Re: [PATCH] null_blk: set dma alignment to logical block size

[Keith Busch <kbusch@kernel.org>]

On Wed, Oct 29, 2025 at 03:32:30PM +0000, Hans Holmberg wrote:
> Hmm, the data is actually copied nullb->dev->blocksize sized chunks
> 
> see copy_to_nullb:
> 
> temp = min_t(size_t, nullb->dev->blocksize, n - count);

That's just saying blocksize is the largest it will transfer. It can be
smaller if there's less data than a block size in the bvec.

> ..
> memcpy_page(t_page->page, offset, source, off + count, temp);
> 
> Just to verify that nullblk was missbehaving I enabled debugging
> for the copy and hit a bug_on in memcpy_page:
> 
> VM_BUG_ON(dst_off + len > PAGE_SIZE || src_off + len > PAGE_SIZE);
> 
> (This was a 4k block size nullblk with the default 511 byte dma limit)

Oh, I see. It's that the source buffer with its weird offsets that don't
align can cause a problem. This should fix it:

---
diff --git a/drivers/block/null_blk/main.c b/drivers/block/null_blk/main.c
index f982027e8c858..385fe6523ec12 100644
--- a/drivers/block/null_blk/main.c
+++ b/drivers/block/null_blk/main.c
@@ -1143,6 +1143,7 @@ static int copy_to_nullb(struct nullb *nullb, struct page *source,
 			null_make_cache_space(nullb, PAGE_SIZE);
 
 		offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
+		temp = min_t(size_t, temp, PAGE_SIZE - offset);
 		t_page = null_insert_page(nullb, sector,
 			!null_cache_active(nullb) || is_fua);
 		if (!t_page)
@@ -1172,6 +1173,7 @@ static int copy_from_nullb(struct nullb *nullb, struct page *dest,
 		temp = min_t(size_t, nullb->dev->blocksize, n - count);
 
 		offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
+		temp = min_t(size_t, temp, PAGE_SIZE - offset);
 		t_page = null_lookup_page(nullb, sector, false,
 			!null_cache_active(nullb));
 
--

After this, you can set dma_alignment to 1.