Skip to content

Conversation

jgunthorpe
Copy link
Member

For some time now modern multi-NIC servers now have very complex topology. Often with NICs, GPUs and NVMe devices that are topologically co-located. These systems tend to come with specialized ACS requirements for PCI Peer to Peer, for instance ACS disable or ACS setup specially for translated traffic.

NVIDIA's latest systems have a novel PCI multipath system that requires special asymmetric ACS.

Introduce a tool to help users configure the ACS on such systems. The tool will be able to parse the PCI topology and identify the topological features then generate the require ACS settings.

Modern kernels support the config_acs kernel command line parameter to allow fine grained settings so the correct ACS for the topology can be fed into Grub and to the kernel command line to configure it at boot

The tool has four functions:
topo - Print out the topology from the RDMA perspective. Indicate what
devices are P2P connected to the NIC.
write-grub-acs - Emit the config_acs kernel command line parameter for
the required ACS configuration
setpci-acs - Use setpci after booting to set the required ACS
configuration. This is not recommended but provided to help
legacy systems without config_acs.
check - Read the live ACS settings and compare them to the required
configuration

This initial version supports two NVIDIA platforms. There is an expectation it will grow to more broadly support more common topologies as well.

For some time now modern multi-NIC servers now have very complex
topology. Often with NICs, GPUs and NVMe devices that are topologically
co-located. These systems tend to come with specialized ACS requirements
for PCI Peer to Peer, for instance ACS disable or ACS setup specially for
translated traffic.

NVIDIA's latest systems have a novel PCI multipath system that requires
special asymmetric ACS.

Introduce a tool to help users configure the ACS on such systems. The tool
will be able to parse the PCI topology and identify the topological
features then generate the require ACS settings.

Modern kernels support the config_acs kernel command line parameter to
allow fine grained settings so the correct ACS for the topology can be fed
into Grub and to the kernel command line to configure it at boot

The tool has four functions:
 topo - Print out the topology from the RDMA perspective. Indicate what
        devices are P2P connected to the NIC.
 write-grub-acs - Emit the config_acs kernel command line parameter for
                  the required ACS configuration
 setpci-acs - Use setpci after booting to set the required ACS
              configuration. This is not recommended but provided to help
              legacy systems without config_acs.
 check - Read the live ACS settings and compare them to the required
         configuration

This initial version supports two NVIDIA platforms. There is an
expectation it will grow to more broadly support more common topologies as
well.

Signed-off-by: Jason Gunthorpe <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants