Re: [PATCH V2] arm: xen: mm: use __GPF_DMA32 for arm64

From: Juergen Gross <jgross@suse.com>
To: Christoph Hellwig <hch@infradead.org>,
	Stefano Stabellini <sstabellini@kernel.org>
Cc: Peng Fan <peng.fan@nxp.com>, "will@kernel.org" <will@kernel.org>,
	Catalin Marinas <Catalin.Marinas@arm.com>,
	"linux@armlinux.org.uk" <linux@armlinux.org.uk>,
	Julien Grall <Julien.Grall@arm.com>,
	dl-linux-imx <linux-imx@nxp.com>,
	"xen-devel@lists.xenproject.org" <xen-devel@lists.xenproject.org>,
	boris.ostrovsky@oracle.com, nd <nd@arm.com>,
	Robin Murphy <Robin.Murphy@arm.com>,
	"linux-arm-kernel@lists.infradead.org"
	<linux-arm-kernel@lists.infradead.org>
Subject: Re: [PATCH V2] arm: xen: mm: use __GPF_DMA32 for arm64
Date: Tue, 3 Sep 2019 13:48:33 +0200	[thread overview]
Message-ID: <8e81051f-ad50-7b4e-2e88-c685ade6b240@suse.com> (raw)
In-Reply-To: <20190902155732.GA8311@infradead.org>

On 02.09.19 17:57, Christoph Hellwig wrote:
> On Fri, Aug 30, 2019 at 07:40:42PM -0700, Stefano Stabellini wrote:
>> + Juergen, Boris
>>
>> On Fri, 30 Aug 2019, Christoph Hellwig wrote:
>>> Can we take a step back and figure out what we want to do here?
>>>
>>> AFAICS this function allocates memory for the swiotlb-xen buffer,
>>> and that means it must be <= 32-bit addressable to satisfy the DMA API
>>> guarantees.  That means we generally want to use GFP_DMA32 everywhere
>>> that exists, but on systems with odd zones we might want to dip into
>>> GFP_DMA.  This also means swiotlb-xen doesn't actually do the right
>>> thing on x86 at the moment.  So shouldn't we just have one common
>>> routine in swiotlb-xen.c that checks if we have CONFIG_ZONE_DMA32
>>> set, then try GFP_DMA32, and if not check if CONFIG_ZONE_DMA is set
>>> and then try that, else default to GFP_KERNEL?
>>
>> Yes, for ARM/ARM64 it makes a lot of sense given that dom0 is 1:1 mapped
>> (pseudo-physical == physical).  I'll let Juergen and Boris comment on
>> the x86 side of things, but on x86 PV Dom0 is not 1:1 mapped so
>> GFP_DMA32 is probably not meaningful.
> 
> But is it actually harmful?  If the GFP_DMA32 doesn't hurt we could
> just use it there.  Or if that seems to ugly we can make the dma
> flags dependents on a XEN_1TO1_MAPPED config option set by arm/arm64.
> 

I'd rather have it only if needed. Especially on X86 PV dom0 I'd like to
avoid GFP_DMA32 as memory below 4GB (guest physical) might be rather
scarce in some configurations.

I think X86 PVH dom0 should need GFP_DMA32, too, as the limit is related
to the address as communicated to the device (before being translated by
the IOMMU), right? This would mean on a X86 kernel configured to support
PV and PVH the test for setting GFP_DMA32 can't depend on a config
option alone, it needs to be dynamic.

BTW, for PV guests the DMA address width is handled via
xen_create_contiguous_region().

Juergen

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel