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