1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2013 Red Hat 4 * Author: Rob Clark <robdclark@gmail.com> 5 */ 6 7 #include <linux/dma-map-ops.h> 8 #include <linux/spinlock.h> 9 #include <linux/shmem_fs.h> 10 #include <linux/dma-buf.h> 11 #include <linux/pfn_t.h> 12 13 #include <drm/drm_prime.h> 14 15 #include "msm_drv.h" 16 #include "msm_fence.h" 17 #include "msm_gem.h" 18 #include "msm_gpu.h" 19 #include "msm_mmu.h" 20 21 static void update_inactive(struct msm_gem_object *msm_obj); 22 23 static dma_addr_t physaddr(struct drm_gem_object *obj) 24 { 25 struct msm_gem_object *msm_obj = to_msm_bo(obj); 26 struct msm_drm_private *priv = obj->dev->dev_private; 27 return (((dma_addr_t)msm_obj->vram_node->start) << PAGE_SHIFT) + 28 priv->vram.paddr; 29 } 30 31 static bool use_pages(struct drm_gem_object *obj) 32 { 33 struct msm_gem_object *msm_obj = to_msm_bo(obj); 34 return !msm_obj->vram_node; 35 } 36 37 /* 38 * Cache sync.. this is a bit over-complicated, to fit dma-mapping 39 * API. Really GPU cache is out of scope here (handled on cmdstream) 40 * and all we need to do is invalidate newly allocated pages before 41 * mapping to CPU as uncached/writecombine. 42 * 43 * On top of this, we have the added headache, that depending on 44 * display generation, the display's iommu may be wired up to either 45 * the toplevel drm device (mdss), or to the mdp sub-node, meaning 46 * that here we either have dma-direct or iommu ops. 47 * 48 * Let this be a cautionary tail of abstraction gone wrong. 49 */ 50 51 static void sync_for_device(struct msm_gem_object *msm_obj) 52 { 53 struct device *dev = msm_obj->base.dev->dev; 54 55 dma_map_sgtable(dev, msm_obj->sgt, DMA_BIDIRECTIONAL, 0); 56 } 57 58 static void sync_for_cpu(struct msm_gem_object *msm_obj) 59 { 60 struct device *dev = msm_obj->base.dev->dev; 61 62 dma_unmap_sgtable(dev, msm_obj->sgt, DMA_BIDIRECTIONAL, 0); 63 } 64 65 /* allocate pages from VRAM carveout, used when no IOMMU: */ 66 static struct page **get_pages_vram(struct drm_gem_object *obj, int npages) 67 { 68 struct msm_gem_object *msm_obj = to_msm_bo(obj); 69 struct msm_drm_private *priv = obj->dev->dev_private; 70 dma_addr_t paddr; 71 struct page **p; 72 int ret, i; 73 74 p = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL); 75 if (!p) 76 return ERR_PTR(-ENOMEM); 77 78 spin_lock(&priv->vram.lock); 79 ret = drm_mm_insert_node(&priv->vram.mm, msm_obj->vram_node, npages); 80 spin_unlock(&priv->vram.lock); 81 if (ret) { 82 kvfree(p); 83 return ERR_PTR(ret); 84 } 85 86 paddr = physaddr(obj); 87 for (i = 0; i < npages; i++) { 88 p[i] = phys_to_page(paddr); 89 paddr += PAGE_SIZE; 90 } 91 92 return p; 93 } 94 95 static struct page **get_pages(struct drm_gem_object *obj) 96 { 97 struct msm_gem_object *msm_obj = to_msm_bo(obj); 98 99 GEM_WARN_ON(!msm_gem_is_locked(obj)); 100 101 if (!msm_obj->pages) { 102 struct drm_device *dev = obj->dev; 103 struct page **p; 104 int npages = obj->size >> PAGE_SHIFT; 105 106 if (use_pages(obj)) 107 p = drm_gem_get_pages(obj); 108 else 109 p = get_pages_vram(obj, npages); 110 111 if (IS_ERR(p)) { 112 DRM_DEV_ERROR(dev->dev, "could not get pages: %ld\n", 113 PTR_ERR(p)); 114 return p; 115 } 116 117 msm_obj->pages = p; 118 119 msm_obj->sgt = drm_prime_pages_to_sg(obj->dev, p, npages); 120 if (IS_ERR(msm_obj->sgt)) { 121 void *ptr = ERR_CAST(msm_obj->sgt); 122 123 DRM_DEV_ERROR(dev->dev, "failed to allocate sgt\n"); 124 msm_obj->sgt = NULL; 125 return ptr; 126 } 127 128 /* For non-cached buffers, ensure the new pages are clean 129 * because display controller, GPU, etc. are not coherent: 130 */ 131 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED)) 132 sync_for_device(msm_obj); 133 } 134 135 return msm_obj->pages; 136 } 137 138 static void put_pages_vram(struct drm_gem_object *obj) 139 { 140 struct msm_gem_object *msm_obj = to_msm_bo(obj); 141 struct msm_drm_private *priv = obj->dev->dev_private; 142 143 spin_lock(&priv->vram.lock); 144 drm_mm_remove_node(msm_obj->vram_node); 145 spin_unlock(&priv->vram.lock); 146 147 kvfree(msm_obj->pages); 148 } 149 150 static void put_pages(struct drm_gem_object *obj) 151 { 152 struct msm_gem_object *msm_obj = to_msm_bo(obj); 153 154 if (msm_obj->pages) { 155 if (msm_obj->sgt) { 156 /* For non-cached buffers, ensure the new 157 * pages are clean because display controller, 158 * GPU, etc. are not coherent: 159 */ 160 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED)) 161 sync_for_cpu(msm_obj); 162 163 sg_free_table(msm_obj->sgt); 164 kfree(msm_obj->sgt); 165 msm_obj->sgt = NULL; 166 } 167 168 if (use_pages(obj)) 169 drm_gem_put_pages(obj, msm_obj->pages, true, false); 170 else 171 put_pages_vram(obj); 172 173 msm_obj->pages = NULL; 174 } 175 } 176 177 struct page **msm_gem_get_pages(struct drm_gem_object *obj) 178 { 179 struct msm_gem_object *msm_obj = to_msm_bo(obj); 180 struct page **p; 181 182 msm_gem_lock(obj); 183 184 if (GEM_WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED)) { 185 msm_gem_unlock(obj); 186 return ERR_PTR(-EBUSY); 187 } 188 189 p = get_pages(obj); 190 msm_gem_unlock(obj); 191 return p; 192 } 193 194 void msm_gem_put_pages(struct drm_gem_object *obj) 195 { 196 /* when we start tracking the pin count, then do something here */ 197 } 198 199 int msm_gem_mmap_obj(struct drm_gem_object *obj, 200 struct vm_area_struct *vma) 201 { 202 struct msm_gem_object *msm_obj = to_msm_bo(obj); 203 204 vma->vm_flags &= ~VM_PFNMAP; 205 vma->vm_flags |= VM_MIXEDMAP; 206 207 if (msm_obj->flags & MSM_BO_WC) { 208 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags)); 209 } else if (msm_obj->flags & MSM_BO_UNCACHED) { 210 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags)); 211 } else { 212 /* 213 * Shunt off cached objs to shmem file so they have their own 214 * address_space (so unmap_mapping_range does what we want, 215 * in particular in the case of mmap'd dmabufs) 216 */ 217 vma->vm_pgoff = 0; 218 vma_set_file(vma, obj->filp); 219 220 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 221 } 222 223 return 0; 224 } 225 226 int msm_gem_mmap(struct file *filp, struct vm_area_struct *vma) 227 { 228 int ret; 229 230 ret = drm_gem_mmap(filp, vma); 231 if (ret) { 232 DBG("mmap failed: %d", ret); 233 return ret; 234 } 235 236 return msm_gem_mmap_obj(vma->vm_private_data, vma); 237 } 238 239 static vm_fault_t msm_gem_fault(struct vm_fault *vmf) 240 { 241 struct vm_area_struct *vma = vmf->vma; 242 struct drm_gem_object *obj = vma->vm_private_data; 243 struct msm_gem_object *msm_obj = to_msm_bo(obj); 244 struct page **pages; 245 unsigned long pfn; 246 pgoff_t pgoff; 247 int err; 248 vm_fault_t ret; 249 250 /* 251 * vm_ops.open/drm_gem_mmap_obj and close get and put 252 * a reference on obj. So, we dont need to hold one here. 253 */ 254 err = msm_gem_lock_interruptible(obj); 255 if (err) { 256 ret = VM_FAULT_NOPAGE; 257 goto out; 258 } 259 260 if (GEM_WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED)) { 261 msm_gem_unlock(obj); 262 return VM_FAULT_SIGBUS; 263 } 264 265 /* make sure we have pages attached now */ 266 pages = get_pages(obj); 267 if (IS_ERR(pages)) { 268 ret = vmf_error(PTR_ERR(pages)); 269 goto out_unlock; 270 } 271 272 /* We don't use vmf->pgoff since that has the fake offset: */ 273 pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT; 274 275 pfn = page_to_pfn(pages[pgoff]); 276 277 VERB("Inserting %p pfn %lx, pa %lx", (void *)vmf->address, 278 pfn, pfn << PAGE_SHIFT); 279 280 ret = vmf_insert_mixed(vma, vmf->address, __pfn_to_pfn_t(pfn, PFN_DEV)); 281 out_unlock: 282 msm_gem_unlock(obj); 283 out: 284 return ret; 285 } 286 287 /** get mmap offset */ 288 static uint64_t mmap_offset(struct drm_gem_object *obj) 289 { 290 struct drm_device *dev = obj->dev; 291 int ret; 292 293 GEM_WARN_ON(!msm_gem_is_locked(obj)); 294 295 /* Make it mmapable */ 296 ret = drm_gem_create_mmap_offset(obj); 297 298 if (ret) { 299 DRM_DEV_ERROR(dev->dev, "could not allocate mmap offset\n"); 300 return 0; 301 } 302 303 return drm_vma_node_offset_addr(&obj->vma_node); 304 } 305 306 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj) 307 { 308 uint64_t offset; 309 310 msm_gem_lock(obj); 311 offset = mmap_offset(obj); 312 msm_gem_unlock(obj); 313 return offset; 314 } 315 316 static struct msm_gem_vma *add_vma(struct drm_gem_object *obj, 317 struct msm_gem_address_space *aspace) 318 { 319 struct msm_gem_object *msm_obj = to_msm_bo(obj); 320 struct msm_gem_vma *vma; 321 322 GEM_WARN_ON(!msm_gem_is_locked(obj)); 323 324 vma = kzalloc(sizeof(*vma), GFP_KERNEL); 325 if (!vma) 326 return ERR_PTR(-ENOMEM); 327 328 vma->aspace = aspace; 329 330 list_add_tail(&vma->list, &msm_obj->vmas); 331 332 return vma; 333 } 334 335 static struct msm_gem_vma *lookup_vma(struct drm_gem_object *obj, 336 struct msm_gem_address_space *aspace) 337 { 338 struct msm_gem_object *msm_obj = to_msm_bo(obj); 339 struct msm_gem_vma *vma; 340 341 GEM_WARN_ON(!msm_gem_is_locked(obj)); 342 343 list_for_each_entry(vma, &msm_obj->vmas, list) { 344 if (vma->aspace == aspace) 345 return vma; 346 } 347 348 return NULL; 349 } 350 351 static void del_vma(struct msm_gem_vma *vma) 352 { 353 if (!vma) 354 return; 355 356 list_del(&vma->list); 357 kfree(vma); 358 } 359 360 /** 361 * If close is true, this also closes the VMA (releasing the allocated 362 * iova range) in addition to removing the iommu mapping. In the eviction 363 * case (!close), we keep the iova allocated, but only remove the iommu 364 * mapping. 365 */ 366 static void 367 put_iova_spaces(struct drm_gem_object *obj, bool close) 368 { 369 struct msm_gem_object *msm_obj = to_msm_bo(obj); 370 struct msm_gem_vma *vma; 371 372 GEM_WARN_ON(!msm_gem_is_locked(obj)); 373 374 list_for_each_entry(vma, &msm_obj->vmas, list) { 375 if (vma->aspace) { 376 msm_gem_purge_vma(vma->aspace, vma); 377 if (close) 378 msm_gem_close_vma(vma->aspace, vma); 379 } 380 } 381 } 382 383 /* Called with msm_obj locked */ 384 static void 385 put_iova_vmas(struct drm_gem_object *obj) 386 { 387 struct msm_gem_object *msm_obj = to_msm_bo(obj); 388 struct msm_gem_vma *vma, *tmp; 389 390 GEM_WARN_ON(!msm_gem_is_locked(obj)); 391 392 list_for_each_entry_safe(vma, tmp, &msm_obj->vmas, list) { 393 del_vma(vma); 394 } 395 } 396 397 static int get_iova_locked(struct drm_gem_object *obj, 398 struct msm_gem_address_space *aspace, uint64_t *iova, 399 u64 range_start, u64 range_end) 400 { 401 struct msm_gem_vma *vma; 402 int ret = 0; 403 404 GEM_WARN_ON(!msm_gem_is_locked(obj)); 405 406 vma = lookup_vma(obj, aspace); 407 408 if (!vma) { 409 vma = add_vma(obj, aspace); 410 if (IS_ERR(vma)) 411 return PTR_ERR(vma); 412 413 ret = msm_gem_init_vma(aspace, vma, obj->size >> PAGE_SHIFT, 414 range_start, range_end); 415 if (ret) { 416 del_vma(vma); 417 return ret; 418 } 419 } 420 421 *iova = vma->iova; 422 return 0; 423 } 424 425 static int msm_gem_pin_iova(struct drm_gem_object *obj, 426 struct msm_gem_address_space *aspace) 427 { 428 struct msm_gem_object *msm_obj = to_msm_bo(obj); 429 struct msm_gem_vma *vma; 430 struct page **pages; 431 int prot = IOMMU_READ; 432 433 if (!(msm_obj->flags & MSM_BO_GPU_READONLY)) 434 prot |= IOMMU_WRITE; 435 436 if (msm_obj->flags & MSM_BO_MAP_PRIV) 437 prot |= IOMMU_PRIV; 438 439 GEM_WARN_ON(!msm_gem_is_locked(obj)); 440 441 if (GEM_WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED)) 442 return -EBUSY; 443 444 vma = lookup_vma(obj, aspace); 445 if (GEM_WARN_ON(!vma)) 446 return -EINVAL; 447 448 pages = get_pages(obj); 449 if (IS_ERR(pages)) 450 return PTR_ERR(pages); 451 452 return msm_gem_map_vma(aspace, vma, prot, 453 msm_obj->sgt, obj->size >> PAGE_SHIFT); 454 } 455 456 static int get_and_pin_iova_range_locked(struct drm_gem_object *obj, 457 struct msm_gem_address_space *aspace, uint64_t *iova, 458 u64 range_start, u64 range_end) 459 { 460 u64 local; 461 int ret; 462 463 GEM_WARN_ON(!msm_gem_is_locked(obj)); 464 465 ret = get_iova_locked(obj, aspace, &local, 466 range_start, range_end); 467 468 if (!ret) 469 ret = msm_gem_pin_iova(obj, aspace); 470 471 if (!ret) 472 *iova = local; 473 474 return ret; 475 } 476 477 /* 478 * get iova and pin it. Should have a matching put 479 * limits iova to specified range (in pages) 480 */ 481 int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj, 482 struct msm_gem_address_space *aspace, uint64_t *iova, 483 u64 range_start, u64 range_end) 484 { 485 int ret; 486 487 msm_gem_lock(obj); 488 ret = get_and_pin_iova_range_locked(obj, aspace, iova, range_start, range_end); 489 msm_gem_unlock(obj); 490 491 return ret; 492 } 493 494 int msm_gem_get_and_pin_iova_locked(struct drm_gem_object *obj, 495 struct msm_gem_address_space *aspace, uint64_t *iova) 496 { 497 return get_and_pin_iova_range_locked(obj, aspace, iova, 0, U64_MAX); 498 } 499 500 /* get iova and pin it. Should have a matching put */ 501 int msm_gem_get_and_pin_iova(struct drm_gem_object *obj, 502 struct msm_gem_address_space *aspace, uint64_t *iova) 503 { 504 return msm_gem_get_and_pin_iova_range(obj, aspace, iova, 0, U64_MAX); 505 } 506 507 /* 508 * Get an iova but don't pin it. Doesn't need a put because iovas are currently 509 * valid for the life of the object 510 */ 511 int msm_gem_get_iova(struct drm_gem_object *obj, 512 struct msm_gem_address_space *aspace, uint64_t *iova) 513 { 514 int ret; 515 516 msm_gem_lock(obj); 517 ret = get_iova_locked(obj, aspace, iova, 0, U64_MAX); 518 msm_gem_unlock(obj); 519 520 return ret; 521 } 522 523 /* get iova without taking a reference, used in places where you have 524 * already done a 'msm_gem_get_and_pin_iova' or 'msm_gem_get_iova' 525 */ 526 uint64_t msm_gem_iova(struct drm_gem_object *obj, 527 struct msm_gem_address_space *aspace) 528 { 529 struct msm_gem_vma *vma; 530 531 msm_gem_lock(obj); 532 vma = lookup_vma(obj, aspace); 533 msm_gem_unlock(obj); 534 GEM_WARN_ON(!vma); 535 536 return vma ? vma->iova : 0; 537 } 538 539 /* 540 * Locked variant of msm_gem_unpin_iova() 541 */ 542 void msm_gem_unpin_iova_locked(struct drm_gem_object *obj, 543 struct msm_gem_address_space *aspace) 544 { 545 struct msm_gem_vma *vma; 546 547 GEM_WARN_ON(!msm_gem_is_locked(obj)); 548 549 vma = lookup_vma(obj, aspace); 550 551 if (!GEM_WARN_ON(!vma)) 552 msm_gem_unmap_vma(aspace, vma); 553 } 554 555 /* 556 * Unpin a iova by updating the reference counts. The memory isn't actually 557 * purged until something else (shrinker, mm_notifier, destroy, etc) decides 558 * to get rid of it 559 */ 560 void msm_gem_unpin_iova(struct drm_gem_object *obj, 561 struct msm_gem_address_space *aspace) 562 { 563 msm_gem_lock(obj); 564 msm_gem_unpin_iova_locked(obj, aspace); 565 msm_gem_unlock(obj); 566 } 567 568 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev, 569 struct drm_mode_create_dumb *args) 570 { 571 args->pitch = align_pitch(args->width, args->bpp); 572 args->size = PAGE_ALIGN(args->pitch * args->height); 573 return msm_gem_new_handle(dev, file, args->size, 574 MSM_BO_SCANOUT | MSM_BO_WC, &args->handle, "dumb"); 575 } 576 577 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev, 578 uint32_t handle, uint64_t *offset) 579 { 580 struct drm_gem_object *obj; 581 int ret = 0; 582 583 /* GEM does all our handle to object mapping */ 584 obj = drm_gem_object_lookup(file, handle); 585 if (obj == NULL) { 586 ret = -ENOENT; 587 goto fail; 588 } 589 590 *offset = msm_gem_mmap_offset(obj); 591 592 drm_gem_object_put(obj); 593 594 fail: 595 return ret; 596 } 597 598 static void *get_vaddr(struct drm_gem_object *obj, unsigned madv) 599 { 600 struct msm_gem_object *msm_obj = to_msm_bo(obj); 601 int ret = 0; 602 603 GEM_WARN_ON(!msm_gem_is_locked(obj)); 604 605 if (obj->import_attach) 606 return ERR_PTR(-ENODEV); 607 608 if (GEM_WARN_ON(msm_obj->madv > madv)) { 609 DRM_DEV_ERROR(obj->dev->dev, "Invalid madv state: %u vs %u\n", 610 msm_obj->madv, madv); 611 return ERR_PTR(-EBUSY); 612 } 613 614 /* increment vmap_count *before* vmap() call, so shrinker can 615 * check vmap_count (is_vunmapable()) outside of msm_obj lock. 616 * This guarantees that we won't try to msm_gem_vunmap() this 617 * same object from within the vmap() call (while we already 618 * hold msm_obj lock) 619 */ 620 msm_obj->vmap_count++; 621 622 if (!msm_obj->vaddr) { 623 struct page **pages = get_pages(obj); 624 if (IS_ERR(pages)) { 625 ret = PTR_ERR(pages); 626 goto fail; 627 } 628 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT, 629 VM_MAP, pgprot_writecombine(PAGE_KERNEL)); 630 if (msm_obj->vaddr == NULL) { 631 ret = -ENOMEM; 632 goto fail; 633 } 634 } 635 636 return msm_obj->vaddr; 637 638 fail: 639 msm_obj->vmap_count--; 640 return ERR_PTR(ret); 641 } 642 643 void *msm_gem_get_vaddr_locked(struct drm_gem_object *obj) 644 { 645 return get_vaddr(obj, MSM_MADV_WILLNEED); 646 } 647 648 void *msm_gem_get_vaddr(struct drm_gem_object *obj) 649 { 650 void *ret; 651 652 msm_gem_lock(obj); 653 ret = msm_gem_get_vaddr_locked(obj); 654 msm_gem_unlock(obj); 655 656 return ret; 657 } 658 659 /* 660 * Don't use this! It is for the very special case of dumping 661 * submits from GPU hangs or faults, were the bo may already 662 * be MSM_MADV_DONTNEED, but we know the buffer is still on the 663 * active list. 664 */ 665 void *msm_gem_get_vaddr_active(struct drm_gem_object *obj) 666 { 667 return get_vaddr(obj, __MSM_MADV_PURGED); 668 } 669 670 void msm_gem_put_vaddr_locked(struct drm_gem_object *obj) 671 { 672 struct msm_gem_object *msm_obj = to_msm_bo(obj); 673 674 GEM_WARN_ON(!msm_gem_is_locked(obj)); 675 GEM_WARN_ON(msm_obj->vmap_count < 1); 676 677 msm_obj->vmap_count--; 678 } 679 680 void msm_gem_put_vaddr(struct drm_gem_object *obj) 681 { 682 msm_gem_lock(obj); 683 msm_gem_put_vaddr_locked(obj); 684 msm_gem_unlock(obj); 685 } 686 687 /* Update madvise status, returns true if not purged, else 688 * false or -errno. 689 */ 690 int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv) 691 { 692 struct msm_gem_object *msm_obj = to_msm_bo(obj); 693 694 msm_gem_lock(obj); 695 696 if (msm_obj->madv != __MSM_MADV_PURGED) 697 msm_obj->madv = madv; 698 699 madv = msm_obj->madv; 700 701 /* If the obj is inactive, we might need to move it 702 * between inactive lists 703 */ 704 if (msm_obj->active_count == 0) 705 update_inactive(msm_obj); 706 707 msm_gem_unlock(obj); 708 709 return (madv != __MSM_MADV_PURGED); 710 } 711 712 void msm_gem_purge(struct drm_gem_object *obj) 713 { 714 struct drm_device *dev = obj->dev; 715 struct msm_gem_object *msm_obj = to_msm_bo(obj); 716 717 GEM_WARN_ON(!is_purgeable(msm_obj)); 718 GEM_WARN_ON(obj->import_attach); 719 720 /* Get rid of any iommu mapping(s): */ 721 put_iova_spaces(obj, true); 722 723 msm_gem_vunmap(obj); 724 725 put_pages(obj); 726 727 put_iova_vmas(obj); 728 729 msm_obj->madv = __MSM_MADV_PURGED; 730 update_inactive(msm_obj); 731 732 drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping); 733 drm_gem_free_mmap_offset(obj); 734 735 /* Our goal here is to return as much of the memory as 736 * is possible back to the system as we are called from OOM. 737 * To do this we must instruct the shmfs to drop all of its 738 * backing pages, *now*. 739 */ 740 shmem_truncate_range(file_inode(obj->filp), 0, (loff_t)-1); 741 742 invalidate_mapping_pages(file_inode(obj->filp)->i_mapping, 743 0, (loff_t)-1); 744 } 745 746 void msm_gem_vunmap(struct drm_gem_object *obj) 747 { 748 struct msm_gem_object *msm_obj = to_msm_bo(obj); 749 750 GEM_WARN_ON(!msm_gem_is_locked(obj)); 751 752 if (!msm_obj->vaddr || GEM_WARN_ON(!is_vunmapable(msm_obj))) 753 return; 754 755 vunmap(msm_obj->vaddr); 756 msm_obj->vaddr = NULL; 757 } 758 759 /* must be called before _move_to_active().. */ 760 int msm_gem_sync_object(struct drm_gem_object *obj, 761 struct msm_fence_context *fctx, bool exclusive) 762 { 763 struct dma_resv_list *fobj; 764 struct dma_fence *fence; 765 int i, ret; 766 767 fobj = dma_resv_get_list(obj->resv); 768 if (!fobj || (fobj->shared_count == 0)) { 769 fence = dma_resv_get_excl(obj->resv); 770 /* don't need to wait on our own fences, since ring is fifo */ 771 if (fence && (fence->context != fctx->context)) { 772 ret = dma_fence_wait(fence, true); 773 if (ret) 774 return ret; 775 } 776 } 777 778 if (!exclusive || !fobj) 779 return 0; 780 781 for (i = 0; i < fobj->shared_count; i++) { 782 fence = rcu_dereference_protected(fobj->shared[i], 783 dma_resv_held(obj->resv)); 784 if (fence->context != fctx->context) { 785 ret = dma_fence_wait(fence, true); 786 if (ret) 787 return ret; 788 } 789 } 790 791 return 0; 792 } 793 794 void msm_gem_active_get(struct drm_gem_object *obj, struct msm_gpu *gpu) 795 { 796 struct msm_gem_object *msm_obj = to_msm_bo(obj); 797 struct msm_drm_private *priv = obj->dev->dev_private; 798 799 might_sleep(); 800 GEM_WARN_ON(!msm_gem_is_locked(obj)); 801 GEM_WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED); 802 GEM_WARN_ON(msm_obj->dontneed); 803 804 if (msm_obj->active_count++ == 0) { 805 mutex_lock(&priv->mm_lock); 806 list_del(&msm_obj->mm_list); 807 list_add_tail(&msm_obj->mm_list, &gpu->active_list); 808 mutex_unlock(&priv->mm_lock); 809 } 810 } 811 812 void msm_gem_active_put(struct drm_gem_object *obj) 813 { 814 struct msm_gem_object *msm_obj = to_msm_bo(obj); 815 816 might_sleep(); 817 GEM_WARN_ON(!msm_gem_is_locked(obj)); 818 819 if (--msm_obj->active_count == 0) { 820 update_inactive(msm_obj); 821 } 822 } 823 824 static void update_inactive(struct msm_gem_object *msm_obj) 825 { 826 struct msm_drm_private *priv = msm_obj->base.dev->dev_private; 827 828 mutex_lock(&priv->mm_lock); 829 GEM_WARN_ON(msm_obj->active_count != 0); 830 831 if (msm_obj->dontneed) 832 mark_unpurgeable(msm_obj); 833 834 list_del(&msm_obj->mm_list); 835 if (msm_obj->madv == MSM_MADV_WILLNEED) { 836 list_add_tail(&msm_obj->mm_list, &priv->inactive_willneed); 837 } else if (msm_obj->madv == MSM_MADV_DONTNEED) { 838 list_add_tail(&msm_obj->mm_list, &priv->inactive_dontneed); 839 mark_purgeable(msm_obj); 840 } else { 841 GEM_WARN_ON(msm_obj->madv != __MSM_MADV_PURGED); 842 list_add_tail(&msm_obj->mm_list, &priv->inactive_purged); 843 } 844 845 mutex_unlock(&priv->mm_lock); 846 } 847 848 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout) 849 { 850 bool write = !!(op & MSM_PREP_WRITE); 851 unsigned long remain = 852 op & MSM_PREP_NOSYNC ? 0 : timeout_to_jiffies(timeout); 853 long ret; 854 855 ret = dma_resv_wait_timeout_rcu(obj->resv, write, 856 true, remain); 857 if (ret == 0) 858 return remain == 0 ? -EBUSY : -ETIMEDOUT; 859 else if (ret < 0) 860 return ret; 861 862 /* TODO cache maintenance */ 863 864 return 0; 865 } 866 867 int msm_gem_cpu_fini(struct drm_gem_object *obj) 868 { 869 /* TODO cache maintenance */ 870 return 0; 871 } 872 873 #ifdef CONFIG_DEBUG_FS 874 static void describe_fence(struct dma_fence *fence, const char *type, 875 struct seq_file *m) 876 { 877 if (!dma_fence_is_signaled(fence)) 878 seq_printf(m, "\t%9s: %s %s seq %llu\n", type, 879 fence->ops->get_driver_name(fence), 880 fence->ops->get_timeline_name(fence), 881 fence->seqno); 882 } 883 884 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m, 885 struct msm_gem_stats *stats) 886 { 887 struct msm_gem_object *msm_obj = to_msm_bo(obj); 888 struct dma_resv *robj = obj->resv; 889 struct dma_resv_list *fobj; 890 struct dma_fence *fence; 891 struct msm_gem_vma *vma; 892 uint64_t off = drm_vma_node_start(&obj->vma_node); 893 const char *madv; 894 895 msm_gem_lock(obj); 896 897 stats->all.count++; 898 stats->all.size += obj->size; 899 900 if (is_active(msm_obj)) { 901 stats->active.count++; 902 stats->active.size += obj->size; 903 } 904 905 if (msm_obj->pages) { 906 stats->resident.count++; 907 stats->resident.size += obj->size; 908 } 909 910 switch (msm_obj->madv) { 911 case __MSM_MADV_PURGED: 912 stats->purged.count++; 913 stats->purged.size += obj->size; 914 madv = " purged"; 915 break; 916 case MSM_MADV_DONTNEED: 917 stats->purgeable.count++; 918 stats->purgeable.size += obj->size; 919 madv = " purgeable"; 920 break; 921 case MSM_MADV_WILLNEED: 922 default: 923 madv = ""; 924 break; 925 } 926 927 seq_printf(m, "%08x: %c %2d (%2d) %08llx %p", 928 msm_obj->flags, is_active(msm_obj) ? 'A' : 'I', 929 obj->name, kref_read(&obj->refcount), 930 off, msm_obj->vaddr); 931 932 seq_printf(m, " %08zu %9s %-32s\n", obj->size, madv, msm_obj->name); 933 934 if (!list_empty(&msm_obj->vmas)) { 935 936 seq_puts(m, " vmas:"); 937 938 list_for_each_entry(vma, &msm_obj->vmas, list) { 939 const char *name, *comm; 940 if (vma->aspace) { 941 struct msm_gem_address_space *aspace = vma->aspace; 942 struct task_struct *task = 943 get_pid_task(aspace->pid, PIDTYPE_PID); 944 if (task) { 945 comm = kstrdup(task->comm, GFP_KERNEL); 946 } else { 947 comm = NULL; 948 } 949 name = aspace->name; 950 } else { 951 name = comm = NULL; 952 } 953 seq_printf(m, " [%s%s%s: aspace=%p, %08llx,%s,inuse=%d]", 954 name, comm ? ":" : "", comm ? comm : "", 955 vma->aspace, vma->iova, 956 vma->mapped ? "mapped" : "unmapped", 957 vma->inuse); 958 kfree(comm); 959 } 960 961 seq_puts(m, "\n"); 962 } 963 964 rcu_read_lock(); 965 fobj = rcu_dereference(robj->fence); 966 if (fobj) { 967 unsigned int i, shared_count = fobj->shared_count; 968 969 for (i = 0; i < shared_count; i++) { 970 fence = rcu_dereference(fobj->shared[i]); 971 describe_fence(fence, "Shared", m); 972 } 973 } 974 975 fence = rcu_dereference(robj->fence_excl); 976 if (fence) 977 describe_fence(fence, "Exclusive", m); 978 rcu_read_unlock(); 979 980 msm_gem_unlock(obj); 981 } 982 983 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m) 984 { 985 struct msm_gem_stats stats = {}; 986 struct msm_gem_object *msm_obj; 987 988 seq_puts(m, " flags id ref offset kaddr size madv name\n"); 989 list_for_each_entry(msm_obj, list, node) { 990 struct drm_gem_object *obj = &msm_obj->base; 991 seq_puts(m, " "); 992 msm_gem_describe(obj, m, &stats); 993 } 994 995 seq_printf(m, "Total: %4d objects, %9zu bytes\n", 996 stats.all.count, stats.all.size); 997 seq_printf(m, "Active: %4d objects, %9zu bytes\n", 998 stats.active.count, stats.active.size); 999 seq_printf(m, "Resident: %4d objects, %9zu bytes\n", 1000 stats.resident.count, stats.resident.size); 1001 seq_printf(m, "Purgeable: %4d objects, %9zu bytes\n", 1002 stats.purgeable.count, stats.purgeable.size); 1003 seq_printf(m, "Purged: %4d objects, %9zu bytes\n", 1004 stats.purged.count, stats.purged.size); 1005 } 1006 #endif 1007 1008 /* don't call directly! Use drm_gem_object_put_locked() and friends */ 1009 void msm_gem_free_object(struct drm_gem_object *obj) 1010 { 1011 struct msm_gem_object *msm_obj = to_msm_bo(obj); 1012 struct drm_device *dev = obj->dev; 1013 struct msm_drm_private *priv = dev->dev_private; 1014 1015 mutex_lock(&priv->obj_lock); 1016 list_del(&msm_obj->node); 1017 mutex_unlock(&priv->obj_lock); 1018 1019 mutex_lock(&priv->mm_lock); 1020 if (msm_obj->dontneed) 1021 mark_unpurgeable(msm_obj); 1022 list_del(&msm_obj->mm_list); 1023 mutex_unlock(&priv->mm_lock); 1024 1025 msm_gem_lock(obj); 1026 1027 /* object should not be on active list: */ 1028 GEM_WARN_ON(is_active(msm_obj)); 1029 1030 put_iova_spaces(obj, true); 1031 1032 if (obj->import_attach) { 1033 GEM_WARN_ON(msm_obj->vaddr); 1034 1035 /* Don't drop the pages for imported dmabuf, as they are not 1036 * ours, just free the array we allocated: 1037 */ 1038 kvfree(msm_obj->pages); 1039 1040 put_iova_vmas(obj); 1041 1042 /* dma_buf_detach() grabs resv lock, so we need to unlock 1043 * prior to drm_prime_gem_destroy 1044 */ 1045 msm_gem_unlock(obj); 1046 1047 drm_prime_gem_destroy(obj, msm_obj->sgt); 1048 } else { 1049 msm_gem_vunmap(obj); 1050 put_pages(obj); 1051 put_iova_vmas(obj); 1052 msm_gem_unlock(obj); 1053 } 1054 1055 drm_gem_object_release(obj); 1056 1057 kfree(msm_obj); 1058 } 1059 1060 /* convenience method to construct a GEM buffer object, and userspace handle */ 1061 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file, 1062 uint32_t size, uint32_t flags, uint32_t *handle, 1063 char *name) 1064 { 1065 struct drm_gem_object *obj; 1066 int ret; 1067 1068 obj = msm_gem_new(dev, size, flags); 1069 1070 if (IS_ERR(obj)) 1071 return PTR_ERR(obj); 1072 1073 if (name) 1074 msm_gem_object_set_name(obj, "%s", name); 1075 1076 ret = drm_gem_handle_create(file, obj, handle); 1077 1078 /* drop reference from allocate - handle holds it now */ 1079 drm_gem_object_put(obj); 1080 1081 return ret; 1082 } 1083 1084 static const struct vm_operations_struct vm_ops = { 1085 .fault = msm_gem_fault, 1086 .open = drm_gem_vm_open, 1087 .close = drm_gem_vm_close, 1088 }; 1089 1090 static const struct drm_gem_object_funcs msm_gem_object_funcs = { 1091 .free = msm_gem_free_object, 1092 .pin = msm_gem_prime_pin, 1093 .unpin = msm_gem_prime_unpin, 1094 .get_sg_table = msm_gem_prime_get_sg_table, 1095 .vmap = msm_gem_prime_vmap, 1096 .vunmap = msm_gem_prime_vunmap, 1097 .vm_ops = &vm_ops, 1098 }; 1099 1100 static int msm_gem_new_impl(struct drm_device *dev, 1101 uint32_t size, uint32_t flags, 1102 struct drm_gem_object **obj) 1103 { 1104 struct msm_gem_object *msm_obj; 1105 1106 switch (flags & MSM_BO_CACHE_MASK) { 1107 case MSM_BO_UNCACHED: 1108 case MSM_BO_CACHED: 1109 case MSM_BO_WC: 1110 break; 1111 default: 1112 DRM_DEV_ERROR(dev->dev, "invalid cache flag: %x\n", 1113 (flags & MSM_BO_CACHE_MASK)); 1114 return -EINVAL; 1115 } 1116 1117 msm_obj = kzalloc(sizeof(*msm_obj), GFP_KERNEL); 1118 if (!msm_obj) 1119 return -ENOMEM; 1120 1121 msm_obj->flags = flags; 1122 msm_obj->madv = MSM_MADV_WILLNEED; 1123 1124 INIT_LIST_HEAD(&msm_obj->submit_entry); 1125 INIT_LIST_HEAD(&msm_obj->vmas); 1126 1127 *obj = &msm_obj->base; 1128 (*obj)->funcs = &msm_gem_object_funcs; 1129 1130 return 0; 1131 } 1132 1133 static struct drm_gem_object *_msm_gem_new(struct drm_device *dev, 1134 uint32_t size, uint32_t flags, bool struct_mutex_locked) 1135 { 1136 struct msm_drm_private *priv = dev->dev_private; 1137 struct msm_gem_object *msm_obj; 1138 struct drm_gem_object *obj = NULL; 1139 bool use_vram = false; 1140 int ret; 1141 1142 size = PAGE_ALIGN(size); 1143 1144 if (!msm_use_mmu(dev)) 1145 use_vram = true; 1146 else if ((flags & (MSM_BO_STOLEN | MSM_BO_SCANOUT)) && priv->vram.size) 1147 use_vram = true; 1148 1149 if (GEM_WARN_ON(use_vram && !priv->vram.size)) 1150 return ERR_PTR(-EINVAL); 1151 1152 /* Disallow zero sized objects as they make the underlying 1153 * infrastructure grumpy 1154 */ 1155 if (size == 0) 1156 return ERR_PTR(-EINVAL); 1157 1158 ret = msm_gem_new_impl(dev, size, flags, &obj); 1159 if (ret) 1160 goto fail; 1161 1162 msm_obj = to_msm_bo(obj); 1163 1164 if (use_vram) { 1165 struct msm_gem_vma *vma; 1166 struct page **pages; 1167 1168 drm_gem_private_object_init(dev, obj, size); 1169 1170 msm_gem_lock(obj); 1171 1172 vma = add_vma(obj, NULL); 1173 msm_gem_unlock(obj); 1174 if (IS_ERR(vma)) { 1175 ret = PTR_ERR(vma); 1176 goto fail; 1177 } 1178 1179 to_msm_bo(obj)->vram_node = &vma->node; 1180 1181 msm_gem_lock(obj); 1182 pages = get_pages(obj); 1183 msm_gem_unlock(obj); 1184 if (IS_ERR(pages)) { 1185 ret = PTR_ERR(pages); 1186 goto fail; 1187 } 1188 1189 vma->iova = physaddr(obj); 1190 } else { 1191 ret = drm_gem_object_init(dev, obj, size); 1192 if (ret) 1193 goto fail; 1194 /* 1195 * Our buffers are kept pinned, so allocating them from the 1196 * MOVABLE zone is a really bad idea, and conflicts with CMA. 1197 * See comments above new_inode() why this is required _and_ 1198 * expected if you're going to pin these pages. 1199 */ 1200 mapping_set_gfp_mask(obj->filp->f_mapping, GFP_HIGHUSER); 1201 } 1202 1203 mutex_lock(&priv->mm_lock); 1204 /* Initially obj is idle, obj->madv == WILLNEED: */ 1205 list_add_tail(&msm_obj->mm_list, &priv->inactive_willneed); 1206 mutex_unlock(&priv->mm_lock); 1207 1208 mutex_lock(&priv->obj_lock); 1209 list_add_tail(&msm_obj->node, &priv->objects); 1210 mutex_unlock(&priv->obj_lock); 1211 1212 return obj; 1213 1214 fail: 1215 if (struct_mutex_locked) { 1216 drm_gem_object_put_locked(obj); 1217 } else { 1218 drm_gem_object_put(obj); 1219 } 1220 return ERR_PTR(ret); 1221 } 1222 1223 struct drm_gem_object *msm_gem_new_locked(struct drm_device *dev, 1224 uint32_t size, uint32_t flags) 1225 { 1226 return _msm_gem_new(dev, size, flags, true); 1227 } 1228 1229 struct drm_gem_object *msm_gem_new(struct drm_device *dev, 1230 uint32_t size, uint32_t flags) 1231 { 1232 return _msm_gem_new(dev, size, flags, false); 1233 } 1234 1235 struct drm_gem_object *msm_gem_import(struct drm_device *dev, 1236 struct dma_buf *dmabuf, struct sg_table *sgt) 1237 { 1238 struct msm_drm_private *priv = dev->dev_private; 1239 struct msm_gem_object *msm_obj; 1240 struct drm_gem_object *obj; 1241 uint32_t size; 1242 int ret, npages; 1243 1244 /* if we don't have IOMMU, don't bother pretending we can import: */ 1245 if (!msm_use_mmu(dev)) { 1246 DRM_DEV_ERROR(dev->dev, "cannot import without IOMMU\n"); 1247 return ERR_PTR(-EINVAL); 1248 } 1249 1250 size = PAGE_ALIGN(dmabuf->size); 1251 1252 ret = msm_gem_new_impl(dev, size, MSM_BO_WC, &obj); 1253 if (ret) 1254 goto fail; 1255 1256 drm_gem_private_object_init(dev, obj, size); 1257 1258 npages = size / PAGE_SIZE; 1259 1260 msm_obj = to_msm_bo(obj); 1261 msm_gem_lock(obj); 1262 msm_obj->sgt = sgt; 1263 msm_obj->pages = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL); 1264 if (!msm_obj->pages) { 1265 msm_gem_unlock(obj); 1266 ret = -ENOMEM; 1267 goto fail; 1268 } 1269 1270 ret = drm_prime_sg_to_page_array(sgt, msm_obj->pages, npages); 1271 if (ret) { 1272 msm_gem_unlock(obj); 1273 goto fail; 1274 } 1275 1276 msm_gem_unlock(obj); 1277 1278 mutex_lock(&priv->mm_lock); 1279 list_add_tail(&msm_obj->mm_list, &priv->inactive_willneed); 1280 mutex_unlock(&priv->mm_lock); 1281 1282 mutex_lock(&priv->obj_lock); 1283 list_add_tail(&msm_obj->node, &priv->objects); 1284 mutex_unlock(&priv->obj_lock); 1285 1286 return obj; 1287 1288 fail: 1289 drm_gem_object_put(obj); 1290 return ERR_PTR(ret); 1291 } 1292 1293 static void *_msm_gem_kernel_new(struct drm_device *dev, uint32_t size, 1294 uint32_t flags, struct msm_gem_address_space *aspace, 1295 struct drm_gem_object **bo, uint64_t *iova, bool locked) 1296 { 1297 void *vaddr; 1298 struct drm_gem_object *obj = _msm_gem_new(dev, size, flags, locked); 1299 int ret; 1300 1301 if (IS_ERR(obj)) 1302 return ERR_CAST(obj); 1303 1304 if (iova) { 1305 ret = msm_gem_get_and_pin_iova(obj, aspace, iova); 1306 if (ret) 1307 goto err; 1308 } 1309 1310 vaddr = msm_gem_get_vaddr(obj); 1311 if (IS_ERR(vaddr)) { 1312 msm_gem_unpin_iova(obj, aspace); 1313 ret = PTR_ERR(vaddr); 1314 goto err; 1315 } 1316 1317 if (bo) 1318 *bo = obj; 1319 1320 return vaddr; 1321 err: 1322 if (locked) 1323 drm_gem_object_put_locked(obj); 1324 else 1325 drm_gem_object_put(obj); 1326 1327 return ERR_PTR(ret); 1328 1329 } 1330 1331 void *msm_gem_kernel_new(struct drm_device *dev, uint32_t size, 1332 uint32_t flags, struct msm_gem_address_space *aspace, 1333 struct drm_gem_object **bo, uint64_t *iova) 1334 { 1335 return _msm_gem_kernel_new(dev, size, flags, aspace, bo, iova, false); 1336 } 1337 1338 void *msm_gem_kernel_new_locked(struct drm_device *dev, uint32_t size, 1339 uint32_t flags, struct msm_gem_address_space *aspace, 1340 struct drm_gem_object **bo, uint64_t *iova) 1341 { 1342 return _msm_gem_kernel_new(dev, size, flags, aspace, bo, iova, true); 1343 } 1344 1345 void msm_gem_kernel_put(struct drm_gem_object *bo, 1346 struct msm_gem_address_space *aspace, bool locked) 1347 { 1348 if (IS_ERR_OR_NULL(bo)) 1349 return; 1350 1351 msm_gem_put_vaddr(bo); 1352 msm_gem_unpin_iova(bo, aspace); 1353 1354 if (locked) 1355 drm_gem_object_put_locked(bo); 1356 else 1357 drm_gem_object_put(bo); 1358 } 1359 1360 void msm_gem_object_set_name(struct drm_gem_object *bo, const char *fmt, ...) 1361 { 1362 struct msm_gem_object *msm_obj = to_msm_bo(bo); 1363 va_list ap; 1364 1365 if (!fmt) 1366 return; 1367 1368 va_start(ap, fmt); 1369 vsnprintf(msm_obj->name, sizeof(msm_obj->name), fmt, ap); 1370 va_end(ap); 1371 } 1372