[PATCH] mm/swap: submit the last readahead batch before unplugging

[PATCH] mm/swap: submit the last readahead batch before unplugging

[Alexandre Ghiti]

Since block swap I/O is batched in struct swap_iocb, swap_add_folio()
only appends a folio to the pending batch when its slot directly follows
the previous one, and submits the pending batch otherwise. Batches
submitted inside the readahead loop go through the plug, where the block
layer can still merge them, but the final batch is submitted after
blk_finish_plug() and always becomes a separate request.

So submit the final batch before finishing the plug.

Fixes: dda8fb68b590 ("mm/swap: also use struct swap_iocb for block I/O")
Signed-off-by: Alexandre Ghiti <alex@ghiti.fr>
---
 mm/swap_state.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/swap_state.c b/mm/swap_state.c
index 2475ba29126d..ebef568cd62e 100644
--- a/mm/swap_state.c
+++ b/mm/swap_state.c
@@ -906,8 +906,8 @@ struct folio *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
 			continue;
 		folio_put(folio);
 	}
-	blk_finish_plug(&plug);
 	swap_read_submit(&ctx);
+	blk_finish_plug(&plug);
 skip:
 	return swap_cache_read_folio_sync(entry, gfp_mask, mpol, ilx);
 }
@@ -1019,8 +1019,8 @@ static struct folio *swap_vma_readahead(swp_entry_t targ_entry, gfp_t gfp_mask,
 	}
 	if (pte)
 		pte_unmap(pte);
-	blk_finish_plug(&plug);
 	swap_read_submit(&ctx);
+	blk_finish_plug(&plug);
 skip:
 	/* The folio was likely read above, so no need for plugging here */
 	return swap_cache_read_folio_sync(targ_entry, gfp_mask, mpol, targ_ilx);
-- 
2.53.0-Meta

Re: [PATCH] mm/swap: submit the last readahead batch before unplugging

[Joshua Hahn]

On Thu,  1 Oct 2026 10:57:12 +0200 Alexandre Ghiti <alex@ghiti.fr> wrote:

> Since block swap I/O is batched in struct swap_iocb, swap_add_folio()
> only appends a folio to the pending batch when its slot directly follows
> the previous one, and submits the pending batch otherwise. Batches
> submitted inside the readahead loop go through the plug, where the block
> layer can still merge them, but the final batch is submitted after
> blk_finish_plug() and always becomes a separate request.
> 
> So submit the final batch before finishing the plug.

Hi Alex,

I read through the code and this looks very reasonable to me.

Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>

> Fixes: dda8fb68b590 ("mm/swap: also use struct swap_iocb for block I/O")

My only question is whether we need the fixes tag, it seems like
dda8fb68b590 is still in mm-stable and not merged to linus yet.

Andrew, what do we usually do in these cases? I am wondering if it will
just become folded into the original commit or if this should still be
standalone.

Thank you both! Have a great day : -)
Joshua

> Signed-off-by: Alexandre Ghiti <alex@ghiti.fr>
> ---
>  mm/swap_state.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/swap_state.c b/mm/swap_state.c
> index 2475ba29126d..ebef568cd62e 100644
> --- a/mm/swap_state.c
> +++ b/mm/swap_state.c
> @@ -906,8 +906,8 @@ struct folio *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
>  			continue;
>  		folio_put(folio);
>  	}
> -	blk_finish_plug(&plug);
>  	swap_read_submit(&ctx);
> +	blk_finish_plug(&plug);
>  skip:
>  	return swap_cache_read_folio_sync(entry, gfp_mask, mpol, ilx);
>  }
> @@ -1019,8 +1019,8 @@ static struct folio *swap_vma_readahead(swp_entry_t targ_entry, gfp_t gfp_mask,
>  	}
>  	if (pte)
>  		pte_unmap(pte);
> -	blk_finish_plug(&plug);
>  	swap_read_submit(&ctx);
> +	blk_finish_plug(&plug);
>  skip:
>  	/* The folio was likely read above, so no need for plugging here */
>  	return swap_cache_read_folio_sync(targ_entry, gfp_mask, mpol, targ_ilx);
> -- 
> 2.53.0-Meta

Re: [PATCH] mm/swap: submit the last readahead batch before unplugging

[Andrew Morton]

On Thu,  1 Oct 2026 02:52:33 -0700 Joshua Hahn <joshua.hahnjy@gmail.com> wrote:

> On Thu,  1 Oct 2026 10:57:12 +0200 Alexandre Ghiti <alex@ghiti.fr> wrote:
> 
> > Since block swap I/O is batched in struct swap_iocb, swap_add_folio()
> > only appends a folio to the pending batch when its slot directly follows
> > the previous one, and submits the pending batch otherwise. Batches
> > submitted inside the readahead loop go through the plug, where the block
> > layer can still merge them, but the final batch is submitted after
> > blk_finish_plug() and always becomes a separate request.
> > 
> > So submit the final batch before finishing the plug.
> 
> Hi Alex,
> 
> I read through the code and this looks very reasonable to me.
> 
> Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>

Thanks.

> > Fixes: dda8fb68b590 ("mm/swap: also use struct swap_iocb for block I/O")
> 
> My only question is whether we need the fixes tag, it seems like
> dda8fb68b590 is still in mm-stable and not merged to linus yet.
> 
> Andrew, what do we usually do in these cases? I am wondering if it will
> just become folded into the original commit or if this should still be
> standalone.

dda8fb68b590 was added to 7.3.rc1 during the most recent merge window.

So the Fixes: is appropriate - if someone chooses to merge dda8fb68b590
into their downstream kernel we're telling them "hey, you need this one
as well".