1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3
4 #include <linux/fs.h>
5 #include <linux/sort.h>
6 #include <linux/slab.h>
7 #include <linux/iversion.h>
8 #include "super.h"
9 #include "mds_client.h"
10 #include <linux/ceph/decode.h>
11
12 /* unused map expires after 5 minutes */
13 #define CEPH_SNAPID_MAP_TIMEOUT (5 * 60 * HZ)
14
15 /*
16 * Snapshots in ceph are driven in large part by cooperation from the
17 * client. In contrast to local file systems or file servers that
18 * implement snapshots at a single point in the system, ceph's
19 * distributed access to storage requires clients to help decide
20 * whether a write logically occurs before or after a recently created
21 * snapshot.
22 *
23 * This provides a perfect instantanous client-wide snapshot. Between
24 * clients, however, snapshots may appear to be applied at slightly
25 * different points in time, depending on delays in delivering the
26 * snapshot notification.
27 *
28 * Snapshots are _not_ file system-wide. Instead, each snapshot
29 * applies to the subdirectory nested beneath some directory. This
30 * effectively divides the hierarchy into multiple "realms," where all
31 * of the files contained by each realm share the same set of
32 * snapshots. An individual realm's snap set contains snapshots
33 * explicitly created on that realm, as well as any snaps in its
34 * parent's snap set _after_ the point at which the parent became it's
35 * parent (due to, say, a rename). Similarly, snaps from prior parents
36 * during the time intervals during which they were the parent are included.
37 *
38 * The client is spared most of this detail, fortunately... it must only
39 * maintains a hierarchy of realms reflecting the current parent/child
40 * realm relationship, and for each realm has an explicit list of snaps
41 * inherited from prior parents.
42 *
43 * A snap_realm struct is maintained for realms containing every inode
44 * with an open cap in the system. (The needed snap realm information is
45 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq'
46 * version number is used to ensure that as realm parameters change (new
47 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
48 *
49 * The realm hierarchy drives the generation of a 'snap context' for each
50 * realm, which simply lists the resulting set of snaps for the realm. This
51 * is attached to any writes sent to OSDs.
52 */
53 /*
54 * Unfortunately error handling is a bit mixed here. If we get a snap
55 * update, but don't have enough memory to update our realm hierarchy,
56 * it's not clear what we can do about it (besides complaining to the
57 * console).
58 */
59
60
61 /*
62 * increase ref count for the realm
63 *
64 * caller must hold snap_rwsem.
65 */
ceph_get_snap_realm(struct ceph_mds_client * mdsc,struct ceph_snap_realm * realm)66 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
67 struct ceph_snap_realm *realm)
68 {
69 lockdep_assert_held(&mdsc->snap_rwsem);
70
71 /*
72 * The 0->1 and 1->0 transitions must take the snap_empty_lock
73 * atomically with the refcount change. Go ahead and bump the
74 * nref here, unless it's 0, in which case we take the spinlock
75 * and then do the increment and remove it from the list.
76 */
77 if (atomic_inc_not_zero(&realm->nref))
78 return;
79
80 spin_lock(&mdsc->snap_empty_lock);
81 if (atomic_inc_return(&realm->nref) == 1)
82 list_del_init(&realm->empty_item);
83 spin_unlock(&mdsc->snap_empty_lock);
84 }
85
__insert_snap_realm(struct rb_root * root,struct ceph_snap_realm * new)86 static void __insert_snap_realm(struct rb_root *root,
87 struct ceph_snap_realm *new)
88 {
89 struct rb_node **p = &root->rb_node;
90 struct rb_node *parent = NULL;
91 struct ceph_snap_realm *r = NULL;
92
93 while (*p) {
94 parent = *p;
95 r = rb_entry(parent, struct ceph_snap_realm, node);
96 if (new->ino < r->ino)
97 p = &(*p)->rb_left;
98 else if (new->ino > r->ino)
99 p = &(*p)->rb_right;
100 else
101 BUG();
102 }
103
104 rb_link_node(&new->node, parent, p);
105 rb_insert_color(&new->node, root);
106 }
107
108 /*
109 * create and get the realm rooted at @ino and bump its ref count.
110 *
111 * caller must hold snap_rwsem for write.
112 */
ceph_create_snap_realm(struct ceph_mds_client * mdsc,u64 ino)113 static struct ceph_snap_realm *ceph_create_snap_realm(
114 struct ceph_mds_client *mdsc,
115 u64 ino)
116 {
117 struct ceph_snap_realm *realm;
118
119 lockdep_assert_held_write(&mdsc->snap_rwsem);
120
121 realm = kzalloc(sizeof(*realm), GFP_NOFS);
122 if (!realm)
123 return ERR_PTR(-ENOMEM);
124
125 /* Do not release the global dummy snaprealm until unmouting */
126 if (ino == CEPH_INO_GLOBAL_SNAPREALM)
127 atomic_set(&realm->nref, 2);
128 else
129 atomic_set(&realm->nref, 1);
130 realm->ino = ino;
131 INIT_LIST_HEAD(&realm->children);
132 INIT_LIST_HEAD(&realm->child_item);
133 INIT_LIST_HEAD(&realm->empty_item);
134 INIT_LIST_HEAD(&realm->dirty_item);
135 INIT_LIST_HEAD(&realm->rebuild_item);
136 INIT_LIST_HEAD(&realm->inodes_with_caps);
137 spin_lock_init(&realm->inodes_with_caps_lock);
138 __insert_snap_realm(&mdsc->snap_realms, realm);
139 mdsc->num_snap_realms++;
140
141 dout("%s %llx %p\n", __func__, realm->ino, realm);
142 return realm;
143 }
144
145 /*
146 * lookup the realm rooted at @ino.
147 *
148 * caller must hold snap_rwsem.
149 */
__lookup_snap_realm(struct ceph_mds_client * mdsc,u64 ino)150 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
151 u64 ino)
152 {
153 struct rb_node *n = mdsc->snap_realms.rb_node;
154 struct ceph_snap_realm *r;
155
156 lockdep_assert_held(&mdsc->snap_rwsem);
157
158 while (n) {
159 r = rb_entry(n, struct ceph_snap_realm, node);
160 if (ino < r->ino)
161 n = n->rb_left;
162 else if (ino > r->ino)
163 n = n->rb_right;
164 else {
165 dout("%s %llx %p\n", __func__, r->ino, r);
166 return r;
167 }
168 }
169 return NULL;
170 }
171
ceph_lookup_snap_realm(struct ceph_mds_client * mdsc,u64 ino)172 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
173 u64 ino)
174 {
175 struct ceph_snap_realm *r;
176 r = __lookup_snap_realm(mdsc, ino);
177 if (r)
178 ceph_get_snap_realm(mdsc, r);
179 return r;
180 }
181
182 static void __put_snap_realm(struct ceph_mds_client *mdsc,
183 struct ceph_snap_realm *realm);
184
185 /*
186 * called with snap_rwsem (write)
187 */
__destroy_snap_realm(struct ceph_mds_client * mdsc,struct ceph_snap_realm * realm)188 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
189 struct ceph_snap_realm *realm)
190 {
191 lockdep_assert_held_write(&mdsc->snap_rwsem);
192
193 dout("%s %p %llx\n", __func__, realm, realm->ino);
194
195 rb_erase(&realm->node, &mdsc->snap_realms);
196 mdsc->num_snap_realms--;
197
198 if (realm->parent) {
199 list_del_init(&realm->child_item);
200 __put_snap_realm(mdsc, realm->parent);
201 }
202
203 kfree(realm->prior_parent_snaps);
204 kfree(realm->snaps);
205 ceph_put_snap_context(realm->cached_context);
206 kfree(realm);
207 }
208
209 /*
210 * caller holds snap_rwsem (write)
211 */
__put_snap_realm(struct ceph_mds_client * mdsc,struct ceph_snap_realm * realm)212 static void __put_snap_realm(struct ceph_mds_client *mdsc,
213 struct ceph_snap_realm *realm)
214 {
215 lockdep_assert_held_write(&mdsc->snap_rwsem);
216
217 /*
218 * We do not require the snap_empty_lock here, as any caller that
219 * increments the value must hold the snap_rwsem.
220 */
221 if (atomic_dec_and_test(&realm->nref))
222 __destroy_snap_realm(mdsc, realm);
223 }
224
225 /*
226 * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
227 */
ceph_put_snap_realm(struct ceph_mds_client * mdsc,struct ceph_snap_realm * realm)228 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
229 struct ceph_snap_realm *realm)
230 {
231 if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
232 return;
233
234 if (down_write_trylock(&mdsc->snap_rwsem)) {
235 spin_unlock(&mdsc->snap_empty_lock);
236 __destroy_snap_realm(mdsc, realm);
237 up_write(&mdsc->snap_rwsem);
238 } else {
239 list_add(&realm->empty_item, &mdsc->snap_empty);
240 spin_unlock(&mdsc->snap_empty_lock);
241 }
242 }
243
244 /*
245 * Clean up any realms whose ref counts have dropped to zero. Note
246 * that this does not include realms who were created but not yet
247 * used.
248 *
249 * Called under snap_rwsem (write)
250 */
__cleanup_empty_realms(struct ceph_mds_client * mdsc)251 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
252 {
253 struct ceph_snap_realm *realm;
254
255 lockdep_assert_held_write(&mdsc->snap_rwsem);
256
257 spin_lock(&mdsc->snap_empty_lock);
258 while (!list_empty(&mdsc->snap_empty)) {
259 realm = list_first_entry(&mdsc->snap_empty,
260 struct ceph_snap_realm, empty_item);
261 list_del(&realm->empty_item);
262 spin_unlock(&mdsc->snap_empty_lock);
263 __destroy_snap_realm(mdsc, realm);
264 spin_lock(&mdsc->snap_empty_lock);
265 }
266 spin_unlock(&mdsc->snap_empty_lock);
267 }
268
ceph_cleanup_global_and_empty_realms(struct ceph_mds_client * mdsc)269 void ceph_cleanup_global_and_empty_realms(struct ceph_mds_client *mdsc)
270 {
271 struct ceph_snap_realm *global_realm;
272
273 down_write(&mdsc->snap_rwsem);
274 global_realm = __lookup_snap_realm(mdsc, CEPH_INO_GLOBAL_SNAPREALM);
275 if (global_realm)
276 ceph_put_snap_realm(mdsc, global_realm);
277 __cleanup_empty_realms(mdsc);
278 up_write(&mdsc->snap_rwsem);
279 }
280
281 /*
282 * adjust the parent realm of a given @realm. adjust child list, and parent
283 * pointers, and ref counts appropriately.
284 *
285 * return true if parent was changed, 0 if unchanged, <0 on error.
286 *
287 * caller must hold snap_rwsem for write.
288 */
adjust_snap_realm_parent(struct ceph_mds_client * mdsc,struct ceph_snap_realm * realm,u64 parentino)289 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
290 struct ceph_snap_realm *realm,
291 u64 parentino)
292 {
293 struct ceph_snap_realm *parent;
294
295 lockdep_assert_held_write(&mdsc->snap_rwsem);
296
297 if (realm->parent_ino == parentino)
298 return 0;
299
300 parent = ceph_lookup_snap_realm(mdsc, parentino);
301 if (!parent) {
302 parent = ceph_create_snap_realm(mdsc, parentino);
303 if (IS_ERR(parent))
304 return PTR_ERR(parent);
305 }
306 dout("%s %llx %p: %llx %p -> %llx %p\n", __func__, realm->ino,
307 realm, realm->parent_ino, realm->parent, parentino, parent);
308 if (realm->parent) {
309 list_del_init(&realm->child_item);
310 ceph_put_snap_realm(mdsc, realm->parent);
311 }
312 realm->parent_ino = parentino;
313 realm->parent = parent;
314 list_add(&realm->child_item, &parent->children);
315 return 1;
316 }
317
318
cmpu64_rev(const void * a,const void * b)319 static int cmpu64_rev(const void *a, const void *b)
320 {
321 if (*(u64 *)a < *(u64 *)b)
322 return 1;
323 if (*(u64 *)a > *(u64 *)b)
324 return -1;
325 return 0;
326 }
327
328
329 /*
330 * build the snap context for a given realm.
331 */
build_snap_context(struct ceph_snap_realm * realm,struct list_head * realm_queue,struct list_head * dirty_realms)332 static int build_snap_context(struct ceph_snap_realm *realm,
333 struct list_head *realm_queue,
334 struct list_head *dirty_realms)
335 {
336 struct ceph_snap_realm *parent = realm->parent;
337 struct ceph_snap_context *snapc;
338 int err = 0;
339 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
340
341 /*
342 * build parent context, if it hasn't been built.
343 * conservatively estimate that all parent snaps might be
344 * included by us.
345 */
346 if (parent) {
347 if (!parent->cached_context) {
348 /* add to the queue head */
349 list_add(&parent->rebuild_item, realm_queue);
350 return 1;
351 }
352 num += parent->cached_context->num_snaps;
353 }
354
355 /* do i actually need to update? not if my context seq
356 matches realm seq, and my parents' does to. (this works
357 because we rebuild_snap_realms() works _downward_ in
358 hierarchy after each update.) */
359 if (realm->cached_context &&
360 realm->cached_context->seq == realm->seq &&
361 (!parent ||
362 realm->cached_context->seq >= parent->cached_context->seq)) {
363 dout("%s %llx %p: %p seq %lld (%u snaps) (unchanged)\n",
364 __func__, realm->ino, realm, realm->cached_context,
365 realm->cached_context->seq,
366 (unsigned int)realm->cached_context->num_snaps);
367 return 0;
368 }
369
370 /* alloc new snap context */
371 err = -ENOMEM;
372 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
373 goto fail;
374 snapc = ceph_create_snap_context(num, GFP_NOFS);
375 if (!snapc)
376 goto fail;
377
378 /* build (reverse sorted) snap vector */
379 num = 0;
380 snapc->seq = realm->seq;
381 if (parent) {
382 u32 i;
383
384 /* include any of parent's snaps occurring _after_ my
385 parent became my parent */
386 for (i = 0; i < parent->cached_context->num_snaps; i++)
387 if (parent->cached_context->snaps[i] >=
388 realm->parent_since)
389 snapc->snaps[num++] =
390 parent->cached_context->snaps[i];
391 if (parent->cached_context->seq > snapc->seq)
392 snapc->seq = parent->cached_context->seq;
393 }
394 memcpy(snapc->snaps + num, realm->snaps,
395 sizeof(u64)*realm->num_snaps);
396 num += realm->num_snaps;
397 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
398 sizeof(u64)*realm->num_prior_parent_snaps);
399 num += realm->num_prior_parent_snaps;
400
401 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
402 snapc->num_snaps = num;
403 dout("%s %llx %p: %p seq %lld (%u snaps)\n", __func__, realm->ino,
404 realm, snapc, snapc->seq, (unsigned int) snapc->num_snaps);
405
406 ceph_put_snap_context(realm->cached_context);
407 realm->cached_context = snapc;
408 /* queue realm for cap_snap creation */
409 list_add_tail(&realm->dirty_item, dirty_realms);
410 return 0;
411
412 fail:
413 /*
414 * if we fail, clear old (incorrect) cached_context... hopefully
415 * we'll have better luck building it later
416 */
417 if (realm->cached_context) {
418 ceph_put_snap_context(realm->cached_context);
419 realm->cached_context = NULL;
420 }
421 pr_err("%s %llx %p fail %d\n", __func__, realm->ino, realm, err);
422 return err;
423 }
424
425 /*
426 * rebuild snap context for the given realm and all of its children.
427 */
rebuild_snap_realms(struct ceph_snap_realm * realm,struct list_head * dirty_realms)428 static void rebuild_snap_realms(struct ceph_snap_realm *realm,
429 struct list_head *dirty_realms)
430 {
431 LIST_HEAD(realm_queue);
432 int last = 0;
433 bool skip = false;
434
435 list_add_tail(&realm->rebuild_item, &realm_queue);
436
437 while (!list_empty(&realm_queue)) {
438 struct ceph_snap_realm *_realm, *child;
439
440 _realm = list_first_entry(&realm_queue,
441 struct ceph_snap_realm,
442 rebuild_item);
443
444 /*
445 * If the last building failed dues to memory
446 * issue, just empty the realm_queue and return
447 * to avoid infinite loop.
448 */
449 if (last < 0) {
450 list_del_init(&_realm->rebuild_item);
451 continue;
452 }
453
454 last = build_snap_context(_realm, &realm_queue, dirty_realms);
455 dout("%s %llx %p, %s\n", __func__, _realm->ino, _realm,
456 last > 0 ? "is deferred" : !last ? "succeeded" : "failed");
457
458 /* is any child in the list ? */
459 list_for_each_entry(child, &_realm->children, child_item) {
460 if (!list_empty(&child->rebuild_item)) {
461 skip = true;
462 break;
463 }
464 }
465
466 if (!skip) {
467 list_for_each_entry(child, &_realm->children, child_item)
468 list_add_tail(&child->rebuild_item, &realm_queue);
469 }
470
471 /* last == 1 means need to build parent first */
472 if (last <= 0)
473 list_del_init(&_realm->rebuild_item);
474 }
475 }
476
477
478 /*
479 * helper to allocate and decode an array of snapids. free prior
480 * instance, if any.
481 */
dup_array(u64 ** dst,__le64 * src,u32 num)482 static int dup_array(u64 **dst, __le64 *src, u32 num)
483 {
484 u32 i;
485
486 kfree(*dst);
487 if (num) {
488 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
489 if (!*dst)
490 return -ENOMEM;
491 for (i = 0; i < num; i++)
492 (*dst)[i] = get_unaligned_le64(src + i);
493 } else {
494 *dst = NULL;
495 }
496 return 0;
497 }
498
has_new_snaps(struct ceph_snap_context * o,struct ceph_snap_context * n)499 static bool has_new_snaps(struct ceph_snap_context *o,
500 struct ceph_snap_context *n)
501 {
502 if (n->num_snaps == 0)
503 return false;
504 /* snaps are in descending order */
505 return n->snaps[0] > o->seq;
506 }
507
508 /*
509 * When a snapshot is applied, the size/mtime inode metadata is queued
510 * in a ceph_cap_snap (one for each snapshot) until writeback
511 * completes and the metadata can be flushed back to the MDS.
512 *
513 * However, if a (sync) write is currently in-progress when we apply
514 * the snapshot, we have to wait until the write succeeds or fails
515 * (and a final size/mtime is known). In this case the
516 * cap_snap->writing = 1, and is said to be "pending." When the write
517 * finishes, we __ceph_finish_cap_snap().
518 *
519 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
520 * change).
521 */
ceph_queue_cap_snap(struct ceph_inode_info * ci,struct ceph_cap_snap ** pcapsnap)522 static void ceph_queue_cap_snap(struct ceph_inode_info *ci,
523 struct ceph_cap_snap **pcapsnap)
524 {
525 struct inode *inode = &ci->netfs.inode;
526 struct ceph_snap_context *old_snapc, *new_snapc;
527 struct ceph_cap_snap *capsnap = *pcapsnap;
528 struct ceph_buffer *old_blob = NULL;
529 int used, dirty;
530
531 spin_lock(&ci->i_ceph_lock);
532 used = __ceph_caps_used(ci);
533 dirty = __ceph_caps_dirty(ci);
534
535 old_snapc = ci->i_head_snapc;
536 new_snapc = ci->i_snap_realm->cached_context;
537
538 /*
539 * If there is a write in progress, treat that as a dirty Fw,
540 * even though it hasn't completed yet; by the time we finish
541 * up this capsnap it will be.
542 */
543 if (used & CEPH_CAP_FILE_WR)
544 dirty |= CEPH_CAP_FILE_WR;
545
546 if (__ceph_have_pending_cap_snap(ci)) {
547 /* there is no point in queuing multiple "pending" cap_snaps,
548 as no new writes are allowed to start when pending, so any
549 writes in progress now were started before the previous
550 cap_snap. lucky us. */
551 dout("%s %p %llx.%llx already pending\n",
552 __func__, inode, ceph_vinop(inode));
553 goto update_snapc;
554 }
555 if (ci->i_wrbuffer_ref_head == 0 &&
556 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
557 dout("%s %p %llx.%llx nothing dirty|writing\n",
558 __func__, inode, ceph_vinop(inode));
559 goto update_snapc;
560 }
561
562 BUG_ON(!old_snapc);
563
564 /*
565 * There is no need to send FLUSHSNAP message to MDS if there is
566 * no new snapshot. But when there is dirty pages or on-going
567 * writes, we still need to create cap_snap. cap_snap is needed
568 * by the write path and page writeback path.
569 *
570 * also see ceph_try_drop_cap_snap()
571 */
572 if (has_new_snaps(old_snapc, new_snapc)) {
573 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
574 capsnap->need_flush = true;
575 } else {
576 if (!(used & CEPH_CAP_FILE_WR) &&
577 ci->i_wrbuffer_ref_head == 0) {
578 dout("%s %p %llx.%llx no new_snap|dirty_page|writing\n",
579 __func__, inode, ceph_vinop(inode));
580 goto update_snapc;
581 }
582 }
583
584 dout("%s %p %llx.%llx cap_snap %p queuing under %p %s %s\n",
585 __func__, inode, ceph_vinop(inode), capsnap, old_snapc,
586 ceph_cap_string(dirty), capsnap->need_flush ? "" : "no_flush");
587 ihold(inode);
588
589 capsnap->follows = old_snapc->seq;
590 capsnap->issued = __ceph_caps_issued(ci, NULL);
591 capsnap->dirty = dirty;
592
593 capsnap->mode = inode->i_mode;
594 capsnap->uid = inode->i_uid;
595 capsnap->gid = inode->i_gid;
596
597 if (dirty & CEPH_CAP_XATTR_EXCL) {
598 old_blob = __ceph_build_xattrs_blob(ci);
599 capsnap->xattr_blob =
600 ceph_buffer_get(ci->i_xattrs.blob);
601 capsnap->xattr_version = ci->i_xattrs.version;
602 } else {
603 capsnap->xattr_blob = NULL;
604 capsnap->xattr_version = 0;
605 }
606
607 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
608
609 /* dirty page count moved from _head to this cap_snap;
610 all subsequent writes page dirties occur _after_ this
611 snapshot. */
612 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
613 ci->i_wrbuffer_ref_head = 0;
614 capsnap->context = old_snapc;
615 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
616
617 if (used & CEPH_CAP_FILE_WR) {
618 dout("%s %p %llx.%llx cap_snap %p snapc %p seq %llu used WR,"
619 " now pending\n", __func__, inode, ceph_vinop(inode),
620 capsnap, old_snapc, old_snapc->seq);
621 capsnap->writing = 1;
622 } else {
623 /* note mtime, size NOW. */
624 __ceph_finish_cap_snap(ci, capsnap);
625 }
626 *pcapsnap = NULL;
627 old_snapc = NULL;
628
629 update_snapc:
630 if (ci->i_wrbuffer_ref_head == 0 &&
631 ci->i_wr_ref == 0 &&
632 ci->i_dirty_caps == 0 &&
633 ci->i_flushing_caps == 0) {
634 ci->i_head_snapc = NULL;
635 } else {
636 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
637 dout(" new snapc is %p\n", new_snapc);
638 }
639 spin_unlock(&ci->i_ceph_lock);
640
641 ceph_buffer_put(old_blob);
642 ceph_put_snap_context(old_snapc);
643 }
644
645 /*
646 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
647 * to be used for the snapshot, to be flushed back to the mds.
648 *
649 * If capsnap can now be flushed, add to snap_flush list, and return 1.
650 *
651 * Caller must hold i_ceph_lock.
652 */
__ceph_finish_cap_snap(struct ceph_inode_info * ci,struct ceph_cap_snap * capsnap)653 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
654 struct ceph_cap_snap *capsnap)
655 {
656 struct inode *inode = &ci->netfs.inode;
657 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
658
659 BUG_ON(capsnap->writing);
660 capsnap->size = i_size_read(inode);
661 capsnap->mtime = inode->i_mtime;
662 capsnap->atime = inode->i_atime;
663 capsnap->ctime = inode->i_ctime;
664 capsnap->btime = ci->i_btime;
665 capsnap->change_attr = inode_peek_iversion_raw(inode);
666 capsnap->time_warp_seq = ci->i_time_warp_seq;
667 capsnap->truncate_size = ci->i_truncate_size;
668 capsnap->truncate_seq = ci->i_truncate_seq;
669 if (capsnap->dirty_pages) {
670 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu "
671 "still has %d dirty pages\n", __func__, inode,
672 ceph_vinop(inode), capsnap, capsnap->context,
673 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
674 capsnap->size, capsnap->dirty_pages);
675 return 0;
676 }
677
678 /* Fb cap still in use, delay it */
679 if (ci->i_wb_ref) {
680 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu "
681 "used WRBUFFER, delaying\n", __func__, inode,
682 ceph_vinop(inode), capsnap, capsnap->context,
683 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
684 capsnap->size);
685 capsnap->writing = 1;
686 return 0;
687 }
688
689 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
690 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu\n",
691 __func__, inode, ceph_vinop(inode), capsnap, capsnap->context,
692 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
693 capsnap->size);
694
695 spin_lock(&mdsc->snap_flush_lock);
696 if (list_empty(&ci->i_snap_flush_item)) {
697 ihold(inode);
698 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
699 }
700 spin_unlock(&mdsc->snap_flush_lock);
701 return 1; /* caller may want to ceph_flush_snaps */
702 }
703
704 /*
705 * Queue cap_snaps for snap writeback for this realm and its children.
706 * Called under snap_rwsem, so realm topology won't change.
707 */
queue_realm_cap_snaps(struct ceph_snap_realm * realm)708 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
709 {
710 struct ceph_inode_info *ci;
711 struct inode *lastinode = NULL;
712 struct ceph_cap_snap *capsnap = NULL;
713
714 dout("%s %p %llx inode\n", __func__, realm, realm->ino);
715
716 spin_lock(&realm->inodes_with_caps_lock);
717 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
718 struct inode *inode = igrab(&ci->netfs.inode);
719 if (!inode)
720 continue;
721 spin_unlock(&realm->inodes_with_caps_lock);
722 iput(lastinode);
723 lastinode = inode;
724
725 /*
726 * Allocate the capsnap memory outside of ceph_queue_cap_snap()
727 * to reduce very possible but unnecessary frequently memory
728 * allocate/free in this loop.
729 */
730 if (!capsnap) {
731 capsnap = kmem_cache_zalloc(ceph_cap_snap_cachep, GFP_NOFS);
732 if (!capsnap) {
733 pr_err("ENOMEM allocating ceph_cap_snap on %p\n",
734 inode);
735 return;
736 }
737 }
738 capsnap->cap_flush.is_capsnap = true;
739 refcount_set(&capsnap->nref, 1);
740 INIT_LIST_HEAD(&capsnap->cap_flush.i_list);
741 INIT_LIST_HEAD(&capsnap->cap_flush.g_list);
742 INIT_LIST_HEAD(&capsnap->ci_item);
743
744 ceph_queue_cap_snap(ci, &capsnap);
745 spin_lock(&realm->inodes_with_caps_lock);
746 }
747 spin_unlock(&realm->inodes_with_caps_lock);
748 iput(lastinode);
749
750 if (capsnap)
751 kmem_cache_free(ceph_cap_snap_cachep, capsnap);
752 dout("%s %p %llx done\n", __func__, realm, realm->ino);
753 }
754
755 /*
756 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
757 * the snap realm parameters from a given realm and all of its ancestors,
758 * up to the root.
759 *
760 * Caller must hold snap_rwsem for write.
761 */
ceph_update_snap_trace(struct ceph_mds_client * mdsc,void * p,void * e,bool deletion,struct ceph_snap_realm ** realm_ret)762 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
763 void *p, void *e, bool deletion,
764 struct ceph_snap_realm **realm_ret)
765 {
766 struct ceph_mds_snap_realm *ri; /* encoded */
767 __le64 *snaps; /* encoded */
768 __le64 *prior_parent_snaps; /* encoded */
769 struct ceph_snap_realm *realm;
770 struct ceph_snap_realm *first_realm = NULL;
771 struct ceph_snap_realm *realm_to_rebuild = NULL;
772 struct ceph_client *client = mdsc->fsc->client;
773 int rebuild_snapcs;
774 int err = -ENOMEM;
775 int ret;
776 LIST_HEAD(dirty_realms);
777
778 lockdep_assert_held_write(&mdsc->snap_rwsem);
779
780 dout("%s deletion=%d\n", __func__, deletion);
781 more:
782 realm = NULL;
783 rebuild_snapcs = 0;
784 ceph_decode_need(&p, e, sizeof(*ri), bad);
785 ri = p;
786 p += sizeof(*ri);
787 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
788 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
789 snaps = p;
790 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
791 prior_parent_snaps = p;
792 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
793
794 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
795 if (!realm) {
796 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
797 if (IS_ERR(realm)) {
798 err = PTR_ERR(realm);
799 goto fail;
800 }
801 }
802
803 /* ensure the parent is correct */
804 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
805 if (err < 0)
806 goto fail;
807 rebuild_snapcs += err;
808
809 if (le64_to_cpu(ri->seq) > realm->seq) {
810 dout("%s updating %llx %p %lld -> %lld\n", __func__,
811 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
812 /* update realm parameters, snap lists */
813 realm->seq = le64_to_cpu(ri->seq);
814 realm->created = le64_to_cpu(ri->created);
815 realm->parent_since = le64_to_cpu(ri->parent_since);
816
817 realm->num_snaps = le32_to_cpu(ri->num_snaps);
818 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
819 if (err < 0)
820 goto fail;
821
822 realm->num_prior_parent_snaps =
823 le32_to_cpu(ri->num_prior_parent_snaps);
824 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
825 realm->num_prior_parent_snaps);
826 if (err < 0)
827 goto fail;
828
829 if (realm->seq > mdsc->last_snap_seq)
830 mdsc->last_snap_seq = realm->seq;
831
832 rebuild_snapcs = 1;
833 } else if (!realm->cached_context) {
834 dout("%s %llx %p seq %lld new\n", __func__,
835 realm->ino, realm, realm->seq);
836 rebuild_snapcs = 1;
837 } else {
838 dout("%s %llx %p seq %lld unchanged\n", __func__,
839 realm->ino, realm, realm->seq);
840 }
841
842 dout("done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
843 realm, rebuild_snapcs, p, e);
844
845 /*
846 * this will always track the uppest parent realm from which
847 * we need to rebuild the snapshot contexts _downward_ in
848 * hierarchy.
849 */
850 if (rebuild_snapcs)
851 realm_to_rebuild = realm;
852
853 /* rebuild_snapcs when we reach the _end_ (root) of the trace */
854 if (realm_to_rebuild && p >= e)
855 rebuild_snap_realms(realm_to_rebuild, &dirty_realms);
856
857 if (!first_realm)
858 first_realm = realm;
859 else
860 ceph_put_snap_realm(mdsc, realm);
861
862 if (p < e)
863 goto more;
864
865 /*
866 * queue cap snaps _after_ we've built the new snap contexts,
867 * so that i_head_snapc can be set appropriately.
868 */
869 while (!list_empty(&dirty_realms)) {
870 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
871 dirty_item);
872 list_del_init(&realm->dirty_item);
873 queue_realm_cap_snaps(realm);
874 }
875
876 if (realm_ret)
877 *realm_ret = first_realm;
878 else
879 ceph_put_snap_realm(mdsc, first_realm);
880
881 __cleanup_empty_realms(mdsc);
882 return 0;
883
884 bad:
885 err = -EIO;
886 fail:
887 if (realm && !IS_ERR(realm))
888 ceph_put_snap_realm(mdsc, realm);
889 if (first_realm)
890 ceph_put_snap_realm(mdsc, first_realm);
891 pr_err("%s error %d\n", __func__, err);
892
893 /*
894 * When receiving a corrupted snap trace we don't know what
895 * exactly has happened in MDS side. And we shouldn't continue
896 * writing to OSD, which may corrupt the snapshot contents.
897 *
898 * Just try to blocklist this kclient and then this kclient
899 * must be remounted to continue after the corrupted metadata
900 * fixed in the MDS side.
901 */
902 WRITE_ONCE(mdsc->fsc->mount_state, CEPH_MOUNT_FENCE_IO);
903 ret = ceph_monc_blocklist_add(&client->monc, &client->msgr.inst.addr);
904 if (ret)
905 pr_err("%s failed to blocklist %s: %d\n", __func__,
906 ceph_pr_addr(&client->msgr.inst.addr), ret);
907
908 WARN(1, "%s: %s%sdo remount to continue%s",
909 __func__, ret ? "" : ceph_pr_addr(&client->msgr.inst.addr),
910 ret ? "" : " was blocklisted, ",
911 err == -EIO ? " after corrupted snaptrace is fixed" : "");
912
913 return err;
914 }
915
916
917 /*
918 * Send any cap_snaps that are queued for flush. Try to carry
919 * s_mutex across multiple snap flushes to avoid locking overhead.
920 *
921 * Caller holds no locks.
922 */
flush_snaps(struct ceph_mds_client * mdsc)923 static void flush_snaps(struct ceph_mds_client *mdsc)
924 {
925 struct ceph_inode_info *ci;
926 struct inode *inode;
927 struct ceph_mds_session *session = NULL;
928
929 dout("%s\n", __func__);
930 spin_lock(&mdsc->snap_flush_lock);
931 while (!list_empty(&mdsc->snap_flush_list)) {
932 ci = list_first_entry(&mdsc->snap_flush_list,
933 struct ceph_inode_info, i_snap_flush_item);
934 inode = &ci->netfs.inode;
935 ihold(inode);
936 spin_unlock(&mdsc->snap_flush_lock);
937 ceph_flush_snaps(ci, &session);
938 iput(inode);
939 spin_lock(&mdsc->snap_flush_lock);
940 }
941 spin_unlock(&mdsc->snap_flush_lock);
942
943 ceph_put_mds_session(session);
944 dout("%s done\n", __func__);
945 }
946
947 /**
948 * ceph_change_snap_realm - change the snap_realm for an inode
949 * @inode: inode to move to new snap realm
950 * @realm: new realm to move inode into (may be NULL)
951 *
952 * Detach an inode from its old snaprealm (if any) and attach it to
953 * the new snaprealm (if any). The old snap realm reference held by
954 * the inode is put. If realm is non-NULL, then the caller's reference
955 * to it is taken over by the inode.
956 */
ceph_change_snap_realm(struct inode * inode,struct ceph_snap_realm * realm)957 void ceph_change_snap_realm(struct inode *inode, struct ceph_snap_realm *realm)
958 {
959 struct ceph_inode_info *ci = ceph_inode(inode);
960 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
961 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
962
963 lockdep_assert_held(&ci->i_ceph_lock);
964
965 if (oldrealm) {
966 spin_lock(&oldrealm->inodes_with_caps_lock);
967 list_del_init(&ci->i_snap_realm_item);
968 if (oldrealm->ino == ci->i_vino.ino)
969 oldrealm->inode = NULL;
970 spin_unlock(&oldrealm->inodes_with_caps_lock);
971 ceph_put_snap_realm(mdsc, oldrealm);
972 }
973
974 ci->i_snap_realm = realm;
975
976 if (realm) {
977 spin_lock(&realm->inodes_with_caps_lock);
978 list_add(&ci->i_snap_realm_item, &realm->inodes_with_caps);
979 if (realm->ino == ci->i_vino.ino)
980 realm->inode = inode;
981 spin_unlock(&realm->inodes_with_caps_lock);
982 }
983 }
984
985 /*
986 * Handle a snap notification from the MDS.
987 *
988 * This can take two basic forms: the simplest is just a snap creation
989 * or deletion notification on an existing realm. This should update the
990 * realm and its children.
991 *
992 * The more difficult case is realm creation, due to snap creation at a
993 * new point in the file hierarchy, or due to a rename that moves a file or
994 * directory into another realm.
995 */
ceph_handle_snap(struct ceph_mds_client * mdsc,struct ceph_mds_session * session,struct ceph_msg * msg)996 void ceph_handle_snap(struct ceph_mds_client *mdsc,
997 struct ceph_mds_session *session,
998 struct ceph_msg *msg)
999 {
1000 struct super_block *sb = mdsc->fsc->sb;
1001 int mds = session->s_mds;
1002 u64 split;
1003 int op;
1004 int trace_len;
1005 struct ceph_snap_realm *realm = NULL;
1006 void *p = msg->front.iov_base;
1007 void *e = p + msg->front.iov_len;
1008 struct ceph_mds_snap_head *h;
1009 int num_split_inos, num_split_realms;
1010 __le64 *split_inos = NULL, *split_realms = NULL;
1011 int i;
1012 int locked_rwsem = 0;
1013 bool close_sessions = false;
1014
1015 if (!ceph_inc_mds_stopping_blocker(mdsc, session))
1016 return;
1017
1018 /* decode */
1019 if (msg->front.iov_len < sizeof(*h))
1020 goto bad;
1021 h = p;
1022 op = le32_to_cpu(h->op);
1023 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
1024 * existing realm */
1025 num_split_inos = le32_to_cpu(h->num_split_inos);
1026 num_split_realms = le32_to_cpu(h->num_split_realms);
1027 trace_len = le32_to_cpu(h->trace_len);
1028 p += sizeof(*h);
1029
1030 dout("%s from mds%d op %s split %llx tracelen %d\n", __func__,
1031 mds, ceph_snap_op_name(op), split, trace_len);
1032
1033 down_write(&mdsc->snap_rwsem);
1034 locked_rwsem = 1;
1035
1036 if (op == CEPH_SNAP_OP_SPLIT) {
1037 struct ceph_mds_snap_realm *ri;
1038
1039 /*
1040 * A "split" breaks part of an existing realm off into
1041 * a new realm. The MDS provides a list of inodes
1042 * (with caps) and child realms that belong to the new
1043 * child.
1044 */
1045 split_inos = p;
1046 p += sizeof(u64) * num_split_inos;
1047 split_realms = p;
1048 p += sizeof(u64) * num_split_realms;
1049 ceph_decode_need(&p, e, sizeof(*ri), bad);
1050 /* we will peek at realm info here, but will _not_
1051 * advance p, as the realm update will occur below in
1052 * ceph_update_snap_trace. */
1053 ri = p;
1054
1055 realm = ceph_lookup_snap_realm(mdsc, split);
1056 if (!realm) {
1057 realm = ceph_create_snap_realm(mdsc, split);
1058 if (IS_ERR(realm))
1059 goto out;
1060 }
1061
1062 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
1063 for (i = 0; i < num_split_inos; i++) {
1064 struct ceph_vino vino = {
1065 .ino = le64_to_cpu(split_inos[i]),
1066 .snap = CEPH_NOSNAP,
1067 };
1068 struct inode *inode = ceph_find_inode(sb, vino);
1069 struct ceph_inode_info *ci;
1070
1071 if (!inode)
1072 continue;
1073 ci = ceph_inode(inode);
1074
1075 spin_lock(&ci->i_ceph_lock);
1076 if (!ci->i_snap_realm)
1077 goto skip_inode;
1078 /*
1079 * If this inode belongs to a realm that was
1080 * created after our new realm, we experienced
1081 * a race (due to another split notifications
1082 * arriving from a different MDS). So skip
1083 * this inode.
1084 */
1085 if (ci->i_snap_realm->created >
1086 le64_to_cpu(ri->created)) {
1087 dout(" leaving %p %llx.%llx in newer realm %llx %p\n",
1088 inode, ceph_vinop(inode), ci->i_snap_realm->ino,
1089 ci->i_snap_realm);
1090 goto skip_inode;
1091 }
1092 dout(" will move %p %llx.%llx to split realm %llx %p\n",
1093 inode, ceph_vinop(inode), realm->ino, realm);
1094
1095 ceph_get_snap_realm(mdsc, realm);
1096 ceph_change_snap_realm(inode, realm);
1097 spin_unlock(&ci->i_ceph_lock);
1098 iput(inode);
1099 continue;
1100
1101 skip_inode:
1102 spin_unlock(&ci->i_ceph_lock);
1103 iput(inode);
1104 }
1105
1106 /* we may have taken some of the old realm's children. */
1107 for (i = 0; i < num_split_realms; i++) {
1108 struct ceph_snap_realm *child =
1109 __lookup_snap_realm(mdsc,
1110 le64_to_cpu(split_realms[i]));
1111 if (!child)
1112 continue;
1113 adjust_snap_realm_parent(mdsc, child, realm->ino);
1114 }
1115 } else {
1116 /*
1117 * In the non-split case both 'num_split_inos' and
1118 * 'num_split_realms' should be 0, making this a no-op.
1119 * However the MDS happens to populate 'split_realms' list
1120 * in one of the UPDATE op cases by mistake.
1121 *
1122 * Skip both lists just in case to ensure that 'p' is
1123 * positioned at the start of realm info, as expected by
1124 * ceph_update_snap_trace().
1125 */
1126 p += sizeof(u64) * num_split_inos;
1127 p += sizeof(u64) * num_split_realms;
1128 }
1129
1130 /*
1131 * update using the provided snap trace. if we are deleting a
1132 * snap, we can avoid queueing cap_snaps.
1133 */
1134 if (ceph_update_snap_trace(mdsc, p, e,
1135 op == CEPH_SNAP_OP_DESTROY,
1136 NULL)) {
1137 close_sessions = true;
1138 goto bad;
1139 }
1140
1141 if (op == CEPH_SNAP_OP_SPLIT)
1142 /* we took a reference when we created the realm, above */
1143 ceph_put_snap_realm(mdsc, realm);
1144
1145 __cleanup_empty_realms(mdsc);
1146
1147 up_write(&mdsc->snap_rwsem);
1148
1149 flush_snaps(mdsc);
1150 ceph_dec_mds_stopping_blocker(mdsc);
1151 return;
1152
1153 bad:
1154 pr_err("%s corrupt snap message from mds%d\n", __func__, mds);
1155 ceph_msg_dump(msg);
1156 out:
1157 if (locked_rwsem)
1158 up_write(&mdsc->snap_rwsem);
1159
1160 ceph_dec_mds_stopping_blocker(mdsc);
1161
1162 if (close_sessions)
1163 ceph_mdsc_close_sessions(mdsc);
1164 return;
1165 }
1166
ceph_get_snapid_map(struct ceph_mds_client * mdsc,u64 snap)1167 struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1168 u64 snap)
1169 {
1170 struct ceph_snapid_map *sm, *exist;
1171 struct rb_node **p, *parent;
1172 int ret;
1173
1174 exist = NULL;
1175 spin_lock(&mdsc->snapid_map_lock);
1176 p = &mdsc->snapid_map_tree.rb_node;
1177 while (*p) {
1178 exist = rb_entry(*p, struct ceph_snapid_map, node);
1179 if (snap > exist->snap) {
1180 p = &(*p)->rb_left;
1181 } else if (snap < exist->snap) {
1182 p = &(*p)->rb_right;
1183 } else {
1184 if (atomic_inc_return(&exist->ref) == 1)
1185 list_del_init(&exist->lru);
1186 break;
1187 }
1188 exist = NULL;
1189 }
1190 spin_unlock(&mdsc->snapid_map_lock);
1191 if (exist) {
1192 dout("%s found snapid map %llx -> %x\n", __func__,
1193 exist->snap, exist->dev);
1194 return exist;
1195 }
1196
1197 sm = kmalloc(sizeof(*sm), GFP_NOFS);
1198 if (!sm)
1199 return NULL;
1200
1201 ret = get_anon_bdev(&sm->dev);
1202 if (ret < 0) {
1203 kfree(sm);
1204 return NULL;
1205 }
1206
1207 INIT_LIST_HEAD(&sm->lru);
1208 atomic_set(&sm->ref, 1);
1209 sm->snap = snap;
1210
1211 exist = NULL;
1212 parent = NULL;
1213 p = &mdsc->snapid_map_tree.rb_node;
1214 spin_lock(&mdsc->snapid_map_lock);
1215 while (*p) {
1216 parent = *p;
1217 exist = rb_entry(*p, struct ceph_snapid_map, node);
1218 if (snap > exist->snap)
1219 p = &(*p)->rb_left;
1220 else if (snap < exist->snap)
1221 p = &(*p)->rb_right;
1222 else
1223 break;
1224 exist = NULL;
1225 }
1226 if (exist) {
1227 if (atomic_inc_return(&exist->ref) == 1)
1228 list_del_init(&exist->lru);
1229 } else {
1230 rb_link_node(&sm->node, parent, p);
1231 rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1232 }
1233 spin_unlock(&mdsc->snapid_map_lock);
1234 if (exist) {
1235 free_anon_bdev(sm->dev);
1236 kfree(sm);
1237 dout("%s found snapid map %llx -> %x\n", __func__,
1238 exist->snap, exist->dev);
1239 return exist;
1240 }
1241
1242 dout("%s create snapid map %llx -> %x\n", __func__,
1243 sm->snap, sm->dev);
1244 return sm;
1245 }
1246
ceph_put_snapid_map(struct ceph_mds_client * mdsc,struct ceph_snapid_map * sm)1247 void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1248 struct ceph_snapid_map *sm)
1249 {
1250 if (!sm)
1251 return;
1252 if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1253 if (!RB_EMPTY_NODE(&sm->node)) {
1254 sm->last_used = jiffies;
1255 list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1256 spin_unlock(&mdsc->snapid_map_lock);
1257 } else {
1258 /* already cleaned up by
1259 * ceph_cleanup_snapid_map() */
1260 spin_unlock(&mdsc->snapid_map_lock);
1261 kfree(sm);
1262 }
1263 }
1264 }
1265
ceph_trim_snapid_map(struct ceph_mds_client * mdsc)1266 void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1267 {
1268 struct ceph_snapid_map *sm;
1269 unsigned long now;
1270 LIST_HEAD(to_free);
1271
1272 spin_lock(&mdsc->snapid_map_lock);
1273 now = jiffies;
1274
1275 while (!list_empty(&mdsc->snapid_map_lru)) {
1276 sm = list_first_entry(&mdsc->snapid_map_lru,
1277 struct ceph_snapid_map, lru);
1278 if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1279 break;
1280
1281 rb_erase(&sm->node, &mdsc->snapid_map_tree);
1282 list_move(&sm->lru, &to_free);
1283 }
1284 spin_unlock(&mdsc->snapid_map_lock);
1285
1286 while (!list_empty(&to_free)) {
1287 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1288 list_del(&sm->lru);
1289 dout("trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1290 free_anon_bdev(sm->dev);
1291 kfree(sm);
1292 }
1293 }
1294
ceph_cleanup_snapid_map(struct ceph_mds_client * mdsc)1295 void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1296 {
1297 struct ceph_snapid_map *sm;
1298 struct rb_node *p;
1299 LIST_HEAD(to_free);
1300
1301 spin_lock(&mdsc->snapid_map_lock);
1302 while ((p = rb_first(&mdsc->snapid_map_tree))) {
1303 sm = rb_entry(p, struct ceph_snapid_map, node);
1304 rb_erase(p, &mdsc->snapid_map_tree);
1305 RB_CLEAR_NODE(p);
1306 list_move(&sm->lru, &to_free);
1307 }
1308 spin_unlock(&mdsc->snapid_map_lock);
1309
1310 while (!list_empty(&to_free)) {
1311 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1312 list_del(&sm->lru);
1313 free_anon_bdev(sm->dev);
1314 if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1315 pr_err("snapid map %llx -> %x still in use\n",
1316 sm->snap, sm->dev);
1317 }
1318 kfree(sm);
1319 }
1320 }
1321