Re: [PATCH 1/2] ext4: mballoc - prefetching for bitmaps

["Theodore Y. Ts'o" <tytso@mit.edu>]

On Fri, May 15, 2020 at 10:07:06AM +0000, Alex Zhuravlev wrote:
> --- a/fs/ext4/mballoc.c
> +++ b/fs/ext4/mballoc.c
> @@ -2127,6 +2127,96 @@ static int ext4_mb_good_group(struct ext4_allocation_context *ac,
>  	return 0;
>  }
>  
> +/*
> + * each allocation context (i.e. a thread doing allocation) has own
> + * sliding prefetch window of @s_mb_prefetch size which starts at the
> + * very first goal and moves ahead of scaning.
> + * a side effect is that subsequent allocations will likely find
> + * the bitmaps in cache or at least in-flight.
> + */
> +static void
> +ext4_mb_prefetch(struct ext4_allocation_context *ac,
> +		    ext4_group_t start)
> +{
> +	struct super_block *sb = ac->ac_sb;
> +	ext4_group_t ngroups = ext4_get_groups_count(sb);
> +	struct ext4_sb_info *sbi = EXT4_SB(sb);
> +	struct ext4_group_info *grp;
> +	ext4_group_t nr, group = start;
> +	struct buffer_head *bh;
> +
> +	/* limit prefetching at cr=0, otherwise mballoc can
> +	 * spend a lot of time loading imperfect groups */
> +	if (ac->ac_criteria < 2 && ac->ac_prefetch_ios >= sbi->s_mb_prefetch_limit)
> +		return;
> +
> +	/* batch prefetching to get few READs in flight */
> +	nr = ac->ac_prefetch - group;
> +	if (ac->ac_prefetch < group)
> +		/* wrapped to the first groups */
> +		nr += ngroups;
> +	if (nr > 0)I
> +		return;
> +	BUG_ON(nr < 0);

What are you trying to do here?  If nr > 0, we return; if nr < 0, we
BUG() --- but nr is an unsigned int, so we never can trigger --- which
was the warning reported by the kbuild test bot.  So we will only get
past this point if ac_prefetch == group.  But ac_prefetch appears to
be the last group that we prefetched, so it's not clear that the logic
is correct here.

						- Ted