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 /* Do not release the global dummy snaprealm until unmouting */
125 if (ino == CEPH_INO_GLOBAL_SNAPREALM)
126 atomic_set(&realm->nref, 2);
128 atomic_set(&realm->nref, 1);
130 INIT_LIST_HEAD(&realm->children);
131 INIT_LIST_HEAD(&realm->child_item);
132 INIT_LIST_HEAD(&realm->empty_item);
133 INIT_LIST_HEAD(&realm->dirty_item);
134 INIT_LIST_HEAD(&realm->rebuild_item);
135 INIT_LIST_HEAD(&realm->inodes_with_caps);
136 spin_lock_init(&realm->inodes_with_caps_lock);
137 __insert_snap_realm(&mdsc->snap_realms, realm);
138 mdsc->num_snap_realms++;
140 dout("%s %llx %p\n", __func__, realm->ino, realm);
145 * lookup the realm rooted at @ino.
147 * caller must hold snap_rwsem.
149 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
152 struct rb_node *n = mdsc->snap_realms.rb_node;
153 struct ceph_snap_realm *r;
155 lockdep_assert_held(&mdsc->snap_rwsem);
158 r = rb_entry(n, struct ceph_snap_realm, node);
161 else if (ino > r->ino)
164 dout("%s %llx %p\n", __func__, r->ino, r);
171 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
174 struct ceph_snap_realm *r;
175 r = __lookup_snap_realm(mdsc, ino);
177 ceph_get_snap_realm(mdsc, r);
181 static void __put_snap_realm(struct ceph_mds_client *mdsc,
182 struct ceph_snap_realm *realm);
185 * called with snap_rwsem (write)
187 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
188 struct ceph_snap_realm *realm)
190 lockdep_assert_held_write(&mdsc->snap_rwsem);
192 dout("%s %p %llx\n", __func__, realm, realm->ino);
194 rb_erase(&realm->node, &mdsc->snap_realms);
195 mdsc->num_snap_realms--;
198 list_del_init(&realm->child_item);
199 __put_snap_realm(mdsc, realm->parent);
202 kfree(realm->prior_parent_snaps);
204 ceph_put_snap_context(realm->cached_context);
209 * caller holds snap_rwsem (write)
211 static void __put_snap_realm(struct ceph_mds_client *mdsc,
212 struct ceph_snap_realm *realm)
214 lockdep_assert_held_write(&mdsc->snap_rwsem);
217 * We do not require the snap_empty_lock here, as any caller that
218 * increments the value must hold the snap_rwsem.
220 if (atomic_dec_and_test(&realm->nref))
221 __destroy_snap_realm(mdsc, realm);
225 * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
227 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
228 struct ceph_snap_realm *realm)
230 if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
233 if (down_write_trylock(&mdsc->snap_rwsem)) {
234 spin_unlock(&mdsc->snap_empty_lock);
235 __destroy_snap_realm(mdsc, realm);
236 up_write(&mdsc->snap_rwsem);
238 list_add(&realm->empty_item, &mdsc->snap_empty);
239 spin_unlock(&mdsc->snap_empty_lock);
244 * Clean up any realms whose ref counts have dropped to zero. Note
245 * that this does not include realms who were created but not yet
248 * Called under snap_rwsem (write)
250 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
252 struct ceph_snap_realm *realm;
254 lockdep_assert_held_write(&mdsc->snap_rwsem);
256 spin_lock(&mdsc->snap_empty_lock);
257 while (!list_empty(&mdsc->snap_empty)) {
258 realm = list_first_entry(&mdsc->snap_empty,
259 struct ceph_snap_realm, empty_item);
260 list_del(&realm->empty_item);
261 spin_unlock(&mdsc->snap_empty_lock);
262 __destroy_snap_realm(mdsc, realm);
263 spin_lock(&mdsc->snap_empty_lock);
265 spin_unlock(&mdsc->snap_empty_lock);
268 void ceph_cleanup_global_and_empty_realms(struct ceph_mds_client *mdsc)
270 struct ceph_snap_realm *global_realm;
272 down_write(&mdsc->snap_rwsem);
273 global_realm = __lookup_snap_realm(mdsc, CEPH_INO_GLOBAL_SNAPREALM);
275 ceph_put_snap_realm(mdsc, global_realm);
276 __cleanup_empty_realms(mdsc);
277 up_write(&mdsc->snap_rwsem);
281 * adjust the parent realm of a given @realm. adjust child list, and parent
282 * pointers, and ref counts appropriately.
284 * return true if parent was changed, 0 if unchanged, <0 on error.
286 * caller must hold snap_rwsem for write.
288 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
289 struct ceph_snap_realm *realm,
292 struct ceph_snap_realm *parent;
294 lockdep_assert_held_write(&mdsc->snap_rwsem);
296 if (realm->parent_ino == parentino)
299 parent = ceph_lookup_snap_realm(mdsc, parentino);
301 parent = ceph_create_snap_realm(mdsc, parentino);
303 return PTR_ERR(parent);
305 dout("%s %llx %p: %llx %p -> %llx %p\n", __func__, realm->ino,
306 realm, realm->parent_ino, realm->parent, parentino, parent);
308 list_del_init(&realm->child_item);
309 ceph_put_snap_realm(mdsc, realm->parent);
311 realm->parent_ino = parentino;
312 realm->parent = parent;
313 list_add(&realm->child_item, &parent->children);
318 static int cmpu64_rev(const void *a, const void *b)
320 if (*(u64 *)a < *(u64 *)b)
322 if (*(u64 *)a > *(u64 *)b)
329 * build the snap context for a given realm.
331 static int build_snap_context(struct ceph_snap_realm *realm,
332 struct list_head *realm_queue,
333 struct list_head *dirty_realms)
335 struct ceph_snap_realm *parent = realm->parent;
336 struct ceph_snap_context *snapc;
338 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
341 * build parent context, if it hasn't been built.
342 * conservatively estimate that all parent snaps might be
346 if (!parent->cached_context) {
347 /* add to the queue head */
348 list_add(&parent->rebuild_item, realm_queue);
351 num += parent->cached_context->num_snaps;
354 /* do i actually need to update? not if my context seq
355 matches realm seq, and my parents' does to. (this works
356 because we rebuild_snap_realms() works _downward_ in
357 hierarchy after each update.) */
358 if (realm->cached_context &&
359 realm->cached_context->seq == realm->seq &&
361 realm->cached_context->seq >= parent->cached_context->seq)) {
362 dout("%s %llx %p: %p seq %lld (%u snaps) (unchanged)\n",
363 __func__, realm->ino, realm, realm->cached_context,
364 realm->cached_context->seq,
365 (unsigned int)realm->cached_context->num_snaps);
369 /* alloc new snap context */
371 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
373 snapc = ceph_create_snap_context(num, GFP_NOFS);
377 /* build (reverse sorted) snap vector */
379 snapc->seq = realm->seq;
383 /* include any of parent's snaps occurring _after_ my
384 parent became my parent */
385 for (i = 0; i < parent->cached_context->num_snaps; i++)
386 if (parent->cached_context->snaps[i] >=
388 snapc->snaps[num++] =
389 parent->cached_context->snaps[i];
390 if (parent->cached_context->seq > snapc->seq)
391 snapc->seq = parent->cached_context->seq;
393 memcpy(snapc->snaps + num, realm->snaps,
394 sizeof(u64)*realm->num_snaps);
395 num += realm->num_snaps;
396 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
397 sizeof(u64)*realm->num_prior_parent_snaps);
398 num += realm->num_prior_parent_snaps;
400 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
401 snapc->num_snaps = num;
402 dout("%s %llx %p: %p seq %lld (%u snaps)\n", __func__, realm->ino,
403 realm, snapc, snapc->seq, (unsigned int) snapc->num_snaps);
405 ceph_put_snap_context(realm->cached_context);
406 realm->cached_context = snapc;
407 /* queue realm for cap_snap creation */
408 list_add_tail(&realm->dirty_item, dirty_realms);
413 * if we fail, clear old (incorrect) cached_context... hopefully
414 * we'll have better luck building it later
416 if (realm->cached_context) {
417 ceph_put_snap_context(realm->cached_context);
418 realm->cached_context = NULL;
420 pr_err("%s %llx %p fail %d\n", __func__, realm->ino, realm, err);
425 * rebuild snap context for the given realm and all of its children.
427 static void rebuild_snap_realms(struct ceph_snap_realm *realm,
428 struct list_head *dirty_realms)
430 LIST_HEAD(realm_queue);
434 list_add_tail(&realm->rebuild_item, &realm_queue);
436 while (!list_empty(&realm_queue)) {
437 struct ceph_snap_realm *_realm, *child;
439 _realm = list_first_entry(&realm_queue,
440 struct ceph_snap_realm,
444 * If the last building failed dues to memory
445 * issue, just empty the realm_queue and return
446 * to avoid infinite loop.
449 list_del_init(&_realm->rebuild_item);
453 last = build_snap_context(_realm, &realm_queue, dirty_realms);
454 dout("%s %llx %p, %s\n", __func__, _realm->ino, _realm,
455 last > 0 ? "is deferred" : !last ? "succeeded" : "failed");
457 /* is any child in the list ? */
458 list_for_each_entry(child, &_realm->children, child_item) {
459 if (!list_empty(&child->rebuild_item)) {
466 list_for_each_entry(child, &_realm->children, child_item)
467 list_add_tail(&child->rebuild_item, &realm_queue);
470 /* last == 1 means need to build parent first */
472 list_del_init(&_realm->rebuild_item);
478 * helper to allocate and decode an array of snapids. free prior
481 static int dup_array(u64 **dst, __le64 *src, u32 num)
487 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
490 for (i = 0; i < num; i++)
491 (*dst)[i] = get_unaligned_le64(src + i);
498 static bool has_new_snaps(struct ceph_snap_context *o,
499 struct ceph_snap_context *n)
501 if (n->num_snaps == 0)
503 /* snaps are in descending order */
504 return n->snaps[0] > o->seq;
508 * When a snapshot is applied, the size/mtime inode metadata is queued
509 * in a ceph_cap_snap (one for each snapshot) until writeback
510 * completes and the metadata can be flushed back to the MDS.
512 * However, if a (sync) write is currently in-progress when we apply
513 * the snapshot, we have to wait until the write succeeds or fails
514 * (and a final size/mtime is known). In this case the
515 * cap_snap->writing = 1, and is said to be "pending." When the write
516 * finishes, we __ceph_finish_cap_snap().
518 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
521 static void ceph_queue_cap_snap(struct ceph_inode_info *ci,
522 struct ceph_cap_snap **pcapsnap)
524 struct inode *inode = &ci->vfs_inode;
525 struct ceph_snap_context *old_snapc, *new_snapc;
526 struct ceph_cap_snap *capsnap = *pcapsnap;
527 struct ceph_buffer *old_blob = NULL;
530 spin_lock(&ci->i_ceph_lock);
531 used = __ceph_caps_used(ci);
532 dirty = __ceph_caps_dirty(ci);
534 old_snapc = ci->i_head_snapc;
535 new_snapc = ci->i_snap_realm->cached_context;
538 * If there is a write in progress, treat that as a dirty Fw,
539 * even though it hasn't completed yet; by the time we finish
540 * up this capsnap it will be.
542 if (used & CEPH_CAP_FILE_WR)
543 dirty |= CEPH_CAP_FILE_WR;
545 if (__ceph_have_pending_cap_snap(ci)) {
546 /* there is no point in queuing multiple "pending" cap_snaps,
547 as no new writes are allowed to start when pending, so any
548 writes in progress now were started before the previous
549 cap_snap. lucky us. */
550 dout("%s %p %llx.%llx already pending\n",
551 __func__, inode, ceph_vinop(inode));
554 if (ci->i_wrbuffer_ref_head == 0 &&
555 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
556 dout("%s %p %llx.%llx nothing dirty|writing\n",
557 __func__, inode, ceph_vinop(inode));
564 * There is no need to send FLUSHSNAP message to MDS if there is
565 * no new snapshot. But when there is dirty pages or on-going
566 * writes, we still need to create cap_snap. cap_snap is needed
567 * by the write path and page writeback path.
569 * also see ceph_try_drop_cap_snap()
571 if (has_new_snaps(old_snapc, new_snapc)) {
572 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
573 capsnap->need_flush = true;
575 if (!(used & CEPH_CAP_FILE_WR) &&
576 ci->i_wrbuffer_ref_head == 0) {
577 dout("%s %p %llx.%llx no new_snap|dirty_page|writing\n",
578 __func__, inode, ceph_vinop(inode));
583 dout("%s %p %llx.%llx cap_snap %p queuing under %p %s %s\n",
584 __func__, inode, ceph_vinop(inode), capsnap, old_snapc,
585 ceph_cap_string(dirty), capsnap->need_flush ? "" : "no_flush");
588 capsnap->follows = old_snapc->seq;
589 capsnap->issued = __ceph_caps_issued(ci, NULL);
590 capsnap->dirty = dirty;
592 capsnap->mode = inode->i_mode;
593 capsnap->uid = inode->i_uid;
594 capsnap->gid = inode->i_gid;
596 if (dirty & CEPH_CAP_XATTR_EXCL) {
597 old_blob = __ceph_build_xattrs_blob(ci);
598 capsnap->xattr_blob =
599 ceph_buffer_get(ci->i_xattrs.blob);
600 capsnap->xattr_version = ci->i_xattrs.version;
602 capsnap->xattr_blob = NULL;
603 capsnap->xattr_version = 0;
606 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
608 /* dirty page count moved from _head to this cap_snap;
609 all subsequent writes page dirties occur _after_ this
611 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
612 ci->i_wrbuffer_ref_head = 0;
613 capsnap->context = old_snapc;
614 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
616 if (used & CEPH_CAP_FILE_WR) {
617 dout("%s %p %llx.%llx cap_snap %p snapc %p seq %llu used WR,"
618 " now pending\n", __func__, inode, ceph_vinop(inode),
619 capsnap, old_snapc, old_snapc->seq);
620 capsnap->writing = 1;
622 /* note mtime, size NOW. */
623 __ceph_finish_cap_snap(ci, capsnap);
629 if (ci->i_wrbuffer_ref_head == 0 &&
631 ci->i_dirty_caps == 0 &&
632 ci->i_flushing_caps == 0) {
633 ci->i_head_snapc = NULL;
635 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
636 dout(" new snapc is %p\n", new_snapc);
638 spin_unlock(&ci->i_ceph_lock);
640 ceph_buffer_put(old_blob);
641 ceph_put_snap_context(old_snapc);
645 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
646 * to be used for the snapshot, to be flushed back to the mds.
648 * If capsnap can now be flushed, add to snap_flush list, and return 1.
650 * Caller must hold i_ceph_lock.
652 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
653 struct ceph_cap_snap *capsnap)
655 struct inode *inode = &ci->vfs_inode;
656 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
658 BUG_ON(capsnap->writing);
659 capsnap->size = i_size_read(inode);
660 capsnap->mtime = inode->i_mtime;
661 capsnap->atime = inode->i_atime;
662 capsnap->ctime = inode->i_ctime;
663 capsnap->btime = ci->i_btime;
664 capsnap->change_attr = inode_peek_iversion_raw(inode);
665 capsnap->time_warp_seq = ci->i_time_warp_seq;
666 capsnap->truncate_size = ci->i_truncate_size;
667 capsnap->truncate_seq = ci->i_truncate_seq;
668 if (capsnap->dirty_pages) {
669 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu "
670 "still has %d dirty pages\n", __func__, inode,
671 ceph_vinop(inode), capsnap, capsnap->context,
672 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
673 capsnap->size, capsnap->dirty_pages);
677 /* Fb cap still in use, delay it */
679 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu "
680 "used WRBUFFER, delaying\n", __func__, inode,
681 ceph_vinop(inode), capsnap, capsnap->context,
682 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
684 capsnap->writing = 1;
688 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
689 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu\n",
690 __func__, inode, ceph_vinop(inode), capsnap, capsnap->context,
691 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
694 spin_lock(&mdsc->snap_flush_lock);
695 if (list_empty(&ci->i_snap_flush_item))
696 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
697 spin_unlock(&mdsc->snap_flush_lock);
698 return 1; /* caller may want to ceph_flush_snaps */
702 * Queue cap_snaps for snap writeback for this realm and its children.
703 * Called under snap_rwsem, so realm topology won't change.
705 static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
707 struct ceph_inode_info *ci;
708 struct inode *lastinode = NULL;
709 struct ceph_cap_snap *capsnap = NULL;
711 dout("%s %p %llx inode\n", __func__, realm, realm->ino);
713 spin_lock(&realm->inodes_with_caps_lock);
714 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
715 struct inode *inode = igrab(&ci->vfs_inode);
718 spin_unlock(&realm->inodes_with_caps_lock);
723 * Allocate the capsnap memory outside of ceph_queue_cap_snap()
724 * to reduce very possible but unnecessary frequently memory
725 * allocate/free in this loop.
728 capsnap = kmem_cache_zalloc(ceph_cap_snap_cachep, GFP_NOFS);
730 pr_err("ENOMEM allocating ceph_cap_snap on %p\n",
735 capsnap->cap_flush.is_capsnap = true;
736 refcount_set(&capsnap->nref, 1);
737 INIT_LIST_HEAD(&capsnap->cap_flush.i_list);
738 INIT_LIST_HEAD(&capsnap->cap_flush.g_list);
739 INIT_LIST_HEAD(&capsnap->ci_item);
741 ceph_queue_cap_snap(ci, &capsnap);
742 spin_lock(&realm->inodes_with_caps_lock);
744 spin_unlock(&realm->inodes_with_caps_lock);
748 kmem_cache_free(ceph_cap_snap_cachep, capsnap);
749 dout("%s %p %llx done\n", __func__, realm, realm->ino);
753 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
754 * the snap realm parameters from a given realm and all of its ancestors,
757 * Caller must hold snap_rwsem for write.
759 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
760 void *p, void *e, bool deletion,
761 struct ceph_snap_realm **realm_ret)
763 struct ceph_mds_snap_realm *ri; /* encoded */
764 __le64 *snaps; /* encoded */
765 __le64 *prior_parent_snaps; /* encoded */
766 struct ceph_snap_realm *realm = NULL;
767 struct ceph_snap_realm *first_realm = NULL;
768 struct ceph_snap_realm *realm_to_rebuild = NULL;
771 LIST_HEAD(dirty_realms);
773 lockdep_assert_held_write(&mdsc->snap_rwsem);
775 dout("%s deletion=%d\n", __func__, deletion);
778 ceph_decode_need(&p, e, sizeof(*ri), bad);
781 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
782 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
784 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
785 prior_parent_snaps = p;
786 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
788 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
790 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
792 err = PTR_ERR(realm);
797 /* ensure the parent is correct */
798 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
801 rebuild_snapcs += err;
803 if (le64_to_cpu(ri->seq) > realm->seq) {
804 dout("%s updating %llx %p %lld -> %lld\n", __func__,
805 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
806 /* update realm parameters, snap lists */
807 realm->seq = le64_to_cpu(ri->seq);
808 realm->created = le64_to_cpu(ri->created);
809 realm->parent_since = le64_to_cpu(ri->parent_since);
811 realm->num_snaps = le32_to_cpu(ri->num_snaps);
812 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
816 realm->num_prior_parent_snaps =
817 le32_to_cpu(ri->num_prior_parent_snaps);
818 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
819 realm->num_prior_parent_snaps);
823 if (realm->seq > mdsc->last_snap_seq)
824 mdsc->last_snap_seq = realm->seq;
827 } else if (!realm->cached_context) {
828 dout("%s %llx %p seq %lld new\n", __func__,
829 realm->ino, realm, realm->seq);
832 dout("%s %llx %p seq %lld unchanged\n", __func__,
833 realm->ino, realm, realm->seq);
836 dout("done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
837 realm, rebuild_snapcs, p, e);
840 * this will always track the uppest parent realm from which
841 * we need to rebuild the snapshot contexts _downward_ in
845 realm_to_rebuild = realm;
847 /* rebuild_snapcs when we reach the _end_ (root) of the trace */
848 if (realm_to_rebuild && p >= e)
849 rebuild_snap_realms(realm_to_rebuild, &dirty_realms);
854 ceph_put_snap_realm(mdsc, realm);
860 * queue cap snaps _after_ we've built the new snap contexts,
861 * so that i_head_snapc can be set appropriately.
863 while (!list_empty(&dirty_realms)) {
864 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
866 list_del_init(&realm->dirty_item);
867 queue_realm_cap_snaps(realm);
871 *realm_ret = first_realm;
873 ceph_put_snap_realm(mdsc, first_realm);
875 __cleanup_empty_realms(mdsc);
881 if (realm && !IS_ERR(realm))
882 ceph_put_snap_realm(mdsc, realm);
884 ceph_put_snap_realm(mdsc, first_realm);
885 pr_err("%s error %d\n", __func__, err);
891 * Send any cap_snaps that are queued for flush. Try to carry
892 * s_mutex across multiple snap flushes to avoid locking overhead.
894 * Caller holds no locks.
896 static void flush_snaps(struct ceph_mds_client *mdsc)
898 struct ceph_inode_info *ci;
900 struct ceph_mds_session *session = NULL;
902 dout("%s\n", __func__);
903 spin_lock(&mdsc->snap_flush_lock);
904 while (!list_empty(&mdsc->snap_flush_list)) {
905 ci = list_first_entry(&mdsc->snap_flush_list,
906 struct ceph_inode_info, i_snap_flush_item);
907 inode = &ci->vfs_inode;
909 spin_unlock(&mdsc->snap_flush_lock);
910 ceph_flush_snaps(ci, &session);
912 spin_lock(&mdsc->snap_flush_lock);
914 spin_unlock(&mdsc->snap_flush_lock);
916 ceph_put_mds_session(session);
917 dout("%s done\n", __func__);
921 * ceph_change_snap_realm - change the snap_realm for an inode
922 * @inode: inode to move to new snap realm
923 * @realm: new realm to move inode into (may be NULL)
925 * Detach an inode from its old snaprealm (if any) and attach it to
926 * the new snaprealm (if any). The old snap realm reference held by
927 * the inode is put. If realm is non-NULL, then the caller's reference
928 * to it is taken over by the inode.
930 void ceph_change_snap_realm(struct inode *inode, struct ceph_snap_realm *realm)
932 struct ceph_inode_info *ci = ceph_inode(inode);
933 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
934 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
936 lockdep_assert_held(&ci->i_ceph_lock);
939 spin_lock(&oldrealm->inodes_with_caps_lock);
940 list_del_init(&ci->i_snap_realm_item);
941 if (oldrealm->ino == ci->i_vino.ino)
942 oldrealm->inode = NULL;
943 spin_unlock(&oldrealm->inodes_with_caps_lock);
944 ceph_put_snap_realm(mdsc, oldrealm);
947 ci->i_snap_realm = realm;
950 spin_lock(&realm->inodes_with_caps_lock);
951 list_add(&ci->i_snap_realm_item, &realm->inodes_with_caps);
952 if (realm->ino == ci->i_vino.ino)
953 realm->inode = inode;
954 spin_unlock(&realm->inodes_with_caps_lock);
959 * Handle a snap notification from the MDS.
961 * This can take two basic forms: the simplest is just a snap creation
962 * or deletion notification on an existing realm. This should update the
963 * realm and its children.
965 * The more difficult case is realm creation, due to snap creation at a
966 * new point in the file hierarchy, or due to a rename that moves a file or
967 * directory into another realm.
969 void ceph_handle_snap(struct ceph_mds_client *mdsc,
970 struct ceph_mds_session *session,
971 struct ceph_msg *msg)
973 struct super_block *sb = mdsc->fsc->sb;
974 int mds = session->s_mds;
978 struct ceph_snap_realm *realm = NULL;
979 void *p = msg->front.iov_base;
980 void *e = p + msg->front.iov_len;
981 struct ceph_mds_snap_head *h;
982 int num_split_inos, num_split_realms;
983 __le64 *split_inos = NULL, *split_realms = NULL;
985 int locked_rwsem = 0;
988 if (msg->front.iov_len < sizeof(*h))
991 op = le32_to_cpu(h->op);
992 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
994 num_split_inos = le32_to_cpu(h->num_split_inos);
995 num_split_realms = le32_to_cpu(h->num_split_realms);
996 trace_len = le32_to_cpu(h->trace_len);
999 dout("%s from mds%d op %s split %llx tracelen %d\n", __func__,
1000 mds, ceph_snap_op_name(op), split, trace_len);
1002 mutex_lock(&session->s_mutex);
1003 inc_session_sequence(session);
1004 mutex_unlock(&session->s_mutex);
1006 down_write(&mdsc->snap_rwsem);
1009 if (op == CEPH_SNAP_OP_SPLIT) {
1010 struct ceph_mds_snap_realm *ri;
1013 * A "split" breaks part of an existing realm off into
1014 * a new realm. The MDS provides a list of inodes
1015 * (with caps) and child realms that belong to the new
1019 p += sizeof(u64) * num_split_inos;
1021 p += sizeof(u64) * num_split_realms;
1022 ceph_decode_need(&p, e, sizeof(*ri), bad);
1023 /* we will peek at realm info here, but will _not_
1024 * advance p, as the realm update will occur below in
1025 * ceph_update_snap_trace. */
1028 realm = ceph_lookup_snap_realm(mdsc, split);
1030 realm = ceph_create_snap_realm(mdsc, split);
1035 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
1036 for (i = 0; i < num_split_inos; i++) {
1037 struct ceph_vino vino = {
1038 .ino = le64_to_cpu(split_inos[i]),
1039 .snap = CEPH_NOSNAP,
1041 struct inode *inode = ceph_find_inode(sb, vino);
1042 struct ceph_inode_info *ci;
1046 ci = ceph_inode(inode);
1048 spin_lock(&ci->i_ceph_lock);
1049 if (!ci->i_snap_realm)
1052 * If this inode belongs to a realm that was
1053 * created after our new realm, we experienced
1054 * a race (due to another split notifications
1055 * arriving from a different MDS). So skip
1058 if (ci->i_snap_realm->created >
1059 le64_to_cpu(ri->created)) {
1060 dout(" leaving %p %llx.%llx in newer realm %llx %p\n",
1061 inode, ceph_vinop(inode), ci->i_snap_realm->ino,
1065 dout(" will move %p %llx.%llx to split realm %llx %p\n",
1066 inode, ceph_vinop(inode), realm->ino, realm);
1068 ceph_get_snap_realm(mdsc, realm);
1069 ceph_change_snap_realm(inode, realm);
1070 spin_unlock(&ci->i_ceph_lock);
1075 spin_unlock(&ci->i_ceph_lock);
1079 /* we may have taken some of the old realm's children. */
1080 for (i = 0; i < num_split_realms; i++) {
1081 struct ceph_snap_realm *child =
1082 __lookup_snap_realm(mdsc,
1083 le64_to_cpu(split_realms[i]));
1086 adjust_snap_realm_parent(mdsc, child, realm->ino);
1091 * update using the provided snap trace. if we are deleting a
1092 * snap, we can avoid queueing cap_snaps.
1094 ceph_update_snap_trace(mdsc, p, e,
1095 op == CEPH_SNAP_OP_DESTROY, NULL);
1097 if (op == CEPH_SNAP_OP_SPLIT)
1098 /* we took a reference when we created the realm, above */
1099 ceph_put_snap_realm(mdsc, realm);
1101 __cleanup_empty_realms(mdsc);
1103 up_write(&mdsc->snap_rwsem);
1109 pr_err("%s corrupt snap message from mds%d\n", __func__, mds);
1113 up_write(&mdsc->snap_rwsem);
1117 struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1120 struct ceph_snapid_map *sm, *exist;
1121 struct rb_node **p, *parent;
1125 spin_lock(&mdsc->snapid_map_lock);
1126 p = &mdsc->snapid_map_tree.rb_node;
1128 exist = rb_entry(*p, struct ceph_snapid_map, node);
1129 if (snap > exist->snap) {
1131 } else if (snap < exist->snap) {
1132 p = &(*p)->rb_right;
1134 if (atomic_inc_return(&exist->ref) == 1)
1135 list_del_init(&exist->lru);
1140 spin_unlock(&mdsc->snapid_map_lock);
1142 dout("%s found snapid map %llx -> %x\n", __func__,
1143 exist->snap, exist->dev);
1147 sm = kmalloc(sizeof(*sm), GFP_NOFS);
1151 ret = get_anon_bdev(&sm->dev);
1157 INIT_LIST_HEAD(&sm->lru);
1158 atomic_set(&sm->ref, 1);
1163 p = &mdsc->snapid_map_tree.rb_node;
1164 spin_lock(&mdsc->snapid_map_lock);
1167 exist = rb_entry(*p, struct ceph_snapid_map, node);
1168 if (snap > exist->snap)
1170 else if (snap < exist->snap)
1171 p = &(*p)->rb_right;
1177 if (atomic_inc_return(&exist->ref) == 1)
1178 list_del_init(&exist->lru);
1180 rb_link_node(&sm->node, parent, p);
1181 rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1183 spin_unlock(&mdsc->snapid_map_lock);
1185 free_anon_bdev(sm->dev);
1187 dout("%s found snapid map %llx -> %x\n", __func__,
1188 exist->snap, exist->dev);
1192 dout("%s create snapid map %llx -> %x\n", __func__,
1197 void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1198 struct ceph_snapid_map *sm)
1202 if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1203 if (!RB_EMPTY_NODE(&sm->node)) {
1204 sm->last_used = jiffies;
1205 list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1206 spin_unlock(&mdsc->snapid_map_lock);
1208 /* already cleaned up by
1209 * ceph_cleanup_snapid_map() */
1210 spin_unlock(&mdsc->snapid_map_lock);
1216 void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1218 struct ceph_snapid_map *sm;
1222 spin_lock(&mdsc->snapid_map_lock);
1225 while (!list_empty(&mdsc->snapid_map_lru)) {
1226 sm = list_first_entry(&mdsc->snapid_map_lru,
1227 struct ceph_snapid_map, lru);
1228 if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1231 rb_erase(&sm->node, &mdsc->snapid_map_tree);
1232 list_move(&sm->lru, &to_free);
1234 spin_unlock(&mdsc->snapid_map_lock);
1236 while (!list_empty(&to_free)) {
1237 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1239 dout("trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1240 free_anon_bdev(sm->dev);
1245 void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1247 struct ceph_snapid_map *sm;
1251 spin_lock(&mdsc->snapid_map_lock);
1252 while ((p = rb_first(&mdsc->snapid_map_tree))) {
1253 sm = rb_entry(p, struct ceph_snapid_map, node);
1254 rb_erase(p, &mdsc->snapid_map_tree);
1256 list_move(&sm->lru, &to_free);
1258 spin_unlock(&mdsc->snapid_map_lock);
1260 while (!list_empty(&to_free)) {
1261 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1263 free_anon_bdev(sm->dev);
1264 if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1265 pr_err("snapid map %llx -> %x still in use\n",