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