Re: [PATCH 2/2] swiotlb: Enforce page alignment in swiotlb_alloc()

[Will Deacon]

On Mon, Jan 29, 2024 at 07:08:53AM +0100, Christoph Hellwig wrote:
> On Fri, Jan 26, 2024 at 03:19:56PM +0000, Will Deacon wrote:
> > diff --git a/kernel/dma/swiotlb.c b/kernel/dma/swiotlb.c
> > index 25febb9e670c..92433ea9f2d2 100644
> > --- a/kernel/dma/swiotlb.c
> > +++ b/kernel/dma/swiotlb.c
> > @@ -1647,6 +1647,12 @@ struct page *swiotlb_alloc(struct device *dev, size_t size)
> >  		return NULL;
> >  
> >  	tlb_addr = slot_addr(pool->start, index);
> > +	if (!PAGE_ALIGNED(tlb_addr)) {
> > +		dev_WARN_ONCE(dev, 1, "Cannot return 'struct page *' for non page-aligned swiotlb addr 0x%pa.\n",
> > +			      &tlb_addr);
> > +		swiotlb_release_slots(dev, tlb_addr);
> > +		return NULL;
> > +	}
> >  
> >  	return pfn_to_page(PFN_DOWN(tlb_addr));
> 
> So PFN_DOWN aligns the address and thus per se converting the unaligned
> address isn't a problem.

Hmm, I'm not sure I follow why it isn't a problem. If the first 2KiB slot
of the 4KiB page has already been allocated to somebody else, isn't it a
big problem to align down like that? Maybe I should word the warning
message a bit better -- how about:

  "Cannot allocate pages from non page-aligned swiotlb addr 0x%pa.\n"

?

> That being said swiotlb obviously should never
> allocate unaligned addresses, but the placement of this check feels
> odd to me.  Also because it only catches swiotlb_alloc and not the
> map side.
> 
> Maybe just throw a WARN_ON_ONCE into slot_addr() ?

Everything is slot-aligned, so I don't think slot_addr() can detect
this. I put the check in swiotlb_alloc() because I think that's the only
place where we assume that a slot address is page-aligned. I don't think
the map path particularly cares, but if you prefer to have the warning
there too then I think we'd have to stick it at the end of
swiotlb_search_pool_area() (effectively just checking that the returned
slot is consistent with the 'alloc_align_mask' parameter).

Will