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.
  63          */
  64         if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
  65                 ret = VM_FAULT_RETRY;
  66                 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
  67                         goto out_unlock;
  68
  69                 ttm_bo_get(bo);
  70                 up_read(&vmf->vma->vm_mm->mmap_sem);
  71                 (void) dma_fence_wait(bo->moving, true);
  72                 dma_resv_unlock(bo->base.resv);
  73                 ttm_bo_put(bo);
  74                 goto out_unlock;
  75         }
  76
  77         /*
  78          * Ordinary wait.
  79          */
  80         err = dma_fence_wait(bo->moving, true);
  81         if (unlikely(err != 0)) {
  82                 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
  83                         VM_FAULT_NOPAGE;
  84                 goto out_unlock;
  85         }
  86
  87 out_clear:
  88         dma_fence_put(bo->moving);
  89         bo->moving = NULL;
  90
  91 out_unlock:
  92         return ret;
  93 }
  94
  95 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
  96                                        unsigned long page_offset)
  97 {
  98         struct ttm_bo_device *bdev = bo->bdev;
  99
 100         if (bdev->driver->io_mem_pfn)
 101                 return bdev->driver->io_mem_pfn(bo, page_offset);
 102
 103         return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
 104                 + page_offset;
 105 }
 106
 107 /**
 108  * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
 109  * @bo: The buffer object
 110  * @vmf: The fault structure handed to the callback
 111  *
 112  * vm callbacks like fault() and *_mkwrite() allow for the mm_sem to be dropped
 113  * during long waits, and after the wait the callback will be restarted. This
 114  * is to allow other threads using the same virtual memory space concurrent
 115  * access to map(), unmap() completely unrelated buffer objects. TTM buffer
 116  * object reservations sometimes wait for GPU and should therefore be
 117  * considered long waits. This function reserves the buffer object interruptibly
 118  * taking this into account. Starvation is avoided by the vm system not
 119  * allowing too many repeated restarts.
 120  * This function is intended to be used in customized fault() and _mkwrite()
 121  * handlers.
 122  *
 123  * Return:
 124  *    0 on success and the bo was reserved.
 125  *    VM_FAULT_RETRY if blocking wait.
 126  *    VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
 127  */
 128 vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
 129                              struct vm_fault *vmf)
 130 {
 131         /*
 132          * Work around locking order reversal in fault / nopfn
 133          * between mmap_sem and bo_reserve: Perform a trylock operation
 134          * for reserve, and if it fails, retry the fault after waiting
 135          * for the buffer to become unreserved.
 136          */
 137         if (unlikely(!dma_resv_trylock(bo->base.resv))) {
 138                 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
 139                         if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 140                                 ttm_bo_get(bo);
 141                                 up_read(&vmf->vma->vm_mm->mmap_sem);
 142                                 if (!dma_resv_lock_interruptible(bo->base.resv,
 143                                                                  NULL))
 144                                         dma_resv_unlock(bo->base.resv);
 145                                 ttm_bo_put(bo);
 146                         }
 147
 148                         return VM_FAULT_RETRY;
 149                 }
 150
 151                 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
 152                         return VM_FAULT_NOPAGE;
 153         }
 154
 155         return 0;
 156 }
 157 EXPORT_SYMBOL(ttm_bo_vm_reserve);
 158
 159 /**
 160  * ttm_bo_vm_fault_reserved - TTM fault helper
 161  * @vmf: The struct vm_fault given as argument to the fault callback
 162  * @prot: The page protection to be used for this memory area.
 163  * @num_prefault: Maximum number of prefault pages. The caller may want to
 164  * specify this based on madvice settings and the size of the GPU object
 165  * backed by the memory.
 166  *
 167  * This function inserts one or more page table entries pointing to the
 168  * memory backing the buffer object, and then returns a return code
 169  * instructing the caller to retry the page access.
 170  *
 171  * Return:
 172  *   VM_FAULT_NOPAGE on success or pending signal
 173  *   VM_FAULT_SIGBUS on unspecified error
 174  *   VM_FAULT_OOM on out-of-memory
 175  *   VM_FAULT_RETRY if retryable wait
 176  */
 177 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
 178                                     pgprot_t prot,
 179                                     pgoff_t num_prefault)
 180 {
 181         struct vm_area_struct *vma = vmf->vma;
 182         struct ttm_buffer_object *bo = vma->vm_private_data;
 183         struct ttm_bo_device *bdev = bo->bdev;
 184         unsigned long page_offset;
 185         unsigned long page_last;
 186         unsigned long pfn;
 187         struct ttm_tt *ttm = NULL;
 188         struct page *page;
 189         int err;
 190         pgoff_t i;
 191         vm_fault_t ret = VM_FAULT_NOPAGE;
 192         unsigned long address = vmf->address;
 193         struct ttm_mem_type_manager *man =
 194                 &bdev->man[bo->mem.mem_type];
 195
 196         /*
 197          * Refuse to fault imported pages. This should be handled
 198          * (if at all) by redirecting mmap to the exporter.
 199          */
 200         if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG))
 201                 return VM_FAULT_SIGBUS;
 202
 203         if (bdev->driver->fault_reserve_notify) {
 204                 struct dma_fence *moving = dma_fence_get(bo->moving);
 205
 206                 err = bdev->driver->fault_reserve_notify(bo);
 207                 switch (err) {
 208                 case 0:
 209                         break;
 210                 case -EBUSY:
 211                 case -ERESTARTSYS:
 212                         return VM_FAULT_NOPAGE;
 213                 default:
 214                         return VM_FAULT_SIGBUS;
 215                 }
 216
 217                 if (bo->moving != moving) {
 218                         spin_lock(&ttm_bo_glob.lru_lock);
 219                         ttm_bo_move_to_lru_tail(bo, NULL);
 220                         spin_unlock(&ttm_bo_glob.lru_lock);
 221                 }
 222                 dma_fence_put(moving);
 223         }
 224
 225         /*
 226          * Wait for buffer data in transit, due to a pipelined
 227          * move.
 228          */
 229         ret = ttm_bo_vm_fault_idle(bo, vmf);
 230         if (unlikely(ret != 0))
 231                 return ret;
 232
 233         err = ttm_mem_io_lock(man, true);
 234         if (unlikely(err != 0))
 235                 return VM_FAULT_NOPAGE;
 236         err = ttm_mem_io_reserve_vm(bo);
 237         if (unlikely(err != 0)) {
 238                 ret = VM_FAULT_SIGBUS;
 239                 goto out_io_unlock;
 240         }
 241
 242         page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
 243                 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
 244         page_last = vma_pages(vma) + vma->vm_pgoff -
 245                 drm_vma_node_start(&bo->base.vma_node);
 246
 247         if (unlikely(page_offset >= bo->num_pages)) {
 248                 ret = VM_FAULT_SIGBUS;
 249                 goto out_io_unlock;
 250         }
 251
 252         prot = ttm_io_prot(bo->mem.placement, prot);
 253         if (!bo->mem.bus.is_iomem) {
 254                 struct ttm_operation_ctx ctx = {
 255                         .interruptible = false,
 256                         .no_wait_gpu = false,
 257                         .flags = TTM_OPT_FLAG_FORCE_ALLOC
 258
 259                 };
 260
 261                 ttm = bo->ttm;
 262                 if (ttm_tt_populate(bo->ttm, &ctx)) {
 263                         ret = VM_FAULT_OOM;
 264                         goto out_io_unlock;
 265                 }
 266         } else {
 267                 /* Iomem should not be marked encrypted */
 268                 prot = pgprot_decrypted(prot);
 269         }
 270
 271         /*
 272          * Speculatively prefault a number of pages. Only error on
 273          * first page.
 274          */
 275         for (i = 0; i < num_prefault; ++i) {
 276                 if (bo->mem.bus.is_iomem) {
 277                         pfn = ttm_bo_io_mem_pfn(bo, page_offset);
 278                 } else {
 279                         page = ttm->pages[page_offset];
 280                         if (unlikely(!page && i == 0)) {
 281                                 ret = VM_FAULT_OOM;
 282                                 goto out_io_unlock;
 283                         } else if (unlikely(!page)) {
 284                                 break;
 285                         }
 286                         page->index = drm_vma_node_start(&bo->base.vma_node) +
 287                                 page_offset;
 288                         pfn = page_to_pfn(page);
 289                 }
 290
 291                 /*
 292                  * Note that the value of @prot at this point may differ from
 293                  * the value of @vma->vm_page_prot in the caching- and
 294                  * encryption bits. This is because the exact location of the
 295                  * data may not be known at mmap() time and may also change
 296                  * at arbitrary times while the data is mmap'ed.
 297                  * See vmf_insert_mixed_prot() for a discussion.
 298                  */
 299                 if (vma->vm_flags & VM_MIXEDMAP)
 300                         ret = vmf_insert_mixed_prot(vma, address,
 301                                                     __pfn_to_pfn_t(pfn, PFN_DEV),
 302                                                     prot);
 303                 else
 304                         ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
 305
 306                 /* Never error on prefaulted PTEs */
 307                 if (unlikely((ret & VM_FAULT_ERROR))) {
 308                         if (i == 0)
 309                                 goto out_io_unlock;
 310                         else
 311                                 break;
 312                 }
 313
 314                 address += PAGE_SIZE;
 315                 if (unlikely(++page_offset >= page_last))
 316                         break;
 317         }
 318         ret = VM_FAULT_NOPAGE;
 319 out_io_unlock:
 320         ttm_mem_io_unlock(man);
 321         return ret;
 322 }
 323 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
 324
 325 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
 326 {
 327         struct vm_area_struct *vma = vmf->vma;
 328         pgprot_t prot;
 329         struct ttm_buffer_object *bo = vma->vm_private_data;
 330         vm_fault_t ret;
 331
 332         ret = ttm_bo_vm_reserve(bo, vmf);
 333         if (ret)
 334                 return ret;
 335
 336         prot = vma->vm_page_prot;
 337         ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
 338         if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
 339                 return ret;
 340
 341         dma_resv_unlock(bo->base.resv);
 342
 343         return ret;
 344 }
 345 EXPORT_SYMBOL(ttm_bo_vm_fault);
 346
 347 void ttm_bo_vm_open(struct vm_area_struct *vma)
 348 {
 349         struct ttm_buffer_object *bo = vma->vm_private_data;
 350
 351         WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
 352
 353         ttm_bo_get(bo);
 354 }
 355 EXPORT_SYMBOL(ttm_bo_vm_open);
 356
 357 void ttm_bo_vm_close(struct vm_area_struct *vma)
 358 {
 359         struct ttm_buffer_object *bo = vma->vm_private_data;
 360
 361         ttm_bo_put(bo);
 362         vma->vm_private_data = NULL;
 363 }
 364 EXPORT_SYMBOL(ttm_bo_vm_close);
 365
 366 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
 367                                  unsigned long offset,
 368                                  uint8_t *buf, int len, int write)
 369 {
 370         unsigned long page = offset >> PAGE_SHIFT;
 371         unsigned long bytes_left = len;
 372         int ret;
 373
 374         /* Copy a page at a time, that way no extra virtual address
 375          * mapping is needed
 376          */
 377         offset -= page << PAGE_SHIFT;
 378         do {
 379                 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
 380                 struct ttm_bo_kmap_obj map;
 381                 void *ptr;
 382                 bool is_iomem;
 383
 384                 ret = ttm_bo_kmap(bo, page, 1, &map);
 385                 if (ret)
 386                         return ret;
 387
 388                 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
 389                 WARN_ON_ONCE(is_iomem);
 390                 if (write)
 391                         memcpy(ptr, buf, bytes);
 392                 else
 393                         memcpy(buf, ptr, bytes);
 394                 ttm_bo_kunmap(&map);
 395
 396                 page++;
 397                 buf += bytes;
 398                 bytes_left -= bytes;
 399                 offset = 0;
 400         } while (bytes_left);
 401
 402         return len;
 403 }
 404
 405 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
 406                      void *buf, int len, int write)
 407 {
 408         unsigned long offset = (addr) - vma->vm_start;
 409         struct ttm_buffer_object *bo = vma->vm_private_data;
 410         int ret;
 411
 412         if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
 413                 return -EIO;
 414
 415         ret = ttm_bo_reserve(bo, true, false, NULL);
 416         if (ret)
 417                 return ret;
 418
 419         switch (bo->mem.mem_type) {
 420         case TTM_PL_SYSTEM:
 421                 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 422                         ret = ttm_tt_swapin(bo->ttm);
 423                         if (unlikely(ret != 0))
 424                                 return ret;
 425                 }
 426                 /* fall through */
 427         case TTM_PL_TT:
 428                 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
 429                 break;
 430         default:
 431                 if (bo->bdev->driver->access_memory)
 432                         ret = bo->bdev->driver->access_memory(
 433                                 bo, offset, buf, len, write);
 434                 else
 435                         ret = -EIO;
 436         }
 437
 438         ttm_bo_unreserve(bo);
 439
 440         return ret;
 441 }
 442 EXPORT_SYMBOL(ttm_bo_vm_access);
 443
 444 static const struct vm_operations_struct ttm_bo_vm_ops = {
 445         .fault = ttm_bo_vm_fault,
 446         .open = ttm_bo_vm_open,
 447         .close = ttm_bo_vm_close,
 448         .access = ttm_bo_vm_access
 449 };
 450
 451 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
 452                                                   unsigned long offset,
 453                                                   unsigned long pages)
 454 {
 455         struct drm_vma_offset_node *node;
 456         struct ttm_buffer_object *bo = NULL;
 457
 458         drm_vma_offset_lock_lookup(bdev->vma_manager);
 459
 460         node = drm_vma_offset_lookup_locked(bdev->vma_manager, offset, pages);
 461         if (likely(node)) {
 462                 bo = container_of(node, struct ttm_buffer_object,
 463                                   base.vma_node);
 464                 bo = ttm_bo_get_unless_zero(bo);
 465         }
 466
 467         drm_vma_offset_unlock_lookup(bdev->vma_manager);
 468
 469         if (!bo)
 470                 pr_err("Could not find buffer object to map\n");
 471
 472         return bo;
 473 }
 474
 475 static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_struct *vma)
 476 {
 477         vma->vm_ops = &ttm_bo_vm_ops;
 478
 479         /*
 480          * Note: We're transferring the bo reference to
 481          * vma->vm_private_data here.
 482          */
 483
 484         vma->vm_private_data = bo;
 485
 486         /*
 487          * We'd like to use VM_PFNMAP on shared mappings, where
 488          * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
 489          * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
 490          * bad for performance. Until that has been sorted out, use
 491          * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
 492          */
 493         vma->vm_flags |= VM_MIXEDMAP;
 494         vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
 495 }
 496
 497 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
 498                 struct ttm_bo_device *bdev)
 499 {
 500         struct ttm_bo_driver *driver;
 501         struct ttm_buffer_object *bo;
 502         int ret;
 503
 504         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
 505                 return -EINVAL;
 506
 507         bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
 508         if (unlikely(!bo))
 509                 return -EINVAL;
 510
 511         driver = bo->bdev->driver;
 512         if (unlikely(!driver->verify_access)) {
 513                 ret = -EPERM;
 514                 goto out_unref;
 515         }
 516         ret = driver->verify_access(bo, filp);
 517         if (unlikely(ret != 0))
 518                 goto out_unref;
 519
 520         ttm_bo_mmap_vma_setup(bo, vma);
 521         return 0;
 522 out_unref:
 523         ttm_bo_put(bo);
 524         return ret;
 525 }
 526 EXPORT_SYMBOL(ttm_bo_mmap);
 527
 528 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 529 {
 530         ttm_bo_get(bo);
 531         ttm_bo_mmap_vma_setup(bo, vma);
 532         return 0;
 533 }
 534 EXPORT_SYMBOL(ttm_bo_mmap_obj);