Re: [PATCH 01/10] swiotlb-xen: use vmalloc_to_page on vmalloc virt addresses

[Julien Grall]

Hi Stefano,

On 22/05/2020 04:54, Stefano Stabellini wrote:
> On Thu, 21 May 2020, Julien Grall wrote:
> > Hi,
> >
> > On 21/05/2020 00:45, Stefano Stabellini wrote:
> > > From: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
> > >
> > > Don't just assume that virt_to_page works on all virtual addresses.
> > > Instead add a is_vmalloc_addr check and use vmalloc_to_page on vmalloc
> > > virt addresses.
> >
> > Can you provide an example where swiotlb is used with vmalloc()?
>
> The issue was reported here happening on the Rasperry Pi 4:
> https://marc.info/?l=xen-devel&m=158862573216800

Thanks, it would be good if the commit message contains a bit more details.

> If you are asking where in the Linux codebase the vmalloc is happening
> specifically, I don't know for sure, my information is limited to the
> stack trace that you see in the link (I don't have a Rasperry Pi 4 yet
> but I shall have one soon.)

Looking at the code there is a comment in xen_swiotlb_alloc_coherent() 
suggesting that xen_alloc_coherent_pages() may return an ioremap'ped 
region on Arm. So it feels to me that commit 
b877ac9815a8fe7e5f6d7fdde3dc34652408840a "xen/swiotlb: remember having 
called xen_create_contiguous_region()" has always been broken on Arm.

As an aside, your commit message also suggests this is an issue for 
every virtual address used in swiotlb. But only the virt_to_page() call 
in xen_swiotlb_free_coherent() is modified. Is it intended? If yes, I 
think you want to update your commit message.

> > > Signed-off-by: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
> > > Signed-off-by: Stefano Stabellini <stefano.stabellini@xxxxxxxxxx>
> > > ---
> > >    drivers/xen/swiotlb-xen.c | 5 ++++-
> > >    1 file changed, 4 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
> > > index b6d27762c6f8..a42129cba36e 100644
> > > --- a/drivers/xen/swiotlb-xen.c
> > > +++ b/drivers/xen/swiotlb-xen.c
> > > @@ -335,6 +335,7 @@ xen_swiotlb_free_coherent(struct device *hwdev, size_t
> > > size, void *vaddr,
> > >    	int order = get_order(size);
> > >    	phys_addr_t phys;
> > >    	u64 dma_mask = DMA_BIT_MASK(32);
> > > +	struct page *pg;
> > >      	if (hwdev && hwdev->coherent_dma_mask)
> > >    		dma_mask = hwdev->coherent_dma_mask;
> > > @@ -346,9 +347,11 @@ xen_swiotlb_free_coherent(struct device *hwdev, size_t
> > > size, void *vaddr,
> > >    	/* Convert the size to actually allocated. */
> > >    	size = 1UL << (order + XEN_PAGE_SHIFT);
> > >    +	pg = is_vmalloc_addr(vaddr) ? vmalloc_to_page(vaddr) :
> > > +				      virt_to_page(vaddr);
> >
> > Common DMA code seems to protect this check with CONFIG_DMA_REMAP. Is it
> > something we want to do it here as well? Or is there any other condition where
> > vmalloc can happen?
>
> I can see it in dma_direct_free_pages:
>
> 	if (IS_ENABLED(CONFIG_DMA_REMAP) && is_vmalloc_addr(cpu_addr))
> 		vunmap(cpu_addr);
>
> I wonder why the common DMA code does that. is_vmalloc_addr should work
> regardless of CONFIG_DMA_REMAP. Maybe just for efficiency?
is_vmalloc_addr() doesn't looks very expensive (although it is not an 
inline function). So I am not sure the reason behind it. It might be 
worth asking the author of the config.

Cheers,

--
Julien Grall