xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c (revision 160b8e75932fd51a49607d32dbfa1d417977b79c)
1 /*
2  * Copyright 2014 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 #include <linux/module.h>
24 #include <linux/fdtable.h>
25 #include <linux/uaccess.h>
26 #include <linux/firmware.h>
27 #include <drm/drmP.h>
28 #include "amdgpu.h"
29 #include "amdgpu_amdkfd.h"
30 #include "amdgpu_ucode.h"
31 #include "gfx_v8_0.h"
32 #include "gca/gfx_8_0_sh_mask.h"
33 #include "gca/gfx_8_0_d.h"
34 #include "gca/gfx_8_0_enum.h"
35 #include "oss/oss_3_0_sh_mask.h"
36 #include "oss/oss_3_0_d.h"
37 #include "gmc/gmc_8_1_sh_mask.h"
38 #include "gmc/gmc_8_1_d.h"
39 #include "vi_structs.h"
40 #include "vid.h"
41 
42 enum hqd_dequeue_request_type {
43 	NO_ACTION = 0,
44 	DRAIN_PIPE,
45 	RESET_WAVES
46 };
47 
48 struct vi_sdma_mqd;
49 
50 /*
51  * Register access functions
52  */
53 
54 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
55 		uint32_t sh_mem_config,
56 		uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit,
57 		uint32_t sh_mem_bases);
58 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
59 		unsigned int vmid);
60 static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
61 		uint32_t hpd_size, uint64_t hpd_gpu_addr);
62 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
63 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
64 			uint32_t queue_id, uint32_t __user *wptr,
65 			uint32_t wptr_shift, uint32_t wptr_mask,
66 			struct mm_struct *mm);
67 static int kgd_hqd_dump(struct kgd_dev *kgd,
68 			uint32_t pipe_id, uint32_t queue_id,
69 			uint32_t (**dump)[2], uint32_t *n_regs);
70 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
71 			     uint32_t __user *wptr, struct mm_struct *mm);
72 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
73 			     uint32_t engine_id, uint32_t queue_id,
74 			     uint32_t (**dump)[2], uint32_t *n_regs);
75 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
76 		uint32_t pipe_id, uint32_t queue_id);
77 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
78 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
79 				enum kfd_preempt_type reset_type,
80 				unsigned int utimeout, uint32_t pipe_id,
81 				uint32_t queue_id);
82 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
83 				unsigned int utimeout);
84 static int kgd_address_watch_disable(struct kgd_dev *kgd);
85 static int kgd_address_watch_execute(struct kgd_dev *kgd,
86 					unsigned int watch_point_id,
87 					uint32_t cntl_val,
88 					uint32_t addr_hi,
89 					uint32_t addr_lo);
90 static int kgd_wave_control_execute(struct kgd_dev *kgd,
91 					uint32_t gfx_index_val,
92 					uint32_t sq_cmd);
93 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
94 					unsigned int watch_point_id,
95 					unsigned int reg_offset);
96 
97 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
98 		uint8_t vmid);
99 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
100 		uint8_t vmid);
101 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
102 static void set_scratch_backing_va(struct kgd_dev *kgd,
103 					uint64_t va, uint32_t vmid);
104 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
105 		uint32_t page_table_base);
106 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
107 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
108 
109 /* Because of REG_GET_FIELD() being used, we put this function in the
110  * asic specific file.
111  */
112 static int get_tile_config(struct kgd_dev *kgd,
113 		struct tile_config *config)
114 {
115 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
116 
117 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
118 	config->num_banks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
119 				MC_ARB_RAMCFG, NOOFBANK);
120 	config->num_ranks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
121 				MC_ARB_RAMCFG, NOOFRANKS);
122 
123 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
124 	config->num_tile_configs =
125 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
126 	config->macro_tile_config_ptr =
127 			adev->gfx.config.macrotile_mode_array;
128 	config->num_macro_tile_configs =
129 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
130 
131 	return 0;
132 }
133 
134 static const struct kfd2kgd_calls kfd2kgd = {
135 	.init_gtt_mem_allocation = alloc_gtt_mem,
136 	.free_gtt_mem = free_gtt_mem,
137 	.get_local_mem_info = get_local_mem_info,
138 	.get_gpu_clock_counter = get_gpu_clock_counter,
139 	.get_max_engine_clock_in_mhz = get_max_engine_clock_in_mhz,
140 	.alloc_pasid = amdgpu_pasid_alloc,
141 	.free_pasid = amdgpu_pasid_free,
142 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
143 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
144 	.init_pipeline = kgd_init_pipeline,
145 	.init_interrupts = kgd_init_interrupts,
146 	.hqd_load = kgd_hqd_load,
147 	.hqd_sdma_load = kgd_hqd_sdma_load,
148 	.hqd_dump = kgd_hqd_dump,
149 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
150 	.hqd_is_occupied = kgd_hqd_is_occupied,
151 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
152 	.hqd_destroy = kgd_hqd_destroy,
153 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
154 	.address_watch_disable = kgd_address_watch_disable,
155 	.address_watch_execute = kgd_address_watch_execute,
156 	.wave_control_execute = kgd_wave_control_execute,
157 	.address_watch_get_offset = kgd_address_watch_get_offset,
158 	.get_atc_vmid_pasid_mapping_pasid =
159 			get_atc_vmid_pasid_mapping_pasid,
160 	.get_atc_vmid_pasid_mapping_valid =
161 			get_atc_vmid_pasid_mapping_valid,
162 	.get_fw_version = get_fw_version,
163 	.set_scratch_backing_va = set_scratch_backing_va,
164 	.get_tile_config = get_tile_config,
165 	.get_cu_info = get_cu_info,
166 	.get_vram_usage = amdgpu_amdkfd_get_vram_usage,
167 	.create_process_vm = amdgpu_amdkfd_gpuvm_create_process_vm,
168 	.destroy_process_vm = amdgpu_amdkfd_gpuvm_destroy_process_vm,
169 	.get_process_page_dir = amdgpu_amdkfd_gpuvm_get_process_page_dir,
170 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
171 	.alloc_memory_of_gpu = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu,
172 	.free_memory_of_gpu = amdgpu_amdkfd_gpuvm_free_memory_of_gpu,
173 	.map_memory_to_gpu = amdgpu_amdkfd_gpuvm_map_memory_to_gpu,
174 	.unmap_memory_to_gpu = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu,
175 	.sync_memory = amdgpu_amdkfd_gpuvm_sync_memory,
176 	.map_gtt_bo_to_kernel = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel,
177 	.restore_process_bos = amdgpu_amdkfd_gpuvm_restore_process_bos,
178 	.invalidate_tlbs = invalidate_tlbs,
179 	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
180 	.submit_ib = amdgpu_amdkfd_submit_ib,
181 };
182 
183 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_8_0_get_functions(void)
184 {
185 	return (struct kfd2kgd_calls *)&kfd2kgd;
186 }
187 
188 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
189 {
190 	return (struct amdgpu_device *)kgd;
191 }
192 
193 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
194 			uint32_t queue, uint32_t vmid)
195 {
196 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
197 	uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
198 
199 	mutex_lock(&adev->srbm_mutex);
200 	WREG32(mmSRBM_GFX_CNTL, value);
201 }
202 
203 static void unlock_srbm(struct kgd_dev *kgd)
204 {
205 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
206 
207 	WREG32(mmSRBM_GFX_CNTL, 0);
208 	mutex_unlock(&adev->srbm_mutex);
209 }
210 
211 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
212 				uint32_t queue_id)
213 {
214 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
215 
216 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
217 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
218 
219 	lock_srbm(kgd, mec, pipe, queue_id, 0);
220 }
221 
222 static void release_queue(struct kgd_dev *kgd)
223 {
224 	unlock_srbm(kgd);
225 }
226 
227 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
228 					uint32_t sh_mem_config,
229 					uint32_t sh_mem_ape1_base,
230 					uint32_t sh_mem_ape1_limit,
231 					uint32_t sh_mem_bases)
232 {
233 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
234 
235 	lock_srbm(kgd, 0, 0, 0, vmid);
236 
237 	WREG32(mmSH_MEM_CONFIG, sh_mem_config);
238 	WREG32(mmSH_MEM_APE1_BASE, sh_mem_ape1_base);
239 	WREG32(mmSH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
240 	WREG32(mmSH_MEM_BASES, sh_mem_bases);
241 
242 	unlock_srbm(kgd);
243 }
244 
245 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
246 					unsigned int vmid)
247 {
248 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
249 
250 	/*
251 	 * We have to assume that there is no outstanding mapping.
252 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
253 	 * a mapping is in progress or because a mapping finished
254 	 * and the SW cleared it.
255 	 * So the protocol is to always wait & clear.
256 	 */
257 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
258 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
259 
260 	WREG32(mmATC_VMID0_PASID_MAPPING + vmid, pasid_mapping);
261 
262 	while (!(RREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))
263 		cpu_relax();
264 	WREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
265 
266 	/* Mapping vmid to pasid also for IH block */
267 	WREG32(mmIH_VMID_0_LUT + vmid, pasid_mapping);
268 
269 	return 0;
270 }
271 
272 static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
273 				uint32_t hpd_size, uint64_t hpd_gpu_addr)
274 {
275 	/* amdgpu owns the per-pipe state */
276 	return 0;
277 }
278 
279 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
280 {
281 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
282 	uint32_t mec;
283 	uint32_t pipe;
284 
285 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
286 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
287 
288 	lock_srbm(kgd, mec, pipe, 0, 0);
289 
290 	WREG32(mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK);
291 
292 	unlock_srbm(kgd);
293 
294 	return 0;
295 }
296 
297 static inline uint32_t get_sdma_base_addr(struct vi_sdma_mqd *m)
298 {
299 	uint32_t retval;
300 
301 	retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
302 		m->sdma_queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
303 	pr_debug("kfd: sdma base address: 0x%x\n", retval);
304 
305 	return retval;
306 }
307 
308 static inline struct vi_mqd *get_mqd(void *mqd)
309 {
310 	return (struct vi_mqd *)mqd;
311 }
312 
313 static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd)
314 {
315 	return (struct vi_sdma_mqd *)mqd;
316 }
317 
318 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
319 			uint32_t queue_id, uint32_t __user *wptr,
320 			uint32_t wptr_shift, uint32_t wptr_mask,
321 			struct mm_struct *mm)
322 {
323 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
324 	struct vi_mqd *m;
325 	uint32_t *mqd_hqd;
326 	uint32_t reg, wptr_val, data;
327 	bool valid_wptr = false;
328 
329 	m = get_mqd(mqd);
330 
331 	acquire_queue(kgd, pipe_id, queue_id);
332 
333 	/* HIQ is set during driver init period with vmid set to 0*/
334 	if (m->cp_hqd_vmid == 0) {
335 		uint32_t value, mec, pipe;
336 
337 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
338 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
339 
340 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
341 			mec, pipe, queue_id);
342 		value = RREG32(mmRLC_CP_SCHEDULERS);
343 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
344 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
345 		WREG32(mmRLC_CP_SCHEDULERS, value);
346 	}
347 
348 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
349 	mqd_hqd = &m->cp_mqd_base_addr_lo;
350 
351 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_CONTROL; reg++)
352 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
353 
354 	/* Tonga errata: EOP RPTR/WPTR should be left unmodified.
355 	 * This is safe since EOP RPTR==WPTR for any inactive HQD
356 	 * on ASICs that do not support context-save.
357 	 * EOP writes/reads can start anywhere in the ring.
358 	 */
359 	if (get_amdgpu_device(kgd)->asic_type != CHIP_TONGA) {
360 		WREG32(mmCP_HQD_EOP_RPTR, m->cp_hqd_eop_rptr);
361 		WREG32(mmCP_HQD_EOP_WPTR, m->cp_hqd_eop_wptr);
362 		WREG32(mmCP_HQD_EOP_WPTR_MEM, m->cp_hqd_eop_wptr_mem);
363 	}
364 
365 	for (reg = mmCP_HQD_EOP_EVENTS; reg <= mmCP_HQD_ERROR; reg++)
366 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
367 
368 	/* Copy userspace write pointer value to register.
369 	 * Activate doorbell logic to monitor subsequent changes.
370 	 */
371 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
372 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
373 	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
374 
375 	/* read_user_ptr may take the mm->mmap_sem.
376 	 * release srbm_mutex to avoid circular dependency between
377 	 * srbm_mutex->mm_sem->reservation_ww_class_mutex->srbm_mutex.
378 	 */
379 	release_queue(kgd);
380 	valid_wptr = read_user_wptr(mm, wptr, wptr_val);
381 	acquire_queue(kgd, pipe_id, queue_id);
382 	if (valid_wptr)
383 		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
384 
385 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
386 	WREG32(mmCP_HQD_ACTIVE, data);
387 
388 	release_queue(kgd);
389 
390 	return 0;
391 }
392 
393 static int kgd_hqd_dump(struct kgd_dev *kgd,
394 			uint32_t pipe_id, uint32_t queue_id,
395 			uint32_t (**dump)[2], uint32_t *n_regs)
396 {
397 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
398 	uint32_t i = 0, reg;
399 #define HQD_N_REGS (54+4)
400 #define DUMP_REG(addr) do {				\
401 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
402 			break;				\
403 		(*dump)[i][0] = (addr) << 2;		\
404 		(*dump)[i++][1] = RREG32(addr);		\
405 	} while (0)
406 
407 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
408 	if (*dump == NULL)
409 		return -ENOMEM;
410 
411 	acquire_queue(kgd, pipe_id, queue_id);
412 
413 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE0);
414 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE1);
415 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE2);
416 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE3);
417 
418 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_DONES; reg++)
419 		DUMP_REG(reg);
420 
421 	release_queue(kgd);
422 
423 	WARN_ON_ONCE(i != HQD_N_REGS);
424 	*n_regs = i;
425 
426 	return 0;
427 }
428 
429 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
430 			     uint32_t __user *wptr, struct mm_struct *mm)
431 {
432 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
433 	struct vi_sdma_mqd *m;
434 	unsigned long end_jiffies;
435 	uint32_t sdma_base_addr;
436 	uint32_t data;
437 
438 	m = get_sdma_mqd(mqd);
439 	sdma_base_addr = get_sdma_base_addr(m);
440 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
441 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
442 
443 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
444 	while (true) {
445 		data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
446 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
447 			break;
448 		if (time_after(jiffies, end_jiffies))
449 			return -ETIME;
450 		usleep_range(500, 1000);
451 	}
452 	if (m->sdma_engine_id) {
453 		data = RREG32(mmSDMA1_GFX_CONTEXT_CNTL);
454 		data = REG_SET_FIELD(data, SDMA1_GFX_CONTEXT_CNTL,
455 				RESUME_CTX, 0);
456 		WREG32(mmSDMA1_GFX_CONTEXT_CNTL, data);
457 	} else {
458 		data = RREG32(mmSDMA0_GFX_CONTEXT_CNTL);
459 		data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
460 				RESUME_CTX, 0);
461 		WREG32(mmSDMA0_GFX_CONTEXT_CNTL, data);
462 	}
463 
464 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
465 			     ENABLE, 1);
466 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
467 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
468 
469 	if (read_user_wptr(mm, wptr, data))
470 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR, data);
471 	else
472 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
473 		       m->sdmax_rlcx_rb_rptr);
474 
475 	WREG32(sdma_base_addr + mmSDMA0_RLC0_VIRTUAL_ADDR,
476 				m->sdmax_rlcx_virtual_addr);
477 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
478 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
479 			m->sdmax_rlcx_rb_base_hi);
480 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
481 			m->sdmax_rlcx_rb_rptr_addr_lo);
482 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
483 			m->sdmax_rlcx_rb_rptr_addr_hi);
484 
485 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
486 			     RB_ENABLE, 1);
487 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
488 
489 	return 0;
490 }
491 
492 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
493 			     uint32_t engine_id, uint32_t queue_id,
494 			     uint32_t (**dump)[2], uint32_t *n_regs)
495 {
496 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
497 	uint32_t sdma_offset = engine_id * SDMA1_REGISTER_OFFSET +
498 		queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
499 	uint32_t i = 0, reg;
500 #undef HQD_N_REGS
501 #define HQD_N_REGS (19+4+2+3+7)
502 
503 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
504 	if (*dump == NULL)
505 		return -ENOMEM;
506 
507 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
508 		DUMP_REG(sdma_offset + reg);
509 	for (reg = mmSDMA0_RLC0_VIRTUAL_ADDR; reg <= mmSDMA0_RLC0_WATERMARK;
510 	     reg++)
511 		DUMP_REG(sdma_offset + reg);
512 	for (reg = mmSDMA0_RLC0_CSA_ADDR_LO; reg <= mmSDMA0_RLC0_CSA_ADDR_HI;
513 	     reg++)
514 		DUMP_REG(sdma_offset + reg);
515 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN; reg <= mmSDMA0_RLC0_DUMMY_REG;
516 	     reg++)
517 		DUMP_REG(sdma_offset + reg);
518 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0; reg <= mmSDMA0_RLC0_MIDCMD_CNTL;
519 	     reg++)
520 		DUMP_REG(sdma_offset + reg);
521 
522 	WARN_ON_ONCE(i != HQD_N_REGS);
523 	*n_regs = i;
524 
525 	return 0;
526 }
527 
528 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
529 				uint32_t pipe_id, uint32_t queue_id)
530 {
531 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
532 	uint32_t act;
533 	bool retval = false;
534 	uint32_t low, high;
535 
536 	acquire_queue(kgd, pipe_id, queue_id);
537 	act = RREG32(mmCP_HQD_ACTIVE);
538 	if (act) {
539 		low = lower_32_bits(queue_address >> 8);
540 		high = upper_32_bits(queue_address >> 8);
541 
542 		if (low == RREG32(mmCP_HQD_PQ_BASE) &&
543 				high == RREG32(mmCP_HQD_PQ_BASE_HI))
544 			retval = true;
545 	}
546 	release_queue(kgd);
547 	return retval;
548 }
549 
550 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
551 {
552 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
553 	struct vi_sdma_mqd *m;
554 	uint32_t sdma_base_addr;
555 	uint32_t sdma_rlc_rb_cntl;
556 
557 	m = get_sdma_mqd(mqd);
558 	sdma_base_addr = get_sdma_base_addr(m);
559 
560 	sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
561 
562 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
563 		return true;
564 
565 	return false;
566 }
567 
568 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
569 				enum kfd_preempt_type reset_type,
570 				unsigned int utimeout, uint32_t pipe_id,
571 				uint32_t queue_id)
572 {
573 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
574 	uint32_t temp;
575 	enum hqd_dequeue_request_type type;
576 	unsigned long flags, end_jiffies;
577 	int retry;
578 	struct vi_mqd *m = get_mqd(mqd);
579 
580 	acquire_queue(kgd, pipe_id, queue_id);
581 
582 	if (m->cp_hqd_vmid == 0)
583 		WREG32_FIELD(RLC_CP_SCHEDULERS, scheduler1, 0);
584 
585 	switch (reset_type) {
586 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
587 		type = DRAIN_PIPE;
588 		break;
589 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
590 		type = RESET_WAVES;
591 		break;
592 	default:
593 		type = DRAIN_PIPE;
594 		break;
595 	}
596 
597 	/* Workaround: If IQ timer is active and the wait time is close to or
598 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
599 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
600 	 * cleared before continuing. Also, ensure wait times are set to at
601 	 * least 0x3.
602 	 */
603 	local_irq_save(flags);
604 	preempt_disable();
605 	retry = 5000; /* wait for 500 usecs at maximum */
606 	while (true) {
607 		temp = RREG32(mmCP_HQD_IQ_TIMER);
608 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
609 			pr_debug("HW is processing IQ\n");
610 			goto loop;
611 		}
612 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
613 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
614 					== 3) /* SEM-rearm is safe */
615 				break;
616 			/* Wait time 3 is safe for CP, but our MMIO read/write
617 			 * time is close to 1 microsecond, so check for 10 to
618 			 * leave more buffer room
619 			 */
620 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
621 					>= 10)
622 				break;
623 			pr_debug("IQ timer is active\n");
624 		} else
625 			break;
626 loop:
627 		if (!retry) {
628 			pr_err("CP HQD IQ timer status time out\n");
629 			break;
630 		}
631 		ndelay(100);
632 		--retry;
633 	}
634 	retry = 1000;
635 	while (true) {
636 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
637 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
638 			break;
639 		pr_debug("Dequeue request is pending\n");
640 
641 		if (!retry) {
642 			pr_err("CP HQD dequeue request time out\n");
643 			break;
644 		}
645 		ndelay(100);
646 		--retry;
647 	}
648 	local_irq_restore(flags);
649 	preempt_enable();
650 
651 	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
652 
653 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
654 	while (true) {
655 		temp = RREG32(mmCP_HQD_ACTIVE);
656 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
657 			break;
658 		if (time_after(jiffies, end_jiffies)) {
659 			pr_err("cp queue preemption time out.\n");
660 			release_queue(kgd);
661 			return -ETIME;
662 		}
663 		usleep_range(500, 1000);
664 	}
665 
666 	release_queue(kgd);
667 	return 0;
668 }
669 
670 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
671 				unsigned int utimeout)
672 {
673 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
674 	struct vi_sdma_mqd *m;
675 	uint32_t sdma_base_addr;
676 	uint32_t temp;
677 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
678 
679 	m = get_sdma_mqd(mqd);
680 	sdma_base_addr = get_sdma_base_addr(m);
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 
702 	return 0;
703 }
704 
705 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
706 							uint8_t vmid)
707 {
708 	uint32_t reg;
709 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
710 
711 	reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
712 	return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
713 }
714 
715 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
716 								uint8_t vmid)
717 {
718 	uint32_t reg;
719 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
720 
721 	reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
722 	return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
723 }
724 
725 static int kgd_address_watch_disable(struct kgd_dev *kgd)
726 {
727 	return 0;
728 }
729 
730 static int kgd_address_watch_execute(struct kgd_dev *kgd,
731 					unsigned int watch_point_id,
732 					uint32_t cntl_val,
733 					uint32_t addr_hi,
734 					uint32_t addr_lo)
735 {
736 	return 0;
737 }
738 
739 static int kgd_wave_control_execute(struct kgd_dev *kgd,
740 					uint32_t gfx_index_val,
741 					uint32_t sq_cmd)
742 {
743 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
744 	uint32_t data = 0;
745 
746 	mutex_lock(&adev->grbm_idx_mutex);
747 
748 	WREG32(mmGRBM_GFX_INDEX, gfx_index_val);
749 	WREG32(mmSQ_CMD, sq_cmd);
750 
751 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
752 		INSTANCE_BROADCAST_WRITES, 1);
753 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
754 		SH_BROADCAST_WRITES, 1);
755 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
756 		SE_BROADCAST_WRITES, 1);
757 
758 	WREG32(mmGRBM_GFX_INDEX, data);
759 	mutex_unlock(&adev->grbm_idx_mutex);
760 
761 	return 0;
762 }
763 
764 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
765 					unsigned int watch_point_id,
766 					unsigned int reg_offset)
767 {
768 	return 0;
769 }
770 
771 static void set_scratch_backing_va(struct kgd_dev *kgd,
772 					uint64_t va, uint32_t vmid)
773 {
774 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
775 
776 	lock_srbm(kgd, 0, 0, 0, vmid);
777 	WREG32(mmSH_HIDDEN_PRIVATE_BASE_VMID, va);
778 	unlock_srbm(kgd);
779 }
780 
781 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
782 {
783 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
784 	const union amdgpu_firmware_header *hdr;
785 
786 	switch (type) {
787 	case KGD_ENGINE_PFP:
788 		hdr = (const union amdgpu_firmware_header *)
789 						adev->gfx.pfp_fw->data;
790 		break;
791 
792 	case KGD_ENGINE_ME:
793 		hdr = (const union amdgpu_firmware_header *)
794 						adev->gfx.me_fw->data;
795 		break;
796 
797 	case KGD_ENGINE_CE:
798 		hdr = (const union amdgpu_firmware_header *)
799 						adev->gfx.ce_fw->data;
800 		break;
801 
802 	case KGD_ENGINE_MEC1:
803 		hdr = (const union amdgpu_firmware_header *)
804 						adev->gfx.mec_fw->data;
805 		break;
806 
807 	case KGD_ENGINE_MEC2:
808 		hdr = (const union amdgpu_firmware_header *)
809 						adev->gfx.mec2_fw->data;
810 		break;
811 
812 	case KGD_ENGINE_RLC:
813 		hdr = (const union amdgpu_firmware_header *)
814 						adev->gfx.rlc_fw->data;
815 		break;
816 
817 	case KGD_ENGINE_SDMA1:
818 		hdr = (const union amdgpu_firmware_header *)
819 						adev->sdma.instance[0].fw->data;
820 		break;
821 
822 	case KGD_ENGINE_SDMA2:
823 		hdr = (const union amdgpu_firmware_header *)
824 						adev->sdma.instance[1].fw->data;
825 		break;
826 
827 	default:
828 		return 0;
829 	}
830 
831 	if (hdr == NULL)
832 		return 0;
833 
834 	/* Only 12 bit in use*/
835 	return hdr->common.ucode_version;
836 }
837 
838 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
839 		uint32_t page_table_base)
840 {
841 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
842 
843 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
844 		pr_err("trying to set page table base for wrong VMID\n");
845 		return;
846 	}
847 	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8, page_table_base);
848 }
849 
850 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
851 {
852 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
853 	int vmid;
854 	unsigned int tmp;
855 
856 	for (vmid = 0; vmid < 16; vmid++) {
857 		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
858 			continue;
859 
860 		tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
861 		if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
862 			(tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
863 			WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
864 			RREG32(mmVM_INVALIDATE_RESPONSE);
865 			break;
866 		}
867 	}
868 
869 	return 0;
870 }
871 
872 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
873 {
874 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
875 
876 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
877 		pr_err("non kfd vmid %d\n", vmid);
878 		return -EINVAL;
879 	}
880 
881 	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
882 	RREG32(mmVM_INVALIDATE_RESPONSE);
883 	return 0;
884 }
885