Re: [PATCH 1/2] ext4: avoid RWM atomic in EXT4_MB_GRP_TEST_AND_SET_READ

[Jan Kara]

On Thu 21-05-26 14:59:28, Bohdan Trach wrote:
> EXT4_MB_GRP_TEST_AND_SET_READ uses test_and_set_bit function which
> issues an atomic write. This can cause high overhead due to cache
> contention when multiple threads iterate over groups in a tight loop,
> as is the case for ext4_mb_prefetch(). We have seen this to be a
> problem for Kunpeng 920b CPUs which uses a single ARM LSE instruction
> for this purpose.
> 
> This change significantly reduces costs of fallocate() operations which
> trigger linear group scans on large multicore machines where
> test_and_set_bit issues an atomic write operation unconditionally.
> 
> Signed-off-by: Bohdan Trach <bohdan.trach@xxxxxxxxxxxxxxx>
...
> diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
> index 56b82d4a15d7..0713207811a6 100644
> --- a/fs/ext4/ext4.h
> +++ b/fs/ext4/ext4.h
> @@ -3551,7 +3551,17 @@ struct ext4_group_info {
>  #define EXT4_MB_GRP_CLEAR_TRIMMED(grp)	\
>  	(clear_bit(EXT4_GROUP_INFO_WAS_TRIMMED_BIT, &((grp)->bb_state)))
>  #define EXT4_MB_GRP_TEST_AND_SET_READ(grp)	\
> -	(test_and_set_bit(EXT4_GROUP_INFO_BBITMAP_READ_BIT, &((grp)->bb_state)))
> +	(ext4_mb_grp_test_and_set_read((grp)))
> +
> +static inline int ext4_mb_grp_test_and_set_read(struct ext4_group_info *grp)
> +{
> +	int r = test_bit_acquire(EXT4_GROUP_INFO_BBITMAP_READ_BIT, &grp->bb_state);
> +
> +	if (!r)
> +		return test_and_set_bit(EXT4_GROUP_INFO_BBITMAP_READ_BIT, &grp->bb_state);
> +	else
> +		return r;
> +}

Good idea but do we really need the 'acquire' barrier here? I don't see
anything that would really need this so I think
EXT4_MB_GRP_TEST_AND_SET_READ() can be just:

test_bit(EXT4_GROUP_INFO_BBITMAP_READ_BIT, &grp->bb_state) || \
  test_and_set_bit(EXT4_GROUP_INFO_BBITMAP_READ_BIT, &grp->bb_state)

or am I missing something?

								Honza
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR