xref: /openbmc/linux/fs/ceph/snap.c (revision 3805e6a1)
1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/sort.h>
4 #include <linux/slab.h>
5 
6 #include "super.h"
7 #include "mds_client.h"
8 
9 #include <linux/ceph/decode.h>
10 
11 /*
12  * Snapshots in ceph are driven in large part by cooperation from the
13  * client.  In contrast to local file systems or file servers that
14  * implement snapshots at a single point in the system, ceph's
15  * distributed access to storage requires clients to help decide
16  * whether a write logically occurs before or after a recently created
17  * snapshot.
18  *
19  * This provides a perfect instantanous client-wide snapshot.  Between
20  * clients, however, snapshots may appear to be applied at slightly
21  * different points in time, depending on delays in delivering the
22  * snapshot notification.
23  *
24  * Snapshots are _not_ file system-wide.  Instead, each snapshot
25  * applies to the subdirectory nested beneath some directory.  This
26  * effectively divides the hierarchy into multiple "realms," where all
27  * of the files contained by each realm share the same set of
28  * snapshots.  An individual realm's snap set contains snapshots
29  * explicitly created on that realm, as well as any snaps in its
30  * parent's snap set _after_ the point at which the parent became it's
31  * parent (due to, say, a rename).  Similarly, snaps from prior parents
32  * during the time intervals during which they were the parent are included.
33  *
34  * The client is spared most of this detail, fortunately... it must only
35  * maintains a hierarchy of realms reflecting the current parent/child
36  * realm relationship, and for each realm has an explicit list of snaps
37  * inherited from prior parents.
38  *
39  * A snap_realm struct is maintained for realms containing every inode
40  * with an open cap in the system.  (The needed snap realm information is
41  * provided by the MDS whenever a cap is issued, i.e., on open.)  A 'seq'
42  * version number is used to ensure that as realm parameters change (new
43  * snapshot, new parent, etc.) the client's realm hierarchy is updated.
44  *
45  * The realm hierarchy drives the generation of a 'snap context' for each
46  * realm, which simply lists the resulting set of snaps for the realm.  This
47  * is attached to any writes sent to OSDs.
48  */
49 /*
50  * Unfortunately error handling is a bit mixed here.  If we get a snap
51  * update, but don't have enough memory to update our realm hierarchy,
52  * it's not clear what we can do about it (besides complaining to the
53  * console).
54  */
55 
56 
57 /*
58  * increase ref count for the realm
59  *
60  * caller must hold snap_rwsem for write.
61  */
62 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
63 			 struct ceph_snap_realm *realm)
64 {
65 	dout("get_realm %p %d -> %d\n", realm,
66 	     atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
67 	/*
68 	 * since we _only_ increment realm refs or empty the empty
69 	 * list with snap_rwsem held, adjusting the empty list here is
70 	 * safe.  we do need to protect against concurrent empty list
71 	 * additions, however.
72 	 */
73 	if (atomic_inc_return(&realm->nref) == 1) {
74 		spin_lock(&mdsc->snap_empty_lock);
75 		list_del_init(&realm->empty_item);
76 		spin_unlock(&mdsc->snap_empty_lock);
77 	}
78 }
79 
80 static void __insert_snap_realm(struct rb_root *root,
81 				struct ceph_snap_realm *new)
82 {
83 	struct rb_node **p = &root->rb_node;
84 	struct rb_node *parent = NULL;
85 	struct ceph_snap_realm *r = NULL;
86 
87 	while (*p) {
88 		parent = *p;
89 		r = rb_entry(parent, struct ceph_snap_realm, node);
90 		if (new->ino < r->ino)
91 			p = &(*p)->rb_left;
92 		else if (new->ino > r->ino)
93 			p = &(*p)->rb_right;
94 		else
95 			BUG();
96 	}
97 
98 	rb_link_node(&new->node, parent, p);
99 	rb_insert_color(&new->node, root);
100 }
101 
102 /*
103  * create and get the realm rooted at @ino and bump its ref count.
104  *
105  * caller must hold snap_rwsem for write.
106  */
107 static struct ceph_snap_realm *ceph_create_snap_realm(
108 	struct ceph_mds_client *mdsc,
109 	u64 ino)
110 {
111 	struct ceph_snap_realm *realm;
112 
113 	realm = kzalloc(sizeof(*realm), GFP_NOFS);
114 	if (!realm)
115 		return ERR_PTR(-ENOMEM);
116 
117 	atomic_set(&realm->nref, 1);    /* for caller */
118 	realm->ino = ino;
119 	INIT_LIST_HEAD(&realm->children);
120 	INIT_LIST_HEAD(&realm->child_item);
121 	INIT_LIST_HEAD(&realm->empty_item);
122 	INIT_LIST_HEAD(&realm->dirty_item);
123 	INIT_LIST_HEAD(&realm->inodes_with_caps);
124 	spin_lock_init(&realm->inodes_with_caps_lock);
125 	__insert_snap_realm(&mdsc->snap_realms, realm);
126 	dout("create_snap_realm %llx %p\n", realm->ino, realm);
127 	return realm;
128 }
129 
130 /*
131  * lookup the realm rooted at @ino.
132  *
133  * caller must hold snap_rwsem for write.
134  */
135 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
136 						   u64 ino)
137 {
138 	struct rb_node *n = mdsc->snap_realms.rb_node;
139 	struct ceph_snap_realm *r;
140 
141 	while (n) {
142 		r = rb_entry(n, struct ceph_snap_realm, node);
143 		if (ino < r->ino)
144 			n = n->rb_left;
145 		else if (ino > r->ino)
146 			n = n->rb_right;
147 		else {
148 			dout("lookup_snap_realm %llx %p\n", r->ino, r);
149 			return r;
150 		}
151 	}
152 	return NULL;
153 }
154 
155 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
156 					       u64 ino)
157 {
158 	struct ceph_snap_realm *r;
159 	r = __lookup_snap_realm(mdsc, ino);
160 	if (r)
161 		ceph_get_snap_realm(mdsc, r);
162 	return r;
163 }
164 
165 static void __put_snap_realm(struct ceph_mds_client *mdsc,
166 			     struct ceph_snap_realm *realm);
167 
168 /*
169  * called with snap_rwsem (write)
170  */
171 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
172 				 struct ceph_snap_realm *realm)
173 {
174 	dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
175 
176 	rb_erase(&realm->node, &mdsc->snap_realms);
177 
178 	if (realm->parent) {
179 		list_del_init(&realm->child_item);
180 		__put_snap_realm(mdsc, realm->parent);
181 	}
182 
183 	kfree(realm->prior_parent_snaps);
184 	kfree(realm->snaps);
185 	ceph_put_snap_context(realm->cached_context);
186 	kfree(realm);
187 }
188 
189 /*
190  * caller holds snap_rwsem (write)
191  */
192 static void __put_snap_realm(struct ceph_mds_client *mdsc,
193 			     struct ceph_snap_realm *realm)
194 {
195 	dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
196 	     atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
197 	if (atomic_dec_and_test(&realm->nref))
198 		__destroy_snap_realm(mdsc, realm);
199 }
200 
201 /*
202  * caller needn't hold any locks
203  */
204 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
205 			 struct ceph_snap_realm *realm)
206 {
207 	dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
208 	     atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
209 	if (!atomic_dec_and_test(&realm->nref))
210 		return;
211 
212 	if (down_write_trylock(&mdsc->snap_rwsem)) {
213 		__destroy_snap_realm(mdsc, realm);
214 		up_write(&mdsc->snap_rwsem);
215 	} else {
216 		spin_lock(&mdsc->snap_empty_lock);
217 		list_add(&realm->empty_item, &mdsc->snap_empty);
218 		spin_unlock(&mdsc->snap_empty_lock);
219 	}
220 }
221 
222 /*
223  * Clean up any realms whose ref counts have dropped to zero.  Note
224  * that this does not include realms who were created but not yet
225  * used.
226  *
227  * Called under snap_rwsem (write)
228  */
229 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
230 {
231 	struct ceph_snap_realm *realm;
232 
233 	spin_lock(&mdsc->snap_empty_lock);
234 	while (!list_empty(&mdsc->snap_empty)) {
235 		realm = list_first_entry(&mdsc->snap_empty,
236 				   struct ceph_snap_realm, empty_item);
237 		list_del(&realm->empty_item);
238 		spin_unlock(&mdsc->snap_empty_lock);
239 		__destroy_snap_realm(mdsc, realm);
240 		spin_lock(&mdsc->snap_empty_lock);
241 	}
242 	spin_unlock(&mdsc->snap_empty_lock);
243 }
244 
245 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
246 {
247 	down_write(&mdsc->snap_rwsem);
248 	__cleanup_empty_realms(mdsc);
249 	up_write(&mdsc->snap_rwsem);
250 }
251 
252 /*
253  * adjust the parent realm of a given @realm.  adjust child list, and parent
254  * pointers, and ref counts appropriately.
255  *
256  * return true if parent was changed, 0 if unchanged, <0 on error.
257  *
258  * caller must hold snap_rwsem for write.
259  */
260 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
261 				    struct ceph_snap_realm *realm,
262 				    u64 parentino)
263 {
264 	struct ceph_snap_realm *parent;
265 
266 	if (realm->parent_ino == parentino)
267 		return 0;
268 
269 	parent = ceph_lookup_snap_realm(mdsc, parentino);
270 	if (!parent) {
271 		parent = ceph_create_snap_realm(mdsc, parentino);
272 		if (IS_ERR(parent))
273 			return PTR_ERR(parent);
274 	}
275 	dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
276 	     realm->ino, realm, realm->parent_ino, realm->parent,
277 	     parentino, parent);
278 	if (realm->parent) {
279 		list_del_init(&realm->child_item);
280 		ceph_put_snap_realm(mdsc, realm->parent);
281 	}
282 	realm->parent_ino = parentino;
283 	realm->parent = parent;
284 	list_add(&realm->child_item, &parent->children);
285 	return 1;
286 }
287 
288 
289 static int cmpu64_rev(const void *a, const void *b)
290 {
291 	if (*(u64 *)a < *(u64 *)b)
292 		return 1;
293 	if (*(u64 *)a > *(u64 *)b)
294 		return -1;
295 	return 0;
296 }
297 
298 
299 /*
300  * build the snap context for a given realm.
301  */
302 static int build_snap_context(struct ceph_snap_realm *realm)
303 {
304 	struct ceph_snap_realm *parent = realm->parent;
305 	struct ceph_snap_context *snapc;
306 	int err = 0;
307 	u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
308 
309 	/*
310 	 * build parent context, if it hasn't been built.
311 	 * conservatively estimate that all parent snaps might be
312 	 * included by us.
313 	 */
314 	if (parent) {
315 		if (!parent->cached_context) {
316 			err = build_snap_context(parent);
317 			if (err)
318 				goto fail;
319 		}
320 		num += parent->cached_context->num_snaps;
321 	}
322 
323 	/* do i actually need to update?  not if my context seq
324 	   matches realm seq, and my parents' does to.  (this works
325 	   because we rebuild_snap_realms() works _downward_ in
326 	   hierarchy after each update.) */
327 	if (realm->cached_context &&
328 	    realm->cached_context->seq == realm->seq &&
329 	    (!parent ||
330 	     realm->cached_context->seq >= parent->cached_context->seq)) {
331 		dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
332 		     " (unchanged)\n",
333 		     realm->ino, realm, realm->cached_context,
334 		     realm->cached_context->seq,
335 		     (unsigned int) realm->cached_context->num_snaps);
336 		return 0;
337 	}
338 
339 	/* alloc new snap context */
340 	err = -ENOMEM;
341 	if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
342 		goto fail;
343 	snapc = ceph_create_snap_context(num, GFP_NOFS);
344 	if (!snapc)
345 		goto fail;
346 
347 	/* build (reverse sorted) snap vector */
348 	num = 0;
349 	snapc->seq = realm->seq;
350 	if (parent) {
351 		u32 i;
352 
353 		/* include any of parent's snaps occurring _after_ my
354 		   parent became my parent */
355 		for (i = 0; i < parent->cached_context->num_snaps; i++)
356 			if (parent->cached_context->snaps[i] >=
357 			    realm->parent_since)
358 				snapc->snaps[num++] =
359 					parent->cached_context->snaps[i];
360 		if (parent->cached_context->seq > snapc->seq)
361 			snapc->seq = parent->cached_context->seq;
362 	}
363 	memcpy(snapc->snaps + num, realm->snaps,
364 	       sizeof(u64)*realm->num_snaps);
365 	num += realm->num_snaps;
366 	memcpy(snapc->snaps + num, realm->prior_parent_snaps,
367 	       sizeof(u64)*realm->num_prior_parent_snaps);
368 	num += realm->num_prior_parent_snaps;
369 
370 	sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
371 	snapc->num_snaps = num;
372 	dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
373 	     realm->ino, realm, snapc, snapc->seq,
374 	     (unsigned int) snapc->num_snaps);
375 
376 	ceph_put_snap_context(realm->cached_context);
377 	realm->cached_context = snapc;
378 	return 0;
379 
380 fail:
381 	/*
382 	 * if we fail, clear old (incorrect) cached_context... hopefully
383 	 * we'll have better luck building it later
384 	 */
385 	if (realm->cached_context) {
386 		ceph_put_snap_context(realm->cached_context);
387 		realm->cached_context = NULL;
388 	}
389 	pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
390 	       realm, err);
391 	return err;
392 }
393 
394 /*
395  * rebuild snap context for the given realm and all of its children.
396  */
397 static void rebuild_snap_realms(struct ceph_snap_realm *realm)
398 {
399 	struct ceph_snap_realm *child;
400 
401 	dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
402 	build_snap_context(realm);
403 
404 	list_for_each_entry(child, &realm->children, child_item)
405 		rebuild_snap_realms(child);
406 }
407 
408 
409 /*
410  * helper to allocate and decode an array of snapids.  free prior
411  * instance, if any.
412  */
413 static int dup_array(u64 **dst, __le64 *src, u32 num)
414 {
415 	u32 i;
416 
417 	kfree(*dst);
418 	if (num) {
419 		*dst = kcalloc(num, sizeof(u64), GFP_NOFS);
420 		if (!*dst)
421 			return -ENOMEM;
422 		for (i = 0; i < num; i++)
423 			(*dst)[i] = get_unaligned_le64(src + i);
424 	} else {
425 		*dst = NULL;
426 	}
427 	return 0;
428 }
429 
430 static bool has_new_snaps(struct ceph_snap_context *o,
431 			  struct ceph_snap_context *n)
432 {
433 	if (n->num_snaps == 0)
434 		return false;
435 	/* snaps are in descending order */
436 	return n->snaps[0] > o->seq;
437 }
438 
439 /*
440  * When a snapshot is applied, the size/mtime inode metadata is queued
441  * in a ceph_cap_snap (one for each snapshot) until writeback
442  * completes and the metadata can be flushed back to the MDS.
443  *
444  * However, if a (sync) write is currently in-progress when we apply
445  * the snapshot, we have to wait until the write succeeds or fails
446  * (and a final size/mtime is known).  In this case the
447  * cap_snap->writing = 1, and is said to be "pending."  When the write
448  * finishes, we __ceph_finish_cap_snap().
449  *
450  * Caller must hold snap_rwsem for read (i.e., the realm topology won't
451  * change).
452  */
453 void ceph_queue_cap_snap(struct ceph_inode_info *ci)
454 {
455 	struct inode *inode = &ci->vfs_inode;
456 	struct ceph_cap_snap *capsnap;
457 	struct ceph_snap_context *old_snapc, *new_snapc;
458 	int used, dirty;
459 
460 	capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
461 	if (!capsnap) {
462 		pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
463 		return;
464 	}
465 
466 	spin_lock(&ci->i_ceph_lock);
467 	used = __ceph_caps_used(ci);
468 	dirty = __ceph_caps_dirty(ci);
469 
470 	old_snapc = ci->i_head_snapc;
471 	new_snapc = ci->i_snap_realm->cached_context;
472 
473 	/*
474 	 * If there is a write in progress, treat that as a dirty Fw,
475 	 * even though it hasn't completed yet; by the time we finish
476 	 * up this capsnap it will be.
477 	 */
478 	if (used & CEPH_CAP_FILE_WR)
479 		dirty |= CEPH_CAP_FILE_WR;
480 
481 	if (__ceph_have_pending_cap_snap(ci)) {
482 		/* there is no point in queuing multiple "pending" cap_snaps,
483 		   as no new writes are allowed to start when pending, so any
484 		   writes in progress now were started before the previous
485 		   cap_snap.  lucky us. */
486 		dout("queue_cap_snap %p already pending\n", inode);
487 		goto update_snapc;
488 	}
489 	if (ci->i_wrbuffer_ref_head == 0 &&
490 	    !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
491 		dout("queue_cap_snap %p nothing dirty|writing\n", inode);
492 		goto update_snapc;
493 	}
494 
495 	BUG_ON(!old_snapc);
496 
497 	/*
498 	 * There is no need to send FLUSHSNAP message to MDS if there is
499 	 * no new snapshot. But when there is dirty pages or on-going
500 	 * writes, we still need to create cap_snap. cap_snap is needed
501 	 * by the write path and page writeback path.
502 	 *
503 	 * also see ceph_try_drop_cap_snap()
504 	 */
505 	if (has_new_snaps(old_snapc, new_snapc)) {
506 		if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
507 			capsnap->need_flush = true;
508 	} else {
509 		if (!(used & CEPH_CAP_FILE_WR) &&
510 		    ci->i_wrbuffer_ref_head == 0) {
511 			dout("queue_cap_snap %p "
512 			     "no new_snap|dirty_page|writing\n", inode);
513 			goto update_snapc;
514 		}
515 	}
516 
517 	dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n",
518 	     inode, capsnap, old_snapc, ceph_cap_string(dirty),
519 	     capsnap->need_flush ? "" : "no_flush");
520 	ihold(inode);
521 
522 	atomic_set(&capsnap->nref, 1);
523 	capsnap->ci = ci;
524 	INIT_LIST_HEAD(&capsnap->ci_item);
525 	INIT_LIST_HEAD(&capsnap->flushing_item);
526 
527 	capsnap->follows = old_snapc->seq;
528 	capsnap->issued = __ceph_caps_issued(ci, NULL);
529 	capsnap->dirty = dirty;
530 
531 	capsnap->mode = inode->i_mode;
532 	capsnap->uid = inode->i_uid;
533 	capsnap->gid = inode->i_gid;
534 
535 	if (dirty & CEPH_CAP_XATTR_EXCL) {
536 		__ceph_build_xattrs_blob(ci);
537 		capsnap->xattr_blob =
538 			ceph_buffer_get(ci->i_xattrs.blob);
539 		capsnap->xattr_version = ci->i_xattrs.version;
540 	} else {
541 		capsnap->xattr_blob = NULL;
542 		capsnap->xattr_version = 0;
543 	}
544 
545 	capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
546 
547 	/* dirty page count moved from _head to this cap_snap;
548 	   all subsequent writes page dirties occur _after_ this
549 	   snapshot. */
550 	capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
551 	ci->i_wrbuffer_ref_head = 0;
552 	capsnap->context = old_snapc;
553 	list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
554 	old_snapc = NULL;
555 
556 	if (used & CEPH_CAP_FILE_WR) {
557 		dout("queue_cap_snap %p cap_snap %p snapc %p"
558 		     " seq %llu used WR, now pending\n", inode,
559 		     capsnap, old_snapc, old_snapc->seq);
560 		capsnap->writing = 1;
561 	} else {
562 		/* note mtime, size NOW. */
563 		__ceph_finish_cap_snap(ci, capsnap);
564 	}
565 	capsnap = NULL;
566 
567 update_snapc:
568 	if (ci->i_head_snapc) {
569 		ci->i_head_snapc = ceph_get_snap_context(new_snapc);
570 		dout(" new snapc is %p\n", new_snapc);
571 	}
572 	spin_unlock(&ci->i_ceph_lock);
573 
574 	kfree(capsnap);
575 	ceph_put_snap_context(old_snapc);
576 }
577 
578 /*
579  * Finalize the size, mtime for a cap_snap.. that is, settle on final values
580  * to be used for the snapshot, to be flushed back to the mds.
581  *
582  * If capsnap can now be flushed, add to snap_flush list, and return 1.
583  *
584  * Caller must hold i_ceph_lock.
585  */
586 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
587 			    struct ceph_cap_snap *capsnap)
588 {
589 	struct inode *inode = &ci->vfs_inode;
590 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
591 
592 	BUG_ON(capsnap->writing);
593 	capsnap->size = inode->i_size;
594 	capsnap->mtime = inode->i_mtime;
595 	capsnap->atime = inode->i_atime;
596 	capsnap->ctime = inode->i_ctime;
597 	capsnap->time_warp_seq = ci->i_time_warp_seq;
598 	if (capsnap->dirty_pages) {
599 		dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
600 		     "still has %d dirty pages\n", inode, capsnap,
601 		     capsnap->context, capsnap->context->seq,
602 		     ceph_cap_string(capsnap->dirty), capsnap->size,
603 		     capsnap->dirty_pages);
604 		return 0;
605 	}
606 	dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
607 	     inode, capsnap, capsnap->context,
608 	     capsnap->context->seq, ceph_cap_string(capsnap->dirty),
609 	     capsnap->size);
610 
611 	spin_lock(&mdsc->snap_flush_lock);
612 	list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
613 	spin_unlock(&mdsc->snap_flush_lock);
614 	return 1;  /* caller may want to ceph_flush_snaps */
615 }
616 
617 /*
618  * Queue cap_snaps for snap writeback for this realm and its children.
619  * Called under snap_rwsem, so realm topology won't change.
620  */
621 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
622 {
623 	struct ceph_inode_info *ci;
624 	struct inode *lastinode = NULL;
625 	struct ceph_snap_realm *child;
626 
627 	dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
628 
629 	spin_lock(&realm->inodes_with_caps_lock);
630 	list_for_each_entry(ci, &realm->inodes_with_caps,
631 			    i_snap_realm_item) {
632 		struct inode *inode = igrab(&ci->vfs_inode);
633 		if (!inode)
634 			continue;
635 		spin_unlock(&realm->inodes_with_caps_lock);
636 		iput(lastinode);
637 		lastinode = inode;
638 		ceph_queue_cap_snap(ci);
639 		spin_lock(&realm->inodes_with_caps_lock);
640 	}
641 	spin_unlock(&realm->inodes_with_caps_lock);
642 	iput(lastinode);
643 
644 	list_for_each_entry(child, &realm->children, child_item) {
645 		dout("queue_realm_cap_snaps %p %llx queue child %p %llx\n",
646 		     realm, realm->ino, child, child->ino);
647 		list_del_init(&child->dirty_item);
648 		list_add(&child->dirty_item, &realm->dirty_item);
649 	}
650 
651 	list_del_init(&realm->dirty_item);
652 	dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
653 }
654 
655 /*
656  * Parse and apply a snapblob "snap trace" from the MDS.  This specifies
657  * the snap realm parameters from a given realm and all of its ancestors,
658  * up to the root.
659  *
660  * Caller must hold snap_rwsem for write.
661  */
662 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
663 			   void *p, void *e, bool deletion,
664 			   struct ceph_snap_realm **realm_ret)
665 {
666 	struct ceph_mds_snap_realm *ri;    /* encoded */
667 	__le64 *snaps;                     /* encoded */
668 	__le64 *prior_parent_snaps;        /* encoded */
669 	struct ceph_snap_realm *realm = NULL;
670 	struct ceph_snap_realm *first_realm = NULL;
671 	int invalidate = 0;
672 	int err = -ENOMEM;
673 	LIST_HEAD(dirty_realms);
674 
675 	dout("update_snap_trace deletion=%d\n", deletion);
676 more:
677 	ceph_decode_need(&p, e, sizeof(*ri), bad);
678 	ri = p;
679 	p += sizeof(*ri);
680 	ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
681 			    le32_to_cpu(ri->num_prior_parent_snaps)), bad);
682 	snaps = p;
683 	p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
684 	prior_parent_snaps = p;
685 	p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
686 
687 	realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
688 	if (!realm) {
689 		realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
690 		if (IS_ERR(realm)) {
691 			err = PTR_ERR(realm);
692 			goto fail;
693 		}
694 	}
695 
696 	/* ensure the parent is correct */
697 	err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
698 	if (err < 0)
699 		goto fail;
700 	invalidate += err;
701 
702 	if (le64_to_cpu(ri->seq) > realm->seq) {
703 		dout("update_snap_trace updating %llx %p %lld -> %lld\n",
704 		     realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
705 		/* update realm parameters, snap lists */
706 		realm->seq = le64_to_cpu(ri->seq);
707 		realm->created = le64_to_cpu(ri->created);
708 		realm->parent_since = le64_to_cpu(ri->parent_since);
709 
710 		realm->num_snaps = le32_to_cpu(ri->num_snaps);
711 		err = dup_array(&realm->snaps, snaps, realm->num_snaps);
712 		if (err < 0)
713 			goto fail;
714 
715 		realm->num_prior_parent_snaps =
716 			le32_to_cpu(ri->num_prior_parent_snaps);
717 		err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
718 				realm->num_prior_parent_snaps);
719 		if (err < 0)
720 			goto fail;
721 
722 		/* queue realm for cap_snap creation */
723 		list_add(&realm->dirty_item, &dirty_realms);
724 		if (realm->seq > mdsc->last_snap_seq)
725 			mdsc->last_snap_seq = realm->seq;
726 
727 		invalidate = 1;
728 	} else if (!realm->cached_context) {
729 		dout("update_snap_trace %llx %p seq %lld new\n",
730 		     realm->ino, realm, realm->seq);
731 		invalidate = 1;
732 	} else {
733 		dout("update_snap_trace %llx %p seq %lld unchanged\n",
734 		     realm->ino, realm, realm->seq);
735 	}
736 
737 	dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
738 	     realm, invalidate, p, e);
739 
740 	/* invalidate when we reach the _end_ (root) of the trace */
741 	if (invalidate && p >= e)
742 		rebuild_snap_realms(realm);
743 
744 	if (!first_realm)
745 		first_realm = realm;
746 	else
747 		ceph_put_snap_realm(mdsc, realm);
748 
749 	if (p < e)
750 		goto more;
751 
752 	/*
753 	 * queue cap snaps _after_ we've built the new snap contexts,
754 	 * so that i_head_snapc can be set appropriately.
755 	 */
756 	while (!list_empty(&dirty_realms)) {
757 		realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
758 					 dirty_item);
759 		queue_realm_cap_snaps(realm);
760 	}
761 
762 	if (realm_ret)
763 		*realm_ret = first_realm;
764 	else
765 		ceph_put_snap_realm(mdsc, first_realm);
766 
767 	__cleanup_empty_realms(mdsc);
768 	return 0;
769 
770 bad:
771 	err = -EINVAL;
772 fail:
773 	if (realm && !IS_ERR(realm))
774 		ceph_put_snap_realm(mdsc, realm);
775 	if (first_realm)
776 		ceph_put_snap_realm(mdsc, first_realm);
777 	pr_err("update_snap_trace error %d\n", err);
778 	return err;
779 }
780 
781 
782 /*
783  * Send any cap_snaps that are queued for flush.  Try to carry
784  * s_mutex across multiple snap flushes to avoid locking overhead.
785  *
786  * Caller holds no locks.
787  */
788 static void flush_snaps(struct ceph_mds_client *mdsc)
789 {
790 	struct ceph_inode_info *ci;
791 	struct inode *inode;
792 	struct ceph_mds_session *session = NULL;
793 
794 	dout("flush_snaps\n");
795 	spin_lock(&mdsc->snap_flush_lock);
796 	while (!list_empty(&mdsc->snap_flush_list)) {
797 		ci = list_first_entry(&mdsc->snap_flush_list,
798 				struct ceph_inode_info, i_snap_flush_item);
799 		inode = &ci->vfs_inode;
800 		ihold(inode);
801 		spin_unlock(&mdsc->snap_flush_lock);
802 		spin_lock(&ci->i_ceph_lock);
803 		__ceph_flush_snaps(ci, &session, 0);
804 		spin_unlock(&ci->i_ceph_lock);
805 		iput(inode);
806 		spin_lock(&mdsc->snap_flush_lock);
807 	}
808 	spin_unlock(&mdsc->snap_flush_lock);
809 
810 	if (session) {
811 		mutex_unlock(&session->s_mutex);
812 		ceph_put_mds_session(session);
813 	}
814 	dout("flush_snaps done\n");
815 }
816 
817 
818 /*
819  * Handle a snap notification from the MDS.
820  *
821  * This can take two basic forms: the simplest is just a snap creation
822  * or deletion notification on an existing realm.  This should update the
823  * realm and its children.
824  *
825  * The more difficult case is realm creation, due to snap creation at a
826  * new point in the file hierarchy, or due to a rename that moves a file or
827  * directory into another realm.
828  */
829 void ceph_handle_snap(struct ceph_mds_client *mdsc,
830 		      struct ceph_mds_session *session,
831 		      struct ceph_msg *msg)
832 {
833 	struct super_block *sb = mdsc->fsc->sb;
834 	int mds = session->s_mds;
835 	u64 split;
836 	int op;
837 	int trace_len;
838 	struct ceph_snap_realm *realm = NULL;
839 	void *p = msg->front.iov_base;
840 	void *e = p + msg->front.iov_len;
841 	struct ceph_mds_snap_head *h;
842 	int num_split_inos, num_split_realms;
843 	__le64 *split_inos = NULL, *split_realms = NULL;
844 	int i;
845 	int locked_rwsem = 0;
846 
847 	/* decode */
848 	if (msg->front.iov_len < sizeof(*h))
849 		goto bad;
850 	h = p;
851 	op = le32_to_cpu(h->op);
852 	split = le64_to_cpu(h->split);   /* non-zero if we are splitting an
853 					  * existing realm */
854 	num_split_inos = le32_to_cpu(h->num_split_inos);
855 	num_split_realms = le32_to_cpu(h->num_split_realms);
856 	trace_len = le32_to_cpu(h->trace_len);
857 	p += sizeof(*h);
858 
859 	dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
860 	     ceph_snap_op_name(op), split, trace_len);
861 
862 	mutex_lock(&session->s_mutex);
863 	session->s_seq++;
864 	mutex_unlock(&session->s_mutex);
865 
866 	down_write(&mdsc->snap_rwsem);
867 	locked_rwsem = 1;
868 
869 	if (op == CEPH_SNAP_OP_SPLIT) {
870 		struct ceph_mds_snap_realm *ri;
871 
872 		/*
873 		 * A "split" breaks part of an existing realm off into
874 		 * a new realm.  The MDS provides a list of inodes
875 		 * (with caps) and child realms that belong to the new
876 		 * child.
877 		 */
878 		split_inos = p;
879 		p += sizeof(u64) * num_split_inos;
880 		split_realms = p;
881 		p += sizeof(u64) * num_split_realms;
882 		ceph_decode_need(&p, e, sizeof(*ri), bad);
883 		/* we will peek at realm info here, but will _not_
884 		 * advance p, as the realm update will occur below in
885 		 * ceph_update_snap_trace. */
886 		ri = p;
887 
888 		realm = ceph_lookup_snap_realm(mdsc, split);
889 		if (!realm) {
890 			realm = ceph_create_snap_realm(mdsc, split);
891 			if (IS_ERR(realm))
892 				goto out;
893 		}
894 
895 		dout("splitting snap_realm %llx %p\n", realm->ino, realm);
896 		for (i = 0; i < num_split_inos; i++) {
897 			struct ceph_vino vino = {
898 				.ino = le64_to_cpu(split_inos[i]),
899 				.snap = CEPH_NOSNAP,
900 			};
901 			struct inode *inode = ceph_find_inode(sb, vino);
902 			struct ceph_inode_info *ci;
903 			struct ceph_snap_realm *oldrealm;
904 
905 			if (!inode)
906 				continue;
907 			ci = ceph_inode(inode);
908 
909 			spin_lock(&ci->i_ceph_lock);
910 			if (!ci->i_snap_realm)
911 				goto skip_inode;
912 			/*
913 			 * If this inode belongs to a realm that was
914 			 * created after our new realm, we experienced
915 			 * a race (due to another split notifications
916 			 * arriving from a different MDS).  So skip
917 			 * this inode.
918 			 */
919 			if (ci->i_snap_realm->created >
920 			    le64_to_cpu(ri->created)) {
921 				dout(" leaving %p in newer realm %llx %p\n",
922 				     inode, ci->i_snap_realm->ino,
923 				     ci->i_snap_realm);
924 				goto skip_inode;
925 			}
926 			dout(" will move %p to split realm %llx %p\n",
927 			     inode, realm->ino, realm);
928 			/*
929 			 * Move the inode to the new realm
930 			 */
931 			spin_lock(&realm->inodes_with_caps_lock);
932 			list_del_init(&ci->i_snap_realm_item);
933 			list_add(&ci->i_snap_realm_item,
934 				 &realm->inodes_with_caps);
935 			oldrealm = ci->i_snap_realm;
936 			ci->i_snap_realm = realm;
937 			spin_unlock(&realm->inodes_with_caps_lock);
938 			spin_unlock(&ci->i_ceph_lock);
939 
940 			ceph_get_snap_realm(mdsc, realm);
941 			ceph_put_snap_realm(mdsc, oldrealm);
942 
943 			iput(inode);
944 			continue;
945 
946 skip_inode:
947 			spin_unlock(&ci->i_ceph_lock);
948 			iput(inode);
949 		}
950 
951 		/* we may have taken some of the old realm's children. */
952 		for (i = 0; i < num_split_realms; i++) {
953 			struct ceph_snap_realm *child =
954 				__lookup_snap_realm(mdsc,
955 					   le64_to_cpu(split_realms[i]));
956 			if (!child)
957 				continue;
958 			adjust_snap_realm_parent(mdsc, child, realm->ino);
959 		}
960 	}
961 
962 	/*
963 	 * update using the provided snap trace. if we are deleting a
964 	 * snap, we can avoid queueing cap_snaps.
965 	 */
966 	ceph_update_snap_trace(mdsc, p, e,
967 			       op == CEPH_SNAP_OP_DESTROY, NULL);
968 
969 	if (op == CEPH_SNAP_OP_SPLIT)
970 		/* we took a reference when we created the realm, above */
971 		ceph_put_snap_realm(mdsc, realm);
972 
973 	__cleanup_empty_realms(mdsc);
974 
975 	up_write(&mdsc->snap_rwsem);
976 
977 	flush_snaps(mdsc);
978 	return;
979 
980 bad:
981 	pr_err("corrupt snap message from mds%d\n", mds);
982 	ceph_msg_dump(msg);
983 out:
984 	if (locked_rwsem)
985 		up_write(&mdsc->snap_rwsem);
986 	return;
987 }
988