diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 396 |
1 files changed, 268 insertions, 128 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index f99d4873bf22..a2879d2b7c8e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -28,6 +28,7 @@ #include "amdgpu.h" #include "amdgpu_gfx.h" #include "amdgpu_dma_buf.h" +#include <drm/ttm/ttm_tt.h> #include <linux/module.h> #include <linux/dma-buf.h> #include "amdgpu_xgmi.h" @@ -53,7 +54,6 @@ int amdgpu_amdkfd_init(void) amdgpu_amdkfd_total_mem_size *= si.mem_unit; ret = kgd2kfd_init(); - amdgpu_amdkfd_gpuvm_init_mem_limits(); kfd_initialized = !ret; return ret; @@ -96,7 +96,7 @@ static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev, size_t *start_offset) { /* - * The first num_doorbells are used by amdgpu. + * The first num_kernel_doorbells are used by amdgpu. * amdkfd takes whatever's left in the aperture. */ if (adev->enable_mes) { @@ -109,11 +109,11 @@ static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev, *aperture_base = adev->doorbell.base; *aperture_size = 0; *start_offset = 0; - } else if (adev->doorbell.size > adev->doorbell.num_doorbells * + } else if (adev->doorbell.size > adev->doorbell.num_kernel_doorbells * sizeof(u32)) { *aperture_base = adev->doorbell.base; *aperture_size = adev->doorbell.size; - *start_offset = adev->doorbell.num_doorbells * sizeof(u32); + *start_offset = adev->doorbell.num_kernel_doorbells * sizeof(u32); } else { *aperture_base = 0; *aperture_size = 0; @@ -133,16 +133,45 @@ static void amdgpu_amdkfd_reset_work(struct work_struct *work) reset_context.method = AMD_RESET_METHOD_NONE; reset_context.reset_req_dev = adev; + reset_context.src = adev->enable_mes ? + AMDGPU_RESET_SRC_MES : + AMDGPU_RESET_SRC_HWS; clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); amdgpu_device_gpu_recover(adev, NULL, &reset_context); } +static const struct drm_client_funcs kfd_client_funcs = { + .unregister = drm_client_release, +}; + +int amdgpu_amdkfd_drm_client_create(struct amdgpu_device *adev) +{ + int ret; + + if (!adev->kfd.init_complete || adev->kfd.client.dev) + return 0; + + ret = drm_client_init(&adev->ddev, &adev->kfd.client, "kfd", + &kfd_client_funcs); + if (ret) { + dev_err(adev->dev, "Failed to init DRM client: %d\n", + ret); + return ret; + } + + drm_client_register(&adev->kfd.client); + + return 0; +} + void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) { int i; int last_valid_bit; + amdgpu_amdkfd_gpuvm_init_mem_limits(); + if (adev->kfd.dev) { struct kgd2kfd_shared_resources gpu_resources = { .compute_vmid_bitmap = @@ -162,8 +191,8 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) * clear */ bitmap_complement(gpu_resources.cp_queue_bitmap, - adev->gfx.mec.queue_bitmap, - KGD_MAX_QUEUES); + adev->gfx.mec_bitmap[0].queue_bitmap, + AMDGPU_MAX_QUEUES); /* According to linux/bitmap.h we shouldn't use bitmap_clear if * nbits is not compile time constant @@ -171,7 +200,7 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) last_valid_bit = 1 /* only first MEC can have compute queues */ * adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_queue_per_pipe; - for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i) + for (i = last_valid_bit; i < AMDGPU_MAX_QUEUES; ++i) clear_bit(i, gpu_resources.cp_queue_bitmap); amdgpu_doorbell_get_kfd_info(adev, @@ -219,38 +248,53 @@ void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev, kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry); } -void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm) +void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool suspend_proc) { - if (adev->kfd.dev) - kgd2kfd_suspend(adev->kfd.dev, run_pm); + if (adev->kfd.dev) { + if (adev->in_s0ix) + kgd2kfd_stop_sched_all_nodes(adev->kfd.dev); + else + kgd2kfd_suspend(adev->kfd.dev, suspend_proc); + } } -int amdgpu_amdkfd_resume_iommu(struct amdgpu_device *adev) +int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool resume_proc) { int r = 0; - if (adev->kfd.dev) - r = kgd2kfd_resume_iommu(adev->kfd.dev); + if (adev->kfd.dev) { + if (adev->in_s0ix) + r = kgd2kfd_start_sched_all_nodes(adev->kfd.dev); + else + r = kgd2kfd_resume(adev->kfd.dev, resume_proc); + } return r; } -int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm) +void amdgpu_amdkfd_suspend_process(struct amdgpu_device *adev) +{ + if (adev->kfd.dev) + kgd2kfd_suspend_process(adev->kfd.dev); +} + +int amdgpu_amdkfd_resume_process(struct amdgpu_device *adev) { int r = 0; if (adev->kfd.dev) - r = kgd2kfd_resume(adev->kfd.dev, run_pm); + r = kgd2kfd_resume_process(adev->kfd.dev); return r; } -int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev) +int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev, + struct amdgpu_reset_context *reset_context) { int r = 0; if (adev->kfd.dev) - r = kgd2kfd_pre_reset(adev->kfd.dev); + r = kgd2kfd_pre_reset(adev->kfd.dev, reset_context); return r; } @@ -344,15 +388,18 @@ allocate_mem_reserve_bo_failed: return r; } -void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void *mem_obj) +void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void **mem_obj) { - struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj; + struct amdgpu_bo **bo = (struct amdgpu_bo **) mem_obj; - amdgpu_bo_reserve(bo, true); - amdgpu_bo_kunmap(bo); - amdgpu_bo_unpin(bo); - amdgpu_bo_unreserve(bo); - amdgpu_bo_unref(&(bo)); + if (!bo || !*bo) + return; + + (void)amdgpu_bo_reserve(*bo, true); + amdgpu_bo_kunmap(*bo); + amdgpu_bo_unpin(*bo); + amdgpu_bo_unreserve(*bo); + amdgpu_bo_unref(bo); } int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size, @@ -427,14 +474,26 @@ uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev, } void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, - struct kfd_local_mem_info *mem_info) + struct kfd_local_mem_info *mem_info, + struct amdgpu_xcp *xcp) { memset(mem_info, 0, sizeof(*mem_info)); - mem_info->local_mem_size_public = adev->gmc.visible_vram_size; - mem_info->local_mem_size_private = adev->gmc.real_vram_size - + if (xcp) { + if (adev->gmc.real_vram_size == adev->gmc.visible_vram_size) + mem_info->local_mem_size_public = + KFD_XCP_MEMORY_SIZE(adev, xcp->id); + else + mem_info->local_mem_size_private = + KFD_XCP_MEMORY_SIZE(adev, xcp->id); + } else if (adev->apu_prefer_gtt) { + mem_info->local_mem_size_public = (ttm_tt_pages_limit() << PAGE_SHIFT); + mem_info->local_mem_size_private = 0; + } else { + mem_info->local_mem_size_public = adev->gmc.visible_vram_size; + mem_info->local_mem_size_private = adev->gmc.real_vram_size - adev->gmc.visible_vram_size; - + } mem_info->vram_width = adev->gmc.vram_width; pr_debug("Address base: %pap public 0x%llx private 0x%llx\n", @@ -442,9 +501,7 @@ void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, mem_info->local_mem_size_public, mem_info->local_mem_size_private); - if (amdgpu_sriov_vf(adev)) - mem_info->mem_clk_max = adev->clock.default_mclk / 100; - else if (adev->pm.dpm_enabled) { + if (adev->pm.dpm_enabled) { if (amdgpu_emu_mode == 1) mem_info->mem_clk_max = 0; else @@ -463,41 +520,17 @@ uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev) uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev) { /* the sclk is in quantas of 10kHz */ - if (amdgpu_sriov_vf(adev)) - return adev->clock.default_sclk / 100; - else if (adev->pm.dpm_enabled) + if (adev->pm.dpm_enabled) return amdgpu_dpm_get_sclk(adev, false) / 100; else return 100; } -void amdgpu_amdkfd_get_cu_info(struct amdgpu_device *adev, struct kfd_cu_info *cu_info) -{ - struct amdgpu_cu_info acu_info = adev->gfx.cu_info; - - memset(cu_info, 0, sizeof(*cu_info)); - if (sizeof(cu_info->cu_bitmap) != sizeof(acu_info.bitmap)) - return; - - cu_info->cu_active_number = acu_info.number; - cu_info->cu_ao_mask = acu_info.ao_cu_mask; - memcpy(&cu_info->cu_bitmap[0], &acu_info.bitmap[0], - sizeof(acu_info.bitmap)); - cu_info->num_shader_engines = adev->gfx.config.max_shader_engines; - cu_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se; - cu_info->num_cu_per_sh = adev->gfx.config.max_cu_per_sh; - cu_info->simd_per_cu = acu_info.simd_per_cu; - cu_info->max_waves_per_simd = acu_info.max_waves_per_simd; - cu_info->wave_front_size = acu_info.wave_front_size; - cu_info->max_scratch_slots_per_cu = acu_info.max_scratch_slots_per_cu; - cu_info->lds_size = acu_info.lds_size; -} - int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd, struct amdgpu_device **dmabuf_adev, uint64_t *bo_size, void *metadata_buffer, size_t buffer_size, uint32_t *metadata_size, - uint32_t *flags) + uint32_t *flags, int8_t *xcp_id) { struct dma_buf *dma_buf; struct drm_gem_object *obj; @@ -541,54 +574,14 @@ int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd, if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) *flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC; } + if (xcp_id) + *xcp_id = bo->xcp_id; out_put: dma_buf_put(dma_buf); return r; } -uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst, - struct amdgpu_device *src) -{ - struct amdgpu_device *peer_adev = src; - struct amdgpu_device *adev = dst; - int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev); - - if (ret < 0) { - DRM_ERROR("amdgpu: failed to get xgmi hops count between node %d and %d. ret = %d\n", - adev->gmc.xgmi.physical_node_id, - peer_adev->gmc.xgmi.physical_node_id, ret); - ret = 0; - } - return (uint8_t)ret; -} - -int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst, - struct amdgpu_device *src, - bool is_min) -{ - struct amdgpu_device *adev = dst, *peer_adev; - int num_links; - - if (adev->asic_type != CHIP_ALDEBARAN) - return 0; - - if (src) - peer_adev = src; - - /* num links returns 0 for indirect peers since indirect route is unknown. */ - num_links = is_min ? 1 : amdgpu_xgmi_get_num_links(adev, peer_adev); - if (num_links < 0) { - DRM_ERROR("amdgpu: failed to get xgmi num links between node %d and %d. ret = %d\n", - adev->gmc.xgmi.physical_node_id, - peer_adev->gmc.xgmi.physical_node_id, num_links); - num_links = 0; - } - - /* Aldebaran xGMI DPM is defeatured so assume x16 x 25Gbps for bandwidth. */ - return (num_links * 16 * 25000)/BITS_PER_BYTE; -} - int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min) { int num_lanes_shift = (is_min ? ffs(adev->pm.pcie_mlw_mask) : @@ -673,7 +666,7 @@ int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev, goto err; } - ret = amdgpu_job_alloc(adev, NULL, NULL, NULL, 1, &job); + ret = amdgpu_job_alloc(adev, NULL, NULL, NULL, 1, &job, 0); if (ret) goto err; @@ -706,12 +699,20 @@ err: void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle) { - /* Temporary workaround to fix issues observed in some - * compute applications when GFXOFF is enabled on GFX11. - */ - if (IP_VERSION_MAJ(adev->ip_versions[GC_HWIP][0]) == 11) { + enum amd_powergating_state state = idle ? AMD_PG_STATE_GATE : AMD_PG_STATE_UNGATE; + if ((IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0)) == 11 && + ((adev->mes.kiq_version & AMDGPU_MES_VERSION_MASK) <= 64)) || + (IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0)) == 12)) { pr_debug("GFXOFF is %s\n", idle ? "enabled" : "disabled"); amdgpu_gfx_off_ctrl(adev, idle); + } else if ((IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0)) == 9) && + (adev->flags & AMD_IS_APU)) { + /* Disable GFXOFF and PG. Temporary workaround + * to fix some compute applications issue on GFX9. + */ + struct amdgpu_ip_block *gfx_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX); + if (gfx_block != NULL) + gfx_block->version->funcs->set_powergating_state((void *)gfx_block, state); } amdgpu_dpm_switch_power_profile(adev, PP_SMC_POWER_PROFILE_COMPUTE, @@ -726,47 +727,186 @@ bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid) return false; } -int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev, - uint16_t vmid) +bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev) { - if (adev->family == AMDGPU_FAMILY_AI) { - int i; + return adev->have_atomics_support; +} - for (i = 0; i < adev->num_vmhubs; i++) - amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0); - } else { - amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0); - } +void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev) +{ + amdgpu_device_flush_hdp(adev, NULL); +} + +bool amdgpu_amdkfd_is_fed(struct amdgpu_device *adev) +{ + return amdgpu_ras_get_fed_status(adev); +} + +void amdgpu_amdkfd_ras_pasid_poison_consumption_handler(struct amdgpu_device *adev, + enum amdgpu_ras_block block, uint16_t pasid, + pasid_notify pasid_fn, void *data, uint32_t reset) +{ + amdgpu_umc_pasid_poison_handler(adev, block, pasid, pasid_fn, data, reset); +} + +void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, + enum amdgpu_ras_block block, uint32_t reset) +{ + amdgpu_umc_pasid_poison_handler(adev, block, 0, NULL, NULL, reset); +} + +int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev, + uint32_t *payload) +{ + int ret; + + /* Device or IH ring is not ready so bail. */ + ret = amdgpu_ih_wait_on_checkpoint_process_ts(adev, &adev->irq.ih); + if (ret) + return ret; + + /* Send payload to fence KFD interrupts */ + amdgpu_amdkfd_interrupt(adev, payload); return 0; } -int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev, - uint16_t pasid, enum TLB_FLUSH_TYPE flush_type) +int amdgpu_amdkfd_check_and_lock_kfd(struct amdgpu_device *adev) { - bool all_hub = false; + return kgd2kfd_check_and_lock_kfd(adev->kfd.dev); +} - if (adev->family == AMDGPU_FAMILY_AI || - adev->family == AMDGPU_FAMILY_RV) - all_hub = true; +void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev) +{ + kgd2kfd_unlock_kfd(adev->kfd.dev); +} - return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub); + +u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id) +{ + s8 mem_id = KFD_XCP_MEM_ID(adev, xcp_id); + u64 tmp; + + if (adev->gmc.num_mem_partitions && xcp_id >= 0 && mem_id >= 0) { + if (adev->gmc.is_app_apu && adev->gmc.num_mem_partitions == 1) { + /* In NPS1 mode, we should restrict the vram reporting + * tied to the ttm_pages_limit which is 1/2 of the system + * memory. For other partition modes, the HBM is uniformly + * divided already per numa node reported. If user wants to + * go beyond the default ttm limit and maximize the ROCm + * allocations, they can go up to max ttm and sysmem limits. + */ + + tmp = (ttm_tt_pages_limit() << PAGE_SHIFT) / num_online_nodes(); + } else { + tmp = adev->gmc.mem_partitions[mem_id].size; + } + do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition); + return ALIGN_DOWN(tmp, PAGE_SIZE); + } else if (adev->apu_prefer_gtt) { + return (ttm_tt_pages_limit() << PAGE_SHIFT); + } else { + return adev->gmc.real_vram_size; + } } -bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev) +int amdgpu_amdkfd_unmap_hiq(struct amdgpu_device *adev, u32 doorbell_off, + u32 inst) { - return adev->have_atomics_support; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst]; + struct amdgpu_ring *kiq_ring = &kiq->ring; + struct amdgpu_ring_funcs *ring_funcs; + struct amdgpu_ring *ring; + int r = 0; + + if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues) + return -EINVAL; + + if (!kiq_ring->sched.ready || amdgpu_in_reset(adev)) + return 0; + + ring_funcs = kzalloc(sizeof(*ring_funcs), GFP_KERNEL); + if (!ring_funcs) + return -ENOMEM; + + ring = kzalloc(sizeof(*ring), GFP_KERNEL); + if (!ring) { + r = -ENOMEM; + goto free_ring_funcs; + } + + ring_funcs->type = AMDGPU_RING_TYPE_COMPUTE; + ring->doorbell_index = doorbell_off; + ring->funcs = ring_funcs; + + spin_lock(&kiq->ring_lock); + + if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size)) { + spin_unlock(&kiq->ring_lock); + r = -ENOMEM; + goto free_ring; + } + + kiq->pmf->kiq_unmap_queues(kiq_ring, ring, RESET_QUEUES, 0, 0); + + /* Submit unmap queue packet */ + amdgpu_ring_commit(kiq_ring); + /* + * Ring test will do a basic scratch register change check. Just run + * this to ensure that unmap queues that is submitted before got + * processed successfully before returning. + */ + r = amdgpu_ring_test_helper(kiq_ring); + + spin_unlock(&kiq->ring_lock); + +free_ring: + kfree(ring); + +free_ring_funcs: + kfree(ring_funcs); + + return r; +} + +/* Stop scheduling on KFD */ +int amdgpu_amdkfd_stop_sched(struct amdgpu_device *adev, uint32_t node_id) +{ + if (!adev->kfd.init_complete) + return 0; + + return kgd2kfd_stop_sched(adev->kfd.dev, node_id); } -void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, bool reset) +/* Start scheduling on KFD */ +int amdgpu_amdkfd_start_sched(struct amdgpu_device *adev, uint32_t node_id) { - amdgpu_umc_poison_handler(adev, reset); + if (!adev->kfd.init_complete) + return 0; + + return kgd2kfd_start_sched(adev->kfd.dev, node_id); } -bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev) +/* check if there are KFD queues active */ +bool amdgpu_amdkfd_compute_active(struct amdgpu_device *adev, uint32_t node_id) { - if (adev->gfx.ras && adev->gfx.ras->query_utcl2_poison_status) - return adev->gfx.ras->query_utcl2_poison_status(adev); - else + if (!adev->kfd.init_complete) return false; + + return kgd2kfd_compute_active(adev->kfd.dev, node_id); +} + +/* Config CGTT_SQ_CLK_CTRL */ +int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id, + bool core_override_enable, bool reg_override_enable, bool perfmon_override_enable) +{ + int r; + + if (!adev->kfd.init_complete) + return 0; + + r = psp_config_sq_perfmon(&adev->psp, xcp_id, core_override_enable, + reg_override_enable, perfmon_override_enable); + + return r; } |