[PATCH RFC 0/7] PCI/HMAT: Describe P2PDMA reachability and performance

Leon Romanovsky posted 7 patches 1 month, 2 weeks ago
drivers/acpi/numa/Kconfig     |  13 ++
drivers/acpi/numa/Makefile    |   1 +
drivers/acpi/numa/hmat.c      | 245 +++++++++++++++++++++-
drivers/acpi/numa/hmat_test.c | 223 ++++++++++++++++++++
drivers/acpi/numa/hmat_test.h |  17 ++
drivers/pci/Kconfig           |  14 ++
drivers/pci/Makefile          |   1 +
drivers/pci/p2pdma.c          | 300 ++++++++++++++++++++------
drivers/pci/p2pdma_test.c     | 477 ++++++++++++++++++++++++++++++++++++++++++
drivers/pci/pci.h             |  23 ++
include/acpi/actbl1.h         |  27 ++-
include/linux/acpi.h          |  26 +++
12 files changed, 1298 insertions(+), 69 deletions(-)
[PATCH RFC 0/7] PCI/HMAT: Describe P2PDMA reachability and performance
Posted by Leon Romanovsky 1 month, 2 weeks ago
Linux derives P2PDMA reachability from visible PCI topology, ACS
configuration, and a host-bridge whitelist. It cannot describe conditional
or slower cross-host-bridge paths; passthrough can hide the physical peer
relationships.

This RFC series implements the proposed ACPI HMAT Type 3 extension below.
Directional Generic Port matrices describe reachability, ordering, latency,
and bandwidth. Linux uses ordered paths to authorize P2PDMA and their
latency/bandwidth to rank providers. A VMM could translate these matrices
into guest domains; that is outside this series.

  HMAT Type 3 structure

  Field               Offset       Bytes  Meaning
  ---------------------------------------------------------------
  Type:                    0           2  value 3
  Length:                  4           4  entire structure
  Flags:                   8           1  traffic classes
  Data-type:               9           1  metric selector
  Initiators:             12           4  Generic Port count (s)
  Targets:                16           4  Generic Port count (t)
  Base-unit:              24           8  ps or MB/s; non-zero
  Initiator-PDs:          32         4*s  proximity-domain list
  Target-PDs:         32+4*s         4*t  proximity-domain list
  Matrix:         32+4*(s+t)       2*s*t  directional Entry[i][j]
  Reserved: bytes 2-3, 10-11, and 20-23

  Flags: encoding
    bit 0  non-UIO
    bit 1  UIO

  Metrics: selector
    0/1/2  access/read/write latency
    3/4/5  access/read/write bandwidth

  Direction: Initiator PD[i] -- Entry[i][j] metric --> Target PD[j]

----------------------------------------------------------------------
Based on [PATCH v3 00/17] PCI/P2PDMA: Fix ACS egress control handling
https://lore.kernel.org/all/20260811-fix-p2p-acs-v3-0-efc488ee7c03@nvidia.com/

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
Leon Romanovsky (7):
      ACPICA: Define PCIe P2P latency and bandwidth information
      ACPI: HMAT: Make PCIe P2P paths available to consumers
      PCI/P2PDMA: Authorize HMAT-described cross-host-bridge P2P
      PCI/P2PDMA: Prefer providers with better HMAT performance
      ACPI: HMAT: Add KUnit tests for PCIe P2P matrix semantics
      PCI/P2PDMA: Add KUnit tests for HMAT policy boundaries
      PCI/P2PDMA: Add KUnit tests for HMAT provider ranking

 drivers/acpi/numa/Kconfig     |  13 ++
 drivers/acpi/numa/Makefile    |   1 +
 drivers/acpi/numa/hmat.c      | 245 +++++++++++++++++++++-
 drivers/acpi/numa/hmat_test.c | 223 ++++++++++++++++++++
 drivers/acpi/numa/hmat_test.h |  17 ++
 drivers/pci/Kconfig           |  14 ++
 drivers/pci/Makefile          |   1 +
 drivers/pci/p2pdma.c          | 300 ++++++++++++++++++++------
 drivers/pci/p2pdma_test.c     | 477 ++++++++++++++++++++++++++++++++++++++++++
 drivers/pci/pci.h             |  23 ++
 include/acpi/actbl1.h         |  27 ++-
 include/linux/acpi.h          |  26 +++
 12 files changed, 1298 insertions(+), 69 deletions(-)
---
base-commit: bc5261fca94dd94ab557cc0852ce465d9c3ab8b0
change-id: 20260715-hmat-p2p-eb4971a9d731

Best regards,
--  
Leon Romanovsky <leonro@nvidia.com>
Re: [PATCH RFC 0/7] PCI/HMAT: Describe P2PDMA reachability and performance
Posted by Alison Schofield 1 month ago
On Wed, Aug 12, 2026 at 10:47:37PM +0300, Leon Romanovsky wrote:
> Linux derives P2PDMA reachability from visible PCI topology, ACS
> configuration, and a host-bridge whitelist. It cannot describe conditional
> or slower cross-host-bridge paths; passthrough can hide the physical peer
> relationships.
> 
> This RFC series implements the proposed ACPI HMAT Type 3 extension below.

Hi Leon,

Is there an ECN document for review alongside this?
I get that the table below is a snapshot of the expected change, but
would appreciate the long version to read.

Please add linux-cxl to CC list on future postings.

-- Alison


> Directional Generic Port matrices describe reachability, ordering, latency,
> and bandwidth. Linux uses ordered paths to authorize P2PDMA and their
> latency/bandwidth to rank providers. A VMM could translate these matrices
> into guest domains; that is outside this series.
> 
>   HMAT Type 3 structure
> 
>   Field               Offset       Bytes  Meaning
>   ---------------------------------------------------------------
>   Type:                    0           2  value 3
>   Length:                  4           4  entire structure
>   Flags:                   8           1  traffic classes
>   Data-type:               9           1  metric selector
>   Initiators:             12           4  Generic Port count (s)
>   Targets:                16           4  Generic Port count (t)
>   Base-unit:              24           8  ps or MB/s; non-zero
>   Initiator-PDs:          32         4*s  proximity-domain list
>   Target-PDs:         32+4*s         4*t  proximity-domain list
>   Matrix:         32+4*(s+t)       2*s*t  directional Entry[i][j]
>   Reserved: bytes 2-3, 10-11, and 20-23
> 
>   Flags: encoding
>     bit 0  non-UIO
>     bit 1  UIO
> 
>   Metrics: selector
>     0/1/2  access/read/write latency
>     3/4/5  access/read/write bandwidth
> 
>   Direction: Initiator PD[i] -- Entry[i][j] metric --> Target PD[j]
> 
> ----------------------------------------------------------------------
> Based on [PATCH v3 00/17] PCI/P2PDMA: Fix ACS egress control handling
> https://lore.kernel.org/all/20260811-fix-p2p-acs-v3-0-efc488ee7c03@nvidia.com/
> 
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> ---
> Leon Romanovsky (7):
>       ACPICA: Define PCIe P2P latency and bandwidth information
>       ACPI: HMAT: Make PCIe P2P paths available to consumers
>       PCI/P2PDMA: Authorize HMAT-described cross-host-bridge P2P
>       PCI/P2PDMA: Prefer providers with better HMAT performance
>       ACPI: HMAT: Add KUnit tests for PCIe P2P matrix semantics
>       PCI/P2PDMA: Add KUnit tests for HMAT policy boundaries
>       PCI/P2PDMA: Add KUnit tests for HMAT provider ranking
> 
>  drivers/acpi/numa/Kconfig     |  13 ++
>  drivers/acpi/numa/Makefile    |   1 +
>  drivers/acpi/numa/hmat.c      | 245 +++++++++++++++++++++-
>  drivers/acpi/numa/hmat_test.c | 223 ++++++++++++++++++++
>  drivers/acpi/numa/hmat_test.h |  17 ++
>  drivers/pci/Kconfig           |  14 ++
>  drivers/pci/Makefile          |   1 +
>  drivers/pci/p2pdma.c          | 300 ++++++++++++++++++++------
>  drivers/pci/p2pdma_test.c     | 477 ++++++++++++++++++++++++++++++++++++++++++
>  drivers/pci/pci.h             |  23 ++
>  include/acpi/actbl1.h         |  27 ++-
>  include/linux/acpi.h          |  26 +++
>  12 files changed, 1298 insertions(+), 69 deletions(-)
> ---
> base-commit: bc5261fca94dd94ab557cc0852ce465d9c3ab8b0
> change-id: 20260715-hmat-p2p-eb4971a9d731
> 
> Best regards,
> --  
> Leon Romanovsky <leonro@nvidia.com>
>
Re: [PATCH RFC 0/7] PCI/HMAT: Describe P2PDMA reachability and performance
Posted by Leon Romanovsky 1 month ago
On Mon, Aug 24, 2026 at 01:15:34PM -0700, Alison Schofield wrote:
> On Wed, Aug 12, 2026 at 10:47:37PM +0300, Leon Romanovsky wrote:
> > Linux derives P2PDMA reachability from visible PCI topology, ACS
> > configuration, and a host-bridge whitelist. It cannot describe conditional
> > or slower cross-host-bridge paths; passthrough can hide the physical peer
> > relationships.
> > 
> > This RFC series implements the proposed ACPI HMAT Type 3 extension below.
> 
> Hi Leon,
> 
> Is there an ECN document for review alongside this?
> I get that the table below is a snapshot of the expected change, but
> would appreciate the long version to read.

Sure, Mahesh sent me the document, which I promised to convert to the
mailing list format. I will send the official RFC to the linux-acpi ML
slightly later.

Thanks

> 
> Please add linux-cxl to CC list on future postings.
> 
> -- Alison
> 
> 
> > Directional Generic Port matrices describe reachability, ordering, latency,
> > and bandwidth. Linux uses ordered paths to authorize P2PDMA and their
> > latency/bandwidth to rank providers. A VMM could translate these matrices
> > into guest domains; that is outside this series.
> > 
> >   HMAT Type 3 structure
> > 
> >   Field               Offset       Bytes  Meaning
> >   ---------------------------------------------------------------
> >   Type:                    0           2  value 3
> >   Length:                  4           4  entire structure
> >   Flags:                   8           1  traffic classes
> >   Data-type:               9           1  metric selector
> >   Initiators:             12           4  Generic Port count (s)
> >   Targets:                16           4  Generic Port count (t)
> >   Base-unit:              24           8  ps or MB/s; non-zero
> >   Initiator-PDs:          32         4*s  proximity-domain list
> >   Target-PDs:         32+4*s         4*t  proximity-domain list
> >   Matrix:         32+4*(s+t)       2*s*t  directional Entry[i][j]
> >   Reserved: bytes 2-3, 10-11, and 20-23
> > 
> >   Flags: encoding
> >     bit 0  non-UIO
> >     bit 1  UIO
> > 
> >   Metrics: selector
> >     0/1/2  access/read/write latency
> >     3/4/5  access/read/write bandwidth
> > 
> >   Direction: Initiator PD[i] -- Entry[i][j] metric --> Target PD[j]
> > 
> > ----------------------------------------------------------------------
> > Based on [PATCH v3 00/17] PCI/P2PDMA: Fix ACS egress control handling
> > https://lore.kernel.org/all/20260811-fix-p2p-acs-v3-0-efc488ee7c03@nvidia.com/
> > 
> > Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> > ---
> > Leon Romanovsky (7):
> >       ACPICA: Define PCIe P2P latency and bandwidth information
> >       ACPI: HMAT: Make PCIe P2P paths available to consumers
> >       PCI/P2PDMA: Authorize HMAT-described cross-host-bridge P2P
> >       PCI/P2PDMA: Prefer providers with better HMAT performance
> >       ACPI: HMAT: Add KUnit tests for PCIe P2P matrix semantics
> >       PCI/P2PDMA: Add KUnit tests for HMAT policy boundaries
> >       PCI/P2PDMA: Add KUnit tests for HMAT provider ranking
> > 
> >  drivers/acpi/numa/Kconfig     |  13 ++
> >  drivers/acpi/numa/Makefile    |   1 +
> >  drivers/acpi/numa/hmat.c      | 245 +++++++++++++++++++++-
> >  drivers/acpi/numa/hmat_test.c | 223 ++++++++++++++++++++
> >  drivers/acpi/numa/hmat_test.h |  17 ++
> >  drivers/pci/Kconfig           |  14 ++
> >  drivers/pci/Makefile          |   1 +
> >  drivers/pci/p2pdma.c          | 300 ++++++++++++++++++++------
> >  drivers/pci/p2pdma_test.c     | 477 ++++++++++++++++++++++++++++++++++++++++++
> >  drivers/pci/pci.h             |  23 ++
> >  include/acpi/actbl1.h         |  27 ++-
> >  include/linux/acpi.h          |  26 +++
> >  12 files changed, 1298 insertions(+), 69 deletions(-)
> > ---
> > base-commit: bc5261fca94dd94ab557cc0852ce465d9c3ab8b0
> > change-id: 20260715-hmat-p2p-eb4971a9d731
> > 
> > Best regards,
> > --  
> > Leon Romanovsky <leonro@nvidia.com>
> > 
>
Re: [PATCH RFC 0/7] PCI/HMAT: Describe P2PDMA reachability and performance
Posted by Natu, Mahesh 1 month ago
attached


Thank you,

Mahesh



________________________________
From: Schofield, Alison <alison.schofield@intel.com>
Sent: Monday, August 24, 2026 1:15 PM
To: Leon Romanovsky <leon@kernel.org>
Cc: Rafael J. Wysocki <rafael@kernel.org>; Len Brown <lenb@kernel.org>; Saket Dumbre <saket.dumbre@intel.com>; Bjorn Helgaas <bhelgaas@google.com>; Logan Gunthorpe <logang@deltatee.com>; linux-acpi@vger.kernel.org <linux-acpi@vger.kernel.org>; acpica-devel@lists.linux.dev <acpica-devel@lists.linux.dev>; linux-kernel@vger.kernel.org <linux-kernel@vger.kernel.org>; linux-pci@vger.kernel.org <linux-pci@vger.kernel.org>; Lukas Wunner <lukas@wunner.de>; Natu, Mahesh <mahesh.natu@intel.com>
Subject: Re: [PATCH RFC 0/7] PCI/HMAT: Describe P2PDMA reachability and performance

On Wed, Aug 12, 2026 at 10:47:37PM +0300, Leon Romanovsky wrote:
> Linux derives P2PDMA reachability from visible PCI topology, ACS
> configuration, and a host-bridge whitelist. It cannot describe conditional
> or slower cross-host-bridge paths; passthrough can hide the physical peer
> relationships.
>
> This RFC series implements the proposed ACPI HMAT Type 3 extension below.

Hi Leon,

Is there an ECN document for review alongside this?
I get that the table below is a snapshot of the expected change, but
would appreciate the long version to read.

Please add linux-cxl to CC list on future postings.

-- Alison


> Directional Generic Port matrices describe reachability, ordering, latency,
> and bandwidth. Linux uses ordered paths to authorize P2PDMA and their
> latency/bandwidth to rank providers. A VMM could translate these matrices
> into guest domains; that is outside this series.
>
>   HMAT Type 3 structure
>
>   Field               Offset       Bytes  Meaning
>   ---------------------------------------------------------------
>   Type:                    0           2  value 3
>   Length:                  4           4  entire structure
>   Flags:                   8           1  traffic classes
>   Data-type:               9           1  metric selector
>   Initiators:             12           4  Generic Port count (s)
>   Targets:                16           4  Generic Port count (t)
>   Base-unit:              24           8  ps or MB/s; non-zero
>   Initiator-PDs:          32         4*s  proximity-domain list
>   Target-PDs:         32+4*s         4*t  proximity-domain list
>   Matrix:         32+4*(s+t)       2*s*t  directional Entry[i][j]
>   Reserved: bytes 2-3, 10-11, and 20-23
>
>   Flags: encoding
>     bit 0  non-UIO
>     bit 1  UIO
>
>   Metrics: selector
>     0/1/2  access/read/write latency
>     3/4/5  access/read/write bandwidth
>
>   Direction: Initiator PD[i] -- Entry[i][j] metric --> Target PD[j]
>
> ----------------------------------------------------------------------
> Based on [PATCH v3 00/17] PCI/P2PDMA: Fix ACS egress control handling
> https://lore.kernel.org/all/20260811-fix-p2p-acs-v3-0-efc488ee7c03@nvidia.com/
>
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> ---
> Leon Romanovsky (7):
>       ACPICA: Define PCIe P2P latency and bandwidth information
>       ACPI: HMAT: Make PCIe P2P paths available to consumers
>       PCI/P2PDMA: Authorize HMAT-described cross-host-bridge P2P
>       PCI/P2PDMA: Prefer providers with better HMAT performance
>       ACPI: HMAT: Add KUnit tests for PCIe P2P matrix semantics
>       PCI/P2PDMA: Add KUnit tests for HMAT policy boundaries
>       PCI/P2PDMA: Add KUnit tests for HMAT provider ranking
>
>  drivers/acpi/numa/Kconfig     |  13 ++
>  drivers/acpi/numa/Makefile    |   1 +
>  drivers/acpi/numa/hmat.c      | 245 +++++++++++++++++++++-
>  drivers/acpi/numa/hmat_test.c | 223 ++++++++++++++++++++
>  drivers/acpi/numa/hmat_test.h |  17 ++
>  drivers/pci/Kconfig           |  14 ++
>  drivers/pci/Makefile          |   1 +
>  drivers/pci/p2pdma.c          | 300 ++++++++++++++++++++------
>  drivers/pci/p2pdma_test.c     | 477 ++++++++++++++++++++++++++++++++++++++++++
>  drivers/pci/pci.h             |  23 ++
>  include/acpi/actbl1.h         |  27 ++-
>  include/linux/acpi.h          |  26 +++
>  12 files changed, 1298 insertions(+), 69 deletions(-)
> ---
> base-commit: bc5261fca94dd94ab557cc0852ce465d9c3ab8b0
> change-id: 20260715-hmat-p2p-eb4971a9d731
>
> Best regards,
> --
> Leon Romanovsky <leonro@nvidia.com>
>