1 #include <linux/ceph/ceph_debug.h>
3 #include <linux/sort.h>
4 #include <linux/slab.h>
7 #include "mds_client.h"
9 #include <linux/ceph/decode.h>
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
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.
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.
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.
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.
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.
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
58 * increase ref count for the realm
60 * caller must hold snap_rwsem for write.
62 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
63 struct ceph_snap_realm *realm)
65 dout("get_realm %p %d -> %d\n", realm,
66 atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
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
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);
79 atomic_inc(&realm->nref);
82 static void __insert_snap_realm(struct rb_root *root,
83 struct ceph_snap_realm *new)
85 struct rb_node **p = &root->rb_node;
86 struct rb_node *parent = NULL;
87 struct ceph_snap_realm *r = NULL;
91 r = rb_entry(parent, struct ceph_snap_realm, node);
92 if (new->ino < r->ino)
94 else if (new->ino > r->ino)
100 rb_link_node(&new->node, parent, p);
101 rb_insert_color(&new->node, root);
105 * create and get the realm rooted at @ino and bump its ref count.
107 * caller must hold snap_rwsem for write.
109 static struct ceph_snap_realm *ceph_create_snap_realm(
110 struct ceph_mds_client *mdsc,
113 struct ceph_snap_realm *realm;
115 realm = kzalloc(sizeof(*realm), GFP_NOFS);
117 return ERR_PTR(-ENOMEM);
119 atomic_set(&realm->nref, 0); /* tree does not take a ref */
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);
133 * lookup the realm rooted at @ino.
135 * caller must hold snap_rwsem for write.
137 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
140 struct rb_node *n = mdsc->snap_realms.rb_node;
141 struct ceph_snap_realm *r;
144 r = rb_entry(n, struct ceph_snap_realm, node);
147 else if (ino > r->ino)
150 dout("lookup_snap_realm %llx %p\n", r->ino, r);
157 static void __put_snap_realm(struct ceph_mds_client *mdsc,
158 struct ceph_snap_realm *realm);
161 * called with snap_rwsem (write)
163 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
164 struct ceph_snap_realm *realm)
166 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
168 rb_erase(&realm->node, &mdsc->snap_realms);
171 list_del_init(&realm->child_item);
172 __put_snap_realm(mdsc, realm->parent);
175 kfree(realm->prior_parent_snaps);
177 ceph_put_snap_context(realm->cached_context);
182 * caller holds snap_rwsem (write)
184 static void __put_snap_realm(struct ceph_mds_client *mdsc,
185 struct ceph_snap_realm *realm)
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);
194 * caller needn't hold any locks
196 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
197 struct ceph_snap_realm *realm)
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))
204 if (down_write_trylock(&mdsc->snap_rwsem)) {
205 __destroy_snap_realm(mdsc, realm);
206 up_write(&mdsc->snap_rwsem);
208 spin_lock(&mdsc->snap_empty_lock);
209 list_add(&realm->empty_item, &mdsc->snap_empty);
210 spin_unlock(&mdsc->snap_empty_lock);
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
219 * Called under snap_rwsem (write)
221 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
223 struct ceph_snap_realm *realm;
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);
234 spin_unlock(&mdsc->snap_empty_lock);
237 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
239 down_write(&mdsc->snap_rwsem);
240 __cleanup_empty_realms(mdsc);
241 up_write(&mdsc->snap_rwsem);
245 * adjust the parent realm of a given @realm. adjust child list, and parent
246 * pointers, and ref counts appropriately.
248 * return true if parent was changed, 0 if unchanged, <0 on error.
250 * caller must hold snap_rwsem for write.
252 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
253 struct ceph_snap_realm *realm,
256 struct ceph_snap_realm *parent;
258 if (realm->parent_ino == parentino)
261 parent = ceph_lookup_snap_realm(mdsc, parentino);
263 parent = ceph_create_snap_realm(mdsc, parentino);
265 return PTR_ERR(parent);
267 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
268 realm->ino, realm, realm->parent_ino, realm->parent,
271 list_del_init(&realm->child_item);
272 ceph_put_snap_realm(mdsc, realm->parent);
274 realm->parent_ino = parentino;
275 realm->parent = parent;
276 ceph_get_snap_realm(mdsc, parent);
277 list_add(&realm->child_item, &parent->children);
282 static int cmpu64_rev(const void *a, const void *b)
284 if (*(u64 *)a < *(u64 *)b)
286 if (*(u64 *)a > *(u64 *)b)
292 static struct ceph_snap_context *empty_snapc;
295 * build the snap context for a given realm.
297 static int build_snap_context(struct ceph_snap_realm *realm)
299 struct ceph_snap_realm *parent = realm->parent;
300 struct ceph_snap_context *snapc;
302 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
305 * build parent context, if it hasn't been built.
306 * conservatively estimate that all parent snaps might be
310 if (!parent->cached_context) {
311 err = build_snap_context(parent);
315 num += parent->cached_context->num_snaps;
318 /* do i actually need to update? not if my context seq
319 matches realm seq, and my parents' does to. (this works
320 because we rebuild_snap_realms() works _downward_ in
321 hierarchy after each update.) */
322 if (realm->cached_context &&
323 realm->cached_context->seq == realm->seq &&
325 realm->cached_context->seq >= parent->cached_context->seq)) {
326 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
328 realm->ino, realm, realm->cached_context,
329 realm->cached_context->seq,
330 (unsigned int) realm->cached_context->num_snaps);
334 if (num == 0 && realm->seq == empty_snapc->seq) {
335 ceph_get_snap_context(empty_snapc);
340 /* alloc new snap context */
342 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
344 snapc = ceph_create_snap_context(num, GFP_NOFS);
348 /* build (reverse sorted) snap vector */
350 snapc->seq = realm->seq;
354 /* include any of parent's snaps occurring _after_ my
355 parent became my parent */
356 for (i = 0; i < parent->cached_context->num_snaps; i++)
357 if (parent->cached_context->snaps[i] >=
359 snapc->snaps[num++] =
360 parent->cached_context->snaps[i];
361 if (parent->cached_context->seq > snapc->seq)
362 snapc->seq = parent->cached_context->seq;
364 memcpy(snapc->snaps + num, realm->snaps,
365 sizeof(u64)*realm->num_snaps);
366 num += realm->num_snaps;
367 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
368 sizeof(u64)*realm->num_prior_parent_snaps);
369 num += realm->num_prior_parent_snaps;
371 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
372 snapc->num_snaps = num;
373 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
374 realm->ino, realm, snapc, snapc->seq,
375 (unsigned int) snapc->num_snaps);
378 ceph_put_snap_context(realm->cached_context);
379 realm->cached_context = snapc;
384 * if we fail, clear old (incorrect) cached_context... hopefully
385 * we'll have better luck building it later
387 if (realm->cached_context) {
388 ceph_put_snap_context(realm->cached_context);
389 realm->cached_context = NULL;
391 pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
397 * rebuild snap context for the given realm and all of its children.
399 static void rebuild_snap_realms(struct ceph_snap_realm *realm)
401 struct ceph_snap_realm *child;
403 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
404 build_snap_context(realm);
406 list_for_each_entry(child, &realm->children, child_item)
407 rebuild_snap_realms(child);
412 * helper to allocate and decode an array of snapids. free prior
415 static int dup_array(u64 **dst, __le64 *src, u32 num)
421 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
424 for (i = 0; i < num; i++)
425 (*dst)[i] = get_unaligned_le64(src + i);
434 * When a snapshot is applied, the size/mtime inode metadata is queued
435 * in a ceph_cap_snap (one for each snapshot) until writeback
436 * completes and the metadata can be flushed back to the MDS.
438 * However, if a (sync) write is currently in-progress when we apply
439 * the snapshot, we have to wait until the write succeeds or fails
440 * (and a final size/mtime is known). In this case the
441 * cap_snap->writing = 1, and is said to be "pending." When the write
442 * finishes, we __ceph_finish_cap_snap().
444 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
447 void ceph_queue_cap_snap(struct ceph_inode_info *ci)
449 struct inode *inode = &ci->vfs_inode;
450 struct ceph_cap_snap *capsnap;
453 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
455 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
459 spin_lock(&ci->i_ceph_lock);
460 used = __ceph_caps_used(ci);
461 dirty = __ceph_caps_dirty(ci);
464 * If there is a write in progress, treat that as a dirty Fw,
465 * even though it hasn't completed yet; by the time we finish
466 * up this capsnap it will be.
468 if (used & CEPH_CAP_FILE_WR)
469 dirty |= CEPH_CAP_FILE_WR;
471 if (__ceph_have_pending_cap_snap(ci)) {
472 /* there is no point in queuing multiple "pending" cap_snaps,
473 as no new writes are allowed to start when pending, so any
474 writes in progress now were started before the previous
475 cap_snap. lucky us. */
476 dout("queue_cap_snap %p already pending\n", inode);
478 } else if (ci->i_snap_realm->cached_context == empty_snapc) {
479 dout("queue_cap_snap %p empty snapc\n", inode);
481 } else if (dirty & (CEPH_CAP_AUTH_EXCL|CEPH_CAP_XATTR_EXCL|
482 CEPH_CAP_FILE_EXCL|CEPH_CAP_FILE_WR)) {
483 struct ceph_snap_context *snapc = ci->i_head_snapc;
486 * if we are a sync write, we may need to go to the snaprealm
487 * to get the current snapc.
490 snapc = ci->i_snap_realm->cached_context;
492 dout("queue_cap_snap %p cap_snap %p queuing under %p %s\n",
493 inode, capsnap, snapc, ceph_cap_string(dirty));
496 atomic_set(&capsnap->nref, 1);
498 INIT_LIST_HEAD(&capsnap->ci_item);
499 INIT_LIST_HEAD(&capsnap->flushing_item);
501 capsnap->follows = snapc->seq;
502 capsnap->issued = __ceph_caps_issued(ci, NULL);
503 capsnap->dirty = dirty;
505 capsnap->mode = inode->i_mode;
506 capsnap->uid = inode->i_uid;
507 capsnap->gid = inode->i_gid;
509 if (dirty & CEPH_CAP_XATTR_EXCL) {
510 __ceph_build_xattrs_blob(ci);
511 capsnap->xattr_blob =
512 ceph_buffer_get(ci->i_xattrs.blob);
513 capsnap->xattr_version = ci->i_xattrs.version;
515 capsnap->xattr_blob = NULL;
516 capsnap->xattr_version = 0;
519 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
521 /* dirty page count moved from _head to this cap_snap;
522 all subsequent writes page dirties occur _after_ this
524 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
525 ci->i_wrbuffer_ref_head = 0;
526 capsnap->context = snapc;
528 ceph_get_snap_context(ci->i_snap_realm->cached_context);
529 dout(" new snapc is %p\n", ci->i_head_snapc);
530 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
532 if (used & CEPH_CAP_FILE_WR) {
533 dout("queue_cap_snap %p cap_snap %p snapc %p"
534 " seq %llu used WR, now pending\n", inode,
535 capsnap, snapc, snapc->seq);
536 capsnap->writing = 1;
538 /* note mtime, size NOW. */
539 __ceph_finish_cap_snap(ci, capsnap);
542 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
546 spin_unlock(&ci->i_ceph_lock);
550 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
551 * to be used for the snapshot, to be flushed back to the mds.
553 * If capsnap can now be flushed, add to snap_flush list, and return 1.
555 * Caller must hold i_ceph_lock.
557 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
558 struct ceph_cap_snap *capsnap)
560 struct inode *inode = &ci->vfs_inode;
561 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
563 BUG_ON(capsnap->writing);
564 capsnap->size = inode->i_size;
565 capsnap->mtime = inode->i_mtime;
566 capsnap->atime = inode->i_atime;
567 capsnap->ctime = inode->i_ctime;
568 capsnap->time_warp_seq = ci->i_time_warp_seq;
569 if (capsnap->dirty_pages) {
570 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
571 "still has %d dirty pages\n", inode, capsnap,
572 capsnap->context, capsnap->context->seq,
573 ceph_cap_string(capsnap->dirty), capsnap->size,
574 capsnap->dirty_pages);
577 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
578 inode, capsnap, capsnap->context,
579 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
582 spin_lock(&mdsc->snap_flush_lock);
583 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
584 spin_unlock(&mdsc->snap_flush_lock);
585 return 1; /* caller may want to ceph_flush_snaps */
589 * Queue cap_snaps for snap writeback for this realm and its children.
590 * Called under snap_rwsem, so realm topology won't change.
592 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
594 struct ceph_inode_info *ci;
595 struct inode *lastinode = NULL;
596 struct ceph_snap_realm *child;
598 dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
600 spin_lock(&realm->inodes_with_caps_lock);
601 list_for_each_entry(ci, &realm->inodes_with_caps,
603 struct inode *inode = igrab(&ci->vfs_inode);
606 spin_unlock(&realm->inodes_with_caps_lock);
609 ceph_queue_cap_snap(ci);
610 spin_lock(&realm->inodes_with_caps_lock);
612 spin_unlock(&realm->inodes_with_caps_lock);
615 list_for_each_entry(child, &realm->children, child_item) {
616 dout("queue_realm_cap_snaps %p %llx queue child %p %llx\n",
617 realm, realm->ino, child, child->ino);
618 list_del_init(&child->dirty_item);
619 list_add(&child->dirty_item, &realm->dirty_item);
622 list_del_init(&realm->dirty_item);
623 dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
627 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
628 * the snap realm parameters from a given realm and all of its ancestors,
631 * Caller must hold snap_rwsem for write.
633 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
634 void *p, void *e, bool deletion)
636 struct ceph_mds_snap_realm *ri; /* encoded */
637 __le64 *snaps; /* encoded */
638 __le64 *prior_parent_snaps; /* encoded */
639 struct ceph_snap_realm *realm;
642 LIST_HEAD(dirty_realms);
644 dout("update_snap_trace deletion=%d\n", deletion);
646 ceph_decode_need(&p, e, sizeof(*ri), bad);
649 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
650 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
652 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
653 prior_parent_snaps = p;
654 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
656 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
658 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
660 err = PTR_ERR(realm);
665 /* ensure the parent is correct */
666 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
671 if (le64_to_cpu(ri->seq) > realm->seq) {
672 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
673 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
674 /* update realm parameters, snap lists */
675 realm->seq = le64_to_cpu(ri->seq);
676 realm->created = le64_to_cpu(ri->created);
677 realm->parent_since = le64_to_cpu(ri->parent_since);
679 realm->num_snaps = le32_to_cpu(ri->num_snaps);
680 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
684 realm->num_prior_parent_snaps =
685 le32_to_cpu(ri->num_prior_parent_snaps);
686 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
687 realm->num_prior_parent_snaps);
691 /* queue realm for cap_snap creation */
692 list_add(&realm->dirty_item, &dirty_realms);
695 } else if (!realm->cached_context) {
696 dout("update_snap_trace %llx %p seq %lld new\n",
697 realm->ino, realm, realm->seq);
700 dout("update_snap_trace %llx %p seq %lld unchanged\n",
701 realm->ino, realm, realm->seq);
704 dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
705 realm, invalidate, p, e);
710 /* invalidate when we reach the _end_ (root) of the trace */
712 rebuild_snap_realms(realm);
715 * queue cap snaps _after_ we've built the new snap contexts,
716 * so that i_head_snapc can be set appropriately.
718 while (!list_empty(&dirty_realms)) {
719 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
721 queue_realm_cap_snaps(realm);
724 __cleanup_empty_realms(mdsc);
730 pr_err("update_snap_trace error %d\n", err);
736 * Send any cap_snaps that are queued for flush. Try to carry
737 * s_mutex across multiple snap flushes to avoid locking overhead.
739 * Caller holds no locks.
741 static void flush_snaps(struct ceph_mds_client *mdsc)
743 struct ceph_inode_info *ci;
745 struct ceph_mds_session *session = NULL;
747 dout("flush_snaps\n");
748 spin_lock(&mdsc->snap_flush_lock);
749 while (!list_empty(&mdsc->snap_flush_list)) {
750 ci = list_first_entry(&mdsc->snap_flush_list,
751 struct ceph_inode_info, i_snap_flush_item);
752 inode = &ci->vfs_inode;
754 spin_unlock(&mdsc->snap_flush_lock);
755 spin_lock(&ci->i_ceph_lock);
756 __ceph_flush_snaps(ci, &session, 0);
757 spin_unlock(&ci->i_ceph_lock);
759 spin_lock(&mdsc->snap_flush_lock);
761 spin_unlock(&mdsc->snap_flush_lock);
764 mutex_unlock(&session->s_mutex);
765 ceph_put_mds_session(session);
767 dout("flush_snaps done\n");
772 * Handle a snap notification from the MDS.
774 * This can take two basic forms: the simplest is just a snap creation
775 * or deletion notification on an existing realm. This should update the
776 * realm and its children.
778 * The more difficult case is realm creation, due to snap creation at a
779 * new point in the file hierarchy, or due to a rename that moves a file or
780 * directory into another realm.
782 void ceph_handle_snap(struct ceph_mds_client *mdsc,
783 struct ceph_mds_session *session,
784 struct ceph_msg *msg)
786 struct super_block *sb = mdsc->fsc->sb;
787 int mds = session->s_mds;
791 struct ceph_snap_realm *realm = NULL;
792 void *p = msg->front.iov_base;
793 void *e = p + msg->front.iov_len;
794 struct ceph_mds_snap_head *h;
795 int num_split_inos, num_split_realms;
796 __le64 *split_inos = NULL, *split_realms = NULL;
798 int locked_rwsem = 0;
801 if (msg->front.iov_len < sizeof(*h))
804 op = le32_to_cpu(h->op);
805 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
807 num_split_inos = le32_to_cpu(h->num_split_inos);
808 num_split_realms = le32_to_cpu(h->num_split_realms);
809 trace_len = le32_to_cpu(h->trace_len);
812 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
813 ceph_snap_op_name(op), split, trace_len);
815 mutex_lock(&session->s_mutex);
817 mutex_unlock(&session->s_mutex);
819 down_write(&mdsc->snap_rwsem);
822 if (op == CEPH_SNAP_OP_SPLIT) {
823 struct ceph_mds_snap_realm *ri;
826 * A "split" breaks part of an existing realm off into
827 * a new realm. The MDS provides a list of inodes
828 * (with caps) and child realms that belong to the new
832 p += sizeof(u64) * num_split_inos;
834 p += sizeof(u64) * num_split_realms;
835 ceph_decode_need(&p, e, sizeof(*ri), bad);
836 /* we will peek at realm info here, but will _not_
837 * advance p, as the realm update will occur below in
838 * ceph_update_snap_trace. */
841 realm = ceph_lookup_snap_realm(mdsc, split);
843 realm = ceph_create_snap_realm(mdsc, split);
847 ceph_get_snap_realm(mdsc, realm);
849 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
850 for (i = 0; i < num_split_inos; i++) {
851 struct ceph_vino vino = {
852 .ino = le64_to_cpu(split_inos[i]),
855 struct inode *inode = ceph_find_inode(sb, vino);
856 struct ceph_inode_info *ci;
857 struct ceph_snap_realm *oldrealm;
861 ci = ceph_inode(inode);
863 spin_lock(&ci->i_ceph_lock);
864 if (!ci->i_snap_realm)
867 * If this inode belongs to a realm that was
868 * created after our new realm, we experienced
869 * a race (due to another split notifications
870 * arriving from a different MDS). So skip
873 if (ci->i_snap_realm->created >
874 le64_to_cpu(ri->created)) {
875 dout(" leaving %p in newer realm %llx %p\n",
876 inode, ci->i_snap_realm->ino,
880 dout(" will move %p to split realm %llx %p\n",
881 inode, realm->ino, realm);
883 * Move the inode to the new realm
885 spin_lock(&realm->inodes_with_caps_lock);
886 list_del_init(&ci->i_snap_realm_item);
887 list_add(&ci->i_snap_realm_item,
888 &realm->inodes_with_caps);
889 oldrealm = ci->i_snap_realm;
890 ci->i_snap_realm = realm;
891 spin_unlock(&realm->inodes_with_caps_lock);
892 spin_unlock(&ci->i_ceph_lock);
894 ceph_get_snap_realm(mdsc, realm);
895 ceph_put_snap_realm(mdsc, oldrealm);
901 spin_unlock(&ci->i_ceph_lock);
905 /* we may have taken some of the old realm's children. */
906 for (i = 0; i < num_split_realms; i++) {
907 struct ceph_snap_realm *child =
908 ceph_lookup_snap_realm(mdsc,
909 le64_to_cpu(split_realms[i]));
912 adjust_snap_realm_parent(mdsc, child, realm->ino);
917 * update using the provided snap trace. if we are deleting a
918 * snap, we can avoid queueing cap_snaps.
920 ceph_update_snap_trace(mdsc, p, e,
921 op == CEPH_SNAP_OP_DESTROY);
923 if (op == CEPH_SNAP_OP_SPLIT)
924 /* we took a reference when we created the realm, above */
925 ceph_put_snap_realm(mdsc, realm);
927 __cleanup_empty_realms(mdsc);
929 up_write(&mdsc->snap_rwsem);
935 pr_err("corrupt snap message from mds%d\n", mds);
939 up_write(&mdsc->snap_rwsem);
943 int __init ceph_snap_init(void)
945 empty_snapc = ceph_create_snap_context(0, GFP_NOFS);
948 empty_snapc->seq = 1;
952 void ceph_snap_exit(void)
954 ceph_put_snap_context(empty_snapc);