Re: [PATCH] iommu/vt-d: avoid sending explicit ATS invalidation request to released device

[Ethan Zhao]

On 2/29/2024 11:31 AM, Ethan Zhao wrote:
> The introduction of per iommu device rbtree also defines the lifetime of
> interoperation between iommu and devices, if the device has been released
> from device rbtree, no need to send ATS invalidation request to it anymore,
> thus avoid the possibility of later ITE fault to be triggered.
>
> This is part of the followup of prior proposed patchset
>
> https://do-db2.lkml.org/lkml/2024/2/22/350
>
> To make sure all the devTLB entries to be invalidated in the device release
> path, do implict invalidation by fapping the E bit of ATS control register.
> see PCIe spec v6.2, sec 10.3.7 implicit invalidation events.
>
> Fixes: 6f7db75e1c46 ("iommu/vt-d: Add second level page table interface")
> Signed-off-by: Ethan Zhao <haifeng.zhao@xxxxxxxxxxxxxxx>
> ---
>   drivers/iommu/intel/iommu.c | 6 ++++++
>   drivers/iommu/intel/pasid.c | 7 +++++++
>   2 files changed, 13 insertions(+)
>
> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
> index 6743fe6c7a36..b85d88ccb4b0 100644
> --- a/drivers/iommu/intel/iommu.c
> +++ b/drivers/iommu/intel/iommu.c
> @@ -1368,6 +1368,12 @@ static void iommu_disable_pci_caps(struct device_domain_info *info)
>   	pdev = to_pci_dev(info->dev);
>   
>   	if (info->ats_enabled) {
> +		pci_disable_ats(pdev);
> +		/*
> +		 * flap the E bit of ATS control register to do implicit
> +		 * ATS invlidation, see PCIe spec v6.2, sec 10.3.7
> +		 */
> +		pci_enable_ats(pdev, VTD_PAGE_SHIFT);
>   		pci_disable_ats(pdev);
>   		info->ats_enabled = 0;
>   		domain_update_iotlb(info->domain);
> diff --git a/drivers/iommu/intel/pasid.c b/drivers/iommu/intel/pasid.c
> index 108158e2b907..5f13e017a12c 100644
> --- a/drivers/iommu/intel/pasid.c
> +++ b/drivers/iommu/intel/pasid.c
> @@ -215,6 +215,13 @@ devtlb_invalidation_with_pasid(struct intel_iommu *iommu,
>   		return;
>   
>   	sid = info->bus << 8 | info->devfn;
> +	/*
> +	 * If device has been released from rbtree, no need to send ATS
> +	 * Invalidation	request anymore, that could cause ITE fault.
> +	 */
> +	if (!device_rbtree_find(iommu, sid))
> +		return;
> +
>   	qdep = info->ats_qdep;
>   	pfsid = info->pfsid;
>   

Given maintainer is going to pick up patchset
https://lore.kernel.org/lkml/2d1788da-521c-4531-a159-81d2fb801d6c@xxxxxxxxxxxxxxx/T/
and this one is mutually exclusive with it, suspend.


Thanks,
Ethan