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_fault *vmf)
  47 {
  48         int ret = 0;
  49
  50         if (likely(!bo->moving))
  51                 goto out_unlock;
  52
  53         /*
  54          * Quick non-stalling check for idle.
  55          */
  56         if (dma_fence_is_signaled(bo->moving))
  57                 goto out_clear;
  58
  59         /*
  60          * If possible, avoid waiting for GPU with mmap_sem
  61          * held.
  62          */
  63         if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
  64                 ret = VM_FAULT_RETRY;
  65                 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
  66                         goto out_unlock;
  67
  68                 ttm_bo_reference(bo);
  69                 up_read(&vmf->vma->vm_mm->mmap_sem);
  70                 (void) dma_fence_wait(bo->moving, true);
  71                 ttm_bo_unreserve(bo);
  72                 ttm_bo_unref(&bo);
  73                 goto out_unlock;
  74         }
  75
  76         /*
  77          * Ordinary wait.
  78          */
  79         ret = dma_fence_wait(bo->moving, true);
  80         if (unlikely(ret != 0)) {
  81                 ret = (ret != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
  82                         VM_FAULT_NOPAGE;
  83                 goto out_unlock;
  84         }
  85
  86 out_clear:
  87         dma_fence_put(bo->moving);
  88         bo->moving = NULL;
  89
  90 out_unlock:
  91         return ret;
  92 }
  93
  94 static int ttm_bo_vm_fault(struct vm_fault *vmf)
  95 {
  96         struct vm_area_struct *vma = vmf->vma;
  97         struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
  98             vma->vm_private_data;
  99         struct ttm_bo_device *bdev = bo->bdev;
 100         unsigned long page_offset;
 101         unsigned long page_last;
 102         unsigned long pfn;
 103         struct ttm_tt *ttm = NULL;
 104         struct page *page;
 105         int ret;
 106         int i;
 107         unsigned long address = vmf->address;
 108         int retval = VM_FAULT_NOPAGE;
 109         struct ttm_mem_type_manager *man =
 110                 &bdev->man[bo->mem.mem_type];
 111         struct vm_area_struct cvma;
 112
 113         /*
 114          * Work around locking order reversal in fault / nopfn
 115          * between mmap_sem and bo_reserve: Perform a trylock operation
 116          * for reserve, and if it fails, retry the fault after waiting
 117          * for the buffer to become unreserved.
 118          */
 119         ret = ttm_bo_reserve(bo, true, true, NULL);
 120         if (unlikely(ret != 0)) {
 121                 if (ret != -EBUSY)
 122                         return VM_FAULT_NOPAGE;
 123
 124                 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
 125                         if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 126                                 ttm_bo_reference(bo);
 127                                 up_read(&vmf->vma->vm_mm->mmap_sem);
 128                                 (void) ttm_bo_wait_unreserved(bo);
 129                                 ttm_bo_unref(&bo);
 130                         }
 131
 132                         return VM_FAULT_RETRY;
 133                 }
 134
 135                 /*
 136                  * If we'd want to change locking order to
 137                  * mmap_sem -> bo::reserve, we'd use a blocking reserve here
 138                  * instead of retrying the fault...
 139                  */
 140                 return VM_FAULT_NOPAGE;
 141         }
 142
 143         /*
 144          * Refuse to fault imported pages. This should be handled
 145          * (if at all) by redirecting mmap to the exporter.
 146          */
 147         if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
 148                 retval = VM_FAULT_SIGBUS;
 149                 goto out_unlock;
 150         }
 151
 152         if (bdev->driver->fault_reserve_notify) {
 153                 ret = bdev->driver->fault_reserve_notify(bo);
 154                 switch (ret) {
 155                 case 0:
 156                         break;
 157                 case -EBUSY:
 158                 case -ERESTARTSYS:
 159                         retval = VM_FAULT_NOPAGE;
 160                         goto out_unlock;
 161                 default:
 162                         retval = VM_FAULT_SIGBUS;
 163                         goto out_unlock;
 164                 }
 165         }
 166
 167         /*
 168          * Wait for buffer data in transit, due to a pipelined
 169          * move.
 170          */
 171         ret = ttm_bo_vm_fault_idle(bo, vmf);
 172         if (unlikely(ret != 0)) {
 173                 retval = ret;
 174
 175                 if (retval == VM_FAULT_RETRY &&
 176                     !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 177                         /* The BO has already been unreserved. */
 178                         return retval;
 179                 }
 180
 181                 goto out_unlock;
 182         }
 183
 184         ret = ttm_mem_io_lock(man, true);
 185         if (unlikely(ret != 0)) {
 186                 retval = VM_FAULT_NOPAGE;
 187                 goto out_unlock;
 188         }
 189         ret = ttm_mem_io_reserve_vm(bo);
 190         if (unlikely(ret != 0)) {
 191                 retval = VM_FAULT_SIGBUS;
 192                 goto out_io_unlock;
 193         }
 194
 195         page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
 196                 vma->vm_pgoff - drm_vma_node_start(&bo->vma_node);
 197         page_last = vma_pages(vma) + vma->vm_pgoff -
 198                 drm_vma_node_start(&bo->vma_node);
 199
 200         if (unlikely(page_offset >= bo->num_pages)) {
 201                 retval = VM_FAULT_SIGBUS;
 202                 goto out_io_unlock;
 203         }
 204
 205         /*
 206          * Make a local vma copy to modify the page_prot member
 207          * and vm_flags if necessary. The vma parameter is protected
 208          * by mmap_sem in write mode.
 209          */
 210         cvma = *vma;
 211         cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
 212
 213         if (bo->mem.bus.is_iomem) {
 214                 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
 215                                                 cvma.vm_page_prot);
 216         } else {
 217                 ttm = bo->ttm;
 218                 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
 219                                                 cvma.vm_page_prot);
 220
 221                 /* Allocate all page at once, most common usage */
 222                 if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
 223                         retval = VM_FAULT_OOM;
 224                         goto out_io_unlock;
 225                 }
 226         }
 227
 228         /*
 229          * Speculatively prefault a number of pages. Only error on
 230          * first page.
 231          */
 232         for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
 233                 if (bo->mem.bus.is_iomem)
 234                         pfn = bdev->driver->io_mem_pfn(bo, page_offset);
 235                 else {
 236                         page = ttm->pages[page_offset];
 237                         if (unlikely(!page && i == 0)) {
 238                                 retval = VM_FAULT_OOM;
 239                                 goto out_io_unlock;
 240                         } else if (unlikely(!page)) {
 241                                 break;
 242                         }
 243                         page->mapping = vma->vm_file->f_mapping;
 244                         page->index = drm_vma_node_start(&bo->vma_node) +
 245                                 page_offset;
 246                         pfn = page_to_pfn(page);
 247                 }
 248
 249                 if (vma->vm_flags & VM_MIXEDMAP)
 250                         ret = vm_insert_mixed(&cvma, address,
 251                                         __pfn_to_pfn_t(pfn, PFN_DEV));
 252                 else
 253                         ret = vm_insert_pfn(&cvma, address, pfn);
 254
 255                 /*
 256                  * Somebody beat us to this PTE or prefaulting to
 257                  * an already populated PTE, or prefaulting error.
 258                  */
 259
 260                 if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0)))
 261                         break;
 262                 else if (unlikely(ret != 0)) {
 263                         retval =
 264                             (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
 265                         goto out_io_unlock;
 266                 }
 267
 268                 address += PAGE_SIZE;
 269                 if (unlikely(++page_offset >= page_last))
 270                         break;
 271         }
 272 out_io_unlock:
 273         ttm_mem_io_unlock(man);
 274 out_unlock:
 275         ttm_bo_unreserve(bo);
 276         return retval;
 277 }
 278
 279 static void ttm_bo_vm_open(struct vm_area_struct *vma)
 280 {
 281         struct ttm_buffer_object *bo =
 282             (struct ttm_buffer_object *)vma->vm_private_data;
 283
 284         WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
 285
 286         (void)ttm_bo_reference(bo);
 287 }
 288
 289 static void ttm_bo_vm_close(struct vm_area_struct *vma)
 290 {
 291         struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
 292
 293         ttm_bo_unref(&bo);
 294         vma->vm_private_data = NULL;
 295 }
 296
 297 static const struct vm_operations_struct ttm_bo_vm_ops = {
 298         .fault = ttm_bo_vm_fault,
 299         .open = ttm_bo_vm_open,
 300         .close = ttm_bo_vm_close
 301 };
 302
 303 static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
 304                                                   unsigned long offset,
 305                                                   unsigned long pages)
 306 {
 307         struct drm_vma_offset_node *node;
 308         struct ttm_buffer_object *bo = NULL;
 309
 310         drm_vma_offset_lock_lookup(&bdev->vma_manager);
 311
 312         node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
 313         if (likely(node)) {
 314                 bo = container_of(node, struct ttm_buffer_object, vma_node);
 315                 if (!kref_get_unless_zero(&bo->kref))
 316                         bo = NULL;
 317         }
 318
 319         drm_vma_offset_unlock_lookup(&bdev->vma_manager);
 320
 321         if (!bo)
 322                 pr_err("Could not find buffer object to map\n");
 323
 324         return bo;
 325 }
 326
 327 unsigned long ttm_bo_default_io_mem_pfn(struct ttm_buffer_object *bo,
 328                                         unsigned long page_offset)
 329 {
 330         return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
 331                 + page_offset;
 332 }
 333 EXPORT_SYMBOL(ttm_bo_default_io_mem_pfn);
 334
 335 int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
 336                 struct ttm_bo_device *bdev)
 337 {
 338         struct ttm_bo_driver *driver;
 339         struct ttm_buffer_object *bo;
 340         int ret;
 341
 342         bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
 343         if (unlikely(!bo))
 344                 return -EINVAL;
 345
 346         driver = bo->bdev->driver;
 347         if (unlikely(!driver->verify_access)) {
 348                 ret = -EPERM;
 349                 goto out_unref;
 350         }
 351         ret = driver->verify_access(bo, filp);
 352         if (unlikely(ret != 0))
 353                 goto out_unref;
 354
 355         vma->vm_ops = &ttm_bo_vm_ops;
 356
 357         /*
 358          * Note: We're transferring the bo reference to
 359          * vma->vm_private_data here.
 360          */
 361
 362         vma->vm_private_data = bo;
 363
 364         /*
 365          * We'd like to use VM_PFNMAP on shared mappings, where
 366          * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
 367          * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
 368          * bad for performance. Until that has been sorted out, use
 369          * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
 370          */
 371         vma->vm_flags |= VM_MIXEDMAP;
 372         vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
 373         return 0;
 374 out_unref:
 375         ttm_bo_unref(&bo);
 376         return ret;
 377 }
 378 EXPORT_SYMBOL(ttm_bo_mmap);
 379
 380 int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 381 {
 382         if (vma->vm_pgoff != 0)
 383                 return -EACCES;
 384
 385         vma->vm_ops = &ttm_bo_vm_ops;
 386         vma->vm_private_data = ttm_bo_reference(bo);
 387         vma->vm_flags |= VM_MIXEDMAP;
 388         vma->vm_flags |= VM_IO | VM_DONTEXPAND;
 389         return 0;
 390 }
 391 EXPORT_SYMBOL(ttm_fbdev_mmap);