On 2022-09-08 10:51, Radovanovic, Aleksandar wrote:

[AMD Official Use Only - General]

quoted

-----Original Message-----
From: Marc Zyngier <maz@kernel.org>
Sent: 08 September 2022 09:08
To: Radovanovic, Aleksandar <redacted>
Cc: Jason Gunthorpe <jgg@nvidia.com>; Gupta, Nipun
[off-list ref]; robh+dt@kernel.org;
krzysztof.kozlowski+dt@linaro.org; gregkh@linuxfoundation.org;
rafael@kernel.org; eric.auger@redhat.com; alex.williamson@redhat.com;
cohuck@redhat.com; Gupta, Puneet (DCG-ENG)
[off-list ref]; song.bao.hua@hisilicon.com;
mchehab+huawei@kernel.org; f.fainelli@gmail.com;
jeffrey.l.hugo@gmail.com; saravanak@google.com;
Michael.Srba@seznam.cz; mani@kernel.org; yishaih@nvidia.com;
robin.murphy@arm.com; will@kernel.org; joro@8bytes.org;
masahiroy@kernel.org; ndesaulniers@google.com; linux-arm-
kernel@lists.infradead.org; linux-kbuild@vger.kernel.org; linux-
kernel@vger.kernel.org; devicetree@vger.kernel.org; kvm@vger.kernel.org;
okaya@kernel.org; Anand, Harpreet [off-list ref]; Agarwal,
Nikhil [off-list ref]; Simek, Michal [off-list ref];
git (AMD-Xilinx) [off-list ref]
Subject: Re: [RFC PATCH v3 4/7] bus/cdx: add cdx-MSI domain with gic-its
domain as parent

[CAUTION: External Email]
  
OK, so you definitely need a mapping, but it cannot be a translation, and it
needs to be in all the possible address spaces. OMG.

Could you elaborate why it needs to be in all the possible address spaces? I'm in no way familiar with kernel IOVA allocation, so not sure I understand this requirement. Note that each CDX device will have its own unique StreamID (in general case, equal to DeviceID sent to the GIC), so, from a SMMU perspective, the mapping can be specific to that device. As long as that IOVA is not allocated to any DMA region for _that_ device, things should be OK? But, I appreciate it might not be that simple from a kernel perspective.

That's the point - any device could could have its own mapping, 
therefore that hole has to be punched in *every* mapping that any of 
those devices could use, so that MSI writes don't unexpectedly fault, or 
corrupt memory if that address is free to be used to map a DMA buffer. 
At least the HiSilicon PCI quirk is functionally similar (for slightly 
different underlying reasons) so there's already precedent and an 
example that you can follow to a reasonable degree.

Robin.

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