drivers/gpu/drm/ttm/ttm_bo_vm.c

   1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
   2 /**************************************************************************
   3  *
   4  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  22  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  23  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  24  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  25  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28 /*
  29  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
  30  */
  31
  32 #define pr_fmt(fmt) "[TTM] " fmt
  33
  34 #include <drm/ttm/ttm_bo_driver.h>
  35 #include <drm/ttm/ttm_placement.h>
  36 #include <drm/drm_vma_manager.h>
  37 #include <drm/drm_drv.h>
  38 #include <drm/drm_managed.h>
  39 #include <linux/mm.h>
  40 #include <linux/pfn_t.h>
  41 #include <linux/rbtree.h>
  42 #include <linux/module.h>
  43 #include <linux/uaccess.h>
  44 #include <linux/mem_encrypt.h>
  45
  46 static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
  47                                 struct vm_fault *vmf)
  48 {
  49         vm_fault_t ret = 0;
  50         int err = 0;
  51
  52         if (likely(!bo->moving))
  53                 goto out_unlock;
  54
  55         /*
  56          * Quick non-stalling check for idle.
  57          */
  58         if (dma_fence_is_signaled(bo->moving))
  59                 goto out_clear;
  60
  61         /*
  62          * If possible, avoid waiting for GPU with mmap_lock
  63          * held.  We only do this if the fault allows retry and this
  64          * is the first attempt.
  65          */
  66         if (fault_flag_allow_retry_first(vmf->flags)) {
  67                 ret = VM_FAULT_RETRY;
  68                 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
  69                         goto out_unlock;
  70
  71                 ttm_bo_get(bo);
  72                 mmap_read_unlock(vmf->vma->vm_mm);
  73                 (void) dma_fence_wait(bo->moving, true);
  74                 dma_resv_unlock(bo->base.resv);
  75                 ttm_bo_put(bo);
  76                 goto out_unlock;
  77         }
  78
  79         /*
  80          * Ordinary wait.
  81          */
  82         err = dma_fence_wait(bo->moving, true);
  83         if (unlikely(err != 0)) {
  84                 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
  85                         VM_FAULT_NOPAGE;
  86                 goto out_unlock;
  87         }
  88
  89 out_clear:
  90         dma_fence_put(bo->moving);
  91         bo->moving = NULL;
  92
  93 out_unlock:
  94         return ret;
  95 }
  96
  97 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
  98                                        unsigned long page_offset)
  99 {
 100         struct ttm_device *bdev = bo->bdev;
 101
 102         if (bdev->funcs->io_mem_pfn)
 103                 return bdev->funcs->io_mem_pfn(bo, page_offset);
 104
 105         return (bo->resource->bus.offset >> PAGE_SHIFT) + page_offset;
 106 }
 107
 108 /**
 109  * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
 110  * @bo: The buffer object
 111  * @vmf: The fault structure handed to the callback
 112  *
 113  * vm callbacks like fault() and *_mkwrite() allow for the mm_sem to be dropped
 114  * during long waits, and after the wait the callback will be restarted. This
 115  * is to allow other threads using the same virtual memory space concurrent
 116  * access to map(), unmap() completely unrelated buffer objects. TTM buffer
 117  * object reservations sometimes wait for GPU and should therefore be
 118  * considered long waits. This function reserves the buffer object interruptibly
 119  * taking this into account. Starvation is avoided by the vm system not
 120  * allowing too many repeated restarts.
 121  * This function is intended to be used in customized fault() and _mkwrite()
 122  * handlers.
 123  *
 124  * Return:
 125  *    0 on success and the bo was reserved.
 126  *    VM_FAULT_RETRY if blocking wait.
 127  *    VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
 128  */
 129 vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
 130                              struct vm_fault *vmf)
 131 {
 132         /*
 133          * Work around locking order reversal in fault / nopfn
 134          * between mmap_lock and bo_reserve: Perform a trylock operation
 135          * for reserve, and if it fails, retry the fault after waiting
 136          * for the buffer to become unreserved.
 137          */
 138         if (unlikely(!dma_resv_trylock(bo->base.resv))) {
 139                 /*
 140                  * If the fault allows retry and this is the first
 141                  * fault attempt, we try to release the mmap_lock
 142                  * before waiting
 143                  */
 144                 if (fault_flag_allow_retry_first(vmf->flags)) {
 145                         if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 146                                 ttm_bo_get(bo);
 147                                 mmap_read_unlock(vmf->vma->vm_mm);
 148                                 if (!dma_resv_lock_interruptible(bo->base.resv,
 149                                                                  NULL))
 150                                         dma_resv_unlock(bo->base.resv);
 151                                 ttm_bo_put(bo);
 152                         }
 153
 154                         return VM_FAULT_RETRY;
 155                 }
 156
 157                 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
 158                         return VM_FAULT_NOPAGE;
 159         }
 160
 161         /*
 162          * Refuse to fault imported pages. This should be handled
 163          * (if at all) by redirecting mmap to the exporter.
 164          */
 165         if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
 166                 dma_resv_unlock(bo->base.resv);
 167                 return VM_FAULT_SIGBUS;
 168         }
 169
 170         return 0;
 171 }
 172 EXPORT_SYMBOL(ttm_bo_vm_reserve);
 173
 174 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 175 /**
 176  * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults
 177  * @vmf: Fault data
 178  * @bo: The buffer object
 179  * @page_offset: Page offset from bo start
 180  * @fault_page_size: The size of the fault in pages.
 181  * @pgprot: The page protections.
 182  * Does additional checking whether it's possible to insert a PUD or PMD
 183  * pfn and performs the insertion.
 184  *
 185  * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if
 186  * a huge fault was not possible, or on insertion error.
 187  */
 188 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
 189                                         struct ttm_buffer_object *bo,
 190                                         pgoff_t page_offset,
 191                                         pgoff_t fault_page_size,
 192                                         pgprot_t pgprot)
 193 {
 194         pgoff_t i;
 195         vm_fault_t ret;
 196         unsigned long pfn;
 197         pfn_t pfnt;
 198         struct ttm_tt *ttm = bo->ttm;
 199         bool write = vmf->flags & FAULT_FLAG_WRITE;
 200
 201         /* Fault should not cross bo boundary. */
 202         page_offset &= ~(fault_page_size - 1);
 203         if (page_offset + fault_page_size > bo->resource->num_pages)
 204                 goto out_fallback;
 205
 206         if (bo->resource->bus.is_iomem)
 207                 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
 208         else
 209                 pfn = page_to_pfn(ttm->pages[page_offset]);
 210
 211         /* pfn must be fault_page_size aligned. */
 212         if ((pfn & (fault_page_size - 1)) != 0)
 213                 goto out_fallback;
 214
 215         /* Check that memory is contiguous. */
 216         if (!bo->resource->bus.is_iomem) {
 217                 for (i = 1; i < fault_page_size; ++i) {
 218                         if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i)
 219                                 goto out_fallback;
 220                 }
 221         } else if (bo->bdev->funcs->io_mem_pfn) {
 222                 for (i = 1; i < fault_page_size; ++i) {
 223                         if (ttm_bo_io_mem_pfn(bo, page_offset + i) != pfn + i)
 224                                 goto out_fallback;
 225                 }
 226         }
 227
 228         pfnt = __pfn_to_pfn_t(pfn, PFN_DEV);
 229         if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT))
 230                 ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write);
 231 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 232         else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT))
 233                 ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write);
 234 #endif
 235         else
 236                 WARN_ON_ONCE(ret = VM_FAULT_FALLBACK);
 237
 238         if (ret != VM_FAULT_NOPAGE)
 239                 goto out_fallback;
 240
 241         return VM_FAULT_NOPAGE;
 242 out_fallback:
 243         count_vm_event(THP_FAULT_FALLBACK);
 244         return VM_FAULT_FALLBACK;
 245 }
 246 #else
 247 static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
 248                                         struct ttm_buffer_object *bo,
 249                                         pgoff_t page_offset,
 250                                         pgoff_t fault_page_size,
 251                                         pgprot_t pgprot)
 252 {
 253         return VM_FAULT_FALLBACK;
 254 }
 255 #endif
 256
 257 /**
 258  * ttm_bo_vm_fault_reserved - TTM fault helper
 259  * @vmf: The struct vm_fault given as argument to the fault callback
 260  * @prot: The page protection to be used for this memory area.
 261  * @num_prefault: Maximum number of prefault pages. The caller may want to
 262  * specify this based on madvice settings and the size of the GPU object
 263  * backed by the memory.
 264  * @fault_page_size: The size of the fault in pages.
 265  *
 266  * This function inserts one or more page table entries pointing to the
 267  * memory backing the buffer object, and then returns a return code
 268  * instructing the caller to retry the page access.
 269  *
 270  * Return:
 271  *   VM_FAULT_NOPAGE on success or pending signal
 272  *   VM_FAULT_SIGBUS on unspecified error
 273  *   VM_FAULT_OOM on out-of-memory
 274  *   VM_FAULT_RETRY if retryable wait
 275  */
 276 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
 277                                     pgprot_t prot,
 278                                     pgoff_t num_prefault,
 279                                     pgoff_t fault_page_size)
 280 {
 281         struct vm_area_struct *vma = vmf->vma;
 282         struct ttm_buffer_object *bo = vma->vm_private_data;
 283         struct ttm_device *bdev = bo->bdev;
 284         unsigned long page_offset;
 285         unsigned long page_last;
 286         unsigned long pfn;
 287         struct ttm_tt *ttm = NULL;
 288         struct page *page;
 289         int err;
 290         pgoff_t i;
 291         vm_fault_t ret = VM_FAULT_NOPAGE;
 292         unsigned long address = vmf->address;
 293
 294         /*
 295          * Wait for buffer data in transit, due to a pipelined
 296          * move.
 297          */
 298         ret = ttm_bo_vm_fault_idle(bo, vmf);
 299         if (unlikely(ret != 0))
 300                 return ret;
 301
 302         err = ttm_mem_io_reserve(bdev, bo->resource);
 303         if (unlikely(err != 0))
 304                 return VM_FAULT_SIGBUS;
 305
 306         page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
 307                 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
 308         page_last = vma_pages(vma) + vma->vm_pgoff -
 309                 drm_vma_node_start(&bo->base.vma_node);
 310
 311         if (unlikely(page_offset >= bo->resource->num_pages))
 312                 return VM_FAULT_SIGBUS;
 313
 314         prot = ttm_io_prot(bo, bo->resource, prot);
 315         if (!bo->resource->bus.is_iomem) {
 316                 struct ttm_operation_ctx ctx = {
 317                         .interruptible = false,
 318                         .no_wait_gpu = false,
 319                         .force_alloc = true
 320                 };
 321
 322                 ttm = bo->ttm;
 323                 if (ttm_tt_populate(bdev, bo->ttm, &ctx))
 324                         return VM_FAULT_OOM;
 325         } else {
 326                 /* Iomem should not be marked encrypted */
 327                 prot = pgprot_decrypted(prot);
 328         }
 329
 330         /* We don't prefault on huge faults. Yet. */
 331         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1)
 332                 return ttm_bo_vm_insert_huge(vmf, bo, page_offset,
 333                                              fault_page_size, prot);
 334
 335         /*
 336          * Speculatively prefault a number of pages. Only error on
 337          * first page.
 338          */
 339         for (i = 0; i < num_prefault; ++i) {
 340                 if (bo->resource->bus.is_iomem) {
 341                         pfn = ttm_bo_io_mem_pfn(bo, page_offset);
 342                 } else {
 343                         page = ttm->pages[page_offset];
 344                         if (unlikely(!page && i == 0)) {
 345                                 return VM_FAULT_OOM;
 346                         } else if (unlikely(!page)) {
 347                                 break;
 348                         }
 349                         page->index = drm_vma_node_start(&bo->base.vma_node) +
 350                                 page_offset;
 351                         pfn = page_to_pfn(page);
 352                 }
 353
 354                 /*
 355                  * Note that the value of @prot at this point may differ from
 356                  * the value of @vma->vm_page_prot in the caching- and
 357                  * encryption bits. This is because the exact location of the
 358                  * data may not be known at mmap() time and may also change
 359                  * at arbitrary times while the data is mmap'ed.
 360                  * See vmf_insert_mixed_prot() for a discussion.
 361                  */
 362                 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
 363
 364                 /* Never error on prefaulted PTEs */
 365                 if (unlikely((ret & VM_FAULT_ERROR))) {
 366                         if (i == 0)
 367                                 return VM_FAULT_NOPAGE;
 368                         else
 369                                 break;
 370                 }
 371
 372                 address += PAGE_SIZE;
 373                 if (unlikely(++page_offset >= page_last))
 374                         break;
 375         }
 376         return ret;
 377 }
 378 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
 379
 380 static void ttm_bo_release_dummy_page(struct drm_device *dev, void *res)
 381 {
 382         struct page *dummy_page = (struct page *)res;
 383
 384         __free_page(dummy_page);
 385 }
 386
 387 vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot)
 388 {
 389         struct vm_area_struct *vma = vmf->vma;
 390         struct ttm_buffer_object *bo = vma->vm_private_data;
 391         struct drm_device *ddev = bo->base.dev;
 392         vm_fault_t ret = VM_FAULT_NOPAGE;
 393         unsigned long address;
 394         unsigned long pfn;
 395         struct page *page;
 396
 397         /* Allocate new dummy page to map all the VA range in this VMA to it*/
 398         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 399         if (!page)
 400                 return VM_FAULT_OOM;
 401
 402         /* Set the page to be freed using drmm release action */
 403         if (drmm_add_action_or_reset(ddev, ttm_bo_release_dummy_page, page))
 404                 return VM_FAULT_OOM;
 405
 406         pfn = page_to_pfn(page);
 407
 408         /* Prefault the entire VMA range right away to avoid further faults */
 409         for (address = vma->vm_start; address < vma->vm_end;
 410              address += PAGE_SIZE)
 411                 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
 412
 413         return ret;
 414 }
 415 EXPORT_SYMBOL(ttm_bo_vm_dummy_page);
 416
 417 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
 418 {
 419         struct vm_area_struct *vma = vmf->vma;
 420         pgprot_t prot;
 421         struct ttm_buffer_object *bo = vma->vm_private_data;
 422         struct drm_device *ddev = bo->base.dev;
 423         vm_fault_t ret;
 424         int idx;
 425
 426         ret = ttm_bo_vm_reserve(bo, vmf);
 427         if (ret)
 428                 return ret;
 429
 430         prot = vma->vm_page_prot;
 431         if (drm_dev_enter(ddev, &idx)) {
 432                 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1);
 433                 drm_dev_exit(idx);
 434         } else {
 435                 ret = ttm_bo_vm_dummy_page(vmf, prot);
 436         }
 437         if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
 438                 return ret;
 439
 440         dma_resv_unlock(bo->base.resv);
 441
 442         return ret;
 443 }
 444 EXPORT_SYMBOL(ttm_bo_vm_fault);
 445
 446 void ttm_bo_vm_open(struct vm_area_struct *vma)
 447 {
 448         struct ttm_buffer_object *bo = vma->vm_private_data;
 449
 450         WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
 451
 452         ttm_bo_get(bo);
 453 }
 454 EXPORT_SYMBOL(ttm_bo_vm_open);
 455
 456 void ttm_bo_vm_close(struct vm_area_struct *vma)
 457 {
 458         struct ttm_buffer_object *bo = vma->vm_private_data;
 459
 460         ttm_bo_put(bo);
 461         vma->vm_private_data = NULL;
 462 }
 463 EXPORT_SYMBOL(ttm_bo_vm_close);
 464
 465 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
 466                                  unsigned long offset,
 467                                  uint8_t *buf, int len, int write)
 468 {
 469         unsigned long page = offset >> PAGE_SHIFT;
 470         unsigned long bytes_left = len;
 471         int ret;
 472
 473         /* Copy a page at a time, that way no extra virtual address
 474          * mapping is needed
 475          */
 476         offset -= page << PAGE_SHIFT;
 477         do {
 478                 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
 479                 struct ttm_bo_kmap_obj map;
 480                 void *ptr;
 481                 bool is_iomem;
 482
 483                 ret = ttm_bo_kmap(bo, page, 1, &map);
 484                 if (ret)
 485                         return ret;
 486
 487                 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
 488                 WARN_ON_ONCE(is_iomem);
 489                 if (write)
 490                         memcpy(ptr, buf, bytes);
 491                 else
 492                         memcpy(buf, ptr, bytes);
 493                 ttm_bo_kunmap(&map);
 494
 495                 page++;
 496                 buf += bytes;
 497                 bytes_left -= bytes;
 498                 offset = 0;
 499         } while (bytes_left);
 500
 501         return len;
 502 }
 503
 504 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
 505                      void *buf, int len, int write)
 506 {
 507         struct ttm_buffer_object *bo = vma->vm_private_data;
 508         unsigned long offset = (addr) - vma->vm_start +
 509                 ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
 510                  << PAGE_SHIFT);
 511         int ret;
 512
 513         if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->resource->num_pages)
 514                 return -EIO;
 515
 516         ret = ttm_bo_reserve(bo, true, false, NULL);
 517         if (ret)
 518                 return ret;
 519
 520         switch (bo->resource->mem_type) {
 521         case TTM_PL_SYSTEM:
 522                 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 523                         ret = ttm_tt_swapin(bo->ttm);
 524                         if (unlikely(ret != 0))
 525                                 return ret;
 526                 }
 527                 fallthrough;
 528         case TTM_PL_TT:
 529                 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
 530                 break;
 531         default:
 532                 if (bo->bdev->funcs->access_memory)
 533                         ret = bo->bdev->funcs->access_memory(
 534                                 bo, offset, buf, len, write);
 535                 else
 536                         ret = -EIO;
 537         }
 538
 539         ttm_bo_unreserve(bo);
 540
 541         return ret;
 542 }
 543 EXPORT_SYMBOL(ttm_bo_vm_access);
 544
 545 static const struct vm_operations_struct ttm_bo_vm_ops = {
 546         .fault = ttm_bo_vm_fault,
 547         .open = ttm_bo_vm_open,
 548         .close = ttm_bo_vm_close,
 549         .access = ttm_bo_vm_access,
 550 };
 551
 552 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 553 {
 554         /* Enforce no COW since would have really strange behavior with it. */
 555         if (is_cow_mapping(vma->vm_flags))
 556                 return -EINVAL;
 557
 558         ttm_bo_get(bo);
 559
 560         /*
 561          * Drivers may want to override the vm_ops field. Otherwise we
 562          * use TTM's default callbacks.
 563          */
 564         if (!vma->vm_ops)
 565                 vma->vm_ops = &ttm_bo_vm_ops;
 566
 567         /*
 568          * Note: We're transferring the bo reference to
 569          * vma->vm_private_data here.
 570          */
 571
 572         vma->vm_private_data = bo;
 573
 574         vma->vm_flags |= VM_PFNMAP;
 575         vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
 576         return 0;
 577 }
 578 EXPORT_SYMBOL(ttm_bo_mmap_obj);