1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/sort.h>
6 #include <linux/slab.h>
7 #include <linux/iversion.h>
9 #include "mds_client.h"
10 #include <linux/ceph/decode.h>
12 /* unused map expires after 5 minutes */
13 #define CEPH_SNAPID_MAP_TIMEOUT (5 * 60 * HZ)
16 * Snapshots in ceph are driven in large part by cooperation from the
17 * client. In contrast to local file systems or file servers that
18 * implement snapshots at a single point in the system, ceph's
19 * distributed access to storage requires clients to help decide
20 * whether a write logically occurs before or after a recently created
23 * This provides a perfect instantanous client-wide snapshot. Between
24 * clients, however, snapshots may appear to be applied at slightly
25 * different points in time, depending on delays in delivering the
26 * snapshot notification.
28 * Snapshots are _not_ file system-wide. Instead, each snapshot
29 * applies to the subdirectory nested beneath some directory. This
30 * effectively divides the hierarchy into multiple "realms," where all
31 * of the files contained by each realm share the same set of
32 * snapshots. An individual realm's snap set contains snapshots
33 * explicitly created on that realm, as well as any snaps in its
34 * parent's snap set _after_ the point at which the parent became it's
35 * parent (due to, say, a rename). Similarly, snaps from prior parents
36 * during the time intervals during which they were the parent are included.
38 * The client is spared most of this detail, fortunately... it must only
39 * maintains a hierarchy of realms reflecting the current parent/child
40 * realm relationship, and for each realm has an explicit list of snaps
41 * inherited from prior parents.
43 * A snap_realm struct is maintained for realms containing every inode
44 * with an open cap in the system. (The needed snap realm information is
45 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq'
46 * version number is used to ensure that as realm parameters change (new
47 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
49 * The realm hierarchy drives the generation of a 'snap context' for each
50 * realm, which simply lists the resulting set of snaps for the realm. This
51 * is attached to any writes sent to OSDs.
54 * Unfortunately error handling is a bit mixed here. If we get a snap
55 * update, but don't have enough memory to update our realm hierarchy,
56 * it's not clear what we can do about it (besides complaining to the
62 * increase ref count for the realm
64 * caller must hold snap_rwsem.
66 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
67 struct ceph_snap_realm *realm)
69 lockdep_assert_held(&mdsc->snap_rwsem);
72 * The 0->1 and 1->0 transitions must take the snap_empty_lock
73 * atomically with the refcount change. Go ahead and bump the
74 * nref here, unless it's 0, in which case we take the spinlock
75 * and then do the increment and remove it from the list.
77 if (atomic_inc_not_zero(&realm->nref))
80 spin_lock(&mdsc->snap_empty_lock);
81 if (atomic_inc_return(&realm->nref) == 1)
82 list_del_init(&realm->empty_item);
83 spin_unlock(&mdsc->snap_empty_lock);
86 static void __insert_snap_realm(struct rb_root *root,
87 struct ceph_snap_realm *new)
89 struct rb_node **p = &root->rb_node;
90 struct rb_node *parent = NULL;
91 struct ceph_snap_realm *r = NULL;
95 r = rb_entry(parent, struct ceph_snap_realm, node);
96 if (new->ino < r->ino)
98 else if (new->ino > r->ino)
104 rb_link_node(&new->node, parent, p);
105 rb_insert_color(&new->node, root);
109 * create and get the realm rooted at @ino and bump its ref count.
111 * caller must hold snap_rwsem for write.
113 static struct ceph_snap_realm *ceph_create_snap_realm(
114 struct ceph_mds_client *mdsc,
117 struct ceph_snap_realm *realm;
119 lockdep_assert_held_write(&mdsc->snap_rwsem);
121 realm = kzalloc(sizeof(*realm), GFP_NOFS);
123 return ERR_PTR(-ENOMEM);
125 /* Do not release the global dummy snaprealm until unmouting */
126 if (ino == CEPH_INO_GLOBAL_SNAPREALM)
127 atomic_set(&realm->nref, 2);
129 atomic_set(&realm->nref, 1);
131 INIT_LIST_HEAD(&realm->children);
132 INIT_LIST_HEAD(&realm->child_item);
133 INIT_LIST_HEAD(&realm->empty_item);
134 INIT_LIST_HEAD(&realm->dirty_item);
135 INIT_LIST_HEAD(&realm->rebuild_item);
136 INIT_LIST_HEAD(&realm->inodes_with_caps);
137 spin_lock_init(&realm->inodes_with_caps_lock);
138 __insert_snap_realm(&mdsc->snap_realms, realm);
139 mdsc->num_snap_realms++;
141 dout("%s %llx %p\n", __func__, realm->ino, realm);
146 * lookup the realm rooted at @ino.
148 * caller must hold snap_rwsem.
150 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
153 struct rb_node *n = mdsc->snap_realms.rb_node;
154 struct ceph_snap_realm *r;
156 lockdep_assert_held(&mdsc->snap_rwsem);
159 r = rb_entry(n, struct ceph_snap_realm, node);
162 else if (ino > r->ino)
165 dout("%s %llx %p\n", __func__, r->ino, r);
172 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
175 struct ceph_snap_realm *r;
176 r = __lookup_snap_realm(mdsc, ino);
178 ceph_get_snap_realm(mdsc, r);
182 static void __put_snap_realm(struct ceph_mds_client *mdsc,
183 struct ceph_snap_realm *realm);
186 * called with snap_rwsem (write)
188 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
189 struct ceph_snap_realm *realm)
191 lockdep_assert_held_write(&mdsc->snap_rwsem);
193 dout("%s %p %llx\n", __func__, realm, realm->ino);
195 rb_erase(&realm->node, &mdsc->snap_realms);
196 mdsc->num_snap_realms--;
199 list_del_init(&realm->child_item);
200 __put_snap_realm(mdsc, realm->parent);
203 kfree(realm->prior_parent_snaps);
205 ceph_put_snap_context(realm->cached_context);
210 * caller holds snap_rwsem (write)
212 static void __put_snap_realm(struct ceph_mds_client *mdsc,
213 struct ceph_snap_realm *realm)
215 lockdep_assert_held_write(&mdsc->snap_rwsem);
218 * We do not require the snap_empty_lock here, as any caller that
219 * increments the value must hold the snap_rwsem.
221 if (atomic_dec_and_test(&realm->nref))
222 __destroy_snap_realm(mdsc, realm);
226 * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
228 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
229 struct ceph_snap_realm *realm)
231 if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
234 if (down_write_trylock(&mdsc->snap_rwsem)) {
235 spin_unlock(&mdsc->snap_empty_lock);
236 __destroy_snap_realm(mdsc, realm);
237 up_write(&mdsc->snap_rwsem);
239 list_add(&realm->empty_item, &mdsc->snap_empty);
240 spin_unlock(&mdsc->snap_empty_lock);
245 * Clean up any realms whose ref counts have dropped to zero. Note
246 * that this does not include realms who were created but not yet
249 * Called under snap_rwsem (write)
251 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
253 struct ceph_snap_realm *realm;
255 lockdep_assert_held_write(&mdsc->snap_rwsem);
257 spin_lock(&mdsc->snap_empty_lock);
258 while (!list_empty(&mdsc->snap_empty)) {
259 realm = list_first_entry(&mdsc->snap_empty,
260 struct ceph_snap_realm, empty_item);
261 list_del(&realm->empty_item);
262 spin_unlock(&mdsc->snap_empty_lock);
263 __destroy_snap_realm(mdsc, realm);
264 spin_lock(&mdsc->snap_empty_lock);
266 spin_unlock(&mdsc->snap_empty_lock);
269 void ceph_cleanup_global_and_empty_realms(struct ceph_mds_client *mdsc)
271 struct ceph_snap_realm *global_realm;
273 down_write(&mdsc->snap_rwsem);
274 global_realm = __lookup_snap_realm(mdsc, CEPH_INO_GLOBAL_SNAPREALM);
276 ceph_put_snap_realm(mdsc, global_realm);
277 __cleanup_empty_realms(mdsc);
278 up_write(&mdsc->snap_rwsem);
282 * adjust the parent realm of a given @realm. adjust child list, and parent
283 * pointers, and ref counts appropriately.
285 * return true if parent was changed, 0 if unchanged, <0 on error.
287 * caller must hold snap_rwsem for write.
289 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
290 struct ceph_snap_realm *realm,
293 struct ceph_snap_realm *parent;
295 lockdep_assert_held_write(&mdsc->snap_rwsem);
297 if (realm->parent_ino == parentino)
300 parent = ceph_lookup_snap_realm(mdsc, parentino);
302 parent = ceph_create_snap_realm(mdsc, parentino);
304 return PTR_ERR(parent);
306 dout("%s %llx %p: %llx %p -> %llx %p\n", __func__, realm->ino,
307 realm, realm->parent_ino, realm->parent, parentino, parent);
309 list_del_init(&realm->child_item);
310 ceph_put_snap_realm(mdsc, realm->parent);
312 realm->parent_ino = parentino;
313 realm->parent = parent;
314 list_add(&realm->child_item, &parent->children);
319 static int cmpu64_rev(const void *a, const void *b)
321 if (*(u64 *)a < *(u64 *)b)
323 if (*(u64 *)a > *(u64 *)b)
330 * build the snap context for a given realm.
332 static int build_snap_context(struct ceph_snap_realm *realm,
333 struct list_head *realm_queue,
334 struct list_head *dirty_realms)
336 struct ceph_snap_realm *parent = realm->parent;
337 struct ceph_snap_context *snapc;
339 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
342 * build parent context, if it hasn't been built.
343 * conservatively estimate that all parent snaps might be
347 if (!parent->cached_context) {
348 /* add to the queue head */
349 list_add(&parent->rebuild_item, realm_queue);
352 num += parent->cached_context->num_snaps;
355 /* do i actually need to update? not if my context seq
356 matches realm seq, and my parents' does to. (this works
357 because we rebuild_snap_realms() works _downward_ in
358 hierarchy after each update.) */
359 if (realm->cached_context &&
360 realm->cached_context->seq == realm->seq &&
362 realm->cached_context->seq >= parent->cached_context->seq)) {
363 dout("%s %llx %p: %p seq %lld (%u snaps) (unchanged)\n",
364 __func__, realm->ino, realm, realm->cached_context,
365 realm->cached_context->seq,
366 (unsigned int)realm->cached_context->num_snaps);
370 /* alloc new snap context */
372 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
374 snapc = ceph_create_snap_context(num, GFP_NOFS);
378 /* build (reverse sorted) snap vector */
380 snapc->seq = realm->seq;
384 /* include any of parent's snaps occurring _after_ my
385 parent became my parent */
386 for (i = 0; i < parent->cached_context->num_snaps; i++)
387 if (parent->cached_context->snaps[i] >=
389 snapc->snaps[num++] =
390 parent->cached_context->snaps[i];
391 if (parent->cached_context->seq > snapc->seq)
392 snapc->seq = parent->cached_context->seq;
394 memcpy(snapc->snaps + num, realm->snaps,
395 sizeof(u64)*realm->num_snaps);
396 num += realm->num_snaps;
397 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
398 sizeof(u64)*realm->num_prior_parent_snaps);
399 num += realm->num_prior_parent_snaps;
401 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
402 snapc->num_snaps = num;
403 dout("%s %llx %p: %p seq %lld (%u snaps)\n", __func__, realm->ino,
404 realm, snapc, snapc->seq, (unsigned int) snapc->num_snaps);
406 ceph_put_snap_context(realm->cached_context);
407 realm->cached_context = snapc;
408 /* queue realm for cap_snap creation */
409 list_add_tail(&realm->dirty_item, dirty_realms);
414 * if we fail, clear old (incorrect) cached_context... hopefully
415 * we'll have better luck building it later
417 if (realm->cached_context) {
418 ceph_put_snap_context(realm->cached_context);
419 realm->cached_context = NULL;
421 pr_err("%s %llx %p fail %d\n", __func__, realm->ino, realm, err);
426 * rebuild snap context for the given realm and all of its children.
428 static void rebuild_snap_realms(struct ceph_snap_realm *realm,
429 struct list_head *dirty_realms)
431 LIST_HEAD(realm_queue);
435 list_add_tail(&realm->rebuild_item, &realm_queue);
437 while (!list_empty(&realm_queue)) {
438 struct ceph_snap_realm *_realm, *child;
440 _realm = list_first_entry(&realm_queue,
441 struct ceph_snap_realm,
445 * If the last building failed dues to memory
446 * issue, just empty the realm_queue and return
447 * to avoid infinite loop.
450 list_del_init(&_realm->rebuild_item);
454 last = build_snap_context(_realm, &realm_queue, dirty_realms);
455 dout("%s %llx %p, %s\n", __func__, _realm->ino, _realm,
456 last > 0 ? "is deferred" : !last ? "succeeded" : "failed");
458 /* is any child in the list ? */
459 list_for_each_entry(child, &_realm->children, child_item) {
460 if (!list_empty(&child->rebuild_item)) {
467 list_for_each_entry(child, &_realm->children, child_item)
468 list_add_tail(&child->rebuild_item, &realm_queue);
471 /* last == 1 means need to build parent first */
473 list_del_init(&_realm->rebuild_item);
479 * helper to allocate and decode an array of snapids. free prior
482 static int dup_array(u64 **dst, __le64 *src, u32 num)
488 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
491 for (i = 0; i < num; i++)
492 (*dst)[i] = get_unaligned_le64(src + i);
499 static bool has_new_snaps(struct ceph_snap_context *o,
500 struct ceph_snap_context *n)
502 if (n->num_snaps == 0)
504 /* snaps are in descending order */
505 return n->snaps[0] > o->seq;
509 * When a snapshot is applied, the size/mtime inode metadata is queued
510 * in a ceph_cap_snap (one for each snapshot) until writeback
511 * completes and the metadata can be flushed back to the MDS.
513 * However, if a (sync) write is currently in-progress when we apply
514 * the snapshot, we have to wait until the write succeeds or fails
515 * (and a final size/mtime is known). In this case the
516 * cap_snap->writing = 1, and is said to be "pending." When the write
517 * finishes, we __ceph_finish_cap_snap().
519 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
522 static void ceph_queue_cap_snap(struct ceph_inode_info *ci,
523 struct ceph_cap_snap **pcapsnap)
525 struct inode *inode = &ci->netfs.inode;
526 struct ceph_snap_context *old_snapc, *new_snapc;
527 struct ceph_cap_snap *capsnap = *pcapsnap;
528 struct ceph_buffer *old_blob = NULL;
531 spin_lock(&ci->i_ceph_lock);
532 used = __ceph_caps_used(ci);
533 dirty = __ceph_caps_dirty(ci);
535 old_snapc = ci->i_head_snapc;
536 new_snapc = ci->i_snap_realm->cached_context;
539 * If there is a write in progress, treat that as a dirty Fw,
540 * even though it hasn't completed yet; by the time we finish
541 * up this capsnap it will be.
543 if (used & CEPH_CAP_FILE_WR)
544 dirty |= CEPH_CAP_FILE_WR;
546 if (__ceph_have_pending_cap_snap(ci)) {
547 /* there is no point in queuing multiple "pending" cap_snaps,
548 as no new writes are allowed to start when pending, so any
549 writes in progress now were started before the previous
550 cap_snap. lucky us. */
551 dout("%s %p %llx.%llx already pending\n",
552 __func__, inode, ceph_vinop(inode));
555 if (ci->i_wrbuffer_ref_head == 0 &&
556 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
557 dout("%s %p %llx.%llx nothing dirty|writing\n",
558 __func__, inode, ceph_vinop(inode));
565 * There is no need to send FLUSHSNAP message to MDS if there is
566 * no new snapshot. But when there is dirty pages or on-going
567 * writes, we still need to create cap_snap. cap_snap is needed
568 * by the write path and page writeback path.
570 * also see ceph_try_drop_cap_snap()
572 if (has_new_snaps(old_snapc, new_snapc)) {
573 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
574 capsnap->need_flush = true;
576 if (!(used & CEPH_CAP_FILE_WR) &&
577 ci->i_wrbuffer_ref_head == 0) {
578 dout("%s %p %llx.%llx no new_snap|dirty_page|writing\n",
579 __func__, inode, ceph_vinop(inode));
584 dout("%s %p %llx.%llx cap_snap %p queuing under %p %s %s\n",
585 __func__, inode, ceph_vinop(inode), capsnap, old_snapc,
586 ceph_cap_string(dirty), capsnap->need_flush ? "" : "no_flush");
589 capsnap->follows = old_snapc->seq;
590 capsnap->issued = __ceph_caps_issued(ci, NULL);
591 capsnap->dirty = dirty;
593 capsnap->mode = inode->i_mode;
594 capsnap->uid = inode->i_uid;
595 capsnap->gid = inode->i_gid;
597 if (dirty & CEPH_CAP_XATTR_EXCL) {
598 old_blob = __ceph_build_xattrs_blob(ci);
599 capsnap->xattr_blob =
600 ceph_buffer_get(ci->i_xattrs.blob);
601 capsnap->xattr_version = ci->i_xattrs.version;
603 capsnap->xattr_blob = NULL;
604 capsnap->xattr_version = 0;
607 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
609 /* dirty page count moved from _head to this cap_snap;
610 all subsequent writes page dirties occur _after_ this
612 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
613 ci->i_wrbuffer_ref_head = 0;
614 capsnap->context = old_snapc;
615 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
617 if (used & CEPH_CAP_FILE_WR) {
618 dout("%s %p %llx.%llx cap_snap %p snapc %p seq %llu used WR,"
619 " now pending\n", __func__, inode, ceph_vinop(inode),
620 capsnap, old_snapc, old_snapc->seq);
621 capsnap->writing = 1;
623 /* note mtime, size NOW. */
624 __ceph_finish_cap_snap(ci, capsnap);
630 if (ci->i_wrbuffer_ref_head == 0 &&
632 ci->i_dirty_caps == 0 &&
633 ci->i_flushing_caps == 0) {
634 ci->i_head_snapc = NULL;
636 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
637 dout(" new snapc is %p\n", new_snapc);
639 spin_unlock(&ci->i_ceph_lock);
641 ceph_buffer_put(old_blob);
642 ceph_put_snap_context(old_snapc);
646 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
647 * to be used for the snapshot, to be flushed back to the mds.
649 * If capsnap can now be flushed, add to snap_flush list, and return 1.
651 * Caller must hold i_ceph_lock.
653 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
654 struct ceph_cap_snap *capsnap)
656 struct inode *inode = &ci->netfs.inode;
657 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
659 BUG_ON(capsnap->writing);
660 capsnap->size = i_size_read(inode);
661 capsnap->mtime = inode->i_mtime;
662 capsnap->atime = inode->i_atime;
663 capsnap->ctime = inode_get_ctime(inode);
664 capsnap->btime = ci->i_btime;
665 capsnap->change_attr = inode_peek_iversion_raw(inode);
666 capsnap->time_warp_seq = ci->i_time_warp_seq;
667 capsnap->truncate_size = ci->i_truncate_size;
668 capsnap->truncate_seq = ci->i_truncate_seq;
669 if (capsnap->dirty_pages) {
670 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu "
671 "still has %d dirty pages\n", __func__, inode,
672 ceph_vinop(inode), capsnap, capsnap->context,
673 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
674 capsnap->size, capsnap->dirty_pages);
679 * Defer flushing the capsnap if the dirty buffer not flushed yet.
680 * And trigger to flush the buffer immediately.
682 if (ci->i_wrbuffer_ref) {
683 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu "
684 "used WRBUFFER, delaying\n", __func__, inode,
685 ceph_vinop(inode), capsnap, capsnap->context,
686 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
688 ceph_queue_writeback(inode);
692 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
693 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu\n",
694 __func__, inode, ceph_vinop(inode), capsnap, capsnap->context,
695 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
698 spin_lock(&mdsc->snap_flush_lock);
699 if (list_empty(&ci->i_snap_flush_item)) {
701 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
703 spin_unlock(&mdsc->snap_flush_lock);
704 return 1; /* caller may want to ceph_flush_snaps */
708 * Queue cap_snaps for snap writeback for this realm and its children.
709 * Called under snap_rwsem, so realm topology won't change.
711 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
713 struct ceph_inode_info *ci;
714 struct inode *lastinode = NULL;
715 struct ceph_cap_snap *capsnap = NULL;
717 dout("%s %p %llx inode\n", __func__, realm, realm->ino);
719 spin_lock(&realm->inodes_with_caps_lock);
720 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
721 struct inode *inode = igrab(&ci->netfs.inode);
724 spin_unlock(&realm->inodes_with_caps_lock);
729 * Allocate the capsnap memory outside of ceph_queue_cap_snap()
730 * to reduce very possible but unnecessary frequently memory
731 * allocate/free in this loop.
734 capsnap = kmem_cache_zalloc(ceph_cap_snap_cachep, GFP_NOFS);
736 pr_err("ENOMEM allocating ceph_cap_snap on %p\n",
741 capsnap->cap_flush.is_capsnap = true;
742 refcount_set(&capsnap->nref, 1);
743 INIT_LIST_HEAD(&capsnap->cap_flush.i_list);
744 INIT_LIST_HEAD(&capsnap->cap_flush.g_list);
745 INIT_LIST_HEAD(&capsnap->ci_item);
747 ceph_queue_cap_snap(ci, &capsnap);
748 spin_lock(&realm->inodes_with_caps_lock);
750 spin_unlock(&realm->inodes_with_caps_lock);
754 kmem_cache_free(ceph_cap_snap_cachep, capsnap);
755 dout("%s %p %llx done\n", __func__, realm, realm->ino);
759 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
760 * the snap realm parameters from a given realm and all of its ancestors,
763 * Caller must hold snap_rwsem for write.
765 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
766 void *p, void *e, bool deletion,
767 struct ceph_snap_realm **realm_ret)
769 struct ceph_mds_snap_realm *ri; /* encoded */
770 __le64 *snaps; /* encoded */
771 __le64 *prior_parent_snaps; /* encoded */
772 struct ceph_snap_realm *realm;
773 struct ceph_snap_realm *first_realm = NULL;
774 struct ceph_snap_realm *realm_to_rebuild = NULL;
775 struct ceph_client *client = mdsc->fsc->client;
779 LIST_HEAD(dirty_realms);
781 lockdep_assert_held_write(&mdsc->snap_rwsem);
783 dout("%s deletion=%d\n", __func__, deletion);
787 ceph_decode_need(&p, e, sizeof(*ri), bad);
790 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
791 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
793 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
794 prior_parent_snaps = p;
795 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
797 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
799 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
801 err = PTR_ERR(realm);
806 /* ensure the parent is correct */
807 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
810 rebuild_snapcs += err;
812 if (le64_to_cpu(ri->seq) > realm->seq) {
813 dout("%s updating %llx %p %lld -> %lld\n", __func__,
814 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
815 /* update realm parameters, snap lists */
816 realm->seq = le64_to_cpu(ri->seq);
817 realm->created = le64_to_cpu(ri->created);
818 realm->parent_since = le64_to_cpu(ri->parent_since);
820 realm->num_snaps = le32_to_cpu(ri->num_snaps);
821 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
825 realm->num_prior_parent_snaps =
826 le32_to_cpu(ri->num_prior_parent_snaps);
827 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
828 realm->num_prior_parent_snaps);
832 if (realm->seq > mdsc->last_snap_seq)
833 mdsc->last_snap_seq = realm->seq;
836 } else if (!realm->cached_context) {
837 dout("%s %llx %p seq %lld new\n", __func__,
838 realm->ino, realm, realm->seq);
841 dout("%s %llx %p seq %lld unchanged\n", __func__,
842 realm->ino, realm, realm->seq);
845 dout("done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
846 realm, rebuild_snapcs, p, e);
849 * this will always track the uppest parent realm from which
850 * we need to rebuild the snapshot contexts _downward_ in
854 realm_to_rebuild = realm;
856 /* rebuild_snapcs when we reach the _end_ (root) of the trace */
857 if (realm_to_rebuild && p >= e)
858 rebuild_snap_realms(realm_to_rebuild, &dirty_realms);
863 ceph_put_snap_realm(mdsc, realm);
869 * queue cap snaps _after_ we've built the new snap contexts,
870 * so that i_head_snapc can be set appropriately.
872 while (!list_empty(&dirty_realms)) {
873 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
875 list_del_init(&realm->dirty_item);
876 queue_realm_cap_snaps(realm);
880 *realm_ret = first_realm;
882 ceph_put_snap_realm(mdsc, first_realm);
884 __cleanup_empty_realms(mdsc);
890 if (realm && !IS_ERR(realm))
891 ceph_put_snap_realm(mdsc, realm);
893 ceph_put_snap_realm(mdsc, first_realm);
894 pr_err("%s error %d\n", __func__, err);
897 * When receiving a corrupted snap trace we don't know what
898 * exactly has happened in MDS side. And we shouldn't continue
899 * writing to OSD, which may corrupt the snapshot contents.
901 * Just try to blocklist this kclient and then this kclient
902 * must be remounted to continue after the corrupted metadata
903 * fixed in the MDS side.
905 WRITE_ONCE(mdsc->fsc->mount_state, CEPH_MOUNT_FENCE_IO);
906 ret = ceph_monc_blocklist_add(&client->monc, &client->msgr.inst.addr);
908 pr_err("%s failed to blocklist %s: %d\n", __func__,
909 ceph_pr_addr(&client->msgr.inst.addr), ret);
911 WARN(1, "%s: %s%sdo remount to continue%s",
912 __func__, ret ? "" : ceph_pr_addr(&client->msgr.inst.addr),
913 ret ? "" : " was blocklisted, ",
914 err == -EIO ? " after corrupted snaptrace is fixed" : "");
921 * Send any cap_snaps that are queued for flush. Try to carry
922 * s_mutex across multiple snap flushes to avoid locking overhead.
924 * Caller holds no locks.
926 static void flush_snaps(struct ceph_mds_client *mdsc)
928 struct ceph_inode_info *ci;
930 struct ceph_mds_session *session = NULL;
932 dout("%s\n", __func__);
933 spin_lock(&mdsc->snap_flush_lock);
934 while (!list_empty(&mdsc->snap_flush_list)) {
935 ci = list_first_entry(&mdsc->snap_flush_list,
936 struct ceph_inode_info, i_snap_flush_item);
937 inode = &ci->netfs.inode;
939 spin_unlock(&mdsc->snap_flush_lock);
940 ceph_flush_snaps(ci, &session);
942 spin_lock(&mdsc->snap_flush_lock);
944 spin_unlock(&mdsc->snap_flush_lock);
946 ceph_put_mds_session(session);
947 dout("%s done\n", __func__);
951 * ceph_change_snap_realm - change the snap_realm for an inode
952 * @inode: inode to move to new snap realm
953 * @realm: new realm to move inode into (may be NULL)
955 * Detach an inode from its old snaprealm (if any) and attach it to
956 * the new snaprealm (if any). The old snap realm reference held by
957 * the inode is put. If realm is non-NULL, then the caller's reference
958 * to it is taken over by the inode.
960 void ceph_change_snap_realm(struct inode *inode, struct ceph_snap_realm *realm)
962 struct ceph_inode_info *ci = ceph_inode(inode);
963 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
964 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
966 lockdep_assert_held(&ci->i_ceph_lock);
969 spin_lock(&oldrealm->inodes_with_caps_lock);
970 list_del_init(&ci->i_snap_realm_item);
971 if (oldrealm->ino == ci->i_vino.ino)
972 oldrealm->inode = NULL;
973 spin_unlock(&oldrealm->inodes_with_caps_lock);
974 ceph_put_snap_realm(mdsc, oldrealm);
977 ci->i_snap_realm = realm;
980 spin_lock(&realm->inodes_with_caps_lock);
981 list_add(&ci->i_snap_realm_item, &realm->inodes_with_caps);
982 if (realm->ino == ci->i_vino.ino)
983 realm->inode = inode;
984 spin_unlock(&realm->inodes_with_caps_lock);
989 * Handle a snap notification from the MDS.
991 * This can take two basic forms: the simplest is just a snap creation
992 * or deletion notification on an existing realm. This should update the
993 * realm and its children.
995 * The more difficult case is realm creation, due to snap creation at a
996 * new point in the file hierarchy, or due to a rename that moves a file or
997 * directory into another realm.
999 void ceph_handle_snap(struct ceph_mds_client *mdsc,
1000 struct ceph_mds_session *session,
1001 struct ceph_msg *msg)
1003 struct super_block *sb = mdsc->fsc->sb;
1004 int mds = session->s_mds;
1008 struct ceph_snap_realm *realm = NULL;
1009 void *p = msg->front.iov_base;
1010 void *e = p + msg->front.iov_len;
1011 struct ceph_mds_snap_head *h;
1012 int num_split_inos, num_split_realms;
1013 __le64 *split_inos = NULL, *split_realms = NULL;
1015 int locked_rwsem = 0;
1016 bool close_sessions = false;
1018 if (!ceph_inc_mds_stopping_blocker(mdsc, session))
1022 if (msg->front.iov_len < sizeof(*h))
1025 op = le32_to_cpu(h->op);
1026 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
1028 num_split_inos = le32_to_cpu(h->num_split_inos);
1029 num_split_realms = le32_to_cpu(h->num_split_realms);
1030 trace_len = le32_to_cpu(h->trace_len);
1033 dout("%s from mds%d op %s split %llx tracelen %d\n", __func__,
1034 mds, ceph_snap_op_name(op), split, trace_len);
1036 down_write(&mdsc->snap_rwsem);
1039 if (op == CEPH_SNAP_OP_SPLIT) {
1040 struct ceph_mds_snap_realm *ri;
1043 * A "split" breaks part of an existing realm off into
1044 * a new realm. The MDS provides a list of inodes
1045 * (with caps) and child realms that belong to the new
1049 p += sizeof(u64) * num_split_inos;
1051 p += sizeof(u64) * num_split_realms;
1052 ceph_decode_need(&p, e, sizeof(*ri), bad);
1053 /* we will peek at realm info here, but will _not_
1054 * advance p, as the realm update will occur below in
1055 * ceph_update_snap_trace. */
1058 realm = ceph_lookup_snap_realm(mdsc, split);
1060 realm = ceph_create_snap_realm(mdsc, split);
1065 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
1066 for (i = 0; i < num_split_inos; i++) {
1067 struct ceph_vino vino = {
1068 .ino = le64_to_cpu(split_inos[i]),
1069 .snap = CEPH_NOSNAP,
1071 struct inode *inode = ceph_find_inode(sb, vino);
1072 struct ceph_inode_info *ci;
1076 ci = ceph_inode(inode);
1078 spin_lock(&ci->i_ceph_lock);
1079 if (!ci->i_snap_realm)
1082 * If this inode belongs to a realm that was
1083 * created after our new realm, we experienced
1084 * a race (due to another split notifications
1085 * arriving from a different MDS). So skip
1088 if (ci->i_snap_realm->created >
1089 le64_to_cpu(ri->created)) {
1090 dout(" leaving %p %llx.%llx in newer realm %llx %p\n",
1091 inode, ceph_vinop(inode), ci->i_snap_realm->ino,
1095 dout(" will move %p %llx.%llx to split realm %llx %p\n",
1096 inode, ceph_vinop(inode), realm->ino, realm);
1098 ceph_get_snap_realm(mdsc, realm);
1099 ceph_change_snap_realm(inode, realm);
1100 spin_unlock(&ci->i_ceph_lock);
1105 spin_unlock(&ci->i_ceph_lock);
1109 /* we may have taken some of the old realm's children. */
1110 for (i = 0; i < num_split_realms; i++) {
1111 struct ceph_snap_realm *child =
1112 __lookup_snap_realm(mdsc,
1113 le64_to_cpu(split_realms[i]));
1116 adjust_snap_realm_parent(mdsc, child, realm->ino);
1120 * In the non-split case both 'num_split_inos' and
1121 * 'num_split_realms' should be 0, making this a no-op.
1122 * However the MDS happens to populate 'split_realms' list
1123 * in one of the UPDATE op cases by mistake.
1125 * Skip both lists just in case to ensure that 'p' is
1126 * positioned at the start of realm info, as expected by
1127 * ceph_update_snap_trace().
1129 p += sizeof(u64) * num_split_inos;
1130 p += sizeof(u64) * num_split_realms;
1134 * update using the provided snap trace. if we are deleting a
1135 * snap, we can avoid queueing cap_snaps.
1137 if (ceph_update_snap_trace(mdsc, p, e,
1138 op == CEPH_SNAP_OP_DESTROY,
1140 close_sessions = true;
1144 if (op == CEPH_SNAP_OP_SPLIT)
1145 /* we took a reference when we created the realm, above */
1146 ceph_put_snap_realm(mdsc, realm);
1148 __cleanup_empty_realms(mdsc);
1150 up_write(&mdsc->snap_rwsem);
1153 ceph_dec_mds_stopping_blocker(mdsc);
1157 pr_err("%s corrupt snap message from mds%d\n", __func__, mds);
1161 up_write(&mdsc->snap_rwsem);
1163 ceph_dec_mds_stopping_blocker(mdsc);
1166 ceph_mdsc_close_sessions(mdsc);
1170 struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1173 struct ceph_snapid_map *sm, *exist;
1174 struct rb_node **p, *parent;
1178 spin_lock(&mdsc->snapid_map_lock);
1179 p = &mdsc->snapid_map_tree.rb_node;
1181 exist = rb_entry(*p, struct ceph_snapid_map, node);
1182 if (snap > exist->snap) {
1184 } else if (snap < exist->snap) {
1185 p = &(*p)->rb_right;
1187 if (atomic_inc_return(&exist->ref) == 1)
1188 list_del_init(&exist->lru);
1193 spin_unlock(&mdsc->snapid_map_lock);
1195 dout("%s found snapid map %llx -> %x\n", __func__,
1196 exist->snap, exist->dev);
1200 sm = kmalloc(sizeof(*sm), GFP_NOFS);
1204 ret = get_anon_bdev(&sm->dev);
1210 INIT_LIST_HEAD(&sm->lru);
1211 atomic_set(&sm->ref, 1);
1216 p = &mdsc->snapid_map_tree.rb_node;
1217 spin_lock(&mdsc->snapid_map_lock);
1220 exist = rb_entry(*p, struct ceph_snapid_map, node);
1221 if (snap > exist->snap)
1223 else if (snap < exist->snap)
1224 p = &(*p)->rb_right;
1230 if (atomic_inc_return(&exist->ref) == 1)
1231 list_del_init(&exist->lru);
1233 rb_link_node(&sm->node, parent, p);
1234 rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1236 spin_unlock(&mdsc->snapid_map_lock);
1238 free_anon_bdev(sm->dev);
1240 dout("%s found snapid map %llx -> %x\n", __func__,
1241 exist->snap, exist->dev);
1245 dout("%s create snapid map %llx -> %x\n", __func__,
1250 void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1251 struct ceph_snapid_map *sm)
1255 if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1256 if (!RB_EMPTY_NODE(&sm->node)) {
1257 sm->last_used = jiffies;
1258 list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1259 spin_unlock(&mdsc->snapid_map_lock);
1261 /* already cleaned up by
1262 * ceph_cleanup_snapid_map() */
1263 spin_unlock(&mdsc->snapid_map_lock);
1269 void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1271 struct ceph_snapid_map *sm;
1275 spin_lock(&mdsc->snapid_map_lock);
1278 while (!list_empty(&mdsc->snapid_map_lru)) {
1279 sm = list_first_entry(&mdsc->snapid_map_lru,
1280 struct ceph_snapid_map, lru);
1281 if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1284 rb_erase(&sm->node, &mdsc->snapid_map_tree);
1285 list_move(&sm->lru, &to_free);
1287 spin_unlock(&mdsc->snapid_map_lock);
1289 while (!list_empty(&to_free)) {
1290 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1292 dout("trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1293 free_anon_bdev(sm->dev);
1298 void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1300 struct ceph_snapid_map *sm;
1304 spin_lock(&mdsc->snapid_map_lock);
1305 while ((p = rb_first(&mdsc->snapid_map_tree))) {
1306 sm = rb_entry(p, struct ceph_snapid_map, node);
1307 rb_erase(p, &mdsc->snapid_map_tree);
1309 list_move(&sm->lru, &to_free);
1311 spin_unlock(&mdsc->snapid_map_lock);
1313 while (!list_empty(&to_free)) {
1314 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1316 free_anon_bdev(sm->dev);
1317 if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1318 pr_err("snapid map %llx -> %x still in use\n",