceph_mds_session *session,
2372 struct ceph_cap *cap)
2373 __releases(inode->i_lock)
2374{
2375 struct ceph_inode_info *ci = ceph_inode(inode);
2376 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
2377 unsigned seq = le32_to_cpu(m->seq);
2378 int dirty = le32_to_cpu(m->dirty);
2379 int cleaned = 0;
2380 int drop = 0;
2381 int i;
2382
2383 for (i = 0; i < CEPH_CAP_BITS; i++)
2384 if ((dirty & (1 << i)) &&
2385 flush_tid == ci->i_cap_flush_tid[i])
2386 cleaned |= 1 << i;
2387
2388 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2389 " flushing %s -> %s\n",
2390 inode, session->s_mds, seq, ceph_cap_string(dirty),
2391 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2392 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2393
2394 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2395 goto out;
2396
2397 ci->i_flushing_caps &= ~cleaned;
2398
2399 spin_lock(&mdsc->cap_dirty_lock);
2400 if (ci->i_flushing_caps == 0) {
2401 list_del_init(&ci->i_flushing_item);
2402 if (!list_empty(&session->s_cap_flushing))
2403 dout(" mds%d still flushing cap on %p\n",
2404 session->s_mds,
2405 &list_entry(session->s_cap_flushing.next,
2406 struct ceph_inode_info,
2407 i_flushing_item)->vfs_inode);
2408 mdsc->num_cap_flushing--;
2409 wake_up(&mdsc->cap_flushing_wq);
2410 dout(" inode %p now !flushing\n", inode);
2411
2412 if (ci->i_dirty_caps == 0) {
2413 dout(" inode %p now clean\n", inode);
2414 BUG_ON(!list_empty(&ci->i_dirty_item));
2415 drop = 1;
2416 } else {
2417 BUG_ON(list_empty(&ci->i_dirty_item));
2418 }
2419 }
2420 spin_unlock(&mdsc->cap_dirty_lock);
2421 wake_up(&ci->i_cap_wq);
2422
2423out:
2424 spin_unlock(&inode->i_lock);
2425 if (drop)
2426 iput(inode);
2427}
2428
2429/*
2430 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2431 * throw away our cap_snap.
2432 *
2433 * Caller hold s_mutex.
2434 */
2435static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2436 struct ceph_mds_caps *m,
2437 struct ceph_mds_session *session)
2438{
2439 struct ceph_inode_info *ci = ceph_inode(inode);
2440 u64 follows = le64_to_cpu(m->snap_follows);
2441 struct ceph_cap_snap *capsnap;
2442 int drop = 0;
2443
2444 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2445 inode, ci, session->s_mds, follows);
2446
2447 spin_lock(&inode->i_lock);
2448 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2449 if (capsnap->follows == follows) {
2450 if (capsnap->flush_tid != flush_tid) {
2451 dout(" cap_snap %p follows %lld tid %lld !="
2452 " %lld\n", capsnap, follows,
2453 flush_tid, capsnap->flush_tid);
2454 break;
2455 }
2456 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2457 dout(" removing cap_snap %p follows %lld\n",
2458 capsnap, follows);
2459 ceph_put_snap_context(capsnap->context);
2460 list_del(&capsnap->ci_item);
2461 list_del(&capsnap->flushing_item);
2462 ceph_put_cap_snap(capsnap);
2463 drop = 1;
2464 break;
2465 } else {
2466 dout(" skipping cap_snap %p follows %lld\n",
2467 capsnap, capsnap->follows);
2468 }
2469 }
2470 spin_unlock(&inode->i_lock);
2471 if (drop)
2472 iput(inode);
2473}
2474
2475/*
2476 * Handle TRUNC from MDS, indicating file truncation.
2477 *
2478 * caller hold s_mutex.
2479 */
2480static void handle_cap_trunc(struct inode *inode,
2481 struct ceph_mds_caps *trunc,
2482 struct ceph_mds_session *session)
2483 __releases(inode->i_lock)
2484{
2485 struct ceph_inode_info *ci = ceph_inode(inode);
2486 int mds = session->s_mds;
2487 int seq = le32_to_cpu(trunc->seq);
2488 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2489 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2490 u64 size = le64_to_cpu(trunc->size);
2491 int implemented = 0;
2492 int dirty = __ceph_caps_dirty(ci);
2493 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2494 int queue_trunc = 0;
2495
2496 issued |= implemented | dirty;
2497
2498 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2499 inode, mds, seq, truncate_size, truncate_seq);
2500 queue_trunc = ceph_fill_file_size(inode, issued,
2501 truncate_seq, truncate_size, size);
2502 spin_unlock(&inode->i_lock);
2503
2504 if (queue_trunc)
2505 ceph_queue_vmtruncate(inode);
2506}
2507
2508/*
2509 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2510 * different one. If we are the most recent migration we've seen (as
2511 * indicated by mseq), make note of the migrating cap bits for the
2512 * duration (until we see the corresponding IMPORT).
2513 *
2514 * caller holds s_mutex
2515 */
2516static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2517 struct ceph_mds_session *session)
2518{
2519 struct ceph_inode_info *ci = ceph_inode(inode);
2520 int mds = session->s_mds;
2521 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2522 struct ceph_cap *cap = NULL, *t;
2523 struct rb_node *p;
2524 int remember = 1;
2525
2526 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n",
2527 inode, ci, mds, mseq);
2528
2529 spin_lock(&inode->i_lock);
2530
2531 /* make sure we haven't seen a higher mseq */
2532 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2533 t = rb_entry(p, struct ceph_cap, ci_node);
2534 if (ceph_seq_cmp(t->mseq, mseq) > 0) {
2535 dout(" higher mseq on cap from mds%d\n",
2536 t->session->s_mds);
2537 remember = 0;
2538 }
2539 if (t->session->s_mds == mds)
2540 cap = t;
2541 }
2542
2543 if (cap) {
2544 if (remember) {
2545 /* make note */
2546 ci->i_cap_exporting_mds = mds;
2547 ci->i_cap_exporting_mseq = mseq;
2548 ci->i_cap_exporting_issued = cap->issued;
2549 }
2550 __ceph_remove_cap(cap);
2551 } else {
2552 WARN_ON(!cap);
2553 }
2554
2555 spin_unlock(&inode->i_lock);
2556}
2557
2558/*
2559 * Handle cap IMPORT. If there are temp bits from an older EXPORT,
2560 * clean them up.
2561 *
2562 * caller holds s_mutex.
2563 */
2564static void handle_cap_import(struct ceph_mds_client *mdsc,
2565 struct inode *inode, struct ceph_mds_caps *im,
2566 struct ceph_mds_session *session,
2567 void *snaptrace, int snaptrace_len)
2568{
2569 struct ceph_inode_info *ci = ceph_inode(inode);
2570 int mds = session->s_mds;
2571 unsigned issued = le32_to_cpu(im->caps);
2572 unsigned wanted = le32_to_cpu(im->wanted);
2573 unsigned seq = le32_to_cpu(im->seq);
2574 unsigned mseq = le32_to_cpu(im->migrate_seq);
2575 u64 realmino = le64_to_cpu(im->realm);
2576 u64 cap_id = le64_to_cpu(im->cap_id);
2577
2578 if (ci->i_cap_exporting_mds >= 0 &&
2579 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) {
2580 dout("handle_cap_import inode %p ci %p mds%d mseq %d"
2581 " - cleared exporting from mds%d\n",
2582 inode, ci, mds, mseq,
2583 ci->i_cap_exporting_mds);
2584 ci->i_cap_exporting_issued = 0;
2585 ci->i_cap_exporting_mseq = 0;
2586 ci->i_cap_exporting_mds = -1;
2587 } else {
2588 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
2589 inode, ci, mds, mseq);
2590 }
2591
2592 down_write(&mdsc->snap_rwsem);
2593 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2594 false);
2595 downgrade_write(&mdsc->snap_rwsem);
2596 ceph_add_cap(inode, session, cap_id, -1,
2597 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2598 NULL /* no caps context */);
2599 try_flush_caps(inode, session, NULL);
2600 up_read(&mdsc->snap_rwsem);
2601}
2602
2603/*
2604 * Handle a caps message from the MDS.
2605 *
2606 * Identify the appropriate session, inode, and call the right handler
2607 * based on the cap op.
2608 */
2609void ceph_handle_caps(struct ceph_mds_session *session,
2610 struct ceph_msg *msg)
2611{
2612 struct ceph_mds_client *mdsc = session->s_mdsc;
2613 struct super_block *sb = mdsc->client->sb;
2614 struct inode *inode;
2615 struct ceph_cap *cap;
2616 struct ceph_mds_caps *h;
2617 int mds = session->s_mds;
2618 int op;
2619 u32 seq;
2620 struct ceph_vino vino;
2621 u64 cap_id;
2622 u64 size, max_size;
2623 u64 tid;
2624 int check_caps = 0;
2625 void *snaptrace;
2626 int r;
2627
2628 dout("handle_caps from mds%d\n", mds);
2629
2630 /* decode */
2631 tid = le64_to_cpu(msg->hdr.tid);
2632 if (msg->front.iov_len < sizeof(*h))
2633 goto bad;
2634 h = msg->front.iov_base;
2635 snaptrace = h + 1;
2636 op = le32_to_cpu(h->op);
2637 vino.ino = le64_to_cpu(h->ino);
2638 vino.snap = CEPH_NOSNAP;
2639 cap_id = le64_to_cpu(h->cap_id);
2640 seq = le32_to_cpu(h->seq);
2641 size = le64_to_cpu(h->size);
2642 max_size = le64_to_cpu(h->max_size);
2643
2644 mutex_lock(&session->s_mutex);
2645 session->s_seq++;
2646 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
2647 (unsigned)seq);
2648
2649 /* lookup ino */
2650 inode = ceph_find_inode(sb, vino);
2651 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
2652 vino.snap, inode);
2653 if (!inode) {
2654 dout(" i don't have ino %llx\n", vino.ino);
2655 goto done;
2656 }
2657
2658 /* these will work even if we don't have a cap yet */
2659 switch (op) {
2660 case CEPH_CAP_OP_FLUSHSNAP_ACK:
2661 handle_cap_flushsnap_ack(inode, tid, h, session);
2662 goto done;
2663
2664 case CEPH_CAP_OP_EXPORT:
2665 handle_cap_export(inode, h, session);
2666 goto done;
2667
2668 case CEPH_CAP_OP_IMPORT:
2669 handle_cap_import(mdsc, inode, h, session,
2670 snaptrace, le32_to_cpu(h->snap_trace_len));
2671 check_caps = 1; /* we may have sent a RELEASE to the old auth */
2672 goto done;
2673 }
2674
2675 /* the rest require a cap */
2676 spin_lock(&inode->i_lock);
2677 cap = __get_cap_for_mds(ceph_inode(inode), mds);
2678 if (!cap) {
2679 dout("no cap on %p ino %llx.%llx from mds%d, releasing\n",
2680 inode, ceph_ino(inode), ceph_snap(inode), mds);
2681 spin_unlock(&inode->i_lock);
2682 goto done;
2683 }
2684
2685 /* note that each of these drops i_lock for us */
2686 switch (op) {
2687 case CEPH_CAP_OP_REVOKE:
2688 case CEPH_CAP_OP_GRANT:
2689 r = handle_cap_grant(inode, h, session, cap, msg->middle);
2690 if (r == 1)
2691 ceph_check_caps(ceph_inode(inode),
2692 CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2693 session);
2694 else if (r == 2)
2695 ceph_check_caps(ceph_inode(inode),
2696 CHECK_CAPS_NODELAY,
2697 session);
2698 break;
2699
2700 case CEPH_CAP_OP_FLUSH_ACK:
2701 handle_cap_flush_ack(inode, tid, h, session, cap);
2702 break;
2703
2704 case CEPH_CAP_OP_TRUNC:
2705 handle_cap_trunc(inode, h, session);
2706 break;
2707
2708 default:
2709 spin_unlock(&inode->i_lock);
2710 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
2711 ceph_cap_op_name(op));
2712 }
2713
2714done:
2715 mutex_unlock(&session->s_mutex);
2716
2717 if (check_caps)
2718 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_NODELAY, NULL);
2719 if (inode)
2720 iput(inode);
2721 return;
2722
2723bad:
2724 pr_err("ceph_handle_caps: corrupt message\n");
2725 ceph_msg_dump(msg);
2726 return;
2727}
2728
2729/*
2730 * Delayed work handler to process end of delayed cap release LRU list.
2731 */
2732void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
2733{
2734 struct ceph_inode_info *ci;
2735 int flags = CHECK_CAPS_NODELAY;
2736
2737 dout("check_delayed_caps\n");
2738 while (1) {
2739 spin_lock(&mdsc->cap_delay_lock);
2740 if (list_empty(&mdsc->cap_delay_list))
2741 break;
2742 ci = list_first_entry(&mdsc->cap_delay_list,
2743 struct ceph_inode_info,
2744 i_cap_delay_list);
2745 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
2746 time_before(jiffies, ci->i_hold_caps_max))
2747 break;
2748 list_del_init(&ci->i_cap_delay_list);
2749 spin_unlock(&mdsc->cap_delay_lock);
2750 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
2751 ceph_check_caps(ci, flags, NULL);
2752 }
2753 spin_unlock(&mdsc->cap_delay_lock);
2754}
2755
2756/*
2757 * Flush all dirty caps to the mds
2758 */
2759void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
2760{
2761 struct ceph_inode_info *ci, *nci = NULL;
2762 struct inode *inode, *ninode = NULL;
2763 struct list_head *p, *n;
2764
2765 dout("flush_dirty_caps\n");
2766 spin_lock(&mdsc->cap_dirty_lock);
2767 list_for_each_safe(p, n, &mdsc->cap_dirty) {
2768 if (nci) {
2769 ci = nci;
2770 inode = ninode;
2771 ci->i_ceph_flags &= ~CEPH_I_NOFLUSH;
2772 dout("flush_dirty_caps inode %p (was next inode)\n",
2773 inode);
2774 } else {
2775 ci = list_entry(p, struct ceph_inode_info,
2776 i_dirty_item);
2777 inode = igrab(&ci->vfs_inode);
2778 BUG_ON(!inode);
2779 dout("flush_dirty_caps inode %p\n", inode);
2780 }
2781 if (n != &mdsc->cap_dirty) {
2782 nci = list_entry(n, struct ceph_inode_info,
2783 i_dirty_item);
2784 ninode = igrab(&nci->vfs_inode);
2785 BUG_ON(!ninode);
2786 nci->i_ceph_flags |= CEPH_I_NOFLUSH;
2787 dout("flush_dirty_caps next inode %p, noflush\n",
2788 ninode);
2789 } else {
2790 nci = NULL;
2791 ninode = NULL;
2792 }
2793 spin_unlock(&mdsc->cap_dirty_lock);
2794 if (inode) {
2795 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH,
2796 NULL);
2797 iput(inode);
2798 }
2799 spin_lock(&mdsc->cap_dirty_lock);
2800 }
2801 spin_unlock(&mdsc->cap_dirty_lock);
2802}
2803
2804/*
2805 * Drop open file reference. If we were the last open file,
2806 * we may need to release capabilities to the MDS (or schedule
2807 * their delayed release).
2808 */
2809void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
2810{
2811 struct inode *inode = &ci->vfs_inode;
2812 int last = 0;
2813
2814 spin_lock(&inode->i_lock);
2815 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
2816 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
2817 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
2818 if (--ci->i_nr_by_mode[fmode] == 0)
2819 last++;
2820 spin_unlock(&inode->i_lock);
2821
2822 if (last && ci->i_vino.snap == CEPH_NOSNAP)
2823 ceph_check_caps(ci, 0, NULL);
2824}
2825
2826/*
2827 * Helpers for embedding cap and dentry lease releases into mds
2828 * requests.
2829 *
2830 * @force is used by dentry_release (below) to force inclusion of a
2831 * record for the directory inode, even when there aren't any caps to
2832 * drop.
2833 */
2834int ceph_encode_inode_release(void **p, struct inode *inode,
2835 int mds, int drop, int unless, int force)
2836{
2837 struct ceph_inode_info *ci = ceph_inode(inode);
2838 struct ceph_cap *cap;
2839 struct ceph_mds_request_release *rel = *p;
2840 int ret = 0;
2841
2842 dout("encode_inode_release %p mds%d drop %s unless %s\n", inode,
2843 mds, ceph_cap_string(drop), ceph_cap_string(unless));
2844
2845 spin_lock(&inode->i_lock);
2846 cap = __get_cap_for_mds(ci, mds);
2847 if (cap && __cap_is_valid(cap)) {
2848 if (force ||
2849 ((cap->issued & drop) &&
2850 (cap->issued & unless) == 0)) {
2851 if ((cap->issued & drop) &&
2852 (cap->issued & unless) == 0) {
2853 dout("encode_inode_release %p cap %p %s -> "
2854 "%s\n", inode, cap,
2855 ceph_cap_string(cap->issued),
2856 ceph_cap_string(cap->issued & ~drop));
2857 cap->issued &= ~drop;
2858 cap->implemented &= ~drop;
2859 if (ci->i_ceph_flags & CEPH_I_NODELAY) {
2860 int wanted = __ceph_caps_wanted(ci);
2861 dout(" wanted %s -> %s (act %s)\n",
2862 ceph_cap_string(cap->mds_wanted),
2863 ceph_cap_string(cap->mds_wanted &
2864 ~wanted),
2865 ceph_cap_string(wanted));
2866 cap->mds_wanted &= wanted;
2867 }
2868 } else {
2869 dout("encode_inode_release %p cap %p %s"
2870 " (force)\n", inode, cap,
2871 ceph_cap_string(cap->issued));
2872 }
2873
2874 rel->ino = cpu_to_le64(ceph_ino(inode));
2875 rel->cap_id = cpu_to_le64(cap->cap_id);
2876 rel->seq = cpu_to_le32(cap->seq);
2877 rel->issue_seq = cpu_to_le32(cap->issue_seq),
2878 rel->mseq = cpu_to_le32(cap->mseq);
2879 rel->caps = cpu_to_le32(cap->issued);
2880 rel->wanted = cpu_to_le32(cap->mds_wanted);
2881 rel->dname_len = 0;
2882 rel->dname_seq = 0;
2883 *p += sizeof(*rel);
2884 ret = 1;
2885 } else {
2886 dout("encode_inode_release %p cap %p %s\n",
2887 inode, cap, ceph_cap_string(cap->issued));
2888 }
2889 }
2890 spin_unlock(&inode->i_lock);
2891 return ret;
2892}
2893
2894int ceph_encode_dentry_release(void **p, struct dentry *dentry,
2895 int mds, int drop, int unless)
2896{
2897 struct inode *dir = dentry->d_parent->d_inode;
2898 struct ceph_mds_request_release *rel = *p;
2899 struct ceph_dentry_info *di = ceph_dentry(dentry);
2900 int force = 0;
2901 int ret;
2902
2903 /*
2904 * force an record for the directory caps if we have a dentry lease.
2905 * this is racy (can't take i_lock and d_lock together), but it
2906 * doesn't have to be perfect; the mds will revoke anything we don't
2907 * release.
2908 */
2909 spin_lock(&dentry->d_lock);
2910 if (di->lease_session && di->lease_session->s_mds == mds)
2911 force = 1;
2912 spin_unlock(&dentry->d_lock);
2913
2914 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
2915
2916 spin_lock(&dentry->d_lock);
2917 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
2918 dout("encode_dentry_release %p mds%d seq %d\n",
2919 dentry, mds, (int)di->lease_seq);
2920 rel->dname_len = cpu_to_le32(dentry->d_name.len);
2921 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
2922 *p += dentry->d_name.len;
2923 rel->dname_seq = cpu_to_le32(di->lease_seq);
2924 }
2925 spin_unlock(&dentry->d_lock);
2926 return ret;
2927}
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-23 13:15:16 -0500