NVIDIA: SAUCE: iommu/arm-smmu-v3: Skip non-PASID devices in SVA conte…

[jamieNguyenNVIDIA]

…xt descriptor updates

On systems with mixed PASID support within an iommu group (i.e. a PASID-capable GPU sharing a group with PCIe bridges that lack PASID capability), arm_smmu_update_ctx_desc_devices() iterates all masters in the domain and calls arm_smmu_write_ctx_desc() for each one. For non-PASID-capable devices with s1cdmax=0, any SSID >= 1 triggers a WARN_ON(ssid >= (1 << cd_table->s1cdmax)) in arm_smmu_write_ctx_desc().

This occurs during GPU process teardown when the MMU notifier release path (uvm_release -> mmput -> exit_mmap -> arm_smmu_mm_release) writes a quiet context descriptor at the process's PASID to all devices in the domain.

Skip the CD write for non-PASID-capable devices when the SSID is not IOMMU_NO_PASID, since these devices were never programmed with that SSID. SSID 0 (IOMMU_NO_PASID) writes are always allowed through, as they are needed by the arm_smmu_share_asid() ASID reassignment path.

Suggested-by: Nicolin Chen nicolinc@nvidia.com

LP: https://bugs.launchpad.net/ubuntu/+source/linux-nvidia-6.8/+bug/2144814

[clsotog]

Acked-by: Carol L Soto <csoto@nvidia.com>

[nirmoy]

Acked-by: Nirmoy Das <nirmoyd@nvidia.com>

[jamieNguyenNVIDIA]

I've been able to verify the fix on a GB300 system.

Setup: Applied pci=config_acs= kernel parameter to recreate the mixed-PASID iommu groups (6 non-PASID PCIe bridges sharing a group with each PASID-capable GPU).

Reproducer: A minimal CUDA program that allocates UVM managed memory and exits, triggering the uvm_release → mmput → exit_mmap → __mmu_notifier_release → arm_smmu_mm_release teardown path.

#include <cuda_runtime.h>
#include <stdio.h>
int main() {
    void *p;
    cudaError_t err = cudaMallocManaged(&p, 4096);
    if (err != cudaSuccess) {
        printf("cudaMallocManaged failed: %s\n", cudaGetErrorString(err));
        return 1;
    }
    printf("UVM allocated, exiting to trigger teardown\n");
    cudaFree(p);
    return 0;
}

Without fix (6.8.0-1049-nvidia-64k): 6 warnings per run -- one for each non-PASID bridge in the GPU's iommu group, exactly matching the customer's report.

With fix: Zero warnings. Same config_acs parameter, same iommu group layout, same test.

[jamieNguyenNVIDIA]

PR sent to Canonical