From 3a87177eb14113bbe8cd95a276af2c412eced6ac Mon Sep 17 00:00:00 2001 From: Harish Kasiviswanathan Date: Fri, 8 Dec 2017 23:08:59 -0500 Subject: [PATCH] drm/amdkfd: Add topology support for dGPUs Generate and parse VCRAT tables for dGPUs in kfd_topology_add_device. Some information that isn't available in the CRAT table is patched into the topology after parsing. HSA_CAP_DOORBELL_TYPE_1_0 is dependent on the ASIC feature CP_HQD_PQ_CONTROL.SLOT_BASED_WPTR, which was not introduced in VI until Carrizo. Report HSA_CAP_DOORBELL_TYPE_PRE_1_0 on Tonga ASICs. v2: Added #include to kfd_crat.c to make it compile Signed-off-by: Harish Kasiviswanathan Signed-off-by: Ben Goz Signed-off-by: Amber Lin Signed-off-by: Jay Cornwall Signed-off-by: Kent Russell Signed-off-by: Felix Kuehling Acked-by: Oded Gabbay Signed-off-by: Oded Gabbay --- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 596 +++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_crat.h | 5 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 + drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 186 +++++-- drivers/gpu/drm/amd/amdkfd/kfd_topology.h | 8 +- 5 files changed, 747 insertions(+), 50 deletions(-) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index c8afbf8015f9..d00061b75193 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -19,11 +19,120 @@ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR * OTHER DEALINGS IN THE SOFTWARE. */ + +#include #include +#include #include "kfd_crat.h" #include "kfd_priv.h" #include "kfd_topology.h" +/* GPU Processor ID base for dGPUs for which VCRAT needs to be created. + * GPU processor ID are expressed with Bit[31]=1. + * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs + * used in the CRAT. + */ +static uint32_t gpu_processor_id_low = 0x80001000; + +/* Return the next available gpu_processor_id and increment it for next GPU + * @total_cu_count - Total CUs present in the GPU including ones + * masked off + */ +static inline unsigned int get_and_inc_gpu_processor_id( + unsigned int total_cu_count) +{ + int current_id = gpu_processor_id_low; + + gpu_processor_id_low += total_cu_count; + return current_id; +} + +/* Static table to describe GPU Cache information */ +struct kfd_gpu_cache_info { + uint32_t cache_size; + uint32_t cache_level; + uint32_t flags; + /* Indicates how many Compute Units share this cache + * Value = 1 indicates the cache is not shared + */ + uint32_t num_cu_shared; +}; + +static struct kfd_gpu_cache_info kaveri_cache_info[] = { + { + /* TCP L1 Cache per CU */ + .cache_size = 16, + .cache_level = 1, + .flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE), + .num_cu_shared = 1, + + }, + { + /* Scalar L1 Instruction Cache (in SQC module) per bank */ + .cache_size = 16, + .cache_level = 1, + .flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_INST_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE), + .num_cu_shared = 2, + }, + { + /* Scalar L1 Data Cache (in SQC module) per bank */ + .cache_size = 8, + .cache_level = 1, + .flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE), + .num_cu_shared = 2, + }, + + /* TODO: Add L2 Cache information */ +}; + + +static struct kfd_gpu_cache_info carrizo_cache_info[] = { + { + /* TCP L1 Cache per CU */ + .cache_size = 16, + .cache_level = 1, + .flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE), + .num_cu_shared = 1, + }, + { + /* Scalar L1 Instruction Cache (in SQC module) per bank */ + .cache_size = 8, + .cache_level = 1, + .flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_INST_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE), + .num_cu_shared = 4, + }, + { + /* Scalar L1 Data Cache (in SQC module) per bank. */ + .cache_size = 4, + .cache_level = 1, + .flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE), + .num_cu_shared = 4, + }, + + /* TODO: Add L2 Cache information */ +}; + +/* NOTE: In future if more information is added to struct kfd_gpu_cache_info + * the following ASICs may need a separate table. + */ +#define hawaii_cache_info kaveri_cache_info +#define tonga_cache_info carrizo_cache_info +#define fiji_cache_info carrizo_cache_info +#define polaris10_cache_info carrizo_cache_info +#define polaris11_cache_info carrizo_cache_info + static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev, struct crat_subtype_computeunit *cu) { @@ -44,7 +153,7 @@ static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev, dev->node_props.lds_size_in_kb = cu->lds_size_in_kb; dev->node_props.max_waves_per_simd = cu->max_waves_simd; dev->node_props.wave_front_size = cu->wave_front_size; - dev->node_props.array_count = cu->num_arrays; + dev->node_props.array_count = cu->array_count; dev->node_props.cu_per_simd_array = cu->num_cu_per_array; dev->node_props.simd_per_cu = cu->num_simd_per_cu; dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu; @@ -94,9 +203,16 @@ static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem, if (!props) return -ENOMEM; - if (dev->node_props.cpu_cores_count == 0) - props->heap_type = HSA_MEM_HEAP_TYPE_FB_PRIVATE; - else + /* We're on GPU node */ + if (dev->node_props.cpu_cores_count == 0) { + /* APU */ + if (mem->visibility_type == 0) + props->heap_type = + HSA_MEM_HEAP_TYPE_FB_PRIVATE; + /* dGPU */ + else + props->heap_type = mem->visibility_type; + } else props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM; if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE) @@ -128,13 +244,29 @@ static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache, struct kfd_cache_properties *props; struct kfd_topology_device *dev; uint32_t id; + uint32_t total_num_of_cu; id = cache->processor_id_low; pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id); - list_for_each_entry(dev, device_list, list) - if (id == dev->node_props.cpu_core_id_base || - id == dev->node_props.simd_id_base) { + list_for_each_entry(dev, device_list, list) { + total_num_of_cu = (dev->node_props.array_count * + dev->node_props.cu_per_simd_array); + + /* Cache infomration in CRAT doesn't have proximity_domain + * information as it is associated with a CPU core or GPU + * Compute Unit. So map the cache using CPU core Id or SIMD + * (GPU) ID. + * TODO: This works because currently we can safely assume that + * Compute Units are parsed before caches are parsed. In + * future, remove this dependency + */ + if ((id >= dev->node_props.cpu_core_id_base && + id <= dev->node_props.cpu_core_id_base + + dev->node_props.cpu_cores_count) || + (id >= dev->node_props.simd_id_base && + id < dev->node_props.simd_id_base + + total_num_of_cu)) { props = kfd_alloc_struct(props); if (!props) return -ENOMEM; @@ -146,6 +278,8 @@ static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache, props->cachelines_per_tag = cache->lines_per_tag; props->cache_assoc = cache->associativity; props->cache_latency = cache->cache_latency; + memcpy(props->sibling_map, cache->sibling_map, + sizeof(props->sibling_map)); if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE) props->cache_type |= HSA_CACHE_TYPE_DATA; @@ -162,6 +296,7 @@ static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache, break; } + } return 0; } @@ -172,8 +307,8 @@ static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache, static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink, struct list_head *device_list) { - struct kfd_iolink_properties *props; - struct kfd_topology_device *dev; + struct kfd_iolink_properties *props = NULL, *props2; + struct kfd_topology_device *dev, *cpu_dev; uint32_t id_from; uint32_t id_to; @@ -192,11 +327,12 @@ static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink, props->node_to = id_to; props->ver_maj = iolink->version_major; props->ver_min = iolink->version_minor; + props->iolink_type = iolink->io_interface_type; - /* - * weight factor (derived from CDIR), currently always 1 - */ - props->weight = 1; + if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) + props->weight = 20; + else + props->weight = node_distance(id_from, id_to); props->min_latency = iolink->minimum_latency; props->max_latency = iolink->maximum_latency; @@ -208,11 +344,29 @@ static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink, dev->io_link_count++; dev->node_props.io_links_count++; list_add_tail(&props->list, &dev->io_link_props); - break; } } + /* CPU topology is created before GPUs are detected, so CPU->GPU + * links are not built at that time. If a PCIe type is discovered, it + * means a GPU is detected and we are adding GPU->CPU to the topology. + * At this time, also add the corresponded CPU->GPU link. + */ + if (props && props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) { + cpu_dev = kfd_topology_device_by_proximity_domain(id_to); + if (!cpu_dev) + return -ENODEV; + /* same everything but the other direction */ + props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL); + props2->node_from = id_to; + props2->node_to = id_from; + props2->kobj = NULL; + cpu_dev->io_link_count++; + cpu_dev->node_props.io_links_count++; + list_add_tail(&props2->list, &cpu_dev->io_link_props); + } + return 0; } @@ -338,6 +492,176 @@ err: return ret; } +/* Helper function. See kfd_fill_gpu_cache_info for parameter description */ +static int fill_in_pcache(struct crat_subtype_cache *pcache, + struct kfd_gpu_cache_info *pcache_info, + struct kfd_cu_info *cu_info, + int mem_available, + int cu_bitmask, + int cache_type, unsigned int cu_processor_id, + int cu_block) +{ + unsigned int cu_sibling_map_mask; + int first_active_cu; + + /* First check if enough memory is available */ + if (sizeof(struct crat_subtype_cache) > mem_available) + return -ENOMEM; + + cu_sibling_map_mask = cu_bitmask; + cu_sibling_map_mask >>= cu_block; + cu_sibling_map_mask &= + ((1 << pcache_info[cache_type].num_cu_shared) - 1); + first_active_cu = ffs(cu_sibling_map_mask); + + /* CU could be inactive. In case of shared cache find the first active + * CU. and incase of non-shared cache check if the CU is inactive. If + * inactive active skip it + */ + if (first_active_cu) { + memset(pcache, 0, sizeof(struct crat_subtype_cache)); + pcache->type = CRAT_SUBTYPE_CACHE_AFFINITY; + pcache->length = sizeof(struct crat_subtype_cache); + pcache->flags = pcache_info[cache_type].flags; + pcache->processor_id_low = cu_processor_id + + (first_active_cu - 1); + pcache->cache_level = pcache_info[cache_type].cache_level; + pcache->cache_size = pcache_info[cache_type].cache_size; + + /* Sibling map is w.r.t processor_id_low, so shift out + * inactive CU + */ + cu_sibling_map_mask = + cu_sibling_map_mask >> (first_active_cu - 1); + + pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF); + pcache->sibling_map[1] = + (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF); + pcache->sibling_map[2] = + (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF); + pcache->sibling_map[3] = + (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF); + return 0; + } + return 1; +} + +/* kfd_fill_gpu_cache_info - Fill GPU cache info using kfd_gpu_cache_info + * tables + * + * @kdev - [IN] GPU device + * @gpu_processor_id - [IN] GPU processor ID to which these caches + * associate + * @available_size - [IN] Amount of memory available in pcache + * @cu_info - [IN] Compute Unit info obtained from KGD + * @pcache - [OUT] memory into which cache data is to be filled in. + * @size_filled - [OUT] amount of data used up in pcache. + * @num_of_entries - [OUT] number of caches added + */ +static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, + int gpu_processor_id, + int available_size, + struct kfd_cu_info *cu_info, + struct crat_subtype_cache *pcache, + int *size_filled, + int *num_of_entries) +{ + struct kfd_gpu_cache_info *pcache_info; + int num_of_cache_types = 0; + int i, j, k; + int ct = 0; + int mem_available = available_size; + unsigned int cu_processor_id; + int ret; + + switch (kdev->device_info->asic_family) { + case CHIP_KAVERI: + pcache_info = kaveri_cache_info; + num_of_cache_types = ARRAY_SIZE(kaveri_cache_info); + break; + case CHIP_HAWAII: + pcache_info = hawaii_cache_info; + num_of_cache_types = ARRAY_SIZE(hawaii_cache_info); + break; + case CHIP_CARRIZO: + pcache_info = carrizo_cache_info; + num_of_cache_types = ARRAY_SIZE(carrizo_cache_info); + break; + case CHIP_TONGA: + pcache_info = tonga_cache_info; + num_of_cache_types = ARRAY_SIZE(tonga_cache_info); + break; + case CHIP_FIJI: + pcache_info = fiji_cache_info; + num_of_cache_types = ARRAY_SIZE(fiji_cache_info); + break; + case CHIP_POLARIS10: + pcache_info = polaris10_cache_info; + num_of_cache_types = ARRAY_SIZE(polaris10_cache_info); + break; + case CHIP_POLARIS11: + pcache_info = polaris11_cache_info; + num_of_cache_types = ARRAY_SIZE(polaris11_cache_info); + break; + default: + return -EINVAL; + } + + *size_filled = 0; + *num_of_entries = 0; + + /* For each type of cache listed in the kfd_gpu_cache_info table, + * go through all available Compute Units. + * The [i,j,k] loop will + * if kfd_gpu_cache_info.num_cu_shared = 1 + * will parse through all available CU + * If (kfd_gpu_cache_info.num_cu_shared != 1) + * then it will consider only one CU from + * the shared unit + */ + + for (ct = 0; ct < num_of_cache_types; ct++) { + cu_processor_id = gpu_processor_id; + for (i = 0; i < cu_info->num_shader_engines; i++) { + for (j = 0; j < cu_info->num_shader_arrays_per_engine; + j++) { + for (k = 0; k < cu_info->num_cu_per_sh; + k += pcache_info[ct].num_cu_shared) { + + ret = fill_in_pcache(pcache, + pcache_info, + cu_info, + mem_available, + cu_info->cu_bitmap[i][j], + ct, + cu_processor_id, + k); + + if (ret < 0) + break; + + if (!ret) { + pcache++; + (*num_of_entries)++; + mem_available -= + sizeof(*pcache); + (*size_filled) += + sizeof(*pcache); + } + + /* Move to next CU block */ + cu_processor_id += + pcache_info[ct].num_cu_shared; + } + } + } + } + + pr_debug("Added [%d] GPU cache entries\n", *num_of_entries); + + return 0; +} + /* * kfd_create_crat_image_acpi - Allocates memory for CRAT image and * copies CRAT from ACPI (if available). @@ -624,6 +948,239 @@ static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size) return 0; } +static int kfd_fill_gpu_memory_affinity(int *avail_size, + struct kfd_dev *kdev, uint8_t type, uint64_t size, + struct crat_subtype_memory *sub_type_hdr, + uint32_t proximity_domain, + const struct kfd_local_mem_info *local_mem_info) +{ + *avail_size -= sizeof(struct crat_subtype_memory); + if (*avail_size < 0) + return -ENOMEM; + + memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory)); + sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY; + sub_type_hdr->length = sizeof(struct crat_subtype_memory); + sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED; + + sub_type_hdr->proximity_domain = proximity_domain; + + pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n", + type, size); + + sub_type_hdr->length_low = lower_32_bits(size); + sub_type_hdr->length_high = upper_32_bits(size); + + sub_type_hdr->width = local_mem_info->vram_width; + sub_type_hdr->visibility_type = type; + + return 0; +} + +/* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU + * to its NUMA node + * @avail_size: Available size in the memory + * @kdev - [IN] GPU device + * @sub_type_hdr: Memory into which io link info will be filled in + * @proximity_domain - proximity domain of the GPU node + * + * Return 0 if successful else return -ve value + */ +static int kfd_fill_gpu_direct_io_link(int *avail_size, + struct kfd_dev *kdev, + struct crat_subtype_iolink *sub_type_hdr, + uint32_t proximity_domain) +{ + *avail_size -= sizeof(struct crat_subtype_iolink); + if (*avail_size < 0) + return -ENOMEM; + + memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink)); + + /* Fill in subtype header data */ + sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY; + sub_type_hdr->length = sizeof(struct crat_subtype_iolink); + sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED; + + /* Fill in IOLINK subtype. + * TODO: Fill-in other fields of iolink subtype + */ + sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS; + 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 + sub_type_hdr->proximity_domain_to = kdev->pdev->dev.numa_node; +#else + sub_type_hdr->proximity_domain_to = 0; +#endif + return 0; +} + +/* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU + * + * @pcrat_image: Fill in VCRAT for GPU + * @size: [IN] allocated size of crat_image. + * [OUT] actual size of data filled in crat_image + */ +static int kfd_create_vcrat_image_gpu(void *pcrat_image, + size_t *size, struct kfd_dev *kdev, + uint32_t proximity_domain) +{ + struct crat_header *crat_table = (struct crat_header *)pcrat_image; + struct crat_subtype_generic *sub_type_hdr; + struct crat_subtype_computeunit *cu; + struct kfd_cu_info cu_info; + struct amd_iommu_device_info iommu_info; + int avail_size = *size; + uint32_t total_num_of_cu; + int num_of_cache_entries = 0; + int cache_mem_filled = 0; + int ret = 0; + const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP | + AMD_IOMMU_DEVICE_FLAG_PRI_SUP | + AMD_IOMMU_DEVICE_FLAG_PASID_SUP; + struct kfd_local_mem_info local_mem_info; + + if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU) + return -EINVAL; + + /* Fill the CRAT Header. + * Modify length and total_entries as subunits are added. + */ + avail_size -= sizeof(struct crat_header); + if (avail_size < 0) + return -ENOMEM; + + memset(crat_table, 0, sizeof(struct crat_header)); + + memcpy(&crat_table->signature, CRAT_SIGNATURE, + sizeof(crat_table->signature)); + /* Change length as we add more subtypes*/ + crat_table->length = sizeof(struct crat_header); + crat_table->num_domains = 1; + crat_table->total_entries = 0; + + /* Fill in Subtype: Compute Unit + * First fill in the sub type header and then sub type data + */ + avail_size -= sizeof(struct crat_subtype_computeunit); + if (avail_size < 0) + return -ENOMEM; + + sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1); + memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit)); + + sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY; + sub_type_hdr->length = sizeof(struct crat_subtype_computeunit); + sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED; + + /* Fill CU subtype data */ + cu = (struct crat_subtype_computeunit *)sub_type_hdr; + cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT; + cu->proximity_domain = proximity_domain; + + kdev->kfd2kgd->get_cu_info(kdev->kgd, &cu_info); + cu->num_simd_per_cu = cu_info.simd_per_cu; + cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number; + cu->max_waves_simd = cu_info.max_waves_per_simd; + + cu->wave_front_size = cu_info.wave_front_size; + cu->array_count = cu_info.num_shader_arrays_per_engine * + cu_info.num_shader_engines; + total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh); + cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu); + cu->num_cu_per_array = cu_info.num_cu_per_sh; + cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu; + cu->num_banks = cu_info.num_shader_engines; + cu->lds_size_in_kb = cu_info.lds_size; + + cu->hsa_capability = 0; + + /* Check if this node supports IOMMU. During parsing this flag will + * translate to HSA_CAP_ATS_PRESENT + */ + iommu_info.flags = 0; + if (amd_iommu_device_info(kdev->pdev, &iommu_info) == 0) { + if ((iommu_info.flags & required_iommu_flags) == + required_iommu_flags) + cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT; + } + + crat_table->length += sub_type_hdr->length; + crat_table->total_entries++; + + /* Fill in Subtype: Memory. Only on systems with large BAR (no + * private FB), report memory as public. On other systems + * report the total FB size (public+private) as a single + * private heap. + */ + kdev->kfd2kgd->get_local_mem_info(kdev->kgd, &local_mem_info); + sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + + sub_type_hdr->length); + + if (local_mem_info.local_mem_size_private == 0) + ret = kfd_fill_gpu_memory_affinity(&avail_size, + kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC, + local_mem_info.local_mem_size_public, + (struct crat_subtype_memory *)sub_type_hdr, + proximity_domain, + &local_mem_info); + else + ret = kfd_fill_gpu_memory_affinity(&avail_size, + kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE, + local_mem_info.local_mem_size_public + + local_mem_info.local_mem_size_private, + (struct crat_subtype_memory *)sub_type_hdr, + proximity_domain, + &local_mem_info); + if (ret < 0) + return ret; + + crat_table->length += sizeof(struct crat_subtype_memory); + crat_table->total_entries++; + + /* TODO: Fill in cache information. This information is NOT readily + * available in KGD + */ + sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + + sub_type_hdr->length); + ret = kfd_fill_gpu_cache_info(kdev, cu->processor_id_low, + avail_size, + &cu_info, + (struct crat_subtype_cache *)sub_type_hdr, + &cache_mem_filled, + &num_of_cache_entries); + + if (ret < 0) + return ret; + + crat_table->length += cache_mem_filled; + crat_table->total_entries += num_of_cache_entries; + avail_size -= cache_mem_filled; + + /* Fill in Subtype: IO_LINKS + * Only direct links are added here which is Link from GPU to + * to its NUMA node. Indirect links are added by userspace. + */ + sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + + cache_mem_filled); + ret = kfd_fill_gpu_direct_io_link(&avail_size, kdev, + (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain); + + if (ret < 0) + return ret; + + crat_table->length += sub_type_hdr->length; + crat_table->total_entries++; + + *size = crat_table->length; + pr_info("Virtual CRAT table created for GPU\n"); + + return ret; +} + /* kfd_create_crat_image_virtual - Allocates memory for CRAT image and * creates a Virtual CRAT (VCRAT) image * @@ -667,9 +1224,14 @@ int kfd_create_crat_image_virtual(void **crat_image, size_t *size, ret = kfd_create_vcrat_image_cpu(pcrat_image, size); break; case COMPUTE_UNIT_GPU: - /* TODO: */ - ret = -EINVAL; - pr_err("VCRAT not implemented for dGPU\n"); + if (!kdev) + return -EINVAL; + pcrat_image = kmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL); + if (!pcrat_image) + return -ENOMEM; + *size = VCRAT_SIZE_FOR_GPU; + ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev, + proximity_domain); break; case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU): /* TODO: */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.h b/drivers/gpu/drm/amd/amdkfd/kfd_crat.h index 1711ab664ec6..b5cd182b9edd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.h @@ -109,7 +109,7 @@ struct crat_subtype_computeunit { uint8_t wave_front_size; uint8_t num_banks; uint16_t micro_engine_id; - uint8_t num_arrays; + uint8_t array_count; uint8_t num_cu_per_array; uint8_t num_simd_per_cu; uint8_t max_slots_scatch_cu; @@ -137,7 +137,8 @@ struct crat_subtype_memory { uint32_t length_low; uint32_t length_high; uint32_t width; - uint8_t reserved2[CRAT_MEMORY_RESERVED_LENGTH]; + uint8_t visibility_type; /* for virtual (dGPU) CRAT */ + uint8_t reserved2[CRAT_MEMORY_RESERVED_LENGTH - 1]; }; /* diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index aeee9d42171b..f0327c219879 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -668,6 +668,8 @@ int kfd_topology_init(void); void kfd_topology_shutdown(void); int kfd_topology_add_device(struct kfd_dev *gpu); int kfd_topology_remove_device(struct kfd_dev *gpu); +struct kfd_topology_device *kfd_topology_device_by_proximity_domain( + uint32_t proximity_domain); struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 9aa600477fff..7fe7ee0670e9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -43,6 +43,25 @@ static struct kfd_system_properties sys_props; static DECLARE_RWSEM(topology_lock); static atomic_t topology_crat_proximity_domain; +struct kfd_topology_device *kfd_topology_device_by_proximity_domain( + uint32_t proximity_domain) +{ + struct kfd_topology_device *top_dev; + struct kfd_topology_device *device = NULL; + + down_read(&topology_lock); + + list_for_each_entry(top_dev, &topology_device_list, list) + if (top_dev->proximity_domain == proximity_domain) { + device = top_dev; + break; + } + + up_read(&topology_lock); + + return device; +} + struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) { struct kfd_topology_device *top_dev; @@ -79,6 +98,7 @@ struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) return device; } +/* Called with write topology_lock acquired */ static void kfd_release_topology_device(struct kfd_topology_device *dev) { struct kfd_mem_properties *mem; @@ -394,8 +414,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, } sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", - dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz( - dev->gpu->kgd)); + dev->node_props.max_engine_clk_fcompute); sysfs_show_64bit_prop(buffer, "local_mem_size", (unsigned long long int) 0); @@ -597,6 +616,7 @@ static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, return 0; } +/* Called with write topology lock acquired */ static int kfd_build_sysfs_node_tree(void) { struct kfd_topology_device *dev; @@ -613,6 +633,7 @@ static int kfd_build_sysfs_node_tree(void) return 0; } +/* Called with write topology lock acquired */ static void kfd_remove_sysfs_node_tree(void) { struct kfd_topology_device *dev; @@ -908,19 +929,26 @@ static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) return hashout; } - +/* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If + * the GPU device is not already present in the topology device + * list then return NULL. This means a new topology device has to + * be created for this GPU. + * TODO: Rather than assiging @gpu to first topology device withtout + * gpu attached, it will better to have more stringent check. + */ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) { struct kfd_topology_device *dev; struct kfd_topology_device *out_dev = NULL; + down_write(&topology_lock); list_for_each_entry(dev, &topology_device_list, list) if (!dev->gpu && (dev->node_props.simd_count > 0)) { dev->gpu = gpu; out_dev = dev; break; } - + up_write(&topology_lock); return out_dev; } @@ -932,6 +960,45 @@ static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) */ } +/* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, + * patch this after CRAT parsing. + */ +static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) +{ + struct kfd_mem_properties *mem; + struct kfd_local_mem_info local_mem_info; + + if (!dev) + return; + + /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with + * single bank of VRAM local memory. + * for dGPUs - VCRAT reports only one bank of Local Memory + * for APUs - If CRAT from ACPI reports more than one bank, then + * all the banks will report the same mem_clk_max information + */ + dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd, + &local_mem_info); + + list_for_each_entry(mem, &dev->mem_props, list) + mem->mem_clk_max = local_mem_info.mem_clk_max; +} + +static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) +{ + struct kfd_iolink_properties *link; + + if (!dev || !dev->gpu) + return; + + /* GPU only creates direck links so apply flags setting to all */ + if (dev->gpu->device_info->asic_family == CHIP_HAWAII) + list_for_each_entry(link, &dev->io_link_props, list) + link->flags = CRAT_IOLINK_FLAGS_ENABLED | + CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | + CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; +} + int kfd_topology_add_device(struct kfd_dev *gpu) { uint32_t gpu_id; @@ -939,6 +1006,9 @@ int kfd_topology_add_device(struct kfd_dev *gpu) struct kfd_cu_info cu_info; int res = 0; struct list_head temp_topology_device_list; + void *crat_image = NULL; + size_t image_size = 0; + int proximity_domain; INIT_LIST_HEAD(&temp_topology_device_list); @@ -946,27 +1016,33 @@ int kfd_topology_add_device(struct kfd_dev *gpu) pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); - /* - * Try to assign the GPU to existing topology device (generated from - * CRAT table + proximity_domain = atomic_inc_return(&topology_crat_proximity_domain); + + /* Check to see if this gpu device exists in the topology_device_list. + * If so, assign the gpu to that device, + * else create a Virtual CRAT for this gpu device and then parse that + * CRAT to create a new topology device. Once created assign the gpu to + * that topology device */ dev = kfd_assign_gpu(gpu); if (!dev) { - pr_info("GPU was not found in the current topology. Extending.\n"); - kfd_debug_print_topology(); - dev = kfd_create_topology_device(&temp_topology_device_list); - if (!dev) { - res = -ENOMEM; + res = kfd_create_crat_image_virtual(&crat_image, &image_size, + COMPUTE_UNIT_GPU, gpu, + proximity_domain); + if (res) { + pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n", + gpu_id); + return res; + } + res = kfd_parse_crat_table(crat_image, + &temp_topology_device_list, + proximity_domain); + if (res) { + pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", + gpu_id); goto err; } - dev->gpu = gpu; - - /* - * TODO: Make a call to retrieve topology information from the - * GPU vBIOS - */ - down_write(&topology_lock); kfd_topology_update_device_list(&temp_topology_device_list, &topology_device_list); @@ -974,34 +1050,86 @@ int kfd_topology_add_device(struct kfd_dev *gpu) /* Update the SYSFS tree, since we added another topology * device */ - if (kfd_topology_update_sysfs() < 0) - kfd_topology_release_sysfs(); - + res = kfd_topology_update_sysfs(); up_write(&topology_lock); + if (!res) + sys_props.generation_count++; + else + pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n", + gpu_id, res); + dev = kfd_assign_gpu(gpu); + if (WARN_ON(!dev)) { + res = -ENODEV; + goto err; + } } dev->gpu_id = gpu_id; gpu->id = gpu_id; + + /* TODO: Move the following lines to function + * kfd_add_non_crat_information + */ + + /* Fill-in additional information that is not available in CRAT but + * needed for the topology + */ + dev->gpu->kfd2kgd->get_cu_info(dev->gpu->kgd, &cu_info); - dev->node_props.simd_count = dev->node_props.simd_per_cu * - cu_info.cu_active_number; + dev->node_props.simd_arrays_per_engine = + cu_info.num_shader_arrays_per_engine; + dev->node_props.vendor_id = gpu->pdev->vendor; dev->node_props.device_id = gpu->pdev->device; dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number, gpu->pdev->devfn); - /* - * TODO: Retrieve max engine clock values from KGD - */ + dev->node_props.max_engine_clk_fcompute = + dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(dev->gpu->kgd); + dev->node_props.max_engine_clk_ccompute = + cpufreq_quick_get_max(0) / 1000; - if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) { - dev->node_props.capability |= HSA_CAP_DOORBELL_PACKET_TYPE; + kfd_fill_mem_clk_max_info(dev); + kfd_fill_iolink_non_crat_info(dev); + + switch (dev->gpu->device_info->asic_family) { + case CHIP_KAVERI: + case CHIP_HAWAII: + case CHIP_TONGA: + dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << + HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & + HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); + break; + case CHIP_CARRIZO: + case CHIP_FIJI: + case CHIP_POLARIS10: + case CHIP_POLARIS11: pr_debug("Adding doorbell packet type capability\n"); + dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << + HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & + HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); + break; + default: + WARN(1, "Unexpected ASIC family %u", + dev->gpu->device_info->asic_family); } + /* Fix errors in CZ CRAT. + * simd_count: Carrizo CRAT reports wrong simd_count, probably + * because it doesn't consider masked out CUs + * capability flag: Carrizo CRAT doesn't report IOMMU + * flags. TODO: Fix this. + */ + if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) + dev->node_props.simd_count = + cu_info.simd_per_cu * cu_info.cu_active_number; + + kfd_debug_print_topology(); + if (!res) kfd_notify_gpu_change(gpu_id, 1); err: + kfd_destroy_crat_image(crat_image); return res; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h index 866818964a9b..55de56fe11c9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h @@ -39,8 +39,12 @@ #define HSA_CAP_WATCH_POINTS_SUPPORTED 0x00000080 #define HSA_CAP_WATCH_POINTS_TOTALBITS_MASK 0x00000f00 #define HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT 8 -#define HSA_CAP_RESERVED 0xfffff000 -#define HSA_CAP_DOORBELL_PACKET_TYPE 0x00001000 +#define HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK 0x00003000 +#define HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT 12 +#define HSA_CAP_RESERVED 0xffffc000 + +#define HSA_CAP_DOORBELL_TYPE_PRE_1_0 0x0 +#define HSA_CAP_DOORBELL_TYPE_1_0 0x1 struct kfd_node_properties { uint32_t cpu_cores_count;