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 
42 enum hqd_dequeue_request_type {
43 	NO_ACTION = 0,
44 	DRAIN_PIPE,
45 	RESET_WAVES
46 };
47 
48 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
49 {
50 	return (struct amdgpu_device *)kgd;
51 }
52 
53 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
54 			uint32_t queue, uint32_t vmid)
55 {
56 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
57 
58 	mutex_lock(&adev->srbm_mutex);
59 	soc15_grbm_select(adev, mec, pipe, queue, vmid);
60 }
61 
62 static void unlock_srbm(struct kgd_dev *kgd)
63 {
64 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
65 
66 	soc15_grbm_select(adev, 0, 0, 0, 0);
67 	mutex_unlock(&adev->srbm_mutex);
68 }
69 
70 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
71 				uint32_t queue_id)
72 {
73 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
74 
75 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
76 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
77 
78 	lock_srbm(kgd, mec, pipe, queue_id, 0);
79 }
80 
81 static uint64_t get_queue_mask(struct amdgpu_device *adev,
82 			       uint32_t pipe_id, uint32_t queue_id)
83 {
84 	unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe +
85 			queue_id;
86 
87 	return 1ull << bit;
88 }
89 
90 static void release_queue(struct kgd_dev *kgd)
91 {
92 	unlock_srbm(kgd);
93 }
94 
95 void kgd_gfx_v9_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
96 					uint32_t sh_mem_config,
97 					uint32_t sh_mem_ape1_base,
98 					uint32_t sh_mem_ape1_limit,
99 					uint32_t sh_mem_bases)
100 {
101 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
102 
103 	lock_srbm(kgd, 0, 0, 0, vmid);
104 
105 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
106 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
107 	/* APE1 no longer exists on GFX9 */
108 
109 	unlock_srbm(kgd);
110 }
111 
112 int kgd_gfx_v9_set_pasid_vmid_mapping(struct kgd_dev *kgd, u32 pasid,
113 					unsigned int vmid)
114 {
115 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
116 
117 	/*
118 	 * We have to assume that there is no outstanding mapping.
119 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
120 	 * a mapping is in progress or because a mapping finished
121 	 * and the SW cleared it.
122 	 * So the protocol is to always wait & clear.
123 	 */
124 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
125 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
126 
127 	/*
128 	 * need to do this twice, once for gfx and once for mmhub
129 	 * for ATC add 16 to VMID for mmhub, for IH different registers.
130 	 * ATC_VMID0..15 registers are separate from ATC_VMID16..31.
131 	 */
132 
133 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
134 	       pasid_mapping);
135 
136 	while (!(RREG32(SOC15_REG_OFFSET(
137 				ATHUB, 0,
138 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
139 		 (1U << vmid)))
140 		cpu_relax();
141 
142 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
143 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
144 	       1U << vmid);
145 
146 	/* Mapping vmid to pasid also for IH block */
147 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
148 	       pasid_mapping);
149 
150 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID16_PASID_MAPPING) + vmid,
151 	       pasid_mapping);
152 
153 	while (!(RREG32(SOC15_REG_OFFSET(
154 				ATHUB, 0,
155 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
156 		 (1U << (vmid + 16))))
157 		cpu_relax();
158 
159 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
160 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
161 	       1U << (vmid + 16));
162 
163 	/* Mapping vmid to pasid also for IH block */
164 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid,
165 	       pasid_mapping);
166 	return 0;
167 }
168 
169 /* TODO - RING0 form of field is obsolete, seems to date back to SI
170  * but still works
171  */
172 
173 int kgd_gfx_v9_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
174 {
175 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
176 	uint32_t mec;
177 	uint32_t pipe;
178 
179 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
180 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
181 
182 	lock_srbm(kgd, mec, pipe, 0, 0);
183 
184 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
185 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
186 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
187 
188 	unlock_srbm(kgd);
189 
190 	return 0;
191 }
192 
193 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
194 				unsigned int engine_id,
195 				unsigned int queue_id)
196 {
197 	uint32_t sdma_engine_reg_base = 0;
198 	uint32_t sdma_rlc_reg_offset;
199 
200 	switch (engine_id) {
201 	default:
202 		dev_warn(adev->dev,
203 			 "Invalid sdma engine id (%d), using engine id 0\n",
204 			 engine_id);
205 		fallthrough;
206 	case 0:
207 		sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA0, 0,
208 				mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL;
209 		break;
210 	case 1:
211 		sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA1, 0,
212 				mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL;
213 		break;
214 	}
215 
216 	sdma_rlc_reg_offset = sdma_engine_reg_base
217 		+ queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
218 
219 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
220 		 queue_id, sdma_rlc_reg_offset);
221 
222 	return sdma_rlc_reg_offset;
223 }
224 
225 static inline struct v9_mqd *get_mqd(void *mqd)
226 {
227 	return (struct v9_mqd *)mqd;
228 }
229 
230 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
231 {
232 	return (struct v9_sdma_mqd *)mqd;
233 }
234 
235 int kgd_gfx_v9_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
236 			uint32_t queue_id, uint32_t __user *wptr,
237 			uint32_t wptr_shift, uint32_t wptr_mask,
238 			struct mm_struct *mm)
239 {
240 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
241 	struct v9_mqd *m;
242 	uint32_t *mqd_hqd;
243 	uint32_t reg, hqd_base, data;
244 
245 	m = get_mqd(mqd);
246 
247 	acquire_queue(kgd, pipe_id, queue_id);
248 
249 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
250 	mqd_hqd = &m->cp_mqd_base_addr_lo;
251 	hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
252 
253 	for (reg = hqd_base;
254 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
255 		WREG32_RLC(reg, mqd_hqd[reg - hqd_base]);
256 
257 
258 	/* Activate doorbell logic before triggering WPTR poll. */
259 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
260 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
261 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
262 
263 	if (wptr) {
264 		/* Don't read wptr with get_user because the user
265 		 * context may not be accessible (if this function
266 		 * runs in a work queue). Instead trigger a one-shot
267 		 * polling read from memory in the CP. This assumes
268 		 * that wptr is GPU-accessible in the queue's VMID via
269 		 * ATC or SVM. WPTR==RPTR before starting the poll so
270 		 * the CP starts fetching new commands from the right
271 		 * place.
272 		 *
273 		 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
274 		 * tricky. Assume that the queue didn't overflow. The
275 		 * number of valid bits in the 32-bit RPTR depends on
276 		 * the queue size. The remaining bits are taken from
277 		 * the saved 64-bit WPTR. If the WPTR wrapped, add the
278 		 * queue size.
279 		 */
280 		uint32_t queue_size =
281 			2 << REG_GET_FIELD(m->cp_hqd_pq_control,
282 					   CP_HQD_PQ_CONTROL, QUEUE_SIZE);
283 		uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
284 
285 		if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
286 			guessed_wptr += queue_size;
287 		guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
288 		guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
289 
290 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
291 		       lower_32_bits(guessed_wptr));
292 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
293 		       upper_32_bits(guessed_wptr));
294 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
295 		       lower_32_bits((uintptr_t)wptr));
296 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
297 		       upper_32_bits((uintptr_t)wptr));
298 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
299 		       (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
300 	}
301 
302 	/* Start the EOP fetcher */
303 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
304 	       REG_SET_FIELD(m->cp_hqd_eop_rptr,
305 			     CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
306 
307 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
308 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
309 
310 	release_queue(kgd);
311 
312 	return 0;
313 }
314 
315 int kgd_gfx_v9_hiq_mqd_load(struct kgd_dev *kgd, void *mqd,
316 			    uint32_t pipe_id, uint32_t queue_id,
317 			    uint32_t doorbell_off)
318 {
319 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
320 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
321 	struct v9_mqd *m;
322 	uint32_t mec, pipe;
323 	int r;
324 
325 	m = get_mqd(mqd);
326 
327 	acquire_queue(kgd, pipe_id, queue_id);
328 
329 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
330 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
331 
332 	pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
333 		 mec, pipe, queue_id);
334 
335 	spin_lock(&adev->gfx.kiq.ring_lock);
336 	r = amdgpu_ring_alloc(kiq_ring, 7);
337 	if (r) {
338 		pr_err("Failed to alloc KIQ (%d).\n", r);
339 		goto out_unlock;
340 	}
341 
342 	amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
343 	amdgpu_ring_write(kiq_ring,
344 			  PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
345 			  PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */
346 			  PACKET3_MAP_QUEUES_QUEUE(queue_id) |
347 			  PACKET3_MAP_QUEUES_PIPE(pipe) |
348 			  PACKET3_MAP_QUEUES_ME((mec - 1)) |
349 			  PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
350 			  PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
351 			  PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */
352 			  PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
353 	amdgpu_ring_write(kiq_ring,
354 			  PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off));
355 	amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo);
356 	amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi);
357 	amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo);
358 	amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi);
359 	amdgpu_ring_commit(kiq_ring);
360 
361 out_unlock:
362 	spin_unlock(&adev->gfx.kiq.ring_lock);
363 	release_queue(kgd);
364 
365 	return r;
366 }
367 
368 int kgd_gfx_v9_hqd_dump(struct kgd_dev *kgd,
369 			uint32_t pipe_id, uint32_t queue_id,
370 			uint32_t (**dump)[2], uint32_t *n_regs)
371 {
372 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
373 	uint32_t i = 0, reg;
374 #define HQD_N_REGS 56
375 #define DUMP_REG(addr) do {				\
376 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
377 			break;				\
378 		(*dump)[i][0] = (addr) << 2;		\
379 		(*dump)[i++][1] = RREG32(addr);		\
380 	} while (0)
381 
382 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
383 	if (*dump == NULL)
384 		return -ENOMEM;
385 
386 	acquire_queue(kgd, pipe_id, queue_id);
387 
388 	for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
389 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
390 		DUMP_REG(reg);
391 
392 	release_queue(kgd);
393 
394 	WARN_ON_ONCE(i != HQD_N_REGS);
395 	*n_regs = i;
396 
397 	return 0;
398 }
399 
400 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
401 			     uint32_t __user *wptr, struct mm_struct *mm)
402 {
403 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
404 	struct v9_sdma_mqd *m;
405 	uint32_t sdma_rlc_reg_offset;
406 	unsigned long end_jiffies;
407 	uint32_t data;
408 	uint64_t data64;
409 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
410 
411 	m = get_sdma_mqd(mqd);
412 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
413 					    m->sdma_queue_id);
414 
415 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
416 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
417 
418 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
419 	while (true) {
420 		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
421 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
422 			break;
423 		if (time_after(jiffies, end_jiffies)) {
424 			pr_err("SDMA RLC not idle in %s\n", __func__);
425 			return -ETIME;
426 		}
427 		usleep_range(500, 1000);
428 	}
429 
430 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
431 	       m->sdmax_rlcx_doorbell_offset);
432 
433 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
434 			     ENABLE, 1);
435 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
436 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
437 				m->sdmax_rlcx_rb_rptr);
438 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
439 				m->sdmax_rlcx_rb_rptr_hi);
440 
441 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
442 	if (read_user_wptr(mm, wptr64, data64)) {
443 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
444 		       lower_32_bits(data64));
445 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
446 		       upper_32_bits(data64));
447 	} else {
448 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
449 		       m->sdmax_rlcx_rb_rptr);
450 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
451 		       m->sdmax_rlcx_rb_rptr_hi);
452 	}
453 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
454 
455 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
456 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
457 			m->sdmax_rlcx_rb_base_hi);
458 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
459 			m->sdmax_rlcx_rb_rptr_addr_lo);
460 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
461 			m->sdmax_rlcx_rb_rptr_addr_hi);
462 
463 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
464 			     RB_ENABLE, 1);
465 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
466 
467 	return 0;
468 }
469 
470 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
471 			     uint32_t engine_id, uint32_t queue_id,
472 			     uint32_t (**dump)[2], uint32_t *n_regs)
473 {
474 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
475 	uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
476 			engine_id, queue_id);
477 	uint32_t i = 0, reg;
478 #undef HQD_N_REGS
479 #define HQD_N_REGS (19+6+7+10)
480 
481 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
482 	if (*dump == NULL)
483 		return -ENOMEM;
484 
485 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
486 		DUMP_REG(sdma_rlc_reg_offset + reg);
487 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
488 		DUMP_REG(sdma_rlc_reg_offset + reg);
489 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
490 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
491 		DUMP_REG(sdma_rlc_reg_offset + reg);
492 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
493 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
494 		DUMP_REG(sdma_rlc_reg_offset + reg);
495 
496 	WARN_ON_ONCE(i != HQD_N_REGS);
497 	*n_regs = i;
498 
499 	return 0;
500 }
501 
502 bool kgd_gfx_v9_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
503 				uint32_t pipe_id, uint32_t queue_id)
504 {
505 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
506 	uint32_t act;
507 	bool retval = false;
508 	uint32_t low, high;
509 
510 	acquire_queue(kgd, pipe_id, queue_id);
511 	act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
512 	if (act) {
513 		low = lower_32_bits(queue_address >> 8);
514 		high = upper_32_bits(queue_address >> 8);
515 
516 		if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
517 		   high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
518 			retval = true;
519 	}
520 	release_queue(kgd);
521 	return retval;
522 }
523 
524 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
525 {
526 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
527 	struct v9_sdma_mqd *m;
528 	uint32_t sdma_rlc_reg_offset;
529 	uint32_t sdma_rlc_rb_cntl;
530 
531 	m = get_sdma_mqd(mqd);
532 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
533 					    m->sdma_queue_id);
534 
535 	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
536 
537 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
538 		return true;
539 
540 	return false;
541 }
542 
543 int kgd_gfx_v9_hqd_destroy(struct kgd_dev *kgd, void *mqd,
544 				enum kfd_preempt_type reset_type,
545 				unsigned int utimeout, uint32_t pipe_id,
546 				uint32_t queue_id)
547 {
548 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
549 	enum hqd_dequeue_request_type type;
550 	unsigned long end_jiffies;
551 	uint32_t temp;
552 	struct v9_mqd *m = get_mqd(mqd);
553 
554 	if (amdgpu_in_reset(adev))
555 		return -EIO;
556 
557 	acquire_queue(kgd, pipe_id, queue_id);
558 
559 	if (m->cp_hqd_vmid == 0)
560 		WREG32_FIELD15_RLC(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
561 
562 	switch (reset_type) {
563 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
564 		type = DRAIN_PIPE;
565 		break;
566 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
567 		type = RESET_WAVES;
568 		break;
569 	default:
570 		type = DRAIN_PIPE;
571 		break;
572 	}
573 
574 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
575 
576 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
577 	while (true) {
578 		temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
579 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
580 			break;
581 		if (time_after(jiffies, end_jiffies)) {
582 			pr_err("cp queue preemption time out.\n");
583 			release_queue(kgd);
584 			return -ETIME;
585 		}
586 		usleep_range(500, 1000);
587 	}
588 
589 	release_queue(kgd);
590 	return 0;
591 }
592 
593 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
594 				unsigned int utimeout)
595 {
596 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
597 	struct v9_sdma_mqd *m;
598 	uint32_t sdma_rlc_reg_offset;
599 	uint32_t temp;
600 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
601 
602 	m = get_sdma_mqd(mqd);
603 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
604 					    m->sdma_queue_id);
605 
606 	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
607 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
608 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
609 
610 	while (true) {
611 		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
612 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
613 			break;
614 		if (time_after(jiffies, end_jiffies)) {
615 			pr_err("SDMA RLC not idle in %s\n", __func__);
616 			return -ETIME;
617 		}
618 		usleep_range(500, 1000);
619 	}
620 
621 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
622 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
623 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
624 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
625 
626 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
627 	m->sdmax_rlcx_rb_rptr_hi =
628 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
629 
630 	return 0;
631 }
632 
633 bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
634 					uint8_t vmid, uint16_t *p_pasid)
635 {
636 	uint32_t value;
637 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
638 
639 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
640 		     + vmid);
641 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
642 
643 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
644 }
645 
646 int kgd_gfx_v9_address_watch_disable(struct kgd_dev *kgd)
647 {
648 	return 0;
649 }
650 
651 int kgd_gfx_v9_address_watch_execute(struct kgd_dev *kgd,
652 					unsigned int watch_point_id,
653 					uint32_t cntl_val,
654 					uint32_t addr_hi,
655 					uint32_t addr_lo)
656 {
657 	return 0;
658 }
659 
660 int kgd_gfx_v9_wave_control_execute(struct kgd_dev *kgd,
661 					uint32_t gfx_index_val,
662 					uint32_t sq_cmd)
663 {
664 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
665 	uint32_t data = 0;
666 
667 	mutex_lock(&adev->grbm_idx_mutex);
668 
669 	WREG32_SOC15_RLC_SHADOW(GC, 0, mmGRBM_GFX_INDEX, gfx_index_val);
670 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
671 
672 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
673 		INSTANCE_BROADCAST_WRITES, 1);
674 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
675 		SH_BROADCAST_WRITES, 1);
676 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
677 		SE_BROADCAST_WRITES, 1);
678 
679 	WREG32_SOC15_RLC_SHADOW(GC, 0, mmGRBM_GFX_INDEX, data);
680 	mutex_unlock(&adev->grbm_idx_mutex);
681 
682 	return 0;
683 }
684 
685 uint32_t kgd_gfx_v9_address_watch_get_offset(struct kgd_dev *kgd,
686 					unsigned int watch_point_id,
687 					unsigned int reg_offset)
688 {
689 	return 0;
690 }
691 
692 void kgd_gfx_v9_set_vm_context_page_table_base(struct kgd_dev *kgd,
693 			uint32_t vmid, uint64_t page_table_base)
694 {
695 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
696 
697 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
698 		pr_err("trying to set page table base for wrong VMID %u\n",
699 		       vmid);
700 		return;
701 	}
702 
703 	adev->mmhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
704 
705 	adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
706 }
707 
708 static void lock_spi_csq_mutexes(struct amdgpu_device *adev)
709 {
710 	mutex_lock(&adev->srbm_mutex);
711 	mutex_lock(&adev->grbm_idx_mutex);
712 
713 }
714 
715 static void unlock_spi_csq_mutexes(struct amdgpu_device *adev)
716 {
717 	mutex_unlock(&adev->grbm_idx_mutex);
718 	mutex_unlock(&adev->srbm_mutex);
719 }
720 
721 /**
722  * @get_wave_count: Read device registers to get number of waves in flight for
723  * a particular queue. The method also returns the VMID associated with the
724  * queue.
725  *
726  * @adev: Handle of device whose registers are to be read
727  * @queue_idx: Index of queue in the queue-map bit-field
728  * @wave_cnt: Output parameter updated with number of waves in flight
729  * @vmid: Output parameter updated with VMID of queue whose wave count
730  * is being collected
731  */
732 static void get_wave_count(struct amdgpu_device *adev, int queue_idx,
733 		int *wave_cnt, int *vmid)
734 {
735 	int pipe_idx;
736 	int queue_slot;
737 	unsigned int reg_val;
738 
739 	/*
740 	 * Program GRBM with appropriate MEID, PIPEID, QUEUEID and VMID
741 	 * parameters to read out waves in flight. Get VMID if there are
742 	 * non-zero waves in flight.
743 	 */
744 	*vmid = 0xFF;
745 	*wave_cnt = 0;
746 	pipe_idx = queue_idx / adev->gfx.mec.num_queue_per_pipe;
747 	queue_slot = queue_idx % adev->gfx.mec.num_queue_per_pipe;
748 	soc15_grbm_select(adev, 1, pipe_idx, queue_slot, 0);
749 	reg_val = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_CSQ_WF_ACTIVE_COUNT_0) +
750 			 queue_slot);
751 	*wave_cnt = reg_val & SPI_CSQ_WF_ACTIVE_COUNT_0__COUNT_MASK;
752 	if (*wave_cnt != 0)
753 		*vmid = (RREG32_SOC15(GC, 0, mmCP_HQD_VMID) &
754 			 CP_HQD_VMID__VMID_MASK) >> CP_HQD_VMID__VMID__SHIFT;
755 }
756 
757 /**
758  * @kgd_gfx_v9_get_cu_occupancy: Reads relevant registers associated with each
759  * shader engine and aggregates the number of waves that are in flight for the
760  * process whose pasid is provided as a parameter. The process could have ZERO
761  * or more queues running and submitting waves to compute units.
762  *
763  * @kgd: Handle of device from which to get number of waves in flight
764  * @pasid: Identifies the process for which this query call is invoked
765  * @wave_cnt: Output parameter updated with number of waves in flight that
766  * belong to process with given pasid
767  * @max_waves_per_cu: Output parameter updated with maximum number of waves
768  * possible per Compute Unit
769  *
770  * @note: It's possible that the device has too many queues (oversubscription)
771  * in which case a VMID could be remapped to a different PASID. This could lead
772  * to an iaccurate wave count. Following is a high-level sequence:
773  *    Time T1: vmid = getVmid(); vmid is associated with Pasid P1
774  *    Time T2: passId = getPasId(vmid); vmid is associated with Pasid P2
775  * In the sequence above wave count obtained from time T1 will be incorrectly
776  * lost or added to total wave count.
777  *
778  * The registers that provide the waves in flight are:
779  *
780  *  SPI_CSQ_WF_ACTIVE_STATUS - bit-map of queues per pipe. The bit is ON if a
781  *  queue is slotted, OFF if there is no queue. A process could have ZERO or
782  *  more queues slotted and submitting waves to be run on compute units. Even
783  *  when there is a queue it is possible there could be zero wave fronts, this
784  *  can happen when queue is waiting on top-of-pipe events - e.g. waitRegMem
785  *  command
786  *
787  *  For each bit that is ON from above:
788  *
789  *    Read (SPI_CSQ_WF_ACTIVE_COUNT_0 + queue_idx) register. It provides the
790  *    number of waves that are in flight for the queue at specified index. The
791  *    index ranges from 0 to 7.
792  *
793  *    If non-zero waves are in flight, read CP_HQD_VMID register to obtain VMID
794  *    of the wave(s).
795  *
796  *    Determine if VMID from above step maps to pasid provided as parameter. If
797  *    it matches agrregate the wave count. That the VMID will not match pasid is
798  *    a normal condition i.e. a device is expected to support multiple queues
799  *    from multiple proceses.
800  *
801  *  Reading registers referenced above involves programming GRBM appropriately
802  */
803 void kgd_gfx_v9_get_cu_occupancy(struct kgd_dev *kgd, int pasid,
804 		int *pasid_wave_cnt, int *max_waves_per_cu)
805 {
806 	int qidx;
807 	int vmid;
808 	int se_idx;
809 	int sh_idx;
810 	int se_cnt;
811 	int sh_cnt;
812 	int wave_cnt;
813 	int queue_map;
814 	int pasid_tmp;
815 	int max_queue_cnt;
816 	int vmid_wave_cnt = 0;
817 	struct amdgpu_device *adev;
818 	DECLARE_BITMAP(cp_queue_bitmap, KGD_MAX_QUEUES);
819 
820 	adev = get_amdgpu_device(kgd);
821 	lock_spi_csq_mutexes(adev);
822 	soc15_grbm_select(adev, 1, 0, 0, 0);
823 
824 	/*
825 	 * Iterate through the shader engines and arrays of the device
826 	 * to get number of waves in flight
827 	 */
828 	bitmap_complement(cp_queue_bitmap, adev->gfx.mec.queue_bitmap,
829 			  KGD_MAX_QUEUES);
830 	max_queue_cnt = adev->gfx.mec.num_pipe_per_mec *
831 			adev->gfx.mec.num_queue_per_pipe;
832 	sh_cnt = adev->gfx.config.max_sh_per_se;
833 	se_cnt = adev->gfx.config.max_shader_engines;
834 	for (se_idx = 0; se_idx < se_cnt; se_idx++) {
835 		for (sh_idx = 0; sh_idx < sh_cnt; sh_idx++) {
836 
837 			gfx_v9_0_select_se_sh(adev, se_idx, sh_idx, 0xffffffff);
838 			queue_map = RREG32(SOC15_REG_OFFSET(GC, 0,
839 					   mmSPI_CSQ_WF_ACTIVE_STATUS));
840 
841 			/*
842 			 * Assumption: queue map encodes following schema: four
843 			 * pipes per each micro-engine, with each pipe mapping
844 			 * eight queues. This schema is true for GFX9 devices
845 			 * and must be verified for newer device families
846 			 */
847 			for (qidx = 0; qidx < max_queue_cnt; qidx++) {
848 
849 				/* Skip qeueus that are not associated with
850 				 * compute functions
851 				 */
852 				if (!test_bit(qidx, cp_queue_bitmap))
853 					continue;
854 
855 				if (!(queue_map & (1 << qidx)))
856 					continue;
857 
858 				/* Get number of waves in flight and aggregate them */
859 				get_wave_count(adev, qidx, &wave_cnt, &vmid);
860 				if (wave_cnt != 0) {
861 					pasid_tmp =
862 					  RREG32(SOC15_REG_OFFSET(OSSSYS, 0,
863 						 mmIH_VMID_0_LUT) + vmid);
864 					if (pasid_tmp == pasid)
865 						vmid_wave_cnt += wave_cnt;
866 				}
867 			}
868 		}
869 	}
870 
871 	gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
872 	soc15_grbm_select(adev, 0, 0, 0, 0);
873 	unlock_spi_csq_mutexes(adev);
874 
875 	/* Update the output parameters and return */
876 	*pasid_wave_cnt = vmid_wave_cnt;
877 	*max_waves_per_cu = adev->gfx.cu_info.simd_per_cu *
878 				adev->gfx.cu_info.max_waves_per_simd;
879 }
880 
881 const struct kfd2kgd_calls gfx_v9_kfd2kgd = {
882 	.program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
883 	.set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping,
884 	.init_interrupts = kgd_gfx_v9_init_interrupts,
885 	.hqd_load = kgd_gfx_v9_hqd_load,
886 	.hiq_mqd_load = kgd_gfx_v9_hiq_mqd_load,
887 	.hqd_sdma_load = kgd_hqd_sdma_load,
888 	.hqd_dump = kgd_gfx_v9_hqd_dump,
889 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
890 	.hqd_is_occupied = kgd_gfx_v9_hqd_is_occupied,
891 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
892 	.hqd_destroy = kgd_gfx_v9_hqd_destroy,
893 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
894 	.address_watch_disable = kgd_gfx_v9_address_watch_disable,
895 	.address_watch_execute = kgd_gfx_v9_address_watch_execute,
896 	.wave_control_execute = kgd_gfx_v9_wave_control_execute,
897 	.address_watch_get_offset = kgd_gfx_v9_address_watch_get_offset,
898 	.get_atc_vmid_pasid_mapping_info =
899 			kgd_gfx_v9_get_atc_vmid_pasid_mapping_info,
900 	.set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base,
901 	.get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,
902 };
903