xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c (revision 3d0fe49454652117522f60bfbefb978ba0e5300b)
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, uint32_t inst)
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, uint32_t inst)
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 				uint32_t inst)
114 {
115 	uint32_t mec;
116 	uint32_t pipe;
117 
118 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
119 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
120 
121 	lock_srbm(adev, mec, pipe, 0, 0);
122 
123 	WREG32(mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
124 			CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
125 
126 	unlock_srbm(adev);
127 
128 	return 0;
129 }
130 
131 static inline uint32_t get_sdma_rlc_reg_offset(struct vi_sdma_mqd *m)
132 {
133 	uint32_t retval;
134 
135 	retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
136 		m->sdma_queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
137 
138 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n",
139 			m->sdma_engine_id, m->sdma_queue_id, retval);
140 
141 	return retval;
142 }
143 
144 static inline struct vi_mqd *get_mqd(void *mqd)
145 {
146 	return (struct vi_mqd *)mqd;
147 }
148 
149 static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd)
150 {
151 	return (struct vi_sdma_mqd *)mqd;
152 }
153 
154 static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd,
155 			uint32_t pipe_id, uint32_t queue_id,
156 			uint32_t __user *wptr, uint32_t wptr_shift,
157 			uint32_t wptr_mask, struct mm_struct *mm, uint32_t inst)
158 {
159 	struct vi_mqd *m;
160 	uint32_t *mqd_hqd;
161 	uint32_t reg, wptr_val, data;
162 	bool valid_wptr = false;
163 
164 	m = get_mqd(mqd);
165 
166 	acquire_queue(adev, pipe_id, queue_id);
167 
168 	/* HIQ is set during driver init period with vmid set to 0*/
169 	if (m->cp_hqd_vmid == 0) {
170 		uint32_t value, mec, pipe;
171 
172 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
173 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
174 
175 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
176 			mec, pipe, queue_id);
177 		value = RREG32(mmRLC_CP_SCHEDULERS);
178 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
179 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
180 		WREG32(mmRLC_CP_SCHEDULERS, value);
181 	}
182 
183 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
184 	mqd_hqd = &m->cp_mqd_base_addr_lo;
185 
186 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_CONTROL; reg++)
187 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
188 
189 	/* Tonga errata: EOP RPTR/WPTR should be left unmodified.
190 	 * This is safe since EOP RPTR==WPTR for any inactive HQD
191 	 * on ASICs that do not support context-save.
192 	 * EOP writes/reads can start anywhere in the ring.
193 	 */
194 	if (adev->asic_type != CHIP_TONGA) {
195 		WREG32(mmCP_HQD_EOP_RPTR, m->cp_hqd_eop_rptr);
196 		WREG32(mmCP_HQD_EOP_WPTR, m->cp_hqd_eop_wptr);
197 		WREG32(mmCP_HQD_EOP_WPTR_MEM, m->cp_hqd_eop_wptr_mem);
198 	}
199 
200 	for (reg = mmCP_HQD_EOP_EVENTS; reg <= mmCP_HQD_ERROR; reg++)
201 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
202 
203 	/* Copy userspace write pointer value to register.
204 	 * Activate doorbell logic to monitor subsequent changes.
205 	 */
206 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
207 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
208 	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
209 
210 	/* read_user_ptr may take the mm->mmap_lock.
211 	 * release srbm_mutex to avoid circular dependency between
212 	 * srbm_mutex->mmap_lock->reservation_ww_class_mutex->srbm_mutex.
213 	 */
214 	release_queue(adev);
215 	valid_wptr = read_user_wptr(mm, wptr, wptr_val);
216 	acquire_queue(adev, pipe_id, queue_id);
217 	if (valid_wptr)
218 		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
219 
220 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
221 	WREG32(mmCP_HQD_ACTIVE, data);
222 
223 	release_queue(adev);
224 
225 	return 0;
226 }
227 
228 static int kgd_hqd_dump(struct amdgpu_device *adev,
229 			uint32_t pipe_id, uint32_t queue_id,
230 			uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst)
231 {
232 	uint32_t i = 0, reg;
233 #define HQD_N_REGS (54+4)
234 #define DUMP_REG(addr) do {				\
235 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
236 			break;				\
237 		(*dump)[i][0] = (addr) << 2;		\
238 		(*dump)[i++][1] = RREG32(addr);		\
239 	} while (0)
240 
241 	*dump = kmalloc_array(HQD_N_REGS, sizeof(**dump), GFP_KERNEL);
242 	if (*dump == NULL)
243 		return -ENOMEM;
244 
245 	acquire_queue(adev, pipe_id, queue_id);
246 
247 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE0);
248 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE1);
249 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE2);
250 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE3);
251 
252 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_DONES; reg++)
253 		DUMP_REG(reg);
254 
255 	release_queue(adev);
256 
257 	WARN_ON_ONCE(i != HQD_N_REGS);
258 	*n_regs = i;
259 
260 	return 0;
261 }
262 
263 static int kgd_hqd_sdma_load(struct amdgpu_device *adev, void *mqd,
264 			     uint32_t __user *wptr, struct mm_struct *mm)
265 {
266 	struct vi_sdma_mqd *m;
267 	unsigned long end_jiffies;
268 	uint32_t sdma_rlc_reg_offset;
269 	uint32_t data;
270 
271 	m = get_sdma_mqd(mqd);
272 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
273 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
274 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
275 
276 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
277 	while (true) {
278 		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
279 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
280 			break;
281 		if (time_after(jiffies, end_jiffies)) {
282 			pr_err("SDMA RLC not idle in %s\n", __func__);
283 			return -ETIME;
284 		}
285 		usleep_range(500, 1000);
286 	}
287 
288 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
289 			     ENABLE, 1);
290 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
291 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
292 				m->sdmax_rlcx_rb_rptr);
293 
294 	if (read_user_wptr(mm, wptr, data))
295 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR, data);
296 	else
297 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
298 		       m->sdmax_rlcx_rb_rptr);
299 
300 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_VIRTUAL_ADDR,
301 				m->sdmax_rlcx_virtual_addr);
302 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
303 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
304 			m->sdmax_rlcx_rb_base_hi);
305 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
306 			m->sdmax_rlcx_rb_rptr_addr_lo);
307 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
308 			m->sdmax_rlcx_rb_rptr_addr_hi);
309 
310 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
311 			     RB_ENABLE, 1);
312 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
313 
314 	return 0;
315 }
316 
317 static int kgd_hqd_sdma_dump(struct amdgpu_device *adev,
318 			     uint32_t engine_id, uint32_t queue_id,
319 			     uint32_t (**dump)[2], uint32_t *n_regs)
320 {
321 	uint32_t sdma_offset = engine_id * SDMA1_REGISTER_OFFSET +
322 		queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
323 	uint32_t i = 0, reg;
324 #undef HQD_N_REGS
325 #define HQD_N_REGS (19+4+2+3+7)
326 
327 	*dump = kmalloc_array(HQD_N_REGS, sizeof(**dump), GFP_KERNEL);
328 	if (*dump == NULL)
329 		return -ENOMEM;
330 
331 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
332 		DUMP_REG(sdma_offset + reg);
333 	for (reg = mmSDMA0_RLC0_VIRTUAL_ADDR; reg <= mmSDMA0_RLC0_WATERMARK;
334 	     reg++)
335 		DUMP_REG(sdma_offset + reg);
336 	for (reg = mmSDMA0_RLC0_CSA_ADDR_LO; reg <= mmSDMA0_RLC0_CSA_ADDR_HI;
337 	     reg++)
338 		DUMP_REG(sdma_offset + reg);
339 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN; reg <= mmSDMA0_RLC0_DUMMY_REG;
340 	     reg++)
341 		DUMP_REG(sdma_offset + reg);
342 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0; reg <= mmSDMA0_RLC0_MIDCMD_CNTL;
343 	     reg++)
344 		DUMP_REG(sdma_offset + reg);
345 
346 	WARN_ON_ONCE(i != HQD_N_REGS);
347 	*n_regs = i;
348 
349 	return 0;
350 }
351 
352 static bool kgd_hqd_is_occupied(struct amdgpu_device *adev,
353 				uint64_t queue_address, uint32_t pipe_id,
354 				uint32_t queue_id, uint32_t inst)
355 {
356 	uint32_t act;
357 	bool retval = false;
358 	uint32_t low, high;
359 
360 	acquire_queue(adev, pipe_id, queue_id);
361 	act = RREG32(mmCP_HQD_ACTIVE);
362 	if (act) {
363 		low = lower_32_bits(queue_address >> 8);
364 		high = upper_32_bits(queue_address >> 8);
365 
366 		if (low == RREG32(mmCP_HQD_PQ_BASE) &&
367 				high == RREG32(mmCP_HQD_PQ_BASE_HI))
368 			retval = true;
369 	}
370 	release_queue(adev);
371 	return retval;
372 }
373 
374 static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd)
375 {
376 	struct vi_sdma_mqd *m;
377 	uint32_t sdma_rlc_reg_offset;
378 	uint32_t sdma_rlc_rb_cntl;
379 
380 	m = get_sdma_mqd(mqd);
381 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
382 
383 	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
384 
385 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
386 		return true;
387 
388 	return false;
389 }
390 
391 static int kgd_hqd_destroy(struct amdgpu_device *adev, void *mqd,
392 				enum kfd_preempt_type reset_type,
393 				unsigned int utimeout, uint32_t pipe_id,
394 				uint32_t queue_id, uint32_t inst)
395 {
396 	uint32_t temp;
397 	enum hqd_dequeue_request_type type;
398 	unsigned long flags, end_jiffies;
399 	int retry;
400 	struct vi_mqd *m = get_mqd(mqd);
401 
402 	if (amdgpu_in_reset(adev))
403 		return -EIO;
404 
405 	acquire_queue(adev, pipe_id, queue_id);
406 
407 	if (m->cp_hqd_vmid == 0)
408 		WREG32_FIELD(RLC_CP_SCHEDULERS, scheduler1, 0);
409 
410 	switch (reset_type) {
411 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
412 		type = DRAIN_PIPE;
413 		break;
414 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
415 		type = RESET_WAVES;
416 		break;
417 	default:
418 		type = DRAIN_PIPE;
419 		break;
420 	}
421 
422 	/* Workaround: If IQ timer is active and the wait time is close to or
423 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
424 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
425 	 * cleared before continuing. Also, ensure wait times are set to at
426 	 * least 0x3.
427 	 */
428 	local_irq_save(flags);
429 	preempt_disable();
430 	retry = 5000; /* wait for 500 usecs at maximum */
431 	while (true) {
432 		temp = RREG32(mmCP_HQD_IQ_TIMER);
433 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
434 			pr_debug("HW is processing IQ\n");
435 			goto loop;
436 		}
437 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
438 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
439 					== 3) /* SEM-rearm is safe */
440 				break;
441 			/* Wait time 3 is safe for CP, but our MMIO read/write
442 			 * time is close to 1 microsecond, so check for 10 to
443 			 * leave more buffer room
444 			 */
445 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
446 					>= 10)
447 				break;
448 			pr_debug("IQ timer is active\n");
449 		} else
450 			break;
451 loop:
452 		if (!retry) {
453 			pr_err("CP HQD IQ timer status time out\n");
454 			break;
455 		}
456 		ndelay(100);
457 		--retry;
458 	}
459 	retry = 1000;
460 	while (true) {
461 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
462 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
463 			break;
464 		pr_debug("Dequeue request is pending\n");
465 
466 		if (!retry) {
467 			pr_err("CP HQD dequeue request time out\n");
468 			break;
469 		}
470 		ndelay(100);
471 		--retry;
472 	}
473 	local_irq_restore(flags);
474 	preempt_enable();
475 
476 	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
477 
478 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
479 	while (true) {
480 		temp = RREG32(mmCP_HQD_ACTIVE);
481 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
482 			break;
483 		if (time_after(jiffies, end_jiffies)) {
484 			pr_err("cp queue preemption time out.\n");
485 			release_queue(adev);
486 			return -ETIME;
487 		}
488 		usleep_range(500, 1000);
489 	}
490 
491 	release_queue(adev);
492 	return 0;
493 }
494 
495 static int kgd_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd,
496 				unsigned int utimeout)
497 {
498 	struct vi_sdma_mqd *m;
499 	uint32_t sdma_rlc_reg_offset;
500 	uint32_t temp;
501 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
502 
503 	m = get_sdma_mqd(mqd);
504 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
505 
506 	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
507 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
508 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
509 
510 	while (true) {
511 		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
512 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
513 			break;
514 		if (time_after(jiffies, end_jiffies)) {
515 			pr_err("SDMA RLC not idle in %s\n", __func__);
516 			return -ETIME;
517 		}
518 		usleep_range(500, 1000);
519 	}
520 
521 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
522 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
523 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
524 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
525 
526 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
527 
528 	return 0;
529 }
530 
531 static bool get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
532 					uint8_t vmid, uint16_t *p_pasid)
533 {
534 	uint32_t value;
535 
536 	value = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
537 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
538 
539 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
540 }
541 
542 static int kgd_wave_control_execute(struct amdgpu_device *adev,
543 					uint32_t gfx_index_val,
544 					uint32_t sq_cmd, uint32_t inst)
545 {
546 	uint32_t data = 0;
547 
548 	mutex_lock(&adev->grbm_idx_mutex);
549 
550 	WREG32(mmGRBM_GFX_INDEX, gfx_index_val);
551 	WREG32(mmSQ_CMD, sq_cmd);
552 
553 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
554 		INSTANCE_BROADCAST_WRITES, 1);
555 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
556 		SH_BROADCAST_WRITES, 1);
557 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
558 		SE_BROADCAST_WRITES, 1);
559 
560 	WREG32(mmGRBM_GFX_INDEX, data);
561 	mutex_unlock(&adev->grbm_idx_mutex);
562 
563 	return 0;
564 }
565 
566 static void set_scratch_backing_va(struct amdgpu_device *adev,
567 					uint64_t va, uint32_t vmid)
568 {
569 	lock_srbm(adev, 0, 0, 0, vmid);
570 	WREG32(mmSH_HIDDEN_PRIVATE_BASE_VMID, va);
571 	unlock_srbm(adev);
572 }
573 
574 static void set_vm_context_page_table_base(struct amdgpu_device *adev,
575 		uint32_t vmid, uint64_t page_table_base)
576 {
577 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
578 		pr_err("trying to set page table base for wrong VMID\n");
579 		return;
580 	}
581 	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8,
582 			lower_32_bits(page_table_base));
583 }
584 
585 const struct kfd2kgd_calls gfx_v8_kfd2kgd = {
586 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
587 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
588 	.init_interrupts = kgd_init_interrupts,
589 	.hqd_load = kgd_hqd_load,
590 	.hqd_sdma_load = kgd_hqd_sdma_load,
591 	.hqd_dump = kgd_hqd_dump,
592 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
593 	.hqd_is_occupied = kgd_hqd_is_occupied,
594 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
595 	.hqd_destroy = kgd_hqd_destroy,
596 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
597 	.wave_control_execute = kgd_wave_control_execute,
598 	.get_atc_vmid_pasid_mapping_info =
599 			get_atc_vmid_pasid_mapping_info,
600 	.set_scratch_backing_va = set_scratch_backing_va,
601 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
602 };
603