diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 79 |
1 files changed, 60 insertions, 19 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 776a947b45df..5d9a34601a1a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -1,3 +1,4 @@ +// SPDX-License-Identifier: MIT /* * Copyright 2014 Advanced Micro Devices, Inc. * @@ -33,6 +34,7 @@ #include <uapi/linux/kfd_ioctl.h> #include "amdgpu_ras.h" #include "amdgpu_umc.h" +#include "amdgpu_reset.h" /* Total memory size in system memory and all GPU VRAM. Used to * estimate worst case amount of memory to reserve for page tables @@ -73,9 +75,6 @@ void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev) return; adev->kfd.dev = kgd2kfd_probe(adev, vf); - - if (adev->kfd.dev) - amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size; } /** @@ -100,7 +99,18 @@ static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev, * The first num_doorbells are used by amdgpu. * amdkfd takes whatever's left in the aperture. */ - if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) { + if (adev->enable_mes) { + /* + * With MES enabled, we only need to initialize + * the base address. The size and offset are + * not initialized as AMDGPU manages the whole + * doorbell space. + */ + *aperture_base = adev->doorbell.base; + *aperture_size = 0; + *start_offset = 0; + } else if (adev->doorbell.size > adev->doorbell.num_doorbells * + sizeof(u32)) { *aperture_base = adev->doorbell.base; *aperture_size = adev->doorbell.size; *start_offset = adev->doorbell.num_doorbells * sizeof(u32); @@ -111,6 +121,23 @@ static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev, } } + +static void amdgpu_amdkfd_reset_work(struct work_struct *work) +{ + struct amdgpu_device *adev = container_of(work, struct amdgpu_device, + kfd.reset_work); + + struct amdgpu_reset_context reset_context; + + memset(&reset_context, 0, sizeof(reset_context)); + + reset_context.method = AMD_RESET_METHOD_NONE; + reset_context.reset_req_dev = adev; + clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); + + amdgpu_device_gpu_recover(adev, NULL, &reset_context); +} + void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) { int i; @@ -128,7 +155,7 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) AMDGPU_GMC_HOLE_START), .drm_render_minor = adev_to_drm(adev)->render->index, .sdma_doorbell_idx = adev->doorbell_index.sdma_engine, - + .enable_mes = adev->enable_mes, }; /* this is going to have a few of the MSBs set that we need to @@ -169,6 +196,10 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) adev->kfd.init_complete = kgd2kfd_device_init(adev->kfd.dev, adev_to_drm(adev), &gpu_resources); + + amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size; + + INIT_WORK(&adev->kfd.reset_work, amdgpu_amdkfd_reset_work); } } @@ -177,6 +208,7 @@ void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev) if (adev->kfd.dev) { kgd2kfd_device_exit(adev->kfd.dev); adev->kfd.dev = NULL; + amdgpu_amdkfd_total_mem_size -= adev->gmc.real_vram_size; } } @@ -236,7 +268,8 @@ int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev) void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev) { if (amdgpu_device_should_recover_gpu(adev)) - amdgpu_device_gpu_recover(adev, NULL); + amdgpu_reset_domain_schedule(adev->reset_domain, + &adev->kfd.reset_work); } int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size, @@ -514,13 +547,6 @@ out_put: return r; } -uint64_t amdgpu_amdkfd_get_vram_usage(struct amdgpu_device *adev) -{ - struct ttm_resource_manager *vram_man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM); - - return amdgpu_vram_mgr_usage(vram_man); -} - uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst, struct amdgpu_device *src) { @@ -659,6 +685,7 @@ int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev, ib->length_dw = ib_len; /* This works for NO_HWS. TODO: need to handle without knowing VMID */ job->vmid = vmid; + job->num_ibs = 1; ret = amdgpu_ib_schedule(ring, 1, ib, job, &f); @@ -667,6 +694,8 @@ int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev, goto err_ib_sched; } + /* Drop the initial kref_init count (see drm_sched_main as example) */ + dma_fence_put(f); ret = dma_fence_wait(f, false); err_ib_sched: @@ -677,6 +706,13 @@ 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) { + pr_debug("GFXOFF is %s\n", idle ? "enabled" : "disabled"); + amdgpu_gfx_off_ctrl(adev, idle); + } amdgpu_dpm_switch_power_profile(adev, PP_SMC_POWER_PROFILE_COMPUTE, !idle); @@ -710,7 +746,8 @@ int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev, { bool all_hub = false; - if (adev->family == AMDGPU_FAMILY_AI) + if (adev->family == AMDGPU_FAMILY_AI || + adev->family == AMDGPU_FAMILY_RV) all_hub = true; return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub); @@ -725,9 +762,13 @@ void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, bo { struct ras_err_data err_data = {0, 0, 0, NULL}; - /* CPU MCA will handle page retirement if connected_to_cpu is 1 */ - if (!adev->gmc.xgmi.connected_to_cpu) - amdgpu_umc_poison_handler(adev, &err_data, reset); - else if (reset) - amdgpu_amdkfd_gpu_reset(adev); + amdgpu_umc_poison_handler(adev, &err_data, reset); +} + +bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev) +{ + if (adev->gfx.ras && adev->gfx.ras->query_utcl2_poison_status) + return adev->gfx.ras->query_utcl2_poison_status(adev); + else + return false; } |