xref: /linux/fs/ceph/snap.c (revision fada1935590f66dc6784981e0d557ca09013c847)
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 
7 #include "super.h"
8 #include "mds_client.h"
9 
10 #include <linux/ceph/decode.h>
11 
12 /*
13  * Snapshots in ceph are driven in large part by cooperation from the
14  * client.  In contrast to local file systems or file servers that
15  * implement snapshots at a single point in the system, ceph's
16  * distributed access to storage requires clients to help decide
17  * whether a write logically occurs before or after a recently created
18  * snapshot.
19  *
20  * This provides a perfect instantanous client-wide snapshot.  Between
21  * clients, however, snapshots may appear to be applied at slightly
22  * different points in time, depending on delays in delivering the
23  * snapshot notification.
24  *
25  * Snapshots are _not_ file system-wide.  Instead, each snapshot
26  * applies to the subdirectory nested beneath some directory.  This
27  * effectively divides the hierarchy into multiple "realms," where all
28  * of the files contained by each realm share the same set of
29  * snapshots.  An individual realm's snap set contains snapshots
30  * explicitly created on that realm, as well as any snaps in its
31  * parent's snap set _after_ the point at which the parent became it's
32  * parent (due to, say, a rename).  Similarly, snaps from prior parents
33  * during the time intervals during which they were the parent are included.
34  *
35  * The client is spared most of this detail, fortunately... it must only
36  * maintains a hierarchy of realms reflecting the current parent/child
37  * realm relationship, and for each realm has an explicit list of snaps
38  * inherited from prior parents.
39  *
40  * A snap_realm struct is maintained for realms containing every inode
41  * with an open cap in the system.  (The needed snap realm information is
42  * provided by the MDS whenever a cap is issued, i.e., on open.)  A 'seq'
43  * version number is used to ensure that as realm parameters change (new
44  * snapshot, new parent, etc.) the client's realm hierarchy is updated.
45  *
46  * The realm hierarchy drives the generation of a 'snap context' for each
47  * realm, which simply lists the resulting set of snaps for the realm.  This
48  * is attached to any writes sent to OSDs.
49  */
50 /*
51  * Unfortunately error handling is a bit mixed here.  If we get a snap
52  * update, but don't have enough memory to update our realm hierarchy,
53  * it's not clear what we can do about it (besides complaining to the
54  * console).
55  */
56 
57 
58 /*
59  * increase ref count for the realm
60  *
61  * caller must hold snap_rwsem for write.
62  */
63 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
64 			 struct ceph_snap_realm *realm)
65 {
66 	dout("get_realm %p %d -> %d\n", realm,
67 	     atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
68 	/*
69 	 * since we _only_ increment realm refs or empty the empty
70 	 * list with snap_rwsem held, adjusting the empty list here is
71 	 * safe.  we do need to protect against concurrent empty list
72 	 * additions, however.
73 	 */
74 	if (atomic_inc_return(&realm->nref) == 1) {
75 		spin_lock(&mdsc->snap_empty_lock);
76 		list_del_init(&realm->empty_item);
77 		spin_unlock(&mdsc->snap_empty_lock);
78 	}
79 }
80 
81 static void __insert_snap_realm(struct rb_root *root,
82 				struct ceph_snap_realm *new)
83 {
84 	struct rb_node **p = &root->rb_node;
85 	struct rb_node *parent = NULL;
86 	struct ceph_snap_realm *r = NULL;
87 
88 	while (*p) {
89 		parent = *p;
90 		r = rb_entry(parent, struct ceph_snap_realm, node);
91 		if (new->ino < r->ino)
92 			p = &(*p)->rb_left;
93 		else if (new->ino > r->ino)
94 			p = &(*p)->rb_right;
95 		else
96 			BUG();
97 	}
98 
99 	rb_link_node(&new->node, parent, p);
100 	rb_insert_color(&new->node, root);
101 }
102 
103 /*
104  * create and get the realm rooted at @ino and bump its ref count.
105  *
106  * caller must hold snap_rwsem for write.
107  */
108 static struct ceph_snap_realm *ceph_create_snap_realm(
109 	struct ceph_mds_client *mdsc,
110 	u64 ino)
111 {
112 	struct ceph_snap_realm *realm;
113 
114 	realm = kzalloc(sizeof(*realm), GFP_NOFS);
115 	if (!realm)
116 		return ERR_PTR(-ENOMEM);
117 
118 	atomic_set(&realm->nref, 1);    /* for caller */
119 	realm->ino = ino;
120 	INIT_LIST_HEAD(&realm->children);
121 	INIT_LIST_HEAD(&realm->child_item);
122 	INIT_LIST_HEAD(&realm->empty_item);
123 	INIT_LIST_HEAD(&realm->dirty_item);
124 	INIT_LIST_HEAD(&realm->inodes_with_caps);
125 	spin_lock_init(&realm->inodes_with_caps_lock);
126 	__insert_snap_realm(&mdsc->snap_realms, realm);
127 	dout("create_snap_realm %llx %p\n", realm->ino, realm);
128 	return realm;
129 }
130 
131 /*
132  * lookup the realm rooted at @ino.
133  *
134  * caller must hold snap_rwsem for write.
135  */
136 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
137 						   u64 ino)
138 {
139 	struct rb_node *n = mdsc->snap_realms.rb_node;
140 	struct ceph_snap_realm *r;
141 
142 	while (n) {
143 		r = rb_entry(n, struct ceph_snap_realm, node);
144 		if (ino < r->ino)
145 			n = n->rb_left;
146 		else if (ino > r->ino)
147 			n = n->rb_right;
148 		else {
149 			dout("lookup_snap_realm %llx %p\n", r->ino, r);
150 			return r;
151 		}
152 	}
153 	return NULL;
154 }
155 
156 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
157 					       u64 ino)
158 {
159 	struct ceph_snap_realm *r;
160 	r = __lookup_snap_realm(mdsc, ino);
161 	if (r)
162 		ceph_get_snap_realm(mdsc, r);
163 	return r;
164 }
165 
166 static void __put_snap_realm(struct ceph_mds_client *mdsc,
167 			     struct ceph_snap_realm *realm);
168 
169 /*
170  * called with snap_rwsem (write)
171  */
172 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
173 				 struct ceph_snap_realm *realm)
174 {
175 	dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
176 
177 	rb_erase(&realm->node, &mdsc->snap_realms);
178 
179 	if (realm->parent) {
180 		list_del_init(&realm->child_item);
181 		__put_snap_realm(mdsc, realm->parent);
182 	}
183 
184 	kfree(realm->prior_parent_snaps);
185 	kfree(realm->snaps);
186 	ceph_put_snap_context(realm->cached_context);
187 	kfree(realm);
188 }
189 
190 /*
191  * caller holds snap_rwsem (write)
192  */
193 static void __put_snap_realm(struct ceph_mds_client *mdsc,
194 			     struct ceph_snap_realm *realm)
195 {
196 	dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
197 	     atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
198 	if (atomic_dec_and_test(&realm->nref))
199 		__destroy_snap_realm(mdsc, realm);
200 }
201 
202 /*
203  * caller needn't hold any locks
204  */
205 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
206 			 struct ceph_snap_realm *realm)
207 {
208 	dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
209 	     atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
210 	if (!atomic_dec_and_test(&realm->nref))
211 		return;
212 
213 	if (down_write_trylock(&mdsc->snap_rwsem)) {
214 		__destroy_snap_realm(mdsc, realm);
215 		up_write(&mdsc->snap_rwsem);
216 	} else {
217 		spin_lock(&mdsc->snap_empty_lock);
218 		list_add(&realm->empty_item, &mdsc->snap_empty);
219 		spin_unlock(&mdsc->snap_empty_lock);
220 	}
221 }
222 
223 /*
224  * Clean up any realms whose ref counts have dropped to zero.  Note
225  * that this does not include realms who were created but not yet
226  * used.
227  *
228  * Called under snap_rwsem (write)
229  */
230 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
231 {
232 	struct ceph_snap_realm *realm;
233 
234 	spin_lock(&mdsc->snap_empty_lock);
235 	while (!list_empty(&mdsc->snap_empty)) {
236 		realm = list_first_entry(&mdsc->snap_empty,
237 				   struct ceph_snap_realm, empty_item);
238 		list_del(&realm->empty_item);
239 		spin_unlock(&mdsc->snap_empty_lock);
240 		__destroy_snap_realm(mdsc, realm);
241 		spin_lock(&mdsc->snap_empty_lock);
242 	}
243 	spin_unlock(&mdsc->snap_empty_lock);
244 }
245 
246 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
247 {
248 	down_write(&mdsc->snap_rwsem);
249 	__cleanup_empty_realms(mdsc);
250 	up_write(&mdsc->snap_rwsem);
251 }
252 
253 /*
254  * adjust the parent realm of a given @realm.  adjust child list, and parent
255  * pointers, and ref counts appropriately.
256  *
257  * return true if parent was changed, 0 if unchanged, <0 on error.
258  *
259  * caller must hold snap_rwsem for write.
260  */
261 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
262 				    struct ceph_snap_realm *realm,
263 				    u64 parentino)
264 {
265 	struct ceph_snap_realm *parent;
266 
267 	if (realm->parent_ino == parentino)
268 		return 0;
269 
270 	parent = ceph_lookup_snap_realm(mdsc, parentino);
271 	if (!parent) {
272 		parent = ceph_create_snap_realm(mdsc, parentino);
273 		if (IS_ERR(parent))
274 			return PTR_ERR(parent);
275 	}
276 	dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
277 	     realm->ino, realm, realm->parent_ino, realm->parent,
278 	     parentino, parent);
279 	if (realm->parent) {
280 		list_del_init(&realm->child_item);
281 		ceph_put_snap_realm(mdsc, realm->parent);
282 	}
283 	realm->parent_ino = parentino;
284 	realm->parent = parent;
285 	list_add(&realm->child_item, &parent->children);
286 	return 1;
287 }
288 
289 
290 static int cmpu64_rev(const void *a, const void *b)
291 {
292 	if (*(u64 *)a < *(u64 *)b)
293 		return 1;
294 	if (*(u64 *)a > *(u64 *)b)
295 		return -1;
296 	return 0;
297 }
298 
299 
300 /*
301  * build the snap context for a given realm.
302  */
303 static int build_snap_context(struct ceph_snap_realm *realm,
304 			      struct list_head* dirty_realms)
305 {
306 	struct ceph_snap_realm *parent = realm->parent;
307 	struct ceph_snap_context *snapc;
308 	int err = 0;
309 	u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
310 
311 	/*
312 	 * build parent context, if it hasn't been built.
313 	 * conservatively estimate that all parent snaps might be
314 	 * included by us.
315 	 */
316 	if (parent) {
317 		if (!parent->cached_context) {
318 			err = build_snap_context(parent, dirty_realms);
319 			if (err)
320 				goto fail;
321 		}
322 		num += parent->cached_context->num_snaps;
323 	}
324 
325 	/* do i actually need to update?  not if my context seq
326 	   matches realm seq, and my parents' does to.  (this works
327 	   because we rebuild_snap_realms() works _downward_ in
328 	   hierarchy after each update.) */
329 	if (realm->cached_context &&
330 	    realm->cached_context->seq == realm->seq &&
331 	    (!parent ||
332 	     realm->cached_context->seq >= parent->cached_context->seq)) {
333 		dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
334 		     " (unchanged)\n",
335 		     realm->ino, realm, realm->cached_context,
336 		     realm->cached_context->seq,
337 		     (unsigned int)realm->cached_context->num_snaps);
338 		return 0;
339 	}
340 
341 	/* alloc new snap context */
342 	err = -ENOMEM;
343 	if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
344 		goto fail;
345 	snapc = ceph_create_snap_context(num, GFP_NOFS);
346 	if (!snapc)
347 		goto fail;
348 
349 	/* build (reverse sorted) snap vector */
350 	num = 0;
351 	snapc->seq = realm->seq;
352 	if (parent) {
353 		u32 i;
354 
355 		/* include any of parent's snaps occurring _after_ my
356 		   parent became my parent */
357 		for (i = 0; i < parent->cached_context->num_snaps; i++)
358 			if (parent->cached_context->snaps[i] >=
359 			    realm->parent_since)
360 				snapc->snaps[num++] =
361 					parent->cached_context->snaps[i];
362 		if (parent->cached_context->seq > snapc->seq)
363 			snapc->seq = parent->cached_context->seq;
364 	}
365 	memcpy(snapc->snaps + num, realm->snaps,
366 	       sizeof(u64)*realm->num_snaps);
367 	num += realm->num_snaps;
368 	memcpy(snapc->snaps + num, realm->prior_parent_snaps,
369 	       sizeof(u64)*realm->num_prior_parent_snaps);
370 	num += realm->num_prior_parent_snaps;
371 
372 	sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
373 	snapc->num_snaps = num;
374 	dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
375 	     realm->ino, realm, snapc, snapc->seq,
376 	     (unsigned int) snapc->num_snaps);
377 
378 	ceph_put_snap_context(realm->cached_context);
379 	realm->cached_context = snapc;
380 	/* queue realm for cap_snap creation */
381 	list_add_tail(&realm->dirty_item, dirty_realms);
382 	return 0;
383 
384 fail:
385 	/*
386 	 * if we fail, clear old (incorrect) cached_context... hopefully
387 	 * we'll have better luck building it later
388 	 */
389 	if (realm->cached_context) {
390 		ceph_put_snap_context(realm->cached_context);
391 		realm->cached_context = NULL;
392 	}
393 	pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
394 	       realm, err);
395 	return err;
396 }
397 
398 /*
399  * rebuild snap context for the given realm and all of its children.
400  */
401 static void rebuild_snap_realms(struct ceph_snap_realm *realm,
402 				struct list_head *dirty_realms)
403 {
404 	struct ceph_snap_realm *child;
405 
406 	dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
407 	build_snap_context(realm, dirty_realms);
408 
409 	list_for_each_entry(child, &realm->children, child_item)
410 		rebuild_snap_realms(child, dirty_realms);
411 }
412 
413 
414 /*
415  * helper to allocate and decode an array of snapids.  free prior
416  * instance, if any.
417  */
418 static int dup_array(u64 **dst, __le64 *src, u32 num)
419 {
420 	u32 i;
421 
422 	kfree(*dst);
423 	if (num) {
424 		*dst = kcalloc(num, sizeof(u64), GFP_NOFS);
425 		if (!*dst)
426 			return -ENOMEM;
427 		for (i = 0; i < num; i++)
428 			(*dst)[i] = get_unaligned_le64(src + i);
429 	} else {
430 		*dst = NULL;
431 	}
432 	return 0;
433 }
434 
435 static bool has_new_snaps(struct ceph_snap_context *o,
436 			  struct ceph_snap_context *n)
437 {
438 	if (n->num_snaps == 0)
439 		return false;
440 	/* snaps are in descending order */
441 	return n->snaps[0] > o->seq;
442 }
443 
444 /*
445  * When a snapshot is applied, the size/mtime inode metadata is queued
446  * in a ceph_cap_snap (one for each snapshot) until writeback
447  * completes and the metadata can be flushed back to the MDS.
448  *
449  * However, if a (sync) write is currently in-progress when we apply
450  * the snapshot, we have to wait until the write succeeds or fails
451  * (and a final size/mtime is known).  In this case the
452  * cap_snap->writing = 1, and is said to be "pending."  When the write
453  * finishes, we __ceph_finish_cap_snap().
454  *
455  * Caller must hold snap_rwsem for read (i.e., the realm topology won't
456  * change).
457  */
458 void ceph_queue_cap_snap(struct ceph_inode_info *ci)
459 {
460 	struct inode *inode = &ci->vfs_inode;
461 	struct ceph_cap_snap *capsnap;
462 	struct ceph_snap_context *old_snapc, *new_snapc;
463 	int used, dirty;
464 
465 	capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
466 	if (!capsnap) {
467 		pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
468 		return;
469 	}
470 
471 	spin_lock(&ci->i_ceph_lock);
472 	used = __ceph_caps_used(ci);
473 	dirty = __ceph_caps_dirty(ci);
474 
475 	old_snapc = ci->i_head_snapc;
476 	new_snapc = ci->i_snap_realm->cached_context;
477 
478 	/*
479 	 * If there is a write in progress, treat that as a dirty Fw,
480 	 * even though it hasn't completed yet; by the time we finish
481 	 * up this capsnap it will be.
482 	 */
483 	if (used & CEPH_CAP_FILE_WR)
484 		dirty |= CEPH_CAP_FILE_WR;
485 
486 	if (__ceph_have_pending_cap_snap(ci)) {
487 		/* there is no point in queuing multiple "pending" cap_snaps,
488 		   as no new writes are allowed to start when pending, so any
489 		   writes in progress now were started before the previous
490 		   cap_snap.  lucky us. */
491 		dout("queue_cap_snap %p already pending\n", inode);
492 		goto update_snapc;
493 	}
494 	if (ci->i_wrbuffer_ref_head == 0 &&
495 	    !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
496 		dout("queue_cap_snap %p nothing dirty|writing\n", inode);
497 		goto update_snapc;
498 	}
499 
500 	BUG_ON(!old_snapc);
501 
502 	/*
503 	 * There is no need to send FLUSHSNAP message to MDS if there is
504 	 * no new snapshot. But when there is dirty pages or on-going
505 	 * writes, we still need to create cap_snap. cap_snap is needed
506 	 * by the write path and page writeback path.
507 	 *
508 	 * also see ceph_try_drop_cap_snap()
509 	 */
510 	if (has_new_snaps(old_snapc, new_snapc)) {
511 		if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
512 			capsnap->need_flush = true;
513 	} else {
514 		if (!(used & CEPH_CAP_FILE_WR) &&
515 		    ci->i_wrbuffer_ref_head == 0) {
516 			dout("queue_cap_snap %p "
517 			     "no new_snap|dirty_page|writing\n", inode);
518 			goto update_snapc;
519 		}
520 	}
521 
522 	dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n",
523 	     inode, capsnap, old_snapc, ceph_cap_string(dirty),
524 	     capsnap->need_flush ? "" : "no_flush");
525 	ihold(inode);
526 
527 	refcount_set(&capsnap->nref, 1);
528 	INIT_LIST_HEAD(&capsnap->ci_item);
529 
530 	capsnap->follows = old_snapc->seq;
531 	capsnap->issued = __ceph_caps_issued(ci, NULL);
532 	capsnap->dirty = dirty;
533 
534 	capsnap->mode = inode->i_mode;
535 	capsnap->uid = inode->i_uid;
536 	capsnap->gid = inode->i_gid;
537 
538 	if (dirty & CEPH_CAP_XATTR_EXCL) {
539 		__ceph_build_xattrs_blob(ci);
540 		capsnap->xattr_blob =
541 			ceph_buffer_get(ci->i_xattrs.blob);
542 		capsnap->xattr_version = ci->i_xattrs.version;
543 	} else {
544 		capsnap->xattr_blob = NULL;
545 		capsnap->xattr_version = 0;
546 	}
547 
548 	capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
549 
550 	/* dirty page count moved from _head to this cap_snap;
551 	   all subsequent writes page dirties occur _after_ this
552 	   snapshot. */
553 	capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
554 	ci->i_wrbuffer_ref_head = 0;
555 	capsnap->context = old_snapc;
556 	list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
557 
558 	if (used & CEPH_CAP_FILE_WR) {
559 		dout("queue_cap_snap %p cap_snap %p snapc %p"
560 		     " seq %llu used WR, now pending\n", inode,
561 		     capsnap, old_snapc, old_snapc->seq);
562 		capsnap->writing = 1;
563 	} else {
564 		/* note mtime, size NOW. */
565 		__ceph_finish_cap_snap(ci, capsnap);
566 	}
567 	capsnap = NULL;
568 	old_snapc = NULL;
569 
570 update_snapc:
571 	if (ci->i_head_snapc) {
572 		ci->i_head_snapc = ceph_get_snap_context(new_snapc);
573 		dout(" new snapc is %p\n", new_snapc);
574 	}
575 	spin_unlock(&ci->i_ceph_lock);
576 
577 	kfree(capsnap);
578 	ceph_put_snap_context(old_snapc);
579 }
580 
581 /*
582  * Finalize the size, mtime for a cap_snap.. that is, settle on final values
583  * to be used for the snapshot, to be flushed back to the mds.
584  *
585  * If capsnap can now be flushed, add to snap_flush list, and return 1.
586  *
587  * Caller must hold i_ceph_lock.
588  */
589 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
590 			    struct ceph_cap_snap *capsnap)
591 {
592 	struct inode *inode = &ci->vfs_inode;
593 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
594 
595 	BUG_ON(capsnap->writing);
596 	capsnap->size = inode->i_size;
597 	capsnap->mtime = inode->i_mtime;
598 	capsnap->atime = inode->i_atime;
599 	capsnap->ctime = inode->i_ctime;
600 	capsnap->time_warp_seq = ci->i_time_warp_seq;
601 	capsnap->truncate_size = ci->i_truncate_size;
602 	capsnap->truncate_seq = ci->i_truncate_seq;
603 	if (capsnap->dirty_pages) {
604 		dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
605 		     "still has %d dirty pages\n", inode, capsnap,
606 		     capsnap->context, capsnap->context->seq,
607 		     ceph_cap_string(capsnap->dirty), capsnap->size,
608 		     capsnap->dirty_pages);
609 		return 0;
610 	}
611 
612 	ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
613 	dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
614 	     inode, capsnap, capsnap->context,
615 	     capsnap->context->seq, ceph_cap_string(capsnap->dirty),
616 	     capsnap->size);
617 
618 	spin_lock(&mdsc->snap_flush_lock);
619 	list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
620 	spin_unlock(&mdsc->snap_flush_lock);
621 	return 1;  /* caller may want to ceph_flush_snaps */
622 }
623 
624 /*
625  * Queue cap_snaps for snap writeback for this realm and its children.
626  * Called under snap_rwsem, so realm topology won't change.
627  */
628 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
629 {
630 	struct ceph_inode_info *ci;
631 	struct inode *lastinode = NULL;
632 
633 	dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
634 
635 	spin_lock(&realm->inodes_with_caps_lock);
636 	list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
637 		struct inode *inode = igrab(&ci->vfs_inode);
638 		if (!inode)
639 			continue;
640 		spin_unlock(&realm->inodes_with_caps_lock);
641 		iput(lastinode);
642 		lastinode = inode;
643 		ceph_queue_cap_snap(ci);
644 		spin_lock(&realm->inodes_with_caps_lock);
645 	}
646 	spin_unlock(&realm->inodes_with_caps_lock);
647 	iput(lastinode);
648 
649 	dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
650 }
651 
652 /*
653  * Parse and apply a snapblob "snap trace" from the MDS.  This specifies
654  * the snap realm parameters from a given realm and all of its ancestors,
655  * up to the root.
656  *
657  * Caller must hold snap_rwsem for write.
658  */
659 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
660 			   void *p, void *e, bool deletion,
661 			   struct ceph_snap_realm **realm_ret)
662 {
663 	struct ceph_mds_snap_realm *ri;    /* encoded */
664 	__le64 *snaps;                     /* encoded */
665 	__le64 *prior_parent_snaps;        /* encoded */
666 	struct ceph_snap_realm *realm = NULL;
667 	struct ceph_snap_realm *first_realm = NULL;
668 	int invalidate = 0;
669 	int err = -ENOMEM;
670 	LIST_HEAD(dirty_realms);
671 
672 	dout("update_snap_trace deletion=%d\n", deletion);
673 more:
674 	ceph_decode_need(&p, e, sizeof(*ri), bad);
675 	ri = p;
676 	p += sizeof(*ri);
677 	ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
678 			    le32_to_cpu(ri->num_prior_parent_snaps)), bad);
679 	snaps = p;
680 	p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
681 	prior_parent_snaps = p;
682 	p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
683 
684 	realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
685 	if (!realm) {
686 		realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
687 		if (IS_ERR(realm)) {
688 			err = PTR_ERR(realm);
689 			goto fail;
690 		}
691 	}
692 
693 	/* ensure the parent is correct */
694 	err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
695 	if (err < 0)
696 		goto fail;
697 	invalidate += err;
698 
699 	if (le64_to_cpu(ri->seq) > realm->seq) {
700 		dout("update_snap_trace updating %llx %p %lld -> %lld\n",
701 		     realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
702 		/* update realm parameters, snap lists */
703 		realm->seq = le64_to_cpu(ri->seq);
704 		realm->created = le64_to_cpu(ri->created);
705 		realm->parent_since = le64_to_cpu(ri->parent_since);
706 
707 		realm->num_snaps = le32_to_cpu(ri->num_snaps);
708 		err = dup_array(&realm->snaps, snaps, realm->num_snaps);
709 		if (err < 0)
710 			goto fail;
711 
712 		realm->num_prior_parent_snaps =
713 			le32_to_cpu(ri->num_prior_parent_snaps);
714 		err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
715 				realm->num_prior_parent_snaps);
716 		if (err < 0)
717 			goto fail;
718 
719 		if (realm->seq > mdsc->last_snap_seq)
720 			mdsc->last_snap_seq = realm->seq;
721 
722 		invalidate = 1;
723 	} else if (!realm->cached_context) {
724 		dout("update_snap_trace %llx %p seq %lld new\n",
725 		     realm->ino, realm, realm->seq);
726 		invalidate = 1;
727 	} else {
728 		dout("update_snap_trace %llx %p seq %lld unchanged\n",
729 		     realm->ino, realm, realm->seq);
730 	}
731 
732 	dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
733 	     realm, invalidate, p, e);
734 
735 	/* invalidate when we reach the _end_ (root) of the trace */
736 	if (invalidate && p >= e)
737 		rebuild_snap_realms(realm, &dirty_realms);
738 
739 	if (!first_realm)
740 		first_realm = realm;
741 	else
742 		ceph_put_snap_realm(mdsc, realm);
743 
744 	if (p < e)
745 		goto more;
746 
747 	/*
748 	 * queue cap snaps _after_ we've built the new snap contexts,
749 	 * so that i_head_snapc can be set appropriately.
750 	 */
751 	while (!list_empty(&dirty_realms)) {
752 		realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
753 					 dirty_item);
754 		list_del_init(&realm->dirty_item);
755 		queue_realm_cap_snaps(realm);
756 	}
757 
758 	if (realm_ret)
759 		*realm_ret = first_realm;
760 	else
761 		ceph_put_snap_realm(mdsc, first_realm);
762 
763 	__cleanup_empty_realms(mdsc);
764 	return 0;
765 
766 bad:
767 	err = -EINVAL;
768 fail:
769 	if (realm && !IS_ERR(realm))
770 		ceph_put_snap_realm(mdsc, realm);
771 	if (first_realm)
772 		ceph_put_snap_realm(mdsc, first_realm);
773 	pr_err("update_snap_trace error %d\n", err);
774 	return err;
775 }
776 
777 
778 /*
779  * Send any cap_snaps that are queued for flush.  Try to carry
780  * s_mutex across multiple snap flushes to avoid locking overhead.
781  *
782  * Caller holds no locks.
783  */
784 static void flush_snaps(struct ceph_mds_client *mdsc)
785 {
786 	struct ceph_inode_info *ci;
787 	struct inode *inode;
788 	struct ceph_mds_session *session = NULL;
789 
790 	dout("flush_snaps\n");
791 	spin_lock(&mdsc->snap_flush_lock);
792 	while (!list_empty(&mdsc->snap_flush_list)) {
793 		ci = list_first_entry(&mdsc->snap_flush_list,
794 				struct ceph_inode_info, i_snap_flush_item);
795 		inode = &ci->vfs_inode;
796 		ihold(inode);
797 		spin_unlock(&mdsc->snap_flush_lock);
798 		ceph_flush_snaps(ci, &session);
799 		iput(inode);
800 		spin_lock(&mdsc->snap_flush_lock);
801 	}
802 	spin_unlock(&mdsc->snap_flush_lock);
803 
804 	if (session) {
805 		mutex_unlock(&session->s_mutex);
806 		ceph_put_mds_session(session);
807 	}
808 	dout("flush_snaps done\n");
809 }
810 
811 
812 /*
813  * Handle a snap notification from the MDS.
814  *
815  * This can take two basic forms: the simplest is just a snap creation
816  * or deletion notification on an existing realm.  This should update the
817  * realm and its children.
818  *
819  * The more difficult case is realm creation, due to snap creation at a
820  * new point in the file hierarchy, or due to a rename that moves a file or
821  * directory into another realm.
822  */
823 void ceph_handle_snap(struct ceph_mds_client *mdsc,
824 		      struct ceph_mds_session *session,
825 		      struct ceph_msg *msg)
826 {
827 	struct super_block *sb = mdsc->fsc->sb;
828 	int mds = session->s_mds;
829 	u64 split;
830 	int op;
831 	int trace_len;
832 	struct ceph_snap_realm *realm = NULL;
833 	void *p = msg->front.iov_base;
834 	void *e = p + msg->front.iov_len;
835 	struct ceph_mds_snap_head *h;
836 	int num_split_inos, num_split_realms;
837 	__le64 *split_inos = NULL, *split_realms = NULL;
838 	int i;
839 	int locked_rwsem = 0;
840 
841 	/* decode */
842 	if (msg->front.iov_len < sizeof(*h))
843 		goto bad;
844 	h = p;
845 	op = le32_to_cpu(h->op);
846 	split = le64_to_cpu(h->split);   /* non-zero if we are splitting an
847 					  * existing realm */
848 	num_split_inos = le32_to_cpu(h->num_split_inos);
849 	num_split_realms = le32_to_cpu(h->num_split_realms);
850 	trace_len = le32_to_cpu(h->trace_len);
851 	p += sizeof(*h);
852 
853 	dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
854 	     ceph_snap_op_name(op), split, trace_len);
855 
856 	mutex_lock(&session->s_mutex);
857 	session->s_seq++;
858 	mutex_unlock(&session->s_mutex);
859 
860 	down_write(&mdsc->snap_rwsem);
861 	locked_rwsem = 1;
862 
863 	if (op == CEPH_SNAP_OP_SPLIT) {
864 		struct ceph_mds_snap_realm *ri;
865 
866 		/*
867 		 * A "split" breaks part of an existing realm off into
868 		 * a new realm.  The MDS provides a list of inodes
869 		 * (with caps) and child realms that belong to the new
870 		 * child.
871 		 */
872 		split_inos = p;
873 		p += sizeof(u64) * num_split_inos;
874 		split_realms = p;
875 		p += sizeof(u64) * num_split_realms;
876 		ceph_decode_need(&p, e, sizeof(*ri), bad);
877 		/* we will peek at realm info here, but will _not_
878 		 * advance p, as the realm update will occur below in
879 		 * ceph_update_snap_trace. */
880 		ri = p;
881 
882 		realm = ceph_lookup_snap_realm(mdsc, split);
883 		if (!realm) {
884 			realm = ceph_create_snap_realm(mdsc, split);
885 			if (IS_ERR(realm))
886 				goto out;
887 		}
888 
889 		dout("splitting snap_realm %llx %p\n", realm->ino, realm);
890 		for (i = 0; i < num_split_inos; i++) {
891 			struct ceph_vino vino = {
892 				.ino = le64_to_cpu(split_inos[i]),
893 				.snap = CEPH_NOSNAP,
894 			};
895 			struct inode *inode = ceph_find_inode(sb, vino);
896 			struct ceph_inode_info *ci;
897 			struct ceph_snap_realm *oldrealm;
898 
899 			if (!inode)
900 				continue;
901 			ci = ceph_inode(inode);
902 
903 			spin_lock(&ci->i_ceph_lock);
904 			if (!ci->i_snap_realm)
905 				goto skip_inode;
906 			/*
907 			 * If this inode belongs to a realm that was
908 			 * created after our new realm, we experienced
909 			 * a race (due to another split notifications
910 			 * arriving from a different MDS).  So skip
911 			 * this inode.
912 			 */
913 			if (ci->i_snap_realm->created >
914 			    le64_to_cpu(ri->created)) {
915 				dout(" leaving %p in newer realm %llx %p\n",
916 				     inode, ci->i_snap_realm->ino,
917 				     ci->i_snap_realm);
918 				goto skip_inode;
919 			}
920 			dout(" will move %p to split realm %llx %p\n",
921 			     inode, realm->ino, realm);
922 			/*
923 			 * Move the inode to the new realm
924 			 */
925 			spin_lock(&realm->inodes_with_caps_lock);
926 			list_del_init(&ci->i_snap_realm_item);
927 			list_add(&ci->i_snap_realm_item,
928 				 &realm->inodes_with_caps);
929 			oldrealm = ci->i_snap_realm;
930 			ci->i_snap_realm = realm;
931 			spin_unlock(&realm->inodes_with_caps_lock);
932 			spin_unlock(&ci->i_ceph_lock);
933 
934 			ceph_get_snap_realm(mdsc, realm);
935 			ceph_put_snap_realm(mdsc, oldrealm);
936 
937 			iput(inode);
938 			continue;
939 
940 skip_inode:
941 			spin_unlock(&ci->i_ceph_lock);
942 			iput(inode);
943 		}
944 
945 		/* we may have taken some of the old realm's children. */
946 		for (i = 0; i < num_split_realms; i++) {
947 			struct ceph_snap_realm *child =
948 				__lookup_snap_realm(mdsc,
949 					   le64_to_cpu(split_realms[i]));
950 			if (!child)
951 				continue;
952 			adjust_snap_realm_parent(mdsc, child, realm->ino);
953 		}
954 	}
955 
956 	/*
957 	 * update using the provided snap trace. if we are deleting a
958 	 * snap, we can avoid queueing cap_snaps.
959 	 */
960 	ceph_update_snap_trace(mdsc, p, e,
961 			       op == CEPH_SNAP_OP_DESTROY, NULL);
962 
963 	if (op == CEPH_SNAP_OP_SPLIT)
964 		/* we took a reference when we created the realm, above */
965 		ceph_put_snap_realm(mdsc, realm);
966 
967 	__cleanup_empty_realms(mdsc);
968 
969 	up_write(&mdsc->snap_rwsem);
970 
971 	flush_snaps(mdsc);
972 	return;
973 
974 bad:
975 	pr_err("corrupt snap message from mds%d\n", mds);
976 	ceph_msg_dump(msg);
977 out:
978 	if (locked_rwsem)
979 		up_write(&mdsc->snap_rwsem);
980 	return;
981 }
982