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 
23 #define pr_fmt(fmt) "kfd2kgd: " fmt
24 
25 #include <linux/module.h>
26 #include <linux/fdtable.h>
27 #include <linux/uaccess.h>
28 #include <linux/mmu_context.h>
29 
30 #include "amdgpu.h"
31 #include "amdgpu_amdkfd.h"
32 #include "soc15_hw_ip.h"
33 #include "gc/gc_9_0_offset.h"
34 #include "gc/gc_9_0_sh_mask.h"
35 #include "vega10_enum.h"
36 #include "sdma0/sdma0_4_0_offset.h"
37 #include "sdma0/sdma0_4_0_sh_mask.h"
38 #include "sdma1/sdma1_4_0_offset.h"
39 #include "sdma1/sdma1_4_0_sh_mask.h"
40 #include "athub/athub_1_0_offset.h"
41 #include "athub/athub_1_0_sh_mask.h"
42 #include "oss/osssys_4_0_offset.h"
43 #include "oss/osssys_4_0_sh_mask.h"
44 #include "soc15_common.h"
45 #include "v9_structs.h"
46 #include "soc15.h"
47 #include "soc15d.h"
48 #include "mmhub_v1_0.h"
49 #include "gfxhub_v1_0.h"
50 
51 
52 #define V9_PIPE_PER_MEC		(4)
53 #define V9_QUEUES_PER_PIPE_MEC	(8)
54 
55 enum hqd_dequeue_request_type {
56 	NO_ACTION = 0,
57 	DRAIN_PIPE,
58 	RESET_WAVES
59 };
60 
61 /*
62  * Register access functions
63  */
64 
65 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
66 		uint32_t sh_mem_config,
67 		uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit,
68 		uint32_t sh_mem_bases);
69 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
70 		unsigned int vmid);
71 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
72 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
73 			uint32_t queue_id, uint32_t __user *wptr,
74 			uint32_t wptr_shift, uint32_t wptr_mask,
75 			struct mm_struct *mm);
76 static int kgd_hqd_dump(struct kgd_dev *kgd,
77 			uint32_t pipe_id, uint32_t queue_id,
78 			uint32_t (**dump)[2], uint32_t *n_regs);
79 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
80 			     uint32_t __user *wptr, struct mm_struct *mm);
81 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
82 			     uint32_t engine_id, uint32_t queue_id,
83 			     uint32_t (**dump)[2], uint32_t *n_regs);
84 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
85 		uint32_t pipe_id, uint32_t queue_id);
86 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
87 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
88 				enum kfd_preempt_type reset_type,
89 				unsigned int utimeout, uint32_t pipe_id,
90 				uint32_t queue_id);
91 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
92 				unsigned int utimeout);
93 static int kgd_address_watch_disable(struct kgd_dev *kgd);
94 static int kgd_address_watch_execute(struct kgd_dev *kgd,
95 					unsigned int watch_point_id,
96 					uint32_t cntl_val,
97 					uint32_t addr_hi,
98 					uint32_t addr_lo);
99 static int kgd_wave_control_execute(struct kgd_dev *kgd,
100 					uint32_t gfx_index_val,
101 					uint32_t sq_cmd);
102 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
103 					unsigned int watch_point_id,
104 					unsigned int reg_offset);
105 
106 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
107 		uint8_t vmid);
108 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
109 		uint8_t vmid);
110 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
111 		uint64_t page_table_base);
112 static void set_scratch_backing_va(struct kgd_dev *kgd,
113 					uint64_t va, uint32_t vmid);
114 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
115 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
116 
117 /* Because of REG_GET_FIELD() being used, we put this function in the
118  * asic specific file.
119  */
120 static int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
121 		struct tile_config *config)
122 {
123 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
124 
125 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
126 
127 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
128 	config->num_tile_configs =
129 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
130 	config->macro_tile_config_ptr =
131 			adev->gfx.config.macrotile_mode_array;
132 	config->num_macro_tile_configs =
133 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
134 
135 	return 0;
136 }
137 
138 static const struct kfd2kgd_calls kfd2kgd = {
139 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
140 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
141 	.init_interrupts = kgd_init_interrupts,
142 	.hqd_load = kgd_hqd_load,
143 	.hqd_sdma_load = kgd_hqd_sdma_load,
144 	.hqd_dump = kgd_hqd_dump,
145 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
146 	.hqd_is_occupied = kgd_hqd_is_occupied,
147 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
148 	.hqd_destroy = kgd_hqd_destroy,
149 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
150 	.address_watch_disable = kgd_address_watch_disable,
151 	.address_watch_execute = kgd_address_watch_execute,
152 	.wave_control_execute = kgd_wave_control_execute,
153 	.address_watch_get_offset = kgd_address_watch_get_offset,
154 	.get_atc_vmid_pasid_mapping_pasid =
155 			get_atc_vmid_pasid_mapping_pasid,
156 	.get_atc_vmid_pasid_mapping_valid =
157 			get_atc_vmid_pasid_mapping_valid,
158 	.set_scratch_backing_va = set_scratch_backing_va,
159 	.get_tile_config = amdgpu_amdkfd_get_tile_config,
160 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
161 	.invalidate_tlbs = invalidate_tlbs,
162 	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
163 	.get_hive_id = amdgpu_amdkfd_get_hive_id,
164 };
165 
166 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_9_0_get_functions(void)
167 {
168 	return (struct kfd2kgd_calls *)&kfd2kgd;
169 }
170 
171 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
172 {
173 	return (struct amdgpu_device *)kgd;
174 }
175 
176 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
177 			uint32_t queue, uint32_t vmid)
178 {
179 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
180 
181 	mutex_lock(&adev->srbm_mutex);
182 	soc15_grbm_select(adev, mec, pipe, queue, vmid);
183 }
184 
185 static void unlock_srbm(struct kgd_dev *kgd)
186 {
187 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
188 
189 	soc15_grbm_select(adev, 0, 0, 0, 0);
190 	mutex_unlock(&adev->srbm_mutex);
191 }
192 
193 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
194 				uint32_t queue_id)
195 {
196 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
197 
198 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
199 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
200 
201 	lock_srbm(kgd, mec, pipe, queue_id, 0);
202 }
203 
204 static uint32_t get_queue_mask(struct amdgpu_device *adev,
205 			       uint32_t pipe_id, uint32_t queue_id)
206 {
207 	unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe +
208 			    queue_id) & 31;
209 
210 	return ((uint32_t)1) << bit;
211 }
212 
213 static void release_queue(struct kgd_dev *kgd)
214 {
215 	unlock_srbm(kgd);
216 }
217 
218 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
219 					uint32_t sh_mem_config,
220 					uint32_t sh_mem_ape1_base,
221 					uint32_t sh_mem_ape1_limit,
222 					uint32_t sh_mem_bases)
223 {
224 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
225 
226 	lock_srbm(kgd, 0, 0, 0, vmid);
227 
228 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
229 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
230 	/* APE1 no longer exists on GFX9 */
231 
232 	unlock_srbm(kgd);
233 }
234 
235 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
236 					unsigned int vmid)
237 {
238 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
239 
240 	/*
241 	 * We have to assume that there is no outstanding mapping.
242 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
243 	 * a mapping is in progress or because a mapping finished
244 	 * and the SW cleared it.
245 	 * So the protocol is to always wait & clear.
246 	 */
247 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
248 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
249 
250 	/*
251 	 * need to do this twice, once for gfx and once for mmhub
252 	 * for ATC add 16 to VMID for mmhub, for IH different registers.
253 	 * ATC_VMID0..15 registers are separate from ATC_VMID16..31.
254 	 */
255 
256 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
257 	       pasid_mapping);
258 
259 	while (!(RREG32(SOC15_REG_OFFSET(
260 				ATHUB, 0,
261 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
262 		 (1U << vmid)))
263 		cpu_relax();
264 
265 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
266 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
267 	       1U << vmid);
268 
269 	/* Mapping vmid to pasid also for IH block */
270 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
271 	       pasid_mapping);
272 
273 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID16_PASID_MAPPING) + vmid,
274 	       pasid_mapping);
275 
276 	while (!(RREG32(SOC15_REG_OFFSET(
277 				ATHUB, 0,
278 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
279 		 (1U << (vmid + 16))))
280 		cpu_relax();
281 
282 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
283 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
284 	       1U << (vmid + 16));
285 
286 	/* Mapping vmid to pasid also for IH block */
287 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid,
288 	       pasid_mapping);
289 	return 0;
290 }
291 
292 /* TODO - RING0 form of field is obsolete, seems to date back to SI
293  * but still works
294  */
295 
296 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
297 {
298 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
299 	uint32_t mec;
300 	uint32_t pipe;
301 
302 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
303 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
304 
305 	lock_srbm(kgd, mec, pipe, 0, 0);
306 
307 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
308 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
309 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
310 
311 	unlock_srbm(kgd);
312 
313 	return 0;
314 }
315 
316 static uint32_t get_sdma_base_addr(struct amdgpu_device *adev,
317 				unsigned int engine_id,
318 				unsigned int queue_id)
319 {
320 	uint32_t base[2] = {
321 		SOC15_REG_OFFSET(SDMA0, 0,
322 				 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
323 		SOC15_REG_OFFSET(SDMA1, 0,
324 				 mmSDMA1_RLC0_RB_CNTL) - mmSDMA1_RLC0_RB_CNTL
325 	};
326 	uint32_t retval;
327 
328 	retval = base[engine_id] + queue_id * (mmSDMA0_RLC1_RB_CNTL -
329 					       mmSDMA0_RLC0_RB_CNTL);
330 
331 	pr_debug("sdma base address: 0x%x\n", retval);
332 
333 	return retval;
334 }
335 
336 static inline struct v9_mqd *get_mqd(void *mqd)
337 {
338 	return (struct v9_mqd *)mqd;
339 }
340 
341 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
342 {
343 	return (struct v9_sdma_mqd *)mqd;
344 }
345 
346 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
347 			uint32_t queue_id, uint32_t __user *wptr,
348 			uint32_t wptr_shift, uint32_t wptr_mask,
349 			struct mm_struct *mm)
350 {
351 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
352 	struct v9_mqd *m;
353 	uint32_t *mqd_hqd;
354 	uint32_t reg, hqd_base, data;
355 
356 	m = get_mqd(mqd);
357 
358 	acquire_queue(kgd, pipe_id, queue_id);
359 
360 	/* HIQ is set during driver init period with vmid set to 0*/
361 	if (m->cp_hqd_vmid == 0) {
362 		uint32_t value, mec, pipe;
363 
364 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
365 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
366 
367 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
368 			mec, pipe, queue_id);
369 		value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS));
370 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
371 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
372 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value);
373 	}
374 
375 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
376 	mqd_hqd = &m->cp_mqd_base_addr_lo;
377 	hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
378 
379 	for (reg = hqd_base;
380 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
381 		WREG32_RLC(reg, mqd_hqd[reg - hqd_base]);
382 
383 
384 	/* Activate doorbell logic before triggering WPTR poll. */
385 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
386 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
387 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
388 
389 	if (wptr) {
390 		/* Don't read wptr with get_user because the user
391 		 * context may not be accessible (if this function
392 		 * runs in a work queue). Instead trigger a one-shot
393 		 * polling read from memory in the CP. This assumes
394 		 * that wptr is GPU-accessible in the queue's VMID via
395 		 * ATC or SVM. WPTR==RPTR before starting the poll so
396 		 * the CP starts fetching new commands from the right
397 		 * place.
398 		 *
399 		 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
400 		 * tricky. Assume that the queue didn't overflow. The
401 		 * number of valid bits in the 32-bit RPTR depends on
402 		 * the queue size. The remaining bits are taken from
403 		 * the saved 64-bit WPTR. If the WPTR wrapped, add the
404 		 * queue size.
405 		 */
406 		uint32_t queue_size =
407 			2 << REG_GET_FIELD(m->cp_hqd_pq_control,
408 					   CP_HQD_PQ_CONTROL, QUEUE_SIZE);
409 		uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
410 
411 		if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
412 			guessed_wptr += queue_size;
413 		guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
414 		guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
415 
416 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
417 		       lower_32_bits(guessed_wptr));
418 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
419 		       upper_32_bits(guessed_wptr));
420 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
421 		       lower_32_bits((uintptr_t)wptr));
422 		WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
423 		       upper_32_bits((uintptr_t)wptr));
424 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
425 		       get_queue_mask(adev, pipe_id, queue_id));
426 	}
427 
428 	/* Start the EOP fetcher */
429 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
430 	       REG_SET_FIELD(m->cp_hqd_eop_rptr,
431 			     CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
432 
433 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
434 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
435 
436 	release_queue(kgd);
437 
438 	return 0;
439 }
440 
441 static int kgd_hqd_dump(struct kgd_dev *kgd,
442 			uint32_t pipe_id, uint32_t queue_id,
443 			uint32_t (**dump)[2], uint32_t *n_regs)
444 {
445 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
446 	uint32_t i = 0, reg;
447 #define HQD_N_REGS 56
448 #define DUMP_REG(addr) do {				\
449 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
450 			break;				\
451 		(*dump)[i][0] = (addr) << 2;		\
452 		(*dump)[i++][1] = RREG32(addr);		\
453 	} while (0)
454 
455 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
456 	if (*dump == NULL)
457 		return -ENOMEM;
458 
459 	acquire_queue(kgd, pipe_id, queue_id);
460 
461 	for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
462 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
463 		DUMP_REG(reg);
464 
465 	release_queue(kgd);
466 
467 	WARN_ON_ONCE(i != HQD_N_REGS);
468 	*n_regs = i;
469 
470 	return 0;
471 }
472 
473 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
474 			     uint32_t __user *wptr, struct mm_struct *mm)
475 {
476 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
477 	struct v9_sdma_mqd *m;
478 	uint32_t sdma_base_addr, sdmax_gfx_context_cntl;
479 	unsigned long end_jiffies;
480 	uint32_t data;
481 	uint64_t data64;
482 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
483 
484 	m = get_sdma_mqd(mqd);
485 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
486 					    m->sdma_queue_id);
487 	sdmax_gfx_context_cntl = m->sdma_engine_id ?
488 		SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_CONTEXT_CNTL) :
489 		SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_CONTEXT_CNTL);
490 
491 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
492 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
493 
494 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
495 	while (true) {
496 		data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
497 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
498 			break;
499 		if (time_after(jiffies, end_jiffies))
500 			return -ETIME;
501 		usleep_range(500, 1000);
502 	}
503 	data = RREG32(sdmax_gfx_context_cntl);
504 	data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
505 			     RESUME_CTX, 0);
506 	WREG32(sdmax_gfx_context_cntl, data);
507 
508 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL_OFFSET,
509 	       m->sdmax_rlcx_doorbell_offset);
510 
511 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
512 			     ENABLE, 1);
513 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
514 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
515 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI,
516 				m->sdmax_rlcx_rb_rptr_hi);
517 
518 	WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
519 	if (read_user_wptr(mm, wptr64, data64)) {
520 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
521 		       lower_32_bits(data64));
522 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
523 		       upper_32_bits(data64));
524 	} else {
525 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
526 		       m->sdmax_rlcx_rb_rptr);
527 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
528 		       m->sdmax_rlcx_rb_rptr_hi);
529 	}
530 	WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
531 
532 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
533 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
534 			m->sdmax_rlcx_rb_base_hi);
535 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
536 			m->sdmax_rlcx_rb_rptr_addr_lo);
537 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
538 			m->sdmax_rlcx_rb_rptr_addr_hi);
539 
540 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
541 			     RB_ENABLE, 1);
542 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
543 
544 	return 0;
545 }
546 
547 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
548 			     uint32_t engine_id, uint32_t queue_id,
549 			     uint32_t (**dump)[2], uint32_t *n_regs)
550 {
551 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
552 	uint32_t sdma_base_addr = get_sdma_base_addr(adev, engine_id, queue_id);
553 	uint32_t i = 0, reg;
554 #undef HQD_N_REGS
555 #define HQD_N_REGS (19+6+7+10)
556 
557 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
558 	if (*dump == NULL)
559 		return -ENOMEM;
560 
561 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
562 		DUMP_REG(sdma_base_addr + reg);
563 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
564 		DUMP_REG(sdma_base_addr + reg);
565 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
566 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
567 		DUMP_REG(sdma_base_addr + reg);
568 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
569 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
570 		DUMP_REG(sdma_base_addr + reg);
571 
572 	WARN_ON_ONCE(i != HQD_N_REGS);
573 	*n_regs = i;
574 
575 	return 0;
576 }
577 
578 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
579 				uint32_t pipe_id, uint32_t queue_id)
580 {
581 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
582 	uint32_t act;
583 	bool retval = false;
584 	uint32_t low, high;
585 
586 	acquire_queue(kgd, pipe_id, queue_id);
587 	act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
588 	if (act) {
589 		low = lower_32_bits(queue_address >> 8);
590 		high = upper_32_bits(queue_address >> 8);
591 
592 		if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
593 		   high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
594 			retval = true;
595 	}
596 	release_queue(kgd);
597 	return retval;
598 }
599 
600 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
601 {
602 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
603 	struct v9_sdma_mqd *m;
604 	uint32_t sdma_base_addr;
605 	uint32_t sdma_rlc_rb_cntl;
606 
607 	m = get_sdma_mqd(mqd);
608 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
609 					    m->sdma_queue_id);
610 
611 	sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
612 
613 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
614 		return true;
615 
616 	return false;
617 }
618 
619 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
620 				enum kfd_preempt_type reset_type,
621 				unsigned int utimeout, uint32_t pipe_id,
622 				uint32_t queue_id)
623 {
624 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
625 	enum hqd_dequeue_request_type type;
626 	unsigned long end_jiffies;
627 	uint32_t temp;
628 	struct v9_mqd *m = get_mqd(mqd);
629 
630 	if (adev->in_gpu_reset)
631 		return -EIO;
632 
633 	acquire_queue(kgd, pipe_id, queue_id);
634 
635 	if (m->cp_hqd_vmid == 0)
636 		WREG32_FIELD15_RLC(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
637 
638 	switch (reset_type) {
639 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
640 		type = DRAIN_PIPE;
641 		break;
642 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
643 		type = RESET_WAVES;
644 		break;
645 	default:
646 		type = DRAIN_PIPE;
647 		break;
648 	}
649 
650 	WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
651 
652 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
653 	while (true) {
654 		temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
655 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
656 			break;
657 		if (time_after(jiffies, end_jiffies)) {
658 			pr_err("cp queue preemption time out.\n");
659 			release_queue(kgd);
660 			return -ETIME;
661 		}
662 		usleep_range(500, 1000);
663 	}
664 
665 	release_queue(kgd);
666 	return 0;
667 }
668 
669 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
670 				unsigned int utimeout)
671 {
672 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
673 	struct v9_sdma_mqd *m;
674 	uint32_t sdma_base_addr;
675 	uint32_t temp;
676 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
677 
678 	m = get_sdma_mqd(mqd);
679 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
680 					    m->sdma_queue_id);
681 
682 	temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
683 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
684 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
685 
686 	while (true) {
687 		temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
688 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
689 			break;
690 		if (time_after(jiffies, end_jiffies))
691 			return -ETIME;
692 		usleep_range(500, 1000);
693 	}
694 
695 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
696 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
697 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
698 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
699 
700 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
701 	m->sdmax_rlcx_rb_rptr_hi =
702 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI);
703 
704 	return 0;
705 }
706 
707 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
708 							uint8_t vmid)
709 {
710 	uint32_t reg;
711 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
712 
713 	reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
714 		     + vmid);
715 	return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
716 }
717 
718 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
719 								uint8_t vmid)
720 {
721 	uint32_t reg;
722 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
723 
724 	reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
725 		     + vmid);
726 	return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
727 }
728 
729 static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid,
730 			uint32_t flush_type)
731 {
732 	signed long r;
733 	uint32_t seq;
734 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
735 
736 	spin_lock(&adev->gfx.kiq.ring_lock);
737 	amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/
738 	amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
739 	amdgpu_ring_write(ring,
740 			PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
741 			PACKET3_INVALIDATE_TLBS_ALL_HUB(1) |
742 			PACKET3_INVALIDATE_TLBS_PASID(pasid) |
743 			PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(flush_type));
744 	amdgpu_fence_emit_polling(ring, &seq);
745 	amdgpu_ring_commit(ring);
746 	spin_unlock(&adev->gfx.kiq.ring_lock);
747 
748 	r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
749 	if (r < 1) {
750 		DRM_ERROR("wait for kiq fence error: %ld.\n", r);
751 		return -ETIME;
752 	}
753 
754 	return 0;
755 }
756 
757 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
758 {
759 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
760 	int vmid;
761 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
762 	uint32_t flush_type = 0;
763 
764 	if (adev->in_gpu_reset)
765 		return -EIO;
766 	if (adev->gmc.xgmi.num_physical_nodes &&
767 		adev->asic_type == CHIP_VEGA20)
768 		flush_type = 2;
769 
770 	if (ring->sched.ready)
771 		return invalidate_tlbs_with_kiq(adev, pasid, flush_type);
772 
773 	for (vmid = 0; vmid < 16; vmid++) {
774 		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
775 			continue;
776 		if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
777 			if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
778 				== pasid) {
779 				amdgpu_gmc_flush_gpu_tlb(adev, vmid,
780 							 flush_type);
781 				break;
782 			}
783 		}
784 	}
785 
786 	return 0;
787 }
788 
789 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
790 {
791 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
792 
793 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
794 		pr_err("non kfd vmid %d\n", vmid);
795 		return 0;
796 	}
797 
798 	/* Use legacy mode tlb invalidation.
799 	 *
800 	 * Currently on Raven the code below is broken for anything but
801 	 * legacy mode due to a MMHUB power gating problem. A workaround
802 	 * is for MMHUB to wait until the condition PER_VMID_INVALIDATE_REQ
803 	 * == PER_VMID_INVALIDATE_ACK instead of simply waiting for the ack
804 	 * bit.
805 	 *
806 	 * TODO 1: agree on the right set of invalidation registers for
807 	 * KFD use. Use the last one for now. Invalidate both GC and
808 	 * MMHUB.
809 	 *
810 	 * TODO 2: support range-based invalidation, requires kfg2kgd
811 	 * interface change
812 	 */
813 	amdgpu_gmc_flush_gpu_tlb(adev, vmid, 0);
814 	return 0;
815 }
816 
817 static int kgd_address_watch_disable(struct kgd_dev *kgd)
818 {
819 	return 0;
820 }
821 
822 static int kgd_address_watch_execute(struct kgd_dev *kgd,
823 					unsigned int watch_point_id,
824 					uint32_t cntl_val,
825 					uint32_t addr_hi,
826 					uint32_t addr_lo)
827 {
828 	return 0;
829 }
830 
831 static int kgd_wave_control_execute(struct kgd_dev *kgd,
832 					uint32_t gfx_index_val,
833 					uint32_t sq_cmd)
834 {
835 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
836 	uint32_t data = 0;
837 
838 	mutex_lock(&adev->grbm_idx_mutex);
839 
840 	WREG32_SOC15_RLC_SHADOW(GC, 0, mmGRBM_GFX_INDEX, gfx_index_val);
841 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
842 
843 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
844 		INSTANCE_BROADCAST_WRITES, 1);
845 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
846 		SH_BROADCAST_WRITES, 1);
847 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
848 		SE_BROADCAST_WRITES, 1);
849 
850 	WREG32_SOC15_RLC_SHADOW(GC, 0, mmGRBM_GFX_INDEX, data);
851 	mutex_unlock(&adev->grbm_idx_mutex);
852 
853 	return 0;
854 }
855 
856 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
857 					unsigned int watch_point_id,
858 					unsigned int reg_offset)
859 {
860 	return 0;
861 }
862 
863 static void set_scratch_backing_va(struct kgd_dev *kgd,
864 					uint64_t va, uint32_t vmid)
865 {
866 	/* No longer needed on GFXv9. The scratch base address is
867 	 * passed to the shader by the CP. It's the user mode driver's
868 	 * responsibility.
869 	 */
870 }
871 
872 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
873 		uint64_t page_table_base)
874 {
875 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
876 
877 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
878 		pr_err("trying to set page table base for wrong VMID %u\n",
879 		       vmid);
880 		return;
881 	}
882 
883 	/* TODO: take advantage of per-process address space size. For
884 	 * now, all processes share the same address space size, like
885 	 * on GFX8 and older.
886 	 */
887 	mmhub_v1_0_setup_vm_pt_regs(adev, vmid, page_table_base);
888 
889 	gfxhub_v1_0_setup_vm_pt_regs(adev, vmid, page_table_base);
890 }
891