On Wed, Oct 07, 2026 at 04:46:58PM +0300, Leon Romanovsky wrote:
On Tue, Oct 06, 2026 at 04:32:05PM -0500, Bjorn Helgaas wrote:
On Thu, Oct 01, 2026 at 02:55:13PM +0300, Leon Romanovsky wrote:
From: Leon Romanovsky leonro@nvidia.com
P2PDMA documentation describes ACS controls as path-wide, although Request and Completion controls apply to different transaction directions and only affect peer-versus-upstream decisions at the path divergence.
Document the fixed client and provider roles, the divergence port checked for each TLP direction, and the conservative handling of unreadable ACS state. Clarify which controls disable_acs_redir changes.
Reviewed-by: Logan Gunthorpe logang@deltatee.com Tested-by: Tushar Dave tdave@nvidia.com Signed-off-by: Leon Romanovsky leonro@nvidia.com
Documentation/admin-guide/kernel-parameters.txt | 15 +++++++++------ Documentation/driver-api/pci/p2pdma.rst | 13 +++++++++++++ 2 files changed, 22 insertions(+), 6 deletions(-)
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt index 68647ff4bdd2..bc83e07dd5fc 100644 --- a/Documentation/admin-guide/kernel-parameters.txt +++ b/Documentation/admin-guide/kernel-parameters.txt @@ -5291,12 +5291,15 @@ Kernel parameters disable_acs_redir=<pci_dev>[; ...] Specify one or more PCI devices (in the format specified above) separated by semicolons.
Each device specified will have the PCI ACSredirect capabilities forced off which willallow P2P traffic between devices throughbridges without forcing it upstream. Note:this removes isolation between devices andmay put more devices in an IOMMU group.
Each device specified will have the PCI ACS P2PRequest Redirect, Completion Redirect, and EgressControl features forced off. This may allow P2Ptraffic through bridges that would otherwise beredirected upstream. This may allow P2P trafficthrough bridges that would otherwise be redirectedupstream and thus this removes isolation betweendevices and may cause affected devices to sharean IOMMU group.diff --git a/Documentation/driver-api/pci/p2pdma.rst b/Documentation/driver-api/pci/p2pdma.rst index 80f8fec9b0e9..42b18610bf7d 100644 --- a/Documentation/driver-api/pci/p2pdma.rst +++ b/Documentation/driver-api/pci/p2pdma.rst @@ -15,6 +15,19 @@ then based on the ACS settings the transaction can route entirely within the PCIe hierarchy and never reach the root port. The kernel will evaluate the PCIe topology and always permit P2P in these well-defined cases. +The client remains the PCIe requester when it reads or writes provider memory. +Where the paths diverge, the kernel therefore evaluates P2P Request Redirect +and Egress Control on the client-side port, and P2P Completion Redirect on the +provider-side port for completions from a read. An enabled Egress Control is +conservatively treated as a Request redirect.
Does "client remains" imply that the client can also be a Completer in other circumstances? I'd like to use PCIe terms when possible since we're talking about PCIe ACS controls.
s/requester/Requester/ since it's defined by PCIe spec. s/completions/Completions/ similarly. s/Request redirect/Request Redirect/ to match spec.
s/and Egress Control/and P2P Egress Control/ to match spec and RR and CR mentions here. The second "Egress Control" matches the "Request Redirect" so I think that's fine.
I think "where the paths diverge" means the point where a port decides whether to route a TLP up through its Upstream Port (or to the RC) or back down via a sibling Downstream Port?
Yes.
Am I right in thinking that there may be several such points e.g., a Request is routed back downstream by any Downstream Port where P2P Request Redirect is not enabled?
I think this "paths diverge" needs some background when it's introduced, e.g. something like this (fix or reword as necessary):
If the Requester and Completer are below the same Root Port, the path taken by Requests depends on ACS settings of the Downstream Ports above the Requester. If ACS P2P Request Redirect is enabled in all of them, the Request is routed upstream; otherwise it will be routed back downstream via the Downstream Port leading to the Completer. The path of Completions likewise depends on ACS P2P Completion Redirect in the Downstream Ports above the Completer.
I'm not sure p2pdma.rst includes the context to interpret "divergence" yet. I think the code comments in [4/18] might also need a little more context about divergence.
IIUC, this divergence point is basically a static point where the ACS controls being applied makes a routing difference.
Right. Previously, we did not account for ACS bits, so any "junction" automatically caused P2P to be considered unsupported.
+Below the divergence, the route toward the other branch is already upstream, +so those P2P redirect controls do not affect it. Redirect controls for the +reverse transaction directions do not affect the mapping. P2P DMA is routed +through the host bridge when either applicable port redirects. If an ACS +Control register cannot be read, P2P DMA is rejected because the kernel cannot +establish a usable route.
Similarly, I don't think "the divergence" is defined yet here. I *think* it means the lowest Switch or Root Port shared between Requester and Completer. That's the first point where a TLP could be either routed back downstream or upstream.
I think a TLP could also be routed back downstream if a higher Switch Downstream Port or the Root Port had Request/Completion Redirect cleared and lower Switch Downstream Ports had it set.
This evaluation assumes clients issue strictly ordered Requests carrying an Untranslated address. Its result is not defined when clients use Relaxed Ordering or issue ATS-translated Requests because those TLP attributes can
-- 2.55.0