1 /*
2  * Copyright 2014-2018 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 #define pr_fmt(fmt) "kfd2kgd: " fmt
24 
25 #include <linux/module.h>
26 #include <linux/fdtable.h>
27 #include <linux/uaccess.h>
28 #include <linux/mmu_context.h>
29 
30 #include "amdgpu.h"
31 #include "amdgpu_amdkfd.h"
32 #include "soc15_hw_ip.h"
33 #include "gc/gc_9_0_offset.h"
34 #include "gc/gc_9_0_sh_mask.h"
35 #include "vega10_enum.h"
36 #include "sdma0/sdma0_4_0_offset.h"
37 #include "sdma0/sdma0_4_0_sh_mask.h"
38 #include "sdma1/sdma1_4_0_offset.h"
39 #include "sdma1/sdma1_4_0_sh_mask.h"
40 #include "athub/athub_1_0_offset.h"
41 #include "athub/athub_1_0_sh_mask.h"
42 #include "oss/osssys_4_0_offset.h"
43 #include "oss/osssys_4_0_sh_mask.h"
44 #include "soc15_common.h"
45 #include "v9_structs.h"
46 #include "soc15.h"
47 #include "soc15d.h"
48 #include "mmhub_v1_0.h"
49 #include "gfxhub_v1_0.h"
50 #include "gmc_v9_0.h"
51 
52 
53 #define V9_PIPE_PER_MEC		(4)
54 #define V9_QUEUES_PER_PIPE_MEC	(8)
55 
56 enum hqd_dequeue_request_type {
57 	NO_ACTION = 0,
58 	DRAIN_PIPE,
59 	RESET_WAVES
60 };
61 
62 
63 /* Because of REG_GET_FIELD() being used, we put this function in the
64  * asic specific file.
65  */
66 int kgd_gfx_v9_get_tile_config(struct kgd_dev *kgd,
67 		struct tile_config *config)
68 {
69 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
70 
71 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
72 
73 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
74 	config->num_tile_configs =
75 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
76 	config->macro_tile_config_ptr =
77 			adev->gfx.config.macrotile_mode_array;
78 	config->num_macro_tile_configs =
79 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
80 
81 	return 0;
82 }
83 
84 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
85 {
86 	return (struct amdgpu_device *)kgd;
87 }
88 
89 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
90 			uint32_t queue, uint32_t vmid)
91 {
92 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
93 
94 	mutex_lock(&adev->srbm_mutex);
95 	soc15_grbm_select(adev, mec, pipe, queue, vmid);
96 }
97 
98 static void unlock_srbm(struct kgd_dev *kgd)
99 {
100 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
101 
102 	soc15_grbm_select(adev, 0, 0, 0, 0);
103 	mutex_unlock(&adev->srbm_mutex);
104 }
105 
106 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
107 				uint32_t queue_id)
108 {
109 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
110 
111 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
112 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
113 
114 	lock_srbm(kgd, mec, pipe, queue_id, 0);
115 }
116 
117 static uint32_t get_queue_mask(struct amdgpu_device *adev,
118 			       uint32_t pipe_id, uint32_t queue_id)
119 {
120 	unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe +
121 			    queue_id) & 31;
122 
123 	return ((uint32_t)1) << bit;
124 }
125 
126 static void release_queue(struct kgd_dev *kgd)
127 {
128 	unlock_srbm(kgd);
129 }
130 
131 void kgd_gfx_v9_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
132 					uint32_t sh_mem_config,
133 					uint32_t sh_mem_ape1_base,
134 					uint32_t sh_mem_ape1_limit,
135 					uint32_t sh_mem_bases)
136 {
137 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
138 
139 	lock_srbm(kgd, 0, 0, 0, vmid);
140 
141 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
142 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
143 	/* APE1 no longer exists on GFX9 */
144 
145 	unlock_srbm(kgd);
146 }
147 
148 int kgd_gfx_v9_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
149 					unsigned int vmid)
150 {
151 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
152 
153 	/*
154 	 * We have to assume that there is no outstanding mapping.
155 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
156 	 * a mapping is in progress or because a mapping finished
157 	 * and the SW cleared it.
158 	 * So the protocol is to always wait & clear.
159 	 */
160 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
161 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
162 
163 	/*
164 	 * need to do this twice, once for gfx and once for mmhub
165 	 * for ATC add 16 to VMID for mmhub, for IH different registers.
166 	 * ATC_VMID0..15 registers are separate from ATC_VMID16..31.
167 	 */
168 
169 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
170 	       pasid_mapping);
171 
172 	while (!(RREG32(SOC15_REG_OFFSET(
173 				ATHUB, 0,
174 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
175 		 (1U << vmid)))
176 		cpu_relax();
177 
178 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
179 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
180 	       1U << vmid);
181 
182 	/* Mapping vmid to pasid also for IH block */
183 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
184 	       pasid_mapping);
185 
186 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID16_PASID_MAPPING) + vmid,
187 	       pasid_mapping);
188 
189 	while (!(RREG32(SOC15_REG_OFFSET(
190 				ATHUB, 0,
191 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
192 		 (1U << (vmid + 16))))
193 		cpu_relax();
194 
195 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
196 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
197 	       1U << (vmid + 16));
198 
199 	/* Mapping vmid to pasid also for IH block */
200 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid,
201 	       pasid_mapping);
202 	return 0;
203 }
204 
205 /* TODO - RING0 form of field is obsolete, seems to date back to SI
206  * but still works
207  */
208 
209 int kgd_gfx_v9_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
210 {
211 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
212 	uint32_t mec;
213 	uint32_t pipe;
214 
215 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
216 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
217 
218 	lock_srbm(kgd, mec, pipe, 0, 0);
219 
220 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
221 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
222 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
223 
224 	unlock_srbm(kgd);
225 
226 	return 0;
227 }
228 
229 static uint32_t get_sdma_base_addr(struct amdgpu_device *adev,
230 				unsigned int engine_id,
231 				unsigned int queue_id)
232 {
233 	uint32_t base[2] = {
234 		SOC15_REG_OFFSET(SDMA0, 0,
235 				 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
236 		SOC15_REG_OFFSET(SDMA1, 0,
237 				 mmSDMA1_RLC0_RB_CNTL) - mmSDMA1_RLC0_RB_CNTL
238 	};
239 	uint32_t retval;
240 
241 	retval = base[engine_id] + queue_id * (mmSDMA0_RLC1_RB_CNTL -
242 					       mmSDMA0_RLC0_RB_CNTL);
243 
244 	pr_debug("sdma base address: 0x%x\n", retval);
245 
246 	return retval;
247 }
248 
249 static inline struct v9_mqd *get_mqd(void *mqd)
250 {
251 	return (struct v9_mqd *)mqd;
252 }
253 
254 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
255 {
256 	return (struct v9_sdma_mqd *)mqd;
257 }
258 
259 int kgd_gfx_v9_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
260 			uint32_t queue_id, uint32_t __user *wptr,
261 			uint32_t wptr_shift, uint32_t wptr_mask,
262 			struct mm_struct *mm)
263 {
264 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
265 	struct v9_mqd *m;
266 	uint32_t *mqd_hqd;
267 	uint32_t reg, hqd_base, data;
268 
269 	m = get_mqd(mqd);
270 
271 	acquire_queue(kgd, pipe_id, queue_id);
272 
273 	/* HIQ is set during driver init period with vmid set to 0*/
274 	if (m->cp_hqd_vmid == 0) {
275 		uint32_t value, mec, pipe;
276 
277 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
278 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
279 
280 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
281 			mec, pipe, queue_id);
282 		value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS));
283 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
284 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
285 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value);
286 	}
287 
288 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
289 	mqd_hqd = &m->cp_mqd_base_addr_lo;
290 	hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
291 
292 	for (reg = hqd_base;
293 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
294 		WREG32_RLC(reg, mqd_hqd[reg - hqd_base]);
295 
296 
297 	/* Activate doorbell logic before triggering WPTR poll. */
298 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
299 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
300 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
301 
302 	if (wptr) {
303 		/* Don't read wptr with get_user because the user
304 		 * context may not be accessible (if this function
305 		 * runs in a work queue). Instead trigger a one-shot
306 		 * polling read from memory in the CP. This assumes
307 		 * that wptr is GPU-accessible in the queue's VMID via
308 		 * ATC or SVM. WPTR==RPTR before starting the poll so
309 		 * the CP starts fetching new commands from the right
310 		 * place.
311 		 *
312 		 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
313 		 * tricky. Assume that the queue didn't overflow. The
314 		 * number of valid bits in the 32-bit RPTR depends on
315 		 * the queue size. The remaining bits are taken from
316 		 * the saved 64-bit WPTR. If the WPTR wrapped, add the
317 		 * queue size.
318 		 */
319 		uint32_t queue_size =
320 			2 << REG_GET_FIELD(m->cp_hqd_pq_control,
321 					   CP_HQD_PQ_CONTROL, QUEUE_SIZE);
322 		uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
323 
324 		if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
325 			guessed_wptr += queue_size;
326 		guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
327 		guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
328 
329 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
330 		       lower_32_bits(guessed_wptr));
331 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
332 		       upper_32_bits(guessed_wptr));
333 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
334 		       lower_32_bits((uintptr_t)wptr));
335 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
336 		       upper_32_bits((uintptr_t)wptr));
337 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
338 		       get_queue_mask(adev, pipe_id, queue_id));
339 	}
340 
341 	/* Start the EOP fetcher */
342 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
343 	       REG_SET_FIELD(m->cp_hqd_eop_rptr,
344 			     CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
345 
346 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
347 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
348 
349 	release_queue(kgd);
350 
351 	return 0;
352 }
353 
354 int kgd_gfx_v9_hqd_dump(struct kgd_dev *kgd,
355 			uint32_t pipe_id, uint32_t queue_id,
356 			uint32_t (**dump)[2], uint32_t *n_regs)
357 {
358 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
359 	uint32_t i = 0, reg;
360 #define HQD_N_REGS 56
361 #define DUMP_REG(addr) do {				\
362 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
363 			break;				\
364 		(*dump)[i][0] = (addr) << 2;		\
365 		(*dump)[i++][1] = RREG32(addr);		\
366 	} while (0)
367 
368 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
369 	if (*dump == NULL)
370 		return -ENOMEM;
371 
372 	acquire_queue(kgd, pipe_id, queue_id);
373 
374 	for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
375 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
376 		DUMP_REG(reg);
377 
378 	release_queue(kgd);
379 
380 	WARN_ON_ONCE(i != HQD_N_REGS);
381 	*n_regs = i;
382 
383 	return 0;
384 }
385 
386 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
387 			     uint32_t __user *wptr, struct mm_struct *mm)
388 {
389 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
390 	struct v9_sdma_mqd *m;
391 	uint32_t sdma_base_addr, sdmax_gfx_context_cntl;
392 	unsigned long end_jiffies;
393 	uint32_t data;
394 	uint64_t data64;
395 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
396 
397 	m = get_sdma_mqd(mqd);
398 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
399 					    m->sdma_queue_id);
400 	sdmax_gfx_context_cntl = m->sdma_engine_id ?
401 		SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_CONTEXT_CNTL) :
402 		SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_CONTEXT_CNTL);
403 
404 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
405 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
406 
407 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
408 	while (true) {
409 		data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
410 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
411 			break;
412 		if (time_after(jiffies, end_jiffies))
413 			return -ETIME;
414 		usleep_range(500, 1000);
415 	}
416 	data = RREG32(sdmax_gfx_context_cntl);
417 	data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
418 			     RESUME_CTX, 0);
419 	WREG32(sdmax_gfx_context_cntl, data);
420 
421 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL_OFFSET,
422 	       m->sdmax_rlcx_doorbell_offset);
423 
424 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
425 			     ENABLE, 1);
426 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
427 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
428 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI,
429 				m->sdmax_rlcx_rb_rptr_hi);
430 
431 	WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
432 	if (read_user_wptr(mm, wptr64, data64)) {
433 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
434 		       lower_32_bits(data64));
435 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
436 		       upper_32_bits(data64));
437 	} else {
438 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
439 		       m->sdmax_rlcx_rb_rptr);
440 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
441 		       m->sdmax_rlcx_rb_rptr_hi);
442 	}
443 	WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
444 
445 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
446 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
447 			m->sdmax_rlcx_rb_base_hi);
448 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
449 			m->sdmax_rlcx_rb_rptr_addr_lo);
450 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
451 			m->sdmax_rlcx_rb_rptr_addr_hi);
452 
453 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
454 			     RB_ENABLE, 1);
455 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
456 
457 	return 0;
458 }
459 
460 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
461 			     uint32_t engine_id, uint32_t queue_id,
462 			     uint32_t (**dump)[2], uint32_t *n_regs)
463 {
464 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
465 	uint32_t sdma_base_addr = get_sdma_base_addr(adev, engine_id, queue_id);
466 	uint32_t i = 0, reg;
467 #undef HQD_N_REGS
468 #define HQD_N_REGS (19+6+7+10)
469 
470 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
471 	if (*dump == NULL)
472 		return -ENOMEM;
473 
474 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
475 		DUMP_REG(sdma_base_addr + reg);
476 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
477 		DUMP_REG(sdma_base_addr + reg);
478 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
479 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
480 		DUMP_REG(sdma_base_addr + reg);
481 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
482 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
483 		DUMP_REG(sdma_base_addr + reg);
484 
485 	WARN_ON_ONCE(i != HQD_N_REGS);
486 	*n_regs = i;
487 
488 	return 0;
489 }
490 
491 bool kgd_gfx_v9_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
492 				uint32_t pipe_id, uint32_t queue_id)
493 {
494 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
495 	uint32_t act;
496 	bool retval = false;
497 	uint32_t low, high;
498 
499 	acquire_queue(kgd, pipe_id, queue_id);
500 	act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
501 	if (act) {
502 		low = lower_32_bits(queue_address >> 8);
503 		high = upper_32_bits(queue_address >> 8);
504 
505 		if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
506 		   high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
507 			retval = true;
508 	}
509 	release_queue(kgd);
510 	return retval;
511 }
512 
513 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
514 {
515 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
516 	struct v9_sdma_mqd *m;
517 	uint32_t sdma_base_addr;
518 	uint32_t sdma_rlc_rb_cntl;
519 
520 	m = get_sdma_mqd(mqd);
521 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
522 					    m->sdma_queue_id);
523 
524 	sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
525 
526 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
527 		return true;
528 
529 	return false;
530 }
531 
532 int kgd_gfx_v9_hqd_destroy(struct kgd_dev *kgd, void *mqd,
533 				enum kfd_preempt_type reset_type,
534 				unsigned int utimeout, uint32_t pipe_id,
535 				uint32_t queue_id)
536 {
537 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
538 	enum hqd_dequeue_request_type type;
539 	unsigned long end_jiffies;
540 	uint32_t temp;
541 	struct v9_mqd *m = get_mqd(mqd);
542 
543 	if (adev->in_gpu_reset)
544 		return -EIO;
545 
546 	acquire_queue(kgd, pipe_id, queue_id);
547 
548 	if (m->cp_hqd_vmid == 0)
549 		WREG32_FIELD15_RLC(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
550 
551 	switch (reset_type) {
552 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
553 		type = DRAIN_PIPE;
554 		break;
555 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
556 		type = RESET_WAVES;
557 		break;
558 	default:
559 		type = DRAIN_PIPE;
560 		break;
561 	}
562 
563 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
564 
565 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
566 	while (true) {
567 		temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
568 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
569 			break;
570 		if (time_after(jiffies, end_jiffies)) {
571 			pr_err("cp queue preemption time out.\n");
572 			release_queue(kgd);
573 			return -ETIME;
574 		}
575 		usleep_range(500, 1000);
576 	}
577 
578 	release_queue(kgd);
579 	return 0;
580 }
581 
582 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
583 				unsigned int utimeout)
584 {
585 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
586 	struct v9_sdma_mqd *m;
587 	uint32_t sdma_base_addr;
588 	uint32_t temp;
589 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
590 
591 	m = get_sdma_mqd(mqd);
592 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
593 					    m->sdma_queue_id);
594 
595 	temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
596 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
597 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
598 
599 	while (true) {
600 		temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
601 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
602 			break;
603 		if (time_after(jiffies, end_jiffies))
604 			return -ETIME;
605 		usleep_range(500, 1000);
606 	}
607 
608 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
609 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
610 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
611 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
612 
613 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
614 	m->sdmax_rlcx_rb_rptr_hi =
615 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI);
616 
617 	return 0;
618 }
619 
620 bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
621 							uint8_t vmid)
622 {
623 	uint32_t reg;
624 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
625 
626 	reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
627 		     + vmid);
628 	return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
629 }
630 
631 uint16_t kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
632 								uint8_t vmid)
633 {
634 	uint32_t reg;
635 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
636 
637 	reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
638 		     + vmid);
639 	return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
640 }
641 
642 static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid,
643 			uint32_t flush_type)
644 {
645 	signed long r;
646 	uint32_t seq;
647 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
648 
649 	spin_lock(&adev->gfx.kiq.ring_lock);
650 	amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/
651 	amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
652 	amdgpu_ring_write(ring,
653 			PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
654 			PACKET3_INVALIDATE_TLBS_ALL_HUB(1) |
655 			PACKET3_INVALIDATE_TLBS_PASID(pasid) |
656 			PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(flush_type));
657 	amdgpu_fence_emit_polling(ring, &seq);
658 	amdgpu_ring_commit(ring);
659 	spin_unlock(&adev->gfx.kiq.ring_lock);
660 
661 	r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
662 	if (r < 1) {
663 		DRM_ERROR("wait for kiq fence error: %ld.\n", r);
664 		return -ETIME;
665 	}
666 
667 	return 0;
668 }
669 
670 int kgd_gfx_v9_invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
671 {
672 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
673 	int vmid, i;
674 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
675 	uint32_t flush_type = 0;
676 
677 	if (adev->in_gpu_reset)
678 		return -EIO;
679 	if (adev->gmc.xgmi.num_physical_nodes &&
680 		adev->asic_type == CHIP_VEGA20)
681 		flush_type = 2;
682 
683 	if (ring->sched.ready)
684 		return invalidate_tlbs_with_kiq(adev, pasid, flush_type);
685 
686 	for (vmid = 0; vmid < 16; vmid++) {
687 		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
688 			continue;
689 		if (kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
690 			if (kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
691 				== pasid) {
692 				for (i = 0; i < adev->num_vmhubs; i++)
693 					amdgpu_gmc_flush_gpu_tlb(adev, vmid,
694 								i, flush_type);
695 				break;
696 			}
697 		}
698 	}
699 
700 	return 0;
701 }
702 
703 int kgd_gfx_v9_invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
704 {
705 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
706 	int i;
707 
708 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
709 		pr_err("non kfd vmid %d\n", vmid);
710 		return 0;
711 	}
712 
713 	/* Use legacy mode tlb invalidation.
714 	 *
715 	 * Currently on Raven the code below is broken for anything but
716 	 * legacy mode due to a MMHUB power gating problem. A workaround
717 	 * is for MMHUB to wait until the condition PER_VMID_INVALIDATE_REQ
718 	 * == PER_VMID_INVALIDATE_ACK instead of simply waiting for the ack
719 	 * bit.
720 	 *
721 	 * TODO 1: agree on the right set of invalidation registers for
722 	 * KFD use. Use the last one for now. Invalidate both GC and
723 	 * MMHUB.
724 	 *
725 	 * TODO 2: support range-based invalidation, requires kfg2kgd
726 	 * interface change
727 	 */
728 	for (i = 0; i < adev->num_vmhubs; i++)
729 		amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
730 
731 	return 0;
732 }
733 
734 int kgd_gfx_v9_address_watch_disable(struct kgd_dev *kgd)
735 {
736 	return 0;
737 }
738 
739 int kgd_gfx_v9_address_watch_execute(struct kgd_dev *kgd,
740 					unsigned int watch_point_id,
741 					uint32_t cntl_val,
742 					uint32_t addr_hi,
743 					uint32_t addr_lo)
744 {
745 	return 0;
746 }
747 
748 int kgd_gfx_v9_wave_control_execute(struct kgd_dev *kgd,
749 					uint32_t gfx_index_val,
750 					uint32_t sq_cmd)
751 {
752 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
753 	uint32_t data = 0;
754 
755 	mutex_lock(&adev->grbm_idx_mutex);
756 
757 	WREG32_SOC15_RLC_SHADOW(GC, 0, mmGRBM_GFX_INDEX, gfx_index_val);
758 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
759 
760 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
761 		INSTANCE_BROADCAST_WRITES, 1);
762 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
763 		SH_BROADCAST_WRITES, 1);
764 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
765 		SE_BROADCAST_WRITES, 1);
766 
767 	WREG32_SOC15_RLC_SHADOW(GC, 0, mmGRBM_GFX_INDEX, data);
768 	mutex_unlock(&adev->grbm_idx_mutex);
769 
770 	return 0;
771 }
772 
773 uint32_t kgd_gfx_v9_address_watch_get_offset(struct kgd_dev *kgd,
774 					unsigned int watch_point_id,
775 					unsigned int reg_offset)
776 {
777 	return 0;
778 }
779 
780 void kgd_gfx_v9_set_scratch_backing_va(struct kgd_dev *kgd,
781 					uint64_t va, uint32_t vmid)
782 {
783 	/* No longer needed on GFXv9. The scratch base address is
784 	 * passed to the shader by the CP. It's the user mode driver's
785 	 * responsibility.
786 	 */
787 }
788 
789 void kgd_gfx_v9_set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
790 		uint64_t page_table_base)
791 {
792 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
793 
794 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
795 		pr_err("trying to set page table base for wrong VMID %u\n",
796 		       vmid);
797 		return;
798 	}
799 
800 	/* TODO: take advantage of per-process address space size. For
801 	 * now, all processes share the same address space size, like
802 	 * on GFX8 and older.
803 	 */
804 	if (adev->asic_type == CHIP_ARCTURUS) {
805 		/* Two MMHUBs */
806 		mmhub_v9_4_setup_vm_pt_regs(adev, 0, vmid, page_table_base);
807 		mmhub_v9_4_setup_vm_pt_regs(adev, 1, vmid, page_table_base);
808 	} else
809 		mmhub_v1_0_setup_vm_pt_regs(adev, vmid, page_table_base);
810 
811 	gfxhub_v1_0_setup_vm_pt_regs(adev, vmid, page_table_base);
812 }
813 
814 static const struct kfd2kgd_calls kfd2kgd = {
815 	.program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
816 	.set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping,
817 	.init_interrupts = kgd_gfx_v9_init_interrupts,
818 	.hqd_load = kgd_gfx_v9_hqd_load,
819 	.hqd_sdma_load = kgd_hqd_sdma_load,
820 	.hqd_dump = kgd_gfx_v9_hqd_dump,
821 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
822 	.hqd_is_occupied = kgd_gfx_v9_hqd_is_occupied,
823 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
824 	.hqd_destroy = kgd_gfx_v9_hqd_destroy,
825 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
826 	.address_watch_disable = kgd_gfx_v9_address_watch_disable,
827 	.address_watch_execute = kgd_gfx_v9_address_watch_execute,
828 	.wave_control_execute = kgd_gfx_v9_wave_control_execute,
829 	.address_watch_get_offset = kgd_gfx_v9_address_watch_get_offset,
830 	.get_atc_vmid_pasid_mapping_pasid =
831 			kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid,
832 	.get_atc_vmid_pasid_mapping_valid =
833 			kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid,
834 	.set_scratch_backing_va = kgd_gfx_v9_set_scratch_backing_va,
835 	.get_tile_config = kgd_gfx_v9_get_tile_config,
836 	.set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base,
837 	.invalidate_tlbs = kgd_gfx_v9_invalidate_tlbs,
838 	.invalidate_tlbs_vmid = kgd_gfx_v9_invalidate_tlbs_vmid,
839 	.get_hive_id = amdgpu_amdkfd_get_hive_id,
840 };
841 
842 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_9_0_get_functions(void)
843 {
844 	return (struct kfd2kgd_calls *)&kfd2kgd;
845 }
846