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