1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2015-2018 Etnaviv Project 4 */ 5 6 #include <linux/clk.h> 7 #include <linux/component.h> 8 #include <linux/delay.h> 9 #include <linux/dma-fence.h> 10 #include <linux/dma-mapping.h> 11 #include <linux/mod_devicetable.h> 12 #include <linux/module.h> 13 #include <linux/platform_device.h> 14 #include <linux/pm_runtime.h> 15 #include <linux/regulator/consumer.h> 16 #include <linux/thermal.h> 17 18 #include "etnaviv_cmdbuf.h" 19 #include "etnaviv_dump.h" 20 #include "etnaviv_gpu.h" 21 #include "etnaviv_gem.h" 22 #include "etnaviv_mmu.h" 23 #include "etnaviv_perfmon.h" 24 #include "etnaviv_sched.h" 25 #include "common.xml.h" 26 #include "state.xml.h" 27 #include "state_hi.xml.h" 28 #include "cmdstream.xml.h" 29 30 static const struct platform_device_id gpu_ids[] = { 31 { .name = "etnaviv-gpu,2d" }, 32 { }, 33 }; 34 35 /* 36 * Driver functions: 37 */ 38 39 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value) 40 { 41 struct etnaviv_drm_private *priv = gpu->drm->dev_private; 42 43 switch (param) { 44 case ETNAVIV_PARAM_GPU_MODEL: 45 *value = gpu->identity.model; 46 break; 47 48 case ETNAVIV_PARAM_GPU_REVISION: 49 *value = gpu->identity.revision; 50 break; 51 52 case ETNAVIV_PARAM_GPU_FEATURES_0: 53 *value = gpu->identity.features; 54 break; 55 56 case ETNAVIV_PARAM_GPU_FEATURES_1: 57 *value = gpu->identity.minor_features0; 58 break; 59 60 case ETNAVIV_PARAM_GPU_FEATURES_2: 61 *value = gpu->identity.minor_features1; 62 break; 63 64 case ETNAVIV_PARAM_GPU_FEATURES_3: 65 *value = gpu->identity.minor_features2; 66 break; 67 68 case ETNAVIV_PARAM_GPU_FEATURES_4: 69 *value = gpu->identity.minor_features3; 70 break; 71 72 case ETNAVIV_PARAM_GPU_FEATURES_5: 73 *value = gpu->identity.minor_features4; 74 break; 75 76 case ETNAVIV_PARAM_GPU_FEATURES_6: 77 *value = gpu->identity.minor_features5; 78 break; 79 80 case ETNAVIV_PARAM_GPU_FEATURES_7: 81 *value = gpu->identity.minor_features6; 82 break; 83 84 case ETNAVIV_PARAM_GPU_FEATURES_8: 85 *value = gpu->identity.minor_features7; 86 break; 87 88 case ETNAVIV_PARAM_GPU_FEATURES_9: 89 *value = gpu->identity.minor_features8; 90 break; 91 92 case ETNAVIV_PARAM_GPU_FEATURES_10: 93 *value = gpu->identity.minor_features9; 94 break; 95 96 case ETNAVIV_PARAM_GPU_FEATURES_11: 97 *value = gpu->identity.minor_features10; 98 break; 99 100 case ETNAVIV_PARAM_GPU_FEATURES_12: 101 *value = gpu->identity.minor_features11; 102 break; 103 104 case ETNAVIV_PARAM_GPU_STREAM_COUNT: 105 *value = gpu->identity.stream_count; 106 break; 107 108 case ETNAVIV_PARAM_GPU_REGISTER_MAX: 109 *value = gpu->identity.register_max; 110 break; 111 112 case ETNAVIV_PARAM_GPU_THREAD_COUNT: 113 *value = gpu->identity.thread_count; 114 break; 115 116 case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE: 117 *value = gpu->identity.vertex_cache_size; 118 break; 119 120 case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT: 121 *value = gpu->identity.shader_core_count; 122 break; 123 124 case ETNAVIV_PARAM_GPU_PIXEL_PIPES: 125 *value = gpu->identity.pixel_pipes; 126 break; 127 128 case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE: 129 *value = gpu->identity.vertex_output_buffer_size; 130 break; 131 132 case ETNAVIV_PARAM_GPU_BUFFER_SIZE: 133 *value = gpu->identity.buffer_size; 134 break; 135 136 case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT: 137 *value = gpu->identity.instruction_count; 138 break; 139 140 case ETNAVIV_PARAM_GPU_NUM_CONSTANTS: 141 *value = gpu->identity.num_constants; 142 break; 143 144 case ETNAVIV_PARAM_GPU_NUM_VARYINGS: 145 *value = gpu->identity.varyings_count; 146 break; 147 148 case ETNAVIV_PARAM_SOFTPIN_START_ADDR: 149 if (priv->mmu_global->version == ETNAVIV_IOMMU_V2) 150 *value = ETNAVIV_SOFTPIN_START_ADDRESS; 151 else 152 *value = ~0ULL; 153 break; 154 155 case ETNAVIV_PARAM_GPU_PRODUCT_ID: 156 *value = gpu->identity.product_id; 157 break; 158 159 case ETNAVIV_PARAM_GPU_CUSTOMER_ID: 160 *value = gpu->identity.customer_id; 161 break; 162 163 case ETNAVIV_PARAM_GPU_ECO_ID: 164 *value = gpu->identity.eco_id; 165 break; 166 167 default: 168 DBG("%s: invalid param: %u", dev_name(gpu->dev), param); 169 return -EINVAL; 170 } 171 172 return 0; 173 } 174 175 static inline bool etnaviv_is_model_rev(struct etnaviv_gpu *gpu, u32 model, u32 revision) 176 { 177 return gpu->identity.model == model && 178 gpu->identity.revision == revision; 179 } 180 181 #define etnaviv_field(val, field) \ 182 (((val) & field##__MASK) >> field##__SHIFT) 183 184 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu) 185 { 186 if (gpu->identity.minor_features0 & 187 chipMinorFeatures0_MORE_MINOR_FEATURES) { 188 u32 specs[4]; 189 unsigned int streams; 190 191 specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS); 192 specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2); 193 specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3); 194 specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4); 195 196 gpu->identity.stream_count = etnaviv_field(specs[0], 197 VIVS_HI_CHIP_SPECS_STREAM_COUNT); 198 gpu->identity.register_max = etnaviv_field(specs[0], 199 VIVS_HI_CHIP_SPECS_REGISTER_MAX); 200 gpu->identity.thread_count = etnaviv_field(specs[0], 201 VIVS_HI_CHIP_SPECS_THREAD_COUNT); 202 gpu->identity.vertex_cache_size = etnaviv_field(specs[0], 203 VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE); 204 gpu->identity.shader_core_count = etnaviv_field(specs[0], 205 VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT); 206 gpu->identity.pixel_pipes = etnaviv_field(specs[0], 207 VIVS_HI_CHIP_SPECS_PIXEL_PIPES); 208 gpu->identity.vertex_output_buffer_size = 209 etnaviv_field(specs[0], 210 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE); 211 212 gpu->identity.buffer_size = etnaviv_field(specs[1], 213 VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE); 214 gpu->identity.instruction_count = etnaviv_field(specs[1], 215 VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT); 216 gpu->identity.num_constants = etnaviv_field(specs[1], 217 VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS); 218 219 gpu->identity.varyings_count = etnaviv_field(specs[2], 220 VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT); 221 222 /* This overrides the value from older register if non-zero */ 223 streams = etnaviv_field(specs[3], 224 VIVS_HI_CHIP_SPECS_4_STREAM_COUNT); 225 if (streams) 226 gpu->identity.stream_count = streams; 227 } 228 229 /* Fill in the stream count if not specified */ 230 if (gpu->identity.stream_count == 0) { 231 if (gpu->identity.model >= 0x1000) 232 gpu->identity.stream_count = 4; 233 else 234 gpu->identity.stream_count = 1; 235 } 236 237 /* Convert the register max value */ 238 if (gpu->identity.register_max) 239 gpu->identity.register_max = 1 << gpu->identity.register_max; 240 else if (gpu->identity.model == chipModel_GC400) 241 gpu->identity.register_max = 32; 242 else 243 gpu->identity.register_max = 64; 244 245 /* Convert thread count */ 246 if (gpu->identity.thread_count) 247 gpu->identity.thread_count = 1 << gpu->identity.thread_count; 248 else if (gpu->identity.model == chipModel_GC400) 249 gpu->identity.thread_count = 64; 250 else if (gpu->identity.model == chipModel_GC500 || 251 gpu->identity.model == chipModel_GC530) 252 gpu->identity.thread_count = 128; 253 else 254 gpu->identity.thread_count = 256; 255 256 if (gpu->identity.vertex_cache_size == 0) 257 gpu->identity.vertex_cache_size = 8; 258 259 if (gpu->identity.shader_core_count == 0) { 260 if (gpu->identity.model >= 0x1000) 261 gpu->identity.shader_core_count = 2; 262 else 263 gpu->identity.shader_core_count = 1; 264 } 265 266 if (gpu->identity.pixel_pipes == 0) 267 gpu->identity.pixel_pipes = 1; 268 269 /* Convert virtex buffer size */ 270 if (gpu->identity.vertex_output_buffer_size) { 271 gpu->identity.vertex_output_buffer_size = 272 1 << gpu->identity.vertex_output_buffer_size; 273 } else if (gpu->identity.model == chipModel_GC400) { 274 if (gpu->identity.revision < 0x4000) 275 gpu->identity.vertex_output_buffer_size = 512; 276 else if (gpu->identity.revision < 0x4200) 277 gpu->identity.vertex_output_buffer_size = 256; 278 else 279 gpu->identity.vertex_output_buffer_size = 128; 280 } else { 281 gpu->identity.vertex_output_buffer_size = 512; 282 } 283 284 switch (gpu->identity.instruction_count) { 285 case 0: 286 if (etnaviv_is_model_rev(gpu, 0x2000, 0x5108) || 287 gpu->identity.model == chipModel_GC880) 288 gpu->identity.instruction_count = 512; 289 else 290 gpu->identity.instruction_count = 256; 291 break; 292 293 case 1: 294 gpu->identity.instruction_count = 1024; 295 break; 296 297 case 2: 298 gpu->identity.instruction_count = 2048; 299 break; 300 301 default: 302 gpu->identity.instruction_count = 256; 303 break; 304 } 305 306 if (gpu->identity.num_constants == 0) 307 gpu->identity.num_constants = 168; 308 309 if (gpu->identity.varyings_count == 0) { 310 if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0) 311 gpu->identity.varyings_count = 12; 312 else 313 gpu->identity.varyings_count = 8; 314 } 315 316 /* 317 * For some cores, two varyings are consumed for position, so the 318 * maximum varying count needs to be reduced by one. 319 */ 320 if (etnaviv_is_model_rev(gpu, 0x5000, 0x5434) || 321 etnaviv_is_model_rev(gpu, 0x4000, 0x5222) || 322 etnaviv_is_model_rev(gpu, 0x4000, 0x5245) || 323 etnaviv_is_model_rev(gpu, 0x4000, 0x5208) || 324 etnaviv_is_model_rev(gpu, 0x3000, 0x5435) || 325 etnaviv_is_model_rev(gpu, 0x2200, 0x5244) || 326 etnaviv_is_model_rev(gpu, 0x2100, 0x5108) || 327 etnaviv_is_model_rev(gpu, 0x2000, 0x5108) || 328 etnaviv_is_model_rev(gpu, 0x1500, 0x5246) || 329 etnaviv_is_model_rev(gpu, 0x880, 0x5107) || 330 etnaviv_is_model_rev(gpu, 0x880, 0x5106)) 331 gpu->identity.varyings_count -= 1; 332 } 333 334 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu) 335 { 336 u32 chipIdentity; 337 338 chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY); 339 340 /* Special case for older graphic cores. */ 341 if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) { 342 gpu->identity.model = chipModel_GC500; 343 gpu->identity.revision = etnaviv_field(chipIdentity, 344 VIVS_HI_CHIP_IDENTITY_REVISION); 345 } else { 346 u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE); 347 348 gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL); 349 gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV); 350 gpu->identity.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID); 351 352 /* 353 * Reading these two registers on GC600 rev 0x19 result in a 354 * unhandled fault: external abort on non-linefetch 355 */ 356 if (!etnaviv_is_model_rev(gpu, 0x600, 0x19)) { 357 gpu->identity.product_id = gpu_read(gpu, VIVS_HI_CHIP_PRODUCT_ID); 358 gpu->identity.eco_id = gpu_read(gpu, VIVS_HI_CHIP_ECO_ID); 359 } 360 361 /* 362 * !!!! HACK ALERT !!!! 363 * Because people change device IDs without letting software 364 * know about it - here is the hack to make it all look the 365 * same. Only for GC400 family. 366 */ 367 if ((gpu->identity.model & 0xff00) == 0x0400 && 368 gpu->identity.model != chipModel_GC420) { 369 gpu->identity.model = gpu->identity.model & 0x0400; 370 } 371 372 /* Another special case */ 373 if (etnaviv_is_model_rev(gpu, 0x300, 0x2201)) { 374 u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME); 375 376 if (chipDate == 0x20080814 && chipTime == 0x12051100) { 377 /* 378 * This IP has an ECO; put the correct 379 * revision in it. 380 */ 381 gpu->identity.revision = 0x1051; 382 } 383 } 384 385 /* 386 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in 387 * reality it's just a re-branded GC3000. We can identify this 388 * core by the upper half of the revision register being all 1. 389 * Fix model/rev here, so all other places can refer to this 390 * core by its real identity. 391 */ 392 if (etnaviv_is_model_rev(gpu, 0x2000, 0xffff5450)) { 393 gpu->identity.model = chipModel_GC3000; 394 gpu->identity.revision &= 0xffff; 395 } 396 397 if (etnaviv_is_model_rev(gpu, 0x1000, 0x5037) && (chipDate == 0x20120617)) 398 gpu->identity.eco_id = 1; 399 400 if (etnaviv_is_model_rev(gpu, 0x320, 0x5303) && (chipDate == 0x20140511)) 401 gpu->identity.eco_id = 1; 402 } 403 404 dev_info(gpu->dev, "model: GC%x, revision: %x\n", 405 gpu->identity.model, gpu->identity.revision); 406 407 gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP; 408 /* 409 * If there is a match in the HWDB, we aren't interested in the 410 * remaining register values, as they might be wrong. 411 */ 412 if (etnaviv_fill_identity_from_hwdb(gpu)) 413 return; 414 415 gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE); 416 417 /* Disable fast clear on GC700. */ 418 if (gpu->identity.model == chipModel_GC700) 419 gpu->identity.features &= ~chipFeatures_FAST_CLEAR; 420 421 /* These models/revisions don't have the 2D pipe bit */ 422 if ((gpu->identity.model == chipModel_GC500 && 423 gpu->identity.revision <= 2) || 424 gpu->identity.model == chipModel_GC300) 425 gpu->identity.features |= chipFeatures_PIPE_2D; 426 427 if ((gpu->identity.model == chipModel_GC500 && 428 gpu->identity.revision < 2) || 429 (gpu->identity.model == chipModel_GC300 && 430 gpu->identity.revision < 0x2000)) { 431 432 /* 433 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these 434 * registers. 435 */ 436 gpu->identity.minor_features0 = 0; 437 gpu->identity.minor_features1 = 0; 438 gpu->identity.minor_features2 = 0; 439 gpu->identity.minor_features3 = 0; 440 gpu->identity.minor_features4 = 0; 441 gpu->identity.minor_features5 = 0; 442 } else 443 gpu->identity.minor_features0 = 444 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0); 445 446 if (gpu->identity.minor_features0 & 447 chipMinorFeatures0_MORE_MINOR_FEATURES) { 448 gpu->identity.minor_features1 = 449 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1); 450 gpu->identity.minor_features2 = 451 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2); 452 gpu->identity.minor_features3 = 453 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3); 454 gpu->identity.minor_features4 = 455 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4); 456 gpu->identity.minor_features5 = 457 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5); 458 } 459 460 /* GC600/300 idle register reports zero bits where modules aren't present */ 461 if (gpu->identity.model == chipModel_GC600 || 462 gpu->identity.model == chipModel_GC300) 463 gpu->idle_mask = VIVS_HI_IDLE_STATE_TX | 464 VIVS_HI_IDLE_STATE_RA | 465 VIVS_HI_IDLE_STATE_SE | 466 VIVS_HI_IDLE_STATE_PA | 467 VIVS_HI_IDLE_STATE_SH | 468 VIVS_HI_IDLE_STATE_PE | 469 VIVS_HI_IDLE_STATE_DE | 470 VIVS_HI_IDLE_STATE_FE; 471 472 etnaviv_hw_specs(gpu); 473 } 474 475 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock) 476 { 477 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock | 478 VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD); 479 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock); 480 } 481 482 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu) 483 { 484 if (gpu->identity.minor_features2 & 485 chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) { 486 clk_set_rate(gpu->clk_core, 487 gpu->base_rate_core >> gpu->freq_scale); 488 clk_set_rate(gpu->clk_shader, 489 gpu->base_rate_shader >> gpu->freq_scale); 490 } else { 491 unsigned int fscale = 1 << (6 - gpu->freq_scale); 492 u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); 493 494 clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK; 495 clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale); 496 etnaviv_gpu_load_clock(gpu, clock); 497 } 498 499 /* 500 * Choose number of wait cycles to target a ~30us (1/32768) max latency 501 * until new work is picked up by the FE when it polls in the idle loop. 502 * If the GPU base frequency is unknown use 200 wait cycles. 503 */ 504 gpu->fe_waitcycles = clamp(gpu->base_rate_core >> (15 - gpu->freq_scale), 505 200UL, 0xffffUL); 506 } 507 508 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu) 509 { 510 u32 control, idle; 511 unsigned long timeout; 512 bool failed = true; 513 514 /* We hope that the GPU resets in under one second */ 515 timeout = jiffies + msecs_to_jiffies(1000); 516 517 while (time_is_after_jiffies(timeout)) { 518 unsigned int fscale = 1 << (6 - gpu->freq_scale); 519 u32 pulse_eater = 0x01590880; 520 521 /* disable clock gating */ 522 gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, 0x0); 523 524 /* disable pulse eater */ 525 pulse_eater |= BIT(17); 526 gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater); 527 pulse_eater |= BIT(0); 528 gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater); 529 530 /* enable clock */ 531 control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale); 532 etnaviv_gpu_load_clock(gpu, control); 533 534 /* isolate the GPU. */ 535 control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU; 536 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 537 538 if (gpu->sec_mode == ETNA_SEC_KERNEL) { 539 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, 540 VIVS_MMUv2_AHB_CONTROL_RESET); 541 } else { 542 /* set soft reset. */ 543 control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET; 544 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 545 } 546 547 /* wait for reset. */ 548 usleep_range(10, 20); 549 550 /* reset soft reset bit. */ 551 control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET; 552 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 553 554 /* reset GPU isolation. */ 555 control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU; 556 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 557 558 /* read idle register. */ 559 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); 560 561 /* try resetting again if FE is not idle */ 562 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) { 563 dev_dbg(gpu->dev, "FE is not idle\n"); 564 continue; 565 } 566 567 /* read reset register. */ 568 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); 569 570 /* is the GPU idle? */ 571 if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) || 572 ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) { 573 dev_dbg(gpu->dev, "GPU is not idle\n"); 574 continue; 575 } 576 577 /* enable debug register access */ 578 control &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS; 579 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 580 581 failed = false; 582 break; 583 } 584 585 if (failed) { 586 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); 587 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); 588 589 dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n", 590 idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ", 591 control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ", 592 control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not "); 593 594 return -EBUSY; 595 } 596 597 /* We rely on the GPU running, so program the clock */ 598 etnaviv_gpu_update_clock(gpu); 599 600 gpu->state = ETNA_GPU_STATE_RESET; 601 gpu->exec_state = -1; 602 if (gpu->mmu_context) 603 etnaviv_iommu_context_put(gpu->mmu_context); 604 gpu->mmu_context = NULL; 605 606 return 0; 607 } 608 609 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu) 610 { 611 u32 pmc, ppc; 612 613 /* enable clock gating */ 614 ppc = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); 615 ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; 616 617 /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */ 618 if (gpu->identity.revision == 0x4301 || 619 gpu->identity.revision == 0x4302) 620 ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING; 621 622 gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, ppc); 623 624 pmc = gpu_read_power(gpu, VIVS_PM_MODULE_CONTROLS); 625 626 /* Disable PA clock gating for GC400+ without bugfix except for GC420 */ 627 if (gpu->identity.model >= chipModel_GC400 && 628 gpu->identity.model != chipModel_GC420 && 629 !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12)) 630 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA; 631 632 /* 633 * Disable PE clock gating on revs < 5.0.0.0 when HZ is 634 * present without a bug fix. 635 */ 636 if (gpu->identity.revision < 0x5000 && 637 gpu->identity.minor_features0 & chipMinorFeatures0_HZ && 638 !(gpu->identity.minor_features1 & 639 chipMinorFeatures1_DISABLE_PE_GATING)) 640 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE; 641 642 if (gpu->identity.revision < 0x5422) 643 pmc |= BIT(15); /* Unknown bit */ 644 645 /* Disable TX clock gating on affected core revisions. */ 646 if (etnaviv_is_model_rev(gpu, 0x4000, 0x5222) || 647 etnaviv_is_model_rev(gpu, 0x2000, 0x5108) || 648 etnaviv_is_model_rev(gpu, 0x7000, 0x6202) || 649 etnaviv_is_model_rev(gpu, 0x7000, 0x6203)) 650 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX; 651 652 /* Disable SE and RA clock gating on affected core revisions. */ 653 if (etnaviv_is_model_rev(gpu, 0x7000, 0x6202)) 654 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SE | 655 VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA; 656 657 /* Disable SH_EU clock gating on affected core revisions. */ 658 if (etnaviv_is_model_rev(gpu, 0x8000, 0x7200) || 659 etnaviv_is_model_rev(gpu, 0x8000, 0x8002) || 660 etnaviv_is_model_rev(gpu, 0x9200, 0x6304)) 661 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SH_EU; 662 663 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ; 664 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ; 665 666 gpu_write_power(gpu, VIVS_PM_MODULE_CONTROLS, pmc); 667 } 668 669 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch) 670 { 671 gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address); 672 gpu_write(gpu, VIVS_FE_COMMAND_CONTROL, 673 VIVS_FE_COMMAND_CONTROL_ENABLE | 674 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch)); 675 676 if (gpu->sec_mode == ETNA_SEC_KERNEL) { 677 gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL, 678 VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE | 679 VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch)); 680 } 681 } 682 683 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu, 684 struct etnaviv_iommu_context *context) 685 { 686 u16 prefetch; 687 u32 address; 688 689 WARN_ON(gpu->state != ETNA_GPU_STATE_INITIALIZED); 690 691 /* setup the MMU */ 692 etnaviv_iommu_restore(gpu, context); 693 694 /* Start command processor */ 695 prefetch = etnaviv_buffer_init(gpu); 696 address = etnaviv_cmdbuf_get_va(&gpu->buffer, 697 &gpu->mmu_context->cmdbuf_mapping); 698 699 etnaviv_gpu_start_fe(gpu, address, prefetch); 700 701 gpu->state = ETNA_GPU_STATE_RUNNING; 702 } 703 704 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu) 705 { 706 /* 707 * Base value for VIVS_PM_PULSE_EATER register on models where it 708 * cannot be read, extracted from vivante kernel driver. 709 */ 710 u32 pulse_eater = 0x01590880; 711 712 if (etnaviv_is_model_rev(gpu, 0x4000, 0x5208) || 713 etnaviv_is_model_rev(gpu, 0x4000, 0x5222)) { 714 pulse_eater |= BIT(23); 715 716 } 717 718 if (etnaviv_is_model_rev(gpu, 0x1000, 0x5039) || 719 etnaviv_is_model_rev(gpu, 0x1000, 0x5040)) { 720 pulse_eater &= ~BIT(16); 721 pulse_eater |= BIT(17); 722 } 723 724 if ((gpu->identity.revision > 0x5420) && 725 (gpu->identity.features & chipFeatures_PIPE_3D)) 726 { 727 /* Performance fix: disable internal DFS */ 728 pulse_eater = gpu_read_power(gpu, VIVS_PM_PULSE_EATER); 729 pulse_eater |= BIT(18); 730 } 731 732 gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater); 733 } 734 735 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu) 736 { 737 WARN_ON(!(gpu->state == ETNA_GPU_STATE_IDENTIFIED || 738 gpu->state == ETNA_GPU_STATE_RESET)); 739 740 if ((etnaviv_is_model_rev(gpu, 0x320, 0x5007) || 741 etnaviv_is_model_rev(gpu, 0x320, 0x5220)) && 742 gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) { 743 u32 mc_memory_debug; 744 745 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff; 746 747 if (gpu->identity.revision == 0x5007) 748 mc_memory_debug |= 0x0c; 749 else 750 mc_memory_debug |= 0x08; 751 752 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug); 753 } 754 755 /* enable module-level clock gating */ 756 etnaviv_gpu_enable_mlcg(gpu); 757 758 /* 759 * Update GPU AXI cache atttribute to "cacheable, no allocate". 760 * This is necessary to prevent the iMX6 SoC locking up. 761 */ 762 gpu_write(gpu, VIVS_HI_AXI_CONFIG, 763 VIVS_HI_AXI_CONFIG_AWCACHE(2) | 764 VIVS_HI_AXI_CONFIG_ARCACHE(2)); 765 766 /* GC2000 rev 5108 needs a special bus config */ 767 if (etnaviv_is_model_rev(gpu, 0x2000, 0x5108)) { 768 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG); 769 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK | 770 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK); 771 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) | 772 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0); 773 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config); 774 } 775 776 if (gpu->sec_mode == ETNA_SEC_KERNEL) { 777 u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL); 778 val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS; 779 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val); 780 } 781 782 /* setup the pulse eater */ 783 etnaviv_gpu_setup_pulse_eater(gpu); 784 785 gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U); 786 787 gpu->state = ETNA_GPU_STATE_INITIALIZED; 788 } 789 790 int etnaviv_gpu_init(struct etnaviv_gpu *gpu) 791 { 792 struct etnaviv_drm_private *priv = gpu->drm->dev_private; 793 dma_addr_t cmdbuf_paddr; 794 int ret, i; 795 796 ret = pm_runtime_get_sync(gpu->dev); 797 if (ret < 0) { 798 dev_err(gpu->dev, "Failed to enable GPU power domain\n"); 799 goto pm_put; 800 } 801 802 etnaviv_hw_identify(gpu); 803 804 if (gpu->identity.model == 0) { 805 dev_err(gpu->dev, "Unknown GPU model\n"); 806 ret = -ENXIO; 807 goto fail; 808 } 809 810 if (gpu->identity.nn_core_count > 0) 811 dev_warn(gpu->dev, "etnaviv has been instantiated on a NPU, " 812 "for which the UAPI is still experimental\n"); 813 814 /* Exclude VG cores with FE2.0 */ 815 if (gpu->identity.features & chipFeatures_PIPE_VG && 816 gpu->identity.features & chipFeatures_FE20) { 817 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n"); 818 ret = -ENXIO; 819 goto fail; 820 } 821 822 /* 823 * On cores with security features supported, we claim control over the 824 * security states. 825 */ 826 if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) && 827 (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB)) 828 gpu->sec_mode = ETNA_SEC_KERNEL; 829 830 gpu->state = ETNA_GPU_STATE_IDENTIFIED; 831 832 ret = etnaviv_hw_reset(gpu); 833 if (ret) { 834 dev_err(gpu->dev, "GPU reset failed\n"); 835 goto fail; 836 } 837 838 ret = etnaviv_iommu_global_init(gpu); 839 if (ret) 840 goto fail; 841 842 /* Create buffer: */ 843 ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer, SZ_4K); 844 if (ret) { 845 dev_err(gpu->dev, "could not create command buffer\n"); 846 goto fail; 847 } 848 849 /* 850 * Set the GPU linear window to cover the cmdbuf region, as the GPU 851 * won't be able to start execution otherwise. The alignment to 128M is 852 * chosen arbitrarily but helps in debugging, as the MMU offset 853 * calculations are much more straight forward this way. 854 * 855 * On MC1.0 cores the linear window offset is ignored by the TS engine, 856 * leading to inconsistent memory views. Avoid using the offset on those 857 * cores if possible, otherwise disable the TS feature. MMUv2 doesn't 858 * expose this issue, as all TS accesses are MMU translated, so the 859 * linear window offset won't be used. 860 */ 861 cmdbuf_paddr = ALIGN_DOWN(etnaviv_cmdbuf_get_pa(&gpu->buffer), SZ_128M); 862 863 if (!(gpu->identity.features & chipFeatures_PIPE_3D) || 864 (gpu->identity.minor_features0 & chipMinorFeatures0_MC20) || 865 (gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION)) { 866 if (cmdbuf_paddr >= SZ_2G) 867 priv->mmu_global->memory_base = SZ_2G; 868 else 869 priv->mmu_global->memory_base = cmdbuf_paddr; 870 } else if (cmdbuf_paddr + SZ_128M >= SZ_2G) { 871 dev_info(gpu->dev, 872 "Need to move linear window on MC1.0, disabling TS\n"); 873 gpu->identity.features &= ~chipFeatures_FAST_CLEAR; 874 priv->mmu_global->memory_base = SZ_2G; 875 } 876 877 /* Setup event management */ 878 spin_lock_init(&gpu->event_spinlock); 879 init_completion(&gpu->event_free); 880 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS); 881 for (i = 0; i < ARRAY_SIZE(gpu->event); i++) 882 complete(&gpu->event_free); 883 884 /* Now program the hardware */ 885 mutex_lock(&gpu->lock); 886 etnaviv_gpu_hw_init(gpu); 887 mutex_unlock(&gpu->lock); 888 889 pm_runtime_mark_last_busy(gpu->dev); 890 pm_runtime_put_autosuspend(gpu->dev); 891 892 return 0; 893 894 fail: 895 pm_runtime_mark_last_busy(gpu->dev); 896 pm_put: 897 pm_runtime_put_autosuspend(gpu->dev); 898 899 return ret; 900 } 901 902 #ifdef CONFIG_DEBUG_FS 903 struct dma_debug { 904 u32 address[2]; 905 u32 state[2]; 906 }; 907 908 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug) 909 { 910 u32 i; 911 912 debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); 913 debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE); 914 915 for (i = 0; i < 500; i++) { 916 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); 917 debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE); 918 919 if (debug->address[0] != debug->address[1]) 920 break; 921 922 if (debug->state[0] != debug->state[1]) 923 break; 924 } 925 } 926 927 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m) 928 { 929 struct dma_debug debug; 930 u32 dma_lo, dma_hi, axi, idle; 931 int ret; 932 933 seq_printf(m, "%s Status:\n", dev_name(gpu->dev)); 934 935 ret = pm_runtime_get_sync(gpu->dev); 936 if (ret < 0) 937 goto pm_put; 938 939 dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW); 940 dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH); 941 axi = gpu_read(gpu, VIVS_HI_AXI_STATUS); 942 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); 943 944 verify_dma(gpu, &debug); 945 946 seq_puts(m, "\tidentity\n"); 947 seq_printf(m, "\t model: 0x%x\n", gpu->identity.model); 948 seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision); 949 seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id); 950 seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id); 951 seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id); 952 953 seq_puts(m, "\tfeatures\n"); 954 seq_printf(m, "\t major_features: 0x%08x\n", 955 gpu->identity.features); 956 seq_printf(m, "\t minor_features0: 0x%08x\n", 957 gpu->identity.minor_features0); 958 seq_printf(m, "\t minor_features1: 0x%08x\n", 959 gpu->identity.minor_features1); 960 seq_printf(m, "\t minor_features2: 0x%08x\n", 961 gpu->identity.minor_features2); 962 seq_printf(m, "\t minor_features3: 0x%08x\n", 963 gpu->identity.minor_features3); 964 seq_printf(m, "\t minor_features4: 0x%08x\n", 965 gpu->identity.minor_features4); 966 seq_printf(m, "\t minor_features5: 0x%08x\n", 967 gpu->identity.minor_features5); 968 seq_printf(m, "\t minor_features6: 0x%08x\n", 969 gpu->identity.minor_features6); 970 seq_printf(m, "\t minor_features7: 0x%08x\n", 971 gpu->identity.minor_features7); 972 seq_printf(m, "\t minor_features8: 0x%08x\n", 973 gpu->identity.minor_features8); 974 seq_printf(m, "\t minor_features9: 0x%08x\n", 975 gpu->identity.minor_features9); 976 seq_printf(m, "\t minor_features10: 0x%08x\n", 977 gpu->identity.minor_features10); 978 seq_printf(m, "\t minor_features11: 0x%08x\n", 979 gpu->identity.minor_features11); 980 981 seq_puts(m, "\tspecs\n"); 982 seq_printf(m, "\t stream_count: %d\n", 983 gpu->identity.stream_count); 984 seq_printf(m, "\t register_max: %d\n", 985 gpu->identity.register_max); 986 seq_printf(m, "\t thread_count: %d\n", 987 gpu->identity.thread_count); 988 seq_printf(m, "\t vertex_cache_size: %d\n", 989 gpu->identity.vertex_cache_size); 990 seq_printf(m, "\t shader_core_count: %d\n", 991 gpu->identity.shader_core_count); 992 seq_printf(m, "\t nn_core_count: %d\n", 993 gpu->identity.nn_core_count); 994 seq_printf(m, "\t pixel_pipes: %d\n", 995 gpu->identity.pixel_pipes); 996 seq_printf(m, "\t vertex_output_buffer_size: %d\n", 997 gpu->identity.vertex_output_buffer_size); 998 seq_printf(m, "\t buffer_size: %d\n", 999 gpu->identity.buffer_size); 1000 seq_printf(m, "\t instruction_count: %d\n", 1001 gpu->identity.instruction_count); 1002 seq_printf(m, "\t num_constants: %d\n", 1003 gpu->identity.num_constants); 1004 seq_printf(m, "\t varyings_count: %d\n", 1005 gpu->identity.varyings_count); 1006 1007 seq_printf(m, "\taxi: 0x%08x\n", axi); 1008 seq_printf(m, "\tidle: 0x%08x\n", idle); 1009 idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP; 1010 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) 1011 seq_puts(m, "\t FE is not idle\n"); 1012 if ((idle & VIVS_HI_IDLE_STATE_DE) == 0) 1013 seq_puts(m, "\t DE is not idle\n"); 1014 if ((idle & VIVS_HI_IDLE_STATE_PE) == 0) 1015 seq_puts(m, "\t PE is not idle\n"); 1016 if ((idle & VIVS_HI_IDLE_STATE_SH) == 0) 1017 seq_puts(m, "\t SH is not idle\n"); 1018 if ((idle & VIVS_HI_IDLE_STATE_PA) == 0) 1019 seq_puts(m, "\t PA is not idle\n"); 1020 if ((idle & VIVS_HI_IDLE_STATE_SE) == 0) 1021 seq_puts(m, "\t SE is not idle\n"); 1022 if ((idle & VIVS_HI_IDLE_STATE_RA) == 0) 1023 seq_puts(m, "\t RA is not idle\n"); 1024 if ((idle & VIVS_HI_IDLE_STATE_TX) == 0) 1025 seq_puts(m, "\t TX is not idle\n"); 1026 if ((idle & VIVS_HI_IDLE_STATE_VG) == 0) 1027 seq_puts(m, "\t VG is not idle\n"); 1028 if ((idle & VIVS_HI_IDLE_STATE_IM) == 0) 1029 seq_puts(m, "\t IM is not idle\n"); 1030 if ((idle & VIVS_HI_IDLE_STATE_FP) == 0) 1031 seq_puts(m, "\t FP is not idle\n"); 1032 if ((idle & VIVS_HI_IDLE_STATE_TS) == 0) 1033 seq_puts(m, "\t TS is not idle\n"); 1034 if ((idle & VIVS_HI_IDLE_STATE_BL) == 0) 1035 seq_puts(m, "\t BL is not idle\n"); 1036 if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0) 1037 seq_puts(m, "\t ASYNCFE is not idle\n"); 1038 if ((idle & VIVS_HI_IDLE_STATE_MC) == 0) 1039 seq_puts(m, "\t MC is not idle\n"); 1040 if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0) 1041 seq_puts(m, "\t PPA is not idle\n"); 1042 if ((idle & VIVS_HI_IDLE_STATE_WD) == 0) 1043 seq_puts(m, "\t WD is not idle\n"); 1044 if ((idle & VIVS_HI_IDLE_STATE_NN) == 0) 1045 seq_puts(m, "\t NN is not idle\n"); 1046 if ((idle & VIVS_HI_IDLE_STATE_TP) == 0) 1047 seq_puts(m, "\t TP is not idle\n"); 1048 if (idle & VIVS_HI_IDLE_STATE_AXI_LP) 1049 seq_puts(m, "\t AXI low power mode\n"); 1050 1051 if (gpu->identity.features & chipFeatures_DEBUG_MODE) { 1052 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0); 1053 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1); 1054 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE); 1055 1056 seq_puts(m, "\tMC\n"); 1057 seq_printf(m, "\t read0: 0x%08x\n", read0); 1058 seq_printf(m, "\t read1: 0x%08x\n", read1); 1059 seq_printf(m, "\t write: 0x%08x\n", write); 1060 } 1061 1062 seq_puts(m, "\tDMA "); 1063 1064 if (debug.address[0] == debug.address[1] && 1065 debug.state[0] == debug.state[1]) { 1066 seq_puts(m, "seems to be stuck\n"); 1067 } else if (debug.address[0] == debug.address[1]) { 1068 seq_puts(m, "address is constant\n"); 1069 } else { 1070 seq_puts(m, "is running\n"); 1071 } 1072 1073 seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]); 1074 seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]); 1075 seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]); 1076 seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]); 1077 seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n", 1078 dma_lo, dma_hi); 1079 1080 ret = 0; 1081 1082 pm_runtime_mark_last_busy(gpu->dev); 1083 pm_put: 1084 pm_runtime_put_autosuspend(gpu->dev); 1085 1086 return ret; 1087 } 1088 #endif 1089 1090 /* fence object management */ 1091 struct etnaviv_fence { 1092 struct etnaviv_gpu *gpu; 1093 struct dma_fence base; 1094 }; 1095 1096 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence) 1097 { 1098 return container_of(fence, struct etnaviv_fence, base); 1099 } 1100 1101 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence) 1102 { 1103 return "etnaviv"; 1104 } 1105 1106 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence) 1107 { 1108 struct etnaviv_fence *f = to_etnaviv_fence(fence); 1109 1110 return dev_name(f->gpu->dev); 1111 } 1112 1113 static bool etnaviv_fence_signaled(struct dma_fence *fence) 1114 { 1115 struct etnaviv_fence *f = to_etnaviv_fence(fence); 1116 1117 return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0; 1118 } 1119 1120 static void etnaviv_fence_release(struct dma_fence *fence) 1121 { 1122 struct etnaviv_fence *f = to_etnaviv_fence(fence); 1123 1124 kfree_rcu(f, base.rcu); 1125 } 1126 1127 static const struct dma_fence_ops etnaviv_fence_ops = { 1128 .get_driver_name = etnaviv_fence_get_driver_name, 1129 .get_timeline_name = etnaviv_fence_get_timeline_name, 1130 .signaled = etnaviv_fence_signaled, 1131 .release = etnaviv_fence_release, 1132 }; 1133 1134 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu) 1135 { 1136 struct etnaviv_fence *f; 1137 1138 /* 1139 * GPU lock must already be held, otherwise fence completion order might 1140 * not match the seqno order assigned here. 1141 */ 1142 lockdep_assert_held(&gpu->lock); 1143 1144 f = kzalloc(sizeof(*f), GFP_KERNEL); 1145 if (!f) 1146 return NULL; 1147 1148 f->gpu = gpu; 1149 1150 dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock, 1151 gpu->fence_context, ++gpu->next_fence); 1152 1153 return &f->base; 1154 } 1155 1156 /* returns true if fence a comes after fence b */ 1157 static inline bool fence_after(u32 a, u32 b) 1158 { 1159 return (s32)(a - b) > 0; 1160 } 1161 1162 /* 1163 * event management: 1164 */ 1165 1166 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events, 1167 unsigned int *events) 1168 { 1169 unsigned long timeout = msecs_to_jiffies(10 * 10000); 1170 unsigned i, acquired = 0, rpm_count = 0; 1171 int ret; 1172 1173 for (i = 0; i < nr_events; i++) { 1174 unsigned long remaining; 1175 1176 remaining = wait_for_completion_timeout(&gpu->event_free, timeout); 1177 1178 if (!remaining) { 1179 dev_err(gpu->dev, "wait_for_completion_timeout failed"); 1180 ret = -EBUSY; 1181 goto out; 1182 } 1183 1184 acquired++; 1185 timeout = remaining; 1186 } 1187 1188 spin_lock(&gpu->event_spinlock); 1189 1190 for (i = 0; i < nr_events; i++) { 1191 int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS); 1192 1193 events[i] = event; 1194 memset(&gpu->event[event], 0, sizeof(struct etnaviv_event)); 1195 set_bit(event, gpu->event_bitmap); 1196 } 1197 1198 spin_unlock(&gpu->event_spinlock); 1199 1200 for (i = 0; i < nr_events; i++) { 1201 ret = pm_runtime_resume_and_get(gpu->dev); 1202 if (ret) 1203 goto out_rpm; 1204 rpm_count++; 1205 } 1206 1207 return 0; 1208 1209 out_rpm: 1210 for (i = 0; i < rpm_count; i++) 1211 pm_runtime_put_autosuspend(gpu->dev); 1212 out: 1213 for (i = 0; i < acquired; i++) 1214 complete(&gpu->event_free); 1215 1216 return ret; 1217 } 1218 1219 static void event_free(struct etnaviv_gpu *gpu, unsigned int event) 1220 { 1221 if (!test_bit(event, gpu->event_bitmap)) { 1222 dev_warn(gpu->dev, "event %u is already marked as free", 1223 event); 1224 } else { 1225 clear_bit(event, gpu->event_bitmap); 1226 complete(&gpu->event_free); 1227 } 1228 1229 pm_runtime_put_autosuspend(gpu->dev); 1230 } 1231 1232 /* 1233 * Cmdstream submission/retirement: 1234 */ 1235 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu, 1236 u32 id, struct drm_etnaviv_timespec *timeout) 1237 { 1238 struct dma_fence *fence; 1239 int ret; 1240 1241 /* 1242 * Look up the fence and take a reference. We might still find a fence 1243 * whose refcount has already dropped to zero. dma_fence_get_rcu 1244 * pretends we didn't find a fence in that case. 1245 */ 1246 rcu_read_lock(); 1247 fence = xa_load(&gpu->user_fences, id); 1248 if (fence) 1249 fence = dma_fence_get_rcu(fence); 1250 rcu_read_unlock(); 1251 1252 if (!fence) 1253 return 0; 1254 1255 if (!timeout) { 1256 /* No timeout was requested: just test for completion */ 1257 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY; 1258 } else { 1259 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout); 1260 1261 ret = dma_fence_wait_timeout(fence, true, remaining); 1262 if (ret == 0) 1263 ret = -ETIMEDOUT; 1264 else if (ret != -ERESTARTSYS) 1265 ret = 0; 1266 1267 } 1268 1269 dma_fence_put(fence); 1270 return ret; 1271 } 1272 1273 /* 1274 * Wait for an object to become inactive. This, on it's own, is not race 1275 * free: the object is moved by the scheduler off the active list, and 1276 * then the iova is put. Moreover, the object could be re-submitted just 1277 * after we notice that it's become inactive. 1278 * 1279 * Although the retirement happens under the gpu lock, we don't want to hold 1280 * that lock in this function while waiting. 1281 */ 1282 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu, 1283 struct etnaviv_gem_object *etnaviv_obj, 1284 struct drm_etnaviv_timespec *timeout) 1285 { 1286 unsigned long remaining; 1287 long ret; 1288 1289 if (!timeout) 1290 return !is_active(etnaviv_obj) ? 0 : -EBUSY; 1291 1292 remaining = etnaviv_timeout_to_jiffies(timeout); 1293 1294 ret = wait_event_interruptible_timeout(gpu->fence_event, 1295 !is_active(etnaviv_obj), 1296 remaining); 1297 if (ret > 0) 1298 return 0; 1299 else if (ret == -ERESTARTSYS) 1300 return -ERESTARTSYS; 1301 else 1302 return -ETIMEDOUT; 1303 } 1304 1305 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu, 1306 struct etnaviv_event *event, unsigned int flags) 1307 { 1308 const struct etnaviv_gem_submit *submit = event->submit; 1309 unsigned int i; 1310 1311 for (i = 0; i < submit->nr_pmrs; i++) { 1312 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i; 1313 1314 if (pmr->flags == flags) 1315 etnaviv_perfmon_process(gpu, pmr, submit->exec_state); 1316 } 1317 } 1318 1319 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu, 1320 struct etnaviv_event *event) 1321 { 1322 u32 val; 1323 1324 mutex_lock(&gpu->lock); 1325 1326 /* disable clock gating */ 1327 val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); 1328 val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; 1329 gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val); 1330 1331 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE); 1332 1333 mutex_unlock(&gpu->lock); 1334 } 1335 1336 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu, 1337 struct etnaviv_event *event) 1338 { 1339 const struct etnaviv_gem_submit *submit = event->submit; 1340 unsigned int i; 1341 u32 val; 1342 1343 mutex_lock(&gpu->lock); 1344 1345 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST); 1346 1347 /* enable clock gating */ 1348 val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); 1349 val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; 1350 gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val); 1351 1352 mutex_unlock(&gpu->lock); 1353 1354 for (i = 0; i < submit->nr_pmrs; i++) { 1355 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i; 1356 1357 *pmr->bo_vma = pmr->sequence; 1358 } 1359 } 1360 1361 1362 /* add bo's to gpu's ring, and kick gpu: */ 1363 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit) 1364 { 1365 struct etnaviv_gpu *gpu = submit->gpu; 1366 struct dma_fence *gpu_fence; 1367 unsigned int i, nr_events = 1, event[3]; 1368 int ret; 1369 1370 /* 1371 * if there are performance monitor requests we need to have 1372 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE 1373 * requests. 1374 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests 1375 * and update the sequence number for userspace. 1376 */ 1377 if (submit->nr_pmrs) 1378 nr_events = 3; 1379 1380 ret = event_alloc(gpu, nr_events, event); 1381 if (ret) { 1382 DRM_ERROR("no free events\n"); 1383 pm_runtime_put_noidle(gpu->dev); 1384 return NULL; 1385 } 1386 1387 mutex_lock(&gpu->lock); 1388 1389 gpu_fence = etnaviv_gpu_fence_alloc(gpu); 1390 if (!gpu_fence) { 1391 for (i = 0; i < nr_events; i++) 1392 event_free(gpu, event[i]); 1393 1394 goto out_unlock; 1395 } 1396 1397 if (gpu->state == ETNA_GPU_STATE_INITIALIZED) 1398 etnaviv_gpu_start_fe_idleloop(gpu, submit->mmu_context); 1399 1400 if (submit->prev_mmu_context) 1401 etnaviv_iommu_context_put(submit->prev_mmu_context); 1402 submit->prev_mmu_context = etnaviv_iommu_context_get(gpu->mmu_context); 1403 1404 if (submit->nr_pmrs) { 1405 gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre; 1406 kref_get(&submit->refcount); 1407 gpu->event[event[1]].submit = submit; 1408 etnaviv_sync_point_queue(gpu, event[1]); 1409 } 1410 1411 gpu->event[event[0]].fence = gpu_fence; 1412 submit->cmdbuf.user_size = submit->cmdbuf.size - 8; 1413 etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context, 1414 event[0], &submit->cmdbuf); 1415 1416 if (submit->nr_pmrs) { 1417 gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post; 1418 kref_get(&submit->refcount); 1419 gpu->event[event[2]].submit = submit; 1420 etnaviv_sync_point_queue(gpu, event[2]); 1421 } 1422 1423 out_unlock: 1424 mutex_unlock(&gpu->lock); 1425 1426 return gpu_fence; 1427 } 1428 1429 static void sync_point_worker(struct work_struct *work) 1430 { 1431 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu, 1432 sync_point_work); 1433 struct etnaviv_event *event = &gpu->event[gpu->sync_point_event]; 1434 u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); 1435 1436 event->sync_point(gpu, event); 1437 etnaviv_submit_put(event->submit); 1438 event_free(gpu, gpu->sync_point_event); 1439 1440 /* restart FE last to avoid GPU and IRQ racing against this worker */ 1441 etnaviv_gpu_start_fe(gpu, addr + 2, 2); 1442 } 1443 1444 void etnaviv_gpu_recover_hang(struct etnaviv_gem_submit *submit) 1445 { 1446 struct etnaviv_gpu *gpu = submit->gpu; 1447 char *comm = NULL, *cmd = NULL; 1448 struct task_struct *task; 1449 unsigned int i; 1450 1451 dev_err(gpu->dev, "recover hung GPU!\n"); 1452 1453 task = get_pid_task(submit->pid, PIDTYPE_PID); 1454 if (task) { 1455 comm = kstrdup(task->comm, GFP_KERNEL); 1456 cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL); 1457 put_task_struct(task); 1458 } 1459 1460 if (comm && cmd) 1461 dev_err(gpu->dev, "offending task: %s (%s)\n", comm, cmd); 1462 1463 kfree(cmd); 1464 kfree(comm); 1465 1466 if (pm_runtime_get_sync(gpu->dev) < 0) 1467 goto pm_put; 1468 1469 mutex_lock(&gpu->lock); 1470 1471 etnaviv_hw_reset(gpu); 1472 1473 /* complete all events, the GPU won't do it after the reset */ 1474 spin_lock(&gpu->event_spinlock); 1475 for_each_set_bit(i, gpu->event_bitmap, ETNA_NR_EVENTS) 1476 event_free(gpu, i); 1477 spin_unlock(&gpu->event_spinlock); 1478 1479 etnaviv_gpu_hw_init(gpu); 1480 1481 mutex_unlock(&gpu->lock); 1482 pm_runtime_mark_last_busy(gpu->dev); 1483 pm_put: 1484 pm_runtime_put_autosuspend(gpu->dev); 1485 } 1486 1487 static void dump_mmu_fault(struct etnaviv_gpu *gpu) 1488 { 1489 static const char *fault_reasons[] = { 1490 "slave not present", 1491 "page not present", 1492 "write violation", 1493 "out of bounds", 1494 "read security violation", 1495 "write security violation", 1496 }; 1497 1498 u32 status_reg, status; 1499 int i; 1500 1501 if (gpu->sec_mode == ETNA_SEC_NONE) 1502 status_reg = VIVS_MMUv2_STATUS; 1503 else 1504 status_reg = VIVS_MMUv2_SEC_STATUS; 1505 1506 status = gpu_read(gpu, status_reg); 1507 dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status); 1508 1509 for (i = 0; i < 4; i++) { 1510 const char *reason = "unknown"; 1511 u32 address_reg; 1512 u32 mmu_status; 1513 1514 mmu_status = (status >> (i * 4)) & VIVS_MMUv2_STATUS_EXCEPTION0__MASK; 1515 if (!mmu_status) 1516 continue; 1517 1518 if ((mmu_status - 1) < ARRAY_SIZE(fault_reasons)) 1519 reason = fault_reasons[mmu_status - 1]; 1520 1521 if (gpu->sec_mode == ETNA_SEC_NONE) 1522 address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i); 1523 else 1524 address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR; 1525 1526 dev_err_ratelimited(gpu->dev, 1527 "MMU %d fault (%s) addr 0x%08x\n", 1528 i, reason, gpu_read(gpu, address_reg)); 1529 } 1530 } 1531 1532 static irqreturn_t irq_handler(int irq, void *data) 1533 { 1534 struct etnaviv_gpu *gpu = data; 1535 irqreturn_t ret = IRQ_NONE; 1536 1537 u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE); 1538 1539 if (intr != 0) { 1540 ktime_t now = ktime_get(); 1541 int event; 1542 1543 pm_runtime_mark_last_busy(gpu->dev); 1544 1545 dev_dbg(gpu->dev, "intr 0x%08x\n", intr); 1546 1547 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) { 1548 dev_err(gpu->dev, "AXI bus error\n"); 1549 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR; 1550 } 1551 1552 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) { 1553 dump_mmu_fault(gpu); 1554 gpu->state = ETNA_GPU_STATE_FAULT; 1555 drm_sched_fault(&gpu->sched); 1556 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION; 1557 } 1558 1559 while ((event = ffs(intr)) != 0) { 1560 struct dma_fence *fence; 1561 1562 event -= 1; 1563 1564 intr &= ~(1 << event); 1565 1566 dev_dbg(gpu->dev, "event %u\n", event); 1567 1568 if (gpu->event[event].sync_point) { 1569 gpu->sync_point_event = event; 1570 queue_work(gpu->wq, &gpu->sync_point_work); 1571 } 1572 1573 fence = gpu->event[event].fence; 1574 if (!fence) 1575 continue; 1576 1577 gpu->event[event].fence = NULL; 1578 1579 /* 1580 * Events can be processed out of order. Eg, 1581 * - allocate and queue event 0 1582 * - allocate event 1 1583 * - event 0 completes, we process it 1584 * - allocate and queue event 0 1585 * - event 1 and event 0 complete 1586 * we can end up processing event 0 first, then 1. 1587 */ 1588 if (fence_after(fence->seqno, gpu->completed_fence)) 1589 gpu->completed_fence = fence->seqno; 1590 dma_fence_signal_timestamp(fence, now); 1591 1592 event_free(gpu, event); 1593 } 1594 1595 ret = IRQ_HANDLED; 1596 } 1597 1598 return ret; 1599 } 1600 1601 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu) 1602 { 1603 int ret; 1604 1605 ret = clk_prepare_enable(gpu->clk_reg); 1606 if (ret) 1607 return ret; 1608 1609 ret = clk_prepare_enable(gpu->clk_bus); 1610 if (ret) 1611 goto disable_clk_reg; 1612 1613 ret = clk_prepare_enable(gpu->clk_core); 1614 if (ret) 1615 goto disable_clk_bus; 1616 1617 ret = clk_prepare_enable(gpu->clk_shader); 1618 if (ret) 1619 goto disable_clk_core; 1620 1621 return 0; 1622 1623 disable_clk_core: 1624 clk_disable_unprepare(gpu->clk_core); 1625 disable_clk_bus: 1626 clk_disable_unprepare(gpu->clk_bus); 1627 disable_clk_reg: 1628 clk_disable_unprepare(gpu->clk_reg); 1629 1630 return ret; 1631 } 1632 1633 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu) 1634 { 1635 clk_disable_unprepare(gpu->clk_shader); 1636 clk_disable_unprepare(gpu->clk_core); 1637 clk_disable_unprepare(gpu->clk_bus); 1638 clk_disable_unprepare(gpu->clk_reg); 1639 1640 return 0; 1641 } 1642 1643 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms) 1644 { 1645 unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms); 1646 1647 do { 1648 u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); 1649 1650 if ((idle & gpu->idle_mask) == gpu->idle_mask) 1651 return 0; 1652 1653 if (time_is_before_jiffies(timeout)) { 1654 dev_warn(gpu->dev, 1655 "timed out waiting for idle: idle=0x%x\n", 1656 idle); 1657 return -ETIMEDOUT; 1658 } 1659 1660 udelay(5); 1661 } while (1); 1662 } 1663 1664 static void etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu) 1665 { 1666 if (gpu->state == ETNA_GPU_STATE_RUNNING) { 1667 /* Replace the last WAIT with END */ 1668 mutex_lock(&gpu->lock); 1669 etnaviv_buffer_end(gpu); 1670 mutex_unlock(&gpu->lock); 1671 1672 /* 1673 * We know that only the FE is busy here, this should 1674 * happen quickly (as the WAIT is only 200 cycles). If 1675 * we fail, just warn and continue. 1676 */ 1677 etnaviv_gpu_wait_idle(gpu, 100); 1678 1679 gpu->state = ETNA_GPU_STATE_INITIALIZED; 1680 } 1681 1682 gpu->exec_state = -1; 1683 } 1684 1685 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu) 1686 { 1687 int ret; 1688 1689 ret = mutex_lock_killable(&gpu->lock); 1690 if (ret) 1691 return ret; 1692 1693 etnaviv_gpu_update_clock(gpu); 1694 etnaviv_gpu_hw_init(gpu); 1695 1696 mutex_unlock(&gpu->lock); 1697 1698 return 0; 1699 } 1700 1701 static int 1702 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev, 1703 unsigned long *state) 1704 { 1705 *state = 6; 1706 1707 return 0; 1708 } 1709 1710 static int 1711 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev, 1712 unsigned long *state) 1713 { 1714 struct etnaviv_gpu *gpu = cdev->devdata; 1715 1716 *state = gpu->freq_scale; 1717 1718 return 0; 1719 } 1720 1721 static int 1722 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev, 1723 unsigned long state) 1724 { 1725 struct etnaviv_gpu *gpu = cdev->devdata; 1726 1727 mutex_lock(&gpu->lock); 1728 gpu->freq_scale = state; 1729 if (!pm_runtime_suspended(gpu->dev)) 1730 etnaviv_gpu_update_clock(gpu); 1731 mutex_unlock(&gpu->lock); 1732 1733 return 0; 1734 } 1735 1736 static const struct thermal_cooling_device_ops cooling_ops = { 1737 .get_max_state = etnaviv_gpu_cooling_get_max_state, 1738 .get_cur_state = etnaviv_gpu_cooling_get_cur_state, 1739 .set_cur_state = etnaviv_gpu_cooling_set_cur_state, 1740 }; 1741 1742 static int etnaviv_gpu_bind(struct device *dev, struct device *master, 1743 void *data) 1744 { 1745 struct drm_device *drm = data; 1746 struct etnaviv_drm_private *priv = drm->dev_private; 1747 struct etnaviv_gpu *gpu = dev_get_drvdata(dev); 1748 int ret; 1749 1750 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) { 1751 gpu->cooling = thermal_of_cooling_device_register(dev->of_node, 1752 (char *)dev_name(dev), gpu, &cooling_ops); 1753 if (IS_ERR(gpu->cooling)) 1754 return PTR_ERR(gpu->cooling); 1755 } 1756 1757 gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0); 1758 if (!gpu->wq) { 1759 ret = -ENOMEM; 1760 goto out_thermal; 1761 } 1762 1763 ret = etnaviv_sched_init(gpu); 1764 if (ret) 1765 goto out_workqueue; 1766 1767 if (!IS_ENABLED(CONFIG_PM)) { 1768 ret = etnaviv_gpu_clk_enable(gpu); 1769 if (ret < 0) 1770 goto out_sched; 1771 } 1772 1773 gpu->drm = drm; 1774 gpu->fence_context = dma_fence_context_alloc(1); 1775 xa_init_flags(&gpu->user_fences, XA_FLAGS_ALLOC); 1776 spin_lock_init(&gpu->fence_spinlock); 1777 1778 INIT_WORK(&gpu->sync_point_work, sync_point_worker); 1779 init_waitqueue_head(&gpu->fence_event); 1780 1781 priv->gpu[priv->num_gpus++] = gpu; 1782 1783 return 0; 1784 1785 out_sched: 1786 etnaviv_sched_fini(gpu); 1787 1788 out_workqueue: 1789 destroy_workqueue(gpu->wq); 1790 1791 out_thermal: 1792 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) 1793 thermal_cooling_device_unregister(gpu->cooling); 1794 1795 return ret; 1796 } 1797 1798 static void etnaviv_gpu_unbind(struct device *dev, struct device *master, 1799 void *data) 1800 { 1801 struct etnaviv_gpu *gpu = dev_get_drvdata(dev); 1802 1803 DBG("%s", dev_name(gpu->dev)); 1804 1805 destroy_workqueue(gpu->wq); 1806 1807 etnaviv_sched_fini(gpu); 1808 1809 if (IS_ENABLED(CONFIG_PM)) { 1810 pm_runtime_get_sync(gpu->dev); 1811 pm_runtime_put_sync_suspend(gpu->dev); 1812 } else { 1813 etnaviv_gpu_hw_suspend(gpu); 1814 etnaviv_gpu_clk_disable(gpu); 1815 } 1816 1817 if (gpu->mmu_context) 1818 etnaviv_iommu_context_put(gpu->mmu_context); 1819 1820 etnaviv_cmdbuf_free(&gpu->buffer); 1821 etnaviv_iommu_global_fini(gpu); 1822 1823 gpu->drm = NULL; 1824 xa_destroy(&gpu->user_fences); 1825 1826 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) 1827 thermal_cooling_device_unregister(gpu->cooling); 1828 gpu->cooling = NULL; 1829 } 1830 1831 static const struct component_ops gpu_ops = { 1832 .bind = etnaviv_gpu_bind, 1833 .unbind = etnaviv_gpu_unbind, 1834 }; 1835 1836 static const struct of_device_id etnaviv_gpu_match[] = { 1837 { 1838 .compatible = "vivante,gc" 1839 }, 1840 { /* sentinel */ } 1841 }; 1842 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match); 1843 1844 static int etnaviv_gpu_platform_probe(struct platform_device *pdev) 1845 { 1846 struct device *dev = &pdev->dev; 1847 struct etnaviv_gpu *gpu; 1848 int err; 1849 1850 gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL); 1851 if (!gpu) 1852 return -ENOMEM; 1853 1854 gpu->dev = dev; 1855 mutex_init(&gpu->lock); 1856 mutex_init(&gpu->sched_lock); 1857 1858 /* Map registers: */ 1859 gpu->mmio = devm_platform_ioremap_resource(pdev, 0); 1860 if (IS_ERR(gpu->mmio)) 1861 return PTR_ERR(gpu->mmio); 1862 1863 /* Get Interrupt: */ 1864 gpu->irq = platform_get_irq(pdev, 0); 1865 if (gpu->irq < 0) 1866 return gpu->irq; 1867 1868 err = devm_request_irq(dev, gpu->irq, irq_handler, 0, 1869 dev_name(dev), gpu); 1870 if (err) { 1871 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err); 1872 return err; 1873 } 1874 1875 /* Get Clocks: */ 1876 gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg"); 1877 DBG("clk_reg: %p", gpu->clk_reg); 1878 if (IS_ERR(gpu->clk_reg)) 1879 return PTR_ERR(gpu->clk_reg); 1880 1881 gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus"); 1882 DBG("clk_bus: %p", gpu->clk_bus); 1883 if (IS_ERR(gpu->clk_bus)) 1884 return PTR_ERR(gpu->clk_bus); 1885 1886 gpu->clk_core = devm_clk_get(&pdev->dev, "core"); 1887 DBG("clk_core: %p", gpu->clk_core); 1888 if (IS_ERR(gpu->clk_core)) 1889 return PTR_ERR(gpu->clk_core); 1890 gpu->base_rate_core = clk_get_rate(gpu->clk_core); 1891 1892 gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader"); 1893 DBG("clk_shader: %p", gpu->clk_shader); 1894 if (IS_ERR(gpu->clk_shader)) 1895 return PTR_ERR(gpu->clk_shader); 1896 gpu->base_rate_shader = clk_get_rate(gpu->clk_shader); 1897 1898 /* TODO: figure out max mapped size */ 1899 dev_set_drvdata(dev, gpu); 1900 1901 /* 1902 * We treat the device as initially suspended. The runtime PM 1903 * autosuspend delay is rather arbitary: no measurements have 1904 * yet been performed to determine an appropriate value. 1905 */ 1906 pm_runtime_use_autosuspend(dev); 1907 pm_runtime_set_autosuspend_delay(dev, 200); 1908 pm_runtime_enable(dev); 1909 1910 err = component_add(dev, &gpu_ops); 1911 if (err < 0) { 1912 dev_err(dev, "failed to register component: %d\n", err); 1913 return err; 1914 } 1915 1916 return 0; 1917 } 1918 1919 static void etnaviv_gpu_platform_remove(struct platform_device *pdev) 1920 { 1921 struct etnaviv_gpu *gpu = dev_get_drvdata(&pdev->dev); 1922 1923 component_del(&pdev->dev, &gpu_ops); 1924 pm_runtime_disable(&pdev->dev); 1925 1926 mutex_destroy(&gpu->lock); 1927 mutex_destroy(&gpu->sched_lock); 1928 } 1929 1930 static int etnaviv_gpu_rpm_suspend(struct device *dev) 1931 { 1932 struct etnaviv_gpu *gpu = dev_get_drvdata(dev); 1933 u32 idle, mask; 1934 1935 /* If there are any jobs in the HW queue, we're not idle */ 1936 if (atomic_read(&gpu->sched.credit_count)) 1937 return -EBUSY; 1938 1939 /* Check whether the hardware (except FE and MC) is idle */ 1940 mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE | 1941 VIVS_HI_IDLE_STATE_MC); 1942 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask; 1943 if (idle != mask) { 1944 dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n", 1945 idle); 1946 return -EBUSY; 1947 } 1948 1949 etnaviv_gpu_hw_suspend(gpu); 1950 1951 gpu->state = ETNA_GPU_STATE_IDENTIFIED; 1952 1953 return etnaviv_gpu_clk_disable(gpu); 1954 } 1955 1956 static int etnaviv_gpu_rpm_resume(struct device *dev) 1957 { 1958 struct etnaviv_gpu *gpu = dev_get_drvdata(dev); 1959 int ret; 1960 1961 ret = etnaviv_gpu_clk_enable(gpu); 1962 if (ret) 1963 return ret; 1964 1965 /* Re-initialise the basic hardware state */ 1966 if (gpu->state == ETNA_GPU_STATE_IDENTIFIED) { 1967 ret = etnaviv_gpu_hw_resume(gpu); 1968 if (ret) { 1969 etnaviv_gpu_clk_disable(gpu); 1970 return ret; 1971 } 1972 } 1973 1974 return 0; 1975 } 1976 1977 static const struct dev_pm_ops etnaviv_gpu_pm_ops = { 1978 RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume, NULL) 1979 }; 1980 1981 struct platform_driver etnaviv_gpu_driver = { 1982 .driver = { 1983 .name = "etnaviv-gpu", 1984 .pm = pm_ptr(&etnaviv_gpu_pm_ops), 1985 .of_match_table = etnaviv_gpu_match, 1986 }, 1987 .probe = etnaviv_gpu_platform_probe, 1988 .remove = etnaviv_gpu_platform_remove, 1989 .id_table = gpu_ids, 1990 }; 1991