On 18.08.21 г. 12:15, Wentao_Liang wrote:
> In line 2955 (#1), "btrfs_put_block_group(cache);" drops the reference to
> cache and may cause the cache to be released. However, in line 3014, the
> cache is dropped again by the same put function (#4). Double putting the
> cache can lead to an incorrect reference count.
>
> Furthermore, according to the definition of btrfs_put_block_group() in fs/
> btrfs/block-group.c, if the reference count of the cache is one at the
> first put, it will be freed by kfree(). Using it again may result in the
> use-after-free flaw. In fact, after the first put (line 2955), the cache
> is also accessed in a few places (#2, #3), e.g., lines 2967, 2973, 2974,
> ….
>
> We believe that the first put of the cache is unnecessary (#1).
> We can fix the above bugs by removing the redundant
> "btrfs_put_block_group(cache);" in line 2955 (#1).
>
> 2951 if (!list_empty(&cache->io_list)) {
> ...
> 2955 btrfs_put_block_group(cache);
> //#1 the first drop to cache (unnecessary)
> ...
> 2957 }
> ...
> 2967 cache_save_setup(cache, trans, path); //#2 use the cache
> ...
> 2972 //#3 use the cache several times
>
> 2973 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
> 2974 cache->io_ctl.inode = NULL;
> 2975 ret = btrfs_write_out_cache(trans, cache, path);
> 2976 if (ret == 0 && cache->io_ctl.inode) {
> 2977 num_started++;
> 2978 should_put = 0;
> 2979 list_add_tail(&cache->io_list, io);
> 2980 } else {
> ...
> 2985 ret = 0;
> 2986 }
> 2987 }
> 2988 if (!ret) {
> 2989 ret = update_block_group_item(trans, path, cache);
> ...
> 3003 if (ret == -ENOENT) {
> ...
> 3006 ret = update_block_group_item(trans, path, cache);
> 3007 }
> ...
> 3010 }
> 3011
> ...
> 3013 if (should_put)
> 3014 btrfs_put_block_group(cache);
> //#4 the second drop to cache
>
> Signed-off-by: Wentao_Liang <[email protected]>
Apart form the patch being buggy you have not demonstrated why doing 2
put block groups is actually given that there are invariant that
guarantee bg will have at least 2 refs held. So it seems you have
produced the patch without considering the big picture of how btrfs'
block group state machine works.