xref: /linux/drivers/gpu/drm/amd/amdgpu/mes_v12_1.c (revision def3488eb0fdb386044aced1a8fb2592b1e68896)
1 /*
2  * Copyright 2025 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 "soc_v1_0.h"
29 #include "gc/gc_12_1_0_offset.h"
30 #include "gc/gc_12_1_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_1_0_mes.bin");
36 MODULE_FIRMWARE("amdgpu/gc_12_1_0_mes1.bin");
37 MODULE_FIRMWARE("amdgpu/gc_12_1_0_uni_mes.bin");
38 
39 static int mes_v12_1_hw_init(struct amdgpu_ip_block *ip_block);
40 static int mes_v12_1_hw_fini(struct amdgpu_ip_block *ip_block);
41 static int mes_v12_1_kiq_hw_init(struct amdgpu_device *adev, uint32_t xcc_id);
42 static int mes_v12_1_kiq_hw_fini(struct amdgpu_device *adev, uint32_t xcc_id);
43 
44 #define MES_EOP_SIZE   2048
45 
46 #define regCP_HQD_IB_CONTROL_MES_12_1_DEFAULT 0x100000
47 
48 static void mes_v12_1_ring_set_wptr(struct amdgpu_ring *ring)
49 {
50 	struct amdgpu_device *adev = ring->adev;
51 
52 	if (ring->use_doorbell) {
53 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
54 			     ring->wptr);
55 		WDOORBELL64(ring->doorbell_index, ring->wptr);
56 	} else {
57 		BUG();
58 	}
59 }
60 
61 static u64 mes_v12_1_ring_get_rptr(struct amdgpu_ring *ring)
62 {
63 	return *ring->rptr_cpu_addr;
64 }
65 
66 static u64 mes_v12_1_ring_get_wptr(struct amdgpu_ring *ring)
67 {
68 	u64 wptr;
69 
70 	if (ring->use_doorbell)
71 		wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
72 	else
73 		BUG();
74 	return wptr;
75 }
76 
77 static const struct amdgpu_ring_funcs mes_v12_1_ring_funcs = {
78 	.type = AMDGPU_RING_TYPE_MES,
79 	.align_mask = 1,
80 	.nop = 0,
81 	.support_64bit_ptrs = true,
82 	.get_rptr = mes_v12_1_ring_get_rptr,
83 	.get_wptr = mes_v12_1_ring_get_wptr,
84 	.set_wptr = mes_v12_1_ring_set_wptr,
85 	.insert_nop = amdgpu_ring_insert_nop,
86 };
87 
88 static const char *mes_v12_1_opcodes[] = {
89 	"SET_HW_RSRC",
90 	"SET_SCHEDULING_CONFIG",
91 	"ADD_QUEUE",
92 	"REMOVE_QUEUE",
93 	"PERFORM_YIELD",
94 	"SET_GANG_PRIORITY_LEVEL",
95 	"SUSPEND",
96 	"RESUME",
97 	"RESET",
98 	"SET_LOG_BUFFER",
99 	"CHANGE_GANG_PRORITY",
100 	"QUERY_SCHEDULER_STATUS",
101 	"unused",
102 	"SET_DEBUG_VMID",
103 	"MISC",
104 	"UPDATE_ROOT_PAGE_TABLE",
105 	"AMD_LOG",
106 	"SET_SE_MODE",
107 	"SET_GANG_SUBMIT",
108 	"SET_HW_RSRC_1",
109 };
110 
111 static const char *mes_v12_1_misc_opcodes[] = {
112 	"WRITE_REG",
113 	"INV_GART",
114 	"QUERY_STATUS",
115 	"READ_REG",
116 	"WAIT_REG_MEM",
117 	"SET_SHADER_DEBUGGER",
118 	"NOTIFY_WORK_ON_UNMAPPED_QUEUE",
119 	"NOTIFY_TO_UNMAP_PROCESSES",
120 };
121 
122 static const char *mes_v12_1_get_op_string(union MESAPI__MISC *x_pkt)
123 {
124 	const char *op_str = NULL;
125 
126 	if (x_pkt->header.opcode < ARRAY_SIZE(mes_v12_1_opcodes))
127 		op_str = mes_v12_1_opcodes[x_pkt->header.opcode];
128 
129 	return op_str;
130 }
131 
132 static const char *mes_v12_1_get_misc_op_string(union MESAPI__MISC *x_pkt)
133 {
134 	const char *op_str = NULL;
135 
136 	if ((x_pkt->header.opcode == MES_SCH_API_MISC) &&
137 	    (x_pkt->opcode < ARRAY_SIZE(mes_v12_1_misc_opcodes)))
138 		op_str = mes_v12_1_misc_opcodes[x_pkt->opcode];
139 
140 	return op_str;
141 }
142 
143 static int mes_v12_1_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
144 					    int xcc_id, int pipe, void *pkt,
145 					    int size, int api_status_off)
146 {
147 	union MESAPI__QUERY_MES_STATUS mes_status_pkt;
148 	signed long timeout = 2100000; /* 2100 ms */
149 	struct amdgpu_device *adev = mes->adev;
150 	struct amdgpu_ring *ring = &mes->ring[MES_PIPE_INST(xcc_id, pipe)];
151 	spinlock_t *ring_lock = &mes->ring_lock[MES_PIPE_INST(xcc_id, pipe)];
152 	struct MES_API_STATUS *api_status;
153 	union MESAPI__MISC *x_pkt = pkt;
154 	const char *op_str, *misc_op_str;
155 	unsigned long flags;
156 	u64 status_gpu_addr;
157 	u32 seq, status_offset;
158 	u64 *status_ptr;
159 	signed long r;
160 	int ret;
161 
162 	if (x_pkt->header.opcode >= MES_SCH_API_MAX)
163 		return -EINVAL;
164 
165 	if (amdgpu_emu_mode) {
166 		timeout *= 1000;
167 	} else if (amdgpu_sriov_vf(adev)) {
168 		/* Worst case in sriov where all other 15 VF timeout, each VF needs about 600ms */
169 		timeout = 15 * 600 * 1000;
170 	}
171 
172 	ret = amdgpu_device_wb_get(adev, &status_offset);
173 	if (ret)
174 		return ret;
175 
176 	status_gpu_addr = adev->wb.gpu_addr + (status_offset * 4);
177 	status_ptr = (u64 *)&adev->wb.wb[status_offset];
178 	*status_ptr = 0;
179 
180 	spin_lock_irqsave(ring_lock, flags);
181 	r = amdgpu_ring_alloc(ring, (size + sizeof(mes_status_pkt)) / 4);
182 	if (r)
183 		goto error_unlock_free;
184 
185 	seq = ++ring->fence_drv.sync_seq;
186 	r = amdgpu_fence_wait_polling(ring,
187 				      seq - ring->fence_drv.num_fences_mask,
188 				      timeout);
189 	if (r < 1)
190 		goto error_undo;
191 
192 	api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
193 	api_status->api_completion_fence_addr = status_gpu_addr;
194 	api_status->api_completion_fence_value = 1;
195 
196 	amdgpu_ring_write_multiple(ring, pkt, size / 4);
197 
198 	memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
199 	mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
200 	mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
201 	mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
202 	mes_status_pkt.api_status.api_completion_fence_addr =
203 		ring->fence_drv.gpu_addr;
204 	mes_status_pkt.api_status.api_completion_fence_value = seq;
205 
206 	amdgpu_ring_write_multiple(ring, &mes_status_pkt,
207 				   sizeof(mes_status_pkt) / 4);
208 
209 	amdgpu_ring_commit(ring);
210 	spin_unlock_irqrestore(ring_lock, flags);
211 
212 	op_str = mes_v12_1_get_op_string(x_pkt);
213 	misc_op_str = mes_v12_1_get_misc_op_string(x_pkt);
214 
215 	if (misc_op_str)
216 		dev_dbg(adev->dev, "MES(%d, %d) msg=%s (%s) was emitted\n",
217 			xcc_id, pipe, op_str, misc_op_str);
218 	else if (op_str)
219 		dev_dbg(adev->dev, "MES(%d, %d) msg=%s was emitted\n",
220 			xcc_id, pipe, op_str);
221 	else
222 		dev_dbg(adev->dev, "MES(%d, %d) msg=%d was emitted\n",
223 			xcc_id, pipe, x_pkt->header.opcode);
224 
225 	r = amdgpu_fence_wait_polling(ring, seq, timeout);
226 	if (r < 1 || !*status_ptr) {
227 		if (misc_op_str)
228 			dev_err(adev->dev,
229 				"MES(%d, %d) failed to respond to msg=%s (%s)\n",
230 				xcc_id, pipe, op_str, misc_op_str);
231 		else if (op_str)
232 			dev_err(adev->dev,
233 				"MES(%d, %d) failed to respond to msg=%s\n",
234 				xcc_id, pipe, op_str);
235 		else
236 			dev_err(adev->dev,
237 				"MES(%d, %d) failed to respond to msg=%d\n",
238 				xcc_id, pipe, x_pkt->header.opcode);
239 
240 		while (halt_if_hws_hang)
241 			schedule();
242 
243 		r = -ETIMEDOUT;
244 		goto error_wb_free;
245 	}
246 
247 	amdgpu_device_wb_free(adev, status_offset);
248 	return 0;
249 
250 error_undo:
251 	dev_err(adev->dev, "MES(%d, %d) ring buffer is full.\n", xcc_id, pipe);
252 	amdgpu_ring_undo(ring);
253 
254 error_unlock_free:
255 	spin_unlock_irqrestore(ring_lock, flags);
256 
257 error_wb_free:
258 	amdgpu_device_wb_free(adev, status_offset);
259 	return r;
260 }
261 
262 static int convert_to_mes_queue_type(int queue_type)
263 {
264 	if (queue_type == AMDGPU_RING_TYPE_GFX)
265 		return MES_QUEUE_TYPE_GFX;
266 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
267 		return MES_QUEUE_TYPE_COMPUTE;
268 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
269 		return MES_QUEUE_TYPE_SDMA;
270 	else if (queue_type == AMDGPU_RING_TYPE_MES)
271 		return MES_QUEUE_TYPE_SCHQ;
272 	else
273 		BUG();
274 	return -1;
275 }
276 
277 static int mes_v12_1_add_hw_queue(struct amdgpu_mes *mes,
278 				  struct mes_add_queue_input *input)
279 {
280 	struct amdgpu_device *adev = mes->adev;
281 	union MESAPI__ADD_QUEUE mes_add_queue_pkt;
282 	struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB(0)];
283 	uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
284 
285 	memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
286 
287 	mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
288 	mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
289 	mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
290 
291 	mes_add_queue_pkt.process_id = input->process_id;
292 	mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
293 	mes_add_queue_pkt.process_va_start = input->process_va_start;
294 	mes_add_queue_pkt.process_va_end = input->process_va_end;
295 	mes_add_queue_pkt.process_quantum = input->process_quantum;
296 	mes_add_queue_pkt.process_context_addr = input->process_context_addr;
297 	mes_add_queue_pkt.gang_quantum = input->gang_quantum;
298 	mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
299 	mes_add_queue_pkt.inprocess_gang_priority =
300 		input->inprocess_gang_priority;
301 	mes_add_queue_pkt.gang_global_priority_level =
302 		input->gang_global_priority_level;
303 	mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
304 	mes_add_queue_pkt.mqd_addr = input->mqd_addr;
305 
306 	mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr;
307 
308 	mes_add_queue_pkt.queue_type =
309 		convert_to_mes_queue_type(input->queue_type);
310 	mes_add_queue_pkt.paging = input->paging;
311 	mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
312 	mes_add_queue_pkt.gws_base = input->gws_base;
313 	mes_add_queue_pkt.gws_size = input->gws_size;
314 	mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
315 	mes_add_queue_pkt.tma_addr = input->tma_addr;
316 	mes_add_queue_pkt.trap_en = input->trap_en;
317 	mes_add_queue_pkt.skip_process_ctx_clear = input->skip_process_ctx_clear;
318 	mes_add_queue_pkt.is_kfd_process = input->is_kfd_process;
319 
320 	/* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
321 	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
322 	mes_add_queue_pkt.gds_size = input->queue_size;
323 
324 	/* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
325 	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
326 	mes_add_queue_pkt.gds_size = input->queue_size;
327 
328 	return mes_v12_1_submit_pkt_and_poll_completion(mes,
329 			input->xcc_id, AMDGPU_MES_SCHED_PIPE,
330 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
331 			offsetof(union MESAPI__ADD_QUEUE, api_status));
332 }
333 
334 static int mes_v12_1_remove_hw_queue(struct amdgpu_mes *mes,
335 				     struct mes_remove_queue_input *input)
336 {
337 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
338 
339 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
340 
341 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
342 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
343 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
344 
345 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
346 	mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
347 
348 	return mes_v12_1_submit_pkt_and_poll_completion(mes,
349 			input->xcc_id, AMDGPU_MES_SCHED_PIPE,
350 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
351 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
352 }
353 
354 static int mes_v12_1_reset_hw_queue(struct amdgpu_mes *mes,
355 				    struct mes_reset_queue_input *input)
356 {
357 	union MESAPI__RESET mes_reset_queue_pkt;
358 	int pipe;
359 
360 	memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
361 
362 	mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
363 	mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
364 	mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
365 
366 	mes_reset_queue_pkt.doorbell_offset = input->doorbell_offset;
367 	/* mes_reset_queue_pkt.gang_context_addr = input->gang_context_addr; */
368 	/*mes_reset_queue_pkt.reset_queue_only = 1;*/
369 
370 	if (mes->adev->enable_uni_mes)
371 		pipe = AMDGPU_MES_KIQ_PIPE;
372 	else
373 		pipe = AMDGPU_MES_SCHED_PIPE;
374 
375 	return mes_v12_1_submit_pkt_and_poll_completion(mes,
376 			input->xcc_id, pipe,
377 			&mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
378 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
379 }
380 
381 static int mes_v12_1_map_legacy_queue(struct amdgpu_mes *mes,
382 				      struct mes_map_legacy_queue_input *input)
383 {
384 	union MESAPI__ADD_QUEUE mes_add_queue_pkt;
385 	int pipe;
386 
387 	memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
388 
389 	mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
390 	mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
391 	mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
392 
393 	mes_add_queue_pkt.pipe_id = input->pipe_id;
394 	mes_add_queue_pkt.queue_id = input->queue_id;
395 	mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
396 	mes_add_queue_pkt.mqd_addr = input->mqd_addr;
397 	mes_add_queue_pkt.wptr_addr = input->wptr_addr;
398 	mes_add_queue_pkt.queue_type =
399 		convert_to_mes_queue_type(input->queue_type);
400 	mes_add_queue_pkt.map_legacy_kq = 1;
401 
402 	if (mes->adev->enable_uni_mes)
403 		pipe = AMDGPU_MES_KIQ_PIPE;
404 	else
405 		pipe = AMDGPU_MES_SCHED_PIPE;
406 
407 	return mes_v12_1_submit_pkt_and_poll_completion(mes,
408 			input->xcc_id, pipe,
409 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
410 			offsetof(union MESAPI__ADD_QUEUE, api_status));
411 }
412 
413 static int mes_v12_1_unmap_legacy_queue(struct amdgpu_mes *mes,
414 			struct mes_unmap_legacy_queue_input *input)
415 {
416 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
417 	int pipe;
418 
419 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
420 
421 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
422 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
423 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
424 
425 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
426 	mes_remove_queue_pkt.gang_context_addr = 0;
427 
428 	mes_remove_queue_pkt.pipe_id = input->pipe_id;
429 	mes_remove_queue_pkt.queue_id = input->queue_id;
430 
431 	if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
432 		mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
433 		mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
434 		mes_remove_queue_pkt.tf_data =
435 			lower_32_bits(input->trail_fence_data);
436 	} else {
437 		mes_remove_queue_pkt.unmap_legacy_queue = 1;
438 		mes_remove_queue_pkt.queue_type =
439 			convert_to_mes_queue_type(input->queue_type);
440 	}
441 
442 	if (mes->adev->enable_uni_mes)
443 		pipe = AMDGPU_MES_KIQ_PIPE;
444 	else
445 		pipe = AMDGPU_MES_SCHED_PIPE;
446 
447 	return mes_v12_1_submit_pkt_and_poll_completion(mes,
448 			input->xcc_id, pipe,
449 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
450 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
451 }
452 
453 static int mes_v12_1_suspend_gang(struct amdgpu_mes *mes,
454 				  struct mes_suspend_gang_input *input)
455 {
456 	return 0;
457 }
458 
459 static int mes_v12_1_resume_gang(struct amdgpu_mes *mes,
460 				 struct mes_resume_gang_input *input)
461 {
462 	return 0;
463 }
464 
465 static int mes_v12_1_query_sched_status(struct amdgpu_mes *mes,
466 					  int pipe, int xcc_id)
467 {
468 	union MESAPI__QUERY_MES_STATUS mes_status_pkt;
469 
470 	memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
471 
472 	mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
473 	mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
474 	mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
475 
476 	return mes_v12_1_submit_pkt_and_poll_completion(mes, xcc_id, pipe,
477 			&mes_status_pkt, sizeof(mes_status_pkt),
478 			offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
479 }
480 
481 static int mes_v12_1_misc_op(struct amdgpu_mes *mes,
482 			     struct mes_misc_op_input *input)
483 {
484 	union MESAPI__MISC misc_pkt;
485 	int pipe;
486 
487 	if (mes->adev->enable_uni_mes)
488 		pipe = AMDGPU_MES_KIQ_PIPE;
489 	else
490 		pipe = AMDGPU_MES_SCHED_PIPE;
491 
492 	memset(&misc_pkt, 0, sizeof(misc_pkt));
493 
494 	misc_pkt.header.type = MES_API_TYPE_SCHEDULER;
495 	misc_pkt.header.opcode = MES_SCH_API_MISC;
496 	misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
497 
498 	switch (input->op) {
499 	case MES_MISC_OP_READ_REG:
500 		misc_pkt.opcode = MESAPI_MISC__READ_REG;
501 		misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset;
502 		misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr;
503 		break;
504 	case MES_MISC_OP_WRITE_REG:
505 		misc_pkt.opcode = MESAPI_MISC__WRITE_REG;
506 		misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset;
507 		misc_pkt.write_reg.reg_value = input->write_reg.reg_value;
508 		break;
509 	case MES_MISC_OP_WRM_REG_WAIT:
510 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
511 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM;
512 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
513 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
514 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
515 		misc_pkt.wait_reg_mem.reg_offset2 = 0;
516 		break;
517 	case MES_MISC_OP_WRM_REG_WR_WAIT:
518 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
519 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG;
520 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
521 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
522 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
523 		misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1;
524 		break;
525 	case MES_MISC_OP_SET_SHADER_DEBUGGER:
526 		pipe = AMDGPU_MES_SCHED_PIPE;
527 		misc_pkt.opcode = MESAPI_MISC__SET_SHADER_DEBUGGER;
528 		misc_pkt.set_shader_debugger.process_context_addr =
529 				input->set_shader_debugger.process_context_addr;
530 		misc_pkt.set_shader_debugger.flags.u32all =
531 				input->set_shader_debugger.flags.u32all;
532 		misc_pkt.set_shader_debugger.spi_gdbg_per_vmid_cntl =
533 				input->set_shader_debugger.spi_gdbg_per_vmid_cntl;
534 		memcpy(misc_pkt.set_shader_debugger.tcp_watch_cntl,
535 				input->set_shader_debugger.tcp_watch_cntl,
536 				sizeof(misc_pkt.set_shader_debugger.tcp_watch_cntl));
537 		misc_pkt.set_shader_debugger.trap_en = input->set_shader_debugger.trap_en;
538 		break;
539 	case MES_MISC_OP_CHANGE_CONFIG:
540 		misc_pkt.opcode = MESAPI_MISC__CHANGE_CONFIG;
541 		misc_pkt.change_config.opcode =
542 			MESAPI_MISC__CHANGE_CONFIG_OPTION_LIMIT_SINGLE_PROCESS;
543 		misc_pkt.change_config.option.bits.limit_single_process =
544 			input->change_config.option.limit_single_process;
545 		break;
546 	default:
547 		DRM_ERROR("unsupported misc op (%d) \n", input->op);
548 		return -EINVAL;
549 	}
550 
551 	return mes_v12_1_submit_pkt_and_poll_completion(mes,
552 			input->xcc_id, pipe,
553 			&misc_pkt, sizeof(misc_pkt),
554 			offsetof(union MESAPI__MISC, api_status));
555 }
556 
557 static int mes_v12_1_set_hw_resources_1(struct amdgpu_mes *mes,
558 					  int pipe, int xcc_id)
559 {
560 	union MESAPI_SET_HW_RESOURCES_1 mes_set_hw_res_1_pkt;
561 
562 	memset(&mes_set_hw_res_1_pkt, 0, sizeof(mes_set_hw_res_1_pkt));
563 
564 	mes_set_hw_res_1_pkt.header.type = MES_API_TYPE_SCHEDULER;
565 	mes_set_hw_res_1_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC_1;
566 	mes_set_hw_res_1_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
567 	mes_set_hw_res_1_pkt.mes_kiq_unmap_timeout = 100;
568 
569 	return mes_v12_1_submit_pkt_and_poll_completion(mes, xcc_id, pipe,
570 			&mes_set_hw_res_1_pkt, sizeof(mes_set_hw_res_1_pkt),
571 			offsetof(union MESAPI_SET_HW_RESOURCES_1, api_status));
572 }
573 
574 static void mes_v12_1_set_gfx_hqd_mask(union MESAPI_SET_HW_RESOURCES *pkt)
575 {
576 	/*
577 	 * GFX V12 has only one GFX pipe, but 8 queues in it.
578 	 * GFX pipe 0 queue 0 is being used by Kernel queue.
579 	 * Set GFX pipe 0 queue 1-7 for MES scheduling
580 	 * mask = 1111 1110b
581 	 */
582 	pkt->gfx_hqd_mask[0] = 0xFE;
583 }
584 
585 static int mes_v12_1_set_hw_resources(struct amdgpu_mes *mes,
586 					int pipe, int xcc_id)
587 {
588 	int i;
589 	struct amdgpu_device *adev = mes->adev;
590 	union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
591 
592 	memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
593 
594 	mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
595 	mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
596 	mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
597 
598 	if (pipe == AMDGPU_MES_SCHED_PIPE) {
599 		mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
600 		mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
601 		mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
602 		mes_set_hw_res_pkt.paging_vmid = 0;
603 
604 		for (i = 0; i < MAX_COMPUTE_PIPES; i++)
605 			mes_set_hw_res_pkt.compute_hqd_mask[i] =
606 				mes->compute_hqd_mask[i];
607 
608 		mes_v12_1_set_gfx_hqd_mask(&mes_set_hw_res_pkt);
609 
610 		for (i = 0; i < MAX_SDMA_PIPES; i++)
611 			mes_set_hw_res_pkt.sdma_hqd_mask[i] =
612 				mes->sdma_hqd_mask[i];
613 
614 		for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
615 			mes_set_hw_res_pkt.aggregated_doorbells[i] =
616 				mes->aggregated_doorbells[i];
617 	}
618 
619 	mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr =
620 		mes->sch_ctx_gpu_addr[pipe];
621 	mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
622 		mes->query_status_fence_gpu_addr[pipe];
623 
624 	for (i = 0; i < 5; i++) {
625 		mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
626 		mes_set_hw_res_pkt.mmhub_base[i] =
627 				adev->reg_offset[MMHUB_HWIP][0][i];
628 		mes_set_hw_res_pkt.osssys_base[i] =
629 		adev->reg_offset[OSSSYS_HWIP][0][i];
630 	}
631 
632 	mes_set_hw_res_pkt.disable_reset = 1;
633 	mes_set_hw_res_pkt.disable_mes_log = 1;
634 	mes_set_hw_res_pkt.use_different_vmid_compute = 1;
635 	mes_set_hw_res_pkt.enable_reg_active_poll = 1;
636 	mes_set_hw_res_pkt.enable_level_process_quantum_check = 1;
637 
638 	/*
639 	 * Keep oversubscribe timer for sdma . When we have unmapped doorbell
640 	 * handling support, other queue will not use the oversubscribe timer.
641 	 * handling  mode - 0: disabled; 1: basic version; 2: basic+ version
642 	 */
643 	mes_set_hw_res_pkt.oversubscription_timer = 50;
644 	mes_set_hw_res_pkt.unmapped_doorbell_handling = 1;
645 
646 	if (amdgpu_mes_log_enable) {
647 		mes_set_hw_res_pkt.enable_mes_event_int_logging = 1;
648 		mes_set_hw_res_pkt.event_intr_history_gpu_mc_ptr =
649 			mes->event_log_gpu_addr + pipe * AMDGPU_MES_LOG_BUFFER_SIZE;
650 	}
651 
652 	if (adev->enforce_isolation[0] == AMDGPU_ENFORCE_ISOLATION_ENABLE)
653 		mes_set_hw_res_pkt.limit_single_process = 1;
654 
655 	return mes_v12_1_submit_pkt_and_poll_completion(mes, xcc_id, pipe,
656 			&mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
657 			offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
658 }
659 
660 static void mes_v12_1_init_aggregated_doorbell(struct amdgpu_mes *mes,
661 						 int xcc_id)
662 {
663 	struct amdgpu_device *adev = mes->adev;
664 	uint32_t data;
665 
666 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL1);
667 	data &= ~(CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET_MASK |
668 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_EN_MASK |
669 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_HIT_MASK);
670 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_LOW] <<
671 		CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET__SHIFT;
672 	data |= 1 << CP_MES_DOORBELL_CONTROL1__DOORBELL_EN__SHIFT;
673 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL1, data);
674 
675 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL2);
676 	data &= ~(CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET_MASK |
677 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_EN_MASK |
678 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_HIT_MASK);
679 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] <<
680 		CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET__SHIFT;
681 	data |= 1 << CP_MES_DOORBELL_CONTROL2__DOORBELL_EN__SHIFT;
682 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL2, data);
683 
684 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL3);
685 	data &= ~(CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET_MASK |
686 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_EN_MASK |
687 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_HIT_MASK);
688 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_MEDIUM] <<
689 		CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET__SHIFT;
690 	data |= 1 << CP_MES_DOORBELL_CONTROL3__DOORBELL_EN__SHIFT;
691 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL3, data);
692 
693 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL4);
694 	data &= ~(CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET_MASK |
695 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_EN_MASK |
696 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_HIT_MASK);
697 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_HIGH] <<
698 		CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET__SHIFT;
699 	data |= 1 << CP_MES_DOORBELL_CONTROL4__DOORBELL_EN__SHIFT;
700 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL4, data);
701 
702 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL5);
703 	data &= ~(CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET_MASK |
704 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_EN_MASK |
705 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_HIT_MASK);
706 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_REALTIME] <<
707 		CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET__SHIFT;
708 	data |= 1 << CP_MES_DOORBELL_CONTROL5__DOORBELL_EN__SHIFT;
709 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_DOORBELL_CONTROL5, data);
710 
711 	data = 1 << CP_HQD_GFX_CONTROL__DB_UPDATED_MSG_EN__SHIFT;
712 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_GFX_CONTROL, data);
713 }
714 
715 
716 static void mes_v12_1_enable_unmapped_doorbell_handling(
717 	struct amdgpu_mes *mes, bool enable, int xcc_id)
718 {
719 	struct amdgpu_device *adev = mes->adev;
720 	uint32_t data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_UNMAPPED_DOORBELL);
721 
722 	/*
723 	 * The default PROC_LSB settng is 0xc which means doorbell
724 	 * addr[16:12] gives the doorbell page number. For kfd, each
725 	 * process will use 2 pages of doorbell, we need to change the
726 	 * setting to 0xd
727 	 */
728 	data &= ~CP_UNMAPPED_DOORBELL__PROC_LSB_MASK;
729 	data |= 0xd <<  CP_UNMAPPED_DOORBELL__PROC_LSB__SHIFT;
730 
731 	data |= (enable ? 1 : 0) << CP_UNMAPPED_DOORBELL__ENABLE__SHIFT;
732 
733 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_UNMAPPED_DOORBELL, data);
734 }
735 
736 #if 0
737 static int mes_v12_1_reset_legacy_queue(struct amdgpu_mes *mes,
738 					struct mes_reset_legacy_queue_input *input)
739 {
740 	union MESAPI__RESET mes_reset_queue_pkt;
741 	int pipe;
742 
743 	memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
744 
745 	mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
746 	mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
747 	mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
748 
749 	mes_reset_queue_pkt.queue_type =
750 		convert_to_mes_queue_type(input->queue_type);
751 
752 	if (mes_reset_queue_pkt.queue_type == MES_QUEUE_TYPE_GFX) {
753 		mes_reset_queue_pkt.reset_legacy_gfx = 1;
754 		mes_reset_queue_pkt.pipe_id_lp = input->pipe_id;
755 		mes_reset_queue_pkt.queue_id_lp = input->queue_id;
756 		mes_reset_queue_pkt.mqd_mc_addr_lp = input->mqd_addr;
757 		mes_reset_queue_pkt.doorbell_offset_lp = input->doorbell_offset;
758 		mes_reset_queue_pkt.wptr_addr_lp = input->wptr_addr;
759 		mes_reset_queue_pkt.vmid_id_lp = input->vmid;
760 	} else {
761 		mes_reset_queue_pkt.reset_queue_only = 1;
762 		mes_reset_queue_pkt.doorbell_offset = input->doorbell_offset;
763 	}
764 
765 	if (mes->adev->enable_uni_mes)
766 		pipe = AMDGPU_MES_KIQ_PIPE;
767 	else
768 		pipe = AMDGPU_MES_SCHED_PIPE;
769 
770 	return mes_v12_1_submit_pkt_and_poll_completion(mes,
771 			input->xcc_id, pipe,
772 			&mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
773 			offsetof(union MESAPI__RESET, api_status));
774 }
775 #endif
776 
777 static const struct amdgpu_mes_funcs mes_v12_1_funcs = {
778 	.add_hw_queue = mes_v12_1_add_hw_queue,
779 	.remove_hw_queue = mes_v12_1_remove_hw_queue,
780 	.map_legacy_queue = mes_v12_1_map_legacy_queue,
781 	.unmap_legacy_queue = mes_v12_1_unmap_legacy_queue,
782 	.suspend_gang = mes_v12_1_suspend_gang,
783 	.resume_gang = mes_v12_1_resume_gang,
784 	.misc_op = mes_v12_1_misc_op,
785 	.reset_hw_queue = mes_v12_1_reset_hw_queue,
786 };
787 
788 static int mes_v12_1_allocate_ucode_buffer(struct amdgpu_device *adev,
789 					     enum amdgpu_mes_pipe pipe,
790 					     int xcc_id)
791 {
792 	int r, inst = MES_PIPE_INST(xcc_id, pipe);
793 	const struct mes_firmware_header_v1_0 *mes_hdr;
794 	const __le32 *fw_data;
795 	unsigned fw_size;
796 
797 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
798 		adev->mes.fw[pipe]->data;
799 
800 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
801 		   le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
802 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
803 
804 	r = amdgpu_bo_create_reserved(adev, fw_size,
805 				      PAGE_SIZE,
806 				      AMDGPU_GEM_DOMAIN_VRAM,
807 				      &adev->mes.ucode_fw_obj[inst],
808 				      &adev->mes.ucode_fw_gpu_addr[inst],
809 				      (void **)&adev->mes.ucode_fw_ptr[inst]);
810 	if (r) {
811 		dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
812 		return r;
813 	}
814 
815 	memcpy(adev->mes.ucode_fw_ptr[inst], fw_data, fw_size);
816 
817 	amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[inst]);
818 	amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[inst]);
819 
820 	return 0;
821 }
822 
823 static int mes_v12_1_allocate_ucode_data_buffer(struct amdgpu_device *adev,
824 						  enum amdgpu_mes_pipe pipe,
825 						  int xcc_id)
826 {
827 	int r, inst = MES_PIPE_INST(xcc_id, pipe);
828 	const struct mes_firmware_header_v1_0 *mes_hdr;
829 	const __le32 *fw_data;
830 	unsigned fw_size;
831 
832 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
833 		adev->mes.fw[pipe]->data;
834 
835 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
836 		   le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
837 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
838 
839 	r = amdgpu_bo_create_reserved(adev, fw_size,
840 				      64 * 1024,
841 				      AMDGPU_GEM_DOMAIN_VRAM,
842 				      &adev->mes.data_fw_obj[inst],
843 				      &adev->mes.data_fw_gpu_addr[inst],
844 				      (void **)&adev->mes.data_fw_ptr[inst]);
845 	if (r) {
846 		dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
847 		return r;
848 	}
849 
850 	memcpy(adev->mes.data_fw_ptr[inst], fw_data, fw_size);
851 
852 	amdgpu_bo_kunmap(adev->mes.data_fw_obj[inst]);
853 	amdgpu_bo_unreserve(adev->mes.data_fw_obj[inst]);
854 
855 	return 0;
856 }
857 
858 static void mes_v12_1_free_ucode_buffers(struct amdgpu_device *adev,
859 					   enum amdgpu_mes_pipe pipe,
860 					   int xcc_id)
861 {
862 	int inst = MES_PIPE_INST(xcc_id, pipe);
863 
864 	amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[inst],
865 			      &adev->mes.data_fw_gpu_addr[inst],
866 			      (void **)&adev->mes.data_fw_ptr[inst]);
867 
868 	amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[inst],
869 			      &adev->mes.ucode_fw_gpu_addr[inst],
870 			      (void **)&adev->mes.ucode_fw_ptr[inst]);
871 }
872 
873 static void mes_v12_1_enable(struct amdgpu_device *adev,
874 			       bool enable, int xcc_id)
875 {
876 	uint64_t ucode_addr;
877 	uint32_t pipe, data = 0;
878 
879 	if (enable) {
880 		data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_CNTL);
881 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
882 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 1);
883 		WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_CNTL, data);
884 
885 		mutex_lock(&adev->srbm_mutex);
886 		for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
887 			soc_v1_0_grbm_select(adev, 3, pipe, 0, 0,
888 					     GET_INST(GC, xcc_id));
889 
890 			ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
891 			WREG32_SOC15(GC, GET_INST(GC, xcc_id),
892 				     regCP_MES_PRGRM_CNTR_START,
893 				     lower_32_bits(ucode_addr));
894 			WREG32_SOC15(GC, GET_INST(GC, xcc_id),
895 				     regCP_MES_PRGRM_CNTR_START_HI,
896 				     upper_32_bits(ucode_addr));
897 		}
898 		soc_v1_0_grbm_select(adev, 0, 0, 0, 0, GET_INST(GC, xcc_id));
899 		mutex_unlock(&adev->srbm_mutex);
900 
901 		/* unhalt MES and activate pipe0 */
902 		data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
903 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 1);
904 		WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_CNTL, data);
905 
906 		if (amdgpu_emu_mode)
907 			msleep(500);
908 		else if (adev->enable_uni_mes)
909 			udelay(500);
910 		else
911 			udelay(50);
912 	} else {
913 		data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_CNTL);
914 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
915 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
916 		data = REG_SET_FIELD(data, CP_MES_CNTL,
917 				     MES_INVALIDATE_ICACHE, 1);
918 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
919 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 1);
920 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
921 		WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_CNTL, data);
922 	}
923 }
924 
925 static void mes_v12_1_set_ucode_start_addr(struct amdgpu_device *adev,
926 					     int xcc_id)
927 {
928 	uint64_t ucode_addr;
929 	int pipe;
930 
931 	mes_v12_1_enable(adev, false, xcc_id);
932 
933 	mutex_lock(&adev->srbm_mutex);
934 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
935 		/* me=3, queue=0 */
936 		soc_v1_0_grbm_select(adev, 3, pipe, 0, 0, GET_INST(GC, xcc_id));
937 
938 		/* set ucode start address */
939 		ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
940 		WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_PRGRM_CNTR_START,
941 				lower_32_bits(ucode_addr));
942 		WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_PRGRM_CNTR_START_HI,
943 				upper_32_bits(ucode_addr));
944 
945 		soc_v1_0_grbm_select(adev, 0, 0, 0, 0, GET_INST(GC, xcc_id));
946 	}
947 	mutex_unlock(&adev->srbm_mutex);
948 }
949 
950 /* This function is for backdoor MES firmware */
951 static int mes_v12_1_load_microcode(struct amdgpu_device *adev,
952 				      enum amdgpu_mes_pipe pipe,
953 				      bool prime_icache, int xcc_id)
954 {
955 	int r, inst = MES_PIPE_INST(xcc_id, pipe);
956 	uint32_t data;
957 
958 	mes_v12_1_enable(adev, false, xcc_id);
959 
960 	if (!adev->mes.fw[pipe])
961 		return -EINVAL;
962 
963 	r = mes_v12_1_allocate_ucode_buffer(adev, pipe, xcc_id);
964 	if (r)
965 		return r;
966 
967 	r = mes_v12_1_allocate_ucode_data_buffer(adev, pipe, xcc_id);
968 	if (r) {
969 		mes_v12_1_free_ucode_buffers(adev, pipe, xcc_id);
970 		return r;
971 	}
972 
973 	mutex_lock(&adev->srbm_mutex);
974 	/* me=3, pipe=0, queue=0 */
975 	soc_v1_0_grbm_select(adev, 3, pipe, 0, 0, GET_INST(GC, xcc_id));
976 
977 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_IC_BASE_CNTL, 0);
978 
979 	/* set ucode fimrware address */
980 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_IC_BASE_LO,
981 		     lower_32_bits(adev->mes.ucode_fw_gpu_addr[inst]));
982 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_IC_BASE_HI,
983 		     upper_32_bits(adev->mes.ucode_fw_gpu_addr[inst]));
984 
985 	/* set ucode instruction cache boundary to 2M-1 */
986 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_MIBOUND_LO, 0x1FFFFF);
987 
988 	/* set ucode data firmware address */
989 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_MDBASE_LO,
990 		     lower_32_bits(adev->mes.data_fw_gpu_addr[inst]));
991 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_MDBASE_HI,
992 		     upper_32_bits(adev->mes.data_fw_gpu_addr[inst]));
993 
994 	/* Set data cache boundary CP_MES_MDBOUND_LO */
995 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_MDBOUND_LO, 0x7FFFF);
996 
997 	if (prime_icache) {
998 		/* invalidate ICACHE */
999 		data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_IC_OP_CNTL);
1000 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
1001 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
1002 		WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_IC_OP_CNTL, data);
1003 
1004 		/* prime the ICACHE. */
1005 		data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_IC_OP_CNTL);
1006 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
1007 		WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_IC_OP_CNTL, data);
1008 	}
1009 
1010 	soc_v1_0_grbm_select(adev, 0, 0, 0, 0, GET_INST(GC, xcc_id));
1011 	mutex_unlock(&adev->srbm_mutex);
1012 
1013 	return 0;
1014 }
1015 
1016 static int mes_v12_1_allocate_eop_buf(struct amdgpu_device *adev,
1017 					enum amdgpu_mes_pipe pipe,
1018 					int xcc_id)
1019 {
1020 	int r, inst = MES_PIPE_INST(xcc_id, pipe);
1021 	u32 *eop;
1022 
1023 	r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
1024 			      AMDGPU_GEM_DOMAIN_GTT,
1025 			      &adev->mes.eop_gpu_obj[inst],
1026 			      &adev->mes.eop_gpu_addr[inst],
1027 			      (void **)&eop);
1028 	if (r) {
1029 		dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
1030 		return r;
1031 	}
1032 
1033 	memset(eop, 0,
1034 	       adev->mes.eop_gpu_obj[inst]->tbo.base.size);
1035 
1036 	amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[inst]);
1037 	amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[inst]);
1038 
1039 	return 0;
1040 }
1041 
1042 static int mes_v12_1_mqd_init(struct amdgpu_ring *ring)
1043 {
1044 	struct v12_1_mes_mqd *mqd = ring->mqd_ptr;
1045 	uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
1046 	uint32_t tmp;
1047 
1048 	mqd->header = 0xC0310800;
1049 	mqd->compute_pipelinestat_enable = 0x00000001;
1050 	mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
1051 	mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
1052 	mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
1053 	mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
1054 	mqd->compute_misc_reserved = 0x00000007;
1055 
1056 	eop_base_addr = ring->eop_gpu_addr >> 8;
1057 
1058 	/* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
1059 	tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
1060 	tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
1061 			(order_base_2(MES_EOP_SIZE / 4) - 1));
1062 
1063 	mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
1064 	mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
1065 	mqd->cp_hqd_eop_control = tmp;
1066 
1067 	/* disable the queue if it's active */
1068 	ring->wptr = 0;
1069 	mqd->cp_hqd_pq_rptr = 0;
1070 	mqd->cp_hqd_pq_wptr_lo = 0;
1071 	mqd->cp_hqd_pq_wptr_hi = 0;
1072 
1073 	/* set the pointer to the MQD */
1074 	mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
1075 	mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
1076 
1077 	/* set MQD vmid to 0 */
1078 	tmp = regCP_MQD_CONTROL_DEFAULT;
1079 	tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
1080 	mqd->cp_mqd_control = tmp;
1081 
1082 	/* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
1083 	hqd_gpu_addr = ring->gpu_addr >> 8;
1084 	mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
1085 	mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
1086 
1087 	/* set the wb address whether it's enabled or not */
1088 	wb_gpu_addr = ring->rptr_gpu_addr;
1089 	mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
1090 	mqd->cp_hqd_pq_rptr_report_addr_hi =
1091 		upper_32_bits(wb_gpu_addr) & 0xffff;
1092 
1093 	/* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
1094 	wb_gpu_addr = ring->wptr_gpu_addr;
1095 	mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
1096 	mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
1097 
1098 	/* set up the HQD, this is similar to CP_RB0_CNTL */
1099 	tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
1100 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
1101 			    (order_base_2(ring->ring_size / 4) - 1));
1102 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
1103 			    ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
1104 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
1105 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
1106 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
1107 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
1108 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
1109 	mqd->cp_hqd_pq_control = tmp;
1110 
1111 	/* enable doorbell */
1112 	tmp = 0;
1113 	if (ring->use_doorbell) {
1114 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1115 				    DOORBELL_OFFSET, ring->doorbell_index);
1116 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1117 				    DOORBELL_EN, 1);
1118 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1119 				    DOORBELL_SOURCE, 0);
1120 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1121 				    DOORBELL_HIT, 0);
1122 	} else {
1123 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1124 				    DOORBELL_EN, 0);
1125 	}
1126 	mqd->cp_hqd_pq_doorbell_control = tmp;
1127 
1128 	mqd->cp_hqd_vmid = 0;
1129 	/* activate the queue */
1130 	mqd->cp_hqd_active = 1;
1131 
1132 	tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
1133 	tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
1134 			    PRELOAD_SIZE, 0x63);
1135 	mqd->cp_hqd_persistent_state = tmp;
1136 
1137 	mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_MES_12_1_DEFAULT;
1138 	mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
1139 	mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
1140 
1141 	/*
1142 	 * Set CP_HQD_GFX_CONTROL.DB_UPDATED_MSG_EN[15] to enable unmapped
1143 	 * doorbell handling. This is a reserved CP internal register can
1144 	 * not be accesss by others
1145 	 */
1146 	//mqd->cp_hqd_gfx_control = BIT(15);
1147 
1148 	return 0;
1149 }
1150 
1151 static void mes_v12_1_queue_init_register(struct amdgpu_ring *ring,
1152 					    int xcc_id)
1153 {
1154 	struct v12_1_mes_mqd *mqd = ring->mqd_ptr;
1155 	struct amdgpu_device *adev = ring->adev;
1156 	uint32_t data = 0;
1157 
1158 	mutex_lock(&adev->srbm_mutex);
1159 	soc_v1_0_grbm_select(adev, 3, ring->pipe, 0, 0, GET_INST(GC, xcc_id));
1160 
1161 	/* set CP_HQD_VMID.VMID = 0. */
1162 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_VMID);
1163 	data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
1164 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_VMID, data);
1165 
1166 	/* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
1167 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_DOORBELL_CONTROL);
1168 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1169 			     DOORBELL_EN, 0);
1170 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_DOORBELL_CONTROL, data);
1171 
1172 	/* set CP_MQD_BASE_ADDR/HI with the MQD base address */
1173 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
1174 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
1175 
1176 	/* set CP_MQD_CONTROL.VMID=0 */
1177 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MQD_CONTROL);
1178 	data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
1179 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MQD_CONTROL, 0);
1180 
1181 	/* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
1182 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
1183 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
1184 
1185 	/* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
1186 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_RPTR_REPORT_ADDR,
1187 		     mqd->cp_hqd_pq_rptr_report_addr_lo);
1188 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
1189 		     mqd->cp_hqd_pq_rptr_report_addr_hi);
1190 
1191 	/* set CP_HQD_PQ_CONTROL */
1192 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
1193 
1194 	/* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
1195 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_WPTR_POLL_ADDR,
1196 		     mqd->cp_hqd_pq_wptr_poll_addr_lo);
1197 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
1198 		     mqd->cp_hqd_pq_wptr_poll_addr_hi);
1199 
1200 	/* set CP_HQD_PQ_DOORBELL_CONTROL */
1201 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_DOORBELL_CONTROL,
1202 		     mqd->cp_hqd_pq_doorbell_control);
1203 
1204 	/* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
1205 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
1206 
1207 	/* set CP_HQD_ACTIVE.ACTIVE=1 */
1208 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_ACTIVE, mqd->cp_hqd_active);
1209 
1210 	soc_v1_0_grbm_select(adev, 0, 0, 0, 0, GET_INST(GC, xcc_id));
1211 	mutex_unlock(&adev->srbm_mutex);
1212 }
1213 
1214 static int mes_v12_1_kiq_enable_queue(struct amdgpu_device *adev, int xcc_id)
1215 {
1216 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
1217 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[xcc_id].ring;
1218 	int r, inst = MES_PIPE_INST(xcc_id, AMDGPU_MES_SCHED_PIPE);
1219 
1220 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
1221 		return -EINVAL;
1222 
1223 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
1224 	if (r) {
1225 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
1226 		return r;
1227 	}
1228 
1229 	kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring[inst]);
1230 
1231 	r = amdgpu_ring_test_ring(kiq_ring);
1232 	if (r) {
1233 		DRM_ERROR("kfq enable failed\n");
1234 		kiq_ring->sched.ready = false;
1235 	}
1236 	return r;
1237 }
1238 
1239 static int mes_v12_1_queue_init(struct amdgpu_device *adev,
1240 				  enum amdgpu_mes_pipe pipe,
1241 				  int xcc_id)
1242 {
1243 	struct amdgpu_ring *ring;
1244 	int r;
1245 
1246 	if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1247 		ring = &adev->gfx.kiq[xcc_id].ring;
1248 	else
1249 		ring = &adev->mes.ring[MES_PIPE_INST(xcc_id, pipe)];
1250 
1251 	if ((adev->enable_uni_mes || pipe == AMDGPU_MES_SCHED_PIPE) &&
1252 	    (amdgpu_in_reset(adev) || adev->in_suspend)) {
1253 		*(ring->wptr_cpu_addr) = 0;
1254 		*(ring->rptr_cpu_addr) = 0;
1255 		amdgpu_ring_clear_ring(ring);
1256 	}
1257 
1258 	r = mes_v12_1_mqd_init(ring);
1259 	if (r)
1260 		return r;
1261 
1262 	if (pipe == AMDGPU_MES_SCHED_PIPE) {
1263 		if (adev->enable_uni_mes)
1264 			r = amdgpu_mes_map_legacy_queue(adev, ring, xcc_id);
1265 		else
1266 			r = mes_v12_1_kiq_enable_queue(adev, xcc_id);
1267 		if (r)
1268 			return r;
1269 	} else {
1270 		mes_v12_1_queue_init_register(ring, xcc_id);
1271 	}
1272 
1273 	/* get MES scheduler/KIQ versions */
1274 	mutex_lock(&adev->srbm_mutex);
1275 	soc_v1_0_grbm_select(adev, 3, pipe, 0, 0, GET_INST(GC, xcc_id));
1276 
1277 	if (pipe == AMDGPU_MES_SCHED_PIPE)
1278 		adev->mes.sched_version = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_GP3_LO);
1279 	else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
1280 		adev->mes.kiq_version = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_MES_GP3_LO);
1281 
1282 	soc_v1_0_grbm_select(adev, 0, 0, 0, 0, GET_INST(GC, xcc_id));
1283 	mutex_unlock(&adev->srbm_mutex);
1284 
1285 	return 0;
1286 }
1287 
1288 static int mes_v12_1_ring_init(struct amdgpu_device *adev,
1289 				 int xcc_id, int pipe)
1290 {
1291 	struct amdgpu_ring *ring;
1292 	int inst = MES_PIPE_INST(xcc_id, pipe);
1293 
1294 	ring = &adev->mes.ring[inst];
1295 
1296 	ring->funcs = &mes_v12_1_ring_funcs;
1297 
1298 	ring->me = 3;
1299 	ring->pipe = pipe;
1300 	ring->queue = 0;
1301 	ring->xcc_id = xcc_id;
1302 	ring->vm_hub = AMDGPU_GFXHUB(xcc_id);
1303 
1304 	ring->ring_obj = NULL;
1305 	ring->use_doorbell = true;
1306 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[inst];
1307 	ring->no_scheduler = true;
1308 	snprintf(ring->name, sizeof(ring->name), "mes_%hhu.%hhu.%hhu.%hhu",
1309 		 (unsigned char)xcc_id, (unsigned char)ring->me,
1310 		 (unsigned char)ring->pipe, (unsigned char)ring->queue);
1311 
1312 	if (pipe == AMDGPU_MES_SCHED_PIPE)
1313 		ring->doorbell_index =
1314 			(adev->doorbell_index.mes_ring0 +
1315 			 xcc_id * adev->doorbell_index.xcc_doorbell_range)
1316 			<< 1;
1317 	else
1318 		ring->doorbell_index =
1319 			(adev->doorbell_index.mes_ring1 +
1320 			 xcc_id * adev->doorbell_index.xcc_doorbell_range)
1321 			<< 1;
1322 
1323 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1324 				AMDGPU_RING_PRIO_DEFAULT, NULL);
1325 }
1326 
1327 static int mes_v12_1_kiq_ring_init(struct amdgpu_device *adev, int xcc_id)
1328 {
1329 	struct amdgpu_ring *ring;
1330 	int inst = MES_PIPE_INST(xcc_id, AMDGPU_MES_KIQ_PIPE);
1331 
1332 	spin_lock_init(&adev->gfx.kiq[xcc_id].ring_lock);
1333 
1334 	ring = &adev->gfx.kiq[xcc_id].ring;
1335 
1336 	ring->me = 3;
1337 	ring->pipe = 1;
1338 	ring->queue = 0;
1339 	ring->xcc_id = xcc_id;
1340 	ring->vm_hub = AMDGPU_GFXHUB(xcc_id);
1341 
1342 	ring->adev = NULL;
1343 	ring->ring_obj = NULL;
1344 	ring->use_doorbell = true;
1345 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[inst];
1346 	ring->no_scheduler = true;
1347 	ring->doorbell_index =
1348 		(adev->doorbell_index.mes_ring1 +
1349 		 xcc_id * adev->doorbell_index.xcc_doorbell_range)
1350 		<< 1;
1351 
1352 	snprintf(ring->name, sizeof(ring->name), "mes_kiq_%hhu.%hhu.%hhu.%hhu",
1353 		 (unsigned char)xcc_id, (unsigned char)ring->me,
1354 		 (unsigned char)ring->pipe, (unsigned char)ring->queue);
1355 
1356 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1357 				AMDGPU_RING_PRIO_DEFAULT, NULL);
1358 }
1359 
1360 static int mes_v12_1_mqd_sw_init(struct amdgpu_device *adev,
1361 				   enum amdgpu_mes_pipe pipe,
1362 				   int xcc_id)
1363 {
1364 	int r, mqd_size = sizeof(struct v12_1_mes_mqd);
1365 	struct amdgpu_ring *ring;
1366 	int inst = MES_PIPE_INST(xcc_id, pipe);
1367 
1368 	if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1369 		ring = &adev->gfx.kiq[xcc_id].ring;
1370 	else
1371 		ring = &adev->mes.ring[inst];
1372 
1373 	if (ring->mqd_obj)
1374 		return 0;
1375 
1376 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
1377 				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
1378 				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
1379 	if (r) {
1380 		dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
1381 		return r;
1382 	}
1383 
1384 	memset(ring->mqd_ptr, 0, mqd_size);
1385 
1386 	/* prepare MQD backup */
1387 	adev->mes.mqd_backup[inst] = kmalloc(mqd_size, GFP_KERNEL);
1388 	if (!adev->mes.mqd_backup[inst])
1389 		dev_warn(adev->dev,
1390 			 "no memory to create MQD backup for ring %s\n",
1391 			 ring->name);
1392 
1393 	return 0;
1394 }
1395 
1396 static int mes_v12_1_sw_init(struct amdgpu_ip_block *ip_block)
1397 {
1398 	struct amdgpu_device *adev = ip_block->adev;
1399 	int pipe, r, xcc_id, num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1400 
1401 	adev->mes.funcs = &mes_v12_1_funcs;
1402 	adev->mes.kiq_hw_init = &mes_v12_1_kiq_hw_init;
1403 	adev->mes.kiq_hw_fini = &mes_v12_1_kiq_hw_fini;
1404 	adev->mes.enable_legacy_queue_map = true;
1405 
1406 	adev->mes.event_log_size =
1407 		adev->enable_uni_mes ? (AMDGPU_MAX_MES_PIPES * AMDGPU_MES_LOG_BUFFER_SIZE * num_xcc) : AMDGPU_MES_LOG_BUFFER_SIZE;
1408 
1409 	r = amdgpu_mes_init(adev);
1410 	if (r)
1411 		return r;
1412 
1413 	for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
1414 		for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1415 			r = mes_v12_1_allocate_eop_buf(adev, pipe, xcc_id);
1416 			if (r)
1417 				return r;
1418 
1419 			r = mes_v12_1_mqd_sw_init(adev, pipe, xcc_id);
1420 			if (r)
1421 				return r;
1422 
1423 			if (!adev->enable_uni_mes && pipe ==
1424 			    AMDGPU_MES_KIQ_PIPE)
1425 				r = mes_v12_1_kiq_ring_init(adev, xcc_id);
1426 			else
1427 				r = mes_v12_1_ring_init(adev, xcc_id, pipe);
1428 			if (r)
1429 				return r;
1430 		}
1431 	}
1432 
1433 	return 0;
1434 }
1435 
1436 static int mes_v12_1_sw_fini(struct amdgpu_ip_block *ip_block)
1437 {
1438 	struct amdgpu_device *adev = ip_block->adev;
1439 	int pipe, inst, xcc_id, num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1440 
1441 	for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
1442 		for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1443 			inst = MES_PIPE_INST(xcc_id, pipe);
1444 
1445 			kfree(adev->mes.mqd_backup[inst]);
1446 
1447 			amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[inst],
1448 					      &adev->mes.eop_gpu_addr[inst],
1449 					      NULL);
1450 			amdgpu_ucode_release(&adev->mes.fw[inst]);
1451 
1452 			if (adev->enable_uni_mes || pipe == AMDGPU_MES_SCHED_PIPE) {
1453 				amdgpu_bo_free_kernel(&adev->mes.ring[inst].mqd_obj,
1454 						      &adev->mes.ring[inst].mqd_gpu_addr,
1455 						      &adev->mes.ring[inst].mqd_ptr);
1456 				amdgpu_ring_fini(&adev->mes.ring[inst]);
1457 			}
1458 		}
1459 	}
1460 
1461 	for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
1462 		if (!adev->enable_uni_mes) {
1463 			amdgpu_bo_free_kernel(&adev->gfx.kiq[xcc_id].ring.mqd_obj,
1464 				      &adev->gfx.kiq[xcc_id].ring.mqd_gpu_addr,
1465 				      &adev->gfx.kiq[xcc_id].ring.mqd_ptr);
1466 			amdgpu_ring_fini(&adev->gfx.kiq[xcc_id].ring);
1467 		}
1468 
1469 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1470 			mes_v12_1_free_ucode_buffers(adev,
1471 				       AMDGPU_MES_KIQ_PIPE, xcc_id);
1472 			mes_v12_1_free_ucode_buffers(adev,
1473 				       AMDGPU_MES_SCHED_PIPE, xcc_id);
1474 		}
1475 	}
1476 
1477 	amdgpu_mes_fini(adev);
1478 	return 0;
1479 }
1480 
1481 static void mes_v12_1_kiq_dequeue_sched(struct amdgpu_device *adev,
1482 					  int xcc_id)
1483 {
1484 	uint32_t data;
1485 	int i;
1486 
1487 	mutex_lock(&adev->srbm_mutex);
1488 	soc_v1_0_grbm_select(adev, 3, AMDGPU_MES_SCHED_PIPE, 0, 0,
1489 			     GET_INST(GC, xcc_id));
1490 
1491 	/* disable the queue if it's active */
1492 	if (RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_ACTIVE) & 1) {
1493 		WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_DEQUEUE_REQUEST, 1);
1494 		for (i = 0; i < adev->usec_timeout; i++) {
1495 			if (!(RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_ACTIVE) & 1))
1496 				break;
1497 			udelay(1);
1498 		}
1499 	}
1500 	data = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_DOORBELL_CONTROL);
1501 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1502 				DOORBELL_EN, 0);
1503 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1504 				DOORBELL_HIT, 1);
1505 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_DOORBELL_CONTROL, data);
1506 
1507 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_DOORBELL_CONTROL, 0);
1508 
1509 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_WPTR_LO, 0);
1510 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_WPTR_HI, 0);
1511 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regCP_HQD_PQ_RPTR, 0);
1512 
1513 	soc_v1_0_grbm_select(adev, 0, 0, 0, 0, GET_INST(GC, xcc_id));
1514 	mutex_unlock(&adev->srbm_mutex);
1515 
1516 	adev->mes.ring[MES_PIPE_INST(xcc_id, 0)].sched.ready = false;
1517 }
1518 
1519 static void mes_v12_1_kiq_setting(struct amdgpu_ring *ring, int xcc_id)
1520 {
1521 	uint32_t tmp;
1522 	struct amdgpu_device *adev = ring->adev;
1523 
1524 	/* tell RLC which is KIQ queue */
1525 	tmp = RREG32_SOC15(GC, GET_INST(GC, xcc_id), regRLC_CP_SCHEDULERS);
1526 	tmp &= 0xffffff00;
1527 	tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1528 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regRLC_CP_SCHEDULERS, tmp);
1529 	tmp |= 0x80;
1530 	WREG32_SOC15(GC, GET_INST(GC, xcc_id), regRLC_CP_SCHEDULERS, tmp);
1531 }
1532 
1533 static int mes_v12_1_kiq_hw_init(struct amdgpu_device *adev, uint32_t xcc_id)
1534 {
1535 	int inst = MES_PIPE_INST(xcc_id, AMDGPU_MES_KIQ_PIPE);
1536 	int r = 0;
1537 	struct amdgpu_ip_block *ip_block;
1538 
1539 	if (adev->enable_uni_mes)
1540 		mes_v12_1_kiq_setting(&adev->mes.ring[inst], xcc_id);
1541 	else
1542 		mes_v12_1_kiq_setting(&adev->gfx.kiq[xcc_id].ring, xcc_id);
1543 
1544 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1545 
1546 		r = mes_v12_1_load_microcode(adev, AMDGPU_MES_SCHED_PIPE,
1547 					       false, xcc_id);
1548 		if (r) {
1549 			DRM_ERROR("failed to load MES fw, r=%d\n", r);
1550 			return r;
1551 		}
1552 
1553 		r = mes_v12_1_load_microcode(adev, AMDGPU_MES_KIQ_PIPE,
1554 					       true, xcc_id);
1555 		if (r) {
1556 			DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1557 			return r;
1558 		}
1559 
1560 		mes_v12_1_set_ucode_start_addr(adev, xcc_id);
1561 
1562 	} else if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO)
1563 		mes_v12_1_set_ucode_start_addr(adev, xcc_id);
1564 
1565 	mes_v12_1_enable(adev, true, xcc_id);
1566 
1567 	ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_MES);
1568 	if (unlikely(!ip_block)) {
1569 		dev_err(adev->dev, "Failed to get MES handle\n");
1570 		return -EINVAL;
1571 	}
1572 
1573 	r = mes_v12_1_queue_init(adev, AMDGPU_MES_KIQ_PIPE, xcc_id);
1574 	if (r)
1575 		goto failure;
1576 
1577 	if (adev->enable_uni_mes) {
1578 		r = mes_v12_1_set_hw_resources(&adev->mes,
1579 						 AMDGPU_MES_KIQ_PIPE, xcc_id);
1580 		if (r)
1581 			goto failure;
1582 
1583 		mes_v12_1_set_hw_resources_1(&adev->mes,
1584 					       AMDGPU_MES_KIQ_PIPE, xcc_id);
1585 	}
1586 
1587 	if (adev->mes.enable_legacy_queue_map) {
1588 		r = mes_v12_1_hw_init(ip_block);
1589 		if (r)
1590 			goto failure;
1591 	}
1592 
1593 	return r;
1594 
1595 failure:
1596 	mes_v12_1_hw_fini(ip_block);
1597 	return r;
1598 }
1599 
1600 static int mes_v12_1_kiq_hw_fini(struct amdgpu_device *adev, uint32_t xcc_id)
1601 {
1602 	int inst = MES_PIPE_INST(xcc_id, AMDGPU_MES_SCHED_PIPE);
1603 
1604 	if (adev->mes.ring[inst].sched.ready) {
1605 		if (adev->enable_uni_mes)
1606 			amdgpu_mes_unmap_legacy_queue(adev,
1607 				      &adev->mes.ring[inst],
1608 				      RESET_QUEUES, 0, 0, xcc_id);
1609 		else
1610 			mes_v12_1_kiq_dequeue_sched(adev, xcc_id);
1611 
1612 		adev->mes.ring[inst].sched.ready = false;
1613 	}
1614 
1615 	mes_v12_1_enable(adev, false, xcc_id);
1616 
1617 	return 0;
1618 }
1619 
1620 static int mes_v12_1_xcc_hw_init(struct amdgpu_ip_block *ip_block, int xcc_id)
1621 {
1622 	int r;
1623 	struct amdgpu_device *adev = ip_block->adev;
1624 
1625 	if (adev->mes.ring[MES_PIPE_INST(xcc_id, 0)].sched.ready)
1626 		goto out;
1627 
1628 	if (!adev->enable_mes_kiq) {
1629 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1630 			r = mes_v12_1_load_microcode(adev,
1631 				       AMDGPU_MES_SCHED_PIPE, true, xcc_id);
1632 			if (r) {
1633 				DRM_ERROR("failed to MES fw, r=%d\n", r);
1634 				return r;
1635 			}
1636 
1637 			mes_v12_1_set_ucode_start_addr(adev, xcc_id);
1638 
1639 		} else if (adev->firmware.load_type ==
1640 			   AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) {
1641 
1642 			mes_v12_1_set_ucode_start_addr(adev, xcc_id);
1643 		}
1644 
1645 		mes_v12_1_enable(adev, true, xcc_id);
1646 	}
1647 
1648 	/* Enable the MES to handle doorbell ring on unmapped queue */
1649 	mes_v12_1_enable_unmapped_doorbell_handling(&adev->mes, true, xcc_id);
1650 
1651 	r = mes_v12_1_queue_init(adev, AMDGPU_MES_SCHED_PIPE, xcc_id);
1652 	if (r)
1653 		goto failure;
1654 
1655 	r = mes_v12_1_set_hw_resources(&adev->mes,
1656 					 AMDGPU_MES_SCHED_PIPE, xcc_id);
1657 	if (r)
1658 		goto failure;
1659 
1660 	if (adev->enable_uni_mes)
1661 		mes_v12_1_set_hw_resources_1(&adev->mes,
1662 					       AMDGPU_MES_SCHED_PIPE, xcc_id);
1663 
1664 	mes_v12_1_init_aggregated_doorbell(&adev->mes, xcc_id);
1665 
1666 	r = mes_v12_1_query_sched_status(&adev->mes,
1667 					   AMDGPU_MES_SCHED_PIPE, xcc_id);
1668 	if (r) {
1669 		DRM_ERROR("MES is busy\n");
1670 		goto failure;
1671 	}
1672 
1673 out:
1674 	/*
1675 	 * Disable KIQ ring usage from the driver once MES is enabled.
1676 	 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1677 	 * with MES enabled.
1678 	 */
1679 	adev->gfx.kiq[xcc_id].ring.sched.ready = false;
1680 	adev->mes.ring[MES_PIPE_INST(xcc_id, 0)].sched.ready = true;
1681 
1682 	return 0;
1683 
1684 failure:
1685 	mes_v12_1_hw_fini(ip_block);
1686 	return r;
1687 }
1688 
1689 static int mes_v12_1_hw_init(struct amdgpu_ip_block *ip_block)
1690 {
1691 	struct amdgpu_device *adev = ip_block->adev;
1692 	int r, xcc_id, num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1693 
1694 	for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
1695 		r = mes_v12_1_xcc_hw_init(ip_block, xcc_id);
1696 		if (r)
1697 			return r;
1698 	}
1699 
1700 	return 0;
1701 }
1702 
1703 static int mes_v12_1_hw_fini(struct amdgpu_ip_block *ip_block)
1704 {
1705 	return 0;
1706 }
1707 
1708 static int mes_v12_1_suspend(struct amdgpu_ip_block *ip_block)
1709 {
1710 	int r;
1711 
1712 	r = amdgpu_mes_suspend(ip_block->adev);
1713 	if (r)
1714 		return r;
1715 
1716 	return mes_v12_1_hw_fini(ip_block);
1717 }
1718 
1719 static int mes_v12_1_resume(struct amdgpu_ip_block *ip_block)
1720 {
1721 	int r;
1722 
1723 	r = mes_v12_1_hw_init(ip_block);
1724 	if (r)
1725 		return r;
1726 
1727 	return amdgpu_mes_resume(ip_block->adev);
1728 }
1729 
1730 static int mes_v12_1_early_init(struct amdgpu_ip_block *ip_block)
1731 {
1732 	struct amdgpu_device *adev = ip_block->adev;
1733 	int pipe, r;
1734 
1735 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1736 		r = amdgpu_mes_init_microcode(adev, pipe);
1737 		if (r)
1738 			return r;
1739 	}
1740 
1741 	return 0;
1742 }
1743 
1744 static const struct amd_ip_funcs mes_v12_1_ip_funcs = {
1745 	.name = "mes_v12_1",
1746 	.early_init = mes_v12_1_early_init,
1747 	.late_init = NULL,
1748 	.sw_init = mes_v12_1_sw_init,
1749 	.sw_fini = mes_v12_1_sw_fini,
1750 	.hw_init = mes_v12_1_hw_init,
1751 	.hw_fini = mes_v12_1_hw_fini,
1752 	.suspend = mes_v12_1_suspend,
1753 	.resume = mes_v12_1_resume,
1754 };
1755 
1756 const struct amdgpu_ip_block_version mes_v12_1_ip_block = {
1757 	.type = AMD_IP_BLOCK_TYPE_MES,
1758 	.major = 12,
1759 	.minor = 1,
1760 	.rev = 0,
1761 	.funcs = &mes_v12_1_ip_funcs,
1762 };
1763