2 * Copyright © 2012-2014 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "i915_trace.h"
29 #include "intel_drv.h"
30 #include <linux/mmu_context.h>
31 #include <linux/mmu_notifier.h>
32 #include <linux/mempolicy.h>
33 #include <linux/swap.h>
35 struct i915_mm_struct {
37 struct drm_device *dev;
38 struct i915_mmu_notifier *mn;
39 struct hlist_node node;
41 struct work_struct work;
44 #if defined(CONFIG_MMU_NOTIFIER)
45 #include <linux/interval_tree.h>
47 struct i915_mmu_notifier {
49 struct hlist_node node;
50 struct mmu_notifier mn;
51 struct rb_root objects;
52 struct list_head linear;
57 struct i915_mmu_object {
58 struct i915_mmu_notifier *mn;
59 struct interval_tree_node it;
60 struct list_head link;
61 struct drm_i915_gem_object *obj;
65 static unsigned long cancel_userptr(struct drm_i915_gem_object *obj)
67 struct drm_device *dev = obj->base.dev;
70 mutex_lock(&dev->struct_mutex);
71 /* Cancel any active worker and force us to re-evaluate gup */
72 obj->userptr.work = NULL;
74 if (obj->pages != NULL) {
75 struct drm_i915_private *dev_priv = to_i915(dev);
76 struct i915_vma *vma, *tmp;
77 bool was_interruptible;
79 was_interruptible = dev_priv->mm.interruptible;
80 dev_priv->mm.interruptible = false;
82 list_for_each_entry_safe(vma, tmp, &obj->vma_list, vma_link) {
83 int ret = i915_vma_unbind(vma);
84 WARN_ON(ret && ret != -EIO);
86 WARN_ON(i915_gem_object_put_pages(obj));
88 dev_priv->mm.interruptible = was_interruptible;
91 end = obj->userptr.ptr + obj->base.size;
93 drm_gem_object_unreference(&obj->base);
94 mutex_unlock(&dev->struct_mutex);
99 static void *invalidate_range__linear(struct i915_mmu_notifier *mn,
100 struct mm_struct *mm,
104 struct i915_mmu_object *mo;
105 unsigned long serial;
109 list_for_each_entry(mo, &mn->linear, link) {
110 struct drm_i915_gem_object *obj;
112 if (mo->it.last < start || mo->it.start > end)
116 drm_gem_object_reference(&obj->base);
117 spin_unlock(&mn->lock);
121 spin_lock(&mn->lock);
122 if (serial != mn->serial)
129 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
130 struct mm_struct *mm,
134 struct i915_mmu_notifier *mn = container_of(_mn, struct i915_mmu_notifier, mn);
135 struct interval_tree_node *it = NULL;
136 unsigned long next = start;
137 unsigned long serial = 0;
139 end--; /* interval ranges are inclusive, but invalidate range is exclusive */
141 struct drm_i915_gem_object *obj = NULL;
143 spin_lock(&mn->lock);
145 it = invalidate_range__linear(mn, mm, start, end);
146 else if (serial == mn->serial)
147 it = interval_tree_iter_next(it, next, end);
149 it = interval_tree_iter_first(&mn->objects, start, end);
151 obj = container_of(it, struct i915_mmu_object, it)->obj;
152 drm_gem_object_reference(&obj->base);
155 spin_unlock(&mn->lock);
159 next = cancel_userptr(obj);
163 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
164 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
167 static struct i915_mmu_notifier *
168 i915_mmu_notifier_create(struct mm_struct *mm)
170 struct i915_mmu_notifier *mn;
173 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
175 return ERR_PTR(-ENOMEM);
177 spin_lock_init(&mn->lock);
178 mn->mn.ops = &i915_gem_userptr_notifier;
179 mn->objects = RB_ROOT;
181 INIT_LIST_HEAD(&mn->linear);
182 mn->has_linear = false;
184 /* Protected by mmap_sem (write-lock) */
185 ret = __mmu_notifier_register(&mn->mn, mm);
194 static void __i915_mmu_notifier_update_serial(struct i915_mmu_notifier *mn)
196 if (++mn->serial == 0)
201 i915_mmu_notifier_add(struct drm_device *dev,
202 struct i915_mmu_notifier *mn,
203 struct i915_mmu_object *mo)
205 struct interval_tree_node *it;
208 ret = i915_mutex_lock_interruptible(dev);
212 /* Make sure we drop the final active reference (and thereby
213 * remove the objects from the interval tree) before we do
214 * the check for overlapping objects.
216 i915_gem_retire_requests(dev);
218 spin_lock(&mn->lock);
219 it = interval_tree_iter_first(&mn->objects,
220 mo->it.start, mo->it.last);
222 struct drm_i915_gem_object *obj;
224 /* We only need to check the first object in the range as it
225 * either has cancelled gup work queued and we need to
226 * return back to the user to give time for the gup-workers
227 * to flush their object references upon which the object will
228 * be removed from the interval-tree, or the the range is
229 * still in use by another client and the overlap is invalid.
231 * If we do have an overlap, we cannot use the interval tree
232 * for fast range invalidation.
235 obj = container_of(it, struct i915_mmu_object, it)->obj;
236 if (!obj->userptr.workers)
237 mn->has_linear = mo->is_linear = true;
241 interval_tree_insert(&mo->it, &mn->objects);
244 list_add(&mo->link, &mn->linear);
245 __i915_mmu_notifier_update_serial(mn);
247 spin_unlock(&mn->lock);
248 mutex_unlock(&dev->struct_mutex);
253 static bool i915_mmu_notifier_has_linear(struct i915_mmu_notifier *mn)
255 struct i915_mmu_object *mo;
257 list_for_each_entry(mo, &mn->linear, link)
265 i915_mmu_notifier_del(struct i915_mmu_notifier *mn,
266 struct i915_mmu_object *mo)
268 spin_lock(&mn->lock);
271 mn->has_linear = i915_mmu_notifier_has_linear(mn);
273 interval_tree_remove(&mo->it, &mn->objects);
274 __i915_mmu_notifier_update_serial(mn);
275 spin_unlock(&mn->lock);
279 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
281 struct i915_mmu_object *mo;
283 mo = obj->userptr.mmu_object;
287 i915_mmu_notifier_del(mo->mn, mo);
290 obj->userptr.mmu_object = NULL;
293 static struct i915_mmu_notifier *
294 i915_mmu_notifier_find(struct i915_mm_struct *mm)
296 struct i915_mmu_notifier *mn = mm->mn;
302 down_write(&mm->mm->mmap_sem);
303 mutex_lock(&to_i915(mm->dev)->mm_lock);
304 if ((mn = mm->mn) == NULL) {
305 mn = i915_mmu_notifier_create(mm->mm);
309 mutex_unlock(&to_i915(mm->dev)->mm_lock);
310 up_write(&mm->mm->mmap_sem);
316 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
319 struct i915_mmu_notifier *mn;
320 struct i915_mmu_object *mo;
323 if (flags & I915_USERPTR_UNSYNCHRONIZED)
324 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
326 if (WARN_ON(obj->userptr.mm == NULL))
329 mn = i915_mmu_notifier_find(obj->userptr.mm);
333 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
338 mo->it.start = obj->userptr.ptr;
339 mo->it.last = mo->it.start + obj->base.size - 1;
342 ret = i915_mmu_notifier_add(obj->base.dev, mn, mo);
348 obj->userptr.mmu_object = mo;
353 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
354 struct mm_struct *mm)
359 mmu_notifier_unregister(&mn->mn, mm);
366 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
371 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
374 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
377 if (!capable(CAP_SYS_ADMIN))
384 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
385 struct mm_struct *mm)
391 static struct i915_mm_struct *
392 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
394 struct i915_mm_struct *mm;
396 /* Protected by dev_priv->mm_lock */
397 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
405 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
407 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
408 struct i915_mm_struct *mm;
411 /* During release of the GEM object we hold the struct_mutex. This
412 * precludes us from calling mmput() at that time as that may be
413 * the last reference and so call exit_mmap(). exit_mmap() will
414 * attempt to reap the vma, and if we were holding a GTT mmap
415 * would then call drm_gem_vm_close() and attempt to reacquire
416 * the struct mutex. So in order to avoid that recursion, we have
417 * to defer releasing the mm reference until after we drop the
418 * struct_mutex, i.e. we need to schedule a worker to do the clean
421 mutex_lock(&dev_priv->mm_lock);
422 mm = __i915_mm_struct_find(dev_priv, current->mm);
424 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
430 kref_init(&mm->kref);
431 mm->dev = obj->base.dev;
433 mm->mm = current->mm;
434 atomic_inc(¤t->mm->mm_count);
438 /* Protected by dev_priv->mm_lock */
439 hash_add(dev_priv->mm_structs,
440 &mm->node, (unsigned long)mm->mm);
444 obj->userptr.mm = mm;
446 mutex_unlock(&dev_priv->mm_lock);
451 __i915_mm_struct_free__worker(struct work_struct *work)
453 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
454 i915_mmu_notifier_free(mm->mn, mm->mm);
460 __i915_mm_struct_free(struct kref *kref)
462 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
464 /* Protected by dev_priv->mm_lock */
466 mutex_unlock(&to_i915(mm->dev)->mm_lock);
468 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
469 schedule_work(&mm->work);
473 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
475 if (obj->userptr.mm == NULL)
478 kref_put_mutex(&obj->userptr.mm->kref,
479 __i915_mm_struct_free,
480 &to_i915(obj->base.dev)->mm_lock);
481 obj->userptr.mm = NULL;
484 struct get_pages_work {
485 struct work_struct work;
486 struct drm_i915_gem_object *obj;
487 struct task_struct *task;
490 #if IS_ENABLED(CONFIG_SWIOTLB)
491 #define swiotlb_active() swiotlb_nr_tbl()
493 #define swiotlb_active() 0
497 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
499 struct scatterlist *sg;
502 *st = kmalloc(sizeof(**st), GFP_KERNEL);
506 if (swiotlb_active()) {
507 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
511 for_each_sg((*st)->sgl, sg, num_pages, n)
512 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
514 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
515 0, num_pages << PAGE_SHIFT,
530 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
532 struct get_pages_work *work = container_of(_work, typeof(*work), work);
533 struct drm_i915_gem_object *obj = work->obj;
534 struct drm_device *dev = obj->base.dev;
535 const int num_pages = obj->base.size >> PAGE_SHIFT;
542 pvec = kmalloc(num_pages*sizeof(struct page *),
543 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
545 pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
547 struct mm_struct *mm = obj->userptr.mm->mm;
549 down_read(&mm->mmap_sem);
550 while (pinned < num_pages) {
551 ret = get_user_pages(work->task, mm,
552 obj->userptr.ptr + pinned * PAGE_SIZE,
554 !obj->userptr.read_only, 0,
555 pvec + pinned, NULL);
561 up_read(&mm->mmap_sem);
564 mutex_lock(&dev->struct_mutex);
565 if (obj->userptr.work != &work->work) {
567 } else if (pinned == num_pages) {
568 ret = st_set_pages(&obj->pages, pvec, num_pages);
570 list_add_tail(&obj->global_list, &to_i915(dev)->mm.unbound_list);
575 obj->userptr.work = ERR_PTR(ret);
576 obj->userptr.workers--;
577 drm_gem_object_unreference(&obj->base);
578 mutex_unlock(&dev->struct_mutex);
580 release_pages(pvec, pinned, 0);
581 drm_free_large(pvec);
583 put_task_struct(work->task);
588 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
590 const int num_pages = obj->base.size >> PAGE_SHIFT;
594 /* If userspace should engineer that these pages are replaced in
595 * the vma between us binding this page into the GTT and completion
596 * of rendering... Their loss. If they change the mapping of their
597 * pages they need to create a new bo to point to the new vma.
599 * However, that still leaves open the possibility of the vma
600 * being copied upon fork. Which falls under the same userspace
601 * synchronisation issue as a regular bo, except that this time
602 * the process may not be expecting that a particular piece of
603 * memory is tied to the GPU.
605 * Fortunately, we can hook into the mmu_notifier in order to
606 * discard the page references prior to anything nasty happening
607 * to the vma (discard or cloning) which should prevent the more
608 * egregious cases from causing harm.
613 if (obj->userptr.mm->mm == current->mm) {
614 pvec = kmalloc(num_pages*sizeof(struct page *),
615 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
617 pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
622 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
623 !obj->userptr.read_only, pvec);
625 if (pinned < num_pages) {
630 /* Spawn a worker so that we can acquire the
631 * user pages without holding our mutex. Access
632 * to the user pages requires mmap_sem, and we have
633 * a strict lock ordering of mmap_sem, struct_mutex -
634 * we already hold struct_mutex here and so cannot
635 * call gup without encountering a lock inversion.
637 * Userspace will keep on repeating the operation
638 * (thanks to EAGAIN) until either we hit the fast
639 * path or the worker completes. If the worker is
640 * cancelled or superseded, the task is still run
641 * but the results ignored. (This leads to
642 * complications that we may have a stray object
643 * refcount that we need to be wary of when
644 * checking for existing objects during creation.)
645 * If the worker encounters an error, it reports
646 * that error back to this function through
647 * obj->userptr.work = ERR_PTR.
650 if (obj->userptr.work == NULL &&
651 obj->userptr.workers < I915_GEM_USERPTR_MAX_WORKERS) {
652 struct get_pages_work *work;
654 work = kmalloc(sizeof(*work), GFP_KERNEL);
656 obj->userptr.work = &work->work;
657 obj->userptr.workers++;
660 drm_gem_object_reference(&obj->base);
662 work->task = current;
663 get_task_struct(work->task);
665 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
666 schedule_work(&work->work);
670 if (IS_ERR(obj->userptr.work)) {
671 ret = PTR_ERR(obj->userptr.work);
672 obj->userptr.work = NULL;
677 ret = st_set_pages(&obj->pages, pvec, num_pages);
679 obj->userptr.work = NULL;
684 release_pages(pvec, pinned, 0);
685 drm_free_large(pvec);
690 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj)
692 struct sg_page_iter sg_iter;
694 BUG_ON(obj->userptr.work != NULL);
696 if (obj->madv != I915_MADV_WILLNEED)
699 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
700 struct page *page = sg_page_iter_page(&sg_iter);
703 set_page_dirty(page);
705 mark_page_accessed(page);
706 page_cache_release(page);
710 sg_free_table(obj->pages);
715 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
717 i915_gem_userptr_release__mmu_notifier(obj);
718 i915_gem_userptr_release__mm_struct(obj);
722 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
724 if (obj->userptr.mmu_object)
727 return i915_gem_userptr_init__mmu_notifier(obj, 0);
730 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
731 .dmabuf_export = i915_gem_userptr_dmabuf_export,
732 .get_pages = i915_gem_userptr_get_pages,
733 .put_pages = i915_gem_userptr_put_pages,
734 .release = i915_gem_userptr_release,
738 * Creates a new mm object that wraps some normal memory from the process
739 * context - user memory.
741 * We impose several restrictions upon the memory being mapped
743 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
744 * 2. It must be normal system memory, not a pointer into another map of IO
745 * space (e.g. it must not be a GTT mmapping of another object).
746 * 3. We only allow a bo as large as we could in theory map into the GTT,
747 * that is we limit the size to the total size of the GTT.
748 * 4. The bo is marked as being snoopable. The backing pages are left
749 * accessible directly by the CPU, but reads and writes by the GPU may
750 * incur the cost of a snoop (unless you have an LLC architecture).
752 * Synchronisation between multiple users and the GPU is left to userspace
753 * through the normal set-domain-ioctl. The kernel will enforce that the
754 * GPU relinquishes the VMA before it is returned back to the system
755 * i.e. upon free(), munmap() or process termination. However, the userspace
756 * malloc() library may not immediately relinquish the VMA after free() and
757 * instead reuse it whilst the GPU is still reading and writing to the VMA.
760 * Also note, that the object created here is not currently a "first class"
761 * object, in that several ioctls are banned. These are the CPU access
762 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
763 * direct access via your pointer rather than use those ioctls.
765 * If you think this is a good interface to use to pass GPU memory between
766 * drivers, please use dma-buf instead. In fact, wherever possible use
770 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
772 struct drm_i915_private *dev_priv = dev->dev_private;
773 struct drm_i915_gem_userptr *args = data;
774 struct drm_i915_gem_object *obj;
778 if (args->flags & ~(I915_USERPTR_READ_ONLY |
779 I915_USERPTR_UNSYNCHRONIZED))
782 if (offset_in_page(args->user_ptr | args->user_size))
785 if (args->user_size > dev_priv->gtt.base.total)
788 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
789 (char __user *)(unsigned long)args->user_ptr, args->user_size))
792 if (args->flags & I915_USERPTR_READ_ONLY) {
793 /* On almost all of the current hw, we cannot tell the GPU that a
794 * page is readonly, so this is just a placeholder in the uAPI.
799 obj = i915_gem_object_alloc(dev);
803 drm_gem_private_object_init(dev, &obj->base, args->user_size);
804 i915_gem_object_init(obj, &i915_gem_userptr_ops);
805 obj->cache_level = I915_CACHE_LLC;
806 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
807 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
809 obj->userptr.ptr = args->user_ptr;
810 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
812 /* And keep a pointer to the current->mm for resolving the user pages
813 * at binding. This means that we need to hook into the mmu_notifier
814 * in order to detect if the mmu is destroyed.
816 ret = i915_gem_userptr_init__mm_struct(obj);
818 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
820 ret = drm_gem_handle_create(file, &obj->base, &handle);
822 /* drop reference from allocate - handle holds it now */
823 drm_gem_object_unreference_unlocked(&obj->base);
827 args->handle = handle;
832 i915_gem_init_userptr(struct drm_device *dev)
834 struct drm_i915_private *dev_priv = to_i915(dev);
835 mutex_init(&dev_priv->mm_lock);
836 hash_init(dev_priv->mm_structs);