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 <linux/mmu_context.h>
23 #include "amdgpu.h"
24 #include "amdgpu_amdkfd.h"
25 #include "gc/gc_10_1_0_offset.h"
26 #include "gc/gc_10_1_0_sh_mask.h"
27 #include "navi10_enum.h"
28 #include "athub/athub_2_0_0_offset.h"
29 #include "athub/athub_2_0_0_sh_mask.h"
30 #include "oss/osssys_5_0_0_offset.h"
31 #include "oss/osssys_5_0_0_sh_mask.h"
32 #include "soc15_common.h"
33 #include "v10_structs.h"
34 #include "nv.h"
35 #include "nvd.h"
36 #include "gfxhub_v2_0.h"
37 
38 enum hqd_dequeue_request_type {
39 	NO_ACTION = 0,
40 	DRAIN_PIPE,
41 	RESET_WAVES,
42 	SAVE_WAVES
43 };
44 
45 /* Because of REG_GET_FIELD() being used, we put this function in the
46  * asic specific file.
47  */
48 static int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
49 		struct tile_config *config)
50 {
51 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
52 
53 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
54 #if 0
55 /* TODO - confirm REG_GET_FIELD x2, should be OK as is... but
56  * MC_ARB_RAMCFG register doesn't exist on Vega10 - initial amdgpu
57  * changes commented out related code, doing the same here for now but
58  * need to sync with Ken et al
59  */
60 	config->num_banks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
61 				MC_ARB_RAMCFG, NOOFBANK);
62 	config->num_ranks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
63 				MC_ARB_RAMCFG, NOOFRANKS);
64 #endif
65 
66 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
67 	config->num_tile_configs =
68 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
69 	config->macro_tile_config_ptr =
70 			adev->gfx.config.macrotile_mode_array;
71 	config->num_macro_tile_configs =
72 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
73 
74 	return 0;
75 }
76 
77 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
78 {
79 	return (struct amdgpu_device *)kgd;
80 }
81 
82 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
83 			uint32_t queue, uint32_t vmid)
84 {
85 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
86 
87 	mutex_lock(&adev->srbm_mutex);
88 	nv_grbm_select(adev, mec, pipe, queue, vmid);
89 }
90 
91 static void unlock_srbm(struct kgd_dev *kgd)
92 {
93 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
94 
95 	nv_grbm_select(adev, 0, 0, 0, 0);
96 	mutex_unlock(&adev->srbm_mutex);
97 }
98 
99 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
100 				uint32_t queue_id)
101 {
102 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
103 
104 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
105 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
106 
107 	lock_srbm(kgd, mec, pipe, queue_id, 0);
108 }
109 
110 static uint32_t get_queue_mask(struct amdgpu_device *adev,
111 			       uint32_t pipe_id, uint32_t queue_id)
112 {
113 	unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe +
114 			    queue_id) & 31;
115 
116 	return ((uint32_t)1) << bit;
117 }
118 
119 static void release_queue(struct kgd_dev *kgd)
120 {
121 	unlock_srbm(kgd);
122 }
123 
124 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
125 					uint32_t sh_mem_config,
126 					uint32_t sh_mem_ape1_base,
127 					uint32_t sh_mem_ape1_limit,
128 					uint32_t sh_mem_bases)
129 {
130 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
131 
132 	lock_srbm(kgd, 0, 0, 0, vmid);
133 
134 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
135 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
136 	/* APE1 no longer exists on GFX9 */
137 
138 	unlock_srbm(kgd);
139 }
140 
141 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
142 					unsigned int vmid)
143 {
144 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
145 
146 	/*
147 	 * We have to assume that there is no outstanding mapping.
148 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
149 	 * a mapping is in progress or because a mapping finished
150 	 * and the SW cleared it.
151 	 * So the protocol is to always wait & clear.
152 	 */
153 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
154 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
155 
156 	pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping);
157 
158 	pr_debug("ATHUB, reg %x\n", SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid);
159 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
160 	       pasid_mapping);
161 
162 #if 0
163 	/* TODO: uncomment this code when the hardware support is ready. */
164 	while (!(RREG32(SOC15_REG_OFFSET(
165 				ATHUB, 0,
166 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
167 		 (1U << vmid)))
168 		cpu_relax();
169 
170 	pr_debug("ATHUB mapping update finished\n");
171 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
172 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
173 	       1U << vmid);
174 #endif
175 
176 	/* Mapping vmid to pasid also for IH block */
177 	pr_debug("update mapping for IH block and mmhub");
178 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
179 	       pasid_mapping);
180 
181 	return 0;
182 }
183 
184 /* TODO - RING0 form of field is obsolete, seems to date back to SI
185  * but still works
186  */
187 
188 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
189 {
190 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
191 	uint32_t mec;
192 	uint32_t pipe;
193 
194 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
195 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
196 
197 	lock_srbm(kgd, mec, pipe, 0, 0);
198 
199 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
200 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
201 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
202 
203 	unlock_srbm(kgd);
204 
205 	return 0;
206 }
207 
208 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
209 				unsigned int engine_id,
210 				unsigned int queue_id)
211 {
212 	uint32_t sdma_engine_reg_base[2] = {
213 		SOC15_REG_OFFSET(SDMA0, 0,
214 				 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
215 		/* On gfx10, mmSDMA1_xxx registers are defined NOT based
216 		 * on SDMA1 base address (dw 0x1860) but based on SDMA0
217 		 * base address (dw 0x1260). Therefore use mmSDMA0_RLC0_RB_CNTL
218 		 * instead of mmSDMA1_RLC0_RB_CNTL for the base address calc
219 		 * below
220 		 */
221 		SOC15_REG_OFFSET(SDMA1, 0,
222 				 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL
223 	};
224 
225 	uint32_t retval = sdma_engine_reg_base[engine_id]
226 		+ queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
227 
228 	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
229 			queue_id, retval);
230 
231 	return retval;
232 }
233 
234 #if 0
235 static uint32_t get_watch_base_addr(struct amdgpu_device *adev)
236 {
237 	uint32_t retval = SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) -
238 			mmTCP_WATCH0_ADDR_H;
239 
240 	pr_debug("kfd: reg watch base address: 0x%x\n", retval);
241 
242 	return retval;
243 }
244 #endif
245 
246 static inline struct v10_compute_mqd *get_mqd(void *mqd)
247 {
248 	return (struct v10_compute_mqd *)mqd;
249 }
250 
251 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
252 {
253 	return (struct v10_sdma_mqd *)mqd;
254 }
255 
256 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
257 			uint32_t queue_id, uint32_t __user *wptr,
258 			uint32_t wptr_shift, uint32_t wptr_mask,
259 			struct mm_struct *mm)
260 {
261 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
262 	struct v10_compute_mqd *m;
263 	uint32_t *mqd_hqd;
264 	uint32_t reg, hqd_base, data;
265 
266 	m = get_mqd(mqd);
267 
268 	pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
269 	acquire_queue(kgd, pipe_id, queue_id);
270 
271 	/* HIQ is set during driver init period with vmid set to 0*/
272 	if (m->cp_hqd_vmid == 0) {
273 		uint32_t value, mec, pipe;
274 
275 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
276 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
277 
278 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
279 			mec, pipe, queue_id);
280 		value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS));
281 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
282 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
283 		WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value);
284 	}
285 
286 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
287 	mqd_hqd = &m->cp_mqd_base_addr_lo;
288 	hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
289 
290 	for (reg = hqd_base;
291 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
292 		WREG32(reg, mqd_hqd[reg - hqd_base]);
293 
294 
295 	/* Activate doorbell logic before triggering WPTR poll. */
296 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
297 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
298 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
299 
300 	if (wptr) {
301 		/* Don't read wptr with get_user because the user
302 		 * context may not be accessible (if this function
303 		 * runs in a work queue). Instead trigger a one-shot
304 		 * polling read from memory in the CP. This assumes
305 		 * that wptr is GPU-accessible in the queue's VMID via
306 		 * ATC or SVM. WPTR==RPTR before starting the poll so
307 		 * the CP starts fetching new commands from the right
308 		 * place.
309 		 *
310 		 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
311 		 * tricky. Assume that the queue didn't overflow. The
312 		 * number of valid bits in the 32-bit RPTR depends on
313 		 * the queue size. The remaining bits are taken from
314 		 * the saved 64-bit WPTR. If the WPTR wrapped, add the
315 		 * queue size.
316 		 */
317 		uint32_t queue_size =
318 			2 << REG_GET_FIELD(m->cp_hqd_pq_control,
319 					   CP_HQD_PQ_CONTROL, QUEUE_SIZE);
320 		uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
321 
322 		if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
323 			guessed_wptr += queue_size;
324 		guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
325 		guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
326 
327 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
328 		       lower_32_bits(guessed_wptr));
329 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
330 		       upper_32_bits(guessed_wptr));
331 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
332 		       lower_32_bits((uint64_t)wptr));
333 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
334 		       upper_32_bits((uint64_t)wptr));
335 		pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__, get_queue_mask(adev, pipe_id, queue_id));
336 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
337 		       get_queue_mask(adev, pipe_id, queue_id));
338 	}
339 
340 	/* Start the EOP fetcher */
341 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
342 	       REG_SET_FIELD(m->cp_hqd_eop_rptr,
343 			     CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
344 
345 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
346 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
347 
348 	release_queue(kgd);
349 
350 	return 0;
351 }
352 
353 static int kgd_hqd_dump(struct kgd_dev *kgd,
354 			uint32_t pipe_id, uint32_t queue_id,
355 			uint32_t (**dump)[2], uint32_t *n_regs)
356 {
357 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
358 	uint32_t i = 0, reg;
359 #define HQD_N_REGS 56
360 #define DUMP_REG(addr) do {				\
361 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
362 			break;				\
363 		(*dump)[i][0] = (addr) << 2;		\
364 		(*dump)[i++][1] = RREG32(addr);		\
365 	} while (0)
366 
367 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
368 	if (*dump == NULL)
369 		return -ENOMEM;
370 
371 	acquire_queue(kgd, pipe_id, queue_id);
372 
373 	for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
374 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
375 		DUMP_REG(reg);
376 
377 	release_queue(kgd);
378 
379 	WARN_ON_ONCE(i != HQD_N_REGS);
380 	*n_regs = i;
381 
382 	return 0;
383 }
384 
385 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
386 			     uint32_t __user *wptr, struct mm_struct *mm)
387 {
388 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
389 	struct v10_sdma_mqd *m;
390 	uint32_t sdma_rlc_reg_offset;
391 	unsigned long end_jiffies;
392 	uint32_t data;
393 	uint64_t data64;
394 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
395 
396 	m = get_sdma_mqd(mqd);
397 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
398 					    m->sdma_queue_id);
399 
400 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
401 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
402 
403 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
404 	while (true) {
405 		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
406 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
407 			break;
408 		if (time_after(jiffies, end_jiffies)) {
409 			pr_err("SDMA RLC not idle in %s\n", __func__);
410 			return -ETIME;
411 		}
412 		usleep_range(500, 1000);
413 	}
414 
415 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
416 	       m->sdmax_rlcx_doorbell_offset);
417 
418 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
419 			     ENABLE, 1);
420 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
421 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
422 				m->sdmax_rlcx_rb_rptr);
423 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
424 				m->sdmax_rlcx_rb_rptr_hi);
425 
426 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
427 	if (read_user_wptr(mm, wptr64, data64)) {
428 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
429 		       lower_32_bits(data64));
430 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
431 		       upper_32_bits(data64));
432 	} else {
433 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
434 		       m->sdmax_rlcx_rb_rptr);
435 		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
436 		       m->sdmax_rlcx_rb_rptr_hi);
437 	}
438 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
439 
440 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
441 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
442 			m->sdmax_rlcx_rb_base_hi);
443 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
444 			m->sdmax_rlcx_rb_rptr_addr_lo);
445 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
446 			m->sdmax_rlcx_rb_rptr_addr_hi);
447 
448 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
449 			     RB_ENABLE, 1);
450 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
451 
452 	return 0;
453 }
454 
455 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
456 			     uint32_t engine_id, uint32_t queue_id,
457 			     uint32_t (**dump)[2], uint32_t *n_regs)
458 {
459 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
460 	uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
461 			engine_id, queue_id);
462 	uint32_t i = 0, reg;
463 #undef HQD_N_REGS
464 #define HQD_N_REGS (19+6+7+10)
465 
466 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
467 	if (*dump == NULL)
468 		return -ENOMEM;
469 
470 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
471 		DUMP_REG(sdma_rlc_reg_offset + reg);
472 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
473 		DUMP_REG(sdma_rlc_reg_offset + reg);
474 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
475 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
476 		DUMP_REG(sdma_rlc_reg_offset + reg);
477 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
478 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
479 		DUMP_REG(sdma_rlc_reg_offset + reg);
480 
481 	WARN_ON_ONCE(i != HQD_N_REGS);
482 	*n_regs = i;
483 
484 	return 0;
485 }
486 
487 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
488 				uint32_t pipe_id, uint32_t queue_id)
489 {
490 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
491 	uint32_t act;
492 	bool retval = false;
493 	uint32_t low, high;
494 
495 	acquire_queue(kgd, pipe_id, queue_id);
496 	act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
497 	if (act) {
498 		low = lower_32_bits(queue_address >> 8);
499 		high = upper_32_bits(queue_address >> 8);
500 
501 		if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
502 		   high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
503 			retval = true;
504 	}
505 	release_queue(kgd);
506 	return retval;
507 }
508 
509 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
510 {
511 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
512 	struct v10_sdma_mqd *m;
513 	uint32_t sdma_rlc_reg_offset;
514 	uint32_t sdma_rlc_rb_cntl;
515 
516 	m = get_sdma_mqd(mqd);
517 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
518 					    m->sdma_queue_id);
519 
520 	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
521 
522 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
523 		return true;
524 
525 	return false;
526 }
527 
528 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
529 				enum kfd_preempt_type reset_type,
530 				unsigned int utimeout, uint32_t pipe_id,
531 				uint32_t queue_id)
532 {
533 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
534 	enum hqd_dequeue_request_type type;
535 	unsigned long end_jiffies;
536 	uint32_t temp;
537 	struct v10_compute_mqd *m = get_mqd(mqd);
538 
539 #if 0
540 	unsigned long flags;
541 	int retry;
542 #endif
543 
544 	acquire_queue(kgd, pipe_id, queue_id);
545 
546 	if (m->cp_hqd_vmid == 0)
547 		WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
548 
549 	switch (reset_type) {
550 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
551 		type = DRAIN_PIPE;
552 		break;
553 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
554 		type = RESET_WAVES;
555 		break;
556 	default:
557 		type = DRAIN_PIPE;
558 		break;
559 	}
560 
561 #if 0 /* Is this still needed? */
562 	/* Workaround: If IQ timer is active and the wait time is close to or
563 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
564 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
565 	 * cleared before continuing. Also, ensure wait times are set to at
566 	 * least 0x3.
567 	 */
568 	local_irq_save(flags);
569 	preempt_disable();
570 	retry = 5000; /* wait for 500 usecs at maximum */
571 	while (true) {
572 		temp = RREG32(mmCP_HQD_IQ_TIMER);
573 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
574 			pr_debug("HW is processing IQ\n");
575 			goto loop;
576 		}
577 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
578 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
579 					== 3) /* SEM-rearm is safe */
580 				break;
581 			/* Wait time 3 is safe for CP, but our MMIO read/write
582 			 * time is close to 1 microsecond, so check for 10 to
583 			 * leave more buffer room
584 			 */
585 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
586 					>= 10)
587 				break;
588 			pr_debug("IQ timer is active\n");
589 		} else
590 			break;
591 loop:
592 		if (!retry) {
593 			pr_err("CP HQD IQ timer status time out\n");
594 			break;
595 		}
596 		ndelay(100);
597 		--retry;
598 	}
599 	retry = 1000;
600 	while (true) {
601 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
602 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
603 			break;
604 		pr_debug("Dequeue request is pending\n");
605 
606 		if (!retry) {
607 			pr_err("CP HQD dequeue request time out\n");
608 			break;
609 		}
610 		ndelay(100);
611 		--retry;
612 	}
613 	local_irq_restore(flags);
614 	preempt_enable();
615 #endif
616 
617 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
618 
619 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
620 	while (true) {
621 		temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
622 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
623 			break;
624 		if (time_after(jiffies, end_jiffies)) {
625 			pr_err("cp queue preemption time out.\n");
626 			release_queue(kgd);
627 			return -ETIME;
628 		}
629 		usleep_range(500, 1000);
630 	}
631 
632 	release_queue(kgd);
633 	return 0;
634 }
635 
636 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
637 				unsigned int utimeout)
638 {
639 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
640 	struct v10_sdma_mqd *m;
641 	uint32_t sdma_rlc_reg_offset;
642 	uint32_t temp;
643 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
644 
645 	m = get_sdma_mqd(mqd);
646 	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
647 					    m->sdma_queue_id);
648 
649 	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
650 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
651 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
652 
653 	while (true) {
654 		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
655 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
656 			break;
657 		if (time_after(jiffies, end_jiffies)) {
658 			pr_err("SDMA RLC not idle in %s\n", __func__);
659 			return -ETIME;
660 		}
661 		usleep_range(500, 1000);
662 	}
663 
664 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
665 	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
666 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
667 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
668 
669 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
670 	m->sdmax_rlcx_rb_rptr_hi =
671 		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
672 
673 	return 0;
674 }
675 
676 static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
677 					uint8_t vmid, uint16_t *p_pasid)
678 {
679 	uint32_t value;
680 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
681 
682 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
683 		     + vmid);
684 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
685 
686 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
687 }
688 
689 static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid)
690 {
691 	signed long r;
692 	uint32_t seq;
693 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
694 
695 	spin_lock(&adev->gfx.kiq.ring_lock);
696 	amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/
697 	amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
698 	amdgpu_ring_write(ring,
699 			PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
700 			PACKET3_INVALIDATE_TLBS_PASID(pasid));
701 	amdgpu_fence_emit_polling(ring, &seq);
702 	amdgpu_ring_commit(ring);
703 	spin_unlock(&adev->gfx.kiq.ring_lock);
704 
705 	r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
706 	if (r < 1) {
707 		DRM_ERROR("wait for kiq fence error: %ld.\n", r);
708 		return -ETIME;
709 	}
710 
711 	return 0;
712 }
713 
714 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
715 {
716 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
717 	int vmid;
718 	uint16_t queried_pasid;
719 	bool ret;
720 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
721 
722 	if (amdgpu_emu_mode == 0 && ring->sched.ready)
723 		return invalidate_tlbs_with_kiq(adev, pasid);
724 
725 	for (vmid = 0; vmid < 16; vmid++) {
726 		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
727 			continue;
728 
729 		ret = get_atc_vmid_pasid_mapping_info(kgd, vmid,
730 				&queried_pasid);
731 		if (ret	&& queried_pasid == pasid) {
732 			amdgpu_gmc_flush_gpu_tlb(adev, vmid,
733 					AMDGPU_GFXHUB_0, 0);
734 			break;
735 		}
736 	}
737 
738 	return 0;
739 }
740 
741 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
742 {
743 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
744 
745 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
746 		pr_err("non kfd vmid %d\n", vmid);
747 		return 0;
748 	}
749 
750 	amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0);
751 	return 0;
752 }
753 
754 static int kgd_address_watch_disable(struct kgd_dev *kgd)
755 {
756 	return 0;
757 }
758 
759 static int kgd_address_watch_execute(struct kgd_dev *kgd,
760 					unsigned int watch_point_id,
761 					uint32_t cntl_val,
762 					uint32_t addr_hi,
763 					uint32_t addr_lo)
764 {
765 	return 0;
766 }
767 
768 static int kgd_wave_control_execute(struct kgd_dev *kgd,
769 					uint32_t gfx_index_val,
770 					uint32_t sq_cmd)
771 {
772 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
773 	uint32_t data = 0;
774 
775 	mutex_lock(&adev->grbm_idx_mutex);
776 
777 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
778 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
779 
780 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
781 		INSTANCE_BROADCAST_WRITES, 1);
782 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
783 		SA_BROADCAST_WRITES, 1);
784 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
785 		SE_BROADCAST_WRITES, 1);
786 
787 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
788 	mutex_unlock(&adev->grbm_idx_mutex);
789 
790 	return 0;
791 }
792 
793 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
794 					unsigned int watch_point_id,
795 					unsigned int reg_offset)
796 {
797 	return 0;
798 }
799 
800 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
801 		uint64_t page_table_base)
802 {
803 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
804 
805 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
806 		pr_err("trying to set page table base for wrong VMID %u\n",
807 		       vmid);
808 		return;
809 	}
810 
811 	/* SDMA is on gfxhub as well for Navi1* series */
812 	gfxhub_v2_0_setup_vm_pt_regs(adev, vmid, page_table_base);
813 }
814 
815 const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
816 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
817 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
818 	.init_interrupts = kgd_init_interrupts,
819 	.hqd_load = kgd_hqd_load,
820 	.hqd_sdma_load = kgd_hqd_sdma_load,
821 	.hqd_dump = kgd_hqd_dump,
822 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
823 	.hqd_is_occupied = kgd_hqd_is_occupied,
824 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
825 	.hqd_destroy = kgd_hqd_destroy,
826 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
827 	.address_watch_disable = kgd_address_watch_disable,
828 	.address_watch_execute = kgd_address_watch_execute,
829 	.wave_control_execute = kgd_wave_control_execute,
830 	.address_watch_get_offset = kgd_address_watch_get_offset,
831 	.get_atc_vmid_pasid_mapping_info =
832 			get_atc_vmid_pasid_mapping_info,
833 	.get_tile_config = amdgpu_amdkfd_get_tile_config,
834 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
835 	.invalidate_tlbs = invalidate_tlbs,
836 	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
837 	.get_hive_id = amdgpu_amdkfd_get_hive_id,
838 };
839