xref: /linux/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c (revision e7d759f31ca295d589f7420719c311870bb3166f)
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_mes_2.bin");
37 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes1.bin");
38 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes.bin");
39 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes_2.bin");
40 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes1.bin");
41 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes.bin");
42 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes_2.bin");
43 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes1.bin");
44 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes.bin");
45 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes_2.bin");
46 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes1.bin");
47 MODULE_FIRMWARE("amdgpu/gc_11_0_4_mes.bin");
48 MODULE_FIRMWARE("amdgpu/gc_11_0_4_mes_2.bin");
49 MODULE_FIRMWARE("amdgpu/gc_11_0_4_mes1.bin");
50 MODULE_FIRMWARE("amdgpu/gc_11_5_0_mes_2.bin");
51 MODULE_FIRMWARE("amdgpu/gc_11_5_0_mes1.bin");
52 
53 
54 static int mes_v11_0_hw_fini(void *handle);
55 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev);
56 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev);
57 
58 #define MES_EOP_SIZE   2048
59 
60 static void mes_v11_0_ring_set_wptr(struct amdgpu_ring *ring)
61 {
62 	struct amdgpu_device *adev = ring->adev;
63 
64 	if (ring->use_doorbell) {
65 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
66 			     ring->wptr);
67 		WDOORBELL64(ring->doorbell_index, ring->wptr);
68 	} else {
69 		BUG();
70 	}
71 }
72 
73 static u64 mes_v11_0_ring_get_rptr(struct amdgpu_ring *ring)
74 {
75 	return *ring->rptr_cpu_addr;
76 }
77 
78 static u64 mes_v11_0_ring_get_wptr(struct amdgpu_ring *ring)
79 {
80 	u64 wptr;
81 
82 	if (ring->use_doorbell)
83 		wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
84 	else
85 		BUG();
86 	return wptr;
87 }
88 
89 static const struct amdgpu_ring_funcs mes_v11_0_ring_funcs = {
90 	.type = AMDGPU_RING_TYPE_MES,
91 	.align_mask = 1,
92 	.nop = 0,
93 	.support_64bit_ptrs = true,
94 	.get_rptr = mes_v11_0_ring_get_rptr,
95 	.get_wptr = mes_v11_0_ring_get_wptr,
96 	.set_wptr = mes_v11_0_ring_set_wptr,
97 	.insert_nop = amdgpu_ring_insert_nop,
98 };
99 
100 static int mes_v11_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
101 						    void *pkt, int size,
102 						    int api_status_off)
103 {
104 	int ndw = size / 4;
105 	signed long r;
106 	union MESAPI__ADD_QUEUE *x_pkt = pkt;
107 	struct MES_API_STATUS *api_status;
108 	struct amdgpu_device *adev = mes->adev;
109 	struct amdgpu_ring *ring = &mes->ring;
110 	unsigned long flags;
111 	signed long timeout = adev->usec_timeout;
112 
113 	if (amdgpu_emu_mode) {
114 		timeout *= 100;
115 	} else if (amdgpu_sriov_vf(adev)) {
116 		/* Worst case in sriov where all other 15 VF timeout, each VF needs about 600ms */
117 		timeout = 15 * 600 * 1000;
118 	}
119 	BUG_ON(size % 4 != 0);
120 
121 	spin_lock_irqsave(&mes->ring_lock, flags);
122 	if (amdgpu_ring_alloc(ring, ndw)) {
123 		spin_unlock_irqrestore(&mes->ring_lock, flags);
124 		return -ENOMEM;
125 	}
126 
127 	api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
128 	api_status->api_completion_fence_addr = mes->ring.fence_drv.gpu_addr;
129 	api_status->api_completion_fence_value = ++mes->ring.fence_drv.sync_seq;
130 
131 	amdgpu_ring_write_multiple(ring, pkt, ndw);
132 	amdgpu_ring_commit(ring);
133 	spin_unlock_irqrestore(&mes->ring_lock, flags);
134 
135 	DRM_DEBUG("MES msg=%d was emitted\n", x_pkt->header.opcode);
136 
137 	r = amdgpu_fence_wait_polling(ring, ring->fence_drv.sync_seq,
138 		      timeout);
139 	if (r < 1) {
140 		DRM_ERROR("MES failed to response msg=%d\n",
141 			  x_pkt->header.opcode);
142 
143 		while (halt_if_hws_hang)
144 			schedule();
145 
146 		return -ETIMEDOUT;
147 	}
148 
149 	return 0;
150 }
151 
152 static int convert_to_mes_queue_type(int queue_type)
153 {
154 	if (queue_type == AMDGPU_RING_TYPE_GFX)
155 		return MES_QUEUE_TYPE_GFX;
156 	else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
157 		return MES_QUEUE_TYPE_COMPUTE;
158 	else if (queue_type == AMDGPU_RING_TYPE_SDMA)
159 		return MES_QUEUE_TYPE_SDMA;
160 	else
161 		BUG();
162 	return -1;
163 }
164 
165 static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
166 				  struct mes_add_queue_input *input)
167 {
168 	struct amdgpu_device *adev = mes->adev;
169 	union MESAPI__ADD_QUEUE mes_add_queue_pkt;
170 	struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB(0)];
171 	uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
172 
173 	memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
174 
175 	mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
176 	mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
177 	mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
178 
179 	mes_add_queue_pkt.process_id = input->process_id;
180 	mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
181 	mes_add_queue_pkt.process_va_start = input->process_va_start;
182 	mes_add_queue_pkt.process_va_end = input->process_va_end;
183 	mes_add_queue_pkt.process_quantum = input->process_quantum;
184 	mes_add_queue_pkt.process_context_addr = input->process_context_addr;
185 	mes_add_queue_pkt.gang_quantum = input->gang_quantum;
186 	mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
187 	mes_add_queue_pkt.inprocess_gang_priority =
188 		input->inprocess_gang_priority;
189 	mes_add_queue_pkt.gang_global_priority_level =
190 		input->gang_global_priority_level;
191 	mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
192 	mes_add_queue_pkt.mqd_addr = input->mqd_addr;
193 
194 	if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
195 			AMDGPU_MES_API_VERSION_SHIFT) >= 2)
196 		mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr;
197 	else
198 		mes_add_queue_pkt.wptr_addr = input->wptr_addr;
199 
200 	mes_add_queue_pkt.queue_type =
201 		convert_to_mes_queue_type(input->queue_type);
202 	mes_add_queue_pkt.paging = input->paging;
203 	mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
204 	mes_add_queue_pkt.gws_base = input->gws_base;
205 	mes_add_queue_pkt.gws_size = input->gws_size;
206 	mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
207 	mes_add_queue_pkt.tma_addr = input->tma_addr;
208 	mes_add_queue_pkt.trap_en = input->trap_en;
209 	mes_add_queue_pkt.skip_process_ctx_clear = input->skip_process_ctx_clear;
210 	mes_add_queue_pkt.is_kfd_process = input->is_kfd_process;
211 
212 	/* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
213 	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
214 	mes_add_queue_pkt.gds_size = input->queue_size;
215 
216 	mes_add_queue_pkt.exclusively_scheduled = input->exclusively_scheduled;
217 
218 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
219 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
220 			offsetof(union MESAPI__ADD_QUEUE, api_status));
221 }
222 
223 static int mes_v11_0_remove_hw_queue(struct amdgpu_mes *mes,
224 				     struct mes_remove_queue_input *input)
225 {
226 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
227 
228 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
229 
230 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
231 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
232 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
233 
234 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
235 	mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
236 
237 	return mes_v11_0_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_v11_0_unmap_legacy_queue(struct amdgpu_mes *mes,
243 			struct mes_unmap_legacy_queue_input *input)
244 {
245 	union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
246 
247 	memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
248 
249 	mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
250 	mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
251 	mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
252 
253 	mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
254 	mes_remove_queue_pkt.gang_context_addr = 0;
255 
256 	mes_remove_queue_pkt.pipe_id = input->pipe_id;
257 	mes_remove_queue_pkt.queue_id = input->queue_id;
258 
259 	if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
260 		mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
261 		mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
262 		mes_remove_queue_pkt.tf_data =
263 			lower_32_bits(input->trail_fence_data);
264 	} else {
265 		mes_remove_queue_pkt.unmap_legacy_queue = 1;
266 		mes_remove_queue_pkt.queue_type =
267 			convert_to_mes_queue_type(input->queue_type);
268 	}
269 
270 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
271 			&mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
272 			offsetof(union MESAPI__REMOVE_QUEUE, api_status));
273 }
274 
275 static int mes_v11_0_suspend_gang(struct amdgpu_mes *mes,
276 				  struct mes_suspend_gang_input *input)
277 {
278 	return 0;
279 }
280 
281 static int mes_v11_0_resume_gang(struct amdgpu_mes *mes,
282 				 struct mes_resume_gang_input *input)
283 {
284 	return 0;
285 }
286 
287 static int mes_v11_0_query_sched_status(struct amdgpu_mes *mes)
288 {
289 	union MESAPI__QUERY_MES_STATUS mes_status_pkt;
290 
291 	memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
292 
293 	mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
294 	mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
295 	mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
296 
297 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
298 			&mes_status_pkt, sizeof(mes_status_pkt),
299 			offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
300 }
301 
302 static int mes_v11_0_misc_op(struct amdgpu_mes *mes,
303 			     struct mes_misc_op_input *input)
304 {
305 	union MESAPI__MISC misc_pkt;
306 
307 	memset(&misc_pkt, 0, sizeof(misc_pkt));
308 
309 	misc_pkt.header.type = MES_API_TYPE_SCHEDULER;
310 	misc_pkt.header.opcode = MES_SCH_API_MISC;
311 	misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
312 
313 	switch (input->op) {
314 	case MES_MISC_OP_READ_REG:
315 		misc_pkt.opcode = MESAPI_MISC__READ_REG;
316 		misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset;
317 		misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr;
318 		break;
319 	case MES_MISC_OP_WRITE_REG:
320 		misc_pkt.opcode = MESAPI_MISC__WRITE_REG;
321 		misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset;
322 		misc_pkt.write_reg.reg_value = input->write_reg.reg_value;
323 		break;
324 	case MES_MISC_OP_WRM_REG_WAIT:
325 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
326 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM;
327 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
328 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
329 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
330 		misc_pkt.wait_reg_mem.reg_offset2 = 0;
331 		break;
332 	case MES_MISC_OP_WRM_REG_WR_WAIT:
333 		misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
334 		misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG;
335 		misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
336 		misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
337 		misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
338 		misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1;
339 		break;
340 	case MES_MISC_OP_SET_SHADER_DEBUGGER:
341 		misc_pkt.opcode = MESAPI_MISC__SET_SHADER_DEBUGGER;
342 		misc_pkt.set_shader_debugger.process_context_addr =
343 				input->set_shader_debugger.process_context_addr;
344 		misc_pkt.set_shader_debugger.flags.u32all =
345 				input->set_shader_debugger.flags.u32all;
346 		misc_pkt.set_shader_debugger.spi_gdbg_per_vmid_cntl =
347 				input->set_shader_debugger.spi_gdbg_per_vmid_cntl;
348 		memcpy(misc_pkt.set_shader_debugger.tcp_watch_cntl,
349 				input->set_shader_debugger.tcp_watch_cntl,
350 				sizeof(misc_pkt.set_shader_debugger.tcp_watch_cntl));
351 		misc_pkt.set_shader_debugger.trap_en = input->set_shader_debugger.trap_en;
352 		break;
353 	default:
354 		DRM_ERROR("unsupported misc op (%d) \n", input->op);
355 		return -EINVAL;
356 	}
357 
358 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
359 			&misc_pkt, sizeof(misc_pkt),
360 			offsetof(union MESAPI__MISC, api_status));
361 }
362 
363 static int mes_v11_0_set_hw_resources(struct amdgpu_mes *mes)
364 {
365 	int i;
366 	struct amdgpu_device *adev = mes->adev;
367 	union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
368 
369 	memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
370 
371 	mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
372 	mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
373 	mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
374 
375 	mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
376 	mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
377 	mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
378 	mes_set_hw_res_pkt.paging_vmid = 0;
379 	mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr = mes->sch_ctx_gpu_addr;
380 	mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
381 		mes->query_status_fence_gpu_addr;
382 
383 	for (i = 0; i < MAX_COMPUTE_PIPES; i++)
384 		mes_set_hw_res_pkt.compute_hqd_mask[i] =
385 			mes->compute_hqd_mask[i];
386 
387 	for (i = 0; i < MAX_GFX_PIPES; i++)
388 		mes_set_hw_res_pkt.gfx_hqd_mask[i] = mes->gfx_hqd_mask[i];
389 
390 	for (i = 0; i < MAX_SDMA_PIPES; i++)
391 		mes_set_hw_res_pkt.sdma_hqd_mask[i] = mes->sdma_hqd_mask[i];
392 
393 	for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
394 		mes_set_hw_res_pkt.aggregated_doorbells[i] =
395 			mes->aggregated_doorbells[i];
396 
397 	for (i = 0; i < 5; i++) {
398 		mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
399 		mes_set_hw_res_pkt.mmhub_base[i] =
400 				adev->reg_offset[MMHUB_HWIP][0][i];
401 		mes_set_hw_res_pkt.osssys_base[i] =
402 		adev->reg_offset[OSSSYS_HWIP][0][i];
403 	}
404 
405 	mes_set_hw_res_pkt.disable_reset = 1;
406 	mes_set_hw_res_pkt.disable_mes_log = 1;
407 	mes_set_hw_res_pkt.use_different_vmid_compute = 1;
408 	mes_set_hw_res_pkt.enable_reg_active_poll = 1;
409 	mes_set_hw_res_pkt.enable_level_process_quantum_check = 1;
410 	mes_set_hw_res_pkt.oversubscription_timer = 50;
411 	mes_set_hw_res_pkt.enable_mes_event_int_logging = 1;
412 	mes_set_hw_res_pkt.event_intr_history_gpu_mc_ptr = mes->event_log_gpu_addr;
413 
414 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
415 			&mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
416 			offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
417 }
418 
419 static const struct amdgpu_mes_funcs mes_v11_0_funcs = {
420 	.add_hw_queue = mes_v11_0_add_hw_queue,
421 	.remove_hw_queue = mes_v11_0_remove_hw_queue,
422 	.unmap_legacy_queue = mes_v11_0_unmap_legacy_queue,
423 	.suspend_gang = mes_v11_0_suspend_gang,
424 	.resume_gang = mes_v11_0_resume_gang,
425 	.misc_op = mes_v11_0_misc_op,
426 };
427 
428 static int mes_v11_0_allocate_ucode_buffer(struct amdgpu_device *adev,
429 					   enum admgpu_mes_pipe pipe)
430 {
431 	int r;
432 	const struct mes_firmware_header_v1_0 *mes_hdr;
433 	const __le32 *fw_data;
434 	unsigned fw_size;
435 
436 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
437 		adev->mes.fw[pipe]->data;
438 
439 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
440 		   le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
441 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
442 
443 	r = amdgpu_bo_create_reserved(adev, fw_size,
444 				      PAGE_SIZE,
445 				      AMDGPU_GEM_DOMAIN_VRAM |
446 				      AMDGPU_GEM_DOMAIN_GTT,
447 				      &adev->mes.ucode_fw_obj[pipe],
448 				      &adev->mes.ucode_fw_gpu_addr[pipe],
449 				      (void **)&adev->mes.ucode_fw_ptr[pipe]);
450 	if (r) {
451 		dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
452 		return r;
453 	}
454 
455 	memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
456 
457 	amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
458 	amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
459 
460 	return 0;
461 }
462 
463 static int mes_v11_0_allocate_ucode_data_buffer(struct amdgpu_device *adev,
464 						enum admgpu_mes_pipe pipe)
465 {
466 	int r;
467 	const struct mes_firmware_header_v1_0 *mes_hdr;
468 	const __le32 *fw_data;
469 	unsigned fw_size;
470 
471 	mes_hdr = (const struct mes_firmware_header_v1_0 *)
472 		adev->mes.fw[pipe]->data;
473 
474 	fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
475 		   le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
476 	fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
477 
478 	r = amdgpu_bo_create_reserved(adev, fw_size,
479 				      64 * 1024,
480 				      AMDGPU_GEM_DOMAIN_VRAM |
481 				      AMDGPU_GEM_DOMAIN_GTT,
482 				      &adev->mes.data_fw_obj[pipe],
483 				      &adev->mes.data_fw_gpu_addr[pipe],
484 				      (void **)&adev->mes.data_fw_ptr[pipe]);
485 	if (r) {
486 		dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
487 		return r;
488 	}
489 
490 	memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
491 
492 	amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
493 	amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
494 
495 	return 0;
496 }
497 
498 static void mes_v11_0_free_ucode_buffers(struct amdgpu_device *adev,
499 					 enum admgpu_mes_pipe pipe)
500 {
501 	amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
502 			      &adev->mes.data_fw_gpu_addr[pipe],
503 			      (void **)&adev->mes.data_fw_ptr[pipe]);
504 
505 	amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
506 			      &adev->mes.ucode_fw_gpu_addr[pipe],
507 			      (void **)&adev->mes.ucode_fw_ptr[pipe]);
508 }
509 
510 static void mes_v11_0_enable(struct amdgpu_device *adev, bool enable)
511 {
512 	uint64_t ucode_addr;
513 	uint32_t pipe, data = 0;
514 
515 	if (enable) {
516 		data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
517 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
518 		data = REG_SET_FIELD(data, CP_MES_CNTL,
519 			     MES_PIPE1_RESET, adev->enable_mes_kiq ? 1 : 0);
520 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
521 
522 		mutex_lock(&adev->srbm_mutex);
523 		for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
524 			if (!adev->enable_mes_kiq &&
525 			    pipe == AMDGPU_MES_KIQ_PIPE)
526 				continue;
527 
528 			soc21_grbm_select(adev, 3, pipe, 0, 0);
529 
530 			ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
531 			WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
532 				     lower_32_bits(ucode_addr));
533 			WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
534 				     upper_32_bits(ucode_addr));
535 		}
536 		soc21_grbm_select(adev, 0, 0, 0, 0);
537 		mutex_unlock(&adev->srbm_mutex);
538 
539 		/* unhalt MES and activate pipe0 */
540 		data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
541 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE,
542 				     adev->enable_mes_kiq ? 1 : 0);
543 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
544 
545 		if (amdgpu_emu_mode)
546 			msleep(100);
547 		else
548 			udelay(50);
549 	} else {
550 		data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
551 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
552 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
553 		data = REG_SET_FIELD(data, CP_MES_CNTL,
554 				     MES_INVALIDATE_ICACHE, 1);
555 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
556 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET,
557 				     adev->enable_mes_kiq ? 1 : 0);
558 		data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
559 		WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
560 	}
561 }
562 
563 /* This function is for backdoor MES firmware */
564 static int mes_v11_0_load_microcode(struct amdgpu_device *adev,
565 				    enum admgpu_mes_pipe pipe, bool prime_icache)
566 {
567 	int r;
568 	uint32_t data;
569 	uint64_t ucode_addr;
570 
571 	mes_v11_0_enable(adev, false);
572 
573 	if (!adev->mes.fw[pipe])
574 		return -EINVAL;
575 
576 	r = mes_v11_0_allocate_ucode_buffer(adev, pipe);
577 	if (r)
578 		return r;
579 
580 	r = mes_v11_0_allocate_ucode_data_buffer(adev, pipe);
581 	if (r) {
582 		mes_v11_0_free_ucode_buffers(adev, pipe);
583 		return r;
584 	}
585 
586 	mutex_lock(&adev->srbm_mutex);
587 	/* me=3, pipe=0, queue=0 */
588 	soc21_grbm_select(adev, 3, pipe, 0, 0);
589 
590 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0);
591 
592 	/* set ucode start address */
593 	ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
594 	WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
595 		     lower_32_bits(ucode_addr));
596 	WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
597 		     upper_32_bits(ucode_addr));
598 
599 	/* set ucode fimrware address */
600 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO,
601 		     lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
602 	WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI,
603 		     upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
604 
605 	/* set ucode instruction cache boundary to 2M-1 */
606 	WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF);
607 
608 	/* set ucode data firmware address */
609 	WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO,
610 		     lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
611 	WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI,
612 		     upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
613 
614 	/* Set 0x3FFFF (256K-1) to CP_MES_MDBOUND_LO */
615 	WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x3FFFF);
616 
617 	if (prime_icache) {
618 		/* invalidate ICACHE */
619 		data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
620 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
621 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
622 		WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
623 
624 		/* prime the ICACHE. */
625 		data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
626 		data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
627 		WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
628 	}
629 
630 	soc21_grbm_select(adev, 0, 0, 0, 0);
631 	mutex_unlock(&adev->srbm_mutex);
632 
633 	return 0;
634 }
635 
636 static int mes_v11_0_allocate_eop_buf(struct amdgpu_device *adev,
637 				      enum admgpu_mes_pipe pipe)
638 {
639 	int r;
640 	u32 *eop;
641 
642 	r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
643 			      AMDGPU_GEM_DOMAIN_GTT,
644 			      &adev->mes.eop_gpu_obj[pipe],
645 			      &adev->mes.eop_gpu_addr[pipe],
646 			      (void **)&eop);
647 	if (r) {
648 		dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
649 		return r;
650 	}
651 
652 	memset(eop, 0,
653 	       adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
654 
655 	amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
656 	amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
657 
658 	return 0;
659 }
660 
661 static int mes_v11_0_mqd_init(struct amdgpu_ring *ring)
662 {
663 	struct v11_compute_mqd *mqd = ring->mqd_ptr;
664 	uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
665 	uint32_t tmp;
666 
667 	memset(mqd, 0, sizeof(*mqd));
668 
669 	mqd->header = 0xC0310800;
670 	mqd->compute_pipelinestat_enable = 0x00000001;
671 	mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
672 	mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
673 	mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
674 	mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
675 	mqd->compute_misc_reserved = 0x00000007;
676 
677 	eop_base_addr = ring->eop_gpu_addr >> 8;
678 
679 	/* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
680 	tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
681 	tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
682 			(order_base_2(MES_EOP_SIZE / 4) - 1));
683 
684 	mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
685 	mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
686 	mqd->cp_hqd_eop_control = tmp;
687 
688 	/* disable the queue if it's active */
689 	ring->wptr = 0;
690 	mqd->cp_hqd_pq_rptr = 0;
691 	mqd->cp_hqd_pq_wptr_lo = 0;
692 	mqd->cp_hqd_pq_wptr_hi = 0;
693 
694 	/* set the pointer to the MQD */
695 	mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
696 	mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
697 
698 	/* set MQD vmid to 0 */
699 	tmp = regCP_MQD_CONTROL_DEFAULT;
700 	tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
701 	mqd->cp_mqd_control = tmp;
702 
703 	/* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
704 	hqd_gpu_addr = ring->gpu_addr >> 8;
705 	mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
706 	mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
707 
708 	/* set the wb address whether it's enabled or not */
709 	wb_gpu_addr = ring->rptr_gpu_addr;
710 	mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
711 	mqd->cp_hqd_pq_rptr_report_addr_hi =
712 		upper_32_bits(wb_gpu_addr) & 0xffff;
713 
714 	/* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
715 	wb_gpu_addr = ring->wptr_gpu_addr;
716 	mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
717 	mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
718 
719 	/* set up the HQD, this is similar to CP_RB0_CNTL */
720 	tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
721 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
722 			    (order_base_2(ring->ring_size / 4) - 1));
723 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
724 			    ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
725 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
726 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
727 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
728 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
729 	tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
730 	mqd->cp_hqd_pq_control = tmp;
731 
732 	/* enable doorbell */
733 	tmp = 0;
734 	if (ring->use_doorbell) {
735 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
736 				    DOORBELL_OFFSET, ring->doorbell_index);
737 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
738 				    DOORBELL_EN, 1);
739 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
740 				    DOORBELL_SOURCE, 0);
741 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
742 				    DOORBELL_HIT, 0);
743 	} else
744 		tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
745 				    DOORBELL_EN, 0);
746 	mqd->cp_hqd_pq_doorbell_control = tmp;
747 
748 	mqd->cp_hqd_vmid = 0;
749 	/* activate the queue */
750 	mqd->cp_hqd_active = 1;
751 
752 	tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
753 	tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
754 			    PRELOAD_SIZE, 0x55);
755 	mqd->cp_hqd_persistent_state = tmp;
756 
757 	mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT;
758 	mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
759 	mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
760 
761 	amdgpu_device_flush_hdp(ring->adev, NULL);
762 	return 0;
763 }
764 
765 static void mes_v11_0_queue_init_register(struct amdgpu_ring *ring)
766 {
767 	struct v11_compute_mqd *mqd = ring->mqd_ptr;
768 	struct amdgpu_device *adev = ring->adev;
769 	uint32_t data = 0;
770 
771 	mutex_lock(&adev->srbm_mutex);
772 	soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
773 
774 	/* set CP_HQD_VMID.VMID = 0. */
775 	data = RREG32_SOC15(GC, 0, regCP_HQD_VMID);
776 	data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
777 	WREG32_SOC15(GC, 0, regCP_HQD_VMID, data);
778 
779 	/* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
780 	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
781 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
782 			     DOORBELL_EN, 0);
783 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
784 
785 	/* set CP_MQD_BASE_ADDR/HI with the MQD base address */
786 	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
787 	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
788 
789 	/* set CP_MQD_CONTROL.VMID=0 */
790 	data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL);
791 	data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
792 	WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0);
793 
794 	/* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
795 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
796 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
797 
798 	/* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
799 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
800 		     mqd->cp_hqd_pq_rptr_report_addr_lo);
801 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
802 		     mqd->cp_hqd_pq_rptr_report_addr_hi);
803 
804 	/* set CP_HQD_PQ_CONTROL */
805 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
806 
807 	/* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
808 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
809 		     mqd->cp_hqd_pq_wptr_poll_addr_lo);
810 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
811 		     mqd->cp_hqd_pq_wptr_poll_addr_hi);
812 
813 	/* set CP_HQD_PQ_DOORBELL_CONTROL */
814 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
815 		     mqd->cp_hqd_pq_doorbell_control);
816 
817 	/* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
818 	WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
819 
820 	/* set CP_HQD_ACTIVE.ACTIVE=1 */
821 	WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active);
822 
823 	soc21_grbm_select(adev, 0, 0, 0, 0);
824 	mutex_unlock(&adev->srbm_mutex);
825 }
826 
827 static int mes_v11_0_kiq_enable_queue(struct amdgpu_device *adev)
828 {
829 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
830 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
831 	int r;
832 
833 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
834 		return -EINVAL;
835 
836 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
837 	if (r) {
838 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
839 		return r;
840 	}
841 
842 	kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring);
843 
844 	return amdgpu_ring_test_helper(kiq_ring);
845 }
846 
847 static int mes_v11_0_queue_init(struct amdgpu_device *adev,
848 				enum admgpu_mes_pipe pipe)
849 {
850 	struct amdgpu_ring *ring;
851 	int r;
852 
853 	if (pipe == AMDGPU_MES_KIQ_PIPE)
854 		ring = &adev->gfx.kiq[0].ring;
855 	else if (pipe == AMDGPU_MES_SCHED_PIPE)
856 		ring = &adev->mes.ring;
857 	else
858 		BUG();
859 
860 	if ((pipe == AMDGPU_MES_SCHED_PIPE) &&
861 	    (amdgpu_in_reset(adev) || adev->in_suspend)) {
862 		*(ring->wptr_cpu_addr) = 0;
863 		*(ring->rptr_cpu_addr) = 0;
864 		amdgpu_ring_clear_ring(ring);
865 	}
866 
867 	r = mes_v11_0_mqd_init(ring);
868 	if (r)
869 		return r;
870 
871 	if (pipe == AMDGPU_MES_SCHED_PIPE) {
872 		r = mes_v11_0_kiq_enable_queue(adev);
873 		if (r)
874 			return r;
875 	} else {
876 		mes_v11_0_queue_init_register(ring);
877 	}
878 
879 	/* get MES scheduler/KIQ versions */
880 	mutex_lock(&adev->srbm_mutex);
881 	soc21_grbm_select(adev, 3, pipe, 0, 0);
882 
883 	if (pipe == AMDGPU_MES_SCHED_PIPE)
884 		adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
885 	else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
886 		adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
887 
888 	soc21_grbm_select(adev, 0, 0, 0, 0);
889 	mutex_unlock(&adev->srbm_mutex);
890 
891 	return 0;
892 }
893 
894 static int mes_v11_0_ring_init(struct amdgpu_device *adev)
895 {
896 	struct amdgpu_ring *ring;
897 
898 	ring = &adev->mes.ring;
899 
900 	ring->funcs = &mes_v11_0_ring_funcs;
901 
902 	ring->me = 3;
903 	ring->pipe = 0;
904 	ring->queue = 0;
905 
906 	ring->ring_obj = NULL;
907 	ring->use_doorbell = true;
908 	ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
909 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_SCHED_PIPE];
910 	ring->no_scheduler = true;
911 	sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
912 
913 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
914 				AMDGPU_RING_PRIO_DEFAULT, NULL);
915 }
916 
917 static int mes_v11_0_kiq_ring_init(struct amdgpu_device *adev)
918 {
919 	struct amdgpu_ring *ring;
920 
921 	spin_lock_init(&adev->gfx.kiq[0].ring_lock);
922 
923 	ring = &adev->gfx.kiq[0].ring;
924 
925 	ring->me = 3;
926 	ring->pipe = 1;
927 	ring->queue = 0;
928 
929 	ring->adev = NULL;
930 	ring->ring_obj = NULL;
931 	ring->use_doorbell = true;
932 	ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
933 	ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
934 	ring->no_scheduler = true;
935 	sprintf(ring->name, "mes_kiq_%d.%d.%d",
936 		ring->me, ring->pipe, ring->queue);
937 
938 	return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
939 				AMDGPU_RING_PRIO_DEFAULT, NULL);
940 }
941 
942 static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev,
943 				 enum admgpu_mes_pipe pipe)
944 {
945 	int r, mqd_size = sizeof(struct v11_compute_mqd);
946 	struct amdgpu_ring *ring;
947 
948 	if (pipe == AMDGPU_MES_KIQ_PIPE)
949 		ring = &adev->gfx.kiq[0].ring;
950 	else if (pipe == AMDGPU_MES_SCHED_PIPE)
951 		ring = &adev->mes.ring;
952 	else
953 		BUG();
954 
955 	if (ring->mqd_obj)
956 		return 0;
957 
958 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
959 				    AMDGPU_GEM_DOMAIN_VRAM |
960 				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
961 				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
962 	if (r) {
963 		dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
964 		return r;
965 	}
966 
967 	memset(ring->mqd_ptr, 0, mqd_size);
968 
969 	/* prepare MQD backup */
970 	adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
971 	if (!adev->mes.mqd_backup[pipe]) {
972 		dev_warn(adev->dev,
973 			 "no memory to create MQD backup for ring %s\n",
974 			 ring->name);
975 		return -ENOMEM;
976 	}
977 
978 	return 0;
979 }
980 
981 static int mes_v11_0_sw_init(void *handle)
982 {
983 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
984 	int pipe, r;
985 
986 	adev->mes.funcs = &mes_v11_0_funcs;
987 	adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init;
988 	adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini;
989 
990 	r = amdgpu_mes_init(adev);
991 	if (r)
992 		return r;
993 
994 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
995 		if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
996 			continue;
997 
998 		r = mes_v11_0_allocate_eop_buf(adev, pipe);
999 		if (r)
1000 			return r;
1001 
1002 		r = mes_v11_0_mqd_sw_init(adev, pipe);
1003 		if (r)
1004 			return r;
1005 	}
1006 
1007 	if (adev->enable_mes_kiq) {
1008 		r = mes_v11_0_kiq_ring_init(adev);
1009 		if (r)
1010 			return r;
1011 	}
1012 
1013 	r = mes_v11_0_ring_init(adev);
1014 	if (r)
1015 		return r;
1016 
1017 	return 0;
1018 }
1019 
1020 static int mes_v11_0_sw_fini(void *handle)
1021 {
1022 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1023 	int pipe;
1024 
1025 	amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
1026 	amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
1027 
1028 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1029 		kfree(adev->mes.mqd_backup[pipe]);
1030 
1031 		amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1032 				      &adev->mes.eop_gpu_addr[pipe],
1033 				      NULL);
1034 		amdgpu_ucode_release(&adev->mes.fw[pipe]);
1035 	}
1036 
1037 	amdgpu_bo_free_kernel(&adev->gfx.kiq[0].ring.mqd_obj,
1038 			      &adev->gfx.kiq[0].ring.mqd_gpu_addr,
1039 			      &adev->gfx.kiq[0].ring.mqd_ptr);
1040 
1041 	amdgpu_bo_free_kernel(&adev->mes.ring.mqd_obj,
1042 			      &adev->mes.ring.mqd_gpu_addr,
1043 			      &adev->mes.ring.mqd_ptr);
1044 
1045 	amdgpu_ring_fini(&adev->gfx.kiq[0].ring);
1046 	amdgpu_ring_fini(&adev->mes.ring);
1047 
1048 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1049 		mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1050 		mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1051 	}
1052 
1053 	amdgpu_mes_fini(adev);
1054 	return 0;
1055 }
1056 
1057 static void mes_v11_0_kiq_dequeue(struct amdgpu_ring *ring)
1058 {
1059 	uint32_t data;
1060 	int i;
1061 	struct amdgpu_device *adev = ring->adev;
1062 
1063 	mutex_lock(&adev->srbm_mutex);
1064 	soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
1065 
1066 	/* disable the queue if it's active */
1067 	if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
1068 		WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
1069 		for (i = 0; i < adev->usec_timeout; i++) {
1070 			if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
1071 				break;
1072 			udelay(1);
1073 		}
1074 	}
1075 	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1076 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1077 				DOORBELL_EN, 0);
1078 	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1079 				DOORBELL_HIT, 1);
1080 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1081 
1082 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0);
1083 
1084 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0);
1085 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0);
1086 	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0);
1087 
1088 	soc21_grbm_select(adev, 0, 0, 0, 0);
1089 	mutex_unlock(&adev->srbm_mutex);
1090 }
1091 
1092 static void mes_v11_0_kiq_setting(struct amdgpu_ring *ring)
1093 {
1094 	uint32_t tmp;
1095 	struct amdgpu_device *adev = ring->adev;
1096 
1097 	/* tell RLC which is KIQ queue */
1098 	tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1099 	tmp &= 0xffffff00;
1100 	tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1101 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1102 	tmp |= 0x80;
1103 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1104 }
1105 
1106 static void mes_v11_0_kiq_clear(struct amdgpu_device *adev)
1107 {
1108 	uint32_t tmp;
1109 
1110 	/* tell RLC which is KIQ dequeue */
1111 	tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1112 	tmp &= ~RLC_CP_SCHEDULERS__scheduler0_MASK;
1113 	WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1114 }
1115 
1116 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev)
1117 {
1118 	int r = 0;
1119 
1120 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1121 
1122 		r = mes_v11_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE, false);
1123 		if (r) {
1124 			DRM_ERROR("failed to load MES fw, r=%d\n", r);
1125 			return r;
1126 		}
1127 
1128 		r = mes_v11_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE, true);
1129 		if (r) {
1130 			DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1131 			return r;
1132 		}
1133 
1134 	}
1135 
1136 	mes_v11_0_enable(adev, true);
1137 
1138 	mes_v11_0_kiq_setting(&adev->gfx.kiq[0].ring);
1139 
1140 	r = mes_v11_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE);
1141 	if (r)
1142 		goto failure;
1143 
1144 	return r;
1145 
1146 failure:
1147 	mes_v11_0_hw_fini(adev);
1148 	return r;
1149 }
1150 
1151 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev)
1152 {
1153 	if (adev->mes.ring.sched.ready) {
1154 		mes_v11_0_kiq_dequeue(&adev->mes.ring);
1155 		adev->mes.ring.sched.ready = false;
1156 	}
1157 
1158 	if (amdgpu_sriov_vf(adev)) {
1159 		mes_v11_0_kiq_dequeue(&adev->gfx.kiq[0].ring);
1160 		mes_v11_0_kiq_clear(adev);
1161 	}
1162 
1163 	mes_v11_0_enable(adev, false);
1164 
1165 	return 0;
1166 }
1167 
1168 static int mes_v11_0_hw_init(void *handle)
1169 {
1170 	int r;
1171 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1172 
1173 	if (!adev->enable_mes_kiq) {
1174 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1175 			r = mes_v11_0_load_microcode(adev,
1176 					     AMDGPU_MES_SCHED_PIPE, true);
1177 			if (r) {
1178 				DRM_ERROR("failed to MES fw, r=%d\n", r);
1179 				return r;
1180 			}
1181 		}
1182 
1183 		mes_v11_0_enable(adev, true);
1184 	}
1185 
1186 	r = mes_v11_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE);
1187 	if (r)
1188 		goto failure;
1189 
1190 	r = mes_v11_0_set_hw_resources(&adev->mes);
1191 	if (r)
1192 		goto failure;
1193 
1194 	r = mes_v11_0_query_sched_status(&adev->mes);
1195 	if (r) {
1196 		DRM_ERROR("MES is busy\n");
1197 		goto failure;
1198 	}
1199 
1200 	/*
1201 	 * Disable KIQ ring usage from the driver once MES is enabled.
1202 	 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1203 	 * with MES enabled.
1204 	 */
1205 	adev->gfx.kiq[0].ring.sched.ready = false;
1206 	adev->mes.ring.sched.ready = true;
1207 
1208 	return 0;
1209 
1210 failure:
1211 	mes_v11_0_hw_fini(adev);
1212 	return r;
1213 }
1214 
1215 static int mes_v11_0_hw_fini(void *handle)
1216 {
1217 	return 0;
1218 }
1219 
1220 static int mes_v11_0_suspend(void *handle)
1221 {
1222 	int r;
1223 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1224 
1225 	r = amdgpu_mes_suspend(adev);
1226 	if (r)
1227 		return r;
1228 
1229 	return mes_v11_0_hw_fini(adev);
1230 }
1231 
1232 static int mes_v11_0_resume(void *handle)
1233 {
1234 	int r;
1235 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1236 
1237 	r = mes_v11_0_hw_init(adev);
1238 	if (r)
1239 		return r;
1240 
1241 	return amdgpu_mes_resume(adev);
1242 }
1243 
1244 static int mes_v11_0_early_init(void *handle)
1245 {
1246 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1247 	int pipe, r;
1248 
1249 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1250 		if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
1251 			continue;
1252 		r = amdgpu_mes_init_microcode(adev, pipe);
1253 		if (r)
1254 			return r;
1255 	}
1256 
1257 	return 0;
1258 }
1259 
1260 static int mes_v11_0_late_init(void *handle)
1261 {
1262 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1263 
1264 	/* it's only intended for use in mes_self_test case, not for s0ix and reset */
1265 	if (!amdgpu_in_reset(adev) && !adev->in_s0ix && !adev->in_suspend &&
1266 	    (amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(11, 0, 3)))
1267 		amdgpu_mes_self_test(adev);
1268 
1269 	return 0;
1270 }
1271 
1272 static const struct amd_ip_funcs mes_v11_0_ip_funcs = {
1273 	.name = "mes_v11_0",
1274 	.early_init = mes_v11_0_early_init,
1275 	.late_init = mes_v11_0_late_init,
1276 	.sw_init = mes_v11_0_sw_init,
1277 	.sw_fini = mes_v11_0_sw_fini,
1278 	.hw_init = mes_v11_0_hw_init,
1279 	.hw_fini = mes_v11_0_hw_fini,
1280 	.suspend = mes_v11_0_suspend,
1281 	.resume = mes_v11_0_resume,
1282 };
1283 
1284 const struct amdgpu_ip_block_version mes_v11_0_ip_block = {
1285 	.type = AMD_IP_BLOCK_TYPE_MES,
1286 	.major = 11,
1287 	.minor = 0,
1288 	.rev = 0,
1289 	.funcs = &mes_v11_0_ip_funcs,
1290 };
1291