1 /* 2 * Copyright (C) 2013 Red Hat 3 * Author: Rob Clark <robdclark@gmail.com> 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 as published by 7 * the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 12 * more details. 13 * 14 * You should have received a copy of the GNU General Public License along with 15 * this program. If not, see <http://www.gnu.org/licenses/>. 16 */ 17 18 #include <linux/spinlock.h> 19 #include <linux/shmem_fs.h> 20 #include <linux/dma-buf.h> 21 #include <linux/pfn_t.h> 22 23 #include "msm_drv.h" 24 #include "msm_fence.h" 25 #include "msm_gem.h" 26 #include "msm_gpu.h" 27 #include "msm_mmu.h" 28 29 static dma_addr_t physaddr(struct drm_gem_object *obj) 30 { 31 struct msm_gem_object *msm_obj = to_msm_bo(obj); 32 struct msm_drm_private *priv = obj->dev->dev_private; 33 return (((dma_addr_t)msm_obj->vram_node->start) << PAGE_SHIFT) + 34 priv->vram.paddr; 35 } 36 37 static bool use_pages(struct drm_gem_object *obj) 38 { 39 struct msm_gem_object *msm_obj = to_msm_bo(obj); 40 return !msm_obj->vram_node; 41 } 42 43 /* allocate pages from VRAM carveout, used when no IOMMU: */ 44 static struct page **get_pages_vram(struct drm_gem_object *obj, 45 int npages) 46 { 47 struct msm_gem_object *msm_obj = to_msm_bo(obj); 48 struct msm_drm_private *priv = obj->dev->dev_private; 49 dma_addr_t paddr; 50 struct page **p; 51 int ret, i; 52 53 p = drm_malloc_ab(npages, sizeof(struct page *)); 54 if (!p) 55 return ERR_PTR(-ENOMEM); 56 57 ret = drm_mm_insert_node(&priv->vram.mm, msm_obj->vram_node, 58 npages, 0, DRM_MM_SEARCH_DEFAULT); 59 if (ret) { 60 drm_free_large(p); 61 return ERR_PTR(ret); 62 } 63 64 paddr = physaddr(obj); 65 for (i = 0; i < npages; i++) { 66 p[i] = phys_to_page(paddr); 67 paddr += PAGE_SIZE; 68 } 69 70 return p; 71 } 72 73 /* called with dev->struct_mutex held */ 74 static struct page **get_pages(struct drm_gem_object *obj) 75 { 76 struct msm_gem_object *msm_obj = to_msm_bo(obj); 77 78 if (!msm_obj->pages) { 79 struct drm_device *dev = obj->dev; 80 struct page **p; 81 int npages = obj->size >> PAGE_SHIFT; 82 83 if (use_pages(obj)) 84 p = drm_gem_get_pages(obj); 85 else 86 p = get_pages_vram(obj, npages); 87 88 if (IS_ERR(p)) { 89 dev_err(dev->dev, "could not get pages: %ld\n", 90 PTR_ERR(p)); 91 return p; 92 } 93 94 msm_obj->sgt = drm_prime_pages_to_sg(p, npages); 95 if (IS_ERR(msm_obj->sgt)) { 96 dev_err(dev->dev, "failed to allocate sgt\n"); 97 return ERR_CAST(msm_obj->sgt); 98 } 99 100 msm_obj->pages = p; 101 102 /* For non-cached buffers, ensure the new pages are clean 103 * because display controller, GPU, etc. are not coherent: 104 */ 105 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED)) 106 dma_map_sg(dev->dev, msm_obj->sgt->sgl, 107 msm_obj->sgt->nents, DMA_BIDIRECTIONAL); 108 } 109 110 return msm_obj->pages; 111 } 112 113 static void put_pages(struct drm_gem_object *obj) 114 { 115 struct msm_gem_object *msm_obj = to_msm_bo(obj); 116 117 if (msm_obj->pages) { 118 /* For non-cached buffers, ensure the new pages are clean 119 * because display controller, GPU, etc. are not coherent: 120 */ 121 if (msm_obj->flags & (MSM_BO_WC|MSM_BO_UNCACHED)) 122 dma_unmap_sg(obj->dev->dev, msm_obj->sgt->sgl, 123 msm_obj->sgt->nents, DMA_BIDIRECTIONAL); 124 sg_free_table(msm_obj->sgt); 125 kfree(msm_obj->sgt); 126 127 if (use_pages(obj)) 128 drm_gem_put_pages(obj, msm_obj->pages, true, false); 129 else { 130 drm_mm_remove_node(msm_obj->vram_node); 131 drm_free_large(msm_obj->pages); 132 } 133 134 msm_obj->pages = NULL; 135 } 136 } 137 138 struct page **msm_gem_get_pages(struct drm_gem_object *obj) 139 { 140 struct drm_device *dev = obj->dev; 141 struct page **p; 142 mutex_lock(&dev->struct_mutex); 143 p = get_pages(obj); 144 mutex_unlock(&dev->struct_mutex); 145 return p; 146 } 147 148 void msm_gem_put_pages(struct drm_gem_object *obj) 149 { 150 /* when we start tracking the pin count, then do something here */ 151 } 152 153 int msm_gem_mmap_obj(struct drm_gem_object *obj, 154 struct vm_area_struct *vma) 155 { 156 struct msm_gem_object *msm_obj = to_msm_bo(obj); 157 158 vma->vm_flags &= ~VM_PFNMAP; 159 vma->vm_flags |= VM_MIXEDMAP; 160 161 if (msm_obj->flags & MSM_BO_WC) { 162 vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags)); 163 } else if (msm_obj->flags & MSM_BO_UNCACHED) { 164 vma->vm_page_prot = pgprot_noncached(vm_get_page_prot(vma->vm_flags)); 165 } else { 166 /* 167 * Shunt off cached objs to shmem file so they have their own 168 * address_space (so unmap_mapping_range does what we want, 169 * in particular in the case of mmap'd dmabufs) 170 */ 171 fput(vma->vm_file); 172 get_file(obj->filp); 173 vma->vm_pgoff = 0; 174 vma->vm_file = obj->filp; 175 176 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 177 } 178 179 return 0; 180 } 181 182 int msm_gem_mmap(struct file *filp, struct vm_area_struct *vma) 183 { 184 int ret; 185 186 ret = drm_gem_mmap(filp, vma); 187 if (ret) { 188 DBG("mmap failed: %d", ret); 189 return ret; 190 } 191 192 return msm_gem_mmap_obj(vma->vm_private_data, vma); 193 } 194 195 int msm_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 196 { 197 struct drm_gem_object *obj = vma->vm_private_data; 198 struct drm_device *dev = obj->dev; 199 struct page **pages; 200 unsigned long pfn; 201 pgoff_t pgoff; 202 int ret; 203 204 /* Make sure we don't parallel update on a fault, nor move or remove 205 * something from beneath our feet 206 */ 207 ret = mutex_lock_interruptible(&dev->struct_mutex); 208 if (ret) 209 goto out; 210 211 /* make sure we have pages attached now */ 212 pages = get_pages(obj); 213 if (IS_ERR(pages)) { 214 ret = PTR_ERR(pages); 215 goto out_unlock; 216 } 217 218 /* We don't use vmf->pgoff since that has the fake offset: */ 219 pgoff = ((unsigned long)vmf->virtual_address - 220 vma->vm_start) >> PAGE_SHIFT; 221 222 pfn = page_to_pfn(pages[pgoff]); 223 224 VERB("Inserting %p pfn %lx, pa %lx", vmf->virtual_address, 225 pfn, pfn << PAGE_SHIFT); 226 227 ret = vm_insert_mixed(vma, (unsigned long)vmf->virtual_address, 228 __pfn_to_pfn_t(pfn, PFN_DEV)); 229 230 out_unlock: 231 mutex_unlock(&dev->struct_mutex); 232 out: 233 switch (ret) { 234 case -EAGAIN: 235 case 0: 236 case -ERESTARTSYS: 237 case -EINTR: 238 case -EBUSY: 239 /* 240 * EBUSY is ok: this just means that another thread 241 * already did the job. 242 */ 243 return VM_FAULT_NOPAGE; 244 case -ENOMEM: 245 return VM_FAULT_OOM; 246 default: 247 return VM_FAULT_SIGBUS; 248 } 249 } 250 251 /** get mmap offset */ 252 static uint64_t mmap_offset(struct drm_gem_object *obj) 253 { 254 struct drm_device *dev = obj->dev; 255 int ret; 256 257 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 258 259 /* Make it mmapable */ 260 ret = drm_gem_create_mmap_offset(obj); 261 262 if (ret) { 263 dev_err(dev->dev, "could not allocate mmap offset\n"); 264 return 0; 265 } 266 267 return drm_vma_node_offset_addr(&obj->vma_node); 268 } 269 270 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj) 271 { 272 uint64_t offset; 273 mutex_lock(&obj->dev->struct_mutex); 274 offset = mmap_offset(obj); 275 mutex_unlock(&obj->dev->struct_mutex); 276 return offset; 277 } 278 279 /* should be called under struct_mutex.. although it can be called 280 * from atomic context without struct_mutex to acquire an extra 281 * iova ref if you know one is already held. 282 * 283 * That means when I do eventually need to add support for unpinning 284 * the refcnt counter needs to be atomic_t. 285 */ 286 int msm_gem_get_iova_locked(struct drm_gem_object *obj, int id, 287 uint32_t *iova) 288 { 289 struct msm_gem_object *msm_obj = to_msm_bo(obj); 290 int ret = 0; 291 292 if (!msm_obj->domain[id].iova) { 293 struct msm_drm_private *priv = obj->dev->dev_private; 294 struct page **pages = get_pages(obj); 295 296 if (IS_ERR(pages)) 297 return PTR_ERR(pages); 298 299 if (iommu_present(&platform_bus_type)) { 300 struct msm_mmu *mmu = priv->mmus[id]; 301 uint32_t offset; 302 303 if (WARN_ON(!mmu)) 304 return -EINVAL; 305 306 offset = (uint32_t)mmap_offset(obj); 307 ret = mmu->funcs->map(mmu, offset, msm_obj->sgt, 308 obj->size, IOMMU_READ | IOMMU_WRITE); 309 msm_obj->domain[id].iova = offset; 310 } else { 311 msm_obj->domain[id].iova = physaddr(obj); 312 } 313 } 314 315 if (!ret) 316 *iova = msm_obj->domain[id].iova; 317 318 return ret; 319 } 320 321 /* get iova, taking a reference. Should have a matching put */ 322 int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint32_t *iova) 323 { 324 struct msm_gem_object *msm_obj = to_msm_bo(obj); 325 int ret; 326 327 /* this is safe right now because we don't unmap until the 328 * bo is deleted: 329 */ 330 if (msm_obj->domain[id].iova) { 331 *iova = msm_obj->domain[id].iova; 332 return 0; 333 } 334 335 mutex_lock(&obj->dev->struct_mutex); 336 ret = msm_gem_get_iova_locked(obj, id, iova); 337 mutex_unlock(&obj->dev->struct_mutex); 338 return ret; 339 } 340 341 /* get iova without taking a reference, used in places where you have 342 * already done a 'msm_gem_get_iova()'. 343 */ 344 uint32_t msm_gem_iova(struct drm_gem_object *obj, int id) 345 { 346 struct msm_gem_object *msm_obj = to_msm_bo(obj); 347 WARN_ON(!msm_obj->domain[id].iova); 348 return msm_obj->domain[id].iova; 349 } 350 351 void msm_gem_put_iova(struct drm_gem_object *obj, int id) 352 { 353 // XXX TODO .. 354 // NOTE: probably don't need a _locked() version.. we wouldn't 355 // normally unmap here, but instead just mark that it could be 356 // unmapped (if the iova refcnt drops to zero), but then later 357 // if another _get_iova_locked() fails we can start unmapping 358 // things that are no longer needed.. 359 } 360 361 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev, 362 struct drm_mode_create_dumb *args) 363 { 364 args->pitch = align_pitch(args->width, args->bpp); 365 args->size = PAGE_ALIGN(args->pitch * args->height); 366 return msm_gem_new_handle(dev, file, args->size, 367 MSM_BO_SCANOUT | MSM_BO_WC, &args->handle); 368 } 369 370 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev, 371 uint32_t handle, uint64_t *offset) 372 { 373 struct drm_gem_object *obj; 374 int ret = 0; 375 376 /* GEM does all our handle to object mapping */ 377 obj = drm_gem_object_lookup(dev, file, handle); 378 if (obj == NULL) { 379 ret = -ENOENT; 380 goto fail; 381 } 382 383 *offset = msm_gem_mmap_offset(obj); 384 385 drm_gem_object_unreference_unlocked(obj); 386 387 fail: 388 return ret; 389 } 390 391 void *msm_gem_vaddr_locked(struct drm_gem_object *obj) 392 { 393 struct msm_gem_object *msm_obj = to_msm_bo(obj); 394 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex)); 395 if (!msm_obj->vaddr) { 396 struct page **pages = get_pages(obj); 397 if (IS_ERR(pages)) 398 return ERR_CAST(pages); 399 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT, 400 VM_MAP, pgprot_writecombine(PAGE_KERNEL)); 401 } 402 return msm_obj->vaddr; 403 } 404 405 void *msm_gem_vaddr(struct drm_gem_object *obj) 406 { 407 void *ret; 408 mutex_lock(&obj->dev->struct_mutex); 409 ret = msm_gem_vaddr_locked(obj); 410 mutex_unlock(&obj->dev->struct_mutex); 411 return ret; 412 } 413 414 void msm_gem_move_to_active(struct drm_gem_object *obj, 415 struct msm_gpu *gpu, bool write, uint32_t fence) 416 { 417 struct msm_gem_object *msm_obj = to_msm_bo(obj); 418 msm_obj->gpu = gpu; 419 if (write) 420 msm_obj->write_fence = fence; 421 else 422 msm_obj->read_fence = fence; 423 list_del_init(&msm_obj->mm_list); 424 list_add_tail(&msm_obj->mm_list, &gpu->active_list); 425 } 426 427 void msm_gem_move_to_inactive(struct drm_gem_object *obj) 428 { 429 struct drm_device *dev = obj->dev; 430 struct msm_drm_private *priv = dev->dev_private; 431 struct msm_gem_object *msm_obj = to_msm_bo(obj); 432 433 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 434 435 msm_obj->gpu = NULL; 436 msm_obj->read_fence = 0; 437 msm_obj->write_fence = 0; 438 list_del_init(&msm_obj->mm_list); 439 list_add_tail(&msm_obj->mm_list, &priv->inactive_list); 440 } 441 442 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout) 443 { 444 struct drm_device *dev = obj->dev; 445 struct msm_drm_private *priv = dev->dev_private; 446 struct msm_gem_object *msm_obj = to_msm_bo(obj); 447 int ret = 0; 448 449 if (is_active(msm_obj)) { 450 uint32_t fence = msm_gem_fence(msm_obj, op); 451 452 if (op & MSM_PREP_NOSYNC) 453 timeout = NULL; 454 455 if (priv->gpu) 456 ret = msm_wait_fence(priv->gpu->fctx, fence, timeout, true); 457 } 458 459 /* TODO cache maintenance */ 460 461 return ret; 462 } 463 464 int msm_gem_cpu_fini(struct drm_gem_object *obj) 465 { 466 /* TODO cache maintenance */ 467 return 0; 468 } 469 470 #ifdef CONFIG_DEBUG_FS 471 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m) 472 { 473 struct drm_device *dev = obj->dev; 474 struct msm_gem_object *msm_obj = to_msm_bo(obj); 475 uint64_t off = drm_vma_node_start(&obj->vma_node); 476 477 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 478 seq_printf(m, "%08x: %c(r=%u,w=%u) %2d (%2d) %08llx %p %zu\n", 479 msm_obj->flags, is_active(msm_obj) ? 'A' : 'I', 480 msm_obj->read_fence, msm_obj->write_fence, 481 obj->name, obj->refcount.refcount.counter, 482 off, msm_obj->vaddr, obj->size); 483 } 484 485 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m) 486 { 487 struct msm_gem_object *msm_obj; 488 int count = 0; 489 size_t size = 0; 490 491 list_for_each_entry(msm_obj, list, mm_list) { 492 struct drm_gem_object *obj = &msm_obj->base; 493 seq_printf(m, " "); 494 msm_gem_describe(obj, m); 495 count++; 496 size += obj->size; 497 } 498 499 seq_printf(m, "Total %d objects, %zu bytes\n", count, size); 500 } 501 #endif 502 503 void msm_gem_free_object(struct drm_gem_object *obj) 504 { 505 struct drm_device *dev = obj->dev; 506 struct msm_drm_private *priv = obj->dev->dev_private; 507 struct msm_gem_object *msm_obj = to_msm_bo(obj); 508 int id; 509 510 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 511 512 /* object should not be on active list: */ 513 WARN_ON(is_active(msm_obj)); 514 515 list_del(&msm_obj->mm_list); 516 517 for (id = 0; id < ARRAY_SIZE(msm_obj->domain); id++) { 518 struct msm_mmu *mmu = priv->mmus[id]; 519 if (mmu && msm_obj->domain[id].iova) { 520 uint32_t offset = msm_obj->domain[id].iova; 521 mmu->funcs->unmap(mmu, offset, msm_obj->sgt, obj->size); 522 } 523 } 524 525 if (obj->import_attach) { 526 if (msm_obj->vaddr) 527 dma_buf_vunmap(obj->import_attach->dmabuf, msm_obj->vaddr); 528 529 /* Don't drop the pages for imported dmabuf, as they are not 530 * ours, just free the array we allocated: 531 */ 532 if (msm_obj->pages) 533 drm_free_large(msm_obj->pages); 534 535 drm_prime_gem_destroy(obj, msm_obj->sgt); 536 } else { 537 vunmap(msm_obj->vaddr); 538 put_pages(obj); 539 } 540 541 if (msm_obj->resv == &msm_obj->_resv) 542 reservation_object_fini(msm_obj->resv); 543 544 drm_gem_object_release(obj); 545 546 kfree(msm_obj); 547 } 548 549 /* convenience method to construct a GEM buffer object, and userspace handle */ 550 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file, 551 uint32_t size, uint32_t flags, uint32_t *handle) 552 { 553 struct drm_gem_object *obj; 554 int ret; 555 556 ret = mutex_lock_interruptible(&dev->struct_mutex); 557 if (ret) 558 return ret; 559 560 obj = msm_gem_new(dev, size, flags); 561 562 mutex_unlock(&dev->struct_mutex); 563 564 if (IS_ERR(obj)) 565 return PTR_ERR(obj); 566 567 ret = drm_gem_handle_create(file, obj, handle); 568 569 /* drop reference from allocate - handle holds it now */ 570 drm_gem_object_unreference_unlocked(obj); 571 572 return ret; 573 } 574 575 static int msm_gem_new_impl(struct drm_device *dev, 576 uint32_t size, uint32_t flags, 577 struct reservation_object *resv, 578 struct drm_gem_object **obj) 579 { 580 struct msm_drm_private *priv = dev->dev_private; 581 struct msm_gem_object *msm_obj; 582 unsigned sz; 583 bool use_vram = false; 584 585 switch (flags & MSM_BO_CACHE_MASK) { 586 case MSM_BO_UNCACHED: 587 case MSM_BO_CACHED: 588 case MSM_BO_WC: 589 break; 590 default: 591 dev_err(dev->dev, "invalid cache flag: %x\n", 592 (flags & MSM_BO_CACHE_MASK)); 593 return -EINVAL; 594 } 595 596 if (!iommu_present(&platform_bus_type)) 597 use_vram = true; 598 else if ((flags & MSM_BO_STOLEN) && priv->vram.size) 599 use_vram = true; 600 601 if (WARN_ON(use_vram && !priv->vram.size)) 602 return -EINVAL; 603 604 sz = sizeof(*msm_obj); 605 if (use_vram) 606 sz += sizeof(struct drm_mm_node); 607 608 msm_obj = kzalloc(sz, GFP_KERNEL); 609 if (!msm_obj) 610 return -ENOMEM; 611 612 if (use_vram) 613 msm_obj->vram_node = (void *)&msm_obj[1]; 614 615 msm_obj->flags = flags; 616 617 if (resv) { 618 msm_obj->resv = resv; 619 } else { 620 msm_obj->resv = &msm_obj->_resv; 621 reservation_object_init(msm_obj->resv); 622 } 623 624 INIT_LIST_HEAD(&msm_obj->submit_entry); 625 list_add_tail(&msm_obj->mm_list, &priv->inactive_list); 626 627 *obj = &msm_obj->base; 628 629 return 0; 630 } 631 632 struct drm_gem_object *msm_gem_new(struct drm_device *dev, 633 uint32_t size, uint32_t flags) 634 { 635 struct drm_gem_object *obj = NULL; 636 int ret; 637 638 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 639 640 size = PAGE_ALIGN(size); 641 642 ret = msm_gem_new_impl(dev, size, flags, NULL, &obj); 643 if (ret) 644 goto fail; 645 646 if (use_pages(obj)) { 647 ret = drm_gem_object_init(dev, obj, size); 648 if (ret) 649 goto fail; 650 } else { 651 drm_gem_private_object_init(dev, obj, size); 652 } 653 654 return obj; 655 656 fail: 657 if (obj) 658 drm_gem_object_unreference(obj); 659 660 return ERR_PTR(ret); 661 } 662 663 struct drm_gem_object *msm_gem_import(struct drm_device *dev, 664 struct dma_buf *dmabuf, struct sg_table *sgt) 665 { 666 struct msm_gem_object *msm_obj; 667 struct drm_gem_object *obj; 668 uint32_t size; 669 int ret, npages; 670 671 /* if we don't have IOMMU, don't bother pretending we can import: */ 672 if (!iommu_present(&platform_bus_type)) { 673 dev_err(dev->dev, "cannot import without IOMMU\n"); 674 return ERR_PTR(-EINVAL); 675 } 676 677 size = PAGE_ALIGN(dmabuf->size); 678 679 ret = msm_gem_new_impl(dev, size, MSM_BO_WC, dmabuf->resv, &obj); 680 if (ret) 681 goto fail; 682 683 drm_gem_private_object_init(dev, obj, size); 684 685 npages = size / PAGE_SIZE; 686 687 msm_obj = to_msm_bo(obj); 688 msm_obj->sgt = sgt; 689 msm_obj->pages = drm_malloc_ab(npages, sizeof(struct page *)); 690 if (!msm_obj->pages) { 691 ret = -ENOMEM; 692 goto fail; 693 } 694 695 ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages); 696 if (ret) 697 goto fail; 698 699 return obj; 700 701 fail: 702 if (obj) 703 drm_gem_object_unreference_unlocked(obj); 704 705 return ERR_PTR(ret); 706 } 707