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 vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
  48                                 struct vm_fault *vmf)
  49 {
  50         vm_fault_t ret = 0;
  51         int err = 0;
  52
  53         if (likely(!bo->moving))
  54                 goto out_unlock;
  55
  56         /*
  57          * Quick non-stalling check for idle.
  58          */
  59         if (dma_fence_is_signaled(bo->moving))
  60                 goto out_clear;
  61
  62         /*
  63          * If possible, avoid waiting for GPU with mmap_sem
  64          * held.
  65          */
  66         if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
  67                 ret = VM_FAULT_RETRY;
  68                 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
  69                         goto out_unlock;
  70
  71                 ttm_bo_reference(bo);
  72                 up_read(&vmf->vma->vm_mm->mmap_sem);
  73                 (void) dma_fence_wait(bo->moving, true);
  74                 ttm_bo_unreserve(bo);
  75                 ttm_bo_unref(&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_bo_device *bdev = bo->bdev;
 101
 102         if (bdev->driver->io_mem_pfn)
 103                 return bdev->driver->io_mem_pfn(bo, page_offset);
 104
 105         return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
 106                 + page_offset;
 107 }
 108
 109 static vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
 110 {
 111         struct vm_area_struct *vma = vmf->vma;
 112         struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
 113             vma->vm_private_data;
 114         struct ttm_bo_device *bdev = bo->bdev;
 115         unsigned long page_offset;
 116         unsigned long page_last;
 117         unsigned long pfn;
 118         struct ttm_tt *ttm = NULL;
 119         struct page *page;
 120         int err;
 121         int i;
 122         vm_fault_t ret = VM_FAULT_NOPAGE;
 123         unsigned long address = vmf->address;
 124         struct ttm_mem_type_manager *man =
 125                 &bdev->man[bo->mem.mem_type];
 126         struct vm_area_struct cvma;
 127
 128         /*
 129          * Work around locking order reversal in fault / nopfn
 130          * between mmap_sem and bo_reserve: Perform a trylock operation
 131          * for reserve, and if it fails, retry the fault after waiting
 132          * for the buffer to become unreserved.
 133          */
 134         err = ttm_bo_reserve(bo, true, true, NULL);
 135         if (unlikely(err != 0)) {
 136                 if (err != -EBUSY)
 137                         return VM_FAULT_NOPAGE;
 138
 139                 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
 140                         if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 141                                 ttm_bo_reference(bo);
 142                                 up_read(&vmf->vma->vm_mm->mmap_sem);
 143                                 (void) ttm_bo_wait_unreserved(bo);
 144                                 ttm_bo_unref(&bo);
 145                         }
 146
 147                         return VM_FAULT_RETRY;
 148                 }
 149
 150                 /*
 151                  * If we'd want to change locking order to
 152                  * mmap_sem -> bo::reserve, we'd use a blocking reserve here
 153                  * instead of retrying the fault...
 154                  */
 155                 return VM_FAULT_NOPAGE;
 156         }
 157
 158         /*
 159          * Refuse to fault imported pages. This should be handled
 160          * (if at all) by redirecting mmap to the exporter.
 161          */
 162         if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
 163                 ret = VM_FAULT_SIGBUS;
 164                 goto out_unlock;
 165         }
 166
 167         if (bdev->driver->fault_reserve_notify) {
 168                 err = bdev->driver->fault_reserve_notify(bo);
 169                 switch (err) {
 170                 case 0:
 171                         break;
 172                 case -EBUSY:
 173                 case -ERESTARTSYS:
 174                         ret = VM_FAULT_NOPAGE;
 175                         goto out_unlock;
 176                 default:
 177                         ret = VM_FAULT_SIGBUS;
 178                         goto out_unlock;
 179                 }
 180         }
 181
 182         /*
 183          * Wait for buffer data in transit, due to a pipelined
 184          * move.
 185          */
 186         ret = ttm_bo_vm_fault_idle(bo, vmf);
 187         if (unlikely(ret != 0)) {
 188                 if (ret == VM_FAULT_RETRY &&
 189                     !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 190                         /* The BO has already been unreserved. */
 191                         return ret;
 192                 }
 193
 194                 goto out_unlock;
 195         }
 196
 197         err = ttm_mem_io_lock(man, true);
 198         if (unlikely(err != 0)) {
 199                 ret = VM_FAULT_NOPAGE;
 200                 goto out_unlock;
 201         }
 202         err = ttm_mem_io_reserve_vm(bo);
 203         if (unlikely(err != 0)) {
 204                 ret = VM_FAULT_SIGBUS;
 205                 goto out_io_unlock;
 206         }
 207
 208         page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
 209                 vma->vm_pgoff - drm_vma_node_start(&bo->vma_node);
 210         page_last = vma_pages(vma) + vma->vm_pgoff -
 211                 drm_vma_node_start(&bo->vma_node);
 212
 213         if (unlikely(page_offset >= bo->num_pages)) {
 214                 ret = VM_FAULT_SIGBUS;
 215                 goto out_io_unlock;
 216         }
 217
 218         /*
 219          * Make a local vma copy to modify the page_prot member
 220          * and vm_flags if necessary. The vma parameter is protected
 221          * by mmap_sem in write mode.
 222          */
 223         cvma = *vma;
 224         cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
 225
 226         if (bo->mem.bus.is_iomem) {
 227                 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
 228                                                 cvma.vm_page_prot);
 229         } else {
 230                 struct ttm_operation_ctx ctx = {
 231                         .interruptible = false,
 232                         .no_wait_gpu = false,
 233                         .flags = TTM_OPT_FLAG_FORCE_ALLOC
 234
 235                 };
 236
 237                 ttm = bo->ttm;
 238                 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
 239                                                 cvma.vm_page_prot);
 240
 241                 /* Allocate all page at once, most common usage */
 242                 if (ttm_tt_populate(ttm, &ctx)) {
 243                         ret = VM_FAULT_OOM;
 244                         goto out_io_unlock;
 245                 }
 246         }
 247
 248         /*
 249          * Speculatively prefault a number of pages. Only error on
 250          * first page.
 251          */
 252         for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
 253                 if (bo->mem.bus.is_iomem) {
 254                         /* Iomem should not be marked encrypted */
 255                         cvma.vm_page_prot = pgprot_decrypted(cvma.vm_page_prot);
 256                         pfn = ttm_bo_io_mem_pfn(bo, page_offset);
 257                 } else {
 258                         page = ttm->pages[page_offset];
 259                         if (unlikely(!page && i == 0)) {
 260                                 ret = VM_FAULT_OOM;
 261                                 goto out_io_unlock;
 262                         } else if (unlikely(!page)) {
 263                                 break;
 264                         }
 265                         page->index = drm_vma_node_start(&bo->vma_node) +
 266                                 page_offset;
 267                         pfn = page_to_pfn(page);
 268                 }
 269
 270                 if (vma->vm_flags & VM_MIXEDMAP)
 271                         ret = vmf_insert_mixed(&cvma, address,
 272                                         __pfn_to_pfn_t(pfn, PFN_DEV));
 273                 else
 274                         ret = vmf_insert_pfn(&cvma, address, pfn);
 275
 276                 /*
 277                  * Somebody beat us to this PTE or prefaulting to
 278                  * an already populated PTE, or prefaulting error.
 279                  */
 280
 281                 if (unlikely((ret == VM_FAULT_NOPAGE && i > 0)))
 282                         break;
 283                 else if (unlikely(ret & VM_FAULT_ERROR))
 284                         goto out_io_unlock;
 285
 286                 address += PAGE_SIZE;
 287                 if (unlikely(++page_offset >= page_last))
 288                         break;
 289         }
 290         ret = VM_FAULT_NOPAGE;
 291 out_io_unlock:
 292         ttm_mem_io_unlock(man);
 293 out_unlock:
 294         ttm_bo_unreserve(bo);
 295         return ret;
 296 }
 297
 298 static void ttm_bo_vm_open(struct vm_area_struct *vma)
 299 {
 300         struct ttm_buffer_object *bo =
 301             (struct ttm_buffer_object *)vma->vm_private_data;
 302
 303         WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
 304
 305         (void)ttm_bo_reference(bo);
 306 }
 307
 308 static void ttm_bo_vm_close(struct vm_area_struct *vma)
 309 {
 310         struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
 311
 312         ttm_bo_unref(&bo);
 313         vma->vm_private_data = NULL;
 314 }
 315
 316 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
 317                                  unsigned long offset,
 318                                  uint8_t *buf, int len, int write)
 319 {
 320         unsigned long page = offset >> PAGE_SHIFT;
 321         unsigned long bytes_left = len;
 322         int ret;
 323
 324         /* Copy a page at a time, that way no extra virtual address
 325          * mapping is needed
 326          */
 327         offset -= page << PAGE_SHIFT;
 328         do {
 329                 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
 330                 struct ttm_bo_kmap_obj map;
 331                 void *ptr;
 332                 bool is_iomem;
 333
 334                 ret = ttm_bo_kmap(bo, page, 1, &map);
 335                 if (ret)
 336                         return ret;
 337
 338                 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
 339                 WARN_ON_ONCE(is_iomem);
 340                 if (write)
 341                         memcpy(ptr, buf, bytes);
 342                 else
 343                         memcpy(buf, ptr, bytes);
 344                 ttm_bo_kunmap(&map);
 345
 346                 page++;
 347                 buf += bytes;
 348                 bytes_left -= bytes;
 349                 offset = 0;
 350         } while (bytes_left);
 351
 352         return len;
 353 }
 354
 355 static int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
 356                             void *buf, int len, int write)
 357 {
 358         unsigned long offset = (addr) - vma->vm_start;
 359         struct ttm_buffer_object *bo = vma->vm_private_data;
 360         int ret;
 361
 362         if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
 363                 return -EIO;
 364
 365         ret = ttm_bo_reserve(bo, true, false, NULL);
 366         if (ret)
 367                 return ret;
 368
 369         switch (bo->mem.mem_type) {
 370         case TTM_PL_SYSTEM:
 371                 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
 372                         ret = ttm_tt_swapin(bo->ttm);
 373                         if (unlikely(ret != 0))
 374                                 return ret;
 375                 }
 376                 /* fall through */
 377         case TTM_PL_TT:
 378                 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
 379                 break;
 380         default:
 381                 if (bo->bdev->driver->access_memory)
 382                         ret = bo->bdev->driver->access_memory(
 383                                 bo, offset, buf, len, write);
 384                 else
 385                         ret = -EIO;
 386         }
 387
 388         ttm_bo_unreserve(bo);
 389
 390         return ret;
 391 }
 392
 393 static const struct vm_operations_struct ttm_bo_vm_ops = {
 394         .fault = ttm_bo_vm_fault,
 395         .open = ttm_bo_vm_open,
 396         .close = ttm_bo_vm_close,
 397         .access = ttm_bo_vm_access
 398 };
 399
 400 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
 401                                                   unsigned long offset,
 402                                                   unsigned long pages)
 403 {
 404         struct drm_vma_offset_node *node;
 405         struct ttm_buffer_object *bo = NULL;
 406
 407         drm_vma_offset_lock_lookup(&bdev->vma_manager);
 408
 409         node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
 410         if (likely(node)) {
 411                 bo = container_of(node, struct ttm_buffer_object, vma_node);
 412                 if (!kref_get_unless_zero(&bo->kref))
 413                         bo = NULL;
 414         }
 415
 416         drm_vma_offset_unlock_lookup(&bdev->vma_manager);
 417
 418         if (!bo)
 419                 pr_err("Could not find buffer object to map\n");
 420
 421         return bo;
 422 }
 423
 424 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
 425                 struct ttm_bo_device *bdev)
 426 {
 427         struct ttm_bo_driver *driver;
 428         struct ttm_buffer_object *bo;
 429         int ret;
 430
 431         bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
 432         if (unlikely(!bo))
 433                 return -EINVAL;
 434
 435         driver = bo->bdev->driver;
 436         if (unlikely(!driver->verify_access)) {
 437                 ret = -EPERM;
 438                 goto out_unref;
 439         }
 440         ret = driver->verify_access(bo, filp);
 441         if (unlikely(ret != 0))
 442                 goto out_unref;
 443
 444         vma->vm_ops = &ttm_bo_vm_ops;
 445
 446         /*
 447          * Note: We're transferring the bo reference to
 448          * vma->vm_private_data here.
 449          */
 450
 451         vma->vm_private_data = bo;
 452
 453         /*
 454          * We'd like to use VM_PFNMAP on shared mappings, where
 455          * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
 456          * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
 457          * bad for performance. Until that has been sorted out, use
 458          * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
 459          */
 460         vma->vm_flags |= VM_MIXEDMAP;
 461         vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
 462         return 0;
 463 out_unref:
 464         ttm_bo_unref(&bo);
 465         return ret;
 466 }
 467 EXPORT_SYMBOL(ttm_bo_mmap);
 468
 469 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 470 {
 471         if (vma->vm_pgoff != 0)
 472                 return -EACCES;
 473
 474         vma->vm_ops = &ttm_bo_vm_ops;
 475         vma->vm_private_data = ttm_bo_reference(bo);
 476         vma->vm_flags |= VM_MIXEDMAP;
 477         vma->vm_flags |= VM_IO | VM_DONTEXPAND;
 478         return 0;
 479 }
 480 EXPORT_SYMBOL(ttm_fbdev_mmap);