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_module.h>
  35 #include <drm/ttm/ttm_bo_driver.h>
  36 #include <drm/ttm/ttm_placement.h>
  37 #include <drm/drm_vma_manager.h>
  38 #include <linux/mm.h>
  39 #include <linux/pfn_t.h>
  40 #include <linux/rbtree.h>
  41 #include <linux/module.h>
  42 #include <linux/uaccess.h>
  43 #include <linux/mem_encrypt.h>
  44
  45 static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
  46                                 struct vm_fault *vmf)
  47 {
  48         vm_fault_t ret = 0;
  49         int err = 0;
  50
  51         if (likely(!bo->moving))
  52                 goto out_unlock;
  53
  54         /*
  55          * Quick non-stalling check for idle.
  56          */
  57         if (dma_fence_is_signaled(bo->moving))
  58                 goto out_clear;
  59
  60         /*
  61          * If possible, avoid waiting for GPU with mmap_sem
  62          * held.  We only do this if the fault allows retry and this
  63          * is the first attempt.
  64          */
  65         if (fault_flag_allow_retry_first(vmf->flags)) {
  66                 ret = VM_FAULT_RETRY;
  67                 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
  68                         goto out_unlock;
  69
  70                 ttm_bo_get(bo);
  71                 up_read(&vmf->vma->vm_mm->mmap_sem);
  72                 (void) dma_fence_wait(bo->moving, true);
  73                 dma_resv_unlock(bo->base.resv);
  74                 ttm_bo_put(bo);
  75                 goto out_unlock;
  76         }
  77
  78         /*
  79          * Ordinary wait.
  80          */
  81         err = dma_fence_wait(bo->moving, true);
  82         if (unlikely(err != 0)) {
  83                 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
  84                         VM_FAULT_NOPAGE;
  85                 goto out_unlock;
  86         }
  87
  88 out_clear:
  89         dma_fence_put(bo->moving);
  90         bo->moving = NULL;
  91
  92 out_unlock:
  93         return ret;
  94 }
  95
  96 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
  97                                        unsigned long page_offset)
  98 {
  99         struct ttm_bo_device *bdev = bo->bdev;
 100
 101         if (bdev->driver->io_mem_pfn)
 102                 return bdev->driver->io_mem_pfn(bo, page_offset);
 103
 104         return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
 105                 + 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_sem 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_sem
 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                                 up_read(&vmf->vma->vm_mm->mmap_sem);
 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         return 0;
 162 }
 163 EXPORT_SYMBOL(ttm_bo_vm_reserve);
 164
 165 /**
 166  * ttm_bo_vm_fault_reserved - TTM fault helper
 167  * @vmf: The struct vm_fault given as argument to the fault callback
 168  * @prot: The page protection to be used for this memory area.
 169  * @num_prefault: Maximum number of prefault pages. The caller may want to
 170  * specify this based on madvice settings and the size of the GPU object
 171  * backed by the memory.
 172  *
 173  * This function inserts one or more page table entries pointing to the
 174  * memory backing the buffer object, and then returns a return code
 175  * instructing the caller to retry the page access.
 176  *
 177  * Return:
 178  *   VM_FAULT_NOPAGE on success or pending signal
 179  *   VM_FAULT_SIGBUS on unspecified error
 180  *   VM_FAULT_OOM on out-of-memory
 181  *   VM_FAULT_RETRY if retryable wait
 182  */
 183 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
 184                                     pgprot_t prot,
 185                                     pgoff_t num_prefault)
 186 {
 187         struct vm_area_struct *vma = vmf->vma;
 188         struct ttm_buffer_object *bo = vma->vm_private_data;
 189         struct ttm_bo_device *bdev = bo->bdev;
 190         unsigned long page_offset;
 191         unsigned long page_last;
 192         unsigned long pfn;
 193         struct ttm_tt *ttm = NULL;
 194         struct page *page;
 195         int err;
 196         pgoff_t i;
 197         vm_fault_t ret = VM_FAULT_NOPAGE;
 198         unsigned long address = vmf->address;
 199         struct ttm_mem_type_manager *man =
 200                 &bdev->man[bo->mem.mem_type];
 201
 202         /*
 203          * Refuse to fault imported pages. This should be handled
 204          * (if at all) by redirecting mmap to the exporter.
 205          */
 206         if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG))
 207                 return VM_FAULT_SIGBUS;
 208
 209         if (bdev->driver->fault_reserve_notify) {
 210                 struct dma_fence *moving = dma_fence_get(bo->moving);
 211
 212                 err = bdev->driver->fault_reserve_notify(bo);
 213                 switch (err) {
 214                 case 0:
 215                         break;
 216                 case -EBUSY:
 217                 case -ERESTARTSYS:
 218                         return VM_FAULT_NOPAGE;
 219                 default:
 220                         return VM_FAULT_SIGBUS;
 221                 }
 222
 223                 if (bo->moving != moving) {
 224                         spin_lock(&ttm_bo_glob.lru_lock);
 225                         ttm_bo_move_to_lru_tail(bo, NULL);
 226                         spin_unlock(&ttm_bo_glob.lru_lock);
 227                 }
 228                 dma_fence_put(moving);
 229         }
 230
 231         /*
 232          * Wait for buffer data in transit, due to a pipelined
 233          * move.
 234          */
 235         ret = ttm_bo_vm_fault_idle(bo, vmf);
 236         if (unlikely(ret != 0))
 237                 return ret;
 238
 239         err = ttm_mem_io_lock(man, true);
 240         if (unlikely(err != 0))
 241                 return VM_FAULT_NOPAGE;
 242         err = ttm_mem_io_reserve_vm(bo);
 243         if (unlikely(err != 0)) {
 244                 ret = VM_FAULT_SIGBUS;
 245                 goto out_io_unlock;
 246         }
 247
 248         page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
 249                 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
 250         page_last = vma_pages(vma) + vma->vm_pgoff -
 251                 drm_vma_node_start(&bo->base.vma_node);
 252
 253         if (unlikely(page_offset >= bo->num_pages)) {
 254                 ret = VM_FAULT_SIGBUS;
 255                 goto out_io_unlock;
 256         }
 257
 258         prot = ttm_io_prot(bo->mem.placement, prot);
 259         if (!bo->mem.bus.is_iomem) {
 260                 struct ttm_operation_ctx ctx = {
 261                         .interruptible = false,
 262                         .no_wait_gpu = false,
 263                         .flags = TTM_OPT_FLAG_FORCE_ALLOC
 264
 265                 };
 266
 267                 ttm = bo->ttm;
 268                 if (ttm_tt_populate(bo->ttm, &ctx)) {
 269                         ret = VM_FAULT_OOM;
 270                         goto out_io_unlock;
 271                 }
 272         } else {
 273                 /* Iomem should not be marked encrypted */
 274                 prot = pgprot_decrypted(prot);
 275         }
 276
 277         /*
 278          * Speculatively prefault a number of pages. Only error on
 279          * first page.
 280          */
 281         for (i = 0; i < num_prefault; ++i) {
 282                 if (bo->mem.bus.is_iomem) {
 283                         pfn = ttm_bo_io_mem_pfn(bo, page_offset);
 284                 } else {
 285                         page = ttm->pages[page_offset];
 286                         if (unlikely(!page && i == 0)) {
 287                                 ret = VM_FAULT_OOM;
 288                                 goto out_io_unlock;
 289                         } else if (unlikely(!page)) {
 290                                 break;
 291                         }
 292                         page->index = drm_vma_node_start(&bo->base.vma_node) +
 293                                 page_offset;
 294                         pfn = page_to_pfn(page);
 295                 }
 296
 297                 /*
 298                  * Note that the value of @prot at this point may differ from
 299                  * the value of @vma->vm_page_prot in the caching- and
 300                  * encryption bits. This is because the exact location of the
 301                  * data may not be known at mmap() time and may also change
 302                  * at arbitrary times while the data is mmap'ed.
 303                  * See vmf_insert_mixed_prot() for a discussion.
 304                  */
 305                 if (vma->vm_flags & VM_MIXEDMAP)
 306                         ret = vmf_insert_mixed_prot(vma, address,
 307                                                     __pfn_to_pfn_t(pfn, PFN_DEV),
 308                                                     prot);
 309                 else
 310                         ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
 311
 312                 /* Never error on prefaulted PTEs */
 313                 if (unlikely((ret & VM_FAULT_ERROR))) {
 314                         if (i == 0)
 315                                 goto out_io_unlock;
 316                         else
 317                                 break;
 318                 }
 319
 320                 address += PAGE_SIZE;
 321                 if (unlikely(++page_offset >= page_last))
 322                         break;
 323         }
 324         ret = VM_FAULT_NOPAGE;
 325 out_io_unlock:
 326         ttm_mem_io_unlock(man);
 327         return ret;
 328 }
 329 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
 330
 331 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
 332 {
 333         struct vm_area_struct *vma = vmf->vma;
 334         pgprot_t prot;
 335         struct ttm_buffer_object *bo = vma->vm_private_data;
 336         vm_fault_t ret;
 337
 338         ret = ttm_bo_vm_reserve(bo, vmf);
 339         if (ret)
 340                 return ret;
 341
 342         prot = vma->vm_page_prot;
 343         ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
 344         if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
 345                 return ret;
 346
 347         dma_resv_unlock(bo->base.resv);
 348
 349         return ret;
 350 }
 351 EXPORT_SYMBOL(ttm_bo_vm_fault);
 352
 353 void ttm_bo_vm_open(struct vm_area_struct *vma)
 354 {
 355         struct ttm_buffer_object *bo = vma->vm_private_data;
 356
 357         WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
 358
 359         ttm_bo_get(bo);
 360 }
 361 EXPORT_SYMBOL(ttm_bo_vm_open);
 362
 363 void ttm_bo_vm_close(struct vm_area_struct *vma)
 364 {
 365         struct ttm_buffer_object *bo = vma->vm_private_data;
 366
 367         ttm_bo_put(bo);
 368         vma->vm_private_data = NULL;
 369 }
 370 EXPORT_SYMBOL(ttm_bo_vm_close);
 371
 372 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
 373                                  unsigned long offset,
 374                                  uint8_t *buf, int len, int write)
 375 {
 376         unsigned long page = offset >> PAGE_SHIFT;
 377         unsigned long bytes_left = len;
 378         int ret;
 379
 380         /* Copy a page at a time, that way no extra virtual address
 381          * mapping is needed
 382          */
 383         offset -= page << PAGE_SHIFT;
 384         do {
 385                 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
 386                 struct ttm_bo_kmap_obj map;
 387                 void *ptr;
 388                 bool is_iomem;
 389
 390                 ret = ttm_bo_kmap(bo, page, 1, &map);
 391                 if (ret)
 392                         return ret;
 393
 394                 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
 395                 WARN_ON_ONCE(is_iomem);
 396                 if (write)
 397                         memcpy(ptr, buf, bytes);
 398                 else
 399                         memcpy(buf, ptr, bytes);
 400                 ttm_bo_kunmap(&map);
 401
 402                 page++;
 403                 buf += bytes;
 404                 bytes_left -= bytes;
 405                 offset = 0;
 406         } while (bytes_left);
 407
 408         return len;
 409 }
 410
 411 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
 412                      void *buf, int len, int write)
 413 {
 414         unsigned long offset = (addr) - vma->vm_start;
 415         struct ttm_buffer_object *bo = vma->vm_private_data;
 416         int ret;
 417
 418         if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
 419                 return -EIO;
 420
 421         ret = ttm_bo_reserve(bo, true, false, NULL);
 422         if (ret)
 423                 return ret;
 424
 425         switch (bo->mem.mem_type) {
 426         case TTM_PL_SYSTEM:
 427                 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 428                         ret = ttm_tt_swapin(bo->ttm);
 429                         if (unlikely(ret != 0))
 430                                 return ret;
 431                 }
 432                 /* fall through */
 433         case TTM_PL_TT:
 434                 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
 435                 break;
 436         default:
 437                 if (bo->bdev->driver->access_memory)
 438                         ret = bo->bdev->driver->access_memory(
 439                                 bo, offset, buf, len, write);
 440                 else
 441                         ret = -EIO;
 442         }
 443
 444         ttm_bo_unreserve(bo);
 445
 446         return ret;
 447 }
 448 EXPORT_SYMBOL(ttm_bo_vm_access);
 449
 450 static const struct vm_operations_struct ttm_bo_vm_ops = {
 451         .fault = ttm_bo_vm_fault,
 452         .open = ttm_bo_vm_open,
 453         .close = ttm_bo_vm_close,
 454         .access = ttm_bo_vm_access
 455 };
 456
 457 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
 458                                                   unsigned long offset,
 459                                                   unsigned long pages)
 460 {
 461         struct drm_vma_offset_node *node;
 462         struct ttm_buffer_object *bo = NULL;
 463
 464         drm_vma_offset_lock_lookup(bdev->vma_manager);
 465
 466         node = drm_vma_offset_lookup_locked(bdev->vma_manager, offset, pages);
 467         if (likely(node)) {
 468                 bo = container_of(node, struct ttm_buffer_object,
 469                                   base.vma_node);
 470                 bo = ttm_bo_get_unless_zero(bo);
 471         }
 472
 473         drm_vma_offset_unlock_lookup(bdev->vma_manager);
 474
 475         if (!bo)
 476                 pr_err("Could not find buffer object to map\n");
 477
 478         return bo;
 479 }
 480
 481 static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_struct *vma)
 482 {
 483         vma->vm_ops = &ttm_bo_vm_ops;
 484
 485         /*
 486          * Note: We're transferring the bo reference to
 487          * vma->vm_private_data here.
 488          */
 489
 490         vma->vm_private_data = bo;
 491
 492         /*
 493          * We'd like to use VM_PFNMAP on shared mappings, where
 494          * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
 495          * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
 496          * bad for performance. Until that has been sorted out, use
 497          * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
 498          */
 499         vma->vm_flags |= VM_MIXEDMAP;
 500         vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
 501 }
 502
 503 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
 504                 struct ttm_bo_device *bdev)
 505 {
 506         struct ttm_bo_driver *driver;
 507         struct ttm_buffer_object *bo;
 508         int ret;
 509
 510         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
 511                 return -EINVAL;
 512
 513         bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
 514         if (unlikely(!bo))
 515                 return -EINVAL;
 516
 517         driver = bo->bdev->driver;
 518         if (unlikely(!driver->verify_access)) {
 519                 ret = -EPERM;
 520                 goto out_unref;
 521         }
 522         ret = driver->verify_access(bo, filp);
 523         if (unlikely(ret != 0))
 524                 goto out_unref;
 525
 526         ttm_bo_mmap_vma_setup(bo, vma);
 527         return 0;
 528 out_unref:
 529         ttm_bo_put(bo);
 530         return ret;
 531 }
 532 EXPORT_SYMBOL(ttm_bo_mmap);
 533
 534 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 535 {
 536         ttm_bo_get(bo);
 537         ttm_bo_mmap_vma_setup(bo, vma);
 538         return 0;
 539 }
 540 EXPORT_SYMBOL(ttm_bo_mmap_obj);