1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright 2018 Marty E. Plummer <hanetzer@startmail.com> */ 3 /* Copyright 2019 Linaro, Ltd., Rob Herring <robh@kernel.org> */ 4 /* Copyright 2019 Collabora ltd. */ 5 #include <linux/bitfield.h> 6 #include <linux/bitmap.h> 7 #include <linux/delay.h> 8 #include <linux/dma-mapping.h> 9 #include <linux/interrupt.h> 10 #include <linux/io.h> 11 #include <linux/iopoll.h> 12 #include <linux/platform_device.h> 13 #include <linux/pm_runtime.h> 14 15 #include <drm/drm_print.h> 16 17 #include "panfrost_device.h" 18 #include "panfrost_features.h" 19 #include "panfrost_issues.h" 20 #include "panfrost_gpu.h" 21 #include "panfrost_perfcnt.h" 22 #include "panfrost_regs.h" 23 24 static irqreturn_t panfrost_gpu_irq_handler(int irq, void *data) 25 { 26 struct panfrost_device *pfdev = data; 27 u32 fault_status, state; 28 29 if (test_bit(PANFROST_COMP_BIT_GPU, pfdev->is_suspended)) 30 return IRQ_NONE; 31 32 fault_status = gpu_read(pfdev, GPU_FAULT_STATUS); 33 state = gpu_read(pfdev, GPU_INT_STAT); 34 if (!state) 35 return IRQ_NONE; 36 37 if (state & GPU_IRQ_MASK_ERROR) { 38 u64 address = (u64) gpu_read(pfdev, GPU_FAULT_ADDRESS_HI) << 32; 39 address |= gpu_read(pfdev, GPU_FAULT_ADDRESS_LO); 40 41 dev_warn(pfdev->base.dev, "GPU Fault 0x%08x (%s) at 0x%016llx\n", 42 fault_status, panfrost_exception_name(fault_status & 0xFF), 43 address); 44 45 if (state & GPU_IRQ_MULTIPLE_FAULT) 46 dev_warn(pfdev->base.dev, "There were multiple GPU faults - some have not been reported\n"); 47 48 gpu_write(pfdev, GPU_INT_MASK, 0); 49 } 50 51 if (state & GPU_IRQ_PERFCNT_SAMPLE_COMPLETED) 52 panfrost_perfcnt_sample_done(pfdev); 53 54 if (state & GPU_IRQ_CLEAN_CACHES_COMPLETED) 55 panfrost_perfcnt_clean_cache_done(pfdev); 56 57 gpu_write(pfdev, GPU_INT_CLEAR, state); 58 59 return IRQ_HANDLED; 60 } 61 62 int panfrost_gpu_soft_reset(struct panfrost_device *pfdev) 63 { 64 int ret; 65 u32 val; 66 67 gpu_write(pfdev, GPU_INT_MASK, 0); 68 gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_RESET_COMPLETED); 69 70 clear_bit(PANFROST_COMP_BIT_GPU, pfdev->is_suspended); 71 72 gpu_write(pfdev, GPU_CMD, GPU_CMD_SOFT_RESET); 73 ret = readl_relaxed_poll_timeout(pfdev->iomem + GPU_INT_RAWSTAT, 74 val, val & GPU_IRQ_RESET_COMPLETED, 10, 10000); 75 76 if (ret) { 77 dev_err(pfdev->base.dev, "gpu soft reset timed out, attempting hard reset\n"); 78 79 gpu_write(pfdev, GPU_CMD, GPU_CMD_HARD_RESET); 80 ret = readl_relaxed_poll_timeout(pfdev->iomem + GPU_INT_RAWSTAT, val, 81 val & GPU_IRQ_RESET_COMPLETED, 100, 10000); 82 if (ret) { 83 dev_err(pfdev->base.dev, "gpu hard reset timed out\n"); 84 return ret; 85 } 86 } 87 88 gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL); 89 90 /* Only enable the interrupts we care about */ 91 gpu_write(pfdev, GPU_INT_MASK, 92 GPU_IRQ_MASK_ERROR | 93 GPU_IRQ_PERFCNT_SAMPLE_COMPLETED | 94 GPU_IRQ_CLEAN_CACHES_COMPLETED); 95 96 /* 97 * All in-flight jobs should have released their cycle 98 * counter references upon reset, but let us make sure 99 */ 100 if (drm_WARN_ON(&pfdev->base, atomic_read(&pfdev->cycle_counter.use_count) != 0)) 101 atomic_set(&pfdev->cycle_counter.use_count, 0); 102 103 return 0; 104 } 105 106 void panfrost_gpu_amlogic_quirk(struct panfrost_device *pfdev) 107 { 108 /* 109 * The Amlogic integrated Mali-T820, Mali-G31 & Mali-G52 needs 110 * these undocumented bits in GPU_PWR_OVERRIDE1 to be set in order 111 * to operate correctly. 112 */ 113 gpu_write(pfdev, GPU_PWR_KEY, GPU_PWR_KEY_UNLOCK); 114 gpu_write(pfdev, GPU_PWR_OVERRIDE1, 0xfff | (0x20 << 16)); 115 } 116 117 static void panfrost_gpu_init_quirks(struct panfrost_device *pfdev) 118 { 119 u32 quirks = 0; 120 121 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8443) || 122 panfrost_has_hw_issue(pfdev, HW_ISSUE_11035)) 123 quirks |= SC_LS_PAUSEBUFFER_DISABLE; 124 125 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10327)) 126 quirks |= SC_SDC_DISABLE_OQ_DISCARD; 127 128 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10797)) 129 quirks |= SC_ENABLE_TEXGRD_FLAGS; 130 131 if (!panfrost_has_hw_issue(pfdev, GPUCORE_1619)) { 132 if (panfrost_model_cmp(pfdev, 0x750) < 0) /* T60x, T62x, T72x */ 133 quirks |= SC_LS_ATTR_CHECK_DISABLE; 134 else if (panfrost_model_cmp(pfdev, 0x880) <= 0) /* T76x, T8xx */ 135 quirks |= SC_LS_ALLOW_ATTR_TYPES; 136 } 137 138 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_TTRX_2968_TTRX_3162)) 139 quirks |= SC_VAR_ALGORITHM; 140 141 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_TLS_HASHING)) 142 quirks |= SC_TLS_HASH_ENABLE; 143 144 if (quirks) 145 gpu_write(pfdev, GPU_SHADER_CONFIG, quirks); 146 147 148 quirks = gpu_read(pfdev, GPU_TILER_CONFIG); 149 150 /* Set tiler clock gate override if required */ 151 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_T76X_3953)) 152 quirks |= TC_CLOCK_GATE_OVERRIDE; 153 154 gpu_write(pfdev, GPU_TILER_CONFIG, quirks); 155 156 157 quirks = 0; 158 if ((panfrost_model_eq(pfdev, 0x860) || panfrost_model_eq(pfdev, 0x880)) && 159 pfdev->features.revision >= 0x2000) 160 quirks |= JM_MAX_JOB_THROTTLE_LIMIT << JM_JOB_THROTTLE_LIMIT_SHIFT; 161 else if (panfrost_model_eq(pfdev, 0x6000) && 162 pfdev->features.coherency_features == COHERENCY_ACE) 163 quirks |= (COHERENCY_ACE_LITE | COHERENCY_ACE) << 164 JM_FORCE_COHERENCY_FEATURES_SHIFT; 165 166 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_IDVS_GROUP_SIZE)) 167 quirks |= JM_DEFAULT_IDVS_GROUP_SIZE << JM_IDVS_GROUP_SIZE_SHIFT; 168 169 if (quirks) 170 gpu_write(pfdev, GPU_JM_CONFIG, quirks); 171 172 /* Here goes platform specific quirks */ 173 if (pfdev->comp->vendor_quirk) 174 pfdev->comp->vendor_quirk(pfdev); 175 } 176 177 #define MAX_HW_REVS 6 178 179 struct panfrost_model { 180 const char *name; 181 u32 id; 182 u64 features; 183 u64 issues; 184 struct { 185 u32 revision; 186 u64 issues; 187 } revs[MAX_HW_REVS]; 188 }; 189 190 #define GPU_MODEL(_name, _id, ...) \ 191 {\ 192 .name = __stringify(_name), \ 193 .id = _id, \ 194 .features = hw_features_##_name, \ 195 .issues = hw_issues_##_name, \ 196 .revs = { __VA_ARGS__ }, \ 197 } 198 199 #define GPU_REV_EXT(name, _rev, _p, _s, stat) \ 200 {\ 201 .revision = (_rev) << 12 | (_p) << 4 | (_s), \ 202 .issues = hw_issues_##name##_r##_rev##p##_p##stat, \ 203 } 204 #define GPU_REV(name, r, p) GPU_REV_EXT(name, r, p, 0, ) 205 206 static const struct panfrost_model gpu_models[] = { 207 /* T60x has an oddball version */ 208 GPU_MODEL(t600, 0x600, 209 GPU_REV_EXT(t600, 0, 0, 1, _15dev0)), 210 GPU_MODEL(t620, 0x620, 211 GPU_REV(t620, 0, 1), GPU_REV(t620, 1, 0)), 212 GPU_MODEL(t720, 0x720), 213 GPU_MODEL(t760, 0x750, 214 GPU_REV(t760, 0, 0), GPU_REV(t760, 0, 1), 215 GPU_REV_EXT(t760, 0, 1, 0, _50rel0), 216 GPU_REV(t760, 0, 2), GPU_REV(t760, 0, 3)), 217 GPU_MODEL(t820, 0x820), 218 GPU_MODEL(t830, 0x830), 219 GPU_MODEL(t860, 0x860), 220 GPU_MODEL(t880, 0x880), 221 222 GPU_MODEL(g71, 0x6000, 223 GPU_REV_EXT(g71, 0, 0, 1, _05dev0)), 224 GPU_MODEL(g72, 0x6001), 225 GPU_MODEL(g51, 0x7000), 226 GPU_MODEL(g76, 0x7001), 227 GPU_MODEL(g52, 0x7002), 228 GPU_MODEL(g31, 0x7003, 229 GPU_REV(g31, 1, 0)), 230 231 GPU_MODEL(g57, 0x9001, 232 GPU_REV(g57, 0, 0)), 233 234 /* MediaTek MT8192 has a Mali-G57 with a different GPU ID from the 235 * standard. Arm's driver does not appear to handle this model. 236 * ChromeOS has a hack downstream for it. Treat it as equivalent to 237 * standard Mali-G57 for now. 238 */ 239 GPU_MODEL(g57, 0x9003, 240 GPU_REV(g57, 0, 0)), 241 242 /* MediaTek MT8188 Mali-G57 MC3 */ 243 GPU_MODEL(g57, 0x9093, 244 GPU_REV(g57, 0, 0)), 245 {0}, 246 }; 247 248 static int panfrost_gpu_init_features(struct panfrost_device *pfdev) 249 { 250 u32 gpu_id, num_js, major, minor, status, rev; 251 const char *name = "unknown"; 252 u64 hw_feat = 0; 253 u64 hw_issues = hw_issues_all; 254 const struct panfrost_model *model; 255 int i; 256 257 pfdev->features.l2_features = gpu_read(pfdev, GPU_L2_FEATURES); 258 pfdev->features.core_features = gpu_read(pfdev, GPU_CORE_FEATURES); 259 pfdev->features.tiler_features = gpu_read(pfdev, GPU_TILER_FEATURES); 260 pfdev->features.mem_features = gpu_read(pfdev, GPU_MEM_FEATURES); 261 pfdev->features.mmu_features = gpu_read(pfdev, GPU_MMU_FEATURES); 262 pfdev->features.thread_features = gpu_read(pfdev, GPU_THREAD_FEATURES); 263 pfdev->features.max_threads = gpu_read(pfdev, GPU_THREAD_MAX_THREADS); 264 pfdev->features.thread_max_workgroup_sz = gpu_read(pfdev, GPU_THREAD_MAX_WORKGROUP_SIZE); 265 pfdev->features.thread_max_barrier_sz = gpu_read(pfdev, GPU_THREAD_MAX_BARRIER_SIZE); 266 pfdev->features.coherency_features = gpu_read(pfdev, GPU_COHERENCY_FEATURES); 267 pfdev->features.afbc_features = gpu_read(pfdev, GPU_AFBC_FEATURES); 268 for (i = 0; i < 4; i++) 269 pfdev->features.texture_features[i] = gpu_read(pfdev, GPU_TEXTURE_FEATURES(i)); 270 271 pfdev->features.as_present = gpu_read(pfdev, GPU_AS_PRESENT); 272 273 pfdev->features.js_present = gpu_read(pfdev, GPU_JS_PRESENT); 274 num_js = hweight32(pfdev->features.js_present); 275 for (i = 0; i < num_js; i++) 276 pfdev->features.js_features[i] = gpu_read(pfdev, GPU_JS_FEATURES(i)); 277 278 pfdev->features.shader_present = gpu_read(pfdev, GPU_SHADER_PRESENT_LO); 279 pfdev->features.shader_present |= (u64)gpu_read(pfdev, GPU_SHADER_PRESENT_HI) << 32; 280 281 pfdev->features.tiler_present = gpu_read(pfdev, GPU_TILER_PRESENT_LO); 282 pfdev->features.tiler_present |= (u64)gpu_read(pfdev, GPU_TILER_PRESENT_HI) << 32; 283 284 pfdev->features.l2_present = gpu_read(pfdev, GPU_L2_PRESENT_LO); 285 pfdev->features.l2_present |= (u64)gpu_read(pfdev, GPU_L2_PRESENT_HI) << 32; 286 pfdev->features.nr_core_groups = hweight64(pfdev->features.l2_present); 287 288 pfdev->features.stack_present = gpu_read(pfdev, GPU_STACK_PRESENT_LO); 289 pfdev->features.stack_present |= (u64)gpu_read(pfdev, GPU_STACK_PRESENT_HI) << 32; 290 291 pfdev->features.thread_tls_alloc = gpu_read(pfdev, GPU_THREAD_TLS_ALLOC); 292 293 gpu_id = gpu_read(pfdev, GPU_ID); 294 pfdev->features.revision = gpu_id & 0xffff; 295 pfdev->features.id = gpu_id >> 16; 296 297 /* The T60x has an oddball ID value. Fix it up to the standard Midgard 298 * format so we (and userspace) don't have to special case it. 299 */ 300 if (pfdev->features.id == 0x6956) 301 pfdev->features.id = 0x0600; 302 303 major = (pfdev->features.revision >> 12) & 0xf; 304 minor = (pfdev->features.revision >> 4) & 0xff; 305 status = pfdev->features.revision & 0xf; 306 rev = pfdev->features.revision; 307 308 gpu_id = pfdev->features.id; 309 310 for (model = gpu_models; model->name; model++) { 311 int best = -1; 312 313 if (!panfrost_model_eq(pfdev, model->id)) 314 continue; 315 316 name = model->name; 317 hw_feat = model->features; 318 hw_issues |= model->issues; 319 for (i = 0; i < MAX_HW_REVS; i++) { 320 if (model->revs[i].revision == rev) { 321 best = i; 322 break; 323 } else if (model->revs[i].revision == (rev & ~0xf)) 324 best = i; 325 } 326 327 if (best >= 0) 328 hw_issues |= model->revs[best].issues; 329 330 break; 331 } 332 333 if (!model->name) { 334 dev_err(pfdev->base.dev, "GPU model not found: mali-%s id rev %#x %#x\n", 335 name, gpu_id, rev); 336 return -ENODEV; 337 } 338 339 bitmap_from_u64(pfdev->features.hw_features, hw_feat); 340 bitmap_from_u64(pfdev->features.hw_issues, hw_issues); 341 342 dev_info(pfdev->base.dev, "mali-%s id 0x%x major 0x%x minor 0x%x status 0x%x", 343 name, gpu_id, major, minor, status); 344 dev_info(pfdev->base.dev, "features: %64pb, issues: %64pb", 345 pfdev->features.hw_features, 346 pfdev->features.hw_issues); 347 348 dev_info(pfdev->base.dev, "Features: L2:0x%08x Shader:0x%08x Tiler:0x%08x Mem:0x%0x MMU:0x%08x AS:0x%x JS:0x%x", 349 pfdev->features.l2_features, 350 pfdev->features.core_features, 351 pfdev->features.tiler_features, 352 pfdev->features.mem_features, 353 pfdev->features.mmu_features, 354 pfdev->features.as_present, 355 pfdev->features.js_present); 356 357 dev_info(pfdev->base.dev, "shader_present=0x%0llx l2_present=0x%0llx", 358 pfdev->features.shader_present, pfdev->features.l2_present); 359 360 return 0; 361 } 362 363 void panfrost_cycle_counter_get(struct panfrost_device *pfdev) 364 { 365 if (atomic_inc_not_zero(&pfdev->cycle_counter.use_count)) 366 return; 367 368 spin_lock(&pfdev->cycle_counter.lock); 369 if (atomic_inc_return(&pfdev->cycle_counter.use_count) == 1) 370 gpu_write(pfdev, GPU_CMD, GPU_CMD_CYCLE_COUNT_START); 371 spin_unlock(&pfdev->cycle_counter.lock); 372 } 373 374 void panfrost_cycle_counter_put(struct panfrost_device *pfdev) 375 { 376 if (atomic_add_unless(&pfdev->cycle_counter.use_count, -1, 1)) 377 return; 378 379 spin_lock(&pfdev->cycle_counter.lock); 380 if (atomic_dec_return(&pfdev->cycle_counter.use_count) == 0) 381 gpu_write(pfdev, GPU_CMD, GPU_CMD_CYCLE_COUNT_STOP); 382 spin_unlock(&pfdev->cycle_counter.lock); 383 } 384 385 unsigned long long panfrost_cycle_counter_read(struct panfrost_device *pfdev) 386 { 387 u32 hi, lo; 388 389 do { 390 hi = gpu_read(pfdev, GPU_CYCLE_COUNT_HI); 391 lo = gpu_read(pfdev, GPU_CYCLE_COUNT_LO); 392 } while (hi != gpu_read(pfdev, GPU_CYCLE_COUNT_HI)); 393 394 return ((u64)hi << 32) | lo; 395 } 396 397 unsigned long long panfrost_timestamp_read(struct panfrost_device *pfdev) 398 { 399 u32 hi, lo; 400 401 do { 402 hi = gpu_read(pfdev, GPU_TIMESTAMP_HI); 403 lo = gpu_read(pfdev, GPU_TIMESTAMP_LO); 404 } while (hi != gpu_read(pfdev, GPU_TIMESTAMP_HI)); 405 406 return ((u64)hi << 32) | lo; 407 } 408 409 static u64 panfrost_get_core_mask(struct panfrost_device *pfdev) 410 { 411 u64 core_mask; 412 413 if (pfdev->features.l2_present == 1) 414 return U64_MAX; 415 416 /* 417 * Only support one core group now. 418 * ~(l2_present - 1) unsets all bits in l2_present except 419 * the bottom bit. (l2_present - 2) has all the bits in 420 * the first core group set. AND them together to generate 421 * a mask of cores in the first core group. 422 */ 423 core_mask = ~(pfdev->features.l2_present - 1) & 424 (pfdev->features.l2_present - 2); 425 dev_info_once(pfdev->base.dev, "using only 1st core group (%lu cores from %lu)\n", 426 hweight64(core_mask), 427 hweight64(pfdev->features.shader_present)); 428 429 return core_mask; 430 } 431 432 void panfrost_gpu_power_on(struct panfrost_device *pfdev) 433 { 434 int ret; 435 u32 val; 436 u64 core_mask; 437 438 panfrost_gpu_init_quirks(pfdev); 439 core_mask = panfrost_get_core_mask(pfdev); 440 441 gpu_write(pfdev, L2_PWRON_LO, pfdev->features.l2_present & core_mask); 442 ret = readl_relaxed_poll_timeout(pfdev->iomem + L2_READY_LO, 443 val, val == (pfdev->features.l2_present & core_mask), 444 10, 20000); 445 if (ret) 446 dev_err(pfdev->base.dev, "error powering up gpu L2"); 447 448 gpu_write(pfdev, SHADER_PWRON_LO, 449 pfdev->features.shader_present & core_mask); 450 ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_READY_LO, 451 val, val == (pfdev->features.shader_present & core_mask), 452 10, 20000); 453 if (ret) 454 dev_err(pfdev->base.dev, "error powering up gpu shader"); 455 456 gpu_write(pfdev, TILER_PWRON_LO, pfdev->features.tiler_present); 457 ret = readl_relaxed_poll_timeout(pfdev->iomem + TILER_READY_LO, 458 val, val == pfdev->features.tiler_present, 10, 1000); 459 if (ret) 460 dev_err(pfdev->base.dev, "error powering up gpu tiler"); 461 } 462 463 void panfrost_gpu_power_off(struct panfrost_device *pfdev) 464 { 465 int ret; 466 u32 val; 467 468 gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present); 469 ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_PWRTRANS_LO, 470 val, !val, 1, 2000); 471 if (ret) 472 dev_err(pfdev->base.dev, "shader power transition timeout"); 473 474 gpu_write(pfdev, TILER_PWROFF_LO, pfdev->features.tiler_present); 475 ret = readl_relaxed_poll_timeout(pfdev->iomem + TILER_PWRTRANS_LO, 476 val, !val, 1, 2000); 477 if (ret) 478 dev_err(pfdev->base.dev, "tiler power transition timeout"); 479 480 gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present); 481 ret = readl_poll_timeout(pfdev->iomem + L2_PWRTRANS_LO, 482 val, !val, 0, 2000); 483 if (ret) 484 dev_err(pfdev->base.dev, "l2 power transition timeout"); 485 } 486 487 void panfrost_gpu_suspend_irq(struct panfrost_device *pfdev) 488 { 489 set_bit(PANFROST_COMP_BIT_GPU, pfdev->is_suspended); 490 491 gpu_write(pfdev, GPU_INT_MASK, 0); 492 synchronize_irq(pfdev->gpu_irq); 493 } 494 495 int panfrost_gpu_init(struct panfrost_device *pfdev) 496 { 497 int err; 498 499 err = panfrost_gpu_soft_reset(pfdev); 500 if (err) 501 return err; 502 503 err = panfrost_gpu_init_features(pfdev); 504 if (err) 505 return err; 506 507 err = dma_set_mask_and_coherent(pfdev->base.dev, 508 DMA_BIT_MASK(FIELD_GET(0xff00, 509 pfdev->features.mmu_features))); 510 if (err) 511 return err; 512 513 dma_set_max_seg_size(pfdev->base.dev, UINT_MAX); 514 515 pfdev->gpu_irq = platform_get_irq_byname(to_platform_device(pfdev->base.dev), "gpu"); 516 if (pfdev->gpu_irq < 0) 517 return pfdev->gpu_irq; 518 519 err = devm_request_irq(pfdev->base.dev, pfdev->gpu_irq, panfrost_gpu_irq_handler, 520 IRQF_SHARED, KBUILD_MODNAME "-gpu", pfdev); 521 if (err) { 522 dev_err(pfdev->base.dev, "failed to request gpu irq"); 523 return err; 524 } 525 526 panfrost_gpu_power_on(pfdev); 527 528 return 0; 529 } 530 531 void panfrost_gpu_fini(struct panfrost_device *pfdev) 532 { 533 panfrost_gpu_power_off(pfdev); 534 } 535 536 u32 panfrost_gpu_get_latest_flush_id(struct panfrost_device *pfdev) 537 { 538 u32 flush_id; 539 540 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION)) { 541 /* Flush reduction only makes sense when the GPU is kept powered on between jobs */ 542 if (pm_runtime_get_if_in_use(pfdev->base.dev)) { 543 flush_id = gpu_read(pfdev, GPU_LATEST_FLUSH_ID); 544 pm_runtime_put(pfdev->base.dev); 545 return flush_id; 546 } 547 } 548 549 return 0; 550 } 551