diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 319 |
1 files changed, 199 insertions, 120 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index d3da9dde4ee1..7077f21f0021 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -21,6 +21,7 @@ */ #include "amdgpu_amdkfd.h" +#include "amd_pcie.h" #include "amd_shared.h" #include "amdgpu.h" @@ -29,44 +30,49 @@ #include <linux/module.h> #include <linux/dma-buf.h> #include "amdgpu_xgmi.h" - -static const unsigned int compute_vmid_bitmap = 0xFF00; +#include <uapi/linux/kfd_ioctl.h> +#include "amdgpu_ras.h" +#include "amdgpu_umc.h" /* Total memory size in system memory and all GPU VRAM. Used to * estimate worst case amount of memory to reserve for page tables */ uint64_t amdgpu_amdkfd_total_mem_size; +static bool kfd_initialized; + int amdgpu_amdkfd_init(void) { struct sysinfo si; int ret; si_meminfo(&si); - amdgpu_amdkfd_total_mem_size = si.totalram - si.totalhigh; + amdgpu_amdkfd_total_mem_size = si.freeram - si.freehigh; amdgpu_amdkfd_total_mem_size *= si.mem_unit; -#ifdef CONFIG_HSA_AMD ret = kgd2kfd_init(); amdgpu_amdkfd_gpuvm_init_mem_limits(); -#else - ret = -ENOENT; -#endif + kfd_initialized = !ret; return ret; } void amdgpu_amdkfd_fini(void) { - kgd2kfd_exit(); + if (kfd_initialized) { + kgd2kfd_exit(); + kfd_initialized = false; + } } void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev) { bool vf = amdgpu_sriov_vf(adev); - adev->kfd.dev = kgd2kfd_probe((struct kgd_dev *)adev, - adev->pdev, adev->asic_type, vf); + if (!kfd_initialized) + return; + + adev->kfd.dev = kgd2kfd_probe((struct kgd_dev *)adev, vf); if (adev->kfd.dev) amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size; @@ -112,13 +118,15 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) if (adev->kfd.dev) { struct kgd2kfd_shared_resources gpu_resources = { - .compute_vmid_bitmap = compute_vmid_bitmap, + .compute_vmid_bitmap = + ((1 << AMDGPU_NUM_VMID) - 1) - + ((1 << adev->vm_manager.first_kfd_vmid) - 1), .num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec, .num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe, .gpuvm_size = min(adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT, AMDGPU_GMC_HOLE_START), - .drm_render_minor = adev->ddev->render->index, + .drm_render_minor = adev_to_drm(adev)->render->index, .sdma_doorbell_idx = adev->doorbell_index.sdma_engine, }; @@ -126,7 +134,7 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) /* this is going to have a few of the MSBs set that we need to * clear */ - bitmap_complement(gpu_resources.queue_bitmap, + bitmap_complement(gpu_resources.cp_queue_bitmap, adev->gfx.mec.queue_bitmap, KGD_MAX_QUEUES); @@ -137,7 +145,7 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) * adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_queue_per_pipe; for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i) - clear_bit(i, gpu_resources.queue_bitmap); + clear_bit(i, gpu_resources.cp_queue_bitmap); amdgpu_doorbell_get_kfd_info(adev, &gpu_resources.doorbell_physical_address, @@ -159,11 +167,12 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) adev->doorbell_index.last_non_cp; } - kgd2kfd_device_init(adev->kfd.dev, adev->ddev, &gpu_resources); + adev->kfd.init_complete = kgd2kfd_device_init(adev->kfd.dev, + adev_to_drm(adev), &gpu_resources); } } -void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev) +void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev) { if (adev->kfd.dev) { kgd2kfd_device_exit(adev->kfd.dev); @@ -178,18 +187,28 @@ void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev, kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry); } -void amdgpu_amdkfd_suspend(struct amdgpu_device *adev) +void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm) { if (adev->kfd.dev) - kgd2kfd_suspend(adev->kfd.dev); + kgd2kfd_suspend(adev->kfd.dev, run_pm); } -int amdgpu_amdkfd_resume(struct amdgpu_device *adev) +int amdgpu_amdkfd_resume_iommu(struct amdgpu_device *adev) { int r = 0; if (adev->kfd.dev) - r = kgd2kfd_resume(adev->kfd.dev); + r = kgd2kfd_resume_iommu(adev->kfd.dev); + + return r; +} + +int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm) +{ + int r = 0; + + if (adev->kfd.dev) + r = kgd2kfd_resume(adev->kfd.dev, run_pm); return r; } @@ -224,7 +243,7 @@ void amdgpu_amdkfd_gpu_reset(struct kgd_dev *kgd) int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size, void **mem_obj, uint64_t *gpu_addr, - void **cpu_ptr, bool mqd_gfx9) + void **cpu_ptr, bool cp_mqd_gfx9) { struct amdgpu_device *adev = (struct amdgpu_device *)kgd; struct amdgpu_bo *bo = NULL; @@ -239,9 +258,10 @@ int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size, bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC; bp.type = ttm_bo_type_kernel; bp.resv = NULL; + bp.bo_ptr_size = sizeof(struct amdgpu_bo); - if (mqd_gfx9) - bp.flags |= AMDGPU_GEM_CREATE_MQD_GFX9; + if (cp_mqd_gfx9) + bp.flags |= AMDGPU_GEM_CREATE_CP_MQD_GFX9; r = amdgpu_bo_create(adev, &bp, &bo); if (r) { @@ -310,6 +330,7 @@ int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size, { struct amdgpu_device *adev = (struct amdgpu_device *)kgd; struct amdgpu_bo *bo = NULL; + struct amdgpu_bo_user *ubo; struct amdgpu_bo_param bp; int r; @@ -320,14 +341,16 @@ int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size, bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS; bp.type = ttm_bo_type_device; bp.resv = NULL; + bp.bo_ptr_size = sizeof(struct amdgpu_bo); - r = amdgpu_bo_create(adev, &bp, &bo); + r = amdgpu_bo_create_user(adev, &bp, &ubo); if (r) { dev_err(adev->dev, "failed to allocate gws BO for amdkfd (%d)\n", r); return r; } + bo = &ubo->bo; *mem_obj = bo; return 0; } @@ -380,29 +403,23 @@ void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd, struct kfd_local_mem_info *mem_info) { struct amdgpu_device *adev = (struct amdgpu_device *)kgd; - uint64_t address_mask = adev->dev->dma_mask ? ~*adev->dev->dma_mask : - ~((1ULL << 32) - 1); - resource_size_t aper_limit = adev->gmc.aper_base + adev->gmc.aper_size; memset(mem_info, 0, sizeof(*mem_info)); - if (!(adev->gmc.aper_base & address_mask || aper_limit & address_mask)) { - 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; - } else { - mem_info->local_mem_size_public = 0; - mem_info->local_mem_size_private = adev->gmc.real_vram_size; - } + + 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 limit %pap public 0x%llx private 0x%llx\n", - &adev->gmc.aper_base, &aper_limit, + pr_debug("Address base: %pap public 0x%llx private 0x%llx\n", + &adev->gmc.aper_base, 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->powerplay.pp_funcs) { + else if (adev->pm.dpm_enabled) { if (amdgpu_emu_mode == 1) mem_info->mem_clk_max = 0; else @@ -427,7 +444,7 @@ uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd) /* the sclk is in quantas of 10kHz */ if (amdgpu_sriov_vf(adev)) return adev->clock.default_sclk / 100; - else if (adev->powerplay.pp_funcs) + else if (adev->pm.dpm_enabled) return amdgpu_dpm_get_sclk(adev, false) / 100; else return 100; @@ -478,11 +495,11 @@ int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd, goto out_put; obj = dma_buf->priv; - if (obj->dev->driver != adev->ddev->driver) + if (obj->dev->driver != adev_to_drm(adev)->driver) /* Can't handle buffers from different drivers */ goto out_put; - adev = obj->dev->dev_private; + adev = drm_to_adev(obj->dev); bo = gem_to_amdgpu_bo(obj); if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT))) @@ -494,17 +511,16 @@ int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd, *dma_buf_kgd = (struct kgd_dev *)adev; if (bo_size) *bo_size = amdgpu_bo_size(bo); - if (metadata_size) - *metadata_size = bo->metadata_size; if (metadata_buffer) r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size, metadata_size, &metadata_flags); if (flags) { *flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? - ALLOC_MEM_FLAGS_VRAM : ALLOC_MEM_FLAGS_GTT; + KFD_IOC_ALLOC_MEM_FLAGS_VRAM + : KFD_IOC_ALLOC_MEM_FLAGS_GTT; if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) - *flags |= ALLOC_MEM_FLAGS_PUBLIC; + *flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC; } out_put: @@ -515,8 +531,9 @@ out_put: uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd) { struct amdgpu_device *adev = (struct amdgpu_device *)kgd; + struct ttm_resource_manager *vram_man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM); - return amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]); + return amdgpu_vram_mgr_usage(vram_man); } uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd) @@ -525,6 +542,14 @@ uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd) return adev->gmc.xgmi.hive_id; } + +uint64_t amdgpu_amdkfd_get_unique_id(struct kgd_dev *kgd) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)kgd; + + return adev->unique_id; +} + uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *src) { struct amdgpu_device *peer_adev = (struct amdgpu_device *)src; @@ -540,6 +565,88 @@ uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *s return (uint8_t)ret; } +int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct kgd_dev *dst, struct kgd_dev *src, bool is_min) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)dst, *peer_adev; + int num_links; + + if (adev->asic_type != CHIP_ALDEBARAN) + return 0; + + if (src) + peer_adev = (struct amdgpu_device *)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 kgd_dev *dev, bool is_min) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)dev; + int num_lanes_shift = (is_min ? ffs(adev->pm.pcie_mlw_mask) : + fls(adev->pm.pcie_mlw_mask)) - 1; + int gen_speed_shift = (is_min ? ffs(adev->pm.pcie_gen_mask & + CAIL_PCIE_LINK_SPEED_SUPPORT_MASK) : + fls(adev->pm.pcie_gen_mask & + CAIL_PCIE_LINK_SPEED_SUPPORT_MASK)) - 1; + uint32_t num_lanes_mask = 1 << num_lanes_shift; + uint32_t gen_speed_mask = 1 << gen_speed_shift; + int num_lanes_factor = 0, gen_speed_mbits_factor = 0; + + switch (num_lanes_mask) { + case CAIL_PCIE_LINK_WIDTH_SUPPORT_X1: + num_lanes_factor = 1; + break; + case CAIL_PCIE_LINK_WIDTH_SUPPORT_X2: + num_lanes_factor = 2; + break; + case CAIL_PCIE_LINK_WIDTH_SUPPORT_X4: + num_lanes_factor = 4; + break; + case CAIL_PCIE_LINK_WIDTH_SUPPORT_X8: + num_lanes_factor = 8; + break; + case CAIL_PCIE_LINK_WIDTH_SUPPORT_X12: + num_lanes_factor = 12; + break; + case CAIL_PCIE_LINK_WIDTH_SUPPORT_X16: + num_lanes_factor = 16; + break; + case CAIL_PCIE_LINK_WIDTH_SUPPORT_X32: + num_lanes_factor = 32; + break; + } + + switch (gen_speed_mask) { + case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1: + gen_speed_mbits_factor = 2500; + break; + case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2: + gen_speed_mbits_factor = 5000; + break; + case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3: + gen_speed_mbits_factor = 8000; + break; + case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4: + gen_speed_mbits_factor = 16000; + break; + case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5: + gen_speed_mbits_factor = 32000; + break; + } + + return (num_lanes_factor * gen_speed_mbits_factor)/BITS_PER_BYTE; +} + uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd) { struct amdgpu_device *adev = (struct amdgpu_device *)kgd; @@ -554,6 +661,20 @@ uint32_t amdgpu_amdkfd_get_num_gws(struct kgd_dev *kgd) return adev->gds.gws_size; } +uint32_t amdgpu_amdkfd_get_asic_rev_id(struct kgd_dev *kgd) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)kgd; + + return adev->rev_id; +} + +int amdgpu_amdkfd_get_noretry(struct kgd_dev *kgd) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)kgd; + + return adev->gmc.noretry; +} + int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine, uint32_t vmid, uint64_t gpu_addr, uint32_t *ib_cmd, uint32_t ib_len) @@ -595,6 +716,7 @@ int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine, job->vmid = vmid; ret = amdgpu_ib_schedule(ring, 1, ib, job, &f); + if (ret) { DRM_ERROR("amdgpu: failed to schedule IB.\n"); goto err_ib_sched; @@ -603,7 +725,6 @@ int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine, ret = dma_fence_wait(f, false); err_ib_sched: - dma_fence_put(f); amdgpu_job_free(job); err: return ret; @@ -613,104 +734,62 @@ void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle) { struct amdgpu_device *adev = (struct amdgpu_device *)kgd; - if (is_support_sw_smu(adev)) - smu_switch_power_profile(&adev->smu, - PP_SMC_POWER_PROFILE_COMPUTE, - !idle); - else if (adev->powerplay.pp_funcs && - adev->powerplay.pp_funcs->switch_power_profile) - amdgpu_dpm_switch_power_profile(adev, - PP_SMC_POWER_PROFILE_COMPUTE, - !idle); + amdgpu_dpm_switch_power_profile(adev, + PP_SMC_POWER_PROFILE_COMPUTE, + !idle); } bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid) { - if (adev->kfd.dev) { - if ((1 << vmid) & compute_vmid_bitmap) - return true; - } + if (adev->kfd.dev) + return vmid >= adev->vm_manager.first_kfd_vmid; return false; } -bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd) +int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct kgd_dev *kgd, uint16_t vmid) { struct amdgpu_device *adev = (struct amdgpu_device *)kgd; - return adev->have_atomics_support; -} - -#ifndef CONFIG_HSA_AMD -bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm) -{ - return false; -} - -void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo) -{ -} - -void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, - struct amdgpu_vm *vm) -{ -} + if (adev->family == AMDGPU_FAMILY_AI) { + int i; -struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f) -{ - return NULL; -} + 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); + } -int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm) -{ return 0; } -struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev, - unsigned int asic_type, bool vf) -{ - return NULL; -} - -bool kgd2kfd_device_init(struct kfd_dev *kfd, - struct drm_device *ddev, - const struct kgd2kfd_shared_resources *gpu_resources) -{ - return false; -} - -void kgd2kfd_device_exit(struct kfd_dev *kfd) -{ -} - -void kgd2kfd_exit(void) +int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct kgd_dev *kgd, uint16_t pasid, + enum TLB_FLUSH_TYPE flush_type) { -} + struct amdgpu_device *adev = (struct amdgpu_device *)kgd; + bool all_hub = false; -void kgd2kfd_suspend(struct kfd_dev *kfd) -{ -} + if (adev->family == AMDGPU_FAMILY_AI) + all_hub = true; -int kgd2kfd_resume(struct kfd_dev *kfd) -{ - return 0; + return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub); } -int kgd2kfd_pre_reset(struct kfd_dev *kfd) +bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd) { - return 0; -} + struct amdgpu_device *adev = (struct amdgpu_device *)kgd; -int kgd2kfd_post_reset(struct kfd_dev *kfd) -{ - return 0; + return adev->have_atomics_support; } -void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) +void amdgpu_amdkfd_ras_poison_consumption_handler(struct kgd_dev *kgd) { -} + struct amdgpu_device *adev = (struct amdgpu_device *)kgd; + struct ras_err_data err_data = {0, 0, 0, NULL}; -void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd) -{ + /* CPU MCA will handle page retirement if connected_to_cpu is 1 */ + if (!adev->gmc.xgmi.connected_to_cpu) + amdgpu_umc_process_ras_data_cb(adev, &err_data, NULL); + else + amdgpu_amdkfd_gpu_reset(kgd); } -#endif |