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 inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
43 {
44 	return (struct amdgpu_device *)kgd;
45 }
46 
47 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
48 			uint32_t queue, uint32_t vmid)
49 {
50 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
51 
52 	mutex_lock(&adev->srbm_mutex);
53 	nv_grbm_select(adev, mec, pipe, queue, vmid);
54 }
55 
56 static void unlock_srbm(struct kgd_dev *kgd)
57 {
58 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
59 
60 	nv_grbm_select(adev, 0, 0, 0, 0);
61 	mutex_unlock(&adev->srbm_mutex);
62 }
63 
64 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
65 				uint32_t queue_id)
66 {
67 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
68 
69 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
70 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
71 
72 	lock_srbm(kgd, mec, pipe, queue_id, 0);
73 }
74 
75 static uint64_t get_queue_mask(struct amdgpu_device *adev,
76 			       uint32_t pipe_id, uint32_t queue_id)
77 {
78 	unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe +
79 			queue_id;
80 
81 	return 1ull << bit;
82 }
83 
84 static void release_queue(struct kgd_dev *kgd)
85 {
86 	unlock_srbm(kgd);
87 }
88 
89 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
90 					uint32_t sh_mem_config,
91 					uint32_t sh_mem_ape1_base,
92 					uint32_t sh_mem_ape1_limit,
93 					uint32_t sh_mem_bases)
94 {
95 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
96 
97 	lock_srbm(kgd, 0, 0, 0, vmid);
98 
99 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
100 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
101 	/* APE1 no longer exists on GFX9 */
102 
103 	unlock_srbm(kgd);
104 }
105 
106 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, u32 pasid,
107 					unsigned int vmid)
108 {
109 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
110 
111 	/*
112 	 * We have to assume that there is no outstanding mapping.
113 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
114 	 * a mapping is in progress or because a mapping finished
115 	 * and the SW cleared it.
116 	 * So the protocol is to always wait & clear.
117 	 */
118 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
119 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
120 
121 	pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping);
122 
123 	pr_debug("ATHUB, reg %x\n", SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid);
124 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
125 	       pasid_mapping);
126 
127 #if 0
128 	/* TODO: uncomment this code when the hardware support is ready. */
129 	while (!(RREG32(SOC15_REG_OFFSET(
130 				ATHUB, 0,
131 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
132 		 (1U << vmid)))
133 		cpu_relax();
134 
135 	pr_debug("ATHUB mapping update finished\n");
136 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
137 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
138 	       1U << vmid);
139 #endif
140 
141 	/* Mapping vmid to pasid also for IH block */
142 	pr_debug("update mapping for IH block and mmhub");
143 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
144 	       pasid_mapping);
145 
146 	return 0;
147 }
148 
149 /* TODO - RING0 form of field is obsolete, seems to date back to SI
150  * but still works
151  */
152 
153 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
154 {
155 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
156 	uint32_t mec;
157 	uint32_t pipe;
158 
159 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
160 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
161 
162 	lock_srbm(kgd, mec, pipe, 0, 0);
163 
164 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
165 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
166 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
167 
168 	unlock_srbm(kgd);
169 
170 	return 0;
171 }
172 
173 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
174 				unsigned int engine_id,
175 				unsigned int queue_id)
176 {
177 	uint32_t sdma_engine_reg_base[2] = {
178 		SOC15_REG_OFFSET(SDMA0, 0,
179 				 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
180 		/* On gfx10, mmSDMA1_xxx registers are defined NOT based
181 		 * on SDMA1 base address (dw 0x1860) but based on SDMA0
182 		 * base address (dw 0x1260). Therefore use mmSDMA0_RLC0_RB_CNTL
183 		 * instead of mmSDMA1_RLC0_RB_CNTL for the base address calc
184 		 * below
185 		 */
186 		SOC15_REG_OFFSET(SDMA1, 0,
187 				 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL
188 	};
189 
190 	uint32_t retval = sdma_engine_reg_base[engine_id]
191 		+ queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
192 
193 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
194 			queue_id, retval);
195 
196 	return retval;
197 }
198 
199 #if 0
200 static uint32_t get_watch_base_addr(struct amdgpu_device *adev)
201 {
202 	uint32_t retval = SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) -
203 			mmTCP_WATCH0_ADDR_H;
204 
205 	pr_debug("kfd: reg watch base address: 0x%x\n", retval);
206 
207 	return retval;
208 }
209 #endif
210 
211 static inline struct v10_compute_mqd *get_mqd(void *mqd)
212 {
213 	return (struct v10_compute_mqd *)mqd;
214 }
215 
216 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
217 {
218 	return (struct v10_sdma_mqd *)mqd;
219 }
220 
221 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
222 			uint32_t queue_id, uint32_t __user *wptr,
223 			uint32_t wptr_shift, uint32_t wptr_mask,
224 			struct mm_struct *mm)
225 {
226 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
227 	struct v10_compute_mqd *m;
228 	uint32_t *mqd_hqd;
229 	uint32_t reg, hqd_base, data;
230 
231 	m = get_mqd(mqd);
232 
233 	pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
234 	acquire_queue(kgd, pipe_id, queue_id);
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, 0, mmCP_MQD_BASE_ADDR);
239 
240 	for (reg = hqd_base;
241 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
242 		WREG32(reg, mqd_hqd[reg - hqd_base]);
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_REG_OFFSET(GC, 0, 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_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
278 		       lower_32_bits(guessed_wptr));
279 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
280 		       upper_32_bits(guessed_wptr));
281 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
282 		       lower_32_bits((uint64_t)wptr));
283 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
284 		       upper_32_bits((uint64_t)wptr));
285 		pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__,
286 			 (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
287 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
288 		       (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
289 	}
290 
291 	/* Start the EOP fetcher */
292 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
293 	       REG_SET_FIELD(m->cp_hqd_eop_rptr,
294 			     CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
295 
296 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
297 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
298 
299 	release_queue(kgd);
300 
301 	return 0;
302 }
303 
304 static int kgd_hiq_mqd_load(struct kgd_dev *kgd, void *mqd,
305 			    uint32_t pipe_id, uint32_t queue_id,
306 			    uint32_t doorbell_off)
307 {
308 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
309 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
310 	struct v10_compute_mqd *m;
311 	uint32_t mec, pipe;
312 	int r;
313 
314 	m = get_mqd(mqd);
315 
316 	acquire_queue(kgd, pipe_id, queue_id);
317 
318 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
319 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
320 
321 	pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
322 		 mec, pipe, queue_id);
323 
324 	spin_lock(&adev->gfx.kiq.ring_lock);
325 	r = amdgpu_ring_alloc(kiq_ring, 7);
326 	if (r) {
327 		pr_err("Failed to alloc KIQ (%d).\n", r);
328 		goto out_unlock;
329 	}
330 
331 	amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
332 	amdgpu_ring_write(kiq_ring,
333 			  PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
334 			  PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */
335 			  PACKET3_MAP_QUEUES_QUEUE(queue_id) |
336 			  PACKET3_MAP_QUEUES_PIPE(pipe) |
337 			  PACKET3_MAP_QUEUES_ME((mec - 1)) |
338 			  PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
339 			  PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
340 			  PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */
341 			  PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
342 	amdgpu_ring_write(kiq_ring,
343 			  PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off));
344 	amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo);
345 	amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi);
346 	amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo);
347 	amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi);
348 	amdgpu_ring_commit(kiq_ring);
349 
350 out_unlock:
351 	spin_unlock(&adev->gfx.kiq.ring_lock);
352 	release_queue(kgd);
353 
354 	return r;
355 }
356 
357 static int kgd_hqd_dump(struct kgd_dev *kgd,
358 			uint32_t pipe_id, uint32_t queue_id,
359 			uint32_t (**dump)[2], uint32_t *n_regs)
360 {
361 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
362 	uint32_t i = 0, reg;
363 #define HQD_N_REGS 56
364 #define DUMP_REG(addr) do {				\
365 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
366 			break;				\
367 		(*dump)[i][0] = (addr) << 2;		\
368 		(*dump)[i++][1] = RREG32(addr);		\
369 	} while (0)
370 
371 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
372 	if (*dump == NULL)
373 		return -ENOMEM;
374 
375 	acquire_queue(kgd, pipe_id, queue_id);
376 
377 	for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
378 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
379 		DUMP_REG(reg);
380 
381 	release_queue(kgd);
382 
383 	WARN_ON_ONCE(i != HQD_N_REGS);
384 	*n_regs = i;
385 
386 	return 0;
387 }
388 
389 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
390 			     uint32_t __user *wptr, struct mm_struct *mm)
391 {
392 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
393 	struct v10_sdma_mqd *m;
394 	uint32_t sdma_rlc_reg_offset;
395 	unsigned long end_jiffies;
396 	uint32_t data;
397 	uint64_t data64;
398 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
399 
400 	m = get_sdma_mqd(mqd);
401 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
402 					    m->sdma_queue_id);
403 
404 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
405 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
406 
407 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
408 	while (true) {
409 		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
410 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
411 			break;
412 		if (time_after(jiffies, end_jiffies)) {
413 			pr_err("SDMA RLC not idle in %s\n", __func__);
414 			return -ETIME;
415 		}
416 		usleep_range(500, 1000);
417 	}
418 
419 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
420 	       m->sdmax_rlcx_doorbell_offset);
421 
422 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
423 			     ENABLE, 1);
424 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
425 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
426 				m->sdmax_rlcx_rb_rptr);
427 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
428 				m->sdmax_rlcx_rb_rptr_hi);
429 
430 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
431 	if (read_user_wptr(mm, wptr64, data64)) {
432 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
433 		       lower_32_bits(data64));
434 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
435 		       upper_32_bits(data64));
436 	} else {
437 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
438 		       m->sdmax_rlcx_rb_rptr);
439 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
440 		       m->sdmax_rlcx_rb_rptr_hi);
441 	}
442 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
443 
444 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
445 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
446 			m->sdmax_rlcx_rb_base_hi);
447 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
448 			m->sdmax_rlcx_rb_rptr_addr_lo);
449 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
450 			m->sdmax_rlcx_rb_rptr_addr_hi);
451 
452 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
453 			     RB_ENABLE, 1);
454 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
455 
456 	return 0;
457 }
458 
459 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
460 			     uint32_t engine_id, uint32_t queue_id,
461 			     uint32_t (**dump)[2], uint32_t *n_regs)
462 {
463 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
464 	uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
465 			engine_id, queue_id);
466 	uint32_t i = 0, reg;
467 #undef HQD_N_REGS
468 #define HQD_N_REGS (19+6+7+10)
469 
470 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
471 	if (*dump == NULL)
472 		return -ENOMEM;
473 
474 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
475 		DUMP_REG(sdma_rlc_reg_offset + reg);
476 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
477 		DUMP_REG(sdma_rlc_reg_offset + reg);
478 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
479 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
480 		DUMP_REG(sdma_rlc_reg_offset + reg);
481 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
482 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
483 		DUMP_REG(sdma_rlc_reg_offset + reg);
484 
485 	WARN_ON_ONCE(i != HQD_N_REGS);
486 	*n_regs = i;
487 
488 	return 0;
489 }
490 
491 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
492 				uint32_t pipe_id, uint32_t queue_id)
493 {
494 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
495 	uint32_t act;
496 	bool retval = false;
497 	uint32_t low, high;
498 
499 	acquire_queue(kgd, pipe_id, queue_id);
500 	act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
501 	if (act) {
502 		low = lower_32_bits(queue_address >> 8);
503 		high = upper_32_bits(queue_address >> 8);
504 
505 		if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
506 		   high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
507 			retval = true;
508 	}
509 	release_queue(kgd);
510 	return retval;
511 }
512 
513 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
514 {
515 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
516 	struct v10_sdma_mqd *m;
517 	uint32_t sdma_rlc_reg_offset;
518 	uint32_t sdma_rlc_rb_cntl;
519 
520 	m = get_sdma_mqd(mqd);
521 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
522 					    m->sdma_queue_id);
523 
524 	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
525 
526 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
527 		return true;
528 
529 	return false;
530 }
531 
532 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
533 				enum kfd_preempt_type reset_type,
534 				unsigned int utimeout, uint32_t pipe_id,
535 				uint32_t queue_id)
536 {
537 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
538 	enum hqd_dequeue_request_type type;
539 	unsigned long end_jiffies;
540 	uint32_t temp;
541 	struct v10_compute_mqd *m = get_mqd(mqd);
542 
543 	if (amdgpu_in_reset(adev))
544 		return -EIO;
545 
546 #if 0
547 	unsigned long flags;
548 	int retry;
549 #endif
550 
551 	acquire_queue(kgd, pipe_id, queue_id);
552 
553 	if (m->cp_hqd_vmid == 0)
554 		WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
555 
556 	switch (reset_type) {
557 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
558 		type = DRAIN_PIPE;
559 		break;
560 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
561 		type = RESET_WAVES;
562 		break;
563 	default:
564 		type = DRAIN_PIPE;
565 		break;
566 	}
567 
568 #if 0 /* Is this still needed? */
569 	/* Workaround: If IQ timer is active and the wait time is close to or
570 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
571 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
572 	 * cleared before continuing. Also, ensure wait times are set to at
573 	 * least 0x3.
574 	 */
575 	local_irq_save(flags);
576 	preempt_disable();
577 	retry = 5000; /* wait for 500 usecs at maximum */
578 	while (true) {
579 		temp = RREG32(mmCP_HQD_IQ_TIMER);
580 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
581 			pr_debug("HW is processing IQ\n");
582 			goto loop;
583 		}
584 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
585 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
586 					== 3) /* SEM-rearm is safe */
587 				break;
588 			/* Wait time 3 is safe for CP, but our MMIO read/write
589 			 * time is close to 1 microsecond, so check for 10 to
590 			 * leave more buffer room
591 			 */
592 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
593 					>= 10)
594 				break;
595 			pr_debug("IQ timer is active\n");
596 		} else
597 			break;
598 loop:
599 		if (!retry) {
600 			pr_err("CP HQD IQ timer status time out\n");
601 			break;
602 		}
603 		ndelay(100);
604 		--retry;
605 	}
606 	retry = 1000;
607 	while (true) {
608 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
609 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
610 			break;
611 		pr_debug("Dequeue request is pending\n");
612 
613 		if (!retry) {
614 			pr_err("CP HQD dequeue request time out\n");
615 			break;
616 		}
617 		ndelay(100);
618 		--retry;
619 	}
620 	local_irq_restore(flags);
621 	preempt_enable();
622 #endif
623 
624 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
625 
626 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
627 	while (true) {
628 		temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
629 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
630 			break;
631 		if (time_after(jiffies, end_jiffies)) {
632 			pr_err("cp queue preemption time out.\n");
633 			release_queue(kgd);
634 			return -ETIME;
635 		}
636 		usleep_range(500, 1000);
637 	}
638 
639 	release_queue(kgd);
640 	return 0;
641 }
642 
643 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
644 				unsigned int utimeout)
645 {
646 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
647 	struct v10_sdma_mqd *m;
648 	uint32_t sdma_rlc_reg_offset;
649 	uint32_t temp;
650 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
651 
652 	m = get_sdma_mqd(mqd);
653 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
654 					    m->sdma_queue_id);
655 
656 	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
657 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
658 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
659 
660 	while (true) {
661 		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
662 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
663 			break;
664 		if (time_after(jiffies, end_jiffies)) {
665 			pr_err("SDMA RLC not idle in %s\n", __func__);
666 			return -ETIME;
667 		}
668 		usleep_range(500, 1000);
669 	}
670 
671 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
672 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
673 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
674 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
675 
676 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
677 	m->sdmax_rlcx_rb_rptr_hi =
678 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
679 
680 	return 0;
681 }
682 
683 static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
684 					uint8_t vmid, uint16_t *p_pasid)
685 {
686 	uint32_t value;
687 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
688 
689 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
690 		     + vmid);
691 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
692 
693 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
694 }
695 
696 static int kgd_address_watch_disable(struct kgd_dev *kgd)
697 {
698 	return 0;
699 }
700 
701 static int kgd_address_watch_execute(struct kgd_dev *kgd,
702 					unsigned int watch_point_id,
703 					uint32_t cntl_val,
704 					uint32_t addr_hi,
705 					uint32_t addr_lo)
706 {
707 	return 0;
708 }
709 
710 static int kgd_wave_control_execute(struct kgd_dev *kgd,
711 					uint32_t gfx_index_val,
712 					uint32_t sq_cmd)
713 {
714 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
715 	uint32_t data = 0;
716 
717 	mutex_lock(&adev->grbm_idx_mutex);
718 
719 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
720 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
721 
722 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
723 		INSTANCE_BROADCAST_WRITES, 1);
724 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
725 		SA_BROADCAST_WRITES, 1);
726 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
727 		SE_BROADCAST_WRITES, 1);
728 
729 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
730 	mutex_unlock(&adev->grbm_idx_mutex);
731 
732 	return 0;
733 }
734 
735 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
736 					unsigned int watch_point_id,
737 					unsigned int reg_offset)
738 {
739 	return 0;
740 }
741 
742 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
743 		uint64_t page_table_base)
744 {
745 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
746 
747 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
748 		pr_err("trying to set page table base for wrong VMID %u\n",
749 		       vmid);
750 		return;
751 	}
752 
753 	/* SDMA is on gfxhub as well for Navi1* series */
754 	adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
755 }
756 
757 const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
758 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
759 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
760 	.init_interrupts = kgd_init_interrupts,
761 	.hqd_load = kgd_hqd_load,
762 	.hiq_mqd_load = kgd_hiq_mqd_load,
763 	.hqd_sdma_load = kgd_hqd_sdma_load,
764 	.hqd_dump = kgd_hqd_dump,
765 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
766 	.hqd_is_occupied = kgd_hqd_is_occupied,
767 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
768 	.hqd_destroy = kgd_hqd_destroy,
769 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
770 	.address_watch_disable = kgd_address_watch_disable,
771 	.address_watch_execute = kgd_address_watch_execute,
772 	.wave_control_execute = kgd_wave_control_execute,
773 	.address_watch_get_offset = kgd_address_watch_get_offset,
774 	.get_atc_vmid_pasid_mapping_info =
775 			get_atc_vmid_pasid_mapping_info,
776 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
777 };
778