drivers/gpu/drm/amd/amdgpu/mes_v11_0.c

   1 /*
   2  * Copyright 2019 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23
  24 #include <linux/firmware.h>
  25 #include <linux/module.h>
  26 #include "amdgpu.h"
  27 #include "soc15_common.h"
  28 #include "soc21.h"
  29 #include "gc/gc_11_0_0_offset.h"
  30 #include "gc/gc_11_0_0_sh_mask.h"
  31 #include "gc/gc_11_0_0_default.h"
  32 #include "v11_structs.h"
  33 #include "mes_v11_api_def.h"
  34
  35 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes.bin");
  36 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes_2.bin");
  37 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mes1.bin");
  38 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes.bin");
  39 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes_2.bin");
  40 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mes1.bin");
  41 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes.bin");
  42 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes_2.bin");
  43 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mes1.bin");
  44 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes.bin");
  45 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes_2.bin");
  46 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes1.bin");
  47 MODULE_FIRMWARE("amdgpu/gc_11_0_4_mes.bin");
  48 MODULE_FIRMWARE("amdgpu/gc_11_0_4_mes_2.bin");
  49 MODULE_FIRMWARE("amdgpu/gc_11_0_4_mes1.bin");
  50 MODULE_FIRMWARE("amdgpu/gc_11_5_0_mes_2.bin");
  51 MODULE_FIRMWARE("amdgpu/gc_11_5_0_mes1.bin");
  52 MODULE_FIRMWARE("amdgpu/gc_11_5_1_mes_2.bin");
  53 MODULE_FIRMWARE("amdgpu/gc_11_5_1_mes1.bin");
  54
  55
  56 static int mes_v11_0_hw_fini(void *handle);
  57 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev);
  58 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev);
  59
  60 #define MES_EOP_SIZE   2048
  61 #define GFX_MES_DRAM_SIZE       0x80000
  62
  63 static void mes_v11_0_ring_set_wptr(struct amdgpu_ring *ring)
  64 {
  65         struct amdgpu_device *adev = ring->adev;
  66
  67         if (ring->use_doorbell) {
  68                 atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
  69                              ring->wptr);
  70                 WDOORBELL64(ring->doorbell_index, ring->wptr);
  71         } else {
  72                 BUG();
  73         }
  74 }
  75
  76 static u64 mes_v11_0_ring_get_rptr(struct amdgpu_ring *ring)
  77 {
  78         return *ring->rptr_cpu_addr;
  79 }
  80
  81 static u64 mes_v11_0_ring_get_wptr(struct amdgpu_ring *ring)
  82 {
  83         u64 wptr;
  84
  85         if (ring->use_doorbell)
  86                 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
  87         else
  88                 BUG();
  89         return wptr;
  90 }
  91
  92 static const struct amdgpu_ring_funcs mes_v11_0_ring_funcs = {
  93         .type = AMDGPU_RING_TYPE_MES,
  94         .align_mask = 1,
  95         .nop = 0,
  96         .support_64bit_ptrs = true,
  97         .get_rptr = mes_v11_0_ring_get_rptr,
  98         .get_wptr = mes_v11_0_ring_get_wptr,
  99         .set_wptr = mes_v11_0_ring_set_wptr,
 100         .insert_nop = amdgpu_ring_insert_nop,
 101 };
 102
 103 static int mes_v11_0_submit_pkt_and_poll_completion(struct amdgpu_mes *mes,
 104                                                     void *pkt, int size,
 105                                                     int api_status_off)
 106 {
 107         int ndw = size / 4;
 108         signed long r;
 109         union MESAPI__ADD_QUEUE *x_pkt = pkt;
 110         struct MES_API_STATUS *api_status;
 111         struct amdgpu_device *adev = mes->adev;
 112         struct amdgpu_ring *ring = &mes->ring;
 113         unsigned long flags;
 114         signed long timeout = adev->usec_timeout;
 115
 116         if (amdgpu_emu_mode) {
 117                 timeout *= 100;
 118         } else if (amdgpu_sriov_vf(adev)) {
 119                 /* Worst case in sriov where all other 15 VF timeout, each VF needs about 600ms */
 120                 timeout = 15 * 600 * 1000;
 121         }
 122         BUG_ON(size % 4 != 0);
 123
 124         spin_lock_irqsave(&mes->ring_lock, flags);
 125         if (amdgpu_ring_alloc(ring, ndw)) {
 126                 spin_unlock_irqrestore(&mes->ring_lock, flags);
 127                 return -ENOMEM;
 128         }
 129
 130         api_status = (struct MES_API_STATUS *)((char *)pkt + api_status_off);
 131         api_status->api_completion_fence_addr = mes->ring.fence_drv.gpu_addr;
 132         api_status->api_completion_fence_value = ++mes->ring.fence_drv.sync_seq;
 133
 134         amdgpu_ring_write_multiple(ring, pkt, ndw);
 135         amdgpu_ring_commit(ring);
 136         spin_unlock_irqrestore(&mes->ring_lock, flags);
 137
 138         DRM_DEBUG("MES msg=%d was emitted\n", x_pkt->header.opcode);
 139
 140         r = amdgpu_fence_wait_polling(ring, ring->fence_drv.sync_seq,
 141                       timeout);
 142         if (r < 1) {
 143                 DRM_ERROR("MES failed to response msg=%d\n",
 144                           x_pkt->header.opcode);
 145
 146                 while (halt_if_hws_hang)
 147                         schedule();
 148
 149                 return -ETIMEDOUT;
 150         }
 151
 152         return 0;
 153 }
 154
 155 static int convert_to_mes_queue_type(int queue_type)
 156 {
 157         if (queue_type == AMDGPU_RING_TYPE_GFX)
 158                 return MES_QUEUE_TYPE_GFX;
 159         else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
 160                 return MES_QUEUE_TYPE_COMPUTE;
 161         else if (queue_type == AMDGPU_RING_TYPE_SDMA)
 162                 return MES_QUEUE_TYPE_SDMA;
 163         else
 164                 BUG();
 165         return -1;
 166 }
 167
 168 static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
 169                                   struct mes_add_queue_input *input)
 170 {
 171         struct amdgpu_device *adev = mes->adev;
 172         union MESAPI__ADD_QUEUE mes_add_queue_pkt;
 173         struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_GFXHUB(0)];
 174         uint32_t vm_cntx_cntl = hub->vm_cntx_cntl;
 175
 176         memset(&mes_add_queue_pkt, 0, sizeof(mes_add_queue_pkt));
 177
 178         mes_add_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
 179         mes_add_queue_pkt.header.opcode = MES_SCH_API_ADD_QUEUE;
 180         mes_add_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
 181
 182         mes_add_queue_pkt.process_id = input->process_id;
 183         mes_add_queue_pkt.page_table_base_addr = input->page_table_base_addr;
 184         mes_add_queue_pkt.process_va_start = input->process_va_start;
 185         mes_add_queue_pkt.process_va_end = input->process_va_end;
 186         mes_add_queue_pkt.process_quantum = input->process_quantum;
 187         mes_add_queue_pkt.process_context_addr = input->process_context_addr;
 188         mes_add_queue_pkt.gang_quantum = input->gang_quantum;
 189         mes_add_queue_pkt.gang_context_addr = input->gang_context_addr;
 190         mes_add_queue_pkt.inprocess_gang_priority =
 191                 input->inprocess_gang_priority;
 192         mes_add_queue_pkt.gang_global_priority_level =
 193                 input->gang_global_priority_level;
 194         mes_add_queue_pkt.doorbell_offset = input->doorbell_offset;
 195         mes_add_queue_pkt.mqd_addr = input->mqd_addr;
 196
 197         if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >>
 198                         AMDGPU_MES_API_VERSION_SHIFT) >= 2)
 199                 mes_add_queue_pkt.wptr_addr = input->wptr_mc_addr;
 200         else
 201                 mes_add_queue_pkt.wptr_addr = input->wptr_addr;
 202
 203         mes_add_queue_pkt.queue_type =
 204                 convert_to_mes_queue_type(input->queue_type);
 205         mes_add_queue_pkt.paging = input->paging;
 206         mes_add_queue_pkt.vm_context_cntl = vm_cntx_cntl;
 207         mes_add_queue_pkt.gws_base = input->gws_base;
 208         mes_add_queue_pkt.gws_size = input->gws_size;
 209         mes_add_queue_pkt.trap_handler_addr = input->tba_addr;
 210         mes_add_queue_pkt.tma_addr = input->tma_addr;
 211         mes_add_queue_pkt.trap_en = input->trap_en;
 212         mes_add_queue_pkt.skip_process_ctx_clear = input->skip_process_ctx_clear;
 213         mes_add_queue_pkt.is_kfd_process = input->is_kfd_process;
 214
 215         /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */
 216         mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
 217         mes_add_queue_pkt.gds_size = input->queue_size;
 218
 219         mes_add_queue_pkt.exclusively_scheduled = input->exclusively_scheduled;
 220
 221         return mes_v11_0_submit_pkt_and_poll_completion(mes,
 222                         &mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
 223                         offsetof(union MESAPI__ADD_QUEUE, api_status));
 224 }
 225
 226 static int mes_v11_0_remove_hw_queue(struct amdgpu_mes *mes,
 227                                      struct mes_remove_queue_input *input)
 228 {
 229         union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
 230
 231         memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
 232
 233         mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
 234         mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
 235         mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
 236
 237         mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
 238         mes_remove_queue_pkt.gang_context_addr = input->gang_context_addr;
 239
 240         return mes_v11_0_submit_pkt_and_poll_completion(mes,
 241                         &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
 242                         offsetof(union MESAPI__REMOVE_QUEUE, api_status));
 243 }
 244
 245 static int mes_v11_0_unmap_legacy_queue(struct amdgpu_mes *mes,
 246                         struct mes_unmap_legacy_queue_input *input)
 247 {
 248         union MESAPI__REMOVE_QUEUE mes_remove_queue_pkt;
 249
 250         memset(&mes_remove_queue_pkt, 0, sizeof(mes_remove_queue_pkt));
 251
 252         mes_remove_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
 253         mes_remove_queue_pkt.header.opcode = MES_SCH_API_REMOVE_QUEUE;
 254         mes_remove_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
 255
 256         mes_remove_queue_pkt.doorbell_offset = input->doorbell_offset;
 257         mes_remove_queue_pkt.gang_context_addr = 0;
 258
 259         mes_remove_queue_pkt.pipe_id = input->pipe_id;
 260         mes_remove_queue_pkt.queue_id = input->queue_id;
 261
 262         if (input->action == PREEMPT_QUEUES_NO_UNMAP) {
 263                 mes_remove_queue_pkt.preempt_legacy_gfx_queue = 1;
 264                 mes_remove_queue_pkt.tf_addr = input->trail_fence_addr;
 265                 mes_remove_queue_pkt.tf_data =
 266                         lower_32_bits(input->trail_fence_data);
 267         } else {
 268                 mes_remove_queue_pkt.unmap_legacy_queue = 1;
 269                 mes_remove_queue_pkt.queue_type =
 270                         convert_to_mes_queue_type(input->queue_type);
 271         }
 272
 273         return mes_v11_0_submit_pkt_and_poll_completion(mes,
 274                         &mes_remove_queue_pkt, sizeof(mes_remove_queue_pkt),
 275                         offsetof(union MESAPI__REMOVE_QUEUE, api_status));
 276 }
 277
 278 static int mes_v11_0_suspend_gang(struct amdgpu_mes *mes,
 279                                   struct mes_suspend_gang_input *input)
 280 {
 281         return 0;
 282 }
 283
 284 static int mes_v11_0_resume_gang(struct amdgpu_mes *mes,
 285                                  struct mes_resume_gang_input *input)
 286 {
 287         return 0;
 288 }
 289
 290 static int mes_v11_0_query_sched_status(struct amdgpu_mes *mes)
 291 {
 292         union MESAPI__QUERY_MES_STATUS mes_status_pkt;
 293
 294         memset(&mes_status_pkt, 0, sizeof(mes_status_pkt));
 295
 296         mes_status_pkt.header.type = MES_API_TYPE_SCHEDULER;
 297         mes_status_pkt.header.opcode = MES_SCH_API_QUERY_SCHEDULER_STATUS;
 298         mes_status_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
 299
 300         return mes_v11_0_submit_pkt_and_poll_completion(mes,
 301                         &mes_status_pkt, sizeof(mes_status_pkt),
 302                         offsetof(union MESAPI__QUERY_MES_STATUS, api_status));
 303 }
 304
 305 static int mes_v11_0_misc_op(struct amdgpu_mes *mes,
 306                              struct mes_misc_op_input *input)
 307 {
 308         union MESAPI__MISC misc_pkt;
 309
 310         memset(&misc_pkt, 0, sizeof(misc_pkt));
 311
 312         misc_pkt.header.type = MES_API_TYPE_SCHEDULER;
 313         misc_pkt.header.opcode = MES_SCH_API_MISC;
 314         misc_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
 315
 316         switch (input->op) {
 317         case MES_MISC_OP_READ_REG:
 318                 misc_pkt.opcode = MESAPI_MISC__READ_REG;
 319                 misc_pkt.read_reg.reg_offset = input->read_reg.reg_offset;
 320                 misc_pkt.read_reg.buffer_addr = input->read_reg.buffer_addr;
 321                 break;
 322         case MES_MISC_OP_WRITE_REG:
 323                 misc_pkt.opcode = MESAPI_MISC__WRITE_REG;
 324                 misc_pkt.write_reg.reg_offset = input->write_reg.reg_offset;
 325                 misc_pkt.write_reg.reg_value = input->write_reg.reg_value;
 326                 break;
 327         case MES_MISC_OP_WRM_REG_WAIT:
 328                 misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
 329                 misc_pkt.wait_reg_mem.op = WRM_OPERATION__WAIT_REG_MEM;
 330                 misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
 331                 misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
 332                 misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
 333                 misc_pkt.wait_reg_mem.reg_offset2 = 0;
 334                 break;
 335         case MES_MISC_OP_WRM_REG_WR_WAIT:
 336                 misc_pkt.opcode = MESAPI_MISC__WAIT_REG_MEM;
 337                 misc_pkt.wait_reg_mem.op = WRM_OPERATION__WR_WAIT_WR_REG;
 338                 misc_pkt.wait_reg_mem.reference = input->wrm_reg.ref;
 339                 misc_pkt.wait_reg_mem.mask = input->wrm_reg.mask;
 340                 misc_pkt.wait_reg_mem.reg_offset1 = input->wrm_reg.reg0;
 341                 misc_pkt.wait_reg_mem.reg_offset2 = input->wrm_reg.reg1;
 342                 break;
 343         case MES_MISC_OP_SET_SHADER_DEBUGGER:
 344                 misc_pkt.opcode = MESAPI_MISC__SET_SHADER_DEBUGGER;
 345                 misc_pkt.set_shader_debugger.process_context_addr =
 346                                 input->set_shader_debugger.process_context_addr;
 347                 misc_pkt.set_shader_debugger.flags.u32all =
 348                                 input->set_shader_debugger.flags.u32all;
 349                 misc_pkt.set_shader_debugger.spi_gdbg_per_vmid_cntl =
 350                                 input->set_shader_debugger.spi_gdbg_per_vmid_cntl;
 351                 memcpy(misc_pkt.set_shader_debugger.tcp_watch_cntl,
 352                                 input->set_shader_debugger.tcp_watch_cntl,
 353                                 sizeof(misc_pkt.set_shader_debugger.tcp_watch_cntl));
 354                 misc_pkt.set_shader_debugger.trap_en = input->set_shader_debugger.trap_en;
 355                 break;
 356         default:
 357                 DRM_ERROR("unsupported misc op (%d) \n", input->op);
 358                 return -EINVAL;
 359         }
 360
 361         return mes_v11_0_submit_pkt_and_poll_completion(mes,
 362                         &misc_pkt, sizeof(misc_pkt),
 363                         offsetof(union MESAPI__MISC, api_status));
 364 }
 365
 366 static int mes_v11_0_set_hw_resources(struct amdgpu_mes *mes)
 367 {
 368         int i;
 369         struct amdgpu_device *adev = mes->adev;
 370         union MESAPI_SET_HW_RESOURCES mes_set_hw_res_pkt;
 371
 372         memset(&mes_set_hw_res_pkt, 0, sizeof(mes_set_hw_res_pkt));
 373
 374         mes_set_hw_res_pkt.header.type = MES_API_TYPE_SCHEDULER;
 375         mes_set_hw_res_pkt.header.opcode = MES_SCH_API_SET_HW_RSRC;
 376         mes_set_hw_res_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
 377
 378         mes_set_hw_res_pkt.vmid_mask_mmhub = mes->vmid_mask_mmhub;
 379         mes_set_hw_res_pkt.vmid_mask_gfxhub = mes->vmid_mask_gfxhub;
 380         mes_set_hw_res_pkt.gds_size = adev->gds.gds_size;
 381         mes_set_hw_res_pkt.paging_vmid = 0;
 382         mes_set_hw_res_pkt.g_sch_ctx_gpu_mc_ptr = mes->sch_ctx_gpu_addr;
 383         mes_set_hw_res_pkt.query_status_fence_gpu_mc_ptr =
 384                 mes->query_status_fence_gpu_addr;
 385
 386         for (i = 0; i < MAX_COMPUTE_PIPES; i++)
 387                 mes_set_hw_res_pkt.compute_hqd_mask[i] =
 388                         mes->compute_hqd_mask[i];
 389
 390         for (i = 0; i < MAX_GFX_PIPES; i++)
 391                 mes_set_hw_res_pkt.gfx_hqd_mask[i] = mes->gfx_hqd_mask[i];
 392
 393         for (i = 0; i < MAX_SDMA_PIPES; i++)
 394                 mes_set_hw_res_pkt.sdma_hqd_mask[i] = mes->sdma_hqd_mask[i];
 395
 396         for (i = 0; i < AMD_PRIORITY_NUM_LEVELS; i++)
 397                 mes_set_hw_res_pkt.aggregated_doorbells[i] =
 398                         mes->aggregated_doorbells[i];
 399
 400         for (i = 0; i < 5; i++) {
 401                 mes_set_hw_res_pkt.gc_base[i] = adev->reg_offset[GC_HWIP][0][i];
 402                 mes_set_hw_res_pkt.mmhub_base[i] =
 403                                 adev->reg_offset[MMHUB_HWIP][0][i];
 404                 mes_set_hw_res_pkt.osssys_base[i] =
 405                 adev->reg_offset[OSSSYS_HWIP][0][i];
 406         }
 407
 408         mes_set_hw_res_pkt.disable_reset = 1;
 409         mes_set_hw_res_pkt.disable_mes_log = 1;
 410         mes_set_hw_res_pkt.use_different_vmid_compute = 1;
 411         mes_set_hw_res_pkt.enable_reg_active_poll = 1;
 412         mes_set_hw_res_pkt.enable_level_process_quantum_check = 1;
 413         mes_set_hw_res_pkt.oversubscription_timer = 50;
 414         mes_set_hw_res_pkt.enable_mes_event_int_logging = 1;
 415         mes_set_hw_res_pkt.event_intr_history_gpu_mc_ptr = mes->event_log_gpu_addr;
 416
 417         return mes_v11_0_submit_pkt_and_poll_completion(mes,
 418                         &mes_set_hw_res_pkt, sizeof(mes_set_hw_res_pkt),
 419                         offsetof(union MESAPI_SET_HW_RESOURCES, api_status));
 420 }
 421
 422 static const struct amdgpu_mes_funcs mes_v11_0_funcs = {
 423         .add_hw_queue = mes_v11_0_add_hw_queue,
 424         .remove_hw_queue = mes_v11_0_remove_hw_queue,
 425         .unmap_legacy_queue = mes_v11_0_unmap_legacy_queue,
 426         .suspend_gang = mes_v11_0_suspend_gang,
 427         .resume_gang = mes_v11_0_resume_gang,
 428         .misc_op = mes_v11_0_misc_op,
 429 };
 430
 431 static int mes_v11_0_allocate_ucode_buffer(struct amdgpu_device *adev,
 432                                            enum admgpu_mes_pipe pipe)
 433 {
 434         int r;
 435         const struct mes_firmware_header_v1_0 *mes_hdr;
 436         const __le32 *fw_data;
 437         unsigned fw_size;
 438
 439         mes_hdr = (const struct mes_firmware_header_v1_0 *)
 440                 adev->mes.fw[pipe]->data;
 441
 442         fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
 443                    le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
 444         fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
 445
 446         r = amdgpu_bo_create_reserved(adev, fw_size,
 447                                       PAGE_SIZE,
 448                                       AMDGPU_GEM_DOMAIN_VRAM |
 449                                       AMDGPU_GEM_DOMAIN_GTT,
 450                                       &adev->mes.ucode_fw_obj[pipe],
 451                                       &adev->mes.ucode_fw_gpu_addr[pipe],
 452                                       (void **)&adev->mes.ucode_fw_ptr[pipe]);
 453         if (r) {
 454                 dev_err(adev->dev, "(%d) failed to create mes fw bo\n", r);
 455                 return r;
 456         }
 457
 458         memcpy(adev->mes.ucode_fw_ptr[pipe], fw_data, fw_size);
 459
 460         amdgpu_bo_kunmap(adev->mes.ucode_fw_obj[pipe]);
 461         amdgpu_bo_unreserve(adev->mes.ucode_fw_obj[pipe]);
 462
 463         return 0;
 464 }
 465
 466 static int mes_v11_0_allocate_ucode_data_buffer(struct amdgpu_device *adev,
 467                                                 enum admgpu_mes_pipe pipe)
 468 {
 469         int r;
 470         const struct mes_firmware_header_v1_0 *mes_hdr;
 471         const __le32 *fw_data;
 472         unsigned fw_size;
 473
 474         mes_hdr = (const struct mes_firmware_header_v1_0 *)
 475                 adev->mes.fw[pipe]->data;
 476
 477         fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
 478                    le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
 479         fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
 480
 481         if (fw_size > GFX_MES_DRAM_SIZE) {
 482                 dev_err(adev->dev, "PIPE%d ucode data fw size (%d) is greater than dram size (%d)\n",
 483                         pipe, fw_size, GFX_MES_DRAM_SIZE);
 484                 return -EINVAL;
 485         }
 486
 487         r = amdgpu_bo_create_reserved(adev, GFX_MES_DRAM_SIZE,
 488                                       64 * 1024,
 489                                       AMDGPU_GEM_DOMAIN_VRAM |
 490                                       AMDGPU_GEM_DOMAIN_GTT,
 491                                       &adev->mes.data_fw_obj[pipe],
 492                                       &adev->mes.data_fw_gpu_addr[pipe],
 493                                       (void **)&adev->mes.data_fw_ptr[pipe]);
 494         if (r) {
 495                 dev_err(adev->dev, "(%d) failed to create mes data fw bo\n", r);
 496                 return r;
 497         }
 498
 499         memcpy(adev->mes.data_fw_ptr[pipe], fw_data, fw_size);
 500
 501         amdgpu_bo_kunmap(adev->mes.data_fw_obj[pipe]);
 502         amdgpu_bo_unreserve(adev->mes.data_fw_obj[pipe]);
 503
 504         return 0;
 505 }
 506
 507 static void mes_v11_0_free_ucode_buffers(struct amdgpu_device *adev,
 508                                          enum admgpu_mes_pipe pipe)
 509 {
 510         amdgpu_bo_free_kernel(&adev->mes.data_fw_obj[pipe],
 511                               &adev->mes.data_fw_gpu_addr[pipe],
 512                               (void **)&adev->mes.data_fw_ptr[pipe]);
 513
 514         amdgpu_bo_free_kernel(&adev->mes.ucode_fw_obj[pipe],
 515                               &adev->mes.ucode_fw_gpu_addr[pipe],
 516                               (void **)&adev->mes.ucode_fw_ptr[pipe]);
 517 }
 518
 519 static void mes_v11_0_enable(struct amdgpu_device *adev, bool enable)
 520 {
 521         uint64_t ucode_addr;
 522         uint32_t pipe, data = 0;
 523
 524         if (enable) {
 525                 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
 526                 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
 527                 data = REG_SET_FIELD(data, CP_MES_CNTL,
 528                              MES_PIPE1_RESET, adev->enable_mes_kiq ? 1 : 0);
 529                 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
 530
 531                 mutex_lock(&adev->srbm_mutex);
 532                 for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
 533                         if (!adev->enable_mes_kiq &&
 534                             pipe == AMDGPU_MES_KIQ_PIPE)
 535                                 continue;
 536
 537                         soc21_grbm_select(adev, 3, pipe, 0, 0);
 538
 539                         ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
 540                         WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
 541                                      lower_32_bits(ucode_addr));
 542                         WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
 543                                      upper_32_bits(ucode_addr));
 544                 }
 545                 soc21_grbm_select(adev, 0, 0, 0, 0);
 546                 mutex_unlock(&adev->srbm_mutex);
 547
 548                 /* unhalt MES and activate pipe0 */
 549                 data = REG_SET_FIELD(0, CP_MES_CNTL, MES_PIPE0_ACTIVE, 1);
 550                 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE,
 551                                      adev->enable_mes_kiq ? 1 : 0);
 552                 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
 553
 554                 if (amdgpu_emu_mode)
 555                         msleep(100);
 556                 else
 557                         udelay(50);
 558         } else {
 559                 data = RREG32_SOC15(GC, 0, regCP_MES_CNTL);
 560                 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_ACTIVE, 0);
 561                 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_ACTIVE, 0);
 562                 data = REG_SET_FIELD(data, CP_MES_CNTL,
 563                                      MES_INVALIDATE_ICACHE, 1);
 564                 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE0_RESET, 1);
 565                 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_PIPE1_RESET,
 566                                      adev->enable_mes_kiq ? 1 : 0);
 567                 data = REG_SET_FIELD(data, CP_MES_CNTL, MES_HALT, 1);
 568                 WREG32_SOC15(GC, 0, regCP_MES_CNTL, data);
 569         }
 570 }
 571
 572 /* This function is for backdoor MES firmware */
 573 static int mes_v11_0_load_microcode(struct amdgpu_device *adev,
 574                                     enum admgpu_mes_pipe pipe, bool prime_icache)
 575 {
 576         int r;
 577         uint32_t data;
 578         uint64_t ucode_addr;
 579
 580         mes_v11_0_enable(adev, false);
 581
 582         if (!adev->mes.fw[pipe])
 583                 return -EINVAL;
 584
 585         r = mes_v11_0_allocate_ucode_buffer(adev, pipe);
 586         if (r)
 587                 return r;
 588
 589         r = mes_v11_0_allocate_ucode_data_buffer(adev, pipe);
 590         if (r) {
 591                 mes_v11_0_free_ucode_buffers(adev, pipe);
 592                 return r;
 593         }
 594
 595         mutex_lock(&adev->srbm_mutex);
 596         /* me=3, pipe=0, queue=0 */
 597         soc21_grbm_select(adev, 3, pipe, 0, 0);
 598
 599         WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_CNTL, 0);
 600
 601         /* set ucode start address */
 602         ucode_addr = adev->mes.uc_start_addr[pipe] >> 2;
 603         WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START,
 604                      lower_32_bits(ucode_addr));
 605         WREG32_SOC15(GC, 0, regCP_MES_PRGRM_CNTR_START_HI,
 606                      upper_32_bits(ucode_addr));
 607
 608         /* set ucode fimrware address */
 609         WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_LO,
 610                      lower_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
 611         WREG32_SOC15(GC, 0, regCP_MES_IC_BASE_HI,
 612                      upper_32_bits(adev->mes.ucode_fw_gpu_addr[pipe]));
 613
 614         /* set ucode instruction cache boundary to 2M-1 */
 615         WREG32_SOC15(GC, 0, regCP_MES_MIBOUND_LO, 0x1FFFFF);
 616
 617         /* set ucode data firmware address */
 618         WREG32_SOC15(GC, 0, regCP_MES_MDBASE_LO,
 619                      lower_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
 620         WREG32_SOC15(GC, 0, regCP_MES_MDBASE_HI,
 621                      upper_32_bits(adev->mes.data_fw_gpu_addr[pipe]));
 622
 623         /* Set 0x7FFFF (512K-1) to CP_MES_MDBOUND_LO */
 624         WREG32_SOC15(GC, 0, regCP_MES_MDBOUND_LO, 0x7FFFF);
 625
 626         if (prime_icache) {
 627                 /* invalidate ICACHE */
 628                 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
 629                 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 0);
 630                 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, INVALIDATE_CACHE, 1);
 631                 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
 632
 633                 /* prime the ICACHE. */
 634                 data = RREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL);
 635                 data = REG_SET_FIELD(data, CP_MES_IC_OP_CNTL, PRIME_ICACHE, 1);
 636                 WREG32_SOC15(GC, 0, regCP_MES_IC_OP_CNTL, data);
 637         }
 638
 639         soc21_grbm_select(adev, 0, 0, 0, 0);
 640         mutex_unlock(&adev->srbm_mutex);
 641
 642         return 0;
 643 }
 644
 645 static int mes_v11_0_allocate_eop_buf(struct amdgpu_device *adev,
 646                                       enum admgpu_mes_pipe pipe)
 647 {
 648         int r;
 649         u32 *eop;
 650
 651         r = amdgpu_bo_create_reserved(adev, MES_EOP_SIZE, PAGE_SIZE,
 652                               AMDGPU_GEM_DOMAIN_GTT,
 653                               &adev->mes.eop_gpu_obj[pipe],
 654                               &adev->mes.eop_gpu_addr[pipe],
 655                               (void **)&eop);
 656         if (r) {
 657                 dev_warn(adev->dev, "(%d) create EOP bo failed\n", r);
 658                 return r;
 659         }
 660
 661         memset(eop, 0,
 662                adev->mes.eop_gpu_obj[pipe]->tbo.base.size);
 663
 664         amdgpu_bo_kunmap(adev->mes.eop_gpu_obj[pipe]);
 665         amdgpu_bo_unreserve(adev->mes.eop_gpu_obj[pipe]);
 666
 667         return 0;
 668 }
 669
 670 static int mes_v11_0_mqd_init(struct amdgpu_ring *ring)
 671 {
 672         struct v11_compute_mqd *mqd = ring->mqd_ptr;
 673         uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
 674         uint32_t tmp;
 675
 676         memset(mqd, 0, sizeof(*mqd));
 677
 678         mqd->header = 0xC0310800;
 679         mqd->compute_pipelinestat_enable = 0x00000001;
 680         mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
 681         mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
 682         mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
 683         mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
 684         mqd->compute_misc_reserved = 0x00000007;
 685
 686         eop_base_addr = ring->eop_gpu_addr >> 8;
 687
 688         /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
 689         tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
 690         tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
 691                         (order_base_2(MES_EOP_SIZE / 4) - 1));
 692
 693         mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_base_addr);
 694         mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
 695         mqd->cp_hqd_eop_control = tmp;
 696
 697         /* disable the queue if it's active */
 698         ring->wptr = 0;
 699         mqd->cp_hqd_pq_rptr = 0;
 700         mqd->cp_hqd_pq_wptr_lo = 0;
 701         mqd->cp_hqd_pq_wptr_hi = 0;
 702
 703         /* set the pointer to the MQD */
 704         mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
 705         mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
 706
 707         /* set MQD vmid to 0 */
 708         tmp = regCP_MQD_CONTROL_DEFAULT;
 709         tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
 710         mqd->cp_mqd_control = tmp;
 711
 712         /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
 713         hqd_gpu_addr = ring->gpu_addr >> 8;
 714         mqd->cp_hqd_pq_base_lo = lower_32_bits(hqd_gpu_addr);
 715         mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
 716
 717         /* set the wb address whether it's enabled or not */
 718         wb_gpu_addr = ring->rptr_gpu_addr;
 719         mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
 720         mqd->cp_hqd_pq_rptr_report_addr_hi =
 721                 upper_32_bits(wb_gpu_addr) & 0xffff;
 722
 723         /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
 724         wb_gpu_addr = ring->wptr_gpu_addr;
 725         mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffff8;
 726         mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
 727
 728         /* set up the HQD, this is similar to CP_RB0_CNTL */
 729         tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
 730         tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
 731                             (order_base_2(ring->ring_size / 4) - 1));
 732         tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
 733                             ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
 734         tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
 735         tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
 736         tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
 737         tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
 738         tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, NO_UPDATE_RPTR, 1);
 739         mqd->cp_hqd_pq_control = tmp;
 740
 741         /* enable doorbell */
 742         tmp = 0;
 743         if (ring->use_doorbell) {
 744                 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
 745                                     DOORBELL_OFFSET, ring->doorbell_index);
 746                 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
 747                                     DOORBELL_EN, 1);
 748                 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
 749                                     DOORBELL_SOURCE, 0);
 750                 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
 751                                     DOORBELL_HIT, 0);
 752         } else
 753                 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
 754                                     DOORBELL_EN, 0);
 755         mqd->cp_hqd_pq_doorbell_control = tmp;
 756
 757         mqd->cp_hqd_vmid = 0;
 758         /* activate the queue */
 759         mqd->cp_hqd_active = 1;
 760
 761         tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
 762         tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE,
 763                             PRELOAD_SIZE, 0x55);
 764         mqd->cp_hqd_persistent_state = tmp;
 765
 766         mqd->cp_hqd_ib_control = regCP_HQD_IB_CONTROL_DEFAULT;
 767         mqd->cp_hqd_iq_timer = regCP_HQD_IQ_TIMER_DEFAULT;
 768         mqd->cp_hqd_quantum = regCP_HQD_QUANTUM_DEFAULT;
 769
 770         amdgpu_device_flush_hdp(ring->adev, NULL);
 771         return 0;
 772 }
 773
 774 static void mes_v11_0_queue_init_register(struct amdgpu_ring *ring)
 775 {
 776         struct v11_compute_mqd *mqd = ring->mqd_ptr;
 777         struct amdgpu_device *adev = ring->adev;
 778         uint32_t data = 0;
 779
 780         mutex_lock(&adev->srbm_mutex);
 781         soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
 782
 783         /* set CP_HQD_VMID.VMID = 0. */
 784         data = RREG32_SOC15(GC, 0, regCP_HQD_VMID);
 785         data = REG_SET_FIELD(data, CP_HQD_VMID, VMID, 0);
 786         WREG32_SOC15(GC, 0, regCP_HQD_VMID, data);
 787
 788         /* set CP_HQD_PQ_DOORBELL_CONTROL.DOORBELL_EN=0 */
 789         data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
 790         data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
 791                              DOORBELL_EN, 0);
 792         WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
 793
 794         /* set CP_MQD_BASE_ADDR/HI with the MQD base address */
 795         WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
 796         WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
 797
 798         /* set CP_MQD_CONTROL.VMID=0 */
 799         data = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL);
 800         data = REG_SET_FIELD(data, CP_MQD_CONTROL, VMID, 0);
 801         WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 0);
 802
 803         /* set CP_HQD_PQ_BASE/HI with the ring buffer base address */
 804         WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
 805         WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
 806
 807         /* set CP_HQD_PQ_RPTR_REPORT_ADDR/HI */
 808         WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
 809                      mqd->cp_hqd_pq_rptr_report_addr_lo);
 810         WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
 811                      mqd->cp_hqd_pq_rptr_report_addr_hi);
 812
 813         /* set CP_HQD_PQ_CONTROL */
 814         WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, mqd->cp_hqd_pq_control);
 815
 816         /* set CP_HQD_PQ_WPTR_POLL_ADDR/HI */
 817         WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
 818                      mqd->cp_hqd_pq_wptr_poll_addr_lo);
 819         WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
 820                      mqd->cp_hqd_pq_wptr_poll_addr_hi);
 821
 822         /* set CP_HQD_PQ_DOORBELL_CONTROL */
 823         WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
 824                      mqd->cp_hqd_pq_doorbell_control);
 825
 826         /* set CP_HQD_PERSISTENT_STATE.PRELOAD_SIZE=0x53 */
 827         WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, mqd->cp_hqd_persistent_state);
 828
 829         /* set CP_HQD_ACTIVE.ACTIVE=1 */
 830         WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, mqd->cp_hqd_active);
 831
 832         soc21_grbm_select(adev, 0, 0, 0, 0);
 833         mutex_unlock(&adev->srbm_mutex);
 834 }
 835
 836 static int mes_v11_0_kiq_enable_queue(struct amdgpu_device *adev)
 837 {
 838         struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
 839         struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
 840         int r;
 841
 842         if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
 843                 return -EINVAL;
 844
 845         r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size);
 846         if (r) {
 847                 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
 848                 return r;
 849         }
 850
 851         kiq->pmf->kiq_map_queues(kiq_ring, &adev->mes.ring);
 852
 853         return amdgpu_ring_test_helper(kiq_ring);
 854 }
 855
 856 static int mes_v11_0_queue_init(struct amdgpu_device *adev,
 857                                 enum admgpu_mes_pipe pipe)
 858 {
 859         struct amdgpu_ring *ring;
 860         int r;
 861
 862         if (pipe == AMDGPU_MES_KIQ_PIPE)
 863                 ring = &adev->gfx.kiq[0].ring;
 864         else if (pipe == AMDGPU_MES_SCHED_PIPE)
 865                 ring = &adev->mes.ring;
 866         else
 867                 BUG();
 868
 869         if ((pipe == AMDGPU_MES_SCHED_PIPE) &&
 870             (amdgpu_in_reset(adev) || adev->in_suspend)) {
 871                 *(ring->wptr_cpu_addr) = 0;
 872                 *(ring->rptr_cpu_addr) = 0;
 873                 amdgpu_ring_clear_ring(ring);
 874         }
 875
 876         r = mes_v11_0_mqd_init(ring);
 877         if (r)
 878                 return r;
 879
 880         if (pipe == AMDGPU_MES_SCHED_PIPE) {
 881                 r = mes_v11_0_kiq_enable_queue(adev);
 882                 if (r)
 883                         return r;
 884         } else {
 885                 mes_v11_0_queue_init_register(ring);
 886         }
 887
 888         /* get MES scheduler/KIQ versions */
 889         mutex_lock(&adev->srbm_mutex);
 890         soc21_grbm_select(adev, 3, pipe, 0, 0);
 891
 892         if (pipe == AMDGPU_MES_SCHED_PIPE)
 893                 adev->mes.sched_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
 894         else if (pipe == AMDGPU_MES_KIQ_PIPE && adev->enable_mes_kiq)
 895                 adev->mes.kiq_version = RREG32_SOC15(GC, 0, regCP_MES_GP3_LO);
 896
 897         soc21_grbm_select(adev, 0, 0, 0, 0);
 898         mutex_unlock(&adev->srbm_mutex);
 899
 900         return 0;
 901 }
 902
 903 static int mes_v11_0_ring_init(struct amdgpu_device *adev)
 904 {
 905         struct amdgpu_ring *ring;
 906
 907         ring = &adev->mes.ring;
 908
 909         ring->funcs = &mes_v11_0_ring_funcs;
 910
 911         ring->me = 3;
 912         ring->pipe = 0;
 913         ring->queue = 0;
 914
 915         ring->ring_obj = NULL;
 916         ring->use_doorbell = true;
 917         ring->doorbell_index = adev->doorbell_index.mes_ring0 << 1;
 918         ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_SCHED_PIPE];
 919         ring->no_scheduler = true;
 920         sprintf(ring->name, "mes_%d.%d.%d", ring->me, ring->pipe, ring->queue);
 921
 922         return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
 923                                 AMDGPU_RING_PRIO_DEFAULT, NULL);
 924 }
 925
 926 static int mes_v11_0_kiq_ring_init(struct amdgpu_device *adev)
 927 {
 928         struct amdgpu_ring *ring;
 929
 930         spin_lock_init(&adev->gfx.kiq[0].ring_lock);
 931
 932         ring = &adev->gfx.kiq[0].ring;
 933
 934         ring->me = 3;
 935         ring->pipe = 1;
 936         ring->queue = 0;
 937
 938         ring->adev = NULL;
 939         ring->ring_obj = NULL;
 940         ring->use_doorbell = true;
 941         ring->doorbell_index = adev->doorbell_index.mes_ring1 << 1;
 942         ring->eop_gpu_addr = adev->mes.eop_gpu_addr[AMDGPU_MES_KIQ_PIPE];
 943         ring->no_scheduler = true;
 944         sprintf(ring->name, "mes_kiq_%d.%d.%d",
 945                 ring->me, ring->pipe, ring->queue);
 946
 947         return amdgpu_ring_init(adev, ring, 1024, NULL, 0,
 948                                 AMDGPU_RING_PRIO_DEFAULT, NULL);
 949 }
 950
 951 static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev,
 952                                  enum admgpu_mes_pipe pipe)
 953 {
 954         int r, mqd_size = sizeof(struct v11_compute_mqd);
 955         struct amdgpu_ring *ring;
 956
 957         if (pipe == AMDGPU_MES_KIQ_PIPE)
 958                 ring = &adev->gfx.kiq[0].ring;
 959         else if (pipe == AMDGPU_MES_SCHED_PIPE)
 960                 ring = &adev->mes.ring;
 961         else
 962                 BUG();
 963
 964         if (ring->mqd_obj)
 965                 return 0;
 966
 967         r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
 968                                     AMDGPU_GEM_DOMAIN_VRAM |
 969                                     AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
 970                                     &ring->mqd_gpu_addr, &ring->mqd_ptr);
 971         if (r) {
 972                 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
 973                 return r;
 974         }
 975
 976         memset(ring->mqd_ptr, 0, mqd_size);
 977
 978         /* prepare MQD backup */
 979         adev->mes.mqd_backup[pipe] = kmalloc(mqd_size, GFP_KERNEL);
 980         if (!adev->mes.mqd_backup[pipe]) {
 981                 dev_warn(adev->dev,
 982                          "no memory to create MQD backup for ring %s\n",
 983                          ring->name);
 984                 return -ENOMEM;
 985         }
 986
 987         return 0;
 988 }
 989
 990 static int mes_v11_0_sw_init(void *handle)
 991 {
 992         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 993         int pipe, r;
 994
 995         adev->mes.funcs = &mes_v11_0_funcs;
 996         adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init;
 997         adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini;
 998
 999         r = amdgpu_mes_init(adev);
1000         if (r)
1001                 return r;
1002
1003         for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1004                 if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
1005                         continue;
1006
1007                 r = mes_v11_0_allocate_eop_buf(adev, pipe);
1008                 if (r)
1009                         return r;
1010
1011                 r = mes_v11_0_mqd_sw_init(adev, pipe);
1012                 if (r)
1013                         return r;
1014         }
1015
1016         if (adev->enable_mes_kiq) {
1017                 r = mes_v11_0_kiq_ring_init(adev);
1018                 if (r)
1019                         return r;
1020         }
1021
1022         r = mes_v11_0_ring_init(adev);
1023         if (r)
1024                 return r;
1025
1026         return 0;
1027 }
1028
1029 static int mes_v11_0_sw_fini(void *handle)
1030 {
1031         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1032         int pipe;
1033
1034         amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
1035         amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
1036
1037         for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1038                 kfree(adev->mes.mqd_backup[pipe]);
1039
1040                 amdgpu_bo_free_kernel(&adev->mes.eop_gpu_obj[pipe],
1041                                       &adev->mes.eop_gpu_addr[pipe],
1042                                       NULL);
1043                 amdgpu_ucode_release(&adev->mes.fw[pipe]);
1044         }
1045
1046         amdgpu_bo_free_kernel(&adev->gfx.kiq[0].ring.mqd_obj,
1047                               &adev->gfx.kiq[0].ring.mqd_gpu_addr,
1048                               &adev->gfx.kiq[0].ring.mqd_ptr);
1049
1050         amdgpu_bo_free_kernel(&adev->mes.ring.mqd_obj,
1051                               &adev->mes.ring.mqd_gpu_addr,
1052                               &adev->mes.ring.mqd_ptr);
1053
1054         amdgpu_ring_fini(&adev->gfx.kiq[0].ring);
1055         amdgpu_ring_fini(&adev->mes.ring);
1056
1057         if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1058                 mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_KIQ_PIPE);
1059                 mes_v11_0_free_ucode_buffers(adev, AMDGPU_MES_SCHED_PIPE);
1060         }
1061
1062         amdgpu_mes_fini(adev);
1063         return 0;
1064 }
1065
1066 static void mes_v11_0_kiq_dequeue(struct amdgpu_ring *ring)
1067 {
1068         uint32_t data;
1069         int i;
1070         struct amdgpu_device *adev = ring->adev;
1071
1072         mutex_lock(&adev->srbm_mutex);
1073         soc21_grbm_select(adev, 3, ring->pipe, 0, 0);
1074
1075         /* disable the queue if it's active */
1076         if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
1077                 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
1078                 for (i = 0; i < adev->usec_timeout; i++) {
1079                         if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
1080                                 break;
1081                         udelay(1);
1082                 }
1083         }
1084         data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
1085         data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1086                                 DOORBELL_EN, 0);
1087         data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
1088                                 DOORBELL_HIT, 1);
1089         WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
1090
1091         WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0);
1092
1093         WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0);
1094         WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0);
1095         WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0);
1096
1097         soc21_grbm_select(adev, 0, 0, 0, 0);
1098         mutex_unlock(&adev->srbm_mutex);
1099 }
1100
1101 static void mes_v11_0_kiq_setting(struct amdgpu_ring *ring)
1102 {
1103         uint32_t tmp;
1104         struct amdgpu_device *adev = ring->adev;
1105
1106         /* tell RLC which is KIQ queue */
1107         tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1108         tmp &= 0xffffff00;
1109         tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1110         WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1111         tmp |= 0x80;
1112         WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1113 }
1114
1115 static void mes_v11_0_kiq_clear(struct amdgpu_device *adev)
1116 {
1117         uint32_t tmp;
1118
1119         /* tell RLC which is KIQ dequeue */
1120         tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
1121         tmp &= ~RLC_CP_SCHEDULERS__scheduler0_MASK;
1122         WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
1123 }
1124
1125 static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev)
1126 {
1127         int r = 0;
1128
1129         if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1130
1131                 r = mes_v11_0_load_microcode(adev, AMDGPU_MES_SCHED_PIPE, false);
1132                 if (r) {
1133                         DRM_ERROR("failed to load MES fw, r=%d\n", r);
1134                         return r;
1135                 }
1136
1137                 r = mes_v11_0_load_microcode(adev, AMDGPU_MES_KIQ_PIPE, true);
1138                 if (r) {
1139                         DRM_ERROR("failed to load MES kiq fw, r=%d\n", r);
1140                         return r;
1141                 }
1142
1143         }
1144
1145         mes_v11_0_enable(adev, true);
1146
1147         mes_v11_0_kiq_setting(&adev->gfx.kiq[0].ring);
1148
1149         r = mes_v11_0_queue_init(adev, AMDGPU_MES_KIQ_PIPE);
1150         if (r)
1151                 goto failure;
1152
1153         return r;
1154
1155 failure:
1156         mes_v11_0_hw_fini(adev);
1157         return r;
1158 }
1159
1160 static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev)
1161 {
1162         if (adev->mes.ring.sched.ready) {
1163                 mes_v11_0_kiq_dequeue(&adev->mes.ring);
1164                 adev->mes.ring.sched.ready = false;
1165         }
1166
1167         if (amdgpu_sriov_vf(adev)) {
1168                 mes_v11_0_kiq_dequeue(&adev->gfx.kiq[0].ring);
1169                 mes_v11_0_kiq_clear(adev);
1170         }
1171
1172         mes_v11_0_enable(adev, false);
1173
1174         return 0;
1175 }
1176
1177 static int mes_v11_0_hw_init(void *handle)
1178 {
1179         int r;
1180         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1181
1182         if (!adev->enable_mes_kiq) {
1183                 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1184                         r = mes_v11_0_load_microcode(adev,
1185                                              AMDGPU_MES_SCHED_PIPE, true);
1186                         if (r) {
1187                                 DRM_ERROR("failed to MES fw, r=%d\n", r);
1188                                 return r;
1189                         }
1190                 }
1191
1192                 mes_v11_0_enable(adev, true);
1193         }
1194
1195         r = mes_v11_0_queue_init(adev, AMDGPU_MES_SCHED_PIPE);
1196         if (r)
1197                 goto failure;
1198
1199         r = mes_v11_0_set_hw_resources(&adev->mes);
1200         if (r)
1201                 goto failure;
1202
1203         r = mes_v11_0_query_sched_status(&adev->mes);
1204         if (r) {
1205                 DRM_ERROR("MES is busy\n");
1206                 goto failure;
1207         }
1208
1209         /*
1210          * Disable KIQ ring usage from the driver once MES is enabled.
1211          * MES uses KIQ ring exclusively so driver cannot access KIQ ring
1212          * with MES enabled.
1213          */
1214         adev->gfx.kiq[0].ring.sched.ready = false;
1215         adev->mes.ring.sched.ready = true;
1216
1217         return 0;
1218
1219 failure:
1220         mes_v11_0_hw_fini(adev);
1221         return r;
1222 }
1223
1224 static int mes_v11_0_hw_fini(void *handle)
1225 {
1226         return 0;
1227 }
1228
1229 static int mes_v11_0_suspend(void *handle)
1230 {
1231         int r;
1232         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1233
1234         r = amdgpu_mes_suspend(adev);
1235         if (r)
1236                 return r;
1237
1238         return mes_v11_0_hw_fini(adev);
1239 }
1240
1241 static int mes_v11_0_resume(void *handle)
1242 {
1243         int r;
1244         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1245
1246         r = mes_v11_0_hw_init(adev);
1247         if (r)
1248                 return r;
1249
1250         return amdgpu_mes_resume(adev);
1251 }
1252
1253 static int mes_v11_0_early_init(void *handle)
1254 {
1255         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1256         int pipe, r;
1257
1258         for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
1259                 if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
1260                         continue;
1261                 r = amdgpu_mes_init_microcode(adev, pipe);
1262                 if (r)
1263                         return r;
1264         }
1265
1266         return 0;
1267 }
1268
1269 static int mes_v11_0_late_init(void *handle)
1270 {
1271         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1272
1273         /* it's only intended for use in mes_self_test case, not for s0ix and reset */
1274         if (!amdgpu_in_reset(adev) && !adev->in_s0ix && !adev->in_suspend &&
1275             (amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(11, 0, 3)))
1276                 amdgpu_mes_self_test(adev);
1277
1278         return 0;
1279 }
1280
1281 static const struct amd_ip_funcs mes_v11_0_ip_funcs = {
1282         .name = "mes_v11_0",
1283         .early_init = mes_v11_0_early_init,
1284         .late_init = mes_v11_0_late_init,
1285         .sw_init = mes_v11_0_sw_init,
1286         .sw_fini = mes_v11_0_sw_fini,
1287         .hw_init = mes_v11_0_hw_init,
1288         .hw_fini = mes_v11_0_hw_fini,
1289         .suspend = mes_v11_0_suspend,
1290         .resume = mes_v11_0_resume,
1291 };
1292
1293 const struct amdgpu_ip_block_version mes_v11_0_ip_block = {
1294         .type = AMD_IP_BLOCK_TYPE_MES,
1295         .major = 11,
1296         .minor = 0,
1297         .rev = 0,
1298         .funcs = &mes_v11_0_ip_funcs,
1299 };