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/mmu_context.h>
27 #include <drm/drmP.h>
28 #include "amdgpu.h"
29 #include "amdgpu_amdkfd.h"
30 #include "gfx_v8_0.h"
31 #include "gca/gfx_8_0_sh_mask.h"
32 #include "gca/gfx_8_0_d.h"
33 #include "gca/gfx_8_0_enum.h"
34 #include "oss/oss_3_0_sh_mask.h"
35 #include "oss/oss_3_0_d.h"
36 #include "gmc/gmc_8_1_sh_mask.h"
37 #include "gmc/gmc_8_1_d.h"
38 #include "vi_structs.h"
39 #include "vid.h"
40 
41 enum hqd_dequeue_request_type {
42 	NO_ACTION = 0,
43 	DRAIN_PIPE,
44 	RESET_WAVES
45 };
46 
47 /*
48  * Register access functions
49  */
50 
51 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
52 		uint32_t sh_mem_config,
53 		uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit,
54 		uint32_t sh_mem_bases);
55 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
56 		unsigned int vmid);
57 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
58 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
59 			uint32_t queue_id, uint32_t __user *wptr,
60 			uint32_t wptr_shift, uint32_t wptr_mask,
61 			struct mm_struct *mm);
62 static int kgd_hqd_dump(struct kgd_dev *kgd,
63 			uint32_t pipe_id, uint32_t queue_id,
64 			uint32_t (**dump)[2], uint32_t *n_regs);
65 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
66 			     uint32_t __user *wptr, struct mm_struct *mm);
67 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
68 			     uint32_t engine_id, uint32_t queue_id,
69 			     uint32_t (**dump)[2], uint32_t *n_regs);
70 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
71 		uint32_t pipe_id, uint32_t queue_id);
72 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
73 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
74 				enum kfd_preempt_type reset_type,
75 				unsigned int utimeout, uint32_t pipe_id,
76 				uint32_t queue_id);
77 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
78 				unsigned int utimeout);
79 static int kgd_address_watch_disable(struct kgd_dev *kgd);
80 static int kgd_address_watch_execute(struct kgd_dev *kgd,
81 					unsigned int watch_point_id,
82 					uint32_t cntl_val,
83 					uint32_t addr_hi,
84 					uint32_t addr_lo);
85 static int kgd_wave_control_execute(struct kgd_dev *kgd,
86 					uint32_t gfx_index_val,
87 					uint32_t sq_cmd);
88 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
89 					unsigned int watch_point_id,
90 					unsigned int reg_offset);
91 
92 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
93 		uint8_t vmid);
94 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
95 		uint8_t vmid);
96 static void set_scratch_backing_va(struct kgd_dev *kgd,
97 					uint64_t va, uint32_t vmid);
98 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
99 		uint64_t page_table_base);
100 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
101 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
102 
103 /* Because of REG_GET_FIELD() being used, we put this function in the
104  * asic specific file.
105  */
106 static int get_tile_config(struct kgd_dev *kgd,
107 		struct tile_config *config)
108 {
109 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
110 
111 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
112 	config->num_banks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
113 				MC_ARB_RAMCFG, NOOFBANK);
114 	config->num_ranks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
115 				MC_ARB_RAMCFG, NOOFRANKS);
116 
117 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
118 	config->num_tile_configs =
119 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
120 	config->macro_tile_config_ptr =
121 			adev->gfx.config.macrotile_mode_array;
122 	config->num_macro_tile_configs =
123 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
124 
125 	return 0;
126 }
127 
128 static const struct kfd2kgd_calls kfd2kgd = {
129 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
130 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
131 	.init_interrupts = kgd_init_interrupts,
132 	.hqd_load = kgd_hqd_load,
133 	.hqd_sdma_load = kgd_hqd_sdma_load,
134 	.hqd_dump = kgd_hqd_dump,
135 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
136 	.hqd_is_occupied = kgd_hqd_is_occupied,
137 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
138 	.hqd_destroy = kgd_hqd_destroy,
139 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
140 	.address_watch_disable = kgd_address_watch_disable,
141 	.address_watch_execute = kgd_address_watch_execute,
142 	.wave_control_execute = kgd_wave_control_execute,
143 	.address_watch_get_offset = kgd_address_watch_get_offset,
144 	.get_atc_vmid_pasid_mapping_pasid =
145 			get_atc_vmid_pasid_mapping_pasid,
146 	.get_atc_vmid_pasid_mapping_valid =
147 			get_atc_vmid_pasid_mapping_valid,
148 	.set_scratch_backing_va = set_scratch_backing_va,
149 	.get_tile_config = get_tile_config,
150 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
151 	.invalidate_tlbs = invalidate_tlbs,
152 	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
153 };
154 
155 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_8_0_get_functions(void)
156 {
157 	return (struct kfd2kgd_calls *)&kfd2kgd;
158 }
159 
160 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
161 {
162 	return (struct amdgpu_device *)kgd;
163 }
164 
165 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
166 			uint32_t queue, uint32_t vmid)
167 {
168 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
169 	uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
170 
171 	mutex_lock(&adev->srbm_mutex);
172 	WREG32(mmSRBM_GFX_CNTL, value);
173 }
174 
175 static void unlock_srbm(struct kgd_dev *kgd)
176 {
177 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
178 
179 	WREG32(mmSRBM_GFX_CNTL, 0);
180 	mutex_unlock(&adev->srbm_mutex);
181 }
182 
183 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
184 				uint32_t queue_id)
185 {
186 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
187 
188 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
189 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
190 
191 	lock_srbm(kgd, mec, pipe, queue_id, 0);
192 }
193 
194 static void release_queue(struct kgd_dev *kgd)
195 {
196 	unlock_srbm(kgd);
197 }
198 
199 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
200 					uint32_t sh_mem_config,
201 					uint32_t sh_mem_ape1_base,
202 					uint32_t sh_mem_ape1_limit,
203 					uint32_t sh_mem_bases)
204 {
205 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
206 
207 	lock_srbm(kgd, 0, 0, 0, vmid);
208 
209 	WREG32(mmSH_MEM_CONFIG, sh_mem_config);
210 	WREG32(mmSH_MEM_APE1_BASE, sh_mem_ape1_base);
211 	WREG32(mmSH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
212 	WREG32(mmSH_MEM_BASES, sh_mem_bases);
213 
214 	unlock_srbm(kgd);
215 }
216 
217 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
218 					unsigned int vmid)
219 {
220 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
221 
222 	/*
223 	 * We have to assume that there is no outstanding mapping.
224 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
225 	 * a mapping is in progress or because a mapping finished
226 	 * and the SW cleared it.
227 	 * So the protocol is to always wait & clear.
228 	 */
229 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
230 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
231 
232 	WREG32(mmATC_VMID0_PASID_MAPPING + vmid, pasid_mapping);
233 
234 	while (!(RREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))
235 		cpu_relax();
236 	WREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
237 
238 	/* Mapping vmid to pasid also for IH block */
239 	WREG32(mmIH_VMID_0_LUT + vmid, pasid_mapping);
240 
241 	return 0;
242 }
243 
244 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
245 {
246 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
247 	uint32_t mec;
248 	uint32_t pipe;
249 
250 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
251 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
252 
253 	lock_srbm(kgd, mec, pipe, 0, 0);
254 
255 	WREG32(mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
256 			CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
257 
258 	unlock_srbm(kgd);
259 
260 	return 0;
261 }
262 
263 static inline uint32_t get_sdma_base_addr(struct vi_sdma_mqd *m)
264 {
265 	uint32_t retval;
266 
267 	retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
268 		m->sdma_queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
269 	pr_debug("kfd: sdma base address: 0x%x\n", retval);
270 
271 	return retval;
272 }
273 
274 static inline struct vi_mqd *get_mqd(void *mqd)
275 {
276 	return (struct vi_mqd *)mqd;
277 }
278 
279 static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd)
280 {
281 	return (struct vi_sdma_mqd *)mqd;
282 }
283 
284 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
285 			uint32_t queue_id, uint32_t __user *wptr,
286 			uint32_t wptr_shift, uint32_t wptr_mask,
287 			struct mm_struct *mm)
288 {
289 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
290 	struct vi_mqd *m;
291 	uint32_t *mqd_hqd;
292 	uint32_t reg, wptr_val, data;
293 	bool valid_wptr = false;
294 
295 	m = get_mqd(mqd);
296 
297 	acquire_queue(kgd, pipe_id, queue_id);
298 
299 	/* HIQ is set during driver init period with vmid set to 0*/
300 	if (m->cp_hqd_vmid == 0) {
301 		uint32_t value, mec, pipe;
302 
303 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
304 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
305 
306 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
307 			mec, pipe, queue_id);
308 		value = RREG32(mmRLC_CP_SCHEDULERS);
309 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
310 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
311 		WREG32(mmRLC_CP_SCHEDULERS, value);
312 	}
313 
314 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
315 	mqd_hqd = &m->cp_mqd_base_addr_lo;
316 
317 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_CONTROL; reg++)
318 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
319 
320 	/* Tonga errata: EOP RPTR/WPTR should be left unmodified.
321 	 * This is safe since EOP RPTR==WPTR for any inactive HQD
322 	 * on ASICs that do not support context-save.
323 	 * EOP writes/reads can start anywhere in the ring.
324 	 */
325 	if (get_amdgpu_device(kgd)->asic_type != CHIP_TONGA) {
326 		WREG32(mmCP_HQD_EOP_RPTR, m->cp_hqd_eop_rptr);
327 		WREG32(mmCP_HQD_EOP_WPTR, m->cp_hqd_eop_wptr);
328 		WREG32(mmCP_HQD_EOP_WPTR_MEM, m->cp_hqd_eop_wptr_mem);
329 	}
330 
331 	for (reg = mmCP_HQD_EOP_EVENTS; reg <= mmCP_HQD_ERROR; reg++)
332 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
333 
334 	/* Copy userspace write pointer value to register.
335 	 * Activate doorbell logic to monitor subsequent changes.
336 	 */
337 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
338 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
339 	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
340 
341 	/* read_user_ptr may take the mm->mmap_sem.
342 	 * release srbm_mutex to avoid circular dependency between
343 	 * srbm_mutex->mm_sem->reservation_ww_class_mutex->srbm_mutex.
344 	 */
345 	release_queue(kgd);
346 	valid_wptr = read_user_wptr(mm, wptr, wptr_val);
347 	acquire_queue(kgd, pipe_id, queue_id);
348 	if (valid_wptr)
349 		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
350 
351 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
352 	WREG32(mmCP_HQD_ACTIVE, data);
353 
354 	release_queue(kgd);
355 
356 	return 0;
357 }
358 
359 static int kgd_hqd_dump(struct kgd_dev *kgd,
360 			uint32_t pipe_id, uint32_t queue_id,
361 			uint32_t (**dump)[2], uint32_t *n_regs)
362 {
363 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
364 	uint32_t i = 0, reg;
365 #define HQD_N_REGS (54+4)
366 #define DUMP_REG(addr) do {				\
367 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
368 			break;				\
369 		(*dump)[i][0] = (addr) << 2;		\
370 		(*dump)[i++][1] = RREG32(addr);		\
371 	} while (0)
372 
373 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
374 	if (*dump == NULL)
375 		return -ENOMEM;
376 
377 	acquire_queue(kgd, pipe_id, queue_id);
378 
379 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE0);
380 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE1);
381 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE2);
382 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE3);
383 
384 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_DONES; reg++)
385 		DUMP_REG(reg);
386 
387 	release_queue(kgd);
388 
389 	WARN_ON_ONCE(i != HQD_N_REGS);
390 	*n_regs = i;
391 
392 	return 0;
393 }
394 
395 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
396 			     uint32_t __user *wptr, struct mm_struct *mm)
397 {
398 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
399 	struct vi_sdma_mqd *m;
400 	unsigned long end_jiffies;
401 	uint32_t sdma_base_addr;
402 	uint32_t data;
403 
404 	m = get_sdma_mqd(mqd);
405 	sdma_base_addr = get_sdma_base_addr(m);
406 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
407 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
408 
409 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
410 	while (true) {
411 		data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
412 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
413 			break;
414 		if (time_after(jiffies, end_jiffies))
415 			return -ETIME;
416 		usleep_range(500, 1000);
417 	}
418 	if (m->sdma_engine_id) {
419 		data = RREG32(mmSDMA1_GFX_CONTEXT_CNTL);
420 		data = REG_SET_FIELD(data, SDMA1_GFX_CONTEXT_CNTL,
421 				RESUME_CTX, 0);
422 		WREG32(mmSDMA1_GFX_CONTEXT_CNTL, data);
423 	} else {
424 		data = RREG32(mmSDMA0_GFX_CONTEXT_CNTL);
425 		data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
426 				RESUME_CTX, 0);
427 		WREG32(mmSDMA0_GFX_CONTEXT_CNTL, data);
428 	}
429 
430 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
431 			     ENABLE, 1);
432 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
433 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
434 
435 	if (read_user_wptr(mm, wptr, data))
436 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR, data);
437 	else
438 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
439 		       m->sdmax_rlcx_rb_rptr);
440 
441 	WREG32(sdma_base_addr + mmSDMA0_RLC0_VIRTUAL_ADDR,
442 				m->sdmax_rlcx_virtual_addr);
443 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
444 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
445 			m->sdmax_rlcx_rb_base_hi);
446 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
447 			m->sdmax_rlcx_rb_rptr_addr_lo);
448 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
449 			m->sdmax_rlcx_rb_rptr_addr_hi);
450 
451 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
452 			     RB_ENABLE, 1);
453 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
454 
455 	return 0;
456 }
457 
458 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
459 			     uint32_t engine_id, uint32_t queue_id,
460 			     uint32_t (**dump)[2], uint32_t *n_regs)
461 {
462 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
463 	uint32_t sdma_offset = engine_id * SDMA1_REGISTER_OFFSET +
464 		queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
465 	uint32_t i = 0, reg;
466 #undef HQD_N_REGS
467 #define HQD_N_REGS (19+4+2+3+7)
468 
469 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
470 	if (*dump == NULL)
471 		return -ENOMEM;
472 
473 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
474 		DUMP_REG(sdma_offset + reg);
475 	for (reg = mmSDMA0_RLC0_VIRTUAL_ADDR; reg <= mmSDMA0_RLC0_WATERMARK;
476 	     reg++)
477 		DUMP_REG(sdma_offset + reg);
478 	for (reg = mmSDMA0_RLC0_CSA_ADDR_LO; reg <= mmSDMA0_RLC0_CSA_ADDR_HI;
479 	     reg++)
480 		DUMP_REG(sdma_offset + reg);
481 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN; reg <= mmSDMA0_RLC0_DUMMY_REG;
482 	     reg++)
483 		DUMP_REG(sdma_offset + reg);
484 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0; reg <= mmSDMA0_RLC0_MIDCMD_CNTL;
485 	     reg++)
486 		DUMP_REG(sdma_offset + reg);
487 
488 	WARN_ON_ONCE(i != HQD_N_REGS);
489 	*n_regs = i;
490 
491 	return 0;
492 }
493 
494 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
495 				uint32_t pipe_id, uint32_t queue_id)
496 {
497 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
498 	uint32_t act;
499 	bool retval = false;
500 	uint32_t low, high;
501 
502 	acquire_queue(kgd, pipe_id, queue_id);
503 	act = RREG32(mmCP_HQD_ACTIVE);
504 	if (act) {
505 		low = lower_32_bits(queue_address >> 8);
506 		high = upper_32_bits(queue_address >> 8);
507 
508 		if (low == RREG32(mmCP_HQD_PQ_BASE) &&
509 				high == RREG32(mmCP_HQD_PQ_BASE_HI))
510 			retval = true;
511 	}
512 	release_queue(kgd);
513 	return retval;
514 }
515 
516 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
517 {
518 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
519 	struct vi_sdma_mqd *m;
520 	uint32_t sdma_base_addr;
521 	uint32_t sdma_rlc_rb_cntl;
522 
523 	m = get_sdma_mqd(mqd);
524 	sdma_base_addr = get_sdma_base_addr(m);
525 
526 	sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
527 
528 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
529 		return true;
530 
531 	return false;
532 }
533 
534 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
535 				enum kfd_preempt_type reset_type,
536 				unsigned int utimeout, uint32_t pipe_id,
537 				uint32_t queue_id)
538 {
539 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
540 	uint32_t temp;
541 	enum hqd_dequeue_request_type type;
542 	unsigned long flags, end_jiffies;
543 	int retry;
544 	struct vi_mqd *m = get_mqd(mqd);
545 
546 	if (adev->in_gpu_reset)
547 		return -EIO;
548 
549 	acquire_queue(kgd, pipe_id, queue_id);
550 
551 	if (m->cp_hqd_vmid == 0)
552 		WREG32_FIELD(RLC_CP_SCHEDULERS, scheduler1, 0);
553 
554 	switch (reset_type) {
555 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
556 		type = DRAIN_PIPE;
557 		break;
558 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
559 		type = RESET_WAVES;
560 		break;
561 	default:
562 		type = DRAIN_PIPE;
563 		break;
564 	}
565 
566 	/* Workaround: If IQ timer is active and the wait time is close to or
567 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
568 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
569 	 * cleared before continuing. Also, ensure wait times are set to at
570 	 * least 0x3.
571 	 */
572 	local_irq_save(flags);
573 	preempt_disable();
574 	retry = 5000; /* wait for 500 usecs at maximum */
575 	while (true) {
576 		temp = RREG32(mmCP_HQD_IQ_TIMER);
577 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
578 			pr_debug("HW is processing IQ\n");
579 			goto loop;
580 		}
581 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
582 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
583 					== 3) /* SEM-rearm is safe */
584 				break;
585 			/* Wait time 3 is safe for CP, but our MMIO read/write
586 			 * time is close to 1 microsecond, so check for 10 to
587 			 * leave more buffer room
588 			 */
589 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
590 					>= 10)
591 				break;
592 			pr_debug("IQ timer is active\n");
593 		} else
594 			break;
595 loop:
596 		if (!retry) {
597 			pr_err("CP HQD IQ timer status time out\n");
598 			break;
599 		}
600 		ndelay(100);
601 		--retry;
602 	}
603 	retry = 1000;
604 	while (true) {
605 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
606 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
607 			break;
608 		pr_debug("Dequeue request is pending\n");
609 
610 		if (!retry) {
611 			pr_err("CP HQD dequeue request time out\n");
612 			break;
613 		}
614 		ndelay(100);
615 		--retry;
616 	}
617 	local_irq_restore(flags);
618 	preempt_enable();
619 
620 	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
621 
622 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
623 	while (true) {
624 		temp = RREG32(mmCP_HQD_ACTIVE);
625 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
626 			break;
627 		if (time_after(jiffies, end_jiffies)) {
628 			pr_err("cp queue preemption time out.\n");
629 			release_queue(kgd);
630 			return -ETIME;
631 		}
632 		usleep_range(500, 1000);
633 	}
634 
635 	release_queue(kgd);
636 	return 0;
637 }
638 
639 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
640 				unsigned int utimeout)
641 {
642 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
643 	struct vi_sdma_mqd *m;
644 	uint32_t sdma_base_addr;
645 	uint32_t temp;
646 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
647 
648 	m = get_sdma_mqd(mqd);
649 	sdma_base_addr = get_sdma_base_addr(m);
650 
651 	temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
652 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
653 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
654 
655 	while (true) {
656 		temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
657 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
658 			break;
659 		if (time_after(jiffies, end_jiffies))
660 			return -ETIME;
661 		usleep_range(500, 1000);
662 	}
663 
664 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
665 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
666 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
667 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
668 
669 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
670 
671 	return 0;
672 }
673 
674 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
675 							uint8_t vmid)
676 {
677 	uint32_t reg;
678 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
679 
680 	reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
681 	return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
682 }
683 
684 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
685 								uint8_t vmid)
686 {
687 	uint32_t reg;
688 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
689 
690 	reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
691 	return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
692 }
693 
694 static int kgd_address_watch_disable(struct kgd_dev *kgd)
695 {
696 	return 0;
697 }
698 
699 static int kgd_address_watch_execute(struct kgd_dev *kgd,
700 					unsigned int watch_point_id,
701 					uint32_t cntl_val,
702 					uint32_t addr_hi,
703 					uint32_t addr_lo)
704 {
705 	return 0;
706 }
707 
708 static int kgd_wave_control_execute(struct kgd_dev *kgd,
709 					uint32_t gfx_index_val,
710 					uint32_t sq_cmd)
711 {
712 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
713 	uint32_t data = 0;
714 
715 	mutex_lock(&adev->grbm_idx_mutex);
716 
717 	WREG32(mmGRBM_GFX_INDEX, gfx_index_val);
718 	WREG32(mmSQ_CMD, sq_cmd);
719 
720 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
721 		INSTANCE_BROADCAST_WRITES, 1);
722 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
723 		SH_BROADCAST_WRITES, 1);
724 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
725 		SE_BROADCAST_WRITES, 1);
726 
727 	WREG32(mmGRBM_GFX_INDEX, data);
728 	mutex_unlock(&adev->grbm_idx_mutex);
729 
730 	return 0;
731 }
732 
733 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
734 					unsigned int watch_point_id,
735 					unsigned int reg_offset)
736 {
737 	return 0;
738 }
739 
740 static void set_scratch_backing_va(struct kgd_dev *kgd,
741 					uint64_t va, uint32_t vmid)
742 {
743 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
744 
745 	lock_srbm(kgd, 0, 0, 0, vmid);
746 	WREG32(mmSH_HIDDEN_PRIVATE_BASE_VMID, va);
747 	unlock_srbm(kgd);
748 }
749 
750 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
751 		uint64_t page_table_base)
752 {
753 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
754 
755 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
756 		pr_err("trying to set page table base for wrong VMID\n");
757 		return;
758 	}
759 	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8,
760 			lower_32_bits(page_table_base));
761 }
762 
763 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
764 {
765 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
766 	int vmid;
767 	unsigned int tmp;
768 
769 	if (adev->in_gpu_reset)
770 		return -EIO;
771 
772 	for (vmid = 0; vmid < 16; vmid++) {
773 		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
774 			continue;
775 
776 		tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
777 		if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
778 			(tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
779 			WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
780 			RREG32(mmVM_INVALIDATE_RESPONSE);
781 			break;
782 		}
783 	}
784 
785 	return 0;
786 }
787 
788 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
789 {
790 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
791 
792 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
793 		pr_err("non kfd vmid %d\n", vmid);
794 		return -EINVAL;
795 	}
796 
797 	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
798 	RREG32(mmVM_INVALIDATE_RESPONSE);
799 	return 0;
800 }
801