[AMD Official Use Only - Internal Distribution Only] Converting using pxm_to_node and then comparing against pxm value looks a bit odd. Shouldn't the comparsion be between equals - node to node or pxm to pxm?
Thanks, Lijo ________________________________ From: Huang, JinHuiEric <jinhuieric.hu...@amd.com> Sent: Tuesday, May 4, 2021 7:30:44 PM To: Lazar, Lijo <lijo.la...@amd.com>; amd-gfx@lists.freedesktop.org <amd-gfx@lists.freedesktop.org> Subject: Re: [PATCH] drm/amdkfd: add ACPI SRAT parsing for topology 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-boun...@lists.freedesktop.org> On Behalf Of Eric Huang > Sent: Wednesday, April 28, 2021 8:42 PM > To: amd-gfx@lists.freedesktop.org > Cc: Huang, JinHuiEric <jinhuieric.hu...@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.hu...@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@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
_______________________________________________ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx