root/drivers/gpu/drm/i915/gem/i915_gem_userptr.c

DEFINITIONS

1. add_object
2. del_object
3. __i915_gem_userptr_set_active
4. userptr_mn_invalidate_range_start
5. i915_mmu_notifier_create
6. i915_gem_userptr_release__mmu_notifier
7. i915_mmu_notifier_find
8. i915_gem_userptr_init__mmu_notifier
9. i915_mmu_notifier_free
10. __i915_gem_userptr_set_active
11. i915_gem_userptr_release__mmu_notifier
12. i915_gem_userptr_init__mmu_notifier
13. i915_mmu_notifier_free
14. __i915_mm_struct_find
15. i915_gem_userptr_init__mm_struct
16. __i915_mm_struct_free__worker
17. __i915_mm_struct_free
18. i915_gem_userptr_release__mm_struct
19. __i915_gem_userptr_alloc_pages
20. __i915_gem_userptr_get_pages_worker
21. __i915_gem_userptr_get_pages_schedule
22. i915_gem_userptr_get_pages
23. i915_gem_userptr_put_pages
24. i915_gem_userptr_release
25. i915_gem_userptr_dmabuf_export
26. i915_gem_userptr_ioctl
27. i915_gem_init_userptr
28. i915_gem_cleanup_userptr

   1 /*
   2  * SPDX-License-Identifier: MIT
   3  *
   4  * Copyright © 2012-2014 Intel Corporation
   5  */
   6 
   7 #include <linux/mmu_context.h>
   8 #include <linux/mmu_notifier.h>
   9 #include <linux/mempolicy.h>
  10 #include <linux/swap.h>
  11 #include <linux/sched/mm.h>
  12 
  13 #include <drm/i915_drm.h>
  14 
  15 #include "i915_drv.h"
  16 #include "i915_gem_ioctls.h"
  17 #include "i915_gem_object.h"
  18 #include "i915_scatterlist.h"
  19 
  20 struct i915_mm_struct {
  21         struct mm_struct *mm;
  22         struct drm_i915_private *i915;
  23         struct i915_mmu_notifier *mn;
  24         struct hlist_node node;
  25         struct kref kref;
  26         struct work_struct work;
  27 };
  28 
  29 #if defined(CONFIG_MMU_NOTIFIER)
  30 #include <linux/interval_tree.h>
  31 
  32 struct i915_mmu_notifier {
  33         spinlock_t lock;
  34         struct hlist_node node;
  35         struct mmu_notifier mn;
  36         struct rb_root_cached objects;
  37         struct i915_mm_struct *mm;
  38 };
  39 
  40 struct i915_mmu_object {
  41         struct i915_mmu_notifier *mn;
  42         struct drm_i915_gem_object *obj;
  43         struct interval_tree_node it;
  44 };
  45 
  46 static void add_object(struct i915_mmu_object *mo)
  47 {
  48         GEM_BUG_ON(!RB_EMPTY_NODE(&mo->it.rb));
  49         interval_tree_insert(&mo->it, &mo->mn->objects);
  50 }
  51 
  52 static void del_object(struct i915_mmu_object *mo)
  53 {
  54         if (RB_EMPTY_NODE(&mo->it.rb))
  55                 return;
  56 
  57         interval_tree_remove(&mo->it, &mo->mn->objects);
  58         RB_CLEAR_NODE(&mo->it.rb);
  59 }
  60 
  61 static void
  62 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj, bool value)
  63 {
  64         struct i915_mmu_object *mo = obj->userptr.mmu_object;
  65 
  66         /*
  67          * During mm_invalidate_range we need to cancel any userptr that
  68          * overlaps the range being invalidated. Doing so requires the
  69          * struct_mutex, and that risks recursion. In order to cause
  70          * recursion, the user must alias the userptr address space with
  71          * a GTT mmapping (possible with a MAP_FIXED) - then when we have
  72          * to invalidate that mmaping, mm_invalidate_range is called with
  73          * the userptr address *and* the struct_mutex held.  To prevent that
  74          * we set a flag under the i915_mmu_notifier spinlock to indicate
  75          * whether this object is valid.
  76          */
  77         if (!mo)
  78                 return;
  79 
  80         spin_lock(&mo->mn->lock);
  81         if (value)
  82                 add_object(mo);
  83         else
  84                 del_object(mo);
  85         spin_unlock(&mo->mn->lock);
  86 }
  87 
  88 static int
  89 userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
  90                                   const struct mmu_notifier_range *range)
  91 {
  92         struct i915_mmu_notifier *mn =
  93                 container_of(_mn, struct i915_mmu_notifier, mn);
  94         struct interval_tree_node *it;
  95         struct mutex *unlock = NULL;
  96         unsigned long end;
  97         int ret = 0;
  98 
  99         if (RB_EMPTY_ROOT(&mn->objects.rb_root))
 100                 return 0;
 101 
 102         /* interval ranges are inclusive, but invalidate range is exclusive */
 103         end = range->end - 1;
 104 
 105         spin_lock(&mn->lock);
 106         it = interval_tree_iter_first(&mn->objects, range->start, end);
 107         while (it) {
 108                 struct drm_i915_gem_object *obj;
 109 
 110                 if (!mmu_notifier_range_blockable(range)) {
 111                         ret = -EAGAIN;
 112                         break;
 113                 }
 114 
 115                 /*
 116                  * The mmu_object is released late when destroying the
 117                  * GEM object so it is entirely possible to gain a
 118                  * reference on an object in the process of being freed
 119                  * since our serialisation is via the spinlock and not
 120                  * the struct_mutex - and consequently use it after it
 121                  * is freed and then double free it. To prevent that
 122                  * use-after-free we only acquire a reference on the
 123                  * object if it is not in the process of being destroyed.
 124                  */
 125                 obj = container_of(it, struct i915_mmu_object, it)->obj;
 126                 if (!kref_get_unless_zero(&obj->base.refcount)) {
 127                         it = interval_tree_iter_next(it, range->start, end);
 128                         continue;
 129                 }
 130                 spin_unlock(&mn->lock);
 131 
 132                 if (!unlock) {
 133                         unlock = &mn->mm->i915->drm.struct_mutex;
 134 
 135                         switch (mutex_trylock_recursive(unlock)) {
 136                         default:
 137                         case MUTEX_TRYLOCK_FAILED:
 138                                 if (mutex_lock_killable_nested(unlock, I915_MM_SHRINKER)) {
 139                                         i915_gem_object_put(obj);
 140                                         return -EINTR;
 141                                 }
 142                                 /* fall through */
 143                         case MUTEX_TRYLOCK_SUCCESS:
 144                                 break;
 145 
 146                         case MUTEX_TRYLOCK_RECURSIVE:
 147                                 unlock = ERR_PTR(-EEXIST);
 148                                 break;
 149                         }
 150                 }
 151 
 152                 ret = i915_gem_object_unbind(obj,
 153                                              I915_GEM_OBJECT_UNBIND_ACTIVE);
 154                 if (ret == 0)
 155                         ret = __i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
 156                 i915_gem_object_put(obj);
 157                 if (ret)
 158                         goto unlock;
 159 
 160                 spin_lock(&mn->lock);
 161 
 162                 /*
 163                  * As we do not (yet) protect the mmu from concurrent insertion
 164                  * over this range, there is no guarantee that this search will
 165                  * terminate given a pathologic workload.
 166                  */
 167                 it = interval_tree_iter_first(&mn->objects, range->start, end);
 168         }
 169         spin_unlock(&mn->lock);
 170 
 171 unlock:
 172         if (!IS_ERR_OR_NULL(unlock))
 173                 mutex_unlock(unlock);
 174 
 175         return ret;
 176 
 177 }
 178 
 179 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
 180         .invalidate_range_start = userptr_mn_invalidate_range_start,
 181 };
 182 
 183 static struct i915_mmu_notifier *
 184 i915_mmu_notifier_create(struct i915_mm_struct *mm)
 185 {
 186         struct i915_mmu_notifier *mn;
 187 
 188         mn = kmalloc(sizeof(*mn), GFP_KERNEL);
 189         if (mn == NULL)
 190                 return ERR_PTR(-ENOMEM);
 191 
 192         spin_lock_init(&mn->lock);
 193         mn->mn.ops = &i915_gem_userptr_notifier;
 194         mn->objects = RB_ROOT_CACHED;
 195         mn->mm = mm;
 196 
 197         return mn;
 198 }
 199 
 200 static void
 201 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
 202 {
 203         struct i915_mmu_object *mo;
 204 
 205         mo = fetch_and_zero(&obj->userptr.mmu_object);
 206         if (!mo)
 207                 return;
 208 
 209         spin_lock(&mo->mn->lock);
 210         del_object(mo);
 211         spin_unlock(&mo->mn->lock);
 212         kfree(mo);
 213 }
 214 
 215 static struct i915_mmu_notifier *
 216 i915_mmu_notifier_find(struct i915_mm_struct *mm)
 217 {
 218         struct i915_mmu_notifier *mn;
 219         int err = 0;
 220 
 221         mn = mm->mn;
 222         if (mn)
 223                 return mn;
 224 
 225         mn = i915_mmu_notifier_create(mm);
 226         if (IS_ERR(mn))
 227                 err = PTR_ERR(mn);
 228 
 229         down_write(&mm->mm->mmap_sem);
 230         mutex_lock(&mm->i915->mm_lock);
 231         if (mm->mn == NULL && !err) {
 232                 /* Protected by mmap_sem (write-lock) */
 233                 err = __mmu_notifier_register(&mn->mn, mm->mm);
 234                 if (!err) {
 235                         /* Protected by mm_lock */
 236                         mm->mn = fetch_and_zero(&mn);
 237                 }
 238         } else if (mm->mn) {
 239                 /*
 240                  * Someone else raced and successfully installed the mmu
 241                  * notifier, we can cancel our own errors.
 242                  */
 243                 err = 0;
 244         }
 245         mutex_unlock(&mm->i915->mm_lock);
 246         up_write(&mm->mm->mmap_sem);
 247 
 248         if (mn && !IS_ERR(mn))
 249                 kfree(mn);
 250 
 251         return err ? ERR_PTR(err) : mm->mn;
 252 }
 253 
 254 static int
 255 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
 256                                     unsigned flags)
 257 {
 258         struct i915_mmu_notifier *mn;
 259         struct i915_mmu_object *mo;
 260 
 261         if (flags & I915_USERPTR_UNSYNCHRONIZED)
 262                 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
 263 
 264         if (WARN_ON(obj->userptr.mm == NULL))
 265                 return -EINVAL;
 266 
 267         mn = i915_mmu_notifier_find(obj->userptr.mm);
 268         if (IS_ERR(mn))
 269                 return PTR_ERR(mn);
 270 
 271         mo = kzalloc(sizeof(*mo), GFP_KERNEL);
 272         if (!mo)
 273                 return -ENOMEM;
 274 
 275         mo->mn = mn;
 276         mo->obj = obj;
 277         mo->it.start = obj->userptr.ptr;
 278         mo->it.last = obj->userptr.ptr + obj->base.size - 1;
 279         RB_CLEAR_NODE(&mo->it.rb);
 280 
 281         obj->userptr.mmu_object = mo;
 282         return 0;
 283 }
 284 
 285 static void
 286 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
 287                        struct mm_struct *mm)
 288 {
 289         if (mn == NULL)
 290                 return;
 291 
 292         mmu_notifier_unregister(&mn->mn, mm);
 293         kfree(mn);
 294 }
 295 
 296 #else
 297 
 298 static void
 299 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj, bool value)
 300 {
 301 }
 302 
 303 static void
 304 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
 305 {
 306 }
 307 
 308 static int
 309 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
 310                                     unsigned flags)
 311 {
 312         if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
 313                 return -ENODEV;
 314 
 315         if (!capable(CAP_SYS_ADMIN))
 316                 return -EPERM;
 317 
 318         return 0;
 319 }
 320 
 321 static void
 322 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
 323                        struct mm_struct *mm)
 324 {
 325 }
 326 
 327 #endif
 328 
 329 static struct i915_mm_struct *
 330 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
 331 {
 332         struct i915_mm_struct *mm;
 333 
 334         /* Protected by dev_priv->mm_lock */
 335         hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
 336                 if (mm->mm == real)
 337                         return mm;
 338 
 339         return NULL;
 340 }
 341 
 342 static int
 343 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
 344 {
 345         struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
 346         struct i915_mm_struct *mm;
 347         int ret = 0;
 348 
 349         /* During release of the GEM object we hold the struct_mutex. This
 350          * precludes us from calling mmput() at that time as that may be
 351          * the last reference and so call exit_mmap(). exit_mmap() will
 352          * attempt to reap the vma, and if we were holding a GTT mmap
 353          * would then call drm_gem_vm_close() and attempt to reacquire
 354          * the struct mutex. So in order to avoid that recursion, we have
 355          * to defer releasing the mm reference until after we drop the
 356          * struct_mutex, i.e. we need to schedule a worker to do the clean
 357          * up.
 358          */
 359         mutex_lock(&dev_priv->mm_lock);
 360         mm = __i915_mm_struct_find(dev_priv, current->mm);
 361         if (mm == NULL) {
 362                 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
 363                 if (mm == NULL) {
 364                         ret = -ENOMEM;
 365                         goto out;
 366                 }
 367 
 368                 kref_init(&mm->kref);
 369                 mm->i915 = to_i915(obj->base.dev);
 370 
 371                 mm->mm = current->mm;
 372                 mmgrab(current->mm);
 373 
 374                 mm->mn = NULL;
 375 
 376                 /* Protected by dev_priv->mm_lock */
 377                 hash_add(dev_priv->mm_structs,
 378                          &mm->node, (unsigned long)mm->mm);
 379         } else
 380                 kref_get(&mm->kref);
 381 
 382         obj->userptr.mm = mm;
 383 out:
 384         mutex_unlock(&dev_priv->mm_lock);
 385         return ret;
 386 }
 387 
 388 static void
 389 __i915_mm_struct_free__worker(struct work_struct *work)
 390 {
 391         struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
 392         i915_mmu_notifier_free(mm->mn, mm->mm);
 393         mmdrop(mm->mm);
 394         kfree(mm);
 395 }
 396 
 397 static void
 398 __i915_mm_struct_free(struct kref *kref)
 399 {
 400         struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
 401 
 402         /* Protected by dev_priv->mm_lock */
 403         hash_del(&mm->node);
 404         mutex_unlock(&mm->i915->mm_lock);
 405 
 406         INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
 407         queue_work(mm->i915->mm.userptr_wq, &mm->work);
 408 }
 409 
 410 static void
 411 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
 412 {
 413         if (obj->userptr.mm == NULL)
 414                 return;
 415 
 416         kref_put_mutex(&obj->userptr.mm->kref,
 417                        __i915_mm_struct_free,
 418                        &to_i915(obj->base.dev)->mm_lock);
 419         obj->userptr.mm = NULL;
 420 }
 421 
 422 struct get_pages_work {
 423         struct work_struct work;
 424         struct drm_i915_gem_object *obj;
 425         struct task_struct *task;
 426 };
 427 
 428 static struct sg_table *
 429 __i915_gem_userptr_alloc_pages(struct drm_i915_gem_object *obj,
 430                                struct page **pvec, unsigned long num_pages)
 431 {
 432         unsigned int max_segment = i915_sg_segment_size();
 433         struct sg_table *st;
 434         unsigned int sg_page_sizes;
 435         int ret;
 436 
 437         st = kmalloc(sizeof(*st), GFP_KERNEL);
 438         if (!st)
 439                 return ERR_PTR(-ENOMEM);
 440 
 441 alloc_table:
 442         ret = __sg_alloc_table_from_pages(st, pvec, num_pages,
 443                                           0, num_pages << PAGE_SHIFT,
 444                                           max_segment,
 445                                           GFP_KERNEL);
 446         if (ret) {
 447                 kfree(st);
 448                 return ERR_PTR(ret);
 449         }
 450 
 451         ret = i915_gem_gtt_prepare_pages(obj, st);
 452         if (ret) {
 453                 sg_free_table(st);
 454 
 455                 if (max_segment > PAGE_SIZE) {
 456                         max_segment = PAGE_SIZE;
 457                         goto alloc_table;
 458                 }
 459 
 460                 kfree(st);
 461                 return ERR_PTR(ret);
 462         }
 463 
 464         sg_page_sizes = i915_sg_page_sizes(st->sgl);
 465 
 466         __i915_gem_object_set_pages(obj, st, sg_page_sizes);
 467 
 468         return st;
 469 }
 470 
 471 static void
 472 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
 473 {
 474         struct get_pages_work *work = container_of(_work, typeof(*work), work);
 475         struct drm_i915_gem_object *obj = work->obj;
 476         const unsigned long npages = obj->base.size >> PAGE_SHIFT;
 477         unsigned long pinned;
 478         struct page **pvec;
 479         int ret;
 480 
 481         ret = -ENOMEM;
 482         pinned = 0;
 483 
 484         pvec = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
 485         if (pvec != NULL) {
 486                 struct mm_struct *mm = obj->userptr.mm->mm;
 487                 unsigned int flags = 0;
 488 
 489                 if (!i915_gem_object_is_readonly(obj))
 490                         flags |= FOLL_WRITE;
 491 
 492                 ret = -EFAULT;
 493                 if (mmget_not_zero(mm)) {
 494                         down_read(&mm->mmap_sem);
 495                         while (pinned < npages) {
 496                                 ret = get_user_pages_remote
 497                                         (work->task, mm,
 498                                          obj->userptr.ptr + pinned * PAGE_SIZE,
 499                                          npages - pinned,
 500                                          flags,
 501                                          pvec + pinned, NULL, NULL);
 502                                 if (ret < 0)
 503                                         break;
 504 
 505                                 pinned += ret;
 506                         }
 507                         up_read(&mm->mmap_sem);
 508                         mmput(mm);
 509                 }
 510         }
 511 
 512         mutex_lock(&obj->mm.lock);
 513         if (obj->userptr.work == &work->work) {
 514                 struct sg_table *pages = ERR_PTR(ret);
 515 
 516                 if (pinned == npages) {
 517                         pages = __i915_gem_userptr_alloc_pages(obj, pvec,
 518                                                                npages);
 519                         if (!IS_ERR(pages)) {
 520                                 pinned = 0;
 521                                 pages = NULL;
 522                         }
 523                 }
 524 
 525                 obj->userptr.work = ERR_CAST(pages);
 526                 if (IS_ERR(pages))
 527                         __i915_gem_userptr_set_active(obj, false);
 528         }
 529         mutex_unlock(&obj->mm.lock);
 530 
 531         release_pages(pvec, pinned);
 532         kvfree(pvec);
 533 
 534         i915_gem_object_put(obj);
 535         put_task_struct(work->task);
 536         kfree(work);
 537 }
 538 
 539 static struct sg_table *
 540 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj)
 541 {
 542         struct get_pages_work *work;
 543 
 544         /* Spawn a worker so that we can acquire the
 545          * user pages without holding our mutex. Access
 546          * to the user pages requires mmap_sem, and we have
 547          * a strict lock ordering of mmap_sem, struct_mutex -
 548          * we already hold struct_mutex here and so cannot
 549          * call gup without encountering a lock inversion.
 550          *
 551          * Userspace will keep on repeating the operation
 552          * (thanks to EAGAIN) until either we hit the fast
 553          * path or the worker completes. If the worker is
 554          * cancelled or superseded, the task is still run
 555          * but the results ignored. (This leads to
 556          * complications that we may have a stray object
 557          * refcount that we need to be wary of when
 558          * checking for existing objects during creation.)
 559          * If the worker encounters an error, it reports
 560          * that error back to this function through
 561          * obj->userptr.work = ERR_PTR.
 562          */
 563         work = kmalloc(sizeof(*work), GFP_KERNEL);
 564         if (work == NULL)
 565                 return ERR_PTR(-ENOMEM);
 566 
 567         obj->userptr.work = &work->work;
 568 
 569         work->obj = i915_gem_object_get(obj);
 570 
 571         work->task = current;
 572         get_task_struct(work->task);
 573 
 574         INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
 575         queue_work(to_i915(obj->base.dev)->mm.userptr_wq, &work->work);
 576 
 577         return ERR_PTR(-EAGAIN);
 578 }
 579 
 580 static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
 581 {
 582         const unsigned long num_pages = obj->base.size >> PAGE_SHIFT;
 583         struct mm_struct *mm = obj->userptr.mm->mm;
 584         struct page **pvec;
 585         struct sg_table *pages;
 586         bool active;
 587         int pinned;
 588 
 589         /* If userspace should engineer that these pages are replaced in
 590          * the vma between us binding this page into the GTT and completion
 591          * of rendering... Their loss. If they change the mapping of their
 592          * pages they need to create a new bo to point to the new vma.
 593          *
 594          * However, that still leaves open the possibility of the vma
 595          * being copied upon fork. Which falls under the same userspace
 596          * synchronisation issue as a regular bo, except that this time
 597          * the process may not be expecting that a particular piece of
 598          * memory is tied to the GPU.
 599          *
 600          * Fortunately, we can hook into the mmu_notifier in order to
 601          * discard the page references prior to anything nasty happening
 602          * to the vma (discard or cloning) which should prevent the more
 603          * egregious cases from causing harm.
 604          */
 605 
 606         if (obj->userptr.work) {
 607                 /* active flag should still be held for the pending work */
 608                 if (IS_ERR(obj->userptr.work))
 609                         return PTR_ERR(obj->userptr.work);
 610                 else
 611                         return -EAGAIN;
 612         }
 613 
 614         pvec = NULL;
 615         pinned = 0;
 616 
 617         if (mm == current->mm) {
 618                 pvec = kvmalloc_array(num_pages, sizeof(struct page *),
 619                                       GFP_KERNEL |
 620                                       __GFP_NORETRY |
 621                                       __GFP_NOWARN);
 622                 /*
 623                  * Using __get_user_pages_fast() with a read-only
 624                  * access is questionable. A read-only page may be
 625                  * COW-broken, and then this might end up giving
 626                  * the wrong side of the COW..
 627                  *
 628                  * We may or may not care.
 629                  */
 630                 if (pvec) /* defer to worker if malloc fails */
 631                         pinned = __get_user_pages_fast(obj->userptr.ptr,
 632                                                        num_pages,
 633                                                        !i915_gem_object_is_readonly(obj),
 634                                                        pvec);
 635         }
 636 
 637         active = false;
 638         if (pinned < 0) {
 639                 pages = ERR_PTR(pinned);
 640                 pinned = 0;
 641         } else if (pinned < num_pages) {
 642                 pages = __i915_gem_userptr_get_pages_schedule(obj);
 643                 active = pages == ERR_PTR(-EAGAIN);
 644         } else {
 645                 pages = __i915_gem_userptr_alloc_pages(obj, pvec, num_pages);
 646                 active = !IS_ERR(pages);
 647         }
 648         if (active)
 649                 __i915_gem_userptr_set_active(obj, true);
 650 
 651         if (IS_ERR(pages))
 652                 release_pages(pvec, pinned);
 653         kvfree(pvec);
 654 
 655         return PTR_ERR_OR_ZERO(pages);
 656 }
 657 
 658 static void
 659 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
 660                            struct sg_table *pages)
 661 {
 662         struct sgt_iter sgt_iter;
 663         struct page *page;
 664 
 665         /* Cancel any inflight work and force them to restart their gup */
 666         obj->userptr.work = NULL;
 667         __i915_gem_userptr_set_active(obj, false);
 668         if (!pages)
 669                 return;
 670 
 671         __i915_gem_object_release_shmem(obj, pages, true);
 672         i915_gem_gtt_finish_pages(obj, pages);
 673 
 674         /*
 675          * We always mark objects as dirty when they are used by the GPU,
 676          * just in case. However, if we set the vma as being read-only we know
 677          * that the object will never have been written to.
 678          */
 679         if (i915_gem_object_is_readonly(obj))
 680                 obj->mm.dirty = false;
 681 
 682         for_each_sgt_page(page, sgt_iter, pages) {
 683                 if (obj->mm.dirty && trylock_page(page)) {
 684                         /*
 685                          * As this may not be anonymous memory (e.g. shmem)
 686                          * but exist on a real mapping, we have to lock
 687                          * the page in order to dirty it -- holding
 688                          * the page reference is not sufficient to
 689                          * prevent the inode from being truncated.
 690                          * Play safe and take the lock.
 691                          *
 692                          * However...!
 693                          *
 694                          * The mmu-notifier can be invalidated for a
 695                          * migrate_page, that is alreadying holding the lock
 696                          * on the page. Such a try_to_unmap() will result
 697                          * in us calling put_pages() and so recursively try
 698                          * to lock the page. We avoid that deadlock with
 699                          * a trylock_page() and in exchange we risk missing
 700                          * some page dirtying.
 701                          */
 702                         set_page_dirty(page);
 703                         unlock_page(page);
 704                 }
 705 
 706                 mark_page_accessed(page);
 707                 put_page(page);
 708         }
 709         obj->mm.dirty = false;
 710 
 711         sg_free_table(pages);
 712         kfree(pages);
 713 }
 714 
 715 static void
 716 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
 717 {
 718         i915_gem_userptr_release__mmu_notifier(obj);
 719         i915_gem_userptr_release__mm_struct(obj);
 720 }
 721 
 722 static int
 723 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
 724 {
 725         if (obj->userptr.mmu_object)
 726                 return 0;
 727 
 728         return i915_gem_userptr_init__mmu_notifier(obj, 0);
 729 }
 730 
 731 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
 732         .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
 733                  I915_GEM_OBJECT_IS_SHRINKABLE |
 734                  I915_GEM_OBJECT_NO_GGTT |
 735                  I915_GEM_OBJECT_ASYNC_CANCEL,
 736         .get_pages = i915_gem_userptr_get_pages,
 737         .put_pages = i915_gem_userptr_put_pages,
 738         .dmabuf_export = i915_gem_userptr_dmabuf_export,
 739         .release = i915_gem_userptr_release,
 740 };
 741 
 742 /*
 743  * Creates a new mm object that wraps some normal memory from the process
 744  * context - user memory.
 745  *
 746  * We impose several restrictions upon the memory being mapped
 747  * into the GPU.
 748  * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
 749  * 2. It must be normal system memory, not a pointer into another map of IO
 750  *    space (e.g. it must not be a GTT mmapping of another object).
 751  * 3. We only allow a bo as large as we could in theory map into the GTT,
 752  *    that is we limit the size to the total size of the GTT.
 753  * 4. The bo is marked as being snoopable. The backing pages are left
 754  *    accessible directly by the CPU, but reads and writes by the GPU may
 755  *    incur the cost of a snoop (unless you have an LLC architecture).
 756  *
 757  * Synchronisation between multiple users and the GPU is left to userspace
 758  * through the normal set-domain-ioctl. The kernel will enforce that the
 759  * GPU relinquishes the VMA before it is returned back to the system
 760  * i.e. upon free(), munmap() or process termination. However, the userspace
 761  * malloc() library may not immediately relinquish the VMA after free() and
 762  * instead reuse it whilst the GPU is still reading and writing to the VMA.
 763  * Caveat emptor.
 764  *
 765  * Also note, that the object created here is not currently a "first class"
 766  * object, in that several ioctls are banned. These are the CPU access
 767  * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
 768  * direct access via your pointer rather than use those ioctls. Another
 769  * restriction is that we do not allow userptr surfaces to be pinned to the
 770  * hardware and so we reject any attempt to create a framebuffer out of a
 771  * userptr.
 772  *
 773  * If you think this is a good interface to use to pass GPU memory between
 774  * drivers, please use dma-buf instead. In fact, wherever possible use
 775  * dma-buf instead.
 776  */
 777 int
 778 i915_gem_userptr_ioctl(struct drm_device *dev,
 779                        void *data,
 780                        struct drm_file *file)
 781 {
 782         struct drm_i915_private *dev_priv = to_i915(dev);
 783         struct drm_i915_gem_userptr *args = data;
 784         struct drm_i915_gem_object *obj;
 785         int ret;
 786         u32 handle;
 787 
 788         if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
 789                 /* We cannot support coherent userptr objects on hw without
 790                  * LLC and broken snooping.
 791                  */
 792                 return -ENODEV;
 793         }
 794 
 795         if (args->flags & ~(I915_USERPTR_READ_ONLY |
 796                             I915_USERPTR_UNSYNCHRONIZED))
 797                 return -EINVAL;
 798 
 799         if (!args->user_size)
 800                 return -EINVAL;
 801 
 802         if (offset_in_page(args->user_ptr | args->user_size))
 803                 return -EINVAL;
 804 
 805         if (!access_ok((char __user *)(unsigned long)args->user_ptr, args->user_size))
 806                 return -EFAULT;
 807 
 808         if (args->flags & I915_USERPTR_READ_ONLY) {
 809                 struct i915_address_space *vm;
 810 
 811                 /*
 812                  * On almost all of the older hw, we cannot tell the GPU that
 813                  * a page is readonly.
 814                  */
 815                 vm = dev_priv->kernel_context->vm;
 816                 if (!vm || !vm->has_read_only)
 817                         return -ENODEV;
 818         }
 819 
 820         obj = i915_gem_object_alloc();
 821         if (obj == NULL)
 822                 return -ENOMEM;
 823 
 824         drm_gem_private_object_init(dev, &obj->base, args->user_size);
 825         i915_gem_object_init(obj, &i915_gem_userptr_ops);
 826         obj->read_domains = I915_GEM_DOMAIN_CPU;
 827         obj->write_domain = I915_GEM_DOMAIN_CPU;
 828         i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
 829 
 830         obj->userptr.ptr = args->user_ptr;
 831         if (args->flags & I915_USERPTR_READ_ONLY)
 832                 i915_gem_object_set_readonly(obj);
 833 
 834         /* And keep a pointer to the current->mm for resolving the user pages
 835          * at binding. This means that we need to hook into the mmu_notifier
 836          * in order to detect if the mmu is destroyed.
 837          */
 838         ret = i915_gem_userptr_init__mm_struct(obj);
 839         if (ret == 0)
 840                 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
 841         if (ret == 0)
 842                 ret = drm_gem_handle_create(file, &obj->base, &handle);
 843 
 844         /* drop reference from allocate - handle holds it now */
 845         i915_gem_object_put(obj);
 846         if (ret)
 847                 return ret;
 848 
 849         args->handle = handle;
 850         return 0;
 851 }
 852 
 853 int i915_gem_init_userptr(struct drm_i915_private *dev_priv)
 854 {
 855         mutex_init(&dev_priv->mm_lock);
 856         hash_init(dev_priv->mm_structs);
 857 
 858         dev_priv->mm.userptr_wq =
 859                 alloc_workqueue("i915-userptr-acquire",
 860                                 WQ_HIGHPRI | WQ_UNBOUND,
 861                                 0);
 862         if (!dev_priv->mm.userptr_wq)
 863                 return -ENOMEM;
 864 
 865         return 0;
 866 }
 867 
 868 void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv)
 869 {
 870         destroy_workqueue(dev_priv->mm.userptr_wq);
 871 }