1 /*
2  * Copyright 2023 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 #include "amdgpu.h"
24 #include "amdgpu_amdkfd.h"
25 #include "gc/gc_12_0_0_offset.h"
26 #include "gc/gc_12_0_0_sh_mask.h"
27 #include "soc24.h"
28 #include <uapi/linux/kfd_ioctl.h>
29 
lock_srbm(struct amdgpu_device * adev,uint32_t mec,uint32_t pipe,uint32_t queue,uint32_t vmid)30 static void lock_srbm(struct amdgpu_device *adev, uint32_t mec, uint32_t pipe,
31 			uint32_t queue, uint32_t vmid)
32 {
33 	mutex_lock(&adev->srbm_mutex);
34 	soc24_grbm_select(adev, mec, pipe, queue, vmid);
35 }
36 
unlock_srbm(struct amdgpu_device * adev)37 static void unlock_srbm(struct amdgpu_device *adev)
38 {
39 	soc24_grbm_select(adev, 0, 0, 0, 0);
40 	mutex_unlock(&adev->srbm_mutex);
41 }
42 
acquire_queue(struct amdgpu_device * adev,uint32_t pipe_id,uint32_t queue_id)43 static void acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id,
44 				uint32_t queue_id)
45 {
46 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
47 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
48 
49 	lock_srbm(adev, mec, pipe, queue_id, 0);
50 }
51 
release_queue(struct amdgpu_device * adev)52 static void release_queue(struct amdgpu_device *adev)
53 {
54 	unlock_srbm(adev);
55 }
56 
init_interrupts_v12(struct amdgpu_device * adev,uint32_t pipe_id,uint32_t inst)57 static int init_interrupts_v12(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t inst)
58 {
59 	uint32_t mec;
60 	uint32_t pipe;
61 
62 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
63 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
64 
65 	lock_srbm(adev, mec, pipe, 0, 0);
66 
67 	WREG32_SOC15(GC, 0, regCPC_INT_CNTL,
68 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
69 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
70 
71 	unlock_srbm(adev);
72 
73 	return 0;
74 }
75 
get_sdma_rlc_reg_offset(struct amdgpu_device * adev,unsigned int engine_id,unsigned int queue_id)76 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
77 				unsigned int engine_id,
78 				unsigned int queue_id)
79 {
80 	uint32_t sdma_engine_reg_base = 0;
81 	uint32_t sdma_rlc_reg_offset;
82 
83 	switch (engine_id) {
84 	case 0:
85 		sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA0, 0,
86 				regSDMA0_QUEUE0_RB_CNTL) - regSDMA0_QUEUE0_RB_CNTL;
87 		break;
88 	case 1:
89 		sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA1, 0,
90 				regSDMA1_QUEUE0_RB_CNTL) - regSDMA0_QUEUE0_RB_CNTL;
91 		break;
92 	default:
93 		BUG();
94 	}
95 
96 	sdma_rlc_reg_offset = sdma_engine_reg_base
97 		+ queue_id * (regSDMA0_QUEUE1_RB_CNTL - regSDMA0_QUEUE0_RB_CNTL);
98 
99 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
100 			queue_id, sdma_rlc_reg_offset);
101 
102 	return sdma_rlc_reg_offset;
103 }
104 
hqd_dump_v12(struct amdgpu_device * adev,uint32_t pipe_id,uint32_t queue_id,uint32_t (** dump)[2],uint32_t * n_regs,uint32_t inst)105 static int hqd_dump_v12(struct amdgpu_device *adev,
106 			uint32_t pipe_id, uint32_t queue_id,
107 			uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst)
108 {
109 	uint32_t i = 0, reg;
110 #define HQD_N_REGS 56
111 #define DUMP_REG(addr) do {				\
112 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
113 			break;				\
114 		(*dump)[i][0] = (addr) << 2;		\
115 		(*dump)[i++][1] = RREG32(addr);		\
116 	} while (0)
117 
118 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
119 	if (*dump == NULL)
120 		return -ENOMEM;
121 
122 	acquire_queue(adev, pipe_id, queue_id);
123 
124 	for (reg = SOC15_REG_OFFSET(GC, 0, regCP_MQD_BASE_ADDR);
125 	     reg <= SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_HI); reg++)
126 		DUMP_REG(reg);
127 
128 	release_queue(adev);
129 
130 	WARN_ON_ONCE(i != HQD_N_REGS);
131 	*n_regs = i;
132 
133 	return 0;
134 }
135 
hqd_sdma_dump_v12(struct amdgpu_device * adev,uint32_t engine_id,uint32_t queue_id,uint32_t (** dump)[2],uint32_t * n_regs)136 static int hqd_sdma_dump_v12(struct amdgpu_device *adev,
137 			     uint32_t engine_id, uint32_t queue_id,
138 			     uint32_t (**dump)[2], uint32_t *n_regs)
139 {
140 	uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
141 			engine_id, queue_id);
142 	uint32_t i = 0, reg;
143 
144 	const uint32_t first_reg = regSDMA0_QUEUE0_RB_CNTL;
145 	const uint32_t last_reg = regSDMA0_QUEUE0_CONTEXT_STATUS;
146 #undef HQD_N_REGS
147 #define HQD_N_REGS (last_reg - first_reg + 1)
148 
149 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
150 	if (*dump == NULL)
151 		return -ENOMEM;
152 
153 	for (reg = first_reg;
154 	     reg <= last_reg; reg++)
155 		DUMP_REG(sdma_rlc_reg_offset + reg);
156 
157 	WARN_ON_ONCE(i != HQD_N_REGS);
158 	*n_regs = i;
159 
160 	return 0;
161 }
162 
wave_control_execute_v12(struct amdgpu_device * adev,uint32_t gfx_index_val,uint32_t sq_cmd,uint32_t inst)163 static int wave_control_execute_v12(struct amdgpu_device *adev,
164 					uint32_t gfx_index_val,
165 					uint32_t sq_cmd, uint32_t inst)
166 {
167 	uint32_t data = 0;
168 
169 	mutex_lock(&adev->grbm_idx_mutex);
170 
171 	WREG32(SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_INDEX), gfx_index_val);
172 	WREG32(SOC15_REG_OFFSET(GC, 0, regSQ_CMD), sq_cmd);
173 
174 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
175 		INSTANCE_BROADCAST_WRITES, 1);
176 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
177 		SA_BROADCAST_WRITES, 1);
178 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
179 		SE_BROADCAST_WRITES, 1);
180 
181 	WREG32(SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_INDEX), data);
182 	mutex_unlock(&adev->grbm_idx_mutex);
183 
184 	return 0;
185 }
186 
187 /* returns TRAP_EN, EXCP_EN and EXCP_REPLACE. */
kgd_gfx_v12_enable_debug_trap(struct amdgpu_device * adev,bool restore_dbg_registers,uint32_t vmid)188 static uint32_t kgd_gfx_v12_enable_debug_trap(struct amdgpu_device *adev,
189 					    bool restore_dbg_registers,
190 					    uint32_t vmid)
191 {
192 	uint32_t data = 0;
193 
194 	data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1);
195 	data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_EN, 0);
196 	data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_REPLACE, 0);
197 
198 	return data;
199 }
200 
201 /* returns TRAP_EN, EXCP_EN and EXCP_REPLACE. */
kgd_gfx_v12_disable_debug_trap(struct amdgpu_device * adev,bool keep_trap_enabled,uint32_t vmid)202 static uint32_t kgd_gfx_v12_disable_debug_trap(struct amdgpu_device *adev,
203 						bool keep_trap_enabled,
204 						uint32_t vmid)
205 {
206 	uint32_t data = 0;
207 
208 	data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1);
209 	data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_EN, 0);
210 	data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_REPLACE, 0);
211 
212 	return data;
213 }
214 
kgd_gfx_v12_validate_trap_override_request(struct amdgpu_device * adev,uint32_t trap_override,uint32_t * trap_mask_supported)215 static int kgd_gfx_v12_validate_trap_override_request(struct amdgpu_device *adev,
216 							uint32_t trap_override,
217 							uint32_t *trap_mask_supported)
218 {
219 	*trap_mask_supported &= KFD_DBG_TRAP_MASK_FP_INVALID |
220 				KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL |
221 				KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO |
222 				KFD_DBG_TRAP_MASK_FP_OVERFLOW |
223 				KFD_DBG_TRAP_MASK_FP_UNDERFLOW |
224 				KFD_DBG_TRAP_MASK_FP_INEXACT |
225 				KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO |
226 				KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH |
227 				KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION |
228 				KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START |
229 				KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END;
230 
231 
232 	if (trap_override != KFD_DBG_TRAP_OVERRIDE_OR &&
233 			trap_override != KFD_DBG_TRAP_OVERRIDE_REPLACE)
234 		return -EPERM;
235 
236 	return 0;
237 }
238 
trap_mask_map_sw_to_hw(uint32_t mask)239 static uint32_t trap_mask_map_sw_to_hw(uint32_t mask)
240 {
241 	uint32_t trap_on_start = (mask & KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START) ? 1 : 0;
242 	uint32_t trap_on_end = (mask & KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END) ? 1 : 0;
243 	uint32_t excp_en = mask & (KFD_DBG_TRAP_MASK_FP_INVALID |
244 			KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL |
245 			KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO |
246 			KFD_DBG_TRAP_MASK_FP_OVERFLOW |
247 			KFD_DBG_TRAP_MASK_FP_UNDERFLOW |
248 			KFD_DBG_TRAP_MASK_FP_INEXACT |
249 			KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO |
250 			KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH |
251 			KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION);
252 	uint32_t ret;
253 
254 	ret = REG_SET_FIELD(0, SPI_GDBG_PER_VMID_CNTL, EXCP_EN, excp_en);
255 	ret = REG_SET_FIELD(ret, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_START, trap_on_start);
256 	ret = REG_SET_FIELD(ret, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_END, trap_on_end);
257 
258 	return ret;
259 }
260 
trap_mask_map_hw_to_sw(uint32_t mask)261 static uint32_t trap_mask_map_hw_to_sw(uint32_t mask)
262 {
263 	uint32_t ret = REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, EXCP_EN);
264 
265 	if (REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_START))
266 		ret |= KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START;
267 
268 	if (REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_END))
269 		ret |= KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END;
270 
271 	return ret;
272 }
273 
274 /* returns TRAP_EN, EXCP_EN and EXCP_REPLACE. */
kgd_gfx_v12_set_wave_launch_trap_override(struct amdgpu_device * adev,uint32_t vmid,uint32_t trap_override,uint32_t trap_mask_bits,uint32_t trap_mask_request,uint32_t * trap_mask_prev,uint32_t kfd_dbg_trap_cntl_prev)275 static uint32_t kgd_gfx_v12_set_wave_launch_trap_override(struct amdgpu_device *adev,
276 					uint32_t vmid,
277 					uint32_t trap_override,
278 					uint32_t trap_mask_bits,
279 					uint32_t trap_mask_request,
280 					uint32_t *trap_mask_prev,
281 					uint32_t kfd_dbg_trap_cntl_prev)
282 
283 {
284 	uint32_t data = 0;
285 
286 	*trap_mask_prev = trap_mask_map_hw_to_sw(kfd_dbg_trap_cntl_prev);
287 
288 	data = (trap_mask_bits & trap_mask_request) | (*trap_mask_prev & ~trap_mask_request);
289 	data = trap_mask_map_sw_to_hw(data);
290 
291 	data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1);
292 	data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_REPLACE, trap_override);
293 
294 	return data;
295 }
296 
297 /* returns STALL_VMID or LAUNCH_MODE. */
kgd_gfx_v12_set_wave_launch_mode(struct amdgpu_device * adev,uint8_t wave_launch_mode,uint32_t vmid)298 static uint32_t kgd_gfx_v12_set_wave_launch_mode(struct amdgpu_device *adev,
299 					uint8_t wave_launch_mode,
300 					uint32_t vmid)
301 {
302 	uint32_t data = 0;
303 	bool is_stall_mode = wave_launch_mode == 4;
304 
305 	if (is_stall_mode)
306 		data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, STALL_VMID,
307 									1);
308 	else
309 		data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, LAUNCH_MODE,
310 							wave_launch_mode);
311 
312 	return data;
313 }
314 
315 #define TCP_WATCH_STRIDE (regTCP_WATCH1_ADDR_H - regTCP_WATCH0_ADDR_H)
kgd_gfx_v12_set_address_watch(struct amdgpu_device * adev,uint64_t watch_address,uint32_t watch_address_mask,uint32_t watch_id,uint32_t watch_mode,uint32_t debug_vmid,uint32_t inst)316 static uint32_t kgd_gfx_v12_set_address_watch(struct amdgpu_device *adev,
317 					uint64_t watch_address,
318 					uint32_t watch_address_mask,
319 					uint32_t watch_id,
320 					uint32_t watch_mode,
321 					uint32_t debug_vmid,
322 					uint32_t inst)
323 {
324 	uint32_t watch_address_high;
325 	uint32_t watch_address_low;
326 	uint32_t watch_address_cntl;
327 
328 	watch_address_cntl = 0;
329 	watch_address_low = lower_32_bits(watch_address);
330 	watch_address_high = upper_32_bits(watch_address) & 0xffff;
331 
332 	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
333 			TCP_WATCH0_CNTL,
334 			MODE,
335 			watch_mode);
336 
337 	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
338 			TCP_WATCH0_CNTL,
339 			MASK,
340 			watch_address_mask >> 7);
341 
342 	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
343 			TCP_WATCH0_CNTL,
344 			VALID,
345 			1);
346 
347 	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_H) +
348 			(watch_id * TCP_WATCH_STRIDE)),
349 			watch_address_high);
350 
351 	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_L) +
352 			(watch_id * TCP_WATCH_STRIDE)),
353 			watch_address_low);
354 
355 	return watch_address_cntl;
356 }
357 
kgd_gfx_v12_clear_address_watch(struct amdgpu_device * adev,uint32_t watch_id)358 static uint32_t kgd_gfx_v12_clear_address_watch(struct amdgpu_device *adev,
359 					uint32_t watch_id)
360 {
361 	return 0;
362 }
363 
364 const struct kfd2kgd_calls gfx_v12_kfd2kgd = {
365 	.init_interrupts = init_interrupts_v12,
366 	.hqd_dump = hqd_dump_v12,
367 	.hqd_sdma_dump = hqd_sdma_dump_v12,
368 	.wave_control_execute = wave_control_execute_v12,
369 	.get_atc_vmid_pasid_mapping_info = NULL,
370 	.enable_debug_trap = kgd_gfx_v12_enable_debug_trap,
371 	.disable_debug_trap = kgd_gfx_v12_disable_debug_trap,
372 	.validate_trap_override_request = kgd_gfx_v12_validate_trap_override_request,
373 	.set_wave_launch_trap_override = kgd_gfx_v12_set_wave_launch_trap_override,
374 	.set_wave_launch_mode = kgd_gfx_v12_set_wave_launch_mode,
375 	.set_address_watch = kgd_gfx_v12_set_address_watch,
376 	.clear_address_watch = kgd_gfx_v12_clear_address_watch,
377 };
378