Re: [PATCH v2 3/3] xfs: Let the max iomap length be consistent with the writeback code

[Dave Chinner]

On Sun, Oct 06, 2024 at 11:28:49PM +0800, Tang Yizhou wrote:
> From: Tang Yizhou <yizhou.tang@xxxxxxxxxx>
> 
> Since commit 1a12d8bd7b29 ("writeback: scale IO chunk size up to half
> device bandwidth"), macro MAX_WRITEBACK_PAGES has been removed from the
> writeback path. Therefore, the MAX_WRITEBACK_PAGES comments in
> xfs_direct_write_iomap_begin() and xfs_buffered_write_iomap_begin() appear
> outdated.
> 
> In addition, Christoph mentioned that the xfs iomap process should be
> similar to writeback, so xfs_max_map_length() was written following the
> logic of writeback_chunk_size().
> 
> v2: Thanks for Christoph's advice. Resync with the writeback code.
> 
> Signed-off-by: Tang Yizhou <yizhou.tang@xxxxxxxxxx>
> ---
>  fs/fs-writeback.c         |  5 ----
>  fs/xfs/xfs_iomap.c        | 52 ++++++++++++++++++++++++---------------
>  include/linux/writeback.h |  5 ++++
>  3 files changed, 37 insertions(+), 25 deletions(-)
> 
> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> index d8bec3c1bb1f..31c72e207e1b 100644
> --- a/fs/fs-writeback.c
> +++ b/fs/fs-writeback.c
> @@ -31,11 +31,6 @@
>  #include <linux/memcontrol.h>
>  #include "internal.h"
>  
> -/*
> - * 4MB minimal write chunk size
> - */
> -#define MIN_WRITEBACK_PAGES	(4096UL >> (PAGE_SHIFT - 10))
> -
>  /*
>   * Passed into wb_writeback(), essentially a subset of writeback_control
>   */
> diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> index 1e11f48814c0..80f759fa9534 100644
> --- a/fs/xfs/xfs_iomap.c
> +++ b/fs/xfs/xfs_iomap.c
> @@ -4,6 +4,8 @@
>   * Copyright (c) 2016-2018 Christoph Hellwig.
>   * All Rights Reserved.
>   */
> +#include <linux/writeback.h>
> +
>  #include "xfs.h"
>  #include "xfs_fs.h"
>  #include "xfs_shared.h"
> @@ -744,6 +746,34 @@ xfs_ilock_for_iomap(
>  	return 0;
>  }
>  
> +/*
> + * We cap the maximum length we map to a sane size to keep the chunks
> + * of work done where somewhat symmetric with the work writeback does.
> + * This is a completely arbitrary number pulled out of thin air as a
> + * best guess for initial testing.
> + *
> + * Following the logic of writeback_chunk_size(), the length will be
> + * rounded to the nearest 4MB boundary.
> + *
> + * Note that the values needs to be less than 32-bits wide until the
> + * lower level functions are updated.
> + */
> +static loff_t
> +xfs_max_map_length(struct inode *inode, loff_t length)
> +{
> +	struct bdi_writeback *wb;
> +	long pages;
> +
> +	spin_lock(&inode->i_lock);
> +	wb = inode_to_wb(wb);
> +	pages = min(wb->avg_write_bandwidth / 2,
> +		    global_wb_domain.dirty_limit / DIRTY_SCOPE);
> +	spin_unlock(&inode->i_lock);
> +	pages = round_down(pages + MIN_WRITEBACK_PAGES, MIN_WRITEBACK_PAGES);
> +
> +	return min_t(loff_t, length, pages * PAGE_SIZE);
> +}

I think this map size limiting is completely unnecessary for
buffered writeback - buffered writes are throttled against writeback
by balance_dirty_pages(), not by extent allocation size. The size of
the delayed allocation or the overwrite map is largely irrelevant -
we're going to map the entire range during a write, do it just
doesn't matter what size the mapping is...

>  /*
>   * Check that the imap we are going to return to the caller spans the entire
>   * range that the caller requested for the IO.
> @@ -878,16 +908,7 @@ xfs_direct_write_iomap_begin(
>  	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
>  		goto out_unlock;
>  
> -	/*
> -	 * We cap the maximum length we map to a sane size  to keep the chunks
> -	 * of work done where somewhat symmetric with the work writeback does.
> -	 * This is a completely arbitrary number pulled out of thin air as a
> -	 * best guess for initial testing.
> -	 *
> -	 * Note that the values needs to be less than 32-bits wide until the
> -	 * lower level functions are updated.
> -	 */
> -	length = min_t(loff_t, length, 1024 * PAGE_SIZE);
> +	length = xfs_max_map_length(inode, length);
>  	end_fsb = xfs_iomap_end_fsb(mp, offset, length);

And I'd just remove this altogether from the direct IO path. The DIO
code will chain as many bios as it takes to issue Io over the entire
mapping that is returned. Given that buffered writeback has done
arbitrarily large writeback for quite some time now, I don't think
there is any need to limit the bio chain lengths in the DIO path
like this anymore, either.

i.e. I'd be looking at removing the "arbitrary size limit" code, not
making it more complex.

-Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx