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 <linux/mmu_context.h>
24 
25 #include "amdgpu.h"
26 #include "amdgpu_amdkfd.h"
27 #include "gfx_v8_0.h"
28 #include "gca/gfx_8_0_sh_mask.h"
29 #include "gca/gfx_8_0_d.h"
30 #include "gca/gfx_8_0_enum.h"
31 #include "oss/oss_3_0_sh_mask.h"
32 #include "oss/oss_3_0_d.h"
33 #include "gmc/gmc_8_1_sh_mask.h"
34 #include "gmc/gmc_8_1_d.h"
35 #include "vi_structs.h"
36 #include "vid.h"
37 
38 enum hqd_dequeue_request_type {
39 	NO_ACTION = 0,
40 	DRAIN_PIPE,
41 	RESET_WAVES
42 };
43 
44 /* Because of REG_GET_FIELD() being used, we put this function in the
45  * asic specific file.
46  */
47 static int get_tile_config(struct kgd_dev *kgd,
48 		struct tile_config *config)
49 {
50 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
51 
52 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
53 	config->num_banks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
54 				MC_ARB_RAMCFG, NOOFBANK);
55 	config->num_ranks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
56 				MC_ARB_RAMCFG, NOOFRANKS);
57 
58 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
59 	config->num_tile_configs =
60 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
61 	config->macro_tile_config_ptr =
62 			adev->gfx.config.macrotile_mode_array;
63 	config->num_macro_tile_configs =
64 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
65 
66 	return 0;
67 }
68 
69 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
70 {
71 	return (struct amdgpu_device *)kgd;
72 }
73 
74 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
75 			uint32_t queue, uint32_t vmid)
76 {
77 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
78 	uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
79 
80 	mutex_lock(&adev->srbm_mutex);
81 	WREG32(mmSRBM_GFX_CNTL, value);
82 }
83 
84 static void unlock_srbm(struct kgd_dev *kgd)
85 {
86 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
87 
88 	WREG32(mmSRBM_GFX_CNTL, 0);
89 	mutex_unlock(&adev->srbm_mutex);
90 }
91 
92 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
93 				uint32_t queue_id)
94 {
95 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
96 
97 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
98 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
99 
100 	lock_srbm(kgd, mec, pipe, queue_id, 0);
101 }
102 
103 static void release_queue(struct kgd_dev *kgd)
104 {
105 	unlock_srbm(kgd);
106 }
107 
108 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
109 					uint32_t sh_mem_config,
110 					uint32_t sh_mem_ape1_base,
111 					uint32_t sh_mem_ape1_limit,
112 					uint32_t sh_mem_bases)
113 {
114 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
115 
116 	lock_srbm(kgd, 0, 0, 0, vmid);
117 
118 	WREG32(mmSH_MEM_CONFIG, sh_mem_config);
119 	WREG32(mmSH_MEM_APE1_BASE, sh_mem_ape1_base);
120 	WREG32(mmSH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
121 	WREG32(mmSH_MEM_BASES, sh_mem_bases);
122 
123 	unlock_srbm(kgd);
124 }
125 
126 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
127 					unsigned int vmid)
128 {
129 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
130 
131 	/*
132 	 * We have to assume that there is no outstanding mapping.
133 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
134 	 * a mapping is in progress or because a mapping finished
135 	 * and the SW cleared it.
136 	 * So the protocol is to always wait & clear.
137 	 */
138 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
139 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
140 
141 	WREG32(mmATC_VMID0_PASID_MAPPING + vmid, pasid_mapping);
142 
143 	while (!(RREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))
144 		cpu_relax();
145 	WREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
146 
147 	/* Mapping vmid to pasid also for IH block */
148 	WREG32(mmIH_VMID_0_LUT + vmid, pasid_mapping);
149 
150 	return 0;
151 }
152 
153 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
154 {
155 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
156 	uint32_t mec;
157 	uint32_t pipe;
158 
159 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
160 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
161 
162 	lock_srbm(kgd, mec, pipe, 0, 0);
163 
164 	WREG32(mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
165 			CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
166 
167 	unlock_srbm(kgd);
168 
169 	return 0;
170 }
171 
172 static inline uint32_t get_sdma_rlc_reg_offset(struct vi_sdma_mqd *m)
173 {
174 	uint32_t retval;
175 
176 	retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
177 		m->sdma_queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
178 
179 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n",
180 			m->sdma_engine_id, m->sdma_queue_id, retval);
181 
182 	return retval;
183 }
184 
185 static inline struct vi_mqd *get_mqd(void *mqd)
186 {
187 	return (struct vi_mqd *)mqd;
188 }
189 
190 static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd)
191 {
192 	return (struct vi_sdma_mqd *)mqd;
193 }
194 
195 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
196 			uint32_t queue_id, uint32_t __user *wptr,
197 			uint32_t wptr_shift, uint32_t wptr_mask,
198 			struct mm_struct *mm)
199 {
200 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
201 	struct vi_mqd *m;
202 	uint32_t *mqd_hqd;
203 	uint32_t reg, wptr_val, data;
204 	bool valid_wptr = false;
205 
206 	m = get_mqd(mqd);
207 
208 	acquire_queue(kgd, pipe_id, queue_id);
209 
210 	/* HIQ is set during driver init period with vmid set to 0*/
211 	if (m->cp_hqd_vmid == 0) {
212 		uint32_t value, mec, pipe;
213 
214 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
215 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
216 
217 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
218 			mec, pipe, queue_id);
219 		value = RREG32(mmRLC_CP_SCHEDULERS);
220 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
221 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
222 		WREG32(mmRLC_CP_SCHEDULERS, value);
223 	}
224 
225 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
226 	mqd_hqd = &m->cp_mqd_base_addr_lo;
227 
228 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_CONTROL; reg++)
229 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
230 
231 	/* Tonga errata: EOP RPTR/WPTR should be left unmodified.
232 	 * This is safe since EOP RPTR==WPTR for any inactive HQD
233 	 * on ASICs that do not support context-save.
234 	 * EOP writes/reads can start anywhere in the ring.
235 	 */
236 	if (get_amdgpu_device(kgd)->asic_type != CHIP_TONGA) {
237 		WREG32(mmCP_HQD_EOP_RPTR, m->cp_hqd_eop_rptr);
238 		WREG32(mmCP_HQD_EOP_WPTR, m->cp_hqd_eop_wptr);
239 		WREG32(mmCP_HQD_EOP_WPTR_MEM, m->cp_hqd_eop_wptr_mem);
240 	}
241 
242 	for (reg = mmCP_HQD_EOP_EVENTS; reg <= mmCP_HQD_ERROR; reg++)
243 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
244 
245 	/* Copy userspace write pointer value to register.
246 	 * Activate doorbell logic to monitor subsequent changes.
247 	 */
248 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
249 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
250 	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
251 
252 	/* read_user_ptr may take the mm->mmap_sem.
253 	 * release srbm_mutex to avoid circular dependency between
254 	 * srbm_mutex->mm_sem->reservation_ww_class_mutex->srbm_mutex.
255 	 */
256 	release_queue(kgd);
257 	valid_wptr = read_user_wptr(mm, wptr, wptr_val);
258 	acquire_queue(kgd, pipe_id, queue_id);
259 	if (valid_wptr)
260 		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
261 
262 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
263 	WREG32(mmCP_HQD_ACTIVE, data);
264 
265 	release_queue(kgd);
266 
267 	return 0;
268 }
269 
270 static int kgd_hqd_dump(struct kgd_dev *kgd,
271 			uint32_t pipe_id, uint32_t queue_id,
272 			uint32_t (**dump)[2], uint32_t *n_regs)
273 {
274 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
275 	uint32_t i = 0, reg;
276 #define HQD_N_REGS (54+4)
277 #define DUMP_REG(addr) do {				\
278 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
279 			break;				\
280 		(*dump)[i][0] = (addr) << 2;		\
281 		(*dump)[i++][1] = RREG32(addr);		\
282 	} while (0)
283 
284 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
285 	if (*dump == NULL)
286 		return -ENOMEM;
287 
288 	acquire_queue(kgd, pipe_id, queue_id);
289 
290 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE0);
291 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE1);
292 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE2);
293 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE3);
294 
295 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_DONES; reg++)
296 		DUMP_REG(reg);
297 
298 	release_queue(kgd);
299 
300 	WARN_ON_ONCE(i != HQD_N_REGS);
301 	*n_regs = i;
302 
303 	return 0;
304 }
305 
306 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
307 			     uint32_t __user *wptr, struct mm_struct *mm)
308 {
309 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
310 	struct vi_sdma_mqd *m;
311 	unsigned long end_jiffies;
312 	uint32_t sdma_rlc_reg_offset;
313 	uint32_t data;
314 
315 	m = get_sdma_mqd(mqd);
316 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
317 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
318 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
319 
320 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
321 	while (true) {
322 		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
323 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
324 			break;
325 		if (time_after(jiffies, end_jiffies)) {
326 			pr_err("SDMA RLC not idle in %s\n", __func__);
327 			return -ETIME;
328 		}
329 		usleep_range(500, 1000);
330 	}
331 
332 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
333 			     ENABLE, 1);
334 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
335 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
336 				m->sdmax_rlcx_rb_rptr);
337 
338 	if (read_user_wptr(mm, wptr, data))
339 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR, data);
340 	else
341 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
342 		       m->sdmax_rlcx_rb_rptr);
343 
344 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_VIRTUAL_ADDR,
345 				m->sdmax_rlcx_virtual_addr);
346 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
347 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
348 			m->sdmax_rlcx_rb_base_hi);
349 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
350 			m->sdmax_rlcx_rb_rptr_addr_lo);
351 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
352 			m->sdmax_rlcx_rb_rptr_addr_hi);
353 
354 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
355 			     RB_ENABLE, 1);
356 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
357 
358 	return 0;
359 }
360 
361 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
362 			     uint32_t engine_id, uint32_t queue_id,
363 			     uint32_t (**dump)[2], uint32_t *n_regs)
364 {
365 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
366 	uint32_t sdma_offset = engine_id * SDMA1_REGISTER_OFFSET +
367 		queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
368 	uint32_t i = 0, reg;
369 #undef HQD_N_REGS
370 #define HQD_N_REGS (19+4+2+3+7)
371 
372 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
373 	if (*dump == NULL)
374 		return -ENOMEM;
375 
376 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
377 		DUMP_REG(sdma_offset + reg);
378 	for (reg = mmSDMA0_RLC0_VIRTUAL_ADDR; reg <= mmSDMA0_RLC0_WATERMARK;
379 	     reg++)
380 		DUMP_REG(sdma_offset + reg);
381 	for (reg = mmSDMA0_RLC0_CSA_ADDR_LO; reg <= mmSDMA0_RLC0_CSA_ADDR_HI;
382 	     reg++)
383 		DUMP_REG(sdma_offset + reg);
384 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN; reg <= mmSDMA0_RLC0_DUMMY_REG;
385 	     reg++)
386 		DUMP_REG(sdma_offset + reg);
387 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0; reg <= mmSDMA0_RLC0_MIDCMD_CNTL;
388 	     reg++)
389 		DUMP_REG(sdma_offset + reg);
390 
391 	WARN_ON_ONCE(i != HQD_N_REGS);
392 	*n_regs = i;
393 
394 	return 0;
395 }
396 
397 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
398 				uint32_t pipe_id, uint32_t queue_id)
399 {
400 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
401 	uint32_t act;
402 	bool retval = false;
403 	uint32_t low, high;
404 
405 	acquire_queue(kgd, pipe_id, queue_id);
406 	act = RREG32(mmCP_HQD_ACTIVE);
407 	if (act) {
408 		low = lower_32_bits(queue_address >> 8);
409 		high = upper_32_bits(queue_address >> 8);
410 
411 		if (low == RREG32(mmCP_HQD_PQ_BASE) &&
412 				high == RREG32(mmCP_HQD_PQ_BASE_HI))
413 			retval = true;
414 	}
415 	release_queue(kgd);
416 	return retval;
417 }
418 
419 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
420 {
421 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
422 	struct vi_sdma_mqd *m;
423 	uint32_t sdma_rlc_reg_offset;
424 	uint32_t sdma_rlc_rb_cntl;
425 
426 	m = get_sdma_mqd(mqd);
427 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
428 
429 	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
430 
431 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
432 		return true;
433 
434 	return false;
435 }
436 
437 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
438 				enum kfd_preempt_type reset_type,
439 				unsigned int utimeout, uint32_t pipe_id,
440 				uint32_t queue_id)
441 {
442 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
443 	uint32_t temp;
444 	enum hqd_dequeue_request_type type;
445 	unsigned long flags, end_jiffies;
446 	int retry;
447 	struct vi_mqd *m = get_mqd(mqd);
448 
449 	if (adev->in_gpu_reset)
450 		return -EIO;
451 
452 	acquire_queue(kgd, pipe_id, queue_id);
453 
454 	if (m->cp_hqd_vmid == 0)
455 		WREG32_FIELD(RLC_CP_SCHEDULERS, scheduler1, 0);
456 
457 	switch (reset_type) {
458 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
459 		type = DRAIN_PIPE;
460 		break;
461 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
462 		type = RESET_WAVES;
463 		break;
464 	default:
465 		type = DRAIN_PIPE;
466 		break;
467 	}
468 
469 	/* Workaround: If IQ timer is active and the wait time is close to or
470 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
471 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
472 	 * cleared before continuing. Also, ensure wait times are set to at
473 	 * least 0x3.
474 	 */
475 	local_irq_save(flags);
476 	preempt_disable();
477 	retry = 5000; /* wait for 500 usecs at maximum */
478 	while (true) {
479 		temp = RREG32(mmCP_HQD_IQ_TIMER);
480 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
481 			pr_debug("HW is processing IQ\n");
482 			goto loop;
483 		}
484 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
485 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
486 					== 3) /* SEM-rearm is safe */
487 				break;
488 			/* Wait time 3 is safe for CP, but our MMIO read/write
489 			 * time is close to 1 microsecond, so check for 10 to
490 			 * leave more buffer room
491 			 */
492 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
493 					>= 10)
494 				break;
495 			pr_debug("IQ timer is active\n");
496 		} else
497 			break;
498 loop:
499 		if (!retry) {
500 			pr_err("CP HQD IQ timer status time out\n");
501 			break;
502 		}
503 		ndelay(100);
504 		--retry;
505 	}
506 	retry = 1000;
507 	while (true) {
508 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
509 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
510 			break;
511 		pr_debug("Dequeue request is pending\n");
512 
513 		if (!retry) {
514 			pr_err("CP HQD dequeue request time out\n");
515 			break;
516 		}
517 		ndelay(100);
518 		--retry;
519 	}
520 	local_irq_restore(flags);
521 	preempt_enable();
522 
523 	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
524 
525 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
526 	while (true) {
527 		temp = RREG32(mmCP_HQD_ACTIVE);
528 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
529 			break;
530 		if (time_after(jiffies, end_jiffies)) {
531 			pr_err("cp queue preemption time out.\n");
532 			release_queue(kgd);
533 			return -ETIME;
534 		}
535 		usleep_range(500, 1000);
536 	}
537 
538 	release_queue(kgd);
539 	return 0;
540 }
541 
542 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
543 				unsigned int utimeout)
544 {
545 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
546 	struct vi_sdma_mqd *m;
547 	uint32_t sdma_rlc_reg_offset;
548 	uint32_t temp;
549 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
550 
551 	m = get_sdma_mqd(mqd);
552 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
553 
554 	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
555 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
556 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
557 
558 	while (true) {
559 		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
560 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
561 			break;
562 		if (time_after(jiffies, end_jiffies)) {
563 			pr_err("SDMA RLC not idle in %s\n", __func__);
564 			return -ETIME;
565 		}
566 		usleep_range(500, 1000);
567 	}
568 
569 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
570 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
571 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
572 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
573 
574 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
575 
576 	return 0;
577 }
578 
579 static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
580 					uint8_t vmid, uint16_t *p_pasid)
581 {
582 	uint32_t value;
583 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
584 
585 	value = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
586 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
587 
588 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
589 }
590 
591 static int kgd_address_watch_disable(struct kgd_dev *kgd)
592 {
593 	return 0;
594 }
595 
596 static int kgd_address_watch_execute(struct kgd_dev *kgd,
597 					unsigned int watch_point_id,
598 					uint32_t cntl_val,
599 					uint32_t addr_hi,
600 					uint32_t addr_lo)
601 {
602 	return 0;
603 }
604 
605 static int kgd_wave_control_execute(struct kgd_dev *kgd,
606 					uint32_t gfx_index_val,
607 					uint32_t sq_cmd)
608 {
609 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
610 	uint32_t data = 0;
611 
612 	mutex_lock(&adev->grbm_idx_mutex);
613 
614 	WREG32(mmGRBM_GFX_INDEX, gfx_index_val);
615 	WREG32(mmSQ_CMD, sq_cmd);
616 
617 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
618 		INSTANCE_BROADCAST_WRITES, 1);
619 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
620 		SH_BROADCAST_WRITES, 1);
621 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
622 		SE_BROADCAST_WRITES, 1);
623 
624 	WREG32(mmGRBM_GFX_INDEX, data);
625 	mutex_unlock(&adev->grbm_idx_mutex);
626 
627 	return 0;
628 }
629 
630 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
631 					unsigned int watch_point_id,
632 					unsigned int reg_offset)
633 {
634 	return 0;
635 }
636 
637 static void set_scratch_backing_va(struct kgd_dev *kgd,
638 					uint64_t va, uint32_t vmid)
639 {
640 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
641 
642 	lock_srbm(kgd, 0, 0, 0, vmid);
643 	WREG32(mmSH_HIDDEN_PRIVATE_BASE_VMID, va);
644 	unlock_srbm(kgd);
645 }
646 
647 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
648 		uint64_t page_table_base)
649 {
650 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
651 
652 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
653 		pr_err("trying to set page table base for wrong VMID\n");
654 		return;
655 	}
656 	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8,
657 			lower_32_bits(page_table_base));
658 }
659 
660 const struct kfd2kgd_calls gfx_v8_kfd2kgd = {
661 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
662 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
663 	.init_interrupts = kgd_init_interrupts,
664 	.hqd_load = kgd_hqd_load,
665 	.hqd_sdma_load = kgd_hqd_sdma_load,
666 	.hqd_dump = kgd_hqd_dump,
667 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
668 	.hqd_is_occupied = kgd_hqd_is_occupied,
669 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
670 	.hqd_destroy = kgd_hqd_destroy,
671 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
672 	.address_watch_disable = kgd_address_watch_disable,
673 	.address_watch_execute = kgd_address_watch_execute,
674 	.wave_control_execute = kgd_wave_control_execute,
675 	.address_watch_get_offset = kgd_address_watch_get_offset,
676 	.get_atc_vmid_pasid_mapping_info =
677 			get_atc_vmid_pasid_mapping_info,
678 	.set_scratch_backing_va = set_scratch_backing_va,
679 	.get_tile_config = get_tile_config,
680 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
681 };
682