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 
35 struct pci_dev;
36 
37 #define KFD_INTERFACE_VERSION 2
38 #define KGD_MAX_QUEUES 128
39 
40 struct kfd_dev;
41 struct kgd_dev;
42 
43 struct kgd_mem;
44 
45 enum kfd_preempt_type {
46 	KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN = 0,
47 	KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
48 };
49 
50 struct 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 */
65 struct 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 
72 enum kgd_memory_pool {
73 	KGD_POOL_SYSTEM_CACHEABLE = 1,
74 	KGD_POOL_SYSTEM_WRITECOMBINE = 2,
75 	KGD_POOL_FRAMEBUFFER = 3,
76 };
77 
78 enum 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 
90 struct 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 
119 struct 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  */
134 #define ALLOC_MEM_FLAGS_VRAM		(1 << 0)
135 #define ALLOC_MEM_FLAGS_GTT		(1 << 1)
136 #define ALLOC_MEM_FLAGS_USERPTR		(1 << 2) /* TODO */
137 #define ALLOC_MEM_FLAGS_DOORBELL	(1 << 3) /* TODO */
138 
139 /*
140  * Allocation flags attributes/access options.
141  */
142 #define ALLOC_MEM_FLAGS_WRITABLE	(1 << 31)
143 #define ALLOC_MEM_FLAGS_EXECUTABLE	(1 << 30)
144 #define ALLOC_MEM_FLAGS_PUBLIC		(1 << 29)
145 #define ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28) /* TODO */
146 #define ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
147 #define ALLOC_MEM_FLAGS_COHERENT	(1 << 26) /* For GFXv9 or later */
148 
149 /**
150  * struct kfd2kgd_calls
151  *
152  * @init_gtt_mem_allocation: Allocate a buffer on the gart aperture.
153  * The buffer can be used for mqds, hpds, kernel queue, fence and runlists
154  *
155  * @free_gtt_mem: Frees a buffer that was allocated on the gart aperture
156  *
157  * @get_local_mem_info: Retrieves information about GPU local memory
158  *
159  * @get_gpu_clock_counter: Retrieves GPU clock counter
160  *
161  * @get_max_engine_clock_in_mhz: Retrieves maximum GPU clock in MHz
162  *
163  * @alloc_pasid: Allocate a PASID
164  * @free_pasid: Free a PASID
165  *
166  * @program_sh_mem_settings: A function that should initiate the memory
167  * properties such as main aperture memory type (cache / non cached) and
168  * secondary aperture base address, size and memory type.
169  * This function is used only for no cp scheduling mode.
170  *
171  * @set_pasid_vmid_mapping: Exposes pasid/vmid pair to the H/W for no cp
172  * scheduling mode. Only used for no cp scheduling mode.
173  *
174  * @init_pipeline: Initialized the compute pipelines.
175  *
176  * @hqd_load: Loads the mqd structure to a H/W hqd slot. used only for no cp
177  * sceduling mode.
178  *
179  * @hqd_sdma_load: Loads the SDMA mqd structure to a H/W SDMA hqd slot.
180  * used only for no HWS mode.
181  *
182  * @hqd_dump: Dumps CPC 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_sdma_dump: Dumps SDMA HQD registers to an array of address-value pairs.
186  * Array is allocated with kmalloc, needs to be freed with kfree by caller.
187  *
188  * @hqd_is_occupies: Checks if a hqd slot is occupied.
189  *
190  * @hqd_destroy: Destructs and preempts the queue assigned to that hqd slot.
191  *
192  * @hqd_sdma_is_occupied: Checks if an SDMA hqd slot is occupied.
193  *
194  * @hqd_sdma_destroy: Destructs and preempts the SDMA queue assigned to that
195  * SDMA hqd slot.
196  *
197  * @get_fw_version: Returns FW versions from the header
198  *
199  * @set_scratch_backing_va: Sets VA for scratch backing memory of a VMID.
200  * Only used for no cp scheduling mode
201  *
202  * @get_tile_config: Returns GPU-specific tiling mode information
203  *
204  * @get_cu_info: Retrieves activated cu info
205  *
206  * @get_vram_usage: Returns current VRAM usage
207  *
208  * @create_process_vm: Create a VM address space for a given process and GPU
209  *
210  * @destroy_process_vm: Destroy a VM
211  *
212  * @get_process_page_dir: Get physical address of a VM page directory
213  *
214  * @set_vm_context_page_table_base: Program page table base for a VMID
215  *
216  * @alloc_memory_of_gpu: Allocate GPUVM memory
217  *
218  * @free_memory_of_gpu: Free GPUVM memory
219  *
220  * @map_memory_to_gpu: Map GPUVM memory into a specific VM address
221  * space. Allocates and updates page tables and page directories as
222  * needed. This function may return before all page table updates have
223  * completed. This allows multiple map operations (on multiple GPUs)
224  * to happen concurrently. Use sync_memory to synchronize with all
225  * pending updates.
226  *
227  * @unmap_memor_to_gpu: Unmap GPUVM memory from a specific VM address space
228  *
229  * @sync_memory: Wait for pending page table updates to complete
230  *
231  * @map_gtt_bo_to_kernel: Map a GTT BO for kernel access
232  * Pins the BO, maps it to kernel address space. Such BOs are never evicted.
233  * The kernel virtual address remains valid until the BO is freed.
234  *
235  * @restore_process_bos: Restore all BOs that belong to the
236  * process. This is intended for restoring memory mappings after a TTM
237  * eviction.
238  *
239  * @invalidate_tlbs: Invalidate TLBs for a specific PASID
240  *
241  * @invalidate_tlbs_vmid: Invalidate TLBs for a specific VMID
242  *
243  * @submit_ib: Submits an IB to the engine specified by inserting the
244  * IB to the corresponding ring (ring type). The IB is executed with the
245  * specified VMID in a user mode context.
246  *
247  * This structure contains function pointers to services that the kgd driver
248  * provides to amdkfd driver.
249  *
250  */
251 struct kfd2kgd_calls {
252 	int (*init_gtt_mem_allocation)(struct kgd_dev *kgd, size_t size,
253 					void **mem_obj, uint64_t *gpu_addr,
254 					void **cpu_ptr);
255 
256 	void (*free_gtt_mem)(struct kgd_dev *kgd, void *mem_obj);
257 
258 	void (*get_local_mem_info)(struct kgd_dev *kgd,
259 			struct kfd_local_mem_info *mem_info);
260 	uint64_t (*get_gpu_clock_counter)(struct kgd_dev *kgd);
261 
262 	uint32_t (*get_max_engine_clock_in_mhz)(struct kgd_dev *kgd);
263 
264 	int (*alloc_pasid)(unsigned int bits);
265 	void (*free_pasid)(unsigned int pasid);
266 
267 	/* Register access functions */
268 	void (*program_sh_mem_settings)(struct kgd_dev *kgd, uint32_t vmid,
269 			uint32_t sh_mem_config,	uint32_t sh_mem_ape1_base,
270 			uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases);
271 
272 	int (*set_pasid_vmid_mapping)(struct kgd_dev *kgd, unsigned int pasid,
273 					unsigned int vmid);
274 
275 	int (*init_pipeline)(struct kgd_dev *kgd, uint32_t pipe_id,
276 				uint32_t hpd_size, uint64_t hpd_gpu_addr);
277 
278 	int (*init_interrupts)(struct kgd_dev *kgd, uint32_t pipe_id);
279 
280 	int (*hqd_load)(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
281 			uint32_t queue_id, uint32_t __user *wptr,
282 			uint32_t wptr_shift, uint32_t wptr_mask,
283 			struct mm_struct *mm);
284 
285 	int (*hqd_sdma_load)(struct kgd_dev *kgd, void *mqd,
286 			     uint32_t __user *wptr, struct mm_struct *mm);
287 
288 	int (*hqd_dump)(struct kgd_dev *kgd,
289 			uint32_t pipe_id, uint32_t queue_id,
290 			uint32_t (**dump)[2], uint32_t *n_regs);
291 
292 	int (*hqd_sdma_dump)(struct kgd_dev *kgd,
293 			     uint32_t engine_id, uint32_t queue_id,
294 			     uint32_t (**dump)[2], uint32_t *n_regs);
295 
296 	bool (*hqd_is_occupied)(struct kgd_dev *kgd, uint64_t queue_address,
297 				uint32_t pipe_id, uint32_t queue_id);
298 
299 	int (*hqd_destroy)(struct kgd_dev *kgd, void *mqd, uint32_t reset_type,
300 				unsigned int timeout, uint32_t pipe_id,
301 				uint32_t queue_id);
302 
303 	bool (*hqd_sdma_is_occupied)(struct kgd_dev *kgd, void *mqd);
304 
305 	int (*hqd_sdma_destroy)(struct kgd_dev *kgd, void *mqd,
306 				unsigned int timeout);
307 
308 	int (*address_watch_disable)(struct kgd_dev *kgd);
309 	int (*address_watch_execute)(struct kgd_dev *kgd,
310 					unsigned int watch_point_id,
311 					uint32_t cntl_val,
312 					uint32_t addr_hi,
313 					uint32_t addr_lo);
314 	int (*wave_control_execute)(struct kgd_dev *kgd,
315 					uint32_t gfx_index_val,
316 					uint32_t sq_cmd);
317 	uint32_t (*address_watch_get_offset)(struct kgd_dev *kgd,
318 					unsigned int watch_point_id,
319 					unsigned int reg_offset);
320 	bool (*get_atc_vmid_pasid_mapping_valid)(
321 					struct kgd_dev *kgd,
322 					uint8_t vmid);
323 	uint16_t (*get_atc_vmid_pasid_mapping_pasid)(
324 					struct kgd_dev *kgd,
325 					uint8_t vmid);
326 
327 	uint16_t (*get_fw_version)(struct kgd_dev *kgd,
328 				enum kgd_engine_type type);
329 	void (*set_scratch_backing_va)(struct kgd_dev *kgd,
330 				uint64_t va, uint32_t vmid);
331 	int (*get_tile_config)(struct kgd_dev *kgd, struct tile_config *config);
332 
333 	void (*get_cu_info)(struct kgd_dev *kgd,
334 			struct kfd_cu_info *cu_info);
335 	uint64_t (*get_vram_usage)(struct kgd_dev *kgd);
336 
337 	int (*create_process_vm)(struct kgd_dev *kgd, void **vm,
338 			void **process_info, struct dma_fence **ef);
339 	void (*destroy_process_vm)(struct kgd_dev *kgd, void *vm);
340 	uint32_t (*get_process_page_dir)(void *vm);
341 	void (*set_vm_context_page_table_base)(struct kgd_dev *kgd,
342 			uint32_t vmid, uint32_t page_table_base);
343 	int (*alloc_memory_of_gpu)(struct kgd_dev *kgd, uint64_t va,
344 			uint64_t size, void *vm,
345 			struct kgd_mem **mem, uint64_t *offset,
346 			uint32_t flags);
347 	int (*free_memory_of_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem);
348 	int (*map_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
349 			void *vm);
350 	int (*unmap_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
351 			void *vm);
352 	int (*sync_memory)(struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
353 	int (*map_gtt_bo_to_kernel)(struct kgd_dev *kgd, struct kgd_mem *mem,
354 			void **kptr, uint64_t *size);
355 	int (*restore_process_bos)(void *process_info, struct dma_fence **ef);
356 
357 	int (*invalidate_tlbs)(struct kgd_dev *kgd, uint16_t pasid);
358 	int (*invalidate_tlbs_vmid)(struct kgd_dev *kgd, uint16_t vmid);
359 
360 	int (*submit_ib)(struct kgd_dev *kgd, enum kgd_engine_type engine,
361 			uint32_t vmid, uint64_t gpu_addr,
362 			uint32_t *ib_cmd, uint32_t ib_len);
363 };
364 
365 /**
366  * struct kgd2kfd_calls
367  *
368  * @exit: Notifies amdkfd that kgd module is unloaded
369  *
370  * @probe: Notifies amdkfd about a probe done on a device in the kgd driver.
371  *
372  * @device_init: Initialize the newly probed device (if it is a device that
373  * amdkfd supports)
374  *
375  * @device_exit: Notifies amdkfd about a removal of a kgd device
376  *
377  * @suspend: Notifies amdkfd about a suspend action done to a kgd device
378  *
379  * @resume: Notifies amdkfd about a resume action done to a kgd device
380  *
381  * @schedule_evict_and_restore_process: Schedules work queue that will prepare
382  * for safe eviction of KFD BOs that belong to the specified process.
383  *
384  * This structure contains function callback pointers so the kgd driver
385  * will notify to the amdkfd about certain status changes.
386  *
387  */
388 struct kgd2kfd_calls {
389 	void (*exit)(void);
390 	struct kfd_dev* (*probe)(struct kgd_dev *kgd, struct pci_dev *pdev,
391 		const struct kfd2kgd_calls *f2g);
392 	bool (*device_init)(struct kfd_dev *kfd,
393 			const struct kgd2kfd_shared_resources *gpu_resources);
394 	void (*device_exit)(struct kfd_dev *kfd);
395 	void (*interrupt)(struct kfd_dev *kfd, const void *ih_ring_entry);
396 	void (*suspend)(struct kfd_dev *kfd);
397 	int (*resume)(struct kfd_dev *kfd);
398 	int (*schedule_evict_and_restore_process)(struct mm_struct *mm,
399 			struct dma_fence *fence);
400 };
401 
402 int kgd2kfd_init(unsigned interface_version,
403 		const struct kgd2kfd_calls **g2f);
404 
405 #endif	/* KGD_KFD_INTERFACE_H_INCLUDED */
406