1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Tegra host1x Job 4 * 5 * Copyright (c) 2010-2015, NVIDIA Corporation. 6 */ 7 8 #include <linux/dma-mapping.h> 9 #include <linux/err.h> 10 #include <linux/host1x.h> 11 #include <linux/iommu.h> 12 #include <linux/kref.h> 13 #include <linux/module.h> 14 #include <linux/scatterlist.h> 15 #include <linux/slab.h> 16 #include <linux/vmalloc.h> 17 #include <trace/events/host1x.h> 18 19 #include "channel.h" 20 #include "dev.h" 21 #include "job.h" 22 #include "syncpt.h" 23 24 #define HOST1X_WAIT_SYNCPT_OFFSET 0x8 25 26 struct host1x_job *host1x_job_alloc(struct host1x_channel *ch, 27 u32 num_cmdbufs, u32 num_relocs) 28 { 29 struct host1x_job *job = NULL; 30 unsigned int num_unpins = num_relocs; 31 u64 total; 32 void *mem; 33 34 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) 35 num_unpins += num_cmdbufs; 36 37 /* Check that we're not going to overflow */ 38 total = sizeof(struct host1x_job) + 39 (u64)num_relocs * sizeof(struct host1x_reloc) + 40 (u64)num_unpins * sizeof(struct host1x_job_unpin_data) + 41 (u64)num_cmdbufs * sizeof(struct host1x_job_gather) + 42 (u64)num_unpins * sizeof(dma_addr_t) + 43 (u64)num_unpins * sizeof(u32 *); 44 if (total > ULONG_MAX) 45 return NULL; 46 47 mem = job = kzalloc(total, GFP_KERNEL); 48 if (!job) 49 return NULL; 50 51 kref_init(&job->ref); 52 job->channel = ch; 53 54 /* Redistribute memory to the structs */ 55 mem += sizeof(struct host1x_job); 56 job->relocs = num_relocs ? mem : NULL; 57 mem += num_relocs * sizeof(struct host1x_reloc); 58 job->unpins = num_unpins ? mem : NULL; 59 mem += num_unpins * sizeof(struct host1x_job_unpin_data); 60 job->gathers = num_cmdbufs ? mem : NULL; 61 mem += num_cmdbufs * sizeof(struct host1x_job_gather); 62 job->addr_phys = num_unpins ? mem : NULL; 63 64 job->reloc_addr_phys = job->addr_phys; 65 job->gather_addr_phys = &job->addr_phys[num_relocs]; 66 67 return job; 68 } 69 EXPORT_SYMBOL(host1x_job_alloc); 70 71 struct host1x_job *host1x_job_get(struct host1x_job *job) 72 { 73 kref_get(&job->ref); 74 return job; 75 } 76 EXPORT_SYMBOL(host1x_job_get); 77 78 static void job_free(struct kref *ref) 79 { 80 struct host1x_job *job = container_of(ref, struct host1x_job, ref); 81 82 kfree(job); 83 } 84 85 void host1x_job_put(struct host1x_job *job) 86 { 87 kref_put(&job->ref, job_free); 88 } 89 EXPORT_SYMBOL(host1x_job_put); 90 91 void host1x_job_add_gather(struct host1x_job *job, struct host1x_bo *bo, 92 unsigned int words, unsigned int offset) 93 { 94 struct host1x_job_gather *gather = &job->gathers[job->num_gathers]; 95 96 gather->words = words; 97 gather->bo = bo; 98 gather->offset = offset; 99 100 job->num_gathers++; 101 } 102 EXPORT_SYMBOL(host1x_job_add_gather); 103 104 static unsigned int pin_job(struct host1x *host, struct host1x_job *job) 105 { 106 struct host1x_client *client = job->client; 107 struct device *dev = client->dev; 108 struct host1x_job_gather *g; 109 struct iommu_domain *domain; 110 unsigned int i; 111 int err; 112 113 domain = iommu_get_domain_for_dev(dev); 114 job->num_unpins = 0; 115 116 for (i = 0; i < job->num_relocs; i++) { 117 struct host1x_reloc *reloc = &job->relocs[i]; 118 dma_addr_t phys_addr, *phys; 119 struct sg_table *sgt; 120 121 reloc->target.bo = host1x_bo_get(reloc->target.bo); 122 if (!reloc->target.bo) { 123 err = -EINVAL; 124 goto unpin; 125 } 126 127 /* 128 * If the client device is not attached to an IOMMU, the 129 * physical address of the buffer object can be used. 130 * 131 * Similarly, when an IOMMU domain is shared between all 132 * host1x clients, the IOVA is already available, so no 133 * need to map the buffer object again. 134 * 135 * XXX Note that this isn't always safe to do because it 136 * relies on an assumption that no cache maintenance is 137 * needed on the buffer objects. 138 */ 139 if (!domain || client->group) 140 phys = &phys_addr; 141 else 142 phys = NULL; 143 144 sgt = host1x_bo_pin(dev, reloc->target.bo, phys); 145 if (IS_ERR(sgt)) { 146 err = PTR_ERR(sgt); 147 goto unpin; 148 } 149 150 if (sgt) { 151 unsigned long mask = HOST1X_RELOC_READ | 152 HOST1X_RELOC_WRITE; 153 enum dma_data_direction dir; 154 155 switch (reloc->flags & mask) { 156 case HOST1X_RELOC_READ: 157 dir = DMA_TO_DEVICE; 158 break; 159 160 case HOST1X_RELOC_WRITE: 161 dir = DMA_FROM_DEVICE; 162 break; 163 164 case HOST1X_RELOC_READ | HOST1X_RELOC_WRITE: 165 dir = DMA_BIDIRECTIONAL; 166 break; 167 168 default: 169 err = -EINVAL; 170 goto unpin; 171 } 172 173 err = dma_map_sg(dev, sgt->sgl, sgt->nents, dir); 174 if (!err) { 175 err = -ENOMEM; 176 goto unpin; 177 } 178 179 job->unpins[job->num_unpins].dev = dev; 180 job->unpins[job->num_unpins].dir = dir; 181 phys_addr = sg_dma_address(sgt->sgl); 182 } 183 184 job->addr_phys[job->num_unpins] = phys_addr; 185 job->unpins[job->num_unpins].bo = reloc->target.bo; 186 job->unpins[job->num_unpins].sgt = sgt; 187 job->num_unpins++; 188 } 189 190 /* 191 * We will copy gathers BO content later, so there is no need to 192 * hold and pin them. 193 */ 194 if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) 195 return 0; 196 197 for (i = 0; i < job->num_gathers; i++) { 198 size_t gather_size = 0; 199 struct scatterlist *sg; 200 struct sg_table *sgt; 201 dma_addr_t phys_addr; 202 unsigned long shift; 203 struct iova *alloc; 204 dma_addr_t *phys; 205 unsigned int j; 206 207 g = &job->gathers[i]; 208 g->bo = host1x_bo_get(g->bo); 209 if (!g->bo) { 210 err = -EINVAL; 211 goto unpin; 212 } 213 214 /** 215 * If the host1x is not attached to an IOMMU, there is no need 216 * to map the buffer object for the host1x, since the physical 217 * address can simply be used. 218 */ 219 if (!iommu_get_domain_for_dev(host->dev)) 220 phys = &phys_addr; 221 else 222 phys = NULL; 223 224 sgt = host1x_bo_pin(host->dev, g->bo, phys); 225 if (IS_ERR(sgt)) { 226 err = PTR_ERR(sgt); 227 goto put; 228 } 229 230 if (host->domain) { 231 for_each_sg(sgt->sgl, sg, sgt->nents, j) 232 gather_size += sg->length; 233 gather_size = iova_align(&host->iova, gather_size); 234 235 shift = iova_shift(&host->iova); 236 alloc = alloc_iova(&host->iova, gather_size >> shift, 237 host->iova_end >> shift, true); 238 if (!alloc) { 239 err = -ENOMEM; 240 goto put; 241 } 242 243 err = iommu_map_sg(host->domain, 244 iova_dma_addr(&host->iova, alloc), 245 sgt->sgl, sgt->nents, IOMMU_READ); 246 if (err == 0) { 247 __free_iova(&host->iova, alloc); 248 err = -EINVAL; 249 goto put; 250 } 251 252 job->unpins[job->num_unpins].size = gather_size; 253 phys_addr = iova_dma_addr(&host->iova, alloc); 254 } else if (sgt) { 255 err = dma_map_sg(host->dev, sgt->sgl, sgt->nents, 256 DMA_TO_DEVICE); 257 if (!err) { 258 err = -ENOMEM; 259 goto put; 260 } 261 262 job->unpins[job->num_unpins].dir = DMA_TO_DEVICE; 263 job->unpins[job->num_unpins].dev = host->dev; 264 phys_addr = sg_dma_address(sgt->sgl); 265 } 266 267 job->addr_phys[job->num_unpins] = phys_addr; 268 job->gather_addr_phys[i] = phys_addr; 269 270 job->unpins[job->num_unpins].bo = g->bo; 271 job->unpins[job->num_unpins].sgt = sgt; 272 job->num_unpins++; 273 } 274 275 return 0; 276 277 put: 278 host1x_bo_put(g->bo); 279 unpin: 280 host1x_job_unpin(job); 281 return err; 282 } 283 284 static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) 285 { 286 void *cmdbuf_addr = NULL; 287 struct host1x_bo *cmdbuf = g->bo; 288 unsigned int i; 289 290 /* pin & patch the relocs for one gather */ 291 for (i = 0; i < job->num_relocs; i++) { 292 struct host1x_reloc *reloc = &job->relocs[i]; 293 u32 reloc_addr = (job->reloc_addr_phys[i] + 294 reloc->target.offset) >> reloc->shift; 295 u32 *target; 296 297 /* skip all other gathers */ 298 if (cmdbuf != reloc->cmdbuf.bo) 299 continue; 300 301 if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) { 302 target = (u32 *)job->gather_copy_mapped + 303 reloc->cmdbuf.offset / sizeof(u32) + 304 g->offset / sizeof(u32); 305 goto patch_reloc; 306 } 307 308 if (!cmdbuf_addr) { 309 cmdbuf_addr = host1x_bo_mmap(cmdbuf); 310 311 if (unlikely(!cmdbuf_addr)) { 312 pr_err("Could not map cmdbuf for relocation\n"); 313 return -ENOMEM; 314 } 315 } 316 317 target = cmdbuf_addr + reloc->cmdbuf.offset; 318 patch_reloc: 319 *target = reloc_addr; 320 } 321 322 if (cmdbuf_addr) 323 host1x_bo_munmap(cmdbuf, cmdbuf_addr); 324 325 return 0; 326 } 327 328 static bool check_reloc(struct host1x_reloc *reloc, struct host1x_bo *cmdbuf, 329 unsigned int offset) 330 { 331 offset *= sizeof(u32); 332 333 if (reloc->cmdbuf.bo != cmdbuf || reloc->cmdbuf.offset != offset) 334 return false; 335 336 /* relocation shift value validation isn't implemented yet */ 337 if (reloc->shift) 338 return false; 339 340 return true; 341 } 342 343 struct host1x_firewall { 344 struct host1x_job *job; 345 struct device *dev; 346 347 unsigned int num_relocs; 348 struct host1x_reloc *reloc; 349 350 struct host1x_bo *cmdbuf; 351 unsigned int offset; 352 353 u32 words; 354 u32 class; 355 u32 reg; 356 u32 mask; 357 u32 count; 358 }; 359 360 static int check_register(struct host1x_firewall *fw, unsigned long offset) 361 { 362 if (!fw->job->is_addr_reg) 363 return 0; 364 365 if (fw->job->is_addr_reg(fw->dev, fw->class, offset)) { 366 if (!fw->num_relocs) 367 return -EINVAL; 368 369 if (!check_reloc(fw->reloc, fw->cmdbuf, fw->offset)) 370 return -EINVAL; 371 372 fw->num_relocs--; 373 fw->reloc++; 374 } 375 376 return 0; 377 } 378 379 static int check_class(struct host1x_firewall *fw, u32 class) 380 { 381 if (!fw->job->is_valid_class) { 382 if (fw->class != class) 383 return -EINVAL; 384 } else { 385 if (!fw->job->is_valid_class(fw->class)) 386 return -EINVAL; 387 } 388 389 return 0; 390 } 391 392 static int check_mask(struct host1x_firewall *fw) 393 { 394 u32 mask = fw->mask; 395 u32 reg = fw->reg; 396 int ret; 397 398 while (mask) { 399 if (fw->words == 0) 400 return -EINVAL; 401 402 if (mask & 1) { 403 ret = check_register(fw, reg); 404 if (ret < 0) 405 return ret; 406 407 fw->words--; 408 fw->offset++; 409 } 410 mask >>= 1; 411 reg++; 412 } 413 414 return 0; 415 } 416 417 static int check_incr(struct host1x_firewall *fw) 418 { 419 u32 count = fw->count; 420 u32 reg = fw->reg; 421 int ret; 422 423 while (count) { 424 if (fw->words == 0) 425 return -EINVAL; 426 427 ret = check_register(fw, reg); 428 if (ret < 0) 429 return ret; 430 431 reg++; 432 fw->words--; 433 fw->offset++; 434 count--; 435 } 436 437 return 0; 438 } 439 440 static int check_nonincr(struct host1x_firewall *fw) 441 { 442 u32 count = fw->count; 443 int ret; 444 445 while (count) { 446 if (fw->words == 0) 447 return -EINVAL; 448 449 ret = check_register(fw, fw->reg); 450 if (ret < 0) 451 return ret; 452 453 fw->words--; 454 fw->offset++; 455 count--; 456 } 457 458 return 0; 459 } 460 461 static int validate(struct host1x_firewall *fw, struct host1x_job_gather *g) 462 { 463 u32 *cmdbuf_base = (u32 *)fw->job->gather_copy_mapped + 464 (g->offset / sizeof(u32)); 465 u32 job_class = fw->class; 466 int err = 0; 467 468 fw->words = g->words; 469 fw->cmdbuf = g->bo; 470 fw->offset = 0; 471 472 while (fw->words && !err) { 473 u32 word = cmdbuf_base[fw->offset]; 474 u32 opcode = (word & 0xf0000000) >> 28; 475 476 fw->mask = 0; 477 fw->reg = 0; 478 fw->count = 0; 479 fw->words--; 480 fw->offset++; 481 482 switch (opcode) { 483 case 0: 484 fw->class = word >> 6 & 0x3ff; 485 fw->mask = word & 0x3f; 486 fw->reg = word >> 16 & 0xfff; 487 err = check_class(fw, job_class); 488 if (!err) 489 err = check_mask(fw); 490 if (err) 491 goto out; 492 break; 493 case 1: 494 fw->reg = word >> 16 & 0xfff; 495 fw->count = word & 0xffff; 496 err = check_incr(fw); 497 if (err) 498 goto out; 499 break; 500 501 case 2: 502 fw->reg = word >> 16 & 0xfff; 503 fw->count = word & 0xffff; 504 err = check_nonincr(fw); 505 if (err) 506 goto out; 507 break; 508 509 case 3: 510 fw->mask = word & 0xffff; 511 fw->reg = word >> 16 & 0xfff; 512 err = check_mask(fw); 513 if (err) 514 goto out; 515 break; 516 case 4: 517 case 14: 518 break; 519 default: 520 err = -EINVAL; 521 break; 522 } 523 } 524 525 out: 526 return err; 527 } 528 529 static inline int copy_gathers(struct device *host, struct host1x_job *job, 530 struct device *dev) 531 { 532 struct host1x_firewall fw; 533 size_t size = 0; 534 size_t offset = 0; 535 unsigned int i; 536 537 fw.job = job; 538 fw.dev = dev; 539 fw.reloc = job->relocs; 540 fw.num_relocs = job->num_relocs; 541 fw.class = job->class; 542 543 for (i = 0; i < job->num_gathers; i++) { 544 struct host1x_job_gather *g = &job->gathers[i]; 545 546 size += g->words * sizeof(u32); 547 } 548 549 /* 550 * Try a non-blocking allocation from a higher priority pools first, 551 * as awaiting for the allocation here is a major performance hit. 552 */ 553 job->gather_copy_mapped = dma_alloc_wc(host, size, &job->gather_copy, 554 GFP_NOWAIT); 555 556 /* the higher priority allocation failed, try the generic-blocking */ 557 if (!job->gather_copy_mapped) 558 job->gather_copy_mapped = dma_alloc_wc(host, size, 559 &job->gather_copy, 560 GFP_KERNEL); 561 if (!job->gather_copy_mapped) 562 return -ENOMEM; 563 564 job->gather_copy_size = size; 565 566 for (i = 0; i < job->num_gathers; i++) { 567 struct host1x_job_gather *g = &job->gathers[i]; 568 void *gather; 569 570 /* Copy the gather */ 571 gather = host1x_bo_mmap(g->bo); 572 memcpy(job->gather_copy_mapped + offset, gather + g->offset, 573 g->words * sizeof(u32)); 574 host1x_bo_munmap(g->bo, gather); 575 576 /* Store the location in the buffer */ 577 g->base = job->gather_copy; 578 g->offset = offset; 579 580 /* Validate the job */ 581 if (validate(&fw, g)) 582 return -EINVAL; 583 584 offset += g->words * sizeof(u32); 585 } 586 587 /* No relocs should remain at this point */ 588 if (fw.num_relocs) 589 return -EINVAL; 590 591 return 0; 592 } 593 594 int host1x_job_pin(struct host1x_job *job, struct device *dev) 595 { 596 int err; 597 unsigned int i, j; 598 struct host1x *host = dev_get_drvdata(dev->parent); 599 600 /* pin memory */ 601 err = pin_job(host, job); 602 if (err) 603 goto out; 604 605 if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) { 606 err = copy_gathers(host->dev, job, dev); 607 if (err) 608 goto out; 609 } 610 611 /* patch gathers */ 612 for (i = 0; i < job->num_gathers; i++) { 613 struct host1x_job_gather *g = &job->gathers[i]; 614 615 /* process each gather mem only once */ 616 if (g->handled) 617 continue; 618 619 /* copy_gathers() sets gathers base if firewall is enabled */ 620 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) 621 g->base = job->gather_addr_phys[i]; 622 623 for (j = i + 1; j < job->num_gathers; j++) { 624 if (job->gathers[j].bo == g->bo) { 625 job->gathers[j].handled = true; 626 job->gathers[j].base = g->base; 627 } 628 } 629 630 err = do_relocs(job, g); 631 if (err) 632 break; 633 } 634 635 out: 636 if (err) 637 host1x_job_unpin(job); 638 wmb(); 639 640 return err; 641 } 642 EXPORT_SYMBOL(host1x_job_pin); 643 644 void host1x_job_unpin(struct host1x_job *job) 645 { 646 struct host1x *host = dev_get_drvdata(job->channel->dev->parent); 647 unsigned int i; 648 649 for (i = 0; i < job->num_unpins; i++) { 650 struct host1x_job_unpin_data *unpin = &job->unpins[i]; 651 struct device *dev = unpin->dev ?: host->dev; 652 struct sg_table *sgt = unpin->sgt; 653 654 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) && 655 unpin->size && host->domain) { 656 iommu_unmap(host->domain, job->addr_phys[i], 657 unpin->size); 658 free_iova(&host->iova, 659 iova_pfn(&host->iova, job->addr_phys[i])); 660 } 661 662 if (unpin->dev && sgt) 663 dma_unmap_sg(unpin->dev, sgt->sgl, sgt->nents, 664 unpin->dir); 665 666 host1x_bo_unpin(dev, unpin->bo, sgt); 667 host1x_bo_put(unpin->bo); 668 } 669 670 job->num_unpins = 0; 671 672 if (job->gather_copy_size) 673 dma_free_wc(host->dev, job->gather_copy_size, 674 job->gather_copy_mapped, job->gather_copy); 675 } 676 EXPORT_SYMBOL(host1x_job_unpin); 677 678 /* 679 * Debug routine used to dump job entries 680 */ 681 void host1x_job_dump(struct device *dev, struct host1x_job *job) 682 { 683 dev_dbg(dev, " SYNCPT_ID %d\n", job->syncpt_id); 684 dev_dbg(dev, " SYNCPT_VAL %d\n", job->syncpt_end); 685 dev_dbg(dev, " FIRST_GET 0x%x\n", job->first_get); 686 dev_dbg(dev, " TIMEOUT %d\n", job->timeout); 687 dev_dbg(dev, " NUM_SLOTS %d\n", job->num_slots); 688 dev_dbg(dev, " NUM_HANDLES %d\n", job->num_unpins); 689 } 690