Re: [PATCH] btrfs: balance RAID1/RAID10 mirror selection

[waxhead <waxhead@dirtcellar.net>]

louis@waffle.tech wrote:
> Balance RAID1/RAID10 mirror selection via plain round-robin scheduling. This should roughly double throughput for large reads.
> 
> Signed-off-by: Louis Jencka <louis@waffle.tech>
> ---
>   fs/btrfs/volumes.c | 6 +++++-
>   1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
> index 58b9c419a..45c581d46 100644
> --- a/fs/btrfs/volumes.c
> +++ b/fs/btrfs/volumes.c
> @@ -333,6 +333,9 @@ struct list_head * __attribute_const__ btrfs_get_fs_uuids(void)
>   	return &fs_uuids;
>   }
>   
> +/* Used for round-robin balancing RAID1/RAID10 reads. */
> +atomic_t rr_counter = ATOMIC_INIT(0);
> +
>   /*
>    * alloc_fs_devices - allocate struct btrfs_fs_devices
>    * @fsid:		if not NULL, copy the UUID to fs_devices::fsid
> @@ -5482,7 +5485,8 @@ static int find_live_mirror(struct btrfs_fs_info *fs_info,
>   	else
>   		num_stripes = map->num_stripes;
>   
> -	preferred_mirror = first + current->pid % num_stripes;
> +	preferred_mirror = first +
> +	    (unsigned)atomic_inc_return(&rr_counter) % num_stripes;
>   
>   	if (dev_replace_is_ongoing &&
>   	    fs_info->dev_replace.cont_reading_from_srcdev_mode ==
> 

I am just a regular user of BTRFS, but the btrfs_fs_info structure 
contains a substructure (dirty_metadata_bytes).

Assuming that writing data leads to a dirty metadata bytes being set to 
non-zero would not something along the lines of this potentially be better?

if(fs_info->dirty_metadata_bytes) atomic_inc_return(&rr_counter);
prefered_mirror = first + (rr_counter+current->pid) % num_stripes;

My knowledge of BTRFS code is near zero and I should know better than 
posting "code" here, but I can't resist getting my point across. There 
must be a simple way to improve slightly over a regular rr scheme.