• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2023 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 
24 #include <linux/firmware.h>
25 #include <linux/module.h>
26 #include "amdgpu.h"
27 #include "soc15_common.h"
28 #include "soc21.h"
29 #include "gc/gc_12_0_0_offset.h"
30 #include "gc/gc_12_0_0_sh_mask.h"
31 #include "gc/gc_11_0_0_default.h"
32 #include "v12_structs.h"
33 #include "mes_v12_api_def.h"
34 
35 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mes.bin");
36 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mes1.bin");
37 MODULE_FIRMWARE("amdgpu/gc_12_0_0_uni_mes.bin");
38 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mes.bin");
39 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mes1.bin");
40 MODULE_FIRMWARE("amdgpu/gc_12_0_1_uni_mes.bin");
41 
42 static int mes_v12_0_hw_init(void *handle);
43 static int mes_v12_0_hw_fini(void *handle);
44 static int mes_v12_0_kiq_hw_init(struct amdgpu_device *adev);
45 static int mes_v12_0_kiq_hw_fini(struct amdgpu_device *adev);
46 
47 #define MES_EOP_SIZE   2048
48 
mes_v12_0_ring_set_wptr(struct amdgpu_ring * ring)49 static void mes_v12_0_ring_set_wptr(struct amdgpu_ring *ring)
50 {
51 	struct amdgpu_device *adev = ring->adev;
52 
53 	if (ring->use_doorbell) {
54 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
55 			     ring->wptr);
56 		WDOORBELL64(ring->doorbell_index, ring->wptr);
57 	} else {
58 		BUG();
59 	}
60 }
61 
mes_v12_0_ring_get_rptr(struct amdgpu_ring * ring)62 static u64 mes_v12_0_ring_get_rptr(struct amdgpu_ring *ring)
63 {
64 	return *ring->rptr_cpu_addr;
65 }
66 
mes_v12_0_ring_get_wptr(struct amdgpu_ring * ring)67 static u64 mes_v12_0_ring_get_wptr(struct amdgpu_ring *ring)
68 {
69 	u64 wptr;
70 
71 	if (ring->use_doorbell)
72 		wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
73 	else
74 		BUG();
75 	return wptr;
76 }
77 
78 static const struct amdgpu_ring_funcs mes_v12_0_ring_funcs = {
79 	.type = AMDGPU_RING_TYPE_MES,
80 	.align_mask = 1,
81 	.nop = 0,
82 	.support_64bit_ptrs = true,
83 	.get_rptr = mes_v12_0_ring_get_rptr,
84 	.get_wptr = mes_v12_0_ring_get_wptr,
85 	.set_wptr = mes_v12_0_ring_set_wptr,
86 	.insert_nop = amdgpu_ring_insert_nop,
87 };
88 
89 static const char *mes_v12_0_opcodes[] = {
90 	"SET_HW_RSRC",
91 	"SET_SCHEDULING_CONFIG",
92 	"ADD_QUEUE",
93 	"REMOVE_QUEUE",
94 	"PERFORM_YIELD",
95 	"SET_GANG_PRIORITY_LEVEL",
96 	"SUSPEND",
97 	"RESUME",
98 	"RESET",
99 	"SET_LOG_BUFFER",
100 	"CHANGE_GANG_PRORITY",
101 	"QUERY_SCHEDULER_STATUS",
102 	"unused",
103 	"SET_DEBUG_VMID",
104 	"MISC",
105 	"UPDATE_ROOT_PAGE_TABLE",
106 	"AMD_LOG",
107 	"SET_SE_MODE",
108 	"SET_GANG_SUBMIT",
109 	"SET_HW_RSRC_1",
110 };
111 
112 static const char *mes_v12_0_misc_opcodes[] = {
113 	"WRITE_REG",
114 	"INV_GART",
115 	"QUERY_STATUS",
116 	"READ_REG",
117 	"WAIT_REG_MEM",
118 	"SET_SHADER_DEBUGGER",
119 	"NOTIFY_WORK_ON_UNMAPPED_QUEUE",
120 	"NOTIFY_TO_UNMAP_PROCESSES",
121 };
122 
mes_v12_0_get_op_string(union MESAPI__MISC * x_pkt)123 static const char *mes_v12_0_get_op_string(union MESAPI__MISC *x_pkt)
124 {
125 	const char *op_str = NULL;
126 
127 	if (x_pkt->header.opcode < ARRAY_SIZE(mes_v12_0_opcodes))
128 		op_str = mes_v12_0_opcodes[x_pkt->header.opcode];
129 
130 	return op_str;
131 }
132 
mes_v12_0_get_misc_op_string(union MESAPI__MISC * x_pkt)133 static const char *mes_v12_0_get_misc_op_string(union MESAPI__MISC *x_pkt)
134 {
135 	const char *op_str = NULL;
136 
137 	if ((x_pkt->header.opcode == MES_SCH_API_MISC) &&
138 	    (x_pkt->opcode < ARRAY_SIZE(mes_v12_0_misc_opcodes)))
139 		op_str = mes_v12_0_misc_opcodes[x_pkt->opcode];
140 
141 	return op_str;
142 }
143 
mes_v12_0_submit_pkt_and_poll_completion(struct amdgpu_mes * mes,int pipe,void * pkt,int size,int api_status_off)144 static int mes_v12_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
145 					    int pipe, void *pkt, int size,
146 					    int api_status_off)
147 {
148 	union MESAPI__QUERY_MES_STATUS mes_status_pkt;
149 	signed long timeout = 2100000; /* 2100 ms */
150 	struct amdgpu_device *adev = mes->adev;
151 	struct amdgpu_ring *ring = &mes->ring[pipe];
152 	spinlock_t *ring_lock = &mes->ring_lock[pipe];
153 	struct MES_API_STATUS *api_status;
154 	union MESAPI__MISC *x_pkt = pkt;
155 	const char *op_str, *misc_op_str;
156 	unsigned long flags;
157 	u64 status_gpu_addr;
158 	u32 seq, status_offset;
159 	u64 *status_ptr;
160 	signed long r;
161 	int ret;
162 
163 	if (x_pkt->header.opcode >= MES_SCH_API_MAX)
164 		return -EINVAL;
165 
166 	if (amdgpu_emu_mode) {
167 		timeout *= 100;
168 	} else if (amdgpu_sriov_vf(adev)) {
169 		/* Worst case in sriov where all other 15 VF timeout, each VF needs about 600ms */
170 		timeout = 15 * 600 * 1000;
171 	}
172 
173 	ret = amdgpu_device_wb_get(adev, &status_offset);
174 	if (ret)
175 		return ret;
176 
177 	status_gpu_addr = adev->wb.gpu_addr + (status_offset * 4);
178 	status_ptr = (u64 *)&adev->wb.wb[status_offset];
179 	*status_ptr = 0;
180 
181 	spin_lock_irqsave(ring_lock, flags);
182 	r = amdgpu_ring_alloc(ring, (size + sizeof(mes_status_pkt)) / 4);
183 	if (r)
184 		goto error_unlock_free;
185 
186 	seq = ++ring->fence_drv.sync_seq;
187 	r = amdgpu_fence_wait_polling(ring,
188 				      seq - ring->fence_drv.num_fences_mask,
189 				      timeout);
190 	if (r < 1)
191 		goto error_undo;
192 
193 	api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
194 	api_status->api_completion_fence_addr = status_gpu_addr;
195 	api_status->api_completion_fence_value = 1;
196 
197 	amdgpu_ring_write_multiple(ring, pkt, size / 4);
198 
199 	memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
200 	mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
201 	mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
202 	mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
203 	mes_status_pkt.api_status.api_completion_fence_addr =
204 		ring->fence_drv.gpu_addr;
205 	mes_status_pkt.api_status.api_completion_fence_value = seq;
206 
207 	amdgpu_ring_write_multiple(ring, &mes_status_pkt,
208 				   sizeof(mes_status_pkt) / 4);
209 
210 	amdgpu_ring_commit(ring);
211 	spin_unlock_irqrestore(ring_lock, flags);
212 
213 	op_str = mes_v12_0_get_op_string(x_pkt);
214 	misc_op_str = mes_v12_0_get_misc_op_string(x_pkt);
215 
216 	if (misc_op_str)
217 		dev_dbg(adev->dev, "MES(%d) msg=%s (%s) was emitted\n",
218 			pipe, op_str, misc_op_str);
219 	else if (op_str)
220 		dev_dbg(adev->dev, "MES(%d) msg=%s was emitted\n",
221 			pipe, op_str);
222 	else
223 		dev_dbg(adev->dev, "MES(%d) msg=%d was emitted\n",
224 			pipe, x_pkt->header.opcode);
225 
226 	r = amdgpu_fence_wait_polling(ring, seq, timeout);
227 	if (r < 1 || !*status_ptr) {
228 
229 		if (misc_op_str)
230 			dev_err(adev->dev, "MES(%d) failed to respond to msg=%s (%s)\n",
231 				pipe, op_str, misc_op_str);
232 		else if (op_str)
233 			dev_err(adev->dev, "MES(%d) failed to respond to msg=%s\n",
234 				pipe, op_str);
235 		else
236 			dev_err(adev->dev, "MES(%d) failed to respond to msg=%d\n",
237 				pipe, x_pkt->header.opcode);
238 
239 		while (halt_if_hws_hang)
240 			schedule();
241 
242 		r = -ETIMEDOUT;
243 		goto error_wb_free;
244 	}
245 
246 	amdgpu_device_wb_free(adev, status_offset);
247 	return 0;
248 
249 error_undo:
250 	dev_err(adev->dev, "MES ring buffer is full.\n");
251 	amdgpu_ring_undo(ring);
252 
253 error_unlock_free:
254 	spin_unlock_irqrestore(ring_lock, flags);
255 
256 error_wb_free:
257 	amdgpu_device_wb_free(adev, status_offset);
258 	return r;
259 }
260 
convert_to_mes_queue_type(int queue_type)261 static int convert_to_mes_queue_type(int queue_type)
262 {
263 	if (queue_type == AMDGPU_RING_TYPE_GFX)
264 		return MES_QUEUE_TYPE_GFX;
265 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
266 		return MES_QUEUE_TYPE_COMPUTE;
267 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
268 		return MES_QUEUE_TYPE_SDMA;
269 	else if (queue_type == AMDGPU_RING_TYPE_MES)
270 		return MES_QUEUE_TYPE_SCHQ;
271 	else
272 		BUG();
273 	return -1;
274 }
275 
mes_v12_0_add_hw_queue(struct amdgpu_mes * mes,struct mes_add_queue_input * input)276 static int mes_v12_0_add_hw_queue(struct amdgpu_mes *mes,
277 				  struct mes_add_queue_input *input)
278 {
279 	struct amdgpu_device *adev = mes->adev;
280 	union MESAPI__ADD_QUEUE mes_add_queue_pkt;
281 	struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB(0)];
282 	uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
283 
284 	memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
285 
286 	mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
287 	mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
288 	mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
289 
290 	mes_add_queue_pkt.process_id = input->process_id;
291 	mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
292 	mes_add_queue_pkt.process_va_start = input->process_va_start;
293 	mes_add_queue_pkt.process_va_end = input->process_va_end;
294 	mes_add_queue_pkt.process_quantum = input->process_quantum;
295 	mes_add_queue_pkt.process_context_addr = input->process_context_addr;
296 	mes_add_queue_pkt.gang_quantum = input->gang_quantum;
297 	mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
298 	mes_add_queue_pkt.inprocess_gang_priority =
299 		input->inprocess_gang_priority;
300 	mes_add_queue_pkt.gang_global_priority_level =
301 		input->gang_global_priority_level;
302 	mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
303 	mes_add_queue_pkt.mqd_addr = input->mqd_addr;
304 
305 	mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr;
306 
307 	mes_add_queue_pkt.queue_type =
308 		convert_to_mes_queue_type(input->queue_type);
309 	mes_add_queue_pkt.paging = input->paging;
310 	mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
311 	mes_add_queue_pkt.gws_base = input->gws_base;
312 	mes_add_queue_pkt.gws_size = input->gws_size;
313 	mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
314 	mes_add_queue_pkt.tma_addr = input->tma_addr;
315 	mes_add_queue_pkt.trap_en = input->trap_en;
316 	mes_add_queue_pkt.skip_process_ctx_clear = input->skip_process_ctx_clear;
317 	mes_add_queue_pkt.is_kfd_process = input->is_kfd_process;
318 
319 	/* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
320 	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
321 	mes_add_queue_pkt.gds_size = input->queue_size;
322 
323 	/* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
324 	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
325 	mes_add_queue_pkt.gds_size = input->queue_size;
326 
327 	return mes_v12_0_submit_pkt_and_poll_completion(mes,
328 			AMDGPU_MES_SCHED_PIPE,
329 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
330 			offsetof(union MESAPI__ADD_QUEUE, api_status));
331 }
332 
mes_v12_0_remove_hw_queue(struct amdgpu_mes * mes,struct mes_remove_queue_input * input)333 static int mes_v12_0_remove_hw_queue(struct amdgpu_mes *mes,
334 				     struct mes_remove_queue_input *input)
335 {
336 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
337 
338 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
339 
340 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
341 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
342 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
343 
344 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
345 	mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
346 
347 	return mes_v12_0_submit_pkt_and_poll_completion(mes,
348 			AMDGPU_MES_SCHED_PIPE,
349 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
350 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
351 }
352 
mes_v12_0_reset_hw_queue(struct amdgpu_mes * mes,struct mes_reset_queue_input * input)353 static int mes_v12_0_reset_hw_queue(struct amdgpu_mes *mes,
354 				    struct mes_reset_queue_input *input)
355 {
356 	union MESAPI__RESET mes_reset_queue_pkt;
357 	int pipe;
358 
359 	memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
360 
361 	mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
362 	mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
363 	mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
364 
365 	mes_reset_queue_pkt.doorbell_offset = input->doorbell_offset;
366 	mes_reset_queue_pkt.gang_context_addr = input->gang_context_addr;
367 	/*mes_reset_queue_pkt.reset_queue_only = 1;*/
368 
369 	if (mes->adev->enable_uni_mes)
370 		pipe = AMDGPU_MES_KIQ_PIPE;
371 	else
372 		pipe = AMDGPU_MES_SCHED_PIPE;
373 
374 	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
375 			&mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
376 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
377 }
378 
mes_v12_0_map_legacy_queue(struct amdgpu_mes * mes,struct mes_map_legacy_queue_input * input)379 static int mes_v12_0_map_legacy_queue(struct amdgpu_mes *mes,
380 				      struct mes_map_legacy_queue_input *input)
381 {
382 	union MESAPI__ADD_QUEUE mes_add_queue_pkt;
383 	int pipe;
384 
385 	memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
386 
387 	mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
388 	mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
389 	mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
390 
391 	mes_add_queue_pkt.pipe_id = input->pipe_id;
392 	mes_add_queue_pkt.queue_id = input->queue_id;
393 	mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
394 	mes_add_queue_pkt.mqd_addr = input->mqd_addr;
395 	mes_add_queue_pkt.wptr_addr = input->wptr_addr;
396 	mes_add_queue_pkt.queue_type =
397 		convert_to_mes_queue_type(input->queue_type);
398 	mes_add_queue_pkt.map_legacy_kq = 1;
399 
400 	if (mes->adev->enable_uni_mes)
401 		pipe = AMDGPU_MES_KIQ_PIPE;
402 	else
403 		pipe = AMDGPU_MES_SCHED_PIPE;
404 
405 	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
406 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
407 			offsetof(union MESAPI__ADD_QUEUE, api_status));
408 }
409 
mes_v12_0_unmap_legacy_queue(struct amdgpu_mes * mes,struct mes_unmap_legacy_queue_input * input)410 static int mes_v12_0_unmap_legacy_queue(struct amdgpu_mes *mes,
411 			struct mes_unmap_legacy_queue_input *input)
412 {
413 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
414 	int pipe;
415 
416 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
417 
418 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
419 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
420 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
421 
422 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
423 	mes_remove_queue_pkt.gang_context_addr = 0;
424 
425 	mes_remove_queue_pkt.pipe_id = input->pipe_id;
426 	mes_remove_queue_pkt.queue_id = input->queue_id;
427 
428 	if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
429 		mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
430 		mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
431 		mes_remove_queue_pkt.tf_data =
432 			lower_32_bits(input->trail_fence_data);
433 	} else {
434 		mes_remove_queue_pkt.unmap_legacy_queue = 1;
435 		mes_remove_queue_pkt.queue_type =
436 			convert_to_mes_queue_type(input->queue_type);
437 	}
438 
439 	if (mes->adev->enable_uni_mes)
440 		pipe = AMDGPU_MES_KIQ_PIPE;
441 	else
442 		pipe = AMDGPU_MES_SCHED_PIPE;
443 
444 	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
445 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
446 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
447 }
448 
mes_v12_0_suspend_gang(struct amdgpu_mes * mes,struct mes_suspend_gang_input * input)449 static int mes_v12_0_suspend_gang(struct amdgpu_mes *mes,
450 				  struct mes_suspend_gang_input *input)
451 {
452 	return 0;
453 }
454 
mes_v12_0_resume_gang(struct amdgpu_mes * mes,struct mes_resume_gang_input * input)455 static int mes_v12_0_resume_gang(struct amdgpu_mes *mes,
456 				 struct mes_resume_gang_input *input)
457 {
458 	return 0;
459 }
460 
mes_v12_0_query_sched_status(struct amdgpu_mes * mes,int pipe)461 static int mes_v12_0_query_sched_status(struct amdgpu_mes *mes, int pipe)
462 {
463 	union MESAPI__QUERY_MES_STATUS mes_status_pkt;
464 
465 	memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
466 
467 	mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
468 	mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
469 	mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
470 
471 	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
472 			&mes_status_pkt, sizeof(mes_status_pkt),
473 			offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
474 }
475 
mes_v12_0_misc_op(struct amdgpu_mes * mes,struct mes_misc_op_input * input)476 static int mes_v12_0_misc_op(struct amdgpu_mes *mes,
477 			     struct mes_misc_op_input *input)
478 {
479 	union MESAPI__MISC misc_pkt;
480 	int pipe;
481 
482 	if (mes->adev->enable_uni_mes)
483 		pipe = AMDGPU_MES_KIQ_PIPE;
484 	else
485 		pipe = AMDGPU_MES_SCHED_PIPE;
486 
487 	memset(&misc_pkt, 0, sizeof(misc_pkt));
488 
489 	misc_pkt.header.type = MES_API_TYPE_SCHEDULER;
490 	misc_pkt.header.opcode = MES_SCH_API_MISC;
491 	misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
492 
493 	switch (input->op) {
494 	case MES_MISC_OP_READ_REG:
495 		misc_pkt.opcode = MESAPI_MISC__READ_REG;
496 		misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset;
497 		misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr;
498 		break;
499 	case MES_MISC_OP_WRITE_REG:
500 		misc_pkt.opcode = MESAPI_MISC__WRITE_REG;
501 		misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset;
502 		misc_pkt.write_reg.reg_value = input->write_reg.reg_value;
503 		break;
504 	case MES_MISC_OP_WRM_REG_WAIT:
505 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
506 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM;
507 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
508 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
509 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
510 		misc_pkt.wait_reg_mem.reg_offset2 = 0;
511 		break;
512 	case MES_MISC_OP_WRM_REG_WR_WAIT:
513 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
514 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG;
515 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
516 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
517 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
518 		misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1;
519 		break;
520 	case MES_MISC_OP_SET_SHADER_DEBUGGER:
521 		pipe = AMDGPU_MES_SCHED_PIPE;
522 		misc_pkt.opcode = MESAPI_MISC__SET_SHADER_DEBUGGER;
523 		misc_pkt.set_shader_debugger.process_context_addr =
524 				input->set_shader_debugger.process_context_addr;
525 		misc_pkt.set_shader_debugger.flags.u32all =
526 				input->set_shader_debugger.flags.u32all;
527 		misc_pkt.set_shader_debugger.spi_gdbg_per_vmid_cntl =
528 				input->set_shader_debugger.spi_gdbg_per_vmid_cntl;
529 		memcpy(misc_pkt.set_shader_debugger.tcp_watch_cntl,
530 				input->set_shader_debugger.tcp_watch_cntl,
531 				sizeof(misc_pkt.set_shader_debugger.tcp_watch_cntl));
532 		misc_pkt.set_shader_debugger.trap_en = input->set_shader_debugger.trap_en;
533 		break;
534 	default:
535 		DRM_ERROR("unsupported misc op (%d) \n", input->op);
536 		return -EINVAL;
537 	}
538 
539 	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
540 			&misc_pkt, sizeof(misc_pkt),
541 			offsetof(union MESAPI__MISC, api_status));
542 }
543 
mes_v12_0_set_hw_resources_1(struct amdgpu_mes * mes,int pipe)544 static int mes_v12_0_set_hw_resources_1(struct amdgpu_mes *mes, int pipe)
545 {
546 	union MESAPI_SET_HW_RESOURCES_1 mes_set_hw_res_1_pkt;
547 
548 	memset(&mes_set_hw_res_1_pkt, 0, sizeof(mes_set_hw_res_1_pkt));
549 
550 	mes_set_hw_res_1_pkt.header.type = MES_API_TYPE_SCHEDULER;
551 	mes_set_hw_res_1_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC_1;
552 	mes_set_hw_res_1_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
553 	mes_set_hw_res_1_pkt.mes_kiq_unmap_timeout = 0xa;
554 
555 	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
556 			&mes_set_hw_res_1_pkt, sizeof(mes_set_hw_res_1_pkt),
557 			offsetof(union MESAPI_SET_HW_RESOURCES_1, api_status));
558 }
559 
mes_v12_0_set_hw_resources(struct amdgpu_mes * mes,int pipe)560 static int mes_v12_0_set_hw_resources(struct amdgpu_mes *mes, int pipe)
561 {
562 	int i;
563 	struct amdgpu_device *adev = mes->adev;
564 	union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
565 
566 	memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
567 
568 	mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
569 	mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
570 	mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
571 
572 	if (pipe == AMDGPU_MES_SCHED_PIPE) {
573 		mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
574 		mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
575 		mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
576 		mes_set_hw_res_pkt.paging_vmid = 0;
577 
578 		for (i = 0; i < MAX_COMPUTE_PIPES; i++)
579 			mes_set_hw_res_pkt.compute_hqd_mask[i] =
580 				mes->compute_hqd_mask[i];
581 
582 		for (i = 0; i < MAX_GFX_PIPES; i++)
583 			mes_set_hw_res_pkt.gfx_hqd_mask[i] =
584 				mes->gfx_hqd_mask[i];
585 
586 		for (i = 0; i < MAX_SDMA_PIPES; i++)
587 			mes_set_hw_res_pkt.sdma_hqd_mask[i] =
588 				mes->sdma_hqd_mask[i];
589 
590 		for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
591 			mes_set_hw_res_pkt.aggregated_doorbells[i] =
592 				mes->aggregated_doorbells[i];
593 	}
594 
595 	mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr =
596 		mes->sch_ctx_gpu_addr[pipe];
597 	mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
598 		mes->query_status_fence_gpu_addr[pipe];
599 
600 	for (i = 0; i < 5; i++) {
601 		mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
602 		mes_set_hw_res_pkt.mmhub_base[i] =
603 				adev->reg_offset[MMHUB_HWIP][0][i];
604 		mes_set_hw_res_pkt.osssys_base[i] =
605 		adev->reg_offset[OSSSYS_HWIP][0][i];
606 	}
607 
608 	mes_set_hw_res_pkt.disable_reset = 1;
609 	mes_set_hw_res_pkt.disable_mes_log = 1;
610 	mes_set_hw_res_pkt.use_different_vmid_compute = 1;
611 	mes_set_hw_res_pkt.enable_reg_active_poll = 1;
612 	mes_set_hw_res_pkt.enable_level_process_quantum_check = 1;
613 	if ((mes->adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 0x82)
614 		mes_set_hw_res_pkt.enable_lr_compute_wa = 1;
615 	else
616 		dev_info_once(adev->dev,
617 			      "MES FW version must be >= 0x82 to enable LR compute workaround.\n");
618 
619 	/*
620 	 * Keep oversubscribe timer for sdma . When we have unmapped doorbell
621 	 * handling support, other queue will not use the oversubscribe timer.
622 	 * handling  mode - 0: disabled; 1: basic version; 2: basic+ version
623 	 */
624 	mes_set_hw_res_pkt.oversubscription_timer = 50;
625 	mes_set_hw_res_pkt.unmapped_doorbell_handling = 1;
626 
627 	if (amdgpu_mes_log_enable) {
628 		mes_set_hw_res_pkt.enable_mes_event_int_logging = 1;
629 		mes_set_hw_res_pkt.event_intr_history_gpu_mc_ptr = mes->event_log_gpu_addr + pipe * AMDGPU_MES_LOG_BUFFER_SIZE;
630 	}
631 
632 	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
633 			&mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
634 			offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
635 }
636 
mes_v12_0_init_aggregated_doorbell(struct amdgpu_mes * mes)637 static void mes_v12_0_init_aggregated_doorbell(struct amdgpu_mes *mes)
638 {
639 	struct amdgpu_device *adev = mes->adev;
640 	uint32_t data;
641 
642 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1);
643 	data &= ~(CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET_MASK |
644 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_EN_MASK |
645 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_HIT_MASK);
646 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_LOW] <<
647 		CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET__SHIFT;
648 	data |= 1 << CP_MES_DOORBELL_CONTROL1__DOORBELL_EN__SHIFT;
649 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1, data);
650 
651 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2);
652 	data &= ~(CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET_MASK |
653 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_EN_MASK |
654 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_HIT_MASK);
655 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] <<
656 		CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET__SHIFT;
657 	data |= 1 << CP_MES_DOORBELL_CONTROL2__DOORBELL_EN__SHIFT;
658 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2, data);
659 
660 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3);
661 	data &= ~(CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET_MASK |
662 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_EN_MASK |
663 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_HIT_MASK);
664 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_MEDIUM] <<
665 		CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET__SHIFT;
666 	data |= 1 << CP_MES_DOORBELL_CONTROL3__DOORBELL_EN__SHIFT;
667 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3, data);
668 
669 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4);
670 	data &= ~(CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET_MASK |
671 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_EN_MASK |
672 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_HIT_MASK);
673 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_HIGH] <<
674 		CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET__SHIFT;
675 	data |= 1 << CP_MES_DOORBELL_CONTROL4__DOORBELL_EN__SHIFT;
676 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4, data);
677 
678 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5);
679 	data &= ~(CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET_MASK |
680 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_EN_MASK |
681 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_HIT_MASK);
682 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_REALTIME] <<
683 		CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET__SHIFT;
684 	data |= 1 << CP_MES_DOORBELL_CONTROL5__DOORBELL_EN__SHIFT;
685 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5, data);
686 
687 	data = 1 << CP_HQD_GFX_CONTROL__DB_UPDATED_MSG_EN__SHIFT;
688 	WREG32_SOC15(GC, 0, regCP_HQD_GFX_CONTROL, data);
689 }
690 
691 
mes_v12_0_enable_unmapped_doorbell_handling(struct amdgpu_mes * mes,bool enable)692 static void mes_v12_0_enable_unmapped_doorbell_handling(
693 		struct amdgpu_mes *mes, bool enable)
694 {
695 	struct amdgpu_device *adev = mes->adev;
696 	uint32_t data = RREG32_SOC15(GC, 0, regCP_UNMAPPED_DOORBELL);
697 
698 	/*
699 	 * The default PROC_LSB settng is 0xc which means doorbell
700 	 * addr[16:12] gives the doorbell page number. For kfd, each
701 	 * process will use 2 pages of doorbell, we need to change the
702 	 * setting to 0xd
703 	 */
704 	data &= ~CP_UNMAPPED_DOORBELL__PROC_LSB_MASK;
705 	data |= 0xd <<  CP_UNMAPPED_DOORBELL__PROC_LSB__SHIFT;
706 
707 	data |= (enable ? 1 : 0) << CP_UNMAPPED_DOORBELL__ENABLE__SHIFT;
708 
709 	WREG32_SOC15(GC, 0, regCP_UNMAPPED_DOORBELL, data);
710 }
711 
mes_v12_0_reset_legacy_queue(struct amdgpu_mes * mes,struct mes_reset_legacy_queue_input * input)712 static int mes_v12_0_reset_legacy_queue(struct amdgpu_mes *mes,
713 					struct mes_reset_legacy_queue_input *input)
714 {
715 	union MESAPI__RESET mes_reset_queue_pkt;
716 	int pipe;
717 
718 	memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
719 
720 	mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
721 	mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
722 	mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
723 
724 	mes_reset_queue_pkt.queue_type =
725 		convert_to_mes_queue_type(input->queue_type);
726 
727 	if (mes_reset_queue_pkt.queue_type == MES_QUEUE_TYPE_GFX) {
728 		mes_reset_queue_pkt.reset_legacy_gfx = 1;
729 		mes_reset_queue_pkt.pipe_id_lp = input->pipe_id;
730 		mes_reset_queue_pkt.queue_id_lp = input->queue_id;
731 		mes_reset_queue_pkt.mqd_mc_addr_lp = input->mqd_addr;
732 		mes_reset_queue_pkt.doorbell_offset_lp = input->doorbell_offset;
733 		mes_reset_queue_pkt.wptr_addr_lp = input->wptr_addr;
734 		mes_reset_queue_pkt.vmid_id_lp = input->vmid;
735 	} else {
736 		mes_reset_queue_pkt.reset_queue_only = 1;
737 		mes_reset_queue_pkt.doorbell_offset = input->doorbell_offset;
738 	}
739 
740 	if (mes->adev->enable_uni_mes)
741 		pipe = AMDGPU_MES_KIQ_PIPE;
742 	else
743 		pipe = AMDGPU_MES_SCHED_PIPE;
744 
745 	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
746 			&mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
747 			offsetof(union MESAPI__RESET, api_status));
748 }
749 
750 static const struct amdgpu_mes_funcs mes_v12_0_funcs = {
751 	.add_hw_queue = mes_v12_0_add_hw_queue,
752 	.remove_hw_queue = mes_v12_0_remove_hw_queue,
753 	.map_legacy_queue = mes_v12_0_map_legacy_queue,
754 	.unmap_legacy_queue = mes_v12_0_unmap_legacy_queue,
755 	.suspend_gang = mes_v12_0_suspend_gang,
756 	.resume_gang = mes_v12_0_resume_gang,
757 	.misc_op = mes_v12_0_misc_op,
758 	.reset_legacy_queue = mes_v12_0_reset_legacy_queue,
759 	.reset_hw_queue = mes_v12_0_reset_hw_queue,
760 };
761 
mes_v12_0_allocate_ucode_buffer(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)762 static int mes_v12_0_allocate_ucode_buffer(struct amdgpu_device *adev,
763 					   enum admgpu_mes_pipe pipe)
764 {
765 	int r;
766 	const struct mes_firmware_header_v1_0 *mes_hdr;
767 	const __le32 *fw_data;
768 	unsigned fw_size;
769 
770 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
771 		adev->mes.fw[pipe]->data;
772 
773 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
774 		   le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
775 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
776 
777 	r = amdgpu_bo_create_reserved(adev, fw_size,
778 				      PAGE_SIZE,
779 				      AMDGPU_GEM_DOMAIN_VRAM,
780 				      &adev->mes.ucode_fw_obj[pipe],
781 				      &adev->mes.ucode_fw_gpu_addr[pipe],
782 				      (void **)&adev->mes.ucode_fw_ptr[pipe]);
783 	if (r) {
784 		dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
785 		return r;
786 	}
787 
788 	memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
789 
790 	amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
791 	amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
792 
793 	return 0;
794 }
795 
mes_v12_0_allocate_ucode_data_buffer(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)796 static int mes_v12_0_allocate_ucode_data_buffer(struct amdgpu_device *adev,
797 						enum admgpu_mes_pipe pipe)
798 {
799 	int r;
800 	const struct mes_firmware_header_v1_0 *mes_hdr;
801 	const __le32 *fw_data;
802 	unsigned fw_size;
803 
804 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
805 		adev->mes.fw[pipe]->data;
806 
807 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
808 		   le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
809 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
810 
811 	r = amdgpu_bo_create_reserved(adev, fw_size,
812 				      64 * 1024,
813 				      AMDGPU_GEM_DOMAIN_VRAM,
814 				      &adev->mes.data_fw_obj[pipe],
815 				      &adev->mes.data_fw_gpu_addr[pipe],
816 				      (void **)&adev->mes.data_fw_ptr[pipe]);
817 	if (r) {
818 		dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
819 		return r;
820 	}
821 
822 	memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
823 
824 	amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
825 	amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
826 
827 	return 0;
828 }
829 
mes_v12_0_free_ucode_buffers(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)830 static void mes_v12_0_free_ucode_buffers(struct amdgpu_device *adev,
831 					 enum admgpu_mes_pipe pipe)
832 {
833 	amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
834 			      &adev->mes.data_fw_gpu_addr[pipe],
835 			      (void **)&adev->mes.data_fw_ptr[pipe]);
836 
837 	amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
838 			      &adev->mes.ucode_fw_gpu_addr[pipe],
839 			      (void **)&adev->mes.ucode_fw_ptr[pipe]);
840 }
841 
mes_v12_0_enable(struct amdgpu_device * adev,bool enable)842 static void mes_v12_0_enable(struct amdgpu_device *adev, bool enable)
843 {
844 	uint64_t ucode_addr;
845 	uint32_t pipe, data = 0;
846 
847 	if (enable) {
848 		data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
849 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
850 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 1);
851 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
852 
853 		mutex_lock(&adev->srbm_mutex);
854 		for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
855 			soc21_grbm_select(adev, 3, pipe, 0, 0);
856 
857 			ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
858 			WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
859 				     lower_32_bits(ucode_addr));
860 			WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
861 				     upper_32_bits(ucode_addr));
862 		}
863 		soc21_grbm_select(adev, 0, 0, 0, 0);
864 		mutex_unlock(&adev->srbm_mutex);
865 
866 		/* unhalt MES and activate pipe0 */
867 		data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
868 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 1);
869 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
870 
871 		if (amdgpu_emu_mode)
872 			msleep(100);
873 		else if (adev->enable_uni_mes)
874 			udelay(500);
875 		else
876 			udelay(50);
877 	} else {
878 		data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
879 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
880 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
881 		data = REG_SET_FIELD(data, CP_MES_CNTL,
882 				     MES_INVALIDATE_ICACHE, 1);
883 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
884 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 1);
885 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
886 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
887 	}
888 }
889 
mes_v12_0_set_ucode_start_addr(struct amdgpu_device * adev)890 static void mes_v12_0_set_ucode_start_addr(struct amdgpu_device *adev)
891 {
892 	uint64_t ucode_addr;
893 	int pipe;
894 
895 	mes_v12_0_enable(adev, false);
896 
897 	mutex_lock(&adev->srbm_mutex);
898 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
899 		/* me=3, queue=0 */
900 		soc21_grbm_select(adev, 3, pipe, 0, 0);
901 
902 		/* set ucode start address */
903 		ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
904 		WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
905 				lower_32_bits(ucode_addr));
906 		WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
907 				upper_32_bits(ucode_addr));
908 
909 		soc21_grbm_select(adev, 0, 0, 0, 0);
910 	}
911 	mutex_unlock(&adev->srbm_mutex);
912 }
913 
914 /* This function is for backdoor MES firmware */
mes_v12_0_load_microcode(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe,bool prime_icache)915 static int mes_v12_0_load_microcode(struct amdgpu_device *adev,
916 				    enum admgpu_mes_pipe pipe, bool prime_icache)
917 {
918 	int r;
919 	uint32_t data;
920 
921 	mes_v12_0_enable(adev, false);
922 
923 	if (!adev->mes.fw[pipe])
924 		return -EINVAL;
925 
926 	r = mes_v12_0_allocate_ucode_buffer(adev, pipe);
927 	if (r)
928 		return r;
929 
930 	r = mes_v12_0_allocate_ucode_data_buffer(adev, pipe);
931 	if (r) {
932 		mes_v12_0_free_ucode_buffers(adev, pipe);
933 		return r;
934 	}
935 
936 	mutex_lock(&adev->srbm_mutex);
937 	/* me=3, pipe=0, queue=0 */
938 	soc21_grbm_select(adev, 3, pipe, 0, 0);
939 
940 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0);
941 
942 	/* set ucode fimrware address */
943 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO,
944 		     lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
945 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI,
946 		     upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
947 
948 	/* set ucode instruction cache boundary to 2M-1 */
949 	WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF);
950 
951 	/* set ucode data firmware address */
952 	WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO,
953 		     lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
954 	WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI,
955 		     upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
956 
957 	/* Set data cache boundary CP_MES_MDBOUND_LO */
958 	WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x7FFFF);
959 
960 	if (prime_icache) {
961 		/* invalidate ICACHE */
962 		data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
963 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
964 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
965 		WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
966 
967 		/* prime the ICACHE. */
968 		data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
969 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
970 		WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
971 	}
972 
973 	soc21_grbm_select(adev, 0, 0, 0, 0);
974 	mutex_unlock(&adev->srbm_mutex);
975 
976 	return 0;
977 }
978 
mes_v12_0_allocate_eop_buf(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)979 static int mes_v12_0_allocate_eop_buf(struct amdgpu_device *adev,
980 				      enum admgpu_mes_pipe pipe)
981 {
982 	int r;
983 	u32 *eop;
984 
985 	r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
986 			      AMDGPU_GEM_DOMAIN_GTT,
987 			      &adev->mes.eop_gpu_obj[pipe],
988 			      &adev->mes.eop_gpu_addr[pipe],
989 			      (void **)&eop);
990 	if (r) {
991 		dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
992 		return r;
993 	}
994 
995 	memset(eop, 0,
996 	       adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
997 
998 	amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
999 	amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
1000 
1001 	return 0;
1002 }
1003 
mes_v12_0_mqd_init(struct amdgpu_ring * ring)1004 static int mes_v12_0_mqd_init(struct amdgpu_ring *ring)
1005 {
1006 	struct v12_compute_mqd *mqd = ring->mqd_ptr;
1007 	uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
1008 	uint32_t tmp;
1009 
1010 	mqd->header = 0xC0310800;
1011 	mqd->compute_pipelinestat_enable = 0x00000001;
1012 	mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
1013 	mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
1014 	mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
1015 	mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
1016 	mqd->compute_misc_reserved = 0x00000007;
1017 
1018 	eop_base_addr = ring->eop_gpu_addr >> 8;
1019 
1020 	/* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
1021 	tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
1022 	tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
1023 			(order_base_2(MES_EOP_SIZE / 4) - 1));
1024 
1025 	mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
1026 	mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
1027 	mqd->cp_hqd_eop_control = tmp;
1028 
1029 	/* disable the queue if it's active */
1030 	ring->wptr = 0;
1031 	mqd->cp_hqd_pq_rptr = 0;
1032 	mqd->cp_hqd_pq_wptr_lo = 0;
1033 	mqd->cp_hqd_pq_wptr_hi = 0;
1034 
1035 	/* set the pointer to the MQD */
1036 	mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
1037 	mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
1038 
1039 	/* set MQD vmid to 0 */
1040 	tmp = regCP_MQD_CONTROL_DEFAULT;
1041 	tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
1042 	mqd->cp_mqd_control = tmp;
1043 
1044 	/* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
1045 	hqd_gpu_addr = ring->gpu_addr >> 8;
1046 	mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
1047 	mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
1048 
1049 	/* set the wb address whether it's enabled or not */
1050 	wb_gpu_addr = ring->rptr_gpu_addr;
1051 	mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
1052 	mqd->cp_hqd_pq_rptr_report_addr_hi =
1053 		upper_32_bits(wb_gpu_addr) & 0xffff;
1054 
1055 	/* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
1056 	wb_gpu_addr = ring->wptr_gpu_addr;
1057 	mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
1058 	mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
1059 
1060 	/* set up the HQD, this is similar to CP_RB0_CNTL */
1061 	tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
1062 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
1063 			    (order_base_2(ring->ring_size / 4) - 1));
1064 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
1065 			    ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
1066 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
1067 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
1068 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
1069 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
1070 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
1071 	mqd->cp_hqd_pq_control = tmp;
1072 
1073 	/* enable doorbell */
1074 	tmp = 0;
1075 	if (ring->use_doorbell) {
1076 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1077 				    DOORBELL_OFFSET, ring->doorbell_index);
1078 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1079 				    DOORBELL_EN, 1);
1080 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1081 				    DOORBELL_SOURCE, 0);
1082 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1083 				    DOORBELL_HIT, 0);
1084 	} else {
1085 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1086 				    DOORBELL_EN, 0);
1087 	}
1088 	mqd->cp_hqd_pq_doorbell_control = tmp;
1089 
1090 	mqd->cp_hqd_vmid = 0;
1091 	/* activate the queue */
1092 	mqd->cp_hqd_active = 1;
1093 
1094 	tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
1095 	tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
1096 			    PRELOAD_SIZE, 0x55);
1097 	mqd->cp_hqd_persistent_state = tmp;
1098 
1099 	mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT;
1100 	mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
1101 	mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
1102 
1103 	/*
1104 	 * Set CP_HQD_GFX_CONTROL.DB_UPDATED_MSG_EN[15] to enable unmapped
1105 	 * doorbell handling. This is a reserved CP internal register can
1106 	 * not be accesss by others
1107 	 */
1108 	mqd->reserved_184 = BIT(15);
1109 
1110 	return 0;
1111 }
1112 
mes_v12_0_queue_init_register(struct amdgpu_ring * ring)1113 static void mes_v12_0_queue_init_register(struct amdgpu_ring *ring)
1114 {
1115 	struct v12_compute_mqd *mqd = ring->mqd_ptr;
1116 	struct amdgpu_device *adev = ring->adev;
1117 	uint32_t data = 0;
1118 
1119 	mutex_lock(&adev->srbm_mutex);
1120 	soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
1121 
1122 	/* set CP_HQD_VMID.VMID = 0. */
1123 	data = RREG32_SOC15(GC, 0, regCP_HQD_VMID);
1124 	data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
1125 	WREG32_SOC15(GC, 0, regCP_HQD_VMID, data);
1126 
1127 	/* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
1128 	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1129 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1130 			     DOORBELL_EN, 0);
1131 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1132 
1133 	/* set CP_MQD_BASE_ADDR/HI with the MQD base address */
1134 	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
1135 	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
1136 
1137 	/* set CP_MQD_CONTROL.VMID=0 */
1138 	data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL);
1139 	data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
1140 	WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0);
1141 
1142 	/* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
1143 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
1144 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
1145 
1146 	/* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
1147 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
1148 		     mqd->cp_hqd_pq_rptr_report_addr_lo);
1149 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
1150 		     mqd->cp_hqd_pq_rptr_report_addr_hi);
1151 
1152 	/* set CP_HQD_PQ_CONTROL */
1153 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
1154 
1155 	/* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
1156 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
1157 		     mqd->cp_hqd_pq_wptr_poll_addr_lo);
1158 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
1159 		     mqd->cp_hqd_pq_wptr_poll_addr_hi);
1160 
1161 	/* set CP_HQD_PQ_DOORBELL_CONTROL */
1162 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
1163 		     mqd->cp_hqd_pq_doorbell_control);
1164 
1165 	/* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
1166 	WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
1167 
1168 	/* set CP_HQD_ACTIVE.ACTIVE=1 */
1169 	WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active);
1170 
1171 	soc21_grbm_select(adev, 0, 0, 0, 0);
1172 	mutex_unlock(&adev->srbm_mutex);
1173 }
1174 
mes_v12_0_kiq_enable_queue(struct amdgpu_device * adev)1175 static int mes_v12_0_kiq_enable_queue(struct amdgpu_device *adev)
1176 {
1177 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
1178 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
1179 	int r;
1180 
1181 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
1182 		return -EINVAL;
1183 
1184 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
1185 	if (r) {
1186 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
1187 		return r;
1188 	}
1189 
1190 	kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring[0]);
1191 
1192 	r = amdgpu_ring_test_ring(kiq_ring);
1193 	if (r) {
1194 		DRM_ERROR("kfq enable failed\n");
1195 		kiq_ring->sched.ready = false;
1196 	}
1197 	return r;
1198 }
1199 
mes_v12_0_queue_init(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)1200 static int mes_v12_0_queue_init(struct amdgpu_device *adev,
1201 				enum admgpu_mes_pipe pipe)
1202 {
1203 	struct amdgpu_ring *ring;
1204 	int r;
1205 
1206 	if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1207 		ring = &adev->gfx.kiq[0].ring;
1208 	else
1209 		ring = &adev->mes.ring[pipe];
1210 
1211 	if ((adev->enable_uni_mes || pipe == AMDGPU_MES_SCHED_PIPE) &&
1212 	    (amdgpu_in_reset(adev) || adev->in_suspend)) {
1213 		*(ring->wptr_cpu_addr) = 0;
1214 		*(ring->rptr_cpu_addr) = 0;
1215 		amdgpu_ring_clear_ring(ring);
1216 	}
1217 
1218 	r = mes_v12_0_mqd_init(ring);
1219 	if (r)
1220 		return r;
1221 
1222 	if (pipe == AMDGPU_MES_SCHED_PIPE) {
1223 		if (adev->enable_uni_mes)
1224 			r = amdgpu_mes_map_legacy_queue(adev, ring);
1225 		else
1226 			r = mes_v12_0_kiq_enable_queue(adev);
1227 		if (r)
1228 			return r;
1229 	} else {
1230 		mes_v12_0_queue_init_register(ring);
1231 	}
1232 
1233 	if (((pipe == AMDGPU_MES_SCHED_PIPE) && !adev->mes.sched_version) ||
1234 	    ((pipe == AMDGPU_MES_KIQ_PIPE) && !adev->mes.kiq_version)) {
1235 		/* get MES scheduler/KIQ versions */
1236 		mutex_lock(&adev->srbm_mutex);
1237 		soc21_grbm_select(adev, 3, pipe, 0, 0);
1238 
1239 		if (pipe == AMDGPU_MES_SCHED_PIPE)
1240 			adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
1241 		else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
1242 			adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
1243 
1244 		soc21_grbm_select(adev, 0, 0, 0, 0);
1245 		mutex_unlock(&adev->srbm_mutex);
1246 	}
1247 
1248 	return 0;
1249 }
1250 
mes_v12_0_ring_init(struct amdgpu_device * adev,int pipe)1251 static int mes_v12_0_ring_init(struct amdgpu_device *adev, int pipe)
1252 {
1253 	struct amdgpu_ring *ring;
1254 
1255 	ring = &adev->mes.ring[pipe];
1256 
1257 	ring->funcs = &mes_v12_0_ring_funcs;
1258 
1259 	ring->me = 3;
1260 	ring->pipe = pipe;
1261 	ring->queue = 0;
1262 
1263 	ring->ring_obj = NULL;
1264 	ring->use_doorbell = true;
1265 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[pipe];
1266 	ring->no_scheduler = true;
1267 	sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
1268 
1269 	if (pipe == AMDGPU_MES_SCHED_PIPE)
1270 		ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
1271 	else
1272 		ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
1273 
1274 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1275 				AMDGPU_RING_PRIO_DEFAULT, NULL);
1276 }
1277 
mes_v12_0_kiq_ring_init(struct amdgpu_device * adev)1278 static int mes_v12_0_kiq_ring_init(struct amdgpu_device *adev)
1279 {
1280 	struct amdgpu_ring *ring;
1281 
1282 	spin_lock_init(&adev->gfx.kiq[0].ring_lock);
1283 
1284 	ring = &adev->gfx.kiq[0].ring;
1285 
1286 	ring->me = 3;
1287 	ring->pipe = 1;
1288 	ring->queue = 0;
1289 
1290 	ring->adev = NULL;
1291 	ring->ring_obj = NULL;
1292 	ring->use_doorbell = true;
1293 	ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
1294 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
1295 	ring->no_scheduler = true;
1296 	sprintf(ring->name, "mes_kiq_%d.%d.%d",
1297 		ring->me, ring->pipe, ring->queue);
1298 
1299 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1300 				AMDGPU_RING_PRIO_DEFAULT, NULL);
1301 }
1302 
mes_v12_0_mqd_sw_init(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)1303 static int mes_v12_0_mqd_sw_init(struct amdgpu_device *adev,
1304 				 enum admgpu_mes_pipe pipe)
1305 {
1306 	int r, mqd_size = sizeof(struct v12_compute_mqd);
1307 	struct amdgpu_ring *ring;
1308 
1309 	if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1310 		ring = &adev->gfx.kiq[0].ring;
1311 	else
1312 		ring = &adev->mes.ring[pipe];
1313 
1314 	if (ring->mqd_obj)
1315 		return 0;
1316 
1317 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
1318 				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
1319 				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
1320 	if (r) {
1321 		dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
1322 		return r;
1323 	}
1324 
1325 	memset(ring->mqd_ptr, 0, mqd_size);
1326 
1327 	/* prepare MQD backup */
1328 	adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
1329 	if (!adev->mes.mqd_backup[pipe])
1330 		dev_warn(adev->dev,
1331 			 "no memory to create MQD backup for ring %s\n",
1332 			 ring->name);
1333 
1334 	return 0;
1335 }
1336 
mes_v12_0_sw_init(void * handle)1337 static int mes_v12_0_sw_init(void *handle)
1338 {
1339 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1340 	int pipe, r;
1341 
1342 	adev->mes.funcs = &mes_v12_0_funcs;
1343 	adev->mes.kiq_hw_init = &mes_v12_0_kiq_hw_init;
1344 	adev->mes.kiq_hw_fini = &mes_v12_0_kiq_hw_fini;
1345 	adev->mes.enable_legacy_queue_map = true;
1346 
1347 	adev->mes.event_log_size = adev->enable_uni_mes ? (AMDGPU_MAX_MES_PIPES * AMDGPU_MES_LOG_BUFFER_SIZE) : AMDGPU_MES_LOG_BUFFER_SIZE;
1348 
1349 	r = amdgpu_mes_init(adev);
1350 	if (r)
1351 		return r;
1352 
1353 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1354 		r = mes_v12_0_allocate_eop_buf(adev, pipe);
1355 		if (r)
1356 			return r;
1357 
1358 		r = mes_v12_0_mqd_sw_init(adev, pipe);
1359 		if (r)
1360 			return r;
1361 
1362 		if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1363 			r = mes_v12_0_kiq_ring_init(adev);
1364 		else
1365 			r = mes_v12_0_ring_init(adev, pipe);
1366 		if (r)
1367 			return r;
1368 	}
1369 
1370 	return 0;
1371 }
1372 
mes_v12_0_sw_fini(void * handle)1373 static int mes_v12_0_sw_fini(void *handle)
1374 {
1375 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1376 	int pipe;
1377 
1378 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1379 		kfree(adev->mes.mqd_backup[pipe]);
1380 
1381 		amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1382 				      &adev->mes.eop_gpu_addr[pipe],
1383 				      NULL);
1384 		amdgpu_ucode_release(&adev->mes.fw[pipe]);
1385 
1386 		if (adev->enable_uni_mes || pipe == AMDGPU_MES_SCHED_PIPE) {
1387 			amdgpu_bo_free_kernel(&adev->mes.ring[pipe].mqd_obj,
1388 					      &adev->mes.ring[pipe].mqd_gpu_addr,
1389 					      &adev->mes.ring[pipe].mqd_ptr);
1390 			amdgpu_ring_fini(&adev->mes.ring[pipe]);
1391 		}
1392 	}
1393 
1394 	if (!adev->enable_uni_mes) {
1395 		amdgpu_bo_free_kernel(&adev->gfx.kiq[0].ring.mqd_obj,
1396 				      &adev->gfx.kiq[0].ring.mqd_gpu_addr,
1397 				      &adev->gfx.kiq[0].ring.mqd_ptr);
1398 		amdgpu_ring_fini(&adev->gfx.kiq[0].ring);
1399 	}
1400 
1401 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1402 		mes_v12_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1403 		mes_v12_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1404 	}
1405 
1406 	amdgpu_mes_fini(adev);
1407 	return 0;
1408 }
1409 
mes_v12_0_kiq_dequeue_sched(struct amdgpu_device * adev)1410 static void mes_v12_0_kiq_dequeue_sched(struct amdgpu_device *adev)
1411 {
1412 	uint32_t data;
1413 	int i;
1414 
1415 	mutex_lock(&adev->srbm_mutex);
1416 	soc21_grbm_select(adev, 3, AMDGPU_MES_SCHED_PIPE, 0, 0);
1417 
1418 	/* disable the queue if it's active */
1419 	if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
1420 		WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
1421 		for (i = 0; i < adev->usec_timeout; i++) {
1422 			if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
1423 				break;
1424 			udelay(1);
1425 		}
1426 	}
1427 	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1428 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1429 				DOORBELL_EN, 0);
1430 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1431 				DOORBELL_HIT, 1);
1432 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1433 
1434 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0);
1435 
1436 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0);
1437 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0);
1438 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0);
1439 
1440 	soc21_grbm_select(adev, 0, 0, 0, 0);
1441 	mutex_unlock(&adev->srbm_mutex);
1442 
1443 	adev->mes.ring[0].sched.ready = false;
1444 }
1445 
mes_v12_0_kiq_setting(struct amdgpu_ring * ring)1446 static void mes_v12_0_kiq_setting(struct amdgpu_ring *ring)
1447 {
1448 	uint32_t tmp;
1449 	struct amdgpu_device *adev = ring->adev;
1450 
1451 	/* tell RLC which is KIQ queue */
1452 	tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1453 	tmp &= 0xffffff00;
1454 	tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1455 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1456 	tmp |= 0x80;
1457 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1458 }
1459 
mes_v12_0_kiq_hw_init(struct amdgpu_device * adev)1460 static int mes_v12_0_kiq_hw_init(struct amdgpu_device *adev)
1461 {
1462 	int r = 0;
1463 
1464 	if (adev->enable_uni_mes)
1465 		mes_v12_0_kiq_setting(&adev->mes.ring[AMDGPU_MES_KIQ_PIPE]);
1466 	else
1467 		mes_v12_0_kiq_setting(&adev->gfx.kiq[0].ring);
1468 
1469 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1470 
1471 		r = mes_v12_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE, false);
1472 		if (r) {
1473 			DRM_ERROR("failed to load MES fw, r=%d\n", r);
1474 			return r;
1475 		}
1476 
1477 		r = mes_v12_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE, true);
1478 		if (r) {
1479 			DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1480 			return r;
1481 		}
1482 
1483 		mes_v12_0_set_ucode_start_addr(adev);
1484 
1485 	} else if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO)
1486 		mes_v12_0_set_ucode_start_addr(adev);
1487 
1488 	mes_v12_0_enable(adev, true);
1489 
1490 	r = mes_v12_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE);
1491 	if (r)
1492 		goto failure;
1493 
1494 	if (adev->enable_uni_mes) {
1495 		r = mes_v12_0_set_hw_resources(&adev->mes, AMDGPU_MES_KIQ_PIPE);
1496 		if (r)
1497 			goto failure;
1498 
1499 		mes_v12_0_set_hw_resources_1(&adev->mes, AMDGPU_MES_KIQ_PIPE);
1500 	}
1501 
1502 	if (adev->mes.enable_legacy_queue_map) {
1503 		r = mes_v12_0_hw_init(adev);
1504 		if (r)
1505 			goto failure;
1506 	}
1507 
1508 	return r;
1509 
1510 failure:
1511 	mes_v12_0_hw_fini(adev);
1512 	return r;
1513 }
1514 
mes_v12_0_kiq_hw_fini(struct amdgpu_device * adev)1515 static int mes_v12_0_kiq_hw_fini(struct amdgpu_device *adev)
1516 {
1517 	if (adev->mes.ring[0].sched.ready) {
1518 		if (adev->enable_uni_mes)
1519 			amdgpu_mes_unmap_legacy_queue(adev,
1520 				      &adev->mes.ring[AMDGPU_MES_SCHED_PIPE],
1521 				      RESET_QUEUES, 0, 0);
1522 		else
1523 			mes_v12_0_kiq_dequeue_sched(adev);
1524 
1525 		adev->mes.ring[0].sched.ready = false;
1526 	}
1527 
1528 	mes_v12_0_enable(adev, false);
1529 
1530 	return 0;
1531 }
1532 
mes_v12_0_hw_init(void * handle)1533 static int mes_v12_0_hw_init(void *handle)
1534 {
1535 	int r;
1536 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1537 
1538 	if (adev->mes.ring[0].sched.ready)
1539 		goto out;
1540 
1541 	if (!adev->enable_mes_kiq) {
1542 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1543 			r = mes_v12_0_load_microcode(adev,
1544 					     AMDGPU_MES_SCHED_PIPE, true);
1545 			if (r) {
1546 				DRM_ERROR("failed to MES fw, r=%d\n", r);
1547 				return r;
1548 			}
1549 
1550 			mes_v12_0_set_ucode_start_addr(adev);
1551 
1552 		} else if (adev->firmware.load_type ==
1553 			   AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) {
1554 
1555 			mes_v12_0_set_ucode_start_addr(adev);
1556 		}
1557 
1558 		mes_v12_0_enable(adev, true);
1559 	}
1560 
1561 	/* Enable the MES to handle doorbell ring on unmapped queue */
1562 	mes_v12_0_enable_unmapped_doorbell_handling(&adev->mes, true);
1563 
1564 	r = mes_v12_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE);
1565 	if (r)
1566 		goto failure;
1567 
1568 	r = mes_v12_0_set_hw_resources(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1569 	if (r)
1570 		goto failure;
1571 
1572 	if (adev->enable_uni_mes)
1573 		mes_v12_0_set_hw_resources_1(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1574 
1575 	mes_v12_0_init_aggregated_doorbell(&adev->mes);
1576 
1577 	r = mes_v12_0_query_sched_status(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1578 	if (r) {
1579 		DRM_ERROR("MES is busy\n");
1580 		goto failure;
1581 	}
1582 
1583 out:
1584 	/*
1585 	 * Disable KIQ ring usage from the driver once MES is enabled.
1586 	 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1587 	 * with MES enabled.
1588 	 */
1589 	adev->gfx.kiq[0].ring.sched.ready = false;
1590 	adev->mes.ring[0].sched.ready = true;
1591 
1592 	return 0;
1593 
1594 failure:
1595 	mes_v12_0_hw_fini(adev);
1596 	return r;
1597 }
1598 
mes_v12_0_hw_fini(void * handle)1599 static int mes_v12_0_hw_fini(void *handle)
1600 {
1601 	return 0;
1602 }
1603 
mes_v12_0_suspend(void * handle)1604 static int mes_v12_0_suspend(void *handle)
1605 {
1606 	int r;
1607 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1608 
1609 	r = amdgpu_mes_suspend(adev);
1610 	if (r)
1611 		return r;
1612 
1613 	return mes_v12_0_hw_fini(adev);
1614 }
1615 
mes_v12_0_resume(void * handle)1616 static int mes_v12_0_resume(void *handle)
1617 {
1618 	int r;
1619 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1620 
1621 	r = mes_v12_0_hw_init(adev);
1622 	if (r)
1623 		return r;
1624 
1625 	return amdgpu_mes_resume(adev);
1626 }
1627 
mes_v12_0_early_init(void * handle)1628 static int mes_v12_0_early_init(void *handle)
1629 {
1630 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1631 	int pipe, r;
1632 
1633 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1634 		r = amdgpu_mes_init_microcode(adev, pipe);
1635 		if (r)
1636 			return r;
1637 	}
1638 
1639 	return 0;
1640 }
1641 
mes_v12_0_late_init(void * handle)1642 static int mes_v12_0_late_init(void *handle)
1643 {
1644 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1645 
1646 	/* it's only intended for use in mes_self_test case, not for s0ix and reset */
1647 	if (!amdgpu_in_reset(adev) && !adev->in_s0ix && !adev->in_suspend)
1648 		amdgpu_mes_self_test(adev);
1649 
1650 	return 0;
1651 }
1652 
1653 static const struct amd_ip_funcs mes_v12_0_ip_funcs = {
1654 	.name = "mes_v12_0",
1655 	.early_init = mes_v12_0_early_init,
1656 	.late_init = mes_v12_0_late_init,
1657 	.sw_init = mes_v12_0_sw_init,
1658 	.sw_fini = mes_v12_0_sw_fini,
1659 	.hw_init = mes_v12_0_hw_init,
1660 	.hw_fini = mes_v12_0_hw_fini,
1661 	.suspend = mes_v12_0_suspend,
1662 	.resume = mes_v12_0_resume,
1663 };
1664 
1665 const struct amdgpu_ip_block_version mes_v12_0_ip_block = {
1666 	.type = AMD_IP_BLOCK_TYPE_MES,
1667 	.major = 12,
1668 	.minor = 0,
1669 	.rev = 0,
1670 	.funcs = &mes_v12_0_ip_funcs,
1671 };
1672