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 "gfx_v12_0.h"
28 #include "soc15_common.h"
29 #include "soc21.h"
30 #include "gc/gc_12_0_0_offset.h"
31 #include "gc/gc_12_0_0_sh_mask.h"
32 #include "gc/gc_11_0_0_default.h"
33 #include "v12_structs.h"
34 #include "mes_v12_api_def.h"
35
36 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mes.bin");
37 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mes1.bin");
38 MODULE_FIRMWARE("amdgpu/gc_12_0_0_uni_mes.bin");
39 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mes.bin");
40 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mes1.bin");
41 MODULE_FIRMWARE("amdgpu/gc_12_0_1_uni_mes.bin");
42
43 static int mes_v12_0_hw_init(struct amdgpu_ip_block *ip_block);
44 static int mes_v12_0_hw_fini(struct amdgpu_ip_block *ip_block);
45 static int mes_v12_0_kiq_hw_init(struct amdgpu_device *adev);
46 static int mes_v12_0_kiq_hw_fini(struct amdgpu_device *adev);
47
48 #define MES_EOP_SIZE 2048
49
mes_v12_0_ring_set_wptr(struct amdgpu_ring * ring)50 static void mes_v12_0_ring_set_wptr(struct amdgpu_ring *ring)
51 {
52 struct amdgpu_device *adev = ring->adev;
53
54 if (ring->use_doorbell) {
55 atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
56 ring->wptr);
57 WDOORBELL64(ring->doorbell_index, ring->wptr);
58 } else {
59 BUG();
60 }
61 }
62
mes_v12_0_ring_get_rptr(struct amdgpu_ring * ring)63 static u64 mes_v12_0_ring_get_rptr(struct amdgpu_ring *ring)
64 {
65 return *ring->rptr_cpu_addr;
66 }
67
mes_v12_0_ring_get_wptr(struct amdgpu_ring * ring)68 static u64 mes_v12_0_ring_get_wptr(struct amdgpu_ring *ring)
69 {
70 u64 wptr;
71
72 if (ring->use_doorbell)
73 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
74 else
75 BUG();
76 return wptr;
77 }
78
79 static const struct amdgpu_ring_funcs mes_v12_0_ring_funcs = {
80 .type = AMDGPU_RING_TYPE_MES,
81 .align_mask = 1,
82 .nop = 0,
83 .support_64bit_ptrs = true,
84 .get_rptr = mes_v12_0_ring_get_rptr,
85 .get_wptr = mes_v12_0_ring_get_wptr,
86 .set_wptr = mes_v12_0_ring_set_wptr,
87 .insert_nop = amdgpu_ring_insert_nop,
88 };
89
90 static const char *mes_v12_0_opcodes[] = {
91 "SET_HW_RSRC",
92 "SET_SCHEDULING_CONFIG",
93 "ADD_QUEUE",
94 "REMOVE_QUEUE",
95 "PERFORM_YIELD",
96 "SET_GANG_PRIORITY_LEVEL",
97 "SUSPEND",
98 "RESUME",
99 "RESET",
100 "SET_LOG_BUFFER",
101 "CHANGE_GANG_PRORITY",
102 "QUERY_SCHEDULER_STATUS",
103 "unused",
104 "SET_DEBUG_VMID",
105 "MISC",
106 "UPDATE_ROOT_PAGE_TABLE",
107 "AMD_LOG",
108 "SET_SE_MODE",
109 "SET_GANG_SUBMIT",
110 "SET_HW_RSRC_1",
111 };
112
113 static const char *mes_v12_0_misc_opcodes[] = {
114 "WRITE_REG",
115 "INV_GART",
116 "QUERY_STATUS",
117 "READ_REG",
118 "WAIT_REG_MEM",
119 "SET_SHADER_DEBUGGER",
120 "NOTIFY_WORK_ON_UNMAPPED_QUEUE",
121 "NOTIFY_TO_UNMAP_PROCESSES",
122 };
123
mes_v12_0_get_op_string(union MESAPI__MISC * x_pkt)124 static const char *mes_v12_0_get_op_string(union MESAPI__MISC *x_pkt)
125 {
126 const char *op_str = NULL;
127
128 if (x_pkt->header.opcode < ARRAY_SIZE(mes_v12_0_opcodes))
129 op_str = mes_v12_0_opcodes[x_pkt->header.opcode];
130
131 return op_str;
132 }
133
mes_v12_0_get_misc_op_string(union MESAPI__MISC * x_pkt)134 static const char *mes_v12_0_get_misc_op_string(union MESAPI__MISC *x_pkt)
135 {
136 const char *op_str = NULL;
137
138 if ((x_pkt->header.opcode == MES_SCH_API_MISC) &&
139 (x_pkt->opcode < ARRAY_SIZE(mes_v12_0_misc_opcodes)))
140 op_str = mes_v12_0_misc_opcodes[x_pkt->opcode];
141
142 return op_str;
143 }
144
mes_v12_0_submit_pkt_and_poll_completion(struct amdgpu_mes * mes,int pipe,void * pkt,int size,int api_status_off)145 static int mes_v12_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
146 int pipe, void *pkt, int size,
147 int api_status_off)
148 {
149 union MESAPI__QUERY_MES_STATUS mes_status_pkt;
150 signed long timeout = 2100000; /* 2100 ms */
151 struct amdgpu_device *adev = mes->adev;
152 struct amdgpu_ring *ring = &mes->ring[pipe];
153 spinlock_t *ring_lock = &mes->ring_lock[pipe];
154 struct MES_API_STATUS *api_status;
155 union MESAPI__MISC *x_pkt = pkt;
156 const char *op_str, *misc_op_str;
157 unsigned long flags;
158 u64 status_gpu_addr;
159 u32 seq, status_offset;
160 u64 *status_ptr;
161 signed long r;
162 int ret;
163
164 if (x_pkt->header.opcode >= MES_SCH_API_MAX)
165 return -EINVAL;
166
167 if (amdgpu_emu_mode) {
168 timeout *= 100;
169 } else if (amdgpu_sriov_vf(adev)) {
170 /* Worst case in sriov where all other 15 VF timeout, each VF needs about 600ms */
171 timeout = 15 * 600 * 1000;
172 }
173
174 ret = amdgpu_device_wb_get(adev, &status_offset);
175 if (ret)
176 return ret;
177
178 status_gpu_addr = adev->wb.gpu_addr + (status_offset * 4);
179 status_ptr = (u64 *)&adev->wb.wb[status_offset];
180 *status_ptr = 0;
181
182 spin_lock_irqsave(ring_lock, flags);
183 r = amdgpu_ring_alloc(ring, (size + sizeof(mes_status_pkt)) / 4);
184 if (r)
185 goto error_unlock_free;
186
187 seq = ++ring->fence_drv.sync_seq;
188 r = amdgpu_fence_wait_polling(ring,
189 seq - ring->fence_drv.num_fences_mask,
190 timeout);
191 if (r < 1)
192 goto error_undo;
193
194 api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
195 api_status->api_completion_fence_addr = status_gpu_addr;
196 api_status->api_completion_fence_value = 1;
197
198 amdgpu_ring_write_multiple(ring, pkt, size / 4);
199
200 memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
201 mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
202 mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
203 mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
204 mes_status_pkt.api_status.api_completion_fence_addr =
205 ring->fence_drv.gpu_addr;
206 mes_status_pkt.api_status.api_completion_fence_value = seq;
207
208 amdgpu_ring_write_multiple(ring, &mes_status_pkt,
209 sizeof(mes_status_pkt) / 4);
210
211 amdgpu_ring_commit(ring);
212 spin_unlock_irqrestore(ring_lock, flags);
213
214 op_str = mes_v12_0_get_op_string(x_pkt);
215 misc_op_str = mes_v12_0_get_misc_op_string(x_pkt);
216
217 if (misc_op_str)
218 dev_dbg(adev->dev, "MES(%d) msg=%s (%s) was emitted\n",
219 pipe, op_str, misc_op_str);
220 else if (op_str)
221 dev_dbg(adev->dev, "MES(%d) msg=%s was emitted\n",
222 pipe, op_str);
223 else
224 dev_dbg(adev->dev, "MES(%d) msg=%d was emitted\n",
225 pipe, x_pkt->header.opcode);
226
227 r = amdgpu_fence_wait_polling(ring, seq, timeout);
228 if (r < 1 || !*status_ptr) {
229
230 if (misc_op_str)
231 dev_err(adev->dev, "MES(%d) failed to respond to msg=%s (%s)\n",
232 pipe, op_str, misc_op_str);
233 else if (op_str)
234 dev_err(adev->dev, "MES(%d) failed to respond to msg=%s\n",
235 pipe, op_str);
236 else
237 dev_err(adev->dev, "MES(%d) failed to respond to msg=%d\n",
238 pipe, x_pkt->header.opcode);
239
240 while (halt_if_hws_hang)
241 schedule();
242
243 r = -ETIMEDOUT;
244 goto error_wb_free;
245 }
246
247 amdgpu_device_wb_free(adev, status_offset);
248 return 0;
249
250 error_undo:
251 dev_err(adev->dev, "MES ring buffer is full.\n");
252 amdgpu_ring_undo(ring);
253
254 error_unlock_free:
255 spin_unlock_irqrestore(ring_lock, flags);
256
257 error_wb_free:
258 amdgpu_device_wb_free(adev, status_offset);
259 return r;
260 }
261
convert_to_mes_queue_type(int queue_type)262 static int convert_to_mes_queue_type(int queue_type)
263 {
264 if (queue_type == AMDGPU_RING_TYPE_GFX)
265 return MES_QUEUE_TYPE_GFX;
266 else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
267 return MES_QUEUE_TYPE_COMPUTE;
268 else if (queue_type == AMDGPU_RING_TYPE_SDMA)
269 return MES_QUEUE_TYPE_SDMA;
270 else if (queue_type == AMDGPU_RING_TYPE_MES)
271 return MES_QUEUE_TYPE_SCHQ;
272 else
273 BUG();
274 return -1;
275 }
276
convert_to_mes_priority_level(int priority_level)277 static int convert_to_mes_priority_level(int priority_level)
278 {
279 switch (priority_level) {
280 case AMDGPU_MES_PRIORITY_LEVEL_LOW:
281 return AMD_PRIORITY_LEVEL_LOW;
282 case AMDGPU_MES_PRIORITY_LEVEL_NORMAL:
283 default:
284 return AMD_PRIORITY_LEVEL_NORMAL;
285 case AMDGPU_MES_PRIORITY_LEVEL_MEDIUM:
286 return AMD_PRIORITY_LEVEL_MEDIUM;
287 case AMDGPU_MES_PRIORITY_LEVEL_HIGH:
288 return AMD_PRIORITY_LEVEL_HIGH;
289 case AMDGPU_MES_PRIORITY_LEVEL_REALTIME:
290 return AMD_PRIORITY_LEVEL_REALTIME;
291 }
292 }
293
mes_v12_0_add_hw_queue(struct amdgpu_mes * mes,struct mes_add_queue_input * input)294 static int mes_v12_0_add_hw_queue(struct amdgpu_mes *mes,
295 struct mes_add_queue_input *input)
296 {
297 struct amdgpu_device *adev = mes->adev;
298 union MESAPI__ADD_QUEUE mes_add_queue_pkt;
299 struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB(0)];
300 uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
301
302 memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
303
304 mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
305 mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
306 mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
307
308 mes_add_queue_pkt.process_id = input->process_id;
309 mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
310 mes_add_queue_pkt.process_va_start = input->process_va_start;
311 mes_add_queue_pkt.process_va_end = input->process_va_end;
312 mes_add_queue_pkt.process_quantum = input->process_quantum;
313 mes_add_queue_pkt.process_context_addr = input->process_context_addr;
314 mes_add_queue_pkt.gang_quantum = input->gang_quantum;
315 mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
316 mes_add_queue_pkt.inprocess_gang_priority =
317 convert_to_mes_priority_level(input->inprocess_gang_priority);
318 mes_add_queue_pkt.gang_global_priority_level =
319 convert_to_mes_priority_level(input->gang_global_priority_level);
320 mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
321 mes_add_queue_pkt.mqd_addr = input->mqd_addr;
322
323 mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr;
324
325 mes_add_queue_pkt.queue_type =
326 convert_to_mes_queue_type(input->queue_type);
327 mes_add_queue_pkt.paging = input->paging;
328 mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
329 mes_add_queue_pkt.gws_base = input->gws_base;
330 mes_add_queue_pkt.gws_size = input->gws_size;
331 mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
332 mes_add_queue_pkt.tma_addr = input->tma_addr;
333 mes_add_queue_pkt.trap_en = input->trap_en;
334 mes_add_queue_pkt.skip_process_ctx_clear = input->skip_process_ctx_clear;
335 mes_add_queue_pkt.is_kfd_process = input->is_kfd_process;
336
337 /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
338 mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
339 mes_add_queue_pkt.gds_size = input->queue_size;
340
341 /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
342 mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
343 mes_add_queue_pkt.gds_size = input->queue_size;
344
345 return mes_v12_0_submit_pkt_and_poll_completion(mes,
346 AMDGPU_MES_SCHED_PIPE,
347 &mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
348 offsetof(union MESAPI__ADD_QUEUE, api_status));
349 }
350
mes_v12_0_remove_hw_queue(struct amdgpu_mes * mes,struct mes_remove_queue_input * input)351 static int mes_v12_0_remove_hw_queue(struct amdgpu_mes *mes,
352 struct mes_remove_queue_input *input)
353 {
354 union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
355
356 memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
357
358 mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
359 mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
360 mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
361
362 mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
363 mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
364
365 return mes_v12_0_submit_pkt_and_poll_completion(mes,
366 AMDGPU_MES_SCHED_PIPE,
367 &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
368 offsetof(union MESAPI__REMOVE_QUEUE, api_status));
369 }
370
gfx_v12_0_request_gfx_index_mutex(struct amdgpu_device * adev,bool req)371 int gfx_v12_0_request_gfx_index_mutex(struct amdgpu_device *adev,
372 bool req)
373 {
374 u32 i, tmp, val;
375
376 for (i = 0; i < adev->usec_timeout; i++) {
377 /* Request with MeId=2, PipeId=0 */
378 tmp = REG_SET_FIELD(0, CP_GFX_INDEX_MUTEX, REQUEST, req);
379 tmp = REG_SET_FIELD(tmp, CP_GFX_INDEX_MUTEX, CLIENTID, 4);
380 WREG32_SOC15(GC, 0, regCP_GFX_INDEX_MUTEX, tmp);
381
382 val = RREG32_SOC15(GC, 0, regCP_GFX_INDEX_MUTEX);
383 if (req) {
384 if (val == tmp)
385 break;
386 } else {
387 tmp = REG_SET_FIELD(tmp, CP_GFX_INDEX_MUTEX,
388 REQUEST, 1);
389
390 /* unlocked or locked by firmware */
391 if (val != tmp)
392 break;
393 }
394 udelay(1);
395 }
396
397 if (i >= adev->usec_timeout)
398 return -EINVAL;
399
400 return 0;
401 }
402
mes_v12_0_reset_queue_mmio(struct amdgpu_mes * mes,uint32_t queue_type,uint32_t me_id,uint32_t pipe_id,uint32_t queue_id,uint32_t vmid)403 static int mes_v12_0_reset_queue_mmio(struct amdgpu_mes *mes, uint32_t queue_type,
404 uint32_t me_id, uint32_t pipe_id,
405 uint32_t queue_id, uint32_t vmid)
406 {
407 struct amdgpu_device *adev = mes->adev;
408 uint32_t value, reg;
409 int i, r = 0;
410
411 amdgpu_gfx_rlc_enter_safe_mode(adev, 0);
412
413 if (queue_type == AMDGPU_RING_TYPE_GFX) {
414 dev_info(adev->dev, "reset gfx queue (%d:%d:%d: vmid:%d)\n",
415 me_id, pipe_id, queue_id, vmid);
416
417 mutex_lock(&adev->gfx.reset_sem_mutex);
418 gfx_v12_0_request_gfx_index_mutex(adev, true);
419 /* all se allow writes */
420 WREG32_SOC15(GC, 0, regGRBM_GFX_INDEX,
421 (uint32_t)(0x1 << GRBM_GFX_INDEX__SE_BROADCAST_WRITES__SHIFT));
422 value = REG_SET_FIELD(0, CP_VMID_RESET, RESET_REQUEST, 1 << vmid);
423 if (pipe_id == 0)
424 value = REG_SET_FIELD(value, CP_VMID_RESET, PIPE0_QUEUES, 1 << queue_id);
425 else
426 value = REG_SET_FIELD(value, CP_VMID_RESET, PIPE1_QUEUES, 1 << queue_id);
427 WREG32_SOC15(GC, 0, regCP_VMID_RESET, value);
428 gfx_v12_0_request_gfx_index_mutex(adev, false);
429 mutex_unlock(&adev->gfx.reset_sem_mutex);
430
431 mutex_lock(&adev->srbm_mutex);
432 soc21_grbm_select(adev, me_id, pipe_id, queue_id, 0);
433 /* wait till dequeue take effects */
434 for (i = 0; i < adev->usec_timeout; i++) {
435 if (!(RREG32_SOC15(GC, 0, regCP_GFX_HQD_ACTIVE) & 1))
436 break;
437 udelay(1);
438 }
439 if (i >= adev->usec_timeout) {
440 dev_err(adev->dev, "failed to wait on gfx hqd deactivate\n");
441 r = -ETIMEDOUT;
442 }
443
444 soc21_grbm_select(adev, 0, 0, 0, 0);
445 mutex_unlock(&adev->srbm_mutex);
446 } else if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
447 dev_info(adev->dev, "reset compute queue (%d:%d:%d)\n",
448 me_id, pipe_id, queue_id);
449 mutex_lock(&adev->srbm_mutex);
450 soc21_grbm_select(adev, me_id, pipe_id, queue_id, 0);
451 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 0x2);
452 WREG32_SOC15(GC, 0, regSPI_COMPUTE_QUEUE_RESET, 0x1);
453
454 /* wait till dequeue take effects */
455 for (i = 0; i < adev->usec_timeout; i++) {
456 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
457 break;
458 udelay(1);
459 }
460 if (i >= adev->usec_timeout) {
461 dev_err(adev->dev, "failed to wait on hqd deactivate\n");
462 r = -ETIMEDOUT;
463 }
464 soc21_grbm_select(adev, 0, 0, 0, 0);
465 mutex_unlock(&adev->srbm_mutex);
466 } else if (queue_type == AMDGPU_RING_TYPE_SDMA) {
467 dev_info(adev->dev, "reset sdma queue (%d:%d:%d)\n",
468 me_id, pipe_id, queue_id);
469 switch (me_id) {
470 case 1:
471 reg = SOC15_REG_OFFSET(GC, 0, regSDMA1_QUEUE_RESET_REQ);
472 break;
473 case 0:
474 default:
475 reg = SOC15_REG_OFFSET(GC, 0, regSDMA0_QUEUE_RESET_REQ);
476 break;
477 }
478
479 value = 1 << queue_id;
480 WREG32(reg, value);
481 /* wait for queue reset done */
482 for (i = 0; i < adev->usec_timeout; i++) {
483 if (!(RREG32(reg) & value))
484 break;
485 udelay(1);
486 }
487 if (i >= adev->usec_timeout) {
488 dev_err(adev->dev, "failed to wait on sdma queue reset done\n");
489 r = -ETIMEDOUT;
490 }
491 }
492
493 amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
494 return r;
495 }
496
mes_v12_0_map_legacy_queue(struct amdgpu_mes * mes,struct mes_map_legacy_queue_input * input)497 static int mes_v12_0_map_legacy_queue(struct amdgpu_mes *mes,
498 struct mes_map_legacy_queue_input *input)
499 {
500 union MESAPI__ADD_QUEUE mes_add_queue_pkt;
501 int pipe;
502
503 memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
504
505 mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
506 mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
507 mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
508
509 mes_add_queue_pkt.pipe_id = input->pipe_id;
510 mes_add_queue_pkt.queue_id = input->queue_id;
511 mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
512 mes_add_queue_pkt.mqd_addr = input->mqd_addr;
513 mes_add_queue_pkt.wptr_addr = input->wptr_addr;
514 mes_add_queue_pkt.queue_type =
515 convert_to_mes_queue_type(input->queue_type);
516 mes_add_queue_pkt.map_legacy_kq = 1;
517
518 if (mes->adev->enable_uni_mes)
519 pipe = AMDGPU_MES_KIQ_PIPE;
520 else
521 pipe = AMDGPU_MES_SCHED_PIPE;
522
523 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
524 &mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
525 offsetof(union MESAPI__ADD_QUEUE, api_status));
526 }
527
mes_v12_0_unmap_legacy_queue(struct amdgpu_mes * mes,struct mes_unmap_legacy_queue_input * input)528 static int mes_v12_0_unmap_legacy_queue(struct amdgpu_mes *mes,
529 struct mes_unmap_legacy_queue_input *input)
530 {
531 union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
532 int pipe;
533
534 memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
535
536 mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
537 mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
538 mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
539
540 mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
541 mes_remove_queue_pkt.gang_context_addr = 0;
542
543 mes_remove_queue_pkt.pipe_id = input->pipe_id;
544 mes_remove_queue_pkt.queue_id = input->queue_id;
545
546 if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
547 mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
548 mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
549 mes_remove_queue_pkt.tf_data =
550 lower_32_bits(input->trail_fence_data);
551 } else {
552 mes_remove_queue_pkt.unmap_legacy_queue = 1;
553 mes_remove_queue_pkt.queue_type =
554 convert_to_mes_queue_type(input->queue_type);
555 }
556
557 if (mes->adev->enable_uni_mes)
558 pipe = AMDGPU_MES_KIQ_PIPE;
559 else
560 pipe = AMDGPU_MES_SCHED_PIPE;
561
562 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
563 &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
564 offsetof(union MESAPI__REMOVE_QUEUE, api_status));
565 }
566
mes_v12_0_suspend_gang(struct amdgpu_mes * mes,struct mes_suspend_gang_input * input)567 static int mes_v12_0_suspend_gang(struct amdgpu_mes *mes,
568 struct mes_suspend_gang_input *input)
569 {
570 return 0;
571 }
572
mes_v12_0_resume_gang(struct amdgpu_mes * mes,struct mes_resume_gang_input * input)573 static int mes_v12_0_resume_gang(struct amdgpu_mes *mes,
574 struct mes_resume_gang_input *input)
575 {
576 return 0;
577 }
578
mes_v12_0_query_sched_status(struct amdgpu_mes * mes,int pipe)579 static int mes_v12_0_query_sched_status(struct amdgpu_mes *mes, int pipe)
580 {
581 union MESAPI__QUERY_MES_STATUS mes_status_pkt;
582
583 memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
584
585 mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
586 mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
587 mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
588
589 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
590 &mes_status_pkt, sizeof(mes_status_pkt),
591 offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
592 }
593
mes_v12_0_misc_op(struct amdgpu_mes * mes,struct mes_misc_op_input * input)594 static int mes_v12_0_misc_op(struct amdgpu_mes *mes,
595 struct mes_misc_op_input *input)
596 {
597 union MESAPI__MISC misc_pkt;
598 int pipe;
599
600 if (mes->adev->enable_uni_mes)
601 pipe = AMDGPU_MES_KIQ_PIPE;
602 else
603 pipe = AMDGPU_MES_SCHED_PIPE;
604
605 memset(&misc_pkt, 0, sizeof(misc_pkt));
606
607 misc_pkt.header.type = MES_API_TYPE_SCHEDULER;
608 misc_pkt.header.opcode = MES_SCH_API_MISC;
609 misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
610
611 switch (input->op) {
612 case MES_MISC_OP_READ_REG:
613 misc_pkt.opcode = MESAPI_MISC__READ_REG;
614 misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset;
615 misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr;
616 break;
617 case MES_MISC_OP_WRITE_REG:
618 misc_pkt.opcode = MESAPI_MISC__WRITE_REG;
619 misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset;
620 misc_pkt.write_reg.reg_value = input->write_reg.reg_value;
621 break;
622 case MES_MISC_OP_WRM_REG_WAIT:
623 misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
624 misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM;
625 misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
626 misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
627 misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
628 misc_pkt.wait_reg_mem.reg_offset2 = 0;
629 break;
630 case MES_MISC_OP_WRM_REG_WR_WAIT:
631 misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
632 misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG;
633 misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
634 misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
635 misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
636 misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1;
637 break;
638 case MES_MISC_OP_SET_SHADER_DEBUGGER:
639 pipe = AMDGPU_MES_SCHED_PIPE;
640 misc_pkt.opcode = MESAPI_MISC__SET_SHADER_DEBUGGER;
641 misc_pkt.set_shader_debugger.process_context_addr =
642 input->set_shader_debugger.process_context_addr;
643 misc_pkt.set_shader_debugger.flags.u32all =
644 input->set_shader_debugger.flags.u32all;
645 misc_pkt.set_shader_debugger.spi_gdbg_per_vmid_cntl =
646 input->set_shader_debugger.spi_gdbg_per_vmid_cntl;
647 memcpy(misc_pkt.set_shader_debugger.tcp_watch_cntl,
648 input->set_shader_debugger.tcp_watch_cntl,
649 sizeof(misc_pkt.set_shader_debugger.tcp_watch_cntl));
650 misc_pkt.set_shader_debugger.trap_en = input->set_shader_debugger.trap_en;
651 break;
652 case MES_MISC_OP_CHANGE_CONFIG:
653 misc_pkt.opcode = MESAPI_MISC__CHANGE_CONFIG;
654 misc_pkt.change_config.opcode =
655 MESAPI_MISC__CHANGE_CONFIG_OPTION_LIMIT_SINGLE_PROCESS;
656 misc_pkt.change_config.option.bits.limit_single_process =
657 input->change_config.option.limit_single_process;
658 break;
659
660 default:
661 DRM_ERROR("unsupported misc op (%d) \n", input->op);
662 return -EINVAL;
663 }
664
665 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
666 &misc_pkt, sizeof(misc_pkt),
667 offsetof(union MESAPI__MISC, api_status));
668 }
669
mes_v12_0_set_hw_resources_1(struct amdgpu_mes * mes,int pipe)670 static int mes_v12_0_set_hw_resources_1(struct amdgpu_mes *mes, int pipe)
671 {
672 union MESAPI_SET_HW_RESOURCES_1 mes_set_hw_res_1_pkt;
673
674 memset(&mes_set_hw_res_1_pkt, 0, sizeof(mes_set_hw_res_1_pkt));
675
676 mes_set_hw_res_1_pkt.header.type = MES_API_TYPE_SCHEDULER;
677 mes_set_hw_res_1_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC_1;
678 mes_set_hw_res_1_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
679 mes_set_hw_res_1_pkt.mes_kiq_unmap_timeout = 0xa;
680 mes_set_hw_res_1_pkt.cleaner_shader_fence_mc_addr =
681 mes->resource_1_gpu_addr[pipe];
682
683 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
684 &mes_set_hw_res_1_pkt, sizeof(mes_set_hw_res_1_pkt),
685 offsetof(union MESAPI_SET_HW_RESOURCES_1, api_status));
686 }
687
mes_v12_0_set_hw_resources(struct amdgpu_mes * mes,int pipe)688 static int mes_v12_0_set_hw_resources(struct amdgpu_mes *mes, int pipe)
689 {
690 int i;
691 struct amdgpu_device *adev = mes->adev;
692 union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
693
694 memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
695
696 mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
697 mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
698 mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
699
700 if (pipe == AMDGPU_MES_SCHED_PIPE) {
701 mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
702 mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
703 mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
704 mes_set_hw_res_pkt.paging_vmid = 0;
705
706 for (i = 0; i < MAX_COMPUTE_PIPES; i++)
707 mes_set_hw_res_pkt.compute_hqd_mask[i] =
708 mes->compute_hqd_mask[i];
709
710 for (i = 0; i < MAX_GFX_PIPES; i++)
711 mes_set_hw_res_pkt.gfx_hqd_mask[i] =
712 mes->gfx_hqd_mask[i];
713
714 for (i = 0; i < MAX_SDMA_PIPES; i++)
715 mes_set_hw_res_pkt.sdma_hqd_mask[i] =
716 mes->sdma_hqd_mask[i];
717
718 for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
719 mes_set_hw_res_pkt.aggregated_doorbells[i] =
720 mes->aggregated_doorbells[i];
721 }
722
723 mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr =
724 mes->sch_ctx_gpu_addr[pipe];
725 mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
726 mes->query_status_fence_gpu_addr[pipe];
727
728 for (i = 0; i < 5; i++) {
729 mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
730 mes_set_hw_res_pkt.mmhub_base[i] =
731 adev->reg_offset[MMHUB_HWIP][0][i];
732 mes_set_hw_res_pkt.osssys_base[i] =
733 adev->reg_offset[OSSSYS_HWIP][0][i];
734 }
735
736 mes_set_hw_res_pkt.disable_reset = 1;
737 mes_set_hw_res_pkt.disable_mes_log = 1;
738 mes_set_hw_res_pkt.use_different_vmid_compute = 1;
739 mes_set_hw_res_pkt.enable_reg_active_poll = 1;
740 mes_set_hw_res_pkt.enable_level_process_quantum_check = 1;
741
742 /*
743 * Keep oversubscribe timer for sdma . When we have unmapped doorbell
744 * handling support, other queue will not use the oversubscribe timer.
745 * handling mode - 0: disabled; 1: basic version; 2: basic+ version
746 */
747 mes_set_hw_res_pkt.oversubscription_timer = 50;
748 mes_set_hw_res_pkt.unmapped_doorbell_handling = 1;
749
750 if (amdgpu_mes_log_enable) {
751 mes_set_hw_res_pkt.enable_mes_event_int_logging = 1;
752 mes_set_hw_res_pkt.event_intr_history_gpu_mc_ptr = mes->event_log_gpu_addr +
753 pipe * (AMDGPU_MES_LOG_BUFFER_SIZE + AMDGPU_MES_MSCRATCH_SIZE);
754 }
755
756 if (adev->enforce_isolation[0] == AMDGPU_ENFORCE_ISOLATION_ENABLE)
757 mes_set_hw_res_pkt.limit_single_process = 1;
758
759 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
760 &mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
761 offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
762 }
763
mes_v12_0_init_aggregated_doorbell(struct amdgpu_mes * mes)764 static void mes_v12_0_init_aggregated_doorbell(struct amdgpu_mes *mes)
765 {
766 struct amdgpu_device *adev = mes->adev;
767 uint32_t data;
768
769 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1);
770 data &= ~(CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET_MASK |
771 CP_MES_DOORBELL_CONTROL1__DOORBELL_EN_MASK |
772 CP_MES_DOORBELL_CONTROL1__DOORBELL_HIT_MASK);
773 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_LOW] <<
774 CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET__SHIFT;
775 data |= 1 << CP_MES_DOORBELL_CONTROL1__DOORBELL_EN__SHIFT;
776 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1, data);
777
778 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2);
779 data &= ~(CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET_MASK |
780 CP_MES_DOORBELL_CONTROL2__DOORBELL_EN_MASK |
781 CP_MES_DOORBELL_CONTROL2__DOORBELL_HIT_MASK);
782 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] <<
783 CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET__SHIFT;
784 data |= 1 << CP_MES_DOORBELL_CONTROL2__DOORBELL_EN__SHIFT;
785 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2, data);
786
787 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3);
788 data &= ~(CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET_MASK |
789 CP_MES_DOORBELL_CONTROL3__DOORBELL_EN_MASK |
790 CP_MES_DOORBELL_CONTROL3__DOORBELL_HIT_MASK);
791 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_MEDIUM] <<
792 CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET__SHIFT;
793 data |= 1 << CP_MES_DOORBELL_CONTROL3__DOORBELL_EN__SHIFT;
794 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3, data);
795
796 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4);
797 data &= ~(CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET_MASK |
798 CP_MES_DOORBELL_CONTROL4__DOORBELL_EN_MASK |
799 CP_MES_DOORBELL_CONTROL4__DOORBELL_HIT_MASK);
800 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_HIGH] <<
801 CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET__SHIFT;
802 data |= 1 << CP_MES_DOORBELL_CONTROL4__DOORBELL_EN__SHIFT;
803 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4, data);
804
805 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5);
806 data &= ~(CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET_MASK |
807 CP_MES_DOORBELL_CONTROL5__DOORBELL_EN_MASK |
808 CP_MES_DOORBELL_CONTROL5__DOORBELL_HIT_MASK);
809 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_REALTIME] <<
810 CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET__SHIFT;
811 data |= 1 << CP_MES_DOORBELL_CONTROL5__DOORBELL_EN__SHIFT;
812 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5, data);
813
814 data = 1 << CP_HQD_GFX_CONTROL__DB_UPDATED_MSG_EN__SHIFT;
815 WREG32_SOC15(GC, 0, regCP_HQD_GFX_CONTROL, data);
816 }
817
818
mes_v12_0_enable_unmapped_doorbell_handling(struct amdgpu_mes * mes,bool enable)819 static void mes_v12_0_enable_unmapped_doorbell_handling(
820 struct amdgpu_mes *mes, bool enable)
821 {
822 struct amdgpu_device *adev = mes->adev;
823 uint32_t data = RREG32_SOC15(GC, 0, regCP_UNMAPPED_DOORBELL);
824
825 /*
826 * The default PROC_LSB settng is 0xc which means doorbell
827 * addr[16:12] gives the doorbell page number. For kfd, each
828 * process will use 2 pages of doorbell, we need to change the
829 * setting to 0xd
830 */
831 data &= ~CP_UNMAPPED_DOORBELL__PROC_LSB_MASK;
832 data |= 0xd << CP_UNMAPPED_DOORBELL__PROC_LSB__SHIFT;
833
834 data |= (enable ? 1 : 0) << CP_UNMAPPED_DOORBELL__ENABLE__SHIFT;
835
836 WREG32_SOC15(GC, 0, regCP_UNMAPPED_DOORBELL, data);
837 }
838
mes_v12_0_reset_hw_queue(struct amdgpu_mes * mes,struct mes_reset_queue_input * input)839 static int mes_v12_0_reset_hw_queue(struct amdgpu_mes *mes,
840 struct mes_reset_queue_input *input)
841 {
842 union MESAPI__RESET mes_reset_queue_pkt;
843 int pipe;
844
845 if (input->use_mmio)
846 return mes_v12_0_reset_queue_mmio(mes, input->queue_type,
847 input->me_id, input->pipe_id,
848 input->queue_id, input->vmid);
849
850 memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
851
852 mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
853 mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
854 mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
855
856 mes_reset_queue_pkt.queue_type =
857 convert_to_mes_queue_type(input->queue_type);
858
859 if (input->legacy_gfx) {
860 mes_reset_queue_pkt.reset_legacy_gfx = 1;
861 mes_reset_queue_pkt.pipe_id_lp = input->pipe_id;
862 mes_reset_queue_pkt.queue_id_lp = input->queue_id;
863 mes_reset_queue_pkt.mqd_mc_addr_lp = input->mqd_addr;
864 mes_reset_queue_pkt.doorbell_offset_lp = input->doorbell_offset;
865 mes_reset_queue_pkt.wptr_addr_lp = input->wptr_addr;
866 mes_reset_queue_pkt.vmid_id_lp = input->vmid;
867 } else {
868 mes_reset_queue_pkt.reset_queue_only = 1;
869 mes_reset_queue_pkt.doorbell_offset = input->doorbell_offset;
870 }
871
872 if (input->is_kq)
873 pipe = AMDGPU_MES_KIQ_PIPE;
874 else
875 pipe = AMDGPU_MES_SCHED_PIPE;
876
877 return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
878 &mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
879 offsetof(union MESAPI__RESET, api_status));
880 }
881
882 static const struct amdgpu_mes_funcs mes_v12_0_funcs = {
883 .add_hw_queue = mes_v12_0_add_hw_queue,
884 .remove_hw_queue = mes_v12_0_remove_hw_queue,
885 .map_legacy_queue = mes_v12_0_map_legacy_queue,
886 .unmap_legacy_queue = mes_v12_0_unmap_legacy_queue,
887 .suspend_gang = mes_v12_0_suspend_gang,
888 .resume_gang = mes_v12_0_resume_gang,
889 .misc_op = mes_v12_0_misc_op,
890 .reset_hw_queue = mes_v12_0_reset_hw_queue,
891 };
892
mes_v12_0_allocate_ucode_buffer(struct amdgpu_device * adev,enum amdgpu_mes_pipe pipe)893 static int mes_v12_0_allocate_ucode_buffer(struct amdgpu_device *adev,
894 enum amdgpu_mes_pipe pipe)
895 {
896 int r;
897 const struct mes_firmware_header_v1_0 *mes_hdr;
898 const __le32 *fw_data;
899 unsigned fw_size;
900
901 mes_hdr = (const struct mes_firmware_header_v1_0 *)
902 adev->mes.fw[pipe]->data;
903
904 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
905 le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
906 fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
907
908 r = amdgpu_bo_create_reserved(adev, fw_size,
909 PAGE_SIZE,
910 AMDGPU_GEM_DOMAIN_VRAM,
911 &adev->mes.ucode_fw_obj[pipe],
912 &adev->mes.ucode_fw_gpu_addr[pipe],
913 (void **)&adev->mes.ucode_fw_ptr[pipe]);
914 if (r) {
915 dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
916 return r;
917 }
918
919 memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
920
921 amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
922 amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
923
924 return 0;
925 }
926
mes_v12_0_allocate_ucode_data_buffer(struct amdgpu_device * adev,enum amdgpu_mes_pipe pipe)927 static int mes_v12_0_allocate_ucode_data_buffer(struct amdgpu_device *adev,
928 enum amdgpu_mes_pipe pipe)
929 {
930 int r;
931 const struct mes_firmware_header_v1_0 *mes_hdr;
932 const __le32 *fw_data;
933 unsigned fw_size;
934
935 mes_hdr = (const struct mes_firmware_header_v1_0 *)
936 adev->mes.fw[pipe]->data;
937
938 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
939 le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
940 fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
941
942 r = amdgpu_bo_create_reserved(adev, fw_size,
943 64 * 1024,
944 AMDGPU_GEM_DOMAIN_VRAM,
945 &adev->mes.data_fw_obj[pipe],
946 &adev->mes.data_fw_gpu_addr[pipe],
947 (void **)&adev->mes.data_fw_ptr[pipe]);
948 if (r) {
949 dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
950 return r;
951 }
952
953 memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
954
955 amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
956 amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
957
958 return 0;
959 }
960
mes_v12_0_free_ucode_buffers(struct amdgpu_device * adev,enum amdgpu_mes_pipe pipe)961 static void mes_v12_0_free_ucode_buffers(struct amdgpu_device *adev,
962 enum amdgpu_mes_pipe pipe)
963 {
964 amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
965 &adev->mes.data_fw_gpu_addr[pipe],
966 (void **)&adev->mes.data_fw_ptr[pipe]);
967
968 amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
969 &adev->mes.ucode_fw_gpu_addr[pipe],
970 (void **)&adev->mes.ucode_fw_ptr[pipe]);
971 }
972
mes_v12_0_enable(struct amdgpu_device * adev,bool enable)973 static void mes_v12_0_enable(struct amdgpu_device *adev, bool enable)
974 {
975 uint64_t ucode_addr;
976 uint32_t pipe, data = 0;
977
978 if (enable) {
979 mutex_lock(&adev->srbm_mutex);
980 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
981 soc21_grbm_select(adev, 3, pipe, 0, 0);
982 if (amdgpu_mes_log_enable) {
983 u32 log_size = AMDGPU_MES_LOG_BUFFER_SIZE + AMDGPU_MES_MSCRATCH_SIZE;
984 /* In case uni mes is not enabled, only program for pipe 0 */
985 if (adev->mes.event_log_size >= (pipe + 1) * log_size) {
986 WREG32_SOC15(GC, 0, regCP_MES_MSCRATCH_LO,
987 lower_32_bits(adev->mes.event_log_gpu_addr +
988 pipe * log_size + AMDGPU_MES_LOG_BUFFER_SIZE));
989 WREG32_SOC15(GC, 0, regCP_MES_MSCRATCH_HI,
990 upper_32_bits(adev->mes.event_log_gpu_addr +
991 pipe * log_size + AMDGPU_MES_LOG_BUFFER_SIZE));
992 dev_info(adev->dev, "Setup CP MES MSCRATCH address : 0x%x. 0x%x\n",
993 RREG32_SOC15(GC, 0, regCP_MES_MSCRATCH_HI),
994 RREG32_SOC15(GC, 0, regCP_MES_MSCRATCH_LO));
995 }
996 }
997
998 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
999 if (pipe == 0)
1000 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
1001 else
1002 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 1);
1003 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
1004
1005 ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
1006 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
1007 lower_32_bits(ucode_addr));
1008 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
1009 upper_32_bits(ucode_addr));
1010
1011 /* unhalt MES and activate one pipe each loop */
1012 data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
1013 if (pipe)
1014 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 1);
1015 dev_info(adev->dev, "program CP_MES_CNTL : 0x%x\n", data);
1016
1017 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
1018
1019 }
1020 soc21_grbm_select(adev, 0, 0, 0, 0);
1021 mutex_unlock(&adev->srbm_mutex);
1022
1023 if (amdgpu_emu_mode)
1024 msleep(100);
1025 else if (adev->enable_uni_mes)
1026 udelay(500);
1027 else
1028 udelay(50);
1029 } else {
1030 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
1031 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
1032 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
1033 data = REG_SET_FIELD(data, CP_MES_CNTL,
1034 MES_INVALIDATE_ICACHE, 1);
1035 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
1036 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 1);
1037 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
1038 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
1039 }
1040 }
1041
mes_v12_0_set_ucode_start_addr(struct amdgpu_device * adev)1042 static void mes_v12_0_set_ucode_start_addr(struct amdgpu_device *adev)
1043 {
1044 uint64_t ucode_addr;
1045 int pipe;
1046
1047 mes_v12_0_enable(adev, false);
1048
1049 mutex_lock(&adev->srbm_mutex);
1050 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1051 /* me=3, queue=0 */
1052 soc21_grbm_select(adev, 3, pipe, 0, 0);
1053
1054 /* set ucode start address */
1055 ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
1056 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
1057 lower_32_bits(ucode_addr));
1058 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
1059 upper_32_bits(ucode_addr));
1060
1061 soc21_grbm_select(adev, 0, 0, 0, 0);
1062 }
1063 mutex_unlock(&adev->srbm_mutex);
1064 }
1065
1066 /* This function is for backdoor MES firmware */
mes_v12_0_load_microcode(struct amdgpu_device * adev,enum amdgpu_mes_pipe pipe,bool prime_icache)1067 static int mes_v12_0_load_microcode(struct amdgpu_device *adev,
1068 enum amdgpu_mes_pipe pipe, bool prime_icache)
1069 {
1070 int r;
1071 uint32_t data;
1072
1073 mes_v12_0_enable(adev, false);
1074
1075 if (!adev->mes.fw[pipe])
1076 return -EINVAL;
1077
1078 r = mes_v12_0_allocate_ucode_buffer(adev, pipe);
1079 if (r)
1080 return r;
1081
1082 r = mes_v12_0_allocate_ucode_data_buffer(adev, pipe);
1083 if (r) {
1084 mes_v12_0_free_ucode_buffers(adev, pipe);
1085 return r;
1086 }
1087
1088 mutex_lock(&adev->srbm_mutex);
1089 /* me=3, pipe=0, queue=0 */
1090 soc21_grbm_select(adev, 3, pipe, 0, 0);
1091
1092 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0);
1093
1094 /* set ucode fimrware address */
1095 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO,
1096 lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
1097 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI,
1098 upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
1099
1100 /* set ucode instruction cache boundary to 2M-1 */
1101 WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF);
1102
1103 /* set ucode data firmware address */
1104 WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO,
1105 lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
1106 WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI,
1107 upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
1108
1109 /* Set data cache boundary CP_MES_MDBOUND_LO */
1110 WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x7FFFF);
1111
1112 if (prime_icache) {
1113 /* invalidate ICACHE */
1114 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
1115 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
1116 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
1117 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
1118
1119 /* prime the ICACHE. */
1120 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
1121 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
1122 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
1123 }
1124
1125 soc21_grbm_select(adev, 0, 0, 0, 0);
1126 mutex_unlock(&adev->srbm_mutex);
1127
1128 return 0;
1129 }
1130
mes_v12_0_allocate_eop_buf(struct amdgpu_device * adev,enum amdgpu_mes_pipe pipe)1131 static int mes_v12_0_allocate_eop_buf(struct amdgpu_device *adev,
1132 enum amdgpu_mes_pipe pipe)
1133 {
1134 int r;
1135 u32 *eop;
1136
1137 r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
1138 AMDGPU_GEM_DOMAIN_GTT,
1139 &adev->mes.eop_gpu_obj[pipe],
1140 &adev->mes.eop_gpu_addr[pipe],
1141 (void **)&eop);
1142 if (r) {
1143 dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
1144 return r;
1145 }
1146
1147 memset(eop, 0,
1148 adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
1149
1150 amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
1151 amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
1152
1153 return 0;
1154 }
1155
mes_v12_0_mqd_init(struct amdgpu_ring * ring)1156 static int mes_v12_0_mqd_init(struct amdgpu_ring *ring)
1157 {
1158 struct v12_compute_mqd *mqd = ring->mqd_ptr;
1159 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
1160 uint32_t tmp;
1161
1162 mqd->header = 0xC0310800;
1163 mqd->compute_pipelinestat_enable = 0x00000001;
1164 mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
1165 mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
1166 mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
1167 mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
1168 mqd->compute_misc_reserved = 0x00000007;
1169
1170 eop_base_addr = ring->eop_gpu_addr >> 8;
1171
1172 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
1173 tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
1174 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
1175 (order_base_2(MES_EOP_SIZE / 4) - 1));
1176
1177 mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
1178 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
1179 mqd->cp_hqd_eop_control = tmp;
1180
1181 /* disable the queue if it's active */
1182 ring->wptr = 0;
1183 mqd->cp_hqd_pq_rptr = 0;
1184 mqd->cp_hqd_pq_wptr_lo = 0;
1185 mqd->cp_hqd_pq_wptr_hi = 0;
1186
1187 /* set the pointer to the MQD */
1188 mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
1189 mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
1190
1191 /* set MQD vmid to 0 */
1192 tmp = regCP_MQD_CONTROL_DEFAULT;
1193 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
1194 mqd->cp_mqd_control = tmp;
1195
1196 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
1197 hqd_gpu_addr = ring->gpu_addr >> 8;
1198 mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
1199 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
1200
1201 /* set the wb address whether it's enabled or not */
1202 wb_gpu_addr = ring->rptr_gpu_addr;
1203 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
1204 mqd->cp_hqd_pq_rptr_report_addr_hi =
1205 upper_32_bits(wb_gpu_addr) & 0xffff;
1206
1207 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
1208 wb_gpu_addr = ring->wptr_gpu_addr;
1209 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
1210 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
1211
1212 /* set up the HQD, this is similar to CP_RB0_CNTL */
1213 tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
1214 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
1215 (order_base_2(ring->ring_size / 4) - 1));
1216 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
1217 ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
1218 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
1219 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
1220 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
1221 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
1222 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
1223 mqd->cp_hqd_pq_control = tmp;
1224
1225 /* enable doorbell */
1226 tmp = 0;
1227 if (ring->use_doorbell) {
1228 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1229 DOORBELL_OFFSET, ring->doorbell_index);
1230 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1231 DOORBELL_EN, 1);
1232 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1233 DOORBELL_SOURCE, 0);
1234 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1235 DOORBELL_HIT, 0);
1236 } else {
1237 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1238 DOORBELL_EN, 0);
1239 }
1240 mqd->cp_hqd_pq_doorbell_control = tmp;
1241
1242 mqd->cp_hqd_vmid = 0;
1243 /* activate the queue */
1244 mqd->cp_hqd_active = 1;
1245
1246 tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
1247 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
1248 PRELOAD_SIZE, 0x55);
1249 mqd->cp_hqd_persistent_state = tmp;
1250
1251 mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT;
1252 mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
1253 mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
1254
1255 /*
1256 * Set CP_HQD_GFX_CONTROL.DB_UPDATED_MSG_EN[15] to enable unmapped
1257 * doorbell handling. This is a reserved CP internal register can
1258 * not be accesss by others
1259 */
1260 mqd->reserved_184 = BIT(15);
1261
1262 return 0;
1263 }
1264
mes_v12_0_queue_init_register(struct amdgpu_ring * ring)1265 static void mes_v12_0_queue_init_register(struct amdgpu_ring *ring)
1266 {
1267 struct v12_compute_mqd *mqd = ring->mqd_ptr;
1268 struct amdgpu_device *adev = ring->adev;
1269 uint32_t data = 0;
1270
1271 mutex_lock(&adev->srbm_mutex);
1272 soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
1273
1274 /* set CP_HQD_VMID.VMID = 0. */
1275 data = RREG32_SOC15(GC, 0, regCP_HQD_VMID);
1276 data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
1277 WREG32_SOC15(GC, 0, regCP_HQD_VMID, data);
1278
1279 /* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
1280 data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1281 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1282 DOORBELL_EN, 0);
1283 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1284
1285 /* set CP_MQD_BASE_ADDR/HI with the MQD base address */
1286 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
1287 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
1288
1289 /* set CP_MQD_CONTROL.VMID=0 */
1290 data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL);
1291 data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
1292 WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0);
1293
1294 /* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
1295 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
1296 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
1297
1298 /* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
1299 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
1300 mqd->cp_hqd_pq_rptr_report_addr_lo);
1301 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
1302 mqd->cp_hqd_pq_rptr_report_addr_hi);
1303
1304 /* set CP_HQD_PQ_CONTROL */
1305 WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
1306
1307 /* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
1308 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
1309 mqd->cp_hqd_pq_wptr_poll_addr_lo);
1310 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
1311 mqd->cp_hqd_pq_wptr_poll_addr_hi);
1312
1313 /* set CP_HQD_PQ_DOORBELL_CONTROL */
1314 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
1315 mqd->cp_hqd_pq_doorbell_control);
1316
1317 /* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
1318 WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
1319
1320 /* set CP_HQD_ACTIVE.ACTIVE=1 */
1321 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active);
1322
1323 soc21_grbm_select(adev, 0, 0, 0, 0);
1324 mutex_unlock(&adev->srbm_mutex);
1325 }
1326
mes_v12_0_kiq_enable_queue(struct amdgpu_device * adev)1327 static int mes_v12_0_kiq_enable_queue(struct amdgpu_device *adev)
1328 {
1329 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
1330 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
1331 int r;
1332
1333 if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
1334 return -EINVAL;
1335
1336 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
1337 if (r) {
1338 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
1339 return r;
1340 }
1341
1342 kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring[0]);
1343
1344 r = amdgpu_ring_test_ring(kiq_ring);
1345 if (r) {
1346 DRM_ERROR("kfq enable failed\n");
1347 kiq_ring->sched.ready = false;
1348 }
1349 return r;
1350 }
1351
mes_v12_0_queue_init(struct amdgpu_device * adev,enum amdgpu_mes_pipe pipe)1352 static int mes_v12_0_queue_init(struct amdgpu_device *adev,
1353 enum amdgpu_mes_pipe pipe)
1354 {
1355 struct amdgpu_ring *ring;
1356 int r;
1357
1358 if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1359 ring = &adev->gfx.kiq[0].ring;
1360 else
1361 ring = &adev->mes.ring[pipe];
1362
1363 if ((adev->enable_uni_mes || pipe == AMDGPU_MES_SCHED_PIPE) &&
1364 (amdgpu_in_reset(adev) || adev->in_suspend)) {
1365 *(ring->wptr_cpu_addr) = 0;
1366 *(ring->rptr_cpu_addr) = 0;
1367 amdgpu_ring_clear_ring(ring);
1368 }
1369
1370 r = mes_v12_0_mqd_init(ring);
1371 if (r)
1372 return r;
1373
1374 if (pipe == AMDGPU_MES_SCHED_PIPE) {
1375 if (adev->enable_uni_mes)
1376 r = amdgpu_mes_map_legacy_queue(adev, ring);
1377 else
1378 r = mes_v12_0_kiq_enable_queue(adev);
1379 if (r)
1380 return r;
1381 } else {
1382 mes_v12_0_queue_init_register(ring);
1383 }
1384
1385 if (((pipe == AMDGPU_MES_SCHED_PIPE) && !adev->mes.sched_version) ||
1386 ((pipe == AMDGPU_MES_KIQ_PIPE) && !adev->mes.kiq_version)) {
1387 /* get MES scheduler/KIQ versions */
1388 mutex_lock(&adev->srbm_mutex);
1389 soc21_grbm_select(adev, 3, pipe, 0, 0);
1390
1391 if (pipe == AMDGPU_MES_SCHED_PIPE)
1392 adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
1393 else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
1394 adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
1395
1396 soc21_grbm_select(adev, 0, 0, 0, 0);
1397 mutex_unlock(&adev->srbm_mutex);
1398 }
1399
1400 return 0;
1401 }
1402
mes_v12_0_ring_init(struct amdgpu_device * adev,int pipe)1403 static int mes_v12_0_ring_init(struct amdgpu_device *adev, int pipe)
1404 {
1405 struct amdgpu_ring *ring;
1406
1407 ring = &adev->mes.ring[pipe];
1408
1409 ring->funcs = &mes_v12_0_ring_funcs;
1410
1411 ring->me = 3;
1412 ring->pipe = pipe;
1413 ring->queue = 0;
1414
1415 ring->ring_obj = NULL;
1416 ring->use_doorbell = true;
1417 ring->eop_gpu_addr = adev->mes.eop_gpu_addr[pipe];
1418 ring->no_scheduler = true;
1419 sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
1420
1421 if (pipe == AMDGPU_MES_SCHED_PIPE)
1422 ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
1423 else
1424 ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
1425
1426 return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1427 AMDGPU_RING_PRIO_DEFAULT, NULL);
1428 }
1429
mes_v12_0_kiq_ring_init(struct amdgpu_device * adev)1430 static int mes_v12_0_kiq_ring_init(struct amdgpu_device *adev)
1431 {
1432 struct amdgpu_ring *ring;
1433
1434 spin_lock_init(&adev->gfx.kiq[0].ring_lock);
1435
1436 ring = &adev->gfx.kiq[0].ring;
1437
1438 ring->me = 3;
1439 ring->pipe = 1;
1440 ring->queue = 0;
1441
1442 ring->adev = NULL;
1443 ring->ring_obj = NULL;
1444 ring->use_doorbell = true;
1445 ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
1446 ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
1447 ring->no_scheduler = true;
1448 sprintf(ring->name, "mes_kiq_%d.%d.%d",
1449 ring->me, ring->pipe, ring->queue);
1450
1451 return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
1452 AMDGPU_RING_PRIO_DEFAULT, NULL);
1453 }
1454
mes_v12_0_mqd_sw_init(struct amdgpu_device * adev,enum amdgpu_mes_pipe pipe)1455 static int mes_v12_0_mqd_sw_init(struct amdgpu_device *adev,
1456 enum amdgpu_mes_pipe pipe)
1457 {
1458 int r, mqd_size = sizeof(struct v12_compute_mqd);
1459 struct amdgpu_ring *ring;
1460
1461 if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE)
1462 ring = &adev->gfx.kiq[0].ring;
1463 else
1464 ring = &adev->mes.ring[pipe];
1465
1466 if (ring->mqd_obj)
1467 return 0;
1468
1469 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
1470 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
1471 &ring->mqd_gpu_addr, &ring->mqd_ptr);
1472 if (r) {
1473 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
1474 return r;
1475 }
1476
1477 memset(ring->mqd_ptr, 0, mqd_size);
1478
1479 /* prepare MQD backup */
1480 adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
1481 if (!adev->mes.mqd_backup[pipe])
1482 dev_warn(adev->dev,
1483 "no memory to create MQD backup for ring %s\n",
1484 ring->name);
1485
1486 return 0;
1487 }
1488
mes_v12_0_sw_init(struct amdgpu_ip_block * ip_block)1489 static int mes_v12_0_sw_init(struct amdgpu_ip_block *ip_block)
1490 {
1491 struct amdgpu_device *adev = ip_block->adev;
1492 int pipe, r;
1493
1494 adev->mes.funcs = &mes_v12_0_funcs;
1495 adev->mes.kiq_hw_init = &mes_v12_0_kiq_hw_init;
1496 adev->mes.kiq_hw_fini = &mes_v12_0_kiq_hw_fini;
1497 adev->mes.enable_legacy_queue_map = true;
1498
1499 adev->mes.event_log_size = adev->enable_uni_mes ?
1500 (AMDGPU_MAX_MES_PIPES * (AMDGPU_MES_LOG_BUFFER_SIZE + AMDGPU_MES_MSCRATCH_SIZE)) :
1501 (AMDGPU_MES_LOG_BUFFER_SIZE + AMDGPU_MES_MSCRATCH_SIZE);
1502 r = amdgpu_mes_init(adev);
1503 if (r)
1504 return r;
1505
1506 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1507 r = mes_v12_0_allocate_eop_buf(adev, pipe);
1508 if (r)
1509 return r;
1510
1511 r = mes_v12_0_mqd_sw_init(adev, pipe);
1512 if (r)
1513 return r;
1514
1515 if (!adev->enable_uni_mes && pipe == AMDGPU_MES_KIQ_PIPE) {
1516 r = mes_v12_0_kiq_ring_init(adev);
1517 }
1518 else {
1519 r = mes_v12_0_ring_init(adev, pipe);
1520 if (r)
1521 return r;
1522 r = amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, PAGE_SIZE,
1523 AMDGPU_GEM_DOMAIN_VRAM,
1524 &adev->mes.resource_1[pipe],
1525 &adev->mes.resource_1_gpu_addr[pipe],
1526 &adev->mes.resource_1_addr[pipe]);
1527 if (r) {
1528 dev_err(adev->dev, "(%d) failed to create mes resource_1 bo pipe[%d]\n", r, pipe);
1529 return r;
1530 }
1531 }
1532 }
1533
1534 return 0;
1535 }
1536
mes_v12_0_sw_fini(struct amdgpu_ip_block * ip_block)1537 static int mes_v12_0_sw_fini(struct amdgpu_ip_block *ip_block)
1538 {
1539 struct amdgpu_device *adev = ip_block->adev;
1540 int pipe;
1541
1542 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1543 amdgpu_bo_free_kernel(&adev->mes.resource_1[pipe],
1544 &adev->mes.resource_1_gpu_addr[pipe],
1545 &adev->mes.resource_1_addr[pipe]);
1546
1547 kfree(adev->mes.mqd_backup[pipe]);
1548
1549 amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1550 &adev->mes.eop_gpu_addr[pipe],
1551 NULL);
1552 amdgpu_ucode_release(&adev->mes.fw[pipe]);
1553
1554 if (adev->enable_uni_mes || pipe == AMDGPU_MES_SCHED_PIPE) {
1555 amdgpu_bo_free_kernel(&adev->mes.ring[pipe].mqd_obj,
1556 &adev->mes.ring[pipe].mqd_gpu_addr,
1557 &adev->mes.ring[pipe].mqd_ptr);
1558 amdgpu_ring_fini(&adev->mes.ring[pipe]);
1559 }
1560 }
1561
1562 if (!adev->enable_uni_mes) {
1563 amdgpu_bo_free_kernel(&adev->gfx.kiq[0].ring.mqd_obj,
1564 &adev->gfx.kiq[0].ring.mqd_gpu_addr,
1565 &adev->gfx.kiq[0].ring.mqd_ptr);
1566 amdgpu_ring_fini(&adev->gfx.kiq[0].ring);
1567 }
1568
1569 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1570 mes_v12_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1571 mes_v12_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1572 }
1573
1574 amdgpu_mes_fini(adev);
1575 return 0;
1576 }
1577
mes_v12_0_kiq_dequeue_sched(struct amdgpu_device * adev)1578 static void mes_v12_0_kiq_dequeue_sched(struct amdgpu_device *adev)
1579 {
1580 uint32_t data;
1581 int i;
1582
1583 mutex_lock(&adev->srbm_mutex);
1584 soc21_grbm_select(adev, 3, AMDGPU_MES_SCHED_PIPE, 0, 0);
1585
1586 /* disable the queue if it's active */
1587 if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
1588 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
1589 for (i = 0; i < adev->usec_timeout; i++) {
1590 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
1591 break;
1592 udelay(1);
1593 }
1594 }
1595 data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1596 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1597 DOORBELL_EN, 0);
1598 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1599 DOORBELL_HIT, 1);
1600 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1601
1602 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0);
1603
1604 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0);
1605 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0);
1606 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0);
1607
1608 soc21_grbm_select(adev, 0, 0, 0, 0);
1609 mutex_unlock(&adev->srbm_mutex);
1610
1611 adev->mes.ring[0].sched.ready = false;
1612 }
1613
mes_v12_0_kiq_setting(struct amdgpu_ring * ring)1614 static void mes_v12_0_kiq_setting(struct amdgpu_ring *ring)
1615 {
1616 uint32_t tmp;
1617 struct amdgpu_device *adev = ring->adev;
1618
1619 /* tell RLC which is KIQ queue */
1620 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1621 tmp &= 0xffffff00;
1622 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1623 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp | 0x80);
1624 }
1625
mes_v12_0_kiq_hw_init(struct amdgpu_device * adev)1626 static int mes_v12_0_kiq_hw_init(struct amdgpu_device *adev)
1627 {
1628 int r = 0;
1629 struct amdgpu_ip_block *ip_block;
1630
1631 if (adev->enable_uni_mes)
1632 mes_v12_0_kiq_setting(&adev->mes.ring[AMDGPU_MES_KIQ_PIPE]);
1633 else
1634 mes_v12_0_kiq_setting(&adev->gfx.kiq[0].ring);
1635
1636 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1637
1638 r = mes_v12_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE, false);
1639 if (r) {
1640 DRM_ERROR("failed to load MES fw, r=%d\n", r);
1641 return r;
1642 }
1643
1644 r = mes_v12_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE, true);
1645 if (r) {
1646 DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1647 return r;
1648 }
1649
1650 mes_v12_0_set_ucode_start_addr(adev);
1651
1652 } else if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO)
1653 mes_v12_0_set_ucode_start_addr(adev);
1654
1655 mes_v12_0_enable(adev, true);
1656
1657 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_MES);
1658 if (unlikely(!ip_block)) {
1659 dev_err(adev->dev, "Failed to get MES handle\n");
1660 return -EINVAL;
1661 }
1662
1663 r = mes_v12_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE);
1664 if (r)
1665 goto failure;
1666
1667 if (adev->enable_uni_mes) {
1668 r = mes_v12_0_set_hw_resources(&adev->mes, AMDGPU_MES_KIQ_PIPE);
1669 if (r)
1670 goto failure;
1671
1672 mes_v12_0_set_hw_resources_1(&adev->mes, AMDGPU_MES_KIQ_PIPE);
1673 }
1674
1675 if (adev->mes.enable_legacy_queue_map) {
1676 r = mes_v12_0_hw_init(ip_block);
1677 if (r)
1678 goto failure;
1679 }
1680
1681 return r;
1682
1683 failure:
1684 mes_v12_0_hw_fini(ip_block);
1685 return r;
1686 }
1687
mes_v12_0_kiq_hw_fini(struct amdgpu_device * adev)1688 static int mes_v12_0_kiq_hw_fini(struct amdgpu_device *adev)
1689 {
1690 if (adev->mes.ring[0].sched.ready) {
1691 if (adev->enable_uni_mes)
1692 amdgpu_mes_unmap_legacy_queue(adev,
1693 &adev->mes.ring[AMDGPU_MES_SCHED_PIPE],
1694 RESET_QUEUES, 0, 0);
1695 else
1696 mes_v12_0_kiq_dequeue_sched(adev);
1697
1698 adev->mes.ring[0].sched.ready = false;
1699 }
1700
1701 mes_v12_0_enable(adev, false);
1702
1703 return 0;
1704 }
1705
mes_v12_0_hw_init(struct amdgpu_ip_block * ip_block)1706 static int mes_v12_0_hw_init(struct amdgpu_ip_block *ip_block)
1707 {
1708 int r;
1709 struct amdgpu_device *adev = ip_block->adev;
1710
1711 if (adev->mes.ring[0].sched.ready)
1712 goto out;
1713
1714 if (!adev->enable_mes_kiq) {
1715 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1716 r = mes_v12_0_load_microcode(adev,
1717 AMDGPU_MES_SCHED_PIPE, true);
1718 if (r) {
1719 DRM_ERROR("failed to MES fw, r=%d\n", r);
1720 return r;
1721 }
1722
1723 mes_v12_0_set_ucode_start_addr(adev);
1724
1725 } else if (adev->firmware.load_type ==
1726 AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) {
1727
1728 mes_v12_0_set_ucode_start_addr(adev);
1729 }
1730
1731 mes_v12_0_enable(adev, true);
1732 }
1733
1734 /* Enable the MES to handle doorbell ring on unmapped queue */
1735 mes_v12_0_enable_unmapped_doorbell_handling(&adev->mes, true);
1736
1737 r = mes_v12_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE);
1738 if (r)
1739 goto failure;
1740
1741 r = mes_v12_0_set_hw_resources(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1742 if (r)
1743 goto failure;
1744
1745 if ((adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 0x4b)
1746 mes_v12_0_set_hw_resources_1(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1747
1748 mes_v12_0_init_aggregated_doorbell(&adev->mes);
1749
1750 r = mes_v12_0_query_sched_status(&adev->mes, AMDGPU_MES_SCHED_PIPE);
1751 if (r) {
1752 DRM_ERROR("MES is busy\n");
1753 goto failure;
1754 }
1755
1756 r = amdgpu_mes_update_enforce_isolation(adev);
1757 if (r)
1758 goto failure;
1759
1760 out:
1761 /*
1762 * Disable KIQ ring usage from the driver once MES is enabled.
1763 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1764 * with MES enabled.
1765 */
1766 adev->gfx.kiq[0].ring.sched.ready = false;
1767 adev->mes.ring[0].sched.ready = true;
1768
1769 return 0;
1770
1771 failure:
1772 mes_v12_0_hw_fini(ip_block);
1773 return r;
1774 }
1775
mes_v12_0_hw_fini(struct amdgpu_ip_block * ip_block)1776 static int mes_v12_0_hw_fini(struct amdgpu_ip_block *ip_block)
1777 {
1778 return 0;
1779 }
1780
mes_v12_0_suspend(struct amdgpu_ip_block * ip_block)1781 static int mes_v12_0_suspend(struct amdgpu_ip_block *ip_block)
1782 {
1783 return mes_v12_0_hw_fini(ip_block);
1784 }
1785
mes_v12_0_resume(struct amdgpu_ip_block * ip_block)1786 static int mes_v12_0_resume(struct amdgpu_ip_block *ip_block)
1787 {
1788 return mes_v12_0_hw_init(ip_block);
1789 }
1790
mes_v12_0_early_init(struct amdgpu_ip_block * ip_block)1791 static int mes_v12_0_early_init(struct amdgpu_ip_block *ip_block)
1792 {
1793 struct amdgpu_device *adev = ip_block->adev;
1794 int pipe, r;
1795
1796 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1797 r = amdgpu_mes_init_microcode(adev, pipe);
1798 if (r)
1799 return r;
1800 }
1801
1802 return 0;
1803 }
1804
1805 static const struct amd_ip_funcs mes_v12_0_ip_funcs = {
1806 .name = "mes_v12_0",
1807 .early_init = mes_v12_0_early_init,
1808 .late_init = NULL,
1809 .sw_init = mes_v12_0_sw_init,
1810 .sw_fini = mes_v12_0_sw_fini,
1811 .hw_init = mes_v12_0_hw_init,
1812 .hw_fini = mes_v12_0_hw_fini,
1813 .suspend = mes_v12_0_suspend,
1814 .resume = mes_v12_0_resume,
1815 };
1816
1817 const struct amdgpu_ip_block_version mes_v12_0_ip_block = {
1818 .type = AMD_IP_BLOCK_TYPE_MES,
1819 .major = 12,
1820 .minor = 0,
1821 .rev = 0,
1822 .funcs = &mes_v12_0_ip_funcs,
1823 };
1824