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 KGD_MAX_QUEUES 128
38 
39 struct kfd_dev;
40 struct kgd_dev;
41 
42 struct kgd_mem;
43 
44 enum kfd_preempt_type {
45 	KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN = 0,
46 	KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
47 };
48 
49 struct 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 struct kfd_cu_info {
61 	uint32_t num_shader_engines;
62 	uint32_t num_shader_arrays_per_engine;
63 	uint32_t num_cu_per_sh;
64 	uint32_t cu_active_number;
65 	uint32_t cu_ao_mask;
66 	uint32_t simd_per_cu;
67 	uint32_t max_waves_per_simd;
68 	uint32_t wave_front_size;
69 	uint32_t max_scratch_slots_per_cu;
70 	uint32_t lds_size;
71 	uint32_t cu_bitmap[4][4];
72 };
73 
74 /* For getting GPU local memory information from KGD */
75 struct kfd_local_mem_info {
76 	uint64_t local_mem_size_private;
77 	uint64_t local_mem_size_public;
78 	uint32_t vram_width;
79 	uint32_t mem_clk_max;
80 };
81 
82 enum kgd_memory_pool {
83 	KGD_POOL_SYSTEM_CACHEABLE = 1,
84 	KGD_POOL_SYSTEM_WRITECOMBINE = 2,
85 	KGD_POOL_FRAMEBUFFER = 3,
86 };
87 
88 /**
89  * enum kfd_sched_policy
90  *
91  * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
92  * scheduling. In this scheduling mode we're using the firmware code to
93  * schedule the user mode queues and kernel queues such as HIQ and DIQ.
94  * the HIQ queue is used as a special queue that dispatches the configuration
95  * to the cp and the user mode queues list that are currently running.
96  * the DIQ queue is a debugging queue that dispatches debugging commands to the
97  * firmware.
98  * in this scheduling mode user mode queues over subscription feature is
99  * enabled.
100  *
101  * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
102  * subscription feature disabled.
103  *
104  * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
105  * set the command processor registers and sets the queues "manually". This
106  * mode is used *ONLY* for debugging proposes.
107  *
108  */
109 enum kfd_sched_policy {
110 	KFD_SCHED_POLICY_HWS = 0,
111 	KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
112 	KFD_SCHED_POLICY_NO_HWS
113 };
114 
115 struct kgd2kfd_shared_resources {
116 	/* Bit n == 1 means VMID n is available for KFD. */
117 	unsigned int compute_vmid_bitmap;
118 
119 	/* number of pipes per mec */
120 	uint32_t num_pipe_per_mec;
121 
122 	/* number of queues per pipe */
123 	uint32_t num_queue_per_pipe;
124 
125 	/* Bit n == 1 means Queue n is available for KFD */
126 	DECLARE_BITMAP(queue_bitmap, KGD_MAX_QUEUES);
127 
128 	/* SDMA doorbell assignments (SOC15 and later chips only). Only
129 	 * specific doorbells are routed to each SDMA engine. Others
130 	 * are routed to IH and VCN. They are not usable by the CP.
131 	 */
132 	uint32_t *sdma_doorbell_idx;
133 
134 	/* From SOC15 onward, the doorbell index range not usable for CP
135 	 * queues.
136 	 */
137 	uint32_t non_cp_doorbells_start;
138 	uint32_t non_cp_doorbells_end;
139 
140 	/* Base address of doorbell aperture. */
141 	phys_addr_t doorbell_physical_address;
142 
143 	/* Size in bytes of doorbell aperture. */
144 	size_t doorbell_aperture_size;
145 
146 	/* Number of bytes at start of aperture reserved for KGD. */
147 	size_t doorbell_start_offset;
148 
149 	/* GPUVM address space size in bytes */
150 	uint64_t gpuvm_size;
151 
152 	/* Minor device number of the render node */
153 	int drm_render_minor;
154 };
155 
156 struct tile_config {
157 	uint32_t *tile_config_ptr;
158 	uint32_t *macro_tile_config_ptr;
159 	uint32_t num_tile_configs;
160 	uint32_t num_macro_tile_configs;
161 
162 	uint32_t gb_addr_config;
163 	uint32_t num_banks;
164 	uint32_t num_ranks;
165 };
166 
167 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096
168 
169 /*
170  * Allocation flag domains
171  * NOTE: This must match the corresponding definitions in kfd_ioctl.h.
172  */
173 #define ALLOC_MEM_FLAGS_VRAM		(1 << 0)
174 #define ALLOC_MEM_FLAGS_GTT		(1 << 1)
175 #define ALLOC_MEM_FLAGS_USERPTR		(1 << 2)
176 #define ALLOC_MEM_FLAGS_DOORBELL	(1 << 3)
177 #define ALLOC_MEM_FLAGS_MMIO_REMAP	(1 << 4)
178 
179 /*
180  * Allocation flags attributes/access options.
181  * NOTE: This must match the corresponding definitions in kfd_ioctl.h.
182  */
183 #define ALLOC_MEM_FLAGS_WRITABLE	(1 << 31)
184 #define ALLOC_MEM_FLAGS_EXECUTABLE	(1 << 30)
185 #define ALLOC_MEM_FLAGS_PUBLIC		(1 << 29)
186 #define ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28) /* TODO */
187 #define ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
188 #define ALLOC_MEM_FLAGS_COHERENT	(1 << 26) /* For GFXv9 or later */
189 
190 /**
191  * struct kfd2kgd_calls
192  *
193  * @program_sh_mem_settings: A function that should initiate the memory
194  * properties such as main aperture memory type (cache / non cached) and
195  * secondary aperture base address, size and memory type.
196  * This function is used only for no cp scheduling mode.
197  *
198  * @set_pasid_vmid_mapping: Exposes pasid/vmid pair to the H/W for no cp
199  * scheduling mode. Only used for no cp scheduling mode.
200  *
201  * @hqd_load: Loads the mqd structure to a H/W hqd slot. used only for no cp
202  * sceduling mode.
203  *
204  * @hqd_sdma_load: Loads the SDMA mqd structure to a H/W SDMA hqd slot.
205  * used only for no HWS mode.
206  *
207  * @hqd_dump: Dumps CPC HQD registers to an array of address-value pairs.
208  * Array is allocated with kmalloc, needs to be freed with kfree by caller.
209  *
210  * @hqd_sdma_dump: Dumps SDMA HQD registers to an array of address-value pairs.
211  * Array is allocated with kmalloc, needs to be freed with kfree by caller.
212  *
213  * @hqd_is_occupies: Checks if a hqd slot is occupied.
214  *
215  * @hqd_destroy: Destructs and preempts the queue assigned to that hqd slot.
216  *
217  * @hqd_sdma_is_occupied: Checks if an SDMA hqd slot is occupied.
218  *
219  * @hqd_sdma_destroy: Destructs and preempts the SDMA queue assigned to that
220  * SDMA hqd slot.
221  *
222  * @set_scratch_backing_va: Sets VA for scratch backing memory of a VMID.
223  * Only used for no cp scheduling mode
224  *
225  * @get_tile_config: Returns GPU-specific tiling mode information
226  *
227  * @set_vm_context_page_table_base: Program page table base for a VMID
228  *
229  * @invalidate_tlbs: Invalidate TLBs for a specific PASID
230  *
231  * @invalidate_tlbs_vmid: Invalidate TLBs for a specific VMID
232  *
233  * @read_vmid_from_vmfault_reg: On Hawaii the VMID is not set in the
234  * IH ring entry. This function allows the KFD ISR to get the VMID
235  * from the fault status register as early as possible.
236  *
237  * @get_hive_id: Returns hive id of current  device,  0 if xgmi is not enabled
238  *
239  * This structure contains function pointers to services that the kgd driver
240  * provides to amdkfd driver.
241  *
242  */
243 struct kfd2kgd_calls {
244 	/* Register access functions */
245 	void (*program_sh_mem_settings)(struct kgd_dev *kgd, uint32_t vmid,
246 			uint32_t sh_mem_config,	uint32_t sh_mem_ape1_base,
247 			uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases);
248 
249 	int (*set_pasid_vmid_mapping)(struct kgd_dev *kgd, unsigned int pasid,
250 					unsigned int vmid);
251 
252 	int (*init_interrupts)(struct kgd_dev *kgd, uint32_t pipe_id);
253 
254 	int (*hqd_load)(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
255 			uint32_t queue_id, uint32_t __user *wptr,
256 			uint32_t wptr_shift, uint32_t wptr_mask,
257 			struct mm_struct *mm);
258 
259 	int (*hqd_sdma_load)(struct kgd_dev *kgd, void *mqd,
260 			     uint32_t __user *wptr, struct mm_struct *mm);
261 
262 	int (*hqd_dump)(struct kgd_dev *kgd,
263 			uint32_t pipe_id, uint32_t queue_id,
264 			uint32_t (**dump)[2], uint32_t *n_regs);
265 
266 	int (*hqd_sdma_dump)(struct kgd_dev *kgd,
267 			     uint32_t engine_id, uint32_t queue_id,
268 			     uint32_t (**dump)[2], uint32_t *n_regs);
269 
270 	bool (*hqd_is_occupied)(struct kgd_dev *kgd, uint64_t queue_address,
271 				uint32_t pipe_id, uint32_t queue_id);
272 
273 	int (*hqd_destroy)(struct kgd_dev *kgd, void *mqd, uint32_t reset_type,
274 				unsigned int timeout, uint32_t pipe_id,
275 				uint32_t queue_id);
276 
277 	bool (*hqd_sdma_is_occupied)(struct kgd_dev *kgd, void *mqd);
278 
279 	int (*hqd_sdma_destroy)(struct kgd_dev *kgd, void *mqd,
280 				unsigned int timeout);
281 
282 	int (*address_watch_disable)(struct kgd_dev *kgd);
283 	int (*address_watch_execute)(struct kgd_dev *kgd,
284 					unsigned int watch_point_id,
285 					uint32_t cntl_val,
286 					uint32_t addr_hi,
287 					uint32_t addr_lo);
288 	int (*wave_control_execute)(struct kgd_dev *kgd,
289 					uint32_t gfx_index_val,
290 					uint32_t sq_cmd);
291 	uint32_t (*address_watch_get_offset)(struct kgd_dev *kgd,
292 					unsigned int watch_point_id,
293 					unsigned int reg_offset);
294 	bool (*get_atc_vmid_pasid_mapping_valid)(
295 					struct kgd_dev *kgd,
296 					uint8_t vmid);
297 	uint16_t (*get_atc_vmid_pasid_mapping_pasid)(
298 					struct kgd_dev *kgd,
299 					uint8_t vmid);
300 
301 	void (*set_scratch_backing_va)(struct kgd_dev *kgd,
302 				uint64_t va, uint32_t vmid);
303 	int (*get_tile_config)(struct kgd_dev *kgd, struct tile_config *config);
304 
305 	void (*set_vm_context_page_table_base)(struct kgd_dev *kgd,
306 			uint32_t vmid, uint64_t page_table_base);
307 	int (*invalidate_tlbs)(struct kgd_dev *kgd, uint16_t pasid);
308 	int (*invalidate_tlbs_vmid)(struct kgd_dev *kgd, uint16_t vmid);
309 	uint32_t (*read_vmid_from_vmfault_reg)(struct kgd_dev *kgd);
310 	uint64_t (*get_hive_id)(struct kgd_dev *kgd);
311 
312 };
313 
314 #endif	/* KGD_KFD_INTERFACE_H_INCLUDED */
315