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_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);
80 static void __insert_snap_realm(struct rb_root *root,
81 struct ceph_snap_realm *new)
83 struct rb_node **p = &root->rb_node;
84 struct rb_node *parent = NULL;
85 struct ceph_snap_realm *r = NULL;
89 r = rb_entry(parent, struct ceph_snap_realm, node);
90 if (new->ino < r->ino)
92 else if (new->ino > r->ino)
98 rb_link_node(&new->node, parent, p);
99 rb_insert_color(&new->node, root);
103 * create and get the realm rooted at @ino and bump its ref count.
105 * caller must hold snap_rwsem for write.
107 static struct ceph_snap_realm *ceph_create_snap_realm(
108 struct ceph_mds_client *mdsc,
111 struct ceph_snap_realm *realm;
113 realm = kzalloc(sizeof(*realm), GFP_NOFS);
115 return ERR_PTR(-ENOMEM);
117 atomic_set(&realm->nref, 1); /* for caller */
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);
131 * lookup the realm rooted at @ino.
133 * caller must hold snap_rwsem for write.
135 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
138 struct rb_node *n = mdsc->snap_realms.rb_node;
139 struct ceph_snap_realm *r;
142 r = rb_entry(n, struct ceph_snap_realm, node);
145 else if (ino > r->ino)
148 dout("lookup_snap_realm %llx %p\n", r->ino, r);
155 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
158 struct ceph_snap_realm *r;
159 r = __lookup_snap_realm(mdsc, ino);
161 ceph_get_snap_realm(mdsc, r);
165 static void __put_snap_realm(struct ceph_mds_client *mdsc,
166 struct ceph_snap_realm *realm);
169 * called with snap_rwsem (write)
171 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
172 struct ceph_snap_realm *realm)
174 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
176 rb_erase(&realm->node, &mdsc->snap_realms);
179 list_del_init(&realm->child_item);
180 __put_snap_realm(mdsc, realm->parent);
183 kfree(realm->prior_parent_snaps);
185 ceph_put_snap_context(realm->cached_context);
190 * caller holds snap_rwsem (write)
192 static void __put_snap_realm(struct ceph_mds_client *mdsc,
193 struct ceph_snap_realm *realm)
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);
202 * caller needn't hold any locks
204 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
205 struct ceph_snap_realm *realm)
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))
212 if (down_write_trylock(&mdsc->snap_rwsem)) {
213 __destroy_snap_realm(mdsc, realm);
214 up_write(&mdsc->snap_rwsem);
216 spin_lock(&mdsc->snap_empty_lock);
217 list_add(&realm->empty_item, &mdsc->snap_empty);
218 spin_unlock(&mdsc->snap_empty_lock);
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
227 * Called under snap_rwsem (write)
229 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
231 struct ceph_snap_realm *realm;
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);
242 spin_unlock(&mdsc->snap_empty_lock);
245 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
247 down_write(&mdsc->snap_rwsem);
248 __cleanup_empty_realms(mdsc);
249 up_write(&mdsc->snap_rwsem);
253 * adjust the parent realm of a given @realm. adjust child list, and parent
254 * pointers, and ref counts appropriately.
256 * return true if parent was changed, 0 if unchanged, <0 on error.
258 * caller must hold snap_rwsem for write.
260 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
261 struct ceph_snap_realm *realm,
264 struct ceph_snap_realm *parent;
266 if (realm->parent_ino == parentino)
269 parent = ceph_lookup_snap_realm(mdsc, parentino);
271 parent = ceph_create_snap_realm(mdsc, parentino);
273 return PTR_ERR(parent);
275 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
276 realm->ino, realm, realm->parent_ino, realm->parent,
279 list_del_init(&realm->child_item);
280 ceph_put_snap_realm(mdsc, realm->parent);
282 realm->parent_ino = parentino;
283 realm->parent = parent;
284 list_add(&realm->child_item, &parent->children);
289 static int cmpu64_rev(const void *a, const void *b)
291 if (*(u64 *)a < *(u64 *)b)
293 if (*(u64 *)a > *(u64 *)b)
300 * build the snap context for a given realm.
302 static int build_snap_context(struct ceph_snap_realm *realm,
303 struct list_head* dirty_realms)
305 struct ceph_snap_realm *parent = realm->parent;
306 struct ceph_snap_context *snapc;
308 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
311 * build parent context, if it hasn't been built.
312 * conservatively estimate that all parent snaps might be
316 if (!parent->cached_context) {
317 err = build_snap_context(parent, dirty_realms);
321 num += parent->cached_context->num_snaps;
324 /* do i actually need to update? not if my context seq
325 matches realm seq, and my parents' does to. (this works
326 because we rebuild_snap_realms() works _downward_ in
327 hierarchy after each update.) */
328 if (realm->cached_context &&
329 realm->cached_context->seq == realm->seq &&
331 realm->cached_context->seq >= parent->cached_context->seq)) {
332 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
334 realm->ino, realm, realm->cached_context,
335 realm->cached_context->seq,
336 (unsigned int)realm->cached_context->num_snaps);
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);
377 ceph_put_snap_context(realm->cached_context);
378 realm->cached_context = snapc;
379 /* queue realm for cap_snap creation */
380 list_add_tail(&realm->dirty_item, dirty_realms);
385 * if we fail, clear old (incorrect) cached_context... hopefully
386 * we'll have better luck building it later
388 if (realm->cached_context) {
389 ceph_put_snap_context(realm->cached_context);
390 realm->cached_context = NULL;
392 pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
398 * rebuild snap context for the given realm and all of its children.
400 static void rebuild_snap_realms(struct ceph_snap_realm *realm,
401 struct list_head *dirty_realms)
403 struct ceph_snap_realm *child;
405 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
406 build_snap_context(realm, dirty_realms);
408 list_for_each_entry(child, &realm->children, child_item)
409 rebuild_snap_realms(child, dirty_realms);
414 * helper to allocate and decode an array of snapids. free prior
417 static int dup_array(u64 **dst, __le64 *src, u32 num)
423 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
426 for (i = 0; i < num; i++)
427 (*dst)[i] = get_unaligned_le64(src + i);
434 static bool has_new_snaps(struct ceph_snap_context *o,
435 struct ceph_snap_context *n)
437 if (n->num_snaps == 0)
439 /* snaps are in descending order */
440 return n->snaps[0] > o->seq;
444 * When a snapshot is applied, the size/mtime inode metadata is queued
445 * in a ceph_cap_snap (one for each snapshot) until writeback
446 * completes and the metadata can be flushed back to the MDS.
448 * However, if a (sync) write is currently in-progress when we apply
449 * the snapshot, we have to wait until the write succeeds or fails
450 * (and a final size/mtime is known). In this case the
451 * cap_snap->writing = 1, and is said to be "pending." When the write
452 * finishes, we __ceph_finish_cap_snap().
454 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
457 void ceph_queue_cap_snap(struct ceph_inode_info *ci)
459 struct inode *inode = &ci->vfs_inode;
460 struct ceph_cap_snap *capsnap;
461 struct ceph_snap_context *old_snapc, *new_snapc;
464 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
466 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
470 spin_lock(&ci->i_ceph_lock);
471 used = __ceph_caps_used(ci);
472 dirty = __ceph_caps_dirty(ci);
474 old_snapc = ci->i_head_snapc;
475 new_snapc = ci->i_snap_realm->cached_context;
478 * If there is a write in progress, treat that as a dirty Fw,
479 * even though it hasn't completed yet; by the time we finish
480 * up this capsnap it will be.
482 if (used & CEPH_CAP_FILE_WR)
483 dirty |= CEPH_CAP_FILE_WR;
485 if (__ceph_have_pending_cap_snap(ci)) {
486 /* there is no point in queuing multiple "pending" cap_snaps,
487 as no new writes are allowed to start when pending, so any
488 writes in progress now were started before the previous
489 cap_snap. lucky us. */
490 dout("queue_cap_snap %p already pending\n", inode);
493 if (ci->i_wrbuffer_ref_head == 0 &&
494 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
495 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
502 * There is no need to send FLUSHSNAP message to MDS if there is
503 * no new snapshot. But when there is dirty pages or on-going
504 * writes, we still need to create cap_snap. cap_snap is needed
505 * by the write path and page writeback path.
507 * also see ceph_try_drop_cap_snap()
509 if (has_new_snaps(old_snapc, new_snapc)) {
510 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
511 capsnap->need_flush = true;
513 if (!(used & CEPH_CAP_FILE_WR) &&
514 ci->i_wrbuffer_ref_head == 0) {
515 dout("queue_cap_snap %p "
516 "no new_snap|dirty_page|writing\n", inode);
521 dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n",
522 inode, capsnap, old_snapc, ceph_cap_string(dirty),
523 capsnap->need_flush ? "" : "no_flush");
526 refcount_set(&capsnap->nref, 1);
527 INIT_LIST_HEAD(&capsnap->ci_item);
529 capsnap->follows = old_snapc->seq;
530 capsnap->issued = __ceph_caps_issued(ci, NULL);
531 capsnap->dirty = dirty;
533 capsnap->mode = inode->i_mode;
534 capsnap->uid = inode->i_uid;
535 capsnap->gid = inode->i_gid;
537 if (dirty & CEPH_CAP_XATTR_EXCL) {
538 __ceph_build_xattrs_blob(ci);
539 capsnap->xattr_blob =
540 ceph_buffer_get(ci->i_xattrs.blob);
541 capsnap->xattr_version = ci->i_xattrs.version;
543 capsnap->xattr_blob = NULL;
544 capsnap->xattr_version = 0;
547 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
549 /* dirty page count moved from _head to this cap_snap;
550 all subsequent writes page dirties occur _after_ this
552 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
553 ci->i_wrbuffer_ref_head = 0;
554 capsnap->context = old_snapc;
555 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
557 if (used & CEPH_CAP_FILE_WR) {
558 dout("queue_cap_snap %p cap_snap %p snapc %p"
559 " seq %llu used WR, now pending\n", inode,
560 capsnap, old_snapc, old_snapc->seq);
561 capsnap->writing = 1;
563 /* note mtime, size NOW. */
564 __ceph_finish_cap_snap(ci, capsnap);
570 if (ci->i_head_snapc) {
571 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
572 dout(" new snapc is %p\n", new_snapc);
574 spin_unlock(&ci->i_ceph_lock);
577 ceph_put_snap_context(old_snapc);
581 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
582 * to be used for the snapshot, to be flushed back to the mds.
584 * If capsnap can now be flushed, add to snap_flush list, and return 1.
586 * Caller must hold i_ceph_lock.
588 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
589 struct ceph_cap_snap *capsnap)
591 struct inode *inode = &ci->vfs_inode;
592 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
594 BUG_ON(capsnap->writing);
595 capsnap->size = inode->i_size;
596 capsnap->mtime = inode->i_mtime;
597 capsnap->atime = inode->i_atime;
598 capsnap->ctime = inode->i_ctime;
599 capsnap->time_warp_seq = ci->i_time_warp_seq;
600 capsnap->truncate_size = ci->i_truncate_size;
601 capsnap->truncate_seq = ci->i_truncate_seq;
602 if (capsnap->dirty_pages) {
603 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
604 "still has %d dirty pages\n", inode, capsnap,
605 capsnap->context, capsnap->context->seq,
606 ceph_cap_string(capsnap->dirty), capsnap->size,
607 capsnap->dirty_pages);
611 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
612 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
613 inode, capsnap, capsnap->context,
614 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
617 spin_lock(&mdsc->snap_flush_lock);
618 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
619 spin_unlock(&mdsc->snap_flush_lock);
620 return 1; /* caller may want to ceph_flush_snaps */
624 * Queue cap_snaps for snap writeback for this realm and its children.
625 * Called under snap_rwsem, so realm topology won't change.
627 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
629 struct ceph_inode_info *ci;
630 struct inode *lastinode = NULL;
632 dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
634 spin_lock(&realm->inodes_with_caps_lock);
635 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
636 struct inode *inode = igrab(&ci->vfs_inode);
639 spin_unlock(&realm->inodes_with_caps_lock);
642 ceph_queue_cap_snap(ci);
643 spin_lock(&realm->inodes_with_caps_lock);
645 spin_unlock(&realm->inodes_with_caps_lock);
648 dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
652 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
653 * the snap realm parameters from a given realm and all of its ancestors,
656 * Caller must hold snap_rwsem for write.
658 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
659 void *p, void *e, bool deletion,
660 struct ceph_snap_realm **realm_ret)
662 struct ceph_mds_snap_realm *ri; /* encoded */
663 __le64 *snaps; /* encoded */
664 __le64 *prior_parent_snaps; /* encoded */
665 struct ceph_snap_realm *realm = NULL;
666 struct ceph_snap_realm *first_realm = NULL;
669 LIST_HEAD(dirty_realms);
671 dout("update_snap_trace deletion=%d\n", deletion);
673 ceph_decode_need(&p, e, sizeof(*ri), bad);
676 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
677 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
679 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
680 prior_parent_snaps = p;
681 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
683 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
685 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
687 err = PTR_ERR(realm);
692 /* ensure the parent is correct */
693 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
698 if (le64_to_cpu(ri->seq) > realm->seq) {
699 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
700 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
701 /* update realm parameters, snap lists */
702 realm->seq = le64_to_cpu(ri->seq);
703 realm->created = le64_to_cpu(ri->created);
704 realm->parent_since = le64_to_cpu(ri->parent_since);
706 realm->num_snaps = le32_to_cpu(ri->num_snaps);
707 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
711 realm->num_prior_parent_snaps =
712 le32_to_cpu(ri->num_prior_parent_snaps);
713 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
714 realm->num_prior_parent_snaps);
718 if (realm->seq > mdsc->last_snap_seq)
719 mdsc->last_snap_seq = realm->seq;
722 } else if (!realm->cached_context) {
723 dout("update_snap_trace %llx %p seq %lld new\n",
724 realm->ino, realm, realm->seq);
727 dout("update_snap_trace %llx %p seq %lld unchanged\n",
728 realm->ino, realm, realm->seq);
731 dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
732 realm, invalidate, p, e);
734 /* invalidate when we reach the _end_ (root) of the trace */
735 if (invalidate && p >= e)
736 rebuild_snap_realms(realm, &dirty_realms);
741 ceph_put_snap_realm(mdsc, realm);
747 * queue cap snaps _after_ we've built the new snap contexts,
748 * so that i_head_snapc can be set appropriately.
750 while (!list_empty(&dirty_realms)) {
751 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
753 list_del_init(&realm->dirty_item);
754 queue_realm_cap_snaps(realm);
758 *realm_ret = first_realm;
760 ceph_put_snap_realm(mdsc, first_realm);
762 __cleanup_empty_realms(mdsc);
768 if (realm && !IS_ERR(realm))
769 ceph_put_snap_realm(mdsc, realm);
771 ceph_put_snap_realm(mdsc, first_realm);
772 pr_err("update_snap_trace error %d\n", err);
778 * Send any cap_snaps that are queued for flush. Try to carry
779 * s_mutex across multiple snap flushes to avoid locking overhead.
781 * Caller holds no locks.
783 static void flush_snaps(struct ceph_mds_client *mdsc)
785 struct ceph_inode_info *ci;
787 struct ceph_mds_session *session = NULL;
789 dout("flush_snaps\n");
790 spin_lock(&mdsc->snap_flush_lock);
791 while (!list_empty(&mdsc->snap_flush_list)) {
792 ci = list_first_entry(&mdsc->snap_flush_list,
793 struct ceph_inode_info, i_snap_flush_item);
794 inode = &ci->vfs_inode;
796 spin_unlock(&mdsc->snap_flush_lock);
797 ceph_flush_snaps(ci, &session);
799 spin_lock(&mdsc->snap_flush_lock);
801 spin_unlock(&mdsc->snap_flush_lock);
804 mutex_unlock(&session->s_mutex);
805 ceph_put_mds_session(session);
807 dout("flush_snaps done\n");
812 * Handle a snap notification from the MDS.
814 * This can take two basic forms: the simplest is just a snap creation
815 * or deletion notification on an existing realm. This should update the
816 * realm and its children.
818 * The more difficult case is realm creation, due to snap creation at a
819 * new point in the file hierarchy, or due to a rename that moves a file or
820 * directory into another realm.
822 void ceph_handle_snap(struct ceph_mds_client *mdsc,
823 struct ceph_mds_session *session,
824 struct ceph_msg *msg)
826 struct super_block *sb = mdsc->fsc->sb;
827 int mds = session->s_mds;
831 struct ceph_snap_realm *realm = NULL;
832 void *p = msg->front.iov_base;
833 void *e = p + msg->front.iov_len;
834 struct ceph_mds_snap_head *h;
835 int num_split_inos, num_split_realms;
836 __le64 *split_inos = NULL, *split_realms = NULL;
838 int locked_rwsem = 0;
841 if (msg->front.iov_len < sizeof(*h))
844 op = le32_to_cpu(h->op);
845 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
847 num_split_inos = le32_to_cpu(h->num_split_inos);
848 num_split_realms = le32_to_cpu(h->num_split_realms);
849 trace_len = le32_to_cpu(h->trace_len);
852 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
853 ceph_snap_op_name(op), split, trace_len);
855 mutex_lock(&session->s_mutex);
857 mutex_unlock(&session->s_mutex);
859 down_write(&mdsc->snap_rwsem);
862 if (op == CEPH_SNAP_OP_SPLIT) {
863 struct ceph_mds_snap_realm *ri;
866 * A "split" breaks part of an existing realm off into
867 * a new realm. The MDS provides a list of inodes
868 * (with caps) and child realms that belong to the new
872 p += sizeof(u64) * num_split_inos;
874 p += sizeof(u64) * num_split_realms;
875 ceph_decode_need(&p, e, sizeof(*ri), bad);
876 /* we will peek at realm info here, but will _not_
877 * advance p, as the realm update will occur below in
878 * ceph_update_snap_trace. */
881 realm = ceph_lookup_snap_realm(mdsc, split);
883 realm = ceph_create_snap_realm(mdsc, split);
888 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
889 for (i = 0; i < num_split_inos; i++) {
890 struct ceph_vino vino = {
891 .ino = le64_to_cpu(split_inos[i]),
894 struct inode *inode = ceph_find_inode(sb, vino);
895 struct ceph_inode_info *ci;
896 struct ceph_snap_realm *oldrealm;
900 ci = ceph_inode(inode);
902 spin_lock(&ci->i_ceph_lock);
903 if (!ci->i_snap_realm)
906 * If this inode belongs to a realm that was
907 * created after our new realm, we experienced
908 * a race (due to another split notifications
909 * arriving from a different MDS). So skip
912 if (ci->i_snap_realm->created >
913 le64_to_cpu(ri->created)) {
914 dout(" leaving %p in newer realm %llx %p\n",
915 inode, ci->i_snap_realm->ino,
919 dout(" will move %p to split realm %llx %p\n",
920 inode, realm->ino, realm);
922 * Move the inode to the new realm
924 spin_lock(&realm->inodes_with_caps_lock);
925 list_del_init(&ci->i_snap_realm_item);
926 list_add(&ci->i_snap_realm_item,
927 &realm->inodes_with_caps);
928 oldrealm = ci->i_snap_realm;
929 ci->i_snap_realm = realm;
930 spin_unlock(&realm->inodes_with_caps_lock);
931 spin_unlock(&ci->i_ceph_lock);
933 ceph_get_snap_realm(mdsc, realm);
934 ceph_put_snap_realm(mdsc, oldrealm);
940 spin_unlock(&ci->i_ceph_lock);
944 /* we may have taken some of the old realm's children. */
945 for (i = 0; i < num_split_realms; i++) {
946 struct ceph_snap_realm *child =
947 __lookup_snap_realm(mdsc,
948 le64_to_cpu(split_realms[i]));
951 adjust_snap_realm_parent(mdsc, child, realm->ino);
956 * update using the provided snap trace. if we are deleting a
957 * snap, we can avoid queueing cap_snaps.
959 ceph_update_snap_trace(mdsc, p, e,
960 op == CEPH_SNAP_OP_DESTROY, NULL);
962 if (op == CEPH_SNAP_OP_SPLIT)
963 /* we took a reference when we created the realm, above */
964 ceph_put_snap_realm(mdsc, realm);
966 __cleanup_empty_realms(mdsc);
968 up_write(&mdsc->snap_rwsem);
974 pr_err("corrupt snap message from mds%d\n", mds);
978 up_write(&mdsc->snap_rwsem);