diff mbox series

[v4,1/2] btrfs: warn on invalid slot in tree mod log rewind

Message ID 742a0f5e01de22aae0ada9504834b0e5e3177349.1685645613.git.boris@bur.io (mailing list archive)
State New, archived
Headers show
Series btrfs: fix logical_to_ino panic in btrfs_map_bio | expand

Commit Message

Boris Burkov June 1, 2023, 6:55 p.m. UTC
The way that tree mod log tracks the ultimate length of the eb, the
variable 'n', eventually turns up the correct value, but at intermediate
steps during the rewind, n can be inaccurate as a representation of the
end of the eb. For example, it doesn't get updated on move rewinds, and
it does get updated for add/remove in the middle of the eb.

To detect cases with invalid moves, introduce a separate variable called
max_slot which tries to track the maximum valid slot in the rewind eb.
We can then warn if we do a move whose src range goes beyond the max
valid slot.

There is a commented caveat that it is possible to have this value be an
overestimate due to the challenge of properly handling 'add' operations
in the middle of the eb, but in practice it doesn't cause enough of a
problem to throw out the max idea in favor of tracking every valid slot.

Signed-off-by: Boris Burkov <boris@bur.io>
---
 fs/btrfs/tree-mod-log.c | 41 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 41 insertions(+)

Comments

Filipe Manana June 2, 2023, 10:09 a.m. UTC | #1
On Thu, Jun 1, 2023 at 7:55 PM Boris Burkov <boris@bur.io> wrote:
>
> The way that tree mod log tracks the ultimate length of the eb, the
> variable 'n', eventually turns up the correct value, but at intermediate
> steps during the rewind, n can be inaccurate as a representation of the
> end of the eb. For example, it doesn't get updated on move rewinds, and
> it does get updated for add/remove in the middle of the eb.
>
> To detect cases with invalid moves, introduce a separate variable called
> max_slot which tries to track the maximum valid slot in the rewind eb.
> We can then warn if we do a move whose src range goes beyond the max
> valid slot.
>
> There is a commented caveat that it is possible to have this value be an
> overestimate due to the challenge of properly handling 'add' operations
> in the middle of the eb, but in practice it doesn't cause enough of a
> problem to throw out the max idea in favor of tracking every valid slot.
>
> Signed-off-by: Boris Burkov <boris@bur.io>

Reviewed-by: Filipe Manana <fdmanana@suse.com>

Looks good, thanks.

> ---
>  fs/btrfs/tree-mod-log.c | 41 +++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 41 insertions(+)
>
> diff --git a/fs/btrfs/tree-mod-log.c b/fs/btrfs/tree-mod-log.c
> index a555baa0143a..157e1a0efab8 100644
> --- a/fs/btrfs/tree-mod-log.c
> +++ b/fs/btrfs/tree-mod-log.c
> @@ -664,10 +664,27 @@ static void tree_mod_log_rewind(struct btrfs_fs_info *fs_info,
>         unsigned long o_dst;
>         unsigned long o_src;
>         unsigned long p_size = sizeof(struct btrfs_key_ptr);
> +       /*
> +        * max_slot tracks the maximum valid slot of the rewind eb at every
> +        * step of the rewind. This is in contrast with 'n' which eventually
> +        * matches the number of items, but can be wrong during moves or if
> +        * removes overlap on already valid slots (which is probably separately
> +        * a bug). We do this to validate the offsets of memmoves for rewinding
> +        * moves and detect invalid memmoves.
> +        *
> +        * Since a rewind eb can start empty, max_slot is a signed integer with
> +        * a special meaning for -1, which is that no slot is valid to move out
> +        * of. Any other negative value is invalid.
> +        */
> +       int max_slot;
> +       int move_src_end_slot;
> +       int move_dst_end_slot;
>
>         n = btrfs_header_nritems(eb);
> +       max_slot = n - 1;
>         read_lock(&fs_info->tree_mod_log_lock);
>         while (tm && tm->seq >= time_seq) {
> +               ASSERT(max_slot >= -1);
>                 /*
>                  * All the operations are recorded with the operator used for
>                  * the modification. As we're going backwards, we do the
> @@ -684,6 +701,8 @@ static void tree_mod_log_rewind(struct btrfs_fs_info *fs_info,
>                         btrfs_set_node_ptr_generation(eb, tm->slot,
>                                                       tm->generation);
>                         n++;
> +                       if (tm->slot > max_slot)
> +                               max_slot = tm->slot;
>                         break;
>                 case BTRFS_MOD_LOG_KEY_REPLACE:
>                         BUG_ON(tm->slot >= n);
> @@ -693,14 +712,36 @@ static void tree_mod_log_rewind(struct btrfs_fs_info *fs_info,
>                                                       tm->generation);
>                         break;
>                 case BTRFS_MOD_LOG_KEY_ADD:
> +                       /*
> +                        * It is possible we could have already removed keys behind the known
> +                        * max slot, so this will be an overestimate. In practice, the copy
> +                        * operation inserts them in increasing order, and overestimating just
> +                        * means we miss some warnings, so it's OK. It isn't worth carefully
> +                        * tracking the full array of valid slots to check against when moving.
> +                        */
> +                       if (tm->slot == max_slot)
> +                               max_slot--;
>                         /* if a move operation is needed it's in the log */
>                         n--;
>                         break;
>                 case BTRFS_MOD_LOG_MOVE_KEYS:
> +                       ASSERT(tm->move.nr_items > 0);
> +                       move_src_end_slot = tm->move.dst_slot + tm->move.nr_items - 1;
> +                       move_dst_end_slot = tm->slot + tm->move.nr_items - 1;
>                         o_dst = btrfs_node_key_ptr_offset(eb, tm->slot);
>                         o_src = btrfs_node_key_ptr_offset(eb, tm->move.dst_slot);
> +                       if (WARN_ON(move_src_end_slot > max_slot ||
> +                                   tm->move.nr_items <= 0)) {
> +                               btrfs_warn(fs_info,
> +                                          "Move from invalid tree mod log slot eb %llu slot %d dst_slot %d nr_items %d seq %llu n %u max_slot %d\n",
> +                                          eb->start, tm->slot,
> +                                          tm->move.dst_slot, tm->move.nr_items,
> +                                          tm->seq, n, max_slot);
> +
> +                       }
>                         memmove_extent_buffer(eb, o_dst, o_src,
>                                               tm->move.nr_items * p_size);
> +                       max_slot = move_dst_end_slot;
>                         break;
>                 case BTRFS_MOD_LOG_ROOT_REPLACE:
>                         /*
> --
> 2.40.1
>
diff mbox series

Patch

diff --git a/fs/btrfs/tree-mod-log.c b/fs/btrfs/tree-mod-log.c
index a555baa0143a..157e1a0efab8 100644
--- a/fs/btrfs/tree-mod-log.c
+++ b/fs/btrfs/tree-mod-log.c
@@ -664,10 +664,27 @@  static void tree_mod_log_rewind(struct btrfs_fs_info *fs_info,
 	unsigned long o_dst;
 	unsigned long o_src;
 	unsigned long p_size = sizeof(struct btrfs_key_ptr);
+	/*
+	 * max_slot tracks the maximum valid slot of the rewind eb at every
+	 * step of the rewind. This is in contrast with 'n' which eventually
+	 * matches the number of items, but can be wrong during moves or if
+	 * removes overlap on already valid slots (which is probably separately
+	 * a bug). We do this to validate the offsets of memmoves for rewinding
+	 * moves and detect invalid memmoves.
+	 *
+	 * Since a rewind eb can start empty, max_slot is a signed integer with
+	 * a special meaning for -1, which is that no slot is valid to move out
+	 * of. Any other negative value is invalid.
+	 */
+	int max_slot;
+	int move_src_end_slot;
+	int move_dst_end_slot;
 
 	n = btrfs_header_nritems(eb);
+	max_slot = n - 1;
 	read_lock(&fs_info->tree_mod_log_lock);
 	while (tm && tm->seq >= time_seq) {
+		ASSERT(max_slot >= -1);
 		/*
 		 * All the operations are recorded with the operator used for
 		 * the modification. As we're going backwards, we do the
@@ -684,6 +701,8 @@  static void tree_mod_log_rewind(struct btrfs_fs_info *fs_info,
 			btrfs_set_node_ptr_generation(eb, tm->slot,
 						      tm->generation);
 			n++;
+			if (tm->slot > max_slot)
+				max_slot = tm->slot;
 			break;
 		case BTRFS_MOD_LOG_KEY_REPLACE:
 			BUG_ON(tm->slot >= n);
@@ -693,14 +712,36 @@  static void tree_mod_log_rewind(struct btrfs_fs_info *fs_info,
 						      tm->generation);
 			break;
 		case BTRFS_MOD_LOG_KEY_ADD:
+			/*
+			 * It is possible we could have already removed keys behind the known
+			 * max slot, so this will be an overestimate. In practice, the copy
+			 * operation inserts them in increasing order, and overestimating just
+			 * means we miss some warnings, so it's OK. It isn't worth carefully
+			 * tracking the full array of valid slots to check against when moving.
+			 */
+			if (tm->slot == max_slot)
+				max_slot--;
 			/* if a move operation is needed it's in the log */
 			n--;
 			break;
 		case BTRFS_MOD_LOG_MOVE_KEYS:
+			ASSERT(tm->move.nr_items > 0);
+			move_src_end_slot = tm->move.dst_slot + tm->move.nr_items - 1;
+			move_dst_end_slot = tm->slot + tm->move.nr_items - 1;
 			o_dst = btrfs_node_key_ptr_offset(eb, tm->slot);
 			o_src = btrfs_node_key_ptr_offset(eb, tm->move.dst_slot);
+			if (WARN_ON(move_src_end_slot > max_slot ||
+				    tm->move.nr_items <= 0)) {
+				btrfs_warn(fs_info,
+					   "Move from invalid tree mod log slot eb %llu slot %d dst_slot %d nr_items %d seq %llu n %u max_slot %d\n",
+					   eb->start, tm->slot,
+					   tm->move.dst_slot, tm->move.nr_items,
+					   tm->seq, n, max_slot);
+
+			}
 			memmove_extent_buffer(eb, o_dst, o_src,
 					      tm->move.nr_items * p_size);
+			max_slot = move_dst_end_slot;
 			break;
 		case BTRFS_MOD_LOG_ROOT_REPLACE:
 			/*