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 
42 static int mes_v11_0_hw_fini(void *handle);
43 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev);
44 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev);
45 
46 #define MES_EOP_SIZE   2048
47 
48 static void mes_v11_0_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_v11_0_ring_get_rptr(struct amdgpu_ring *ring)
62 {
63 	return *ring->rptr_cpu_addr;
64 }
65 
66 static u64 mes_v11_0_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_v11_0_ring_funcs = {
78 	.type = AMDGPU_RING_TYPE_MES,
79 	.align_mask = 1,
80 	.nop = 0,
81 	.support_64bit_ptrs = true,
82 	.get_rptr = mes_v11_0_ring_get_rptr,
83 	.get_wptr = mes_v11_0_ring_get_wptr,
84 	.set_wptr = mes_v11_0_ring_set_wptr,
85 	.insert_nop = amdgpu_ring_insert_nop,
86 };
87 
88 static int mes_v11_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
89 						    void *pkt, int size,
90 						    int api_status_off)
91 {
92 	int ndw = size / 4;
93 	signed long r;
94 	union MESAPI__ADD_QUEUE *x_pkt = pkt;
95 	struct MES_API_STATUS *api_status;
96 	struct amdgpu_device *adev = mes->adev;
97 	struct amdgpu_ring *ring = &mes->ring;
98 	unsigned long flags;
99 
100 	BUG_ON(size % 4 != 0);
101 
102 	spin_lock_irqsave(&mes->ring_lock, flags);
103 	if (amdgpu_ring_alloc(ring, ndw)) {
104 		spin_unlock_irqrestore(&mes->ring_lock, flags);
105 		return -ENOMEM;
106 	}
107 
108 	api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
109 	api_status->api_completion_fence_addr = mes->ring.fence_drv.gpu_addr;
110 	api_status->api_completion_fence_value = ++mes->ring.fence_drv.sync_seq;
111 
112 	amdgpu_ring_write_multiple(ring, pkt, ndw);
113 	amdgpu_ring_commit(ring);
114 	spin_unlock_irqrestore(&mes->ring_lock, flags);
115 
116 	DRM_DEBUG("MES msg=%d was emitted\n", x_pkt->header.opcode);
117 
118 	r = amdgpu_fence_wait_polling(ring, ring->fence_drv.sync_seq,
119 		      adev->usec_timeout * (amdgpu_emu_mode ? 100 : 1));
120 	if (r < 1) {
121 		DRM_ERROR("MES failed to response msg=%d\n",
122 			  x_pkt->header.opcode);
123 		return -ETIMEDOUT;
124 	}
125 
126 	return 0;
127 }
128 
129 static int convert_to_mes_queue_type(int queue_type)
130 {
131 	if (queue_type == AMDGPU_RING_TYPE_GFX)
132 		return MES_QUEUE_TYPE_GFX;
133 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
134 		return MES_QUEUE_TYPE_COMPUTE;
135 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
136 		return MES_QUEUE_TYPE_SDMA;
137 	else
138 		BUG();
139 	return -1;
140 }
141 
142 static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
143 				  struct mes_add_queue_input *input)
144 {
145 	struct amdgpu_device *adev = mes->adev;
146 	union MESAPI__ADD_QUEUE mes_add_queue_pkt;
147 	struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB_0];
148 	uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
149 
150 	memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
151 
152 	mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
153 	mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
154 	mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
155 
156 	mes_add_queue_pkt.process_id = input->process_id;
157 	mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
158 	mes_add_queue_pkt.process_va_start = input->process_va_start;
159 	mes_add_queue_pkt.process_va_end = input->process_va_end;
160 	mes_add_queue_pkt.process_quantum = input->process_quantum;
161 	mes_add_queue_pkt.process_context_addr = input->process_context_addr;
162 	mes_add_queue_pkt.gang_quantum = input->gang_quantum;
163 	mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
164 	mes_add_queue_pkt.inprocess_gang_priority =
165 		input->inprocess_gang_priority;
166 	mes_add_queue_pkt.gang_global_priority_level =
167 		input->gang_global_priority_level;
168 	mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
169 	mes_add_queue_pkt.mqd_addr = input->mqd_addr;
170 
171 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
172 			AMDGPU_MES_API_VERSION_SHIFT) >= 2)
173 		mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr;
174 	else
175 		mes_add_queue_pkt.wptr_addr = input->wptr_addr;
176 
177 	mes_add_queue_pkt.queue_type =
178 		convert_to_mes_queue_type(input->queue_type);
179 	mes_add_queue_pkt.paging = input->paging;
180 	mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
181 	mes_add_queue_pkt.gws_base = input->gws_base;
182 	mes_add_queue_pkt.gws_size = input->gws_size;
183 	mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
184 	mes_add_queue_pkt.tma_addr = input->tma_addr;
185 	mes_add_queue_pkt.is_kfd_process = input->is_kfd_process;
186 
187 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
188 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
189 			offsetof(union MESAPI__ADD_QUEUE, api_status));
190 }
191 
192 static int mes_v11_0_remove_hw_queue(struct amdgpu_mes *mes,
193 				     struct mes_remove_queue_input *input)
194 {
195 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
196 
197 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
198 
199 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
200 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
201 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
202 
203 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
204 	mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
205 
206 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
207 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
208 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
209 }
210 
211 static int mes_v11_0_unmap_legacy_queue(struct amdgpu_mes *mes,
212 			struct mes_unmap_legacy_queue_input *input)
213 {
214 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
215 
216 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
217 
218 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
219 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
220 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
221 
222 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
223 	mes_remove_queue_pkt.gang_context_addr = 0;
224 
225 	mes_remove_queue_pkt.pipe_id = input->pipe_id;
226 	mes_remove_queue_pkt.queue_id = input->queue_id;
227 
228 	if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
229 		mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
230 		mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
231 		mes_remove_queue_pkt.tf_data =
232 			lower_32_bits(input->trail_fence_data);
233 	} else {
234 		mes_remove_queue_pkt.unmap_legacy_queue = 1;
235 		mes_remove_queue_pkt.queue_type =
236 			convert_to_mes_queue_type(input->queue_type);
237 	}
238 
239 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
240 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
241 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
242 }
243 
244 static int mes_v11_0_suspend_gang(struct amdgpu_mes *mes,
245 				  struct mes_suspend_gang_input *input)
246 {
247 	return 0;
248 }
249 
250 static int mes_v11_0_resume_gang(struct amdgpu_mes *mes,
251 				 struct mes_resume_gang_input *input)
252 {
253 	return 0;
254 }
255 
256 static int mes_v11_0_query_sched_status(struct amdgpu_mes *mes)
257 {
258 	union MESAPI__QUERY_MES_STATUS mes_status_pkt;
259 
260 	memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
261 
262 	mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
263 	mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
264 	mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
265 
266 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
267 			&mes_status_pkt, sizeof(mes_status_pkt),
268 			offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
269 }
270 
271 static int mes_v11_0_misc_op(struct amdgpu_mes *mes,
272 			     struct mes_misc_op_input *input)
273 {
274 	union MESAPI__MISC misc_pkt;
275 
276 	memset(&misc_pkt, 0, sizeof(misc_pkt));
277 
278 	misc_pkt.header.type = MES_API_TYPE_SCHEDULER;
279 	misc_pkt.header.opcode = MES_SCH_API_MISC;
280 	misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
281 
282 	switch (input->op) {
283 	case MES_MISC_OP_READ_REG:
284 		misc_pkt.opcode = MESAPI_MISC__READ_REG;
285 		misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset;
286 		misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr;
287 		break;
288 	case MES_MISC_OP_WRITE_REG:
289 		misc_pkt.opcode = MESAPI_MISC__WRITE_REG;
290 		misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset;
291 		misc_pkt.write_reg.reg_value = input->write_reg.reg_value;
292 		break;
293 	case MES_MISC_OP_WRM_REG_WAIT:
294 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
295 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM;
296 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
297 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
298 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
299 		misc_pkt.wait_reg_mem.reg_offset2 = 0;
300 		break;
301 	case MES_MISC_OP_WRM_REG_WR_WAIT:
302 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
303 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG;
304 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
305 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
306 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
307 		misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1;
308 		break;
309 	default:
310 		DRM_ERROR("unsupported misc op (%d) \n", input->op);
311 		return -EINVAL;
312 	}
313 
314 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
315 			&misc_pkt, sizeof(misc_pkt),
316 			offsetof(union MESAPI__MISC, api_status));
317 }
318 
319 static int mes_v11_0_set_hw_resources(struct amdgpu_mes *mes)
320 {
321 	int i;
322 	struct amdgpu_device *adev = mes->adev;
323 	union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
324 
325 	memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
326 
327 	mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
328 	mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
329 	mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
330 
331 	mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
332 	mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
333 	mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
334 	mes_set_hw_res_pkt.paging_vmid = 0;
335 	mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr = mes->sch_ctx_gpu_addr;
336 	mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
337 		mes->query_status_fence_gpu_addr;
338 
339 	for (i = 0; i < MAX_COMPUTE_PIPES; i++)
340 		mes_set_hw_res_pkt.compute_hqd_mask[i] =
341 			mes->compute_hqd_mask[i];
342 
343 	for (i = 0; i < MAX_GFX_PIPES; i++)
344 		mes_set_hw_res_pkt.gfx_hqd_mask[i] = mes->gfx_hqd_mask[i];
345 
346 	for (i = 0; i < MAX_SDMA_PIPES; i++)
347 		mes_set_hw_res_pkt.sdma_hqd_mask[i] = mes->sdma_hqd_mask[i];
348 
349 	for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
350 		mes_set_hw_res_pkt.aggregated_doorbells[i] =
351 			mes->aggregated_doorbells[i];
352 
353 	for (i = 0; i < 5; i++) {
354 		mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
355 		mes_set_hw_res_pkt.mmhub_base[i] =
356 				adev->reg_offset[MMHUB_HWIP][0][i];
357 		mes_set_hw_res_pkt.osssys_base[i] =
358 		adev->reg_offset[OSSSYS_HWIP][0][i];
359 	}
360 
361 	mes_set_hw_res_pkt.disable_reset = 1;
362 	mes_set_hw_res_pkt.disable_mes_log = 1;
363 	mes_set_hw_res_pkt.use_different_vmid_compute = 1;
364 	mes_set_hw_res_pkt.oversubscription_timer = 50;
365 
366 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
367 			&mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
368 			offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
369 }
370 
371 static void mes_v11_0_init_aggregated_doorbell(struct amdgpu_mes *mes)
372 {
373 	struct amdgpu_device *adev = mes->adev;
374 	uint32_t data;
375 
376 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1);
377 	data &= ~(CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET_MASK |
378 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_EN_MASK |
379 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_HIT_MASK);
380 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_LOW] <<
381 		CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET__SHIFT;
382 	data |= 1 << CP_MES_DOORBELL_CONTROL1__DOORBELL_EN__SHIFT;
383 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1, data);
384 
385 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2);
386 	data &= ~(CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET_MASK |
387 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_EN_MASK |
388 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_HIT_MASK);
389 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] <<
390 		CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET__SHIFT;
391 	data |= 1 << CP_MES_DOORBELL_CONTROL2__DOORBELL_EN__SHIFT;
392 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2, data);
393 
394 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3);
395 	data &= ~(CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET_MASK |
396 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_EN_MASK |
397 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_HIT_MASK);
398 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_MEDIUM] <<
399 		CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET__SHIFT;
400 	data |= 1 << CP_MES_DOORBELL_CONTROL3__DOORBELL_EN__SHIFT;
401 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3, data);
402 
403 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4);
404 	data &= ~(CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET_MASK |
405 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_EN_MASK |
406 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_HIT_MASK);
407 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_HIGH] <<
408 		CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET__SHIFT;
409 	data |= 1 << CP_MES_DOORBELL_CONTROL4__DOORBELL_EN__SHIFT;
410 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4, data);
411 
412 	data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5);
413 	data &= ~(CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET_MASK |
414 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_EN_MASK |
415 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_HIT_MASK);
416 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_REALTIME] <<
417 		CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET__SHIFT;
418 	data |= 1 << CP_MES_DOORBELL_CONTROL5__DOORBELL_EN__SHIFT;
419 	WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5, data);
420 
421 	data = 1 << CP_HQD_GFX_CONTROL__DB_UPDATED_MSG_EN__SHIFT;
422 	WREG32_SOC15(GC, 0, regCP_HQD_GFX_CONTROL, data);
423 }
424 
425 static const struct amdgpu_mes_funcs mes_v11_0_funcs = {
426 	.add_hw_queue = mes_v11_0_add_hw_queue,
427 	.remove_hw_queue = mes_v11_0_remove_hw_queue,
428 	.unmap_legacy_queue = mes_v11_0_unmap_legacy_queue,
429 	.suspend_gang = mes_v11_0_suspend_gang,
430 	.resume_gang = mes_v11_0_resume_gang,
431 	.misc_op = mes_v11_0_misc_op,
432 };
433 
434 static int mes_v11_0_init_microcode(struct amdgpu_device *adev,
435 				    enum admgpu_mes_pipe pipe)
436 {
437 	char fw_name[30];
438 	char ucode_prefix[30];
439 	int err;
440 	const struct mes_firmware_header_v1_0 *mes_hdr;
441 	struct amdgpu_firmware_info *info;
442 
443 	amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix, sizeof(ucode_prefix));
444 
445 	if (pipe == AMDGPU_MES_SCHED_PIPE)
446 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes.bin",
447 			 ucode_prefix);
448 	else
449 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes1.bin",
450 			 ucode_prefix);
451 
452 	err = request_firmware(&adev->mes.fw[pipe], fw_name, adev->dev);
453 	if (err)
454 		return err;
455 
456 	err = amdgpu_ucode_validate(adev->mes.fw[pipe]);
457 	if (err) {
458 		release_firmware(adev->mes.fw[pipe]);
459 		adev->mes.fw[pipe] = NULL;
460 		return err;
461 	}
462 
463 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
464 		adev->mes.fw[pipe]->data;
465 	adev->mes.ucode_fw_version[pipe] =
466 		le32_to_cpu(mes_hdr->mes_ucode_version);
467 	adev->mes.ucode_fw_version[pipe] =
468 		le32_to_cpu(mes_hdr->mes_ucode_data_version);
469 	adev->mes.uc_start_addr[pipe] =
470 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
471 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
472 	adev->mes.data_start_addr[pipe] =
473 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
474 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
475 
476 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
477 		int ucode, ucode_data;
478 
479 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
480 			ucode = AMDGPU_UCODE_ID_CP_MES;
481 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
482 		} else {
483 			ucode = AMDGPU_UCODE_ID_CP_MES1;
484 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
485 		}
486 
487 		info = &adev->firmware.ucode[ucode];
488 		info->ucode_id = ucode;
489 		info->fw = adev->mes.fw[pipe];
490 		adev->firmware.fw_size +=
491 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
492 			      PAGE_SIZE);
493 
494 		info = &adev->firmware.ucode[ucode_data];
495 		info->ucode_id = ucode_data;
496 		info->fw = adev->mes.fw[pipe];
497 		adev->firmware.fw_size +=
498 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
499 			      PAGE_SIZE);
500 	}
501 
502 	return 0;
503 }
504 
505 static void mes_v11_0_free_microcode(struct amdgpu_device *adev,
506 				     enum admgpu_mes_pipe pipe)
507 {
508 	release_firmware(adev->mes.fw[pipe]);
509 	adev->mes.fw[pipe] = NULL;
510 }
511 
512 static int mes_v11_0_allocate_ucode_buffer(struct amdgpu_device *adev,
513 					   enum admgpu_mes_pipe pipe)
514 {
515 	int r;
516 	const struct mes_firmware_header_v1_0 *mes_hdr;
517 	const __le32 *fw_data;
518 	unsigned fw_size;
519 
520 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
521 		adev->mes.fw[pipe]->data;
522 
523 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
524 		   le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
525 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
526 
527 	r = amdgpu_bo_create_reserved(adev, fw_size,
528 				      PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
529 				      &adev->mes.ucode_fw_obj[pipe],
530 				      &adev->mes.ucode_fw_gpu_addr[pipe],
531 				      (void **)&adev->mes.ucode_fw_ptr[pipe]);
532 	if (r) {
533 		dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
534 		return r;
535 	}
536 
537 	memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
538 
539 	amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
540 	amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
541 
542 	return 0;
543 }
544 
545 static int mes_v11_0_allocate_ucode_data_buffer(struct amdgpu_device *adev,
546 						enum admgpu_mes_pipe pipe)
547 {
548 	int r;
549 	const struct mes_firmware_header_v1_0 *mes_hdr;
550 	const __le32 *fw_data;
551 	unsigned fw_size;
552 
553 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
554 		adev->mes.fw[pipe]->data;
555 
556 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
557 		   le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
558 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
559 
560 	r = amdgpu_bo_create_reserved(adev, fw_size,
561 				      64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
562 				      &adev->mes.data_fw_obj[pipe],
563 				      &adev->mes.data_fw_gpu_addr[pipe],
564 				      (void **)&adev->mes.data_fw_ptr[pipe]);
565 	if (r) {
566 		dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
567 		return r;
568 	}
569 
570 	memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
571 
572 	amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
573 	amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
574 
575 	return 0;
576 }
577 
578 static void mes_v11_0_free_ucode_buffers(struct amdgpu_device *adev,
579 					 enum admgpu_mes_pipe pipe)
580 {
581 	amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
582 			      &adev->mes.data_fw_gpu_addr[pipe],
583 			      (void **)&adev->mes.data_fw_ptr[pipe]);
584 
585 	amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
586 			      &adev->mes.ucode_fw_gpu_addr[pipe],
587 			      (void **)&adev->mes.ucode_fw_ptr[pipe]);
588 }
589 
590 static void mes_v11_0_enable(struct amdgpu_device *adev, bool enable)
591 {
592 	uint64_t ucode_addr;
593 	uint32_t pipe, data = 0;
594 
595 	if (enable) {
596 		data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
597 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
598 		data = REG_SET_FIELD(data, CP_MES_CNTL,
599 			     MES_PIPE1_RESET, adev->enable_mes_kiq ? 1 : 0);
600 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
601 
602 		mutex_lock(&adev->srbm_mutex);
603 		for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
604 			if (!adev->enable_mes_kiq &&
605 			    pipe == AMDGPU_MES_KIQ_PIPE)
606 				continue;
607 
608 			soc21_grbm_select(adev, 3, pipe, 0, 0);
609 
610 			ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
611 			WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
612 				     lower_32_bits(ucode_addr));
613 			WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
614 				     upper_32_bits(ucode_addr));
615 		}
616 		soc21_grbm_select(adev, 0, 0, 0, 0);
617 		mutex_unlock(&adev->srbm_mutex);
618 
619 		/* unhalt MES and activate pipe0 */
620 		data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
621 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE,
622 				     adev->enable_mes_kiq ? 1 : 0);
623 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
624 
625 		if (amdgpu_emu_mode)
626 			msleep(100);
627 		else
628 			udelay(50);
629 	} else {
630 		data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
631 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
632 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
633 		data = REG_SET_FIELD(data, CP_MES_CNTL,
634 				     MES_INVALIDATE_ICACHE, 1);
635 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
636 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET,
637 				     adev->enable_mes_kiq ? 1 : 0);
638 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
639 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
640 	}
641 }
642 
643 /* This function is for backdoor MES firmware */
644 static int mes_v11_0_load_microcode(struct amdgpu_device *adev,
645 				    enum admgpu_mes_pipe pipe, bool prime_icache)
646 {
647 	int r;
648 	uint32_t data;
649 	uint64_t ucode_addr;
650 
651 	mes_v11_0_enable(adev, false);
652 
653 	if (!adev->mes.fw[pipe])
654 		return -EINVAL;
655 
656 	r = mes_v11_0_allocate_ucode_buffer(adev, pipe);
657 	if (r)
658 		return r;
659 
660 	r = mes_v11_0_allocate_ucode_data_buffer(adev, pipe);
661 	if (r) {
662 		mes_v11_0_free_ucode_buffers(adev, pipe);
663 		return r;
664 	}
665 
666 	mutex_lock(&adev->srbm_mutex);
667 	/* me=3, pipe=0, queue=0 */
668 	soc21_grbm_select(adev, 3, pipe, 0, 0);
669 
670 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0);
671 
672 	/* set ucode start address */
673 	ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
674 	WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
675 		     lower_32_bits(ucode_addr));
676 	WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
677 		     upper_32_bits(ucode_addr));
678 
679 	/* set ucode fimrware address */
680 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO,
681 		     lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
682 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI,
683 		     upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
684 
685 	/* set ucode instruction cache boundary to 2M-1 */
686 	WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF);
687 
688 	/* set ucode data firmware address */
689 	WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO,
690 		     lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
691 	WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI,
692 		     upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
693 
694 	/* Set 0x3FFFF (256K-1) to CP_MES_MDBOUND_LO */
695 	WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x3FFFF);
696 
697 	if (prime_icache) {
698 		/* invalidate ICACHE */
699 		data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
700 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
701 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
702 		WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
703 
704 		/* prime the ICACHE. */
705 		data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
706 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
707 		WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
708 	}
709 
710 	soc21_grbm_select(adev, 0, 0, 0, 0);
711 	mutex_unlock(&adev->srbm_mutex);
712 
713 	return 0;
714 }
715 
716 static int mes_v11_0_allocate_eop_buf(struct amdgpu_device *adev,
717 				      enum admgpu_mes_pipe pipe)
718 {
719 	int r;
720 	u32 *eop;
721 
722 	r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
723 			      AMDGPU_GEM_DOMAIN_GTT,
724 			      &adev->mes.eop_gpu_obj[pipe],
725 			      &adev->mes.eop_gpu_addr[pipe],
726 			      (void **)&eop);
727 	if (r) {
728 		dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
729 		return r;
730 	}
731 
732 	memset(eop, 0,
733 	       adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
734 
735 	amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
736 	amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
737 
738 	return 0;
739 }
740 
741 static int mes_v11_0_mqd_init(struct amdgpu_ring *ring)
742 {
743 	struct v11_compute_mqd *mqd = ring->mqd_ptr;
744 	uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
745 	uint32_t tmp;
746 
747 	mqd->header = 0xC0310800;
748 	mqd->compute_pipelinestat_enable = 0x00000001;
749 	mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
750 	mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
751 	mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
752 	mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
753 	mqd->compute_misc_reserved = 0x00000007;
754 
755 	eop_base_addr = ring->eop_gpu_addr >> 8;
756 
757 	/* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
758 	tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
759 	tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
760 			(order_base_2(MES_EOP_SIZE / 4) - 1));
761 
762 	mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
763 	mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
764 	mqd->cp_hqd_eop_control = tmp;
765 
766 	/* disable the queue if it's active */
767 	ring->wptr = 0;
768 	mqd->cp_hqd_pq_rptr = 0;
769 	mqd->cp_hqd_pq_wptr_lo = 0;
770 	mqd->cp_hqd_pq_wptr_hi = 0;
771 
772 	/* set the pointer to the MQD */
773 	mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
774 	mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
775 
776 	/* set MQD vmid to 0 */
777 	tmp = regCP_MQD_CONTROL_DEFAULT;
778 	tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
779 	mqd->cp_mqd_control = tmp;
780 
781 	/* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
782 	hqd_gpu_addr = ring->gpu_addr >> 8;
783 	mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
784 	mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
785 
786 	/* set the wb address whether it's enabled or not */
787 	wb_gpu_addr = ring->rptr_gpu_addr;
788 	mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
789 	mqd->cp_hqd_pq_rptr_report_addr_hi =
790 		upper_32_bits(wb_gpu_addr) & 0xffff;
791 
792 	/* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
793 	wb_gpu_addr = ring->wptr_gpu_addr;
794 	mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
795 	mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
796 
797 	/* set up the HQD, this is similar to CP_RB0_CNTL */
798 	tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
799 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
800 			    (order_base_2(ring->ring_size / 4) - 1));
801 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
802 			    ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
803 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
804 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
805 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
806 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
807 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
808 	mqd->cp_hqd_pq_control = tmp;
809 
810 	/* enable doorbell */
811 	tmp = 0;
812 	if (ring->use_doorbell) {
813 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
814 				    DOORBELL_OFFSET, ring->doorbell_index);
815 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
816 				    DOORBELL_EN, 1);
817 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
818 				    DOORBELL_SOURCE, 0);
819 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
820 				    DOORBELL_HIT, 0);
821 	}
822 	else
823 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
824 				    DOORBELL_EN, 0);
825 	mqd->cp_hqd_pq_doorbell_control = tmp;
826 
827 	mqd->cp_hqd_vmid = 0;
828 	/* activate the queue */
829 	mqd->cp_hqd_active = 1;
830 
831 	tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
832 	tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
833 			    PRELOAD_SIZE, 0x55);
834 	mqd->cp_hqd_persistent_state = tmp;
835 
836 	mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT;
837 	mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
838 	mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
839 
840 	return 0;
841 }
842 
843 static void mes_v11_0_queue_init_register(struct amdgpu_ring *ring)
844 {
845 	struct v11_compute_mqd *mqd = ring->mqd_ptr;
846 	struct amdgpu_device *adev = ring->adev;
847 	uint32_t data = 0;
848 
849 	mutex_lock(&adev->srbm_mutex);
850 	soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
851 
852 	/* set CP_HQD_VMID.VMID = 0. */
853 	data = RREG32_SOC15(GC, 0, regCP_HQD_VMID);
854 	data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
855 	WREG32_SOC15(GC, 0, regCP_HQD_VMID, data);
856 
857 	/* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
858 	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
859 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
860 			     DOORBELL_EN, 0);
861 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
862 
863 	/* set CP_MQD_BASE_ADDR/HI with the MQD base address */
864 	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
865 	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
866 
867 	/* set CP_MQD_CONTROL.VMID=0 */
868 	data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL);
869 	data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
870 	WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0);
871 
872 	/* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
873 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
874 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
875 
876 	/* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
877 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
878 		     mqd->cp_hqd_pq_rptr_report_addr_lo);
879 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
880 		     mqd->cp_hqd_pq_rptr_report_addr_hi);
881 
882 	/* set CP_HQD_PQ_CONTROL */
883 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
884 
885 	/* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
886 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
887 		     mqd->cp_hqd_pq_wptr_poll_addr_lo);
888 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
889 		     mqd->cp_hqd_pq_wptr_poll_addr_hi);
890 
891 	/* set CP_HQD_PQ_DOORBELL_CONTROL */
892 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
893 		     mqd->cp_hqd_pq_doorbell_control);
894 
895 	/* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
896 	WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
897 
898 	/* set CP_HQD_ACTIVE.ACTIVE=1 */
899 	WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active);
900 
901 	soc21_grbm_select(adev, 0, 0, 0, 0);
902 	mutex_unlock(&adev->srbm_mutex);
903 }
904 
905 static int mes_v11_0_kiq_enable_queue(struct amdgpu_device *adev)
906 {
907 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
908 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
909 	int r;
910 
911 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
912 		return -EINVAL;
913 
914 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
915 	if (r) {
916 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
917 		return r;
918 	}
919 
920 	kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring);
921 
922 	r = amdgpu_ring_test_ring(kiq_ring);
923 	if (r) {
924 		DRM_ERROR("kfq enable failed\n");
925 		kiq_ring->sched.ready = false;
926 	}
927 	return r;
928 }
929 
930 static int mes_v11_0_queue_init(struct amdgpu_device *adev,
931 				enum admgpu_mes_pipe pipe)
932 {
933 	struct amdgpu_ring *ring;
934 	int r;
935 
936 	if (pipe == AMDGPU_MES_KIQ_PIPE)
937 		ring = &adev->gfx.kiq.ring;
938 	else if (pipe == AMDGPU_MES_SCHED_PIPE)
939 		ring = &adev->mes.ring;
940 	else
941 		BUG();
942 
943 	if ((pipe == AMDGPU_MES_SCHED_PIPE) &&
944 	    (amdgpu_in_reset(adev) || adev->in_suspend)) {
945 		*(ring->wptr_cpu_addr) = 0;
946 		*(ring->rptr_cpu_addr) = 0;
947 		amdgpu_ring_clear_ring(ring);
948 	}
949 
950 	r = mes_v11_0_mqd_init(ring);
951 	if (r)
952 		return r;
953 
954 	if (pipe == AMDGPU_MES_SCHED_PIPE) {
955 		r = mes_v11_0_kiq_enable_queue(adev);
956 		if (r)
957 			return r;
958 	} else {
959 		mes_v11_0_queue_init_register(ring);
960 	}
961 
962 	/* get MES scheduler/KIQ versions */
963 	mutex_lock(&adev->srbm_mutex);
964 	soc21_grbm_select(adev, 3, pipe, 0, 0);
965 
966 	if (pipe == AMDGPU_MES_SCHED_PIPE)
967 		adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
968 	else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
969 		adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
970 
971 	soc21_grbm_select(adev, 0, 0, 0, 0);
972 	mutex_unlock(&adev->srbm_mutex);
973 
974 	return 0;
975 }
976 
977 static int mes_v11_0_ring_init(struct amdgpu_device *adev)
978 {
979 	struct amdgpu_ring *ring;
980 
981 	ring = &adev->mes.ring;
982 
983 	ring->funcs = &mes_v11_0_ring_funcs;
984 
985 	ring->me = 3;
986 	ring->pipe = 0;
987 	ring->queue = 0;
988 
989 	ring->ring_obj = NULL;
990 	ring->use_doorbell = true;
991 	ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
992 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_SCHED_PIPE];
993 	ring->no_scheduler = true;
994 	sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
995 
996 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
997 				AMDGPU_RING_PRIO_DEFAULT, NULL);
998 }
999 
1000 static int mes_v11_0_kiq_ring_init(struct amdgpu_device *adev)
1001 {
1002 	struct amdgpu_ring *ring;
1003 
1004 	spin_lock_init(&adev->gfx.kiq.ring_lock);
1005 
1006 	ring = &adev->gfx.kiq.ring;
1007 
1008 	ring->me = 3;
1009 	ring->pipe = 1;
1010 	ring->queue = 0;
1011 
1012 	ring->adev = NULL;
1013 	ring->ring_obj = NULL;
1014 	ring->use_doorbell = true;
1015 	ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
1016 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
1017 	ring->no_scheduler = true;
1018 	sprintf(ring->name, "mes_kiq_%d.%d.%d",
1019 		ring->me, ring->pipe, ring->queue);
1020 
1021 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1022 				AMDGPU_RING_PRIO_DEFAULT, NULL);
1023 }
1024 
1025 static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev,
1026 				 enum admgpu_mes_pipe pipe)
1027 {
1028 	int r, mqd_size = sizeof(struct v11_compute_mqd);
1029 	struct amdgpu_ring *ring;
1030 
1031 	if (pipe == AMDGPU_MES_KIQ_PIPE)
1032 		ring = &adev->gfx.kiq.ring;
1033 	else if (pipe == AMDGPU_MES_SCHED_PIPE)
1034 		ring = &adev->mes.ring;
1035 	else
1036 		BUG();
1037 
1038 	if (ring->mqd_obj)
1039 		return 0;
1040 
1041 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
1042 				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
1043 				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
1044 	if (r) {
1045 		dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
1046 		return r;
1047 	}
1048 
1049 	memset(ring->mqd_ptr, 0, mqd_size);
1050 
1051 	/* prepare MQD backup */
1052 	adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
1053 	if (!adev->mes.mqd_backup[pipe])
1054 		dev_warn(adev->dev,
1055 			 "no memory to create MQD backup for ring %s\n",
1056 			 ring->name);
1057 
1058 	return 0;
1059 }
1060 
1061 static int mes_v11_0_sw_init(void *handle)
1062 {
1063 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1064 	int pipe, r;
1065 
1066 	adev->mes.adev = adev;
1067 	adev->mes.funcs = &mes_v11_0_funcs;
1068 	adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init;
1069 	adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini;
1070 
1071 	r = amdgpu_mes_init(adev);
1072 	if (r)
1073 		return r;
1074 
1075 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1076 		if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
1077 			continue;
1078 
1079 		r = mes_v11_0_init_microcode(adev, pipe);
1080 		if (r)
1081 			return r;
1082 
1083 		r = mes_v11_0_allocate_eop_buf(adev, pipe);
1084 		if (r)
1085 			return r;
1086 
1087 		r = mes_v11_0_mqd_sw_init(adev, pipe);
1088 		if (r)
1089 			return r;
1090 	}
1091 
1092 	if (adev->enable_mes_kiq) {
1093 		r = mes_v11_0_kiq_ring_init(adev);
1094 		if (r)
1095 			return r;
1096 	}
1097 
1098 	r = mes_v11_0_ring_init(adev);
1099 	if (r)
1100 		return r;
1101 
1102 	return 0;
1103 }
1104 
1105 static int mes_v11_0_sw_fini(void *handle)
1106 {
1107 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1108 	int pipe;
1109 
1110 	amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
1111 	amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
1112 
1113 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1114 		kfree(adev->mes.mqd_backup[pipe]);
1115 
1116 		amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1117 				      &adev->mes.eop_gpu_addr[pipe],
1118 				      NULL);
1119 
1120 		mes_v11_0_free_microcode(adev, pipe);
1121 	}
1122 
1123 	amdgpu_bo_free_kernel(&adev->gfx.kiq.ring.mqd_obj,
1124 			      &adev->gfx.kiq.ring.mqd_gpu_addr,
1125 			      &adev->gfx.kiq.ring.mqd_ptr);
1126 
1127 	amdgpu_bo_free_kernel(&adev->mes.ring.mqd_obj,
1128 			      &adev->mes.ring.mqd_gpu_addr,
1129 			      &adev->mes.ring.mqd_ptr);
1130 
1131 	amdgpu_ring_fini(&adev->gfx.kiq.ring);
1132 	amdgpu_ring_fini(&adev->mes.ring);
1133 
1134 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1135 		mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1136 		mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1137 	}
1138 
1139 	amdgpu_mes_fini(adev);
1140 	return 0;
1141 }
1142 
1143 static void mes_v11_0_kiq_setting(struct amdgpu_ring *ring)
1144 {
1145 	uint32_t tmp;
1146 	struct amdgpu_device *adev = ring->adev;
1147 
1148 	/* tell RLC which is KIQ queue */
1149 	tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1150 	tmp &= 0xffffff00;
1151 	tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1152 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1153 	tmp |= 0x80;
1154 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1155 }
1156 
1157 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev)
1158 {
1159 	int r = 0;
1160 
1161 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1162 
1163 		r = mes_v11_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE, false);
1164 		if (r) {
1165 			DRM_ERROR("failed to load MES fw, r=%d\n", r);
1166 			return r;
1167 		}
1168 
1169 		r = mes_v11_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE, true);
1170 		if (r) {
1171 			DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1172 			return r;
1173 		}
1174 
1175 	}
1176 
1177 	mes_v11_0_enable(adev, true);
1178 
1179 	mes_v11_0_kiq_setting(&adev->gfx.kiq.ring);
1180 
1181 	r = mes_v11_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE);
1182 	if (r)
1183 		goto failure;
1184 
1185 	return r;
1186 
1187 failure:
1188 	mes_v11_0_hw_fini(adev);
1189 	return r;
1190 }
1191 
1192 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev)
1193 {
1194 	mes_v11_0_enable(adev, false);
1195 	return 0;
1196 }
1197 
1198 static int mes_v11_0_hw_init(void *handle)
1199 {
1200 	int r;
1201 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1202 
1203 	if (!adev->enable_mes_kiq) {
1204 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1205 			r = mes_v11_0_load_microcode(adev,
1206 					     AMDGPU_MES_SCHED_PIPE, true);
1207 			if (r) {
1208 				DRM_ERROR("failed to MES fw, r=%d\n", r);
1209 				return r;
1210 			}
1211 		}
1212 
1213 		mes_v11_0_enable(adev, true);
1214 	}
1215 
1216 	r = mes_v11_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE);
1217 	if (r)
1218 		goto failure;
1219 
1220 	r = mes_v11_0_set_hw_resources(&adev->mes);
1221 	if (r)
1222 		goto failure;
1223 
1224 	mes_v11_0_init_aggregated_doorbell(&adev->mes);
1225 
1226 	r = mes_v11_0_query_sched_status(&adev->mes);
1227 	if (r) {
1228 		DRM_ERROR("MES is busy\n");
1229 		goto failure;
1230 	}
1231 
1232 	/*
1233 	 * Disable KIQ ring usage from the driver once MES is enabled.
1234 	 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1235 	 * with MES enabled.
1236 	 */
1237 	adev->gfx.kiq.ring.sched.ready = false;
1238 	adev->mes.ring.sched.ready = true;
1239 
1240 	return 0;
1241 
1242 failure:
1243 	mes_v11_0_hw_fini(adev);
1244 	return r;
1245 }
1246 
1247 static int mes_v11_0_hw_fini(void *handle)
1248 {
1249 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1250 
1251 	adev->mes.ring.sched.ready = false;
1252 	return 0;
1253 }
1254 
1255 static int mes_v11_0_suspend(void *handle)
1256 {
1257 	int r;
1258 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1259 
1260 	r = amdgpu_mes_suspend(adev);
1261 	if (r)
1262 		return r;
1263 
1264 	return mes_v11_0_hw_fini(adev);
1265 }
1266 
1267 static int mes_v11_0_resume(void *handle)
1268 {
1269 	int r;
1270 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1271 
1272 	r = mes_v11_0_hw_init(adev);
1273 	if (r)
1274 		return r;
1275 
1276 	return amdgpu_mes_resume(adev);
1277 }
1278 
1279 static int mes_v11_0_late_init(void *handle)
1280 {
1281 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1282 
1283 	if (!amdgpu_in_reset(adev))
1284 		amdgpu_mes_self_test(adev);
1285 
1286 	return 0;
1287 }
1288 
1289 static const struct amd_ip_funcs mes_v11_0_ip_funcs = {
1290 	.name = "mes_v11_0",
1291 	.late_init = mes_v11_0_late_init,
1292 	.sw_init = mes_v11_0_sw_init,
1293 	.sw_fini = mes_v11_0_sw_fini,
1294 	.hw_init = mes_v11_0_hw_init,
1295 	.hw_fini = mes_v11_0_hw_fini,
1296 	.suspend = mes_v11_0_suspend,
1297 	.resume = mes_v11_0_resume,
1298 };
1299 
1300 const struct amdgpu_ip_block_version mes_v11_0_ip_block = {
1301 	.type = AMD_IP_BLOCK_TYPE_MES,
1302 	.major = 11,
1303 	.minor = 0,
1304 	.rev = 0,
1305 	.funcs = &mes_v11_0_ip_funcs,
1306 };
1307