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/firmware.h>
29 #include <linux/mmu_context.h>
30 #include <drm/drmP.h>
31 #include "amdgpu.h"
32 #include "amdgpu_amdkfd.h"
33 #include "amdgpu_ucode.h"
34 #include "soc15_hw_ip.h"
35 #include "gc/gc_9_0_offset.h"
36 #include "gc/gc_9_0_sh_mask.h"
37 #include "vega10_enum.h"
38 #include "sdma0/sdma0_4_0_offset.h"
39 #include "sdma0/sdma0_4_0_sh_mask.h"
40 #include "sdma1/sdma1_4_0_offset.h"
41 #include "sdma1/sdma1_4_0_sh_mask.h"
42 #include "athub/athub_1_0_offset.h"
43 #include "athub/athub_1_0_sh_mask.h"
44 #include "oss/osssys_4_0_offset.h"
45 #include "oss/osssys_4_0_sh_mask.h"
46 #include "soc15_common.h"
47 #include "v9_structs.h"
48 #include "soc15.h"
49 #include "soc15d.h"
50 #include "mmhub_v1_0.h"
51 #include "gfxhub_v1_0.h"
52 
53 
54 #define V9_PIPE_PER_MEC		(4)
55 #define V9_QUEUES_PER_PIPE_MEC	(8)
56 
57 enum hqd_dequeue_request_type {
58 	NO_ACTION = 0,
59 	DRAIN_PIPE,
60 	RESET_WAVES
61 };
62 
63 /*
64  * Register access functions
65  */
66 
67 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
68 		uint32_t sh_mem_config,
69 		uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit,
70 		uint32_t sh_mem_bases);
71 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
72 		unsigned int vmid);
73 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
74 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
75 			uint32_t queue_id, uint32_t __user *wptr,
76 			uint32_t wptr_shift, uint32_t wptr_mask,
77 			struct mm_struct *mm);
78 static int kgd_hqd_dump(struct kgd_dev *kgd,
79 			uint32_t pipe_id, uint32_t queue_id,
80 			uint32_t (**dump)[2], uint32_t *n_regs);
81 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
82 			     uint32_t __user *wptr, struct mm_struct *mm);
83 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
84 			     uint32_t engine_id, uint32_t queue_id,
85 			     uint32_t (**dump)[2], uint32_t *n_regs);
86 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
87 		uint32_t pipe_id, uint32_t queue_id);
88 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
89 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
90 				enum kfd_preempt_type reset_type,
91 				unsigned int utimeout, uint32_t pipe_id,
92 				uint32_t queue_id);
93 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
94 				unsigned int utimeout);
95 static int kgd_address_watch_disable(struct kgd_dev *kgd);
96 static int kgd_address_watch_execute(struct kgd_dev *kgd,
97 					unsigned int watch_point_id,
98 					uint32_t cntl_val,
99 					uint32_t addr_hi,
100 					uint32_t addr_lo);
101 static int kgd_wave_control_execute(struct kgd_dev *kgd,
102 					uint32_t gfx_index_val,
103 					uint32_t sq_cmd);
104 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
105 					unsigned int watch_point_id,
106 					unsigned int reg_offset);
107 
108 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
109 		uint8_t vmid);
110 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
111 		uint8_t vmid);
112 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
113 		uint64_t page_table_base);
114 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
115 static void set_scratch_backing_va(struct kgd_dev *kgd,
116 					uint64_t va, uint32_t vmid);
117 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
118 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
119 
120 /* Because of REG_GET_FIELD() being used, we put this function in the
121  * asic specific file.
122  */
123 static int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
124 		struct tile_config *config)
125 {
126 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
127 
128 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
129 
130 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
131 	config->num_tile_configs =
132 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
133 	config->macro_tile_config_ptr =
134 			adev->gfx.config.macrotile_mode_array;
135 	config->num_macro_tile_configs =
136 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
137 
138 	return 0;
139 }
140 
141 static const struct kfd2kgd_calls kfd2kgd = {
142 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
143 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
144 	.init_interrupts = kgd_init_interrupts,
145 	.hqd_load = kgd_hqd_load,
146 	.hqd_sdma_load = kgd_hqd_sdma_load,
147 	.hqd_dump = kgd_hqd_dump,
148 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
149 	.hqd_is_occupied = kgd_hqd_is_occupied,
150 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
151 	.hqd_destroy = kgd_hqd_destroy,
152 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
153 	.address_watch_disable = kgd_address_watch_disable,
154 	.address_watch_execute = kgd_address_watch_execute,
155 	.wave_control_execute = kgd_wave_control_execute,
156 	.address_watch_get_offset = kgd_address_watch_get_offset,
157 	.get_atc_vmid_pasid_mapping_pasid =
158 			get_atc_vmid_pasid_mapping_pasid,
159 	.get_atc_vmid_pasid_mapping_valid =
160 			get_atc_vmid_pasid_mapping_valid,
161 	.get_fw_version = get_fw_version,
162 	.set_scratch_backing_va = set_scratch_backing_va,
163 	.get_tile_config = amdgpu_amdkfd_get_tile_config,
164 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
165 	.invalidate_tlbs = invalidate_tlbs,
166 	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
167 	.get_hive_id = amdgpu_amdkfd_get_hive_id,
168 };
169 
170 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_9_0_get_functions(void)
171 {
172 	return (struct kfd2kgd_calls *)&kfd2kgd;
173 }
174 
175 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
176 {
177 	return (struct amdgpu_device *)kgd;
178 }
179 
180 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
181 			uint32_t queue, uint32_t vmid)
182 {
183 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
184 
185 	mutex_lock(&adev->srbm_mutex);
186 	soc15_grbm_select(adev, mec, pipe, queue, vmid);
187 }
188 
189 static void unlock_srbm(struct kgd_dev *kgd)
190 {
191 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
192 
193 	soc15_grbm_select(adev, 0, 0, 0, 0);
194 	mutex_unlock(&adev->srbm_mutex);
195 }
196 
197 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
198 				uint32_t queue_id)
199 {
200 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
201 
202 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
203 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
204 
205 	lock_srbm(kgd, mec, pipe, queue_id, 0);
206 }
207 
208 static uint32_t get_queue_mask(struct amdgpu_device *adev,
209 			       uint32_t pipe_id, uint32_t queue_id)
210 {
211 	unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe +
212 			    queue_id) & 31;
213 
214 	return ((uint32_t)1) << bit;
215 }
216 
217 static void release_queue(struct kgd_dev *kgd)
218 {
219 	unlock_srbm(kgd);
220 }
221 
222 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
223 					uint32_t sh_mem_config,
224 					uint32_t sh_mem_ape1_base,
225 					uint32_t sh_mem_ape1_limit,
226 					uint32_t sh_mem_bases)
227 {
228 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
229 
230 	lock_srbm(kgd, 0, 0, 0, vmid);
231 
232 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
233 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
234 	/* APE1 no longer exists on GFX9 */
235 
236 	unlock_srbm(kgd);
237 }
238 
239 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
240 					unsigned int vmid)
241 {
242 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
243 
244 	/*
245 	 * We have to assume that there is no outstanding mapping.
246 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
247 	 * a mapping is in progress or because a mapping finished
248 	 * and the SW cleared it.
249 	 * So the protocol is to always wait & clear.
250 	 */
251 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
252 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
253 
254 	/*
255 	 * need to do this twice, once for gfx and once for mmhub
256 	 * for ATC add 16 to VMID for mmhub, for IH different registers.
257 	 * ATC_VMID0..15 registers are separate from ATC_VMID16..31.
258 	 */
259 
260 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
261 	       pasid_mapping);
262 
263 	while (!(RREG32(SOC15_REG_OFFSET(
264 				ATHUB, 0,
265 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
266 		 (1U << vmid)))
267 		cpu_relax();
268 
269 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
270 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
271 	       1U << vmid);
272 
273 	/* Mapping vmid to pasid also for IH block */
274 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
275 	       pasid_mapping);
276 
277 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID16_PASID_MAPPING) + vmid,
278 	       pasid_mapping);
279 
280 	while (!(RREG32(SOC15_REG_OFFSET(
281 				ATHUB, 0,
282 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
283 		 (1U << (vmid + 16))))
284 		cpu_relax();
285 
286 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
287 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
288 	       1U << (vmid + 16));
289 
290 	/* Mapping vmid to pasid also for IH block */
291 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid,
292 	       pasid_mapping);
293 	return 0;
294 }
295 
296 /* TODO - RING0 form of field is obsolete, seems to date back to SI
297  * but still works
298  */
299 
300 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
301 {
302 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
303 	uint32_t mec;
304 	uint32_t pipe;
305 
306 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
307 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
308 
309 	lock_srbm(kgd, mec, pipe, 0, 0);
310 
311 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
312 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
313 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
314 
315 	unlock_srbm(kgd);
316 
317 	return 0;
318 }
319 
320 static uint32_t get_sdma_base_addr(struct amdgpu_device *adev,
321 				unsigned int engine_id,
322 				unsigned int queue_id)
323 {
324 	uint32_t base[2] = {
325 		SOC15_REG_OFFSET(SDMA0, 0,
326 				 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
327 		SOC15_REG_OFFSET(SDMA1, 0,
328 				 mmSDMA1_RLC0_RB_CNTL) - mmSDMA1_RLC0_RB_CNTL
329 	};
330 	uint32_t retval;
331 
332 	retval = base[engine_id] + queue_id * (mmSDMA0_RLC1_RB_CNTL -
333 					       mmSDMA0_RLC0_RB_CNTL);
334 
335 	pr_debug("sdma base address: 0x%x\n", retval);
336 
337 	return retval;
338 }
339 
340 static inline struct v9_mqd *get_mqd(void *mqd)
341 {
342 	return (struct v9_mqd *)mqd;
343 }
344 
345 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
346 {
347 	return (struct v9_sdma_mqd *)mqd;
348 }
349 
350 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
351 			uint32_t queue_id, uint32_t __user *wptr,
352 			uint32_t wptr_shift, uint32_t wptr_mask,
353 			struct mm_struct *mm)
354 {
355 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
356 	struct v9_mqd *m;
357 	uint32_t *mqd_hqd;
358 	uint32_t reg, hqd_base, data;
359 
360 	m = get_mqd(mqd);
361 
362 	acquire_queue(kgd, pipe_id, queue_id);
363 
364 	/* HIQ is set during driver init period with vmid set to 0*/
365 	if (m->cp_hqd_vmid == 0) {
366 		uint32_t value, mec, pipe;
367 
368 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
369 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
370 
371 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
372 			mec, pipe, queue_id);
373 		value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS));
374 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
375 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
376 		WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value);
377 	}
378 
379 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
380 	mqd_hqd = &m->cp_mqd_base_addr_lo;
381 	hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
382 
383 	for (reg = hqd_base;
384 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
385 		WREG32(reg, mqd_hqd[reg - hqd_base]);
386 
387 
388 	/* Activate doorbell logic before triggering WPTR poll. */
389 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
390 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
391 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
392 
393 	if (wptr) {
394 		/* Don't read wptr with get_user because the user
395 		 * context may not be accessible (if this function
396 		 * runs in a work queue). Instead trigger a one-shot
397 		 * polling read from memory in the CP. This assumes
398 		 * that wptr is GPU-accessible in the queue's VMID via
399 		 * ATC or SVM. WPTR==RPTR before starting the poll so
400 		 * the CP starts fetching new commands from the right
401 		 * place.
402 		 *
403 		 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
404 		 * tricky. Assume that the queue didn't overflow. The
405 		 * number of valid bits in the 32-bit RPTR depends on
406 		 * the queue size. The remaining bits are taken from
407 		 * the saved 64-bit WPTR. If the WPTR wrapped, add the
408 		 * queue size.
409 		 */
410 		uint32_t queue_size =
411 			2 << REG_GET_FIELD(m->cp_hqd_pq_control,
412 					   CP_HQD_PQ_CONTROL, QUEUE_SIZE);
413 		uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
414 
415 		if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
416 			guessed_wptr += queue_size;
417 		guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
418 		guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
419 
420 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
421 		       lower_32_bits(guessed_wptr));
422 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
423 		       upper_32_bits(guessed_wptr));
424 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
425 		       lower_32_bits((uintptr_t)wptr));
426 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
427 		       upper_32_bits((uintptr_t)wptr));
428 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
429 		       get_queue_mask(adev, pipe_id, queue_id));
430 	}
431 
432 	/* Start the EOP fetcher */
433 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
434 	       REG_SET_FIELD(m->cp_hqd_eop_rptr,
435 			     CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
436 
437 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
438 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
439 
440 	release_queue(kgd);
441 
442 	return 0;
443 }
444 
445 static int kgd_hqd_dump(struct kgd_dev *kgd,
446 			uint32_t pipe_id, uint32_t queue_id,
447 			uint32_t (**dump)[2], uint32_t *n_regs)
448 {
449 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
450 	uint32_t i = 0, reg;
451 #define HQD_N_REGS 56
452 #define DUMP_REG(addr) do {				\
453 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
454 			break;				\
455 		(*dump)[i][0] = (addr) << 2;		\
456 		(*dump)[i++][1] = RREG32(addr);		\
457 	} while (0)
458 
459 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
460 	if (*dump == NULL)
461 		return -ENOMEM;
462 
463 	acquire_queue(kgd, pipe_id, queue_id);
464 
465 	for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
466 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
467 		DUMP_REG(reg);
468 
469 	release_queue(kgd);
470 
471 	WARN_ON_ONCE(i != HQD_N_REGS);
472 	*n_regs = i;
473 
474 	return 0;
475 }
476 
477 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
478 			     uint32_t __user *wptr, struct mm_struct *mm)
479 {
480 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
481 	struct v9_sdma_mqd *m;
482 	uint32_t sdma_base_addr, sdmax_gfx_context_cntl;
483 	unsigned long end_jiffies;
484 	uint32_t data;
485 	uint64_t data64;
486 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
487 
488 	m = get_sdma_mqd(mqd);
489 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
490 					    m->sdma_queue_id);
491 	sdmax_gfx_context_cntl = m->sdma_engine_id ?
492 		SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_CONTEXT_CNTL) :
493 		SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_CONTEXT_CNTL);
494 
495 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
496 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
497 
498 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
499 	while (true) {
500 		data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
501 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
502 			break;
503 		if (time_after(jiffies, end_jiffies))
504 			return -ETIME;
505 		usleep_range(500, 1000);
506 	}
507 	data = RREG32(sdmax_gfx_context_cntl);
508 	data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
509 			     RESUME_CTX, 0);
510 	WREG32(sdmax_gfx_context_cntl, data);
511 
512 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL_OFFSET,
513 	       m->sdmax_rlcx_doorbell_offset);
514 
515 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
516 			     ENABLE, 1);
517 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
518 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
519 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI,
520 				m->sdmax_rlcx_rb_rptr_hi);
521 
522 	WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
523 	if (read_user_wptr(mm, wptr64, data64)) {
524 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
525 		       lower_32_bits(data64));
526 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
527 		       upper_32_bits(data64));
528 	} else {
529 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
530 		       m->sdmax_rlcx_rb_rptr);
531 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
532 		       m->sdmax_rlcx_rb_rptr_hi);
533 	}
534 	WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
535 
536 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
537 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
538 			m->sdmax_rlcx_rb_base_hi);
539 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
540 			m->sdmax_rlcx_rb_rptr_addr_lo);
541 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
542 			m->sdmax_rlcx_rb_rptr_addr_hi);
543 
544 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
545 			     RB_ENABLE, 1);
546 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
547 
548 	return 0;
549 }
550 
551 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
552 			     uint32_t engine_id, uint32_t queue_id,
553 			     uint32_t (**dump)[2], uint32_t *n_regs)
554 {
555 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
556 	uint32_t sdma_base_addr = get_sdma_base_addr(adev, engine_id, queue_id);
557 	uint32_t i = 0, reg;
558 #undef HQD_N_REGS
559 #define HQD_N_REGS (19+6+7+10)
560 
561 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
562 	if (*dump == NULL)
563 		return -ENOMEM;
564 
565 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
566 		DUMP_REG(sdma_base_addr + reg);
567 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
568 		DUMP_REG(sdma_base_addr + reg);
569 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
570 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
571 		DUMP_REG(sdma_base_addr + reg);
572 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
573 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
574 		DUMP_REG(sdma_base_addr + reg);
575 
576 	WARN_ON_ONCE(i != HQD_N_REGS);
577 	*n_regs = i;
578 
579 	return 0;
580 }
581 
582 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
583 				uint32_t pipe_id, uint32_t queue_id)
584 {
585 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
586 	uint32_t act;
587 	bool retval = false;
588 	uint32_t low, high;
589 
590 	acquire_queue(kgd, pipe_id, queue_id);
591 	act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
592 	if (act) {
593 		low = lower_32_bits(queue_address >> 8);
594 		high = upper_32_bits(queue_address >> 8);
595 
596 		if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
597 		   high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
598 			retval = true;
599 	}
600 	release_queue(kgd);
601 	return retval;
602 }
603 
604 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
605 {
606 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
607 	struct v9_sdma_mqd *m;
608 	uint32_t sdma_base_addr;
609 	uint32_t sdma_rlc_rb_cntl;
610 
611 	m = get_sdma_mqd(mqd);
612 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
613 					    m->sdma_queue_id);
614 
615 	sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
616 
617 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
618 		return true;
619 
620 	return false;
621 }
622 
623 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
624 				enum kfd_preempt_type reset_type,
625 				unsigned int utimeout, uint32_t pipe_id,
626 				uint32_t queue_id)
627 {
628 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
629 	enum hqd_dequeue_request_type type;
630 	unsigned long end_jiffies;
631 	uint32_t temp;
632 	struct v9_mqd *m = get_mqd(mqd);
633 
634 	if (adev->in_gpu_reset)
635 		return -EIO;
636 
637 	acquire_queue(kgd, pipe_id, queue_id);
638 
639 	if (m->cp_hqd_vmid == 0)
640 		WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
641 
642 	switch (reset_type) {
643 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
644 		type = DRAIN_PIPE;
645 		break;
646 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
647 		type = RESET_WAVES;
648 		break;
649 	default:
650 		type = DRAIN_PIPE;
651 		break;
652 	}
653 
654 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
655 
656 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
657 	while (true) {
658 		temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
659 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
660 			break;
661 		if (time_after(jiffies, end_jiffies)) {
662 			pr_err("cp queue preemption time out.\n");
663 			release_queue(kgd);
664 			return -ETIME;
665 		}
666 		usleep_range(500, 1000);
667 	}
668 
669 	release_queue(kgd);
670 	return 0;
671 }
672 
673 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
674 				unsigned int utimeout)
675 {
676 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
677 	struct v9_sdma_mqd *m;
678 	uint32_t sdma_base_addr;
679 	uint32_t temp;
680 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
681 
682 	m = get_sdma_mqd(mqd);
683 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
684 					    m->sdma_queue_id);
685 
686 	temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
687 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
688 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
689 
690 	while (true) {
691 		temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
692 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
693 			break;
694 		if (time_after(jiffies, end_jiffies))
695 			return -ETIME;
696 		usleep_range(500, 1000);
697 	}
698 
699 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
700 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
701 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
702 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
703 
704 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
705 	m->sdmax_rlcx_rb_rptr_hi =
706 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI);
707 
708 	return 0;
709 }
710 
711 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
712 							uint8_t vmid)
713 {
714 	uint32_t reg;
715 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
716 
717 	reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
718 		     + vmid);
719 	return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
720 }
721 
722 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
723 								uint8_t vmid)
724 {
725 	uint32_t reg;
726 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
727 
728 	reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
729 		     + vmid);
730 	return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
731 }
732 
733 static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid)
734 {
735 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
736 
737 	/* Use legacy mode tlb invalidation.
738 	 *
739 	 * Currently on Raven the code below is broken for anything but
740 	 * legacy mode due to a MMHUB power gating problem. A workaround
741 	 * is for MMHUB to wait until the condition PER_VMID_INVALIDATE_REQ
742 	 * == PER_VMID_INVALIDATE_ACK instead of simply waiting for the ack
743 	 * bit.
744 	 *
745 	 * TODO 1: agree on the right set of invalidation registers for
746 	 * KFD use. Use the last one for now. Invalidate both GC and
747 	 * MMHUB.
748 	 *
749 	 * TODO 2: support range-based invalidation, requires kfg2kgd
750 	 * interface change
751 	 */
752 	amdgpu_gmc_flush_gpu_tlb(adev, vmid, 0);
753 }
754 
755 static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid)
756 {
757 	signed long r;
758 	uint32_t seq;
759 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
760 
761 	spin_lock(&adev->gfx.kiq.ring_lock);
762 	amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/
763 	amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
764 	amdgpu_ring_write(ring,
765 			PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
766 			PACKET3_INVALIDATE_TLBS_ALL_HUB(1) |
767 			PACKET3_INVALIDATE_TLBS_PASID(pasid) |
768 			PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(0)); /* legacy */
769 	amdgpu_fence_emit_polling(ring, &seq);
770 	amdgpu_ring_commit(ring);
771 	spin_unlock(&adev->gfx.kiq.ring_lock);
772 
773 	r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
774 	if (r < 1) {
775 		DRM_ERROR("wait for kiq fence error: %ld.\n", r);
776 		return -ETIME;
777 	}
778 
779 	return 0;
780 }
781 
782 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
783 {
784 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
785 	int vmid;
786 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
787 
788 	if (adev->in_gpu_reset)
789 		return -EIO;
790 
791 	if (ring->sched.ready)
792 		return invalidate_tlbs_with_kiq(adev, pasid);
793 
794 	for (vmid = 0; vmid < 16; vmid++) {
795 		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
796 			continue;
797 		if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
798 			if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
799 				== pasid) {
800 				write_vmid_invalidate_request(kgd, vmid);
801 				break;
802 			}
803 		}
804 	}
805 
806 	return 0;
807 }
808 
809 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
810 {
811 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
812 
813 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
814 		pr_err("non kfd vmid %d\n", vmid);
815 		return 0;
816 	}
817 
818 	write_vmid_invalidate_request(kgd, vmid);
819 	return 0;
820 }
821 
822 static int kgd_address_watch_disable(struct kgd_dev *kgd)
823 {
824 	return 0;
825 }
826 
827 static int kgd_address_watch_execute(struct kgd_dev *kgd,
828 					unsigned int watch_point_id,
829 					uint32_t cntl_val,
830 					uint32_t addr_hi,
831 					uint32_t addr_lo)
832 {
833 	return 0;
834 }
835 
836 static int kgd_wave_control_execute(struct kgd_dev *kgd,
837 					uint32_t gfx_index_val,
838 					uint32_t sq_cmd)
839 {
840 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
841 	uint32_t data = 0;
842 
843 	mutex_lock(&adev->grbm_idx_mutex);
844 
845 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
846 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
847 
848 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
849 		INSTANCE_BROADCAST_WRITES, 1);
850 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
851 		SH_BROADCAST_WRITES, 1);
852 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
853 		SE_BROADCAST_WRITES, 1);
854 
855 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
856 	mutex_unlock(&adev->grbm_idx_mutex);
857 
858 	return 0;
859 }
860 
861 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
862 					unsigned int watch_point_id,
863 					unsigned int reg_offset)
864 {
865 	return 0;
866 }
867 
868 static void set_scratch_backing_va(struct kgd_dev *kgd,
869 					uint64_t va, uint32_t vmid)
870 {
871 	/* No longer needed on GFXv9. The scratch base address is
872 	 * passed to the shader by the CP. It's the user mode driver's
873 	 * responsibility.
874 	 */
875 }
876 
877 /* FIXME: Does this need to be ASIC-specific code? */
878 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
879 {
880 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
881 	const union amdgpu_firmware_header *hdr;
882 
883 	switch (type) {
884 	case KGD_ENGINE_PFP:
885 		hdr = (const union amdgpu_firmware_header *)adev->gfx.pfp_fw->data;
886 		break;
887 
888 	case KGD_ENGINE_ME:
889 		hdr = (const union amdgpu_firmware_header *)adev->gfx.me_fw->data;
890 		break;
891 
892 	case KGD_ENGINE_CE:
893 		hdr = (const union amdgpu_firmware_header *)adev->gfx.ce_fw->data;
894 		break;
895 
896 	case KGD_ENGINE_MEC1:
897 		hdr = (const union amdgpu_firmware_header *)adev->gfx.mec_fw->data;
898 		break;
899 
900 	case KGD_ENGINE_MEC2:
901 		hdr = (const union amdgpu_firmware_header *)adev->gfx.mec2_fw->data;
902 		break;
903 
904 	case KGD_ENGINE_RLC:
905 		hdr = (const union amdgpu_firmware_header *)adev->gfx.rlc_fw->data;
906 		break;
907 
908 	case KGD_ENGINE_SDMA1:
909 		hdr = (const union amdgpu_firmware_header *)adev->sdma.instance[0].fw->data;
910 		break;
911 
912 	case KGD_ENGINE_SDMA2:
913 		hdr = (const union amdgpu_firmware_header *)adev->sdma.instance[1].fw->data;
914 		break;
915 
916 	default:
917 		return 0;
918 	}
919 
920 	if (hdr == NULL)
921 		return 0;
922 
923 	/* Only 12 bit in use*/
924 	return hdr->common.ucode_version;
925 }
926 
927 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
928 		uint64_t page_table_base)
929 {
930 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
931 
932 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
933 		pr_err("trying to set page table base for wrong VMID %u\n",
934 		       vmid);
935 		return;
936 	}
937 
938 	/* TODO: take advantage of per-process address space size. For
939 	 * now, all processes share the same address space size, like
940 	 * on GFX8 and older.
941 	 */
942 	mmhub_v1_0_setup_vm_pt_regs(adev, vmid, page_table_base);
943 
944 	gfxhub_v1_0_setup_vm_pt_regs(adev, vmid, page_table_base);
945 }
946