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 #define TTM_BO_VM_NUM_PREFAULT 16
  46
  47 static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
  48                                 struct vm_fault *vmf)
  49 {
  50         int ret = 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_sem
  63          * held.
  64          */
  65         if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
  66                 ret = VM_FAULT_RETRY;
  67                 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
  68                         goto out_unlock;
  69
  70                 ttm_bo_reference(bo);
  71                 up_read(&vmf->vma->vm_mm->mmap_sem);
  72                 (void) dma_fence_wait(bo->moving, true);
  73                 ttm_bo_unreserve(bo);
  74                 ttm_bo_unref(&bo);
  75                 goto out_unlock;
  76         }
  77
  78         /*
  79          * Ordinary wait.
  80          */
  81         ret = dma_fence_wait(bo->moving, true);
  82         if (unlikely(ret != 0)) {
  83                 ret = (ret != -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 static int ttm_bo_vm_fault(struct vm_fault *vmf)
 109 {
 110         struct vm_area_struct *vma = vmf->vma;
 111         struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
 112             vma->vm_private_data;
 113         struct ttm_bo_device *bdev = bo->bdev;
 114         unsigned long page_offset;
 115         unsigned long page_last;
 116         unsigned long pfn;
 117         struct ttm_tt *ttm = NULL;
 118         struct page *page;
 119         int ret;
 120         int i;
 121         unsigned long address = vmf->address;
 122         struct ttm_mem_type_manager *man =
 123                 &bdev->man[bo->mem.mem_type];
 124         struct vm_area_struct cvma;
 125
 126         /*
 127          * Work around locking order reversal in fault / nopfn
 128          * between mmap_sem and bo_reserve: Perform a trylock operation
 129          * for reserve, and if it fails, retry the fault after waiting
 130          * for the buffer to become unreserved.
 131          */
 132         ret = ttm_bo_reserve(bo, true, true, NULL);
 133         if (unlikely(ret != 0)) {
 134                 if (ret != -EBUSY)
 135                         return VM_FAULT_NOPAGE;
 136
 137                 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
 138                         if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 139                                 ttm_bo_reference(bo);
 140                                 up_read(&vmf->vma->vm_mm->mmap_sem);
 141                                 (void) ttm_bo_wait_unreserved(bo);
 142                                 ttm_bo_unref(&bo);
 143                         }
 144
 145                         return VM_FAULT_RETRY;
 146                 }
 147
 148                 /*
 149                  * If we'd want to change locking order to
 150                  * mmap_sem -> bo::reserve, we'd use a blocking reserve here
 151                  * instead of retrying the fault...
 152                  */
 153                 return VM_FAULT_NOPAGE;
 154         }
 155
 156         /*
 157          * Refuse to fault imported pages. This should be handled
 158          * (if at all) by redirecting mmap to the exporter.
 159          */
 160         if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
 161                 ret = VM_FAULT_SIGBUS;
 162                 goto out_unlock;
 163         }
 164
 165         if (bdev->driver->fault_reserve_notify) {
 166                 ret = bdev->driver->fault_reserve_notify(bo);
 167                 switch (ret) {
 168                 case 0:
 169                         break;
 170                 case -EBUSY:
 171                 case -ERESTARTSYS:
 172                         ret = VM_FAULT_NOPAGE;
 173                         goto out_unlock;
 174                 default:
 175                         ret = VM_FAULT_SIGBUS;
 176                         goto out_unlock;
 177                 }
 178         }
 179
 180         /*
 181          * Wait for buffer data in transit, due to a pipelined
 182          * move.
 183          */
 184         ret = ttm_bo_vm_fault_idle(bo, vmf);
 185         if (unlikely(ret != 0)) {
 186                 if (ret == VM_FAULT_RETRY &&
 187                     !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 188                         /* The BO has already been unreserved. */
 189                         return ret;
 190                 }
 191
 192                 goto out_unlock;
 193         }
 194
 195         ret = ttm_mem_io_lock(man, true);
 196         if (unlikely(ret != 0)) {
 197                 ret = VM_FAULT_NOPAGE;
 198                 goto out_unlock;
 199         }
 200         ret = ttm_mem_io_reserve_vm(bo);
 201         if (unlikely(ret != 0)) {
 202                 ret = VM_FAULT_SIGBUS;
 203                 goto out_io_unlock;
 204         }
 205
 206         page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
 207                 vma->vm_pgoff - drm_vma_node_start(&bo->vma_node);
 208         page_last = vma_pages(vma) + vma->vm_pgoff -
 209                 drm_vma_node_start(&bo->vma_node);
 210
 211         if (unlikely(page_offset >= bo->num_pages)) {
 212                 ret = VM_FAULT_SIGBUS;
 213                 goto out_io_unlock;
 214         }
 215
 216         /*
 217          * Make a local vma copy to modify the page_prot member
 218          * and vm_flags if necessary. The vma parameter is protected
 219          * by mmap_sem in write mode.
 220          */
 221         cvma = *vma;
 222         cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
 223
 224         if (bo->mem.bus.is_iomem) {
 225                 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
 226                                                 cvma.vm_page_prot);
 227         } else {
 228                 struct ttm_operation_ctx ctx = {
 229                         .interruptible = false,
 230                         .no_wait_gpu = false,
 231                         .flags = TTM_OPT_FLAG_FORCE_ALLOC
 232
 233                 };
 234
 235                 ttm = bo->ttm;
 236                 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
 237                                                 cvma.vm_page_prot);
 238
 239                 /* Allocate all page at once, most common usage */
 240                 if (ttm_tt_populate(ttm, &ctx)) {
 241                         ret = VM_FAULT_OOM;
 242                         goto out_io_unlock;
 243                 }
 244         }
 245
 246         /*
 247          * Speculatively prefault a number of pages. Only error on
 248          * first page.
 249          */
 250         for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
 251                 if (bo->mem.bus.is_iomem) {
 252                         /* Iomem should not be marked encrypted */
 253                         cvma.vm_page_prot = pgprot_decrypted(cvma.vm_page_prot);
 254                         pfn = ttm_bo_io_mem_pfn(bo, page_offset);
 255                 } else {
 256                         page = ttm->pages[page_offset];
 257                         if (unlikely(!page && i == 0)) {
 258                                 ret = VM_FAULT_OOM;
 259                                 goto out_io_unlock;
 260                         } else if (unlikely(!page)) {
 261                                 break;
 262                         }
 263                         page->index = drm_vma_node_start(&bo->vma_node) +
 264                                 page_offset;
 265                         pfn = page_to_pfn(page);
 266                 }
 267
 268                 if (vma->vm_flags & VM_MIXEDMAP)
 269                         ret = vm_insert_mixed(&cvma, address,
 270                                         __pfn_to_pfn_t(pfn, PFN_DEV));
 271                 else
 272                         ret = vm_insert_pfn(&cvma, address, pfn);
 273
 274                 /*
 275                  * Somebody beat us to this PTE or prefaulting to
 276                  * an already populated PTE, or prefaulting error.
 277                  */
 278
 279                 if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0)))
 280                         break;
 281                 else if (unlikely(ret != 0)) {
 282                         ret =
 283                             (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
 284                         goto out_io_unlock;
 285                 }
 286
 287                 address += PAGE_SIZE;
 288                 if (unlikely(++page_offset >= page_last))
 289                         break;
 290         }
 291         ret = VM_FAULT_NOPAGE;
 292 out_io_unlock:
 293         ttm_mem_io_unlock(man);
 294 out_unlock:
 295         ttm_bo_unreserve(bo);
 296         return ret;
 297 }
 298
 299 static void ttm_bo_vm_open(struct vm_area_struct *vma)
 300 {
 301         struct ttm_buffer_object *bo =
 302             (struct ttm_buffer_object *)vma->vm_private_data;
 303
 304         WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
 305
 306         (void)ttm_bo_reference(bo);
 307 }
 308
 309 static void ttm_bo_vm_close(struct vm_area_struct *vma)
 310 {
 311         struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
 312
 313         ttm_bo_unref(&bo);
 314         vma->vm_private_data = NULL;
 315 }
 316
 317 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
 318                                  unsigned long offset,
 319                                  uint8_t *buf, int len, int write)
 320 {
 321         unsigned long page = offset >> PAGE_SHIFT;
 322         unsigned long bytes_left = len;
 323         int ret;
 324
 325         /* Copy a page at a time, that way no extra virtual address
 326          * mapping is needed
 327          */
 328         offset -= page << PAGE_SHIFT;
 329         do {
 330                 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
 331                 struct ttm_bo_kmap_obj map;
 332                 void *ptr;
 333                 bool is_iomem;
 334
 335                 ret = ttm_bo_kmap(bo, page, 1, &map);
 336                 if (ret)
 337                         return ret;
 338
 339                 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
 340                 WARN_ON_ONCE(is_iomem);
 341                 if (write)
 342                         memcpy(ptr, buf, bytes);
 343                 else
 344                         memcpy(buf, ptr, bytes);
 345                 ttm_bo_kunmap(&map);
 346
 347                 page++;
 348                 buf += bytes;
 349                 bytes_left -= bytes;
 350                 offset = 0;
 351         } while (bytes_left);
 352
 353         return len;
 354 }
 355
 356 static int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
 357                             void *buf, int len, int write)
 358 {
 359         unsigned long offset = (addr) - vma->vm_start;
 360         struct ttm_buffer_object *bo = vma->vm_private_data;
 361         int ret;
 362
 363         if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
 364                 return -EIO;
 365
 366         ret = ttm_bo_reserve(bo, true, false, NULL);
 367         if (ret)
 368                 return ret;
 369
 370         switch (bo->mem.mem_type) {
 371         case TTM_PL_SYSTEM:
 372                 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 373                         ret = ttm_tt_swapin(bo->ttm);
 374                         if (unlikely(ret != 0))
 375                                 return ret;
 376                 }
 377                 /* fall through */
 378         case TTM_PL_TT:
 379                 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
 380                 break;
 381         default:
 382                 if (bo->bdev->driver->access_memory)
 383                         ret = bo->bdev->driver->access_memory(
 384                                 bo, offset, buf, len, write);
 385                 else
 386                         ret = -EIO;
 387         }
 388
 389         ttm_bo_unreserve(bo);
 390
 391         return ret;
 392 }
 393
 394 static const struct vm_operations_struct ttm_bo_vm_ops = {
 395         .fault = ttm_bo_vm_fault,
 396         .open = ttm_bo_vm_open,
 397         .close = ttm_bo_vm_close,
 398         .access = ttm_bo_vm_access
 399 };
 400
 401 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
 402                                                   unsigned long offset,
 403                                                   unsigned long pages)
 404 {
 405         struct drm_vma_offset_node *node;
 406         struct ttm_buffer_object *bo = NULL;
 407
 408         drm_vma_offset_lock_lookup(&bdev->vma_manager);
 409
 410         node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
 411         if (likely(node)) {
 412                 bo = container_of(node, struct ttm_buffer_object, vma_node);
 413                 if (!kref_get_unless_zero(&bo->kref))
 414                         bo = NULL;
 415         }
 416
 417         drm_vma_offset_unlock_lookup(&bdev->vma_manager);
 418
 419         if (!bo)
 420                 pr_err("Could not find buffer object to map\n");
 421
 422         return bo;
 423 }
 424
 425 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
 426                 struct ttm_bo_device *bdev)
 427 {
 428         struct ttm_bo_driver *driver;
 429         struct ttm_buffer_object *bo;
 430         int ret;
 431
 432         bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
 433         if (unlikely(!bo))
 434                 return -EINVAL;
 435
 436         driver = bo->bdev->driver;
 437         if (unlikely(!driver->verify_access)) {
 438                 ret = -EPERM;
 439                 goto out_unref;
 440         }
 441         ret = driver->verify_access(bo, filp);
 442         if (unlikely(ret != 0))
 443                 goto out_unref;
 444
 445         vma->vm_ops = &ttm_bo_vm_ops;
 446
 447         /*
 448          * Note: We're transferring the bo reference to
 449          * vma->vm_private_data here.
 450          */
 451
 452         vma->vm_private_data = bo;
 453
 454         /*
 455          * We'd like to use VM_PFNMAP on shared mappings, where
 456          * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
 457          * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
 458          * bad for performance. Until that has been sorted out, use
 459          * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
 460          */
 461         vma->vm_flags |= VM_MIXEDMAP;
 462         vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
 463         return 0;
 464 out_unref:
 465         ttm_bo_unref(&bo);
 466         return ret;
 467 }
 468 EXPORT_SYMBOL(ttm_bo_mmap);
 469
 470 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 471 {
 472         if (vma->vm_pgoff != 0)
 473                 return -EACCES;
 474
 475         vma->vm_ops = &ttm_bo_vm_ops;
 476         vma->vm_private_data = ttm_bo_reference(bo);
 477         vma->vm_flags |= VM_MIXEDMAP;
 478         vma->vm_flags |= VM_IO | VM_DONTEXPAND;
 479         return 0;
 480 }
 481 EXPORT_SYMBOL(ttm_fbdev_mmap);