1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
4 #include <linux/sort.h>
5 #include <linux/slab.h>
6 #include <linux/iversion.h>
8 #include "mds_client.h"
9 #include <linux/ceph/decode.h>
11 /* unused map expires after 5 minutes */
12 #define CEPH_SNAPID_MAP_TIMEOUT (5 * 60 * HZ)
15 * Snapshots in ceph are driven in large part by cooperation from the
16 * client. In contrast to local file systems or file servers that
17 * implement snapshots at a single point in the system, ceph's
18 * distributed access to storage requires clients to help decide
19 * whether a write logically occurs before or after a recently created
22 * This provides a perfect instantanous client-wide snapshot. Between
23 * clients, however, snapshots may appear to be applied at slightly
24 * different points in time, depending on delays in delivering the
25 * snapshot notification.
27 * Snapshots are _not_ file system-wide. Instead, each snapshot
28 * applies to the subdirectory nested beneath some directory. This
29 * effectively divides the hierarchy into multiple "realms," where all
30 * of the files contained by each realm share the same set of
31 * snapshots. An individual realm's snap set contains snapshots
32 * explicitly created on that realm, as well as any snaps in its
33 * parent's snap set _after_ the point at which the parent became it's
34 * parent (due to, say, a rename). Similarly, snaps from prior parents
35 * during the time intervals during which they were the parent are included.
37 * The client is spared most of this detail, fortunately... it must only
38 * maintains a hierarchy of realms reflecting the current parent/child
39 * realm relationship, and for each realm has an explicit list of snaps
40 * inherited from prior parents.
42 * A snap_realm struct is maintained for realms containing every inode
43 * with an open cap in the system. (The needed snap realm information is
44 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq'
45 * version number is used to ensure that as realm parameters change (new
46 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
48 * The realm hierarchy drives the generation of a 'snap context' for each
49 * realm, which simply lists the resulting set of snaps for the realm. This
50 * is attached to any writes sent to OSDs.
53 * Unfortunately error handling is a bit mixed here. If we get a snap
54 * update, but don't have enough memory to update our realm hierarchy,
55 * it's not clear what we can do about it (besides complaining to the
61 * increase ref count for the realm
63 * caller must hold snap_rwsem.
65 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
66 struct ceph_snap_realm *realm)
68 lockdep_assert_held(&mdsc->snap_rwsem);
71 * The 0->1 and 1->0 transitions must take the snap_empty_lock
72 * atomically with the refcount change. Go ahead and bump the
73 * nref here, unless it's 0, in which case we take the spinlock
74 * and then do the increment and remove it from the list.
76 if (atomic_inc_not_zero(&realm->nref))
79 spin_lock(&mdsc->snap_empty_lock);
80 if (atomic_inc_return(&realm->nref) == 1)
81 list_del_init(&realm->empty_item);
82 spin_unlock(&mdsc->snap_empty_lock);
85 static void __insert_snap_realm(struct rb_root *root,
86 struct ceph_snap_realm *new)
88 struct rb_node **p = &root->rb_node;
89 struct rb_node *parent = NULL;
90 struct ceph_snap_realm *r = NULL;
94 r = rb_entry(parent, struct ceph_snap_realm, node);
95 if (new->ino < r->ino)
97 else if (new->ino > r->ino)
103 rb_link_node(&new->node, parent, p);
104 rb_insert_color(&new->node, root);
108 * create and get the realm rooted at @ino and bump its ref count.
110 * caller must hold snap_rwsem for write.
112 static struct ceph_snap_realm *ceph_create_snap_realm(
113 struct ceph_mds_client *mdsc,
116 struct ceph_snap_realm *realm;
118 lockdep_assert_held_write(&mdsc->snap_rwsem);
120 realm = kzalloc(sizeof(*realm), GFP_NOFS);
122 return ERR_PTR(-ENOMEM);
124 atomic_set(&realm->nref, 1); /* for caller */
126 INIT_LIST_HEAD(&realm->children);
127 INIT_LIST_HEAD(&realm->child_item);
128 INIT_LIST_HEAD(&realm->empty_item);
129 INIT_LIST_HEAD(&realm->dirty_item);
130 INIT_LIST_HEAD(&realm->inodes_with_caps);
131 spin_lock_init(&realm->inodes_with_caps_lock);
132 __insert_snap_realm(&mdsc->snap_realms, realm);
133 mdsc->num_snap_realms++;
135 dout("create_snap_realm %llx %p\n", realm->ino, realm);
140 * lookup the realm rooted at @ino.
142 * caller must hold snap_rwsem.
144 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
147 struct rb_node *n = mdsc->snap_realms.rb_node;
148 struct ceph_snap_realm *r;
150 lockdep_assert_held(&mdsc->snap_rwsem);
153 r = rb_entry(n, struct ceph_snap_realm, node);
156 else if (ino > r->ino)
159 dout("lookup_snap_realm %llx %p\n", r->ino, r);
166 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
169 struct ceph_snap_realm *r;
170 r = __lookup_snap_realm(mdsc, ino);
172 ceph_get_snap_realm(mdsc, r);
176 static void __put_snap_realm(struct ceph_mds_client *mdsc,
177 struct ceph_snap_realm *realm);
180 * called with snap_rwsem (write)
182 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
183 struct ceph_snap_realm *realm)
185 lockdep_assert_held_write(&mdsc->snap_rwsem);
187 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
189 rb_erase(&realm->node, &mdsc->snap_realms);
190 mdsc->num_snap_realms--;
193 list_del_init(&realm->child_item);
194 __put_snap_realm(mdsc, realm->parent);
197 kfree(realm->prior_parent_snaps);
199 ceph_put_snap_context(realm->cached_context);
204 * caller holds snap_rwsem (write)
206 static void __put_snap_realm(struct ceph_mds_client *mdsc,
207 struct ceph_snap_realm *realm)
209 lockdep_assert_held_write(&mdsc->snap_rwsem);
212 * We do not require the snap_empty_lock here, as any caller that
213 * increments the value must hold the snap_rwsem.
215 if (atomic_dec_and_test(&realm->nref))
216 __destroy_snap_realm(mdsc, realm);
220 * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
222 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
223 struct ceph_snap_realm *realm)
225 if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
228 if (down_write_trylock(&mdsc->snap_rwsem)) {
229 spin_unlock(&mdsc->snap_empty_lock);
230 __destroy_snap_realm(mdsc, realm);
231 up_write(&mdsc->snap_rwsem);
233 list_add(&realm->empty_item, &mdsc->snap_empty);
234 spin_unlock(&mdsc->snap_empty_lock);
239 * Clean up any realms whose ref counts have dropped to zero. Note
240 * that this does not include realms who were created but not yet
243 * Called under snap_rwsem (write)
245 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
247 struct ceph_snap_realm *realm;
249 lockdep_assert_held_write(&mdsc->snap_rwsem);
251 spin_lock(&mdsc->snap_empty_lock);
252 while (!list_empty(&mdsc->snap_empty)) {
253 realm = list_first_entry(&mdsc->snap_empty,
254 struct ceph_snap_realm, empty_item);
255 list_del(&realm->empty_item);
256 spin_unlock(&mdsc->snap_empty_lock);
257 __destroy_snap_realm(mdsc, realm);
258 spin_lock(&mdsc->snap_empty_lock);
260 spin_unlock(&mdsc->snap_empty_lock);
263 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
265 down_write(&mdsc->snap_rwsem);
266 __cleanup_empty_realms(mdsc);
267 up_write(&mdsc->snap_rwsem);
271 * adjust the parent realm of a given @realm. adjust child list, and parent
272 * pointers, and ref counts appropriately.
274 * return true if parent was changed, 0 if unchanged, <0 on error.
276 * caller must hold snap_rwsem for write.
278 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
279 struct ceph_snap_realm *realm,
282 struct ceph_snap_realm *parent;
284 lockdep_assert_held_write(&mdsc->snap_rwsem);
286 if (realm->parent_ino == parentino)
289 parent = ceph_lookup_snap_realm(mdsc, parentino);
291 parent = ceph_create_snap_realm(mdsc, parentino);
293 return PTR_ERR(parent);
295 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
296 realm->ino, realm, realm->parent_ino, realm->parent,
299 list_del_init(&realm->child_item);
300 ceph_put_snap_realm(mdsc, realm->parent);
302 realm->parent_ino = parentino;
303 realm->parent = parent;
304 list_add(&realm->child_item, &parent->children);
309 static int cmpu64_rev(const void *a, const void *b)
311 if (*(u64 *)a < *(u64 *)b)
313 if (*(u64 *)a > *(u64 *)b)
320 * build the snap context for a given realm.
322 static int build_snap_context(struct ceph_snap_realm *realm,
323 struct list_head* dirty_realms)
325 struct ceph_snap_realm *parent = realm->parent;
326 struct ceph_snap_context *snapc;
328 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
331 * build parent context, if it hasn't been built.
332 * conservatively estimate that all parent snaps might be
336 if (!parent->cached_context) {
337 err = build_snap_context(parent, dirty_realms);
341 num += parent->cached_context->num_snaps;
344 /* do i actually need to update? not if my context seq
345 matches realm seq, and my parents' does to. (this works
346 because we rebuild_snap_realms() works _downward_ in
347 hierarchy after each update.) */
348 if (realm->cached_context &&
349 realm->cached_context->seq == realm->seq &&
351 realm->cached_context->seq >= parent->cached_context->seq)) {
352 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
354 realm->ino, realm, realm->cached_context,
355 realm->cached_context->seq,
356 (unsigned int)realm->cached_context->num_snaps);
360 /* alloc new snap context */
362 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
364 snapc = ceph_create_snap_context(num, GFP_NOFS);
368 /* build (reverse sorted) snap vector */
370 snapc->seq = realm->seq;
374 /* include any of parent's snaps occurring _after_ my
375 parent became my parent */
376 for (i = 0; i < parent->cached_context->num_snaps; i++)
377 if (parent->cached_context->snaps[i] >=
379 snapc->snaps[num++] =
380 parent->cached_context->snaps[i];
381 if (parent->cached_context->seq > snapc->seq)
382 snapc->seq = parent->cached_context->seq;
384 memcpy(snapc->snaps + num, realm->snaps,
385 sizeof(u64)*realm->num_snaps);
386 num += realm->num_snaps;
387 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
388 sizeof(u64)*realm->num_prior_parent_snaps);
389 num += realm->num_prior_parent_snaps;
391 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
392 snapc->num_snaps = num;
393 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
394 realm->ino, realm, snapc, snapc->seq,
395 (unsigned int) snapc->num_snaps);
397 ceph_put_snap_context(realm->cached_context);
398 realm->cached_context = snapc;
399 /* queue realm for cap_snap creation */
400 list_add_tail(&realm->dirty_item, dirty_realms);
405 * if we fail, clear old (incorrect) cached_context... hopefully
406 * we'll have better luck building it later
408 if (realm->cached_context) {
409 ceph_put_snap_context(realm->cached_context);
410 realm->cached_context = NULL;
412 pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
418 * rebuild snap context for the given realm and all of its children.
420 static void rebuild_snap_realms(struct ceph_snap_realm *realm,
421 struct list_head *dirty_realms)
423 struct ceph_snap_realm *child;
425 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
426 build_snap_context(realm, dirty_realms);
428 list_for_each_entry(child, &realm->children, child_item)
429 rebuild_snap_realms(child, dirty_realms);
434 * helper to allocate and decode an array of snapids. free prior
437 static int dup_array(u64 **dst, __le64 *src, u32 num)
443 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
446 for (i = 0; i < num; i++)
447 (*dst)[i] = get_unaligned_le64(src + i);
454 static bool has_new_snaps(struct ceph_snap_context *o,
455 struct ceph_snap_context *n)
457 if (n->num_snaps == 0)
459 /* snaps are in descending order */
460 return n->snaps[0] > o->seq;
464 * When a snapshot is applied, the size/mtime inode metadata is queued
465 * in a ceph_cap_snap (one for each snapshot) until writeback
466 * completes and the metadata can be flushed back to the MDS.
468 * However, if a (sync) write is currently in-progress when we apply
469 * the snapshot, we have to wait until the write succeeds or fails
470 * (and a final size/mtime is known). In this case the
471 * cap_snap->writing = 1, and is said to be "pending." When the write
472 * finishes, we __ceph_finish_cap_snap().
474 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
477 void ceph_queue_cap_snap(struct ceph_inode_info *ci)
479 struct inode *inode = &ci->vfs_inode;
480 struct ceph_cap_snap *capsnap;
481 struct ceph_snap_context *old_snapc, *new_snapc;
482 struct ceph_buffer *old_blob = NULL;
485 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
487 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
491 spin_lock(&ci->i_ceph_lock);
492 used = __ceph_caps_used(ci);
493 dirty = __ceph_caps_dirty(ci);
495 old_snapc = ci->i_head_snapc;
496 new_snapc = ci->i_snap_realm->cached_context;
499 * If there is a write in progress, treat that as a dirty Fw,
500 * even though it hasn't completed yet; by the time we finish
501 * up this capsnap it will be.
503 if (used & CEPH_CAP_FILE_WR)
504 dirty |= CEPH_CAP_FILE_WR;
506 if (__ceph_have_pending_cap_snap(ci)) {
507 /* there is no point in queuing multiple "pending" cap_snaps,
508 as no new writes are allowed to start when pending, so any
509 writes in progress now were started before the previous
510 cap_snap. lucky us. */
511 dout("queue_cap_snap %p already pending\n", inode);
514 if (ci->i_wrbuffer_ref_head == 0 &&
515 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
516 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
523 * There is no need to send FLUSHSNAP message to MDS if there is
524 * no new snapshot. But when there is dirty pages or on-going
525 * writes, we still need to create cap_snap. cap_snap is needed
526 * by the write path and page writeback path.
528 * also see ceph_try_drop_cap_snap()
530 if (has_new_snaps(old_snapc, new_snapc)) {
531 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
532 capsnap->need_flush = true;
534 if (!(used & CEPH_CAP_FILE_WR) &&
535 ci->i_wrbuffer_ref_head == 0) {
536 dout("queue_cap_snap %p "
537 "no new_snap|dirty_page|writing\n", inode);
542 dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n",
543 inode, capsnap, old_snapc, ceph_cap_string(dirty),
544 capsnap->need_flush ? "" : "no_flush");
547 refcount_set(&capsnap->nref, 1);
548 INIT_LIST_HEAD(&capsnap->ci_item);
550 capsnap->follows = old_snapc->seq;
551 capsnap->issued = __ceph_caps_issued(ci, NULL);
552 capsnap->dirty = dirty;
554 capsnap->mode = inode->i_mode;
555 capsnap->uid = inode->i_uid;
556 capsnap->gid = inode->i_gid;
558 if (dirty & CEPH_CAP_XATTR_EXCL) {
559 old_blob = __ceph_build_xattrs_blob(ci);
560 capsnap->xattr_blob =
561 ceph_buffer_get(ci->i_xattrs.blob);
562 capsnap->xattr_version = ci->i_xattrs.version;
564 capsnap->xattr_blob = NULL;
565 capsnap->xattr_version = 0;
568 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
570 /* dirty page count moved from _head to this cap_snap;
571 all subsequent writes page dirties occur _after_ this
573 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
574 ci->i_wrbuffer_ref_head = 0;
575 capsnap->context = old_snapc;
576 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
578 if (used & CEPH_CAP_FILE_WR) {
579 dout("queue_cap_snap %p cap_snap %p snapc %p"
580 " seq %llu used WR, now pending\n", inode,
581 capsnap, old_snapc, old_snapc->seq);
582 capsnap->writing = 1;
584 /* note mtime, size NOW. */
585 __ceph_finish_cap_snap(ci, capsnap);
591 if (ci->i_wrbuffer_ref_head == 0 &&
593 ci->i_dirty_caps == 0 &&
594 ci->i_flushing_caps == 0) {
595 ci->i_head_snapc = NULL;
597 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
598 dout(" new snapc is %p\n", new_snapc);
600 spin_unlock(&ci->i_ceph_lock);
602 ceph_buffer_put(old_blob);
604 ceph_put_snap_context(old_snapc);
608 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
609 * to be used for the snapshot, to be flushed back to the mds.
611 * If capsnap can now be flushed, add to snap_flush list, and return 1.
613 * Caller must hold i_ceph_lock.
615 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
616 struct ceph_cap_snap *capsnap)
618 struct inode *inode = &ci->vfs_inode;
619 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
621 BUG_ON(capsnap->writing);
622 capsnap->size = inode->i_size;
623 capsnap->mtime = inode->i_mtime;
624 capsnap->atime = inode->i_atime;
625 capsnap->ctime = inode->i_ctime;
626 capsnap->btime = ci->i_btime;
627 capsnap->change_attr = inode_peek_iversion_raw(inode);
628 capsnap->time_warp_seq = ci->i_time_warp_seq;
629 capsnap->truncate_size = ci->i_truncate_size;
630 capsnap->truncate_seq = ci->i_truncate_seq;
631 if (capsnap->dirty_pages) {
632 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
633 "still has %d dirty pages\n", inode, capsnap,
634 capsnap->context, capsnap->context->seq,
635 ceph_cap_string(capsnap->dirty), capsnap->size,
636 capsnap->dirty_pages);
640 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
641 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
642 inode, capsnap, capsnap->context,
643 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
646 spin_lock(&mdsc->snap_flush_lock);
647 if (list_empty(&ci->i_snap_flush_item)) {
649 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
651 spin_unlock(&mdsc->snap_flush_lock);
652 return 1; /* caller may want to ceph_flush_snaps */
656 * Queue cap_snaps for snap writeback for this realm and its children.
657 * Called under snap_rwsem, so realm topology won't change.
659 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
661 struct ceph_inode_info *ci;
662 struct inode *lastinode = NULL;
664 dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
666 spin_lock(&realm->inodes_with_caps_lock);
667 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
668 struct inode *inode = igrab(&ci->vfs_inode);
671 spin_unlock(&realm->inodes_with_caps_lock);
672 /* avoid calling iput_final() while holding
673 * mdsc->snap_rwsem or in mds dispatch threads */
674 ceph_async_iput(lastinode);
676 ceph_queue_cap_snap(ci);
677 spin_lock(&realm->inodes_with_caps_lock);
679 spin_unlock(&realm->inodes_with_caps_lock);
680 ceph_async_iput(lastinode);
682 dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
686 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
687 * the snap realm parameters from a given realm and all of its ancestors,
690 * Caller must hold snap_rwsem for write.
692 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
693 void *p, void *e, bool deletion,
694 struct ceph_snap_realm **realm_ret)
696 struct ceph_mds_snap_realm *ri; /* encoded */
697 __le64 *snaps; /* encoded */
698 __le64 *prior_parent_snaps; /* encoded */
699 struct ceph_snap_realm *realm;
700 struct ceph_snap_realm *first_realm = NULL;
701 struct ceph_snap_realm *realm_to_rebuild = NULL;
704 LIST_HEAD(dirty_realms);
706 lockdep_assert_held_write(&mdsc->snap_rwsem);
708 dout("update_snap_trace deletion=%d\n", deletion);
712 ceph_decode_need(&p, e, sizeof(*ri), bad);
715 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
716 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
718 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
719 prior_parent_snaps = p;
720 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
722 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
724 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
726 err = PTR_ERR(realm);
731 /* ensure the parent is correct */
732 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
735 rebuild_snapcs += err;
737 if (le64_to_cpu(ri->seq) > realm->seq) {
738 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
739 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
740 /* update realm parameters, snap lists */
741 realm->seq = le64_to_cpu(ri->seq);
742 realm->created = le64_to_cpu(ri->created);
743 realm->parent_since = le64_to_cpu(ri->parent_since);
745 realm->num_snaps = le32_to_cpu(ri->num_snaps);
746 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
750 realm->num_prior_parent_snaps =
751 le32_to_cpu(ri->num_prior_parent_snaps);
752 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
753 realm->num_prior_parent_snaps);
757 if (realm->seq > mdsc->last_snap_seq)
758 mdsc->last_snap_seq = realm->seq;
761 } else if (!realm->cached_context) {
762 dout("update_snap_trace %llx %p seq %lld new\n",
763 realm->ino, realm, realm->seq);
766 dout("update_snap_trace %llx %p seq %lld unchanged\n",
767 realm->ino, realm, realm->seq);
770 dout("done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
771 realm, rebuild_snapcs, p, e);
774 * this will always track the uppest parent realm from which
775 * we need to rebuild the snapshot contexts _downward_ in
779 realm_to_rebuild = realm;
781 /* rebuild_snapcs when we reach the _end_ (root) of the trace */
782 if (realm_to_rebuild && p >= e)
783 rebuild_snap_realms(realm_to_rebuild, &dirty_realms);
788 ceph_put_snap_realm(mdsc, realm);
794 * queue cap snaps _after_ we've built the new snap contexts,
795 * so that i_head_snapc can be set appropriately.
797 while (!list_empty(&dirty_realms)) {
798 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
800 list_del_init(&realm->dirty_item);
801 queue_realm_cap_snaps(realm);
805 *realm_ret = first_realm;
807 ceph_put_snap_realm(mdsc, first_realm);
809 __cleanup_empty_realms(mdsc);
815 if (realm && !IS_ERR(realm))
816 ceph_put_snap_realm(mdsc, realm);
818 ceph_put_snap_realm(mdsc, first_realm);
819 pr_err("update_snap_trace error %d\n", err);
825 * Send any cap_snaps that are queued for flush. Try to carry
826 * s_mutex across multiple snap flushes to avoid locking overhead.
828 * Caller holds no locks.
830 static void flush_snaps(struct ceph_mds_client *mdsc)
832 struct ceph_inode_info *ci;
834 struct ceph_mds_session *session = NULL;
836 dout("flush_snaps\n");
837 spin_lock(&mdsc->snap_flush_lock);
838 while (!list_empty(&mdsc->snap_flush_list)) {
839 ci = list_first_entry(&mdsc->snap_flush_list,
840 struct ceph_inode_info, i_snap_flush_item);
841 inode = &ci->vfs_inode;
843 spin_unlock(&mdsc->snap_flush_lock);
844 ceph_flush_snaps(ci, &session);
845 /* avoid calling iput_final() while holding
846 * session->s_mutex or in mds dispatch threads */
847 ceph_async_iput(inode);
848 spin_lock(&mdsc->snap_flush_lock);
850 spin_unlock(&mdsc->snap_flush_lock);
853 mutex_unlock(&session->s_mutex);
854 ceph_put_mds_session(session);
856 dout("flush_snaps done\n");
861 * Handle a snap notification from the MDS.
863 * This can take two basic forms: the simplest is just a snap creation
864 * or deletion notification on an existing realm. This should update the
865 * realm and its children.
867 * The more difficult case is realm creation, due to snap creation at a
868 * new point in the file hierarchy, or due to a rename that moves a file or
869 * directory into another realm.
871 void ceph_handle_snap(struct ceph_mds_client *mdsc,
872 struct ceph_mds_session *session,
873 struct ceph_msg *msg)
875 struct super_block *sb = mdsc->fsc->sb;
876 int mds = session->s_mds;
880 struct ceph_snap_realm *realm = NULL;
881 void *p = msg->front.iov_base;
882 void *e = p + msg->front.iov_len;
883 struct ceph_mds_snap_head *h;
884 int num_split_inos, num_split_realms;
885 __le64 *split_inos = NULL, *split_realms = NULL;
887 int locked_rwsem = 0;
890 if (msg->front.iov_len < sizeof(*h))
893 op = le32_to_cpu(h->op);
894 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
896 num_split_inos = le32_to_cpu(h->num_split_inos);
897 num_split_realms = le32_to_cpu(h->num_split_realms);
898 trace_len = le32_to_cpu(h->trace_len);
901 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
902 ceph_snap_op_name(op), split, trace_len);
904 mutex_lock(&session->s_mutex);
906 mutex_unlock(&session->s_mutex);
908 down_write(&mdsc->snap_rwsem);
911 if (op == CEPH_SNAP_OP_SPLIT) {
912 struct ceph_mds_snap_realm *ri;
915 * A "split" breaks part of an existing realm off into
916 * a new realm. The MDS provides a list of inodes
917 * (with caps) and child realms that belong to the new
921 p += sizeof(u64) * num_split_inos;
923 p += sizeof(u64) * num_split_realms;
924 ceph_decode_need(&p, e, sizeof(*ri), bad);
925 /* we will peek at realm info here, but will _not_
926 * advance p, as the realm update will occur below in
927 * ceph_update_snap_trace. */
930 realm = ceph_lookup_snap_realm(mdsc, split);
932 realm = ceph_create_snap_realm(mdsc, split);
937 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
938 for (i = 0; i < num_split_inos; i++) {
939 struct ceph_vino vino = {
940 .ino = le64_to_cpu(split_inos[i]),
943 struct inode *inode = ceph_find_inode(sb, vino);
944 struct ceph_inode_info *ci;
945 struct ceph_snap_realm *oldrealm;
949 ci = ceph_inode(inode);
951 spin_lock(&ci->i_ceph_lock);
952 if (!ci->i_snap_realm)
955 * If this inode belongs to a realm that was
956 * created after our new realm, we experienced
957 * a race (due to another split notifications
958 * arriving from a different MDS). So skip
961 if (ci->i_snap_realm->created >
962 le64_to_cpu(ri->created)) {
963 dout(" leaving %p in newer realm %llx %p\n",
964 inode, ci->i_snap_realm->ino,
968 dout(" will move %p to split realm %llx %p\n",
969 inode, realm->ino, realm);
971 * Move the inode to the new realm
973 oldrealm = ci->i_snap_realm;
974 spin_lock(&oldrealm->inodes_with_caps_lock);
975 list_del_init(&ci->i_snap_realm_item);
976 spin_unlock(&oldrealm->inodes_with_caps_lock);
978 spin_lock(&realm->inodes_with_caps_lock);
979 list_add(&ci->i_snap_realm_item,
980 &realm->inodes_with_caps);
981 ci->i_snap_realm = realm;
982 if (realm->ino == ci->i_vino.ino)
983 realm->inode = inode;
984 spin_unlock(&realm->inodes_with_caps_lock);
986 spin_unlock(&ci->i_ceph_lock);
988 ceph_get_snap_realm(mdsc, realm);
989 ceph_put_snap_realm(mdsc, oldrealm);
991 /* avoid calling iput_final() while holding
992 * mdsc->snap_rwsem or mds in dispatch threads */
993 ceph_async_iput(inode);
997 spin_unlock(&ci->i_ceph_lock);
998 ceph_async_iput(inode);
1001 /* we may have taken some of the old realm's children. */
1002 for (i = 0; i < num_split_realms; i++) {
1003 struct ceph_snap_realm *child =
1004 __lookup_snap_realm(mdsc,
1005 le64_to_cpu(split_realms[i]));
1008 adjust_snap_realm_parent(mdsc, child, realm->ino);
1012 * In the non-split case both 'num_split_inos' and
1013 * 'num_split_realms' should be 0, making this a no-op.
1014 * However the MDS happens to populate 'split_realms' list
1015 * in one of the UPDATE op cases by mistake.
1017 * Skip both lists just in case to ensure that 'p' is
1018 * positioned at the start of realm info, as expected by
1019 * ceph_update_snap_trace().
1021 p += sizeof(u64) * num_split_inos;
1022 p += sizeof(u64) * num_split_realms;
1026 * update using the provided snap trace. if we are deleting a
1027 * snap, we can avoid queueing cap_snaps.
1029 ceph_update_snap_trace(mdsc, p, e,
1030 op == CEPH_SNAP_OP_DESTROY, NULL);
1032 if (op == CEPH_SNAP_OP_SPLIT)
1033 /* we took a reference when we created the realm, above */
1034 ceph_put_snap_realm(mdsc, realm);
1036 __cleanup_empty_realms(mdsc);
1038 up_write(&mdsc->snap_rwsem);
1044 pr_err("corrupt snap message from mds%d\n", mds);
1048 up_write(&mdsc->snap_rwsem);
1052 struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1055 struct ceph_snapid_map *sm, *exist;
1056 struct rb_node **p, *parent;
1060 spin_lock(&mdsc->snapid_map_lock);
1061 p = &mdsc->snapid_map_tree.rb_node;
1063 exist = rb_entry(*p, struct ceph_snapid_map, node);
1064 if (snap > exist->snap) {
1066 } else if (snap < exist->snap) {
1067 p = &(*p)->rb_right;
1069 if (atomic_inc_return(&exist->ref) == 1)
1070 list_del_init(&exist->lru);
1075 spin_unlock(&mdsc->snapid_map_lock);
1077 dout("found snapid map %llx -> %x\n", exist->snap, exist->dev);
1081 sm = kmalloc(sizeof(*sm), GFP_NOFS);
1085 ret = get_anon_bdev(&sm->dev);
1091 INIT_LIST_HEAD(&sm->lru);
1092 atomic_set(&sm->ref, 1);
1097 p = &mdsc->snapid_map_tree.rb_node;
1098 spin_lock(&mdsc->snapid_map_lock);
1101 exist = rb_entry(*p, struct ceph_snapid_map, node);
1102 if (snap > exist->snap)
1104 else if (snap < exist->snap)
1105 p = &(*p)->rb_right;
1111 if (atomic_inc_return(&exist->ref) == 1)
1112 list_del_init(&exist->lru);
1114 rb_link_node(&sm->node, parent, p);
1115 rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1117 spin_unlock(&mdsc->snapid_map_lock);
1119 free_anon_bdev(sm->dev);
1121 dout("found snapid map %llx -> %x\n", exist->snap, exist->dev);
1125 dout("create snapid map %llx -> %x\n", sm->snap, sm->dev);
1129 void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1130 struct ceph_snapid_map *sm)
1134 if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1135 if (!RB_EMPTY_NODE(&sm->node)) {
1136 sm->last_used = jiffies;
1137 list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1138 spin_unlock(&mdsc->snapid_map_lock);
1140 /* already cleaned up by
1141 * ceph_cleanup_snapid_map() */
1142 spin_unlock(&mdsc->snapid_map_lock);
1148 void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1150 struct ceph_snapid_map *sm;
1154 spin_lock(&mdsc->snapid_map_lock);
1157 while (!list_empty(&mdsc->snapid_map_lru)) {
1158 sm = list_first_entry(&mdsc->snapid_map_lru,
1159 struct ceph_snapid_map, lru);
1160 if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1163 rb_erase(&sm->node, &mdsc->snapid_map_tree);
1164 list_move(&sm->lru, &to_free);
1166 spin_unlock(&mdsc->snapid_map_lock);
1168 while (!list_empty(&to_free)) {
1169 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1171 dout("trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1172 free_anon_bdev(sm->dev);
1177 void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1179 struct ceph_snapid_map *sm;
1183 spin_lock(&mdsc->snapid_map_lock);
1184 while ((p = rb_first(&mdsc->snapid_map_tree))) {
1185 sm = rb_entry(p, struct ceph_snapid_map, node);
1186 rb_erase(p, &mdsc->snapid_map_tree);
1188 list_move(&sm->lru, &to_free);
1190 spin_unlock(&mdsc->snapid_map_lock);
1192 while (!list_empty(&to_free)) {
1193 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1195 free_anon_bdev(sm->dev);
1196 if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1197 pr_err("snapid map %llx -> %x still in use\n",