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/fdtable.h>
24 #include <linux/uaccess.h>
25 #include <linux/mmu_context.h>
26 #include <drm/drmP.h>
27 #include "amdgpu.h"
28 #include "amdgpu_amdkfd.h"
29 #include "cikd.h"
30 #include "cik_sdma.h"
31 #include "gfx_v7_0.h"
32 #include "gca/gfx_7_2_d.h"
33 #include "gca/gfx_7_2_enum.h"
34 #include "gca/gfx_7_2_sh_mask.h"
35 #include "oss/oss_2_0_d.h"
36 #include "oss/oss_2_0_sh_mask.h"
37 #include "gmc/gmc_7_1_d.h"
38 #include "gmc/gmc_7_1_sh_mask.h"
39 #include "cik_structs.h"
40 
41 enum hqd_dequeue_request_type {
42 	NO_ACTION = 0,
43 	DRAIN_PIPE,
44 	RESET_WAVES
45 };
46 
47 enum {
48 	MAX_TRAPID = 8,		/* 3 bits in the bitfield. */
49 	MAX_WATCH_ADDRESSES = 4
50 };
51 
52 enum {
53 	ADDRESS_WATCH_REG_ADDR_HI = 0,
54 	ADDRESS_WATCH_REG_ADDR_LO,
55 	ADDRESS_WATCH_REG_CNTL,
56 	ADDRESS_WATCH_REG_MAX
57 };
58 
59 /*  not defined in the CI/KV reg file  */
60 enum {
61 	ADDRESS_WATCH_REG_CNTL_ATC_BIT = 0x10000000UL,
62 	ADDRESS_WATCH_REG_CNTL_DEFAULT_MASK = 0x00FFFFFF,
63 	ADDRESS_WATCH_REG_ADDLOW_MASK_EXTENSION = 0x03000000,
64 	/* extend the mask to 26 bits to match the low address field */
65 	ADDRESS_WATCH_REG_ADDLOW_SHIFT = 6,
66 	ADDRESS_WATCH_REG_ADDHIGH_MASK = 0xFFFF
67 };
68 
69 static const uint32_t watchRegs[MAX_WATCH_ADDRESSES * ADDRESS_WATCH_REG_MAX] = {
70 	mmTCP_WATCH0_ADDR_H, mmTCP_WATCH0_ADDR_L, mmTCP_WATCH0_CNTL,
71 	mmTCP_WATCH1_ADDR_H, mmTCP_WATCH1_ADDR_L, mmTCP_WATCH1_CNTL,
72 	mmTCP_WATCH2_ADDR_H, mmTCP_WATCH2_ADDR_L, mmTCP_WATCH2_CNTL,
73 	mmTCP_WATCH3_ADDR_H, mmTCP_WATCH3_ADDR_L, mmTCP_WATCH3_CNTL
74 };
75 
76 union TCP_WATCH_CNTL_BITS {
77 	struct {
78 		uint32_t mask:24;
79 		uint32_t vmid:4;
80 		uint32_t atc:1;
81 		uint32_t mode:2;
82 		uint32_t valid:1;
83 	} bitfields, bits;
84 	uint32_t u32All;
85 	signed int i32All;
86 	float f32All;
87 };
88 
89 /*
90  * Register access functions
91  */
92 
93 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
94 		uint32_t sh_mem_config,	uint32_t sh_mem_ape1_base,
95 		uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases);
96 
97 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
98 					unsigned int vmid);
99 
100 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
101 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
102 			uint32_t queue_id, uint32_t __user *wptr,
103 			uint32_t wptr_shift, uint32_t wptr_mask,
104 			struct mm_struct *mm);
105 static int kgd_hqd_dump(struct kgd_dev *kgd,
106 			uint32_t pipe_id, uint32_t queue_id,
107 			uint32_t (**dump)[2], uint32_t *n_regs);
108 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
109 			     uint32_t __user *wptr, struct mm_struct *mm);
110 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
111 			     uint32_t engine_id, uint32_t queue_id,
112 			     uint32_t (**dump)[2], uint32_t *n_regs);
113 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
114 				uint32_t pipe_id, uint32_t queue_id);
115 
116 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
117 				enum kfd_preempt_type reset_type,
118 				unsigned int utimeout, uint32_t pipe_id,
119 				uint32_t queue_id);
120 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
121 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
122 				unsigned int utimeout);
123 static int kgd_address_watch_disable(struct kgd_dev *kgd);
124 static int kgd_address_watch_execute(struct kgd_dev *kgd,
125 					unsigned int watch_point_id,
126 					uint32_t cntl_val,
127 					uint32_t addr_hi,
128 					uint32_t addr_lo);
129 static int kgd_wave_control_execute(struct kgd_dev *kgd,
130 					uint32_t gfx_index_val,
131 					uint32_t sq_cmd);
132 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
133 					unsigned int watch_point_id,
134 					unsigned int reg_offset);
135 
136 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, uint8_t vmid);
137 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
138 							uint8_t vmid);
139 
140 static void set_scratch_backing_va(struct kgd_dev *kgd,
141 					uint64_t va, uint32_t vmid);
142 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
143 		uint64_t page_table_base);
144 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
145 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
146 static uint32_t read_vmid_from_vmfault_reg(struct kgd_dev *kgd);
147 
148 /* Because of REG_GET_FIELD() being used, we put this function in the
149  * asic specific file.
150  */
151 static int get_tile_config(struct kgd_dev *kgd,
152 		struct tile_config *config)
153 {
154 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
155 
156 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
157 	config->num_banks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
158 				MC_ARB_RAMCFG, NOOFBANK);
159 	config->num_ranks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
160 				MC_ARB_RAMCFG, NOOFRANKS);
161 
162 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
163 	config->num_tile_configs =
164 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
165 	config->macro_tile_config_ptr =
166 			adev->gfx.config.macrotile_mode_array;
167 	config->num_macro_tile_configs =
168 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
169 
170 	return 0;
171 }
172 
173 static const struct kfd2kgd_calls kfd2kgd = {
174 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
175 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
176 	.init_interrupts = kgd_init_interrupts,
177 	.hqd_load = kgd_hqd_load,
178 	.hqd_sdma_load = kgd_hqd_sdma_load,
179 	.hqd_dump = kgd_hqd_dump,
180 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
181 	.hqd_is_occupied = kgd_hqd_is_occupied,
182 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
183 	.hqd_destroy = kgd_hqd_destroy,
184 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
185 	.address_watch_disable = kgd_address_watch_disable,
186 	.address_watch_execute = kgd_address_watch_execute,
187 	.wave_control_execute = kgd_wave_control_execute,
188 	.address_watch_get_offset = kgd_address_watch_get_offset,
189 	.get_atc_vmid_pasid_mapping_pasid = get_atc_vmid_pasid_mapping_pasid,
190 	.get_atc_vmid_pasid_mapping_valid = get_atc_vmid_pasid_mapping_valid,
191 	.set_scratch_backing_va = set_scratch_backing_va,
192 	.get_tile_config = get_tile_config,
193 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
194 	.invalidate_tlbs = invalidate_tlbs,
195 	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
196 	.read_vmid_from_vmfault_reg = read_vmid_from_vmfault_reg,
197 };
198 
199 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_7_get_functions(void)
200 {
201 	return (struct kfd2kgd_calls *)&kfd2kgd;
202 }
203 
204 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
205 {
206 	return (struct amdgpu_device *)kgd;
207 }
208 
209 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
210 			uint32_t queue, uint32_t vmid)
211 {
212 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
213 	uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
214 
215 	mutex_lock(&adev->srbm_mutex);
216 	WREG32(mmSRBM_GFX_CNTL, value);
217 }
218 
219 static void unlock_srbm(struct kgd_dev *kgd)
220 {
221 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
222 
223 	WREG32(mmSRBM_GFX_CNTL, 0);
224 	mutex_unlock(&adev->srbm_mutex);
225 }
226 
227 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
228 				uint32_t queue_id)
229 {
230 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
231 
232 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
233 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
234 
235 	lock_srbm(kgd, mec, pipe, queue_id, 0);
236 }
237 
238 static void release_queue(struct kgd_dev *kgd)
239 {
240 	unlock_srbm(kgd);
241 }
242 
243 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
244 					uint32_t sh_mem_config,
245 					uint32_t sh_mem_ape1_base,
246 					uint32_t sh_mem_ape1_limit,
247 					uint32_t sh_mem_bases)
248 {
249 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
250 
251 	lock_srbm(kgd, 0, 0, 0, vmid);
252 
253 	WREG32(mmSH_MEM_CONFIG, sh_mem_config);
254 	WREG32(mmSH_MEM_APE1_BASE, sh_mem_ape1_base);
255 	WREG32(mmSH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
256 	WREG32(mmSH_MEM_BASES, sh_mem_bases);
257 
258 	unlock_srbm(kgd);
259 }
260 
261 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
262 					unsigned int vmid)
263 {
264 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
265 
266 	/*
267 	 * We have to assume that there is no outstanding mapping.
268 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
269 	 * a mapping is in progress or because a mapping finished and the
270 	 * SW cleared it. So the protocol is to always wait & clear.
271 	 */
272 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
273 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
274 
275 	WREG32(mmATC_VMID0_PASID_MAPPING + vmid, pasid_mapping);
276 
277 	while (!(RREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))
278 		cpu_relax();
279 	WREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
280 
281 	/* Mapping vmid to pasid also for IH block */
282 	WREG32(mmIH_VMID_0_LUT + vmid, pasid_mapping);
283 
284 	return 0;
285 }
286 
287 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
288 {
289 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
290 	uint32_t mec;
291 	uint32_t pipe;
292 
293 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
294 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
295 
296 	lock_srbm(kgd, mec, pipe, 0, 0);
297 
298 	WREG32(mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
299 			CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
300 
301 	unlock_srbm(kgd);
302 
303 	return 0;
304 }
305 
306 static inline uint32_t get_sdma_base_addr(struct cik_sdma_rlc_registers *m)
307 {
308 	uint32_t retval;
309 
310 	retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
311 			m->sdma_queue_id * KFD_CIK_SDMA_QUEUE_OFFSET;
312 
313 	pr_debug("kfd: sdma base address: 0x%x\n", retval);
314 
315 	return retval;
316 }
317 
318 static inline struct cik_mqd *get_mqd(void *mqd)
319 {
320 	return (struct cik_mqd *)mqd;
321 }
322 
323 static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd)
324 {
325 	return (struct cik_sdma_rlc_registers *)mqd;
326 }
327 
328 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
329 			uint32_t queue_id, uint32_t __user *wptr,
330 			uint32_t wptr_shift, uint32_t wptr_mask,
331 			struct mm_struct *mm)
332 {
333 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
334 	struct cik_mqd *m;
335 	uint32_t *mqd_hqd;
336 	uint32_t reg, wptr_val, data;
337 	bool valid_wptr = false;
338 
339 	m = get_mqd(mqd);
340 
341 	acquire_queue(kgd, pipe_id, queue_id);
342 
343 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_MQD_CONTROL. */
344 	mqd_hqd = &m->cp_mqd_base_addr_lo;
345 
346 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_MQD_CONTROL; reg++)
347 		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
348 
349 	/* Copy userspace write pointer value to register.
350 	 * Activate doorbell logic to monitor subsequent changes.
351 	 */
352 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
353 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
354 	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
355 
356 	/* read_user_ptr may take the mm->mmap_sem.
357 	 * release srbm_mutex to avoid circular dependency between
358 	 * srbm_mutex->mm_sem->reservation_ww_class_mutex->srbm_mutex.
359 	 */
360 	release_queue(kgd);
361 	valid_wptr = read_user_wptr(mm, wptr, wptr_val);
362 	acquire_queue(kgd, pipe_id, queue_id);
363 	if (valid_wptr)
364 		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
365 
366 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
367 	WREG32(mmCP_HQD_ACTIVE, data);
368 
369 	release_queue(kgd);
370 
371 	return 0;
372 }
373 
374 static int kgd_hqd_dump(struct kgd_dev *kgd,
375 			uint32_t pipe_id, uint32_t queue_id,
376 			uint32_t (**dump)[2], uint32_t *n_regs)
377 {
378 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
379 	uint32_t i = 0, reg;
380 #define HQD_N_REGS (35+4)
381 #define DUMP_REG(addr) do {				\
382 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
383 			break;				\
384 		(*dump)[i][0] = (addr) << 2;		\
385 		(*dump)[i++][1] = RREG32(addr);		\
386 	} while (0)
387 
388 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
389 	if (*dump == NULL)
390 		return -ENOMEM;
391 
392 	acquire_queue(kgd, pipe_id, queue_id);
393 
394 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE0);
395 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE1);
396 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE2);
397 	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE3);
398 
399 	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_MQD_CONTROL; reg++)
400 		DUMP_REG(reg);
401 
402 	release_queue(kgd);
403 
404 	WARN_ON_ONCE(i != HQD_N_REGS);
405 	*n_regs = i;
406 
407 	return 0;
408 }
409 
410 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
411 			     uint32_t __user *wptr, struct mm_struct *mm)
412 {
413 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
414 	struct cik_sdma_rlc_registers *m;
415 	unsigned long end_jiffies;
416 	uint32_t sdma_base_addr;
417 	uint32_t data;
418 
419 	m = get_sdma_mqd(mqd);
420 	sdma_base_addr = get_sdma_base_addr(m);
421 
422 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
423 		m->sdma_rlc_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
424 
425 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
426 	while (true) {
427 		data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
428 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
429 			break;
430 		if (time_after(jiffies, end_jiffies))
431 			return -ETIME;
432 		usleep_range(500, 1000);
433 	}
434 	if (m->sdma_engine_id) {
435 		data = RREG32(mmSDMA1_GFX_CONTEXT_CNTL);
436 		data = REG_SET_FIELD(data, SDMA1_GFX_CONTEXT_CNTL,
437 				RESUME_CTX, 0);
438 		WREG32(mmSDMA1_GFX_CONTEXT_CNTL, data);
439 	} else {
440 		data = RREG32(mmSDMA0_GFX_CONTEXT_CNTL);
441 		data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
442 				RESUME_CTX, 0);
443 		WREG32(mmSDMA0_GFX_CONTEXT_CNTL, data);
444 	}
445 
446 	data = REG_SET_FIELD(m->sdma_rlc_doorbell, SDMA0_RLC0_DOORBELL,
447 			     ENABLE, 1);
448 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
449 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdma_rlc_rb_rptr);
450 
451 	if (read_user_wptr(mm, wptr, data))
452 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR, data);
453 	else
454 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
455 		       m->sdma_rlc_rb_rptr);
456 
457 	WREG32(sdma_base_addr + mmSDMA0_RLC0_VIRTUAL_ADDR,
458 				m->sdma_rlc_virtual_addr);
459 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdma_rlc_rb_base);
460 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
461 			m->sdma_rlc_rb_base_hi);
462 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
463 			m->sdma_rlc_rb_rptr_addr_lo);
464 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
465 			m->sdma_rlc_rb_rptr_addr_hi);
466 
467 	data = REG_SET_FIELD(m->sdma_rlc_rb_cntl, SDMA0_RLC0_RB_CNTL,
468 			     RB_ENABLE, 1);
469 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
470 
471 	return 0;
472 }
473 
474 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
475 			     uint32_t engine_id, uint32_t queue_id,
476 			     uint32_t (**dump)[2], uint32_t *n_regs)
477 {
478 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
479 	uint32_t sdma_offset = engine_id * SDMA1_REGISTER_OFFSET +
480 		queue_id * KFD_CIK_SDMA_QUEUE_OFFSET;
481 	uint32_t i = 0, reg;
482 #undef HQD_N_REGS
483 #define HQD_N_REGS (19+4)
484 
485 	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
486 	if (*dump == NULL)
487 		return -ENOMEM;
488 
489 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
490 		DUMP_REG(sdma_offset + reg);
491 	for (reg = mmSDMA0_RLC0_VIRTUAL_ADDR; reg <= mmSDMA0_RLC0_WATERMARK;
492 	     reg++)
493 		DUMP_REG(sdma_offset + reg);
494 
495 	WARN_ON_ONCE(i != HQD_N_REGS);
496 	*n_regs = i;
497 
498 	return 0;
499 }
500 
501 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
502 				uint32_t pipe_id, uint32_t queue_id)
503 {
504 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
505 	uint32_t act;
506 	bool retval = false;
507 	uint32_t low, high;
508 
509 	acquire_queue(kgd, pipe_id, queue_id);
510 	act = RREG32(mmCP_HQD_ACTIVE);
511 	if (act) {
512 		low = lower_32_bits(queue_address >> 8);
513 		high = upper_32_bits(queue_address >> 8);
514 
515 		if (low == RREG32(mmCP_HQD_PQ_BASE) &&
516 				high == RREG32(mmCP_HQD_PQ_BASE_HI))
517 			retval = true;
518 	}
519 	release_queue(kgd);
520 	return retval;
521 }
522 
523 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
524 {
525 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
526 	struct cik_sdma_rlc_registers *m;
527 	uint32_t sdma_base_addr;
528 	uint32_t sdma_rlc_rb_cntl;
529 
530 	m = get_sdma_mqd(mqd);
531 	sdma_base_addr = get_sdma_base_addr(m);
532 
533 	sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
534 
535 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
536 		return true;
537 
538 	return false;
539 }
540 
541 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
542 				enum kfd_preempt_type reset_type,
543 				unsigned int utimeout, uint32_t pipe_id,
544 				uint32_t queue_id)
545 {
546 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
547 	uint32_t temp;
548 	enum hqd_dequeue_request_type type;
549 	unsigned long flags, end_jiffies;
550 	int retry;
551 
552 	if (adev->in_gpu_reset)
553 		return -EIO;
554 
555 	acquire_queue(kgd, pipe_id, queue_id);
556 	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, 0);
557 
558 	switch (reset_type) {
559 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
560 		type = DRAIN_PIPE;
561 		break;
562 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
563 		type = RESET_WAVES;
564 		break;
565 	default:
566 		type = DRAIN_PIPE;
567 		break;
568 	}
569 
570 	/* Workaround: If IQ timer is active and the wait time is close to or
571 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
572 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
573 	 * cleared before continuing. Also, ensure wait times are set to at
574 	 * least 0x3.
575 	 */
576 	local_irq_save(flags);
577 	preempt_disable();
578 	retry = 5000; /* wait for 500 usecs at maximum */
579 	while (true) {
580 		temp = RREG32(mmCP_HQD_IQ_TIMER);
581 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
582 			pr_debug("HW is processing IQ\n");
583 			goto loop;
584 		}
585 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
586 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
587 					== 3) /* SEM-rearm is safe */
588 				break;
589 			/* Wait time 3 is safe for CP, but our MMIO read/write
590 			 * time is close to 1 microsecond, so check for 10 to
591 			 * leave more buffer room
592 			 */
593 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
594 					>= 10)
595 				break;
596 			pr_debug("IQ timer is active\n");
597 		} else
598 			break;
599 loop:
600 		if (!retry) {
601 			pr_err("CP HQD IQ timer status time out\n");
602 			break;
603 		}
604 		ndelay(100);
605 		--retry;
606 	}
607 	retry = 1000;
608 	while (true) {
609 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
610 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
611 			break;
612 		pr_debug("Dequeue request is pending\n");
613 
614 		if (!retry) {
615 			pr_err("CP HQD dequeue request time out\n");
616 			break;
617 		}
618 		ndelay(100);
619 		--retry;
620 	}
621 	local_irq_restore(flags);
622 	preempt_enable();
623 
624 	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
625 
626 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
627 	while (true) {
628 		temp = RREG32(mmCP_HQD_ACTIVE);
629 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
630 			break;
631 		if (time_after(jiffies, end_jiffies)) {
632 			pr_err("cp queue preemption time out\n");
633 			release_queue(kgd);
634 			return -ETIME;
635 		}
636 		usleep_range(500, 1000);
637 	}
638 
639 	release_queue(kgd);
640 	return 0;
641 }
642 
643 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
644 				unsigned int utimeout)
645 {
646 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
647 	struct cik_sdma_rlc_registers *m;
648 	uint32_t sdma_base_addr;
649 	uint32_t temp;
650 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
651 
652 	m = get_sdma_mqd(mqd);
653 	sdma_base_addr = get_sdma_base_addr(m);
654 
655 	temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
656 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
657 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
658 
659 	while (true) {
660 		temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
661 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
662 			break;
663 		if (time_after(jiffies, end_jiffies))
664 			return -ETIME;
665 		usleep_range(500, 1000);
666 	}
667 
668 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
669 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
670 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
671 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
672 
673 	m->sdma_rlc_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
674 
675 	return 0;
676 }
677 
678 static int kgd_address_watch_disable(struct kgd_dev *kgd)
679 {
680 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
681 	union TCP_WATCH_CNTL_BITS cntl;
682 	unsigned int i;
683 
684 	cntl.u32All = 0;
685 
686 	cntl.bitfields.valid = 0;
687 	cntl.bitfields.mask = ADDRESS_WATCH_REG_CNTL_DEFAULT_MASK;
688 	cntl.bitfields.atc = 1;
689 
690 	/* Turning off this address until we set all the registers */
691 	for (i = 0; i < MAX_WATCH_ADDRESSES; i++)
692 		WREG32(watchRegs[i * ADDRESS_WATCH_REG_MAX +
693 			ADDRESS_WATCH_REG_CNTL], cntl.u32All);
694 
695 	return 0;
696 }
697 
698 static int kgd_address_watch_execute(struct kgd_dev *kgd,
699 					unsigned int watch_point_id,
700 					uint32_t cntl_val,
701 					uint32_t addr_hi,
702 					uint32_t addr_lo)
703 {
704 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
705 	union TCP_WATCH_CNTL_BITS cntl;
706 
707 	cntl.u32All = cntl_val;
708 
709 	/* Turning off this watch point until we set all the registers */
710 	cntl.bitfields.valid = 0;
711 	WREG32(watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
712 		ADDRESS_WATCH_REG_CNTL], cntl.u32All);
713 
714 	WREG32(watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
715 		ADDRESS_WATCH_REG_ADDR_HI], addr_hi);
716 
717 	WREG32(watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
718 		ADDRESS_WATCH_REG_ADDR_LO], addr_lo);
719 
720 	/* Enable the watch point */
721 	cntl.bitfields.valid = 1;
722 
723 	WREG32(watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
724 		ADDRESS_WATCH_REG_CNTL], cntl.u32All);
725 
726 	return 0;
727 }
728 
729 static int kgd_wave_control_execute(struct kgd_dev *kgd,
730 					uint32_t gfx_index_val,
731 					uint32_t sq_cmd)
732 {
733 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
734 	uint32_t data;
735 
736 	mutex_lock(&adev->grbm_idx_mutex);
737 
738 	WREG32(mmGRBM_GFX_INDEX, gfx_index_val);
739 	WREG32(mmSQ_CMD, sq_cmd);
740 
741 	/*  Restore the GRBM_GFX_INDEX register  */
742 
743 	data = GRBM_GFX_INDEX__INSTANCE_BROADCAST_WRITES_MASK |
744 		GRBM_GFX_INDEX__SH_BROADCAST_WRITES_MASK |
745 		GRBM_GFX_INDEX__SE_BROADCAST_WRITES_MASK;
746 
747 	WREG32(mmGRBM_GFX_INDEX, data);
748 
749 	mutex_unlock(&adev->grbm_idx_mutex);
750 
751 	return 0;
752 }
753 
754 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
755 					unsigned int watch_point_id,
756 					unsigned int reg_offset)
757 {
758 	return watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX + reg_offset];
759 }
760 
761 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
762 							uint8_t vmid)
763 {
764 	uint32_t reg;
765 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
766 
767 	reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
768 	return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
769 }
770 
771 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
772 								uint8_t vmid)
773 {
774 	uint32_t reg;
775 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
776 
777 	reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
778 	return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
779 }
780 
781 static void set_scratch_backing_va(struct kgd_dev *kgd,
782 					uint64_t va, uint32_t vmid)
783 {
784 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
785 
786 	lock_srbm(kgd, 0, 0, 0, vmid);
787 	WREG32(mmSH_HIDDEN_PRIVATE_BASE_VMID, va);
788 	unlock_srbm(kgd);
789 }
790 
791 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
792 			uint64_t page_table_base)
793 {
794 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
795 
796 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
797 		pr_err("trying to set page table base for wrong VMID\n");
798 		return;
799 	}
800 	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8,
801 		lower_32_bits(page_table_base));
802 }
803 
804 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
805 {
806 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
807 	int vmid;
808 	unsigned int tmp;
809 
810 	if (adev->in_gpu_reset)
811 		return -EIO;
812 
813 	for (vmid = 0; vmid < 16; vmid++) {
814 		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
815 			continue;
816 
817 		tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
818 		if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
819 			(tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
820 			WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
821 			RREG32(mmVM_INVALIDATE_RESPONSE);
822 			break;
823 		}
824 	}
825 
826 	return 0;
827 }
828 
829 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
830 {
831 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
832 
833 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
834 		pr_err("non kfd vmid\n");
835 		return 0;
836 	}
837 
838 	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
839 	RREG32(mmVM_INVALIDATE_RESPONSE);
840 	return 0;
841 }
842 
843  /**
844   * read_vmid_from_vmfault_reg - read vmid from register
845   *
846   * adev: amdgpu_device pointer
847   * @vmid: vmid pointer
848   * read vmid from register (CIK).
849   */
850 static uint32_t read_vmid_from_vmfault_reg(struct kgd_dev *kgd)
851 {
852 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
853 
854 	uint32_t status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS);
855 
856 	return REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID);
857 }
858