xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 #include "amdgpu.h"
23 #include "amdgpu_amdkfd.h"
24 #include "gc/gc_9_0_offset.h"
25 #include "gc/gc_9_0_sh_mask.h"
26 #include "vega10_enum.h"
27 #include "sdma0/sdma0_4_0_offset.h"
28 #include "sdma0/sdma0_4_0_sh_mask.h"
29 #include "sdma1/sdma1_4_0_offset.h"
30 #include "sdma1/sdma1_4_0_sh_mask.h"
31 #include "athub/athub_1_0_offset.h"
32 #include "athub/athub_1_0_sh_mask.h"
33 #include "oss/osssys_4_0_offset.h"
34 #include "oss/osssys_4_0_sh_mask.h"
35 #include "soc15_common.h"
36 #include "v9_structs.h"
37 #include "soc15.h"
38 #include "soc15d.h"
39 #include "gfx_v9_0.h"
40 #include "amdgpu_amdkfd_gfx_v9.h"
41 #include <uapi/linux/kfd_ioctl.h>
42 
43 enum hqd_dequeue_request_type {
44 	NO_ACTION = 0,
45 	DRAIN_PIPE,
46 	RESET_WAVES,
47 	SAVE_WAVES
48 };
49 
50 static void kgd_gfx_v9_lock_srbm(struct amdgpu_device *adev, uint32_t mec, uint32_t pipe,
51 			uint32_t queue, uint32_t vmid, uint32_t inst)
52 {
53 	mutex_lock(&adev->srbm_mutex);
54 	soc15_grbm_select(adev, mec, pipe, queue, vmid, GET_INST(GC, inst));
55 }
56 
57 static void kgd_gfx_v9_unlock_srbm(struct amdgpu_device *adev, uint32_t inst)
58 {
59 	soc15_grbm_select(adev, 0, 0, 0, 0, GET_INST(GC, inst));
60 	mutex_unlock(&adev->srbm_mutex);
61 }
62 
63 void kgd_gfx_v9_acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id,
64 				uint32_t queue_id, uint32_t inst)
65 {
66 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
67 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
68 
69 	kgd_gfx_v9_lock_srbm(adev, mec, pipe, queue_id, 0, inst);
70 }
71 
72 uint64_t kgd_gfx_v9_get_queue_mask(struct amdgpu_device *adev,
73 			       uint32_t pipe_id, uint32_t queue_id)
74 {
75 	unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe +
76 			queue_id;
77 
78 	return 1ull << bit;
79 }
80 
81 void kgd_gfx_v9_release_queue(struct amdgpu_device *adev, uint32_t inst)
82 {
83 	kgd_gfx_v9_unlock_srbm(adev, inst);
84 }
85 
86 void kgd_gfx_v9_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmid,
87 					uint32_t sh_mem_config,
88 					uint32_t sh_mem_ape1_base,
89 					uint32_t sh_mem_ape1_limit,
90 					uint32_t sh_mem_bases, uint32_t inst)
91 {
92 	kgd_gfx_v9_lock_srbm(adev, 0, 0, 0, vmid, inst);
93 
94 	WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmSH_MEM_CONFIG, sh_mem_config);
95 	WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmSH_MEM_BASES, sh_mem_bases);
96 	/* APE1 no longer exists on GFX9 */
97 
98 	kgd_gfx_v9_unlock_srbm(adev, inst);
99 }
100 
101 int kgd_gfx_v9_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid,
102 					unsigned int vmid, uint32_t inst)
103 {
104 	/*
105 	 * We have to assume that there is no outstanding mapping.
106 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
107 	 * a mapping is in progress or because a mapping finished
108 	 * and the SW cleared it.
109 	 * So the protocol is to always wait & clear.
110 	 */
111 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
112 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
113 
114 	/*
115 	 * need to do this twice, once for gfx and once for mmhub
116 	 * for ATC add 16 to VMID for mmhub, for IH different registers.
117 	 * ATC_VMID0..15 registers are separate from ATC_VMID16..31.
118 	 */
119 
120 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
121 	       pasid_mapping);
122 
123 	while (!(RREG32(SOC15_REG_OFFSET(
124 				ATHUB, 0,
125 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
126 		 (1U << vmid)))
127 		cpu_relax();
128 
129 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
130 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
131 	       1U << vmid);
132 
133 	/* Mapping vmid to pasid also for IH block */
134 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
135 	       pasid_mapping);
136 
137 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID16_PASID_MAPPING) + vmid,
138 	       pasid_mapping);
139 
140 	while (!(RREG32(SOC15_REG_OFFSET(
141 				ATHUB, 0,
142 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
143 		 (1U << (vmid + 16))))
144 		cpu_relax();
145 
146 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
147 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
148 	       1U << (vmid + 16));
149 
150 	/* Mapping vmid to pasid also for IH block */
151 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid,
152 	       pasid_mapping);
153 	return 0;
154 }
155 
156 /* TODO - RING0 form of field is obsolete, seems to date back to SI
157  * but still works
158  */
159 
160 int kgd_gfx_v9_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id,
161 				uint32_t inst)
162 {
163 	uint32_t mec;
164 	uint32_t pipe;
165 
166 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
167 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
168 
169 	kgd_gfx_v9_lock_srbm(adev, mec, pipe, 0, 0, inst);
170 
171 	WREG32_SOC15(GC, GET_INST(GC, inst), mmCPC_INT_CNTL,
172 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
173 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
174 
175 	kgd_gfx_v9_unlock_srbm(adev, inst);
176 
177 	return 0;
178 }
179 
180 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
181 				unsigned int engine_id,
182 				unsigned int queue_id)
183 {
184 	uint32_t sdma_engine_reg_base = 0;
185 	uint32_t sdma_rlc_reg_offset;
186 
187 	switch (engine_id) {
188 	default:
189 		dev_warn(adev->dev,
190 			 "Invalid sdma engine id (%d), using engine id 0\n",
191 			 engine_id);
192 		fallthrough;
193 	case 0:
194 		sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA0, 0,
195 				mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL;
196 		break;
197 	case 1:
198 		sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA1, 0,
199 				mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL;
200 		break;
201 	}
202 
203 	sdma_rlc_reg_offset = sdma_engine_reg_base
204 		+ queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
205 
206 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
207 		 queue_id, sdma_rlc_reg_offset);
208 
209 	return sdma_rlc_reg_offset;
210 }
211 
212 static inline struct v9_mqd *get_mqd(void *mqd)
213 {
214 	return (struct v9_mqd *)mqd;
215 }
216 
217 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
218 {
219 	return (struct v9_sdma_mqd *)mqd;
220 }
221 
222 int kgd_gfx_v9_hqd_load(struct amdgpu_device *adev, void *mqd,
223 			uint32_t pipe_id, uint32_t queue_id,
224 			uint32_t __user *wptr, uint32_t wptr_shift,
225 			uint32_t wptr_mask, struct mm_struct *mm,
226 			uint32_t inst)
227 {
228 	struct v9_mqd *m;
229 	uint32_t *mqd_hqd;
230 	uint32_t reg, hqd_base, data;
231 
232 	m = get_mqd(mqd);
233 
234 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
235 
236 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
237 	mqd_hqd = &m->cp_mqd_base_addr_lo;
238 	hqd_base = SOC15_REG_OFFSET(GC, GET_INST(GC, inst), mmCP_MQD_BASE_ADDR);
239 
240 	for (reg = hqd_base;
241 	     reg <= SOC15_REG_OFFSET(GC, GET_INST(GC, inst), mmCP_HQD_PQ_WPTR_HI); reg++)
242 		WREG32_XCC(reg, mqd_hqd[reg - hqd_base], inst);
243 
244 
245 	/* Activate doorbell logic before triggering WPTR poll. */
246 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
247 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
248 	WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_HQD_PQ_DOORBELL_CONTROL, data);
249 
250 	if (wptr) {
251 		/* Don't read wptr with get_user because the user
252 		 * context may not be accessible (if this function
253 		 * runs in a work queue). Instead trigger a one-shot
254 		 * polling read from memory in the CP. This assumes
255 		 * that wptr is GPU-accessible in the queue's VMID via
256 		 * ATC or SVM. WPTR==RPTR before starting the poll so
257 		 * the CP starts fetching new commands from the right
258 		 * place.
259 		 *
260 		 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
261 		 * tricky. Assume that the queue didn't overflow. The
262 		 * number of valid bits in the 32-bit RPTR depends on
263 		 * the queue size. The remaining bits are taken from
264 		 * the saved 64-bit WPTR. If the WPTR wrapped, add the
265 		 * queue size.
266 		 */
267 		uint32_t queue_size =
268 			2 << REG_GET_FIELD(m->cp_hqd_pq_control,
269 					   CP_HQD_PQ_CONTROL, QUEUE_SIZE);
270 		uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
271 
272 		if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
273 			guessed_wptr += queue_size;
274 		guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
275 		guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
276 
277 		WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_HQD_PQ_WPTR_LO,
278 			lower_32_bits(guessed_wptr));
279 		WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_HQD_PQ_WPTR_HI,
280 			upper_32_bits(guessed_wptr));
281 		WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_HQD_PQ_WPTR_POLL_ADDR,
282 			lower_32_bits((uintptr_t)wptr));
283 		WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_HQD_PQ_WPTR_POLL_ADDR_HI,
284 			upper_32_bits((uintptr_t)wptr));
285 		WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_PQ_WPTR_POLL_CNTL1,
286 			(uint32_t)kgd_gfx_v9_get_queue_mask(adev, pipe_id, queue_id));
287 	}
288 
289 	/* Start the EOP fetcher */
290 	WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_HQD_EOP_RPTR,
291 	       REG_SET_FIELD(m->cp_hqd_eop_rptr, CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
292 
293 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
294 	WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE, data);
295 
296 	kgd_gfx_v9_release_queue(adev, inst);
297 
298 	return 0;
299 }
300 
301 int kgd_gfx_v9_hiq_mqd_load(struct amdgpu_device *adev, void *mqd,
302 			    uint32_t pipe_id, uint32_t queue_id,
303 			    uint32_t doorbell_off, uint32_t inst)
304 {
305 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[inst].ring;
306 	struct v9_mqd *m;
307 	uint32_t mec, pipe;
308 	int r;
309 
310 	m = get_mqd(mqd);
311 
312 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
313 
314 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
315 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
316 
317 	pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
318 		 mec, pipe, queue_id);
319 
320 	spin_lock(&adev->gfx.kiq[inst].ring_lock);
321 	r = amdgpu_ring_alloc(kiq_ring, 7);
322 	if (r) {
323 		pr_err("Failed to alloc KIQ (%d).\n", r);
324 		goto out_unlock;
325 	}
326 
327 	amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
328 	amdgpu_ring_write(kiq_ring,
329 			  PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
330 			  PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */
331 			  PACKET3_MAP_QUEUES_QUEUE(queue_id) |
332 			  PACKET3_MAP_QUEUES_PIPE(pipe) |
333 			  PACKET3_MAP_QUEUES_ME((mec - 1)) |
334 			  PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
335 			  PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
336 			  PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */
337 			  PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
338 	amdgpu_ring_write(kiq_ring,
339 			  PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off));
340 	amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo);
341 	amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi);
342 	amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo);
343 	amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi);
344 	amdgpu_ring_commit(kiq_ring);
345 
346 out_unlock:
347 	spin_unlock(&adev->gfx.kiq[inst].ring_lock);
348 	kgd_gfx_v9_release_queue(adev, inst);
349 
350 	return r;
351 }
352 
353 int kgd_gfx_v9_hqd_dump(struct amdgpu_device *adev,
354 			uint32_t pipe_id, uint32_t queue_id,
355 			uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst)
356 {
357 	uint32_t i = 0, reg;
358 #define HQD_N_REGS 56
359 #define DUMP_REG(addr) do {				\
360 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
361 			break;				\
362 		(*dump)[i][0] = (addr) << 2;		\
363 		(*dump)[i++][1] = RREG32(addr);		\
364 	} while (0)
365 
366 	*dump = kmalloc_array(HQD_N_REGS, sizeof(**dump), GFP_KERNEL);
367 	if (*dump == NULL)
368 		return -ENOMEM;
369 
370 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
371 
372 	for (reg = SOC15_REG_OFFSET(GC, GET_INST(GC, inst), mmCP_MQD_BASE_ADDR);
373 	     reg <= SOC15_REG_OFFSET(GC, GET_INST(GC, inst), mmCP_HQD_PQ_WPTR_HI); reg++)
374 		DUMP_REG(reg);
375 
376 	kgd_gfx_v9_release_queue(adev, inst);
377 
378 	WARN_ON_ONCE(i != HQD_N_REGS);
379 	*n_regs = i;
380 
381 	return 0;
382 }
383 
384 static int kgd_hqd_sdma_load(struct amdgpu_device *adev, void *mqd,
385 			     uint32_t __user *wptr, struct mm_struct *mm)
386 {
387 	struct v9_sdma_mqd *m;
388 	uint32_t sdma_rlc_reg_offset;
389 	unsigned long end_jiffies;
390 	uint32_t data;
391 	uint64_t data64;
392 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
393 
394 	m = get_sdma_mqd(mqd);
395 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
396 					    m->sdma_queue_id);
397 
398 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
399 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
400 
401 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
402 	while (true) {
403 		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
404 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
405 			break;
406 		if (time_after(jiffies, end_jiffies)) {
407 			pr_err("SDMA RLC not idle in %s\n", __func__);
408 			return -ETIME;
409 		}
410 		usleep_range(500, 1000);
411 	}
412 
413 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
414 	       m->sdmax_rlcx_doorbell_offset);
415 
416 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
417 			     ENABLE, 1);
418 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
419 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
420 				m->sdmax_rlcx_rb_rptr);
421 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
422 				m->sdmax_rlcx_rb_rptr_hi);
423 
424 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
425 	if (read_user_wptr(mm, wptr64, data64)) {
426 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
427 		       lower_32_bits(data64));
428 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
429 		       upper_32_bits(data64));
430 	} else {
431 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
432 		       m->sdmax_rlcx_rb_rptr);
433 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
434 		       m->sdmax_rlcx_rb_rptr_hi);
435 	}
436 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
437 
438 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
439 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
440 			m->sdmax_rlcx_rb_base_hi);
441 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
442 			m->sdmax_rlcx_rb_rptr_addr_lo);
443 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
444 			m->sdmax_rlcx_rb_rptr_addr_hi);
445 
446 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
447 			     RB_ENABLE, 1);
448 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
449 
450 	return 0;
451 }
452 
453 static int kgd_hqd_sdma_dump(struct amdgpu_device *adev,
454 			     uint32_t engine_id, uint32_t queue_id,
455 			     uint32_t (**dump)[2], uint32_t *n_regs)
456 {
457 	uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
458 			engine_id, queue_id);
459 	uint32_t i = 0, reg;
460 #undef HQD_N_REGS
461 #define HQD_N_REGS (19+6+7+10)
462 
463 	*dump = kmalloc_array(HQD_N_REGS, sizeof(**dump), GFP_KERNEL);
464 	if (*dump == NULL)
465 		return -ENOMEM;
466 
467 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
468 		DUMP_REG(sdma_rlc_reg_offset + reg);
469 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
470 		DUMP_REG(sdma_rlc_reg_offset + reg);
471 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
472 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
473 		DUMP_REG(sdma_rlc_reg_offset + reg);
474 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
475 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
476 		DUMP_REG(sdma_rlc_reg_offset + reg);
477 
478 	WARN_ON_ONCE(i != HQD_N_REGS);
479 	*n_regs = i;
480 
481 	return 0;
482 }
483 
484 bool kgd_gfx_v9_hqd_is_occupied(struct amdgpu_device *adev,
485 				uint64_t queue_address, uint32_t pipe_id,
486 				uint32_t queue_id, uint32_t inst)
487 {
488 	uint32_t act;
489 	bool retval = false;
490 	uint32_t low, high;
491 
492 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
493 	act = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE);
494 	if (act) {
495 		low = lower_32_bits(queue_address >> 8);
496 		high = upper_32_bits(queue_address >> 8);
497 
498 		if (low == RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE) &&
499 		   high == RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE_HI))
500 			retval = true;
501 	}
502 	kgd_gfx_v9_release_queue(adev, inst);
503 	return retval;
504 }
505 
506 static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd)
507 {
508 	struct v9_sdma_mqd *m;
509 	uint32_t sdma_rlc_reg_offset;
510 	uint32_t sdma_rlc_rb_cntl;
511 
512 	m = get_sdma_mqd(mqd);
513 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
514 					    m->sdma_queue_id);
515 
516 	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
517 
518 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
519 		return true;
520 
521 	return false;
522 }
523 
524 int kgd_gfx_v9_hqd_destroy(struct amdgpu_device *adev, void *mqd,
525 				enum kfd_preempt_type reset_type,
526 				unsigned int utimeout, uint32_t pipe_id,
527 				uint32_t queue_id, uint32_t inst)
528 {
529 	enum hqd_dequeue_request_type type;
530 	unsigned long end_jiffies;
531 	uint32_t temp;
532 	struct v9_mqd *m = get_mqd(mqd);
533 
534 	if (amdgpu_in_reset(adev))
535 		return -EIO;
536 
537 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
538 
539 	if (m->cp_hqd_vmid == 0)
540 		WREG32_FIELD15_RLC(GC, GET_INST(GC, inst), RLC_CP_SCHEDULERS, scheduler1, 0);
541 
542 	switch (reset_type) {
543 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
544 		type = DRAIN_PIPE;
545 		break;
546 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
547 		type = RESET_WAVES;
548 		break;
549 	case KFD_PREEMPT_TYPE_WAVEFRONT_SAVE:
550 		type = SAVE_WAVES;
551 		break;
552 	default:
553 		type = DRAIN_PIPE;
554 		break;
555 	}
556 
557 	WREG32_SOC15_RLC(GC, GET_INST(GC, inst), mmCP_HQD_DEQUEUE_REQUEST, type);
558 
559 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
560 	while (true) {
561 		temp = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE);
562 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
563 			break;
564 		if (time_after(jiffies, end_jiffies)) {
565 			pr_err("cp queue preemption time out.\n");
566 			kgd_gfx_v9_release_queue(adev, inst);
567 			return -ETIME;
568 		}
569 		usleep_range(500, 1000);
570 	}
571 
572 	kgd_gfx_v9_release_queue(adev, inst);
573 	return 0;
574 }
575 
576 static int kgd_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd,
577 				unsigned int utimeout)
578 {
579 	struct v9_sdma_mqd *m;
580 	uint32_t sdma_rlc_reg_offset;
581 	uint32_t temp;
582 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
583 
584 	m = get_sdma_mqd(mqd);
585 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
586 					    m->sdma_queue_id);
587 
588 	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
589 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
590 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
591 
592 	while (true) {
593 		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
594 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
595 			break;
596 		if (time_after(jiffies, end_jiffies)) {
597 			pr_err("SDMA RLC not idle in %s\n", __func__);
598 			return -ETIME;
599 		}
600 		usleep_range(500, 1000);
601 	}
602 
603 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
604 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
605 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
606 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
607 
608 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
609 	m->sdmax_rlcx_rb_rptr_hi =
610 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
611 
612 	return 0;
613 }
614 
615 bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
616 					uint8_t vmid, uint16_t *p_pasid)
617 {
618 	uint32_t value;
619 
620 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
621 		     + vmid);
622 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
623 
624 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
625 }
626 
627 int kgd_gfx_v9_wave_control_execute(struct amdgpu_device *adev,
628 					uint32_t gfx_index_val,
629 					uint32_t sq_cmd, uint32_t inst)
630 {
631 	uint32_t data = 0;
632 
633 	mutex_lock(&adev->grbm_idx_mutex);
634 
635 	WREG32_SOC15_RLC_SHADOW(GC, GET_INST(GC, inst), mmGRBM_GFX_INDEX, gfx_index_val);
636 	WREG32_SOC15(GC, GET_INST(GC, inst), mmSQ_CMD, sq_cmd);
637 
638 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
639 		INSTANCE_BROADCAST_WRITES, 1);
640 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
641 		SH_BROADCAST_WRITES, 1);
642 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
643 		SE_BROADCAST_WRITES, 1);
644 
645 	WREG32_SOC15_RLC_SHADOW(GC, GET_INST(GC, inst), mmGRBM_GFX_INDEX, data);
646 	mutex_unlock(&adev->grbm_idx_mutex);
647 
648 	return 0;
649 }
650 
651 /*
652  * GFX9 helper for wave launch stall requirements on debug trap setting.
653  *
654  * vmid:
655  *   Target VMID to stall/unstall.
656  *
657  * stall:
658  *   0-unstall wave launch (enable), 1-stall wave launch (disable).
659  *   After wavefront launch has been stalled, allocated waves must drain from
660  *   SPI in order for debug trap settings to take effect on those waves.
661  *   This is roughly a ~96 clock cycle wait on SPI where a read on
662  *   SPI_GDBG_WAVE_CNTL translates to ~32 clock cycles.
663  *   KGD_GFX_V9_WAVE_LAUNCH_SPI_DRAIN_LATENCY indicates the number of reads required.
664  *
665  *   NOTE: We can afford to clear the entire STALL_VMID field on unstall
666  *   because GFX9.4.1 cannot support multi-process debugging due to trap
667  *   configuration and masking being limited to global scope.  Always assume
668  *   single process conditions.
669  */
670 #define KGD_GFX_V9_WAVE_LAUNCH_SPI_DRAIN_LATENCY	3
671 void kgd_gfx_v9_set_wave_launch_stall(struct amdgpu_device *adev,
672 					uint32_t vmid,
673 					bool stall)
674 {
675 	int i;
676 	uint32_t data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
677 
678 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1))
679 		data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL, STALL_VMID,
680 							stall ? 1 << vmid : 0);
681 	else
682 		data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL, STALL_RA,
683 							stall ? 1 : 0);
684 
685 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), data);
686 
687 	if (!stall)
688 		return;
689 
690 	for (i = 0; i < KGD_GFX_V9_WAVE_LAUNCH_SPI_DRAIN_LATENCY; i++)
691 		RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
692 }
693 
694 /*
695  * restore_dbg_registers is ignored here but is a general interface requirement
696  * for devices that support GFXOFF and where the RLC save/restore list
697  * does not support hw registers for debugging i.e. the driver has to manually
698  * initialize the debug mode registers after it has disabled GFX off during the
699  * debug session.
700  */
701 uint32_t kgd_gfx_v9_enable_debug_trap(struct amdgpu_device *adev,
702 				bool restore_dbg_registers,
703 				uint32_t vmid)
704 {
705 	mutex_lock(&adev->grbm_idx_mutex);
706 
707 	kgd_gfx_v9_set_wave_launch_stall(adev, vmid, true);
708 
709 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0);
710 
711 	kgd_gfx_v9_set_wave_launch_stall(adev, vmid, false);
712 
713 	mutex_unlock(&adev->grbm_idx_mutex);
714 
715 	return 0;
716 }
717 
718 /*
719  * keep_trap_enabled is ignored here but is a general interface requirement
720  * for devices that support multi-process debugging where the performance
721  * overhead from trap temporary setup needs to be bypassed when the debug
722  * session has ended.
723  */
724 uint32_t kgd_gfx_v9_disable_debug_trap(struct amdgpu_device *adev,
725 					bool keep_trap_enabled,
726 					uint32_t vmid)
727 {
728 	mutex_lock(&adev->grbm_idx_mutex);
729 
730 	kgd_gfx_v9_set_wave_launch_stall(adev, vmid, true);
731 
732 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0);
733 
734 	kgd_gfx_v9_set_wave_launch_stall(adev, vmid, false);
735 
736 	mutex_unlock(&adev->grbm_idx_mutex);
737 
738 	return 0;
739 }
740 
741 int kgd_gfx_v9_validate_trap_override_request(struct amdgpu_device *adev,
742 					uint32_t trap_override,
743 					uint32_t *trap_mask_supported)
744 {
745 	*trap_mask_supported &= KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH;
746 
747 	/* The SPI_GDBG_TRAP_MASK register is global and affects all
748 	 * processes. Only allow OR-ing the address-watch bit, since
749 	 * this only affects processes under the debugger. Other bits
750 	 * should stay 0 to avoid the debugger interfering with other
751 	 * processes.
752 	 */
753 	if (trap_override != KFD_DBG_TRAP_OVERRIDE_OR)
754 		return -EINVAL;
755 
756 	return 0;
757 }
758 
759 uint32_t kgd_gfx_v9_set_wave_launch_trap_override(struct amdgpu_device *adev,
760 					     uint32_t vmid,
761 					     uint32_t trap_override,
762 					     uint32_t trap_mask_bits,
763 					     uint32_t trap_mask_request,
764 					     uint32_t *trap_mask_prev,
765 					     uint32_t kfd_dbg_cntl_prev)
766 {
767 	uint32_t data, wave_cntl_prev;
768 
769 	mutex_lock(&adev->grbm_idx_mutex);
770 
771 	wave_cntl_prev = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
772 
773 	kgd_gfx_v9_set_wave_launch_stall(adev, vmid, true);
774 
775 	data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK));
776 	*trap_mask_prev = REG_GET_FIELD(data, SPI_GDBG_TRAP_MASK, EXCP_EN);
777 
778 	trap_mask_bits = (trap_mask_bits & trap_mask_request) |
779 		(*trap_mask_prev & ~trap_mask_request);
780 
781 	data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK, EXCP_EN, trap_mask_bits);
782 	data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK, REPLACE, trap_override);
783 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), data);
784 
785 	/* We need to preserve wave launch mode stall settings. */
786 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), wave_cntl_prev);
787 
788 	mutex_unlock(&adev->grbm_idx_mutex);
789 
790 	return 0;
791 }
792 
793 uint32_t kgd_gfx_v9_set_wave_launch_mode(struct amdgpu_device *adev,
794 					uint8_t wave_launch_mode,
795 					uint32_t vmid)
796 {
797 	uint32_t data = 0;
798 	bool is_mode_set = !!wave_launch_mode;
799 
800 	mutex_lock(&adev->grbm_idx_mutex);
801 
802 	kgd_gfx_v9_set_wave_launch_stall(adev, vmid, true);
803 
804 	data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2,
805 		VMID_MASK, is_mode_set ? 1 << vmid : 0);
806 	data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2,
807 		MODE, is_mode_set ? wave_launch_mode : 0);
808 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL2), data);
809 
810 	kgd_gfx_v9_set_wave_launch_stall(adev, vmid, false);
811 
812 	mutex_unlock(&adev->grbm_idx_mutex);
813 
814 	return 0;
815 }
816 
817 #define TCP_WATCH_STRIDE (mmTCP_WATCH1_ADDR_H - mmTCP_WATCH0_ADDR_H)
818 uint32_t kgd_gfx_v9_set_address_watch(struct amdgpu_device *adev,
819 					uint64_t watch_address,
820 					uint32_t watch_address_mask,
821 					uint32_t watch_id,
822 					uint32_t watch_mode,
823 					uint32_t debug_vmid,
824 					uint32_t inst)
825 {
826 	uint32_t watch_address_high;
827 	uint32_t watch_address_low;
828 	uint32_t watch_address_cntl;
829 
830 	watch_address_cntl = 0;
831 
832 	watch_address_low = lower_32_bits(watch_address);
833 	watch_address_high = upper_32_bits(watch_address) & 0xffff;
834 
835 	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
836 			TCP_WATCH0_CNTL,
837 			VMID,
838 			debug_vmid);
839 	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
840 			TCP_WATCH0_CNTL,
841 			MODE,
842 			watch_mode);
843 	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
844 			TCP_WATCH0_CNTL,
845 			MASK,
846 			watch_address_mask >> 6);
847 
848 	/* Turning off this watch point until we set all the registers */
849 	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
850 			TCP_WATCH0_CNTL,
851 			VALID,
852 			0);
853 
854 	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
855 			(watch_id * TCP_WATCH_STRIDE)),
856 			watch_address_cntl);
857 
858 	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) +
859 			(watch_id * TCP_WATCH_STRIDE)),
860 			watch_address_high);
861 
862 	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_L) +
863 			(watch_id * TCP_WATCH_STRIDE)),
864 			watch_address_low);
865 
866 	/* Enable the watch point */
867 	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
868 			TCP_WATCH0_CNTL,
869 			VALID,
870 			1);
871 
872 	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
873 			(watch_id * TCP_WATCH_STRIDE)),
874 			watch_address_cntl);
875 
876 	return 0;
877 }
878 
879 uint32_t kgd_gfx_v9_clear_address_watch(struct amdgpu_device *adev,
880 					uint32_t watch_id)
881 {
882 	uint32_t watch_address_cntl;
883 
884 	watch_address_cntl = 0;
885 
886 	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
887 			(watch_id * TCP_WATCH_STRIDE)),
888 			watch_address_cntl);
889 
890 	return 0;
891 }
892 
893 /* kgd_gfx_v9_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values
894  * The values read are:
895  *     ib_offload_wait_time     -- Wait Count for Indirect Buffer Offloads.
896  *     atomic_offload_wait_time -- Wait Count for L2 and GDS Atomics Offloads.
897  *     wrm_offload_wait_time    -- Wait Count for WAIT_REG_MEM Offloads.
898  *     gws_wait_time            -- Wait Count for Global Wave Syncs.
899  *     que_sleep_wait_time      -- Wait Count for Dequeue Retry.
900  *     sch_wave_wait_time       -- Wait Count for Scheduling Wave Message.
901  *     sem_rearm_wait_time      -- Wait Count for Semaphore re-arm.
902  *     deq_retry_wait_time      -- Wait Count for Global Wave Syncs.
903  */
904 void kgd_gfx_v9_get_iq_wait_times(struct amdgpu_device *adev,
905 					uint32_t *wait_times,
906 					uint32_t inst)
907 
908 {
909 	*wait_times = RREG32_SOC15_RLC(GC, GET_INST(GC, inst),
910 			mmCP_IQ_WAIT_TIME2);
911 }
912 
913 void kgd_gfx_v9_set_vm_context_page_table_base(struct amdgpu_device *adev,
914 			uint32_t vmid, uint64_t page_table_base)
915 {
916 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
917 		pr_err("trying to set page table base for wrong VMID %u\n",
918 		       vmid);
919 		return;
920 	}
921 
922 	adev->mmhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
923 
924 	adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
925 }
926 
927 static void lock_spi_csq_mutexes(struct amdgpu_device *adev)
928 {
929 	mutex_lock(&adev->srbm_mutex);
930 	mutex_lock(&adev->grbm_idx_mutex);
931 
932 }
933 
934 static void unlock_spi_csq_mutexes(struct amdgpu_device *adev)
935 {
936 	mutex_unlock(&adev->grbm_idx_mutex);
937 	mutex_unlock(&adev->srbm_mutex);
938 }
939 
940 /**
941  * get_wave_count: Read device registers to get number of waves in flight for
942  * a particular queue. The method also returns the VMID associated with the
943  * queue.
944  *
945  * @adev: Handle of device whose registers are to be read
946  * @queue_idx: Index of queue in the queue-map bit-field
947  * @wave_cnt: Output parameter updated with number of waves in flight
948  * @vmid: Output parameter updated with VMID of queue whose wave count
949  *        is being collected
950  * @inst: xcc's instance number on a multi-XCC setup
951  */
952 static void get_wave_count(struct amdgpu_device *adev, int queue_idx,
953 		struct kfd_cu_occupancy *queue_cnt, uint32_t inst)
954 {
955 	int pipe_idx;
956 	int queue_slot;
957 	unsigned int reg_val;
958 	unsigned int wave_cnt;
959 	/*
960 	 * Program GRBM with appropriate MEID, PIPEID, QUEUEID and VMID
961 	 * parameters to read out waves in flight. Get VMID if there are
962 	 * non-zero waves in flight.
963 	 */
964 	pipe_idx = queue_idx / adev->gfx.mec.num_queue_per_pipe;
965 	queue_slot = queue_idx % adev->gfx.mec.num_queue_per_pipe;
966 	soc15_grbm_select(adev, 1, pipe_idx, queue_slot, 0, GET_INST(GC, inst));
967 	reg_val = RREG32_SOC15_IP(GC, SOC15_REG_OFFSET(GC, GET_INST(GC, inst),
968 				  mmSPI_CSQ_WF_ACTIVE_COUNT_0) + queue_slot);
969 	wave_cnt = reg_val & SPI_CSQ_WF_ACTIVE_COUNT_0__COUNT_MASK;
970 	if (wave_cnt != 0) {
971 		queue_cnt->wave_cnt += wave_cnt;
972 		queue_cnt->doorbell_off =
973 			(RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_DOORBELL_CONTROL) &
974 			 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET_MASK) >>
975 			 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
976 	}
977 }
978 
979 /**
980  * kgd_gfx_v9_get_cu_occupancy: Reads relevant registers associated with each
981  * shader engine and aggregates the number of waves that are in flight for the
982  * process whose pasid is provided as a parameter. The process could have ZERO
983  * or more queues running and submitting waves to compute units.
984  *
985  * @adev: Handle of device from which to get number of waves in flight
986  * @cu_occupancy: Array that gets filled with wave_cnt and doorbell offset
987  *		  for comparison later.
988  * @max_waves_per_cu: Output parameter updated with maximum number of waves
989  *                    possible per Compute Unit
990  * @inst: xcc's instance number on a multi-XCC setup
991  *
992  * Note: It's possible that the device has too many queues (oversubscription)
993  * in which case a VMID could be remapped to a different PASID. This could lead
994  * to an inaccurate wave count. Following is a high-level sequence:
995  *    Time T1: vmid = getVmid(); vmid is associated with Pasid P1
996  *    Time T2: passId = getPasId(vmid); vmid is associated with Pasid P2
997  * In the sequence above wave count obtained from time T1 will be incorrectly
998  * lost or added to total wave count.
999  *
1000  * The registers that provide the waves in flight are:
1001  *
1002  *  SPI_CSQ_WF_ACTIVE_STATUS - bit-map of queues per pipe. The bit is ON if a
1003  *  queue is slotted, OFF if there is no queue. A process could have ZERO or
1004  *  more queues slotted and submitting waves to be run on compute units. Even
1005  *  when there is a queue it is possible there could be zero wave fronts, this
1006  *  can happen when queue is waiting on top-of-pipe events - e.g. waitRegMem
1007  *  command
1008  *
1009  *  For each bit that is ON from above:
1010  *
1011  *    Read (SPI_CSQ_WF_ACTIVE_COUNT_0 + queue_idx) register. It provides the
1012  *    number of waves that are in flight for the queue at specified index. The
1013  *    index ranges from 0 to 7.
1014  *
1015  *    If non-zero waves are in flight, store the corresponding doorbell offset
1016  *    of the queue, along with the wave count.
1017  *
1018  *    Determine if the queue belongs to the process by comparing the doorbell
1019  *    offset against the process's queues. If it matches, aggregate the wave
1020  *    count for the process.
1021  *
1022  *  Reading registers referenced above involves programming GRBM appropriately
1023  */
1024 void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev,
1025 				 struct kfd_cu_occupancy *cu_occupancy,
1026 				 int *max_waves_per_cu, uint32_t inst)
1027 {
1028 	int qidx;
1029 	int se_idx;
1030 	int se_cnt;
1031 	int queue_map;
1032 	int max_queue_cnt;
1033 	DECLARE_BITMAP(cp_queue_bitmap, AMDGPU_MAX_QUEUES);
1034 
1035 	lock_spi_csq_mutexes(adev);
1036 	soc15_grbm_select(adev, 1, 0, 0, 0, GET_INST(GC, inst));
1037 
1038 	/*
1039 	 * Iterate through the shader engines and arrays of the device
1040 	 * to get number of waves in flight
1041 	 */
1042 	bitmap_complement(cp_queue_bitmap, adev->gfx.mec_bitmap[0].queue_bitmap,
1043 			  AMDGPU_MAX_QUEUES);
1044 	max_queue_cnt = adev->gfx.mec.num_pipe_per_mec *
1045 			adev->gfx.mec.num_queue_per_pipe;
1046 	se_cnt = adev->gfx.config.max_shader_engines;
1047 	for (se_idx = 0; se_idx < se_cnt; se_idx++) {
1048 		amdgpu_gfx_select_se_sh(adev, se_idx, 0, 0xffffffff, inst);
1049 		queue_map = RREG32_SOC15(GC, GET_INST(GC, inst), mmSPI_CSQ_WF_ACTIVE_STATUS);
1050 
1051 		/*
1052 		 * Assumption: queue map encodes following schema: four
1053 		 * pipes per each micro-engine, with each pipe mapping
1054 		 * eight queues. This schema is true for GFX9 devices
1055 		 * and must be verified for newer device families
1056 		 */
1057 		for (qidx = 0; qidx < max_queue_cnt; qidx++) {
1058 			/* Skip qeueus that are not associated with
1059 			 * compute functions
1060 			 */
1061 			if (!test_bit(qidx, cp_queue_bitmap))
1062 				continue;
1063 
1064 			if (!(queue_map & (1 << qidx)))
1065 				continue;
1066 
1067 			/* Get number of waves in flight and aggregate them */
1068 			get_wave_count(adev, qidx, &cu_occupancy[qidx],
1069 					inst);
1070 		}
1071 	}
1072 
1073 	amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff, inst);
1074 	soc15_grbm_select(adev, 0, 0, 0, 0, GET_INST(GC, inst));
1075 	unlock_spi_csq_mutexes(adev);
1076 
1077 	/* Update the output parameters and return */
1078 	*max_waves_per_cu = adev->gfx.cu_info.simd_per_cu *
1079 				adev->gfx.cu_info.max_waves_per_simd;
1080 }
1081 
1082 void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev,
1083 		uint32_t wait_times,
1084 		uint32_t grace_period,
1085 		uint32_t *reg_offset,
1086 		uint32_t *reg_data)
1087 {
1088 	*reg_data = wait_times;
1089 
1090 	/*
1091 	 * The CP cannot handle a 0 grace period input and will result in
1092 	 * an infinite grace period being set so set to 1 to prevent this.
1093 	 */
1094 	if (grace_period == 0)
1095 		grace_period = 1;
1096 
1097 	*reg_data = REG_SET_FIELD(*reg_data,
1098 			CP_IQ_WAIT_TIME2,
1099 			SCH_WAVE,
1100 			grace_period);
1101 
1102 	*reg_offset = SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2);
1103 }
1104 
1105 void kgd_gfx_v9_program_trap_handler_settings(struct amdgpu_device *adev,
1106 		uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, uint32_t inst)
1107 {
1108 	kgd_gfx_v9_lock_srbm(adev, 0, 0, 0, vmid, inst);
1109 
1110 	/*
1111 	 * Program TBA registers
1112 	 */
1113 	WREG32_SOC15(GC, GET_INST(GC, inst), mmSQ_SHADER_TBA_LO,
1114 			lower_32_bits(tba_addr >> 8));
1115 	WREG32_SOC15(GC, GET_INST(GC, inst), mmSQ_SHADER_TBA_HI,
1116 			upper_32_bits(tba_addr >> 8));
1117 
1118 	/*
1119 	 * Program TMA registers
1120 	 */
1121 	WREG32_SOC15(GC, GET_INST(GC, inst), mmSQ_SHADER_TMA_LO,
1122 			lower_32_bits(tma_addr >> 8));
1123 	WREG32_SOC15(GC, GET_INST(GC, inst), mmSQ_SHADER_TMA_HI,
1124 			upper_32_bits(tma_addr >> 8));
1125 
1126 	kgd_gfx_v9_unlock_srbm(adev, inst);
1127 }
1128 
1129 uint64_t kgd_gfx_v9_hqd_get_pq_addr(struct amdgpu_device *adev,
1130 				    uint32_t pipe_id, uint32_t queue_id,
1131 				    uint32_t inst)
1132 {
1133 	uint32_t low, high;
1134 	uint64_t queue_addr = 0;
1135 
1136 	if (!adev->debug_exp_resets &&
1137 	    !adev->gfx.num_gfx_rings)
1138 		return 0;
1139 
1140 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
1141 	amdgpu_gfx_rlc_enter_safe_mode(adev, inst);
1142 
1143 	if (!RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE))
1144 		goto unlock_out;
1145 
1146 	low = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE);
1147 	high = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE_HI);
1148 
1149 	/* only concerned with user queues. */
1150 	if (!high)
1151 		goto unlock_out;
1152 
1153 	queue_addr = (((queue_addr | high) << 32) | low) << 8;
1154 
1155 unlock_out:
1156 	amdgpu_gfx_rlc_exit_safe_mode(adev, inst);
1157 	kgd_gfx_v9_release_queue(adev, inst);
1158 
1159 	return queue_addr;
1160 }
1161 
1162 /* assume queue acquired  */
1163 static int kgd_gfx_v9_hqd_dequeue_wait(struct amdgpu_device *adev, uint32_t inst,
1164 				       unsigned int utimeout)
1165 {
1166 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
1167 
1168 	while (true) {
1169 		uint32_t temp = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE);
1170 
1171 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
1172 			return 0;
1173 
1174 		if (time_after(jiffies, end_jiffies))
1175 			return -ETIME;
1176 
1177 		usleep_range(500, 1000);
1178 	}
1179 }
1180 
1181 uint64_t kgd_gfx_v9_hqd_reset(struct amdgpu_device *adev,
1182 			      uint32_t pipe_id, uint32_t queue_id,
1183 			      uint32_t inst, unsigned int utimeout)
1184 {
1185 	uint32_t low, high, pipe_reset_data = 0;
1186 	uint64_t queue_addr = 0;
1187 
1188 	kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
1189 	amdgpu_gfx_rlc_enter_safe_mode(adev, inst);
1190 
1191 	if (!RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE))
1192 		goto unlock_out;
1193 
1194 	low = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE);
1195 	high = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE_HI);
1196 
1197 	/* only concerned with user queues. */
1198 	if (!high)
1199 		goto unlock_out;
1200 
1201 	queue_addr = (((queue_addr | high) << 32) | low) << 8;
1202 
1203 	pr_debug("Attempting queue reset on XCC %i pipe id %i queue id %i\n",
1204 		 inst, pipe_id, queue_id);
1205 
1206 	/* assume previous dequeue request issued will take affect after reset */
1207 	WREG32_SOC15(GC, GET_INST(GC, inst), mmSPI_COMPUTE_QUEUE_RESET, 0x1);
1208 
1209 	if (!kgd_gfx_v9_hqd_dequeue_wait(adev, inst, utimeout))
1210 		goto unlock_out;
1211 
1212 	pr_debug("Attempting pipe reset on XCC %i pipe id %i\n", inst, pipe_id);
1213 
1214 	pipe_reset_data = REG_SET_FIELD(pipe_reset_data, CP_MEC_CNTL, MEC_ME1_PIPE0_RESET, 1);
1215 	pipe_reset_data = pipe_reset_data << pipe_id;
1216 
1217 	WREG32_SOC15(GC, GET_INST(GC, inst), mmCP_MEC_CNTL, pipe_reset_data);
1218 	WREG32_SOC15(GC, GET_INST(GC, inst), mmCP_MEC_CNTL, 0);
1219 
1220 	if (kgd_gfx_v9_hqd_dequeue_wait(adev, inst, utimeout))
1221 		queue_addr = 0;
1222 
1223 unlock_out:
1224 	pr_debug("queue reset on XCC %i pipe id %i queue id %i %s\n",
1225 		 inst, pipe_id, queue_id, !!queue_addr ? "succeeded!" : "failed!");
1226 	amdgpu_gfx_rlc_exit_safe_mode(adev, inst);
1227 	kgd_gfx_v9_release_queue(adev, inst);
1228 
1229 	return queue_addr;
1230 }
1231 
1232 const struct kfd2kgd_calls gfx_v9_kfd2kgd = {
1233 	.program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
1234 	.set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping,
1235 	.init_interrupts = kgd_gfx_v9_init_interrupts,
1236 	.hqd_load = kgd_gfx_v9_hqd_load,
1237 	.hiq_mqd_load = kgd_gfx_v9_hiq_mqd_load,
1238 	.hqd_sdma_load = kgd_hqd_sdma_load,
1239 	.hqd_dump = kgd_gfx_v9_hqd_dump,
1240 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
1241 	.hqd_is_occupied = kgd_gfx_v9_hqd_is_occupied,
1242 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
1243 	.hqd_destroy = kgd_gfx_v9_hqd_destroy,
1244 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
1245 	.wave_control_execute = kgd_gfx_v9_wave_control_execute,
1246 	.get_atc_vmid_pasid_mapping_info =
1247 			kgd_gfx_v9_get_atc_vmid_pasid_mapping_info,
1248 	.set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base,
1249 	.enable_debug_trap = kgd_gfx_v9_enable_debug_trap,
1250 	.disable_debug_trap = kgd_gfx_v9_disable_debug_trap,
1251 	.validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request,
1252 	.set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override,
1253 	.set_wave_launch_mode = kgd_gfx_v9_set_wave_launch_mode,
1254 	.set_address_watch = kgd_gfx_v9_set_address_watch,
1255 	.clear_address_watch = kgd_gfx_v9_clear_address_watch,
1256 	.get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times,
1257 	.build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info,
1258 	.get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,
1259 	.program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
1260 	.hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
1261 	.hqd_reset = kgd_gfx_v9_hqd_reset
1262 };
1263