1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
4 #include <linux/sort.h>
5 #include <linux/slab.h>
8 #include "mds_client.h"
10 #include <linux/ceph/decode.h>
13 * Snapshots in ceph are driven in large part by cooperation from the
14 * client. In contrast to local file systems or file servers that
15 * implement snapshots at a single point in the system, ceph's
16 * distributed access to storage requires clients to help decide
17 * whether a write logically occurs before or after a recently created
20 * This provides a perfect instantanous client-wide snapshot. Between
21 * clients, however, snapshots may appear to be applied at slightly
22 * different points in time, depending on delays in delivering the
23 * snapshot notification.
25 * Snapshots are _not_ file system-wide. Instead, each snapshot
26 * applies to the subdirectory nested beneath some directory. This
27 * effectively divides the hierarchy into multiple "realms," where all
28 * of the files contained by each realm share the same set of
29 * snapshots. An individual realm's snap set contains snapshots
30 * explicitly created on that realm, as well as any snaps in its
31 * parent's snap set _after_ the point at which the parent became it's
32 * parent (due to, say, a rename). Similarly, snaps from prior parents
33 * during the time intervals during which they were the parent are included.
35 * The client is spared most of this detail, fortunately... it must only
36 * maintains a hierarchy of realms reflecting the current parent/child
37 * realm relationship, and for each realm has an explicit list of snaps
38 * inherited from prior parents.
40 * A snap_realm struct is maintained for realms containing every inode
41 * with an open cap in the system. (The needed snap realm information is
42 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq'
43 * version number is used to ensure that as realm parameters change (new
44 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
46 * The realm hierarchy drives the generation of a 'snap context' for each
47 * realm, which simply lists the resulting set of snaps for the realm. This
48 * is attached to any writes sent to OSDs.
51 * Unfortunately error handling is a bit mixed here. If we get a snap
52 * update, but don't have enough memory to update our realm hierarchy,
53 * it's not clear what we can do about it (besides complaining to the
59 * increase ref count for the realm
61 * caller must hold snap_rwsem for write.
63 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
64 struct ceph_snap_realm *realm)
66 dout("get_realm %p %d -> %d\n", realm,
67 atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
69 * since we _only_ increment realm refs or empty the empty
70 * list with snap_rwsem held, adjusting the empty list here is
71 * safe. we do need to protect against concurrent empty list
74 if (atomic_inc_return(&realm->nref) == 1) {
75 spin_lock(&mdsc->snap_empty_lock);
76 list_del_init(&realm->empty_item);
77 spin_unlock(&mdsc->snap_empty_lock);
81 static void __insert_snap_realm(struct rb_root *root,
82 struct ceph_snap_realm *new)
84 struct rb_node **p = &root->rb_node;
85 struct rb_node *parent = NULL;
86 struct ceph_snap_realm *r = NULL;
90 r = rb_entry(parent, struct ceph_snap_realm, node);
91 if (new->ino < r->ino)
93 else if (new->ino > r->ino)
99 rb_link_node(&new->node, parent, p);
100 rb_insert_color(&new->node, root);
104 * create and get the realm rooted at @ino and bump its ref count.
106 * caller must hold snap_rwsem for write.
108 static struct ceph_snap_realm *ceph_create_snap_realm(
109 struct ceph_mds_client *mdsc,
112 struct ceph_snap_realm *realm;
114 realm = kzalloc(sizeof(*realm), GFP_NOFS);
116 return ERR_PTR(-ENOMEM);
118 atomic_set(&realm->nref, 1); /* for caller */
120 INIT_LIST_HEAD(&realm->children);
121 INIT_LIST_HEAD(&realm->child_item);
122 INIT_LIST_HEAD(&realm->empty_item);
123 INIT_LIST_HEAD(&realm->dirty_item);
124 INIT_LIST_HEAD(&realm->inodes_with_caps);
125 spin_lock_init(&realm->inodes_with_caps_lock);
126 __insert_snap_realm(&mdsc->snap_realms, realm);
127 dout("create_snap_realm %llx %p\n", realm->ino, realm);
132 * lookup the realm rooted at @ino.
134 * caller must hold snap_rwsem for write.
136 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
139 struct rb_node *n = mdsc->snap_realms.rb_node;
140 struct ceph_snap_realm *r;
143 r = rb_entry(n, struct ceph_snap_realm, node);
146 else if (ino > r->ino)
149 dout("lookup_snap_realm %llx %p\n", r->ino, r);
156 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
159 struct ceph_snap_realm *r;
160 r = __lookup_snap_realm(mdsc, ino);
162 ceph_get_snap_realm(mdsc, r);
166 static void __put_snap_realm(struct ceph_mds_client *mdsc,
167 struct ceph_snap_realm *realm);
170 * called with snap_rwsem (write)
172 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
173 struct ceph_snap_realm *realm)
175 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
177 rb_erase(&realm->node, &mdsc->snap_realms);
180 list_del_init(&realm->child_item);
181 __put_snap_realm(mdsc, realm->parent);
184 kfree(realm->prior_parent_snaps);
186 ceph_put_snap_context(realm->cached_context);
191 * caller holds snap_rwsem (write)
193 static void __put_snap_realm(struct ceph_mds_client *mdsc,
194 struct ceph_snap_realm *realm)
196 dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
197 atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
198 if (atomic_dec_and_test(&realm->nref))
199 __destroy_snap_realm(mdsc, realm);
203 * caller needn't hold any locks
205 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
206 struct ceph_snap_realm *realm)
208 dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
209 atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
210 if (!atomic_dec_and_test(&realm->nref))
213 if (down_write_trylock(&mdsc->snap_rwsem)) {
214 __destroy_snap_realm(mdsc, realm);
215 up_write(&mdsc->snap_rwsem);
217 spin_lock(&mdsc->snap_empty_lock);
218 list_add(&realm->empty_item, &mdsc->snap_empty);
219 spin_unlock(&mdsc->snap_empty_lock);
224 * Clean up any realms whose ref counts have dropped to zero. Note
225 * that this does not include realms who were created but not yet
228 * Called under snap_rwsem (write)
230 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
232 struct ceph_snap_realm *realm;
234 spin_lock(&mdsc->snap_empty_lock);
235 while (!list_empty(&mdsc->snap_empty)) {
236 realm = list_first_entry(&mdsc->snap_empty,
237 struct ceph_snap_realm, empty_item);
238 list_del(&realm->empty_item);
239 spin_unlock(&mdsc->snap_empty_lock);
240 __destroy_snap_realm(mdsc, realm);
241 spin_lock(&mdsc->snap_empty_lock);
243 spin_unlock(&mdsc->snap_empty_lock);
246 void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
248 down_write(&mdsc->snap_rwsem);
249 __cleanup_empty_realms(mdsc);
250 up_write(&mdsc->snap_rwsem);
254 * adjust the parent realm of a given @realm. adjust child list, and parent
255 * pointers, and ref counts appropriately.
257 * return true if parent was changed, 0 if unchanged, <0 on error.
259 * caller must hold snap_rwsem for write.
261 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
262 struct ceph_snap_realm *realm,
265 struct ceph_snap_realm *parent;
267 if (realm->parent_ino == parentino)
270 parent = ceph_lookup_snap_realm(mdsc, parentino);
272 parent = ceph_create_snap_realm(mdsc, parentino);
274 return PTR_ERR(parent);
276 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
277 realm->ino, realm, realm->parent_ino, realm->parent,
280 list_del_init(&realm->child_item);
281 ceph_put_snap_realm(mdsc, realm->parent);
283 realm->parent_ino = parentino;
284 realm->parent = parent;
285 list_add(&realm->child_item, &parent->children);
290 static int cmpu64_rev(const void *a, const void *b)
292 if (*(u64 *)a < *(u64 *)b)
294 if (*(u64 *)a > *(u64 *)b)
301 * build the snap context for a given realm.
303 static int build_snap_context(struct ceph_snap_realm *realm,
304 struct list_head* dirty_realms)
306 struct ceph_snap_realm *parent = realm->parent;
307 struct ceph_snap_context *snapc;
309 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
312 * build parent context, if it hasn't been built.
313 * conservatively estimate that all parent snaps might be
317 if (!parent->cached_context) {
318 err = build_snap_context(parent, dirty_realms);
322 num += parent->cached_context->num_snaps;
325 /* do i actually need to update? not if my context seq
326 matches realm seq, and my parents' does to. (this works
327 because we rebuild_snap_realms() works _downward_ in
328 hierarchy after each update.) */
329 if (realm->cached_context &&
330 realm->cached_context->seq == realm->seq &&
332 realm->cached_context->seq >= parent->cached_context->seq)) {
333 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
335 realm->ino, realm, realm->cached_context,
336 realm->cached_context->seq,
337 (unsigned int)realm->cached_context->num_snaps);
341 /* alloc new snap context */
343 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
345 snapc = ceph_create_snap_context(num, GFP_NOFS);
349 /* build (reverse sorted) snap vector */
351 snapc->seq = realm->seq;
355 /* include any of parent's snaps occurring _after_ my
356 parent became my parent */
357 for (i = 0; i < parent->cached_context->num_snaps; i++)
358 if (parent->cached_context->snaps[i] >=
360 snapc->snaps[num++] =
361 parent->cached_context->snaps[i];
362 if (parent->cached_context->seq > snapc->seq)
363 snapc->seq = parent->cached_context->seq;
365 memcpy(snapc->snaps + num, realm->snaps,
366 sizeof(u64)*realm->num_snaps);
367 num += realm->num_snaps;
368 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
369 sizeof(u64)*realm->num_prior_parent_snaps);
370 num += realm->num_prior_parent_snaps;
372 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
373 snapc->num_snaps = num;
374 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
375 realm->ino, realm, snapc, snapc->seq,
376 (unsigned int) snapc->num_snaps);
378 ceph_put_snap_context(realm->cached_context);
379 realm->cached_context = snapc;
380 /* queue realm for cap_snap creation */
381 list_add_tail(&realm->dirty_item, dirty_realms);
386 * if we fail, clear old (incorrect) cached_context... hopefully
387 * we'll have better luck building it later
389 if (realm->cached_context) {
390 ceph_put_snap_context(realm->cached_context);
391 realm->cached_context = NULL;
393 pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
399 * rebuild snap context for the given realm and all of its children.
401 static void rebuild_snap_realms(struct ceph_snap_realm *realm,
402 struct list_head *dirty_realms)
404 struct ceph_snap_realm *child;
406 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
407 build_snap_context(realm, dirty_realms);
409 list_for_each_entry(child, &realm->children, child_item)
410 rebuild_snap_realms(child, dirty_realms);
415 * helper to allocate and decode an array of snapids. free prior
418 static int dup_array(u64 **dst, __le64 *src, u32 num)
424 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
427 for (i = 0; i < num; i++)
428 (*dst)[i] = get_unaligned_le64(src + i);
435 static bool has_new_snaps(struct ceph_snap_context *o,
436 struct ceph_snap_context *n)
438 if (n->num_snaps == 0)
440 /* snaps are in descending order */
441 return n->snaps[0] > o->seq;
445 * When a snapshot is applied, the size/mtime inode metadata is queued
446 * in a ceph_cap_snap (one for each snapshot) until writeback
447 * completes and the metadata can be flushed back to the MDS.
449 * However, if a (sync) write is currently in-progress when we apply
450 * the snapshot, we have to wait until the write succeeds or fails
451 * (and a final size/mtime is known). In this case the
452 * cap_snap->writing = 1, and is said to be "pending." When the write
453 * finishes, we __ceph_finish_cap_snap().
455 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
458 void ceph_queue_cap_snap(struct ceph_inode_info *ci)
460 struct inode *inode = &ci->vfs_inode;
461 struct ceph_cap_snap *capsnap;
462 struct ceph_snap_context *old_snapc, *new_snapc;
463 struct ceph_buffer *old_blob = NULL;
466 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
468 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
472 spin_lock(&ci->i_ceph_lock);
473 used = __ceph_caps_used(ci);
474 dirty = __ceph_caps_dirty(ci);
476 old_snapc = ci->i_head_snapc;
477 new_snapc = ci->i_snap_realm->cached_context;
480 * If there is a write in progress, treat that as a dirty Fw,
481 * even though it hasn't completed yet; by the time we finish
482 * up this capsnap it will be.
484 if (used & CEPH_CAP_FILE_WR)
485 dirty |= CEPH_CAP_FILE_WR;
487 if (__ceph_have_pending_cap_snap(ci)) {
488 /* there is no point in queuing multiple "pending" cap_snaps,
489 as no new writes are allowed to start when pending, so any
490 writes in progress now were started before the previous
491 cap_snap. lucky us. */
492 dout("queue_cap_snap %p already pending\n", inode);
495 if (ci->i_wrbuffer_ref_head == 0 &&
496 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
497 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
504 * There is no need to send FLUSHSNAP message to MDS if there is
505 * no new snapshot. But when there is dirty pages or on-going
506 * writes, we still need to create cap_snap. cap_snap is needed
507 * by the write path and page writeback path.
509 * also see ceph_try_drop_cap_snap()
511 if (has_new_snaps(old_snapc, new_snapc)) {
512 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
513 capsnap->need_flush = true;
515 if (!(used & CEPH_CAP_FILE_WR) &&
516 ci->i_wrbuffer_ref_head == 0) {
517 dout("queue_cap_snap %p "
518 "no new_snap|dirty_page|writing\n", inode);
523 dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n",
524 inode, capsnap, old_snapc, ceph_cap_string(dirty),
525 capsnap->need_flush ? "" : "no_flush");
528 refcount_set(&capsnap->nref, 1);
529 INIT_LIST_HEAD(&capsnap->ci_item);
531 capsnap->follows = old_snapc->seq;
532 capsnap->issued = __ceph_caps_issued(ci, NULL);
533 capsnap->dirty = dirty;
535 capsnap->mode = inode->i_mode;
536 capsnap->uid = inode->i_uid;
537 capsnap->gid = inode->i_gid;
539 if (dirty & CEPH_CAP_XATTR_EXCL) {
540 old_blob = __ceph_build_xattrs_blob(ci);
541 capsnap->xattr_blob =
542 ceph_buffer_get(ci->i_xattrs.blob);
543 capsnap->xattr_version = ci->i_xattrs.version;
545 capsnap->xattr_blob = NULL;
546 capsnap->xattr_version = 0;
549 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
551 /* dirty page count moved from _head to this cap_snap;
552 all subsequent writes page dirties occur _after_ this
554 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
555 ci->i_wrbuffer_ref_head = 0;
556 capsnap->context = old_snapc;
557 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
559 if (used & CEPH_CAP_FILE_WR) {
560 dout("queue_cap_snap %p cap_snap %p snapc %p"
561 " seq %llu used WR, now pending\n", inode,
562 capsnap, old_snapc, old_snapc->seq);
563 capsnap->writing = 1;
565 /* note mtime, size NOW. */
566 __ceph_finish_cap_snap(ci, capsnap);
572 if (ci->i_wrbuffer_ref_head == 0 &&
574 ci->i_dirty_caps == 0 &&
575 ci->i_flushing_caps == 0) {
576 ci->i_head_snapc = NULL;
578 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
579 dout(" new snapc is %p\n", new_snapc);
581 spin_unlock(&ci->i_ceph_lock);
583 ceph_buffer_put(old_blob);
585 ceph_put_snap_context(old_snapc);
589 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
590 * to be used for the snapshot, to be flushed back to the mds.
592 * If capsnap can now be flushed, add to snap_flush list, and return 1.
594 * Caller must hold i_ceph_lock.
596 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
597 struct ceph_cap_snap *capsnap)
599 struct inode *inode = &ci->vfs_inode;
600 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
602 BUG_ON(capsnap->writing);
603 capsnap->size = inode->i_size;
604 capsnap->mtime = inode->i_mtime;
605 capsnap->atime = inode->i_atime;
606 capsnap->ctime = inode->i_ctime;
607 capsnap->time_warp_seq = ci->i_time_warp_seq;
608 capsnap->truncate_size = ci->i_truncate_size;
609 capsnap->truncate_seq = ci->i_truncate_seq;
610 if (capsnap->dirty_pages) {
611 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
612 "still has %d dirty pages\n", inode, capsnap,
613 capsnap->context, capsnap->context->seq,
614 ceph_cap_string(capsnap->dirty), capsnap->size,
615 capsnap->dirty_pages);
619 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
620 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
621 inode, capsnap, capsnap->context,
622 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
625 spin_lock(&mdsc->snap_flush_lock);
626 if (list_empty(&ci->i_snap_flush_item)) {
628 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
630 spin_unlock(&mdsc->snap_flush_lock);
631 return 1; /* caller may want to ceph_flush_snaps */
635 * Queue cap_snaps for snap writeback for this realm and its children.
636 * Called under snap_rwsem, so realm topology won't change.
638 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
640 struct ceph_inode_info *ci;
641 struct inode *lastinode = NULL;
643 dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
645 spin_lock(&realm->inodes_with_caps_lock);
646 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
647 struct inode *inode = igrab(&ci->vfs_inode);
650 spin_unlock(&realm->inodes_with_caps_lock);
653 ceph_queue_cap_snap(ci);
654 spin_lock(&realm->inodes_with_caps_lock);
656 spin_unlock(&realm->inodes_with_caps_lock);
659 dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
663 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
664 * the snap realm parameters from a given realm and all of its ancestors,
667 * Caller must hold snap_rwsem for write.
669 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
670 void *p, void *e, bool deletion,
671 struct ceph_snap_realm **realm_ret)
673 struct ceph_mds_snap_realm *ri; /* encoded */
674 __le64 *snaps; /* encoded */
675 __le64 *prior_parent_snaps; /* encoded */
676 struct ceph_snap_realm *realm;
677 struct ceph_snap_realm *first_realm = NULL;
678 struct ceph_snap_realm *realm_to_rebuild = NULL;
681 LIST_HEAD(dirty_realms);
683 dout("update_snap_trace deletion=%d\n", deletion);
687 ceph_decode_need(&p, e, sizeof(*ri), bad);
690 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
691 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
693 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
694 prior_parent_snaps = p;
695 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
697 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
699 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
701 err = PTR_ERR(realm);
706 /* ensure the parent is correct */
707 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
710 rebuild_snapcs += err;
712 if (le64_to_cpu(ri->seq) > realm->seq) {
713 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
714 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
715 /* update realm parameters, snap lists */
716 realm->seq = le64_to_cpu(ri->seq);
717 realm->created = le64_to_cpu(ri->created);
718 realm->parent_since = le64_to_cpu(ri->parent_since);
720 realm->num_snaps = le32_to_cpu(ri->num_snaps);
721 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
725 realm->num_prior_parent_snaps =
726 le32_to_cpu(ri->num_prior_parent_snaps);
727 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
728 realm->num_prior_parent_snaps);
732 if (realm->seq > mdsc->last_snap_seq)
733 mdsc->last_snap_seq = realm->seq;
736 } else if (!realm->cached_context) {
737 dout("update_snap_trace %llx %p seq %lld new\n",
738 realm->ino, realm, realm->seq);
741 dout("update_snap_trace %llx %p seq %lld unchanged\n",
742 realm->ino, realm, realm->seq);
745 dout("done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
746 realm, rebuild_snapcs, p, e);
749 * this will always track the uppest parent realm from which
750 * we need to rebuild the snapshot contexts _downward_ in
754 realm_to_rebuild = realm;
756 /* rebuild_snapcs when we reach the _end_ (root) of the trace */
757 if (realm_to_rebuild && p >= e)
758 rebuild_snap_realms(realm_to_rebuild, &dirty_realms);
763 ceph_put_snap_realm(mdsc, realm);
769 * queue cap snaps _after_ we've built the new snap contexts,
770 * so that i_head_snapc can be set appropriately.
772 while (!list_empty(&dirty_realms)) {
773 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
775 list_del_init(&realm->dirty_item);
776 queue_realm_cap_snaps(realm);
780 *realm_ret = first_realm;
782 ceph_put_snap_realm(mdsc, first_realm);
784 __cleanup_empty_realms(mdsc);
790 if (realm && !IS_ERR(realm))
791 ceph_put_snap_realm(mdsc, realm);
793 ceph_put_snap_realm(mdsc, first_realm);
794 pr_err("update_snap_trace error %d\n", err);
800 * Send any cap_snaps that are queued for flush. Try to carry
801 * s_mutex across multiple snap flushes to avoid locking overhead.
803 * Caller holds no locks.
805 static void flush_snaps(struct ceph_mds_client *mdsc)
807 struct ceph_inode_info *ci;
809 struct ceph_mds_session *session = NULL;
811 dout("flush_snaps\n");
812 spin_lock(&mdsc->snap_flush_lock);
813 while (!list_empty(&mdsc->snap_flush_list)) {
814 ci = list_first_entry(&mdsc->snap_flush_list,
815 struct ceph_inode_info, i_snap_flush_item);
816 inode = &ci->vfs_inode;
818 spin_unlock(&mdsc->snap_flush_lock);
819 ceph_flush_snaps(ci, &session);
821 spin_lock(&mdsc->snap_flush_lock);
823 spin_unlock(&mdsc->snap_flush_lock);
826 mutex_unlock(&session->s_mutex);
827 ceph_put_mds_session(session);
829 dout("flush_snaps done\n");
834 * Handle a snap notification from the MDS.
836 * This can take two basic forms: the simplest is just a snap creation
837 * or deletion notification on an existing realm. This should update the
838 * realm and its children.
840 * The more difficult case is realm creation, due to snap creation at a
841 * new point in the file hierarchy, or due to a rename that moves a file or
842 * directory into another realm.
844 void ceph_handle_snap(struct ceph_mds_client *mdsc,
845 struct ceph_mds_session *session,
846 struct ceph_msg *msg)
848 struct super_block *sb = mdsc->fsc->sb;
849 int mds = session->s_mds;
853 struct ceph_snap_realm *realm = NULL;
854 void *p = msg->front.iov_base;
855 void *e = p + msg->front.iov_len;
856 struct ceph_mds_snap_head *h;
857 int num_split_inos, num_split_realms;
858 __le64 *split_inos = NULL, *split_realms = NULL;
860 int locked_rwsem = 0;
863 if (msg->front.iov_len < sizeof(*h))
866 op = le32_to_cpu(h->op);
867 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
869 num_split_inos = le32_to_cpu(h->num_split_inos);
870 num_split_realms = le32_to_cpu(h->num_split_realms);
871 trace_len = le32_to_cpu(h->trace_len);
874 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
875 ceph_snap_op_name(op), split, trace_len);
877 mutex_lock(&session->s_mutex);
879 mutex_unlock(&session->s_mutex);
881 down_write(&mdsc->snap_rwsem);
884 if (op == CEPH_SNAP_OP_SPLIT) {
885 struct ceph_mds_snap_realm *ri;
888 * A "split" breaks part of an existing realm off into
889 * a new realm. The MDS provides a list of inodes
890 * (with caps) and child realms that belong to the new
894 p += sizeof(u64) * num_split_inos;
896 p += sizeof(u64) * num_split_realms;
897 ceph_decode_need(&p, e, sizeof(*ri), bad);
898 /* we will peek at realm info here, but will _not_
899 * advance p, as the realm update will occur below in
900 * ceph_update_snap_trace. */
903 realm = ceph_lookup_snap_realm(mdsc, split);
905 realm = ceph_create_snap_realm(mdsc, split);
910 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
911 for (i = 0; i < num_split_inos; i++) {
912 struct ceph_vino vino = {
913 .ino = le64_to_cpu(split_inos[i]),
916 struct inode *inode = ceph_find_inode(sb, vino);
917 struct ceph_inode_info *ci;
918 struct ceph_snap_realm *oldrealm;
922 ci = ceph_inode(inode);
924 spin_lock(&ci->i_ceph_lock);
925 if (!ci->i_snap_realm)
928 * If this inode belongs to a realm that was
929 * created after our new realm, we experienced
930 * a race (due to another split notifications
931 * arriving from a different MDS). So skip
934 if (ci->i_snap_realm->created >
935 le64_to_cpu(ri->created)) {
936 dout(" leaving %p in newer realm %llx %p\n",
937 inode, ci->i_snap_realm->ino,
941 dout(" will move %p to split realm %llx %p\n",
942 inode, realm->ino, realm);
944 * Move the inode to the new realm
946 oldrealm = ci->i_snap_realm;
947 spin_lock(&oldrealm->inodes_with_caps_lock);
948 list_del_init(&ci->i_snap_realm_item);
949 spin_unlock(&oldrealm->inodes_with_caps_lock);
951 spin_lock(&realm->inodes_with_caps_lock);
952 list_add(&ci->i_snap_realm_item,
953 &realm->inodes_with_caps);
954 ci->i_snap_realm = realm;
955 if (realm->ino == ci->i_vino.ino)
956 realm->inode = inode;
957 spin_unlock(&realm->inodes_with_caps_lock);
959 spin_unlock(&ci->i_ceph_lock);
961 ceph_get_snap_realm(mdsc, realm);
962 ceph_put_snap_realm(mdsc, oldrealm);
968 spin_unlock(&ci->i_ceph_lock);
972 /* we may have taken some of the old realm's children. */
973 for (i = 0; i < num_split_realms; i++) {
974 struct ceph_snap_realm *child =
975 __lookup_snap_realm(mdsc,
976 le64_to_cpu(split_realms[i]));
979 adjust_snap_realm_parent(mdsc, child, realm->ino);
983 * In the non-split case both 'num_split_inos' and
984 * 'num_split_realms' should be 0, making this a no-op.
985 * However the MDS happens to populate 'split_realms' list
986 * in one of the UPDATE op cases by mistake.
988 * Skip both lists just in case to ensure that 'p' is
989 * positioned at the start of realm info, as expected by
990 * ceph_update_snap_trace().
992 p += sizeof(u64) * num_split_inos;
993 p += sizeof(u64) * num_split_realms;
997 * update using the provided snap trace. if we are deleting a
998 * snap, we can avoid queueing cap_snaps.
1000 ceph_update_snap_trace(mdsc, p, e,
1001 op == CEPH_SNAP_OP_DESTROY, NULL);
1003 if (op == CEPH_SNAP_OP_SPLIT)
1004 /* we took a reference when we created the realm, above */
1005 ceph_put_snap_realm(mdsc, realm);
1007 __cleanup_empty_realms(mdsc);
1009 up_write(&mdsc->snap_rwsem);
1015 pr_err("corrupt snap message from mds%d\n", mds);
1019 up_write(&mdsc->snap_rwsem);