xref: /linux/fs/ceph/snap.c (revision 0af298e62e5cff4ebf9bec47b8711177a7e2f115)
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  */
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 
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  */
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_obj(*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 	doutc(mdsc->fsc->client, "%llx %p\n", realm->ino, realm);
142 	return realm;
143 }
144 
145 /*
146  * lookup the realm rooted at @ino.
147  *
148  * caller must hold snap_rwsem.
149  */
150 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
151 						   u64 ino)
152 {
153 	struct ceph_client *cl = mdsc->fsc->client;
154 	struct rb_node *n = mdsc->snap_realms.rb_node;
155 	struct ceph_snap_realm *r;
156 
157 	lockdep_assert_held(&mdsc->snap_rwsem);
158 
159 	while (n) {
160 		r = rb_entry(n, struct ceph_snap_realm, node);
161 		if (ino < r->ino)
162 			n = n->rb_left;
163 		else if (ino > r->ino)
164 			n = n->rb_right;
165 		else {
166 			doutc(cl, "%llx %p\n", r->ino, r);
167 			return r;
168 		}
169 	}
170 	return NULL;
171 }
172 
173 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
174 					       u64 ino)
175 {
176 	struct ceph_snap_realm *r;
177 	r = __lookup_snap_realm(mdsc, ino);
178 	if (r)
179 		ceph_get_snap_realm(mdsc, r);
180 	return r;
181 }
182 
183 static void __put_snap_realm(struct ceph_mds_client *mdsc,
184 			     struct ceph_snap_realm *realm);
185 
186 /*
187  * called with snap_rwsem (write)
188  */
189 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
190 				 struct ceph_snap_realm *realm)
191 {
192 	struct ceph_client *cl = mdsc->fsc->client;
193 	lockdep_assert_held_write(&mdsc->snap_rwsem);
194 
195 	doutc(cl, "%p %llx\n", realm, realm->ino);
196 
197 	rb_erase(&realm->node, &mdsc->snap_realms);
198 	mdsc->num_snap_realms--;
199 
200 	if (realm->parent) {
201 		list_del_init(&realm->child_item);
202 		__put_snap_realm(mdsc, realm->parent);
203 	}
204 
205 	kfree(realm->prior_parent_snaps);
206 	kfree(realm->snaps);
207 	ceph_put_snap_context(realm->cached_context);
208 	kfree(realm);
209 }
210 
211 /*
212  * caller holds snap_rwsem (write)
213  */
214 static void __put_snap_realm(struct ceph_mds_client *mdsc,
215 			     struct ceph_snap_realm *realm)
216 {
217 	lockdep_assert_held_write(&mdsc->snap_rwsem);
218 
219 	/*
220 	 * We do not require the snap_empty_lock here, as any caller that
221 	 * increments the value must hold the snap_rwsem.
222 	 */
223 	if (atomic_dec_and_test(&realm->nref))
224 		__destroy_snap_realm(mdsc, realm);
225 }
226 
227 /*
228  * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
229  */
230 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
231 			 struct ceph_snap_realm *realm)
232 {
233 	if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
234 		return;
235 
236 	if (down_write_trylock(&mdsc->snap_rwsem)) {
237 		spin_unlock(&mdsc->snap_empty_lock);
238 		__destroy_snap_realm(mdsc, realm);
239 		up_write(&mdsc->snap_rwsem);
240 	} else {
241 		list_add(&realm->empty_item, &mdsc->snap_empty);
242 		spin_unlock(&mdsc->snap_empty_lock);
243 	}
244 }
245 
246 /*
247  * Clean up any realms whose ref counts have dropped to zero.  Note
248  * that this does not include realms who were created but not yet
249  * used.
250  *
251  * Called under snap_rwsem (write)
252  */
253 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
254 {
255 	struct ceph_snap_realm *realm;
256 
257 	lockdep_assert_held_write(&mdsc->snap_rwsem);
258 
259 	spin_lock(&mdsc->snap_empty_lock);
260 	while (!list_empty(&mdsc->snap_empty)) {
261 		realm = list_first_entry(&mdsc->snap_empty,
262 				   struct ceph_snap_realm, empty_item);
263 		list_del(&realm->empty_item);
264 		spin_unlock(&mdsc->snap_empty_lock);
265 		__destroy_snap_realm(mdsc, realm);
266 		spin_lock(&mdsc->snap_empty_lock);
267 	}
268 	spin_unlock(&mdsc->snap_empty_lock);
269 }
270 
271 void ceph_cleanup_global_and_empty_realms(struct ceph_mds_client *mdsc)
272 {
273 	struct ceph_snap_realm *global_realm;
274 
275 	down_write(&mdsc->snap_rwsem);
276 	global_realm = __lookup_snap_realm(mdsc, CEPH_INO_GLOBAL_SNAPREALM);
277 	if (global_realm)
278 		ceph_put_snap_realm(mdsc, global_realm);
279 	__cleanup_empty_realms(mdsc);
280 	up_write(&mdsc->snap_rwsem);
281 }
282 
283 /*
284  * adjust the parent realm of a given @realm.  adjust child list, and parent
285  * pointers, and ref counts appropriately.
286  *
287  * return true if parent was changed, 0 if unchanged, <0 on error.
288  *
289  * caller must hold snap_rwsem for write.
290  */
291 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
292 				    struct ceph_snap_realm *realm,
293 				    u64 parentino)
294 {
295 	struct ceph_client *cl = mdsc->fsc->client;
296 	struct ceph_snap_realm *parent;
297 
298 	lockdep_assert_held_write(&mdsc->snap_rwsem);
299 
300 	if (realm->parent_ino == parentino)
301 		return 0;
302 
303 	parent = ceph_lookup_snap_realm(mdsc, parentino);
304 	if (!parent) {
305 		parent = ceph_create_snap_realm(mdsc, parentino);
306 		if (IS_ERR(parent))
307 			return PTR_ERR(parent);
308 	}
309 	doutc(cl, "%llx %p: %llx %p -> %llx %p\n", realm->ino, realm,
310 	      realm->parent_ino, realm->parent, parentino, parent);
311 	if (realm->parent) {
312 		list_del_init(&realm->child_item);
313 		ceph_put_snap_realm(mdsc, realm->parent);
314 	}
315 	realm->parent_ino = parentino;
316 	realm->parent = parent;
317 	list_add(&realm->child_item, &parent->children);
318 	return 1;
319 }
320 
321 
322 static int cmpu64_rev(const void *a, const void *b)
323 {
324 	if (*(u64 *)a < *(u64 *)b)
325 		return 1;
326 	if (*(u64 *)a > *(u64 *)b)
327 		return -1;
328 	return 0;
329 }
330 
331 
332 /*
333  * build the snap context for a given realm.
334  */
335 static int build_snap_context(struct ceph_mds_client *mdsc,
336 			      struct ceph_snap_realm *realm,
337 			      struct list_head *realm_queue,
338 			      struct list_head *dirty_realms)
339 {
340 	struct ceph_client *cl = mdsc->fsc->client;
341 	struct ceph_snap_realm *parent = realm->parent;
342 	struct ceph_snap_context *snapc;
343 	int err = 0;
344 	u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
345 
346 	/*
347 	 * build parent context, if it hasn't been built.
348 	 * conservatively estimate that all parent snaps might be
349 	 * included by us.
350 	 */
351 	if (parent) {
352 		if (!parent->cached_context) {
353 			/* add to the queue head */
354 			list_add(&parent->rebuild_item, realm_queue);
355 			return 1;
356 		}
357 		num += parent->cached_context->num_snaps;
358 	}
359 
360 	/* do i actually need to update?  not if my context seq
361 	   matches realm seq, and my parents' does to.  (this works
362 	   because we rebuild_snap_realms() works _downward_ in
363 	   hierarchy after each update.) */
364 	if (realm->cached_context &&
365 	    realm->cached_context->seq == realm->seq &&
366 	    (!parent ||
367 	     realm->cached_context->seq >= parent->cached_context->seq)) {
368 		doutc(cl, "%llx %p: %p seq %lld (%u snaps) (unchanged)\n",
369 		      realm->ino, realm, realm->cached_context,
370 		      realm->cached_context->seq,
371 		      (unsigned int)realm->cached_context->num_snaps);
372 		return 0;
373 	}
374 
375 	/* alloc new snap context */
376 	err = -ENOMEM;
377 	if ((size_t)num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
378 		goto fail;
379 	snapc = ceph_create_snap_context(num, GFP_NOFS);
380 	if (!snapc)
381 		goto fail;
382 
383 	/* build (reverse sorted) snap vector */
384 	num = 0;
385 	snapc->seq = realm->seq;
386 	if (parent) {
387 		u32 i;
388 
389 		/* include any of parent's snaps occurring _after_ my
390 		   parent became my parent */
391 		for (i = 0; i < parent->cached_context->num_snaps; i++)
392 			if (parent->cached_context->snaps[i] >=
393 			    realm->parent_since)
394 				snapc->snaps[num++] =
395 					parent->cached_context->snaps[i];
396 		if (parent->cached_context->seq > snapc->seq)
397 			snapc->seq = parent->cached_context->seq;
398 	}
399 	memcpy(snapc->snaps + num, realm->snaps,
400 	       sizeof(u64)*realm->num_snaps);
401 	num += realm->num_snaps;
402 	memcpy(snapc->snaps + num, realm->prior_parent_snaps,
403 	       sizeof(u64)*realm->num_prior_parent_snaps);
404 	num += realm->num_prior_parent_snaps;
405 
406 	sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
407 	snapc->num_snaps = num;
408 	doutc(cl, "%llx %p: %p seq %lld (%u snaps)\n", realm->ino, realm,
409 	      snapc, snapc->seq, (unsigned int) snapc->num_snaps);
410 
411 	ceph_put_snap_context(realm->cached_context);
412 	realm->cached_context = snapc;
413 	/* queue realm for cap_snap creation */
414 	list_add_tail(&realm->dirty_item, dirty_realms);
415 	return 0;
416 
417 fail:
418 	/*
419 	 * if we fail, clear old (incorrect) cached_context... hopefully
420 	 * we'll have better luck building it later
421 	 */
422 	if (realm->cached_context) {
423 		ceph_put_snap_context(realm->cached_context);
424 		realm->cached_context = NULL;
425 	}
426 	pr_err_client(cl, "%llx %p fail %d\n", realm->ino, realm, err);
427 	return err;
428 }
429 
430 /*
431  * rebuild snap context for the given realm and all of its children.
432  */
433 static void rebuild_snap_realms(struct ceph_mds_client *mdsc,
434 				struct ceph_snap_realm *realm,
435 				struct list_head *dirty_realms)
436 {
437 	struct ceph_client *cl = mdsc->fsc->client;
438 	LIST_HEAD(realm_queue);
439 	int last = 0;
440 	bool skip = false;
441 
442 	list_add_tail(&realm->rebuild_item, &realm_queue);
443 
444 	while (!list_empty(&realm_queue)) {
445 		struct ceph_snap_realm *_realm, *child;
446 
447 		_realm = list_first_entry(&realm_queue,
448 					  struct ceph_snap_realm,
449 					  rebuild_item);
450 
451 		/*
452 		 * If the last building failed dues to memory
453 		 * issue, just empty the realm_queue and return
454 		 * to avoid infinite loop.
455 		 */
456 		if (last < 0) {
457 			list_del_init(&_realm->rebuild_item);
458 			continue;
459 		}
460 
461 		last = build_snap_context(mdsc, _realm, &realm_queue,
462 					  dirty_realms);
463 		doutc(cl, "%llx %p, %s\n", realm->ino, realm,
464 		      last > 0 ? "is deferred" : !last ? "succeeded" : "failed");
465 
466 		/* is any child in the list ? */
467 		list_for_each_entry(child, &_realm->children, child_item) {
468 			if (!list_empty(&child->rebuild_item)) {
469 				skip = true;
470 				break;
471 			}
472 		}
473 
474 		if (!skip) {
475 			list_for_each_entry(child, &_realm->children, child_item)
476 				list_add_tail(&child->rebuild_item, &realm_queue);
477 		}
478 
479 		/* last == 1 means need to build parent first */
480 		if (last <= 0)
481 			list_del_init(&_realm->rebuild_item);
482 	}
483 }
484 
485 
486 /*
487  * helper to allocate and decode an array of snapids.  free prior
488  * instance, if any.
489  */
490 static int dup_array(u64 **dst, __le64 *src, u32 num)
491 {
492 	u32 i;
493 
494 	kfree(*dst);
495 	if (num) {
496 		*dst = kzalloc_objs(u64, num, GFP_NOFS);
497 		if (!*dst)
498 			return -ENOMEM;
499 		for (i = 0; i < num; i++)
500 			(*dst)[i] = get_unaligned_le64(src + i);
501 	} else {
502 		*dst = NULL;
503 	}
504 	return 0;
505 }
506 
507 static bool has_new_snaps(struct ceph_snap_context *o,
508 			  struct ceph_snap_context *n)
509 {
510 	if (n->num_snaps == 0)
511 		return false;
512 	/* snaps are in descending order */
513 	return n->snaps[0] > o->seq;
514 }
515 
516 /*
517  * When a snapshot is applied, the size/mtime inode metadata is queued
518  * in a ceph_cap_snap (one for each snapshot) until writeback
519  * completes and the metadata can be flushed back to the MDS.
520  *
521  * However, if a (sync) write is currently in-progress when we apply
522  * the snapshot, we have to wait until the write succeeds or fails
523  * (and a final size/mtime is known).  In this case the
524  * cap_snap->writing = true, and is said to be "pending."  When the write
525  * finishes, we __ceph_finish_cap_snap().
526  *
527  * Caller must hold snap_rwsem for read (i.e., the realm topology won't
528  * change).
529  */
530 static void ceph_queue_cap_snap(struct ceph_inode_info *ci,
531 				struct ceph_cap_snap **pcapsnap)
532 {
533 	struct inode *inode = &ci->netfs.inode;
534 	struct ceph_client *cl = ceph_inode_to_client(inode);
535 	struct ceph_snap_context *old_snapc, *new_snapc;
536 	struct ceph_cap_snap *capsnap = *pcapsnap;
537 	struct ceph_buffer *old_blob = NULL;
538 	int used, dirty;
539 
540 	spin_lock(&ci->i_ceph_lock);
541 	used = __ceph_caps_used(ci);
542 	dirty = __ceph_caps_dirty(ci);
543 
544 	old_snapc = ci->i_head_snapc;
545 	new_snapc = ci->i_snap_realm->cached_context;
546 
547 	/*
548 	 * If there is a write in progress, treat that as a dirty Fw,
549 	 * even though it hasn't completed yet; by the time we finish
550 	 * up this capsnap it will be.
551 	 */
552 	if (used & CEPH_CAP_FILE_WR)
553 		dirty |= CEPH_CAP_FILE_WR;
554 
555 	if (__ceph_have_pending_cap_snap(ci)) {
556 		/* there is no point in queuing multiple "pending" cap_snaps,
557 		   as no new writes are allowed to start when pending, so any
558 		   writes in progress now were started before the previous
559 		   cap_snap.  lucky us. */
560 		doutc(cl, "%p %llx.%llx already pending\n", inode,
561 		      ceph_vinop(inode));
562 		goto update_snapc;
563 	}
564 	if (ci->i_wrbuffer_ref_head == 0 &&
565 	    !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
566 		doutc(cl, "%p %llx.%llx nothing dirty|writing\n", inode,
567 		      ceph_vinop(inode));
568 		goto update_snapc;
569 	}
570 
571 	BUG_ON(!old_snapc);
572 
573 	/*
574 	 * There is no need to send FLUSHSNAP message to MDS if there is
575 	 * no new snapshot. But when there is dirty pages or on-going
576 	 * writes, we still need to create cap_snap. cap_snap is needed
577 	 * by the write path and page writeback path.
578 	 *
579 	 * also see ceph_try_drop_cap_snap()
580 	 */
581 	if (has_new_snaps(old_snapc, new_snapc)) {
582 		if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
583 			capsnap->need_flush = true;
584 	} else {
585 		if (!(used & CEPH_CAP_FILE_WR) &&
586 		    ci->i_wrbuffer_ref_head == 0) {
587 			doutc(cl, "%p %llx.%llx no new_snap|dirty_page|writing\n",
588 			      inode, ceph_vinop(inode));
589 			goto update_snapc;
590 		}
591 	}
592 
593 	doutc(cl, "%p %llx.%llx cap_snap %p queuing under %p %s %s\n",
594 	      inode, ceph_vinop(inode), capsnap, old_snapc,
595 	      ceph_cap_string(dirty), capsnap->need_flush ? "" : "no_flush");
596 	ihold(inode);
597 
598 	capsnap->follows = old_snapc->seq;
599 	capsnap->issued = __ceph_caps_issued(ci, NULL);
600 	capsnap->dirty = dirty;
601 
602 	capsnap->mode = inode->i_mode;
603 	capsnap->uid = inode->i_uid;
604 	capsnap->gid = inode->i_gid;
605 
606 	if (dirty & CEPH_CAP_XATTR_EXCL) {
607 		old_blob = __ceph_build_xattrs_blob(ci);
608 		capsnap->xattr_blob =
609 			ceph_buffer_get(ci->i_xattrs.blob);
610 		capsnap->xattr_version = ci->i_xattrs.version;
611 	} else {
612 		capsnap->xattr_blob = NULL;
613 		capsnap->xattr_version = 0;
614 	}
615 
616 	capsnap->inline_data = ceph_inline_version(ci) != CEPH_INLINE_NONE;
617 
618 	/* dirty page count moved from _head to this cap_snap;
619 	   all subsequent writes page dirties occur _after_ this
620 	   snapshot. */
621 	capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
622 	ci->i_wrbuffer_ref_head = 0;
623 	capsnap->context = old_snapc;
624 	list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
625 
626 	if (used & CEPH_CAP_FILE_WR) {
627 		doutc(cl, "%p %llx.%llx cap_snap %p snapc %p seq %llu used WR,"
628 		      " now pending\n", inode, ceph_vinop(inode), capsnap,
629 		      old_snapc, old_snapc->seq);
630 		capsnap->writing = true;
631 	} else {
632 		/* note mtime, size NOW. */
633 		__ceph_finish_cap_snap(ci, capsnap);
634 	}
635 	*pcapsnap = NULL;
636 	old_snapc = NULL;
637 
638 update_snapc:
639 	if (ci->i_wrbuffer_ref_head == 0 &&
640 	    ci->i_wr_ref == 0 &&
641 	    ci->i_dirty_caps == 0 &&
642 	    ci->i_flushing_caps == 0) {
643 		ci->i_head_snapc = NULL;
644 	} else {
645 		ci->i_head_snapc = ceph_get_snap_context(new_snapc);
646 		doutc(cl, " new snapc is %p\n", new_snapc);
647 	}
648 	spin_unlock(&ci->i_ceph_lock);
649 
650 	ceph_buffer_put(old_blob);
651 	ceph_put_snap_context(old_snapc);
652 }
653 
654 /*
655  * Finalize the size, mtime for a cap_snap.. that is, settle on final values
656  * to be used for the snapshot, to be flushed back to the mds.
657  *
658  * If capsnap can now be flushed, add to snap_flush list, and return 1.
659  *
660  * Caller must hold i_ceph_lock.
661  */
662 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
663 			    struct ceph_cap_snap *capsnap)
664 {
665 	struct inode *inode = &ci->netfs.inode;
666 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
667 	struct ceph_client *cl = mdsc->fsc->client;
668 
669 	BUG_ON(capsnap->writing);
670 	capsnap->size = i_size_read(inode);
671 	capsnap->mtime = inode_get_mtime(inode);
672 	capsnap->atime = inode_get_atime(inode);
673 	capsnap->ctime = inode_get_ctime(inode);
674 	capsnap->btime = ci->i_btime;
675 	capsnap->change_attr = inode_peek_iversion_raw(inode);
676 	capsnap->time_warp_seq = ci->i_time_warp_seq;
677 	capsnap->truncate_size = ci->i_truncate_size;
678 	capsnap->truncate_seq = ci->i_truncate_seq;
679 	if (capsnap->dirty_pages) {
680 		doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s "
681 		      "s=%llu still has %d dirty pages\n", inode,
682 		      ceph_vinop(inode), capsnap, capsnap->context,
683 		      capsnap->context->seq,
684 		      ceph_cap_string(capsnap->dirty),
685 		      capsnap->size, capsnap->dirty_pages);
686 		return 0;
687 	}
688 
689 	/*
690 	 * Defer flushing the capsnap if the dirty buffer not flushed yet.
691 	 * And trigger to flush the buffer immediately.
692 	 */
693 	if (ci->i_wrbuffer_ref) {
694 		doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s "
695 		      "s=%llu used WRBUFFER, delaying\n", inode,
696 		      ceph_vinop(inode), capsnap, capsnap->context,
697 		      capsnap->context->seq, ceph_cap_string(capsnap->dirty),
698 		      capsnap->size);
699 		ceph_queue_writeback(inode);
700 		return 0;
701 	}
702 
703 	set_bit(CEPH_I_FLUSH_SNAPS_BIT, &ci->i_ceph_flags);
704 	doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu\n",
705 	      inode, ceph_vinop(inode), capsnap, capsnap->context,
706 	      capsnap->context->seq, ceph_cap_string(capsnap->dirty),
707 	      capsnap->size);
708 
709 	spin_lock(&mdsc->snap_flush_lock);
710 	if (list_empty(&ci->i_snap_flush_item)) {
711 		ihold(inode);
712 		list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
713 	}
714 	spin_unlock(&mdsc->snap_flush_lock);
715 	return 1;  /* caller may want to ceph_flush_snaps */
716 }
717 
718 /*
719  * Queue cap_snaps for snap writeback for this realm and its children.
720  * Called under snap_rwsem, so realm topology won't change.
721  */
722 static void queue_realm_cap_snaps(struct ceph_mds_client *mdsc,
723 				  struct ceph_snap_realm *realm)
724 {
725 	struct ceph_client *cl = mdsc->fsc->client;
726 	struct ceph_inode_info *ci;
727 	struct inode *lastinode = NULL;
728 	struct ceph_cap_snap *capsnap = NULL;
729 
730 	doutc(cl, "%p %llx inode\n", realm, realm->ino);
731 
732 	spin_lock(&realm->inodes_with_caps_lock);
733 	list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
734 		struct inode *inode = igrab(&ci->netfs.inode);
735 		if (!inode)
736 			continue;
737 		spin_unlock(&realm->inodes_with_caps_lock);
738 		iput(lastinode);
739 		lastinode = inode;
740 
741 		/*
742 		 * Allocate the capsnap memory outside of ceph_queue_cap_snap()
743 		 * to reduce very possible but unnecessary frequently memory
744 		 * allocate/free in this loop.
745 		 */
746 		if (!capsnap) {
747 			capsnap = kmem_cache_zalloc(ceph_cap_snap_cachep, GFP_NOFS);
748 			if (!capsnap) {
749 				pr_err_client(cl,
750 					"ENOMEM allocating ceph_cap_snap on %p\n",
751 					inode);
752 				return;
753 			}
754 		}
755 		capsnap->cap_flush.is_capsnap = true;
756 		refcount_set(&capsnap->nref, 1);
757 		INIT_LIST_HEAD(&capsnap->cap_flush.i_list);
758 		INIT_LIST_HEAD(&capsnap->cap_flush.g_list);
759 		INIT_LIST_HEAD(&capsnap->ci_item);
760 
761 		ceph_queue_cap_snap(ci, &capsnap);
762 		spin_lock(&realm->inodes_with_caps_lock);
763 	}
764 	spin_unlock(&realm->inodes_with_caps_lock);
765 	iput(lastinode);
766 
767 	if (capsnap)
768 		kmem_cache_free(ceph_cap_snap_cachep, capsnap);
769 	doutc(cl, "%p %llx done\n", realm, realm->ino);
770 }
771 
772 /*
773  * Parse and apply a snapblob "snap trace" from the MDS.  This specifies
774  * the snap realm parameters from a given realm and all of its ancestors,
775  * up to the root.
776  *
777  * Caller must hold snap_rwsem for write.
778  */
779 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
780 			   void *p, void *e, bool deletion,
781 			   struct ceph_snap_realm **realm_ret)
782 {
783 	struct ceph_client *cl = mdsc->fsc->client;
784 	struct ceph_mds_snap_realm *ri;    /* encoded */
785 	__le64 *snaps;                     /* encoded */
786 	__le64 *prior_parent_snaps;        /* encoded */
787 	struct ceph_snap_realm *realm;
788 	struct ceph_snap_realm *first_realm = NULL;
789 	struct ceph_snap_realm *realm_to_rebuild = NULL;
790 	struct ceph_client *client = mdsc->fsc->client;
791 	size_t snaps_len;
792 	int rebuild_snapcs;
793 	int err = -ENOMEM;
794 	int ret;
795 	LIST_HEAD(dirty_realms);
796 
797 	lockdep_assert_held_write(&mdsc->snap_rwsem);
798 
799 	doutc(cl, "deletion=%d\n", deletion);
800 more:
801 	realm = NULL;
802 	rebuild_snapcs = 0;
803 	ceph_decode_need(&p, e, sizeof(*ri), bad);
804 	ri = p;
805 	p += sizeof(*ri);
806 	snaps_len = size_mul(size_add(le32_to_cpu(ri->num_snaps),
807 				      le32_to_cpu(ri->num_prior_parent_snaps)),
808 			     sizeof(u64));
809 	ceph_decode_need(&p, e, snaps_len, bad);
810 	snaps = p;
811 	p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
812 	prior_parent_snaps = p;
813 	p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
814 
815 	realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
816 	if (!realm) {
817 		realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
818 		if (IS_ERR(realm)) {
819 			err = PTR_ERR(realm);
820 			goto fail;
821 		}
822 	}
823 
824 	/* ensure the parent is correct */
825 	err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
826 	if (err < 0)
827 		goto fail;
828 	rebuild_snapcs += err;
829 
830 	if (le64_to_cpu(ri->seq) > realm->seq) {
831 		doutc(cl, "updating %llx %p %lld -> %lld\n", realm->ino,
832 		      realm, realm->seq, le64_to_cpu(ri->seq));
833 		/* update realm parameters, snap lists */
834 		realm->seq = le64_to_cpu(ri->seq);
835 		realm->created = le64_to_cpu(ri->created);
836 		realm->parent_since = le64_to_cpu(ri->parent_since);
837 
838 		realm->num_snaps = le32_to_cpu(ri->num_snaps);
839 		err = dup_array(&realm->snaps, snaps, realm->num_snaps);
840 		if (err < 0)
841 			goto fail;
842 
843 		realm->num_prior_parent_snaps =
844 			le32_to_cpu(ri->num_prior_parent_snaps);
845 		err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
846 				realm->num_prior_parent_snaps);
847 		if (err < 0)
848 			goto fail;
849 
850 		if (realm->seq > mdsc->last_snap_seq)
851 			mdsc->last_snap_seq = realm->seq;
852 
853 		rebuild_snapcs = 1;
854 	} else if (!realm->cached_context) {
855 		doutc(cl, "%llx %p seq %lld new\n", realm->ino, realm,
856 		      realm->seq);
857 		rebuild_snapcs = 1;
858 	} else {
859 		doutc(cl, "%llx %p seq %lld unchanged\n", realm->ino, realm,
860 		      realm->seq);
861 	}
862 
863 	doutc(cl, "done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
864 	      realm, rebuild_snapcs, p, e);
865 
866 	/*
867 	 * this will always track the uppest parent realm from which
868 	 * we need to rebuild the snapshot contexts _downward_ in
869 	 * hierarchy.
870 	 */
871 	if (rebuild_snapcs)
872 		realm_to_rebuild = realm;
873 
874 	/* rebuild_snapcs when we reach the _end_ (root) of the trace */
875 	if (realm_to_rebuild && p >= e)
876 		rebuild_snap_realms(mdsc, realm_to_rebuild, &dirty_realms);
877 
878 	if (!first_realm)
879 		first_realm = realm;
880 	else
881 		ceph_put_snap_realm(mdsc, realm);
882 
883 	if (p < e)
884 		goto more;
885 
886 	/*
887 	 * queue cap snaps _after_ we've built the new snap contexts,
888 	 * so that i_head_snapc can be set appropriately.
889 	 */
890 	while (!list_empty(&dirty_realms)) {
891 		realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
892 					 dirty_item);
893 		list_del_init(&realm->dirty_item);
894 		queue_realm_cap_snaps(mdsc, realm);
895 	}
896 
897 	if (realm_ret)
898 		*realm_ret = first_realm;
899 	else
900 		ceph_put_snap_realm(mdsc, first_realm);
901 
902 	__cleanup_empty_realms(mdsc);
903 	return 0;
904 
905 bad:
906 	err = -EIO;
907 fail:
908 	if (realm && !IS_ERR(realm))
909 		ceph_put_snap_realm(mdsc, realm);
910 	if (first_realm)
911 		ceph_put_snap_realm(mdsc, first_realm);
912 	pr_err_client(cl, "error %d\n", err);
913 
914 	/*
915 	 * When receiving a corrupted snap trace we don't know what
916 	 * exactly has happened in MDS side. And we shouldn't continue
917 	 * writing to OSD, which may corrupt the snapshot contents.
918 	 *
919 	 * Just try to blocklist this kclient and then this kclient
920 	 * must be remounted to continue after the corrupted metadata
921 	 * fixed in the MDS side.
922 	 */
923 	WRITE_ONCE(mdsc->fsc->mount_state, CEPH_MOUNT_FENCE_IO);
924 	ret = ceph_monc_blocklist_add(&client->monc, &client->msgr.inst.addr);
925 	if (ret)
926 		pr_err_client(cl, "failed to blocklist %s: %d\n",
927 			      ceph_pr_addr(&client->msgr.inst.addr), ret);
928 
929 	WARN(1, "[client.%lld] %s %s%sdo remount to continue%s",
930 	     client->monc.auth->global_id, __func__,
931 	     ret ? "" : ceph_pr_addr(&client->msgr.inst.addr),
932 	     ret ? "" : " was blocklisted, ",
933 	     err == -EIO ? " after corrupted snaptrace is fixed" : "");
934 
935 	return err;
936 }
937 
938 
939 /*
940  * Send any cap_snaps that are queued for flush.  Try to carry
941  * s_mutex across multiple snap flushes to avoid locking overhead.
942  *
943  * Caller holds no locks.
944  */
945 static void flush_snaps(struct ceph_mds_client *mdsc)
946 {
947 	struct ceph_client *cl = mdsc->fsc->client;
948 	struct ceph_inode_info *ci;
949 	struct inode *inode;
950 	struct ceph_mds_session *session = NULL;
951 
952 	doutc(cl, "begin\n");
953 	spin_lock(&mdsc->snap_flush_lock);
954 	while (!list_empty(&mdsc->snap_flush_list)) {
955 		ci = list_first_entry(&mdsc->snap_flush_list,
956 				struct ceph_inode_info, i_snap_flush_item);
957 		inode = &ci->netfs.inode;
958 		ihold(inode);
959 		spin_unlock(&mdsc->snap_flush_lock);
960 		ceph_flush_snaps(ci, &session);
961 		iput(inode);
962 		spin_lock(&mdsc->snap_flush_lock);
963 	}
964 	spin_unlock(&mdsc->snap_flush_lock);
965 
966 	ceph_put_mds_session(session);
967 	doutc(cl, "done\n");
968 }
969 
970 /**
971  * ceph_change_snap_realm - change the snap_realm for an inode
972  * @inode: inode to move to new snap realm
973  * @realm: new realm to move inode into (may be NULL)
974  *
975  * Detach an inode from its old snaprealm (if any) and attach it to
976  * the new snaprealm (if any). The old snap realm reference held by
977  * the inode is put. If realm is non-NULL, then the caller's reference
978  * to it is taken over by the inode.
979  */
980 void ceph_change_snap_realm(struct inode *inode, struct ceph_snap_realm *realm)
981 {
982 	struct ceph_inode_info *ci = ceph_inode(inode);
983 	struct ceph_mds_client *mdsc = ceph_inode_to_fs_client(inode)->mdsc;
984 	struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
985 
986 	lockdep_assert_held(&ci->i_ceph_lock);
987 
988 	if (oldrealm) {
989 		spin_lock(&oldrealm->inodes_with_caps_lock);
990 		list_del_init(&ci->i_snap_realm_item);
991 		if (oldrealm->ino == ceph_ino(inode))
992 			oldrealm->inode = NULL;
993 		spin_unlock(&oldrealm->inodes_with_caps_lock);
994 		ceph_put_snap_realm(mdsc, oldrealm);
995 	}
996 
997 	ci->i_snap_realm = realm;
998 
999 	if (realm) {
1000 		spin_lock(&realm->inodes_with_caps_lock);
1001 		list_add(&ci->i_snap_realm_item, &realm->inodes_with_caps);
1002 		if (realm->ino == ceph_ino(inode))
1003 			realm->inode = inode;
1004 		spin_unlock(&realm->inodes_with_caps_lock);
1005 	}
1006 }
1007 
1008 /*
1009  * Handle a snap notification from the MDS.
1010  *
1011  * This can take two basic forms: the simplest is just a snap creation
1012  * or deletion notification on an existing realm.  This should update the
1013  * realm and its children.
1014  *
1015  * The more difficult case is realm creation, due to snap creation at a
1016  * new point in the file hierarchy, or due to a rename that moves a file or
1017  * directory into another realm.
1018  */
1019 void ceph_handle_snap(struct ceph_mds_client *mdsc,
1020 		      struct ceph_mds_session *session,
1021 		      struct ceph_msg *msg)
1022 {
1023 	struct ceph_client *cl = mdsc->fsc->client;
1024 	struct super_block *sb = mdsc->fsc->sb;
1025 	int mds = session->s_mds;
1026 	u64 split;
1027 	int op;
1028 	int trace_len;
1029 	struct ceph_snap_realm *realm = NULL;
1030 	void *p = msg->front.iov_base;
1031 	void *e = p + msg->front.iov_len;
1032 	struct ceph_mds_snap_head *h;
1033 	int num_split_inos, num_split_realms;
1034 	__le64 *split_inos = NULL, *split_realms = NULL;
1035 	int i;
1036 	int locked_rwsem = 0;
1037 	bool close_sessions = false;
1038 
1039 	if (!ceph_inc_mds_stopping_blocker(mdsc, session))
1040 		return;
1041 
1042 	/* decode */
1043 	if (msg->front.iov_len < sizeof(*h))
1044 		goto bad;
1045 	h = p;
1046 	op = le32_to_cpu(h->op);
1047 	split = le64_to_cpu(h->split);   /* non-zero if we are splitting an
1048 					  * existing realm */
1049 	num_split_inos = le32_to_cpu(h->num_split_inos);
1050 	num_split_realms = le32_to_cpu(h->num_split_realms);
1051 	trace_len = le32_to_cpu(h->trace_len);
1052 	p += sizeof(*h);
1053 
1054 	doutc(cl, "from mds%d op %s split %llx tracelen %d\n", mds,
1055 	      ceph_snap_op_name(op), split, trace_len);
1056 
1057 	down_write(&mdsc->snap_rwsem);
1058 	locked_rwsem = 1;
1059 
1060 	if (op == CEPH_SNAP_OP_SPLIT) {
1061 		struct ceph_mds_snap_realm *ri;
1062 
1063 		/*
1064 		 * A "split" breaks part of an existing realm off into
1065 		 * a new realm.  The MDS provides a list of inodes
1066 		 * (with caps) and child realms that belong to the new
1067 		 * child.
1068 		 */
1069 		split_inos = p;
1070 		p += sizeof(u64) * num_split_inos;
1071 		split_realms = p;
1072 		p += sizeof(u64) * num_split_realms;
1073 		ceph_decode_need(&p, e, sizeof(*ri), bad);
1074 		/* we will peek at realm info here, but will _not_
1075 		 * advance p, as the realm update will occur below in
1076 		 * ceph_update_snap_trace. */
1077 		ri = p;
1078 
1079 		realm = ceph_lookup_snap_realm(mdsc, split);
1080 		if (!realm) {
1081 			realm = ceph_create_snap_realm(mdsc, split);
1082 			if (IS_ERR(realm))
1083 				goto out;
1084 		}
1085 
1086 		doutc(cl, "splitting snap_realm %llx %p\n", realm->ino, realm);
1087 		for (i = 0; i < num_split_inos; i++) {
1088 			struct ceph_vino vino = {
1089 				.ino = le64_to_cpu(split_inos[i]),
1090 				.snap = CEPH_NOSNAP,
1091 			};
1092 			struct inode *inode = ceph_find_inode(sb, vino);
1093 			struct ceph_inode_info *ci;
1094 
1095 			if (!inode)
1096 				continue;
1097 			ci = ceph_inode(inode);
1098 
1099 			spin_lock(&ci->i_ceph_lock);
1100 			if (!ci->i_snap_realm)
1101 				goto skip_inode;
1102 			/*
1103 			 * If this inode belongs to a realm that was
1104 			 * created after our new realm, we experienced
1105 			 * a race (due to another split notifications
1106 			 * arriving from a different MDS).  So skip
1107 			 * this inode.
1108 			 */
1109 			if (ci->i_snap_realm->created >
1110 			    le64_to_cpu(ri->created)) {
1111 				doutc(cl, " leaving %p %llx.%llx in newer realm %llx %p\n",
1112 				      inode, ceph_vinop(inode), ci->i_snap_realm->ino,
1113 				      ci->i_snap_realm);
1114 				goto skip_inode;
1115 			}
1116 			doutc(cl, " will move %p %llx.%llx to split realm %llx %p\n",
1117 			      inode, ceph_vinop(inode), realm->ino, realm);
1118 
1119 			ceph_get_snap_realm(mdsc, realm);
1120 			ceph_change_snap_realm(inode, realm);
1121 			spin_unlock(&ci->i_ceph_lock);
1122 			iput(inode);
1123 			continue;
1124 
1125 skip_inode:
1126 			spin_unlock(&ci->i_ceph_lock);
1127 			iput(inode);
1128 		}
1129 
1130 		/* we may have taken some of the old realm's children. */
1131 		for (i = 0; i < num_split_realms; i++) {
1132 			struct ceph_snap_realm *child =
1133 				__lookup_snap_realm(mdsc,
1134 					   le64_to_cpu(split_realms[i]));
1135 			if (!child)
1136 				continue;
1137 			adjust_snap_realm_parent(mdsc, child, realm->ino);
1138 		}
1139 	} else {
1140 		/*
1141 		 * In the non-split case both 'num_split_inos' and
1142 		 * 'num_split_realms' should be 0, making this a no-op.
1143 		 * However the MDS happens to populate 'split_realms' list
1144 		 * in one of the UPDATE op cases by mistake.
1145 		 *
1146 		 * Skip both lists just in case to ensure that 'p' is
1147 		 * positioned at the start of realm info, as expected by
1148 		 * ceph_update_snap_trace().
1149 		 */
1150 		p += sizeof(u64) * num_split_inos;
1151 		p += sizeof(u64) * num_split_realms;
1152 	}
1153 
1154 	/*
1155 	 * update using the provided snap trace. if we are deleting a
1156 	 * snap, we can avoid queueing cap_snaps.
1157 	 */
1158 	if (ceph_update_snap_trace(mdsc, p, e,
1159 				   op == CEPH_SNAP_OP_DESTROY,
1160 				   NULL)) {
1161 		close_sessions = true;
1162 		goto bad;
1163 	}
1164 
1165 	if (op == CEPH_SNAP_OP_SPLIT)
1166 		/* we took a reference when we created the realm, above */
1167 		ceph_put_snap_realm(mdsc, realm);
1168 
1169 	__cleanup_empty_realms(mdsc);
1170 
1171 	up_write(&mdsc->snap_rwsem);
1172 
1173 	flush_snaps(mdsc);
1174 	ceph_dec_mds_stopping_blocker(mdsc);
1175 	return;
1176 
1177 bad:
1178 	pr_err_client(cl, "corrupt snap message from mds%d\n", mds);
1179 	ceph_msg_dump(msg);
1180 out:
1181 	if (locked_rwsem)
1182 		up_write(&mdsc->snap_rwsem);
1183 
1184 	ceph_dec_mds_stopping_blocker(mdsc);
1185 
1186 	if (close_sessions)
1187 		ceph_mdsc_close_sessions(mdsc);
1188 	return;
1189 }
1190 
1191 struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1192 					    u64 snap)
1193 {
1194 	struct ceph_client *cl = mdsc->fsc->client;
1195 	struct ceph_snapid_map *sm, *exist;
1196 	struct rb_node **p, *parent;
1197 	int ret;
1198 
1199 	exist = NULL;
1200 	spin_lock(&mdsc->snapid_map_lock);
1201 	p = &mdsc->snapid_map_tree.rb_node;
1202 	while (*p) {
1203 		exist = rb_entry(*p, struct ceph_snapid_map, node);
1204 		if (snap > exist->snap) {
1205 			p = &(*p)->rb_left;
1206 		} else if (snap < exist->snap) {
1207 			p = &(*p)->rb_right;
1208 		} else {
1209 			if (atomic_inc_return(&exist->ref) == 1)
1210 				list_del_init(&exist->lru);
1211 			break;
1212 		}
1213 		exist = NULL;
1214 	}
1215 	spin_unlock(&mdsc->snapid_map_lock);
1216 	if (exist) {
1217 		doutc(cl, "found snapid map %llx -> %x\n", exist->snap,
1218 		      exist->dev);
1219 		return exist;
1220 	}
1221 
1222 	sm = kmalloc_obj(*sm, GFP_NOFS);
1223 	if (!sm)
1224 		return NULL;
1225 
1226 	ret = get_anon_bdev(&sm->dev);
1227 	if (ret < 0) {
1228 		kfree(sm);
1229 		return NULL;
1230 	}
1231 
1232 	INIT_LIST_HEAD(&sm->lru);
1233 	atomic_set(&sm->ref, 1);
1234 	sm->snap = snap;
1235 
1236 	exist = NULL;
1237 	parent = NULL;
1238 	p = &mdsc->snapid_map_tree.rb_node;
1239 	spin_lock(&mdsc->snapid_map_lock);
1240 	while (*p) {
1241 		parent = *p;
1242 		exist = rb_entry(*p, struct ceph_snapid_map, node);
1243 		if (snap > exist->snap)
1244 			p = &(*p)->rb_left;
1245 		else if (snap < exist->snap)
1246 			p = &(*p)->rb_right;
1247 		else
1248 			break;
1249 		exist = NULL;
1250 	}
1251 	if (exist) {
1252 		if (atomic_inc_return(&exist->ref) == 1)
1253 			list_del_init(&exist->lru);
1254 	} else {
1255 		rb_link_node(&sm->node, parent, p);
1256 		rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1257 	}
1258 	spin_unlock(&mdsc->snapid_map_lock);
1259 	if (exist) {
1260 		free_anon_bdev(sm->dev);
1261 		kfree(sm);
1262 		doutc(cl, "found snapid map %llx -> %x\n", exist->snap,
1263 		      exist->dev);
1264 		return exist;
1265 	}
1266 
1267 	doutc(cl, "create snapid map %llx -> %x\n", sm->snap, sm->dev);
1268 	return sm;
1269 }
1270 
1271 void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1272 			 struct ceph_snapid_map *sm)
1273 {
1274 	if (!sm)
1275 		return;
1276 	if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1277 		if (!RB_EMPTY_NODE(&sm->node)) {
1278 			sm->last_used = jiffies;
1279 			list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1280 			spin_unlock(&mdsc->snapid_map_lock);
1281 		} else {
1282 			/* already cleaned up by
1283 			 * ceph_cleanup_snapid_map() */
1284 			spin_unlock(&mdsc->snapid_map_lock);
1285 			kfree(sm);
1286 		}
1287 	}
1288 }
1289 
1290 void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1291 {
1292 	struct ceph_client *cl = mdsc->fsc->client;
1293 	struct ceph_snapid_map *sm;
1294 	unsigned long now;
1295 	LIST_HEAD(to_free);
1296 
1297 	spin_lock(&mdsc->snapid_map_lock);
1298 	now = jiffies;
1299 
1300 	while (!list_empty(&mdsc->snapid_map_lru)) {
1301 		sm = list_first_entry(&mdsc->snapid_map_lru,
1302 				      struct ceph_snapid_map, lru);
1303 		if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1304 			break;
1305 
1306 		rb_erase(&sm->node, &mdsc->snapid_map_tree);
1307 		list_move(&sm->lru, &to_free);
1308 	}
1309 	spin_unlock(&mdsc->snapid_map_lock);
1310 
1311 	while (!list_empty(&to_free)) {
1312 		sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1313 		list_del(&sm->lru);
1314 		doutc(cl, "trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1315 		free_anon_bdev(sm->dev);
1316 		kfree(sm);
1317 	}
1318 }
1319 
1320 void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1321 {
1322 	struct ceph_client *cl = mdsc->fsc->client;
1323 	struct ceph_snapid_map *sm;
1324 	struct rb_node *p;
1325 	LIST_HEAD(to_free);
1326 
1327 	spin_lock(&mdsc->snapid_map_lock);
1328 	while ((p = rb_first(&mdsc->snapid_map_tree))) {
1329 		sm = rb_entry(p, struct ceph_snapid_map, node);
1330 		rb_erase(p, &mdsc->snapid_map_tree);
1331 		RB_CLEAR_NODE(p);
1332 		list_move(&sm->lru, &to_free);
1333 	}
1334 	spin_unlock(&mdsc->snapid_map_lock);
1335 
1336 	while (!list_empty(&to_free)) {
1337 		sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1338 		list_del(&sm->lru);
1339 		free_anon_bdev(sm->dev);
1340 		if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1341 			pr_err_client(cl, "snapid map %llx -> %x still in use\n",
1342 				      sm->snap, sm->dev);
1343 		}
1344 		kfree(sm);
1345 	}
1346 }
1347