drivers/gpu/drm/ttm/ttm_bo_vm.c

   1 /**************************************************************************
   2  *
   3  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27 /*
  28  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
  29  */
  30
  31 #define pr_fmt(fmt) "[TTM] " fmt
  32
  33 #include <ttm/ttm_module.h>
  34 #include <ttm/ttm_bo_driver.h>
  35 #include <ttm/ttm_placement.h>
  36 #include <drm/drm_vma_manager.h>
  37 #include <linux/mm.h>
  38 #include <linux/pfn_t.h>
  39 #include <linux/rbtree.h>
  40 #include <linux/module.h>
  41 #include <linux/uaccess.h>
  42
  43 #define TTM_BO_VM_NUM_PREFAULT 16
  44
  45 static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
  46                                 struct vm_area_struct *vma,
  47                                 struct vm_fault *vmf)
  48 {
  49         int ret = 0;
  50
  51         if (likely(!test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)))
  52                 goto out_unlock;
  53
  54         /*
  55          * Quick non-stalling check for idle.
  56          */
  57         ret = ttm_bo_wait(bo, false, true);
  58         if (likely(ret == 0))
  59                 goto out_unlock;
  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                 up_read(&vma->vm_mm->mmap_sem);
  71                 (void) ttm_bo_wait(bo, true, false);
  72                 goto out_unlock;
  73         }
  74
  75         /*
  76          * Ordinary wait.
  77          */
  78         ret = ttm_bo_wait(bo, true, false);
  79         if (unlikely(ret != 0))
  80                 ret = (ret != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
  81                         VM_FAULT_NOPAGE;
  82
  83 out_unlock:
  84         return ret;
  85 }
  86
  87 static int ttm_bo_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
  88 {
  89         struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
  90             vma->vm_private_data;
  91         struct ttm_bo_device *bdev = bo->bdev;
  92         unsigned long page_offset;
  93         unsigned long page_last;
  94         unsigned long pfn;
  95         struct ttm_tt *ttm = NULL;
  96         struct page *page;
  97         int ret;
  98         int i;
  99         unsigned long address = (unsigned long)vmf->virtual_address;
 100         int retval = VM_FAULT_NOPAGE;
 101         struct ttm_mem_type_manager *man =
 102                 &bdev->man[bo->mem.mem_type];
 103         struct vm_area_struct cvma;
 104
 105         /*
 106          * Work around locking order reversal in fault / nopfn
 107          * between mmap_sem and bo_reserve: Perform a trylock operation
 108          * for reserve, and if it fails, retry the fault after waiting
 109          * for the buffer to become unreserved.
 110          */
 111         ret = ttm_bo_reserve(bo, true, true, NULL);
 112         if (unlikely(ret != 0)) {
 113                 if (ret != -EBUSY)
 114                         return VM_FAULT_NOPAGE;
 115
 116                 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
 117                         if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 118                                 up_read(&vma->vm_mm->mmap_sem);
 119                                 (void) ttm_bo_wait_unreserved(bo);
 120                         }
 121
 122                         return VM_FAULT_RETRY;
 123                 }
 124
 125                 /*
 126                  * If we'd want to change locking order to
 127                  * mmap_sem -> bo::reserve, we'd use a blocking reserve here
 128                  * instead of retrying the fault...
 129                  */
 130                 return VM_FAULT_NOPAGE;
 131         }
 132
 133         /*
 134          * Refuse to fault imported pages. This should be handled
 135          * (if at all) by redirecting mmap to the exporter.
 136          */
 137         if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
 138                 retval = VM_FAULT_SIGBUS;
 139                 goto out_unlock;
 140         }
 141
 142         if (bdev->driver->fault_reserve_notify) {
 143                 ret = bdev->driver->fault_reserve_notify(bo);
 144                 switch (ret) {
 145                 case 0:
 146                         break;
 147                 case -EBUSY:
 148                 case -ERESTARTSYS:
 149                         retval = VM_FAULT_NOPAGE;
 150                         goto out_unlock;
 151                 default:
 152                         retval = VM_FAULT_SIGBUS;
 153                         goto out_unlock;
 154                 }
 155         }
 156
 157         /*
 158          * Wait for buffer data in transit, due to a pipelined
 159          * move.
 160          */
 161         ret = ttm_bo_vm_fault_idle(bo, vma, vmf);
 162         if (unlikely(ret != 0)) {
 163                 retval = ret;
 164                 goto out_unlock;
 165         }
 166
 167         ret = ttm_mem_io_lock(man, true);
 168         if (unlikely(ret != 0)) {
 169                 retval = VM_FAULT_NOPAGE;
 170                 goto out_unlock;
 171         }
 172         ret = ttm_mem_io_reserve_vm(bo);
 173         if (unlikely(ret != 0)) {
 174                 retval = VM_FAULT_SIGBUS;
 175                 goto out_io_unlock;
 176         }
 177
 178         page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
 179                 vma->vm_pgoff - drm_vma_node_start(&bo->vma_node);
 180         page_last = vma_pages(vma) + vma->vm_pgoff -
 181                 drm_vma_node_start(&bo->vma_node);
 182
 183         if (unlikely(page_offset >= bo->num_pages)) {
 184                 retval = VM_FAULT_SIGBUS;
 185                 goto out_io_unlock;
 186         }
 187
 188         /*
 189          * Make a local vma copy to modify the page_prot member
 190          * and vm_flags if necessary. The vma parameter is protected
 191          * by mmap_sem in write mode.
 192          */
 193         cvma = *vma;
 194         cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
 195
 196         if (bo->mem.bus.is_iomem) {
 197                 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
 198                                                 cvma.vm_page_prot);
 199         } else {
 200                 ttm = bo->ttm;
 201                 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
 202                                                 cvma.vm_page_prot);
 203
 204                 /* Allocate all page at once, most common usage */
 205                 if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
 206                         retval = VM_FAULT_OOM;
 207                         goto out_io_unlock;
 208                 }
 209         }
 210
 211         /*
 212          * Speculatively prefault a number of pages. Only error on
 213          * first page.
 214          */
 215         for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
 216                 if (bo->mem.bus.is_iomem)
 217                         pfn = ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT) + page_offset;
 218                 else {
 219                         page = ttm->pages[page_offset];
 220                         if (unlikely(!page && i == 0)) {
 221                                 retval = VM_FAULT_OOM;
 222                                 goto out_io_unlock;
 223                         } else if (unlikely(!page)) {
 224                                 break;
 225                         }
 226                         page->mapping = vma->vm_file->f_mapping;
 227                         page->index = drm_vma_node_start(&bo->vma_node) +
 228                                 page_offset;
 229                         pfn = page_to_pfn(page);
 230                 }
 231
 232                 if (vma->vm_flags & VM_MIXEDMAP)
 233                         ret = vm_insert_mixed(&cvma, address,
 234                                         __pfn_to_pfn_t(pfn, PFN_DEV));
 235                 else
 236                         ret = vm_insert_pfn(&cvma, address, pfn);
 237
 238                 /*
 239                  * Somebody beat us to this PTE or prefaulting to
 240                  * an already populated PTE, or prefaulting error.
 241                  */
 242
 243                 if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0)))
 244                         break;
 245                 else if (unlikely(ret != 0)) {
 246                         retval =
 247                             (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
 248                         goto out_io_unlock;
 249                 }
 250
 251                 address += PAGE_SIZE;
 252                 if (unlikely(++page_offset >= page_last))
 253                         break;
 254         }
 255 out_io_unlock:
 256         ttm_mem_io_unlock(man);
 257 out_unlock:
 258         ttm_bo_unreserve(bo);
 259         return retval;
 260 }
 261
 262 static void ttm_bo_vm_open(struct vm_area_struct *vma)
 263 {
 264         struct ttm_buffer_object *bo =
 265             (struct ttm_buffer_object *)vma->vm_private_data;
 266
 267         WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
 268
 269         (void)ttm_bo_reference(bo);
 270 }
 271
 272 static void ttm_bo_vm_close(struct vm_area_struct *vma)
 273 {
 274         struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
 275
 276         ttm_bo_unref(&bo);
 277         vma->vm_private_data = NULL;
 278 }
 279
 280 static const struct vm_operations_struct ttm_bo_vm_ops = {
 281         .fault = ttm_bo_vm_fault,
 282         .open = ttm_bo_vm_open,
 283         .close = ttm_bo_vm_close
 284 };
 285
 286 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
 287                                                   unsigned long offset,
 288                                                   unsigned long pages)
 289 {
 290         struct drm_vma_offset_node *node;
 291         struct ttm_buffer_object *bo = NULL;
 292
 293         drm_vma_offset_lock_lookup(&bdev->vma_manager);
 294
 295         node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
 296         if (likely(node)) {
 297                 bo = container_of(node, struct ttm_buffer_object, vma_node);
 298                 if (!kref_get_unless_zero(&bo->kref))
 299                         bo = NULL;
 300         }
 301
 302         drm_vma_offset_unlock_lookup(&bdev->vma_manager);
 303
 304         if (!bo)
 305                 pr_err("Could not find buffer object to map\n");
 306
 307         return bo;
 308 }
 309
 310 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
 311                 struct ttm_bo_device *bdev)
 312 {
 313         struct ttm_bo_driver *driver;
 314         struct ttm_buffer_object *bo;
 315         int ret;
 316
 317         bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
 318         if (unlikely(!bo))
 319                 return -EINVAL;
 320
 321         driver = bo->bdev->driver;
 322         if (unlikely(!driver->verify_access)) {
 323                 ret = -EPERM;
 324                 goto out_unref;
 325         }
 326         ret = driver->verify_access(bo, filp);
 327         if (unlikely(ret != 0))
 328                 goto out_unref;
 329
 330         vma->vm_ops = &ttm_bo_vm_ops;
 331
 332         /*
 333          * Note: We're transferring the bo reference to
 334          * vma->vm_private_data here.
 335          */
 336
 337         vma->vm_private_data = bo;
 338
 339         /*
 340          * We'd like to use VM_PFNMAP on shared mappings, where
 341          * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
 342          * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
 343          * bad for performance. Until that has been sorted out, use
 344          * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
 345          */
 346         vma->vm_flags |= VM_MIXEDMAP;
 347         vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
 348         return 0;
 349 out_unref:
 350         ttm_bo_unref(&bo);
 351         return ret;
 352 }
 353 EXPORT_SYMBOL(ttm_bo_mmap);
 354
 355 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 356 {
 357         if (vma->vm_pgoff != 0)
 358                 return -EACCES;
 359
 360         vma->vm_ops = &ttm_bo_vm_ops;
 361         vma->vm_private_data = ttm_bo_reference(bo);
 362         vma->vm_flags |= VM_MIXEDMAP;
 363         vma->vm_flags |= VM_IO | VM_DONTEXPAND;
 364         return 0;
 365 }
 366 EXPORT_SYMBOL(ttm_fbdev_mmap);