1 /* 2 * Copyright (C) 2015 Etnaviv Project 3 * 4 * This program is free software; you can redistribute it and/or modify it 5 * under the terms of the GNU General Public License version 2 as published by 6 * the Free Software Foundation. 7 * 8 * This program is distributed in the hope that it will be useful, but WITHOUT 9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 11 * more details. 12 * 13 * You should have received a copy of the GNU General Public License along with 14 * this program. If not, see <http://www.gnu.org/licenses/>. 15 */ 16 17 #include <linux/component.h> 18 #include <linux/dma-fence.h> 19 #include <linux/moduleparam.h> 20 #include <linux/of_device.h> 21 #include <linux/thermal.h> 22 23 #include "etnaviv_cmdbuf.h" 24 #include "etnaviv_dump.h" 25 #include "etnaviv_gpu.h" 26 #include "etnaviv_gem.h" 27 #include "etnaviv_mmu.h" 28 #include "common.xml.h" 29 #include "state.xml.h" 30 #include "state_hi.xml.h" 31 #include "cmdstream.xml.h" 32 33 static const struct platform_device_id gpu_ids[] = { 34 { .name = "etnaviv-gpu,2d" }, 35 { }, 36 }; 37 38 static bool etnaviv_dump_core = true; 39 module_param_named(dump_core, etnaviv_dump_core, bool, 0600); 40 41 /* 42 * Driver functions: 43 */ 44 45 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value) 46 { 47 switch (param) { 48 case ETNAVIV_PARAM_GPU_MODEL: 49 *value = gpu->identity.model; 50 break; 51 52 case ETNAVIV_PARAM_GPU_REVISION: 53 *value = gpu->identity.revision; 54 break; 55 56 case ETNAVIV_PARAM_GPU_FEATURES_0: 57 *value = gpu->identity.features; 58 break; 59 60 case ETNAVIV_PARAM_GPU_FEATURES_1: 61 *value = gpu->identity.minor_features0; 62 break; 63 64 case ETNAVIV_PARAM_GPU_FEATURES_2: 65 *value = gpu->identity.minor_features1; 66 break; 67 68 case ETNAVIV_PARAM_GPU_FEATURES_3: 69 *value = gpu->identity.minor_features2; 70 break; 71 72 case ETNAVIV_PARAM_GPU_FEATURES_4: 73 *value = gpu->identity.minor_features3; 74 break; 75 76 case ETNAVIV_PARAM_GPU_FEATURES_5: 77 *value = gpu->identity.minor_features4; 78 break; 79 80 case ETNAVIV_PARAM_GPU_FEATURES_6: 81 *value = gpu->identity.minor_features5; 82 break; 83 84 case ETNAVIV_PARAM_GPU_STREAM_COUNT: 85 *value = gpu->identity.stream_count; 86 break; 87 88 case ETNAVIV_PARAM_GPU_REGISTER_MAX: 89 *value = gpu->identity.register_max; 90 break; 91 92 case ETNAVIV_PARAM_GPU_THREAD_COUNT: 93 *value = gpu->identity.thread_count; 94 break; 95 96 case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE: 97 *value = gpu->identity.vertex_cache_size; 98 break; 99 100 case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT: 101 *value = gpu->identity.shader_core_count; 102 break; 103 104 case ETNAVIV_PARAM_GPU_PIXEL_PIPES: 105 *value = gpu->identity.pixel_pipes; 106 break; 107 108 case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE: 109 *value = gpu->identity.vertex_output_buffer_size; 110 break; 111 112 case ETNAVIV_PARAM_GPU_BUFFER_SIZE: 113 *value = gpu->identity.buffer_size; 114 break; 115 116 case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT: 117 *value = gpu->identity.instruction_count; 118 break; 119 120 case ETNAVIV_PARAM_GPU_NUM_CONSTANTS: 121 *value = gpu->identity.num_constants; 122 break; 123 124 case ETNAVIV_PARAM_GPU_NUM_VARYINGS: 125 *value = gpu->identity.varyings_count; 126 break; 127 128 default: 129 DBG("%s: invalid param: %u", dev_name(gpu->dev), param); 130 return -EINVAL; 131 } 132 133 return 0; 134 } 135 136 137 #define etnaviv_is_model_rev(gpu, mod, rev) \ 138 ((gpu)->identity.model == chipModel_##mod && \ 139 (gpu)->identity.revision == rev) 140 #define etnaviv_field(val, field) \ 141 (((val) & field##__MASK) >> field##__SHIFT) 142 143 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu) 144 { 145 if (gpu->identity.minor_features0 & 146 chipMinorFeatures0_MORE_MINOR_FEATURES) { 147 u32 specs[4]; 148 unsigned int streams; 149 150 specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS); 151 specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2); 152 specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3); 153 specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4); 154 155 gpu->identity.stream_count = etnaviv_field(specs[0], 156 VIVS_HI_CHIP_SPECS_STREAM_COUNT); 157 gpu->identity.register_max = etnaviv_field(specs[0], 158 VIVS_HI_CHIP_SPECS_REGISTER_MAX); 159 gpu->identity.thread_count = etnaviv_field(specs[0], 160 VIVS_HI_CHIP_SPECS_THREAD_COUNT); 161 gpu->identity.vertex_cache_size = etnaviv_field(specs[0], 162 VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE); 163 gpu->identity.shader_core_count = etnaviv_field(specs[0], 164 VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT); 165 gpu->identity.pixel_pipes = etnaviv_field(specs[0], 166 VIVS_HI_CHIP_SPECS_PIXEL_PIPES); 167 gpu->identity.vertex_output_buffer_size = 168 etnaviv_field(specs[0], 169 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE); 170 171 gpu->identity.buffer_size = etnaviv_field(specs[1], 172 VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE); 173 gpu->identity.instruction_count = etnaviv_field(specs[1], 174 VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT); 175 gpu->identity.num_constants = etnaviv_field(specs[1], 176 VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS); 177 178 gpu->identity.varyings_count = etnaviv_field(specs[2], 179 VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT); 180 181 /* This overrides the value from older register if non-zero */ 182 streams = etnaviv_field(specs[3], 183 VIVS_HI_CHIP_SPECS_4_STREAM_COUNT); 184 if (streams) 185 gpu->identity.stream_count = streams; 186 } 187 188 /* Fill in the stream count if not specified */ 189 if (gpu->identity.stream_count == 0) { 190 if (gpu->identity.model >= 0x1000) 191 gpu->identity.stream_count = 4; 192 else 193 gpu->identity.stream_count = 1; 194 } 195 196 /* Convert the register max value */ 197 if (gpu->identity.register_max) 198 gpu->identity.register_max = 1 << gpu->identity.register_max; 199 else if (gpu->identity.model == chipModel_GC400) 200 gpu->identity.register_max = 32; 201 else 202 gpu->identity.register_max = 64; 203 204 /* Convert thread count */ 205 if (gpu->identity.thread_count) 206 gpu->identity.thread_count = 1 << gpu->identity.thread_count; 207 else if (gpu->identity.model == chipModel_GC400) 208 gpu->identity.thread_count = 64; 209 else if (gpu->identity.model == chipModel_GC500 || 210 gpu->identity.model == chipModel_GC530) 211 gpu->identity.thread_count = 128; 212 else 213 gpu->identity.thread_count = 256; 214 215 if (gpu->identity.vertex_cache_size == 0) 216 gpu->identity.vertex_cache_size = 8; 217 218 if (gpu->identity.shader_core_count == 0) { 219 if (gpu->identity.model >= 0x1000) 220 gpu->identity.shader_core_count = 2; 221 else 222 gpu->identity.shader_core_count = 1; 223 } 224 225 if (gpu->identity.pixel_pipes == 0) 226 gpu->identity.pixel_pipes = 1; 227 228 /* Convert virtex buffer size */ 229 if (gpu->identity.vertex_output_buffer_size) { 230 gpu->identity.vertex_output_buffer_size = 231 1 << gpu->identity.vertex_output_buffer_size; 232 } else if (gpu->identity.model == chipModel_GC400) { 233 if (gpu->identity.revision < 0x4000) 234 gpu->identity.vertex_output_buffer_size = 512; 235 else if (gpu->identity.revision < 0x4200) 236 gpu->identity.vertex_output_buffer_size = 256; 237 else 238 gpu->identity.vertex_output_buffer_size = 128; 239 } else { 240 gpu->identity.vertex_output_buffer_size = 512; 241 } 242 243 switch (gpu->identity.instruction_count) { 244 case 0: 245 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) || 246 gpu->identity.model == chipModel_GC880) 247 gpu->identity.instruction_count = 512; 248 else 249 gpu->identity.instruction_count = 256; 250 break; 251 252 case 1: 253 gpu->identity.instruction_count = 1024; 254 break; 255 256 case 2: 257 gpu->identity.instruction_count = 2048; 258 break; 259 260 default: 261 gpu->identity.instruction_count = 256; 262 break; 263 } 264 265 if (gpu->identity.num_constants == 0) 266 gpu->identity.num_constants = 168; 267 268 if (gpu->identity.varyings_count == 0) { 269 if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0) 270 gpu->identity.varyings_count = 12; 271 else 272 gpu->identity.varyings_count = 8; 273 } 274 275 /* 276 * For some cores, two varyings are consumed for position, so the 277 * maximum varying count needs to be reduced by one. 278 */ 279 if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) || 280 etnaviv_is_model_rev(gpu, GC4000, 0x5222) || 281 etnaviv_is_model_rev(gpu, GC4000, 0x5245) || 282 etnaviv_is_model_rev(gpu, GC4000, 0x5208) || 283 etnaviv_is_model_rev(gpu, GC3000, 0x5435) || 284 etnaviv_is_model_rev(gpu, GC2200, 0x5244) || 285 etnaviv_is_model_rev(gpu, GC2100, 0x5108) || 286 etnaviv_is_model_rev(gpu, GC2000, 0x5108) || 287 etnaviv_is_model_rev(gpu, GC1500, 0x5246) || 288 etnaviv_is_model_rev(gpu, GC880, 0x5107) || 289 etnaviv_is_model_rev(gpu, GC880, 0x5106)) 290 gpu->identity.varyings_count -= 1; 291 } 292 293 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu) 294 { 295 u32 chipIdentity; 296 297 chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY); 298 299 /* Special case for older graphic cores. */ 300 if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) { 301 gpu->identity.model = chipModel_GC500; 302 gpu->identity.revision = etnaviv_field(chipIdentity, 303 VIVS_HI_CHIP_IDENTITY_REVISION); 304 } else { 305 306 gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL); 307 gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV); 308 309 /* 310 * !!!! HACK ALERT !!!! 311 * Because people change device IDs without letting software 312 * know about it - here is the hack to make it all look the 313 * same. Only for GC400 family. 314 */ 315 if ((gpu->identity.model & 0xff00) == 0x0400 && 316 gpu->identity.model != chipModel_GC420) { 317 gpu->identity.model = gpu->identity.model & 0x0400; 318 } 319 320 /* Another special case */ 321 if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) { 322 u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE); 323 u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME); 324 325 if (chipDate == 0x20080814 && chipTime == 0x12051100) { 326 /* 327 * This IP has an ECO; put the correct 328 * revision in it. 329 */ 330 gpu->identity.revision = 0x1051; 331 } 332 } 333 334 /* 335 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in 336 * reality it's just a re-branded GC3000. We can identify this 337 * core by the upper half of the revision register being all 1. 338 * Fix model/rev here, so all other places can refer to this 339 * core by its real identity. 340 */ 341 if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) { 342 gpu->identity.model = chipModel_GC3000; 343 gpu->identity.revision &= 0xffff; 344 } 345 } 346 347 dev_info(gpu->dev, "model: GC%x, revision: %x\n", 348 gpu->identity.model, gpu->identity.revision); 349 350 gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE); 351 352 /* Disable fast clear on GC700. */ 353 if (gpu->identity.model == chipModel_GC700) 354 gpu->identity.features &= ~chipFeatures_FAST_CLEAR; 355 356 if ((gpu->identity.model == chipModel_GC500 && 357 gpu->identity.revision < 2) || 358 (gpu->identity.model == chipModel_GC300 && 359 gpu->identity.revision < 0x2000)) { 360 361 /* 362 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these 363 * registers. 364 */ 365 gpu->identity.minor_features0 = 0; 366 gpu->identity.minor_features1 = 0; 367 gpu->identity.minor_features2 = 0; 368 gpu->identity.minor_features3 = 0; 369 gpu->identity.minor_features4 = 0; 370 gpu->identity.minor_features5 = 0; 371 } else 372 gpu->identity.minor_features0 = 373 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0); 374 375 if (gpu->identity.minor_features0 & 376 chipMinorFeatures0_MORE_MINOR_FEATURES) { 377 gpu->identity.minor_features1 = 378 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1); 379 gpu->identity.minor_features2 = 380 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2); 381 gpu->identity.minor_features3 = 382 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3); 383 gpu->identity.minor_features4 = 384 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4); 385 gpu->identity.minor_features5 = 386 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5); 387 } 388 389 /* GC600 idle register reports zero bits where modules aren't present */ 390 if (gpu->identity.model == chipModel_GC600) { 391 gpu->idle_mask = VIVS_HI_IDLE_STATE_TX | 392 VIVS_HI_IDLE_STATE_RA | 393 VIVS_HI_IDLE_STATE_SE | 394 VIVS_HI_IDLE_STATE_PA | 395 VIVS_HI_IDLE_STATE_SH | 396 VIVS_HI_IDLE_STATE_PE | 397 VIVS_HI_IDLE_STATE_DE | 398 VIVS_HI_IDLE_STATE_FE; 399 } else { 400 gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP; 401 } 402 403 etnaviv_hw_specs(gpu); 404 } 405 406 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock) 407 { 408 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock | 409 VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD); 410 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock); 411 } 412 413 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu) 414 { 415 if (gpu->identity.minor_features2 & 416 chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) { 417 clk_set_rate(gpu->clk_core, 418 gpu->base_rate_core >> gpu->freq_scale); 419 clk_set_rate(gpu->clk_shader, 420 gpu->base_rate_shader >> gpu->freq_scale); 421 } else { 422 unsigned int fscale = 1 << (6 - gpu->freq_scale); 423 u32 clock = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS | 424 VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale); 425 426 etnaviv_gpu_load_clock(gpu, clock); 427 } 428 } 429 430 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu) 431 { 432 u32 control, idle; 433 unsigned long timeout; 434 bool failed = true; 435 436 /* TODO 437 * 438 * - clock gating 439 * - puls eater 440 * - what about VG? 441 */ 442 443 /* We hope that the GPU resets in under one second */ 444 timeout = jiffies + msecs_to_jiffies(1000); 445 446 while (time_is_after_jiffies(timeout)) { 447 /* enable clock */ 448 etnaviv_gpu_update_clock(gpu); 449 450 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); 451 452 /* Wait for stable clock. Vivante's code waited for 1ms */ 453 usleep_range(1000, 10000); 454 455 /* isolate the GPU. */ 456 control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU; 457 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 458 459 /* set soft reset. */ 460 control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET; 461 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 462 463 /* wait for reset. */ 464 msleep(1); 465 466 /* reset soft reset bit. */ 467 control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET; 468 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 469 470 /* reset GPU isolation. */ 471 control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU; 472 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); 473 474 /* read idle register. */ 475 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); 476 477 /* try reseting again if FE it not idle */ 478 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) { 479 dev_dbg(gpu->dev, "FE is not idle\n"); 480 continue; 481 } 482 483 /* read reset register. */ 484 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); 485 486 /* is the GPU idle? */ 487 if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) || 488 ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) { 489 dev_dbg(gpu->dev, "GPU is not idle\n"); 490 continue; 491 } 492 493 failed = false; 494 break; 495 } 496 497 if (failed) { 498 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); 499 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); 500 501 dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n", 502 idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ", 503 control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ", 504 control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not "); 505 506 return -EBUSY; 507 } 508 509 /* We rely on the GPU running, so program the clock */ 510 etnaviv_gpu_update_clock(gpu); 511 512 return 0; 513 } 514 515 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu) 516 { 517 u32 pmc, ppc; 518 519 /* enable clock gating */ 520 ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS); 521 ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; 522 523 /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */ 524 if (gpu->identity.revision == 0x4301 || 525 gpu->identity.revision == 0x4302) 526 ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING; 527 528 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc); 529 530 pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS); 531 532 /* Disable PA clock gating for GC400+ without bugfix except for GC420 */ 533 if (gpu->identity.model >= chipModel_GC400 && 534 gpu->identity.model != chipModel_GC420 && 535 !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12)) 536 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA; 537 538 /* 539 * Disable PE clock gating on revs < 5.0.0.0 when HZ is 540 * present without a bug fix. 541 */ 542 if (gpu->identity.revision < 0x5000 && 543 gpu->identity.minor_features0 & chipMinorFeatures0_HZ && 544 !(gpu->identity.minor_features1 & 545 chipMinorFeatures1_DISABLE_PE_GATING)) 546 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE; 547 548 if (gpu->identity.revision < 0x5422) 549 pmc |= BIT(15); /* Unknown bit */ 550 551 /* Disable TX clock gating on affected core revisions. */ 552 if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) || 553 etnaviv_is_model_rev(gpu, GC2000, 0x5108)) 554 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX; 555 556 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ; 557 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ; 558 559 gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc); 560 } 561 562 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch) 563 { 564 gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address); 565 gpu_write(gpu, VIVS_FE_COMMAND_CONTROL, 566 VIVS_FE_COMMAND_CONTROL_ENABLE | 567 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch)); 568 } 569 570 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu) 571 { 572 /* 573 * Base value for VIVS_PM_PULSE_EATER register on models where it 574 * cannot be read, extracted from vivante kernel driver. 575 */ 576 u32 pulse_eater = 0x01590880; 577 578 if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) || 579 etnaviv_is_model_rev(gpu, GC4000, 0x5222)) { 580 pulse_eater |= BIT(23); 581 582 } 583 584 if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) || 585 etnaviv_is_model_rev(gpu, GC1000, 0x5040)) { 586 pulse_eater &= ~BIT(16); 587 pulse_eater |= BIT(17); 588 } 589 590 if ((gpu->identity.revision > 0x5420) && 591 (gpu->identity.features & chipFeatures_PIPE_3D)) 592 { 593 /* Performance fix: disable internal DFS */ 594 pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER); 595 pulse_eater |= BIT(18); 596 } 597 598 gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater); 599 } 600 601 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu) 602 { 603 u16 prefetch; 604 605 if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) || 606 etnaviv_is_model_rev(gpu, GC320, 0x5220)) && 607 gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) { 608 u32 mc_memory_debug; 609 610 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff; 611 612 if (gpu->identity.revision == 0x5007) 613 mc_memory_debug |= 0x0c; 614 else 615 mc_memory_debug |= 0x08; 616 617 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug); 618 } 619 620 /* enable module-level clock gating */ 621 etnaviv_gpu_enable_mlcg(gpu); 622 623 /* 624 * Update GPU AXI cache atttribute to "cacheable, no allocate". 625 * This is necessary to prevent the iMX6 SoC locking up. 626 */ 627 gpu_write(gpu, VIVS_HI_AXI_CONFIG, 628 VIVS_HI_AXI_CONFIG_AWCACHE(2) | 629 VIVS_HI_AXI_CONFIG_ARCACHE(2)); 630 631 /* GC2000 rev 5108 needs a special bus config */ 632 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) { 633 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG); 634 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK | 635 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK); 636 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) | 637 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0); 638 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config); 639 } 640 641 /* setup the pulse eater */ 642 etnaviv_gpu_setup_pulse_eater(gpu); 643 644 /* setup the MMU */ 645 etnaviv_iommu_restore(gpu); 646 647 /* Start command processor */ 648 prefetch = etnaviv_buffer_init(gpu); 649 650 gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U); 651 etnaviv_gpu_start_fe(gpu, etnaviv_cmdbuf_get_va(gpu->buffer), 652 prefetch); 653 } 654 655 int etnaviv_gpu_init(struct etnaviv_gpu *gpu) 656 { 657 int ret, i; 658 659 ret = pm_runtime_get_sync(gpu->dev); 660 if (ret < 0) { 661 dev_err(gpu->dev, "Failed to enable GPU power domain\n"); 662 return ret; 663 } 664 665 etnaviv_hw_identify(gpu); 666 667 if (gpu->identity.model == 0) { 668 dev_err(gpu->dev, "Unknown GPU model\n"); 669 ret = -ENXIO; 670 goto fail; 671 } 672 673 /* Exclude VG cores with FE2.0 */ 674 if (gpu->identity.features & chipFeatures_PIPE_VG && 675 gpu->identity.features & chipFeatures_FE20) { 676 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n"); 677 ret = -ENXIO; 678 goto fail; 679 } 680 681 /* 682 * Set the GPU linear window to be at the end of the DMA window, where 683 * the CMA area is likely to reside. This ensures that we are able to 684 * map the command buffers while having the linear window overlap as 685 * much RAM as possible, so we can optimize mappings for other buffers. 686 * 687 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads 688 * to different views of the memory on the individual engines. 689 */ 690 if (!(gpu->identity.features & chipFeatures_PIPE_3D) || 691 (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) { 692 u32 dma_mask = (u32)dma_get_required_mask(gpu->dev); 693 if (dma_mask < PHYS_OFFSET + SZ_2G) 694 gpu->memory_base = PHYS_OFFSET; 695 else 696 gpu->memory_base = dma_mask - SZ_2G + 1; 697 } else if (PHYS_OFFSET >= SZ_2G) { 698 dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n"); 699 gpu->memory_base = PHYS_OFFSET; 700 gpu->identity.features &= ~chipFeatures_FAST_CLEAR; 701 } 702 703 ret = etnaviv_hw_reset(gpu); 704 if (ret) { 705 dev_err(gpu->dev, "GPU reset failed\n"); 706 goto fail; 707 } 708 709 gpu->mmu = etnaviv_iommu_new(gpu); 710 if (IS_ERR(gpu->mmu)) { 711 dev_err(gpu->dev, "Failed to instantiate GPU IOMMU\n"); 712 ret = PTR_ERR(gpu->mmu); 713 goto fail; 714 } 715 716 gpu->cmdbuf_suballoc = etnaviv_cmdbuf_suballoc_new(gpu); 717 if (IS_ERR(gpu->cmdbuf_suballoc)) { 718 dev_err(gpu->dev, "Failed to create cmdbuf suballocator\n"); 719 ret = PTR_ERR(gpu->cmdbuf_suballoc); 720 goto fail; 721 } 722 723 /* Create buffer: */ 724 gpu->buffer = etnaviv_cmdbuf_new(gpu->cmdbuf_suballoc, PAGE_SIZE, 0); 725 if (!gpu->buffer) { 726 ret = -ENOMEM; 727 dev_err(gpu->dev, "could not create command buffer\n"); 728 goto destroy_iommu; 729 } 730 731 if (gpu->mmu->version == ETNAVIV_IOMMU_V1 && 732 etnaviv_cmdbuf_get_va(gpu->buffer) > 0x80000000) { 733 ret = -EINVAL; 734 dev_err(gpu->dev, 735 "command buffer outside valid memory window\n"); 736 goto free_buffer; 737 } 738 739 /* Setup event management */ 740 spin_lock_init(&gpu->event_spinlock); 741 init_completion(&gpu->event_free); 742 for (i = 0; i < ARRAY_SIZE(gpu->event); i++) { 743 gpu->event[i].used = false; 744 complete(&gpu->event_free); 745 } 746 747 /* Now program the hardware */ 748 mutex_lock(&gpu->lock); 749 etnaviv_gpu_hw_init(gpu); 750 gpu->exec_state = -1; 751 mutex_unlock(&gpu->lock); 752 753 pm_runtime_mark_last_busy(gpu->dev); 754 pm_runtime_put_autosuspend(gpu->dev); 755 756 return 0; 757 758 free_buffer: 759 etnaviv_cmdbuf_free(gpu->buffer); 760 gpu->buffer = NULL; 761 destroy_iommu: 762 etnaviv_iommu_destroy(gpu->mmu); 763 gpu->mmu = NULL; 764 fail: 765 pm_runtime_mark_last_busy(gpu->dev); 766 pm_runtime_put_autosuspend(gpu->dev); 767 768 return ret; 769 } 770 771 #ifdef CONFIG_DEBUG_FS 772 struct dma_debug { 773 u32 address[2]; 774 u32 state[2]; 775 }; 776 777 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug) 778 { 779 u32 i; 780 781 debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); 782 debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE); 783 784 for (i = 0; i < 500; i++) { 785 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); 786 debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE); 787 788 if (debug->address[0] != debug->address[1]) 789 break; 790 791 if (debug->state[0] != debug->state[1]) 792 break; 793 } 794 } 795 796 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m) 797 { 798 struct dma_debug debug; 799 u32 dma_lo, dma_hi, axi, idle; 800 int ret; 801 802 seq_printf(m, "%s Status:\n", dev_name(gpu->dev)); 803 804 ret = pm_runtime_get_sync(gpu->dev); 805 if (ret < 0) 806 return ret; 807 808 dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW); 809 dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH); 810 axi = gpu_read(gpu, VIVS_HI_AXI_STATUS); 811 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); 812 813 verify_dma(gpu, &debug); 814 815 seq_puts(m, "\tfeatures\n"); 816 seq_printf(m, "\t minor_features0: 0x%08x\n", 817 gpu->identity.minor_features0); 818 seq_printf(m, "\t minor_features1: 0x%08x\n", 819 gpu->identity.minor_features1); 820 seq_printf(m, "\t minor_features2: 0x%08x\n", 821 gpu->identity.minor_features2); 822 seq_printf(m, "\t minor_features3: 0x%08x\n", 823 gpu->identity.minor_features3); 824 seq_printf(m, "\t minor_features4: 0x%08x\n", 825 gpu->identity.minor_features4); 826 seq_printf(m, "\t minor_features5: 0x%08x\n", 827 gpu->identity.minor_features5); 828 829 seq_puts(m, "\tspecs\n"); 830 seq_printf(m, "\t stream_count: %d\n", 831 gpu->identity.stream_count); 832 seq_printf(m, "\t register_max: %d\n", 833 gpu->identity.register_max); 834 seq_printf(m, "\t thread_count: %d\n", 835 gpu->identity.thread_count); 836 seq_printf(m, "\t vertex_cache_size: %d\n", 837 gpu->identity.vertex_cache_size); 838 seq_printf(m, "\t shader_core_count: %d\n", 839 gpu->identity.shader_core_count); 840 seq_printf(m, "\t pixel_pipes: %d\n", 841 gpu->identity.pixel_pipes); 842 seq_printf(m, "\t vertex_output_buffer_size: %d\n", 843 gpu->identity.vertex_output_buffer_size); 844 seq_printf(m, "\t buffer_size: %d\n", 845 gpu->identity.buffer_size); 846 seq_printf(m, "\t instruction_count: %d\n", 847 gpu->identity.instruction_count); 848 seq_printf(m, "\t num_constants: %d\n", 849 gpu->identity.num_constants); 850 seq_printf(m, "\t varyings_count: %d\n", 851 gpu->identity.varyings_count); 852 853 seq_printf(m, "\taxi: 0x%08x\n", axi); 854 seq_printf(m, "\tidle: 0x%08x\n", idle); 855 idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP; 856 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) 857 seq_puts(m, "\t FE is not idle\n"); 858 if ((idle & VIVS_HI_IDLE_STATE_DE) == 0) 859 seq_puts(m, "\t DE is not idle\n"); 860 if ((idle & VIVS_HI_IDLE_STATE_PE) == 0) 861 seq_puts(m, "\t PE is not idle\n"); 862 if ((idle & VIVS_HI_IDLE_STATE_SH) == 0) 863 seq_puts(m, "\t SH is not idle\n"); 864 if ((idle & VIVS_HI_IDLE_STATE_PA) == 0) 865 seq_puts(m, "\t PA is not idle\n"); 866 if ((idle & VIVS_HI_IDLE_STATE_SE) == 0) 867 seq_puts(m, "\t SE is not idle\n"); 868 if ((idle & VIVS_HI_IDLE_STATE_RA) == 0) 869 seq_puts(m, "\t RA is not idle\n"); 870 if ((idle & VIVS_HI_IDLE_STATE_TX) == 0) 871 seq_puts(m, "\t TX is not idle\n"); 872 if ((idle & VIVS_HI_IDLE_STATE_VG) == 0) 873 seq_puts(m, "\t VG is not idle\n"); 874 if ((idle & VIVS_HI_IDLE_STATE_IM) == 0) 875 seq_puts(m, "\t IM is not idle\n"); 876 if ((idle & VIVS_HI_IDLE_STATE_FP) == 0) 877 seq_puts(m, "\t FP is not idle\n"); 878 if ((idle & VIVS_HI_IDLE_STATE_TS) == 0) 879 seq_puts(m, "\t TS is not idle\n"); 880 if (idle & VIVS_HI_IDLE_STATE_AXI_LP) 881 seq_puts(m, "\t AXI low power mode\n"); 882 883 if (gpu->identity.features & chipFeatures_DEBUG_MODE) { 884 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0); 885 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1); 886 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE); 887 888 seq_puts(m, "\tMC\n"); 889 seq_printf(m, "\t read0: 0x%08x\n", read0); 890 seq_printf(m, "\t read1: 0x%08x\n", read1); 891 seq_printf(m, "\t write: 0x%08x\n", write); 892 } 893 894 seq_puts(m, "\tDMA "); 895 896 if (debug.address[0] == debug.address[1] && 897 debug.state[0] == debug.state[1]) { 898 seq_puts(m, "seems to be stuck\n"); 899 } else if (debug.address[0] == debug.address[1]) { 900 seq_puts(m, "address is constant\n"); 901 } else { 902 seq_puts(m, "is running\n"); 903 } 904 905 seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]); 906 seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]); 907 seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]); 908 seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]); 909 seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n", 910 dma_lo, dma_hi); 911 912 ret = 0; 913 914 pm_runtime_mark_last_busy(gpu->dev); 915 pm_runtime_put_autosuspend(gpu->dev); 916 917 return ret; 918 } 919 #endif 920 921 /* 922 * Hangcheck detection for locked gpu: 923 */ 924 static void recover_worker(struct work_struct *work) 925 { 926 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu, 927 recover_work); 928 unsigned long flags; 929 unsigned int i; 930 931 dev_err(gpu->dev, "hangcheck recover!\n"); 932 933 if (pm_runtime_get_sync(gpu->dev) < 0) 934 return; 935 936 mutex_lock(&gpu->lock); 937 938 /* Only catch the first event, or when manually re-armed */ 939 if (etnaviv_dump_core) { 940 etnaviv_core_dump(gpu); 941 etnaviv_dump_core = false; 942 } 943 944 etnaviv_hw_reset(gpu); 945 946 /* complete all events, the GPU won't do it after the reset */ 947 spin_lock_irqsave(&gpu->event_spinlock, flags); 948 for (i = 0; i < ARRAY_SIZE(gpu->event); i++) { 949 if (!gpu->event[i].used) 950 continue; 951 dma_fence_signal(gpu->event[i].fence); 952 gpu->event[i].fence = NULL; 953 gpu->event[i].used = false; 954 complete(&gpu->event_free); 955 } 956 spin_unlock_irqrestore(&gpu->event_spinlock, flags); 957 gpu->completed_fence = gpu->active_fence; 958 959 etnaviv_gpu_hw_init(gpu); 960 gpu->lastctx = NULL; 961 gpu->exec_state = -1; 962 963 mutex_unlock(&gpu->lock); 964 pm_runtime_mark_last_busy(gpu->dev); 965 pm_runtime_put_autosuspend(gpu->dev); 966 967 /* Retire the buffer objects in a work */ 968 etnaviv_queue_work(gpu->drm, &gpu->retire_work); 969 } 970 971 static void hangcheck_timer_reset(struct etnaviv_gpu *gpu) 972 { 973 DBG("%s", dev_name(gpu->dev)); 974 mod_timer(&gpu->hangcheck_timer, 975 round_jiffies_up(jiffies + DRM_ETNAVIV_HANGCHECK_JIFFIES)); 976 } 977 978 static void hangcheck_handler(unsigned long data) 979 { 980 struct etnaviv_gpu *gpu = (struct etnaviv_gpu *)data; 981 u32 fence = gpu->completed_fence; 982 bool progress = false; 983 984 if (fence != gpu->hangcheck_fence) { 985 gpu->hangcheck_fence = fence; 986 progress = true; 987 } 988 989 if (!progress) { 990 u32 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); 991 int change = dma_addr - gpu->hangcheck_dma_addr; 992 993 if (change < 0 || change > 16) { 994 gpu->hangcheck_dma_addr = dma_addr; 995 progress = true; 996 } 997 } 998 999 if (!progress && fence_after(gpu->active_fence, fence)) { 1000 dev_err(gpu->dev, "hangcheck detected gpu lockup!\n"); 1001 dev_err(gpu->dev, " completed fence: %u\n", fence); 1002 dev_err(gpu->dev, " active fence: %u\n", 1003 gpu->active_fence); 1004 etnaviv_queue_work(gpu->drm, &gpu->recover_work); 1005 } 1006 1007 /* if still more pending work, reset the hangcheck timer: */ 1008 if (fence_after(gpu->active_fence, gpu->hangcheck_fence)) 1009 hangcheck_timer_reset(gpu); 1010 } 1011 1012 static void hangcheck_disable(struct etnaviv_gpu *gpu) 1013 { 1014 del_timer_sync(&gpu->hangcheck_timer); 1015 cancel_work_sync(&gpu->recover_work); 1016 } 1017 1018 /* fence object management */ 1019 struct etnaviv_fence { 1020 struct etnaviv_gpu *gpu; 1021 struct dma_fence base; 1022 }; 1023 1024 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence) 1025 { 1026 return container_of(fence, struct etnaviv_fence, base); 1027 } 1028 1029 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence) 1030 { 1031 return "etnaviv"; 1032 } 1033 1034 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence) 1035 { 1036 struct etnaviv_fence *f = to_etnaviv_fence(fence); 1037 1038 return dev_name(f->gpu->dev); 1039 } 1040 1041 static bool etnaviv_fence_enable_signaling(struct dma_fence *fence) 1042 { 1043 return true; 1044 } 1045 1046 static bool etnaviv_fence_signaled(struct dma_fence *fence) 1047 { 1048 struct etnaviv_fence *f = to_etnaviv_fence(fence); 1049 1050 return fence_completed(f->gpu, f->base.seqno); 1051 } 1052 1053 static void etnaviv_fence_release(struct dma_fence *fence) 1054 { 1055 struct etnaviv_fence *f = to_etnaviv_fence(fence); 1056 1057 kfree_rcu(f, base.rcu); 1058 } 1059 1060 static const struct dma_fence_ops etnaviv_fence_ops = { 1061 .get_driver_name = etnaviv_fence_get_driver_name, 1062 .get_timeline_name = etnaviv_fence_get_timeline_name, 1063 .enable_signaling = etnaviv_fence_enable_signaling, 1064 .signaled = etnaviv_fence_signaled, 1065 .wait = dma_fence_default_wait, 1066 .release = etnaviv_fence_release, 1067 }; 1068 1069 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu) 1070 { 1071 struct etnaviv_fence *f; 1072 1073 /* 1074 * GPU lock must already be held, otherwise fence completion order might 1075 * not match the seqno order assigned here. 1076 */ 1077 lockdep_assert_held(&gpu->lock); 1078 1079 f = kzalloc(sizeof(*f), GFP_KERNEL); 1080 if (!f) 1081 return NULL; 1082 1083 f->gpu = gpu; 1084 1085 dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock, 1086 gpu->fence_context, ++gpu->next_fence); 1087 1088 return &f->base; 1089 } 1090 1091 int etnaviv_gpu_fence_sync_obj(struct etnaviv_gem_object *etnaviv_obj, 1092 unsigned int context, bool exclusive, bool explicit) 1093 { 1094 struct reservation_object *robj = etnaviv_obj->resv; 1095 struct reservation_object_list *fobj; 1096 struct dma_fence *fence; 1097 int i, ret; 1098 1099 if (!exclusive) { 1100 ret = reservation_object_reserve_shared(robj); 1101 if (ret) 1102 return ret; 1103 } 1104 1105 if (explicit) 1106 return 0; 1107 1108 /* 1109 * If we have any shared fences, then the exclusive fence 1110 * should be ignored as it will already have been signalled. 1111 */ 1112 fobj = reservation_object_get_list(robj); 1113 if (!fobj || fobj->shared_count == 0) { 1114 /* Wait on any existing exclusive fence which isn't our own */ 1115 fence = reservation_object_get_excl(robj); 1116 if (fence && fence->context != context) { 1117 ret = dma_fence_wait(fence, true); 1118 if (ret) 1119 return ret; 1120 } 1121 } 1122 1123 if (!exclusive || !fobj) 1124 return 0; 1125 1126 for (i = 0; i < fobj->shared_count; i++) { 1127 fence = rcu_dereference_protected(fobj->shared[i], 1128 reservation_object_held(robj)); 1129 if (fence->context != context) { 1130 ret = dma_fence_wait(fence, true); 1131 if (ret) 1132 return ret; 1133 } 1134 } 1135 1136 return 0; 1137 } 1138 1139 /* 1140 * event management: 1141 */ 1142 1143 static unsigned int event_alloc(struct etnaviv_gpu *gpu) 1144 { 1145 unsigned long ret, flags; 1146 unsigned int i, event = ~0U; 1147 1148 ret = wait_for_completion_timeout(&gpu->event_free, 1149 msecs_to_jiffies(10 * 10000)); 1150 if (!ret) 1151 dev_err(gpu->dev, "wait_for_completion_timeout failed"); 1152 1153 spin_lock_irqsave(&gpu->event_spinlock, flags); 1154 1155 /* find first free event */ 1156 for (i = 0; i < ARRAY_SIZE(gpu->event); i++) { 1157 if (gpu->event[i].used == false) { 1158 gpu->event[i].used = true; 1159 event = i; 1160 break; 1161 } 1162 } 1163 1164 spin_unlock_irqrestore(&gpu->event_spinlock, flags); 1165 1166 return event; 1167 } 1168 1169 static void event_free(struct etnaviv_gpu *gpu, unsigned int event) 1170 { 1171 unsigned long flags; 1172 1173 spin_lock_irqsave(&gpu->event_spinlock, flags); 1174 1175 if (gpu->event[event].used == false) { 1176 dev_warn(gpu->dev, "event %u is already marked as free", 1177 event); 1178 spin_unlock_irqrestore(&gpu->event_spinlock, flags); 1179 } else { 1180 gpu->event[event].used = false; 1181 spin_unlock_irqrestore(&gpu->event_spinlock, flags); 1182 1183 complete(&gpu->event_free); 1184 } 1185 } 1186 1187 /* 1188 * Cmdstream submission/retirement: 1189 */ 1190 1191 static void retire_worker(struct work_struct *work) 1192 { 1193 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu, 1194 retire_work); 1195 u32 fence = gpu->completed_fence; 1196 struct etnaviv_cmdbuf *cmdbuf, *tmp; 1197 unsigned int i; 1198 1199 mutex_lock(&gpu->lock); 1200 list_for_each_entry_safe(cmdbuf, tmp, &gpu->active_cmd_list, node) { 1201 if (!dma_fence_is_signaled(cmdbuf->fence)) 1202 break; 1203 1204 list_del(&cmdbuf->node); 1205 dma_fence_put(cmdbuf->fence); 1206 1207 for (i = 0; i < cmdbuf->nr_bos; i++) { 1208 struct etnaviv_vram_mapping *mapping = cmdbuf->bo_map[i]; 1209 struct etnaviv_gem_object *etnaviv_obj = mapping->object; 1210 1211 atomic_dec(&etnaviv_obj->gpu_active); 1212 /* drop the refcount taken in etnaviv_gpu_submit */ 1213 etnaviv_gem_mapping_unreference(mapping); 1214 } 1215 1216 etnaviv_cmdbuf_free(cmdbuf); 1217 /* 1218 * We need to balance the runtime PM count caused by 1219 * each submission. Upon submission, we increment 1220 * the runtime PM counter, and allocate one event. 1221 * So here, we put the runtime PM count for each 1222 * completed event. 1223 */ 1224 pm_runtime_put_autosuspend(gpu->dev); 1225 } 1226 1227 gpu->retired_fence = fence; 1228 1229 mutex_unlock(&gpu->lock); 1230 1231 wake_up_all(&gpu->fence_event); 1232 } 1233 1234 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu, 1235 u32 fence, struct timespec *timeout) 1236 { 1237 int ret; 1238 1239 if (fence_after(fence, gpu->next_fence)) { 1240 DRM_ERROR("waiting on invalid fence: %u (of %u)\n", 1241 fence, gpu->next_fence); 1242 return -EINVAL; 1243 } 1244 1245 if (!timeout) { 1246 /* No timeout was requested: just test for completion */ 1247 ret = fence_completed(gpu, fence) ? 0 : -EBUSY; 1248 } else { 1249 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout); 1250 1251 ret = wait_event_interruptible_timeout(gpu->fence_event, 1252 fence_completed(gpu, fence), 1253 remaining); 1254 if (ret == 0) { 1255 DBG("timeout waiting for fence: %u (retired: %u completed: %u)", 1256 fence, gpu->retired_fence, 1257 gpu->completed_fence); 1258 ret = -ETIMEDOUT; 1259 } else if (ret != -ERESTARTSYS) { 1260 ret = 0; 1261 } 1262 } 1263 1264 return ret; 1265 } 1266 1267 /* 1268 * Wait for an object to become inactive. This, on it's own, is not race 1269 * free: the object is moved by the retire worker off the active list, and 1270 * then the iova is put. Moreover, the object could be re-submitted just 1271 * after we notice that it's become inactive. 1272 * 1273 * Although the retirement happens under the gpu lock, we don't want to hold 1274 * that lock in this function while waiting. 1275 */ 1276 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu, 1277 struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout) 1278 { 1279 unsigned long remaining; 1280 long ret; 1281 1282 if (!timeout) 1283 return !is_active(etnaviv_obj) ? 0 : -EBUSY; 1284 1285 remaining = etnaviv_timeout_to_jiffies(timeout); 1286 1287 ret = wait_event_interruptible_timeout(gpu->fence_event, 1288 !is_active(etnaviv_obj), 1289 remaining); 1290 if (ret > 0) { 1291 struct etnaviv_drm_private *priv = gpu->drm->dev_private; 1292 1293 /* Synchronise with the retire worker */ 1294 flush_workqueue(priv->wq); 1295 return 0; 1296 } else if (ret == -ERESTARTSYS) { 1297 return -ERESTARTSYS; 1298 } else { 1299 return -ETIMEDOUT; 1300 } 1301 } 1302 1303 int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu) 1304 { 1305 return pm_runtime_get_sync(gpu->dev); 1306 } 1307 1308 void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu) 1309 { 1310 pm_runtime_mark_last_busy(gpu->dev); 1311 pm_runtime_put_autosuspend(gpu->dev); 1312 } 1313 1314 /* add bo's to gpu's ring, and kick gpu: */ 1315 int etnaviv_gpu_submit(struct etnaviv_gpu *gpu, 1316 struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf) 1317 { 1318 struct dma_fence *fence; 1319 unsigned int event, i; 1320 int ret; 1321 1322 ret = etnaviv_gpu_pm_get_sync(gpu); 1323 if (ret < 0) 1324 return ret; 1325 1326 /* 1327 * TODO 1328 * 1329 * - flush 1330 * - data endian 1331 * - prefetch 1332 * 1333 */ 1334 1335 event = event_alloc(gpu); 1336 if (unlikely(event == ~0U)) { 1337 DRM_ERROR("no free event\n"); 1338 ret = -EBUSY; 1339 goto out_pm_put; 1340 } 1341 1342 mutex_lock(&gpu->lock); 1343 1344 fence = etnaviv_gpu_fence_alloc(gpu); 1345 if (!fence) { 1346 event_free(gpu, event); 1347 ret = -ENOMEM; 1348 goto out_unlock; 1349 } 1350 1351 gpu->event[event].fence = fence; 1352 submit->fence = dma_fence_get(fence); 1353 gpu->active_fence = submit->fence->seqno; 1354 1355 if (gpu->lastctx != cmdbuf->ctx) { 1356 gpu->mmu->need_flush = true; 1357 gpu->switch_context = true; 1358 gpu->lastctx = cmdbuf->ctx; 1359 } 1360 1361 etnaviv_buffer_queue(gpu, event, cmdbuf); 1362 1363 cmdbuf->fence = fence; 1364 list_add_tail(&cmdbuf->node, &gpu->active_cmd_list); 1365 1366 /* We're committed to adding this command buffer, hold a PM reference */ 1367 pm_runtime_get_noresume(gpu->dev); 1368 1369 for (i = 0; i < submit->nr_bos; i++) { 1370 struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj; 1371 1372 /* Each cmdbuf takes a refcount on the mapping */ 1373 etnaviv_gem_mapping_reference(submit->bos[i].mapping); 1374 cmdbuf->bo_map[i] = submit->bos[i].mapping; 1375 atomic_inc(&etnaviv_obj->gpu_active); 1376 1377 if (submit->bos[i].flags & ETNA_SUBMIT_BO_WRITE) 1378 reservation_object_add_excl_fence(etnaviv_obj->resv, 1379 fence); 1380 else 1381 reservation_object_add_shared_fence(etnaviv_obj->resv, 1382 fence); 1383 } 1384 cmdbuf->nr_bos = submit->nr_bos; 1385 hangcheck_timer_reset(gpu); 1386 ret = 0; 1387 1388 out_unlock: 1389 mutex_unlock(&gpu->lock); 1390 1391 out_pm_put: 1392 etnaviv_gpu_pm_put(gpu); 1393 1394 return ret; 1395 } 1396 1397 /* 1398 * Init/Cleanup: 1399 */ 1400 static irqreturn_t irq_handler(int irq, void *data) 1401 { 1402 struct etnaviv_gpu *gpu = data; 1403 irqreturn_t ret = IRQ_NONE; 1404 1405 u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE); 1406 1407 if (intr != 0) { 1408 int event; 1409 1410 pm_runtime_mark_last_busy(gpu->dev); 1411 1412 dev_dbg(gpu->dev, "intr 0x%08x\n", intr); 1413 1414 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) { 1415 dev_err(gpu->dev, "AXI bus error\n"); 1416 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR; 1417 } 1418 1419 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) { 1420 int i; 1421 1422 dev_err_ratelimited(gpu->dev, 1423 "MMU fault status 0x%08x\n", 1424 gpu_read(gpu, VIVS_MMUv2_STATUS)); 1425 for (i = 0; i < 4; i++) { 1426 dev_err_ratelimited(gpu->dev, 1427 "MMU %d fault addr 0x%08x\n", 1428 i, gpu_read(gpu, 1429 VIVS_MMUv2_EXCEPTION_ADDR(i))); 1430 } 1431 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION; 1432 } 1433 1434 while ((event = ffs(intr)) != 0) { 1435 struct dma_fence *fence; 1436 1437 event -= 1; 1438 1439 intr &= ~(1 << event); 1440 1441 dev_dbg(gpu->dev, "event %u\n", event); 1442 1443 fence = gpu->event[event].fence; 1444 gpu->event[event].fence = NULL; 1445 dma_fence_signal(fence); 1446 1447 /* 1448 * Events can be processed out of order. Eg, 1449 * - allocate and queue event 0 1450 * - allocate event 1 1451 * - event 0 completes, we process it 1452 * - allocate and queue event 0 1453 * - event 1 and event 0 complete 1454 * we can end up processing event 0 first, then 1. 1455 */ 1456 if (fence_after(fence->seqno, gpu->completed_fence)) 1457 gpu->completed_fence = fence->seqno; 1458 1459 event_free(gpu, event); 1460 } 1461 1462 /* Retire the buffer objects in a work */ 1463 etnaviv_queue_work(gpu->drm, &gpu->retire_work); 1464 1465 ret = IRQ_HANDLED; 1466 } 1467 1468 return ret; 1469 } 1470 1471 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu) 1472 { 1473 int ret; 1474 1475 if (gpu->clk_bus) { 1476 ret = clk_prepare_enable(gpu->clk_bus); 1477 if (ret) 1478 return ret; 1479 } 1480 1481 if (gpu->clk_core) { 1482 ret = clk_prepare_enable(gpu->clk_core); 1483 if (ret) 1484 goto disable_clk_bus; 1485 } 1486 1487 if (gpu->clk_shader) { 1488 ret = clk_prepare_enable(gpu->clk_shader); 1489 if (ret) 1490 goto disable_clk_core; 1491 } 1492 1493 return 0; 1494 1495 disable_clk_core: 1496 if (gpu->clk_core) 1497 clk_disable_unprepare(gpu->clk_core); 1498 disable_clk_bus: 1499 if (gpu->clk_bus) 1500 clk_disable_unprepare(gpu->clk_bus); 1501 1502 return ret; 1503 } 1504 1505 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu) 1506 { 1507 if (gpu->clk_shader) 1508 clk_disable_unprepare(gpu->clk_shader); 1509 if (gpu->clk_core) 1510 clk_disable_unprepare(gpu->clk_core); 1511 if (gpu->clk_bus) 1512 clk_disable_unprepare(gpu->clk_bus); 1513 1514 return 0; 1515 } 1516 1517 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms) 1518 { 1519 unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms); 1520 1521 do { 1522 u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); 1523 1524 if ((idle & gpu->idle_mask) == gpu->idle_mask) 1525 return 0; 1526 1527 if (time_is_before_jiffies(timeout)) { 1528 dev_warn(gpu->dev, 1529 "timed out waiting for idle: idle=0x%x\n", 1530 idle); 1531 return -ETIMEDOUT; 1532 } 1533 1534 udelay(5); 1535 } while (1); 1536 } 1537 1538 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu) 1539 { 1540 if (gpu->buffer) { 1541 /* Replace the last WAIT with END */ 1542 etnaviv_buffer_end(gpu); 1543 1544 /* 1545 * We know that only the FE is busy here, this should 1546 * happen quickly (as the WAIT is only 200 cycles). If 1547 * we fail, just warn and continue. 1548 */ 1549 etnaviv_gpu_wait_idle(gpu, 100); 1550 } 1551 1552 return etnaviv_gpu_clk_disable(gpu); 1553 } 1554 1555 #ifdef CONFIG_PM 1556 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu) 1557 { 1558 int ret; 1559 1560 ret = mutex_lock_killable(&gpu->lock); 1561 if (ret) 1562 return ret; 1563 1564 etnaviv_gpu_update_clock(gpu); 1565 etnaviv_gpu_hw_init(gpu); 1566 1567 gpu->switch_context = true; 1568 gpu->exec_state = -1; 1569 1570 mutex_unlock(&gpu->lock); 1571 1572 return 0; 1573 } 1574 #endif 1575 1576 static int 1577 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev, 1578 unsigned long *state) 1579 { 1580 *state = 6; 1581 1582 return 0; 1583 } 1584 1585 static int 1586 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev, 1587 unsigned long *state) 1588 { 1589 struct etnaviv_gpu *gpu = cdev->devdata; 1590 1591 *state = gpu->freq_scale; 1592 1593 return 0; 1594 } 1595 1596 static int 1597 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev, 1598 unsigned long state) 1599 { 1600 struct etnaviv_gpu *gpu = cdev->devdata; 1601 1602 mutex_lock(&gpu->lock); 1603 gpu->freq_scale = state; 1604 if (!pm_runtime_suspended(gpu->dev)) 1605 etnaviv_gpu_update_clock(gpu); 1606 mutex_unlock(&gpu->lock); 1607 1608 return 0; 1609 } 1610 1611 static struct thermal_cooling_device_ops cooling_ops = { 1612 .get_max_state = etnaviv_gpu_cooling_get_max_state, 1613 .get_cur_state = etnaviv_gpu_cooling_get_cur_state, 1614 .set_cur_state = etnaviv_gpu_cooling_set_cur_state, 1615 }; 1616 1617 static int etnaviv_gpu_bind(struct device *dev, struct device *master, 1618 void *data) 1619 { 1620 struct drm_device *drm = data; 1621 struct etnaviv_drm_private *priv = drm->dev_private; 1622 struct etnaviv_gpu *gpu = dev_get_drvdata(dev); 1623 int ret; 1624 1625 if (IS_ENABLED(CONFIG_THERMAL)) { 1626 gpu->cooling = thermal_of_cooling_device_register(dev->of_node, 1627 (char *)dev_name(dev), gpu, &cooling_ops); 1628 if (IS_ERR(gpu->cooling)) 1629 return PTR_ERR(gpu->cooling); 1630 } 1631 1632 #ifdef CONFIG_PM 1633 ret = pm_runtime_get_sync(gpu->dev); 1634 #else 1635 ret = etnaviv_gpu_clk_enable(gpu); 1636 #endif 1637 if (ret < 0) { 1638 thermal_cooling_device_unregister(gpu->cooling); 1639 return ret; 1640 } 1641 1642 gpu->drm = drm; 1643 gpu->fence_context = dma_fence_context_alloc(1); 1644 spin_lock_init(&gpu->fence_spinlock); 1645 1646 INIT_LIST_HEAD(&gpu->active_cmd_list); 1647 INIT_WORK(&gpu->retire_work, retire_worker); 1648 INIT_WORK(&gpu->recover_work, recover_worker); 1649 init_waitqueue_head(&gpu->fence_event); 1650 1651 setup_deferrable_timer(&gpu->hangcheck_timer, hangcheck_handler, 1652 (unsigned long)gpu); 1653 1654 priv->gpu[priv->num_gpus++] = gpu; 1655 1656 pm_runtime_mark_last_busy(gpu->dev); 1657 pm_runtime_put_autosuspend(gpu->dev); 1658 1659 return 0; 1660 } 1661 1662 static void etnaviv_gpu_unbind(struct device *dev, struct device *master, 1663 void *data) 1664 { 1665 struct etnaviv_gpu *gpu = dev_get_drvdata(dev); 1666 1667 DBG("%s", dev_name(gpu->dev)); 1668 1669 hangcheck_disable(gpu); 1670 1671 #ifdef CONFIG_PM 1672 pm_runtime_get_sync(gpu->dev); 1673 pm_runtime_put_sync_suspend(gpu->dev); 1674 #else 1675 etnaviv_gpu_hw_suspend(gpu); 1676 #endif 1677 1678 if (gpu->buffer) { 1679 etnaviv_cmdbuf_free(gpu->buffer); 1680 gpu->buffer = NULL; 1681 } 1682 1683 if (gpu->cmdbuf_suballoc) { 1684 etnaviv_cmdbuf_suballoc_destroy(gpu->cmdbuf_suballoc); 1685 gpu->cmdbuf_suballoc = NULL; 1686 } 1687 1688 if (gpu->mmu) { 1689 etnaviv_iommu_destroy(gpu->mmu); 1690 gpu->mmu = NULL; 1691 } 1692 1693 gpu->drm = NULL; 1694 1695 thermal_cooling_device_unregister(gpu->cooling); 1696 gpu->cooling = NULL; 1697 } 1698 1699 static const struct component_ops gpu_ops = { 1700 .bind = etnaviv_gpu_bind, 1701 .unbind = etnaviv_gpu_unbind, 1702 }; 1703 1704 static const struct of_device_id etnaviv_gpu_match[] = { 1705 { 1706 .compatible = "vivante,gc" 1707 }, 1708 { /* sentinel */ } 1709 }; 1710 1711 static int etnaviv_gpu_platform_probe(struct platform_device *pdev) 1712 { 1713 struct device *dev = &pdev->dev; 1714 struct etnaviv_gpu *gpu; 1715 int err; 1716 1717 gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL); 1718 if (!gpu) 1719 return -ENOMEM; 1720 1721 gpu->dev = &pdev->dev; 1722 mutex_init(&gpu->lock); 1723 1724 /* Map registers: */ 1725 gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev)); 1726 if (IS_ERR(gpu->mmio)) 1727 return PTR_ERR(gpu->mmio); 1728 1729 /* Get Interrupt: */ 1730 gpu->irq = platform_get_irq(pdev, 0); 1731 if (gpu->irq < 0) { 1732 dev_err(dev, "failed to get irq: %d\n", gpu->irq); 1733 return gpu->irq; 1734 } 1735 1736 err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0, 1737 dev_name(gpu->dev), gpu); 1738 if (err) { 1739 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err); 1740 return err; 1741 } 1742 1743 /* Get Clocks: */ 1744 gpu->clk_bus = devm_clk_get(&pdev->dev, "bus"); 1745 DBG("clk_bus: %p", gpu->clk_bus); 1746 if (IS_ERR(gpu->clk_bus)) 1747 gpu->clk_bus = NULL; 1748 1749 gpu->clk_core = devm_clk_get(&pdev->dev, "core"); 1750 DBG("clk_core: %p", gpu->clk_core); 1751 if (IS_ERR(gpu->clk_core)) 1752 gpu->clk_core = NULL; 1753 gpu->base_rate_core = clk_get_rate(gpu->clk_core); 1754 1755 gpu->clk_shader = devm_clk_get(&pdev->dev, "shader"); 1756 DBG("clk_shader: %p", gpu->clk_shader); 1757 if (IS_ERR(gpu->clk_shader)) 1758 gpu->clk_shader = NULL; 1759 gpu->base_rate_shader = clk_get_rate(gpu->clk_shader); 1760 1761 /* TODO: figure out max mapped size */ 1762 dev_set_drvdata(dev, gpu); 1763 1764 /* 1765 * We treat the device as initially suspended. The runtime PM 1766 * autosuspend delay is rather arbitary: no measurements have 1767 * yet been performed to determine an appropriate value. 1768 */ 1769 pm_runtime_use_autosuspend(gpu->dev); 1770 pm_runtime_set_autosuspend_delay(gpu->dev, 200); 1771 pm_runtime_enable(gpu->dev); 1772 1773 err = component_add(&pdev->dev, &gpu_ops); 1774 if (err < 0) { 1775 dev_err(&pdev->dev, "failed to register component: %d\n", err); 1776 return err; 1777 } 1778 1779 return 0; 1780 } 1781 1782 static int etnaviv_gpu_platform_remove(struct platform_device *pdev) 1783 { 1784 component_del(&pdev->dev, &gpu_ops); 1785 pm_runtime_disable(&pdev->dev); 1786 return 0; 1787 } 1788 1789 #ifdef CONFIG_PM 1790 static int etnaviv_gpu_rpm_suspend(struct device *dev) 1791 { 1792 struct etnaviv_gpu *gpu = dev_get_drvdata(dev); 1793 u32 idle, mask; 1794 1795 /* If we have outstanding fences, we're not idle */ 1796 if (gpu->completed_fence != gpu->active_fence) 1797 return -EBUSY; 1798 1799 /* Check whether the hardware (except FE) is idle */ 1800 mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE; 1801 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask; 1802 if (idle != mask) 1803 return -EBUSY; 1804 1805 return etnaviv_gpu_hw_suspend(gpu); 1806 } 1807 1808 static int etnaviv_gpu_rpm_resume(struct device *dev) 1809 { 1810 struct etnaviv_gpu *gpu = dev_get_drvdata(dev); 1811 int ret; 1812 1813 ret = etnaviv_gpu_clk_enable(gpu); 1814 if (ret) 1815 return ret; 1816 1817 /* Re-initialise the basic hardware state */ 1818 if (gpu->drm && gpu->buffer) { 1819 ret = etnaviv_gpu_hw_resume(gpu); 1820 if (ret) { 1821 etnaviv_gpu_clk_disable(gpu); 1822 return ret; 1823 } 1824 } 1825 1826 return 0; 1827 } 1828 #endif 1829 1830 static const struct dev_pm_ops etnaviv_gpu_pm_ops = { 1831 SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume, 1832 NULL) 1833 }; 1834 1835 struct platform_driver etnaviv_gpu_driver = { 1836 .driver = { 1837 .name = "etnaviv-gpu", 1838 .owner = THIS_MODULE, 1839 .pm = &etnaviv_gpu_pm_ops, 1840 .of_match_table = etnaviv_gpu_match, 1841 }, 1842 .probe = etnaviv_gpu_platform_probe, 1843 .remove = etnaviv_gpu_platform_remove, 1844 .id_table = gpu_ids, 1845 }; 1846