Re: [PATCH v2] readahead: Correct the start and size in ondemand_readahead()

[Mike Kravetz <mike.kravetz@oracle.com>]

On 06/28/23 12:43, Yin Fengwei wrote:
> The commit
> 9425c591e06a ("page cache: fix page_cache_next/prev_miss off by one")
> updated the page_cache_next_miss() to return the index beyond
> range.
> 
> But it breaks the start/size of ra in ondemand_readahead() because
> the offset by one is accumulated to readahead_index. As a consequence,
> not best readahead order is picked.
> 
> Tracing of the order parameter of filemap_alloc_folio() showed:
>      page order    : count     distribution
>         0          : 892073   |                                        |
>         1          : 0        |                                        |
>         2          : 65120457 |****************************************|
>         3          : 32914005 |********************                    |
>         4          : 33020991 |********************                    |
> with 9425c591e06a9.
> 
> With parent commit:
>      page order    : count     distribution
>         0          : 3417288  |****                                    |
>         1          : 0        |                                        |
>         2          : 877012   |*                                       |
>         3          : 288      |                                        |
>         4          : 5607522  |*******                                 |
>         5          : 29974228 |****************************************|
> 
> Fix the issue by removing the offset by one when page_cache_next_miss()
> returns no gaps in the range.
> 
> After the fix:
>     page order     : count     distribution
>         0          : 2598561  |***                                     |
>         1          : 0        |                                        |
>         2          : 687739   |                                        |
>         3          : 288      |                                        |
>         4          : 207210   |                                        |
>         5          : 32628260 |****************************************|
> 

Thank you for your detailed analysis!

When the regression was initially discovered, I sent a patch to revert
commit 9425c591e06a.  Andrew has picked up this change.  And, Andrew has
also picked up this patch.

I have not verified yet, but I suspect that this patch is going to cause
a regression because it depends on the behavior of page_cache_next_miss
in 9425c591e06a which has been reverted.

Sorry for the delay in responding as I was traveling.
-- 
Mike Kravetz



> Reported-by: kernel test robot <oliver.sang@intel.com>
> Closes: https://lore.kernel.org/oe-lkp/202306211346.1e9ff03e-oliver.sang@intel.com
> Fixes: 9425c591e06a ("page cache: fix page_cache_next/prev_miss off by one")
> Signed-off-by: Yin Fengwei <fengwei.yin@intel.com>
> ---
> Changes from v1:
>   - only removing offset by one when there is no gaps found by
>     page_cache_next_miss()
>   - Update commit message to include the histogram of page order
>     after fix
> 
>  mm/readahead.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/readahead.c b/mm/readahead.c
> index 47afbca1d122..a93af773686f 100644
> --- a/mm/readahead.c
> +++ b/mm/readahead.c
> @@ -614,9 +614,17 @@ static void ondemand_readahead(struct readahead_control *ractl,
>  				max_pages);
>  		rcu_read_unlock();
>  
> -		if (!start || start - index > max_pages)
> +		if (!start || start - index - 1 > max_pages)
>  			return;
>  
> +		/*
> +		 * If no gaps in the range, page_cache_next_miss() returns
> +		 * index beyond range. Adjust it back to make sure
> +		 * ractl->_index is updated correctly later.
> +		 */
> +		if ((start - index - 1) == max_pages)
> +			start--;
> +
>  		ra->start = start;
>  		ra->size = start - index;	/* old async_size */
>  		ra->size += req_size;
> -- 
> 2.39.2
>