2 * Copyright 2019 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
24 #include <linux/firmware.h>
25 #include <linux/module.h>
27 #include "soc15_common.h"
29 #include "gc/gc_11_0_0_offset.h"
30 #include "gc/gc_11_0_0_sh_mask.h"
31 #include "gc/gc_11_0_0_default.h"
32 #include "v11_structs.h"
33 #include "mes_v11_api_def.h"
35 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes.bin");
36 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes1.bin");
37 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes.bin");
38 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes1.bin");
39 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes.bin");
40 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes1.bin");
42 static int mes_v11_0_hw_fini(void *handle);
43 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev);
44 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev);
46 #define MES_EOP_SIZE 2048
48 static void mes_v11_0_ring_set_wptr(struct amdgpu_ring *ring)
50 struct amdgpu_device *adev = ring->adev;
52 if (ring->use_doorbell) {
53 atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
55 WDOORBELL64(ring->doorbell_index, ring->wptr);
61 static u64 mes_v11_0_ring_get_rptr(struct amdgpu_ring *ring)
63 return *ring->rptr_cpu_addr;
66 static u64 mes_v11_0_ring_get_wptr(struct amdgpu_ring *ring)
70 if (ring->use_doorbell)
71 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
77 static const struct amdgpu_ring_funcs mes_v11_0_ring_funcs = {
78 .type = AMDGPU_RING_TYPE_MES,
81 .support_64bit_ptrs = true,
82 .get_rptr = mes_v11_0_ring_get_rptr,
83 .get_wptr = mes_v11_0_ring_get_wptr,
84 .set_wptr = mes_v11_0_ring_set_wptr,
85 .insert_nop = amdgpu_ring_insert_nop,
88 static int mes_v11_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
93 union MESAPI__ADD_QUEUE *x_pkt = pkt;
94 struct amdgpu_device *adev = mes->adev;
95 struct amdgpu_ring *ring = &mes->ring;
97 BUG_ON(size % 4 != 0);
99 if (amdgpu_ring_alloc(ring, ndw))
102 amdgpu_ring_write_multiple(ring, pkt, ndw);
103 amdgpu_ring_commit(ring);
105 DRM_DEBUG("MES msg=%d was emitted\n", x_pkt->header.opcode);
107 r = amdgpu_fence_wait_polling(ring, ring->fence_drv.sync_seq,
108 adev->usec_timeout * (amdgpu_emu_mode ? 100 : 1));
110 DRM_ERROR("MES failed to response msg=%d\n",
111 x_pkt->header.opcode);
118 static int convert_to_mes_queue_type(int queue_type)
120 if (queue_type == AMDGPU_RING_TYPE_GFX)
121 return MES_QUEUE_TYPE_GFX;
122 else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
123 return MES_QUEUE_TYPE_COMPUTE;
124 else if (queue_type == AMDGPU_RING_TYPE_SDMA)
125 return MES_QUEUE_TYPE_SDMA;
131 static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
132 struct mes_add_queue_input *input)
134 struct amdgpu_device *adev = mes->adev;
135 union MESAPI__ADD_QUEUE mes_add_queue_pkt;
136 struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB_0];
137 uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
139 memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
141 mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
142 mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
143 mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
145 mes_add_queue_pkt.process_id = input->process_id;
146 mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
147 mes_add_queue_pkt.process_va_start = input->process_va_start;
148 mes_add_queue_pkt.process_va_end = input->process_va_end;
149 mes_add_queue_pkt.process_quantum = input->process_quantum;
150 mes_add_queue_pkt.process_context_addr = input->process_context_addr;
151 mes_add_queue_pkt.gang_quantum = input->gang_quantum;
152 mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
153 mes_add_queue_pkt.inprocess_gang_priority =
154 input->inprocess_gang_priority;
155 mes_add_queue_pkt.gang_global_priority_level =
156 input->gang_global_priority_level;
157 mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
158 mes_add_queue_pkt.mqd_addr = input->mqd_addr;
159 mes_add_queue_pkt.wptr_addr = input->wptr_addr;
160 mes_add_queue_pkt.queue_type =
161 convert_to_mes_queue_type(input->queue_type);
162 mes_add_queue_pkt.paging = input->paging;
163 mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
164 mes_add_queue_pkt.gws_base = input->gws_base;
165 mes_add_queue_pkt.gws_size = input->gws_size;
166 mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
167 mes_add_queue_pkt.tma_addr = input->tma_addr;
169 mes_add_queue_pkt.api_status.api_completion_fence_addr =
170 mes->ring.fence_drv.gpu_addr;
171 mes_add_queue_pkt.api_status.api_completion_fence_value =
172 ++mes->ring.fence_drv.sync_seq;
174 return mes_v11_0_submit_pkt_and_poll_completion(mes,
175 &mes_add_queue_pkt, sizeof(mes_add_queue_pkt));
178 static int mes_v11_0_remove_hw_queue(struct amdgpu_mes *mes,
179 struct mes_remove_queue_input *input)
181 union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
183 memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
185 mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
186 mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
187 mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
189 mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
190 mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
192 mes_remove_queue_pkt.api_status.api_completion_fence_addr =
193 mes->ring.fence_drv.gpu_addr;
194 mes_remove_queue_pkt.api_status.api_completion_fence_value =
195 ++mes->ring.fence_drv.sync_seq;
197 return mes_v11_0_submit_pkt_and_poll_completion(mes,
198 &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt));
201 static int mes_v11_0_unmap_legacy_queue(struct amdgpu_mes *mes,
202 struct mes_unmap_legacy_queue_input *input)
204 union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
206 memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
208 mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
209 mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
210 mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
212 mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset << 2;
213 mes_remove_queue_pkt.gang_context_addr = 0;
215 mes_remove_queue_pkt.pipe_id = input->pipe_id;
216 mes_remove_queue_pkt.queue_id = input->queue_id;
218 if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
219 mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
220 mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
221 mes_remove_queue_pkt.tf_data =
222 lower_32_bits(input->trail_fence_data);
224 if (input->queue_type == AMDGPU_RING_TYPE_GFX)
225 mes_remove_queue_pkt.unmap_legacy_gfx_queue = 1;
227 mes_remove_queue_pkt.unmap_kiq_utility_queue = 1;
230 mes_remove_queue_pkt.api_status.api_completion_fence_addr =
231 mes->ring.fence_drv.gpu_addr;
232 mes_remove_queue_pkt.api_status.api_completion_fence_value =
233 ++mes->ring.fence_drv.sync_seq;
235 return mes_v11_0_submit_pkt_and_poll_completion(mes,
236 &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt));
239 static int mes_v11_0_suspend_gang(struct amdgpu_mes *mes,
240 struct mes_suspend_gang_input *input)
245 static int mes_v11_0_resume_gang(struct amdgpu_mes *mes,
246 struct mes_resume_gang_input *input)
251 static int mes_v11_0_query_sched_status(struct amdgpu_mes *mes)
253 union MESAPI__QUERY_MES_STATUS mes_status_pkt;
255 memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
257 mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
258 mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
259 mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
261 mes_status_pkt.api_status.api_completion_fence_addr =
262 mes->ring.fence_drv.gpu_addr;
263 mes_status_pkt.api_status.api_completion_fence_value =
264 ++mes->ring.fence_drv.sync_seq;
266 return mes_v11_0_submit_pkt_and_poll_completion(mes,
267 &mes_status_pkt, sizeof(mes_status_pkt));
270 static int mes_v11_0_set_hw_resources(struct amdgpu_mes *mes)
273 struct amdgpu_device *adev = mes->adev;
274 union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
276 memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
278 mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
279 mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
280 mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
282 mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
283 mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
284 mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
285 mes_set_hw_res_pkt.paging_vmid = 0;
286 mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr = mes->sch_ctx_gpu_addr;
287 mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
288 mes->query_status_fence_gpu_addr;
290 for (i = 0; i < MAX_COMPUTE_PIPES; i++)
291 mes_set_hw_res_pkt.compute_hqd_mask[i] =
292 mes->compute_hqd_mask[i];
294 for (i = 0; i < MAX_GFX_PIPES; i++)
295 mes_set_hw_res_pkt.gfx_hqd_mask[i] = mes->gfx_hqd_mask[i];
297 for (i = 0; i < MAX_SDMA_PIPES; i++)
298 mes_set_hw_res_pkt.sdma_hqd_mask[i] = mes->sdma_hqd_mask[i];
300 for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
301 mes_set_hw_res_pkt.aggregated_doorbells[i] =
302 mes->agreegated_doorbells[i];
304 for (i = 0; i < 5; i++) {
305 mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
306 mes_set_hw_res_pkt.mmhub_base[i] =
307 adev->reg_offset[MMHUB_HWIP][0][i];
308 mes_set_hw_res_pkt.osssys_base[i] =
309 adev->reg_offset[OSSSYS_HWIP][0][i];
312 mes_set_hw_res_pkt.disable_reset = 1;
313 mes_set_hw_res_pkt.disable_mes_log = 1;
314 mes_set_hw_res_pkt.use_different_vmid_compute = 1;
316 mes_set_hw_res_pkt.api_status.api_completion_fence_addr =
317 mes->ring.fence_drv.gpu_addr;
318 mes_set_hw_res_pkt.api_status.api_completion_fence_value =
319 ++mes->ring.fence_drv.sync_seq;
321 return mes_v11_0_submit_pkt_and_poll_completion(mes,
322 &mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt));
325 static const struct amdgpu_mes_funcs mes_v11_0_funcs = {
326 .add_hw_queue = mes_v11_0_add_hw_queue,
327 .remove_hw_queue = mes_v11_0_remove_hw_queue,
328 .unmap_legacy_queue = mes_v11_0_unmap_legacy_queue,
329 .suspend_gang = mes_v11_0_suspend_gang,
330 .resume_gang = mes_v11_0_resume_gang,
333 static int mes_v11_0_init_microcode(struct amdgpu_device *adev,
334 enum admgpu_mes_pipe pipe)
337 char ucode_prefix[30];
339 const struct mes_firmware_header_v1_0 *mes_hdr;
340 struct amdgpu_firmware_info *info;
342 amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix, sizeof(ucode_prefix));
344 if (pipe == AMDGPU_MES_SCHED_PIPE)
345 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes.bin",
348 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes1.bin",
351 err = request_firmware(&adev->mes.fw[pipe], fw_name, adev->dev);
355 err = amdgpu_ucode_validate(adev->mes.fw[pipe]);
357 release_firmware(adev->mes.fw[pipe]);
358 adev->mes.fw[pipe] = NULL;
362 mes_hdr = (const struct mes_firmware_header_v1_0 *)
363 adev->mes.fw[pipe]->data;
364 adev->mes.ucode_fw_version[pipe] =
365 le32_to_cpu(mes_hdr->mes_ucode_version);
366 adev->mes.ucode_fw_version[pipe] =
367 le32_to_cpu(mes_hdr->mes_ucode_data_version);
368 adev->mes.uc_start_addr[pipe] =
369 le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) |
370 ((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32);
371 adev->mes.data_start_addr[pipe] =
372 le32_to_cpu(mes_hdr->mes_data_start_addr_lo) |
373 ((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32);
375 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
376 int ucode, ucode_data;
378 if (pipe == AMDGPU_MES_SCHED_PIPE) {
379 ucode = AMDGPU_UCODE_ID_CP_MES;
380 ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA;
382 ucode = AMDGPU_UCODE_ID_CP_MES1;
383 ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA;
386 info = &adev->firmware.ucode[ucode];
387 info->ucode_id = ucode;
388 info->fw = adev->mes.fw[pipe];
389 adev->firmware.fw_size +=
390 ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes),
393 info = &adev->firmware.ucode[ucode_data];
394 info->ucode_id = ucode_data;
395 info->fw = adev->mes.fw[pipe];
396 adev->firmware.fw_size +=
397 ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes),
404 static void mes_v11_0_free_microcode(struct amdgpu_device *adev,
405 enum admgpu_mes_pipe pipe)
407 release_firmware(adev->mes.fw[pipe]);
408 adev->mes.fw[pipe] = NULL;
411 static int mes_v11_0_allocate_ucode_buffer(struct amdgpu_device *adev,
412 enum admgpu_mes_pipe pipe)
415 const struct mes_firmware_header_v1_0 *mes_hdr;
416 const __le32 *fw_data;
419 mes_hdr = (const struct mes_firmware_header_v1_0 *)
420 adev->mes.fw[pipe]->data;
422 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
423 le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
424 fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
426 r = amdgpu_bo_create_reserved(adev, fw_size,
427 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
428 &adev->mes.ucode_fw_obj[pipe],
429 &adev->mes.ucode_fw_gpu_addr[pipe],
430 (void **)&adev->mes.ucode_fw_ptr[pipe]);
432 dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
436 memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
438 amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
439 amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
444 static int mes_v11_0_allocate_ucode_data_buffer(struct amdgpu_device *adev,
445 enum admgpu_mes_pipe pipe)
448 const struct mes_firmware_header_v1_0 *mes_hdr;
449 const __le32 *fw_data;
452 mes_hdr = (const struct mes_firmware_header_v1_0 *)
453 adev->mes.fw[pipe]->data;
455 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
456 le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
457 fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
459 r = amdgpu_bo_create_reserved(adev, fw_size,
460 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
461 &adev->mes.data_fw_obj[pipe],
462 &adev->mes.data_fw_gpu_addr[pipe],
463 (void **)&adev->mes.data_fw_ptr[pipe]);
465 dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
469 memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
471 amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
472 amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
477 static void mes_v11_0_free_ucode_buffers(struct amdgpu_device *adev,
478 enum admgpu_mes_pipe pipe)
480 amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
481 &adev->mes.data_fw_gpu_addr[pipe],
482 (void **)&adev->mes.data_fw_ptr[pipe]);
484 amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
485 &adev->mes.ucode_fw_gpu_addr[pipe],
486 (void **)&adev->mes.ucode_fw_ptr[pipe]);
489 static void mes_v11_0_enable(struct amdgpu_device *adev, bool enable)
492 uint32_t pipe, data = 0;
495 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
496 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
497 data = REG_SET_FIELD(data, CP_MES_CNTL,
498 MES_PIPE1_RESET, adev->enable_mes_kiq ? 1 : 0);
499 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
501 mutex_lock(&adev->srbm_mutex);
502 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
503 if (!adev->enable_mes_kiq &&
504 pipe == AMDGPU_MES_KIQ_PIPE)
507 soc21_grbm_select(adev, 3, pipe, 0, 0);
509 ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
510 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
511 lower_32_bits(ucode_addr));
512 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
513 upper_32_bits(ucode_addr));
515 soc21_grbm_select(adev, 0, 0, 0, 0);
516 mutex_unlock(&adev->srbm_mutex);
518 /* unhalt MES and activate pipe0 */
519 data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
520 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE,
521 adev->enable_mes_kiq ? 1 : 0);
522 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
529 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
530 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
531 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
532 data = REG_SET_FIELD(data, CP_MES_CNTL,
533 MES_INVALIDATE_ICACHE, 1);
534 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
535 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET,
536 adev->enable_mes_kiq ? 1 : 0);
537 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
538 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
542 /* This function is for backdoor MES firmware */
543 static int mes_v11_0_load_microcode(struct amdgpu_device *adev,
544 enum admgpu_mes_pipe pipe)
550 mes_v11_0_enable(adev, false);
552 if (!adev->mes.fw[pipe])
555 r = mes_v11_0_allocate_ucode_buffer(adev, pipe);
559 r = mes_v11_0_allocate_ucode_data_buffer(adev, pipe);
561 mes_v11_0_free_ucode_buffers(adev, pipe);
565 mutex_lock(&adev->srbm_mutex);
566 /* me=3, pipe=0, queue=0 */
567 soc21_grbm_select(adev, 3, pipe, 0, 0);
569 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0);
571 /* set ucode start address */
572 ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
573 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
574 lower_32_bits(ucode_addr));
575 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
576 upper_32_bits(ucode_addr));
578 /* set ucode fimrware address */
579 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO,
580 lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
581 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI,
582 upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
584 /* set ucode instruction cache boundary to 2M-1 */
585 WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF);
587 /* set ucode data firmware address */
588 WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO,
589 lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
590 WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI,
591 upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
593 /* Set 0x3FFFF (256K-1) to CP_MES_MDBOUND_LO */
594 WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x3FFFF);
596 /* invalidate ICACHE */
597 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
598 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
599 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
600 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
602 /* prime the ICACHE. */
603 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
604 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
605 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
607 soc21_grbm_select(adev, 0, 0, 0, 0);
608 mutex_unlock(&adev->srbm_mutex);
613 static int mes_v11_0_allocate_eop_buf(struct amdgpu_device *adev,
614 enum admgpu_mes_pipe pipe)
619 r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
620 AMDGPU_GEM_DOMAIN_GTT,
621 &adev->mes.eop_gpu_obj[pipe],
622 &adev->mes.eop_gpu_addr[pipe],
625 dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
630 adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
632 amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
633 amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
638 static int mes_v11_0_mqd_init(struct amdgpu_ring *ring)
640 struct v11_compute_mqd *mqd = ring->mqd_ptr;
641 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
644 mqd->header = 0xC0310800;
645 mqd->compute_pipelinestat_enable = 0x00000001;
646 mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
647 mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
648 mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
649 mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
650 mqd->compute_misc_reserved = 0x00000007;
652 eop_base_addr = ring->eop_gpu_addr >> 8;
654 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
655 tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
656 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
657 (order_base_2(MES_EOP_SIZE / 4) - 1));
659 mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
660 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
661 mqd->cp_hqd_eop_control = tmp;
663 /* disable the queue if it's active */
665 mqd->cp_hqd_pq_rptr = 0;
666 mqd->cp_hqd_pq_wptr_lo = 0;
667 mqd->cp_hqd_pq_wptr_hi = 0;
669 /* set the pointer to the MQD */
670 mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
671 mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
673 /* set MQD vmid to 0 */
674 tmp = regCP_MQD_CONTROL_DEFAULT;
675 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
676 mqd->cp_mqd_control = tmp;
678 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
679 hqd_gpu_addr = ring->gpu_addr >> 8;
680 mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
681 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
683 /* set the wb address whether it's enabled or not */
684 wb_gpu_addr = ring->rptr_gpu_addr;
685 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
686 mqd->cp_hqd_pq_rptr_report_addr_hi =
687 upper_32_bits(wb_gpu_addr) & 0xffff;
689 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
690 wb_gpu_addr = ring->wptr_gpu_addr;
691 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
692 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
694 /* set up the HQD, this is similar to CP_RB0_CNTL */
695 tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
696 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
697 (order_base_2(ring->ring_size / 4) - 1));
698 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
699 ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
700 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
701 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
702 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
703 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
704 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
705 mqd->cp_hqd_pq_control = tmp;
707 /* enable doorbell */
709 if (ring->use_doorbell) {
710 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
711 DOORBELL_OFFSET, ring->doorbell_index);
712 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
714 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
716 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
720 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
722 mqd->cp_hqd_pq_doorbell_control = tmp;
724 mqd->cp_hqd_vmid = 0;
725 /* activate the queue */
726 mqd->cp_hqd_active = 1;
728 tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
729 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
731 mqd->cp_hqd_persistent_state = tmp;
733 mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT;
734 mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
735 mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
740 static void mes_v11_0_queue_init_register(struct amdgpu_ring *ring)
742 struct v11_compute_mqd *mqd = ring->mqd_ptr;
743 struct amdgpu_device *adev = ring->adev;
746 mutex_lock(&adev->srbm_mutex);
747 soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
749 /* set CP_HQD_VMID.VMID = 0. */
750 data = RREG32_SOC15(GC, 0, regCP_HQD_VMID);
751 data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
752 WREG32_SOC15(GC, 0, regCP_HQD_VMID, data);
754 /* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
755 data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
756 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
758 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
760 /* set CP_MQD_BASE_ADDR/HI with the MQD base address */
761 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
762 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
764 /* set CP_MQD_CONTROL.VMID=0 */
765 data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL);
766 data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
767 WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0);
769 /* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
770 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
771 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
773 /* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
774 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
775 mqd->cp_hqd_pq_rptr_report_addr_lo);
776 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
777 mqd->cp_hqd_pq_rptr_report_addr_hi);
779 /* set CP_HQD_PQ_CONTROL */
780 WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
782 /* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
783 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
784 mqd->cp_hqd_pq_wptr_poll_addr_lo);
785 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
786 mqd->cp_hqd_pq_wptr_poll_addr_hi);
788 /* set CP_HQD_PQ_DOORBELL_CONTROL */
789 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
790 mqd->cp_hqd_pq_doorbell_control);
792 /* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
793 WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
795 /* set CP_HQD_ACTIVE.ACTIVE=1 */
796 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active);
798 soc21_grbm_select(adev, 0, 0, 0, 0);
799 mutex_unlock(&adev->srbm_mutex);
802 static int mes_v11_0_kiq_enable_queue(struct amdgpu_device *adev)
804 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
805 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
808 if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
811 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
813 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
817 kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring);
819 r = amdgpu_ring_test_ring(kiq_ring);
821 DRM_ERROR("kfq enable failed\n");
822 kiq_ring->sched.ready = false;
827 static int mes_v11_0_queue_init(struct amdgpu_device *adev,
828 enum admgpu_mes_pipe pipe)
830 struct amdgpu_ring *ring;
833 if (pipe == AMDGPU_MES_KIQ_PIPE)
834 ring = &adev->gfx.kiq.ring;
835 else if (pipe == AMDGPU_MES_SCHED_PIPE)
836 ring = &adev->mes.ring;
840 if ((pipe == AMDGPU_MES_SCHED_PIPE) &&
841 (amdgpu_in_reset(adev) || adev->in_suspend)) {
842 *(ring->wptr_cpu_addr) = 0;
843 *(ring->rptr_cpu_addr) = 0;
844 amdgpu_ring_clear_ring(ring);
847 r = mes_v11_0_mqd_init(ring);
851 if (pipe == AMDGPU_MES_SCHED_PIPE) {
852 r = mes_v11_0_kiq_enable_queue(adev);
856 mes_v11_0_queue_init_register(ring);
862 static int mes_v11_0_ring_init(struct amdgpu_device *adev)
864 struct amdgpu_ring *ring;
866 ring = &adev->mes.ring;
868 ring->funcs = &mes_v11_0_ring_funcs;
874 ring->ring_obj = NULL;
875 ring->use_doorbell = true;
876 ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
877 ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_SCHED_PIPE];
878 ring->no_scheduler = true;
879 sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
881 return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
882 AMDGPU_RING_PRIO_DEFAULT, NULL);
885 static int mes_v11_0_kiq_ring_init(struct amdgpu_device *adev)
887 struct amdgpu_ring *ring;
889 spin_lock_init(&adev->gfx.kiq.ring_lock);
891 ring = &adev->gfx.kiq.ring;
898 ring->ring_obj = NULL;
899 ring->use_doorbell = true;
900 ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
901 ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
902 ring->no_scheduler = true;
903 sprintf(ring->name, "mes_kiq_%d.%d.%d",
904 ring->me, ring->pipe, ring->queue);
906 return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
907 AMDGPU_RING_PRIO_DEFAULT, NULL);
910 static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev,
911 enum admgpu_mes_pipe pipe)
913 int r, mqd_size = sizeof(struct v11_compute_mqd);
914 struct amdgpu_ring *ring;
916 if (pipe == AMDGPU_MES_KIQ_PIPE)
917 ring = &adev->gfx.kiq.ring;
918 else if (pipe == AMDGPU_MES_SCHED_PIPE)
919 ring = &adev->mes.ring;
926 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
927 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
928 &ring->mqd_gpu_addr, &ring->mqd_ptr);
930 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
934 memset(ring->mqd_ptr, 0, mqd_size);
936 /* prepare MQD backup */
937 adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
938 if (!adev->mes.mqd_backup[pipe])
940 "no memory to create MQD backup for ring %s\n",
946 static int mes_v11_0_sw_init(void *handle)
948 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
951 adev->mes.adev = adev;
952 adev->mes.funcs = &mes_v11_0_funcs;
953 adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init;
954 adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini;
956 r = amdgpu_mes_init(adev);
960 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
961 if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
964 r = mes_v11_0_init_microcode(adev, pipe);
968 r = mes_v11_0_allocate_eop_buf(adev, pipe);
972 r = mes_v11_0_mqd_sw_init(adev, pipe);
977 if (adev->enable_mes_kiq) {
978 r = mes_v11_0_kiq_ring_init(adev);
983 r = mes_v11_0_ring_init(adev);
990 static int mes_v11_0_sw_fini(void *handle)
992 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
995 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
996 amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
998 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
999 kfree(adev->mes.mqd_backup[pipe]);
1001 amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1002 &adev->mes.eop_gpu_addr[pipe],
1005 mes_v11_0_free_microcode(adev, pipe);
1008 amdgpu_bo_free_kernel(&adev->gfx.kiq.ring.mqd_obj,
1009 &adev->gfx.kiq.ring.mqd_gpu_addr,
1010 &adev->gfx.kiq.ring.mqd_ptr);
1012 amdgpu_bo_free_kernel(&adev->mes.ring.mqd_obj,
1013 &adev->mes.ring.mqd_gpu_addr,
1014 &adev->mes.ring.mqd_ptr);
1016 amdgpu_ring_fini(&adev->gfx.kiq.ring);
1017 amdgpu_ring_fini(&adev->mes.ring);
1019 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1020 mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1021 mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1024 amdgpu_mes_fini(adev);
1028 static void mes_v11_0_kiq_setting(struct amdgpu_ring *ring)
1031 struct amdgpu_device *adev = ring->adev;
1033 /* tell RLC which is KIQ queue */
1034 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1036 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1037 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1039 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1042 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev)
1046 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1047 r = mes_v11_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE);
1049 DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1053 r = mes_v11_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE);
1055 DRM_ERROR("failed to load MES fw, r=%d\n", r);
1060 mes_v11_0_enable(adev, true);
1062 mes_v11_0_kiq_setting(&adev->gfx.kiq.ring);
1064 r = mes_v11_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE);
1071 mes_v11_0_hw_fini(adev);
1075 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev)
1077 mes_v11_0_enable(adev, false);
1081 static int mes_v11_0_hw_init(void *handle)
1084 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1086 if (!adev->enable_mes_kiq) {
1087 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1088 r = mes_v11_0_load_microcode(adev,
1089 AMDGPU_MES_SCHED_PIPE);
1091 DRM_ERROR("failed to MES fw, r=%d\n", r);
1096 mes_v11_0_enable(adev, true);
1099 r = mes_v11_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE);
1103 r = mes_v11_0_set_hw_resources(&adev->mes);
1107 r = mes_v11_0_query_sched_status(&adev->mes);
1109 DRM_ERROR("MES is busy\n");
1114 * Disable KIQ ring usage from the driver once MES is enabled.
1115 * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1118 adev->gfx.kiq.ring.sched.ready = false;
1123 mes_v11_0_hw_fini(adev);
1127 static int mes_v11_0_hw_fini(void *handle)
1132 static int mes_v11_0_suspend(void *handle)
1135 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1137 r = amdgpu_mes_suspend(adev);
1141 return mes_v11_0_hw_fini(adev);
1144 static int mes_v11_0_resume(void *handle)
1147 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1149 r = mes_v11_0_hw_init(adev);
1153 return amdgpu_mes_resume(adev);
1156 static int mes_v11_0_late_init(void *handle)
1158 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1160 amdgpu_mes_self_test(adev);
1165 static const struct amd_ip_funcs mes_v11_0_ip_funcs = {
1166 .name = "mes_v11_0",
1167 .late_init = mes_v11_0_late_init,
1168 .sw_init = mes_v11_0_sw_init,
1169 .sw_fini = mes_v11_0_sw_fini,
1170 .hw_init = mes_v11_0_hw_init,
1171 .hw_fini = mes_v11_0_hw_fini,
1172 .suspend = mes_v11_0_suspend,
1173 .resume = mes_v11_0_resume,
1176 const struct amdgpu_ip_block_version mes_v11_0_ip_block = {
1177 .type = AMD_IP_BLOCK_TYPE_MES,
1181 .funcs = &mes_v11_0_ip_funcs,