1130e0371SOded Gabbay /* 2130e0371SOded Gabbay * Copyright 2014 Advanced Micro Devices, Inc. 3130e0371SOded Gabbay * 4130e0371SOded Gabbay * Permission is hereby granted, free of charge, to any person obtaining a 5130e0371SOded Gabbay * copy of this software and associated documentation files (the "Software"), 6130e0371SOded Gabbay * to deal in the Software without restriction, including without limitation 7130e0371SOded Gabbay * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8130e0371SOded Gabbay * and/or sell copies of the Software, and to permit persons to whom the 9130e0371SOded Gabbay * Software is furnished to do so, subject to the following conditions: 10130e0371SOded Gabbay * 11130e0371SOded Gabbay * The above copyright notice and this permission notice shall be included in 12130e0371SOded Gabbay * all copies or substantial portions of the Software. 13130e0371SOded Gabbay * 14130e0371SOded Gabbay * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15130e0371SOded Gabbay * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16130e0371SOded Gabbay * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17130e0371SOded Gabbay * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18130e0371SOded Gabbay * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19130e0371SOded Gabbay * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20130e0371SOded Gabbay * OTHER DEALINGS IN THE SOFTWARE. 21130e0371SOded Gabbay */ 22130e0371SOded Gabbay 23130e0371SOded Gabbay /* amdgpu_amdkfd.h defines the private interface between amdgpu and amdkfd. */ 24130e0371SOded Gabbay 25130e0371SOded Gabbay #ifndef AMDGPU_AMDKFD_H_INCLUDED 26130e0371SOded Gabbay #define AMDGPU_AMDKFD_H_INCLUDED 27130e0371SOded Gabbay 28130e0371SOded Gabbay #include <linux/types.h> 297420f482SOded Gabbay #include <linux/mm.h> 309bf5b9ebSChristoph Hellwig #include <linux/kthread.h> 315ae0283eSFelix Kuehling #include <linux/workqueue.h> 32130e0371SOded Gabbay #include <kgd_kfd_interface.h> 33a46a2cd1SFelix Kuehling #include <drm/ttm/ttm_execbuf_util.h> 34a46a2cd1SFelix Kuehling #include "amdgpu_sync.h" 35a46a2cd1SFelix Kuehling #include "amdgpu_vm.h" 36130e0371SOded Gabbay 37611736d8SFelix Kuehling extern uint64_t amdgpu_amdkfd_total_mem_size; 38d8d019ccSFelix Kuehling 39130e0371SOded Gabbay struct amdgpu_device; 40130e0371SOded Gabbay 41a46a2cd1SFelix Kuehling struct kfd_bo_va_list { 42a46a2cd1SFelix Kuehling struct list_head bo_list; 43a46a2cd1SFelix Kuehling struct amdgpu_bo_va *bo_va; 44a46a2cd1SFelix Kuehling void *kgd_dev; 45a46a2cd1SFelix Kuehling bool is_mapped; 46a46a2cd1SFelix Kuehling uint64_t va; 47a46a2cd1SFelix Kuehling uint64_t pte_flags; 48a46a2cd1SFelix Kuehling }; 49a46a2cd1SFelix Kuehling 50130e0371SOded Gabbay struct kgd_mem { 51a46a2cd1SFelix Kuehling struct mutex lock; 52130e0371SOded Gabbay struct amdgpu_bo *bo; 53a46a2cd1SFelix Kuehling struct list_head bo_va_list; 54a46a2cd1SFelix Kuehling /* protected by amdkfd_process_info.lock */ 55a46a2cd1SFelix Kuehling struct ttm_validate_buffer validate_list; 56a46a2cd1SFelix Kuehling struct ttm_validate_buffer resv_list; 57a46a2cd1SFelix Kuehling uint32_t domain; 58a46a2cd1SFelix Kuehling unsigned int mapped_to_gpu_memory; 59a46a2cd1SFelix Kuehling uint64_t va; 60a46a2cd1SFelix Kuehling 61d0ba51b1SFelix Kuehling uint32_t alloc_flags; 62a46a2cd1SFelix Kuehling 635ae0283eSFelix Kuehling atomic_t invalid; 64a46a2cd1SFelix Kuehling struct amdkfd_process_info *process_info; 65a46a2cd1SFelix Kuehling 66a46a2cd1SFelix Kuehling struct amdgpu_sync sync; 67a46a2cd1SFelix Kuehling 68a46a2cd1SFelix Kuehling bool aql_queue; 69d4566deeSMukul Joshi bool is_imported; 70130e0371SOded Gabbay }; 71130e0371SOded Gabbay 72d8d019ccSFelix Kuehling /* KFD Memory Eviction */ 73d8d019ccSFelix Kuehling struct amdgpu_amdkfd_fence { 74d8d019ccSFelix Kuehling struct dma_fence base; 75d8d019ccSFelix Kuehling struct mm_struct *mm; 76d8d019ccSFelix Kuehling spinlock_t lock; 77d8d019ccSFelix Kuehling char timeline_name[TASK_COMM_LEN]; 78d8d019ccSFelix Kuehling }; 79d8d019ccSFelix Kuehling 80611736d8SFelix Kuehling struct amdgpu_kfd_dev { 81611736d8SFelix Kuehling struct kfd_dev *dev; 82611736d8SFelix Kuehling uint64_t vram_used; 83611736d8SFelix Kuehling }; 84611736d8SFelix Kuehling 850da8b10eSAmber Lin enum kgd_engine_type { 860da8b10eSAmber Lin KGD_ENGINE_PFP = 1, 870da8b10eSAmber Lin KGD_ENGINE_ME, 880da8b10eSAmber Lin KGD_ENGINE_CE, 890da8b10eSAmber Lin KGD_ENGINE_MEC1, 900da8b10eSAmber Lin KGD_ENGINE_MEC2, 910da8b10eSAmber Lin KGD_ENGINE_RLC, 920da8b10eSAmber Lin KGD_ENGINE_SDMA1, 930da8b10eSAmber Lin KGD_ENGINE_SDMA2, 940da8b10eSAmber Lin KGD_ENGINE_MAX 950da8b10eSAmber Lin }; 960da8b10eSAmber Lin 97d8d019ccSFelix Kuehling struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context, 98d8d019ccSFelix Kuehling struct mm_struct *mm); 99d8d019ccSFelix Kuehling bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm); 100d8d019ccSFelix Kuehling struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f); 101f4a3c42bSxinhui pan int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo); 102d8d019ccSFelix Kuehling 103a46a2cd1SFelix Kuehling struct amdkfd_process_info { 104a46a2cd1SFelix Kuehling /* List head of all VMs that belong to a KFD process */ 105a46a2cd1SFelix Kuehling struct list_head vm_list_head; 106a46a2cd1SFelix Kuehling /* List head for all KFD BOs that belong to a KFD process. */ 107a46a2cd1SFelix Kuehling struct list_head kfd_bo_list; 1085ae0283eSFelix Kuehling /* List of userptr BOs that are valid or invalid */ 1095ae0283eSFelix Kuehling struct list_head userptr_valid_list; 1105ae0283eSFelix Kuehling struct list_head userptr_inval_list; 111a46a2cd1SFelix Kuehling /* Lock to protect kfd_bo_list */ 112a46a2cd1SFelix Kuehling struct mutex lock; 113a46a2cd1SFelix Kuehling 114a46a2cd1SFelix Kuehling /* Number of VMs */ 115a46a2cd1SFelix Kuehling unsigned int n_vms; 116a46a2cd1SFelix Kuehling /* Eviction Fence */ 117a46a2cd1SFelix Kuehling struct amdgpu_amdkfd_fence *eviction_fence; 1185ae0283eSFelix Kuehling 1195ae0283eSFelix Kuehling /* MMU-notifier related fields */ 1205ae0283eSFelix Kuehling atomic_t evicted_bos; 1215ae0283eSFelix Kuehling struct delayed_work restore_userptr_work; 1225ae0283eSFelix Kuehling struct pid *pid; 123a46a2cd1SFelix Kuehling }; 124a46a2cd1SFelix Kuehling 125efb1c658SOded Gabbay int amdgpu_amdkfd_init(void); 126130e0371SOded Gabbay void amdgpu_amdkfd_fini(void); 127130e0371SOded Gabbay 1289593f4d6SRajneesh Bhardwaj void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm); 1299593f4d6SRajneesh Bhardwaj int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm); 130dc102c43SAndres Rodriguez void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev, 131130e0371SOded Gabbay const void *ih_ring_entry); 132dc102c43SAndres Rodriguez void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev); 133dc102c43SAndres Rodriguez void amdgpu_amdkfd_device_init(struct amdgpu_device *adev); 134dc102c43SAndres Rodriguez void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev); 135130e0371SOded Gabbay 136e52482deSFelix Kuehling int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm); 1374c660c8fSFelix Kuehling int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine, 1384c660c8fSFelix Kuehling uint32_t vmid, uint64_t gpu_addr, 1394c660c8fSFelix Kuehling uint32_t *ib_cmd, uint32_t ib_len); 14001c097dbSFelix Kuehling void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle); 141aabf3a95SJack Xiao bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd); 142ffa02269SAlex Sierra int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct kgd_dev *kgd, uint16_t vmid); 143ffa02269SAlex Sierra int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct kgd_dev *kgd, uint16_t pasid); 1444c660c8fSFelix Kuehling 145155494dbSFelix Kuehling bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid); 146155494dbSFelix Kuehling 1475c6dd71eSShaoyun Liu int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev); 1485c6dd71eSShaoyun Liu 1495c6dd71eSShaoyun Liu int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev); 1505c6dd71eSShaoyun Liu 15124da5a9cSShaoyun Liu void amdgpu_amdkfd_gpu_reset(struct kgd_dev *kgd); 15224da5a9cSShaoyun Liu 153d09f85d5SYong Zhao int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev, 154d09f85d5SYong Zhao int queue_bit); 155d09f85d5SYong Zhao 156130e0371SOded Gabbay /* Shared API */ 1577cd52c91SAmber Lin int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size, 158130e0371SOded Gabbay void **mem_obj, uint64_t *gpu_addr, 15915426dbbSYong Zhao void **cpu_ptr, bool mqd_gfx9); 1607cd52c91SAmber Lin void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj); 161ca66fb8fSOak Zeng int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size, void **mem_obj); 162ca66fb8fSOak Zeng void amdgpu_amdkfd_free_gws(struct kgd_dev *kgd, void *mem_obj); 16371efab6aSOak Zeng int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem); 16471efab6aSOak Zeng int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem); 1650da8b10eSAmber Lin uint32_t amdgpu_amdkfd_get_fw_version(struct kgd_dev *kgd, 1660da8b10eSAmber Lin enum kgd_engine_type type); 1677cd52c91SAmber Lin void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd, 16830f1c042SHarish Kasiviswanathan struct kfd_local_mem_info *mem_info); 1697cd52c91SAmber Lin uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd); 170130e0371SOded Gabbay 1717cd52c91SAmber Lin uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd); 1727cd52c91SAmber Lin void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info); 1731dde0ea9SFelix Kuehling int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd, 1741dde0ea9SFelix Kuehling struct kgd_dev **dmabuf_kgd, 1751dde0ea9SFelix Kuehling uint64_t *bo_size, void *metadata_buffer, 1761dde0ea9SFelix Kuehling size_t buffer_size, uint32_t *metadata_size, 1771dde0ea9SFelix Kuehling uint32_t *flags); 1789f0a0b41SKent Russell uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd); 179db8b62c0SShaoyun Liu uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd); 1800c663695SDivya Shikre uint64_t amdgpu_amdkfd_get_unique_id(struct kgd_dev *kgd); 181d8e408a8SOak Zeng uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd); 18229e76462SOak Zeng uint32_t amdgpu_amdkfd_get_num_gws(struct kgd_dev *kgd); 183c6d1ec41SJoseph Greathouse uint32_t amdgpu_amdkfd_get_asic_rev_id(struct kgd_dev *kgd); 184da361dd1Sshaoyunl uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *src); 185130e0371SOded Gabbay 186cd05c865SFelix Kuehling /* Read user wptr from a specified user address space with page fault 187cd05c865SFelix Kuehling * disabled. The memory must be pinned and mapped to the hardware when 188cd05c865SFelix Kuehling * this is called in hqd_load functions, so it should never fault in 189cd05c865SFelix Kuehling * the first place. This resolves a circular lock dependency involving 190c1e8d7c6SMichel Lespinasse * four locks, including the DQM lock and mmap_lock. 191cd05c865SFelix Kuehling */ 19270539bd7SFelix Kuehling #define read_user_wptr(mmptr, wptr, dst) \ 19370539bd7SFelix Kuehling ({ \ 19470539bd7SFelix Kuehling bool valid = false; \ 19570539bd7SFelix Kuehling if ((mmptr) && (wptr)) { \ 196cd05c865SFelix Kuehling pagefault_disable(); \ 19770539bd7SFelix Kuehling if ((mmptr) == current->mm) { \ 19870539bd7SFelix Kuehling valid = !get_user((dst), (wptr)); \ 199*8449d150SChristoph Hellwig } else if (current->flags & PF_KTHREAD) { \ 200f5678e7fSChristoph Hellwig kthread_use_mm(mmptr); \ 20170539bd7SFelix Kuehling valid = !get_user((dst), (wptr)); \ 202f5678e7fSChristoph Hellwig kthread_unuse_mm(mmptr); \ 20370539bd7SFelix Kuehling } \ 204cd05c865SFelix Kuehling pagefault_enable(); \ 20570539bd7SFelix Kuehling } \ 20670539bd7SFelix Kuehling valid; \ 20770539bd7SFelix Kuehling }) 20870539bd7SFelix Kuehling 209a46a2cd1SFelix Kuehling /* GPUVM API */ 2101685b01aSOak Zeng int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, unsigned int pasid, 2111685b01aSOak Zeng void **vm, void **process_info, 212a46a2cd1SFelix Kuehling struct dma_fence **ef); 213ede0dd86SFelix Kuehling int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd, 2141685b01aSOak Zeng struct file *filp, unsigned int pasid, 215ede0dd86SFelix Kuehling void **vm, void **process_info, 216ede0dd86SFelix Kuehling struct dma_fence **ef); 217ede0dd86SFelix Kuehling void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, 218ede0dd86SFelix Kuehling struct amdgpu_vm *vm); 219a46a2cd1SFelix Kuehling void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm); 220bf47afbaSOak Zeng void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *vm); 221e715c6d0SShaoyun Liu uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm); 222a46a2cd1SFelix Kuehling int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( 223a46a2cd1SFelix Kuehling struct kgd_dev *kgd, uint64_t va, uint64_t size, 224a46a2cd1SFelix Kuehling void *vm, struct kgd_mem **mem, 225a46a2cd1SFelix Kuehling uint64_t *offset, uint32_t flags); 226a46a2cd1SFelix Kuehling int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( 227d4566deeSMukul Joshi struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *size); 228a46a2cd1SFelix Kuehling int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( 229a46a2cd1SFelix Kuehling struct kgd_dev *kgd, struct kgd_mem *mem, void *vm); 230a46a2cd1SFelix Kuehling int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( 231a46a2cd1SFelix Kuehling struct kgd_dev *kgd, struct kgd_mem *mem, void *vm); 232a46a2cd1SFelix Kuehling int amdgpu_amdkfd_gpuvm_sync_memory( 233a46a2cd1SFelix Kuehling struct kgd_dev *kgd, struct kgd_mem *mem, bool intr); 234a46a2cd1SFelix Kuehling int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd, 235a46a2cd1SFelix Kuehling struct kgd_mem *mem, void **kptr, uint64_t *size); 236a46a2cd1SFelix Kuehling int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info, 237a46a2cd1SFelix Kuehling struct dma_fence **ef); 238a46a2cd1SFelix Kuehling 239b97dfa27Sshaoyunl int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd, 240b97dfa27Sshaoyunl struct kfd_vm_fault_info *info); 241b97dfa27Sshaoyunl 2421dde0ea9SFelix Kuehling int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd, 2431dde0ea9SFelix Kuehling struct dma_buf *dmabuf, 2441dde0ea9SFelix Kuehling uint64_t va, void *vm, 2451dde0ea9SFelix Kuehling struct kgd_mem **mem, uint64_t *size, 2461dde0ea9SFelix Kuehling uint64_t *mmap_offset); 2471dde0ea9SFelix Kuehling 248a46a2cd1SFelix Kuehling void amdgpu_amdkfd_gpuvm_init_mem_limits(void); 249611736d8SFelix Kuehling void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo); 250a46a2cd1SFelix Kuehling 251fd7d08baSYong Zhao int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd, 252fd7d08baSYong Zhao struct tile_config *config); 253fd7d08baSYong Zhao 2542d3d25b6SAmber Lin /* KGD2KFD callbacks */ 255308176d6SAmber Lin int kgd2kfd_init(void); 2562d3d25b6SAmber Lin void kgd2kfd_exit(void); 2572d3d25b6SAmber Lin struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev, 258050091abSYong Zhao unsigned int asic_type, bool vf); 2592d3d25b6SAmber Lin bool kgd2kfd_device_init(struct kfd_dev *kfd, 2603a0c3423SHarish Kasiviswanathan struct drm_device *ddev, 2612d3d25b6SAmber Lin const struct kgd2kfd_shared_resources *gpu_resources); 2622d3d25b6SAmber Lin void kgd2kfd_device_exit(struct kfd_dev *kfd); 2639593f4d6SRajneesh Bhardwaj void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm); 2649593f4d6SRajneesh Bhardwaj int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm); 2652d3d25b6SAmber Lin int kgd2kfd_pre_reset(struct kfd_dev *kfd); 2662d3d25b6SAmber Lin int kgd2kfd_post_reset(struct kfd_dev *kfd); 2672d3d25b6SAmber Lin void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); 2682d3d25b6SAmber Lin int kgd2kfd_quiesce_mm(struct mm_struct *mm); 2692d3d25b6SAmber Lin int kgd2kfd_resume_mm(struct mm_struct *mm); 2702d3d25b6SAmber Lin int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm, 2712d3d25b6SAmber Lin struct dma_fence *fence); 2729b54d201SEric Huang void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd); 2732d3d25b6SAmber Lin 274130e0371SOded Gabbay #endif /* AMDGPU_AMDKFD_H_INCLUDED */ 275