1 /* 2 * Copyright 2019 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24 #include <linux/firmware.h> 25 #include <linux/module.h> 26 #include "amdgpu.h" 27 #include "soc15_common.h" 28 #include "soc21.h" 29 #include "gc/gc_11_0_0_offset.h" 30 #include "gc/gc_11_0_0_sh_mask.h" 31 #include "gc/gc_11_0_0_default.h" 32 #include "v11_structs.h" 33 #include "mes_v11_api_def.h" 34 35 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes.bin"); 36 MODULE_FIRMWARE("amdgpu/gc_11_0_0_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"); 41 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes.bin"); 42 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes1.bin"); 43 44 static int mes_v11_0_hw_fini(void *handle); 45 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev); 46 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev); 47 48 #define MES_EOP_SIZE 2048 49 50 static void mes_v11_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 63 static u64 mes_v11_0_ring_get_rptr(struct amdgpu_ring *ring) 64 { 65 return *ring->rptr_cpu_addr; 66 } 67 68 static u64 mes_v11_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_v11_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_v11_0_ring_get_rptr, 85 .get_wptr = mes_v11_0_ring_get_wptr, 86 .set_wptr = mes_v11_0_ring_set_wptr, 87 .insert_nop = amdgpu_ring_insert_nop, 88 }; 89 90 static int mes_v11_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes, 91 void *pkt, int size, 92 int api_status_off) 93 { 94 int ndw = size / 4; 95 signed long r; 96 union MESAPI__ADD_QUEUE *x_pkt = pkt; 97 struct MES_API_STATUS *api_status; 98 struct amdgpu_device *adev = mes->adev; 99 struct amdgpu_ring *ring = &mes->ring; 100 unsigned long flags; 101 signed long timeout = adev->usec_timeout; 102 103 if (amdgpu_emu_mode) { 104 timeout *= 100; 105 } else if (amdgpu_sriov_vf(adev)) { 106 /* Worst case in sriov where all other 15 VF timeout, each VF needs about 600ms */ 107 timeout = 15 * 600 * 1000; 108 } 109 BUG_ON(size % 4 != 0); 110 111 spin_lock_irqsave(&mes->ring_lock, flags); 112 if (amdgpu_ring_alloc(ring, ndw)) { 113 spin_unlock_irqrestore(&mes->ring_lock, flags); 114 return -ENOMEM; 115 } 116 117 api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off); 118 api_status->api_completion_fence_addr = mes->ring.fence_drv.gpu_addr; 119 api_status->api_completion_fence_value = ++mes->ring.fence_drv.sync_seq; 120 121 amdgpu_ring_write_multiple(ring, pkt, ndw); 122 amdgpu_ring_commit(ring); 123 spin_unlock_irqrestore(&mes->ring_lock, flags); 124 125 DRM_DEBUG("MES msg=%d was emitted\n", x_pkt->header.opcode); 126 127 r = amdgpu_fence_wait_polling(ring, ring->fence_drv.sync_seq, 128 timeout); 129 if (r < 1) { 130 DRM_ERROR("MES failed to response msg=%d\n", 131 x_pkt->header.opcode); 132 return -ETIMEDOUT; 133 } 134 135 return 0; 136 } 137 138 static int convert_to_mes_queue_type(int queue_type) 139 { 140 if (queue_type == AMDGPU_RING_TYPE_GFX) 141 return MES_QUEUE_TYPE_GFX; 142 else if (queue_type == AMDGPU_RING_TYPE_COMPUTE) 143 return MES_QUEUE_TYPE_COMPUTE; 144 else if (queue_type == AMDGPU_RING_TYPE_SDMA) 145 return MES_QUEUE_TYPE_SDMA; 146 else 147 BUG(); 148 return -1; 149 } 150 151 static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes, 152 struct mes_add_queue_input *input) 153 { 154 struct amdgpu_device *adev = mes->adev; 155 union MESAPI__ADD_QUEUE mes_add_queue_pkt; 156 struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB_0]; 157 uint32_t vm_cntx_cntl = hub->vm_cntx_cntl; 158 159 memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt)); 160 161 mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER; 162 mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE; 163 mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS; 164 165 mes_add_queue_pkt.process_id = input->process_id; 166 mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr; 167 mes_add_queue_pkt.process_va_start = input->process_va_start; 168 mes_add_queue_pkt.process_va_end = input->process_va_end; 169 mes_add_queue_pkt.process_quantum = input->process_quantum; 170 mes_add_queue_pkt.process_context_addr = input->process_context_addr; 171 mes_add_queue_pkt.gang_quantum = input->gang_quantum; 172 mes_add_queue_pkt.gang_context_addr = input->gang_context_addr; 173 mes_add_queue_pkt.inprocess_gang_priority = 174 input->inprocess_gang_priority; 175 mes_add_queue_pkt.gang_global_priority_level = 176 input->gang_global_priority_level; 177 mes_add_queue_pkt.doorbell_offset = input->doorbell_offset; 178 mes_add_queue_pkt.mqd_addr = input->mqd_addr; 179 180 if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >> 181 AMDGPU_MES_API_VERSION_SHIFT) >= 2) 182 mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr; 183 else 184 mes_add_queue_pkt.wptr_addr = input->wptr_addr; 185 186 mes_add_queue_pkt.queue_type = 187 convert_to_mes_queue_type(input->queue_type); 188 mes_add_queue_pkt.paging = input->paging; 189 mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl; 190 mes_add_queue_pkt.gws_base = input->gws_base; 191 mes_add_queue_pkt.gws_size = input->gws_size; 192 mes_add_queue_pkt.trap_handler_addr = input->tba_addr; 193 mes_add_queue_pkt.tma_addr = input->tma_addr; 194 mes_add_queue_pkt.is_kfd_process = input->is_kfd_process; 195 mes_add_queue_pkt.trap_en = 1; 196 197 /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */ 198 mes_add_queue_pkt.is_aql_queue = input->is_aql_queue; 199 mes_add_queue_pkt.gds_size = input->queue_size; 200 201 if (!(((adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 4) && 202 (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 0)) && 203 (adev->ip_versions[GC_HWIP][0] <= IP_VERSION(11, 0, 3)))) 204 mes_add_queue_pkt.trap_en = 1; 205 206 /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */ 207 mes_add_queue_pkt.is_aql_queue = input->is_aql_queue; 208 mes_add_queue_pkt.gds_size = input->queue_size; 209 210 return mes_v11_0_submit_pkt_and_poll_completion(mes, 211 &mes_add_queue_pkt, sizeof(mes_add_queue_pkt), 212 offsetof(union MESAPI__ADD_QUEUE, api_status)); 213 } 214 215 static int mes_v11_0_remove_hw_queue(struct amdgpu_mes *mes, 216 struct mes_remove_queue_input *input) 217 { 218 union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt; 219 220 memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt)); 221 222 mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER; 223 mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE; 224 mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS; 225 226 mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset; 227 mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr; 228 229 return mes_v11_0_submit_pkt_and_poll_completion(mes, 230 &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt), 231 offsetof(union MESAPI__REMOVE_QUEUE, api_status)); 232 } 233 234 static int mes_v11_0_unmap_legacy_queue(struct amdgpu_mes *mes, 235 struct mes_unmap_legacy_queue_input *input) 236 { 237 union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt; 238 239 memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt)); 240 241 mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER; 242 mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE; 243 mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS; 244 245 mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset; 246 mes_remove_queue_pkt.gang_context_addr = 0; 247 248 mes_remove_queue_pkt.pipe_id = input->pipe_id; 249 mes_remove_queue_pkt.queue_id = input->queue_id; 250 251 if (input->action == PREEMPT_QUEUES_NO_UNMAP) { 252 mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1; 253 mes_remove_queue_pkt.tf_addr = input->trail_fence_addr; 254 mes_remove_queue_pkt.tf_data = 255 lower_32_bits(input->trail_fence_data); 256 } else { 257 mes_remove_queue_pkt.unmap_legacy_queue = 1; 258 mes_remove_queue_pkt.queue_type = 259 convert_to_mes_queue_type(input->queue_type); 260 } 261 262 return mes_v11_0_submit_pkt_and_poll_completion(mes, 263 &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt), 264 offsetof(union MESAPI__REMOVE_QUEUE, api_status)); 265 } 266 267 static int mes_v11_0_suspend_gang(struct amdgpu_mes *mes, 268 struct mes_suspend_gang_input *input) 269 { 270 return 0; 271 } 272 273 static int mes_v11_0_resume_gang(struct amdgpu_mes *mes, 274 struct mes_resume_gang_input *input) 275 { 276 return 0; 277 } 278 279 static int mes_v11_0_query_sched_status(struct amdgpu_mes *mes) 280 { 281 union MESAPI__QUERY_MES_STATUS mes_status_pkt; 282 283 memset(&mes_status_pkt, 0, sizeof(mes_status_pkt)); 284 285 mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER; 286 mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS; 287 mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS; 288 289 return mes_v11_0_submit_pkt_and_poll_completion(mes, 290 &mes_status_pkt, sizeof(mes_status_pkt), 291 offsetof(union MESAPI__QUERY_MES_STATUS, api_status)); 292 } 293 294 static int mes_v11_0_misc_op(struct amdgpu_mes *mes, 295 struct mes_misc_op_input *input) 296 { 297 union MESAPI__MISC misc_pkt; 298 299 memset(&misc_pkt, 0, sizeof(misc_pkt)); 300 301 misc_pkt.header.type = MES_API_TYPE_SCHEDULER; 302 misc_pkt.header.opcode = MES_SCH_API_MISC; 303 misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS; 304 305 switch (input->op) { 306 case MES_MISC_OP_READ_REG: 307 misc_pkt.opcode = MESAPI_MISC__READ_REG; 308 misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset; 309 misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr; 310 break; 311 case MES_MISC_OP_WRITE_REG: 312 misc_pkt.opcode = MESAPI_MISC__WRITE_REG; 313 misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset; 314 misc_pkt.write_reg.reg_value = input->write_reg.reg_value; 315 break; 316 case MES_MISC_OP_WRM_REG_WAIT: 317 misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM; 318 misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM; 319 misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref; 320 misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask; 321 misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0; 322 misc_pkt.wait_reg_mem.reg_offset2 = 0; 323 break; 324 case MES_MISC_OP_WRM_REG_WR_WAIT: 325 misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM; 326 misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG; 327 misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref; 328 misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask; 329 misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0; 330 misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1; 331 break; 332 default: 333 DRM_ERROR("unsupported misc op (%d) \n", input->op); 334 return -EINVAL; 335 } 336 337 return mes_v11_0_submit_pkt_and_poll_completion(mes, 338 &misc_pkt, sizeof(misc_pkt), 339 offsetof(union MESAPI__MISC, api_status)); 340 } 341 342 static int mes_v11_0_set_hw_resources(struct amdgpu_mes *mes) 343 { 344 int i; 345 struct amdgpu_device *adev = mes->adev; 346 union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt; 347 348 memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt)); 349 350 mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER; 351 mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC; 352 mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS; 353 354 mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub; 355 mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub; 356 mes_set_hw_res_pkt.gds_size = adev->gds.gds_size; 357 mes_set_hw_res_pkt.paging_vmid = 0; 358 mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr = mes->sch_ctx_gpu_addr; 359 mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr = 360 mes->query_status_fence_gpu_addr; 361 362 for (i = 0; i < MAX_COMPUTE_PIPES; i++) 363 mes_set_hw_res_pkt.compute_hqd_mask[i] = 364 mes->compute_hqd_mask[i]; 365 366 for (i = 0; i < MAX_GFX_PIPES; i++) 367 mes_set_hw_res_pkt.gfx_hqd_mask[i] = mes->gfx_hqd_mask[i]; 368 369 for (i = 0; i < MAX_SDMA_PIPES; i++) 370 mes_set_hw_res_pkt.sdma_hqd_mask[i] = mes->sdma_hqd_mask[i]; 371 372 for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++) 373 mes_set_hw_res_pkt.aggregated_doorbells[i] = 374 mes->aggregated_doorbells[i]; 375 376 for (i = 0; i < 5; i++) { 377 mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i]; 378 mes_set_hw_res_pkt.mmhub_base[i] = 379 adev->reg_offset[MMHUB_HWIP][0][i]; 380 mes_set_hw_res_pkt.osssys_base[i] = 381 adev->reg_offset[OSSSYS_HWIP][0][i]; 382 } 383 384 mes_set_hw_res_pkt.disable_reset = 1; 385 mes_set_hw_res_pkt.disable_mes_log = 1; 386 mes_set_hw_res_pkt.use_different_vmid_compute = 1; 387 mes_set_hw_res_pkt.oversubscription_timer = 50; 388 389 return mes_v11_0_submit_pkt_and_poll_completion(mes, 390 &mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt), 391 offsetof(union MESAPI_SET_HW_RESOURCES, api_status)); 392 } 393 394 static void mes_v11_0_init_aggregated_doorbell(struct amdgpu_mes *mes) 395 { 396 struct amdgpu_device *adev = mes->adev; 397 uint32_t data; 398 399 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1); 400 data &= ~(CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET_MASK | 401 CP_MES_DOORBELL_CONTROL1__DOORBELL_EN_MASK | 402 CP_MES_DOORBELL_CONTROL1__DOORBELL_HIT_MASK); 403 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_LOW] << 404 CP_MES_DOORBELL_CONTROL1__DOORBELL_OFFSET__SHIFT; 405 data |= 1 << CP_MES_DOORBELL_CONTROL1__DOORBELL_EN__SHIFT; 406 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL1, data); 407 408 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2); 409 data &= ~(CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET_MASK | 410 CP_MES_DOORBELL_CONTROL2__DOORBELL_EN_MASK | 411 CP_MES_DOORBELL_CONTROL2__DOORBELL_HIT_MASK); 412 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_NORMAL] << 413 CP_MES_DOORBELL_CONTROL2__DOORBELL_OFFSET__SHIFT; 414 data |= 1 << CP_MES_DOORBELL_CONTROL2__DOORBELL_EN__SHIFT; 415 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL2, data); 416 417 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3); 418 data &= ~(CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET_MASK | 419 CP_MES_DOORBELL_CONTROL3__DOORBELL_EN_MASK | 420 CP_MES_DOORBELL_CONTROL3__DOORBELL_HIT_MASK); 421 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_MEDIUM] << 422 CP_MES_DOORBELL_CONTROL3__DOORBELL_OFFSET__SHIFT; 423 data |= 1 << CP_MES_DOORBELL_CONTROL3__DOORBELL_EN__SHIFT; 424 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL3, data); 425 426 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4); 427 data &= ~(CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET_MASK | 428 CP_MES_DOORBELL_CONTROL4__DOORBELL_EN_MASK | 429 CP_MES_DOORBELL_CONTROL4__DOORBELL_HIT_MASK); 430 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_HIGH] << 431 CP_MES_DOORBELL_CONTROL4__DOORBELL_OFFSET__SHIFT; 432 data |= 1 << CP_MES_DOORBELL_CONTROL4__DOORBELL_EN__SHIFT; 433 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL4, data); 434 435 data = RREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5); 436 data &= ~(CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET_MASK | 437 CP_MES_DOORBELL_CONTROL5__DOORBELL_EN_MASK | 438 CP_MES_DOORBELL_CONTROL5__DOORBELL_HIT_MASK); 439 data |= mes->aggregated_doorbells[AMDGPU_MES_PRIORITY_LEVEL_REALTIME] << 440 CP_MES_DOORBELL_CONTROL5__DOORBELL_OFFSET__SHIFT; 441 data |= 1 << CP_MES_DOORBELL_CONTROL5__DOORBELL_EN__SHIFT; 442 WREG32_SOC15(GC, 0, regCP_MES_DOORBELL_CONTROL5, data); 443 444 data = 1 << CP_HQD_GFX_CONTROL__DB_UPDATED_MSG_EN__SHIFT; 445 WREG32_SOC15(GC, 0, regCP_HQD_GFX_CONTROL, data); 446 } 447 448 static const struct amdgpu_mes_funcs mes_v11_0_funcs = { 449 .add_hw_queue = mes_v11_0_add_hw_queue, 450 .remove_hw_queue = mes_v11_0_remove_hw_queue, 451 .unmap_legacy_queue = mes_v11_0_unmap_legacy_queue, 452 .suspend_gang = mes_v11_0_suspend_gang, 453 .resume_gang = mes_v11_0_resume_gang, 454 .misc_op = mes_v11_0_misc_op, 455 }; 456 457 static int mes_v11_0_init_microcode(struct amdgpu_device *adev, 458 enum admgpu_mes_pipe pipe) 459 { 460 char fw_name[30]; 461 char ucode_prefix[30]; 462 int err; 463 const struct mes_firmware_header_v1_0 *mes_hdr; 464 struct amdgpu_firmware_info *info; 465 466 amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix, sizeof(ucode_prefix)); 467 468 if (pipe == AMDGPU_MES_SCHED_PIPE) 469 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes.bin", 470 ucode_prefix); 471 else 472 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mes1.bin", 473 ucode_prefix); 474 475 err = request_firmware(&adev->mes.fw[pipe], fw_name, adev->dev); 476 if (err) 477 return err; 478 479 err = amdgpu_ucode_validate(adev->mes.fw[pipe]); 480 if (err) { 481 release_firmware(adev->mes.fw[pipe]); 482 adev->mes.fw[pipe] = NULL; 483 return err; 484 } 485 486 mes_hdr = (const struct mes_firmware_header_v1_0 *) 487 adev->mes.fw[pipe]->data; 488 adev->mes.uc_start_addr[pipe] = 489 le32_to_cpu(mes_hdr->mes_uc_start_addr_lo) | 490 ((uint64_t)(le32_to_cpu(mes_hdr->mes_uc_start_addr_hi)) << 32); 491 adev->mes.data_start_addr[pipe] = 492 le32_to_cpu(mes_hdr->mes_data_start_addr_lo) | 493 ((uint64_t)(le32_to_cpu(mes_hdr->mes_data_start_addr_hi)) << 32); 494 495 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) { 496 int ucode, ucode_data; 497 498 if (pipe == AMDGPU_MES_SCHED_PIPE) { 499 ucode = AMDGPU_UCODE_ID_CP_MES; 500 ucode_data = AMDGPU_UCODE_ID_CP_MES_DATA; 501 } else { 502 ucode = AMDGPU_UCODE_ID_CP_MES1; 503 ucode_data = AMDGPU_UCODE_ID_CP_MES1_DATA; 504 } 505 506 info = &adev->firmware.ucode[ucode]; 507 info->ucode_id = ucode; 508 info->fw = adev->mes.fw[pipe]; 509 adev->firmware.fw_size += 510 ALIGN(le32_to_cpu(mes_hdr->mes_ucode_size_bytes), 511 PAGE_SIZE); 512 513 info = &adev->firmware.ucode[ucode_data]; 514 info->ucode_id = ucode_data; 515 info->fw = adev->mes.fw[pipe]; 516 adev->firmware.fw_size += 517 ALIGN(le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes), 518 PAGE_SIZE); 519 } 520 521 return 0; 522 } 523 524 static void mes_v11_0_free_microcode(struct amdgpu_device *adev, 525 enum admgpu_mes_pipe pipe) 526 { 527 release_firmware(adev->mes.fw[pipe]); 528 adev->mes.fw[pipe] = NULL; 529 } 530 531 static int mes_v11_0_allocate_ucode_buffer(struct amdgpu_device *adev, 532 enum admgpu_mes_pipe pipe) 533 { 534 int r; 535 const struct mes_firmware_header_v1_0 *mes_hdr; 536 const __le32 *fw_data; 537 unsigned fw_size; 538 539 mes_hdr = (const struct mes_firmware_header_v1_0 *) 540 adev->mes.fw[pipe]->data; 541 542 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data + 543 le32_to_cpu(mes_hdr->mes_ucode_offset_bytes)); 544 fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes); 545 546 r = amdgpu_bo_create_reserved(adev, fw_size, 547 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM, 548 &adev->mes.ucode_fw_obj[pipe], 549 &adev->mes.ucode_fw_gpu_addr[pipe], 550 (void **)&adev->mes.ucode_fw_ptr[pipe]); 551 if (r) { 552 dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r); 553 return r; 554 } 555 556 memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size); 557 558 amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]); 559 amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]); 560 561 return 0; 562 } 563 564 static int mes_v11_0_allocate_ucode_data_buffer(struct amdgpu_device *adev, 565 enum admgpu_mes_pipe pipe) 566 { 567 int r; 568 const struct mes_firmware_header_v1_0 *mes_hdr; 569 const __le32 *fw_data; 570 unsigned fw_size; 571 572 mes_hdr = (const struct mes_firmware_header_v1_0 *) 573 adev->mes.fw[pipe]->data; 574 575 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data + 576 le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes)); 577 fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes); 578 579 r = amdgpu_bo_create_reserved(adev, fw_size, 580 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 581 &adev->mes.data_fw_obj[pipe], 582 &adev->mes.data_fw_gpu_addr[pipe], 583 (void **)&adev->mes.data_fw_ptr[pipe]); 584 if (r) { 585 dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r); 586 return r; 587 } 588 589 memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size); 590 591 amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]); 592 amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]); 593 594 return 0; 595 } 596 597 static void mes_v11_0_free_ucode_buffers(struct amdgpu_device *adev, 598 enum admgpu_mes_pipe pipe) 599 { 600 amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe], 601 &adev->mes.data_fw_gpu_addr[pipe], 602 (void **)&adev->mes.data_fw_ptr[pipe]); 603 604 amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe], 605 &adev->mes.ucode_fw_gpu_addr[pipe], 606 (void **)&adev->mes.ucode_fw_ptr[pipe]); 607 } 608 609 static void mes_v11_0_enable(struct amdgpu_device *adev, bool enable) 610 { 611 uint64_t ucode_addr; 612 uint32_t pipe, data = 0; 613 614 if (enable) { 615 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL); 616 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1); 617 data = REG_SET_FIELD(data, CP_MES_CNTL, 618 MES_PIPE1_RESET, adev->enable_mes_kiq ? 1 : 0); 619 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data); 620 621 mutex_lock(&adev->srbm_mutex); 622 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) { 623 if (!adev->enable_mes_kiq && 624 pipe == AMDGPU_MES_KIQ_PIPE) 625 continue; 626 627 soc21_grbm_select(adev, 3, pipe, 0, 0); 628 629 ucode_addr = adev->mes.uc_start_addr[pipe] >> 2; 630 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START, 631 lower_32_bits(ucode_addr)); 632 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI, 633 upper_32_bits(ucode_addr)); 634 } 635 soc21_grbm_select(adev, 0, 0, 0, 0); 636 mutex_unlock(&adev->srbm_mutex); 637 638 /* unhalt MES and activate pipe0 */ 639 data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1); 640 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 641 adev->enable_mes_kiq ? 1 : 0); 642 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data); 643 644 if (amdgpu_emu_mode) 645 msleep(100); 646 else 647 udelay(50); 648 } else { 649 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL); 650 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0); 651 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0); 652 data = REG_SET_FIELD(data, CP_MES_CNTL, 653 MES_INVALIDATE_ICACHE, 1); 654 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1); 655 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET, 656 adev->enable_mes_kiq ? 1 : 0); 657 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1); 658 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data); 659 } 660 } 661 662 /* This function is for backdoor MES firmware */ 663 static int mes_v11_0_load_microcode(struct amdgpu_device *adev, 664 enum admgpu_mes_pipe pipe, bool prime_icache) 665 { 666 int r; 667 uint32_t data; 668 uint64_t ucode_addr; 669 670 mes_v11_0_enable(adev, false); 671 672 if (!adev->mes.fw[pipe]) 673 return -EINVAL; 674 675 r = mes_v11_0_allocate_ucode_buffer(adev, pipe); 676 if (r) 677 return r; 678 679 r = mes_v11_0_allocate_ucode_data_buffer(adev, pipe); 680 if (r) { 681 mes_v11_0_free_ucode_buffers(adev, pipe); 682 return r; 683 } 684 685 mutex_lock(&adev->srbm_mutex); 686 /* me=3, pipe=0, queue=0 */ 687 soc21_grbm_select(adev, 3, pipe, 0, 0); 688 689 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0); 690 691 /* set ucode start address */ 692 ucode_addr = adev->mes.uc_start_addr[pipe] >> 2; 693 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START, 694 lower_32_bits(ucode_addr)); 695 WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI, 696 upper_32_bits(ucode_addr)); 697 698 /* set ucode fimrware address */ 699 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO, 700 lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe])); 701 WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI, 702 upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe])); 703 704 /* set ucode instruction cache boundary to 2M-1 */ 705 WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF); 706 707 /* set ucode data firmware address */ 708 WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO, 709 lower_32_bits(adev->mes.data_fw_gpu_addr[pipe])); 710 WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI, 711 upper_32_bits(adev->mes.data_fw_gpu_addr[pipe])); 712 713 /* Set 0x3FFFF (256K-1) to CP_MES_MDBOUND_LO */ 714 WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x3FFFF); 715 716 if (prime_icache) { 717 /* invalidate ICACHE */ 718 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL); 719 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0); 720 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1); 721 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data); 722 723 /* prime the ICACHE. */ 724 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL); 725 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1); 726 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data); 727 } 728 729 soc21_grbm_select(adev, 0, 0, 0, 0); 730 mutex_unlock(&adev->srbm_mutex); 731 732 return 0; 733 } 734 735 static int mes_v11_0_allocate_eop_buf(struct amdgpu_device *adev, 736 enum admgpu_mes_pipe pipe) 737 { 738 int r; 739 u32 *eop; 740 741 r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE, 742 AMDGPU_GEM_DOMAIN_GTT, 743 &adev->mes.eop_gpu_obj[pipe], 744 &adev->mes.eop_gpu_addr[pipe], 745 (void **)&eop); 746 if (r) { 747 dev_warn(adev->dev, "(%d) create EOP bo failed\n", r); 748 return r; 749 } 750 751 memset(eop, 0, 752 adev->mes.eop_gpu_obj[pipe]->tbo.base.size); 753 754 amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]); 755 amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]); 756 757 return 0; 758 } 759 760 static int mes_v11_0_mqd_init(struct amdgpu_ring *ring) 761 { 762 struct v11_compute_mqd *mqd = ring->mqd_ptr; 763 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr; 764 uint32_t tmp; 765 766 mqd->header = 0xC0310800; 767 mqd->compute_pipelinestat_enable = 0x00000001; 768 mqd->compute_static_thread_mgmt_se0 = 0xffffffff; 769 mqd->compute_static_thread_mgmt_se1 = 0xffffffff; 770 mqd->compute_static_thread_mgmt_se2 = 0xffffffff; 771 mqd->compute_static_thread_mgmt_se3 = 0xffffffff; 772 mqd->compute_misc_reserved = 0x00000007; 773 774 eop_base_addr = ring->eop_gpu_addr >> 8; 775 776 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */ 777 tmp = regCP_HQD_EOP_CONTROL_DEFAULT; 778 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE, 779 (order_base_2(MES_EOP_SIZE / 4) - 1)); 780 781 mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr); 782 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr); 783 mqd->cp_hqd_eop_control = tmp; 784 785 /* disable the queue if it's active */ 786 ring->wptr = 0; 787 mqd->cp_hqd_pq_rptr = 0; 788 mqd->cp_hqd_pq_wptr_lo = 0; 789 mqd->cp_hqd_pq_wptr_hi = 0; 790 791 /* set the pointer to the MQD */ 792 mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc; 793 mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr); 794 795 /* set MQD vmid to 0 */ 796 tmp = regCP_MQD_CONTROL_DEFAULT; 797 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0); 798 mqd->cp_mqd_control = tmp; 799 800 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */ 801 hqd_gpu_addr = ring->gpu_addr >> 8; 802 mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr); 803 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr); 804 805 /* set the wb address whether it's enabled or not */ 806 wb_gpu_addr = ring->rptr_gpu_addr; 807 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc; 808 mqd->cp_hqd_pq_rptr_report_addr_hi = 809 upper_32_bits(wb_gpu_addr) & 0xffff; 810 811 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */ 812 wb_gpu_addr = ring->wptr_gpu_addr; 813 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8; 814 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff; 815 816 /* set up the HQD, this is similar to CP_RB0_CNTL */ 817 tmp = regCP_HQD_PQ_CONTROL_DEFAULT; 818 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE, 819 (order_base_2(ring->ring_size / 4) - 1)); 820 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE, 821 ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8)); 822 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1); 823 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0); 824 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1); 825 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1); 826 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1); 827 mqd->cp_hqd_pq_control = tmp; 828 829 /* enable doorbell */ 830 tmp = 0; 831 if (ring->use_doorbell) { 832 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 833 DOORBELL_OFFSET, ring->doorbell_index); 834 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 835 DOORBELL_EN, 1); 836 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 837 DOORBELL_SOURCE, 0); 838 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 839 DOORBELL_HIT, 0); 840 } 841 else 842 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 843 DOORBELL_EN, 0); 844 mqd->cp_hqd_pq_doorbell_control = tmp; 845 846 mqd->cp_hqd_vmid = 0; 847 /* activate the queue */ 848 mqd->cp_hqd_active = 1; 849 850 tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT; 851 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE, 852 PRELOAD_SIZE, 0x55); 853 mqd->cp_hqd_persistent_state = tmp; 854 855 mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT; 856 mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT; 857 mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT; 858 859 return 0; 860 } 861 862 static void mes_v11_0_queue_init_register(struct amdgpu_ring *ring) 863 { 864 struct v11_compute_mqd *mqd = ring->mqd_ptr; 865 struct amdgpu_device *adev = ring->adev; 866 uint32_t data = 0; 867 868 mutex_lock(&adev->srbm_mutex); 869 soc21_grbm_select(adev, 3, ring->pipe, 0, 0); 870 871 /* set CP_HQD_VMID.VMID = 0. */ 872 data = RREG32_SOC15(GC, 0, regCP_HQD_VMID); 873 data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0); 874 WREG32_SOC15(GC, 0, regCP_HQD_VMID, data); 875 876 /* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */ 877 data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL); 878 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL, 879 DOORBELL_EN, 0); 880 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data); 881 882 /* set CP_MQD_BASE_ADDR/HI with the MQD base address */ 883 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo); 884 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi); 885 886 /* set CP_MQD_CONTROL.VMID=0 */ 887 data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL); 888 data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0); 889 WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0); 890 891 /* set CP_HQD_PQ_BASE/HI with the ring buffer base address */ 892 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo); 893 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi); 894 895 /* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */ 896 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR, 897 mqd->cp_hqd_pq_rptr_report_addr_lo); 898 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI, 899 mqd->cp_hqd_pq_rptr_report_addr_hi); 900 901 /* set CP_HQD_PQ_CONTROL */ 902 WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control); 903 904 /* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */ 905 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR, 906 mqd->cp_hqd_pq_wptr_poll_addr_lo); 907 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI, 908 mqd->cp_hqd_pq_wptr_poll_addr_hi); 909 910 /* set CP_HQD_PQ_DOORBELL_CONTROL */ 911 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 912 mqd->cp_hqd_pq_doorbell_control); 913 914 /* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */ 915 WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state); 916 917 /* set CP_HQD_ACTIVE.ACTIVE=1 */ 918 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active); 919 920 soc21_grbm_select(adev, 0, 0, 0, 0); 921 mutex_unlock(&adev->srbm_mutex); 922 } 923 924 static int mes_v11_0_kiq_enable_queue(struct amdgpu_device *adev) 925 { 926 struct amdgpu_kiq *kiq = &adev->gfx.kiq; 927 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring; 928 int r; 929 930 if (!kiq->pmf || !kiq->pmf->kiq_map_queues) 931 return -EINVAL; 932 933 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size); 934 if (r) { 935 DRM_ERROR("Failed to lock KIQ (%d).\n", r); 936 return r; 937 } 938 939 kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring); 940 941 r = amdgpu_ring_test_ring(kiq_ring); 942 if (r) { 943 DRM_ERROR("kfq enable failed\n"); 944 kiq_ring->sched.ready = false; 945 } 946 return r; 947 } 948 949 static int mes_v11_0_queue_init(struct amdgpu_device *adev, 950 enum admgpu_mes_pipe pipe) 951 { 952 struct amdgpu_ring *ring; 953 int r; 954 955 if (pipe == AMDGPU_MES_KIQ_PIPE) 956 ring = &adev->gfx.kiq.ring; 957 else if (pipe == AMDGPU_MES_SCHED_PIPE) 958 ring = &adev->mes.ring; 959 else 960 BUG(); 961 962 if ((pipe == AMDGPU_MES_SCHED_PIPE) && 963 (amdgpu_in_reset(adev) || adev->in_suspend)) { 964 *(ring->wptr_cpu_addr) = 0; 965 *(ring->rptr_cpu_addr) = 0; 966 amdgpu_ring_clear_ring(ring); 967 } 968 969 r = mes_v11_0_mqd_init(ring); 970 if (r) 971 return r; 972 973 if (pipe == AMDGPU_MES_SCHED_PIPE) { 974 r = mes_v11_0_kiq_enable_queue(adev); 975 if (r) 976 return r; 977 } else { 978 mes_v11_0_queue_init_register(ring); 979 } 980 981 /* get MES scheduler/KIQ versions */ 982 mutex_lock(&adev->srbm_mutex); 983 soc21_grbm_select(adev, 3, pipe, 0, 0); 984 985 if (pipe == AMDGPU_MES_SCHED_PIPE) 986 adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO); 987 else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq) 988 adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO); 989 990 soc21_grbm_select(adev, 0, 0, 0, 0); 991 mutex_unlock(&adev->srbm_mutex); 992 993 return 0; 994 } 995 996 static int mes_v11_0_ring_init(struct amdgpu_device *adev) 997 { 998 struct amdgpu_ring *ring; 999 1000 ring = &adev->mes.ring; 1001 1002 ring->funcs = &mes_v11_0_ring_funcs; 1003 1004 ring->me = 3; 1005 ring->pipe = 0; 1006 ring->queue = 0; 1007 1008 ring->ring_obj = NULL; 1009 ring->use_doorbell = true; 1010 ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1; 1011 ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_SCHED_PIPE]; 1012 ring->no_scheduler = true; 1013 sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue); 1014 1015 return amdgpu_ring_init(adev, ring, 1024, NULL, 0, 1016 AMDGPU_RING_PRIO_DEFAULT, NULL); 1017 } 1018 1019 static int mes_v11_0_kiq_ring_init(struct amdgpu_device *adev) 1020 { 1021 struct amdgpu_ring *ring; 1022 1023 spin_lock_init(&adev->gfx.kiq.ring_lock); 1024 1025 ring = &adev->gfx.kiq.ring; 1026 1027 ring->me = 3; 1028 ring->pipe = 1; 1029 ring->queue = 0; 1030 1031 ring->adev = NULL; 1032 ring->ring_obj = NULL; 1033 ring->use_doorbell = true; 1034 ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1; 1035 ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE]; 1036 ring->no_scheduler = true; 1037 sprintf(ring->name, "mes_kiq_%d.%d.%d", 1038 ring->me, ring->pipe, ring->queue); 1039 1040 return amdgpu_ring_init(adev, ring, 1024, NULL, 0, 1041 AMDGPU_RING_PRIO_DEFAULT, NULL); 1042 } 1043 1044 static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev, 1045 enum admgpu_mes_pipe pipe) 1046 { 1047 int r, mqd_size = sizeof(struct v11_compute_mqd); 1048 struct amdgpu_ring *ring; 1049 1050 if (pipe == AMDGPU_MES_KIQ_PIPE) 1051 ring = &adev->gfx.kiq.ring; 1052 else if (pipe == AMDGPU_MES_SCHED_PIPE) 1053 ring = &adev->mes.ring; 1054 else 1055 BUG(); 1056 1057 if (ring->mqd_obj) 1058 return 0; 1059 1060 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE, 1061 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj, 1062 &ring->mqd_gpu_addr, &ring->mqd_ptr); 1063 if (r) { 1064 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r); 1065 return r; 1066 } 1067 1068 memset(ring->mqd_ptr, 0, mqd_size); 1069 1070 /* prepare MQD backup */ 1071 adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL); 1072 if (!adev->mes.mqd_backup[pipe]) 1073 dev_warn(adev->dev, 1074 "no memory to create MQD backup for ring %s\n", 1075 ring->name); 1076 1077 return 0; 1078 } 1079 1080 static int mes_v11_0_sw_init(void *handle) 1081 { 1082 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1083 int pipe, r; 1084 1085 adev->mes.adev = adev; 1086 adev->mes.funcs = &mes_v11_0_funcs; 1087 adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init; 1088 adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini; 1089 1090 r = amdgpu_mes_init(adev); 1091 if (r) 1092 return r; 1093 1094 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) { 1095 if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE) 1096 continue; 1097 1098 r = mes_v11_0_init_microcode(adev, pipe); 1099 if (r) 1100 return r; 1101 1102 r = mes_v11_0_allocate_eop_buf(adev, pipe); 1103 if (r) 1104 return r; 1105 1106 r = mes_v11_0_mqd_sw_init(adev, pipe); 1107 if (r) 1108 return r; 1109 } 1110 1111 if (adev->enable_mes_kiq) { 1112 r = mes_v11_0_kiq_ring_init(adev); 1113 if (r) 1114 return r; 1115 } 1116 1117 r = mes_v11_0_ring_init(adev); 1118 if (r) 1119 return r; 1120 1121 return 0; 1122 } 1123 1124 static int mes_v11_0_sw_fini(void *handle) 1125 { 1126 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1127 int pipe; 1128 1129 amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs); 1130 amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs); 1131 1132 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) { 1133 kfree(adev->mes.mqd_backup[pipe]); 1134 1135 amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe], 1136 &adev->mes.eop_gpu_addr[pipe], 1137 NULL); 1138 1139 mes_v11_0_free_microcode(adev, pipe); 1140 } 1141 1142 amdgpu_bo_free_kernel(&adev->gfx.kiq.ring.mqd_obj, 1143 &adev->gfx.kiq.ring.mqd_gpu_addr, 1144 &adev->gfx.kiq.ring.mqd_ptr); 1145 1146 amdgpu_bo_free_kernel(&adev->mes.ring.mqd_obj, 1147 &adev->mes.ring.mqd_gpu_addr, 1148 &adev->mes.ring.mqd_ptr); 1149 1150 amdgpu_ring_fini(&adev->gfx.kiq.ring); 1151 amdgpu_ring_fini(&adev->mes.ring); 1152 1153 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 1154 mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE); 1155 mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE); 1156 } 1157 1158 amdgpu_mes_fini(adev); 1159 return 0; 1160 } 1161 1162 static void mes_v11_0_kiq_dequeue_sched(struct amdgpu_device *adev) 1163 { 1164 uint32_t data; 1165 int i; 1166 1167 mutex_lock(&adev->srbm_mutex); 1168 soc21_grbm_select(adev, 3, AMDGPU_MES_SCHED_PIPE, 0, 0); 1169 1170 /* disable the queue if it's active */ 1171 if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) { 1172 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1); 1173 for (i = 0; i < adev->usec_timeout; i++) { 1174 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1)) 1175 break; 1176 udelay(1); 1177 } 1178 } 1179 data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL); 1180 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL, 1181 DOORBELL_EN, 0); 1182 data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL, 1183 DOORBELL_HIT, 1); 1184 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data); 1185 1186 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0); 1187 1188 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0); 1189 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0); 1190 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0); 1191 1192 soc21_grbm_select(adev, 0, 0, 0, 0); 1193 mutex_unlock(&adev->srbm_mutex); 1194 1195 adev->mes.ring.sched.ready = false; 1196 } 1197 1198 static void mes_v11_0_kiq_setting(struct amdgpu_ring *ring) 1199 { 1200 uint32_t tmp; 1201 struct amdgpu_device *adev = ring->adev; 1202 1203 /* tell RLC which is KIQ queue */ 1204 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS); 1205 tmp &= 0xffffff00; 1206 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue); 1207 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp); 1208 tmp |= 0x80; 1209 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp); 1210 } 1211 1212 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev) 1213 { 1214 int r = 0; 1215 1216 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 1217 1218 r = mes_v11_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE, false); 1219 if (r) { 1220 DRM_ERROR("failed to load MES fw, r=%d\n", r); 1221 return r; 1222 } 1223 1224 r = mes_v11_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE, true); 1225 if (r) { 1226 DRM_ERROR("failed to load MES kiq fw, r=%d\n", r); 1227 return r; 1228 } 1229 1230 } 1231 1232 mes_v11_0_enable(adev, true); 1233 1234 mes_v11_0_kiq_setting(&adev->gfx.kiq.ring); 1235 1236 r = mes_v11_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE); 1237 if (r) 1238 goto failure; 1239 1240 return r; 1241 1242 failure: 1243 mes_v11_0_hw_fini(adev); 1244 return r; 1245 } 1246 1247 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev) 1248 { 1249 if (adev->mes.ring.sched.ready) 1250 mes_v11_0_kiq_dequeue_sched(adev); 1251 1252 mes_v11_0_enable(adev, false); 1253 return 0; 1254 } 1255 1256 static int mes_v11_0_hw_init(void *handle) 1257 { 1258 int r; 1259 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1260 1261 if (!adev->enable_mes_kiq) { 1262 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 1263 r = mes_v11_0_load_microcode(adev, 1264 AMDGPU_MES_SCHED_PIPE, true); 1265 if (r) { 1266 DRM_ERROR("failed to MES fw, r=%d\n", r); 1267 return r; 1268 } 1269 } 1270 1271 mes_v11_0_enable(adev, true); 1272 } 1273 1274 r = mes_v11_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE); 1275 if (r) 1276 goto failure; 1277 1278 r = mes_v11_0_set_hw_resources(&adev->mes); 1279 if (r) 1280 goto failure; 1281 1282 mes_v11_0_init_aggregated_doorbell(&adev->mes); 1283 1284 r = mes_v11_0_query_sched_status(&adev->mes); 1285 if (r) { 1286 DRM_ERROR("MES is busy\n"); 1287 goto failure; 1288 } 1289 1290 /* 1291 * Disable KIQ ring usage from the driver once MES is enabled. 1292 * MES uses KIQ ring exclusively so driver cannot access KIQ ring 1293 * with MES enabled. 1294 */ 1295 adev->gfx.kiq.ring.sched.ready = false; 1296 adev->mes.ring.sched.ready = true; 1297 1298 return 0; 1299 1300 failure: 1301 mes_v11_0_hw_fini(adev); 1302 return r; 1303 } 1304 1305 static int mes_v11_0_hw_fini(void *handle) 1306 { 1307 return 0; 1308 } 1309 1310 static int mes_v11_0_suspend(void *handle) 1311 { 1312 int r; 1313 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1314 1315 r = amdgpu_mes_suspend(adev); 1316 if (r) 1317 return r; 1318 1319 return mes_v11_0_hw_fini(adev); 1320 } 1321 1322 static int mes_v11_0_resume(void *handle) 1323 { 1324 int r; 1325 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1326 1327 r = mes_v11_0_hw_init(adev); 1328 if (r) 1329 return r; 1330 1331 return amdgpu_mes_resume(adev); 1332 } 1333 1334 static int mes_v11_0_late_init(void *handle) 1335 { 1336 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1337 1338 if (!amdgpu_in_reset(adev) && 1339 (adev->ip_versions[GC_HWIP][0] != IP_VERSION(11, 0, 3))) 1340 amdgpu_mes_self_test(adev); 1341 1342 return 0; 1343 } 1344 1345 static const struct amd_ip_funcs mes_v11_0_ip_funcs = { 1346 .name = "mes_v11_0", 1347 .late_init = mes_v11_0_late_init, 1348 .sw_init = mes_v11_0_sw_init, 1349 .sw_fini = mes_v11_0_sw_fini, 1350 .hw_init = mes_v11_0_hw_init, 1351 .hw_fini = mes_v11_0_hw_fini, 1352 .suspend = mes_v11_0_suspend, 1353 .resume = mes_v11_0_resume, 1354 }; 1355 1356 const struct amdgpu_ip_block_version mes_v11_0_ip_block = { 1357 .type = AMD_IP_BLOCK_TYPE_MES, 1358 .major = 11, 1359 .minor = 0, 1360 .rev = 0, 1361 .funcs = &mes_v11_0_ip_funcs, 1362 }; 1363