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_vm_fault_info {
51 	uint64_t	page_addr;
52 	uint32_t	vmid;
53 	uint32_t	mc_id;
54 	uint32_t	status;
55 	bool		prot_valid;
56 	bool		prot_read;
57 	bool		prot_write;
58 	bool		prot_exec;
59 };
60 
61 struct kfd_cu_info {
62 	uint32_t num_shader_engines;
63 	uint32_t num_shader_arrays_per_engine;
64 	uint32_t num_cu_per_sh;
65 	uint32_t cu_active_number;
66 	uint32_t cu_ao_mask;
67 	uint32_t simd_per_cu;
68 	uint32_t max_waves_per_simd;
69 	uint32_t wave_front_size;
70 	uint32_t max_scratch_slots_per_cu;
71 	uint32_t lds_size;
72 	uint32_t cu_bitmap[4][4];
73 };
74 
75 /* For getting GPU local memory information from KGD */
76 struct kfd_local_mem_info {
77 	uint64_t local_mem_size_private;
78 	uint64_t local_mem_size_public;
79 	uint32_t vram_width;
80 	uint32_t mem_clk_max;
81 };
82 
83 enum kgd_memory_pool {
84 	KGD_POOL_SYSTEM_CACHEABLE = 1,
85 	KGD_POOL_SYSTEM_WRITECOMBINE = 2,
86 	KGD_POOL_FRAMEBUFFER = 3,
87 };
88 
89 enum kgd_engine_type {
90 	KGD_ENGINE_PFP = 1,
91 	KGD_ENGINE_ME,
92 	KGD_ENGINE_CE,
93 	KGD_ENGINE_MEC1,
94 	KGD_ENGINE_MEC2,
95 	KGD_ENGINE_RLC,
96 	KGD_ENGINE_SDMA1,
97 	KGD_ENGINE_SDMA2,
98 	KGD_ENGINE_MAX
99 };
100 
101 struct kgd2kfd_shared_resources {
102 	/* Bit n == 1 means VMID n is available for KFD. */
103 	unsigned int compute_vmid_bitmap;
104 
105 	/* number of pipes per mec */
106 	uint32_t num_pipe_per_mec;
107 
108 	/* number of queues per pipe */
109 	uint32_t num_queue_per_pipe;
110 
111 	/* Bit n == 1 means Queue n is available for KFD */
112 	DECLARE_BITMAP(queue_bitmap, KGD_MAX_QUEUES);
113 
114 	/* Doorbell assignments (SOC15 and later chips only). Only
115 	 * specific doorbells are routed to each SDMA engine. Others
116 	 * are routed to IH and VCN. They are not usable by the CP.
117 	 *
118 	 * Any doorbell number D that satisfies the following condition
119 	 * is reserved: (D & reserved_doorbell_mask) == reserved_doorbell_val
120 	 *
121 	 * KFD currently uses 1024 (= 0x3ff) doorbells per process. If
122 	 * doorbells 0x0f0-0x0f7 and 0x2f-0x2f7 are reserved, that means
123 	 * mask would be set to 0x1f8 and val set to 0x0f0.
124 	 */
125 	unsigned int sdma_doorbell[2][2];
126 	unsigned int reserved_doorbell_mask;
127 	unsigned int reserved_doorbell_val;
128 
129 	/* Base address of doorbell aperture. */
130 	phys_addr_t doorbell_physical_address;
131 
132 	/* Size in bytes of doorbell aperture. */
133 	size_t doorbell_aperture_size;
134 
135 	/* Number of bytes at start of aperture reserved for KGD. */
136 	size_t doorbell_start_offset;
137 
138 	/* GPUVM address space size in bytes */
139 	uint64_t gpuvm_size;
140 
141 	/* Minor device number of the render node */
142 	int drm_render_minor;
143 };
144 
145 struct tile_config {
146 	uint32_t *tile_config_ptr;
147 	uint32_t *macro_tile_config_ptr;
148 	uint32_t num_tile_configs;
149 	uint32_t num_macro_tile_configs;
150 
151 	uint32_t gb_addr_config;
152 	uint32_t num_banks;
153 	uint32_t num_ranks;
154 };
155 
156 
157 /*
158  * Allocation flag domains
159  * NOTE: This must match the corresponding definitions in kfd_ioctl.h.
160  */
161 #define ALLOC_MEM_FLAGS_VRAM		(1 << 0)
162 #define ALLOC_MEM_FLAGS_GTT		(1 << 1)
163 #define ALLOC_MEM_FLAGS_USERPTR		(1 << 2) /* TODO */
164 #define ALLOC_MEM_FLAGS_DOORBELL	(1 << 3) /* TODO */
165 
166 /*
167  * Allocation flags attributes/access options.
168  * NOTE: This must match the corresponding definitions in kfd_ioctl.h.
169  */
170 #define ALLOC_MEM_FLAGS_WRITABLE	(1 << 31)
171 #define ALLOC_MEM_FLAGS_EXECUTABLE	(1 << 30)
172 #define ALLOC_MEM_FLAGS_PUBLIC		(1 << 29)
173 #define ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28) /* TODO */
174 #define ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
175 #define ALLOC_MEM_FLAGS_COHERENT	(1 << 26) /* For GFXv9 or later */
176 
177 /**
178  * struct kfd2kgd_calls
179  *
180  * @init_gtt_mem_allocation: Allocate a buffer on the gart aperture.
181  * The buffer can be used for mqds, hpds, kernel queue, fence and runlists
182  *
183  * @free_gtt_mem: Frees a buffer that was allocated on the gart aperture
184  *
185  * @get_local_mem_info: Retrieves information about GPU local memory
186  *
187  * @get_gpu_clock_counter: Retrieves GPU clock counter
188  *
189  * @get_max_engine_clock_in_mhz: Retrieves maximum GPU clock in MHz
190  *
191  * @alloc_pasid: Allocate a PASID
192  * @free_pasid: Free a PASID
193  *
194  * @program_sh_mem_settings: A function that should initiate the memory
195  * properties such as main aperture memory type (cache / non cached) and
196  * secondary aperture base address, size and memory type.
197  * This function is used only for no cp scheduling mode.
198  *
199  * @set_pasid_vmid_mapping: Exposes pasid/vmid pair to the H/W for no cp
200  * scheduling mode. Only used for no cp scheduling mode.
201  *
202  * @hqd_load: Loads the mqd structure to a H/W hqd slot. used only for no cp
203  * sceduling mode.
204  *
205  * @hqd_sdma_load: Loads the SDMA mqd structure to a H/W SDMA hqd slot.
206  * used only for no HWS mode.
207  *
208  * @hqd_dump: Dumps CPC HQD registers to an array of address-value pairs.
209  * Array is allocated with kmalloc, needs to be freed with kfree by caller.
210  *
211  * @hqd_sdma_dump: Dumps SDMA HQD registers to an array of address-value pairs.
212  * Array is allocated with kmalloc, needs to be freed with kfree by caller.
213  *
214  * @hqd_is_occupies: Checks if a hqd slot is occupied.
215  *
216  * @hqd_destroy: Destructs and preempts the queue assigned to that hqd slot.
217  *
218  * @hqd_sdma_is_occupied: Checks if an SDMA hqd slot is occupied.
219  *
220  * @hqd_sdma_destroy: Destructs and preempts the SDMA queue assigned to that
221  * SDMA hqd slot.
222  *
223  * @get_fw_version: Returns FW versions from the header
224  *
225  * @set_scratch_backing_va: Sets VA for scratch backing memory of a VMID.
226  * Only used for no cp scheduling mode
227  *
228  * @get_tile_config: Returns GPU-specific tiling mode information
229  *
230  * @get_cu_info: Retrieves activated cu info
231  *
232  * @get_vram_usage: Returns current VRAM usage
233  *
234  * @create_process_vm: Create a VM address space for a given process and GPU
235  *
236  * @destroy_process_vm: Destroy a VM
237  *
238  * @get_process_page_dir: Get physical address of a VM page directory
239  *
240  * @set_vm_context_page_table_base: Program page table base for a VMID
241  *
242  * @alloc_memory_of_gpu: Allocate GPUVM memory
243  *
244  * @free_memory_of_gpu: Free GPUVM memory
245  *
246  * @map_memory_to_gpu: Map GPUVM memory into a specific VM address
247  * space. Allocates and updates page tables and page directories as
248  * needed. This function may return before all page table updates have
249  * completed. This allows multiple map operations (on multiple GPUs)
250  * to happen concurrently. Use sync_memory to synchronize with all
251  * pending updates.
252  *
253  * @unmap_memor_to_gpu: Unmap GPUVM memory from a specific VM address space
254  *
255  * @sync_memory: Wait for pending page table updates to complete
256  *
257  * @map_gtt_bo_to_kernel: Map a GTT BO for kernel access
258  * Pins the BO, maps it to kernel address space. Such BOs are never evicted.
259  * The kernel virtual address remains valid until the BO is freed.
260  *
261  * @restore_process_bos: Restore all BOs that belong to the
262  * process. This is intended for restoring memory mappings after a TTM
263  * eviction.
264  *
265  * @invalidate_tlbs: Invalidate TLBs for a specific PASID
266  *
267  * @invalidate_tlbs_vmid: Invalidate TLBs for a specific VMID
268  *
269  * @submit_ib: Submits an IB to the engine specified by inserting the
270  * IB to the corresponding ring (ring type). The IB is executed with the
271  * specified VMID in a user mode context.
272  *
273  * @get_vm_fault_info: Return information about a recent VM fault on
274  * GFXv7 and v8. If multiple VM faults occurred since the last call of
275  * this function, it will return information about the first of those
276  * faults. On GFXv9 VM fault information is fully contained in the IH
277  * packet and this function is not needed.
278  *
279  * @read_vmid_from_vmfault_reg: On Hawaii the VMID is not set in the
280  * IH ring entry. This function allows the KFD ISR to get the VMID
281  * from the fault status register as early as possible.
282  *
283  * @gpu_recover: let kgd reset gpu after kfd detect CPC hang
284  *
285  * @set_compute_idle: Indicates that compute is idle on a device. This
286  * can be used to change power profiles depending on compute activity.
287  *
288  * This structure contains function pointers to services that the kgd driver
289  * provides to amdkfd driver.
290  *
291  */
292 struct kfd2kgd_calls {
293 	int (*init_gtt_mem_allocation)(struct kgd_dev *kgd, size_t size,
294 					void **mem_obj, uint64_t *gpu_addr,
295 					void **cpu_ptr, bool mqd_gfx9);
296 
297 	void (*free_gtt_mem)(struct kgd_dev *kgd, void *mem_obj);
298 
299 	void (*get_local_mem_info)(struct kgd_dev *kgd,
300 			struct kfd_local_mem_info *mem_info);
301 	uint64_t (*get_gpu_clock_counter)(struct kgd_dev *kgd);
302 
303 	uint32_t (*get_max_engine_clock_in_mhz)(struct kgd_dev *kgd);
304 
305 	int (*alloc_pasid)(unsigned int bits);
306 	void (*free_pasid)(unsigned int pasid);
307 
308 	/* Register access functions */
309 	void (*program_sh_mem_settings)(struct kgd_dev *kgd, uint32_t vmid,
310 			uint32_t sh_mem_config,	uint32_t sh_mem_ape1_base,
311 			uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases);
312 
313 	int (*set_pasid_vmid_mapping)(struct kgd_dev *kgd, unsigned int pasid,
314 					unsigned int vmid);
315 
316 	int (*init_interrupts)(struct kgd_dev *kgd, uint32_t pipe_id);
317 
318 	int (*hqd_load)(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
319 			uint32_t queue_id, uint32_t __user *wptr,
320 			uint32_t wptr_shift, uint32_t wptr_mask,
321 			struct mm_struct *mm);
322 
323 	int (*hqd_sdma_load)(struct kgd_dev *kgd, void *mqd,
324 			     uint32_t __user *wptr, struct mm_struct *mm);
325 
326 	int (*hqd_dump)(struct kgd_dev *kgd,
327 			uint32_t pipe_id, uint32_t queue_id,
328 			uint32_t (**dump)[2], uint32_t *n_regs);
329 
330 	int (*hqd_sdma_dump)(struct kgd_dev *kgd,
331 			     uint32_t engine_id, uint32_t queue_id,
332 			     uint32_t (**dump)[2], uint32_t *n_regs);
333 
334 	bool (*hqd_is_occupied)(struct kgd_dev *kgd, uint64_t queue_address,
335 				uint32_t pipe_id, uint32_t queue_id);
336 
337 	int (*hqd_destroy)(struct kgd_dev *kgd, void *mqd, uint32_t reset_type,
338 				unsigned int timeout, uint32_t pipe_id,
339 				uint32_t queue_id);
340 
341 	bool (*hqd_sdma_is_occupied)(struct kgd_dev *kgd, void *mqd);
342 
343 	int (*hqd_sdma_destroy)(struct kgd_dev *kgd, void *mqd,
344 				unsigned int timeout);
345 
346 	int (*address_watch_disable)(struct kgd_dev *kgd);
347 	int (*address_watch_execute)(struct kgd_dev *kgd,
348 					unsigned int watch_point_id,
349 					uint32_t cntl_val,
350 					uint32_t addr_hi,
351 					uint32_t addr_lo);
352 	int (*wave_control_execute)(struct kgd_dev *kgd,
353 					uint32_t gfx_index_val,
354 					uint32_t sq_cmd);
355 	uint32_t (*address_watch_get_offset)(struct kgd_dev *kgd,
356 					unsigned int watch_point_id,
357 					unsigned int reg_offset);
358 	bool (*get_atc_vmid_pasid_mapping_valid)(
359 					struct kgd_dev *kgd,
360 					uint8_t vmid);
361 	uint16_t (*get_atc_vmid_pasid_mapping_pasid)(
362 					struct kgd_dev *kgd,
363 					uint8_t vmid);
364 
365 	uint16_t (*get_fw_version)(struct kgd_dev *kgd,
366 				enum kgd_engine_type type);
367 	void (*set_scratch_backing_va)(struct kgd_dev *kgd,
368 				uint64_t va, uint32_t vmid);
369 	int (*get_tile_config)(struct kgd_dev *kgd, struct tile_config *config);
370 
371 	void (*get_cu_info)(struct kgd_dev *kgd,
372 			struct kfd_cu_info *cu_info);
373 	uint64_t (*get_vram_usage)(struct kgd_dev *kgd);
374 
375 	int (*create_process_vm)(struct kgd_dev *kgd, void **vm,
376 			void **process_info, struct dma_fence **ef);
377 	int (*acquire_process_vm)(struct kgd_dev *kgd, struct file *filp,
378 			void **vm, void **process_info, struct dma_fence **ef);
379 	void (*destroy_process_vm)(struct kgd_dev *kgd, void *vm);
380 	uint32_t (*get_process_page_dir)(void *vm);
381 	void (*set_vm_context_page_table_base)(struct kgd_dev *kgd,
382 			uint32_t vmid, uint32_t page_table_base);
383 	int (*alloc_memory_of_gpu)(struct kgd_dev *kgd, uint64_t va,
384 			uint64_t size, void *vm,
385 			struct kgd_mem **mem, uint64_t *offset,
386 			uint32_t flags);
387 	int (*free_memory_of_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem);
388 	int (*map_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
389 			void *vm);
390 	int (*unmap_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
391 			void *vm);
392 	int (*sync_memory)(struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
393 	int (*map_gtt_bo_to_kernel)(struct kgd_dev *kgd, struct kgd_mem *mem,
394 			void **kptr, uint64_t *size);
395 	int (*restore_process_bos)(void *process_info, struct dma_fence **ef);
396 
397 	int (*invalidate_tlbs)(struct kgd_dev *kgd, uint16_t pasid);
398 	int (*invalidate_tlbs_vmid)(struct kgd_dev *kgd, uint16_t vmid);
399 
400 	int (*submit_ib)(struct kgd_dev *kgd, enum kgd_engine_type engine,
401 			uint32_t vmid, uint64_t gpu_addr,
402 			uint32_t *ib_cmd, uint32_t ib_len);
403 
404 	int (*get_vm_fault_info)(struct kgd_dev *kgd,
405 			struct kfd_vm_fault_info *info);
406 	uint32_t (*read_vmid_from_vmfault_reg)(struct kgd_dev *kgd);
407 
408 	void (*gpu_recover)(struct kgd_dev *kgd);
409 
410 	void (*set_compute_idle)(struct kgd_dev *kgd, bool idle);
411 };
412 
413 /**
414  * struct kgd2kfd_calls
415  *
416  * @exit: Notifies amdkfd that kgd module is unloaded
417  *
418  * @probe: Notifies amdkfd about a probe done on a device in the kgd driver.
419  *
420  * @device_init: Initialize the newly probed device (if it is a device that
421  * amdkfd supports)
422  *
423  * @device_exit: Notifies amdkfd about a removal of a kgd device
424  *
425  * @suspend: Notifies amdkfd about a suspend action done to a kgd device
426  *
427  * @resume: Notifies amdkfd about a resume action done to a kgd device
428  *
429  * @quiesce_mm: Quiesce all user queue access to specified MM address space
430  *
431  * @resume_mm: Resume user queue access to specified MM address space
432  *
433  * @schedule_evict_and_restore_process: Schedules work queue that will prepare
434  * for safe eviction of KFD BOs that belong to the specified process.
435  *
436  * @pre_reset: Notifies amdkfd that amdgpu about to reset the gpu
437  *
438  * @post_reset: Notify amdkfd that amgpu successfully reseted the gpu
439  *
440  * This structure contains function callback pointers so the kgd driver
441  * will notify to the amdkfd about certain status changes.
442  *
443  */
444 struct kgd2kfd_calls {
445 	void (*exit)(void);
446 	struct kfd_dev* (*probe)(struct kgd_dev *kgd, struct pci_dev *pdev,
447 		const struct kfd2kgd_calls *f2g);
448 	bool (*device_init)(struct kfd_dev *kfd,
449 			const struct kgd2kfd_shared_resources *gpu_resources);
450 	void (*device_exit)(struct kfd_dev *kfd);
451 	void (*interrupt)(struct kfd_dev *kfd, const void *ih_ring_entry);
452 	void (*suspend)(struct kfd_dev *kfd);
453 	int (*resume)(struct kfd_dev *kfd);
454 	int (*quiesce_mm)(struct mm_struct *mm);
455 	int (*resume_mm)(struct mm_struct *mm);
456 	int (*schedule_evict_and_restore_process)(struct mm_struct *mm,
457 			struct dma_fence *fence);
458 	int  (*pre_reset)(struct kfd_dev *kfd);
459 	int  (*post_reset)(struct kfd_dev *kfd);
460 };
461 
462 int kgd2kfd_init(unsigned interface_version,
463 		const struct kgd2kfd_calls **g2f);
464 
465 #endif	/* KGD_KFD_INTERFACE_H_INCLUDED */
466