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 bool skip_firewall) 29 { 30 struct host1x_job *job = NULL; 31 unsigned int num_unpins = num_relocs; 32 bool enable_firewall; 33 u64 total; 34 void *mem; 35 36 enable_firewall = IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) && !skip_firewall; 37 38 if (!enable_firewall) 39 num_unpins += num_cmdbufs; 40 41 /* Check that we're not going to overflow */ 42 total = sizeof(struct host1x_job) + 43 (u64)num_relocs * sizeof(struct host1x_reloc) + 44 (u64)num_unpins * sizeof(struct host1x_job_unpin_data) + 45 (u64)num_cmdbufs * sizeof(struct host1x_job_cmd) + 46 (u64)num_unpins * sizeof(dma_addr_t) + 47 (u64)num_unpins * sizeof(u32 *); 48 if (total > ULONG_MAX) 49 return NULL; 50 51 mem = job = kzalloc(total, GFP_KERNEL); 52 if (!job) 53 return NULL; 54 55 job->enable_firewall = enable_firewall; 56 57 kref_init(&job->ref); 58 job->channel = ch; 59 60 /* Redistribute memory to the structs */ 61 mem += sizeof(struct host1x_job); 62 job->relocs = num_relocs ? mem : NULL; 63 mem += num_relocs * sizeof(struct host1x_reloc); 64 job->unpins = num_unpins ? mem : NULL; 65 mem += num_unpins * sizeof(struct host1x_job_unpin_data); 66 job->cmds = num_cmdbufs ? mem : NULL; 67 mem += num_cmdbufs * sizeof(struct host1x_job_cmd); 68 job->addr_phys = num_unpins ? mem : NULL; 69 70 job->reloc_addr_phys = job->addr_phys; 71 job->gather_addr_phys = &job->addr_phys[num_relocs]; 72 73 return job; 74 } 75 EXPORT_SYMBOL(host1x_job_alloc); 76 77 struct host1x_job *host1x_job_get(struct host1x_job *job) 78 { 79 kref_get(&job->ref); 80 return job; 81 } 82 EXPORT_SYMBOL(host1x_job_get); 83 84 static void job_free(struct kref *ref) 85 { 86 struct host1x_job *job = container_of(ref, struct host1x_job, ref); 87 88 if (job->release) 89 job->release(job); 90 91 if (job->waiter) 92 host1x_intr_put_ref(job->syncpt->host, job->syncpt->id, 93 job->waiter, false); 94 95 if (job->syncpt) 96 host1x_syncpt_put(job->syncpt); 97 98 kfree(job); 99 } 100 101 void host1x_job_put(struct host1x_job *job) 102 { 103 kref_put(&job->ref, job_free); 104 } 105 EXPORT_SYMBOL(host1x_job_put); 106 107 void host1x_job_add_gather(struct host1x_job *job, struct host1x_bo *bo, 108 unsigned int words, unsigned int offset) 109 { 110 struct host1x_job_gather *gather = &job->cmds[job->num_cmds].gather; 111 112 gather->words = words; 113 gather->bo = bo; 114 gather->offset = offset; 115 116 job->num_cmds++; 117 } 118 EXPORT_SYMBOL(host1x_job_add_gather); 119 120 void host1x_job_add_wait(struct host1x_job *job, u32 id, u32 thresh, 121 bool relative, u32 next_class) 122 { 123 struct host1x_job_cmd *cmd = &job->cmds[job->num_cmds]; 124 125 cmd->is_wait = true; 126 cmd->wait.id = id; 127 cmd->wait.threshold = thresh; 128 cmd->wait.next_class = next_class; 129 cmd->wait.relative = relative; 130 131 job->num_cmds++; 132 } 133 EXPORT_SYMBOL(host1x_job_add_wait); 134 135 static unsigned int pin_job(struct host1x *host, struct host1x_job *job) 136 { 137 struct host1x_client *client = job->client; 138 struct device *dev = client->dev; 139 struct host1x_job_gather *g; 140 struct iommu_domain *domain; 141 struct sg_table *sgt; 142 unsigned int i; 143 int err; 144 145 domain = iommu_get_domain_for_dev(dev); 146 job->num_unpins = 0; 147 148 for (i = 0; i < job->num_relocs; i++) { 149 struct host1x_reloc *reloc = &job->relocs[i]; 150 dma_addr_t phys_addr, *phys; 151 152 reloc->target.bo = host1x_bo_get(reloc->target.bo); 153 if (!reloc->target.bo) { 154 err = -EINVAL; 155 goto unpin; 156 } 157 158 /* 159 * If the client device is not attached to an IOMMU, the 160 * physical address of the buffer object can be used. 161 * 162 * Similarly, when an IOMMU domain is shared between all 163 * host1x clients, the IOVA is already available, so no 164 * need to map the buffer object again. 165 * 166 * XXX Note that this isn't always safe to do because it 167 * relies on an assumption that no cache maintenance is 168 * needed on the buffer objects. 169 */ 170 if (!domain || client->group) 171 phys = &phys_addr; 172 else 173 phys = NULL; 174 175 sgt = host1x_bo_pin(dev, reloc->target.bo, phys); 176 if (IS_ERR(sgt)) { 177 err = PTR_ERR(sgt); 178 goto unpin; 179 } 180 181 if (sgt) { 182 unsigned long mask = HOST1X_RELOC_READ | 183 HOST1X_RELOC_WRITE; 184 enum dma_data_direction dir; 185 186 switch (reloc->flags & mask) { 187 case HOST1X_RELOC_READ: 188 dir = DMA_TO_DEVICE; 189 break; 190 191 case HOST1X_RELOC_WRITE: 192 dir = DMA_FROM_DEVICE; 193 break; 194 195 case HOST1X_RELOC_READ | HOST1X_RELOC_WRITE: 196 dir = DMA_BIDIRECTIONAL; 197 break; 198 199 default: 200 err = -EINVAL; 201 goto unpin; 202 } 203 204 err = dma_map_sgtable(dev, sgt, dir, 0); 205 if (err) 206 goto unpin; 207 208 job->unpins[job->num_unpins].dev = dev; 209 job->unpins[job->num_unpins].dir = dir; 210 phys_addr = sg_dma_address(sgt->sgl); 211 } 212 213 job->addr_phys[job->num_unpins] = phys_addr; 214 job->unpins[job->num_unpins].bo = reloc->target.bo; 215 job->unpins[job->num_unpins].sgt = sgt; 216 job->num_unpins++; 217 } 218 219 /* 220 * We will copy gathers BO content later, so there is no need to 221 * hold and pin them. 222 */ 223 if (job->enable_firewall) 224 return 0; 225 226 for (i = 0; i < job->num_cmds; i++) { 227 size_t gather_size = 0; 228 struct scatterlist *sg; 229 dma_addr_t phys_addr; 230 unsigned long shift; 231 struct iova *alloc; 232 dma_addr_t *phys; 233 unsigned int j; 234 235 if (job->cmds[i].is_wait) 236 continue; 237 238 g = &job->cmds[i].gather; 239 240 g->bo = host1x_bo_get(g->bo); 241 if (!g->bo) { 242 err = -EINVAL; 243 goto unpin; 244 } 245 246 /** 247 * If the host1x is not attached to an IOMMU, there is no need 248 * to map the buffer object for the host1x, since the physical 249 * address can simply be used. 250 */ 251 if (!iommu_get_domain_for_dev(host->dev)) 252 phys = &phys_addr; 253 else 254 phys = NULL; 255 256 sgt = host1x_bo_pin(host->dev, g->bo, phys); 257 if (IS_ERR(sgt)) { 258 err = PTR_ERR(sgt); 259 goto put; 260 } 261 262 if (host->domain) { 263 for_each_sgtable_sg(sgt, sg, j) 264 gather_size += sg->length; 265 gather_size = iova_align(&host->iova, gather_size); 266 267 shift = iova_shift(&host->iova); 268 alloc = alloc_iova(&host->iova, gather_size >> shift, 269 host->iova_end >> shift, true); 270 if (!alloc) { 271 err = -ENOMEM; 272 goto put; 273 } 274 275 err = iommu_map_sgtable(host->domain, 276 iova_dma_addr(&host->iova, alloc), 277 sgt, IOMMU_READ); 278 if (err == 0) { 279 __free_iova(&host->iova, alloc); 280 err = -EINVAL; 281 goto put; 282 } 283 284 job->unpins[job->num_unpins].size = gather_size; 285 phys_addr = iova_dma_addr(&host->iova, alloc); 286 } else if (sgt) { 287 err = dma_map_sgtable(host->dev, sgt, DMA_TO_DEVICE, 0); 288 if (err) 289 goto put; 290 291 job->unpins[job->num_unpins].dir = DMA_TO_DEVICE; 292 job->unpins[job->num_unpins].dev = host->dev; 293 phys_addr = sg_dma_address(sgt->sgl); 294 } 295 296 job->addr_phys[job->num_unpins] = phys_addr; 297 job->gather_addr_phys[i] = phys_addr; 298 299 job->unpins[job->num_unpins].bo = g->bo; 300 job->unpins[job->num_unpins].sgt = sgt; 301 job->num_unpins++; 302 } 303 304 return 0; 305 306 put: 307 host1x_bo_put(g->bo); 308 unpin: 309 host1x_job_unpin(job); 310 return err; 311 } 312 313 static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) 314 { 315 void *cmdbuf_addr = NULL; 316 struct host1x_bo *cmdbuf = g->bo; 317 unsigned int i; 318 319 /* pin & patch the relocs for one gather */ 320 for (i = 0; i < job->num_relocs; i++) { 321 struct host1x_reloc *reloc = &job->relocs[i]; 322 u32 reloc_addr = (job->reloc_addr_phys[i] + 323 reloc->target.offset) >> reloc->shift; 324 u32 *target; 325 326 /* skip all other gathers */ 327 if (cmdbuf != reloc->cmdbuf.bo) 328 continue; 329 330 if (job->enable_firewall) { 331 target = (u32 *)job->gather_copy_mapped + 332 reloc->cmdbuf.offset / sizeof(u32) + 333 g->offset / sizeof(u32); 334 goto patch_reloc; 335 } 336 337 if (!cmdbuf_addr) { 338 cmdbuf_addr = host1x_bo_mmap(cmdbuf); 339 340 if (unlikely(!cmdbuf_addr)) { 341 pr_err("Could not map cmdbuf for relocation\n"); 342 return -ENOMEM; 343 } 344 } 345 346 target = cmdbuf_addr + reloc->cmdbuf.offset; 347 patch_reloc: 348 *target = reloc_addr; 349 } 350 351 if (cmdbuf_addr) 352 host1x_bo_munmap(cmdbuf, cmdbuf_addr); 353 354 return 0; 355 } 356 357 static bool check_reloc(struct host1x_reloc *reloc, struct host1x_bo *cmdbuf, 358 unsigned int offset) 359 { 360 offset *= sizeof(u32); 361 362 if (reloc->cmdbuf.bo != cmdbuf || reloc->cmdbuf.offset != offset) 363 return false; 364 365 /* relocation shift value validation isn't implemented yet */ 366 if (reloc->shift) 367 return false; 368 369 return true; 370 } 371 372 struct host1x_firewall { 373 struct host1x_job *job; 374 struct device *dev; 375 376 unsigned int num_relocs; 377 struct host1x_reloc *reloc; 378 379 struct host1x_bo *cmdbuf; 380 unsigned int offset; 381 382 u32 words; 383 u32 class; 384 u32 reg; 385 u32 mask; 386 u32 count; 387 }; 388 389 static int check_register(struct host1x_firewall *fw, unsigned long offset) 390 { 391 if (!fw->job->is_addr_reg) 392 return 0; 393 394 if (fw->job->is_addr_reg(fw->dev, fw->class, offset)) { 395 if (!fw->num_relocs) 396 return -EINVAL; 397 398 if (!check_reloc(fw->reloc, fw->cmdbuf, fw->offset)) 399 return -EINVAL; 400 401 fw->num_relocs--; 402 fw->reloc++; 403 } 404 405 return 0; 406 } 407 408 static int check_class(struct host1x_firewall *fw, u32 class) 409 { 410 if (!fw->job->is_valid_class) { 411 if (fw->class != class) 412 return -EINVAL; 413 } else { 414 if (!fw->job->is_valid_class(fw->class)) 415 return -EINVAL; 416 } 417 418 return 0; 419 } 420 421 static int check_mask(struct host1x_firewall *fw) 422 { 423 u32 mask = fw->mask; 424 u32 reg = fw->reg; 425 int ret; 426 427 while (mask) { 428 if (fw->words == 0) 429 return -EINVAL; 430 431 if (mask & 1) { 432 ret = check_register(fw, reg); 433 if (ret < 0) 434 return ret; 435 436 fw->words--; 437 fw->offset++; 438 } 439 mask >>= 1; 440 reg++; 441 } 442 443 return 0; 444 } 445 446 static int check_incr(struct host1x_firewall *fw) 447 { 448 u32 count = fw->count; 449 u32 reg = fw->reg; 450 int ret; 451 452 while (count) { 453 if (fw->words == 0) 454 return -EINVAL; 455 456 ret = check_register(fw, reg); 457 if (ret < 0) 458 return ret; 459 460 reg++; 461 fw->words--; 462 fw->offset++; 463 count--; 464 } 465 466 return 0; 467 } 468 469 static int check_nonincr(struct host1x_firewall *fw) 470 { 471 u32 count = fw->count; 472 int ret; 473 474 while (count) { 475 if (fw->words == 0) 476 return -EINVAL; 477 478 ret = check_register(fw, fw->reg); 479 if (ret < 0) 480 return ret; 481 482 fw->words--; 483 fw->offset++; 484 count--; 485 } 486 487 return 0; 488 } 489 490 static int validate(struct host1x_firewall *fw, struct host1x_job_gather *g) 491 { 492 u32 *cmdbuf_base = (u32 *)fw->job->gather_copy_mapped + 493 (g->offset / sizeof(u32)); 494 u32 job_class = fw->class; 495 int err = 0; 496 497 fw->words = g->words; 498 fw->cmdbuf = g->bo; 499 fw->offset = 0; 500 501 while (fw->words && !err) { 502 u32 word = cmdbuf_base[fw->offset]; 503 u32 opcode = (word & 0xf0000000) >> 28; 504 505 fw->mask = 0; 506 fw->reg = 0; 507 fw->count = 0; 508 fw->words--; 509 fw->offset++; 510 511 switch (opcode) { 512 case 0: 513 fw->class = word >> 6 & 0x3ff; 514 fw->mask = word & 0x3f; 515 fw->reg = word >> 16 & 0xfff; 516 err = check_class(fw, job_class); 517 if (!err) 518 err = check_mask(fw); 519 if (err) 520 goto out; 521 break; 522 case 1: 523 fw->reg = word >> 16 & 0xfff; 524 fw->count = word & 0xffff; 525 err = check_incr(fw); 526 if (err) 527 goto out; 528 break; 529 530 case 2: 531 fw->reg = word >> 16 & 0xfff; 532 fw->count = word & 0xffff; 533 err = check_nonincr(fw); 534 if (err) 535 goto out; 536 break; 537 538 case 3: 539 fw->mask = word & 0xffff; 540 fw->reg = word >> 16 & 0xfff; 541 err = check_mask(fw); 542 if (err) 543 goto out; 544 break; 545 case 4: 546 case 14: 547 break; 548 default: 549 err = -EINVAL; 550 break; 551 } 552 } 553 554 out: 555 return err; 556 } 557 558 static inline int copy_gathers(struct device *host, struct host1x_job *job, 559 struct device *dev) 560 { 561 struct host1x_firewall fw; 562 size_t size = 0; 563 size_t offset = 0; 564 unsigned int i; 565 566 fw.job = job; 567 fw.dev = dev; 568 fw.reloc = job->relocs; 569 fw.num_relocs = job->num_relocs; 570 fw.class = job->class; 571 572 for (i = 0; i < job->num_cmds; i++) { 573 struct host1x_job_gather *g; 574 575 if (job->cmds[i].is_wait) 576 continue; 577 578 g = &job->cmds[i].gather; 579 580 size += g->words * sizeof(u32); 581 } 582 583 /* 584 * Try a non-blocking allocation from a higher priority pools first, 585 * as awaiting for the allocation here is a major performance hit. 586 */ 587 job->gather_copy_mapped = dma_alloc_wc(host, size, &job->gather_copy, 588 GFP_NOWAIT); 589 590 /* the higher priority allocation failed, try the generic-blocking */ 591 if (!job->gather_copy_mapped) 592 job->gather_copy_mapped = dma_alloc_wc(host, size, 593 &job->gather_copy, 594 GFP_KERNEL); 595 if (!job->gather_copy_mapped) 596 return -ENOMEM; 597 598 job->gather_copy_size = size; 599 600 for (i = 0; i < job->num_cmds; i++) { 601 struct host1x_job_gather *g; 602 void *gather; 603 604 if (job->cmds[i].is_wait) 605 continue; 606 g = &job->cmds[i].gather; 607 608 /* Copy the gather */ 609 gather = host1x_bo_mmap(g->bo); 610 memcpy(job->gather_copy_mapped + offset, gather + g->offset, 611 g->words * sizeof(u32)); 612 host1x_bo_munmap(g->bo, gather); 613 614 /* Store the location in the buffer */ 615 g->base = job->gather_copy; 616 g->offset = offset; 617 618 /* Validate the job */ 619 if (validate(&fw, g)) 620 return -EINVAL; 621 622 offset += g->words * sizeof(u32); 623 } 624 625 /* No relocs should remain at this point */ 626 if (fw.num_relocs) 627 return -EINVAL; 628 629 return 0; 630 } 631 632 int host1x_job_pin(struct host1x_job *job, struct device *dev) 633 { 634 int err; 635 unsigned int i, j; 636 struct host1x *host = dev_get_drvdata(dev->parent); 637 638 /* pin memory */ 639 err = pin_job(host, job); 640 if (err) 641 goto out; 642 643 if (job->enable_firewall) { 644 err = copy_gathers(host->dev, job, dev); 645 if (err) 646 goto out; 647 } 648 649 /* patch gathers */ 650 for (i = 0; i < job->num_cmds; i++) { 651 struct host1x_job_gather *g; 652 653 if (job->cmds[i].is_wait) 654 continue; 655 g = &job->cmds[i].gather; 656 657 /* process each gather mem only once */ 658 if (g->handled) 659 continue; 660 661 /* copy_gathers() sets gathers base if firewall is enabled */ 662 if (!job->enable_firewall) 663 g->base = job->gather_addr_phys[i]; 664 665 for (j = i + 1; j < job->num_cmds; j++) { 666 if (!job->cmds[j].is_wait && 667 job->cmds[j].gather.bo == g->bo) { 668 job->cmds[j].gather.handled = true; 669 job->cmds[j].gather.base = g->base; 670 } 671 } 672 673 err = do_relocs(job, g); 674 if (err) 675 break; 676 } 677 678 out: 679 if (err) 680 host1x_job_unpin(job); 681 wmb(); 682 683 return err; 684 } 685 EXPORT_SYMBOL(host1x_job_pin); 686 687 void host1x_job_unpin(struct host1x_job *job) 688 { 689 struct host1x *host = dev_get_drvdata(job->channel->dev->parent); 690 unsigned int i; 691 692 for (i = 0; i < job->num_unpins; i++) { 693 struct host1x_job_unpin_data *unpin = &job->unpins[i]; 694 struct device *dev = unpin->dev ?: host->dev; 695 struct sg_table *sgt = unpin->sgt; 696 697 if (!job->enable_firewall && unpin->size && host->domain) { 698 iommu_unmap(host->domain, job->addr_phys[i], 699 unpin->size); 700 free_iova(&host->iova, 701 iova_pfn(&host->iova, job->addr_phys[i])); 702 } 703 704 if (unpin->dev && sgt) 705 dma_unmap_sgtable(unpin->dev, sgt, unpin->dir, 0); 706 707 host1x_bo_unpin(dev, unpin->bo, sgt); 708 host1x_bo_put(unpin->bo); 709 } 710 711 job->num_unpins = 0; 712 713 if (job->gather_copy_size) 714 dma_free_wc(host->dev, job->gather_copy_size, 715 job->gather_copy_mapped, job->gather_copy); 716 } 717 EXPORT_SYMBOL(host1x_job_unpin); 718 719 /* 720 * Debug routine used to dump job entries 721 */ 722 void host1x_job_dump(struct device *dev, struct host1x_job *job) 723 { 724 dev_dbg(dev, " SYNCPT_ID %d\n", job->syncpt->id); 725 dev_dbg(dev, " SYNCPT_VAL %d\n", job->syncpt_end); 726 dev_dbg(dev, " FIRST_GET 0x%x\n", job->first_get); 727 dev_dbg(dev, " TIMEOUT %d\n", job->timeout); 728 dev_dbg(dev, " NUM_SLOTS %d\n", job->num_slots); 729 dev_dbg(dev, " NUM_HANDLES %d\n", job->num_unpins); 730 } 731