Re: [PATCH v2 1/2] xen/swiotlb: add alignment check for dma buffers

["Thierry Escande" <thierry.escande@xxxxxxxxxx>]

Hi,

On 16/09/2024 08:47, Juergen Gross wrote:
> diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
> index 35155258a7e2..ddf5b1df632e 100644
> --- a/drivers/xen/swiotlb-xen.c
> +++ b/drivers/xen/swiotlb-xen.c
> @@ -78,9 +78,15 @@ static inline int 
> range_straddles_page_boundary(phys_addr_t p, size_t size)
>  {
>       unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p);
>       unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size);
> +     phys_addr_t algn = 1ULL << (get_order(size) + PAGE_SHIFT);
>  
>       next_bfn = pfn_to_bfn(xen_pfn);
>  
> +     /* If buffer is physically aligned, ensure DMA alignment. */
> +     if (IS_ALIGNED(p, algn) &&
> +         !IS_ALIGNED(next_bfn << XEN_PAGE_SHIFT, algn))
> +             return 1;

There is a regression in the mpt3sas driver because of this change.
When, in a dom0, this driver creates its DMA pool at probe time and
calls dma_pool_zalloc() (see [1]), the call to
range_straddles_page_boundary() (from xen_swiotlb_alloc_coherent())
returns 1 because of the alignment checks added by this patch. Then the
call to xen_create_contiguous_region() in xen_swiotlb_alloc_coherent()
fails because the passed size order is too big (> MAX_CONTIG_ORDER).
This driver sets the pool allocation block size to 2.3+ MBytes.

>From previous discussions on the v1 patch, these checks are not
necessary from xen_swiotlb_alloc_coherent() that already ensures
alignment, right?

[1]
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/scsi/mpt3sas/mpt3sas_base.c?h=v6.12.1#n6227

Regards,
Thierry