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 #include "amdgpu.h"
24 #include "amdgpu_amdkfd.h"
25 #include "gfx_v8_0.h"
26 #include "gca/gfx_8_0_sh_mask.h"
27 #include "gca/gfx_8_0_d.h"
28 #include "gca/gfx_8_0_enum.h"
29 #include "oss/oss_3_0_sh_mask.h"
30 #include "oss/oss_3_0_d.h"
31 #include "gmc/gmc_8_1_sh_mask.h"
32 #include "gmc/gmc_8_1_d.h"
33 #include "vi_structs.h"
34 #include "vid.h"
35 
36 enum hqd_dequeue_request_type {
37 	NO_ACTION = 0,
38 	DRAIN_PIPE,
39 	RESET_WAVES
40 };
41 
42 static void lock_srbm(struct amdgpu_device *adev, uint32_t mec, uint32_t pipe,
43 			uint32_t queue, uint32_t vmid)
44 {
45 	uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
46 
47 	mutex_lock(&adev->srbm_mutex);
48 	WREG32(mmSRBM_GFX_CNTL, value);
49 }
50 
51 static void unlock_srbm(struct amdgpu_device *adev)
52 {
53 	WREG32(mmSRBM_GFX_CNTL, 0);
54 	mutex_unlock(&adev->srbm_mutex);
55 }
56 
57 static void acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id,
58 				uint32_t queue_id)
59 {
60 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
61 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
62 
63 	lock_srbm(adev, mec, pipe, queue_id, 0);
64 }
65 
66 static void release_queue(struct amdgpu_device *adev)
67 {
68 	unlock_srbm(adev);
69 }
70 
71 static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmid,
72 					uint32_t sh_mem_config,
73 					uint32_t sh_mem_ape1_base,
74 					uint32_t sh_mem_ape1_limit,
75 					uint32_t sh_mem_bases)
76 {
77 	lock_srbm(adev, 0, 0, 0, vmid);
78 
79 	WREG32(mmSH_MEM_CONFIG, sh_mem_config);
80 	WREG32(mmSH_MEM_APE1_BASE, sh_mem_ape1_base);
81 	WREG32(mmSH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
82 	WREG32(mmSH_MEM_BASES, sh_mem_bases);
83 
84 	unlock_srbm(adev);
85 }
86 
87 static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid,
88 					unsigned int vmid)
89 {
90 	/*
91 	 * We have to assume that there is no outstanding mapping.
92 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
93 	 * a mapping is in progress or because a mapping finished
94 	 * and the SW cleared it.
95 	 * So the protocol is to always wait & clear.
96 	 */
97 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
98 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
99 
100 	WREG32(mmATC_VMID0_PASID_MAPPING + vmid, pasid_mapping);
101 
102 	while (!(RREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))
103 		cpu_relax();
104 	WREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
105 
106 	/* Mapping vmid to pasid also for IH block */
107 	WREG32(mmIH_VMID_0_LUT + vmid, pasid_mapping);
108 
109 	return 0;
110 }
111 
112 static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id)
113 {
114 	uint32_t mec;
115 	uint32_t pipe;
116 
117 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
118 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
119 
120 	lock_srbm(adev, mec, pipe, 0, 0);
121 
122 	WREG32(mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
123 			CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
124 
125 	unlock_srbm(adev);
126 
127 	return 0;
128 }
129 
130 static inline uint32_t get_sdma_rlc_reg_offset(struct vi_sdma_mqd *m)
131 {
132 	uint32_t retval;
133 
134 	retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
135 		m->sdma_queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
136 
137 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n",
138 			m->sdma_engine_id, m->sdma_queue_id, retval);
139 
140 	return retval;
141 }
142 
143 static inline struct vi_mqd *get_mqd(void *mqd)
144 {
145 	return (struct vi_mqd *)mqd;
146 }
147 
148 static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd)
149 {
150 	return (struct vi_sdma_mqd *)mqd;
151 }
152 
153 static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd,
154 			uint32_t pipe_id, uint32_t queue_id,
155 			uint32_t __user *wptr, uint32_t wptr_shift,
156 			uint32_t wptr_mask, struct mm_struct *mm)
157 {
158 	struct vi_mqd *m;
159 	uint32_t *mqd_hqd;
160 	uint32_t reg, wptr_val, data;
161 	bool valid_wptr = false;
162 
163 	m = get_mqd(mqd);
164 
165 	acquire_queue(adev, pipe_id, queue_id);
166 
167 	/* HIQ is set during driver init period with vmid set to 0*/
168 	if (m->cp_hqd_vmid == 0) {
169 		uint32_t value, mec, pipe;
170 
171 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
172 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
173 
174 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
175 			mec, pipe, queue_id);
176 		value = RREG32(mmRLC_CP_SCHEDULERS);
177 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
178 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
179 		WREG32(mmRLC_CP_SCHEDULERS, value);
180 	}
181 
182 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
183 	mqd_hqd = &m->cp_mqd_base_addr_lo;
184 
185 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_CONTROL; reg++)
186 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
187 
188 	/* Tonga errata: EOP RPTR/WPTR should be left unmodified.
189 	 * This is safe since EOP RPTR==WPTR for any inactive HQD
190 	 * on ASICs that do not support context-save.
191 	 * EOP writes/reads can start anywhere in the ring.
192 	 */
193 	if (adev->asic_type != CHIP_TONGA) {
194 		WREG32(mmCP_HQD_EOP_RPTR, m->cp_hqd_eop_rptr);
195 		WREG32(mmCP_HQD_EOP_WPTR, m->cp_hqd_eop_wptr);
196 		WREG32(mmCP_HQD_EOP_WPTR_MEM, m->cp_hqd_eop_wptr_mem);
197 	}
198 
199 	for (reg = mmCP_HQD_EOP_EVENTS; reg <= mmCP_HQD_ERROR; reg++)
200 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
201 
202 	/* Copy userspace write pointer value to register.
203 	 * Activate doorbell logic to monitor subsequent changes.
204 	 */
205 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
206 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
207 	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
208 
209 	/* read_user_ptr may take the mm->mmap_lock.
210 	 * release srbm_mutex to avoid circular dependency between
211 	 * srbm_mutex->mm_sem->reservation_ww_class_mutex->srbm_mutex.
212 	 */
213 	release_queue(adev);
214 	valid_wptr = read_user_wptr(mm, wptr, wptr_val);
215 	acquire_queue(adev, pipe_id, queue_id);
216 	if (valid_wptr)
217 		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
218 
219 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
220 	WREG32(mmCP_HQD_ACTIVE, data);
221 
222 	release_queue(adev);
223 
224 	return 0;
225 }
226 
227 static int kgd_hqd_dump(struct amdgpu_device *adev,
228 			uint32_t pipe_id, uint32_t queue_id,
229 			uint32_t (**dump)[2], uint32_t *n_regs)
230 {
231 	uint32_t i = 0, reg;
232 #define HQD_N_REGS (54+4)
233 #define DUMP_REG(addr) do {				\
234 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
235 			break;				\
236 		(*dump)[i][0] = (addr) << 2;		\
237 		(*dump)[i++][1] = RREG32(addr);		\
238 	} while (0)
239 
240 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
241 	if (*dump == NULL)
242 		return -ENOMEM;
243 
244 	acquire_queue(adev, pipe_id, queue_id);
245 
246 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE0);
247 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE1);
248 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE2);
249 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE3);
250 
251 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_DONES; reg++)
252 		DUMP_REG(reg);
253 
254 	release_queue(adev);
255 
256 	WARN_ON_ONCE(i != HQD_N_REGS);
257 	*n_regs = i;
258 
259 	return 0;
260 }
261 
262 static int kgd_hqd_sdma_load(struct amdgpu_device *adev, void *mqd,
263 			     uint32_t __user *wptr, struct mm_struct *mm)
264 {
265 	struct vi_sdma_mqd *m;
266 	unsigned long end_jiffies;
267 	uint32_t sdma_rlc_reg_offset;
268 	uint32_t data;
269 
270 	m = get_sdma_mqd(mqd);
271 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
272 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
273 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
274 
275 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
276 	while (true) {
277 		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
278 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
279 			break;
280 		if (time_after(jiffies, end_jiffies)) {
281 			pr_err("SDMA RLC not idle in %s\n", __func__);
282 			return -ETIME;
283 		}
284 		usleep_range(500, 1000);
285 	}
286 
287 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
288 			     ENABLE, 1);
289 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
290 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
291 				m->sdmax_rlcx_rb_rptr);
292 
293 	if (read_user_wptr(mm, wptr, data))
294 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR, data);
295 	else
296 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
297 		       m->sdmax_rlcx_rb_rptr);
298 
299 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_VIRTUAL_ADDR,
300 				m->sdmax_rlcx_virtual_addr);
301 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
302 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
303 			m->sdmax_rlcx_rb_base_hi);
304 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
305 			m->sdmax_rlcx_rb_rptr_addr_lo);
306 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
307 			m->sdmax_rlcx_rb_rptr_addr_hi);
308 
309 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
310 			     RB_ENABLE, 1);
311 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
312 
313 	return 0;
314 }
315 
316 static int kgd_hqd_sdma_dump(struct amdgpu_device *adev,
317 			     uint32_t engine_id, uint32_t queue_id,
318 			     uint32_t (**dump)[2], uint32_t *n_regs)
319 {
320 	uint32_t sdma_offset = engine_id * SDMA1_REGISTER_OFFSET +
321 		queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
322 	uint32_t i = 0, reg;
323 #undef HQD_N_REGS
324 #define HQD_N_REGS (19+4+2+3+7)
325 
326 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
327 	if (*dump == NULL)
328 		return -ENOMEM;
329 
330 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
331 		DUMP_REG(sdma_offset + reg);
332 	for (reg = mmSDMA0_RLC0_VIRTUAL_ADDR; reg <= mmSDMA0_RLC0_WATERMARK;
333 	     reg++)
334 		DUMP_REG(sdma_offset + reg);
335 	for (reg = mmSDMA0_RLC0_CSA_ADDR_LO; reg <= mmSDMA0_RLC0_CSA_ADDR_HI;
336 	     reg++)
337 		DUMP_REG(sdma_offset + reg);
338 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN; reg <= mmSDMA0_RLC0_DUMMY_REG;
339 	     reg++)
340 		DUMP_REG(sdma_offset + reg);
341 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0; reg <= mmSDMA0_RLC0_MIDCMD_CNTL;
342 	     reg++)
343 		DUMP_REG(sdma_offset + reg);
344 
345 	WARN_ON_ONCE(i != HQD_N_REGS);
346 	*n_regs = i;
347 
348 	return 0;
349 }
350 
351 static bool kgd_hqd_is_occupied(struct amdgpu_device *adev,
352 				uint64_t queue_address, uint32_t pipe_id,
353 				uint32_t queue_id)
354 {
355 	uint32_t act;
356 	bool retval = false;
357 	uint32_t low, high;
358 
359 	acquire_queue(adev, pipe_id, queue_id);
360 	act = RREG32(mmCP_HQD_ACTIVE);
361 	if (act) {
362 		low = lower_32_bits(queue_address >> 8);
363 		high = upper_32_bits(queue_address >> 8);
364 
365 		if (low == RREG32(mmCP_HQD_PQ_BASE) &&
366 				high == RREG32(mmCP_HQD_PQ_BASE_HI))
367 			retval = true;
368 	}
369 	release_queue(adev);
370 	return retval;
371 }
372 
373 static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd)
374 {
375 	struct vi_sdma_mqd *m;
376 	uint32_t sdma_rlc_reg_offset;
377 	uint32_t sdma_rlc_rb_cntl;
378 
379 	m = get_sdma_mqd(mqd);
380 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
381 
382 	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
383 
384 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
385 		return true;
386 
387 	return false;
388 }
389 
390 static int kgd_hqd_destroy(struct amdgpu_device *adev, void *mqd,
391 				enum kfd_preempt_type reset_type,
392 				unsigned int utimeout, uint32_t pipe_id,
393 				uint32_t queue_id)
394 {
395 	uint32_t temp;
396 	enum hqd_dequeue_request_type type;
397 	unsigned long flags, end_jiffies;
398 	int retry;
399 	struct vi_mqd *m = get_mqd(mqd);
400 
401 	if (amdgpu_in_reset(adev))
402 		return -EIO;
403 
404 	acquire_queue(adev, pipe_id, queue_id);
405 
406 	if (m->cp_hqd_vmid == 0)
407 		WREG32_FIELD(RLC_CP_SCHEDULERS, scheduler1, 0);
408 
409 	switch (reset_type) {
410 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
411 		type = DRAIN_PIPE;
412 		break;
413 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
414 		type = RESET_WAVES;
415 		break;
416 	default:
417 		type = DRAIN_PIPE;
418 		break;
419 	}
420 
421 	/* Workaround: If IQ timer is active and the wait time is close to or
422 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
423 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
424 	 * cleared before continuing. Also, ensure wait times are set to at
425 	 * least 0x3.
426 	 */
427 	local_irq_save(flags);
428 	preempt_disable();
429 	retry = 5000; /* wait for 500 usecs at maximum */
430 	while (true) {
431 		temp = RREG32(mmCP_HQD_IQ_TIMER);
432 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
433 			pr_debug("HW is processing IQ\n");
434 			goto loop;
435 		}
436 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
437 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
438 					== 3) /* SEM-rearm is safe */
439 				break;
440 			/* Wait time 3 is safe for CP, but our MMIO read/write
441 			 * time is close to 1 microsecond, so check for 10 to
442 			 * leave more buffer room
443 			 */
444 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
445 					>= 10)
446 				break;
447 			pr_debug("IQ timer is active\n");
448 		} else
449 			break;
450 loop:
451 		if (!retry) {
452 			pr_err("CP HQD IQ timer status time out\n");
453 			break;
454 		}
455 		ndelay(100);
456 		--retry;
457 	}
458 	retry = 1000;
459 	while (true) {
460 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
461 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
462 			break;
463 		pr_debug("Dequeue request is pending\n");
464 
465 		if (!retry) {
466 			pr_err("CP HQD dequeue request time out\n");
467 			break;
468 		}
469 		ndelay(100);
470 		--retry;
471 	}
472 	local_irq_restore(flags);
473 	preempt_enable();
474 
475 	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
476 
477 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
478 	while (true) {
479 		temp = RREG32(mmCP_HQD_ACTIVE);
480 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
481 			break;
482 		if (time_after(jiffies, end_jiffies)) {
483 			pr_err("cp queue preemption time out.\n");
484 			release_queue(adev);
485 			return -ETIME;
486 		}
487 		usleep_range(500, 1000);
488 	}
489 
490 	release_queue(adev);
491 	return 0;
492 }
493 
494 static int kgd_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd,
495 				unsigned int utimeout)
496 {
497 	struct vi_sdma_mqd *m;
498 	uint32_t sdma_rlc_reg_offset;
499 	uint32_t temp;
500 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
501 
502 	m = get_sdma_mqd(mqd);
503 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
504 
505 	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
506 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
507 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
508 
509 	while (true) {
510 		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
511 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
512 			break;
513 		if (time_after(jiffies, end_jiffies)) {
514 			pr_err("SDMA RLC not idle in %s\n", __func__);
515 			return -ETIME;
516 		}
517 		usleep_range(500, 1000);
518 	}
519 
520 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
521 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
522 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
523 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
524 
525 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
526 
527 	return 0;
528 }
529 
530 static bool get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
531 					uint8_t vmid, uint16_t *p_pasid)
532 {
533 	uint32_t value;
534 
535 	value = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
536 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
537 
538 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
539 }
540 
541 static int kgd_wave_control_execute(struct amdgpu_device *adev,
542 					uint32_t gfx_index_val,
543 					uint32_t sq_cmd)
544 {
545 	uint32_t data = 0;
546 
547 	mutex_lock(&adev->grbm_idx_mutex);
548 
549 	WREG32(mmGRBM_GFX_INDEX, gfx_index_val);
550 	WREG32(mmSQ_CMD, sq_cmd);
551 
552 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
553 		INSTANCE_BROADCAST_WRITES, 1);
554 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
555 		SH_BROADCAST_WRITES, 1);
556 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
557 		SE_BROADCAST_WRITES, 1);
558 
559 	WREG32(mmGRBM_GFX_INDEX, data);
560 	mutex_unlock(&adev->grbm_idx_mutex);
561 
562 	return 0;
563 }
564 
565 static void set_scratch_backing_va(struct amdgpu_device *adev,
566 					uint64_t va, uint32_t vmid)
567 {
568 	lock_srbm(adev, 0, 0, 0, vmid);
569 	WREG32(mmSH_HIDDEN_PRIVATE_BASE_VMID, va);
570 	unlock_srbm(adev);
571 }
572 
573 static void set_vm_context_page_table_base(struct amdgpu_device *adev,
574 		uint32_t vmid, uint64_t page_table_base)
575 {
576 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
577 		pr_err("trying to set page table base for wrong VMID\n");
578 		return;
579 	}
580 	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8,
581 			lower_32_bits(page_table_base));
582 }
583 
584 const struct kfd2kgd_calls gfx_v8_kfd2kgd = {
585 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
586 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
587 	.init_interrupts = kgd_init_interrupts,
588 	.hqd_load = kgd_hqd_load,
589 	.hqd_sdma_load = kgd_hqd_sdma_load,
590 	.hqd_dump = kgd_hqd_dump,
591 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
592 	.hqd_is_occupied = kgd_hqd_is_occupied,
593 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
594 	.hqd_destroy = kgd_hqd_destroy,
595 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
596 	.wave_control_execute = kgd_wave_control_execute,
597 	.get_atc_vmid_pasid_mapping_info =
598 			get_atc_vmid_pasid_mapping_info,
599 	.set_scratch_backing_va = set_scratch_backing_va,
600 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
601 };
602