1 /*
2  * Copyright 2019 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_11_0_0_offset.h"
30 #include "gc/gc_11_0_0_sh_mask.h"
31 #include "gc/gc_11_0_0_default.h"
32 #include "v11_structs.h"
33 #include "mes_v11_api_def.h"
34 
35 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes.bin");
36 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes1.bin");
37 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes.bin");
38 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes1.bin");
39 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes.bin");
40 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes1.bin");
41 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes.bin");
42 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes1.bin");
43 
44 static int mes_v11_0_hw_fini(void *handle);
45 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev);
46 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev);
47 
48 #define MES_EOP_SIZE   2048
49 
50 static void mes_v11_0_ring_set_wptr(struct amdgpu_ring *ring)
51 {
52 	struct amdgpu_device *adev = ring->adev;
53 
54 	if (ring->use_doorbell) {
55 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
56 			     ring->wptr);
57 		WDOORBELL64(ring->doorbell_index, ring->wptr);
58 	} else {
59 		BUG();
60 	}
61 }
62 
63 static u64 mes_v11_0_ring_get_rptr(struct amdgpu_ring *ring)
64 {
65 	return *ring->rptr_cpu_addr;
66 }
67 
68 static u64 mes_v11_0_ring_get_wptr(struct amdgpu_ring *ring)
69 {
70 	u64 wptr;
71 
72 	if (ring->use_doorbell)
73 		wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
74 	else
75 		BUG();
76 	return wptr;
77 }
78 
79 static const struct amdgpu_ring_funcs mes_v11_0_ring_funcs = {
80 	.type = AMDGPU_RING_TYPE_MES,
81 	.align_mask = 1,
82 	.nop = 0,
83 	.support_64bit_ptrs = true,
84 	.get_rptr = mes_v11_0_ring_get_rptr,
85 	.get_wptr = mes_v11_0_ring_get_wptr,
86 	.set_wptr = mes_v11_0_ring_set_wptr,
87 	.insert_nop = amdgpu_ring_insert_nop,
88 };
89 
90 static int mes_v11_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
91 						    void *pkt, int size,
92 						    int api_status_off)
93 {
94 	int ndw = size / 4;
95 	signed long r;
96 	union MESAPI__ADD_QUEUE *x_pkt = pkt;
97 	struct MES_API_STATUS *api_status;
98 	struct amdgpu_device *adev = mes->adev;
99 	struct amdgpu_ring *ring = &mes->ring;
100 	unsigned long flags;
101 	signed long timeout = adev->usec_timeout;
102 
103 	if (amdgpu_emu_mode) {
104 		timeout *= 100;
105 	} else if (amdgpu_sriov_vf(adev)) {
106 		/* Worst case in sriov where all other 15 VF timeout, each VF needs about 600ms */
107 		timeout = 15 * 600 * 1000;
108 	}
109 	BUG_ON(size % 4 != 0);
110 
111 	spin_lock_irqsave(&mes->ring_lock, flags);
112 	if (amdgpu_ring_alloc(ring, ndw)) {
113 		spin_unlock_irqrestore(&mes->ring_lock, flags);
114 		return -ENOMEM;
115 	}
116 
117 	api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
118 	api_status->api_completion_fence_addr = mes->ring.fence_drv.gpu_addr;
119 	api_status->api_completion_fence_value = ++mes->ring.fence_drv.sync_seq;
120 
121 	amdgpu_ring_write_multiple(ring, pkt, ndw);
122 	amdgpu_ring_commit(ring);
123 	spin_unlock_irqrestore(&mes->ring_lock, flags);
124 
125 	DRM_DEBUG("MES msg=%d was emitted\n", x_pkt->header.opcode);
126 
127 	r = amdgpu_fence_wait_polling(ring, ring->fence_drv.sync_seq,
128 		      timeout);
129 	if (r < 1) {
130 		DRM_ERROR("MES failed to response msg=%d\n",
131 			  x_pkt->header.opcode);
132 
133 		while (halt_if_hws_hang)
134 			schedule();
135 
136 		return -ETIMEDOUT;
137 	}
138 
139 	return 0;
140 }
141 
142 static int convert_to_mes_queue_type(int queue_type)
143 {
144 	if (queue_type == AMDGPU_RING_TYPE_GFX)
145 		return MES_QUEUE_TYPE_GFX;
146 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
147 		return MES_QUEUE_TYPE_COMPUTE;
148 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
149 		return MES_QUEUE_TYPE_SDMA;
150 	else
151 		BUG();
152 	return -1;
153 }
154 
155 static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
156 				  struct mes_add_queue_input *input)
157 {
158 	struct amdgpu_device *adev = mes->adev;
159 	union MESAPI__ADD_QUEUE mes_add_queue_pkt;
160 	struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB_0];
161 	uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
162 
163 	memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
164 
165 	mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
166 	mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
167 	mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
168 
169 	mes_add_queue_pkt.process_id = input->process_id;
170 	mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
171 	mes_add_queue_pkt.process_va_start = input->process_va_start;
172 	mes_add_queue_pkt.process_va_end = input->process_va_end;
173 	mes_add_queue_pkt.process_quantum = input->process_quantum;
174 	mes_add_queue_pkt.process_context_addr = input->process_context_addr;
175 	mes_add_queue_pkt.gang_quantum = input->gang_quantum;
176 	mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
177 	mes_add_queue_pkt.inprocess_gang_priority =
178 		input->inprocess_gang_priority;
179 	mes_add_queue_pkt.gang_global_priority_level =
180 		input->gang_global_priority_level;
181 	mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
182 	mes_add_queue_pkt.mqd_addr = input->mqd_addr;
183 
184 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
185 			AMDGPU_MES_API_VERSION_SHIFT) >= 2)
186 		mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr;
187 	else
188 		mes_add_queue_pkt.wptr_addr = input->wptr_addr;
189 
190 	mes_add_queue_pkt.queue_type =
191 		convert_to_mes_queue_type(input->queue_type);
192 	mes_add_queue_pkt.paging = input->paging;
193 	mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
194 	mes_add_queue_pkt.gws_base = input->gws_base;
195 	mes_add_queue_pkt.gws_size = input->gws_size;
196 	mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
197 	mes_add_queue_pkt.tma_addr = input->tma_addr;
198 	mes_add_queue_pkt.is_kfd_process = input->is_kfd_process;
199 	mes_add_queue_pkt.trap_en = 1;
200 
201 	/* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
202 	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
203 	mes_add_queue_pkt.gds_size = input->queue_size;
204 
205 	if (!(((adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 4) &&
206 		  (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 0)) &&
207 		  (adev->ip_versions[GC_HWIP][0] <= IP_VERSION(11, 0, 3))))
208 		mes_add_queue_pkt.trap_en = 1;
209 
210 	/* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
211 	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
212 	mes_add_queue_pkt.gds_size = input->queue_size;
213 
214 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
215 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
216 			offsetof(union MESAPI__ADD_QUEUE, api_status));
217 }
218 
219 static int mes_v11_0_remove_hw_queue(struct amdgpu_mes *mes,
220 				     struct mes_remove_queue_input *input)
221 {
222 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
223 
224 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
225 
226 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
227 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
228 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
229 
230 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
231 	mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
232 
233 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
234 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
235 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
236 }
237 
238 static int mes_v11_0_unmap_legacy_queue(struct amdgpu_mes *mes,
239 			struct mes_unmap_legacy_queue_input *input)
240 {
241 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
242 
243 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
244 
245 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
246 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
247 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
248 
249 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
250 	mes_remove_queue_pkt.gang_context_addr = 0;
251 
252 	mes_remove_queue_pkt.pipe_id = input->pipe_id;
253 	mes_remove_queue_pkt.queue_id = input->queue_id;
254 
255 	if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
256 		mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
257 		mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
258 		mes_remove_queue_pkt.tf_data =
259 			lower_32_bits(input->trail_fence_data);
260 	} else {
261 		mes_remove_queue_pkt.unmap_legacy_queue = 1;
262 		mes_remove_queue_pkt.queue_type =
263 			convert_to_mes_queue_type(input->queue_type);
264 	}
265 
266 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
267 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
268 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
269 }
270 
271 static int mes_v11_0_suspend_gang(struct amdgpu_mes *mes,
272 				  struct mes_suspend_gang_input *input)
273 {
274 	return 0;
275 }
276 
277 static int mes_v11_0_resume_gang(struct amdgpu_mes *mes,
278 				 struct mes_resume_gang_input *input)
279 {
280 	return 0;
281 }
282 
283 static int mes_v11_0_query_sched_status(struct amdgpu_mes *mes)
284 {
285 	union MESAPI__QUERY_MES_STATUS mes_status_pkt;
286 
287 	memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
288 
289 	mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
290 	mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
291 	mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
292 
293 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
294 			&mes_status_pkt, sizeof(mes_status_pkt),
295 			offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
296 }
297 
298 static int mes_v11_0_misc_op(struct amdgpu_mes *mes,
299 			     struct mes_misc_op_input *input)
300 {
301 	union MESAPI__MISC misc_pkt;
302 
303 	memset(&misc_pkt, 0, sizeof(misc_pkt));
304 
305 	misc_pkt.header.type = MES_API_TYPE_SCHEDULER;
306 	misc_pkt.header.opcode = MES_SCH_API_MISC;
307 	misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
308 
309 	switch (input->op) {
310 	case MES_MISC_OP_READ_REG:
311 		misc_pkt.opcode = MESAPI_MISC__READ_REG;
312 		misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset;
313 		misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr;
314 		break;
315 	case MES_MISC_OP_WRITE_REG:
316 		misc_pkt.opcode = MESAPI_MISC__WRITE_REG;
317 		misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset;
318 		misc_pkt.write_reg.reg_value = input->write_reg.reg_value;
319 		break;
320 	case MES_MISC_OP_WRM_REG_WAIT:
321 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
322 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM;
323 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
324 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
325 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
326 		misc_pkt.wait_reg_mem.reg_offset2 = 0;
327 		break;
328 	case MES_MISC_OP_WRM_REG_WR_WAIT:
329 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
330 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG;
331 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
332 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
333 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
334 		misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1;
335 		break;
336 	default:
337 		DRM_ERROR("unsupported misc op (%d) \n", input->op);
338 		return -EINVAL;
339 	}
340 
341 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
342 			&misc_pkt, sizeof(misc_pkt),
343 			offsetof(union MESAPI__MISC, api_status));
344 }
345 
346 static int mes_v11_0_set_hw_resources(struct amdgpu_mes *mes)
347 {
348 	int i;
349 	struct amdgpu_device *adev = mes->adev;
350 	union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
351 
352 	memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
353 
354 	mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
355 	mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
356 	mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
357 
358 	mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
359 	mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
360 	mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
361 	mes_set_hw_res_pkt.paging_vmid = 0;
362 	mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr = mes->sch_ctx_gpu_addr;
363 	mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
364 		mes->query_status_fence_gpu_addr;
365 
366 	for (i = 0; i < MAX_COMPUTE_PIPES; i++)
367 		mes_set_hw_res_pkt.compute_hqd_mask[i] =
368 			mes->compute_hqd_mask[i];
369 
370 	for (i = 0; i < MAX_GFX_PIPES; i++)
371 		mes_set_hw_res_pkt.gfx_hqd_mask[i] = mes->gfx_hqd_mask[i];
372 
373 	for (i = 0; i < MAX_SDMA_PIPES; i++)
374 		mes_set_hw_res_pkt.sdma_hqd_mask[i] = mes->sdma_hqd_mask[i];
375 
376 	for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
377 		mes_set_hw_res_pkt.aggregated_doorbells[i] =
378 			mes->aggregated_doorbells[i];
379 
380 	for (i = 0; i < 5; i++) {
381 		mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
382 		mes_set_hw_res_pkt.mmhub_base[i] =
383 				adev->reg_offset[MMHUB_HWIP][0][i];
384 		mes_set_hw_res_pkt.osssys_base[i] =
385 		adev->reg_offset[OSSSYS_HWIP][0][i];
386 	}
387 
388 	mes_set_hw_res_pkt.disable_reset = 1;
389 	mes_set_hw_res_pkt.disable_mes_log = 1;
390 	mes_set_hw_res_pkt.use_different_vmid_compute = 1;
391 	mes_set_hw_res_pkt.enable_reg_active_poll = 1;
392 	mes_set_hw_res_pkt.oversubscription_timer = 50;
393 
394 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
395 			&mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
396 			offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
397 }
398 
399 static void mes_v11_0_init_aggregated_doorbell(struct amdgpu_mes *mes)
400 {
401 	struct amdgpu_device *adev = mes->adev;
402 	uint32_t data;
403 
404 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1);
405 	data &= ~(CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET_MASK |
406 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_EN_MASK |
407 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_HIT_MASK);
408 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_LOW] <<
409 		CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET__SHIFT;
410 	data |= 1 << CP_MES_DOORBELL_CONTROL1__DOORBELL_EN__SHIFT;
411 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1, data);
412 
413 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2);
414 	data &= ~(CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET_MASK |
415 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_EN_MASK |
416 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_HIT_MASK);
417 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] <<
418 		CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET__SHIFT;
419 	data |= 1 << CP_MES_DOORBELL_CONTROL2__DOORBELL_EN__SHIFT;
420 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2, data);
421 
422 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3);
423 	data &= ~(CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET_MASK |
424 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_EN_MASK |
425 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_HIT_MASK);
426 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_MEDIUM] <<
427 		CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET__SHIFT;
428 	data |= 1 << CP_MES_DOORBELL_CONTROL3__DOORBELL_EN__SHIFT;
429 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3, data);
430 
431 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4);
432 	data &= ~(CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET_MASK |
433 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_EN_MASK |
434 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_HIT_MASK);
435 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_HIGH] <<
436 		CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET__SHIFT;
437 	data |= 1 << CP_MES_DOORBELL_CONTROL4__DOORBELL_EN__SHIFT;
438 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4, data);
439 
440 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5);
441 	data &= ~(CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET_MASK |
442 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_EN_MASK |
443 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_HIT_MASK);
444 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_REALTIME] <<
445 		CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET__SHIFT;
446 	data |= 1 << CP_MES_DOORBELL_CONTROL5__DOORBELL_EN__SHIFT;
447 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5, data);
448 
449 	data = 1 << CP_HQD_GFX_CONTROL__DB_UPDATED_MSG_EN__SHIFT;
450 	WREG32_SOC15(GC, 0, regCP_HQD_GFX_CONTROL, data);
451 }
452 
453 static const struct amdgpu_mes_funcs mes_v11_0_funcs = {
454 	.add_hw_queue = mes_v11_0_add_hw_queue,
455 	.remove_hw_queue = mes_v11_0_remove_hw_queue,
456 	.unmap_legacy_queue = mes_v11_0_unmap_legacy_queue,
457 	.suspend_gang = mes_v11_0_suspend_gang,
458 	.resume_gang = mes_v11_0_resume_gang,
459 	.misc_op = mes_v11_0_misc_op,
460 };
461 
462 static int mes_v11_0_init_microcode(struct amdgpu_device *adev,
463 				    enum admgpu_mes_pipe pipe)
464 {
465 	char fw_name[30];
466 	char ucode_prefix[30];
467 	int err;
468 	const struct mes_firmware_header_v1_0 *mes_hdr;
469 	struct amdgpu_firmware_info *info;
470 
471 	amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix, sizeof(ucode_prefix));
472 
473 	if (pipe == AMDGPU_MES_SCHED_PIPE)
474 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes.bin",
475 			 ucode_prefix);
476 	else
477 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes1.bin",
478 			 ucode_prefix);
479 
480 	err = request_firmware(&adev->mes.fw[pipe], fw_name, adev->dev);
481 	if (err)
482 		return err;
483 
484 	err = amdgpu_ucode_validate(adev->mes.fw[pipe]);
485 	if (err) {
486 		release_firmware(adev->mes.fw[pipe]);
487 		adev->mes.fw[pipe] = NULL;
488 		return err;
489 	}
490 
491 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
492 		adev->mes.fw[pipe]->data;
493 	adev->mes.uc_start_addr[pipe] =
494 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
495 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
496 	adev->mes.data_start_addr[pipe] =
497 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
498 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
499 
500 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
501 		int ucode, ucode_data;
502 
503 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
504 			ucode = AMDGPU_UCODE_ID_CP_MES;
505 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
506 		} else {
507 			ucode = AMDGPU_UCODE_ID_CP_MES1;
508 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
509 		}
510 
511 		info = &adev->firmware.ucode[ucode];
512 		info->ucode_id = ucode;
513 		info->fw = adev->mes.fw[pipe];
514 		adev->firmware.fw_size +=
515 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
516 			      PAGE_SIZE);
517 
518 		info = &adev->firmware.ucode[ucode_data];
519 		info->ucode_id = ucode_data;
520 		info->fw = adev->mes.fw[pipe];
521 		adev->firmware.fw_size +=
522 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
523 			      PAGE_SIZE);
524 	}
525 
526 	return 0;
527 }
528 
529 static void mes_v11_0_free_microcode(struct amdgpu_device *adev,
530 				     enum admgpu_mes_pipe pipe)
531 {
532 	release_firmware(adev->mes.fw[pipe]);
533 	adev->mes.fw[pipe] = NULL;
534 }
535 
536 static int mes_v11_0_allocate_ucode_buffer(struct amdgpu_device *adev,
537 					   enum admgpu_mes_pipe pipe)
538 {
539 	int r;
540 	const struct mes_firmware_header_v1_0 *mes_hdr;
541 	const __le32 *fw_data;
542 	unsigned fw_size;
543 
544 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
545 		adev->mes.fw[pipe]->data;
546 
547 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
548 		   le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
549 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
550 
551 	r = amdgpu_bo_create_reserved(adev, fw_size,
552 				      PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
553 				      &adev->mes.ucode_fw_obj[pipe],
554 				      &adev->mes.ucode_fw_gpu_addr[pipe],
555 				      (void **)&adev->mes.ucode_fw_ptr[pipe]);
556 	if (r) {
557 		dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
558 		return r;
559 	}
560 
561 	memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
562 
563 	amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
564 	amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
565 
566 	return 0;
567 }
568 
569 static int mes_v11_0_allocate_ucode_data_buffer(struct amdgpu_device *adev,
570 						enum admgpu_mes_pipe pipe)
571 {
572 	int r;
573 	const struct mes_firmware_header_v1_0 *mes_hdr;
574 	const __le32 *fw_data;
575 	unsigned fw_size;
576 
577 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
578 		adev->mes.fw[pipe]->data;
579 
580 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
581 		   le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
582 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
583 
584 	r = amdgpu_bo_create_reserved(adev, fw_size,
585 				      64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
586 				      &adev->mes.data_fw_obj[pipe],
587 				      &adev->mes.data_fw_gpu_addr[pipe],
588 				      (void **)&adev->mes.data_fw_ptr[pipe]);
589 	if (r) {
590 		dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
591 		return r;
592 	}
593 
594 	memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
595 
596 	amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
597 	amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
598 
599 	return 0;
600 }
601 
602 static void mes_v11_0_free_ucode_buffers(struct amdgpu_device *adev,
603 					 enum admgpu_mes_pipe pipe)
604 {
605 	amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
606 			      &adev->mes.data_fw_gpu_addr[pipe],
607 			      (void **)&adev->mes.data_fw_ptr[pipe]);
608 
609 	amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
610 			      &adev->mes.ucode_fw_gpu_addr[pipe],
611 			      (void **)&adev->mes.ucode_fw_ptr[pipe]);
612 }
613 
614 static void mes_v11_0_enable(struct amdgpu_device *adev, bool enable)
615 {
616 	uint64_t ucode_addr;
617 	uint32_t pipe, data = 0;
618 
619 	if (enable) {
620 		data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
621 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
622 		data = REG_SET_FIELD(data, CP_MES_CNTL,
623 			     MES_PIPE1_RESET, adev->enable_mes_kiq ? 1 : 0);
624 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
625 
626 		mutex_lock(&adev->srbm_mutex);
627 		for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
628 			if (!adev->enable_mes_kiq &&
629 			    pipe == AMDGPU_MES_KIQ_PIPE)
630 				continue;
631 
632 			soc21_grbm_select(adev, 3, pipe, 0, 0);
633 
634 			ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
635 			WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
636 				     lower_32_bits(ucode_addr));
637 			WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
638 				     upper_32_bits(ucode_addr));
639 		}
640 		soc21_grbm_select(adev, 0, 0, 0, 0);
641 		mutex_unlock(&adev->srbm_mutex);
642 
643 		/* unhalt MES and activate pipe0 */
644 		data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
645 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE,
646 				     adev->enable_mes_kiq ? 1 : 0);
647 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
648 
649 		if (amdgpu_emu_mode)
650 			msleep(100);
651 		else
652 			udelay(50);
653 	} else {
654 		data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
655 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
656 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
657 		data = REG_SET_FIELD(data, CP_MES_CNTL,
658 				     MES_INVALIDATE_ICACHE, 1);
659 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
660 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET,
661 				     adev->enable_mes_kiq ? 1 : 0);
662 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
663 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
664 	}
665 }
666 
667 /* This function is for backdoor MES firmware */
668 static int mes_v11_0_load_microcode(struct amdgpu_device *adev,
669 				    enum admgpu_mes_pipe pipe, bool prime_icache)
670 {
671 	int r;
672 	uint32_t data;
673 	uint64_t ucode_addr;
674 
675 	mes_v11_0_enable(adev, false);
676 
677 	if (!adev->mes.fw[pipe])
678 		return -EINVAL;
679 
680 	r = mes_v11_0_allocate_ucode_buffer(adev, pipe);
681 	if (r)
682 		return r;
683 
684 	r = mes_v11_0_allocate_ucode_data_buffer(adev, pipe);
685 	if (r) {
686 		mes_v11_0_free_ucode_buffers(adev, pipe);
687 		return r;
688 	}
689 
690 	mutex_lock(&adev->srbm_mutex);
691 	/* me=3, pipe=0, queue=0 */
692 	soc21_grbm_select(adev, 3, pipe, 0, 0);
693 
694 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0);
695 
696 	/* set ucode start address */
697 	ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
698 	WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
699 		     lower_32_bits(ucode_addr));
700 	WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
701 		     upper_32_bits(ucode_addr));
702 
703 	/* set ucode fimrware address */
704 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO,
705 		     lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
706 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI,
707 		     upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
708 
709 	/* set ucode instruction cache boundary to 2M-1 */
710 	WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF);
711 
712 	/* set ucode data firmware address */
713 	WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO,
714 		     lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
715 	WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI,
716 		     upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
717 
718 	/* Set 0x3FFFF (256K-1) to CP_MES_MDBOUND_LO */
719 	WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x3FFFF);
720 
721 	if (prime_icache) {
722 		/* invalidate ICACHE */
723 		data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
724 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
725 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
726 		WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
727 
728 		/* prime the ICACHE. */
729 		data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
730 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
731 		WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
732 	}
733 
734 	soc21_grbm_select(adev, 0, 0, 0, 0);
735 	mutex_unlock(&adev->srbm_mutex);
736 
737 	return 0;
738 }
739 
740 static int mes_v11_0_allocate_eop_buf(struct amdgpu_device *adev,
741 				      enum admgpu_mes_pipe pipe)
742 {
743 	int r;
744 	u32 *eop;
745 
746 	r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
747 			      AMDGPU_GEM_DOMAIN_GTT,
748 			      &adev->mes.eop_gpu_obj[pipe],
749 			      &adev->mes.eop_gpu_addr[pipe],
750 			      (void **)&eop);
751 	if (r) {
752 		dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
753 		return r;
754 	}
755 
756 	memset(eop, 0,
757 	       adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
758 
759 	amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
760 	amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
761 
762 	return 0;
763 }
764 
765 static int mes_v11_0_mqd_init(struct amdgpu_ring *ring)
766 {
767 	struct v11_compute_mqd *mqd = ring->mqd_ptr;
768 	uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
769 	uint32_t tmp;
770 
771 	mqd->header = 0xC0310800;
772 	mqd->compute_pipelinestat_enable = 0x00000001;
773 	mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
774 	mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
775 	mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
776 	mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
777 	mqd->compute_misc_reserved = 0x00000007;
778 
779 	eop_base_addr = ring->eop_gpu_addr >> 8;
780 
781 	/* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
782 	tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
783 	tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
784 			(order_base_2(MES_EOP_SIZE / 4) - 1));
785 
786 	mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
787 	mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
788 	mqd->cp_hqd_eop_control = tmp;
789 
790 	/* disable the queue if it's active */
791 	ring->wptr = 0;
792 	mqd->cp_hqd_pq_rptr = 0;
793 	mqd->cp_hqd_pq_wptr_lo = 0;
794 	mqd->cp_hqd_pq_wptr_hi = 0;
795 
796 	/* set the pointer to the MQD */
797 	mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
798 	mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
799 
800 	/* set MQD vmid to 0 */
801 	tmp = regCP_MQD_CONTROL_DEFAULT;
802 	tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
803 	mqd->cp_mqd_control = tmp;
804 
805 	/* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
806 	hqd_gpu_addr = ring->gpu_addr >> 8;
807 	mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
808 	mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
809 
810 	/* set the wb address whether it's enabled or not */
811 	wb_gpu_addr = ring->rptr_gpu_addr;
812 	mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
813 	mqd->cp_hqd_pq_rptr_report_addr_hi =
814 		upper_32_bits(wb_gpu_addr) & 0xffff;
815 
816 	/* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
817 	wb_gpu_addr = ring->wptr_gpu_addr;
818 	mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
819 	mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
820 
821 	/* set up the HQD, this is similar to CP_RB0_CNTL */
822 	tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
823 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
824 			    (order_base_2(ring->ring_size / 4) - 1));
825 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
826 			    ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
827 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
828 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
829 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
830 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
831 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
832 	mqd->cp_hqd_pq_control = tmp;
833 
834 	/* enable doorbell */
835 	tmp = 0;
836 	if (ring->use_doorbell) {
837 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
838 				    DOORBELL_OFFSET, ring->doorbell_index);
839 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
840 				    DOORBELL_EN, 1);
841 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
842 				    DOORBELL_SOURCE, 0);
843 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
844 				    DOORBELL_HIT, 0);
845 	}
846 	else
847 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
848 				    DOORBELL_EN, 0);
849 	mqd->cp_hqd_pq_doorbell_control = tmp;
850 
851 	mqd->cp_hqd_vmid = 0;
852 	/* activate the queue */
853 	mqd->cp_hqd_active = 1;
854 
855 	tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
856 	tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
857 			    PRELOAD_SIZE, 0x55);
858 	mqd->cp_hqd_persistent_state = tmp;
859 
860 	mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT;
861 	mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
862 	mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
863 
864 	return 0;
865 }
866 
867 static void mes_v11_0_queue_init_register(struct amdgpu_ring *ring)
868 {
869 	struct v11_compute_mqd *mqd = ring->mqd_ptr;
870 	struct amdgpu_device *adev = ring->adev;
871 	uint32_t data = 0;
872 
873 	mutex_lock(&adev->srbm_mutex);
874 	soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
875 
876 	/* set CP_HQD_VMID.VMID = 0. */
877 	data = RREG32_SOC15(GC, 0, regCP_HQD_VMID);
878 	data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
879 	WREG32_SOC15(GC, 0, regCP_HQD_VMID, data);
880 
881 	/* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
882 	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
883 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
884 			     DOORBELL_EN, 0);
885 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
886 
887 	/* set CP_MQD_BASE_ADDR/HI with the MQD base address */
888 	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
889 	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
890 
891 	/* set CP_MQD_CONTROL.VMID=0 */
892 	data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL);
893 	data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
894 	WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0);
895 
896 	/* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
897 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
898 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
899 
900 	/* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
901 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
902 		     mqd->cp_hqd_pq_rptr_report_addr_lo);
903 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
904 		     mqd->cp_hqd_pq_rptr_report_addr_hi);
905 
906 	/* set CP_HQD_PQ_CONTROL */
907 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
908 
909 	/* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
910 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
911 		     mqd->cp_hqd_pq_wptr_poll_addr_lo);
912 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
913 		     mqd->cp_hqd_pq_wptr_poll_addr_hi);
914 
915 	/* set CP_HQD_PQ_DOORBELL_CONTROL */
916 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
917 		     mqd->cp_hqd_pq_doorbell_control);
918 
919 	/* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
920 	WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
921 
922 	/* set CP_HQD_ACTIVE.ACTIVE=1 */
923 	WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active);
924 
925 	soc21_grbm_select(adev, 0, 0, 0, 0);
926 	mutex_unlock(&adev->srbm_mutex);
927 }
928 
929 static int mes_v11_0_kiq_enable_queue(struct amdgpu_device *adev)
930 {
931 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
932 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
933 	int r;
934 
935 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
936 		return -EINVAL;
937 
938 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
939 	if (r) {
940 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
941 		return r;
942 	}
943 
944 	kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring);
945 
946 	r = amdgpu_ring_test_ring(kiq_ring);
947 	if (r) {
948 		DRM_ERROR("kfq enable failed\n");
949 		kiq_ring->sched.ready = false;
950 	}
951 	return r;
952 }
953 
954 static int mes_v11_0_queue_init(struct amdgpu_device *adev,
955 				enum admgpu_mes_pipe pipe)
956 {
957 	struct amdgpu_ring *ring;
958 	int r;
959 
960 	if (pipe == AMDGPU_MES_KIQ_PIPE)
961 		ring = &adev->gfx.kiq.ring;
962 	else if (pipe == AMDGPU_MES_SCHED_PIPE)
963 		ring = &adev->mes.ring;
964 	else
965 		BUG();
966 
967 	if ((pipe == AMDGPU_MES_SCHED_PIPE) &&
968 	    (amdgpu_in_reset(adev) || adev->in_suspend)) {
969 		*(ring->wptr_cpu_addr) = 0;
970 		*(ring->rptr_cpu_addr) = 0;
971 		amdgpu_ring_clear_ring(ring);
972 	}
973 
974 	r = mes_v11_0_mqd_init(ring);
975 	if (r)
976 		return r;
977 
978 	if (pipe == AMDGPU_MES_SCHED_PIPE) {
979 		r = mes_v11_0_kiq_enable_queue(adev);
980 		if (r)
981 			return r;
982 	} else {
983 		mes_v11_0_queue_init_register(ring);
984 	}
985 
986 	/* get MES scheduler/KIQ versions */
987 	mutex_lock(&adev->srbm_mutex);
988 	soc21_grbm_select(adev, 3, pipe, 0, 0);
989 
990 	if (pipe == AMDGPU_MES_SCHED_PIPE)
991 		adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
992 	else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
993 		adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
994 
995 	soc21_grbm_select(adev, 0, 0, 0, 0);
996 	mutex_unlock(&adev->srbm_mutex);
997 
998 	return 0;
999 }
1000 
1001 static int mes_v11_0_ring_init(struct amdgpu_device *adev)
1002 {
1003 	struct amdgpu_ring *ring;
1004 
1005 	ring = &adev->mes.ring;
1006 
1007 	ring->funcs = &mes_v11_0_ring_funcs;
1008 
1009 	ring->me = 3;
1010 	ring->pipe = 0;
1011 	ring->queue = 0;
1012 
1013 	ring->ring_obj = NULL;
1014 	ring->use_doorbell = true;
1015 	ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
1016 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_SCHED_PIPE];
1017 	ring->no_scheduler = true;
1018 	sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
1019 
1020 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1021 				AMDGPU_RING_PRIO_DEFAULT, NULL);
1022 }
1023 
1024 static int mes_v11_0_kiq_ring_init(struct amdgpu_device *adev)
1025 {
1026 	struct amdgpu_ring *ring;
1027 
1028 	spin_lock_init(&adev->gfx.kiq.ring_lock);
1029 
1030 	ring = &adev->gfx.kiq.ring;
1031 
1032 	ring->me = 3;
1033 	ring->pipe = 1;
1034 	ring->queue = 0;
1035 
1036 	ring->adev = NULL;
1037 	ring->ring_obj = NULL;
1038 	ring->use_doorbell = true;
1039 	ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
1040 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
1041 	ring->no_scheduler = true;
1042 	sprintf(ring->name, "mes_kiq_%d.%d.%d",
1043 		ring->me, ring->pipe, ring->queue);
1044 
1045 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1046 				AMDGPU_RING_PRIO_DEFAULT, NULL);
1047 }
1048 
1049 static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev,
1050 				 enum admgpu_mes_pipe pipe)
1051 {
1052 	int r, mqd_size = sizeof(struct v11_compute_mqd);
1053 	struct amdgpu_ring *ring;
1054 
1055 	if (pipe == AMDGPU_MES_KIQ_PIPE)
1056 		ring = &adev->gfx.kiq.ring;
1057 	else if (pipe == AMDGPU_MES_SCHED_PIPE)
1058 		ring = &adev->mes.ring;
1059 	else
1060 		BUG();
1061 
1062 	if (ring->mqd_obj)
1063 		return 0;
1064 
1065 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
1066 				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
1067 				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
1068 	if (r) {
1069 		dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
1070 		return r;
1071 	}
1072 
1073 	memset(ring->mqd_ptr, 0, mqd_size);
1074 
1075 	/* prepare MQD backup */
1076 	adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
1077 	if (!adev->mes.mqd_backup[pipe])
1078 		dev_warn(adev->dev,
1079 			 "no memory to create MQD backup for ring %s\n",
1080 			 ring->name);
1081 
1082 	return 0;
1083 }
1084 
1085 static int mes_v11_0_sw_init(void *handle)
1086 {
1087 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1088 	int pipe, r;
1089 
1090 	adev->mes.adev = adev;
1091 	adev->mes.funcs = &mes_v11_0_funcs;
1092 	adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init;
1093 	adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini;
1094 
1095 	r = amdgpu_mes_init(adev);
1096 	if (r)
1097 		return r;
1098 
1099 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1100 		if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
1101 			continue;
1102 
1103 		r = mes_v11_0_init_microcode(adev, pipe);
1104 		if (r)
1105 			return r;
1106 
1107 		r = mes_v11_0_allocate_eop_buf(adev, pipe);
1108 		if (r)
1109 			return r;
1110 
1111 		r = mes_v11_0_mqd_sw_init(adev, pipe);
1112 		if (r)
1113 			return r;
1114 	}
1115 
1116 	if (adev->enable_mes_kiq) {
1117 		r = mes_v11_0_kiq_ring_init(adev);
1118 		if (r)
1119 			return r;
1120 	}
1121 
1122 	r = mes_v11_0_ring_init(adev);
1123 	if (r)
1124 		return r;
1125 
1126 	return 0;
1127 }
1128 
1129 static int mes_v11_0_sw_fini(void *handle)
1130 {
1131 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1132 	int pipe;
1133 
1134 	amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
1135 	amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
1136 
1137 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1138 		kfree(adev->mes.mqd_backup[pipe]);
1139 
1140 		amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1141 				      &adev->mes.eop_gpu_addr[pipe],
1142 				      NULL);
1143 
1144 		mes_v11_0_free_microcode(adev, pipe);
1145 	}
1146 
1147 	amdgpu_bo_free_kernel(&adev->gfx.kiq.ring.mqd_obj,
1148 			      &adev->gfx.kiq.ring.mqd_gpu_addr,
1149 			      &adev->gfx.kiq.ring.mqd_ptr);
1150 
1151 	amdgpu_bo_free_kernel(&adev->mes.ring.mqd_obj,
1152 			      &adev->mes.ring.mqd_gpu_addr,
1153 			      &adev->mes.ring.mqd_ptr);
1154 
1155 	amdgpu_ring_fini(&adev->gfx.kiq.ring);
1156 	amdgpu_ring_fini(&adev->mes.ring);
1157 
1158 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1159 		mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1160 		mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1161 	}
1162 
1163 	amdgpu_mes_fini(adev);
1164 	return 0;
1165 }
1166 
1167 static void mes_v11_0_kiq_dequeue_sched(struct amdgpu_device *adev)
1168 {
1169 	uint32_t data;
1170 	int i;
1171 
1172 	mutex_lock(&adev->srbm_mutex);
1173 	soc21_grbm_select(adev, 3, AMDGPU_MES_SCHED_PIPE, 0, 0);
1174 
1175 	/* disable the queue if it's active */
1176 	if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
1177 		WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
1178 		for (i = 0; i < adev->usec_timeout; i++) {
1179 			if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
1180 				break;
1181 			udelay(1);
1182 		}
1183 	}
1184 	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1185 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1186 				DOORBELL_EN, 0);
1187 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1188 				DOORBELL_HIT, 1);
1189 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1190 
1191 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0);
1192 
1193 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0);
1194 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0);
1195 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0);
1196 
1197 	soc21_grbm_select(adev, 0, 0, 0, 0);
1198 	mutex_unlock(&adev->srbm_mutex);
1199 
1200 	adev->mes.ring.sched.ready = false;
1201 }
1202 
1203 static void mes_v11_0_kiq_setting(struct amdgpu_ring *ring)
1204 {
1205 	uint32_t tmp;
1206 	struct amdgpu_device *adev = ring->adev;
1207 
1208 	/* tell RLC which is KIQ queue */
1209 	tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1210 	tmp &= 0xffffff00;
1211 	tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1212 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1213 	tmp |= 0x80;
1214 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1215 }
1216 
1217 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev)
1218 {
1219 	int r = 0;
1220 
1221 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1222 
1223 		r = mes_v11_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE, false);
1224 		if (r) {
1225 			DRM_ERROR("failed to load MES fw, r=%d\n", r);
1226 			return r;
1227 		}
1228 
1229 		r = mes_v11_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE, true);
1230 		if (r) {
1231 			DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1232 			return r;
1233 		}
1234 
1235 	}
1236 
1237 	mes_v11_0_enable(adev, true);
1238 
1239 	mes_v11_0_kiq_setting(&adev->gfx.kiq.ring);
1240 
1241 	r = mes_v11_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE);
1242 	if (r)
1243 		goto failure;
1244 
1245 	return r;
1246 
1247 failure:
1248 	mes_v11_0_hw_fini(adev);
1249 	return r;
1250 }
1251 
1252 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev)
1253 {
1254 	if (adev->mes.ring.sched.ready)
1255 		mes_v11_0_kiq_dequeue_sched(adev);
1256 
1257 	if (!amdgpu_sriov_vf(adev))
1258 		mes_v11_0_enable(adev, false);
1259 
1260 	return 0;
1261 }
1262 
1263 static int mes_v11_0_hw_init(void *handle)
1264 {
1265 	int r;
1266 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1267 
1268 	if (!adev->enable_mes_kiq) {
1269 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1270 			r = mes_v11_0_load_microcode(adev,
1271 					     AMDGPU_MES_SCHED_PIPE, true);
1272 			if (r) {
1273 				DRM_ERROR("failed to MES fw, r=%d\n", r);
1274 				return r;
1275 			}
1276 		}
1277 
1278 		mes_v11_0_enable(adev, true);
1279 	}
1280 
1281 	r = mes_v11_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE);
1282 	if (r)
1283 		goto failure;
1284 
1285 	r = mes_v11_0_set_hw_resources(&adev->mes);
1286 	if (r)
1287 		goto failure;
1288 
1289 	mes_v11_0_init_aggregated_doorbell(&adev->mes);
1290 
1291 	r = mes_v11_0_query_sched_status(&adev->mes);
1292 	if (r) {
1293 		DRM_ERROR("MES is busy\n");
1294 		goto failure;
1295 	}
1296 
1297 	/*
1298 	 * Disable KIQ ring usage from the driver once MES is enabled.
1299 	 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1300 	 * with MES enabled.
1301 	 */
1302 	adev->gfx.kiq.ring.sched.ready = false;
1303 	adev->mes.ring.sched.ready = true;
1304 
1305 	return 0;
1306 
1307 failure:
1308 	mes_v11_0_hw_fini(adev);
1309 	return r;
1310 }
1311 
1312 static int mes_v11_0_hw_fini(void *handle)
1313 {
1314 	return 0;
1315 }
1316 
1317 static int mes_v11_0_suspend(void *handle)
1318 {
1319 	int r;
1320 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1321 
1322 	r = amdgpu_mes_suspend(adev);
1323 	if (r)
1324 		return r;
1325 
1326 	return mes_v11_0_hw_fini(adev);
1327 }
1328 
1329 static int mes_v11_0_resume(void *handle)
1330 {
1331 	int r;
1332 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1333 
1334 	r = mes_v11_0_hw_init(adev);
1335 	if (r)
1336 		return r;
1337 
1338 	return amdgpu_mes_resume(adev);
1339 }
1340 
1341 static int mes_v11_0_late_init(void *handle)
1342 {
1343 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1344 
1345 	if (!amdgpu_in_reset(adev) &&
1346 	    (adev->ip_versions[GC_HWIP][0] != IP_VERSION(11, 0, 3)))
1347 		amdgpu_mes_self_test(adev);
1348 
1349 	return 0;
1350 }
1351 
1352 static const struct amd_ip_funcs mes_v11_0_ip_funcs = {
1353 	.name = "mes_v11_0",
1354 	.late_init = mes_v11_0_late_init,
1355 	.sw_init = mes_v11_0_sw_init,
1356 	.sw_fini = mes_v11_0_sw_fini,
1357 	.hw_init = mes_v11_0_hw_init,
1358 	.hw_fini = mes_v11_0_hw_fini,
1359 	.suspend = mes_v11_0_suspend,
1360 	.resume = mes_v11_0_resume,
1361 };
1362 
1363 const struct amdgpu_ip_block_version mes_v11_0_ip_block = {
1364 	.type = AMD_IP_BLOCK_TYPE_MES,
1365 	.major = 11,
1366 	.minor = 0,
1367 	.rev = 0,
1368 	.funcs = &mes_v11_0_ip_funcs,
1369 };
1370