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 "nv.h"
29 #include "gc/gc_10_1_0_offset.h"
30 #include "gc/gc_10_1_0_sh_mask.h"
31 #include "gc/gc_10_1_0_default.h"
32 #include "v10_structs.h"
33 #include "mes_api_def.h"
34 
35 #define mmCP_MES_IC_OP_CNTL_Sienna_Cichlid               0x2820
36 #define mmCP_MES_IC_OP_CNTL_Sienna_Cichlid_BASE_IDX      1
37 #define mmRLC_CP_SCHEDULERS_Sienna_Cichlid		0x4ca1
38 #define mmRLC_CP_SCHEDULERS_Sienna_Cichlid_BASE_IDX	1
39 
40 MODULE_FIRMWARE("amdgpu/navi10_mes.bin");
41 MODULE_FIRMWARE("amdgpu/sienna_cichlid_mes.bin");
42 MODULE_FIRMWARE("amdgpu/sienna_cichlid_mes1.bin");
43 
44 static int mes_v10_1_hw_fini(void *handle);
45 static int mes_v10_1_kiq_hw_init(struct amdgpu_device *adev);
46 
47 #define MES_EOP_SIZE   2048
48 
49 static void mes_v10_1_ring_set_wptr(struct amdgpu_ring *ring)
50 {
51 	struct amdgpu_device *adev = ring->adev;
52 
53 	if (ring->use_doorbell) {
54 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
55 			     ring->wptr);
56 		WDOORBELL64(ring->doorbell_index, ring->wptr);
57 	} else {
58 		BUG();
59 	}
60 }
61 
62 static u64 mes_v10_1_ring_get_rptr(struct amdgpu_ring *ring)
63 {
64 	return *ring->rptr_cpu_addr;
65 }
66 
67 static u64 mes_v10_1_ring_get_wptr(struct amdgpu_ring *ring)
68 {
69 	u64 wptr;
70 
71 	if (ring->use_doorbell)
72 		wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
73 	else
74 		BUG();
75 	return wptr;
76 }
77 
78 static const struct amdgpu_ring_funcs mes_v10_1_ring_funcs = {
79 	.type = AMDGPU_RING_TYPE_MES,
80 	.align_mask = 1,
81 	.nop = 0,
82 	.support_64bit_ptrs = true,
83 	.get_rptr = mes_v10_1_ring_get_rptr,
84 	.get_wptr = mes_v10_1_ring_get_wptr,
85 	.set_wptr = mes_v10_1_ring_set_wptr,
86 	.insert_nop = amdgpu_ring_insert_nop,
87 };
88 
89 static int mes_v10_1_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
90 						    void *pkt, int size,
91 						    int api_status_off)
92 {
93 	int ndw = size / 4;
94 	signed long r;
95 	union MESAPI__ADD_QUEUE *x_pkt = pkt;
96 	struct MES_API_STATUS *api_status;
97 	struct amdgpu_device *adev = mes->adev;
98 	struct amdgpu_ring *ring = &mes->ring;
99 	unsigned long flags;
100 
101 	BUG_ON(size % 4 != 0);
102 
103 	spin_lock_irqsave(&mes->ring_lock, flags);
104 	if (amdgpu_ring_alloc(ring, ndw)) {
105 		spin_unlock_irqrestore(&mes->ring_lock, flags);
106 		return -ENOMEM;
107 	}
108 
109 	api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
110 	api_status->api_completion_fence_addr = mes->ring.fence_drv.gpu_addr;
111 	api_status->api_completion_fence_value = ++mes->ring.fence_drv.sync_seq;
112 
113 	amdgpu_ring_write_multiple(ring, pkt, ndw);
114 	amdgpu_ring_commit(ring);
115 	spin_unlock_irqrestore(&mes->ring_lock, flags);
116 
117 	DRM_DEBUG("MES msg=%d was emitted\n", x_pkt->header.opcode);
118 
119 	r = amdgpu_fence_wait_polling(ring, ring->fence_drv.sync_seq,
120 				      adev->usec_timeout);
121 	if (r < 1) {
122 		DRM_ERROR("MES failed to response msg=%d\n",
123 			  x_pkt->header.opcode);
124 
125 		while (halt_if_hws_hang)
126 			schedule();
127 
128 		return -ETIMEDOUT;
129 	}
130 
131 	return 0;
132 }
133 
134 static int convert_to_mes_queue_type(int queue_type)
135 {
136 	if (queue_type == AMDGPU_RING_TYPE_GFX)
137 		return MES_QUEUE_TYPE_GFX;
138 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
139 		return MES_QUEUE_TYPE_COMPUTE;
140 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
141 		return MES_QUEUE_TYPE_SDMA;
142 	else
143 		BUG();
144 	return -1;
145 }
146 
147 static int mes_v10_1_add_hw_queue(struct amdgpu_mes *mes,
148 				  struct mes_add_queue_input *input)
149 {
150 	struct amdgpu_device *adev = mes->adev;
151 	union MESAPI__ADD_QUEUE mes_add_queue_pkt;
152 	struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB_0];
153 	uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
154 
155 	memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
156 
157 	mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
158 	mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
159 	mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
160 
161 	mes_add_queue_pkt.process_id = input->process_id;
162 	mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
163 	mes_add_queue_pkt.process_va_start = input->process_va_start;
164 	mes_add_queue_pkt.process_va_end = input->process_va_end;
165 	mes_add_queue_pkt.process_quantum = input->process_quantum;
166 	mes_add_queue_pkt.process_context_addr = input->process_context_addr;
167 	mes_add_queue_pkt.gang_quantum = input->gang_quantum;
168 	mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
169 	mes_add_queue_pkt.inprocess_gang_priority =
170 		input->inprocess_gang_priority;
171 	mes_add_queue_pkt.gang_global_priority_level =
172 		input->gang_global_priority_level;
173 	mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
174 	mes_add_queue_pkt.mqd_addr = input->mqd_addr;
175 	mes_add_queue_pkt.wptr_addr = input->wptr_addr;
176 	mes_add_queue_pkt.queue_type =
177 		convert_to_mes_queue_type(input->queue_type);
178 	mes_add_queue_pkt.paging = input->paging;
179 	mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
180 	mes_add_queue_pkt.gws_base = input->gws_base;
181 	mes_add_queue_pkt.gws_size = input->gws_size;
182 	mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
183 
184 	return mes_v10_1_submit_pkt_and_poll_completion(mes,
185 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
186 			offsetof(union MESAPI__ADD_QUEUE, api_status));
187 }
188 
189 static int mes_v10_1_remove_hw_queue(struct amdgpu_mes *mes,
190 				     struct mes_remove_queue_input *input)
191 {
192 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
193 
194 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
195 
196 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
197 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
198 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
199 
200 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
201 	mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
202 
203 	return mes_v10_1_submit_pkt_and_poll_completion(mes,
204 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
205 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
206 }
207 
208 static int mes_v10_1_unmap_legacy_queue(struct amdgpu_mes *mes,
209 				 struct mes_unmap_legacy_queue_input *input)
210 {
211 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
212 
213 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
214 
215 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
216 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
217 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
218 
219 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
220 	mes_remove_queue_pkt.gang_context_addr = 0;
221 
222 	mes_remove_queue_pkt.pipe_id = input->pipe_id;
223 	mes_remove_queue_pkt.queue_id = input->queue_id;
224 
225 	if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
226 		mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
227 		mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
228 		mes_remove_queue_pkt.tf_data =
229 			lower_32_bits(input->trail_fence_data);
230 	} else {
231 		if (input->queue_type == AMDGPU_RING_TYPE_GFX)
232 			mes_remove_queue_pkt.unmap_legacy_gfx_queue = 1;
233 		else
234 			mes_remove_queue_pkt.unmap_kiq_utility_queue = 1;
235 	}
236 
237 	return mes_v10_1_submit_pkt_and_poll_completion(mes,
238 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
239 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
240 }
241 
242 static int mes_v10_1_suspend_gang(struct amdgpu_mes *mes,
243 				  struct mes_suspend_gang_input *input)
244 {
245 	return 0;
246 }
247 
248 static int mes_v10_1_resume_gang(struct amdgpu_mes *mes,
249 				 struct mes_resume_gang_input *input)
250 {
251 	return 0;
252 }
253 
254 static int mes_v10_1_query_sched_status(struct amdgpu_mes *mes)
255 {
256 	union MESAPI__QUERY_MES_STATUS mes_status_pkt;
257 
258 	memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
259 
260 	mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
261 	mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
262 	mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
263 
264 	return mes_v10_1_submit_pkt_and_poll_completion(mes,
265 			&mes_status_pkt, sizeof(mes_status_pkt),
266 			offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
267 }
268 
269 static int mes_v10_1_set_hw_resources(struct amdgpu_mes *mes)
270 {
271 	int i;
272 	struct amdgpu_device *adev = mes->adev;
273 	union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
274 
275 	memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
276 
277 	mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
278 	mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
279 	mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
280 
281 	mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
282 	mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
283 	mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
284 	mes_set_hw_res_pkt.paging_vmid = 0;
285 	mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr = mes->sch_ctx_gpu_addr;
286 	mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
287 		mes->query_status_fence_gpu_addr;
288 
289 	for (i = 0; i < MAX_COMPUTE_PIPES; i++)
290 		mes_set_hw_res_pkt.compute_hqd_mask[i] =
291 			mes->compute_hqd_mask[i];
292 
293 	for (i = 0; i < MAX_GFX_PIPES; i++)
294 		mes_set_hw_res_pkt.gfx_hqd_mask[i] = mes->gfx_hqd_mask[i];
295 
296 	for (i = 0; i < MAX_SDMA_PIPES; i++)
297 		mes_set_hw_res_pkt.sdma_hqd_mask[i] = mes->sdma_hqd_mask[i];
298 
299 	for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
300 		mes_set_hw_res_pkt.aggregated_doorbells[i] =
301 			mes->aggregated_doorbells[i];
302 
303 	for (i = 0; i < 5; i++) {
304 		mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
305 		mes_set_hw_res_pkt.mmhub_base[i] =
306 			adev->reg_offset[MMHUB_HWIP][0][i];
307 		mes_set_hw_res_pkt.osssys_base[i] =
308 			adev->reg_offset[OSSSYS_HWIP][0][i];
309 	}
310 
311 	mes_set_hw_res_pkt.disable_reset = 1;
312 	mes_set_hw_res_pkt.disable_mes_log = 1;
313 	mes_set_hw_res_pkt.use_different_vmid_compute = 1;
314 
315 	return mes_v10_1_submit_pkt_and_poll_completion(mes,
316 			&mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
317 			offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
318 }
319 
320 static void mes_v10_1_init_aggregated_doorbell(struct amdgpu_mes *mes)
321 {
322 	struct amdgpu_device *adev = mes->adev;
323 	uint32_t data;
324 
325 	data = RREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL1);
326 	data &= ~(CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET_MASK |
327 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_EN_MASK |
328 		  CP_MES_DOORBELL_CONTROL1__DOORBELL_HIT_MASK);
329 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_LOW] <<
330 		CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET__SHIFT;
331 	data |= 1 << CP_MES_DOORBELL_CONTROL1__DOORBELL_EN__SHIFT;
332 	WREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL1, data);
333 
334 	data = RREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL2);
335 	data &= ~(CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET_MASK |
336 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_EN_MASK |
337 		  CP_MES_DOORBELL_CONTROL2__DOORBELL_HIT_MASK);
338 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] <<
339 		CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET__SHIFT;
340 	data |= 1 << CP_MES_DOORBELL_CONTROL2__DOORBELL_EN__SHIFT;
341 	WREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL2, data);
342 
343 	data = RREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL3);
344 	data &= ~(CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET_MASK |
345 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_EN_MASK |
346 		  CP_MES_DOORBELL_CONTROL3__DOORBELL_HIT_MASK);
347 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_MEDIUM] <<
348 		CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET__SHIFT;
349 	data |= 1 << CP_MES_DOORBELL_CONTROL3__DOORBELL_EN__SHIFT;
350 	WREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL3, data);
351 
352 	data = RREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL4);
353 	data &= ~(CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET_MASK |
354 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_EN_MASK |
355 		  CP_MES_DOORBELL_CONTROL4__DOORBELL_HIT_MASK);
356 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_HIGH] <<
357 		CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET__SHIFT;
358 	data |= 1 << CP_MES_DOORBELL_CONTROL4__DOORBELL_EN__SHIFT;
359 	WREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL4, data);
360 
361 	data = RREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL5);
362 	data &= ~(CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET_MASK |
363 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_EN_MASK |
364 		  CP_MES_DOORBELL_CONTROL5__DOORBELL_HIT_MASK);
365 	data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_REALTIME] <<
366 		CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET__SHIFT;
367 	data |= 1 << CP_MES_DOORBELL_CONTROL5__DOORBELL_EN__SHIFT;
368 	WREG32_SOC15(GC, 0, mmCP_MES_DOORBELL_CONTROL5, data);
369 
370 	data = 1 << CP_HQD_GFX_CONTROL__DB_UPDATED_MSG_EN__SHIFT;
371 	WREG32_SOC15(GC, 0, mmCP_HQD_GFX_CONTROL, data);
372 }
373 
374 static const struct amdgpu_mes_funcs mes_v10_1_funcs = {
375 	.add_hw_queue = mes_v10_1_add_hw_queue,
376 	.remove_hw_queue = mes_v10_1_remove_hw_queue,
377 	.unmap_legacy_queue = mes_v10_1_unmap_legacy_queue,
378 	.suspend_gang = mes_v10_1_suspend_gang,
379 	.resume_gang = mes_v10_1_resume_gang,
380 };
381 
382 static int mes_v10_1_init_microcode(struct amdgpu_device *adev,
383 				    enum admgpu_mes_pipe pipe)
384 {
385 	const char *chip_name;
386 	char fw_name[30];
387 	int err;
388 	const struct mes_firmware_header_v1_0 *mes_hdr;
389 	struct amdgpu_firmware_info *info;
390 
391 	switch (adev->ip_versions[GC_HWIP][0]) {
392 	case IP_VERSION(10, 1, 10):
393 		chip_name = "navi10";
394 		break;
395 	case IP_VERSION(10, 3, 0):
396 		chip_name = "sienna_cichlid";
397 		break;
398 	default:
399 		BUG();
400 	}
401 
402 	if (pipe == AMDGPU_MES_SCHED_PIPE)
403 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes.bin",
404 			 chip_name);
405 	else
406 		snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes1.bin",
407 			 chip_name);
408 
409 	err = request_firmware(&adev->mes.fw[pipe], fw_name, adev->dev);
410 	if (err)
411 		return err;
412 
413 	err = amdgpu_ucode_validate(adev->mes.fw[pipe]);
414 	if (err) {
415 		release_firmware(adev->mes.fw[pipe]);
416 		adev->mes.fw[pipe] = NULL;
417 		return err;
418 	}
419 
420 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
421 		adev->mes.fw[pipe]->data;
422 	adev->mes.uc_start_addr[pipe] =
423 		le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
424 		((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
425 	adev->mes.data_start_addr[pipe] =
426 		le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
427 		((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
428 
429 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
430 		int ucode, ucode_data;
431 
432 		if (pipe == AMDGPU_MES_SCHED_PIPE) {
433 			ucode = AMDGPU_UCODE_ID_CP_MES;
434 			ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
435 		} else {
436 			ucode = AMDGPU_UCODE_ID_CP_MES1;
437 			ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
438 		}
439 
440 		info = &adev->firmware.ucode[ucode];
441 		info->ucode_id = ucode;
442 		info->fw = adev->mes.fw[pipe];
443 		adev->firmware.fw_size +=
444 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
445 			      PAGE_SIZE);
446 
447 		info = &adev->firmware.ucode[ucode_data];
448 		info->ucode_id = ucode_data;
449 		info->fw = adev->mes.fw[pipe];
450 		adev->firmware.fw_size +=
451 			ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
452 			      PAGE_SIZE);
453 	}
454 
455 	return 0;
456 }
457 
458 static void mes_v10_1_free_microcode(struct amdgpu_device *adev,
459 				     enum admgpu_mes_pipe pipe)
460 {
461 	release_firmware(adev->mes.fw[pipe]);
462 	adev->mes.fw[pipe] = NULL;
463 }
464 
465 static int mes_v10_1_allocate_ucode_buffer(struct amdgpu_device *adev,
466 					   enum admgpu_mes_pipe pipe)
467 {
468 	int r;
469 	const struct mes_firmware_header_v1_0 *mes_hdr;
470 	const __le32 *fw_data;
471 	unsigned fw_size;
472 
473 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
474 		adev->mes.fw[pipe]->data;
475 
476 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
477 		   le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
478 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
479 
480 	r = amdgpu_bo_create_reserved(adev, fw_size,
481 				      PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
482 				      &adev->mes.ucode_fw_obj[pipe],
483 				      &adev->mes.ucode_fw_gpu_addr[pipe],
484 				      (void **)&adev->mes.ucode_fw_ptr[pipe]);
485 	if (r) {
486 		dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
487 		return r;
488 	}
489 
490 	memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
491 
492 	amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
493 	amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
494 
495 	return 0;
496 }
497 
498 static int mes_v10_1_allocate_ucode_data_buffer(struct amdgpu_device *adev,
499 						enum admgpu_mes_pipe pipe)
500 {
501 	int r;
502 	const struct mes_firmware_header_v1_0 *mes_hdr;
503 	const __le32 *fw_data;
504 	unsigned fw_size;
505 
506 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
507 		adev->mes.fw[pipe]->data;
508 
509 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
510 		   le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
511 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
512 
513 	r = amdgpu_bo_create_reserved(adev, fw_size,
514 				      64 * 1024, AMDGPU_GEM_DOMAIN_GTT,
515 				      &adev->mes.data_fw_obj[pipe],
516 				      &adev->mes.data_fw_gpu_addr[pipe],
517 				      (void **)&adev->mes.data_fw_ptr[pipe]);
518 	if (r) {
519 		dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
520 		return r;
521 	}
522 
523 	memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
524 
525 	amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
526 	amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
527 
528 	return 0;
529 }
530 
531 static void mes_v10_1_free_ucode_buffers(struct amdgpu_device *adev,
532 					 enum admgpu_mes_pipe pipe)
533 {
534 	amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
535 			      &adev->mes.data_fw_gpu_addr[pipe],
536 			      (void **)&adev->mes.data_fw_ptr[pipe]);
537 
538 	amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
539 			      &adev->mes.ucode_fw_gpu_addr[pipe],
540 			      (void **)&adev->mes.ucode_fw_ptr[pipe]);
541 }
542 
543 static void mes_v10_1_enable(struct amdgpu_device *adev, bool enable)
544 {
545 	uint32_t pipe, data = 0;
546 
547 	if (enable) {
548 		data = RREG32_SOC15(GC, 0, mmCP_MES_CNTL);
549 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
550 		data = REG_SET_FIELD(data, CP_MES_CNTL,
551 			     MES_PIPE1_RESET, adev->enable_mes_kiq ? 1 : 0);
552 		WREG32_SOC15(GC, 0, mmCP_MES_CNTL, data);
553 
554 		mutex_lock(&adev->srbm_mutex);
555 		for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
556 			if (!adev->enable_mes_kiq &&
557 			    pipe == AMDGPU_MES_KIQ_PIPE)
558 				continue;
559 
560 			nv_grbm_select(adev, 3, pipe, 0, 0);
561 			WREG32_SOC15(GC, 0, mmCP_MES_PRGRM_CNTR_START,
562 			     (uint32_t)(adev->mes.uc_start_addr[pipe]) >> 2);
563 		}
564 		nv_grbm_select(adev, 0, 0, 0, 0);
565 		mutex_unlock(&adev->srbm_mutex);
566 
567 		/* clear BYPASS_UNCACHED to avoid hangs after interrupt. */
568 		data = RREG32_SOC15(GC, 0, mmCP_MES_DC_OP_CNTL);
569 		data = REG_SET_FIELD(data, CP_MES_DC_OP_CNTL,
570 				     BYPASS_UNCACHED, 0);
571 		WREG32_SOC15(GC, 0, mmCP_MES_DC_OP_CNTL, data);
572 
573 		/* unhalt MES and activate pipe0 */
574 		data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
575 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE,
576 				     adev->enable_mes_kiq ? 1 : 0);
577 		WREG32_SOC15(GC, 0, mmCP_MES_CNTL, data);
578 		udelay(100);
579 	} else {
580 		data = RREG32_SOC15(GC, 0, mmCP_MES_CNTL);
581 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
582 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
583 		data = REG_SET_FIELD(data, CP_MES_CNTL,
584 				     MES_INVALIDATE_ICACHE, 1);
585 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
586 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET,
587 				     adev->enable_mes_kiq ? 1 : 0);
588 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
589 		WREG32_SOC15(GC, 0, mmCP_MES_CNTL, data);
590 	}
591 }
592 
593 /* This function is for backdoor MES firmware */
594 static int mes_v10_1_load_microcode(struct amdgpu_device *adev,
595 				    enum admgpu_mes_pipe pipe)
596 {
597 	int r;
598 	uint32_t data;
599 
600 	mes_v10_1_enable(adev, false);
601 
602 	if (!adev->mes.fw[pipe])
603 		return -EINVAL;
604 
605 	r = mes_v10_1_allocate_ucode_buffer(adev, pipe);
606 	if (r)
607 		return r;
608 
609 	r = mes_v10_1_allocate_ucode_data_buffer(adev, pipe);
610 	if (r) {
611 		mes_v10_1_free_ucode_buffers(adev, pipe);
612 		return r;
613 	}
614 
615 	WREG32_SOC15(GC, 0, mmCP_MES_IC_BASE_CNTL, 0);
616 
617 	mutex_lock(&adev->srbm_mutex);
618 	/* me=3, pipe=0, queue=0 */
619 	nv_grbm_select(adev, 3, pipe, 0, 0);
620 
621 	/* set ucode start address */
622 	WREG32_SOC15(GC, 0, mmCP_MES_PRGRM_CNTR_START,
623 		     (uint32_t)(adev->mes.uc_start_addr[pipe]) >> 2);
624 
625 	/* set ucode fimrware address */
626 	WREG32_SOC15(GC, 0, mmCP_MES_IC_BASE_LO,
627 		     lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
628 	WREG32_SOC15(GC, 0, mmCP_MES_IC_BASE_HI,
629 		     upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
630 
631 	/* set ucode instruction cache boundary to 2M-1 */
632 	WREG32_SOC15(GC, 0, mmCP_MES_MIBOUND_LO, 0x1FFFFF);
633 
634 	/* set ucode data firmware address */
635 	WREG32_SOC15(GC, 0, mmCP_MES_MDBASE_LO,
636 		     lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
637 	WREG32_SOC15(GC, 0, mmCP_MES_MDBASE_HI,
638 		     upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
639 
640 	/* Set 0x3FFFF (256K-1) to CP_MES_MDBOUND_LO */
641 	WREG32_SOC15(GC, 0, mmCP_MES_MDBOUND_LO, 0x3FFFF);
642 
643 	/* invalidate ICACHE */
644 	switch (adev->ip_versions[GC_HWIP][0]) {
645 	case IP_VERSION(10, 3, 0):
646 		data = RREG32_SOC15(GC, 0, mmCP_MES_IC_OP_CNTL_Sienna_Cichlid);
647 		break;
648 	default:
649 		data = RREG32_SOC15(GC, 0, mmCP_MES_IC_OP_CNTL);
650 		break;
651 	}
652 	data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
653 	data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
654 	switch (adev->ip_versions[GC_HWIP][0]) {
655 	case IP_VERSION(10, 3, 0):
656 		WREG32_SOC15(GC, 0, mmCP_MES_IC_OP_CNTL_Sienna_Cichlid, data);
657 		break;
658 	default:
659 		WREG32_SOC15(GC, 0, mmCP_MES_IC_OP_CNTL, data);
660 		break;
661 	}
662 
663 	/* prime the ICACHE. */
664 	switch (adev->ip_versions[GC_HWIP][0]) {
665 	case IP_VERSION(10, 3, 0):
666 		data = RREG32_SOC15(GC, 0, mmCP_MES_IC_OP_CNTL_Sienna_Cichlid);
667 		break;
668 	default:
669 		data = RREG32_SOC15(GC, 0, mmCP_MES_IC_OP_CNTL);
670 		break;
671 	}
672 	data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
673 	switch (adev->ip_versions[GC_HWIP][0]) {
674 	case IP_VERSION(10, 3, 0):
675 		WREG32_SOC15(GC, 0, mmCP_MES_IC_OP_CNTL_Sienna_Cichlid, data);
676 		break;
677 	default:
678 		WREG32_SOC15(GC, 0, mmCP_MES_IC_OP_CNTL, data);
679 		break;
680 	}
681 
682 	nv_grbm_select(adev, 0, 0, 0, 0);
683 	mutex_unlock(&adev->srbm_mutex);
684 
685 	return 0;
686 }
687 
688 static int mes_v10_1_allocate_eop_buf(struct amdgpu_device *adev,
689 				      enum admgpu_mes_pipe pipe)
690 {
691 	int r;
692 	u32 *eop;
693 
694 	r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
695 			      AMDGPU_GEM_DOMAIN_GTT,
696 			      &adev->mes.eop_gpu_obj[pipe],
697 			      &adev->mes.eop_gpu_addr[pipe],
698 			      (void **)&eop);
699 	if (r) {
700 		dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
701 		return r;
702 	}
703 
704 	memset(eop, 0, adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
705 
706 	amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
707 	amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
708 
709 	return 0;
710 }
711 
712 static int mes_v10_1_mqd_init(struct amdgpu_ring *ring)
713 {
714 	struct v10_compute_mqd *mqd = ring->mqd_ptr;
715 	uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
716 	uint32_t tmp;
717 
718 	mqd->header = 0xC0310800;
719 	mqd->compute_pipelinestat_enable = 0x00000001;
720 	mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
721 	mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
722 	mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
723 	mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
724 	mqd->compute_misc_reserved = 0x00000003;
725 
726 	eop_base_addr = ring->eop_gpu_addr >> 8;
727 
728 	/* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
729 	tmp = mmCP_HQD_EOP_CONTROL_DEFAULT;
730 	tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
731 			(order_base_2(MES_EOP_SIZE / 4) - 1));
732 
733 	mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
734 	mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
735 	mqd->cp_hqd_eop_control = tmp;
736 
737 	/* disable the queue if it's active */
738 	ring->wptr = 0;
739 	mqd->cp_hqd_pq_rptr = 0;
740 	mqd->cp_hqd_pq_wptr_lo = 0;
741 	mqd->cp_hqd_pq_wptr_hi = 0;
742 
743 	/* set the pointer to the MQD */
744 	mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
745 	mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
746 
747 	/* set MQD vmid to 0 */
748 	tmp = mmCP_MQD_CONTROL_DEFAULT;
749 	tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
750 	mqd->cp_mqd_control = tmp;
751 
752 	/* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
753 	hqd_gpu_addr = ring->gpu_addr >> 8;
754 	mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
755 	mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
756 
757 	/* set the wb address whether it's enabled or not */
758 	wb_gpu_addr = ring->rptr_gpu_addr;
759 	mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
760 	mqd->cp_hqd_pq_rptr_report_addr_hi =
761 		upper_32_bits(wb_gpu_addr) & 0xffff;
762 
763 	/* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
764 	wb_gpu_addr = ring->wptr_gpu_addr;
765 	mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
766 	mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
767 
768 	/* set up the HQD, this is similar to CP_RB0_CNTL */
769 	tmp = mmCP_HQD_PQ_CONTROL_DEFAULT;
770 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
771 			    (order_base_2(ring->ring_size / 4) - 1));
772 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
773 			    ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
774 #ifdef __BIG_ENDIAN
775 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, ENDIAN_SWAP, 1);
776 #endif
777 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
778 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
779 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
780 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
781 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
782 	mqd->cp_hqd_pq_control = tmp;
783 
784 	/* enable doorbell? */
785 	tmp = 0;
786 	if (ring->use_doorbell) {
787 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
788 				    DOORBELL_OFFSET, ring->doorbell_index);
789 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
790 				    DOORBELL_EN, 1);
791 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
792 				    DOORBELL_SOURCE, 0);
793 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
794 				    DOORBELL_HIT, 0);
795 	}
796 	else
797 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
798 				    DOORBELL_EN, 0);
799 	mqd->cp_hqd_pq_doorbell_control = tmp;
800 
801 	mqd->cp_hqd_vmid = 0;
802 	/* activate the queue */
803 	mqd->cp_hqd_active = 1;
804 	mqd->cp_hqd_persistent_state = mmCP_HQD_PERSISTENT_STATE_DEFAULT;
805 	mqd->cp_hqd_ib_control = mmCP_HQD_IB_CONTROL_DEFAULT;
806 	mqd->cp_hqd_iq_timer = mmCP_HQD_IQ_TIMER_DEFAULT;
807 	mqd->cp_hqd_quantum = mmCP_HQD_QUANTUM_DEFAULT;
808 
809 	tmp = mmCP_HQD_GFX_CONTROL_DEFAULT;
810 	tmp = REG_SET_FIELD(tmp, CP_HQD_GFX_CONTROL, DB_UPDATED_MSG_EN, 1);
811 	/* offset: 184 - this is used for CP_HQD_GFX_CONTROL */
812 	mqd->cp_hqd_suspend_cntl_stack_offset = tmp;
813 
814 	return 0;
815 }
816 
817 #if 0
818 static void mes_v10_1_queue_init_register(struct amdgpu_ring *ring)
819 {
820 	struct v10_compute_mqd *mqd = ring->mqd_ptr;
821 	struct amdgpu_device *adev = ring->adev;
822 	uint32_t data = 0;
823 
824 	mutex_lock(&adev->srbm_mutex);
825 	nv_grbm_select(adev, 3, ring->pipe, 0, 0);
826 
827 	/* set CP_HQD_VMID.VMID = 0. */
828 	data = RREG32_SOC15(GC, 0, mmCP_HQD_VMID);
829 	data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
830 	WREG32_SOC15(GC, 0, mmCP_HQD_VMID, data);
831 
832 	/* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
833 	data = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL);
834 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
835 			     DOORBELL_EN, 0);
836 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL, data);
837 
838 	/* set CP_MQD_BASE_ADDR/HI with the MQD base address */
839 	WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
840 	WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
841 
842 	/* set CP_MQD_CONTROL.VMID=0 */
843 	data = RREG32_SOC15(GC, 0, mmCP_MQD_CONTROL);
844 	data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
845 	WREG32_SOC15(GC, 0, mmCP_MQD_CONTROL, 0);
846 
847 	/* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
848 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
849 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
850 
851 	/* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
852 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR_REPORT_ADDR,
853 		     mqd->cp_hqd_pq_rptr_report_addr_lo);
854 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
855 		     mqd->cp_hqd_pq_rptr_report_addr_hi);
856 
857 	/* set CP_HQD_PQ_CONTROL */
858 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
859 
860 	/* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
861 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR,
862 		     mqd->cp_hqd_pq_wptr_poll_addr_lo);
863 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI,
864 		     mqd->cp_hqd_pq_wptr_poll_addr_hi);
865 
866 	/* set CP_HQD_PQ_DOORBELL_CONTROL */
867 	WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL,
868 		     mqd->cp_hqd_pq_doorbell_control);
869 
870 	/* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
871 	WREG32_SOC15(GC, 0, mmCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
872 
873 	/* set CP_HQD_ACTIVE.ACTIVE=1 */
874 	WREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE, mqd->cp_hqd_active);
875 
876 	nv_grbm_select(adev, 0, 0, 0, 0);
877 	mutex_unlock(&adev->srbm_mutex);
878 }
879 #endif
880 
881 static int mes_v10_1_kiq_enable_queue(struct amdgpu_device *adev)
882 {
883 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
884 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
885 	int r;
886 
887 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
888 		return -EINVAL;
889 
890 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
891 	if (r) {
892 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
893 		return r;
894 	}
895 
896 	kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring);
897 
898 	r = amdgpu_ring_test_ring(kiq_ring);
899 	if (r) {
900 		DRM_ERROR("kfq enable failed\n");
901 		kiq_ring->sched.ready = false;
902 	}
903 
904 	return r;
905 }
906 
907 static int mes_v10_1_queue_init(struct amdgpu_device *adev)
908 {
909 	int r;
910 
911 	r = mes_v10_1_mqd_init(&adev->mes.ring);
912 	if (r)
913 		return r;
914 
915 	r = mes_v10_1_kiq_enable_queue(adev);
916 	if (r)
917 		return r;
918 
919 	return 0;
920 }
921 
922 static int mes_v10_1_ring_init(struct amdgpu_device *adev)
923 {
924 	struct amdgpu_ring *ring;
925 
926 	ring = &adev->mes.ring;
927 
928 	ring->funcs = &mes_v10_1_ring_funcs;
929 
930 	ring->me = 3;
931 	ring->pipe = 0;
932 	ring->queue = 0;
933 
934 	ring->ring_obj = NULL;
935 	ring->use_doorbell = true;
936 	ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
937 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_SCHED_PIPE];
938 	ring->no_scheduler = true;
939 	sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
940 
941 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
942 				AMDGPU_RING_PRIO_DEFAULT, NULL);
943 }
944 
945 static int mes_v10_1_kiq_ring_init(struct amdgpu_device *adev)
946 {
947 	struct amdgpu_ring *ring;
948 
949 	spin_lock_init(&adev->gfx.kiq.ring_lock);
950 
951 	ring = &adev->gfx.kiq.ring;
952 
953 	ring->me = 3;
954 	ring->pipe = 1;
955 	ring->queue = 0;
956 
957 	ring->adev = NULL;
958 	ring->ring_obj = NULL;
959 	ring->use_doorbell = true;
960 	ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
961 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
962 	ring->no_scheduler = true;
963 	sprintf(ring->name, "mes_kiq_%d.%d.%d",
964 		ring->me, ring->pipe, ring->queue);
965 
966 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
967 				AMDGPU_RING_PRIO_DEFAULT, NULL);
968 }
969 
970 static int mes_v10_1_mqd_sw_init(struct amdgpu_device *adev,
971 				 enum admgpu_mes_pipe pipe)
972 {
973 	int r, mqd_size = sizeof(struct v10_compute_mqd);
974 	struct amdgpu_ring *ring;
975 
976 	if (pipe == AMDGPU_MES_KIQ_PIPE)
977 		ring = &adev->gfx.kiq.ring;
978 	else if (pipe == AMDGPU_MES_SCHED_PIPE)
979 		ring = &adev->mes.ring;
980 	else
981 		BUG();
982 
983 	if (ring->mqd_obj)
984 		return 0;
985 
986 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
987 				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
988 				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
989 	if (r) {
990 		dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
991 		return r;
992 	}
993 	memset(ring->mqd_ptr, 0, mqd_size);
994 
995 	/* prepare MQD backup */
996 	adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
997 	if (!adev->mes.mqd_backup[pipe])
998 		dev_warn(adev->dev,
999 			 "no memory to create MQD backup for ring %s\n",
1000 			 ring->name);
1001 
1002 	return 0;
1003 }
1004 
1005 static int mes_v10_1_sw_init(void *handle)
1006 {
1007 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1008 	int pipe, r;
1009 
1010 	adev->mes.adev = adev;
1011 	adev->mes.funcs = &mes_v10_1_funcs;
1012 	adev->mes.kiq_hw_init = &mes_v10_1_kiq_hw_init;
1013 
1014 	r = amdgpu_mes_init(adev);
1015 	if (r)
1016 		return r;
1017 
1018 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1019 		if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
1020 			continue;
1021 
1022 		r = mes_v10_1_init_microcode(adev, pipe);
1023 		if (r)
1024 			return r;
1025 
1026 		r = mes_v10_1_allocate_eop_buf(adev, pipe);
1027 		if (r)
1028 			return r;
1029 
1030 		r = mes_v10_1_mqd_sw_init(adev, pipe);
1031 		if (r)
1032 			return r;
1033 	}
1034 
1035 	if (adev->enable_mes_kiq) {
1036 		r = mes_v10_1_kiq_ring_init(adev);
1037 		if (r)
1038 			return r;
1039 	}
1040 
1041 	r = mes_v10_1_ring_init(adev);
1042 	if (r)
1043 		return r;
1044 
1045 	return 0;
1046 }
1047 
1048 static int mes_v10_1_sw_fini(void *handle)
1049 {
1050 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1051 	int pipe;
1052 
1053 	amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
1054 	amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
1055 
1056 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1057 		kfree(adev->mes.mqd_backup[pipe]);
1058 
1059 		amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1060 				      &adev->mes.eop_gpu_addr[pipe],
1061 				      NULL);
1062 
1063 		mes_v10_1_free_microcode(adev, pipe);
1064 	}
1065 
1066 	amdgpu_bo_free_kernel(&adev->gfx.kiq.ring.mqd_obj,
1067 			      &adev->gfx.kiq.ring.mqd_gpu_addr,
1068 			      &adev->gfx.kiq.ring.mqd_ptr);
1069 
1070 	amdgpu_bo_free_kernel(&adev->mes.ring.mqd_obj,
1071 			      &adev->mes.ring.mqd_gpu_addr,
1072 			      &adev->mes.ring.mqd_ptr);
1073 
1074 	amdgpu_ring_fini(&adev->gfx.kiq.ring);
1075 	amdgpu_ring_fini(&adev->mes.ring);
1076 
1077 	amdgpu_mes_fini(adev);
1078 	return 0;
1079 }
1080 
1081 static void mes_v10_1_kiq_setting(struct amdgpu_ring *ring)
1082 {
1083 	uint32_t tmp;
1084 	struct amdgpu_device *adev = ring->adev;
1085 
1086 	/* tell RLC which is KIQ queue */
1087 	switch (adev->ip_versions[GC_HWIP][0]) {
1088 	case IP_VERSION(10, 3, 0):
1089 	case IP_VERSION(10, 3, 2):
1090 	case IP_VERSION(10, 3, 1):
1091 	case IP_VERSION(10, 3, 4):
1092 		tmp = RREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS_Sienna_Cichlid);
1093 		tmp &= 0xffffff00;
1094 		tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1095 		WREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS_Sienna_Cichlid, tmp);
1096 		tmp |= 0x80;
1097 		WREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS_Sienna_Cichlid, tmp);
1098 		break;
1099 	default:
1100 		tmp = RREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS);
1101 		tmp &= 0xffffff00;
1102 		tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1103 		WREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS, tmp);
1104 		tmp |= 0x80;
1105 		WREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS, tmp);
1106 		break;
1107 	}
1108 }
1109 
1110 static int mes_v10_1_kiq_hw_init(struct amdgpu_device *adev)
1111 {
1112 	int r = 0;
1113 
1114 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1115 		r = mes_v10_1_load_microcode(adev, AMDGPU_MES_KIQ_PIPE);
1116 		if (r) {
1117 			DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1118 			return r;
1119 		}
1120 
1121 		r = mes_v10_1_load_microcode(adev, AMDGPU_MES_SCHED_PIPE);
1122 		if (r) {
1123 			DRM_ERROR("failed to load MES fw, r=%d\n", r);
1124 			return r;
1125 		}
1126 	}
1127 
1128 	mes_v10_1_enable(adev, true);
1129 
1130 	mes_v10_1_kiq_setting(&adev->gfx.kiq.ring);
1131 
1132 	r = mes_v10_1_queue_init(adev);
1133 	if (r)
1134 		goto failure;
1135 
1136 	return r;
1137 
1138 failure:
1139 	mes_v10_1_hw_fini(adev);
1140 	return r;
1141 }
1142 
1143 static int mes_v10_1_hw_init(void *handle)
1144 {
1145 	int r;
1146 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1147 
1148 	if (!adev->enable_mes_kiq) {
1149 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1150 			r = mes_v10_1_load_microcode(adev,
1151 					     AMDGPU_MES_SCHED_PIPE);
1152 			if (r) {
1153 				DRM_ERROR("failed to MES fw, r=%d\n", r);
1154 				return r;
1155 			}
1156 		}
1157 
1158 		mes_v10_1_enable(adev, true);
1159 	}
1160 
1161 	r = mes_v10_1_queue_init(adev);
1162 	if (r)
1163 		goto failure;
1164 
1165 	r = mes_v10_1_set_hw_resources(&adev->mes);
1166 	if (r)
1167 		goto failure;
1168 
1169 	mes_v10_1_init_aggregated_doorbell(&adev->mes);
1170 
1171 	r = mes_v10_1_query_sched_status(&adev->mes);
1172 	if (r) {
1173 		DRM_ERROR("MES is busy\n");
1174 		goto failure;
1175 	}
1176 
1177 	/*
1178 	 * Disable KIQ ring usage from the driver once MES is enabled.
1179 	 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1180 	 * with MES enabled.
1181 	 */
1182 	adev->gfx.kiq.ring.sched.ready = false;
1183 	adev->mes.ring.sched.ready = true;
1184 
1185 	return 0;
1186 
1187 failure:
1188 	mes_v10_1_hw_fini(adev);
1189 	return r;
1190 }
1191 
1192 static int mes_v10_1_hw_fini(void *handle)
1193 {
1194 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1195 
1196 	adev->mes.ring.sched.ready = false;
1197 
1198 	mes_v10_1_enable(adev, false);
1199 
1200 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1201 		mes_v10_1_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1202 		mes_v10_1_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1203 	}
1204 
1205 	return 0;
1206 }
1207 
1208 static int mes_v10_1_suspend(void *handle)
1209 {
1210 	int r;
1211 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1212 
1213 	r = amdgpu_mes_suspend(adev);
1214 	if (r)
1215 		return r;
1216 
1217 	return mes_v10_1_hw_fini(adev);
1218 }
1219 
1220 static int mes_v10_1_resume(void *handle)
1221 {
1222 	int r;
1223 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1224 
1225 	r = mes_v10_1_hw_init(adev);
1226 	if (r)
1227 		return r;
1228 
1229 	return amdgpu_mes_resume(adev);
1230 }
1231 
1232 static int mes_v10_0_late_init(void *handle)
1233 {
1234 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1235 
1236 	if (!amdgpu_in_reset(adev))
1237 		amdgpu_mes_self_test(adev);
1238 
1239 	return 0;
1240 }
1241 
1242 static const struct amd_ip_funcs mes_v10_1_ip_funcs = {
1243 	.name = "mes_v10_1",
1244 	.late_init = mes_v10_0_late_init,
1245 	.sw_init = mes_v10_1_sw_init,
1246 	.sw_fini = mes_v10_1_sw_fini,
1247 	.hw_init = mes_v10_1_hw_init,
1248 	.hw_fini = mes_v10_1_hw_fini,
1249 	.suspend = mes_v10_1_suspend,
1250 	.resume = mes_v10_1_resume,
1251 };
1252 
1253 const struct amdgpu_ip_block_version mes_v10_1_ip_block = {
1254 	.type = AMD_IP_BLOCK_TYPE_MES,
1255 	.major = 10,
1256 	.minor = 1,
1257 	.rev = 0,
1258 	.funcs = &mes_v10_1_ip_funcs,
1259 };
1260