Re: [PATCH v2 4/7] DMA-API: Add dma_(un)map_resource() documentation

From: "joro@8bytes.org" <joro@8bytes.org>
To: Mark Hounschell <markh@compro.net>
Cc: William Davis <wdavis@nvidia.com>,
	Bjorn Helgaas <bhelgaas@google.com>,
	"iommu@lists.linux-foundation.org"
	<iommu@lists.linux-foundation.org>,
	"linux-pci@vger.kernel.org" <linux-pci@vger.kernel.org>,
	Terence Ripperda <TRipperda@nvidia.com>,
	John Hubbard <jhubbard@nvidia.com>,
	"jglisse@redhat.com" <jglisse@redhat.com>,
	"konrad.wilk@oracle.com" <konrad.wilk@oracle.com>,
	Jonathan Corbet <corbet@lwn.net>,
	"David S. Miller" <davem@davemloft.net>
Subject: Re: [PATCH v2 4/7] DMA-API: Add dma_(un)map_resource() documentation
Date: Fri, 29 May 2015 10:24:44 +0200	[thread overview]
Message-ID: <20150529082444.GB13130@8bytes.org> (raw)
In-Reply-To: <555CDD6F.40304@compro.net>

On Wed, May 20, 2015 at 03:15:59PM -0400, Mark Hounschell wrote:
> On 05/20/2015 01:30 PM, William Davis wrote:
> >In an IOMMU environment, the DMA ops would be one of the IOMMU
> >implementations, so these APIs would create a mapping for the peer device
> >resource, even if it's on the same bus. Would a transaction targeting that
> >mapping be forwarded upstream until it hits an IOMMU, which would then send
> >the translated request back downstream? Or is my understanding of this
> >configuration incorrect?
> >
> 
> It's my understanding of the IOMMU that is lacking here. I have no
> idea if that is actually what would happen. Does it?

On the x86 side at least the IOMMU drivers enable ACS when available,
forcing all perr-to-peer DMA between PCIe devices to go through the
IOMMU. This only works when the peers are in different iommu groups,
though.

	Joerg