Linux Audio

Check our new training course

Loading...
v6.13.7
  1/*
  2 * Copyright 2014 Advanced Micro Devices, Inc.
  3 *
  4 * Permission is hereby granted, free of charge, to any person obtaining a
  5 * copy of this software and associated documentation files (the "Software"),
  6 * to deal in the Software without restriction, including without limitation
  7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  8 * and/or sell copies of the Software, and to permit persons to whom the
  9 * Software is furnished to do so, subject to the following conditions:
 10 *
 11 * The above copyright notice and this permission notice shall be included in
 12 * all copies or substantial portions of the Software.
 13 *
 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 20 * OTHER DEALINGS IN THE SOFTWARE.
 21 */
 22
 23/*
 24 * This file defines the private interface between the
 25 * AMD kernel graphics drivers and the AMD KFD.
 26 */
 27
 28#ifndef KGD_KFD_INTERFACE_H_INCLUDED
 29#define KGD_KFD_INTERFACE_H_INCLUDED
 30
 31#include <linux/types.h>
 32#include <linux/bitmap.h>
 33#include <linux/dma-fence.h>
 34#include "amdgpu_irq.h"
 35#include "amdgpu_gfx.h"
 36
 37struct pci_dev;
 38struct amdgpu_device;
 
 
 39
 40struct kfd_dev;
 
 
 41struct kgd_mem;
 42
 43enum kfd_preempt_type {
 44	KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN = 0,
 45	KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
 46	KFD_PREEMPT_TYPE_WAVEFRONT_SAVE
 47};
 48
 49struct kfd_vm_fault_info {
 50	uint64_t	page_addr;
 51	uint32_t	vmid;
 52	uint32_t	mc_id;
 53	uint32_t	status;
 54	bool		prot_valid;
 55	bool		prot_read;
 56	bool		prot_write;
 57	bool		prot_exec;
 
 
 
 58};
 59
 60/* For getting GPU local memory information from KGD */
 61struct kfd_local_mem_info {
 62	uint64_t local_mem_size_private;
 63	uint64_t local_mem_size_public;
 64	uint32_t vram_width;
 65	uint32_t mem_clk_max;
 66};
 67
 68enum kgd_memory_pool {
 69	KGD_POOL_SYSTEM_CACHEABLE = 1,
 70	KGD_POOL_SYSTEM_WRITECOMBINE = 2,
 71	KGD_POOL_FRAMEBUFFER = 3,
 72};
 73
 74struct kfd_cu_occupancy {
 75	u32 wave_cnt;
 76	u32 doorbell_off;
 77};
 78
 79/**
 80 * enum kfd_sched_policy
 81 *
 82 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
 83 * scheduling. In this scheduling mode we're using the firmware code to
 84 * schedule the user mode queues and kernel queues such as HIQ and DIQ.
 85 * the HIQ queue is used as a special queue that dispatches the configuration
 86 * to the cp and the user mode queues list that are currently running.
 87 * the DIQ queue is a debugging queue that dispatches debugging commands to the
 88 * firmware.
 89 * in this scheduling mode user mode queues over subscription feature is
 90 * enabled.
 91 *
 92 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
 93 * subscription feature disabled.
 94 *
 95 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
 96 * set the command processor registers and sets the queues "manually". This
 97 * mode is used *ONLY* for debugging proposes.
 98 *
 99 */
100enum kfd_sched_policy {
101	KFD_SCHED_POLICY_HWS = 0,
102	KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
103	KFD_SCHED_POLICY_NO_HWS
104};
105
106struct kgd2kfd_shared_resources {
107	/* Bit n == 1 means VMID n is available for KFD. */
108	unsigned int compute_vmid_bitmap;
109
110	/* number of pipes per mec */
111	uint32_t num_pipe_per_mec;
112
113	/* number of queues per pipe */
114	uint32_t num_queue_per_pipe;
115
116	/* Bit n == 1 means Queue n is available for KFD */
117	DECLARE_BITMAP(cp_queue_bitmap, AMDGPU_MAX_QUEUES);
118
119	/* SDMA doorbell assignments (SOC15 and later chips only). Only
120	 * specific doorbells are routed to each SDMA engine. Others
121	 * are routed to IH and VCN. They are not usable by the CP.
122	 */
123	uint32_t *sdma_doorbell_idx;
124
125	/* From SOC15 onward, the doorbell index range not usable for CP
126	 * queues.
127	 */
128	uint32_t non_cp_doorbells_start;
129	uint32_t non_cp_doorbells_end;
130
131	/* Base address of doorbell aperture. */
132	phys_addr_t doorbell_physical_address;
133
134	/* Size in bytes of doorbell aperture. */
135	size_t doorbell_aperture_size;
136
137	/* Number of bytes at start of aperture reserved for KGD. */
138	size_t doorbell_start_offset;
139
140	/* GPUVM address space size in bytes */
141	uint64_t gpuvm_size;
142
143	/* Minor device number of the render node */
144	int drm_render_minor;
145
146	bool enable_mes;
147};
148
149struct tile_config {
150	uint32_t *tile_config_ptr;
151	uint32_t *macro_tile_config_ptr;
152	uint32_t num_tile_configs;
153	uint32_t num_macro_tile_configs;
154
155	uint32_t gb_addr_config;
156	uint32_t num_banks;
157	uint32_t num_ranks;
158};
159
160#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
161
162/**
163 * struct kfd2kgd_calls
164 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
165 * @program_sh_mem_settings: A function that should initiate the memory
166 * properties such as main aperture memory type (cache / non cached) and
167 * secondary aperture base address, size and memory type.
168 * This function is used only for no cp scheduling mode.
169 *
170 * @set_pasid_vmid_mapping: Exposes pasid/vmid pair to the H/W for no cp
171 * scheduling mode. Only used for no cp scheduling mode.
172 *
 
 
173 * @hqd_load: Loads the mqd structure to a H/W hqd slot. used only for no cp
174 * sceduling mode.
175 *
176 * @hqd_sdma_load: Loads the SDMA mqd structure to a H/W SDMA hqd slot.
177 * used only for no HWS mode.
178 *
179 * @hqd_dump: Dumps CPC HQD registers to an array of address-value pairs.
180 * Array is allocated with kmalloc, needs to be freed with kfree by caller.
181 *
182 * @hqd_sdma_dump: Dumps SDMA HQD registers to an array of address-value pairs.
183 * Array is allocated with kmalloc, needs to be freed with kfree by caller.
184 *
185 * @hqd_is_occupies: Checks if a hqd slot is occupied.
186 *
187 * @hqd_destroy: Destructs and preempts the queue assigned to that hqd slot.
188 *
189 * @hqd_sdma_is_occupied: Checks if an SDMA hqd slot is occupied.
190 *
191 * @hqd_sdma_destroy: Destructs and preempts the SDMA queue assigned to that
192 * SDMA hqd slot.
193 *
 
 
194 * @set_scratch_backing_va: Sets VA for scratch backing memory of a VMID.
195 * Only used for no cp scheduling mode
196 *
 
 
 
 
 
 
 
 
 
 
 
 
197 * @set_vm_context_page_table_base: Program page table base for a VMID
198 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
199 * @invalidate_tlbs: Invalidate TLBs for a specific PASID
200 *
201 * @invalidate_tlbs_vmid: Invalidate TLBs for a specific VMID
202 *
203 * @read_vmid_from_vmfault_reg: On Hawaii the VMID is not set in the
204 * IH ring entry. This function allows the KFD ISR to get the VMID
205 * from the fault status register as early as possible.
206 *
207 * @get_cu_occupancy: Function pointer that returns to caller the number
208 * of wave fronts that are in flight for all of the queues of a process
209 * as identified by its pasid. It is important to note that the value
210 * returned by this function is a snapshot of current moment and cannot
211 * guarantee any minimum for the number of waves in-flight. This function
212 * is defined for devices that belong to GFX9 and later GFX families. Care
213 * must be taken in calling this function as it is not defined for devices
214 * that belong to GFX8 and below GFX families.
215 *
216 * This structure contains function pointers to services that the kgd driver
217 * provides to amdkfd driver.
218 *
219 */
220struct kfd2kgd_calls {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
221	/* Register access functions */
222	void (*program_sh_mem_settings)(struct amdgpu_device *adev, uint32_t vmid,
223			uint32_t sh_mem_config,	uint32_t sh_mem_ape1_base,
224			uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases,
225			uint32_t inst);
226
227	int (*set_pasid_vmid_mapping)(struct amdgpu_device *adev, u32 pasid,
228					unsigned int vmid, uint32_t inst);
229
230	int (*init_interrupts)(struct amdgpu_device *adev, uint32_t pipe_id,
231			uint32_t inst);
232
233	int (*hqd_load)(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id,
 
 
234			uint32_t queue_id, uint32_t __user *wptr,
235			uint32_t wptr_shift, uint32_t wptr_mask,
236			struct mm_struct *mm, uint32_t inst);
237
238	int (*hiq_mqd_load)(struct amdgpu_device *adev, void *mqd,
239			    uint32_t pipe_id, uint32_t queue_id,
240			    uint32_t doorbell_off, uint32_t inst);
241
242	int (*hqd_sdma_load)(struct amdgpu_device *adev, void *mqd,
243			     uint32_t __user *wptr, struct mm_struct *mm);
244
245	int (*hqd_dump)(struct amdgpu_device *adev,
246			uint32_t pipe_id, uint32_t queue_id,
247			uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst);
248
249	int (*hqd_sdma_dump)(struct amdgpu_device *adev,
250			     uint32_t engine_id, uint32_t queue_id,
251			     uint32_t (**dump)[2], uint32_t *n_regs);
252
253	bool (*hqd_is_occupied)(struct amdgpu_device *adev,
254				uint64_t queue_address, uint32_t pipe_id,
255				uint32_t queue_id, uint32_t inst);
256
257	int (*hqd_destroy)(struct amdgpu_device *adev, void *mqd,
258				enum kfd_preempt_type reset_type,
259				unsigned int timeout, uint32_t pipe_id,
260				uint32_t queue_id, uint32_t inst);
261
262	bool (*hqd_sdma_is_occupied)(struct amdgpu_device *adev, void *mqd);
263
264	int (*hqd_sdma_destroy)(struct amdgpu_device *adev, void *mqd,
265				unsigned int timeout);
266
267	int (*wave_control_execute)(struct amdgpu_device *adev,
 
 
 
 
 
 
268					uint32_t gfx_index_val,
269					uint32_t sq_cmd, uint32_t inst);
270	bool (*get_atc_vmid_pasid_mapping_info)(struct amdgpu_device *adev,
271					uint8_t vmid,
272					uint16_t *p_pasid);
273
274	/* No longer needed from GFXv9 onward. The scratch base address is
275	 * passed to the shader by the CP. It's the user mode driver's
276	 * responsibility.
277	 */
278	void (*set_scratch_backing_va)(struct amdgpu_device *adev,
 
 
 
 
279				uint64_t va, uint32_t vmid);
 
280
281	void (*set_vm_context_page_table_base)(struct amdgpu_device *adev,
282			uint32_t vmid, uint64_t page_table_base);
283	uint32_t (*read_vmid_from_vmfault_reg)(struct amdgpu_device *adev);
284
285	uint32_t (*enable_debug_trap)(struct amdgpu_device *adev,
286					bool restore_dbg_registers,
287					uint32_t vmid);
288	uint32_t (*disable_debug_trap)(struct amdgpu_device *adev,
289					bool keep_trap_enabled,
290					uint32_t vmid);
291	int (*validate_trap_override_request)(struct amdgpu_device *adev,
292					uint32_t trap_override,
293					uint32_t *trap_mask_supported);
294	uint32_t (*set_wave_launch_trap_override)(struct amdgpu_device *adev,
295					     uint32_t vmid,
296					     uint32_t trap_override,
297					     uint32_t trap_mask_bits,
298					     uint32_t trap_mask_request,
299					     uint32_t *trap_mask_prev,
300					     uint32_t kfd_dbg_trap_cntl_prev);
301	uint32_t (*set_wave_launch_mode)(struct amdgpu_device *adev,
302					uint8_t wave_launch_mode,
303					uint32_t vmid);
304	uint32_t (*set_address_watch)(struct amdgpu_device *adev,
305					uint64_t watch_address,
306					uint32_t watch_address_mask,
307					uint32_t watch_id,
308					uint32_t watch_mode,
309					uint32_t debug_vmid,
310					uint32_t inst);
311	uint32_t (*clear_address_watch)(struct amdgpu_device *adev,
312			uint32_t watch_id);
313	void (*get_iq_wait_times)(struct amdgpu_device *adev,
314			uint32_t *wait_times,
315			uint32_t inst);
316	void (*build_grace_period_packet_info)(struct amdgpu_device *adev,
317			uint32_t wait_times,
318			uint32_t grace_period,
319			uint32_t *reg_offset,
320			uint32_t *reg_data);
321	void (*get_cu_occupancy)(struct amdgpu_device *adev,
322				 struct kfd_cu_occupancy *cu_occupancy,
323				 int *max_waves_per_cu, uint32_t inst);
324	void (*program_trap_handler_settings)(struct amdgpu_device *adev,
325			uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr,
326			uint32_t inst);
327	uint64_t (*hqd_get_pq_addr)(struct amdgpu_device *adev,
328				    uint32_t pipe_id, uint32_t queue_id,
329				    uint32_t inst);
330	uint64_t (*hqd_reset)(struct amdgpu_device *adev,
331			      uint32_t pipe_id, uint32_t queue_id,
332			      uint32_t inst, unsigned int utimeout);
333};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
334
335#endif	/* KGD_KFD_INTERFACE_H_INCLUDED */
v4.17
  1/*
  2 * Copyright 2014 Advanced Micro Devices, Inc.
  3 *
  4 * Permission is hereby granted, free of charge, to any person obtaining a
  5 * copy of this software and associated documentation files (the "Software"),
  6 * to deal in the Software without restriction, including without limitation
  7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  8 * and/or sell copies of the Software, and to permit persons to whom the
  9 * Software is furnished to do so, subject to the following conditions:
 10 *
 11 * The above copyright notice and this permission notice shall be included in
 12 * all copies or substantial portions of the Software.
 13 *
 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 20 * OTHER DEALINGS IN THE SOFTWARE.
 21 */
 22
 23/*
 24 * This file defines the private interface between the
 25 * AMD kernel graphics drivers and the AMD KFD.
 26 */
 27
 28#ifndef KGD_KFD_INTERFACE_H_INCLUDED
 29#define KGD_KFD_INTERFACE_H_INCLUDED
 30
 31#include <linux/types.h>
 32#include <linux/bitmap.h>
 33#include <linux/dma-fence.h>
 
 
 34
 35struct pci_dev;
 36
 37#define KFD_INTERFACE_VERSION 2
 38#define KGD_MAX_QUEUES 128
 39
 40struct kfd_dev;
 41struct kgd_dev;
 42
 43struct kgd_mem;
 44
 45enum kfd_preempt_type {
 46	KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN = 0,
 47	KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
 
 48};
 49
 50struct kfd_cu_info {
 51	uint32_t num_shader_engines;
 52	uint32_t num_shader_arrays_per_engine;
 53	uint32_t num_cu_per_sh;
 54	uint32_t cu_active_number;
 55	uint32_t cu_ao_mask;
 56	uint32_t simd_per_cu;
 57	uint32_t max_waves_per_simd;
 58	uint32_t wave_front_size;
 59	uint32_t max_scratch_slots_per_cu;
 60	uint32_t lds_size;
 61	uint32_t cu_bitmap[4][4];
 62};
 63
 64/* For getting GPU local memory information from KGD */
 65struct kfd_local_mem_info {
 66	uint64_t local_mem_size_private;
 67	uint64_t local_mem_size_public;
 68	uint32_t vram_width;
 69	uint32_t mem_clk_max;
 70};
 71
 72enum kgd_memory_pool {
 73	KGD_POOL_SYSTEM_CACHEABLE = 1,
 74	KGD_POOL_SYSTEM_WRITECOMBINE = 2,
 75	KGD_POOL_FRAMEBUFFER = 3,
 76};
 77
 78enum kgd_engine_type {
 79	KGD_ENGINE_PFP = 1,
 80	KGD_ENGINE_ME,
 81	KGD_ENGINE_CE,
 82	KGD_ENGINE_MEC1,
 83	KGD_ENGINE_MEC2,
 84	KGD_ENGINE_RLC,
 85	KGD_ENGINE_SDMA1,
 86	KGD_ENGINE_SDMA2,
 87	KGD_ENGINE_MAX
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 88};
 89
 90struct kgd2kfd_shared_resources {
 91	/* Bit n == 1 means VMID n is available for KFD. */
 92	unsigned int compute_vmid_bitmap;
 93
 94	/* number of pipes per mec */
 95	uint32_t num_pipe_per_mec;
 96
 97	/* number of queues per pipe */
 98	uint32_t num_queue_per_pipe;
 99
100	/* Bit n == 1 means Queue n is available for KFD */
101	DECLARE_BITMAP(queue_bitmap, KGD_MAX_QUEUES);
 
 
 
 
 
 
 
 
 
 
 
 
102
103	/* Base address of doorbell aperture. */
104	phys_addr_t doorbell_physical_address;
105
106	/* Size in bytes of doorbell aperture. */
107	size_t doorbell_aperture_size;
108
109	/* Number of bytes at start of aperture reserved for KGD. */
110	size_t doorbell_start_offset;
111
112	/* GPUVM address space size in bytes */
113	uint64_t gpuvm_size;
114
115	/* Minor device number of the render node */
116	int drm_render_minor;
 
 
117};
118
119struct tile_config {
120	uint32_t *tile_config_ptr;
121	uint32_t *macro_tile_config_ptr;
122	uint32_t num_tile_configs;
123	uint32_t num_macro_tile_configs;
124
125	uint32_t gb_addr_config;
126	uint32_t num_banks;
127	uint32_t num_ranks;
128};
129
130
131/*
132 * Allocation flag domains
133 * NOTE: This must match the corresponding definitions in kfd_ioctl.h.
134 */
135#define ALLOC_MEM_FLAGS_VRAM		(1 << 0)
136#define ALLOC_MEM_FLAGS_GTT		(1 << 1)
137#define ALLOC_MEM_FLAGS_USERPTR		(1 << 2) /* TODO */
138#define ALLOC_MEM_FLAGS_DOORBELL	(1 << 3) /* TODO */
139
140/*
141 * Allocation flags attributes/access options.
142 * NOTE: This must match the corresponding definitions in kfd_ioctl.h.
143 */
144#define ALLOC_MEM_FLAGS_WRITABLE	(1 << 31)
145#define ALLOC_MEM_FLAGS_EXECUTABLE	(1 << 30)
146#define ALLOC_MEM_FLAGS_PUBLIC		(1 << 29)
147#define ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28) /* TODO */
148#define ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
149#define ALLOC_MEM_FLAGS_COHERENT	(1 << 26) /* For GFXv9 or later */
150
151/**
152 * struct kfd2kgd_calls
153 *
154 * @init_gtt_mem_allocation: Allocate a buffer on the gart aperture.
155 * The buffer can be used for mqds, hpds, kernel queue, fence and runlists
156 *
157 * @free_gtt_mem: Frees a buffer that was allocated on the gart aperture
158 *
159 * @get_local_mem_info: Retrieves information about GPU local memory
160 *
161 * @get_gpu_clock_counter: Retrieves GPU clock counter
162 *
163 * @get_max_engine_clock_in_mhz: Retrieves maximum GPU clock in MHz
164 *
165 * @alloc_pasid: Allocate a PASID
166 * @free_pasid: Free a PASID
167 *
168 * @program_sh_mem_settings: A function that should initiate the memory
169 * properties such as main aperture memory type (cache / non cached) and
170 * secondary aperture base address, size and memory type.
171 * This function is used only for no cp scheduling mode.
172 *
173 * @set_pasid_vmid_mapping: Exposes pasid/vmid pair to the H/W for no cp
174 * scheduling mode. Only used for no cp scheduling mode.
175 *
176 * @init_pipeline: Initialized the compute pipelines.
177 *
178 * @hqd_load: Loads the mqd structure to a H/W hqd slot. used only for no cp
179 * sceduling mode.
180 *
181 * @hqd_sdma_load: Loads the SDMA mqd structure to a H/W SDMA hqd slot.
182 * used only for no HWS mode.
183 *
184 * @hqd_dump: Dumps CPC HQD registers to an array of address-value pairs.
185 * Array is allocated with kmalloc, needs to be freed with kfree by caller.
186 *
187 * @hqd_sdma_dump: Dumps SDMA HQD registers to an array of address-value pairs.
188 * Array is allocated with kmalloc, needs to be freed with kfree by caller.
189 *
190 * @hqd_is_occupies: Checks if a hqd slot is occupied.
191 *
192 * @hqd_destroy: Destructs and preempts the queue assigned to that hqd slot.
193 *
194 * @hqd_sdma_is_occupied: Checks if an SDMA hqd slot is occupied.
195 *
196 * @hqd_sdma_destroy: Destructs and preempts the SDMA queue assigned to that
197 * SDMA hqd slot.
198 *
199 * @get_fw_version: Returns FW versions from the header
200 *
201 * @set_scratch_backing_va: Sets VA for scratch backing memory of a VMID.
202 * Only used for no cp scheduling mode
203 *
204 * @get_tile_config: Returns GPU-specific tiling mode information
205 *
206 * @get_cu_info: Retrieves activated cu info
207 *
208 * @get_vram_usage: Returns current VRAM usage
209 *
210 * @create_process_vm: Create a VM address space for a given process and GPU
211 *
212 * @destroy_process_vm: Destroy a VM
213 *
214 * @get_process_page_dir: Get physical address of a VM page directory
215 *
216 * @set_vm_context_page_table_base: Program page table base for a VMID
217 *
218 * @alloc_memory_of_gpu: Allocate GPUVM memory
219 *
220 * @free_memory_of_gpu: Free GPUVM memory
221 *
222 * @map_memory_to_gpu: Map GPUVM memory into a specific VM address
223 * space. Allocates and updates page tables and page directories as
224 * needed. This function may return before all page table updates have
225 * completed. This allows multiple map operations (on multiple GPUs)
226 * to happen concurrently. Use sync_memory to synchronize with all
227 * pending updates.
228 *
229 * @unmap_memor_to_gpu: Unmap GPUVM memory from a specific VM address space
230 *
231 * @sync_memory: Wait for pending page table updates to complete
232 *
233 * @map_gtt_bo_to_kernel: Map a GTT BO for kernel access
234 * Pins the BO, maps it to kernel address space. Such BOs are never evicted.
235 * The kernel virtual address remains valid until the BO is freed.
236 *
237 * @restore_process_bos: Restore all BOs that belong to the
238 * process. This is intended for restoring memory mappings after a TTM
239 * eviction.
240 *
241 * @invalidate_tlbs: Invalidate TLBs for a specific PASID
242 *
243 * @invalidate_tlbs_vmid: Invalidate TLBs for a specific VMID
244 *
245 * @submit_ib: Submits an IB to the engine specified by inserting the
246 * IB to the corresponding ring (ring type). The IB is executed with the
247 * specified VMID in a user mode context.
 
 
 
 
 
 
 
 
 
248 *
249 * This structure contains function pointers to services that the kgd driver
250 * provides to amdkfd driver.
251 *
252 */
253struct kfd2kgd_calls {
254	int (*init_gtt_mem_allocation)(struct kgd_dev *kgd, size_t size,
255					void **mem_obj, uint64_t *gpu_addr,
256					void **cpu_ptr);
257
258	void (*free_gtt_mem)(struct kgd_dev *kgd, void *mem_obj);
259
260	void (*get_local_mem_info)(struct kgd_dev *kgd,
261			struct kfd_local_mem_info *mem_info);
262	uint64_t (*get_gpu_clock_counter)(struct kgd_dev *kgd);
263
264	uint32_t (*get_max_engine_clock_in_mhz)(struct kgd_dev *kgd);
265
266	int (*alloc_pasid)(unsigned int bits);
267	void (*free_pasid)(unsigned int pasid);
268
269	/* Register access functions */
270	void (*program_sh_mem_settings)(struct kgd_dev *kgd, uint32_t vmid,
271			uint32_t sh_mem_config,	uint32_t sh_mem_ape1_base,
272			uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases);
 
273
274	int (*set_pasid_vmid_mapping)(struct kgd_dev *kgd, unsigned int pasid,
275					unsigned int vmid);
276
277	int (*init_pipeline)(struct kgd_dev *kgd, uint32_t pipe_id,
278				uint32_t hpd_size, uint64_t hpd_gpu_addr);
279
280	int (*init_interrupts)(struct kgd_dev *kgd, uint32_t pipe_id);
281
282	int (*hqd_load)(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
283			uint32_t queue_id, uint32_t __user *wptr,
284			uint32_t wptr_shift, uint32_t wptr_mask,
285			struct mm_struct *mm);
 
 
 
 
286
287	int (*hqd_sdma_load)(struct kgd_dev *kgd, void *mqd,
288			     uint32_t __user *wptr, struct mm_struct *mm);
289
290	int (*hqd_dump)(struct kgd_dev *kgd,
291			uint32_t pipe_id, uint32_t queue_id,
292			uint32_t (**dump)[2], uint32_t *n_regs);
293
294	int (*hqd_sdma_dump)(struct kgd_dev *kgd,
295			     uint32_t engine_id, uint32_t queue_id,
296			     uint32_t (**dump)[2], uint32_t *n_regs);
297
298	bool (*hqd_is_occupied)(struct kgd_dev *kgd, uint64_t queue_address,
299				uint32_t pipe_id, uint32_t queue_id);
 
300
301	int (*hqd_destroy)(struct kgd_dev *kgd, void *mqd, uint32_t reset_type,
 
302				unsigned int timeout, uint32_t pipe_id,
303				uint32_t queue_id);
304
305	bool (*hqd_sdma_is_occupied)(struct kgd_dev *kgd, void *mqd);
306
307	int (*hqd_sdma_destroy)(struct kgd_dev *kgd, void *mqd,
308				unsigned int timeout);
309
310	int (*address_watch_disable)(struct kgd_dev *kgd);
311	int (*address_watch_execute)(struct kgd_dev *kgd,
312					unsigned int watch_point_id,
313					uint32_t cntl_val,
314					uint32_t addr_hi,
315					uint32_t addr_lo);
316	int (*wave_control_execute)(struct kgd_dev *kgd,
317					uint32_t gfx_index_val,
318					uint32_t sq_cmd);
319	uint32_t (*address_watch_get_offset)(struct kgd_dev *kgd,
320					unsigned int watch_point_id,
321					unsigned int reg_offset);
322	bool (*get_atc_vmid_pasid_mapping_valid)(
323					struct kgd_dev *kgd,
324					uint8_t vmid);
325	uint16_t (*get_atc_vmid_pasid_mapping_pasid)(
326					struct kgd_dev *kgd,
327					uint8_t vmid);
328
329	uint16_t (*get_fw_version)(struct kgd_dev *kgd,
330				enum kgd_engine_type type);
331	void (*set_scratch_backing_va)(struct kgd_dev *kgd,
332				uint64_t va, uint32_t vmid);
333	int (*get_tile_config)(struct kgd_dev *kgd, struct tile_config *config);
334
335	void (*get_cu_info)(struct kgd_dev *kgd,
336			struct kfd_cu_info *cu_info);
337	uint64_t (*get_vram_usage)(struct kgd_dev *kgd);
338
339	int (*create_process_vm)(struct kgd_dev *kgd, void **vm,
340			void **process_info, struct dma_fence **ef);
341	int (*acquire_process_vm)(struct kgd_dev *kgd, struct file *filp,
342			void **vm, void **process_info, struct dma_fence **ef);
343	void (*destroy_process_vm)(struct kgd_dev *kgd, void *vm);
344	uint32_t (*get_process_page_dir)(void *vm);
345	void (*set_vm_context_page_table_base)(struct kgd_dev *kgd,
346			uint32_t vmid, uint32_t page_table_base);
347	int (*alloc_memory_of_gpu)(struct kgd_dev *kgd, uint64_t va,
348			uint64_t size, void *vm,
349			struct kgd_mem **mem, uint64_t *offset,
350			uint32_t flags);
351	int (*free_memory_of_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem);
352	int (*map_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
353			void *vm);
354	int (*unmap_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
355			void *vm);
356	int (*sync_memory)(struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
357	int (*map_gtt_bo_to_kernel)(struct kgd_dev *kgd, struct kgd_mem *mem,
358			void **kptr, uint64_t *size);
359	int (*restore_process_bos)(void *process_info, struct dma_fence **ef);
360
361	int (*invalidate_tlbs)(struct kgd_dev *kgd, uint16_t pasid);
362	int (*invalidate_tlbs_vmid)(struct kgd_dev *kgd, uint16_t vmid);
363
364	int (*submit_ib)(struct kgd_dev *kgd, enum kgd_engine_type engine,
365			uint32_t vmid, uint64_t gpu_addr,
366			uint32_t *ib_cmd, uint32_t ib_len);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
367};
368
369/**
370 * struct kgd2kfd_calls
371 *
372 * @exit: Notifies amdkfd that kgd module is unloaded
373 *
374 * @probe: Notifies amdkfd about a probe done on a device in the kgd driver.
375 *
376 * @device_init: Initialize the newly probed device (if it is a device that
377 * amdkfd supports)
378 *
379 * @device_exit: Notifies amdkfd about a removal of a kgd device
380 *
381 * @suspend: Notifies amdkfd about a suspend action done to a kgd device
382 *
383 * @resume: Notifies amdkfd about a resume action done to a kgd device
384 *
385 * @schedule_evict_and_restore_process: Schedules work queue that will prepare
386 * for safe eviction of KFD BOs that belong to the specified process.
387 *
388 * This structure contains function callback pointers so the kgd driver
389 * will notify to the amdkfd about certain status changes.
390 *
391 */
392struct kgd2kfd_calls {
393	void (*exit)(void);
394	struct kfd_dev* (*probe)(struct kgd_dev *kgd, struct pci_dev *pdev,
395		const struct kfd2kgd_calls *f2g);
396	bool (*device_init)(struct kfd_dev *kfd,
397			const struct kgd2kfd_shared_resources *gpu_resources);
398	void (*device_exit)(struct kfd_dev *kfd);
399	void (*interrupt)(struct kfd_dev *kfd, const void *ih_ring_entry);
400	void (*suspend)(struct kfd_dev *kfd);
401	int (*resume)(struct kfd_dev *kfd);
402	int (*schedule_evict_and_restore_process)(struct mm_struct *mm,
403			struct dma_fence *fence);
404};
405
406int kgd2kfd_init(unsigned interface_version,
407		const struct kgd2kfd_calls **g2f);
408
409#endif	/* KGD_KFD_INTERFACE_H_INCLUDED */