[PATCH] drm/amdkfd: add ACPI SRAT parsing for topology

[Eric Huang]

Like I answer Oak's question,
"For GCD parsing, the relation of GCD to CCD is defined by AMD, generic 
parsing in srat.c is considering a GCD as a new numa node which is not 
suitable for our need."

GCD's pxm domain will get a wrong numa node which may be bigger than CCD 
domains, so I have to do a sanity check to correct it.

Regards,
Eric

On 2021-05-04 3:46 a.m., Lazar, Lijo wrote:
> [AMD Public Use]
>
>> *numa_node > max_pxm
> Why numa node number is compared to a proximity domain? Since you are already using pxm_to_node() API, assume that should take care.
>
> That also will avoid parsing ACPI_SRAT_TYPE_CPU_AFFINITY structs.
>
> Thanks,
> Lijo
>
>
> -----Original Message-----
> From: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> On Behalf Of Eric Huang
> Sent: Wednesday, April 28, 2021 8:42 PM
> To: amd-gfx at lists.freedesktop.org
> Cc: Huang, JinHuiEric <JinHuiEric.Huang at amd.com>
> Subject: [PATCH] drm/amdkfd: add ACPI SRAT parsing for topology
>
> In NPS4 BIOS we need to find the closest numa node when creating topology io link between cpu and gpu, if PCI driver doesn't set it.
>
> Signed-off-by: Eric Huang <jinhuieric.huang at amd.com>
> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 94 ++++++++++++++++++++++++++-
>   1 file changed, 91 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
> index 38d45711675f..57518136c7d7 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
> @@ -1759,6 +1759,87 @@ static int kfd_fill_gpu_memory_affinity(int *avail_size,
>   	return 0;
>   }
>   
> +#ifdef CONFIG_ACPI
> +static void kfd_find_numa_node_in_srat(struct kfd_dev *kdev,
> +		int *numa_node)
> +{
> +	struct acpi_table_header *table_header = NULL;
> +	struct acpi_subtable_header *sub_header = NULL;
> +	unsigned long table_end, subtable_len;
> +	u32 pci_id = pci_domain_nr(kdev->pdev->bus) << 16 |
> +			pci_dev_id(kdev->pdev);
> +	u32 bdf;
> +	acpi_status status;
> +	struct acpi_srat_cpu_affinity *cpu;
> +	struct acpi_srat_generic_affinity *gpu;
> +	int pxm = 0, max_pxm = 0;
> +	bool found = false;
> +
> +	/* Fetch the SRAT table from ACPI */
> +	status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header);
> +	if (status == AE_NOT_FOUND) {
> +		pr_warn("SRAT table not found\n");
> +		return;
> +	} else if (ACPI_FAILURE(status)) {
> +		const char *err = acpi_format_exception(status);
> +		pr_err("SRAT table error: %s\n", err);
> +		return;
> +	}
> +
> +	table_end = (unsigned long)table_header + table_header->length;
> +
> +	/* Parse all entries looking for a match. */
> +
> +	sub_header = (struct acpi_subtable_header *)
> +			((unsigned long)table_header +
> +			sizeof(struct acpi_table_srat));
> +	subtable_len = sub_header->length;
> +
> +	while (((unsigned long)sub_header) + subtable_len  < table_end) {
> +		/*
> +		 * If length is 0, break from this loop to avoid
> +		 * infinite loop.
> +		 */
> +		if (subtable_len == 0) {
> +			pr_err("SRAT invalid zero length\n");
> +			break;
> +		}
> +
> +		switch (sub_header->type) {
> +		case ACPI_SRAT_TYPE_CPU_AFFINITY:
> +			cpu = (struct acpi_srat_cpu_affinity *)sub_header;
> +			pxm = *((u32 *)cpu->proximity_domain_hi) << 8 |
> +					cpu->proximity_domain_lo;
> +			if (pxm > max_pxm)
> +				max_pxm = pxm;
> +			break;
> +		case ACPI_SRAT_TYPE_GENERIC_AFFINITY:
> +			gpu = (struct acpi_srat_generic_affinity *)sub_header;
> +			bdf = *((u16 *)(&gpu->device_handle[0])) << 16 |
> +					*((u16 *)(&gpu->device_handle[2]));
> +			if (bdf == pci_id) {
> +				found = true;
> +				*numa_node = pxm_to_node(gpu->proximity_domain);
> +			}
> +			break;
> +		default:
> +			break;
> +		}
> +
> +		if (found)
> +			break;
> +
> +		sub_header = (struct acpi_subtable_header *)
> +				((unsigned long)sub_header + subtable_len);
> +		subtable_len = sub_header->length;
> +	}
> +
> +	/* workaround bad cpu-gpu binding case */
> +	if (found && (*numa_node < 0 || *numa_node > max_pxm))
> +		*numa_node = 0;
> +}
> +#endif
> +
>   /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
>    * to its NUMA node
>    *	@avail_size: Available size in the memory
> @@ -1774,6 +1855,9 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
>   			uint32_t proximity_domain)
>   {
>   	struct amdgpu_device *adev = (struct amdgpu_device *)kdev->kgd;
> +#ifdef CONFIG_NUMA
> +	int numa_node = 0;
> +#endif
>   
>   	*avail_size -= sizeof(struct crat_subtype_iolink);
>   	if (*avail_size < 0)
> @@ -1805,9 +1889,13 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
>   
>   	sub_type_hdr->proximity_domain_from = proximity_domain;  #ifdef CONFIG_NUMA
> -	if (kdev->pdev->dev.numa_node == NUMA_NO_NODE)
> -		sub_type_hdr->proximity_domain_to = 0;
> -	else
> +	if (kdev->pdev->dev.numa_node == NUMA_NO_NODE) { #ifdef CONFIG_ACPI
> +		kfd_find_numa_node_in_srat(kdev, &numa_node); #endif
> +		sub_type_hdr->proximity_domain_to = numa_node;
> +		set_dev_node(&kdev->pdev->dev, numa_node);
> +	} else
>   		sub_type_hdr->proximity_domain_to = kdev->pdev->dev.numa_node;  #else
>   	sub_type_hdr->proximity_domain_to = 0;
> --
> 2.17.1
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Clijo.lazar%40amd.com%7C96808a6aab7b40861eeb08d90a580524%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637552195438132467%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=ipBmGTX%2Fokto1zRuQ8jlDA8p%2B8BOjHZa5WGGKNJszEY%3D&reserved=0