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 page **pages = get_pages(obj); 286 287 if (IS_ERR(pages)) 288 return PTR_ERR(pages); 289 290 if (iommu_present(&platform_bus_type)) { 291 struct msm_mmu *mmu = priv->mmus[id]; 292 uint32_t offset; 293 294 if (WARN_ON(!mmu)) 295 return -EINVAL; 296 297 offset = (uint32_t)mmap_offset(obj); 298 ret = mmu->funcs->map(mmu, offset, msm_obj->sgt, 299 obj->size, IOMMU_READ | IOMMU_WRITE); 300 msm_obj->domain[id].iova = offset; 301 } else { 302 msm_obj->domain[id].iova = physaddr(obj); 303 } 304 } 305 306 if (!ret) 307 *iova = msm_obj->domain[id].iova; 308 309 return ret; 310 } 311 312 /* get iova, taking a reference. Should have a matching put */ 313 int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint32_t *iova) 314 { 315 struct msm_gem_object *msm_obj = to_msm_bo(obj); 316 int ret; 317 318 /* this is safe right now because we don't unmap until the 319 * bo is deleted: 320 */ 321 if (msm_obj->domain[id].iova) { 322 *iova = msm_obj->domain[id].iova; 323 return 0; 324 } 325 326 mutex_lock(&obj->dev->struct_mutex); 327 ret = msm_gem_get_iova_locked(obj, id, iova); 328 mutex_unlock(&obj->dev->struct_mutex); 329 return ret; 330 } 331 332 /* get iova without taking a reference, used in places where you have 333 * already done a 'msm_gem_get_iova()'. 334 */ 335 uint32_t msm_gem_iova(struct drm_gem_object *obj, int id) 336 { 337 struct msm_gem_object *msm_obj = to_msm_bo(obj); 338 WARN_ON(!msm_obj->domain[id].iova); 339 return msm_obj->domain[id].iova; 340 } 341 342 void msm_gem_put_iova(struct drm_gem_object *obj, int id) 343 { 344 // XXX TODO .. 345 // NOTE: probably don't need a _locked() version.. we wouldn't 346 // normally unmap here, but instead just mark that it could be 347 // unmapped (if the iova refcnt drops to zero), but then later 348 // if another _get_iova_locked() fails we can start unmapping 349 // things that are no longer needed.. 350 } 351 352 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev, 353 struct drm_mode_create_dumb *args) 354 { 355 args->pitch = align_pitch(args->width, args->bpp); 356 args->size = PAGE_ALIGN(args->pitch * args->height); 357 return msm_gem_new_handle(dev, file, args->size, 358 MSM_BO_SCANOUT | MSM_BO_WC, &args->handle); 359 } 360 361 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev, 362 uint32_t handle, uint64_t *offset) 363 { 364 struct drm_gem_object *obj; 365 int ret = 0; 366 367 /* GEM does all our handle to object mapping */ 368 obj = drm_gem_object_lookup(dev, file, handle); 369 if (obj == NULL) { 370 ret = -ENOENT; 371 goto fail; 372 } 373 374 *offset = msm_gem_mmap_offset(obj); 375 376 drm_gem_object_unreference_unlocked(obj); 377 378 fail: 379 return ret; 380 } 381 382 void *msm_gem_vaddr_locked(struct drm_gem_object *obj) 383 { 384 struct msm_gem_object *msm_obj = to_msm_bo(obj); 385 WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex)); 386 if (!msm_obj->vaddr) { 387 struct page **pages = get_pages(obj); 388 if (IS_ERR(pages)) 389 return ERR_CAST(pages); 390 msm_obj->vaddr = vmap(pages, obj->size >> PAGE_SHIFT, 391 VM_MAP, pgprot_writecombine(PAGE_KERNEL)); 392 } 393 return msm_obj->vaddr; 394 } 395 396 void *msm_gem_vaddr(struct drm_gem_object *obj) 397 { 398 void *ret; 399 mutex_lock(&obj->dev->struct_mutex); 400 ret = msm_gem_vaddr_locked(obj); 401 mutex_unlock(&obj->dev->struct_mutex); 402 return ret; 403 } 404 405 /* setup callback for when bo is no longer busy.. 406 * TODO probably want to differentiate read vs write.. 407 */ 408 int msm_gem_queue_inactive_cb(struct drm_gem_object *obj, 409 struct msm_fence_cb *cb) 410 { 411 struct msm_gem_object *msm_obj = to_msm_bo(obj); 412 uint32_t fence = msm_gem_fence(msm_obj, 413 MSM_PREP_READ | MSM_PREP_WRITE); 414 return msm_queue_fence_cb(obj->dev, cb, fence); 415 } 416 417 void msm_gem_move_to_active(struct drm_gem_object *obj, 418 struct msm_gpu *gpu, bool write, uint32_t fence) 419 { 420 struct msm_gem_object *msm_obj = to_msm_bo(obj); 421 msm_obj->gpu = gpu; 422 if (write) 423 msm_obj->write_fence = fence; 424 else 425 msm_obj->read_fence = fence; 426 list_del_init(&msm_obj->mm_list); 427 list_add_tail(&msm_obj->mm_list, &gpu->active_list); 428 } 429 430 void msm_gem_move_to_inactive(struct drm_gem_object *obj) 431 { 432 struct drm_device *dev = obj->dev; 433 struct msm_drm_private *priv = dev->dev_private; 434 struct msm_gem_object *msm_obj = to_msm_bo(obj); 435 436 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 437 438 msm_obj->gpu = NULL; 439 msm_obj->read_fence = 0; 440 msm_obj->write_fence = 0; 441 list_del_init(&msm_obj->mm_list); 442 list_add_tail(&msm_obj->mm_list, &priv->inactive_list); 443 } 444 445 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, 446 struct timespec *timeout) 447 { 448 struct drm_device *dev = obj->dev; 449 struct msm_gem_object *msm_obj = to_msm_bo(obj); 450 int ret = 0; 451 452 if (is_active(msm_obj)) { 453 uint32_t fence = msm_gem_fence(msm_obj, op); 454 455 if (op & MSM_PREP_NOSYNC) 456 timeout = NULL; 457 458 ret = msm_wait_fence_interruptable(dev, fence, timeout); 459 } 460 461 /* TODO cache maintenance */ 462 463 return ret; 464 } 465 466 int msm_gem_cpu_fini(struct drm_gem_object *obj) 467 { 468 /* TODO cache maintenance */ 469 return 0; 470 } 471 472 #ifdef CONFIG_DEBUG_FS 473 void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m) 474 { 475 struct drm_device *dev = obj->dev; 476 struct msm_gem_object *msm_obj = to_msm_bo(obj); 477 uint64_t off = drm_vma_node_start(&obj->vma_node); 478 479 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 480 seq_printf(m, "%08x: %c(r=%u,w=%u) %2d (%2d) %08llx %p %d\n", 481 msm_obj->flags, is_active(msm_obj) ? 'A' : 'I', 482 msm_obj->read_fence, msm_obj->write_fence, 483 obj->name, obj->refcount.refcount.counter, 484 off, msm_obj->vaddr, obj->size); 485 } 486 487 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m) 488 { 489 struct msm_gem_object *msm_obj; 490 int count = 0; 491 size_t size = 0; 492 493 list_for_each_entry(msm_obj, list, mm_list) { 494 struct drm_gem_object *obj = &msm_obj->base; 495 seq_printf(m, " "); 496 msm_gem_describe(obj, m); 497 count++; 498 size += obj->size; 499 } 500 501 seq_printf(m, "Total %d objects, %zu bytes\n", count, size); 502 } 503 #endif 504 505 void msm_gem_free_object(struct drm_gem_object *obj) 506 { 507 struct drm_device *dev = obj->dev; 508 struct msm_drm_private *priv = obj->dev->dev_private; 509 struct msm_gem_object *msm_obj = to_msm_bo(obj); 510 int id; 511 512 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 513 514 /* object should not be on active list: */ 515 WARN_ON(is_active(msm_obj)); 516 517 list_del(&msm_obj->mm_list); 518 519 for (id = 0; id < ARRAY_SIZE(msm_obj->domain); id++) { 520 struct msm_mmu *mmu = priv->mmus[id]; 521 if (mmu && msm_obj->domain[id].iova) { 522 uint32_t offset = msm_obj->domain[id].iova; 523 mmu->funcs->unmap(mmu, offset, msm_obj->sgt, obj->size); 524 } 525 } 526 527 if (obj->import_attach) { 528 if (msm_obj->vaddr) 529 dma_buf_vunmap(obj->import_attach->dmabuf, msm_obj->vaddr); 530 531 /* Don't drop the pages for imported dmabuf, as they are not 532 * ours, just free the array we allocated: 533 */ 534 if (msm_obj->pages) 535 drm_free_large(msm_obj->pages); 536 537 } else { 538 vunmap(msm_obj->vaddr); 539 put_pages(obj); 540 } 541 542 if (msm_obj->resv == &msm_obj->_resv) 543 reservation_object_fini(msm_obj->resv); 544 545 drm_gem_object_release(obj); 546 547 kfree(msm_obj); 548 } 549 550 /* convenience method to construct a GEM buffer object, and userspace handle */ 551 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file, 552 uint32_t size, uint32_t flags, uint32_t *handle) 553 { 554 struct drm_gem_object *obj; 555 int ret; 556 557 ret = mutex_lock_interruptible(&dev->struct_mutex); 558 if (ret) 559 return ret; 560 561 obj = msm_gem_new(dev, size, flags); 562 563 mutex_unlock(&dev->struct_mutex); 564 565 if (IS_ERR(obj)) 566 return PTR_ERR(obj); 567 568 ret = drm_gem_handle_create(file, obj, handle); 569 570 /* drop reference from allocate - handle holds it now */ 571 drm_gem_object_unreference_unlocked(obj); 572 573 return ret; 574 } 575 576 static int msm_gem_new_impl(struct drm_device *dev, 577 uint32_t size, uint32_t flags, 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; 583 584 switch (flags & MSM_BO_CACHE_MASK) { 585 case MSM_BO_UNCACHED: 586 case MSM_BO_CACHED: 587 case MSM_BO_WC: 588 break; 589 default: 590 dev_err(dev->dev, "invalid cache flag: %x\n", 591 (flags & MSM_BO_CACHE_MASK)); 592 return -EINVAL; 593 } 594 595 sz = sizeof(*msm_obj); 596 if (!iommu_present(&platform_bus_type)) 597 sz += sizeof(struct drm_mm_node); 598 599 msm_obj = kzalloc(sz, GFP_KERNEL); 600 if (!msm_obj) 601 return -ENOMEM; 602 603 if (!iommu_present(&platform_bus_type)) 604 msm_obj->vram_node = (void *)&msm_obj[1]; 605 606 msm_obj->flags = flags; 607 608 msm_obj->resv = &msm_obj->_resv; 609 reservation_object_init(msm_obj->resv); 610 611 INIT_LIST_HEAD(&msm_obj->submit_entry); 612 list_add_tail(&msm_obj->mm_list, &priv->inactive_list); 613 614 *obj = &msm_obj->base; 615 616 return 0; 617 } 618 619 struct drm_gem_object *msm_gem_new(struct drm_device *dev, 620 uint32_t size, uint32_t flags) 621 { 622 struct drm_gem_object *obj = NULL; 623 int ret; 624 625 WARN_ON(!mutex_is_locked(&dev->struct_mutex)); 626 627 size = PAGE_ALIGN(size); 628 629 ret = msm_gem_new_impl(dev, size, flags, &obj); 630 if (ret) 631 goto fail; 632 633 if (iommu_present(&platform_bus_type)) { 634 ret = drm_gem_object_init(dev, obj, size); 635 if (ret) 636 goto fail; 637 } else { 638 drm_gem_private_object_init(dev, obj, size); 639 } 640 641 return obj; 642 643 fail: 644 if (obj) 645 drm_gem_object_unreference(obj); 646 647 return ERR_PTR(ret); 648 } 649 650 struct drm_gem_object *msm_gem_import(struct drm_device *dev, 651 uint32_t size, struct sg_table *sgt) 652 { 653 struct msm_gem_object *msm_obj; 654 struct drm_gem_object *obj; 655 int ret, npages; 656 657 /* if we don't have IOMMU, don't bother pretending we can import: */ 658 if (!iommu_present(&platform_bus_type)) { 659 dev_err(dev->dev, "cannot import without IOMMU\n"); 660 return ERR_PTR(-EINVAL); 661 } 662 663 size = PAGE_ALIGN(size); 664 665 ret = msm_gem_new_impl(dev, size, MSM_BO_WC, &obj); 666 if (ret) 667 goto fail; 668 669 drm_gem_private_object_init(dev, obj, size); 670 671 npages = size / PAGE_SIZE; 672 673 msm_obj = to_msm_bo(obj); 674 msm_obj->sgt = sgt; 675 msm_obj->pages = drm_malloc_ab(npages, sizeof(struct page *)); 676 if (!msm_obj->pages) { 677 ret = -ENOMEM; 678 goto fail; 679 } 680 681 ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages); 682 if (ret) 683 goto fail; 684 685 return obj; 686 687 fail: 688 if (obj) 689 drm_gem_object_unreference_unlocked(obj); 690 691 return ERR_PTR(ret); 692 } 693