xref: /openbmc/linux/fs/ceph/snap.c (revision 464a5728)
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  */
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 
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  */
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 	atomic_set(&realm->nref, 1);    /* for caller */
125 	realm->ino = ino;
126 	INIT_LIST_HEAD(&realm->children);
127 	INIT_LIST_HEAD(&realm->child_item);
128 	INIT_LIST_HEAD(&realm->empty_item);
129 	INIT_LIST_HEAD(&realm->dirty_item);
130 	INIT_LIST_HEAD(&realm->inodes_with_caps);
131 	spin_lock_init(&realm->inodes_with_caps_lock);
132 	__insert_snap_realm(&mdsc->snap_realms, realm);
133 	mdsc->num_snap_realms++;
134 
135 	dout("create_snap_realm %llx %p\n", realm->ino, realm);
136 	return realm;
137 }
138 
139 /*
140  * lookup the realm rooted at @ino.
141  *
142  * caller must hold snap_rwsem.
143  */
144 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
145 						   u64 ino)
146 {
147 	struct rb_node *n = mdsc->snap_realms.rb_node;
148 	struct ceph_snap_realm *r;
149 
150 	lockdep_assert_held(&mdsc->snap_rwsem);
151 
152 	while (n) {
153 		r = rb_entry(n, struct ceph_snap_realm, node);
154 		if (ino < r->ino)
155 			n = n->rb_left;
156 		else if (ino > r->ino)
157 			n = n->rb_right;
158 		else {
159 			dout("lookup_snap_realm %llx %p\n", r->ino, r);
160 			return r;
161 		}
162 	}
163 	return NULL;
164 }
165 
166 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
167 					       u64 ino)
168 {
169 	struct ceph_snap_realm *r;
170 	r = __lookup_snap_realm(mdsc, ino);
171 	if (r)
172 		ceph_get_snap_realm(mdsc, r);
173 	return r;
174 }
175 
176 static void __put_snap_realm(struct ceph_mds_client *mdsc,
177 			     struct ceph_snap_realm *realm);
178 
179 /*
180  * called with snap_rwsem (write)
181  */
182 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
183 				 struct ceph_snap_realm *realm)
184 {
185 	lockdep_assert_held_write(&mdsc->snap_rwsem);
186 
187 	dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
188 
189 	rb_erase(&realm->node, &mdsc->snap_realms);
190 	mdsc->num_snap_realms--;
191 
192 	if (realm->parent) {
193 		list_del_init(&realm->child_item);
194 		__put_snap_realm(mdsc, realm->parent);
195 	}
196 
197 	kfree(realm->prior_parent_snaps);
198 	kfree(realm->snaps);
199 	ceph_put_snap_context(realm->cached_context);
200 	kfree(realm);
201 }
202 
203 /*
204  * caller holds snap_rwsem (write)
205  */
206 static void __put_snap_realm(struct ceph_mds_client *mdsc,
207 			     struct ceph_snap_realm *realm)
208 {
209 	lockdep_assert_held_write(&mdsc->snap_rwsem);
210 
211 	/*
212 	 * We do not require the snap_empty_lock here, as any caller that
213 	 * increments the value must hold the snap_rwsem.
214 	 */
215 	if (atomic_dec_and_test(&realm->nref))
216 		__destroy_snap_realm(mdsc, realm);
217 }
218 
219 /*
220  * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
221  */
222 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
223 			 struct ceph_snap_realm *realm)
224 {
225 	if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
226 		return;
227 
228 	if (down_write_trylock(&mdsc->snap_rwsem)) {
229 		spin_unlock(&mdsc->snap_empty_lock);
230 		__destroy_snap_realm(mdsc, realm);
231 		up_write(&mdsc->snap_rwsem);
232 	} else {
233 		list_add(&realm->empty_item, &mdsc->snap_empty);
234 		spin_unlock(&mdsc->snap_empty_lock);
235 	}
236 }
237 
238 /*
239  * Clean up any realms whose ref counts have dropped to zero.  Note
240  * that this does not include realms who were created but not yet
241  * used.
242  *
243  * Called under snap_rwsem (write)
244  */
245 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
246 {
247 	struct ceph_snap_realm *realm;
248 
249 	lockdep_assert_held_write(&mdsc->snap_rwsem);
250 
251 	spin_lock(&mdsc->snap_empty_lock);
252 	while (!list_empty(&mdsc->snap_empty)) {
253 		realm = list_first_entry(&mdsc->snap_empty,
254 				   struct ceph_snap_realm, empty_item);
255 		list_del(&realm->empty_item);
256 		spin_unlock(&mdsc->snap_empty_lock);
257 		__destroy_snap_realm(mdsc, realm);
258 		spin_lock(&mdsc->snap_empty_lock);
259 	}
260 	spin_unlock(&mdsc->snap_empty_lock);
261 }
262 
263 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
264 {
265 	down_write(&mdsc->snap_rwsem);
266 	__cleanup_empty_realms(mdsc);
267 	up_write(&mdsc->snap_rwsem);
268 }
269 
270 /*
271  * adjust the parent realm of a given @realm.  adjust child list, and parent
272  * pointers, and ref counts appropriately.
273  *
274  * return true if parent was changed, 0 if unchanged, <0 on error.
275  *
276  * caller must hold snap_rwsem for write.
277  */
278 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
279 				    struct ceph_snap_realm *realm,
280 				    u64 parentino)
281 {
282 	struct ceph_snap_realm *parent;
283 
284 	lockdep_assert_held_write(&mdsc->snap_rwsem);
285 
286 	if (realm->parent_ino == parentino)
287 		return 0;
288 
289 	parent = ceph_lookup_snap_realm(mdsc, parentino);
290 	if (!parent) {
291 		parent = ceph_create_snap_realm(mdsc, parentino);
292 		if (IS_ERR(parent))
293 			return PTR_ERR(parent);
294 	}
295 	dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
296 	     realm->ino, realm, realm->parent_ino, realm->parent,
297 	     parentino, parent);
298 	if (realm->parent) {
299 		list_del_init(&realm->child_item);
300 		ceph_put_snap_realm(mdsc, realm->parent);
301 	}
302 	realm->parent_ino = parentino;
303 	realm->parent = parent;
304 	list_add(&realm->child_item, &parent->children);
305 	return 1;
306 }
307 
308 
309 static int cmpu64_rev(const void *a, const void *b)
310 {
311 	if (*(u64 *)a < *(u64 *)b)
312 		return 1;
313 	if (*(u64 *)a > *(u64 *)b)
314 		return -1;
315 	return 0;
316 }
317 
318 
319 /*
320  * build the snap context for a given realm.
321  */
322 static int build_snap_context(struct ceph_snap_realm *realm,
323 			      struct list_head* dirty_realms)
324 {
325 	struct ceph_snap_realm *parent = realm->parent;
326 	struct ceph_snap_context *snapc;
327 	int err = 0;
328 	u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
329 
330 	/*
331 	 * build parent context, if it hasn't been built.
332 	 * conservatively estimate that all parent snaps might be
333 	 * included by us.
334 	 */
335 	if (parent) {
336 		if (!parent->cached_context) {
337 			err = build_snap_context(parent, dirty_realms);
338 			if (err)
339 				goto fail;
340 		}
341 		num += parent->cached_context->num_snaps;
342 	}
343 
344 	/* do i actually need to update?  not if my context seq
345 	   matches realm seq, and my parents' does to.  (this works
346 	   because we rebuild_snap_realms() works _downward_ in
347 	   hierarchy after each update.) */
348 	if (realm->cached_context &&
349 	    realm->cached_context->seq == realm->seq &&
350 	    (!parent ||
351 	     realm->cached_context->seq >= parent->cached_context->seq)) {
352 		dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
353 		     " (unchanged)\n",
354 		     realm->ino, realm, realm->cached_context,
355 		     realm->cached_context->seq,
356 		     (unsigned int)realm->cached_context->num_snaps);
357 		return 0;
358 	}
359 
360 	/* alloc new snap context */
361 	err = -ENOMEM;
362 	if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
363 		goto fail;
364 	snapc = ceph_create_snap_context(num, GFP_NOFS);
365 	if (!snapc)
366 		goto fail;
367 
368 	/* build (reverse sorted) snap vector */
369 	num = 0;
370 	snapc->seq = realm->seq;
371 	if (parent) {
372 		u32 i;
373 
374 		/* include any of parent's snaps occurring _after_ my
375 		   parent became my parent */
376 		for (i = 0; i < parent->cached_context->num_snaps; i++)
377 			if (parent->cached_context->snaps[i] >=
378 			    realm->parent_since)
379 				snapc->snaps[num++] =
380 					parent->cached_context->snaps[i];
381 		if (parent->cached_context->seq > snapc->seq)
382 			snapc->seq = parent->cached_context->seq;
383 	}
384 	memcpy(snapc->snaps + num, realm->snaps,
385 	       sizeof(u64)*realm->num_snaps);
386 	num += realm->num_snaps;
387 	memcpy(snapc->snaps + num, realm->prior_parent_snaps,
388 	       sizeof(u64)*realm->num_prior_parent_snaps);
389 	num += realm->num_prior_parent_snaps;
390 
391 	sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
392 	snapc->num_snaps = num;
393 	dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
394 	     realm->ino, realm, snapc, snapc->seq,
395 	     (unsigned int) snapc->num_snaps);
396 
397 	ceph_put_snap_context(realm->cached_context);
398 	realm->cached_context = snapc;
399 	/* queue realm for cap_snap creation */
400 	list_add_tail(&realm->dirty_item, dirty_realms);
401 	return 0;
402 
403 fail:
404 	/*
405 	 * if we fail, clear old (incorrect) cached_context... hopefully
406 	 * we'll have better luck building it later
407 	 */
408 	if (realm->cached_context) {
409 		ceph_put_snap_context(realm->cached_context);
410 		realm->cached_context = NULL;
411 	}
412 	pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
413 	       realm, err);
414 	return err;
415 }
416 
417 /*
418  * rebuild snap context for the given realm and all of its children.
419  */
420 static void rebuild_snap_realms(struct ceph_snap_realm *realm,
421 				struct list_head *dirty_realms)
422 {
423 	struct ceph_snap_realm *child;
424 
425 	dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
426 	build_snap_context(realm, dirty_realms);
427 
428 	list_for_each_entry(child, &realm->children, child_item)
429 		rebuild_snap_realms(child, dirty_realms);
430 }
431 
432 
433 /*
434  * helper to allocate and decode an array of snapids.  free prior
435  * instance, if any.
436  */
437 static int dup_array(u64 **dst, __le64 *src, u32 num)
438 {
439 	u32 i;
440 
441 	kfree(*dst);
442 	if (num) {
443 		*dst = kcalloc(num, sizeof(u64), GFP_NOFS);
444 		if (!*dst)
445 			return -ENOMEM;
446 		for (i = 0; i < num; i++)
447 			(*dst)[i] = get_unaligned_le64(src + i);
448 	} else {
449 		*dst = NULL;
450 	}
451 	return 0;
452 }
453 
454 static bool has_new_snaps(struct ceph_snap_context *o,
455 			  struct ceph_snap_context *n)
456 {
457 	if (n->num_snaps == 0)
458 		return false;
459 	/* snaps are in descending order */
460 	return n->snaps[0] > o->seq;
461 }
462 
463 /*
464  * When a snapshot is applied, the size/mtime inode metadata is queued
465  * in a ceph_cap_snap (one for each snapshot) until writeback
466  * completes and the metadata can be flushed back to the MDS.
467  *
468  * However, if a (sync) write is currently in-progress when we apply
469  * the snapshot, we have to wait until the write succeeds or fails
470  * (and a final size/mtime is known).  In this case the
471  * cap_snap->writing = 1, and is said to be "pending."  When the write
472  * finishes, we __ceph_finish_cap_snap().
473  *
474  * Caller must hold snap_rwsem for read (i.e., the realm topology won't
475  * change).
476  */
477 static void ceph_queue_cap_snap(struct ceph_inode_info *ci)
478 {
479 	struct inode *inode = &ci->vfs_inode;
480 	struct ceph_cap_snap *capsnap;
481 	struct ceph_snap_context *old_snapc, *new_snapc;
482 	struct ceph_buffer *old_blob = NULL;
483 	int used, dirty;
484 
485 	capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
486 	if (!capsnap) {
487 		pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
488 		return;
489 	}
490 	capsnap->cap_flush.is_capsnap = true;
491 	INIT_LIST_HEAD(&capsnap->cap_flush.i_list);
492 	INIT_LIST_HEAD(&capsnap->cap_flush.g_list);
493 
494 	spin_lock(&ci->i_ceph_lock);
495 	used = __ceph_caps_used(ci);
496 	dirty = __ceph_caps_dirty(ci);
497 
498 	old_snapc = ci->i_head_snapc;
499 	new_snapc = ci->i_snap_realm->cached_context;
500 
501 	/*
502 	 * If there is a write in progress, treat that as a dirty Fw,
503 	 * even though it hasn't completed yet; by the time we finish
504 	 * up this capsnap it will be.
505 	 */
506 	if (used & CEPH_CAP_FILE_WR)
507 		dirty |= CEPH_CAP_FILE_WR;
508 
509 	if (__ceph_have_pending_cap_snap(ci)) {
510 		/* there is no point in queuing multiple "pending" cap_snaps,
511 		   as no new writes are allowed to start when pending, so any
512 		   writes in progress now were started before the previous
513 		   cap_snap.  lucky us. */
514 		dout("queue_cap_snap %p already pending\n", inode);
515 		goto update_snapc;
516 	}
517 	if (ci->i_wrbuffer_ref_head == 0 &&
518 	    !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
519 		dout("queue_cap_snap %p nothing dirty|writing\n", inode);
520 		goto update_snapc;
521 	}
522 
523 	BUG_ON(!old_snapc);
524 
525 	/*
526 	 * There is no need to send FLUSHSNAP message to MDS if there is
527 	 * no new snapshot. But when there is dirty pages or on-going
528 	 * writes, we still need to create cap_snap. cap_snap is needed
529 	 * by the write path and page writeback path.
530 	 *
531 	 * also see ceph_try_drop_cap_snap()
532 	 */
533 	if (has_new_snaps(old_snapc, new_snapc)) {
534 		if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
535 			capsnap->need_flush = true;
536 	} else {
537 		if (!(used & CEPH_CAP_FILE_WR) &&
538 		    ci->i_wrbuffer_ref_head == 0) {
539 			dout("queue_cap_snap %p "
540 			     "no new_snap|dirty_page|writing\n", inode);
541 			goto update_snapc;
542 		}
543 	}
544 
545 	dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n",
546 	     inode, capsnap, old_snapc, ceph_cap_string(dirty),
547 	     capsnap->need_flush ? "" : "no_flush");
548 	ihold(inode);
549 
550 	refcount_set(&capsnap->nref, 1);
551 	INIT_LIST_HEAD(&capsnap->ci_item);
552 
553 	capsnap->follows = old_snapc->seq;
554 	capsnap->issued = __ceph_caps_issued(ci, NULL);
555 	capsnap->dirty = dirty;
556 
557 	capsnap->mode = inode->i_mode;
558 	capsnap->uid = inode->i_uid;
559 	capsnap->gid = inode->i_gid;
560 
561 	if (dirty & CEPH_CAP_XATTR_EXCL) {
562 		old_blob = __ceph_build_xattrs_blob(ci);
563 		capsnap->xattr_blob =
564 			ceph_buffer_get(ci->i_xattrs.blob);
565 		capsnap->xattr_version = ci->i_xattrs.version;
566 	} else {
567 		capsnap->xattr_blob = NULL;
568 		capsnap->xattr_version = 0;
569 	}
570 
571 	capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
572 
573 	/* dirty page count moved from _head to this cap_snap;
574 	   all subsequent writes page dirties occur _after_ this
575 	   snapshot. */
576 	capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
577 	ci->i_wrbuffer_ref_head = 0;
578 	capsnap->context = old_snapc;
579 	list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
580 
581 	if (used & CEPH_CAP_FILE_WR) {
582 		dout("queue_cap_snap %p cap_snap %p snapc %p"
583 		     " seq %llu used WR, now pending\n", inode,
584 		     capsnap, old_snapc, old_snapc->seq);
585 		capsnap->writing = 1;
586 	} else {
587 		/* note mtime, size NOW. */
588 		__ceph_finish_cap_snap(ci, capsnap);
589 	}
590 	capsnap = NULL;
591 	old_snapc = NULL;
592 
593 update_snapc:
594        if (ci->i_wrbuffer_ref_head == 0 &&
595            ci->i_wr_ref == 0 &&
596            ci->i_dirty_caps == 0 &&
597            ci->i_flushing_caps == 0) {
598                ci->i_head_snapc = NULL;
599        } else {
600 		ci->i_head_snapc = ceph_get_snap_context(new_snapc);
601 		dout(" new snapc is %p\n", new_snapc);
602 	}
603 	spin_unlock(&ci->i_ceph_lock);
604 
605 	ceph_buffer_put(old_blob);
606 	kfree(capsnap);
607 	ceph_put_snap_context(old_snapc);
608 }
609 
610 /*
611  * Finalize the size, mtime for a cap_snap.. that is, settle on final values
612  * to be used for the snapshot, to be flushed back to the mds.
613  *
614  * If capsnap can now be flushed, add to snap_flush list, and return 1.
615  *
616  * Caller must hold i_ceph_lock.
617  */
618 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
619 			    struct ceph_cap_snap *capsnap)
620 {
621 	struct inode *inode = &ci->vfs_inode;
622 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
623 
624 	BUG_ON(capsnap->writing);
625 	capsnap->size = i_size_read(inode);
626 	capsnap->mtime = inode->i_mtime;
627 	capsnap->atime = inode->i_atime;
628 	capsnap->ctime = inode->i_ctime;
629 	capsnap->btime = ci->i_btime;
630 	capsnap->change_attr = inode_peek_iversion_raw(inode);
631 	capsnap->time_warp_seq = ci->i_time_warp_seq;
632 	capsnap->truncate_size = ci->i_truncate_size;
633 	capsnap->truncate_seq = ci->i_truncate_seq;
634 	if (capsnap->dirty_pages) {
635 		dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
636 		     "still has %d dirty pages\n", inode, capsnap,
637 		     capsnap->context, capsnap->context->seq,
638 		     ceph_cap_string(capsnap->dirty), capsnap->size,
639 		     capsnap->dirty_pages);
640 		return 0;
641 	}
642 
643 	/* Fb cap still in use, delay it */
644 	if (ci->i_wb_ref) {
645 		dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
646 		     "used WRBUFFER, delaying\n", inode, capsnap,
647 		     capsnap->context, capsnap->context->seq,
648 		     ceph_cap_string(capsnap->dirty), capsnap->size);
649 		capsnap->writing = 1;
650 		return 0;
651 	}
652 
653 	ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
654 	dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
655 	     inode, capsnap, capsnap->context,
656 	     capsnap->context->seq, ceph_cap_string(capsnap->dirty),
657 	     capsnap->size);
658 
659 	spin_lock(&mdsc->snap_flush_lock);
660 	if (list_empty(&ci->i_snap_flush_item))
661 		list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
662 	spin_unlock(&mdsc->snap_flush_lock);
663 	return 1;  /* caller may want to ceph_flush_snaps */
664 }
665 
666 /*
667  * Queue cap_snaps for snap writeback for this realm and its children.
668  * Called under snap_rwsem, so realm topology won't change.
669  */
670 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
671 {
672 	struct ceph_inode_info *ci;
673 	struct inode *lastinode = NULL;
674 
675 	dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
676 
677 	spin_lock(&realm->inodes_with_caps_lock);
678 	list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
679 		struct inode *inode = igrab(&ci->vfs_inode);
680 		if (!inode)
681 			continue;
682 		spin_unlock(&realm->inodes_with_caps_lock);
683 		iput(lastinode);
684 		lastinode = inode;
685 		ceph_queue_cap_snap(ci);
686 		spin_lock(&realm->inodes_with_caps_lock);
687 	}
688 	spin_unlock(&realm->inodes_with_caps_lock);
689 	iput(lastinode);
690 
691 	dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
692 }
693 
694 /*
695  * Parse and apply a snapblob "snap trace" from the MDS.  This specifies
696  * the snap realm parameters from a given realm and all of its ancestors,
697  * up to the root.
698  *
699  * Caller must hold snap_rwsem for write.
700  */
701 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
702 			   void *p, void *e, bool deletion,
703 			   struct ceph_snap_realm **realm_ret)
704 {
705 	struct ceph_mds_snap_realm *ri;    /* encoded */
706 	__le64 *snaps;                     /* encoded */
707 	__le64 *prior_parent_snaps;        /* encoded */
708 	struct ceph_snap_realm *realm = NULL;
709 	struct ceph_snap_realm *first_realm = NULL;
710 	int invalidate = 0;
711 	int err = -ENOMEM;
712 	LIST_HEAD(dirty_realms);
713 
714 	lockdep_assert_held_write(&mdsc->snap_rwsem);
715 
716 	dout("update_snap_trace deletion=%d\n", deletion);
717 more:
718 	ceph_decode_need(&p, e, sizeof(*ri), bad);
719 	ri = p;
720 	p += sizeof(*ri);
721 	ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
722 			    le32_to_cpu(ri->num_prior_parent_snaps)), bad);
723 	snaps = p;
724 	p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
725 	prior_parent_snaps = p;
726 	p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
727 
728 	realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
729 	if (!realm) {
730 		realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
731 		if (IS_ERR(realm)) {
732 			err = PTR_ERR(realm);
733 			goto fail;
734 		}
735 	}
736 
737 	/* ensure the parent is correct */
738 	err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
739 	if (err < 0)
740 		goto fail;
741 	invalidate += err;
742 
743 	if (le64_to_cpu(ri->seq) > realm->seq) {
744 		dout("update_snap_trace updating %llx %p %lld -> %lld\n",
745 		     realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
746 		/* update realm parameters, snap lists */
747 		realm->seq = le64_to_cpu(ri->seq);
748 		realm->created = le64_to_cpu(ri->created);
749 		realm->parent_since = le64_to_cpu(ri->parent_since);
750 
751 		realm->num_snaps = le32_to_cpu(ri->num_snaps);
752 		err = dup_array(&realm->snaps, snaps, realm->num_snaps);
753 		if (err < 0)
754 			goto fail;
755 
756 		realm->num_prior_parent_snaps =
757 			le32_to_cpu(ri->num_prior_parent_snaps);
758 		err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
759 				realm->num_prior_parent_snaps);
760 		if (err < 0)
761 			goto fail;
762 
763 		if (realm->seq > mdsc->last_snap_seq)
764 			mdsc->last_snap_seq = realm->seq;
765 
766 		invalidate = 1;
767 	} else if (!realm->cached_context) {
768 		dout("update_snap_trace %llx %p seq %lld new\n",
769 		     realm->ino, realm, realm->seq);
770 		invalidate = 1;
771 	} else {
772 		dout("update_snap_trace %llx %p seq %lld unchanged\n",
773 		     realm->ino, realm, realm->seq);
774 	}
775 
776 	dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
777 	     realm, invalidate, p, e);
778 
779 	/* invalidate when we reach the _end_ (root) of the trace */
780 	if (invalidate && p >= e)
781 		rebuild_snap_realms(realm, &dirty_realms);
782 
783 	if (!first_realm)
784 		first_realm = realm;
785 	else
786 		ceph_put_snap_realm(mdsc, realm);
787 
788 	if (p < e)
789 		goto more;
790 
791 	/*
792 	 * queue cap snaps _after_ we've built the new snap contexts,
793 	 * so that i_head_snapc can be set appropriately.
794 	 */
795 	while (!list_empty(&dirty_realms)) {
796 		realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
797 					 dirty_item);
798 		list_del_init(&realm->dirty_item);
799 		queue_realm_cap_snaps(realm);
800 	}
801 
802 	if (realm_ret)
803 		*realm_ret = first_realm;
804 	else
805 		ceph_put_snap_realm(mdsc, first_realm);
806 
807 	__cleanup_empty_realms(mdsc);
808 	return 0;
809 
810 bad:
811 	err = -EIO;
812 fail:
813 	if (realm && !IS_ERR(realm))
814 		ceph_put_snap_realm(mdsc, realm);
815 	if (first_realm)
816 		ceph_put_snap_realm(mdsc, first_realm);
817 	pr_err("update_snap_trace error %d\n", err);
818 	return err;
819 }
820 
821 
822 /*
823  * Send any cap_snaps that are queued for flush.  Try to carry
824  * s_mutex across multiple snap flushes to avoid locking overhead.
825  *
826  * Caller holds no locks.
827  */
828 static void flush_snaps(struct ceph_mds_client *mdsc)
829 {
830 	struct ceph_inode_info *ci;
831 	struct inode *inode;
832 	struct ceph_mds_session *session = NULL;
833 
834 	dout("flush_snaps\n");
835 	spin_lock(&mdsc->snap_flush_lock);
836 	while (!list_empty(&mdsc->snap_flush_list)) {
837 		ci = list_first_entry(&mdsc->snap_flush_list,
838 				struct ceph_inode_info, i_snap_flush_item);
839 		inode = &ci->vfs_inode;
840 		ihold(inode);
841 		spin_unlock(&mdsc->snap_flush_lock);
842 		ceph_flush_snaps(ci, &session);
843 		iput(inode);
844 		spin_lock(&mdsc->snap_flush_lock);
845 	}
846 	spin_unlock(&mdsc->snap_flush_lock);
847 
848 	ceph_put_mds_session(session);
849 	dout("flush_snaps done\n");
850 }
851 
852 
853 /*
854  * Handle a snap notification from the MDS.
855  *
856  * This can take two basic forms: the simplest is just a snap creation
857  * or deletion notification on an existing realm.  This should update the
858  * realm and its children.
859  *
860  * The more difficult case is realm creation, due to snap creation at a
861  * new point in the file hierarchy, or due to a rename that moves a file or
862  * directory into another realm.
863  */
864 void ceph_handle_snap(struct ceph_mds_client *mdsc,
865 		      struct ceph_mds_session *session,
866 		      struct ceph_msg *msg)
867 {
868 	struct super_block *sb = mdsc->fsc->sb;
869 	int mds = session->s_mds;
870 	u64 split;
871 	int op;
872 	int trace_len;
873 	struct ceph_snap_realm *realm = NULL;
874 	void *p = msg->front.iov_base;
875 	void *e = p + msg->front.iov_len;
876 	struct ceph_mds_snap_head *h;
877 	int num_split_inos, num_split_realms;
878 	__le64 *split_inos = NULL, *split_realms = NULL;
879 	int i;
880 	int locked_rwsem = 0;
881 
882 	/* decode */
883 	if (msg->front.iov_len < sizeof(*h))
884 		goto bad;
885 	h = p;
886 	op = le32_to_cpu(h->op);
887 	split = le64_to_cpu(h->split);   /* non-zero if we are splitting an
888 					  * existing realm */
889 	num_split_inos = le32_to_cpu(h->num_split_inos);
890 	num_split_realms = le32_to_cpu(h->num_split_realms);
891 	trace_len = le32_to_cpu(h->trace_len);
892 	p += sizeof(*h);
893 
894 	dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
895 	     ceph_snap_op_name(op), split, trace_len);
896 
897 	mutex_lock(&session->s_mutex);
898 	inc_session_sequence(session);
899 	mutex_unlock(&session->s_mutex);
900 
901 	down_write(&mdsc->snap_rwsem);
902 	locked_rwsem = 1;
903 
904 	if (op == CEPH_SNAP_OP_SPLIT) {
905 		struct ceph_mds_snap_realm *ri;
906 
907 		/*
908 		 * A "split" breaks part of an existing realm off into
909 		 * a new realm.  The MDS provides a list of inodes
910 		 * (with caps) and child realms that belong to the new
911 		 * child.
912 		 */
913 		split_inos = p;
914 		p += sizeof(u64) * num_split_inos;
915 		split_realms = p;
916 		p += sizeof(u64) * num_split_realms;
917 		ceph_decode_need(&p, e, sizeof(*ri), bad);
918 		/* we will peek at realm info here, but will _not_
919 		 * advance p, as the realm update will occur below in
920 		 * ceph_update_snap_trace. */
921 		ri = p;
922 
923 		realm = ceph_lookup_snap_realm(mdsc, split);
924 		if (!realm) {
925 			realm = ceph_create_snap_realm(mdsc, split);
926 			if (IS_ERR(realm))
927 				goto out;
928 		}
929 
930 		dout("splitting snap_realm %llx %p\n", realm->ino, realm);
931 		for (i = 0; i < num_split_inos; i++) {
932 			struct ceph_vino vino = {
933 				.ino = le64_to_cpu(split_inos[i]),
934 				.snap = CEPH_NOSNAP,
935 			};
936 			struct inode *inode = ceph_find_inode(sb, vino);
937 			struct ceph_inode_info *ci;
938 			struct ceph_snap_realm *oldrealm;
939 
940 			if (!inode)
941 				continue;
942 			ci = ceph_inode(inode);
943 
944 			spin_lock(&ci->i_ceph_lock);
945 			if (!ci->i_snap_realm)
946 				goto skip_inode;
947 			/*
948 			 * If this inode belongs to a realm that was
949 			 * created after our new realm, we experienced
950 			 * a race (due to another split notifications
951 			 * arriving from a different MDS).  So skip
952 			 * this inode.
953 			 */
954 			if (ci->i_snap_realm->created >
955 			    le64_to_cpu(ri->created)) {
956 				dout(" leaving %p in newer realm %llx %p\n",
957 				     inode, ci->i_snap_realm->ino,
958 				     ci->i_snap_realm);
959 				goto skip_inode;
960 			}
961 			dout(" will move %p to split realm %llx %p\n",
962 			     inode, realm->ino, realm);
963 			/*
964 			 * Move the inode to the new realm
965 			 */
966 			oldrealm = ci->i_snap_realm;
967 			spin_lock(&oldrealm->inodes_with_caps_lock);
968 			list_del_init(&ci->i_snap_realm_item);
969 			spin_unlock(&oldrealm->inodes_with_caps_lock);
970 
971 			spin_lock(&realm->inodes_with_caps_lock);
972 			list_add(&ci->i_snap_realm_item,
973 				 &realm->inodes_with_caps);
974 			ci->i_snap_realm = realm;
975 			if (realm->ino == ci->i_vino.ino)
976                                 realm->inode = inode;
977 			spin_unlock(&realm->inodes_with_caps_lock);
978 
979 			spin_unlock(&ci->i_ceph_lock);
980 
981 			ceph_get_snap_realm(mdsc, realm);
982 			ceph_put_snap_realm(mdsc, oldrealm);
983 
984 			iput(inode);
985 			continue;
986 
987 skip_inode:
988 			spin_unlock(&ci->i_ceph_lock);
989 			iput(inode);
990 		}
991 
992 		/* we may have taken some of the old realm's children. */
993 		for (i = 0; i < num_split_realms; i++) {
994 			struct ceph_snap_realm *child =
995 				__lookup_snap_realm(mdsc,
996 					   le64_to_cpu(split_realms[i]));
997 			if (!child)
998 				continue;
999 			adjust_snap_realm_parent(mdsc, child, realm->ino);
1000 		}
1001 	}
1002 
1003 	/*
1004 	 * update using the provided snap trace. if we are deleting a
1005 	 * snap, we can avoid queueing cap_snaps.
1006 	 */
1007 	ceph_update_snap_trace(mdsc, p, e,
1008 			       op == CEPH_SNAP_OP_DESTROY, NULL);
1009 
1010 	if (op == CEPH_SNAP_OP_SPLIT)
1011 		/* we took a reference when we created the realm, above */
1012 		ceph_put_snap_realm(mdsc, realm);
1013 
1014 	__cleanup_empty_realms(mdsc);
1015 
1016 	up_write(&mdsc->snap_rwsem);
1017 
1018 	flush_snaps(mdsc);
1019 	return;
1020 
1021 bad:
1022 	pr_err("corrupt snap message from mds%d\n", mds);
1023 	ceph_msg_dump(msg);
1024 out:
1025 	if (locked_rwsem)
1026 		up_write(&mdsc->snap_rwsem);
1027 	return;
1028 }
1029 
1030 struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1031 					    u64 snap)
1032 {
1033 	struct ceph_snapid_map *sm, *exist;
1034 	struct rb_node **p, *parent;
1035 	int ret;
1036 
1037 	exist = NULL;
1038 	spin_lock(&mdsc->snapid_map_lock);
1039 	p = &mdsc->snapid_map_tree.rb_node;
1040 	while (*p) {
1041 		exist = rb_entry(*p, struct ceph_snapid_map, node);
1042 		if (snap > exist->snap) {
1043 			p = &(*p)->rb_left;
1044 		} else if (snap < exist->snap) {
1045 			p = &(*p)->rb_right;
1046 		} else {
1047 			if (atomic_inc_return(&exist->ref) == 1)
1048 				list_del_init(&exist->lru);
1049 			break;
1050 		}
1051 		exist = NULL;
1052 	}
1053 	spin_unlock(&mdsc->snapid_map_lock);
1054 	if (exist) {
1055 		dout("found snapid map %llx -> %x\n", exist->snap, exist->dev);
1056 		return exist;
1057 	}
1058 
1059 	sm = kmalloc(sizeof(*sm), GFP_NOFS);
1060 	if (!sm)
1061 		return NULL;
1062 
1063 	ret = get_anon_bdev(&sm->dev);
1064 	if (ret < 0) {
1065 		kfree(sm);
1066 		return NULL;
1067 	}
1068 
1069 	INIT_LIST_HEAD(&sm->lru);
1070 	atomic_set(&sm->ref, 1);
1071 	sm->snap = snap;
1072 
1073 	exist = NULL;
1074 	parent = NULL;
1075 	p = &mdsc->snapid_map_tree.rb_node;
1076 	spin_lock(&mdsc->snapid_map_lock);
1077 	while (*p) {
1078 		parent = *p;
1079 		exist = rb_entry(*p, struct ceph_snapid_map, node);
1080 		if (snap > exist->snap)
1081 			p = &(*p)->rb_left;
1082 		else if (snap < exist->snap)
1083 			p = &(*p)->rb_right;
1084 		else
1085 			break;
1086 		exist = NULL;
1087 	}
1088 	if (exist) {
1089 		if (atomic_inc_return(&exist->ref) == 1)
1090 			list_del_init(&exist->lru);
1091 	} else {
1092 		rb_link_node(&sm->node, parent, p);
1093 		rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1094 	}
1095 	spin_unlock(&mdsc->snapid_map_lock);
1096 	if (exist) {
1097 		free_anon_bdev(sm->dev);
1098 		kfree(sm);
1099 		dout("found snapid map %llx -> %x\n", exist->snap, exist->dev);
1100 		return exist;
1101 	}
1102 
1103 	dout("create snapid map %llx -> %x\n", sm->snap, sm->dev);
1104 	return sm;
1105 }
1106 
1107 void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1108 			 struct ceph_snapid_map *sm)
1109 {
1110 	if (!sm)
1111 		return;
1112 	if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1113 		if (!RB_EMPTY_NODE(&sm->node)) {
1114 			sm->last_used = jiffies;
1115 			list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1116 			spin_unlock(&mdsc->snapid_map_lock);
1117 		} else {
1118 			/* already cleaned up by
1119 			 * ceph_cleanup_snapid_map() */
1120 			spin_unlock(&mdsc->snapid_map_lock);
1121 			kfree(sm);
1122 		}
1123 	}
1124 }
1125 
1126 void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1127 {
1128 	struct ceph_snapid_map *sm;
1129 	unsigned long now;
1130 	LIST_HEAD(to_free);
1131 
1132 	spin_lock(&mdsc->snapid_map_lock);
1133 	now = jiffies;
1134 
1135 	while (!list_empty(&mdsc->snapid_map_lru)) {
1136 		sm = list_first_entry(&mdsc->snapid_map_lru,
1137 				      struct ceph_snapid_map, lru);
1138 		if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1139 			break;
1140 
1141 		rb_erase(&sm->node, &mdsc->snapid_map_tree);
1142 		list_move(&sm->lru, &to_free);
1143 	}
1144 	spin_unlock(&mdsc->snapid_map_lock);
1145 
1146 	while (!list_empty(&to_free)) {
1147 		sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1148 		list_del(&sm->lru);
1149 		dout("trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1150 		free_anon_bdev(sm->dev);
1151 		kfree(sm);
1152 	}
1153 }
1154 
1155 void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1156 {
1157 	struct ceph_snapid_map *sm;
1158 	struct rb_node *p;
1159 	LIST_HEAD(to_free);
1160 
1161 	spin_lock(&mdsc->snapid_map_lock);
1162 	while ((p = rb_first(&mdsc->snapid_map_tree))) {
1163 		sm = rb_entry(p, struct ceph_snapid_map, node);
1164 		rb_erase(p, &mdsc->snapid_map_tree);
1165 		RB_CLEAR_NODE(p);
1166 		list_move(&sm->lru, &to_free);
1167 	}
1168 	spin_unlock(&mdsc->snapid_map_lock);
1169 
1170 	while (!list_empty(&to_free)) {
1171 		sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1172 		list_del(&sm->lru);
1173 		free_anon_bdev(sm->dev);
1174 		if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1175 			pr_err("snapid map %llx -> %x still in use\n",
1176 			       sm->snap, sm->dev);
1177 		}
1178 		kfree(sm);
1179 	}
1180 }
1181