diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd')
23 files changed, 1865 insertions, 249 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/Makefile b/drivers/gpu/drm/amd/amdkfd/Makefile index 19cfbf9577b4..e758c2a24cd0 100644 --- a/drivers/gpu/drm/amd/amdkfd/Makefile +++ b/drivers/gpu/drm/amd/amdkfd/Makefile @@ -37,6 +37,7 @@ AMDKFD_FILES := $(AMDKFD_PATH)/kfd_module.o \ $(AMDKFD_PATH)/kfd_mqd_manager_vi.o \ $(AMDKFD_PATH)/kfd_mqd_manager_v9.o \ $(AMDKFD_PATH)/kfd_mqd_manager_v10.o \ + $(AMDKFD_PATH)/kfd_mqd_manager_v11.o \ $(AMDKFD_PATH)/kfd_kernel_queue.o \ $(AMDKFD_PATH)/kfd_packet_manager.o \ $(AMDKFD_PATH)/kfd_packet_manager_vi.o \ @@ -47,10 +48,12 @@ AMDKFD_FILES := $(AMDKFD_PATH)/kfd_module.o \ $(AMDKFD_PATH)/kfd_device_queue_manager_vi.o \ $(AMDKFD_PATH)/kfd_device_queue_manager_v9.o \ $(AMDKFD_PATH)/kfd_device_queue_manager_v10.o \ + $(AMDKFD_PATH)/kfd_device_queue_manager_v11.o \ $(AMDKFD_PATH)/kfd_interrupt.o \ $(AMDKFD_PATH)/kfd_events.o \ $(AMDKFD_PATH)/cik_event_interrupt.o \ $(AMDKFD_PATH)/kfd_int_process_v9.o \ + $(AMDKFD_PATH)/kfd_int_process_v11.o \ $(AMDKFD_PATH)/kfd_smi_events.o \ $(AMDKFD_PATH)/kfd_crat.o diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 607f65ab39ac..1c7016958d6d 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -944,8 +944,6 @@ err_drm_file: bool kfd_dev_is_large_bar(struct kfd_dev *dev) { - struct kfd_local_mem_info mem_info; - if (debug_largebar) { pr_debug("Simulate large-bar allocation on non large-bar machine\n"); return true; @@ -954,9 +952,8 @@ bool kfd_dev_is_large_bar(struct kfd_dev *dev) if (dev->use_iommu_v2) return false; - amdgpu_amdkfd_get_local_mem_info(dev->adev, &mem_info); - if (mem_info.local_mem_size_private == 0 && - mem_info.local_mem_size_public > 0) + if (dev->local_mem_info.local_mem_size_private == 0 && + dev->local_mem_info.local_mem_size_public > 0) return true; return false; } @@ -1128,14 +1125,6 @@ err_pdd: return ret; } -static bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev) -{ - return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2) || - (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && - dev->adev->sdma.instance[0].fw_version >= 18) || - KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0); -} - static int kfd_ioctl_map_memory_to_gpu(struct file *filep, struct kfd_process *p, void *data) { @@ -1146,7 +1135,6 @@ static int kfd_ioctl_map_memory_to_gpu(struct file *filep, long err = 0; int i; uint32_t *devices_arr = NULL; - bool table_freed = false; if (!args->n_devices) { pr_debug("Device IDs array empty\n"); @@ -1208,7 +1196,7 @@ static int kfd_ioctl_map_memory_to_gpu(struct file *filep, err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu( peer_pdd->dev->adev, (struct kgd_mem *)mem, - peer_pdd->drm_priv, &table_freed); + peer_pdd->drm_priv); if (err) { struct pci_dev *pdev = peer_pdd->dev->adev->pdev; @@ -1233,13 +1221,11 @@ static int kfd_ioctl_map_memory_to_gpu(struct file *filep, } /* Flush TLBs after waiting for the page table updates to complete */ - if (table_freed || !kfd_flush_tlb_after_unmap(dev)) { - for (i = 0; i < args->n_devices; i++) { - peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]); - if (WARN_ON_ONCE(!peer_pdd)) - continue; - kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY); - } + for (i = 0; i < args->n_devices; i++) { + peer_pdd = kfd_process_device_data_by_id(p, devices_arr[i]); + if (WARN_ON_ONCE(!peer_pdd)) + continue; + kfd_flush_tlb(peer_pdd, TLB_FLUSH_LEGACY); } kfree(devices_arr); @@ -2206,8 +2192,8 @@ static int criu_restore_bo(struct kfd_process *p, if (IS_ERR(peer_pdd)) return PTR_ERR(peer_pdd); - ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem, peer_pdd->drm_priv, - NULL); + ret = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(peer->adev, kgd_mem, + peer_pdd->drm_priv); if (ret) { pr_err("Failed to map to gpu %d/%d\n", j, p->n_pdds); return ret; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index 1eaabd2cb41b..5e9adbc71bbd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -1056,7 +1056,7 @@ static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink, * table, add corresponded reversed direction link now. */ if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) { - to_dev = kfd_topology_device_by_proximity_domain(id_to); + to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to); if (!to_dev) return -ENODEV; /* same everything but the other direction */ @@ -1315,6 +1315,80 @@ static int fill_in_l2_l3_pcache(struct crat_subtype_cache *pcache, return 1; } +#define KFD_MAX_CACHE_TYPES 6 + +static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev, + struct kfd_gpu_cache_info *pcache_info) +{ + struct amdgpu_device *adev = kdev->adev; + int i = 0; + + /* TCP L1 Cache per CU */ + if (adev->gfx.config.gc_tcp_l1_size) { + pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size; + pcache_info[i].cache_level = 1; + pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE); + pcache_info[0].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2; + i++; + } + /* Scalar L1 Instruction Cache per SQC */ + if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) { + pcache_info[i].cache_size = + adev->gfx.config.gc_l1_instruction_cache_size_per_sqc; + pcache_info[i].cache_level = 1; + pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_INST_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE); + pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2; + i++; + } + /* Scalar L1 Data Cache per SQC */ + if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) { + pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc; + pcache_info[i].cache_level = 1; + pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE); + pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2; + i++; + } + /* GL1 Data Cache per SA */ + if (adev->gfx.config.gc_gl1c_per_sa && + adev->gfx.config.gc_gl1c_size_per_instance) { + pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa * + adev->gfx.config.gc_gl1c_size_per_instance; + pcache_info[i].cache_level = 1; + pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE); + pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; + i++; + } + /* L2 Data Cache per GPU (Total Tex Cache) */ + if (adev->gfx.config.gc_gl2c_per_gpu) { + pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu; + pcache_info[i].cache_level = 2; + pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE); + pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; + i++; + } + /* L3 Data Cache per GPU */ + if (adev->gmc.mall_size) { + pcache_info[i].cache_size = adev->gmc.mall_size / 1024; + pcache_info[i].cache_level = 3; + pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED | + CRAT_CACHE_FLAGS_DATA_CACHE | + CRAT_CACHE_FLAGS_SIMD_CACHE); + pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh; + i++; + } + return i; +} + /* kfd_fill_gpu_cache_info - Fill GPU cache info using kfd_gpu_cache_info * tables * @@ -1336,6 +1410,7 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, int *num_of_entries) { struct kfd_gpu_cache_info *pcache_info; + struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES]; int num_of_cache_types = 0; int i, j, k; int ct = 0; @@ -1444,6 +1519,13 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, pcache_info = yellow_carp_cache_info; num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info); break; + case IP_VERSION(11, 0, 0): + case IP_VERSION(11, 0, 1): + case IP_VERSION(11, 0, 2): + pcache_info = cache_info; + num_of_cache_types = + kfd_fill_gpu_cache_info_from_gfx_config(kdev, pcache_info); + break; default: return -EINVAL; } @@ -2152,7 +2234,7 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image, * report the total FB size (public+private) as a single * private heap. */ - amdgpu_amdkfd_get_local_mem_info(kdev->adev, &local_mem_info); + local_mem_info = kdev->local_mem_info; sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + sub_type_hdr->length); @@ -2225,7 +2307,7 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image, */ if (kdev->hive_id) { for (nid = 0; nid < proximity_domain; ++nid) { - peer_dev = kfd_topology_device_by_proximity_domain(nid); + peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid); if (!peer_dev->gpu) continue; if (peer_dev->gpu->hive_id != kdev->hive_id) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 62aa6c9d5123..f1a225a20719 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -53,6 +53,7 @@ extern const struct kfd2kgd_calls arcturus_kfd2kgd; extern const struct kfd2kgd_calls aldebaran_kfd2kgd; extern const struct kfd2kgd_calls gfx_v10_kfd2kgd; extern const struct kfd2kgd_calls gfx_v10_3_kfd2kgd; +extern const struct kfd2kgd_calls gfx_v11_kfd2kgd; static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, unsigned int chunk_size); @@ -60,7 +61,7 @@ static void kfd_gtt_sa_fini(struct kfd_dev *kfd); static int kfd_resume(struct kfd_dev *kfd); -static void kfd_device_info_set_sdma_queue_num(struct kfd_dev *kfd) +static void kfd_device_info_set_sdma_info(struct kfd_dev *kfd) { uint32_t sdma_version = kfd->adev->ip_versions[SDMA0_HWIP][0]; @@ -72,6 +73,7 @@ static void kfd_device_info_set_sdma_queue_num(struct kfd_dev *kfd) case IP_VERSION(4, 1, 2):/* RENOIR */ case IP_VERSION(5, 2, 1):/* VANGOGH */ case IP_VERSION(5, 2, 3):/* YELLOW_CARP */ + case IP_VERSION(6, 0, 1): kfd->device_info.num_sdma_queues_per_engine = 2; break; case IP_VERSION(4, 2, 0):/* VEGA20 */ @@ -85,6 +87,8 @@ static void kfd_device_info_set_sdma_queue_num(struct kfd_dev *kfd) case IP_VERSION(5, 2, 2):/* NAVY_FLOUNDER */ case IP_VERSION(5, 2, 4):/* DIMGREY_CAVEFISH */ case IP_VERSION(5, 2, 5):/* BEIGE_GOBY */ + case IP_VERSION(6, 0, 0): + case IP_VERSION(6, 0, 2): kfd->device_info.num_sdma_queues_per_engine = 8; break; default: @@ -93,6 +97,19 @@ static void kfd_device_info_set_sdma_queue_num(struct kfd_dev *kfd) sdma_version); kfd->device_info.num_sdma_queues_per_engine = 8; } + + switch (sdma_version) { + case IP_VERSION(6, 0, 0): + case IP_VERSION(6, 0, 1): + case IP_VERSION(6, 0, 2): + /* Reserve 1 for paging and 1 for gfx */ + kfd->device_info.num_reserved_sdma_queues_per_engine = 2; + /* BIT(0)=engine-0 queue-0; BIT(1)=engine-1 queue-0; BIT(2)=engine-0 queue-1; ... */ + kfd->device_info.reserved_sdma_queues_bitmap = 0xFULL; + break; + default: + break; + } } static void kfd_device_info_set_event_interrupt_class(struct kfd_dev *kfd) @@ -121,6 +138,11 @@ static void kfd_device_info_set_event_interrupt_class(struct kfd_dev *kfd) case IP_VERSION(10, 3, 5): /* BEIGE_GOBY */ kfd->device_info.event_interrupt_class = &event_interrupt_class_v9; break; + case IP_VERSION(11, 0, 0): + case IP_VERSION(11, 0, 1): + case IP_VERSION(11, 0, 2): + kfd->device_info.event_interrupt_class = &event_interrupt_class_v11; + break; default: dev_warn(kfd_device, "v9 event interrupt handler is set due to " "mismatch of gc ip block(GC_HWIP:0x%x).\n", gc_version); @@ -145,7 +167,7 @@ static void kfd_device_info_init(struct kfd_dev *kfd, kfd->device_info.ih_ring_entry_size = 8 * sizeof(uint32_t); kfd->device_info.supports_cwsr = true; - kfd_device_info_set_sdma_queue_num(kfd); + kfd_device_info_set_sdma_info(kfd); kfd_device_info_set_event_interrupt_class(kfd); @@ -346,6 +368,18 @@ struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf) if (!vf) f2g = &gfx_v10_3_kfd2kgd; break; + case IP_VERSION(11, 0, 0): + gfx_target_version = 110000; + f2g = &gfx_v11_kfd2kgd; + break; + case IP_VERSION(11, 0, 1): + gfx_target_version = 110003; + f2g = &gfx_v11_kfd2kgd; + break; + case IP_VERSION(11, 0, 2): + gfx_target_version = 110002; + f2g = &gfx_v11_kfd2kgd; + break; default: break; } @@ -575,6 +609,8 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, if (kfd_resume(kfd)) goto kfd_resume_error; + amdgpu_amdkfd_get_local_mem_info(kfd->adev, &kfd->local_mem_info); + if (kfd_topology_add_device(kfd)) { dev_err(kfd_device, "Error adding device to topology\n"); goto kfd_topology_add_device_error; @@ -873,8 +909,6 @@ out: static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, unsigned int chunk_size) { - unsigned int num_of_longs; - if (WARN_ON(buf_size < chunk_size)) return -EINVAL; if (WARN_ON(buf_size == 0)) @@ -885,11 +919,8 @@ static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, kfd->gtt_sa_chunk_size = chunk_size; kfd->gtt_sa_num_of_chunks = buf_size / chunk_size; - num_of_longs = (kfd->gtt_sa_num_of_chunks + BITS_PER_LONG - 1) / - BITS_PER_LONG; - - kfd->gtt_sa_bitmap = kcalloc(num_of_longs, sizeof(long), GFP_KERNEL); - + kfd->gtt_sa_bitmap = bitmap_zalloc(kfd->gtt_sa_num_of_chunks, + GFP_KERNEL); if (!kfd->gtt_sa_bitmap) return -ENOMEM; @@ -899,13 +930,12 @@ static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, mutex_init(&kfd->gtt_sa_lock); return 0; - } static void kfd_gtt_sa_fini(struct kfd_dev *kfd) { mutex_destroy(&kfd->gtt_sa_lock); - kfree(kfd->gtt_sa_bitmap); + bitmap_free(kfd->gtt_sa_bitmap); } static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr, @@ -973,7 +1003,7 @@ kfd_gtt_restart_search: /* If we need only one chunk, mark it as allocated and get out */ if (size <= kfd->gtt_sa_chunk_size) { pr_debug("Single bit\n"); - set_bit(found, kfd->gtt_sa_bitmap); + __set_bit(found, kfd->gtt_sa_bitmap); goto kfd_gtt_out; } @@ -1011,10 +1041,8 @@ kfd_gtt_restart_search: (*mem_obj)->range_start, (*mem_obj)->range_end); /* Mark the chunks as allocated */ - for (found = (*mem_obj)->range_start; - found <= (*mem_obj)->range_end; - found++) - set_bit(found, kfd->gtt_sa_bitmap); + bitmap_set(kfd->gtt_sa_bitmap, (*mem_obj)->range_start, + (*mem_obj)->range_end - (*mem_obj)->range_start + 1); kfd_gtt_out: mutex_unlock(&kfd->gtt_sa_lock); @@ -1029,8 +1057,6 @@ kfd_gtt_no_free_chunk: int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj) { - unsigned int bit; - /* Act like kfree when trying to free a NULL object */ if (!mem_obj) return 0; @@ -1041,10 +1067,8 @@ int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj) mutex_lock(&kfd->gtt_sa_lock); /* Mark the chunks as free */ - for (bit = mem_obj->range_start; - bit <= mem_obj->range_end; - bit++) - clear_bit(bit, kfd->gtt_sa_bitmap); + bitmap_clear(kfd->gtt_sa_bitmap, mem_obj->range_start, + mem_obj->range_end - mem_obj->range_start + 1); mutex_unlock(&kfd->gtt_sa_lock); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 198672264492..e1797657b04c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -35,6 +35,7 @@ #include "cik_regs.h" #include "kfd_kernel_queue.h" #include "amdgpu_amdkfd.h" +#include "mes_api_def.h" /* Size of the per-pipe EOP queue */ #define CIK_HPD_EOP_BYTES_LOG2 11 @@ -118,6 +119,11 @@ unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm) dqm->dev->device_info.num_sdma_queues_per_engine; } +static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manager *dqm) +{ + return dqm->dev->device_info.reserved_sdma_queues_bitmap; +} + void program_sh_mem_settings(struct device_queue_manager *dqm, struct qcm_process_device *qpd) { @@ -129,6 +135,152 @@ void program_sh_mem_settings(struct device_queue_manager *dqm, qpd->sh_mem_bases); } +static void kfd_hws_hang(struct device_queue_manager *dqm) +{ + /* + * Issue a GPU reset if HWS is unresponsive + */ + dqm->is_hws_hang = true; + + /* It's possible we're detecting a HWS hang in the + * middle of a GPU reset. No need to schedule another + * reset in this case. + */ + if (!dqm->is_resetting) + schedule_work(&dqm->hw_exception_work); +} + +static int convert_to_mes_queue_type(int queue_type) +{ + int mes_queue_type; + + switch (queue_type) { + case KFD_QUEUE_TYPE_COMPUTE: + mes_queue_type = MES_QUEUE_TYPE_COMPUTE; + break; + case KFD_QUEUE_TYPE_SDMA: + mes_queue_type = MES_QUEUE_TYPE_SDMA; + break; + default: + WARN(1, "Invalid queue type %d", queue_type); + mes_queue_type = -EINVAL; + break; + } + + return mes_queue_type; +} + +static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q, + struct qcm_process_device *qpd) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev; + struct kfd_process_device *pdd = qpd_to_pdd(qpd); + struct mes_add_queue_input queue_input; + int r, queue_type; + + if (dqm->is_hws_hang) + return -EIO; + + memset(&queue_input, 0x0, sizeof(struct mes_add_queue_input)); + queue_input.process_id = qpd->pqm->process->pasid; + queue_input.page_table_base_addr = qpd->page_table_base; + queue_input.process_va_start = 0; + queue_input.process_va_end = adev->vm_manager.max_pfn - 1; + /* MES unit for quantum is 100ns */ + queue_input.process_quantum = KFD_MES_PROCESS_QUANTUM; /* Equivalent to 10ms. */ + queue_input.process_context_addr = pdd->proc_ctx_gpu_addr; + queue_input.gang_quantum = KFD_MES_GANG_QUANTUM; /* Equivalent to 1ms */ + queue_input.gang_context_addr = q->gang_ctx_gpu_addr; + queue_input.inprocess_gang_priority = q->properties.priority; + queue_input.gang_global_priority_level = + AMDGPU_MES_PRIORITY_LEVEL_NORMAL; + queue_input.doorbell_offset = q->properties.doorbell_off; + queue_input.mqd_addr = q->gart_mqd_addr; + queue_input.wptr_addr = (uint64_t)q->properties.write_ptr; + queue_input.paging = false; + queue_input.tba_addr = qpd->tba_addr; + queue_input.tma_addr = qpd->tma_addr; + + queue_type = convert_to_mes_queue_type(q->properties.type); + if (queue_type < 0) { + pr_err("Queue type not supported with MES, queue:%d\n", + q->properties.type); + return -EINVAL; + } + queue_input.queue_type = (uint32_t)queue_type; + + if (q->gws) { + queue_input.gws_base = 0; + queue_input.gws_size = qpd->num_gws; + } + + amdgpu_mes_lock(&adev->mes); + r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input); + amdgpu_mes_unlock(&adev->mes); + if (r) { + pr_err("failed to add hardware queue to MES, doorbell=0x%x\n", + q->properties.doorbell_off); + pr_err("MES might be in unrecoverable state, issue a GPU reset\n"); + kfd_hws_hang(dqm); +} + + return r; +} + +static int remove_queue_mes(struct device_queue_manager *dqm, struct queue *q, + struct qcm_process_device *qpd) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev; + int r; + struct mes_remove_queue_input queue_input; + + if (dqm->is_hws_hang) + return -EIO; + + memset(&queue_input, 0x0, sizeof(struct mes_remove_queue_input)); + queue_input.doorbell_offset = q->properties.doorbell_off; + queue_input.gang_context_addr = q->gang_ctx_gpu_addr; + + amdgpu_mes_lock(&adev->mes); + r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input); + amdgpu_mes_unlock(&adev->mes); + + if (r) { + pr_err("failed to remove hardware queue from MES, doorbell=0x%x\n", + q->properties.doorbell_off); + pr_err("MES might be in unrecoverable state, issue a GPU reset\n"); + kfd_hws_hang(dqm); + } + + return r; +} + +static int remove_all_queues_mes(struct device_queue_manager *dqm) +{ + struct device_process_node *cur; + struct qcm_process_device *qpd; + struct queue *q; + int retval = 0; + + list_for_each_entry(cur, &dqm->queues, list) { + qpd = cur->qpd; + list_for_each_entry(q, &qpd->queues_list, list) { + if (q->properties.is_active) { + retval = remove_queue_mes(dqm, q, qpd); + if (retval) { + pr_err("%s: Failed to remove queue %d for dev %d", + __func__, + q->properties.queue_id, + dqm->dev->id); + return retval; + } + } + } + } + + return retval; +} + static void increment_queue_count(struct device_queue_manager *dqm, struct qcm_process_device *qpd, struct queue *q) @@ -659,6 +811,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, struct mqd_manager *mqd_mgr; struct kfd_process_device *pdd; bool prev_active = false; + bool add_queue = false; dqm_lock(dqm); pdd = kfd_get_process_device_data(q->device, q->process); @@ -674,8 +827,12 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, /* Make sure the queue is unmapped before updating the MQD */ if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { - retval = unmap_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false); + if (!dqm->dev->shared_resources.enable_mes) + retval = unmap_queues_cpsch(dqm, + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false); + else if (prev_active) + retval = remove_queue_mes(dqm, q, &pdd->qpd); + if (retval) { pr_err("unmap queue failed\n"); goto out_unlock; @@ -727,9 +884,12 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, q->properties.is_gws = false; } - if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) - retval = map_queues_cpsch(dqm); - else if (q->properties.is_active && + if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { + if (!dqm->dev->shared_resources.enable_mes) + retval = map_queues_cpsch(dqm); + else if (add_queue) + retval = add_queue_mes(dqm, q, &pdd->qpd); + } else if (q->properties.is_active && (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || q->properties.type == KFD_QUEUE_TYPE_SDMA || q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { @@ -822,12 +982,22 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, q->properties.is_active = false; decrement_queue_count(dqm, qpd, q); + + if (dqm->dev->shared_resources.enable_mes) { + retval = remove_queue_mes(dqm, q, qpd); + if (retval) { + pr_err("Failed to evict queue %d\n", + q->properties.queue_id); + goto out; + } + } } pdd->last_evict_timestamp = get_jiffies_64(); - retval = execute_queues_cpsch(dqm, - qpd->is_debug ? - KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + if (!dqm->dev->shared_resources.enable_mes) + retval = execute_queues_cpsch(dqm, + qpd->is_debug ? + KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); out: dqm_unlock(dqm); @@ -951,9 +1121,19 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, q->properties.is_active = true; increment_queue_count(dqm, &pdd->qpd, q); + + if (dqm->dev->shared_resources.enable_mes) { + retval = add_queue_mes(dqm, q, qpd); + if (retval) { + pr_err("Failed to restore queue %d\n", + q->properties.queue_id); + goto out; + } + } } - retval = execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + if (!dqm->dev->shared_resources.enable_mes) + retval = execute_queues_cpsch(dqm, + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); qpd->evicted = 0; eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp; atomic64_add(eviction_duration, &pdd->evict_duration_counter); @@ -1081,6 +1261,9 @@ static int initialize_nocpsch(struct device_queue_manager *dqm) memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid)); dqm->sdma_bitmap = ~0ULL >> (64 - get_num_sdma_queues(dqm)); + dqm->sdma_bitmap &= ~(get_reserved_sdma_queues_bitmap(dqm)); + pr_info("sdma_bitmap: %llx\n", dqm->sdma_bitmap); + dqm->xgmi_sdma_bitmap = ~0ULL >> (64 - get_num_xgmi_sdma_queues(dqm)); return 0; @@ -1277,6 +1460,9 @@ static int initialize_cpsch(struct device_queue_manager *dqm) else dqm->sdma_bitmap = (BIT_ULL(num_sdma_queues) - 1); + dqm->sdma_bitmap &= ~(get_reserved_sdma_queues_bitmap(dqm)); + pr_info("sdma_bitmap: %llx\n", dqm->sdma_bitmap); + num_xgmi_sdma_queues = get_num_xgmi_sdma_queues(dqm); if (num_xgmi_sdma_queues >= BITS_PER_TYPE(dqm->xgmi_sdma_bitmap)) dqm->xgmi_sdma_bitmap = ULLONG_MAX; @@ -1295,14 +1481,16 @@ static int start_cpsch(struct device_queue_manager *dqm) retval = 0; dqm_lock(dqm); - retval = pm_init(&dqm->packet_mgr, dqm); - if (retval) - goto fail_packet_manager_init; - retval = set_sched_resources(dqm); - if (retval) - goto fail_set_sched_resources; + if (!dqm->dev->shared_resources.enable_mes) { + retval = pm_init(&dqm->packet_mgr, dqm); + if (retval) + goto fail_packet_manager_init; + retval = set_sched_resources(dqm); + if (retval) + goto fail_set_sched_resources; + } pr_debug("Allocating fence memory\n"); /* allocate fence memory on the gart */ @@ -1321,13 +1509,15 @@ static int start_cpsch(struct device_queue_manager *dqm) dqm->is_hws_hang = false; dqm->is_resetting = false; dqm->sched_running = true; - execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + if (!dqm->dev->shared_resources.enable_mes) + execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); dqm_unlock(dqm); return 0; fail_allocate_vidmem: fail_set_sched_resources: - pm_uninit(&dqm->packet_mgr, false); + if (!dqm->dev->shared_resources.enable_mes) + pm_uninit(&dqm->packet_mgr, false); fail_packet_manager_init: dqm_unlock(dqm); return retval; @@ -1343,15 +1533,22 @@ static int stop_cpsch(struct device_queue_manager *dqm) return 0; } - if (!dqm->is_hws_hang) - unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, false); + if (!dqm->is_hws_hang) { + if (!dqm->dev->shared_resources.enable_mes) + unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, false); + else + remove_all_queues_mes(dqm); + } + hanging = dqm->is_hws_hang || dqm->is_resetting; dqm->sched_running = false; - pm_release_ib(&dqm->packet_mgr); + if (!dqm->dev->shared_resources.enable_mes) + pm_release_ib(&dqm->packet_mgr); kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); - pm_uninit(&dqm->packet_mgr, hanging); + if (!dqm->dev->shared_resources.enable_mes) + pm_uninit(&dqm->packet_mgr, hanging); dqm_unlock(dqm); return 0; @@ -1469,8 +1666,14 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, if (q->properties.is_active) { increment_queue_count(dqm, qpd, q); - execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + if (!dqm->dev->shared_resources.enable_mes) { + retval = execute_queues_cpsch(dqm, + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + } else { + retval = add_queue_mes(dqm, q, qpd); + if (retval) + goto cleanup_queue; + } } /* @@ -1485,6 +1688,13 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, dqm_unlock(dqm); return retval; +cleanup_queue: + qpd->queue_count--; + list_del(&q->list); + if (q->properties.is_active) + decrement_queue_count(dqm, qpd, q); + mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); + dqm_unlock(dqm); out_deallocate_doorbell: deallocate_doorbell(qpd, q); out_deallocate_sdma_queue: @@ -1572,13 +1782,7 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, queue_preemption_timeout_ms); if (retval) { pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n"); - dqm->is_hws_hang = true; - /* It's possible we're detecting a HWS hang in the - * middle of a GPU reset. No need to schedule another - * reset in this case. - */ - if (!dqm->is_resetting) - schedule_work(&dqm->hw_exception_work); + kfd_hws_hang(dqm); return retval; } @@ -1683,11 +1887,15 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm, list_del(&q->list); qpd->queue_count--; if (q->properties.is_active) { - decrement_queue_count(dqm, qpd, q); - retval = execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); - if (retval == -ETIME) - qpd->reset_wavefronts = true; + if (!dqm->dev->shared_resources.enable_mes) { + decrement_queue_count(dqm, qpd, q); + retval = execute_queues_cpsch(dqm, + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + if (retval == -ETIME) + qpd->reset_wavefronts = true; + } else { + retval = remove_queue_mes(dqm, q, qpd); + } } /* @@ -1941,9 +2149,17 @@ static int process_termination_cpsch(struct device_queue_manager *dqm, else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) deallocate_sdma_queue(dqm, q); - if (q->properties.is_active) + if (q->properties.is_active) { decrement_queue_count(dqm, qpd, q); + if (dqm->dev->shared_resources.enable_mes) { + retval = remove_queue_mes(dqm, q, qpd); + if (retval) + pr_err("Failed to remove queue %d\n", + q->properties.queue_id); + } + } + dqm->total_queue_count--; } @@ -1958,7 +2174,9 @@ static int process_termination_cpsch(struct device_queue_manager *dqm, } } - retval = execute_queues_cpsch(dqm, filter, 0); + if (!dqm->dev->shared_resources.enable_mes) + retval = execute_queues_cpsch(dqm, filter, 0); + if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) { pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev); dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process); @@ -2133,7 +2351,9 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev) break; default: - if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1)) + if (KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0)) + device_queue_manager_init_v11(&dqm->asic_ops); + else if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1)) device_queue_manager_init_v10_navi10(&dqm->asic_ops); else if (KFD_GC_VERSION(dev) >= IP_VERSION(9, 0, 1)) device_queue_manager_init_v9(&dqm->asic_ops); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index 3d539d6483e0..a537b9ef3e16 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -35,6 +35,9 @@ #define VMID_NUM 16 +#define KFD_MES_PROCESS_QUANTUM 100000 +#define KFD_MES_GANG_QUANTUM 10000 + struct device_process_node { struct qcm_process_device *qpd; struct list_head list; @@ -267,6 +270,8 @@ void device_queue_manager_init_v9( struct device_queue_manager_asic_ops *asic_ops); void device_queue_manager_init_v10_navi10( struct device_queue_manager_asic_ops *asic_ops); +void device_queue_manager_init_v11( + struct device_queue_manager_asic_ops *asic_ops); void program_sh_mem_settings(struct device_queue_manager *dqm, struct qcm_process_device *qpd); unsigned int get_cp_queues_num(struct device_queue_manager *dqm); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v11.c new file mode 100644 index 000000000000..2e129da7acb4 --- /dev/null +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v11.c @@ -0,0 +1,81 @@ +/* + * Copyright 2021 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR + * OTHER DEALINGS IN THE SOFTWARE. + * + */ + +#include "kfd_device_queue_manager.h" +#include "gc/gc_11_0_0_offset.h" +#include "gc/gc_11_0_0_sh_mask.h" +#include "soc21_enum.h" + +static int update_qpd_v11(struct device_queue_manager *dqm, + struct qcm_process_device *qpd); +static void init_sdma_vm_v11(struct device_queue_manager *dqm, struct queue *q, + struct qcm_process_device *qpd); + +void device_queue_manager_init_v11( + struct device_queue_manager_asic_ops *asic_ops) +{ + asic_ops->update_qpd = update_qpd_v11; + asic_ops->init_sdma_vm = init_sdma_vm_v11; + asic_ops->mqd_manager_init = mqd_manager_init_v11; +} + +static uint32_t compute_sh_mem_bases_64bit(struct kfd_process_device *pdd) +{ + uint32_t shared_base = pdd->lds_base >> 48; + uint32_t private_base = pdd->scratch_base >> 48; + + return (shared_base << SH_MEM_BASES__SHARED_BASE__SHIFT) | + private_base; +} + +static int update_qpd_v11(struct device_queue_manager *dqm, + struct qcm_process_device *qpd) +{ + struct kfd_process_device *pdd; + + pdd = qpd_to_pdd(qpd); + + /* check if sh_mem_config register already configured */ + if (qpd->sh_mem_config == 0) { + qpd->sh_mem_config = + (SH_MEM_ALIGNMENT_MODE_UNALIGNED << + SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT) | + (3 << SH_MEM_CONFIG__INITIAL_INST_PREFETCH__SHIFT); + + qpd->sh_mem_ape1_limit = 0; + qpd->sh_mem_ape1_base = 0; + } + + qpd->sh_mem_bases = compute_sh_mem_bases_64bit(pdd); + + pr_debug("sh_mem_bases 0x%X\n", qpd->sh_mem_bases); + + return 0; +} + +static void init_sdma_vm_v11(struct device_queue_manager *dqm, struct queue *q, + struct qcm_process_device *qpd) +{ + /* Not needed on SDMAv4 onwards any more */ + q->properties.sdma_vm_addr = 0; +} diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c index 5401b6317f25..cb3d2ccc5100 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c @@ -49,9 +49,13 @@ /* # of doorbell bytes allocated for each process. */ size_t kfd_doorbell_process_slice(struct kfd_dev *kfd) { - return roundup(kfd->device_info.doorbell_size * - KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, - PAGE_SIZE); + if (!kfd->shared_resources.enable_mes) + return roundup(kfd->device_info.doorbell_size * + KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, + PAGE_SIZE); + else + return amdgpu_mes_doorbell_process_slice( + (struct amdgpu_device *)kfd->adev); } /* Doorbell calculations for device init. */ @@ -62,6 +66,16 @@ int kfd_doorbell_init(struct kfd_dev *kfd) size_t doorbell_process_limit; /* + * With MES enabled, just set the doorbell base as it is needed + * to calculate doorbell physical address. + */ + if (kfd->shared_resources.enable_mes) { + kfd->doorbell_base = + kfd->shared_resources.doorbell_physical_address; + return 0; + } + + /* * We start with calculations in bytes because the input data might * only be byte-aligned. * Only after we have done the rounding can we assume any alignment. @@ -237,10 +251,16 @@ unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, * the process's doorbells. The offset returned is in dword * units regardless of the ASIC-dependent doorbell size. */ - return kfd->doorbell_base_dw_offset + - pdd->doorbell_index - * kfd_doorbell_process_slice(kfd) / sizeof(u32) + - doorbell_id * kfd->device_info.doorbell_size / sizeof(u32); + if (!kfd->shared_resources.enable_mes) + return kfd->doorbell_base_dw_offset + + pdd->doorbell_index + * kfd_doorbell_process_slice(kfd) / sizeof(u32) + + doorbell_id * + kfd->device_info.doorbell_size / sizeof(u32); + else + return amdgpu_mes_get_doorbell_dw_offset_in_bar( + (struct amdgpu_device *)kfd->adev, + pdd->doorbell_index, doorbell_id); } uint64_t kfd_get_number_elems(struct kfd_dev *kfd) @@ -261,8 +281,16 @@ phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd) int kfd_alloc_process_doorbells(struct kfd_dev *kfd, unsigned int *doorbell_index) { - int r = ida_simple_get(&kfd->doorbell_ida, 1, kfd->max_doorbell_slices, - GFP_KERNEL); + int r = 0; + + if (!kfd->shared_resources.enable_mes) + r = ida_simple_get(&kfd->doorbell_ida, 1, + kfd->max_doorbell_slices, GFP_KERNEL); + else + r = amdgpu_mes_alloc_process_doorbells( + (struct amdgpu_device *)kfd->adev, + doorbell_index); + if (r > 0) *doorbell_index = r; @@ -271,6 +299,12 @@ int kfd_alloc_process_doorbells(struct kfd_dev *kfd, unsigned int *doorbell_inde void kfd_free_process_doorbells(struct kfd_dev *kfd, unsigned int doorbell_index) { - if (doorbell_index) - ida_simple_remove(&kfd->doorbell_ida, doorbell_index); + if (doorbell_index) { + if (!kfd->shared_resources.enable_mes) + ida_simple_remove(&kfd->doorbell_ida, doorbell_index); + else + amdgpu_mes_free_process_doorbells( + (struct amdgpu_device *)kfd->adev, + doorbell_index); + } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_events.c index 64f4a51cc880..4df9c36146ba 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_events.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.c @@ -128,8 +128,8 @@ static int allocate_event_notification_slot(struct kfd_process *p, } /* - * Assumes that p->event_mutex is held and of course that p is not going - * away (current or locked). + * Assumes that p->event_mutex or rcu_readlock is held and of course that p is + * not going away. */ static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id) { @@ -238,12 +238,24 @@ static int create_other_event(struct kfd_process *p, struct kfd_event *ev, const return 0; } -void kfd_event_init_process(struct kfd_process *p) +int kfd_event_init_process(struct kfd_process *p) { + int id; + mutex_init(&p->event_mutex); idr_init(&p->event_idr); p->signal_page = NULL; - p->signal_event_count = 0; + p->signal_event_count = 1; + /* Allocate event ID 0. It is used for a fast path to ignore bogus events + * that are sent by the CP without a context ID + */ + id = idr_alloc(&p->event_idr, NULL, 0, 1, GFP_KERNEL); + if (id < 0) { + idr_destroy(&p->event_idr); + mutex_destroy(&p->event_mutex); + return id; + } + return 0; } static void destroy_event(struct kfd_process *p, struct kfd_event *ev) @@ -251,16 +263,18 @@ static void destroy_event(struct kfd_process *p, struct kfd_event *ev) struct kfd_event_waiter *waiter; /* Wake up pending waiters. They will return failure */ + spin_lock(&ev->lock); list_for_each_entry(waiter, &ev->wq.head, wait.entry) - waiter->event = NULL; + WRITE_ONCE(waiter->event, NULL); wake_up_all(&ev->wq); + spin_unlock(&ev->lock); if (ev->type == KFD_EVENT_TYPE_SIGNAL || ev->type == KFD_EVENT_TYPE_DEBUG) p->signal_event_count--; idr_remove(&p->event_idr, ev->event_id); - kfree(ev); + kfree_rcu(ev, rcu); } static void destroy_events(struct kfd_process *p) @@ -269,8 +283,10 @@ static void destroy_events(struct kfd_process *p) uint32_t id; idr_for_each_entry(&p->event_idr, ev, id) - destroy_event(p, ev); + if (ev) + destroy_event(p, ev); idr_destroy(&p->event_idr); + mutex_destroy(&p->event_mutex); } /* @@ -392,6 +408,7 @@ int kfd_event_create(struct file *devkfd, struct kfd_process *p, ev->auto_reset = auto_reset; ev->signaled = false; + spin_lock_init(&ev->lock); init_waitqueue_head(&ev->wq); *event_page_offset = 0; @@ -466,6 +483,7 @@ int kfd_criu_restore_event(struct file *devkfd, ev->auto_reset = ev_priv->auto_reset; ev->signaled = ev_priv->signaled; + spin_lock_init(&ev->lock); init_waitqueue_head(&ev->wq); mutex_lock(&p->event_mutex); @@ -609,13 +627,13 @@ static void set_event(struct kfd_event *ev) /* Auto reset if the list is non-empty and we're waking * someone. waitqueue_active is safe here because we're - * protected by the p->event_mutex, which is also held when + * protected by the ev->lock, which is also held when * updating the wait queues in kfd_wait_on_events. */ ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq); list_for_each_entry(waiter, &ev->wq.head, wait.entry) - waiter->activated = true; + WRITE_ONCE(waiter->activated, true); wake_up_all(&ev->wq); } @@ -626,16 +644,23 @@ int kfd_set_event(struct kfd_process *p, uint32_t event_id) int ret = 0; struct kfd_event *ev; - mutex_lock(&p->event_mutex); + rcu_read_lock(); ev = lookup_event_by_id(p, event_id); + if (!ev) { + ret = -EINVAL; + goto unlock_rcu; + } + spin_lock(&ev->lock); - if (ev && event_can_be_cpu_signaled(ev)) + if (event_can_be_cpu_signaled(ev)) set_event(ev); else ret = -EINVAL; - mutex_unlock(&p->event_mutex); + spin_unlock(&ev->lock); +unlock_rcu: + rcu_read_unlock(); return ret; } @@ -650,23 +675,30 @@ int kfd_reset_event(struct kfd_process *p, uint32_t event_id) int ret = 0; struct kfd_event *ev; - mutex_lock(&p->event_mutex); + rcu_read_lock(); ev = lookup_event_by_id(p, event_id); + if (!ev) { + ret = -EINVAL; + goto unlock_rcu; + } + spin_lock(&ev->lock); - if (ev && event_can_be_cpu_signaled(ev)) + if (event_can_be_cpu_signaled(ev)) reset_event(ev); else ret = -EINVAL; - mutex_unlock(&p->event_mutex); + spin_unlock(&ev->lock); +unlock_rcu: + rcu_read_unlock(); return ret; } static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev) { - page_slots(p->signal_page)[ev->event_id] = UNSIGNALED_EVENT_SLOT; + WRITE_ONCE(page_slots(p->signal_page)[ev->event_id], UNSIGNALED_EVENT_SLOT); } static void set_event_from_interrupt(struct kfd_process *p, @@ -674,7 +706,9 @@ static void set_event_from_interrupt(struct kfd_process *p, { if (ev && event_can_be_gpu_signaled(ev)) { acknowledge_signal(p, ev); + spin_lock(&ev->lock); set_event(ev); + spin_unlock(&ev->lock); } } @@ -693,7 +727,7 @@ void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, if (!p) return; /* Presumably process exited. */ - mutex_lock(&p->event_mutex); + rcu_read_lock(); if (valid_id_bits) ev = lookup_signaled_event_by_partial_id(p, partial_id, @@ -721,7 +755,7 @@ void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, if (id >= KFD_SIGNAL_EVENT_LIMIT) break; - if (slots[id] != UNSIGNALED_EVENT_SLOT) + if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) set_event_from_interrupt(p, ev); } } else { @@ -729,15 +763,15 @@ void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, * iterate over the signal slots and lookup * only signaled events from the IDR. */ - for (id = 0; id < KFD_SIGNAL_EVENT_LIMIT; id++) - if (slots[id] != UNSIGNALED_EVENT_SLOT) { + for (id = 1; id < KFD_SIGNAL_EVENT_LIMIT; id++) + if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) { ev = lookup_event_by_id(p, id); set_event_from_interrupt(p, ev); } } } - mutex_unlock(&p->event_mutex); + rcu_read_unlock(); kfd_unref_process(p); } @@ -760,7 +794,7 @@ static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events) return event_waiters; } -static int init_event_waiter_get_status(struct kfd_process *p, +static int init_event_waiter(struct kfd_process *p, struct kfd_event_waiter *waiter, uint32_t event_id) { @@ -769,22 +803,15 @@ static int init_event_waiter_get_status(struct kfd_process *p, if (!ev) return -EINVAL; + spin_lock(&ev->lock); waiter->event = ev; waiter->activated = ev->signaled; ev->signaled = ev->signaled && !ev->auto_reset; - - return 0; -} - -static void init_event_waiter_add_to_waitlist(struct kfd_event_waiter *waiter) -{ - struct kfd_event *ev = waiter->event; - - /* Only add to the wait list if we actually need to - * wait on this event. - */ if (!waiter->activated) add_wait_queue(&ev->wq, &waiter->wait); + spin_unlock(&ev->lock); + + return 0; } /* test_event_condition - Test condition of events being waited for @@ -804,10 +831,10 @@ static uint32_t test_event_condition(bool all, uint32_t num_events, uint32_t activated_count = 0; for (i = 0; i < num_events; i++) { - if (!event_waiters[i].event) + if (!READ_ONCE(event_waiters[i].event)) return KFD_IOC_WAIT_RESULT_FAIL; - if (event_waiters[i].activated) { + if (READ_ONCE(event_waiters[i].activated)) { if (!all) return KFD_IOC_WAIT_RESULT_COMPLETE; @@ -836,6 +863,8 @@ static int copy_signaled_event_data(uint32_t num_events, for (i = 0; i < num_events; i++) { waiter = &event_waiters[i]; event = waiter->event; + if (!event) + return -EINVAL; /* event was destroyed */ if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) { dst = &data[i].memory_exception_data; src = &event->memory_exception_data; @@ -846,11 +875,8 @@ static int copy_signaled_event_data(uint32_t num_events, } return 0; - } - - static long user_timeout_to_jiffies(uint32_t user_timeout_ms) { if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE) @@ -874,9 +900,12 @@ static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters) uint32_t i; for (i = 0; i < num_events; i++) - if (waiters[i].event) + if (waiters[i].event) { + spin_lock(&waiters[i].event->lock); remove_wait_queue(&waiters[i].event->wq, &waiters[i].wait); + spin_unlock(&waiters[i].event->lock); + } kfree(waiters); } @@ -900,6 +929,9 @@ int kfd_wait_on_events(struct kfd_process *p, goto out; } + /* Use p->event_mutex here to protect against concurrent creation and + * destruction of events while we initialize event_waiters. + */ mutex_lock(&p->event_mutex); for (i = 0; i < num_events; i++) { @@ -911,8 +943,8 @@ int kfd_wait_on_events(struct kfd_process *p, goto out_unlock; } - ret = init_event_waiter_get_status(p, &event_waiters[i], - event_data.event_id); + ret = init_event_waiter(p, &event_waiters[i], + event_data.event_id); if (ret) goto out_unlock; } @@ -930,10 +962,6 @@ int kfd_wait_on_events(struct kfd_process *p, goto out_unlock; } - /* Add to wait lists if we need to wait. */ - for (i = 0; i < num_events; i++) - init_event_waiter_add_to_waitlist(&event_waiters[i]); - mutex_unlock(&p->event_mutex); while (true) { @@ -978,14 +1006,19 @@ int kfd_wait_on_events(struct kfd_process *p, } __set_current_state(TASK_RUNNING); + mutex_lock(&p->event_mutex); /* copy_signaled_event_data may sleep. So this has to happen * after the task state is set back to RUNNING. + * + * The event may also have been destroyed after signaling. So + * copy_signaled_event_data also must confirm that the event + * still exists. Therefore this must be under the p->event_mutex + * which is also held when events are destroyed. */ if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) ret = copy_signaled_event_data(num_events, event_waiters, events); - mutex_lock(&p->event_mutex); out_unlock: free_waiters(num_events, event_waiters); mutex_unlock(&p->event_mutex); @@ -1044,8 +1077,7 @@ int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma) } /* - * Assumes that p->event_mutex is held and of course - * that p is not going away (current or locked). + * Assumes that p is not going away. */ static void lookup_events_by_type_and_signal(struct kfd_process *p, int type, void *event_data) @@ -1057,6 +1089,8 @@ static void lookup_events_by_type_and_signal(struct kfd_process *p, ev_data = (struct kfd_hsa_memory_exception_data *) event_data; + rcu_read_lock(); + id = KFD_FIRST_NONSIGNAL_EVENT_ID; idr_for_each_entry_continue(&p->event_idr, ev, id) if (ev->type == type) { @@ -1064,9 +1098,11 @@ static void lookup_events_by_type_and_signal(struct kfd_process *p, dev_dbg(kfd_device, "Event found: id %X type %d", ev->event_id, ev->type); + spin_lock(&ev->lock); set_event(ev); if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data) ev->memory_exception_data = *ev_data; + spin_unlock(&ev->lock); } if (type == KFD_EVENT_TYPE_MEMORY) { @@ -1089,6 +1125,8 @@ static void lookup_events_by_type_and_signal(struct kfd_process *p, p->lead_thread->pid, p->pasid); } } + + rcu_read_unlock(); } #ifdef KFD_SUPPORT_IOMMU_V2 @@ -1164,16 +1202,10 @@ void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid, if (KFD_GC_VERSION(dev) != IP_VERSION(9, 1, 0) && KFD_GC_VERSION(dev) != IP_VERSION(9, 2, 2) && - KFD_GC_VERSION(dev) != IP_VERSION(9, 3, 0)) { - mutex_lock(&p->event_mutex); - - /* Lookup events by type and signal them */ + KFD_GC_VERSION(dev) != IP_VERSION(9, 3, 0)) lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY, &memory_exception_data); - mutex_unlock(&p->event_mutex); - } - kfd_unref_process(p); } #endif /* KFD_SUPPORT_IOMMU_V2 */ @@ -1190,12 +1222,7 @@ void kfd_signal_hw_exception_event(u32 pasid) if (!p) return; /* Presumably process exited. */ - mutex_lock(&p->event_mutex); - - /* Lookup events by type and signal them */ lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL); - - mutex_unlock(&p->event_mutex); kfd_unref_process(p); } @@ -1231,16 +1258,19 @@ void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, info->prot_write ? 1 : 0; memory_exception_data.failure.imprecise = 0; } - mutex_lock(&p->event_mutex); + + rcu_read_lock(); id = KFD_FIRST_NONSIGNAL_EVENT_ID; idr_for_each_entry_continue(&p->event_idr, ev, id) if (ev->type == KFD_EVENT_TYPE_MEMORY) { + spin_lock(&ev->lock); ev->memory_exception_data = memory_exception_data; set_event(ev); + spin_unlock(&ev->lock); } - mutex_unlock(&p->event_mutex); + rcu_read_unlock(); kfd_unref_process(p); } @@ -1274,22 +1304,28 @@ void kfd_signal_reset_event(struct kfd_dev *dev) continue; } - mutex_lock(&p->event_mutex); + rcu_read_lock(); + id = KFD_FIRST_NONSIGNAL_EVENT_ID; idr_for_each_entry_continue(&p->event_idr, ev, id) { if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) { + spin_lock(&ev->lock); ev->hw_exception_data = hw_exception_data; ev->hw_exception_data.gpu_id = user_gpu_id; set_event(ev); + spin_unlock(&ev->lock); } if (ev->type == KFD_EVENT_TYPE_MEMORY && reset_cause == KFD_HW_EXCEPTION_ECC) { + spin_lock(&ev->lock); ev->memory_exception_data = memory_exception_data; ev->memory_exception_data.gpu_id = user_gpu_id; set_event(ev); + spin_unlock(&ev->lock); } } - mutex_unlock(&p->event_mutex); + + rcu_read_unlock(); } srcu_read_unlock(&kfd_processes_srcu, idx); } @@ -1322,19 +1358,25 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid) memory_exception_data.gpu_id = user_gpu_id; memory_exception_data.failure.imprecise = true; - mutex_lock(&p->event_mutex); + rcu_read_lock(); + idr_for_each_entry_continue(&p->event_idr, ev, id) { if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) { + spin_lock(&ev->lock); ev->hw_exception_data = hw_exception_data; set_event(ev); + spin_unlock(&ev->lock); } if (ev->type == KFD_EVENT_TYPE_MEMORY) { + spin_lock(&ev->lock); ev->memory_exception_data = memory_exception_data; set_event(ev); + spin_unlock(&ev->lock); } } - mutex_unlock(&p->event_mutex); + + rcu_read_unlock(); /* user application will handle SIGBUS signal */ send_sig(SIGBUS, p->lead_thread, 0); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.h b/drivers/gpu/drm/amd/amdkfd/kfd_events.h index 1238af11916e..1c62c8dd6460 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_events.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.h @@ -59,6 +59,7 @@ struct kfd_event { int type; + spinlock_t lock; wait_queue_head_t wq; /* List of event waiters. */ /* Only for signal events. */ @@ -69,6 +70,8 @@ struct kfd_event { struct kfd_hsa_memory_exception_data memory_exception_data; struct kfd_hsa_hw_exception_data hw_exception_data; }; + + struct rcu_head rcu; /* for asynchronous kfree_rcu */ }; #define KFD_EVENT_TIMEOUT_IMMEDIATE 0 diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c new file mode 100644 index 000000000000..a6fcbeeb7428 --- /dev/null +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c @@ -0,0 +1,384 @@ +/* + * Copyright 2021 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR + * OTHER DEALINGS IN THE SOFTWARE. + */ + +#include "kfd_priv.h" +#include "kfd_events.h" +#include "soc15_int.h" +#include "kfd_device_queue_manager.h" +#include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" +#include "kfd_smi_events.h" + +/* + * GFX11 SQ Interrupts + * + * There are 3 encoding types of interrupts sourced from SQ sent as a 44-bit + * packet to the Interrupt Handler: + * Auto - Generated by the SQG (various cmd overflows, timestamps etc) + * Wave - Generated by S_SENDMSG through a shader program + * Error - HW generated errors (Illegal instructions, Memviols, EDC etc) + * + * The 44-bit packet is mapped as {context_id1[7:0],context_id0[31:0]} plus + * 4-bits for VMID (SOC15_VMID_FROM_IH_ENTRY) as such: + * + * - context_id1[7:6] + * Encoding type (0 = Auto, 1 = Wave, 2 = Error) + * + * - context_id0[26] + * PRIV bit indicates that Wave S_SEND or error occurred within trap + * + * - context_id0[24:0] + * 25-bit data with the following layout per encoding type: + * Auto - only context_id0[8:0] is used, which reports various interrupts + * generated by SQG. The rest is 0. + * Wave - user data sent from m0 via S_SENDMSG (context_id0[23:0]) + * Error - Error Type (context_id0[24:21]), Error Details (context_id0[20:0]) + * + * The other context_id bits show coordinates (SE/SH/CU/SIMD/WGP) for wave + * S_SENDMSG and Errors. These are 0 for Auto. + */ + +enum SQ_INTERRUPT_WORD_ENCODING { + SQ_INTERRUPT_WORD_ENCODING_AUTO = 0x0, + SQ_INTERRUPT_WORD_ENCODING_INST, + SQ_INTERRUPT_WORD_ENCODING_ERROR, +}; + +enum SQ_INTERRUPT_ERROR_TYPE { + SQ_INTERRUPT_ERROR_TYPE_EDC_FUE = 0x0, + SQ_INTERRUPT_ERROR_TYPE_ILLEGAL_INST, + SQ_INTERRUPT_ERROR_TYPE_MEMVIOL, + SQ_INTERRUPT_ERROR_TYPE_EDC_FED, +}; + +/* SQ_INTERRUPT_WORD_AUTO_CTXID */ +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE__SHIFT 0 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__WLT__SHIFT 1 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_BUF_FULL__SHIFT 2 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__REG_TIMESTAMP__SHIFT 3 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__CMD_TIMESTAMP__SHIFT 4 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__HOST_CMD_OVERFLOW__SHIFT 5 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__HOST_REG_OVERFLOW__SHIFT 6 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__IMMED_OVERFLOW__SHIFT 7 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_UTC_ERROR__SHIFT 8 +#define SQ_INTERRUPT_WORD_AUTO_CTXID1__ENCODING__SHIFT 6 + +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_MASK 0x00000001 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__WLT_MASK 0x00000002 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_BUF_FULL_MASK 0x00000004 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__REG_TIMESTAMP_MASK 0x00000008 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__CMD_TIMESTAMP_MASK 0x00000010 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__HOST_CMD_OVERFLOW_MASK 0x00000020 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__HOST_REG_OVERFLOW_MASK 0x00000040 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__IMMED_OVERFLOW_MASK 0x00000080 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_UTC_ERROR_MASK 0x00000100 +#define SQ_INTERRUPT_WORD_AUTO_CTXID1__ENCODING_MASK 0x000000c0 + +/* SQ_INTERRUPT_WORD_WAVE_CTXID */ +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__DATA__SHIFT 0 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__SH_ID__SHIFT 25 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__PRIV__SHIFT 26 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__WAVE_ID__SHIFT 27 +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__SIMD_ID__SHIFT 0 +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__WGP_ID__SHIFT 2 +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__ENCODING__SHIFT 6 + +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__DATA_MASK 0x00ffffff /* [23:0] */ +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__SH_ID_MASK 0x02000000 /* [25] */ +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__PRIV_MASK 0x04000000 /* [26] */ +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__WAVE_ID_MASK 0xf8000000 /* [31:27] */ +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__SIMD_ID_MASK 0x00000003 /* [33:32] */ +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__WGP_ID_MASK 0x0000003c /* [37:34] */ +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__ENCODING_MASK 0x000000c0 /* [39:38] */ + +/* SQ_INTERRUPT_WORD_ERROR_CTXID */ +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__DETAIL__SHIFT 0 +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__TYPE__SHIFT 21 +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__SH_ID__SHIFT 25 +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__PRIV__SHIFT 26 +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__WAVE_ID__SHIFT 27 +#define SQ_INTERRUPT_WORD_ERROR_CTXID1__SIMD_ID__SHIFT 0 +#define SQ_INTERRUPT_WORD_ERROR_CTXID1__WGP_ID__SHIFT 2 +#define SQ_INTERRUPT_WORD_ERROR_CTXID1__ENCODING__SHIFT 6 + +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__DETAIL_MASK 0x001fffff /* [20:0] */ +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__TYPE_MASK 0x01e00000 /* [24:21] */ +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__SH_ID_MASK 0x02000000 /* [25] */ +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__PRIV_MASK 0x04000000 /* [26] */ +#define SQ_INTERRUPT_WORD_ERROR_CTXID0__WAVE_ID_MASK 0xf8000000 /* [31:27] */ +#define SQ_INTERRUPT_WORD_ERROR_CTXID1__SIMD_ID_MASK 0x00000003 /* [33:32] */ +#define SQ_INTERRUPT_WORD_ERROR_CTXID1__WGP_ID_MASK 0x0000003c /* [37:34] */ +#define SQ_INTERRUPT_WORD_ERROR_CTXID1__ENCODING_MASK 0x000000c0 /* [39:38] */ + +/* + * The debugger will send user data(m0) with PRIV=1 to indicate it requires + * notification from the KFD with the following queue id (DOORBELL_ID) and + * trap code (TRAP_CODE). + */ +#define KFD_CTXID0_TRAP_CODE_SHIFT 10 +#define KFD_CTXID0_TRAP_CODE_MASK 0xfffc00 +#define KFD_CTXID0_CP_BAD_OP_ECODE_MASK 0x3ffffff +#define KFD_CTXID0_DOORBELL_ID_MASK 0x0003ff + +#define KFD_CTXID0_TRAP_CODE(ctxid0) (((ctxid0) & \ + KFD_CTXID0_TRAP_CODE_MASK) >> \ + KFD_CTXID0_TRAP_CODE_SHIFT) +#define KFD_CTXID0_CP_BAD_OP_ECODE(ctxid0) (((ctxid0) & \ + KFD_CTXID0_CP_BAD_OP_ECODE_MASK) >> \ + KFD_CTXID0_TRAP_CODE_SHIFT) +#define KFD_CTXID0_DOORBELL_ID(ctxid0) ((ctxid0) & \ + KFD_CTXID0_DOORBELL_ID_MASK) + +static void print_sq_intr_info_auto(uint32_t context_id0, uint32_t context_id1) +{ + pr_debug( + "sq_intr: auto, ttrace %d, wlt %d, ttrace_buf_full %d, reg_tms %d, cmd_tms %d, host_cmd_ovf %d, host_reg_ovf %d, immed_ovf %d, ttrace_utc_err %d\n", + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, THREAD_TRACE), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, WLT), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, THREAD_TRACE_BUF_FULL), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, REG_TIMESTAMP), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, CMD_TIMESTAMP), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, HOST_CMD_OVERFLOW), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, HOST_REG_OVERFLOW), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, IMMED_OVERFLOW), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, THREAD_TRACE_UTC_ERROR)); +} + +static void print_sq_intr_info_inst(uint32_t context_id0, uint32_t context_id1) +{ + pr_debug( + "sq_intr: inst, data 0x%08x, sh %d, priv %d, wave_id %d, simd_id %d, wgp_id %d\n", + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, DATA), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, SH_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, PRIV), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, WAVE_ID), + REG_GET_FIELD(context_id1, SQ_INTERRUPT_WORD_WAVE_CTXID1, SIMD_ID), + REG_GET_FIELD(context_id1, SQ_INTERRUPT_WORD_WAVE_CTXID1, WGP_ID)); +} + +static void print_sq_intr_info_error(uint32_t context_id0, uint32_t context_id1) +{ + pr_warn( + "sq_intr: error, detail 0x%08x, type %d, sh %d, priv %d, wave_id %d, simd_id %d, wgp_id %d\n", + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_ERROR_CTXID0, DETAIL), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_ERROR_CTXID0, TYPE), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_ERROR_CTXID0, SH_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_ERROR_CTXID0, PRIV), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_ERROR_CTXID0, WAVE_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_ERROR_CTXID1, SIMD_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_ERROR_CTXID1, WGP_ID)); +} + +static void event_interrupt_poison_consumption_v11(struct kfd_dev *dev, + uint16_t pasid, uint16_t source_id) +{ + int ret = -EINVAL; + struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); + + if (!p) + return; + + /* all queues of a process will be unmapped in one time */ + if (atomic_read(&p->poison)) { + kfd_unref_process(p); + return; + } + + atomic_set(&p->poison, 1); + kfd_unref_process(p); + + switch (source_id) { + case SOC15_INTSRC_SQ_INTERRUPT_MSG: + if (dev->dqm->ops.reset_queues) + ret = dev->dqm->ops.reset_queues(dev->dqm, pasid); + break; + case SOC21_INTSRC_SDMA_ECC: + default: + break; + } + + kfd_signal_poison_consumed_event(dev, pasid); + + /* resetting queue passes, do page retirement without gpu reset + resetting queue fails, fallback to gpu reset solution */ + if (!ret) + amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, false); + else + amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, true); +} + +static bool event_interrupt_isr_v11(struct kfd_dev *dev, + const uint32_t *ih_ring_entry, + uint32_t *patched_ihre, + bool *patched_flag) +{ + uint16_t source_id, client_id, pasid, vmid; + const uint32_t *data = ih_ring_entry; + uint32_t context_id0; + + source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry); + client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); + /* Only handle interrupts from KFD VMIDs */ + vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); + if (/*!KFD_IRQ_IS_FENCE(client_id, source_id) &&*/ + (vmid < dev->vm_info.first_vmid_kfd || + vmid > dev->vm_info.last_vmid_kfd)) + return false; + + pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry); + context_id0 = SOC15_CONTEXT_ID0_FROM_IH_ENTRY(ih_ring_entry); + + if ((source_id == SOC15_INTSRC_CP_END_OF_PIPE) && + (context_id0 & AMDGPU_FENCE_MES_QUEUE_FLAG)) + return false; + + pr_debug("client id 0x%x, source id %d, vmid %d, pasid 0x%x. raw data:\n", + client_id, source_id, vmid, pasid); + pr_debug("%8X, %8X, %8X, %8X, %8X, %8X, %8X, %8X.\n", + data[0], data[1], data[2], data[3], + data[4], data[5], data[6], data[7]); + + /* If there is no valid PASID, it's likely a bug */ + if (WARN_ONCE(pasid == 0, "Bug: No PASID in KFD interrupt")) + return false; + + /* Interrupt types we care about: various signals and faults. + * They will be forwarded to a work queue (see below). + */ + return source_id == SOC15_INTSRC_CP_END_OF_PIPE || + source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG || + source_id == SOC15_INTSRC_CP_BAD_OPCODE || + source_id == SOC21_INTSRC_SDMA_TRAP || + /* KFD_IRQ_IS_FENCE(client_id, source_id) || */ + (((client_id == SOC21_IH_CLIENTID_VMC) || + ((client_id == SOC21_IH_CLIENTID_GFX) && + (source_id == UTCL2_1_0__SRCID__FAULT))) && + !amdgpu_no_queue_eviction_on_vm_fault); +} + +static void event_interrupt_wq_v11(struct kfd_dev *dev, + const uint32_t *ih_ring_entry) +{ + uint16_t source_id, client_id, ring_id, pasid, vmid; + uint32_t context_id0, context_id1; + uint8_t sq_int_enc, sq_int_errtype, sq_int_priv; + struct kfd_vm_fault_info info = {0}; + struct kfd_hsa_memory_exception_data exception_data; + + source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry); + client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); + ring_id = SOC15_RING_ID_FROM_IH_ENTRY(ih_ring_entry); + pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry); + vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); + context_id0 = SOC15_CONTEXT_ID0_FROM_IH_ENTRY(ih_ring_entry); + context_id1 = SOC15_CONTEXT_ID1_FROM_IH_ENTRY(ih_ring_entry); + + /* VMC, UTCL2 */ + if (client_id == SOC21_IH_CLIENTID_VMC || + ((client_id == SOC21_IH_CLIENTID_GFX) && + (source_id == UTCL2_1_0__SRCID__FAULT))) { + + info.vmid = vmid; + info.mc_id = client_id; + info.page_addr = ih_ring_entry[4] | + (uint64_t)(ih_ring_entry[5] & 0xf) << 32; + info.prot_valid = ring_id & 0x08; + info.prot_read = ring_id & 0x10; + info.prot_write = ring_id & 0x20; + + memset(&exception_data, 0, sizeof(exception_data)); + exception_data.gpu_id = dev->id; + exception_data.va = (info.page_addr) << PAGE_SHIFT; + exception_data.failure.NotPresent = info.prot_valid ? 1 : 0; + exception_data.failure.NoExecute = info.prot_exec ? 1 : 0; + exception_data.failure.ReadOnly = info.prot_write ? 1 : 0; + exception_data.failure.imprecise = 0; + + /*kfd_set_dbg_ev_from_interrupt(dev, pasid, -1, + KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION), + &exception_data, sizeof(exception_data));*/ + kfd_smi_event_update_vmfault(dev, pasid); + + /* GRBM, SDMA, SE, PMM */ + } else if (client_id == SOC21_IH_CLIENTID_GRBM_CP || + client_id == SOC21_IH_CLIENTID_GFX) { + + /* CP */ + if (source_id == SOC15_INTSRC_CP_END_OF_PIPE) + kfd_signal_event_interrupt(pasid, context_id0, 32); + /*else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE) + kfd_set_dbg_ev_from_interrupt(dev, pasid, + KFD_CTXID0_DOORBELL_ID(context_id0), + KFD_EC_MASK(KFD_CTXID0_CP_BAD_OP_ECODE(context_id0)), + NULL, 0);*/ + + /* SDMA */ + else if (source_id == SOC21_INTSRC_SDMA_TRAP) + kfd_signal_event_interrupt(pasid, context_id0 & 0xfffffff, 28); + else if (source_id == SOC21_INTSRC_SDMA_ECC) { + event_interrupt_poison_consumption_v11(dev, pasid, source_id); + return; + } + + /* SQ */ + else if (source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG) { + sq_int_enc = REG_GET_FIELD(context_id1, + SQ_INTERRUPT_WORD_WAVE_CTXID1, ENCODING); + switch (sq_int_enc) { + case SQ_INTERRUPT_WORD_ENCODING_AUTO: + print_sq_intr_info_auto(context_id0, context_id1); + break; + case SQ_INTERRUPT_WORD_ENCODING_INST: + print_sq_intr_info_inst(context_id0, context_id1); + sq_int_priv = REG_GET_FIELD(context_id0, + SQ_INTERRUPT_WORD_WAVE_CTXID0, PRIV); + if (sq_int_priv /*&& (kfd_set_dbg_ev_from_interrupt(dev, pasid, + KFD_CTXID0_DOORBELL_ID(context_id0), + KFD_CTXID0_TRAP_CODE(context_id0), + NULL, 0))*/) + return; + break; + case SQ_INTERRUPT_WORD_ENCODING_ERROR: + print_sq_intr_info_error(context_id0, context_id1); + sq_int_errtype = REG_GET_FIELD(context_id0, + SQ_INTERRUPT_WORD_ERROR_CTXID0, TYPE); + if (sq_int_errtype != SQ_INTERRUPT_ERROR_TYPE_ILLEGAL_INST && + sq_int_errtype != SQ_INTERRUPT_ERROR_TYPE_MEMVIOL) { + event_interrupt_poison_consumption_v11( + dev, pasid, source_id); + return; + } + break; + default: + break; + } + kfd_signal_event_interrupt(pasid, context_id0 & 0xffffff, 24); + } + + /*} else if (KFD_IRQ_IS_FENCE(client_id, source_id)) { + kfd_process_close_interrupt_drain(pasid);*/ + } +} + +const struct kfd_event_interrupt_class event_interrupt_class_v11 = { + .interrupt_isr = event_interrupt_isr_v11, + .interrupt_wq = event_interrupt_wq_v11, +}; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c index 7eedbcd14828..0b75a37b689b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c @@ -90,29 +90,35 @@ enum SQ_INTERRUPT_ERROR_TYPE { #define KFD_SQ_INT_DATA__ERR_TYPE_MASK 0xF00000 #define KFD_SQ_INT_DATA__ERR_TYPE__SHIFT 20 -static void event_interrupt_poison_consumption(struct kfd_dev *dev, - uint16_t pasid, uint16_t source_id) +static void event_interrupt_poison_consumption_v9(struct kfd_dev *dev, + uint16_t pasid, uint16_t client_id) { - int ret = -EINVAL; + int old_poison, ret = -EINVAL; struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); if (!p) return; /* all queues of a process will be unmapped in one time */ - if (atomic_read(&p->poison)) { - kfd_unref_process(p); - return; - } - - atomic_set(&p->poison, 1); + old_poison = atomic_cmpxchg(&p->poison, 0, 1); kfd_unref_process(p); + if (old_poison) + return; - switch (source_id) { - case SOC15_INTSRC_SQ_INTERRUPT_MSG: + switch (client_id) { + case SOC15_IH_CLIENTID_SE0SH: + case SOC15_IH_CLIENTID_SE1SH: + case SOC15_IH_CLIENTID_SE2SH: + case SOC15_IH_CLIENTID_SE3SH: + case SOC15_IH_CLIENTID_UTCL2: ret = kfd_dqm_evict_pasid(dev->dqm, pasid); break; - case SOC15_INTSRC_SDMA_ECC: + case SOC15_IH_CLIENTID_SDMA0: + case SOC15_IH_CLIENTID_SDMA1: + case SOC15_IH_CLIENTID_SDMA2: + case SOC15_IH_CLIENTID_SDMA3: + case SOC15_IH_CLIENTID_SDMA4: + break; default: break; } @@ -122,10 +128,36 @@ static void event_interrupt_poison_consumption(struct kfd_dev *dev, /* resetting queue passes, do page retirement without gpu reset * resetting queue fails, fallback to gpu reset solution */ - if (!ret) + if (!ret) { + dev_warn(dev->adev->dev, + "RAS poison consumption, unmap queue flow succeeded: client id %d\n", + client_id); amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, false); - else + } else { + dev_warn(dev->adev->dev, + "RAS poison consumption, fall back to gpu reset flow: client id %d\n", + client_id); amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, true); + } +} + +static bool context_id_expected(struct kfd_dev *dev) +{ + switch (KFD_GC_VERSION(dev)) { + case IP_VERSION(9, 0, 1): + return dev->mec_fw_version >= 0x817a; + case IP_VERSION(9, 1, 0): + case IP_VERSION(9, 2, 1): + case IP_VERSION(9, 2, 2): + case IP_VERSION(9, 3, 0): + case IP_VERSION(9, 4, 0): + return dev->mec_fw_version >= 0x17a; + default: + /* Other GFXv9 and later GPUs always sent valid context IDs + * on legitimate events + */ + return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 1); + } } static bool event_interrupt_isr_v9(struct kfd_dev *dev, @@ -193,6 +225,20 @@ static bool event_interrupt_isr_v9(struct kfd_dev *dev, if (WARN_ONCE(pasid == 0, "Bug: No PASID in KFD interrupt")) return false; + /* Workaround CP firmware sending bogus signals with 0 context_id. + * Those can be safely ignored on hardware and firmware versions that + * include a valid context_id on legitimate signals. This avoids the + * slow path in kfd_signal_event_interrupt that scans all event slots + * for signaled events. + */ + if (source_id == SOC15_INTSRC_CP_END_OF_PIPE) { + uint32_t context_id = + SOC15_CONTEXT_ID0_FROM_IH_ENTRY(ih_ring_entry); + + if (context_id == 0 && context_id_expected(dev)) + return false; + } + /* Interrupt types we care about: various signals and faults. * They will be forwarded to a work queue (see below). */ @@ -270,7 +316,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev, sq_intr_err); if (sq_intr_err != SQ_INTERRUPT_ERROR_TYPE_ILLEGAL_INST && sq_intr_err != SQ_INTERRUPT_ERROR_TYPE_MEMVIOL) { - event_interrupt_poison_consumption(dev, pasid, source_id); + event_interrupt_poison_consumption_v9(dev, pasid, client_id); return; } break; @@ -291,7 +337,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev, if (source_id == SOC15_INTSRC_SDMA_TRAP) { kfd_signal_event_interrupt(pasid, context_id0 & 0xfffffff, 28); } else if (source_id == SOC15_INTSRC_SDMA_ECC) { - event_interrupt_poison_consumption(dev, pasid, source_id); + event_interrupt_poison_consumption_v9(dev, pasid, client_id); return; } } else if (client_id == SOC15_IH_CLIENTID_VMC || @@ -300,6 +346,12 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev, struct kfd_vm_fault_info info = {0}; uint16_t ring_id = SOC15_RING_ID_FROM_IH_ENTRY(ih_ring_entry); + if (client_id == SOC15_IH_CLIENTID_UTCL2 && + amdgpu_amdkfd_ras_query_utcl2_poison_status(dev->adev)) { + event_interrupt_poison_consumption_v9(dev, pasid, client_id); + return; + } + info.vmid = vmid; info.mc_id = client_id; info.page_addr = ih_ring_entry[4] | diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c index 9178cfe34f20..a9466d154395 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c @@ -146,15 +146,24 @@ static void interrupt_wq(struct work_struct *work) struct kfd_dev *dev = container_of(work, struct kfd_dev, interrupt_work); uint32_t ih_ring_entry[KFD_MAX_RING_ENTRY_SIZE]; + long start_jiffies = jiffies; if (dev->device_info.ih_ring_entry_size > sizeof(ih_ring_entry)) { dev_err_once(dev->adev->dev, "Ring entry too small\n"); return; } - while (dequeue_ih_ring_entry(dev, ih_ring_entry)) + while (dequeue_ih_ring_entry(dev, ih_ring_entry)) { dev->device_info.event_interrupt_class->interrupt_wq(dev, ih_ring_entry); + if (jiffies - start_jiffies > HZ) { + /* If we spent more than a second processing signals, + * reschedule the worker to avoid soft-lockup warnings + */ + queue_work(dev->ih_wq, &dev->interrupt_work); + break; + } + } } bool interrupt_is_wanted(struct kfd_dev *dev, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 7e3a7fcb9fe6..997650d597ec 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -410,7 +410,6 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, struct migrate_vma migrate; unsigned long cpages = 0; dma_addr_t *scratch; - size_t size; void *buf; int r = -ENOMEM; @@ -421,9 +420,9 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, migrate.flags = MIGRATE_VMA_SELECT_SYSTEM; migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); - size = 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t); - size *= npages; - buf = kvmalloc(size, GFP_KERNEL | __GFP_ZERO); + buf = kvcalloc(npages, + 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), + GFP_KERNEL); if (!buf) goto out; @@ -665,7 +664,6 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, struct dma_fence *mfence = NULL; struct migrate_vma migrate; dma_addr_t *scratch; - size_t size; void *buf; int r = -ENOMEM; @@ -676,9 +674,10 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE; migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); - size = 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t); - size *= npages; - buf = kvmalloc(size, GFP_KERNEL | __GFP_ZERO); + buf = kvcalloc(npages, + 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), + GFP_KERNEL); + if (!buf) goto out; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c index 5ac209209613..49a283be6b57 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c @@ -100,7 +100,7 @@ void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm, { struct kfd_cu_info cu_info; uint32_t cu_per_sh[KFD_MAX_NUM_SE][KFD_MAX_NUM_SH_PER_SE] = {0}; - int i, se, sh, cu; + int i, se, sh, cu, cu_bitmap_sh_mul; amdgpu_amdkfd_get_cu_info(mm->dev->adev, &cu_info); @@ -120,6 +120,10 @@ void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm, cu_info.num_shader_arrays_per_engine * cu_info.num_shader_engines); return; } + + cu_bitmap_sh_mul = (KFD_GC_VERSION(mm->dev) >= IP_VERSION(11, 0, 0) && + KFD_GC_VERSION(mm->dev) < IP_VERSION(12, 0, 0)) ? 2 : 1; + /* Count active CUs per SH. * * Some CUs in an SH may be disabled. HW expects disabled CUs to be @@ -129,10 +133,12 @@ void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm, * Each half of se_mask must be filled only on bits 0-cu_per_sh[se][sh]-1. * * See note on Arcturus cu_bitmap layout in gfx_v9_0_get_cu_info. + * See note on GFX11 cu_bitmap layout in gfx_v11_0_get_cu_info. */ for (se = 0; se < cu_info.num_shader_engines; se++) for (sh = 0; sh < cu_info.num_shader_arrays_per_engine; sh++) - cu_per_sh[se][sh] = hweight32(cu_info.cu_bitmap[se % 4][sh + (se / 4)]); + cu_per_sh[se][sh] = hweight32( + cu_info.cu_bitmap[se % 4][sh + (se / 4) * cu_bitmap_sh_mul]); /* Symmetrically map cu_mask to all SEs & SHs: * se_mask programs up to 2 SH in the upper and lower 16 bits. diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c new file mode 100644 index 000000000000..4e0387f591be --- /dev/null +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c @@ -0,0 +1,508 @@ +/* + * Copyright 2021 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR + * OTHER DEALINGS IN THE SOFTWARE. + * + */ + +#include <linux/printk.h> +#include <linux/slab.h> +#include <linux/uaccess.h> +#include "kfd_priv.h" +#include "kfd_mqd_manager.h" +#include "v11_structs.h" +#include "gc/gc_11_0_0_offset.h" +#include "gc/gc_11_0_0_sh_mask.h" +#include "amdgpu_amdkfd.h" + +static inline struct v11_compute_mqd *get_mqd(void *mqd) +{ + return (struct v11_compute_mqd *)mqd; +} + +static inline struct v11_sdma_mqd *get_sdma_mqd(void *mqd) +{ + return (struct v11_sdma_mqd *)mqd; +} + +static void update_cu_mask(struct mqd_manager *mm, void *mqd, + struct mqd_update_info *minfo) +{ + struct v11_compute_mqd *m; + uint32_t se_mask[KFD_MAX_NUM_SE] = {0}; + + if (!minfo || (minfo->update_flag != UPDATE_FLAG_CU_MASK) || + !minfo->cu_mask.ptr) + return; + + mqd_symmetrically_map_cu_mask(mm, + minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask); + + m = get_mqd(mqd); + m->compute_static_thread_mgmt_se0 = se_mask[0]; + m->compute_static_thread_mgmt_se1 = se_mask[1]; + m->compute_static_thread_mgmt_se2 = se_mask[2]; + m->compute_static_thread_mgmt_se3 = se_mask[3]; + m->compute_static_thread_mgmt_se4 = se_mask[4]; + m->compute_static_thread_mgmt_se5 = se_mask[5]; + m->compute_static_thread_mgmt_se6 = se_mask[6]; + m->compute_static_thread_mgmt_se7 = se_mask[7]; + + pr_debug("update cu mask to %#x %#x %#x %#x %#x %#x %#x %#x\n", + m->compute_static_thread_mgmt_se0, + m->compute_static_thread_mgmt_se1, + m->compute_static_thread_mgmt_se2, + m->compute_static_thread_mgmt_se3, + m->compute_static_thread_mgmt_se4, + m->compute_static_thread_mgmt_se5, + m->compute_static_thread_mgmt_se6, + m->compute_static_thread_mgmt_se7); +} + +static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q) +{ + m->cp_hqd_pipe_priority = pipe_priority_map[q->priority]; + m->cp_hqd_queue_priority = q->priority; +} + +static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, + struct queue_properties *q) +{ + struct kfd_mem_obj *mqd_mem_obj; + int size; + + /* + * MES write to areas beyond MQD size. So allocate + * 1 PAGE_SIZE memory for MQD is MES is enabled. + */ + if (kfd->shared_resources.enable_mes) + size = PAGE_SIZE; + else + size = sizeof(struct v11_compute_mqd); + + if (kfd_gtt_sa_allocate(kfd, size, &mqd_mem_obj)) + return NULL; + + return mqd_mem_obj; +} + +static void init_mqd(struct mqd_manager *mm, void **mqd, + struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, + struct queue_properties *q) +{ + uint64_t addr; + struct v11_compute_mqd *m; + int size; + + m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr; + addr = mqd_mem_obj->gpu_addr; + + if (mm->dev->shared_resources.enable_mes) + size = PAGE_SIZE; + else + size = sizeof(struct v11_compute_mqd); + + memset(m, 0, size); + + m->header = 0xC0310800; + m->compute_pipelinestat_enable = 1; + m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF; + m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF; + m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF; + m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF; + + m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK | + 0x55 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT; + + m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT; + + m->cp_mqd_base_addr_lo = lower_32_bits(addr); + m->cp_mqd_base_addr_hi = upper_32_bits(addr); + + m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT | + 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | + 1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; + + if (q->format == KFD_QUEUE_FORMAT_AQL) { + m->cp_hqd_aql_control = + 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; + } + + if (mm->dev->cwsr_enabled) { + m->cp_hqd_persistent_state |= + (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT); + m->cp_hqd_ctx_save_base_addr_lo = + lower_32_bits(q->ctx_save_restore_area_address); + m->cp_hqd_ctx_save_base_addr_hi = + upper_32_bits(q->ctx_save_restore_area_address); + m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size; + m->cp_hqd_cntl_stack_size = q->ctl_stack_size; + m->cp_hqd_cntl_stack_offset = q->ctl_stack_size; + m->cp_hqd_wg_state_offset = q->ctl_stack_size; + } + + *mqd = m; + if (gart_addr) + *gart_addr = addr; + mm->update_mqd(mm, m, q, NULL); +} + +static int load_mqd(struct mqd_manager *mm, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + struct queue_properties *p, struct mm_struct *mms) +{ + int r = 0; + /* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */ + uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0); + + r = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id, + (uint32_t __user *)p->write_ptr, + wptr_shift, 0, mms); + return r; +} + +static int hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + struct queue_properties *p, struct mm_struct *mms) +{ + return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, mqd, pipe_id, + queue_id, p->doorbell_off); +} + +static void update_mqd(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, + struct mqd_update_info *minfo) +{ + struct v11_compute_mqd *m; + + m = get_mqd(mqd); + + m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT; + m->cp_hqd_pq_control |= + ffs(q->queue_size / sizeof(unsigned int)) - 1 - 1; + pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control); + + m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8); + m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8); + + m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr); + m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr); + m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr); + m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr); + + m->cp_hqd_pq_doorbell_control = + q->doorbell_off << + CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT; + pr_debug("cp_hqd_pq_doorbell_control 0x%x\n", + m->cp_hqd_pq_doorbell_control); + + m->cp_hqd_ib_control = 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT; + + /* + * HW does not clamp this field correctly. Maximum EOP queue size + * is constrained by per-SE EOP done signal count, which is 8-bit. + * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit + * more than (EOP entry count - 1) so a queue size of 0x800 dwords + * is safe, giving a maximum field value of 0xA. + */ + m->cp_hqd_eop_control = min(0xA, + ffs(q->eop_ring_buffer_size / sizeof(unsigned int)) - 1 - 1); + m->cp_hqd_eop_base_addr_lo = + lower_32_bits(q->eop_ring_buffer_address >> 8); + m->cp_hqd_eop_base_addr_hi = + upper_32_bits(q->eop_ring_buffer_address >> 8); + + m->cp_hqd_iq_timer = 0; + + m->cp_hqd_vmid = q->vmid; + + if (q->format == KFD_QUEUE_FORMAT_AQL) { + /* GC 10 removed WPP_CLAMP from PQ Control */ + m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK | + 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT | + 1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT ; + m->cp_hqd_pq_doorbell_control |= + 1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT; + } + if (mm->dev->cwsr_enabled) + m->cp_hqd_ctx_save_control = 0; + + update_cu_mask(mm, mqd, minfo); + set_priority(m, q); + + q->is_active = QUEUE_IS_ACTIVE(*q); +} + +static uint32_t read_doorbell_id(void *mqd) +{ + struct v11_compute_mqd *m = (struct v11_compute_mqd *)mqd; + + return m->queue_doorbell_id0; +} + +static int destroy_mqd(struct mqd_manager *mm, void *mqd, + enum kfd_preempt_type type, + unsigned int timeout, uint32_t pipe_id, + uint32_t queue_id) +{ + return mm->dev->kfd2kgd->hqd_destroy + (mm->dev->adev, mqd, type, timeout, + pipe_id, queue_id); +} + +static void free_mqd(struct mqd_manager *mm, void *mqd, + struct kfd_mem_obj *mqd_mem_obj) +{ + kfd_gtt_sa_free(mm->dev, mqd_mem_obj); +} + +static bool is_occupied(struct mqd_manager *mm, void *mqd, + uint64_t queue_address, uint32_t pipe_id, + uint32_t queue_id) +{ + return mm->dev->kfd2kgd->hqd_is_occupied( + mm->dev->adev, queue_address, + pipe_id, queue_id); +} + +static int get_wave_state(struct mqd_manager *mm, void *mqd, + void __user *ctl_stack, + u32 *ctl_stack_used_size, + u32 *save_area_used_size) +{ + struct v11_compute_mqd *m; + /*struct mqd_user_context_save_area_header header;*/ + + m = get_mqd(mqd); + + /* Control stack is written backwards, while workgroup context data + * is written forwards. Both starts from m->cp_hqd_cntl_stack_size. + * Current position is at m->cp_hqd_cntl_stack_offset and + * m->cp_hqd_wg_state_offset, respectively. + */ + *ctl_stack_used_size = m->cp_hqd_cntl_stack_size - + m->cp_hqd_cntl_stack_offset; + *save_area_used_size = m->cp_hqd_wg_state_offset - + m->cp_hqd_cntl_stack_size; + + /* Control stack is not copied to user mode for GFXv11 because + * it's part of the context save area that is already + * accessible to user mode + */ +/* + header.control_stack_size = *ctl_stack_used_size; + header.wave_state_size = *save_area_used_size; + + header.wave_state_offset = m->cp_hqd_wg_state_offset; + header.control_stack_offset = m->cp_hqd_cntl_stack_offset; + + if (copy_to_user(ctl_stack, &header, sizeof(header))) + return -EFAULT; +*/ + return 0; +} + +static void init_mqd_hiq(struct mqd_manager *mm, void **mqd, + struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, + struct queue_properties *q) +{ + struct v11_compute_mqd *m; + + init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q); + + m = get_mqd(*mqd); + + m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT | + 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT; +} + +static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, + struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, + struct queue_properties *q) +{ + struct v11_sdma_mqd *m; + + m = (struct v11_sdma_mqd *) mqd_mem_obj->cpu_ptr; + + memset(m, 0, sizeof(struct v11_sdma_mqd)); + + *mqd = m; + if (gart_addr) + *gart_addr = mqd_mem_obj->gpu_addr; + + mm->update_mqd(mm, m, q, NULL); +} + +static int load_mqd_sdma(struct mqd_manager *mm, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + struct queue_properties *p, struct mm_struct *mms) +{ + return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->adev, mqd, + (uint32_t __user *)p->write_ptr, + mms); +} + +#define SDMA_RLC_DUMMY_DEFAULT 0xf + +static void update_mqd_sdma(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, + struct mqd_update_info *minfo) +{ + struct v11_sdma_mqd *m; + + m = get_sdma_mqd(mqd); + m->sdmax_rlcx_rb_cntl = (ffs(q->queue_size / sizeof(unsigned int)) - 1) + << SDMA0_QUEUE0_RB_CNTL__RB_SIZE__SHIFT | + q->vmid << SDMA0_QUEUE0_RB_CNTL__RB_VMID__SHIFT | + 1 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | + 6 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT; + + m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8); + m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8); + m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr); + m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr); + m->sdmax_rlcx_doorbell_offset = + q->doorbell_off << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT; + + m->sdma_engine_id = q->sdma_engine_id; + m->sdma_queue_id = q->sdma_queue_id; + m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT; + + q->is_active = QUEUE_IS_ACTIVE(*q); +} + +/* + * * preempt type here is ignored because there is only one way + * * to preempt sdma queue + */ +static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd, + enum kfd_preempt_type type, + unsigned int timeout, uint32_t pipe_id, + uint32_t queue_id) +{ + return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->adev, mqd, timeout); +} + +static bool is_occupied_sdma(struct mqd_manager *mm, void *mqd, + uint64_t queue_address, uint32_t pipe_id, + uint32_t queue_id) +{ + return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->adev, mqd); +} + +#if defined(CONFIG_DEBUG_FS) + +static int debugfs_show_mqd(struct seq_file *m, void *data) +{ + seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, + data, sizeof(struct v11_compute_mqd), false); + return 0; +} + +static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) +{ + seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4, + data, sizeof(struct v11_sdma_mqd), false); + return 0; +} + +#endif + +struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type, + struct kfd_dev *dev) +{ + struct mqd_manager *mqd; + + if (WARN_ON(type >= KFD_MQD_TYPE_MAX)) + return NULL; + + mqd = kzalloc(sizeof(*mqd), GFP_KERNEL); + if (!mqd) + return NULL; + + mqd->dev = dev; + + switch (type) { + case KFD_MQD_TYPE_CP: + pr_debug("%s@%i\n", __func__, __LINE__); + mqd->allocate_mqd = allocate_mqd; + mqd->init_mqd = init_mqd; + mqd->free_mqd = free_mqd; + mqd->load_mqd = load_mqd; + mqd->update_mqd = update_mqd; + mqd->destroy_mqd = destroy_mqd; + mqd->is_occupied = is_occupied; + mqd->mqd_size = sizeof(struct v11_compute_mqd); + mqd->get_wave_state = get_wave_state; +#if defined(CONFIG_DEBUG_FS) + mqd->debugfs_show_mqd = debugfs_show_mqd; +#endif + pr_debug("%s@%i\n", __func__, __LINE__); + break; + case KFD_MQD_TYPE_HIQ: + pr_debug("%s@%i\n", __func__, __LINE__); + mqd->allocate_mqd = allocate_hiq_mqd; + mqd->init_mqd = init_mqd_hiq; + mqd->free_mqd = free_mqd_hiq_sdma; + mqd->load_mqd = hiq_load_mqd_kiq; + mqd->update_mqd = update_mqd; + mqd->destroy_mqd = destroy_mqd; + mqd->is_occupied = is_occupied; + mqd->mqd_size = sizeof(struct v11_compute_mqd); +#if defined(CONFIG_DEBUG_FS) + mqd->debugfs_show_mqd = debugfs_show_mqd; +#endif + mqd->read_doorbell_id = read_doorbell_id; + pr_debug("%s@%i\n", __func__, __LINE__); + break; + case KFD_MQD_TYPE_DIQ: + mqd->allocate_mqd = allocate_mqd; + mqd->init_mqd = init_mqd_hiq; + mqd->free_mqd = free_mqd; + mqd->load_mqd = load_mqd; + mqd->update_mqd = update_mqd; + mqd->destroy_mqd = destroy_mqd; + mqd->is_occupied = is_occupied; + mqd->mqd_size = sizeof(struct v11_compute_mqd); +#if defined(CONFIG_DEBUG_FS) + mqd->debugfs_show_mqd = debugfs_show_mqd; +#endif + break; + case KFD_MQD_TYPE_SDMA: + pr_debug("%s@%i\n", __func__, __LINE__); + mqd->allocate_mqd = allocate_sdma_mqd; + mqd->init_mqd = init_mqd_sdma; + mqd->free_mqd = free_mqd_hiq_sdma; + mqd->load_mqd = load_mqd_sdma; + mqd->update_mqd = update_mqd_sdma; + mqd->destroy_mqd = destroy_mqd_sdma; + mqd->is_occupied = is_occupied_sdma; + mqd->mqd_size = sizeof(struct v11_sdma_mqd); +#if defined(CONFIG_DEBUG_FS) + mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; +#endif + pr_debug("%s@%i\n", __func__, __LINE__); + break; + default: + kfree(mqd); + return NULL; + } + + return mqd; +} diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 8f58fc491b28..2585d6e61d42 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -229,6 +229,8 @@ struct kfd_device_info { bool needs_pci_atomics; uint32_t no_atomic_fw_version; unsigned int num_sdma_queues_per_engine; + unsigned int num_reserved_sdma_queues_per_engine; + uint64_t reserved_sdma_queues_bitmap; }; unsigned int kfd_get_num_sdma_engines(struct kfd_dev *kdev); @@ -272,6 +274,7 @@ struct kfd_dev { struct kgd2kfd_shared_resources shared_resources; struct kfd_vmid_info vm_info; + struct kfd_local_mem_info local_mem_info; const struct kfd2kgd_calls *kfd2kgd; struct mutex doorbell_mutex; @@ -564,6 +567,10 @@ struct queue { /* procfs */ struct kobject kobj; + + void *gang_ctx_bo; + uint64_t gang_ctx_gpu_addr; + void *gang_ctx_cpu_ptr; }; enum KFD_MQD_TYPE { @@ -706,6 +713,7 @@ struct kfd_process_device { /* VM context for GPUVM allocations */ struct file *drm_file; void *drm_priv; + atomic64_t tlb_seq; /* GPUVM allocations storage */ struct idr alloc_idr; @@ -778,6 +786,10 @@ struct kfd_process_device { * checkpointed node to refer to this device. */ uint32_t user_gpu_id; + + void *proc_ctx_bo; + uint64_t proc_ctx_gpu_addr; + void *proc_ctx_cpu_ptr; }; #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) @@ -1016,6 +1028,8 @@ 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_topology_device *kfd_topology_device_by_proximity_domain_no_lock( + uint32_t proximity_domain); struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); @@ -1167,6 +1181,8 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, struct kfd_dev *dev); struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, struct kfd_dev *dev); +struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type, + struct kfd_dev *dev); struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); void device_queue_manager_uninit(struct device_queue_manager *dqm); struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, @@ -1289,10 +1305,11 @@ uint64_t kfd_get_number_elems(struct kfd_dev *kfd); /* Events */ extern const struct kfd_event_interrupt_class event_interrupt_class_cik; extern const struct kfd_event_interrupt_class event_interrupt_class_v9; +extern const struct kfd_event_interrupt_class event_interrupt_class_v11; extern const struct kfd_device_global_init_class device_global_init_class_cik; -void kfd_event_init_process(struct kfd_process *p); +int kfd_event_init_process(struct kfd_process *p); void kfd_event_free_process(struct kfd_process *p); int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); int kfd_wait_on_events(struct kfd_process *p, @@ -1326,6 +1343,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid); void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev) +{ + return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2) || + (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && + dev->adev->sdma.instance[0].fw_version >= 18) || + KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0); +} + bool kfd_is_locked(void); /* Compute profile */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 59c04b2d383b..e3d64ec8c353 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -722,7 +722,7 @@ static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd, goto err_alloc_mem; err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->adev, *mem, - pdd->drm_priv, NULL); + pdd->drm_priv); if (err) goto err_map_mem; @@ -1041,6 +1041,9 @@ static void kfd_process_destroy_pdds(struct kfd_process *p) kfd_free_process_doorbells(pdd->dev, pdd->doorbell_index); + if (pdd->dev->shared_resources.enable_mes) + amdgpu_amdkfd_free_gtt_mem(pdd->dev->adev, + pdd->proc_ctx_bo); /* * before destroying pdd, make sure to report availability * for auto suspend @@ -1370,12 +1373,16 @@ static struct kfd_process *create_process(const struct task_struct *thread) INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker); INIT_DELAYED_WORK(&process->restore_work, restore_process_worker); process->last_restore_timestamp = get_jiffies_64(); - kfd_event_init_process(process); + err = kfd_event_init_process(process); + if (err) + goto err_event_init; process->is_32bit_user_mode = in_compat_syscall(); process->pasid = kfd_pasid_alloc(); - if (process->pasid == 0) + if (process->pasid == 0) { + err = -ENOSPC; goto err_alloc_pasid; + } err = pqm_init(&process->pqm, process); if (err != 0) @@ -1424,6 +1431,8 @@ err_init_apertures: err_process_pqm_init: kfd_pasid_free(process->pasid); err_alloc_pasid: + kfd_event_free_process(process); +err_event_init: mutex_destroy(&process->mutex); kfree(process); err_alloc_process: @@ -1478,6 +1487,7 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, struct kfd_process *p) { struct kfd_process_device *pdd = NULL; + int retval = 0; if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE)) return NULL; @@ -1510,6 +1520,21 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, pdd->sdma_past_activity_counter = 0; pdd->user_gpu_id = dev->id; atomic64_set(&pdd->evict_duration_counter, 0); + + if (dev->shared_resources.enable_mes) { + retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, + AMDGPU_MES_PROC_CTX_SIZE, + &pdd->proc_ctx_bo, + &pdd->proc_ctx_gpu_addr, + &pdd->proc_ctx_cpu_ptr, + false); + if (retval) { + pr_err("failed to allocate process context bo\n"); + goto err_free_pdd; + } + memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE); + } + p->pdds[p->n_pdds++] = pdd; /* Init idr used for memory handle translation */ @@ -1560,6 +1585,7 @@ int kfd_process_device_init_vm(struct kfd_process_device *pdd, return ret; } pdd->drm_priv = drm_file->private_data; + atomic64_set(&pdd->tlb_seq, 0); ret = kfd_process_device_reserve_ib_mem(pdd); if (ret) @@ -1949,8 +1975,18 @@ int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type) { + struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); + uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm); struct kfd_dev *dev = pdd->dev; + /* + * It can be that we race and lose here, but that is extremely unlikely + * and the worst thing which could happen is that we flush the changes + * into the TLB once more which is harmless. + */ + if (atomic64_xchg(&pdd->tlb_seq, tlb_seq) == tlb_seq) + return; + if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { /* Nothing to flush until a VMID is assigned, which * only happens when the first queue is created. diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index 4f58e671d39b..dc00484ff484 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -198,8 +198,26 @@ static int init_user_queue(struct process_queue_manager *pqm, (*q)->device = dev; (*q)->process = pqm->process; + if (dev->shared_resources.enable_mes) { + retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, + AMDGPU_MES_GANG_CTX_SIZE, + &(*q)->gang_ctx_bo, + &(*q)->gang_ctx_gpu_addr, + &(*q)->gang_ctx_cpu_ptr, + false); + if (retval) { + pr_err("failed to allocate gang context bo\n"); + goto cleanup; + } + memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE); + } + pr_debug("PQM After init queue"); + return 0; +cleanup: + if (dev->shared_resources.enable_mes) + uninit_queue(*q); return retval; } @@ -418,6 +436,9 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) pdd->qpd.num_gws = 0; } + if (dev->shared_resources.enable_mes) + amdgpu_amdkfd_free_gtt_mem(dev->adev, + pqn->q->gang_ctx_bo); uninit_queue(pqn->q); } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 3b8856b4cece..29e9ebf6d8d5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -149,8 +149,7 @@ svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange, int i, r; if (!addr) { - addr = kvmalloc_array(prange->npages, sizeof(*addr), - GFP_KERNEL | __GFP_ZERO); + addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL); if (!addr) return -ENOMEM; prange->dma_addr[gpuidx] = addr; @@ -548,7 +547,7 @@ svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, goto reserve_bo_failed; } - r = dma_resv_reserve_shared(bo->tbo.base.resv, 1); + r = dma_resv_reserve_fences(bo->tbo.base.resv, 1); if (r) { pr_debug("failed %d to reserve bo\n", r); amdgpu_bo_unreserve(bo); @@ -686,7 +685,8 @@ svm_range_check_attr(struct kfd_process *p, static void svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, - uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) + uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs, + bool *update_mapping) { uint32_t i; int gpuidx; @@ -702,6 +702,7 @@ svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, case KFD_IOCTL_SVM_ATTR_ACCESS: case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: case KFD_IOCTL_SVM_ATTR_NO_ACCESS: + *update_mapping = true; gpuidx = kfd_process_gpuidx_from_gpuid(p, attrs[i].value); if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { @@ -716,9 +717,11 @@ svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, } break; case KFD_IOCTL_SVM_ATTR_SET_FLAGS: + *update_mapping = true; prange->flags |= attrs[i].value; break; case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: + *update_mapping = true; prange->flags &= ~attrs[i].value; break; case KFD_IOCTL_SVM_ATTR_GRANULARITY: @@ -951,6 +954,7 @@ svm_range_split_adjust(struct svm_range *new, struct svm_range *old, new->prefetch_loc = old->prefetch_loc; new->actual_loc = old->actual_loc; new->granularity = old->granularity; + new->mapped_to_gpu = old->mapped_to_gpu; bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); @@ -1188,9 +1192,9 @@ svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm, pr_debug("[0x%llx 0x%llx]\n", start, last); - return amdgpu_vm_bo_update_mapping(adev, adev, vm, false, true, NULL, - start, last, init_pte_value, 0, - NULL, NULL, fence, NULL); + return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start, + last, init_pte_value, 0, 0, NULL, NULL, + fence); } static int @@ -1204,6 +1208,17 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, uint32_t gpuidx; int r = 0; + if (!prange->mapped_to_gpu) { + pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n", + prange, prange->start, prange->last); + return 0; + } + + if (prange->start == start && prange->last == last) { + pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange); + prange->mapped_to_gpu = false; + } + bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip, MAX_GPU_INSTANCE); p = container_of(prange->svms, struct kfd_process, svms); @@ -1239,11 +1254,10 @@ static int svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, unsigned long offset, unsigned long npages, bool readonly, dma_addr_t *dma_addr, struct amdgpu_device *bo_adev, - struct dma_fence **fence) + struct dma_fence **fence, bool flush_tlb) { struct amdgpu_device *adev = pdd->dev->adev; struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); - bool table_freed = false; uint64_t pte_flags; unsigned long last_start; int last_domain; @@ -1278,13 +1292,12 @@ svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0, pte_flags); - r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false, - NULL, last_start, - prange->start + i, pte_flags, - last_start - prange->start, - NULL, dma_addr, - &vm->last_update, - &table_freed); + r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL, + last_start, prange->start + i, + pte_flags, + last_start - prange->start, + bo_adev ? bo_adev->vm_manager.vram_base_offset : 0, + NULL, dma_addr, &vm->last_update); for (j = last_start - prange->start; j <= i; j++) dma_addr[j] |= last_domain; @@ -1306,8 +1319,6 @@ svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, if (fence) *fence = dma_fence_get(vm->last_update); - if (table_freed) - kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY); out: return r; } @@ -1315,7 +1326,7 @@ out: static int svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset, unsigned long npages, bool readonly, - unsigned long *bitmap, bool wait) + unsigned long *bitmap, bool wait, bool flush_tlb) { struct kfd_process_device *pdd; struct amdgpu_device *bo_adev; @@ -1350,7 +1361,8 @@ svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset, r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly, prange->dma_addr[gpuidx], - bo_adev, wait ? &fence : NULL); + bo_adev, wait ? &fence : NULL, + flush_tlb); if (r) break; @@ -1363,6 +1375,8 @@ svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset, break; } } + + kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY); } return r; @@ -1372,7 +1386,7 @@ struct svm_validate_context { struct kfd_process *process; struct svm_range *prange; bool intr; - unsigned long bitmap[MAX_GPU_INSTANCE]; + DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); struct ttm_validate_buffer tv[MAX_GPU_INSTANCE]; struct list_head validate_list; struct ww_acquire_ctx ticket; @@ -1469,8 +1483,8 @@ static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx) * 5. Release page table (and SVM BO) reservation */ static int svm_range_validate_and_map(struct mm_struct *mm, - struct svm_range *prange, - int32_t gpuidx, bool intr, bool wait) + struct svm_range *prange, int32_t gpuidx, + bool intr, bool wait, bool flush_tlb) { struct svm_validate_context ctx; unsigned long start, end, addr; @@ -1509,8 +1523,12 @@ static int svm_range_validate_and_map(struct mm_struct *mm, prange->bitmap_aip, MAX_GPU_INSTANCE); } - if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE)) - return 0; + if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE)) { + if (!prange->mapped_to_gpu) + return 0; + + bitmap_copy(ctx.bitmap, prange->bitmap_access, MAX_GPU_INSTANCE); + } if (prange->actual_loc && !prange->ttm_res) { /* This should never happen. actual_loc gets set by @@ -1582,7 +1600,7 @@ static int svm_range_validate_and_map(struct mm_struct *mm, } r = svm_range_map_to_gpus(prange, offset, npages, readonly, - ctx.bitmap, wait); + ctx.bitmap, wait, flush_tlb); unlock_out: svm_range_unlock(prange); @@ -1590,8 +1608,10 @@ unlock_out: addr = next; } - if (addr == end) + if (addr == end) { prange->validated_once = true; + prange->mapped_to_gpu = true; + } unreserve_out: svm_range_unreserve_bos(&ctx); @@ -1676,7 +1696,7 @@ static void svm_range_restore_work(struct work_struct *work) mutex_lock(&prange->migrate_mutex); r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE, - false, true); + false, true, false); if (r) pr_debug("failed %d to map 0x%lx to gpus\n", r, prange->start); @@ -1822,6 +1842,7 @@ static struct svm_range *svm_range_clone(struct svm_range *old) new->prefetch_loc = old->prefetch_loc; new->actual_loc = old->actual_loc; new->granularity = old->granularity; + new->mapped_to_gpu = old->mapped_to_gpu; bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); @@ -2687,11 +2708,6 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, pr_debug("kfd process not founded pasid 0x%x\n", pasid); return 0; } - if (!p->xnack_enabled) { - pr_debug("XNACK not enabled for pasid 0x%x\n", pasid); - r = -EFAULT; - goto out; - } svms = &p->svms; pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr); @@ -2702,6 +2718,12 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, goto out; } + if (!p->xnack_enabled) { + pr_debug("XNACK not enabled for pasid 0x%x\n", pasid); + r = -EFAULT; + goto out; + } + /* p->lead_thread is available as kfd_process_wq_release flush the work * before releasing task ref. */ @@ -2812,7 +2834,7 @@ retry_write_locked: } } - r = svm_range_validate_and_map(mm, prange, gpuidx, false, false); + r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false); if (r) pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n", r, svms, prange->start, prange->last); @@ -3225,6 +3247,8 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, struct svm_range_list *svms; struct svm_range *prange; struct svm_range *next; + bool update_mapping = false; + bool flush_tlb; int r = 0; pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n", @@ -3263,7 +3287,7 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, svm_range_add_notifier_locked(mm, prange); } list_for_each_entry(prange, &update_list, update_list) { - svm_range_apply_attrs(p, prange, nattr, attrs); + svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping); /* TODO: unmap ranges from GPU that lost access */ } list_for_each_entry_safe(prange, next, &remove_list, update_list) { @@ -3296,8 +3320,15 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, continue; } + if (!migrated && !update_mapping) { + mutex_unlock(&prange->migrate_mutex); + continue; + } + + flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu; + r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE, - true, true); + true, true, flush_tlb); if (r) pr_debug("failed %d to map svm range\n", r); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index 66c77f00ac3e..2d54147b4dda 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -133,6 +133,7 @@ struct svm_range { DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE); DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); bool validated_once; + bool mapped_to_gpu; }; static inline void svm_range_lock(struct svm_range *prange) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 3bdcae239bc0..2e20f54bb147 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -46,22 +46,32 @@ static struct list_head topology_device_list; static struct kfd_system_properties sys_props; static DECLARE_RWSEM(topology_lock); -static atomic_t topology_crat_proximity_domain; +static uint32_t topology_crat_proximity_domain; -struct kfd_topology_device *kfd_topology_device_by_proximity_domain( +struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock( 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; } + return device; +} + +struct kfd_topology_device *kfd_topology_device_by_proximity_domain( + uint32_t proximity_domain) +{ + struct kfd_topology_device *device = NULL; + + down_read(&topology_lock); + + device = kfd_topology_device_by_proximity_domain_no_lock( + proximity_domain); up_read(&topology_lock); return device; @@ -1060,7 +1070,7 @@ int kfd_topology_init(void) down_write(&topology_lock); kfd_topology_update_device_list(&temp_topology_device_list, &topology_device_list); - atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1); + topology_crat_proximity_domain = sys_props.num_devices-1; ret = kfd_topology_update_sysfs(); up_write(&topology_lock); @@ -1102,15 +1112,12 @@ static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) uint32_t buf[7]; uint64_t local_mem_size; int i; - struct kfd_local_mem_info local_mem_info; if (!gpu) return 0; - amdgpu_amdkfd_get_local_mem_info(gpu->adev, &local_mem_info); - - local_mem_size = local_mem_info.local_mem_size_private + - local_mem_info.local_mem_size_public; + local_mem_size = gpu->local_mem_info.local_mem_size_private + + gpu->local_mem_info.local_mem_size_public; buf[0] = gpu->pdev->devfn; buf[1] = gpu->pdev->subsystem_vendor | @@ -1295,8 +1302,6 @@ int kfd_topology_add_device(struct kfd_dev *gpu) pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); - proximity_domain = atomic_inc_return(&topology_crat_proximity_domain); - /* Include the CPU in xGMI hive if xGMI connected by assigning it the hive ID. */ if (gpu->hive_id && gpu->adev->gmc.xgmi.connected_to_cpu) { struct kfd_topology_device *top_dev; @@ -1321,12 +1326,16 @@ int kfd_topology_add_device(struct kfd_dev *gpu) */ dev = kfd_assign_gpu(gpu); if (!dev) { + down_write(&topology_lock); + proximity_domain = ++topology_crat_proximity_domain; + 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); + topology_crat_proximity_domain--; return res; } res = kfd_parse_crat_table(crat_image, @@ -1335,10 +1344,10 @@ int kfd_topology_add_device(struct kfd_dev *gpu) if (res) { pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", gpu_id); + topology_crat_proximity_domain--; goto err; } - down_write(&topology_lock); kfd_topology_update_device_list(&temp_topology_device_list, &topology_device_list); @@ -1403,7 +1412,8 @@ int kfd_topology_add_device(struct kfd_dev *gpu) dev->node_props.num_sdma_xgmi_engines = kfd_get_num_xgmi_sdma_engines(gpu); dev->node_props.num_sdma_queues_per_engine = - gpu->device_info.num_sdma_queues_per_engine; + gpu->device_info.num_sdma_queues_per_engine - + gpu->device_info.num_reserved_sdma_queues_per_engine; dev->node_props.num_gws = (dev->gpu->gws && dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ? dev->gpu->adev->gds.gws_size : 0; @@ -1485,25 +1495,78 @@ err: return res; } +/** + * kfd_topology_update_io_links() - Update IO links after device removal. + * @proximity_domain: Proximity domain value of the dev being removed. + * + * The topology list currently is arranged in increasing order of + * proximity domain. + * + * Two things need to be done when a device is removed: + * 1. All the IO links to this device need to be removed. + * 2. All nodes after the current device node need to move + * up once this device node is removed from the topology + * list. As a result, the proximity domain values for + * all nodes after the node being deleted reduce by 1. + * This would also cause the proximity domain values for + * io links to be updated based on new proximity domain + * values. + * + * Context: The caller must hold write topology_lock. + */ +static void kfd_topology_update_io_links(int proximity_domain) +{ + struct kfd_topology_device *dev; + struct kfd_iolink_properties *iolink, *tmp; + + list_for_each_entry(dev, &topology_device_list, list) { + if (dev->proximity_domain > proximity_domain) + dev->proximity_domain--; + + list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) { + /* + * If there is an io link to the dev being deleted + * then remove that IO link also. + */ + if (iolink->node_to == proximity_domain) { + list_del(&iolink->list); + dev->io_link_count--; + dev->node_props.io_links_count--; + } else { + if (iolink->node_from > proximity_domain) + iolink->node_from--; + if (iolink->node_to > proximity_domain) + iolink->node_to--; + } + } + } +} + int kfd_topology_remove_device(struct kfd_dev *gpu) { struct kfd_topology_device *dev, *tmp; uint32_t gpu_id; int res = -ENODEV; + int i = 0; down_write(&topology_lock); - list_for_each_entry_safe(dev, tmp, &topology_device_list, list) + list_for_each_entry_safe(dev, tmp, &topology_device_list, list) { if (dev->gpu == gpu) { gpu_id = dev->gpu_id; kfd_remove_sysfs_node_entry(dev); kfd_release_topology_device(dev); sys_props.num_devices--; + kfd_topology_update_io_links(i); + topology_crat_proximity_domain = sys_props.num_devices-1; + sys_props.generation_count++; res = 0; if (kfd_topology_update_sysfs() < 0) kfd_topology_release_sysfs(); break; } + i++; + } up_write(&topology_lock); diff --git a/drivers/gpu/drm/amd/amdkfd/soc15_int.h b/drivers/gpu/drm/amd/amdkfd/soc15_int.h index daf3c44547d3..e3f3b0b93a59 100644 --- a/drivers/gpu/drm/amd/amdkfd/soc15_int.h +++ b/drivers/gpu/drm/amd/amdkfd/soc15_int.h @@ -31,7 +31,8 @@ #define SOC15_INTSRC_VMC_FAULT 0 #define SOC15_INTSRC_SDMA_TRAP 224 #define SOC15_INTSRC_SDMA_ECC 220 - +#define SOC21_INTSRC_SDMA_TRAP 49 +#define SOC21_INTSRC_SDMA_ECC 62 #define SOC15_CLIENT_ID_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[0]) & 0xff) #define SOC15_SOURCE_ID_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[0]) >> 8 & 0xff) |