1 /*
2  * Copyright 2019 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_10_1_0_offset.h"
25 #include "gc/gc_10_1_0_sh_mask.h"
26 #include "athub/athub_2_0_0_offset.h"
27 #include "athub/athub_2_0_0_sh_mask.h"
28 #include "oss/osssys_5_0_0_offset.h"
29 #include "oss/osssys_5_0_0_sh_mask.h"
30 #include "soc15_common.h"
31 #include "v10_structs.h"
32 #include "nv.h"
33 #include "nvd.h"
34 
35 enum hqd_dequeue_request_type {
36 	NO_ACTION = 0,
37 	DRAIN_PIPE,
38 	RESET_WAVES,
39 	SAVE_WAVES
40 };
41 
42 static void lock_srbm(struct amdgpu_device *adev, uint32_t mec, uint32_t pipe,
43 			uint32_t queue, uint32_t vmid)
44 {
45 	mutex_lock(&adev->srbm_mutex);
46 	nv_grbm_select(adev, mec, pipe, queue, vmid);
47 }
48 
49 static void unlock_srbm(struct amdgpu_device *adev)
50 {
51 	nv_grbm_select(adev, 0, 0, 0, 0);
52 	mutex_unlock(&adev->srbm_mutex);
53 }
54 
55 static void acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id,
56 				uint32_t queue_id)
57 {
58 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
59 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
60 
61 	lock_srbm(adev, mec, pipe, queue_id, 0);
62 }
63 
64 static uint64_t get_queue_mask(struct amdgpu_device *adev,
65 			       uint32_t pipe_id, uint32_t queue_id)
66 {
67 	unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe +
68 			queue_id;
69 
70 	return 1ull << bit;
71 }
72 
73 static void release_queue(struct amdgpu_device *adev)
74 {
75 	unlock_srbm(adev);
76 }
77 
78 static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmid,
79 					uint32_t sh_mem_config,
80 					uint32_t sh_mem_ape1_base,
81 					uint32_t sh_mem_ape1_limit,
82 					uint32_t sh_mem_bases)
83 {
84 	lock_srbm(adev, 0, 0, 0, vmid);
85 
86 	WREG32_SOC15(GC, 0, mmSH_MEM_CONFIG, sh_mem_config);
87 	WREG32_SOC15(GC, 0, mmSH_MEM_BASES, sh_mem_bases);
88 	/* APE1 no longer exists on GFX9 */
89 
90 	unlock_srbm(adev);
91 }
92 
93 static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid,
94 					unsigned int vmid)
95 {
96 	/*
97 	 * We have to assume that there is no outstanding mapping.
98 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
99 	 * a mapping is in progress or because a mapping finished
100 	 * and the SW cleared it.
101 	 * So the protocol is to always wait & clear.
102 	 */
103 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
104 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
105 
106 	pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping);
107 
108 	pr_debug("ATHUB, reg %x\n", SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid);
109 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
110 	       pasid_mapping);
111 
112 #if 0
113 	/* TODO: uncomment this code when the hardware support is ready. */
114 	while (!(RREG32(SOC15_REG_OFFSET(
115 				ATHUB, 0,
116 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
117 		 (1U << vmid)))
118 		cpu_relax();
119 
120 	pr_debug("ATHUB mapping update finished\n");
121 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
122 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
123 	       1U << vmid);
124 #endif
125 
126 	/* Mapping vmid to pasid also for IH block */
127 	pr_debug("update mapping for IH block and mmhub");
128 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
129 	       pasid_mapping);
130 
131 	return 0;
132 }
133 
134 /* TODO - RING0 form of field is obsolete, seems to date back to SI
135  * but still works
136  */
137 
138 static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id)
139 {
140 	uint32_t mec;
141 	uint32_t pipe;
142 
143 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
144 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
145 
146 	lock_srbm(adev, mec, pipe, 0, 0);
147 
148 	WREG32_SOC15(GC, 0, mmCPC_INT_CNTL,
149 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
150 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
151 
152 	unlock_srbm(adev);
153 
154 	return 0;
155 }
156 
157 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
158 				unsigned int engine_id,
159 				unsigned int queue_id)
160 {
161 	uint32_t sdma_engine_reg_base[2] = {
162 		SOC15_REG_OFFSET(SDMA0, 0,
163 				 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
164 		/* On gfx10, mmSDMA1_xxx registers are defined NOT based
165 		 * on SDMA1 base address (dw 0x1860) but based on SDMA0
166 		 * base address (dw 0x1260). Therefore use mmSDMA0_RLC0_RB_CNTL
167 		 * instead of mmSDMA1_RLC0_RB_CNTL for the base address calc
168 		 * below
169 		 */
170 		SOC15_REG_OFFSET(SDMA1, 0,
171 				 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL
172 	};
173 
174 	uint32_t retval = sdma_engine_reg_base[engine_id]
175 		+ queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
176 
177 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
178 			queue_id, retval);
179 
180 	return retval;
181 }
182 
183 #if 0
184 static uint32_t get_watch_base_addr(struct amdgpu_device *adev)
185 {
186 	uint32_t retval = SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) -
187 			mmTCP_WATCH0_ADDR_H;
188 
189 	pr_debug("kfd: reg watch base address: 0x%x\n", retval);
190 
191 	return retval;
192 }
193 #endif
194 
195 static inline struct v10_compute_mqd *get_mqd(void *mqd)
196 {
197 	return (struct v10_compute_mqd *)mqd;
198 }
199 
200 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
201 {
202 	return (struct v10_sdma_mqd *)mqd;
203 }
204 
205 static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd,
206 			uint32_t pipe_id, uint32_t queue_id,
207 			uint32_t __user *wptr, uint32_t wptr_shift,
208 			uint32_t wptr_mask, struct mm_struct *mm)
209 {
210 	struct v10_compute_mqd *m;
211 	uint32_t *mqd_hqd;
212 	uint32_t reg, hqd_base, data;
213 
214 	m = get_mqd(mqd);
215 
216 	pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
217 	acquire_queue(adev, pipe_id, queue_id);
218 
219 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
220 	mqd_hqd = &m->cp_mqd_base_addr_lo;
221 	hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
222 
223 	for (reg = hqd_base;
224 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
225 		WREG32_SOC15_IP(GC, reg, mqd_hqd[reg - hqd_base]);
226 
227 
228 	/* Activate doorbell logic before triggering WPTR poll. */
229 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
230 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
231 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL, data);
232 
233 	if (wptr) {
234 		/* Don't read wptr with get_user because the user
235 		 * context may not be accessible (if this function
236 		 * runs in a work queue). Instead trigger a one-shot
237 		 * polling read from memory in the CP. This assumes
238 		 * that wptr is GPU-accessible in the queue's VMID via
239 		 * ATC or SVM. WPTR==RPTR before starting the poll so
240 		 * the CP starts fetching new commands from the right
241 		 * place.
242 		 *
243 		 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
244 		 * tricky. Assume that the queue didn't overflow. The
245 		 * number of valid bits in the 32-bit RPTR depends on
246 		 * the queue size. The remaining bits are taken from
247 		 * the saved 64-bit WPTR. If the WPTR wrapped, add the
248 		 * queue size.
249 		 */
250 		uint32_t queue_size =
251 			2 << REG_GET_FIELD(m->cp_hqd_pq_control,
252 					   CP_HQD_PQ_CONTROL, QUEUE_SIZE);
253 		uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
254 
255 		if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
256 			guessed_wptr += queue_size;
257 		guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
258 		guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
259 
260 		WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_LO,
261 		       lower_32_bits(guessed_wptr));
262 		WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_HI,
263 		       upper_32_bits(guessed_wptr));
264 		WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR,
265 		       lower_32_bits((uint64_t)wptr));
266 		WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI,
267 		       upper_32_bits((uint64_t)wptr));
268 		pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__,
269 			 (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
270 		WREG32_SOC15(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1,
271 		       (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
272 	}
273 
274 	/* Start the EOP fetcher */
275 	WREG32_SOC15(GC, 0, mmCP_HQD_EOP_RPTR,
276 	       REG_SET_FIELD(m->cp_hqd_eop_rptr,
277 			     CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
278 
279 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
280 	WREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE, data);
281 
282 	release_queue(adev);
283 
284 	return 0;
285 }
286 
287 static int kgd_hiq_mqd_load(struct amdgpu_device *adev, void *mqd,
288 			    uint32_t pipe_id, uint32_t queue_id,
289 			    uint32_t doorbell_off)
290 {
291 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
292 	struct v10_compute_mqd *m;
293 	uint32_t mec, pipe;
294 	int r;
295 
296 	m = get_mqd(mqd);
297 
298 	acquire_queue(adev, pipe_id, queue_id);
299 
300 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
301 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
302 
303 	pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
304 		 mec, pipe, queue_id);
305 
306 	spin_lock(&adev->gfx.kiq.ring_lock);
307 	r = amdgpu_ring_alloc(kiq_ring, 7);
308 	if (r) {
309 		pr_err("Failed to alloc KIQ (%d).\n", r);
310 		goto out_unlock;
311 	}
312 
313 	amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
314 	amdgpu_ring_write(kiq_ring,
315 			  PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
316 			  PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */
317 			  PACKET3_MAP_QUEUES_QUEUE(queue_id) |
318 			  PACKET3_MAP_QUEUES_PIPE(pipe) |
319 			  PACKET3_MAP_QUEUES_ME((mec - 1)) |
320 			  PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
321 			  PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
322 			  PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */
323 			  PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
324 	amdgpu_ring_write(kiq_ring,
325 			  PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off));
326 	amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo);
327 	amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi);
328 	amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo);
329 	amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi);
330 	amdgpu_ring_commit(kiq_ring);
331 
332 out_unlock:
333 	spin_unlock(&adev->gfx.kiq.ring_lock);
334 	release_queue(adev);
335 
336 	return r;
337 }
338 
339 static int kgd_hqd_dump(struct amdgpu_device *adev,
340 			uint32_t pipe_id, uint32_t queue_id,
341 			uint32_t (**dump)[2], uint32_t *n_regs)
342 {
343 	uint32_t i = 0, reg;
344 #define HQD_N_REGS 56
345 #define DUMP_REG(addr) do {				\
346 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
347 			break;				\
348 		(*dump)[i][0] = (addr) << 2;		\
349 		(*dump)[i++][1] = RREG32_SOC15_IP(GC, addr);		\
350 	} while (0)
351 
352 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
353 	if (*dump == NULL)
354 		return -ENOMEM;
355 
356 	acquire_queue(adev, pipe_id, queue_id);
357 
358 	for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
359 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
360 		DUMP_REG(reg);
361 
362 	release_queue(adev);
363 
364 	WARN_ON_ONCE(i != HQD_N_REGS);
365 	*n_regs = i;
366 
367 	return 0;
368 }
369 
370 static int kgd_hqd_sdma_load(struct amdgpu_device *adev, void *mqd,
371 			     uint32_t __user *wptr, struct mm_struct *mm)
372 {
373 	struct v10_sdma_mqd *m;
374 	uint32_t sdma_rlc_reg_offset;
375 	unsigned long end_jiffies;
376 	uint32_t data;
377 	uint64_t data64;
378 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
379 
380 	m = get_sdma_mqd(mqd);
381 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
382 					    m->sdma_queue_id);
383 
384 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
385 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
386 
387 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
388 	while (true) {
389 		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
390 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
391 			break;
392 		if (time_after(jiffies, end_jiffies)) {
393 			pr_err("SDMA RLC not idle in %s\n", __func__);
394 			return -ETIME;
395 		}
396 		usleep_range(500, 1000);
397 	}
398 
399 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
400 	       m->sdmax_rlcx_doorbell_offset);
401 
402 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
403 			     ENABLE, 1);
404 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
405 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
406 				m->sdmax_rlcx_rb_rptr);
407 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
408 				m->sdmax_rlcx_rb_rptr_hi);
409 
410 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
411 	if (read_user_wptr(mm, wptr64, data64)) {
412 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
413 		       lower_32_bits(data64));
414 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
415 		       upper_32_bits(data64));
416 	} else {
417 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
418 		       m->sdmax_rlcx_rb_rptr);
419 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
420 		       m->sdmax_rlcx_rb_rptr_hi);
421 	}
422 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
423 
424 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
425 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
426 			m->sdmax_rlcx_rb_base_hi);
427 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
428 			m->sdmax_rlcx_rb_rptr_addr_lo);
429 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
430 			m->sdmax_rlcx_rb_rptr_addr_hi);
431 
432 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
433 			     RB_ENABLE, 1);
434 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
435 
436 	return 0;
437 }
438 
439 static int kgd_hqd_sdma_dump(struct amdgpu_device *adev,
440 			     uint32_t engine_id, uint32_t queue_id,
441 			     uint32_t (**dump)[2], uint32_t *n_regs)
442 {
443 	uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
444 			engine_id, queue_id);
445 	uint32_t i = 0, reg;
446 #undef HQD_N_REGS
447 #define HQD_N_REGS (19+6+7+10)
448 
449 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
450 	if (*dump == NULL)
451 		return -ENOMEM;
452 
453 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
454 		DUMP_REG(sdma_rlc_reg_offset + reg);
455 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
456 		DUMP_REG(sdma_rlc_reg_offset + reg);
457 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
458 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
459 		DUMP_REG(sdma_rlc_reg_offset + reg);
460 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
461 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
462 		DUMP_REG(sdma_rlc_reg_offset + reg);
463 
464 	WARN_ON_ONCE(i != HQD_N_REGS);
465 	*n_regs = i;
466 
467 	return 0;
468 }
469 
470 static bool kgd_hqd_is_occupied(struct amdgpu_device *adev,
471 				uint64_t queue_address, uint32_t pipe_id,
472 				uint32_t queue_id)
473 {
474 	uint32_t act;
475 	bool retval = false;
476 	uint32_t low, high;
477 
478 	acquire_queue(adev, pipe_id, queue_id);
479 	act = RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE);
480 	if (act) {
481 		low = lower_32_bits(queue_address >> 8);
482 		high = upper_32_bits(queue_address >> 8);
483 
484 		if (low == RREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE) &&
485 		   high == RREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE_HI))
486 			retval = true;
487 	}
488 	release_queue(adev);
489 	return retval;
490 }
491 
492 static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd)
493 {
494 	struct v10_sdma_mqd *m;
495 	uint32_t sdma_rlc_reg_offset;
496 	uint32_t sdma_rlc_rb_cntl;
497 
498 	m = get_sdma_mqd(mqd);
499 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
500 					    m->sdma_queue_id);
501 
502 	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
503 
504 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
505 		return true;
506 
507 	return false;
508 }
509 
510 static int kgd_hqd_destroy(struct amdgpu_device *adev, void *mqd,
511 				enum kfd_preempt_type reset_type,
512 				unsigned int utimeout, uint32_t pipe_id,
513 				uint32_t queue_id)
514 {
515 	enum hqd_dequeue_request_type type;
516 	unsigned long end_jiffies;
517 	uint32_t temp;
518 	struct v10_compute_mqd *m = get_mqd(mqd);
519 
520 	if (amdgpu_in_reset(adev))
521 		return -EIO;
522 
523 #if 0
524 	unsigned long flags;
525 	int retry;
526 #endif
527 
528 	acquire_queue(adev, pipe_id, queue_id);
529 
530 	if (m->cp_hqd_vmid == 0)
531 		WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
532 
533 	switch (reset_type) {
534 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
535 		type = DRAIN_PIPE;
536 		break;
537 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
538 		type = RESET_WAVES;
539 		break;
540 	case KFD_PREEMPT_TYPE_WAVEFRONT_SAVE:
541 		type = SAVE_WAVES;
542 		break;
543 	default:
544 		type = DRAIN_PIPE;
545 		break;
546 	}
547 
548 #if 0 /* Is this still needed? */
549 	/* Workaround: If IQ timer is active and the wait time is close to or
550 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
551 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
552 	 * cleared before continuing. Also, ensure wait times are set to at
553 	 * least 0x3.
554 	 */
555 	local_irq_save(flags);
556 	preempt_disable();
557 	retry = 5000; /* wait for 500 usecs at maximum */
558 	while (true) {
559 		temp = RREG32(mmCP_HQD_IQ_TIMER);
560 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
561 			pr_debug("HW is processing IQ\n");
562 			goto loop;
563 		}
564 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
565 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
566 					== 3) /* SEM-rearm is safe */
567 				break;
568 			/* Wait time 3 is safe for CP, but our MMIO read/write
569 			 * time is close to 1 microsecond, so check for 10 to
570 			 * leave more buffer room
571 			 */
572 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
573 					>= 10)
574 				break;
575 			pr_debug("IQ timer is active\n");
576 		} else
577 			break;
578 loop:
579 		if (!retry) {
580 			pr_err("CP HQD IQ timer status time out\n");
581 			break;
582 		}
583 		ndelay(100);
584 		--retry;
585 	}
586 	retry = 1000;
587 	while (true) {
588 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
589 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
590 			break;
591 		pr_debug("Dequeue request is pending\n");
592 
593 		if (!retry) {
594 			pr_err("CP HQD dequeue request time out\n");
595 			break;
596 		}
597 		ndelay(100);
598 		--retry;
599 	}
600 	local_irq_restore(flags);
601 	preempt_enable();
602 #endif
603 
604 	WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, type);
605 
606 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
607 	while (true) {
608 		temp = RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE);
609 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
610 			break;
611 		if (time_after(jiffies, end_jiffies)) {
612 			pr_err("cp queue preemption time out.\n");
613 			release_queue(adev);
614 			return -ETIME;
615 		}
616 		usleep_range(500, 1000);
617 	}
618 
619 	release_queue(adev);
620 	return 0;
621 }
622 
623 static int kgd_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd,
624 				unsigned int utimeout)
625 {
626 	struct v10_sdma_mqd *m;
627 	uint32_t sdma_rlc_reg_offset;
628 	uint32_t temp;
629 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
630 
631 	m = get_sdma_mqd(mqd);
632 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
633 					    m->sdma_queue_id);
634 
635 	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
636 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
637 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
638 
639 	while (true) {
640 		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
641 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
642 			break;
643 		if (time_after(jiffies, end_jiffies)) {
644 			pr_err("SDMA RLC not idle in %s\n", __func__);
645 			return -ETIME;
646 		}
647 		usleep_range(500, 1000);
648 	}
649 
650 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
651 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
652 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
653 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
654 
655 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
656 	m->sdmax_rlcx_rb_rptr_hi =
657 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
658 
659 	return 0;
660 }
661 
662 static bool get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
663 					uint8_t vmid, uint16_t *p_pasid)
664 {
665 	uint32_t value;
666 
667 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
668 		     + vmid);
669 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
670 
671 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
672 }
673 
674 static int kgd_wave_control_execute(struct amdgpu_device *adev,
675 					uint32_t gfx_index_val,
676 					uint32_t sq_cmd)
677 {
678 	uint32_t data = 0;
679 
680 	mutex_lock(&adev->grbm_idx_mutex);
681 
682 	WREG32_SOC15(GC, 0, mmGRBM_GFX_INDEX, gfx_index_val);
683 	WREG32_SOC15(GC, 0, mmSQ_CMD, sq_cmd);
684 
685 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
686 		INSTANCE_BROADCAST_WRITES, 1);
687 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
688 		SA_BROADCAST_WRITES, 1);
689 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
690 		SE_BROADCAST_WRITES, 1);
691 
692 	WREG32_SOC15(GC, 0, mmGRBM_GFX_INDEX, data);
693 	mutex_unlock(&adev->grbm_idx_mutex);
694 
695 	return 0;
696 }
697 
698 static void set_vm_context_page_table_base(struct amdgpu_device *adev,
699 		uint32_t vmid, uint64_t page_table_base)
700 {
701 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
702 		pr_err("trying to set page table base for wrong VMID %u\n",
703 		       vmid);
704 		return;
705 	}
706 
707 	/* SDMA is on gfxhub as well for Navi1* series */
708 	adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
709 }
710 
711 static void program_trap_handler_settings(struct amdgpu_device *adev,
712 		uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr)
713 {
714 	lock_srbm(adev, 0, 0, 0, vmid);
715 
716 	/*
717 	 * Program TBA registers
718 	 */
719 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_LO),
720 			lower_32_bits(tba_addr >> 8));
721 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_HI),
722 			upper_32_bits(tba_addr >> 8) |
723 			(1 << SQ_SHADER_TBA_HI__TRAP_EN__SHIFT));
724 
725 	/*
726 	 * Program TMA registers
727 	 */
728 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_LO),
729 			lower_32_bits(tma_addr >> 8));
730 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_HI),
731 			upper_32_bits(tma_addr >> 8));
732 
733 	unlock_srbm(adev);
734 }
735 
736 const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
737 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
738 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
739 	.init_interrupts = kgd_init_interrupts,
740 	.hqd_load = kgd_hqd_load,
741 	.hiq_mqd_load = kgd_hiq_mqd_load,
742 	.hqd_sdma_load = kgd_hqd_sdma_load,
743 	.hqd_dump = kgd_hqd_dump,
744 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
745 	.hqd_is_occupied = kgd_hqd_is_occupied,
746 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
747 	.hqd_destroy = kgd_hqd_destroy,
748 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
749 	.wave_control_execute = kgd_wave_control_execute,
750 	.get_atc_vmid_pasid_mapping_info =
751 			get_atc_vmid_pasid_mapping_info,
752 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
753 	.program_trap_handler_settings = program_trap_handler_settings,
754 };
755