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_bo_driver.h>
  35 #include <drm/ttm/ttm_placement.h>
  36 #include <drm/drm_vma_manager.h>
  37 #include <drm/drm_drv.h>
  38 #include <drm/drm_managed.h>
  39 #include <linux/mm.h>
  40 #include <linux/pfn_t.h>
  41 #include <linux/rbtree.h>
  42 #include <linux/module.h>
  43 #include <linux/uaccess.h>
  44 #include <linux/mem_encrypt.h>
  45
  46 static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
  47                                 struct vm_fault *vmf)
  48 {
  49         vm_fault_t ret = 0;
  50         int err = 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_lock
  63          * held.  We only do this if the fault allows retry and this
  64          * is the first attempt.
  65          */
  66         if (fault_flag_allow_retry_first(vmf->flags)) {
  67                 ret = VM_FAULT_RETRY;
  68                 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
  69                         goto out_unlock;
  70
  71                 ttm_bo_get(bo);
  72                 mmap_read_unlock(vmf->vma->vm_mm);
  73                 (void) dma_fence_wait(bo->moving, true);
  74                 dma_resv_unlock(bo->base.resv);
  75                 ttm_bo_put(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_device *bdev = bo->bdev;
 101
 102         if (bdev->funcs->io_mem_pfn)
 103                 return bdev->funcs->io_mem_pfn(bo, page_offset);
 104
 105         return (bo->resource->bus.offset >> PAGE_SHIFT) + page_offset;
 106 }
 107
 108 /**
 109  * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
 110  * @bo: The buffer object
 111  * @vmf: The fault structure handed to the callback
 112  *
 113  * vm callbacks like fault() and *_mkwrite() allow for the mm_sem to be dropped
 114  * during long waits, and after the wait the callback will be restarted. This
 115  * is to allow other threads using the same virtual memory space concurrent
 116  * access to map(), unmap() completely unrelated buffer objects. TTM buffer
 117  * object reservations sometimes wait for GPU and should therefore be
 118  * considered long waits. This function reserves the buffer object interruptibly
 119  * taking this into account. Starvation is avoided by the vm system not
 120  * allowing too many repeated restarts.
 121  * This function is intended to be used in customized fault() and _mkwrite()
 122  * handlers.
 123  *
 124  * Return:
 125  *    0 on success and the bo was reserved.
 126  *    VM_FAULT_RETRY if blocking wait.
 127  *    VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
 128  */
 129 vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
 130                              struct vm_fault *vmf)
 131 {
 132         /*
 133          * Work around locking order reversal in fault / nopfn
 134          * between mmap_lock and bo_reserve: Perform a trylock operation
 135          * for reserve, and if it fails, retry the fault after waiting
 136          * for the buffer to become unreserved.
 137          */
 138         if (unlikely(!dma_resv_trylock(bo->base.resv))) {
 139                 /*
 140                  * If the fault allows retry and this is the first
 141                  * fault attempt, we try to release the mmap_lock
 142                  * before waiting
 143                  */
 144                 if (fault_flag_allow_retry_first(vmf->flags)) {
 145                         if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
 146                                 ttm_bo_get(bo);
 147                                 mmap_read_unlock(vmf->vma->vm_mm);
 148                                 if (!dma_resv_lock_interruptible(bo->base.resv,
 149                                                                  NULL))
 150                                         dma_resv_unlock(bo->base.resv);
 151                                 ttm_bo_put(bo);
 152                         }
 153
 154                         return VM_FAULT_RETRY;
 155                 }
 156
 157                 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
 158                         return VM_FAULT_NOPAGE;
 159         }
 160
 161         /*
 162          * Refuse to fault imported pages. This should be handled
 163          * (if at all) by redirecting mmap to the exporter.
 164          */
 165         if (bo->ttm && (bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL)) {
 166                 if (!(bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL_MAPPABLE)) {
 167                         dma_resv_unlock(bo->base.resv);
 168                         return VM_FAULT_SIGBUS;
 169                 }
 170         }
 171
 172         return 0;
 173 }
 174 EXPORT_SYMBOL(ttm_bo_vm_reserve);
 175
 176 /**
 177  * ttm_bo_vm_fault_reserved - TTM fault helper
 178  * @vmf: The struct vm_fault given as argument to the fault callback
 179  * @prot: The page protection to be used for this memory area.
 180  * @num_prefault: Maximum number of prefault pages. The caller may want to
 181  * specify this based on madvice settings and the size of the GPU object
 182  * backed by the memory.
 183  *
 184  * This function inserts one or more page table entries pointing to the
 185  * memory backing the buffer object, and then returns a return code
 186  * instructing the caller to retry the page access.
 187  *
 188  * Return:
 189  *   VM_FAULT_NOPAGE on success or pending signal
 190  *   VM_FAULT_SIGBUS on unspecified error
 191  *   VM_FAULT_OOM on out-of-memory
 192  *   VM_FAULT_RETRY if retryable wait
 193  */
 194 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
 195                                     pgprot_t prot,
 196                                     pgoff_t num_prefault)
 197 {
 198         struct vm_area_struct *vma = vmf->vma;
 199         struct ttm_buffer_object *bo = vma->vm_private_data;
 200         struct ttm_device *bdev = bo->bdev;
 201         unsigned long page_offset;
 202         unsigned long page_last;
 203         unsigned long pfn;
 204         struct ttm_tt *ttm = NULL;
 205         struct page *page;
 206         int err;
 207         pgoff_t i;
 208         vm_fault_t ret = VM_FAULT_NOPAGE;
 209         unsigned long address = vmf->address;
 210
 211         /*
 212          * Wait for buffer data in transit, due to a pipelined
 213          * move.
 214          */
 215         ret = ttm_bo_vm_fault_idle(bo, vmf);
 216         if (unlikely(ret != 0))
 217                 return ret;
 218
 219         err = ttm_mem_io_reserve(bdev, bo->resource);
 220         if (unlikely(err != 0))
 221                 return VM_FAULT_SIGBUS;
 222
 223         page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
 224                 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
 225         page_last = vma_pages(vma) + vma->vm_pgoff -
 226                 drm_vma_node_start(&bo->base.vma_node);
 227
 228         if (unlikely(page_offset >= bo->resource->num_pages))
 229                 return VM_FAULT_SIGBUS;
 230
 231         prot = ttm_io_prot(bo, bo->resource, prot);
 232         if (!bo->resource->bus.is_iomem) {
 233                 struct ttm_operation_ctx ctx = {
 234                         .interruptible = false,
 235                         .no_wait_gpu = false,
 236                         .force_alloc = true
 237                 };
 238
 239                 ttm = bo->ttm;
 240                 if (ttm_tt_populate(bdev, bo->ttm, &ctx))
 241                         return VM_FAULT_OOM;
 242         } else {
 243                 /* Iomem should not be marked encrypted */
 244                 prot = pgprot_decrypted(prot);
 245         }
 246
 247         /*
 248          * Speculatively prefault a number of pages. Only error on
 249          * first page.
 250          */
 251         for (i = 0; i < num_prefault; ++i) {
 252                 if (bo->resource->bus.is_iomem) {
 253                         pfn = ttm_bo_io_mem_pfn(bo, page_offset);
 254                 } else {
 255                         page = ttm->pages[page_offset];
 256                         if (unlikely(!page && i == 0)) {
 257                                 return VM_FAULT_OOM;
 258                         } else if (unlikely(!page)) {
 259                                 break;
 260                         }
 261                         pfn = page_to_pfn(page);
 262                 }
 263
 264                 /*
 265                  * Note that the value of @prot at this point may differ from
 266                  * the value of @vma->vm_page_prot in the caching- and
 267                  * encryption bits. This is because the exact location of the
 268                  * data may not be known at mmap() time and may also change
 269                  * at arbitrary times while the data is mmap'ed.
 270                  * See vmf_insert_mixed_prot() for a discussion.
 271                  */
 272                 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
 273
 274                 /* Never error on prefaulted PTEs */
 275                 if (unlikely((ret & VM_FAULT_ERROR))) {
 276                         if (i == 0)
 277                                 return VM_FAULT_NOPAGE;
 278                         else
 279                                 break;
 280                 }
 281
 282                 address += PAGE_SIZE;
 283                 if (unlikely(++page_offset >= page_last))
 284                         break;
 285         }
 286         return ret;
 287 }
 288 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
 289
 290 static void ttm_bo_release_dummy_page(struct drm_device *dev, void *res)
 291 {
 292         struct page *dummy_page = (struct page *)res;
 293
 294         __free_page(dummy_page);
 295 }
 296
 297 vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot)
 298 {
 299         struct vm_area_struct *vma = vmf->vma;
 300         struct ttm_buffer_object *bo = vma->vm_private_data;
 301         struct drm_device *ddev = bo->base.dev;
 302         vm_fault_t ret = VM_FAULT_NOPAGE;
 303         unsigned long address;
 304         unsigned long pfn;
 305         struct page *page;
 306
 307         /* Allocate new dummy page to map all the VA range in this VMA to it*/
 308         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 309         if (!page)
 310                 return VM_FAULT_OOM;
 311
 312         /* Set the page to be freed using drmm release action */
 313         if (drmm_add_action_or_reset(ddev, ttm_bo_release_dummy_page, page))
 314                 return VM_FAULT_OOM;
 315
 316         pfn = page_to_pfn(page);
 317
 318         /* Prefault the entire VMA range right away to avoid further faults */
 319         for (address = vma->vm_start; address < vma->vm_end;
 320              address += PAGE_SIZE)
 321                 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
 322
 323         return ret;
 324 }
 325 EXPORT_SYMBOL(ttm_bo_vm_dummy_page);
 326
 327 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
 328 {
 329         struct vm_area_struct *vma = vmf->vma;
 330         pgprot_t prot;
 331         struct ttm_buffer_object *bo = vma->vm_private_data;
 332         struct drm_device *ddev = bo->base.dev;
 333         vm_fault_t ret;
 334         int idx;
 335
 336         ret = ttm_bo_vm_reserve(bo, vmf);
 337         if (ret)
 338                 return ret;
 339
 340         prot = vma->vm_page_prot;
 341         if (drm_dev_enter(ddev, &idx)) {
 342                 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
 343                 drm_dev_exit(idx);
 344         } else {
 345                 ret = ttm_bo_vm_dummy_page(vmf, prot);
 346         }
 347         if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
 348                 return ret;
 349
 350         dma_resv_unlock(bo->base.resv);
 351
 352         return ret;
 353 }
 354 EXPORT_SYMBOL(ttm_bo_vm_fault);
 355
 356 void ttm_bo_vm_open(struct vm_area_struct *vma)
 357 {
 358         struct ttm_buffer_object *bo = vma->vm_private_data;
 359
 360         WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
 361
 362         ttm_bo_get(bo);
 363 }
 364 EXPORT_SYMBOL(ttm_bo_vm_open);
 365
 366 void ttm_bo_vm_close(struct vm_area_struct *vma)
 367 {
 368         struct ttm_buffer_object *bo = vma->vm_private_data;
 369
 370         ttm_bo_put(bo);
 371         vma->vm_private_data = NULL;
 372 }
 373 EXPORT_SYMBOL(ttm_bo_vm_close);
 374
 375 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
 376                                  unsigned long offset,
 377                                  uint8_t *buf, int len, int write)
 378 {
 379         unsigned long page = offset >> PAGE_SHIFT;
 380         unsigned long bytes_left = len;
 381         int ret;
 382
 383         /* Copy a page at a time, that way no extra virtual address
 384          * mapping is needed
 385          */
 386         offset -= page << PAGE_SHIFT;
 387         do {
 388                 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
 389                 struct ttm_bo_kmap_obj map;
 390                 void *ptr;
 391                 bool is_iomem;
 392
 393                 ret = ttm_bo_kmap(bo, page, 1, &map);
 394                 if (ret)
 395                         return ret;
 396
 397                 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
 398                 WARN_ON_ONCE(is_iomem);
 399                 if (write)
 400                         memcpy(ptr, buf, bytes);
 401                 else
 402                         memcpy(buf, ptr, bytes);
 403                 ttm_bo_kunmap(&map);
 404
 405                 page++;
 406                 buf += bytes;
 407                 bytes_left -= bytes;
 408                 offset = 0;
 409         } while (bytes_left);
 410
 411         return len;
 412 }
 413
 414 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
 415                      void *buf, int len, int write)
 416 {
 417         struct ttm_buffer_object *bo = vma->vm_private_data;
 418         unsigned long offset = (addr) - vma->vm_start +
 419                 ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
 420                  << PAGE_SHIFT);
 421         int ret;
 422
 423         if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->resource->num_pages)
 424                 return -EIO;
 425
 426         ret = ttm_bo_reserve(bo, true, false, NULL);
 427         if (ret)
 428                 return ret;
 429
 430         switch (bo->resource->mem_type) {
 431         case TTM_PL_SYSTEM:
 432                 fallthrough;
 433         case TTM_PL_TT:
 434                 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
 435                 break;
 436         default:
 437                 if (bo->bdev->funcs->access_memory)
 438                         ret = bo->bdev->funcs->access_memory(
 439                                 bo, offset, buf, len, write);
 440                 else
 441                         ret = -EIO;
 442         }
 443
 444         ttm_bo_unreserve(bo);
 445
 446         return ret;
 447 }
 448 EXPORT_SYMBOL(ttm_bo_vm_access);
 449
 450 static const struct vm_operations_struct ttm_bo_vm_ops = {
 451         .fault = ttm_bo_vm_fault,
 452         .open = ttm_bo_vm_open,
 453         .close = ttm_bo_vm_close,
 454         .access = ttm_bo_vm_access,
 455 };
 456
 457 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
 458 {
 459         /* Enforce no COW since would have really strange behavior with it. */
 460         if (is_cow_mapping(vma->vm_flags))
 461                 return -EINVAL;
 462
 463         ttm_bo_get(bo);
 464
 465         /*
 466          * Drivers may want to override the vm_ops field. Otherwise we
 467          * use TTM's default callbacks.
 468          */
 469         if (!vma->vm_ops)
 470                 vma->vm_ops = &ttm_bo_vm_ops;
 471
 472         /*
 473          * Note: We're transferring the bo reference to
 474          * vma->vm_private_data here.
 475          */
 476
 477         vma->vm_private_data = bo;
 478
 479         vma->vm_flags |= VM_PFNMAP;
 480         vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
 481         return 0;
 482 }
 483 EXPORT_SYMBOL(ttm_bo_mmap_obj);