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