1 /*
2 * Copyright 2023 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_12_0_0_offset.h"
30 #include "gc/gc_12_0_0_sh_mask.h"
31 #include "gc/gc_11_0_0_default.h"
32 #include "v12_structs.h"
33 #include "mes_v12_api_def.h"
34
35 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mes.bin");
36 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mes1.bin");
37 MODULE_FIRMWARE("amdgpu/gc_12_0_0_uni_mes.bin");
38 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mes.bin");
39 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mes1.bin");
40 MODULE_FIRMWARE("amdgpu/gc_12_0_1_uni_mes.bin");
41
42 static int mes_v12_0_hw_init(struct amdgpu_ip_block *ip_block);
43 static int mes_v12_0_hw_fini(struct amdgpu_ip_block *ip_block);
44 static int mes_v12_0_kiq_hw_init(struct amdgpu_device *adev);
45 static int mes_v12_0_kiq_hw_fini(struct amdgpu_device *adev);
46
47 #define MES_EOP_SIZE 2048
48
mes_v12_0_ring_set_wptr(struct amdgpu_ring * ring)49 static void mes_v12_0_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
mes_v12_0_ring_get_rptr(struct amdgpu_ring * ring)62 static u64 mes_v12_0_ring_get_rptr(struct amdgpu_ring *ring)
63 {
64 return *ring->rptr_cpu_addr;
65 }
66
mes_v12_0_ring_get_wptr(struct amdgpu_ring * ring)67 static u64 mes_v12_0_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_v12_0_ring_funcs = {
79 .type = AMDGPU_RING_TYPE_MES,
80 .align_mask = 1,
81 .nop = 0,
82 .support_64bit_ptrs = true,
83 .get_rptr = mes_v12_0_ring_get_rptr,
84 .get_wptr = mes_v12_0_ring_get_wptr,
85 .set_wptr = mes_v12_0_ring_set_wptr,
86 .insert_nop = amdgpu_ring_insert_nop,
87 };
88
89 static const char *mes_v12_0_opcodes[] = {
90 "SET_HW_RSRC",
91 "SET_SCHEDULING_CONFIG",
92 "ADD_QUEUE",
93 "REMOVE_QUEUE",
94 "PERFORM_YIELD",
95 "SET_GANG_PRIORITY_LEVEL",
96 "SUSPEND",
97 "RESUME",
98 "RESET",
99 "SET_LOG_BUFFER",
100 "CHANGE_GANG_PRORITY",
101 "QUERY_SCHEDULER_STATUS",
102 "unused",
103 "SET_DEBUG_VMID",
104 "MISC",
105 "UPDATE_ROOT_PAGE_TABLE",
106 "AMD_LOG",
107 "SET_SE_MODE",
108 "SET_GANG_SUBMIT",
109 "SET_HW_RSRC_1",
110 };
111
112 static const char *mes_v12_0_misc_opcodes[] = {
113 "WRITE_REG",
114 "INV_GART",
115 "QUERY_STATUS",
116 "READ_REG",
117 "WAIT_REG_MEM",
118 "SET_SHADER_DEBUGGER",
119 "NOTIFY_WORK_ON_UNMAPPED_QUEUE",
120 "NOTIFY_TO_UNMAP_PROCESSES",
121 };
122
mes_v12_0_get_op_string(union MESAPI__MISC * x_pkt)123 static const char *mes_v12_0_get_op_string(union MESAPI__MISC *x_pkt)
124 {
125 const char *op_str = NULL;
126
127 if (x_pkt->header.opcode < ARRAY_SIZE(mes_v12_0_opcodes))
128 op_str = mes_v12_0_opcodes[x_pkt->header.opcode];
129
130 return op_str;
131 }
132
mes_v12_0_get_misc_op_string(union MESAPI__MISC * x_pkt)133 static const char *mes_v12_0_get_misc_op_string(union MESAPI__MISC *x_pkt)
134 {
135 const char *op_str = NULL;
136
137 if ((x_pkt->header.opcode == MES_SCH_API_MISC) &&
138 (x_pkt->opcode < ARRAY_SIZE(mes_v12_0_misc_opcodes)))
139 op_str = mes_v12_0_misc_opcodes[x_pkt->opcode];
140
141 return op_str;
142 }
143
mes_v12_0_submit_pkt_and_poll_completion(struct amdgpu_mes * mes,int pipe,void * pkt,int size,int api_status_off)144 static int mes_v12_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
145 int pipe, void *pkt, int size,
146 int api_status_off)
147 {
148 union MESAPI__QUERY_MES_STATUS mes_status_pkt;
149 signed long timeout = 2100000; /* 2100 ms */
150 struct amdgpu_device *adev = mes->adev;
151 struct amdgpu_ring *ring = &mes->ring[pipe];
152 spinlock_t *ring_lock = &mes->ring_lock[pipe];
153 struct MES_API_STATUS *api_status;
154 union MESAPI__MISC *x_pkt = pkt;
155 const char *op_str, *misc_op_str;
156 unsigned long flags;
157 u64 status_gpu_addr;
158 u32 seq, status_offset;
159 u64 *status_ptr;
160 signed long r;
161 int ret;
162
163 if (x_pkt->header.opcode >= MES_SCH_API_MAX)
164 return -EINVAL;
165
166 if (amdgpu_emu_mode) {
167 timeout *= 100;
168 } else if (amdgpu_sriov_vf(adev)) {
169 /* Worst case in sriov where all other 15 VF timeout, each VF needs about 600ms */
170 timeout = 15 * 600 * 1000;
171 }
172
173 ret = amdgpu_device_wb_get(adev, &status_offset);
174 if (ret)
175 return ret;
176
177 status_gpu_addr = adev->wb.gpu_addr + (status_offset * 4);
178 status_ptr = (u64 *)&adev->wb.wb[status_offset];
179 *status_ptr = 0;
180
181 spin_lock_irqsave(ring_lock, flags);
182 r = amdgpu_ring_alloc(ring, (size + sizeof(mes_status_pkt)) / 4);
183 if (r)
184 goto error_unlock_free;
185
186 seq = ++ring->fence_drv.sync_seq;
187 r = amdgpu_fence_wait_polling(ring,
188 seq - ring->fence_drv.num_fences_mask,
189 timeout);
190 if (r < 1)
191 goto error_undo;
192
193 api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
194 api_status->api_completion_fence_addr = status_gpu_addr;
195 api_status->api_completion_fence_value = 1;
196
197 amdgpu_ring_write_multiple(ring, pkt, size / 4);
198
199 memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
200 mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
201 mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
202 mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
203 mes_status_pkt.api_status.api_completion_fence_addr =
204 ring->fence_drv.gpu_addr;
205 mes_status_pkt.api_status.api_completion_fence_value = seq;
206
207 amdgpu_ring_write_multiple(ring, &mes_status_pkt,
208 sizeof(mes_status_pkt) / 4);
209
210 amdgpu_ring_commit(ring);
211 spin_unlock_irqrestore(ring_lock, flags);
212
213 op_str = mes_v12_0_get_op_string(x_pkt);
214 misc_op_str = mes_v12_0_get_misc_op_string(x_pkt);
215
216 if (misc_op_str)
217 dev_dbg(adev->dev, "MES(%d) msg=%s (%s) was emitted\n",
218 pipe, op_str, misc_op_str);
219 else if (op_str)
220 dev_dbg(adev->dev, "MES(%d) msg=%s was emitted\n",
221 pipe, op_str);
222 else
223 dev_dbg(adev->dev, "MES(%d) msg=%d was emitted\n",
224 pipe, x_pkt->header.opcode);
225
226 r = amdgpu_fence_wait_polling(ring, seq, timeout);
227 if (r < 1 || !*status_ptr) {
228
229 if (misc_op_str)
230 dev_err(adev->dev, "MES(%d) failed to respond to msg=%s (%s)\n",
231 pipe, op_str, misc_op_str);
232 else if (op_str)
233 dev_err(adev->dev, "MES(%d) failed to respond to msg=%s\n",
234 pipe, op_str);
235 else
236 dev_err(adev->dev, "MES(%d) failed to respond to msg=%d\n",
237 pipe, x_pkt->header.opcode);
238
239 while (halt_if_hws_hang)
240 schedule();
241
242 r = -ETIMEDOUT;
243 goto error_wb_free;
244 }
245
246 amdgpu_device_wb_free(adev, status_offset);
247 return 0;
248
249 error_undo:
250 dev_err(adev->dev, "MES ring buffer is full.\n");
251 amdgpu_ring_undo(ring);
252
253 error_unlock_free:
254 spin_unlock_irqrestore(ring_lock, flags);
255
256 error_wb_free:
257 amdgpu_device_wb_free(adev, status_offset);
258 return r;
259 }
260
convert_to_mes_queue_type(int queue_type)261 static int convert_to_mes_queue_type(int queue_type)
262 {
263 if (queue_type == AMDGPU_RING_TYPE_GFX)
264 return MES_QUEUE_TYPE_GFX;
265 else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
266 return MES_QUEUE_TYPE_COMPUTE;
267 else if (queue_type == AMDGPU_RING_TYPE_SDMA)
268 return MES_QUEUE_TYPE_SDMA;
269 else if (queue_type == AMDGPU_RING_TYPE_MES)
270 return MES_QUEUE_TYPE_SCHQ;
271 else
272 BUG();
273 return -1;
274 }
275
mes_v12_0_add_hw_queue(struct amdgpu_mes * mes,struct mes_add_queue_input * input)276 static int mes_v12_0_add_hw_queue(struct amdgpu_mes *mes,
277 struct mes_add_queue_input *input)
278 {
279 struct amdgpu_device *adev = mes->adev;
280 union MESAPI__ADD_QUEUE mes_add_queue_pkt;
281 struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB(0)];
282 uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
283
284 memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
285
286 mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
287 mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
288 mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
289
290 mes_add_queue_pkt.process_id = input->process_id;
291 mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
292 mes_add_queue_pkt.process_va_start = input->process_va_start;
293 mes_add_queue_pkt.process_va_end = input->process_va_end;
294 mes_add_queue_pkt.process_quantum = input->process_quantum;
295 mes_add_queue_pkt.process_context_addr = input->process_context_addr;
296 mes_add_queue_pkt.gang_quantum = input->gang_quantum;
297 mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
298 mes_add_queue_pkt.inprocess_gang_priority =
299 input->inprocess_gang_priority;
300 mes_add_queue_pkt.gang_global_priority_level =
301 input->gang_global_priority_level;
302 mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
303 mes_add_queue_pkt.mqd_addr = input->mqd_addr;
304
305 mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr;
306
307 mes_add_queue_pkt.queue_type =
308 convert_to_mes_queue_type(input->queue_type);
309 mes_add_queue_pkt.paging = input->paging;
310 mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
311 mes_add_queue_pkt.gws_base = input->gws_base;
312 mes_add_queue_pkt.gws_size = input->gws_size;
313 mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
314 mes_add_queue_pkt.tma_addr = input->tma_addr;
315 mes_add_queue_pkt.trap_en = input->trap_en;
316 mes_add_queue_pkt.skip_process_ctx_clear = input->skip_process_ctx_clear;
317 mes_add_queue_pkt.is_kfd_process = input->is_kfd_process;
318
319 /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
320 mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
321 mes_add_queue_pkt.gds_size = input->queue_size;
322
323 /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
324 mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
325 mes_add_queue_pkt.gds_size = input->queue_size;
326
327 return mes_v12_0_submit_pkt_and_poll_completion(mes,
328 AMDGPU_MES_SCHED_PIPE,
329 &mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
330 offsetof(union MESAPI__ADD_QUEUE, api_status));
331 }
332
mes_v12_0_remove_hw_queue(struct amdgpu_mes * mes,struct mes_remove_queue_input * input)333 static int mes_v12_0_remove_hw_queue(struct amdgpu_mes *mes,
334 struct mes_remove_queue_input *input)
335 {
336 union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
337
338 memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
339
340 mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
341 mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
342 mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
343
344 mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
345 mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
346
347 return mes_v12_0_submit_pkt_and_poll_completion(mes,
348 AMDGPU_MES_SCHED_PIPE,
349 &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
350 offsetof(union MESAPI__REMOVE_QUEUE, api_status));
351 }
352
mes_v12_0_reset_hw_queue(struct amdgpu_mes * mes,struct mes_reset_queue_input * input)353 static int mes_v12_0_reset_hw_queue(struct amdgpu_mes *mes,
354 struct mes_reset_queue_input *input)
355 {
356 union MESAPI__RESET mes_reset_queue_pkt;
357 int pipe;
358
359 memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
360
361 mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
362 mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
363 mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
364
365 mes_reset_queue_pkt.doorbell_offset = input->doorbell_offset;
366 mes_reset_queue_pkt.gang_context_addr = input->gang_context_addr;
367 /*mes_reset_queue_pkt.reset_queue_only = 1;*/
368
369 if (mes->adev->enable_uni_mes)
370 pipe = AMDGPU_MES_KIQ_PIPE;
371 else
372 pipe = AMDGPU_MES_SCHED_PIPE;
373
374 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
375 &mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
376 offsetof(union MESAPI__REMOVE_QUEUE, api_status));
377 }
378
mes_v12_0_map_legacy_queue(struct amdgpu_mes * mes,struct mes_map_legacy_queue_input * input)379 static int mes_v12_0_map_legacy_queue(struct amdgpu_mes *mes,
380 struct mes_map_legacy_queue_input *input)
381 {
382 union MESAPI__ADD_QUEUE mes_add_queue_pkt;
383 int pipe;
384
385 memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
386
387 mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
388 mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
389 mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
390
391 mes_add_queue_pkt.pipe_id = input->pipe_id;
392 mes_add_queue_pkt.queue_id = input->queue_id;
393 mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
394 mes_add_queue_pkt.mqd_addr = input->mqd_addr;
395 mes_add_queue_pkt.wptr_addr = input->wptr_addr;
396 mes_add_queue_pkt.queue_type =
397 convert_to_mes_queue_type(input->queue_type);
398 mes_add_queue_pkt.map_legacy_kq = 1;
399
400 if (mes->adev->enable_uni_mes)
401 pipe = AMDGPU_MES_KIQ_PIPE;
402 else
403 pipe = AMDGPU_MES_SCHED_PIPE;
404
405 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
406 &mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
407 offsetof(union MESAPI__ADD_QUEUE, api_status));
408 }
409
mes_v12_0_unmap_legacy_queue(struct amdgpu_mes * mes,struct mes_unmap_legacy_queue_input * input)410 static int mes_v12_0_unmap_legacy_queue(struct amdgpu_mes *mes,
411 struct mes_unmap_legacy_queue_input *input)
412 {
413 union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
414 int pipe;
415
416 memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
417
418 mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
419 mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
420 mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
421
422 mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
423 mes_remove_queue_pkt.gang_context_addr = 0;
424
425 mes_remove_queue_pkt.pipe_id = input->pipe_id;
426 mes_remove_queue_pkt.queue_id = input->queue_id;
427
428 if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
429 mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
430 mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
431 mes_remove_queue_pkt.tf_data =
432 lower_32_bits(input->trail_fence_data);
433 } else {
434 mes_remove_queue_pkt.unmap_legacy_queue = 1;
435 mes_remove_queue_pkt.queue_type =
436 convert_to_mes_queue_type(input->queue_type);
437 }
438
439 if (mes->adev->enable_uni_mes)
440 pipe = AMDGPU_MES_KIQ_PIPE;
441 else
442 pipe = AMDGPU_MES_SCHED_PIPE;
443
444 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
445 &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
446 offsetof(union MESAPI__REMOVE_QUEUE, api_status));
447 }
448
mes_v12_0_suspend_gang(struct amdgpu_mes * mes,struct mes_suspend_gang_input * input)449 static int mes_v12_0_suspend_gang(struct amdgpu_mes *mes,
450 struct mes_suspend_gang_input *input)
451 {
452 return 0;
453 }
454
mes_v12_0_resume_gang(struct amdgpu_mes * mes,struct mes_resume_gang_input * input)455 static int mes_v12_0_resume_gang(struct amdgpu_mes *mes,
456 struct mes_resume_gang_input *input)
457 {
458 return 0;
459 }
460
mes_v12_0_query_sched_status(struct amdgpu_mes * mes,int pipe)461 static int mes_v12_0_query_sched_status(struct amdgpu_mes *mes, int pipe)
462 {
463 union MESAPI__QUERY_MES_STATUS mes_status_pkt;
464
465 memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
466
467 mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
468 mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
469 mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
470
471 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
472 &mes_status_pkt, sizeof(mes_status_pkt),
473 offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
474 }
475
mes_v12_0_misc_op(struct amdgpu_mes * mes,struct mes_misc_op_input * input)476 static int mes_v12_0_misc_op(struct amdgpu_mes *mes,
477 struct mes_misc_op_input *input)
478 {
479 union MESAPI__MISC misc_pkt;
480 int pipe;
481
482 if (mes->adev->enable_uni_mes)
483 pipe = AMDGPU_MES_KIQ_PIPE;
484 else
485 pipe = AMDGPU_MES_SCHED_PIPE;
486
487 memset(&misc_pkt, 0, sizeof(misc_pkt));
488
489 misc_pkt.header.type = MES_API_TYPE_SCHEDULER;
490 misc_pkt.header.opcode = MES_SCH_API_MISC;
491 misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
492
493 switch (input->op) {
494 case MES_MISC_OP_READ_REG:
495 misc_pkt.opcode = MESAPI_MISC__READ_REG;
496 misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset;
497 misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr;
498 break;
499 case MES_MISC_OP_WRITE_REG:
500 misc_pkt.opcode = MESAPI_MISC__WRITE_REG;
501 misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset;
502 misc_pkt.write_reg.reg_value = input->write_reg.reg_value;
503 break;
504 case MES_MISC_OP_WRM_REG_WAIT:
505 misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
506 misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM;
507 misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
508 misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
509 misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
510 misc_pkt.wait_reg_mem.reg_offset2 = 0;
511 break;
512 case MES_MISC_OP_WRM_REG_WR_WAIT:
513 misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
514 misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG;
515 misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
516 misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
517 misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
518 misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1;
519 break;
520 case MES_MISC_OP_SET_SHADER_DEBUGGER:
521 pipe = AMDGPU_MES_SCHED_PIPE;
522 misc_pkt.opcode = MESAPI_MISC__SET_SHADER_DEBUGGER;
523 misc_pkt.set_shader_debugger.process_context_addr =
524 input->set_shader_debugger.process_context_addr;
525 misc_pkt.set_shader_debugger.flags.u32all =
526 input->set_shader_debugger.flags.u32all;
527 misc_pkt.set_shader_debugger.spi_gdbg_per_vmid_cntl =
528 input->set_shader_debugger.spi_gdbg_per_vmid_cntl;
529 memcpy(misc_pkt.set_shader_debugger.tcp_watch_cntl,
530 input->set_shader_debugger.tcp_watch_cntl,
531 sizeof(misc_pkt.set_shader_debugger.tcp_watch_cntl));
532 misc_pkt.set_shader_debugger.trap_en = input->set_shader_debugger.trap_en;
533 break;
534 case MES_MISC_OP_CHANGE_CONFIG:
535 misc_pkt.opcode = MESAPI_MISC__CHANGE_CONFIG;
536 misc_pkt.change_config.opcode =
537 MESAPI_MISC__CHANGE_CONFIG_OPTION_LIMIT_SINGLE_PROCESS;
538 misc_pkt.change_config.option.bits.limit_single_process =
539 input->change_config.option.limit_single_process;
540 break;
541
542 default:
543 DRM_ERROR("unsupported misc op (%d) \n", input->op);
544 return -EINVAL;
545 }
546
547 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
548 &misc_pkt, sizeof(misc_pkt),
549 offsetof(union MESAPI__MISC, api_status));
550 }
551
mes_v12_0_set_hw_resources_1(struct amdgpu_mes * mes,int pipe)552 static int mes_v12_0_set_hw_resources_1(struct amdgpu_mes *mes, int pipe)
553 {
554 union MESAPI_SET_HW_RESOURCES_1 mes_set_hw_res_1_pkt;
555
556 memset(&mes_set_hw_res_1_pkt, 0, sizeof(mes_set_hw_res_1_pkt));
557
558 mes_set_hw_res_1_pkt.header.type = MES_API_TYPE_SCHEDULER;
559 mes_set_hw_res_1_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC_1;
560 mes_set_hw_res_1_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
561 mes_set_hw_res_1_pkt.mes_kiq_unmap_timeout = 0xa;
562
563 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
564 &mes_set_hw_res_1_pkt, sizeof(mes_set_hw_res_1_pkt),
565 offsetof(union MESAPI_SET_HW_RESOURCES_1, api_status));
566 }
567
mes_v12_0_set_hw_resources(struct amdgpu_mes * mes,int pipe)568 static int mes_v12_0_set_hw_resources(struct amdgpu_mes *mes, int pipe)
569 {
570 int i;
571 struct amdgpu_device *adev = mes->adev;
572 union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
573
574 memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
575
576 mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
577 mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
578 mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
579
580 if (pipe == AMDGPU_MES_SCHED_PIPE) {
581 mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
582 mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
583 mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
584 mes_set_hw_res_pkt.paging_vmid = 0;
585
586 for (i = 0; i < MAX_COMPUTE_PIPES; i++)
587 mes_set_hw_res_pkt.compute_hqd_mask[i] =
588 mes->compute_hqd_mask[i];
589
590 for (i = 0; i < MAX_GFX_PIPES; i++)
591 mes_set_hw_res_pkt.gfx_hqd_mask[i] =
592 mes->gfx_hqd_mask[i];
593
594 for (i = 0; i < MAX_SDMA_PIPES; i++)
595 mes_set_hw_res_pkt.sdma_hqd_mask[i] =
596 mes->sdma_hqd_mask[i];
597
598 for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
599 mes_set_hw_res_pkt.aggregated_doorbells[i] =
600 mes->aggregated_doorbells[i];
601 }
602
603 mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr =
604 mes->sch_ctx_gpu_addr[pipe];
605 mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
606 mes->query_status_fence_gpu_addr[pipe];
607
608 for (i = 0; i < 5; i++) {
609 mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
610 mes_set_hw_res_pkt.mmhub_base[i] =
611 adev->reg_offset[MMHUB_HWIP][0][i];
612 mes_set_hw_res_pkt.osssys_base[i] =
613 adev->reg_offset[OSSSYS_HWIP][0][i];
614 }
615
616 mes_set_hw_res_pkt.disable_reset = 1;
617 mes_set_hw_res_pkt.disable_mes_log = 1;
618 mes_set_hw_res_pkt.use_different_vmid_compute = 1;
619 mes_set_hw_res_pkt.enable_reg_active_poll = 1;
620 mes_set_hw_res_pkt.enable_level_process_quantum_check = 1;
621
622 /*
623 * Keep oversubscribe timer for sdma . When we have unmapped doorbell
624 * handling support, other queue will not use the oversubscribe timer.
625 * handling mode - 0: disabled; 1: basic version; 2: basic+ version
626 */
627 mes_set_hw_res_pkt.oversubscription_timer = 50;
628 mes_set_hw_res_pkt.unmapped_doorbell_handling = 1;
629
630 if (amdgpu_mes_log_enable) {
631 mes_set_hw_res_pkt.enable_mes_event_int_logging = 1;
632 mes_set_hw_res_pkt.event_intr_history_gpu_mc_ptr = mes->event_log_gpu_addr + pipe * AMDGPU_MES_LOG_BUFFER_SIZE;
633 }
634
635 if (enforce_isolation)
636 mes_set_hw_res_pkt.limit_single_process = 1;
637
638 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
639 &mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
640 offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
641 }
642
mes_v12_0_init_aggregated_doorbell(struct amdgpu_mes * mes)643 static void mes_v12_0_init_aggregated_doorbell(struct amdgpu_mes *mes)
644 {
645 struct amdgpu_device *adev = mes->adev;
646 uint32_t data;
647
648 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1);
649 data &= ~(CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET_MASK |
650 CP_MES_DOORBELL_CONTROL1__DOORBELL_EN_MASK |
651 CP_MES_DOORBELL_CONTROL1__DOORBELL_HIT_MASK);
652 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_LOW] <<
653 CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET__SHIFT;
654 data |= 1 << CP_MES_DOORBELL_CONTROL1__DOORBELL_EN__SHIFT;
655 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1, data);
656
657 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2);
658 data &= ~(CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET_MASK |
659 CP_MES_DOORBELL_CONTROL2__DOORBELL_EN_MASK |
660 CP_MES_DOORBELL_CONTROL2__DOORBELL_HIT_MASK);
661 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] <<
662 CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET__SHIFT;
663 data |= 1 << CP_MES_DOORBELL_CONTROL2__DOORBELL_EN__SHIFT;
664 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2, data);
665
666 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3);
667 data &= ~(CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET_MASK |
668 CP_MES_DOORBELL_CONTROL3__DOORBELL_EN_MASK |
669 CP_MES_DOORBELL_CONTROL3__DOORBELL_HIT_MASK);
670 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_MEDIUM] <<
671 CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET__SHIFT;
672 data |= 1 << CP_MES_DOORBELL_CONTROL3__DOORBELL_EN__SHIFT;
673 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3, data);
674
675 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4);
676 data &= ~(CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET_MASK |
677 CP_MES_DOORBELL_CONTROL4__DOORBELL_EN_MASK |
678 CP_MES_DOORBELL_CONTROL4__DOORBELL_HIT_MASK);
679 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_HIGH] <<
680 CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET__SHIFT;
681 data |= 1 << CP_MES_DOORBELL_CONTROL4__DOORBELL_EN__SHIFT;
682 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4, data);
683
684 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5);
685 data &= ~(CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET_MASK |
686 CP_MES_DOORBELL_CONTROL5__DOORBELL_EN_MASK |
687 CP_MES_DOORBELL_CONTROL5__DOORBELL_HIT_MASK);
688 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_REALTIME] <<
689 CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET__SHIFT;
690 data |= 1 << CP_MES_DOORBELL_CONTROL5__DOORBELL_EN__SHIFT;
691 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5, data);
692
693 data = 1 << CP_HQD_GFX_CONTROL__DB_UPDATED_MSG_EN__SHIFT;
694 WREG32_SOC15(GC, 0, regCP_HQD_GFX_CONTROL, data);
695 }
696
697
mes_v12_0_enable_unmapped_doorbell_handling(struct amdgpu_mes * mes,bool enable)698 static void mes_v12_0_enable_unmapped_doorbell_handling(
699 struct amdgpu_mes *mes, bool enable)
700 {
701 struct amdgpu_device *adev = mes->adev;
702 uint32_t data = RREG32_SOC15(GC, 0, regCP_UNMAPPED_DOORBELL);
703
704 /*
705 * The default PROC_LSB settng is 0xc which means doorbell
706 * addr[16:12] gives the doorbell page number. For kfd, each
707 * process will use 2 pages of doorbell, we need to change the
708 * setting to 0xd
709 */
710 data &= ~CP_UNMAPPED_DOORBELL__PROC_LSB_MASK;
711 data |= 0xd << CP_UNMAPPED_DOORBELL__PROC_LSB__SHIFT;
712
713 data |= (enable ? 1 : 0) << CP_UNMAPPED_DOORBELL__ENABLE__SHIFT;
714
715 WREG32_SOC15(GC, 0, regCP_UNMAPPED_DOORBELL, data);
716 }
717
mes_v12_0_reset_legacy_queue(struct amdgpu_mes * mes,struct mes_reset_legacy_queue_input * input)718 static int mes_v12_0_reset_legacy_queue(struct amdgpu_mes *mes,
719 struct mes_reset_legacy_queue_input *input)
720 {
721 union MESAPI__RESET mes_reset_queue_pkt;
722 int pipe;
723
724 memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
725
726 mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
727 mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
728 mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
729
730 mes_reset_queue_pkt.queue_type =
731 convert_to_mes_queue_type(input->queue_type);
732
733 if (mes_reset_queue_pkt.queue_type == MES_QUEUE_TYPE_GFX) {
734 mes_reset_queue_pkt.reset_legacy_gfx = 1;
735 mes_reset_queue_pkt.pipe_id_lp = input->pipe_id;
736 mes_reset_queue_pkt.queue_id_lp = input->queue_id;
737 mes_reset_queue_pkt.mqd_mc_addr_lp = input->mqd_addr;
738 mes_reset_queue_pkt.doorbell_offset_lp = input->doorbell_offset;
739 mes_reset_queue_pkt.wptr_addr_lp = input->wptr_addr;
740 mes_reset_queue_pkt.vmid_id_lp = input->vmid;
741 } else {
742 mes_reset_queue_pkt.reset_queue_only = 1;
743 mes_reset_queue_pkt.doorbell_offset = input->doorbell_offset;
744 }
745
746 if (mes->adev->enable_uni_mes)
747 pipe = AMDGPU_MES_KIQ_PIPE;
748 else
749 pipe = AMDGPU_MES_SCHED_PIPE;
750
751 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
752 &mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
753 offsetof(union MESAPI__RESET, api_status));
754 }
755
756 static const struct amdgpu_mes_funcs mes_v12_0_funcs = {
757 .add_hw_queue = mes_v12_0_add_hw_queue,
758 .remove_hw_queue = mes_v12_0_remove_hw_queue,
759 .map_legacy_queue = mes_v12_0_map_legacy_queue,
760 .unmap_legacy_queue = mes_v12_0_unmap_legacy_queue,
761 .suspend_gang = mes_v12_0_suspend_gang,
762 .resume_gang = mes_v12_0_resume_gang,
763 .misc_op = mes_v12_0_misc_op,
764 .reset_legacy_queue = mes_v12_0_reset_legacy_queue,
765 .reset_hw_queue = mes_v12_0_reset_hw_queue,
766 };
767
mes_v12_0_allocate_ucode_buffer(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)768 static int mes_v12_0_allocate_ucode_buffer(struct amdgpu_device *adev,
769 enum admgpu_mes_pipe pipe)
770 {
771 int r;
772 const struct mes_firmware_header_v1_0 *mes_hdr;
773 const __le32 *fw_data;
774 unsigned fw_size;
775
776 mes_hdr = (const struct mes_firmware_header_v1_0 *)
777 adev->mes.fw[pipe]->data;
778
779 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
780 le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
781 fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
782
783 r = amdgpu_bo_create_reserved(adev, fw_size,
784 PAGE_SIZE,
785 AMDGPU_GEM_DOMAIN_VRAM,
786 &adev->mes.ucode_fw_obj[pipe],
787 &adev->mes.ucode_fw_gpu_addr[pipe],
788 (void **)&adev->mes.ucode_fw_ptr[pipe]);
789 if (r) {
790 dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
791 return r;
792 }
793
794 memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
795
796 amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
797 amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
798
799 return 0;
800 }
801
mes_v12_0_allocate_ucode_data_buffer(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)802 static int mes_v12_0_allocate_ucode_data_buffer(struct amdgpu_device *adev,
803 enum admgpu_mes_pipe pipe)
804 {
805 int r;
806 const struct mes_firmware_header_v1_0 *mes_hdr;
807 const __le32 *fw_data;
808 unsigned fw_size;
809
810 mes_hdr = (const struct mes_firmware_header_v1_0 *)
811 adev->mes.fw[pipe]->data;
812
813 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
814 le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
815 fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
816
817 r = amdgpu_bo_create_reserved(adev, fw_size,
818 64 * 1024,
819 AMDGPU_GEM_DOMAIN_VRAM,
820 &adev->mes.data_fw_obj[pipe],
821 &adev->mes.data_fw_gpu_addr[pipe],
822 (void **)&adev->mes.data_fw_ptr[pipe]);
823 if (r) {
824 dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
825 return r;
826 }
827
828 memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
829
830 amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
831 amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
832
833 return 0;
834 }
835
mes_v12_0_free_ucode_buffers(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)836 static void mes_v12_0_free_ucode_buffers(struct amdgpu_device *adev,
837 enum admgpu_mes_pipe pipe)
838 {
839 amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
840 &adev->mes.data_fw_gpu_addr[pipe],
841 (void **)&adev->mes.data_fw_ptr[pipe]);
842
843 amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
844 &adev->mes.ucode_fw_gpu_addr[pipe],
845 (void **)&adev->mes.ucode_fw_ptr[pipe]);
846 }
847
mes_v12_0_enable(struct amdgpu_device * adev,bool enable)848 static void mes_v12_0_enable(struct amdgpu_device *adev, bool enable)
849 {
850 uint64_t ucode_addr;
851 uint32_t pipe, data = 0;
852
853 if (enable) {
854 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
855 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
856 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 1);
857 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
858
859 mutex_lock(&adev->srbm_mutex);
860 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
861 soc21_grbm_select(adev, 3, pipe, 0, 0);
862
863 ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
864 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
865 lower_32_bits(ucode_addr));
866 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
867 upper_32_bits(ucode_addr));
868 }
869 soc21_grbm_select(adev, 0, 0, 0, 0);
870 mutex_unlock(&adev->srbm_mutex);
871
872 /* unhalt MES and activate pipe0 */
873 data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
874 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 1);
875 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
876
877 if (amdgpu_emu_mode)
878 msleep(100);
879 else if (adev->enable_uni_mes)
880 udelay(500);
881 else
882 udelay(50);
883 } else {
884 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
885 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
886 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
887 data = REG_SET_FIELD(data, CP_MES_CNTL,
888 MES_INVALIDATE_ICACHE, 1);
889 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
890 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 1);
891 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
892 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
893 }
894 }
895
mes_v12_0_set_ucode_start_addr(struct amdgpu_device * adev)896 static void mes_v12_0_set_ucode_start_addr(struct amdgpu_device *adev)
897 {
898 uint64_t ucode_addr;
899 int pipe;
900
901 mes_v12_0_enable(adev, false);
902
903 mutex_lock(&adev->srbm_mutex);
904 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
905 /* me=3, queue=0 */
906 soc21_grbm_select(adev, 3, pipe, 0, 0);
907
908 /* set ucode start address */
909 ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
910 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
911 lower_32_bits(ucode_addr));
912 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
913 upper_32_bits(ucode_addr));
914
915 soc21_grbm_select(adev, 0, 0, 0, 0);
916 }
917 mutex_unlock(&adev->srbm_mutex);
918 }
919
920 /* This function is for backdoor MES firmware */
mes_v12_0_load_microcode(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe,bool prime_icache)921 static int mes_v12_0_load_microcode(struct amdgpu_device *adev,
922 enum admgpu_mes_pipe pipe, bool prime_icache)
923 {
924 int r;
925 uint32_t data;
926
927 mes_v12_0_enable(adev, false);
928
929 if (!adev->mes.fw[pipe])
930 return -EINVAL;
931
932 r = mes_v12_0_allocate_ucode_buffer(adev, pipe);
933 if (r)
934 return r;
935
936 r = mes_v12_0_allocate_ucode_data_buffer(adev, pipe);
937 if (r) {
938 mes_v12_0_free_ucode_buffers(adev, pipe);
939 return r;
940 }
941
942 mutex_lock(&adev->srbm_mutex);
943 /* me=3, pipe=0, queue=0 */
944 soc21_grbm_select(adev, 3, pipe, 0, 0);
945
946 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0);
947
948 /* set ucode fimrware address */
949 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO,
950 lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
951 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI,
952 upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
953
954 /* set ucode instruction cache boundary to 2M-1 */
955 WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF);
956
957 /* set ucode data firmware address */
958 WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO,
959 lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
960 WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI,
961 upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
962
963 /* Set data cache boundary CP_MES_MDBOUND_LO */
964 WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x7FFFF);
965
966 if (prime_icache) {
967 /* invalidate ICACHE */
968 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
969 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
970 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
971 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
972
973 /* prime the ICACHE. */
974 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
975 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
976 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
977 }
978
979 soc21_grbm_select(adev, 0, 0, 0, 0);
980 mutex_unlock(&adev->srbm_mutex);
981
982 return 0;
983 }
984
mes_v12_0_allocate_eop_buf(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)985 static int mes_v12_0_allocate_eop_buf(struct amdgpu_device *adev,
986 enum admgpu_mes_pipe pipe)
987 {
988 int r;
989 u32 *eop;
990
991 r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
992 AMDGPU_GEM_DOMAIN_GTT,
993 &adev->mes.eop_gpu_obj[pipe],
994 &adev->mes.eop_gpu_addr[pipe],
995 (void **)&eop);
996 if (r) {
997 dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
998 return r;
999 }
1000
1001 memset(eop, 0,
1002 adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
1003
1004 amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
1005 amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
1006
1007 return 0;
1008 }
1009
mes_v12_0_mqd_init(struct amdgpu_ring * ring)1010 static int mes_v12_0_mqd_init(struct amdgpu_ring *ring)
1011 {
1012 struct v12_compute_mqd *mqd = ring->mqd_ptr;
1013 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
1014 uint32_t tmp;
1015
1016 mqd->header = 0xC0310800;
1017 mqd->compute_pipelinestat_enable = 0x00000001;
1018 mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
1019 mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
1020 mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
1021 mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
1022 mqd->compute_misc_reserved = 0x00000007;
1023
1024 eop_base_addr = ring->eop_gpu_addr >> 8;
1025
1026 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
1027 tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
1028 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
1029 (order_base_2(MES_EOP_SIZE / 4) - 1));
1030
1031 mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
1032 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
1033 mqd->cp_hqd_eop_control = tmp;
1034
1035 /* disable the queue if it's active */
1036 ring->wptr = 0;
1037 mqd->cp_hqd_pq_rptr = 0;
1038 mqd->cp_hqd_pq_wptr_lo = 0;
1039 mqd->cp_hqd_pq_wptr_hi = 0;
1040
1041 /* set the pointer to the MQD */
1042 mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
1043 mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
1044
1045 /* set MQD vmid to 0 */
1046 tmp = regCP_MQD_CONTROL_DEFAULT;
1047 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
1048 mqd->cp_mqd_control = tmp;
1049
1050 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
1051 hqd_gpu_addr = ring->gpu_addr >> 8;
1052 mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
1053 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
1054
1055 /* set the wb address whether it's enabled or not */
1056 wb_gpu_addr = ring->rptr_gpu_addr;
1057 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
1058 mqd->cp_hqd_pq_rptr_report_addr_hi =
1059 upper_32_bits(wb_gpu_addr) & 0xffff;
1060
1061 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
1062 wb_gpu_addr = ring->wptr_gpu_addr;
1063 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
1064 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
1065
1066 /* set up the HQD, this is similar to CP_RB0_CNTL */
1067 tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
1068 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
1069 (order_base_2(ring->ring_size / 4) - 1));
1070 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
1071 ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
1072 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
1073 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
1074 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
1075 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
1076 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
1077 mqd->cp_hqd_pq_control = tmp;
1078
1079 /* enable doorbell */
1080 tmp = 0;
1081 if (ring->use_doorbell) {
1082 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1083 DOORBELL_OFFSET, ring->doorbell_index);
1084 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1085 DOORBELL_EN, 1);
1086 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1087 DOORBELL_SOURCE, 0);
1088 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1089 DOORBELL_HIT, 0);
1090 } else {
1091 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1092 DOORBELL_EN, 0);
1093 }
1094 mqd->cp_hqd_pq_doorbell_control = tmp;
1095
1096 mqd->cp_hqd_vmid = 0;
1097 /* activate the queue */
1098 mqd->cp_hqd_active = 1;
1099
1100 tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
1101 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
1102 PRELOAD_SIZE, 0x55);
1103 mqd->cp_hqd_persistent_state = tmp;
1104
1105 mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT;
1106 mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
1107 mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
1108
1109 /*
1110 * Set CP_HQD_GFX_CONTROL.DB_UPDATED_MSG_EN[15] to enable unmapped
1111 * doorbell handling. This is a reserved CP internal register can
1112 * not be accesss by others
1113 */
1114 mqd->reserved_184 = BIT(15);
1115
1116 return 0;
1117 }
1118
mes_v12_0_queue_init_register(struct amdgpu_ring * ring)1119 static void mes_v12_0_queue_init_register(struct amdgpu_ring *ring)
1120 {
1121 struct v12_compute_mqd *mqd = ring->mqd_ptr;
1122 struct amdgpu_device *adev = ring->adev;
1123 uint32_t data = 0;
1124
1125 mutex_lock(&adev->srbm_mutex);
1126 soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
1127
1128 /* set CP_HQD_VMID.VMID = 0. */
1129 data = RREG32_SOC15(GC, 0, regCP_HQD_VMID);
1130 data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
1131 WREG32_SOC15(GC, 0, regCP_HQD_VMID, data);
1132
1133 /* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
1134 data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1135 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1136 DOORBELL_EN, 0);
1137 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1138
1139 /* set CP_MQD_BASE_ADDR/HI with the MQD base address */
1140 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
1141 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
1142
1143 /* set CP_MQD_CONTROL.VMID=0 */
1144 data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL);
1145 data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
1146 WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0);
1147
1148 /* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
1149 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
1150 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
1151
1152 /* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
1153 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
1154 mqd->cp_hqd_pq_rptr_report_addr_lo);
1155 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
1156 mqd->cp_hqd_pq_rptr_report_addr_hi);
1157
1158 /* set CP_HQD_PQ_CONTROL */
1159 WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
1160
1161 /* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
1162 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
1163 mqd->cp_hqd_pq_wptr_poll_addr_lo);
1164 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
1165 mqd->cp_hqd_pq_wptr_poll_addr_hi);
1166
1167 /* set CP_HQD_PQ_DOORBELL_CONTROL */
1168 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
1169 mqd->cp_hqd_pq_doorbell_control);
1170
1171 /* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
1172 WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
1173
1174 /* set CP_HQD_ACTIVE.ACTIVE=1 */
1175 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active);
1176
1177 soc21_grbm_select(adev, 0, 0, 0, 0);
1178 mutex_unlock(&adev->srbm_mutex);
1179 }
1180
mes_v12_0_kiq_enable_queue(struct amdgpu_device * adev)1181 static int mes_v12_0_kiq_enable_queue(struct amdgpu_device *adev)
1182 {
1183 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
1184 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
1185 int r;
1186
1187 if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
1188 return -EINVAL;
1189
1190 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
1191 if (r) {
1192 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
1193 return r;
1194 }
1195
1196 kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring[0]);
1197
1198 r = amdgpu_ring_test_ring(kiq_ring);
1199 if (r) {
1200 DRM_ERROR("kfq enable failed\n");
1201 kiq_ring->sched.ready = false;
1202 }
1203 return r;
1204 }
1205
mes_v12_0_queue_init(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)1206 static int mes_v12_0_queue_init(struct amdgpu_device *adev,
1207 enum admgpu_mes_pipe pipe)
1208 {
1209 struct amdgpu_ring *ring;
1210 int r;
1211
1212 if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1213 ring = &adev->gfx.kiq[0].ring;
1214 else
1215 ring = &adev->mes.ring[pipe];
1216
1217 if ((adev->enable_uni_mes || pipe == AMDGPU_MES_SCHED_PIPE) &&
1218 (amdgpu_in_reset(adev) || adev->in_suspend)) {
1219 *(ring->wptr_cpu_addr) = 0;
1220 *(ring->rptr_cpu_addr) = 0;
1221 amdgpu_ring_clear_ring(ring);
1222 }
1223
1224 r = mes_v12_0_mqd_init(ring);
1225 if (r)
1226 return r;
1227
1228 if (pipe == AMDGPU_MES_SCHED_PIPE) {
1229 if (adev->enable_uni_mes)
1230 r = amdgpu_mes_map_legacy_queue(adev, ring);
1231 else
1232 r = mes_v12_0_kiq_enable_queue(adev);
1233 if (r)
1234 return r;
1235 } else {
1236 mes_v12_0_queue_init_register(ring);
1237 }
1238
1239 /* get MES scheduler/KIQ versions */
1240 mutex_lock(&adev->srbm_mutex);
1241 soc21_grbm_select(adev, 3, pipe, 0, 0);
1242
1243 if (pipe == AMDGPU_MES_SCHED_PIPE)
1244 adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
1245 else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
1246 adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
1247
1248 soc21_grbm_select(adev, 0, 0, 0, 0);
1249 mutex_unlock(&adev->srbm_mutex);
1250
1251 return 0;
1252 }
1253
mes_v12_0_ring_init(struct amdgpu_device * adev,int pipe)1254 static int mes_v12_0_ring_init(struct amdgpu_device *adev, int pipe)
1255 {
1256 struct amdgpu_ring *ring;
1257
1258 ring = &adev->mes.ring[pipe];
1259
1260 ring->funcs = &mes_v12_0_ring_funcs;
1261
1262 ring->me = 3;
1263 ring->pipe = pipe;
1264 ring->queue = 0;
1265
1266 ring->ring_obj = NULL;
1267 ring->use_doorbell = true;
1268 ring->eop_gpu_addr = adev->mes.eop_gpu_addr[pipe];
1269 ring->no_scheduler = true;
1270 sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
1271
1272 if (pipe == AMDGPU_MES_SCHED_PIPE)
1273 ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
1274 else
1275 ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
1276
1277 return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1278 AMDGPU_RING_PRIO_DEFAULT, NULL);
1279 }
1280
mes_v12_0_kiq_ring_init(struct amdgpu_device * adev)1281 static int mes_v12_0_kiq_ring_init(struct amdgpu_device *adev)
1282 {
1283 struct amdgpu_ring *ring;
1284
1285 spin_lock_init(&adev->gfx.kiq[0].ring_lock);
1286
1287 ring = &adev->gfx.kiq[0].ring;
1288
1289 ring->me = 3;
1290 ring->pipe = 1;
1291 ring->queue = 0;
1292
1293 ring->adev = NULL;
1294 ring->ring_obj = NULL;
1295 ring->use_doorbell = true;
1296 ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
1297 ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
1298 ring->no_scheduler = true;
1299 sprintf(ring->name, "mes_kiq_%d.%d.%d",
1300 ring->me, ring->pipe, ring->queue);
1301
1302 return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1303 AMDGPU_RING_PRIO_DEFAULT, NULL);
1304 }
1305
mes_v12_0_mqd_sw_init(struct amdgpu_device * adev,enum admgpu_mes_pipe pipe)1306 static int mes_v12_0_mqd_sw_init(struct amdgpu_device *adev,
1307 enum admgpu_mes_pipe pipe)
1308 {
1309 int r, mqd_size = sizeof(struct v12_compute_mqd);
1310 struct amdgpu_ring *ring;
1311
1312 if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1313 ring = &adev->gfx.kiq[0].ring;
1314 else
1315 ring = &adev->mes.ring[pipe];
1316
1317 if (ring->mqd_obj)
1318 return 0;
1319
1320 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
1321 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
1322 &ring->mqd_gpu_addr, &ring->mqd_ptr);
1323 if (r) {
1324 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
1325 return r;
1326 }
1327
1328 memset(ring->mqd_ptr, 0, mqd_size);
1329
1330 /* prepare MQD backup */
1331 adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
1332 if (!adev->mes.mqd_backup[pipe])
1333 dev_warn(adev->dev,
1334 "no memory to create MQD backup for ring %s\n",
1335 ring->name);
1336
1337 return 0;
1338 }
1339
mes_v12_0_sw_init(struct amdgpu_ip_block * ip_block)1340 static int mes_v12_0_sw_init(struct amdgpu_ip_block *ip_block)
1341 {
1342 struct amdgpu_device *adev = ip_block->adev;
1343 int pipe, r;
1344
1345 adev->mes.funcs = &mes_v12_0_funcs;
1346 adev->mes.kiq_hw_init = &mes_v12_0_kiq_hw_init;
1347 adev->mes.kiq_hw_fini = &mes_v12_0_kiq_hw_fini;
1348 adev->mes.enable_legacy_queue_map = true;
1349
1350 adev->mes.event_log_size = adev->enable_uni_mes ? (AMDGPU_MAX_MES_PIPES * AMDGPU_MES_LOG_BUFFER_SIZE) : AMDGPU_MES_LOG_BUFFER_SIZE;
1351
1352 r = amdgpu_mes_init(adev);
1353 if (r)
1354 return r;
1355
1356 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1357 r = mes_v12_0_allocate_eop_buf(adev, pipe);
1358 if (r)
1359 return r;
1360
1361 r = mes_v12_0_mqd_sw_init(adev, pipe);
1362 if (r)
1363 return r;
1364
1365 if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1366 r = mes_v12_0_kiq_ring_init(adev);
1367 else
1368 r = mes_v12_0_ring_init(adev, pipe);
1369 if (r)
1370 return r;
1371 }
1372
1373 return 0;
1374 }
1375
mes_v12_0_sw_fini(struct amdgpu_ip_block * ip_block)1376 static int mes_v12_0_sw_fini(struct amdgpu_ip_block *ip_block)
1377 {
1378 struct amdgpu_device *adev = ip_block->adev;
1379 int pipe;
1380
1381 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1382 kfree(adev->mes.mqd_backup[pipe]);
1383
1384 amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1385 &adev->mes.eop_gpu_addr[pipe],
1386 NULL);
1387 amdgpu_ucode_release(&adev->mes.fw[pipe]);
1388
1389 if (adev->enable_uni_mes || pipe == AMDGPU_MES_SCHED_PIPE) {
1390 amdgpu_bo_free_kernel(&adev->mes.ring[pipe].mqd_obj,
1391 &adev->mes.ring[pipe].mqd_gpu_addr,
1392 &adev->mes.ring[pipe].mqd_ptr);
1393 amdgpu_ring_fini(&adev->mes.ring[pipe]);
1394 }
1395 }
1396
1397 if (!adev->enable_uni_mes) {
1398 amdgpu_bo_free_kernel(&adev->gfx.kiq[0].ring.mqd_obj,
1399 &adev->gfx.kiq[0].ring.mqd_gpu_addr,
1400 &adev->gfx.kiq[0].ring.mqd_ptr);
1401 amdgpu_ring_fini(&adev->gfx.kiq[0].ring);
1402 }
1403
1404 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1405 mes_v12_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1406 mes_v12_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1407 }
1408
1409 amdgpu_mes_fini(adev);
1410 return 0;
1411 }
1412
mes_v12_0_kiq_dequeue_sched(struct amdgpu_device * adev)1413 static void mes_v12_0_kiq_dequeue_sched(struct amdgpu_device *adev)
1414 {
1415 uint32_t data;
1416 int i;
1417
1418 mutex_lock(&adev->srbm_mutex);
1419 soc21_grbm_select(adev, 3, AMDGPU_MES_SCHED_PIPE, 0, 0);
1420
1421 /* disable the queue if it's active */
1422 if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
1423 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
1424 for (i = 0; i < adev->usec_timeout; i++) {
1425 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
1426 break;
1427 udelay(1);
1428 }
1429 }
1430 data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1431 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1432 DOORBELL_EN, 0);
1433 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1434 DOORBELL_HIT, 1);
1435 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1436
1437 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0);
1438
1439 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0);
1440 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0);
1441 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0);
1442
1443 soc21_grbm_select(adev, 0, 0, 0, 0);
1444 mutex_unlock(&adev->srbm_mutex);
1445
1446 adev->mes.ring[0].sched.ready = false;
1447 }
1448
mes_v12_0_kiq_setting(struct amdgpu_ring * ring)1449 static void mes_v12_0_kiq_setting(struct amdgpu_ring *ring)
1450 {
1451 uint32_t tmp;
1452 struct amdgpu_device *adev = ring->adev;
1453
1454 /* tell RLC which is KIQ queue */
1455 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1456 tmp &= 0xffffff00;
1457 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1458 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1459 tmp |= 0x80;
1460 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1461 }
1462
mes_v12_0_kiq_hw_init(struct amdgpu_device * adev)1463 static int mes_v12_0_kiq_hw_init(struct amdgpu_device *adev)
1464 {
1465 int r = 0;
1466 struct amdgpu_ip_block *ip_block;
1467
1468 if (adev->enable_uni_mes)
1469 mes_v12_0_kiq_setting(&adev->mes.ring[AMDGPU_MES_KIQ_PIPE]);
1470 else
1471 mes_v12_0_kiq_setting(&adev->gfx.kiq[0].ring);
1472
1473 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1474
1475 r = mes_v12_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE, false);
1476 if (r) {
1477 DRM_ERROR("failed to load MES fw, r=%d\n", r);
1478 return r;
1479 }
1480
1481 r = mes_v12_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE, true);
1482 if (r) {
1483 DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1484 return r;
1485 }
1486
1487 mes_v12_0_set_ucode_start_addr(adev);
1488
1489 } else if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO)
1490 mes_v12_0_set_ucode_start_addr(adev);
1491
1492 mes_v12_0_enable(adev, true);
1493
1494 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_MES);
1495 if (unlikely(!ip_block)) {
1496 dev_err(adev->dev, "Failed to get MES handle\n");
1497 return -EINVAL;
1498 }
1499
1500 r = mes_v12_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE);
1501 if (r)
1502 goto failure;
1503
1504 if (adev->enable_uni_mes) {
1505 r = mes_v12_0_set_hw_resources(&adev->mes, AMDGPU_MES_KIQ_PIPE);
1506 if (r)
1507 goto failure;
1508
1509 mes_v12_0_set_hw_resources_1(&adev->mes, AMDGPU_MES_KIQ_PIPE);
1510 }
1511
1512 if (adev->mes.enable_legacy_queue_map) {
1513 r = mes_v12_0_hw_init(ip_block);
1514 if (r)
1515 goto failure;
1516 }
1517
1518 return r;
1519
1520 failure:
1521 mes_v12_0_hw_fini(ip_block);
1522 return r;
1523 }
1524
mes_v12_0_kiq_hw_fini(struct amdgpu_device * adev)1525 static int mes_v12_0_kiq_hw_fini(struct amdgpu_device *adev)
1526 {
1527 if (adev->mes.ring[0].sched.ready) {
1528 if (adev->enable_uni_mes)
1529 amdgpu_mes_unmap_legacy_queue(adev,
1530 &adev->mes.ring[AMDGPU_MES_SCHED_PIPE],
1531 RESET_QUEUES, 0, 0);
1532 else
1533 mes_v12_0_kiq_dequeue_sched(adev);
1534
1535 adev->mes.ring[0].sched.ready = false;
1536 }
1537
1538 mes_v12_0_enable(adev, false);
1539
1540 return 0;
1541 }
1542
mes_v12_0_hw_init(struct amdgpu_ip_block * ip_block)1543 static int mes_v12_0_hw_init(struct amdgpu_ip_block *ip_block)
1544 {
1545 int r;
1546 struct amdgpu_device *adev = ip_block->adev;
1547
1548 if (adev->mes.ring[0].sched.ready)
1549 goto out;
1550
1551 if (!adev->enable_mes_kiq) {
1552 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1553 r = mes_v12_0_load_microcode(adev,
1554 AMDGPU_MES_SCHED_PIPE, true);
1555 if (r) {
1556 DRM_ERROR("failed to MES fw, r=%d\n", r);
1557 return r;
1558 }
1559
1560 mes_v12_0_set_ucode_start_addr(adev);
1561
1562 } else if (adev->firmware.load_type ==
1563 AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) {
1564
1565 mes_v12_0_set_ucode_start_addr(adev);
1566 }
1567
1568 mes_v12_0_enable(adev, true);
1569 }
1570
1571 /* Enable the MES to handle doorbell ring on unmapped queue */
1572 mes_v12_0_enable_unmapped_doorbell_handling(&adev->mes, true);
1573
1574 r = mes_v12_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE);
1575 if (r)
1576 goto failure;
1577
1578 r = mes_v12_0_set_hw_resources(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1579 if (r)
1580 goto failure;
1581
1582 if (adev->enable_uni_mes)
1583 mes_v12_0_set_hw_resources_1(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1584
1585 mes_v12_0_init_aggregated_doorbell(&adev->mes);
1586
1587 r = mes_v12_0_query_sched_status(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1588 if (r) {
1589 DRM_ERROR("MES is busy\n");
1590 goto failure;
1591 }
1592
1593 out:
1594 /*
1595 * Disable KIQ ring usage from the driver once MES is enabled.
1596 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1597 * with MES enabled.
1598 */
1599 adev->gfx.kiq[0].ring.sched.ready = false;
1600 adev->mes.ring[0].sched.ready = true;
1601
1602 return 0;
1603
1604 failure:
1605 mes_v12_0_hw_fini(ip_block);
1606 return r;
1607 }
1608
mes_v12_0_hw_fini(struct amdgpu_ip_block * ip_block)1609 static int mes_v12_0_hw_fini(struct amdgpu_ip_block *ip_block)
1610 {
1611 return 0;
1612 }
1613
mes_v12_0_suspend(struct amdgpu_ip_block * ip_block)1614 static int mes_v12_0_suspend(struct amdgpu_ip_block *ip_block)
1615 {
1616 int r;
1617
1618 r = amdgpu_mes_suspend(ip_block->adev);
1619 if (r)
1620 return r;
1621
1622 return mes_v12_0_hw_fini(ip_block);
1623 }
1624
mes_v12_0_resume(struct amdgpu_ip_block * ip_block)1625 static int mes_v12_0_resume(struct amdgpu_ip_block *ip_block)
1626 {
1627 int r;
1628
1629 r = mes_v12_0_hw_init(ip_block);
1630 if (r)
1631 return r;
1632
1633 return amdgpu_mes_resume(ip_block->adev);
1634 }
1635
mes_v12_0_early_init(struct amdgpu_ip_block * ip_block)1636 static int mes_v12_0_early_init(struct amdgpu_ip_block *ip_block)
1637 {
1638 struct amdgpu_device *adev = ip_block->adev;
1639 int pipe, r;
1640
1641 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1642 r = amdgpu_mes_init_microcode(adev, pipe);
1643 if (r)
1644 return r;
1645 }
1646
1647 return 0;
1648 }
1649
mes_v12_0_late_init(struct amdgpu_ip_block * ip_block)1650 static int mes_v12_0_late_init(struct amdgpu_ip_block *ip_block)
1651 {
1652 struct amdgpu_device *adev = ip_block->adev;
1653
1654 /* it's only intended for use in mes_self_test case, not for s0ix and reset */
1655 if (!amdgpu_in_reset(adev) && !adev->in_s0ix && !adev->in_suspend)
1656 amdgpu_mes_self_test(adev);
1657
1658 return 0;
1659 }
1660
1661 static const struct amd_ip_funcs mes_v12_0_ip_funcs = {
1662 .name = "mes_v12_0",
1663 .early_init = mes_v12_0_early_init,
1664 .late_init = mes_v12_0_late_init,
1665 .sw_init = mes_v12_0_sw_init,
1666 .sw_fini = mes_v12_0_sw_fini,
1667 .hw_init = mes_v12_0_hw_init,
1668 .hw_fini = mes_v12_0_hw_fini,
1669 .suspend = mes_v12_0_suspend,
1670 .resume = mes_v12_0_resume,
1671 };
1672
1673 const struct amdgpu_ip_block_version mes_v12_0_ip_block = {
1674 .type = AMD_IP_BLOCK_TYPE_MES,
1675 .major = 12,
1676 .minor = 0,
1677 .rev = 0,
1678 .funcs = &mes_v12_0_ip_funcs,
1679 };
1680