[00/10] PCI/ATS: Device-tree support and other improvements

Message

Jean-Philippe Brucker Feb. 13, 2020, 4:50 p.m. UTC

Enable ATS on device-tree based systems, and factor the common ATS
enablement checks into pci_enable_ats().

ATS support in PCIe endpoints is discovered through the ATS capability,
but there is no common method for discovering whether the host bridge
supports ATS. Each vendor provides their own ACPI method:
* DMAR (Intel) reports ATS support per domain or per root port.
* IVRS (AMD) reports negative ATS support for a range of devices.
* IORT (ARM) reports ATS support for a root complex.

In my opinion it's important that we only enable ATS if the host bridge
supports it, but I don't think a finer granularity is useful. If the
host bridge ignores the Address Translation field of TLP headers, it
could in theory treat a Translation Request as a Memory Read Request,
and a Translated Request as a normal memory access, which could result
in silent memory corruption.

Patches 1-3 add a device-tree property that declares ATS support in host
controllers. We only add it to the generic host, but the property can
easily be reused by other PCI hosts. Alternatively, the host drivers can
directly set the new ats_supported property of the host bridge
introduced in patch 1.

Patch 4 uses the new ats_supported host bridge property for IORT. Patch
9 removes the old method that set a flag in each endpoint's fwspec.

Patches 5-8 put all checks required for enabling ATS in common, along
with the new host bridge check.

Jean-Philippe Brucker (11):
  dt-bindings: PCI: generic: Add ats-supported property
  PCI: Add ats_supported host bridge flag
  PCI: OF: Check whether the host bridge supports ATS
  ACPI/IORT: Check ATS capability in root complex node
  PCI/ATS: Gather checks into pci_ats_supported()
  iommu/amd: Use pci_ats_supported()
  iommu/arm-smmu-v3: Use pci_ats_supported()
  iommu/vt-d: Use pci_ats_supported()
  ACPI/IORT: Drop ATS fwspec flag
  arm64: dts: fast models: Enable PCIe ATS for Base RevC FVP
  Documentation: Generalize the "pci=noats" boot parameter

 .../admin-guide/kernel-parameters.txt         |  4 +-
 .../bindings/pci/host-generic-pci.yaml        |  6 +++
 arch/arm64/boot/dts/arm/fvp-base-revc.dts     |  1 +
 drivers/acpi/arm64/iort.c                     | 38 +++++++++++++------
 drivers/acpi/pci_root.c                       |  3 ++
 drivers/iommu/amd_iommu.c                     | 12 ++----
 drivers/iommu/arm-smmu-v3.c                   | 18 ++-------
 drivers/iommu/intel-iommu.c                   |  9 ++---
 drivers/pci/ats.c                             | 30 ++++++++++++++-
 drivers/pci/controller/pci-host-common.c      |  1 +
 drivers/pci/of.c                              |  9 +++++
 drivers/pci/probe.c                           |  7 ++++
 include/linux/acpi_iort.h                     |  8 ++++
 include/linux/iommu.h                         |  4 --
 include/linux/of_pci.h                        |  3 ++
 include/linux/pci-ats.h                       |  3 ++
 include/linux/pci.h                           |  1 +
 17 files changed, 110 insertions(+), 47 deletions(-)

Comments

Joerg Roedel Feb. 19, 2020, 10:42 a.m. UTC | #1

Hi Jean-Philippe,


On Thu, Feb 13, 2020 at 05:50:38PM +0100, Jean-Philippe Brucker wrote:
> Jean-Philippe Brucker (11):
>   dt-bindings: PCI: generic: Add ats-supported property
>   PCI: Add ats_supported host bridge flag
>   PCI: OF: Check whether the host bridge supports ATS
>   ACPI/IORT: Check ATS capability in root complex node
>   PCI/ATS: Gather checks into pci_ats_supported()
>   iommu/amd: Use pci_ats_supported()
>   iommu/arm-smmu-v3: Use pci_ats_supported()
>   iommu/vt-d: Use pci_ats_supported()
>   ACPI/IORT: Drop ATS fwspec flag
>   arm64: dts: fast models: Enable PCIe ATS for Base RevC FVP
>   Documentation: Generalize the "pci=noats" boot parameter 

Nice patch-set! Thanks for the consolidation work. I can take it into
the iommu-tree, given that the non-iommu parts get the necessary Acks.


Regards,

	Joerg

Hanjun Guo March 6, 2020, 9:32 a.m. UTC | #2

On 2020/2/14 0:50, Jean-Philippe Brucker wrote:
> Enable ATS on device-tree based systems, and factor the common ATS
> enablement checks into pci_enable_ats().
> 
> ATS support in PCIe endpoints is discovered through the ATS capability,
> but there is no common method for discovering whether the host bridge
> supports ATS. Each vendor provides their own ACPI method:
> * DMAR (Intel) reports ATS support per domain or per root port.
> * IVRS (AMD) reports negative ATS support for a range of devices.
> * IORT (ARM) reports ATS support for a root complex.

Tested this patch set against 5.6-rc2 on a Kunpeng920 ARM server,
I just confirmed that this patch set didn't break anything in
my box with ACPI booting, PCI devices work as expected, FWIW,

Tested-by: Hanjun Guo <[e-mail]>

Thanks
Hanjun