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 */
 62void 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
 80static 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 */
107static 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 */
135static 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
155struct 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
165static void __put_snap_realm(struct ceph_mds_client *mdsc,
166			     struct ceph_snap_realm *realm);
167
168/*
169 * called with snap_rwsem (write)
170 */
171static 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 */
192static 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 */
204void 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 */
229static 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
245void 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 */
260static 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
289static 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 */
302static 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
380fail:
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 */
397static 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 */
413static 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
430static 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 */
453void 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
567update_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 */
586int __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 */
621static 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 */
662int 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);
676more:
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
770bad:
771	err = -EINVAL;
772fail:
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 */
788static 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 */
829void 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
946skip_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
980bad:
981	pr_err("corrupt snap message from mds%d\n", mds);
982	ceph_msg_dump(msg);
983out:
984	if (locked_rwsem)
985		up_write(&mdsc->snap_rwsem);
986	return;
987}