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
26 #include <drm/i915_drm.h>
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>
34 #include <linux/sched/mm.h>
36 struct i915_mm_struct {
38 struct drm_i915_private *i915;
39 struct i915_mmu_notifier *mn;
40 struct hlist_node node;
42 struct work_struct work;
45 #if defined(CONFIG_MMU_NOTIFIER)
46 #include <linux/interval_tree.h>
48 struct i915_mmu_notifier {
50 struct hlist_node node;
51 struct mmu_notifier mn;
52 struct rb_root_cached objects;
53 struct workqueue_struct *wq;
56 struct i915_mmu_object {
57 struct i915_mmu_notifier *mn;
58 struct drm_i915_gem_object *obj;
59 struct interval_tree_node it;
60 struct list_head link;
61 struct work_struct work;
65 static void cancel_userptr(struct work_struct *work)
67 struct i915_mmu_object *mo = container_of(work, typeof(*mo), work);
68 struct drm_i915_gem_object *obj = mo->obj;
69 struct work_struct *active;
71 /* Cancel any active worker and force us to re-evaluate gup */
72 mutex_lock(&obj->mm.lock);
73 active = fetch_and_zero(&obj->userptr.work);
74 mutex_unlock(&obj->mm.lock);
78 i915_gem_object_wait(obj, I915_WAIT_ALL, MAX_SCHEDULE_TIMEOUT, NULL);
80 mutex_lock(&obj->base.dev->struct_mutex);
82 /* We are inside a kthread context and can't be interrupted */
83 if (i915_gem_object_unbind(obj) == 0)
84 __i915_gem_object_put_pages(obj, I915_MM_NORMAL);
85 WARN_ONCE(i915_gem_object_has_pages(obj),
86 "Failed to release pages: bind_count=%d, pages_pin_count=%d, pin_global=%d\n",
88 atomic_read(&obj->mm.pages_pin_count),
91 mutex_unlock(&obj->base.dev->struct_mutex);
94 i915_gem_object_put(obj);
97 static void add_object(struct i915_mmu_object *mo)
102 interval_tree_insert(&mo->it, &mo->mn->objects);
106 static void del_object(struct i915_mmu_object *mo)
111 interval_tree_remove(&mo->it, &mo->mn->objects);
112 mo->attached = false;
115 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
116 struct mm_struct *mm,
120 struct i915_mmu_notifier *mn =
121 container_of(_mn, struct i915_mmu_notifier, mn);
122 struct i915_mmu_object *mo;
123 struct interval_tree_node *it;
124 LIST_HEAD(cancelled);
126 if (RB_EMPTY_ROOT(&mn->objects.rb_root))
129 /* interval ranges are inclusive, but invalidate range is exclusive */
132 spin_lock(&mn->lock);
133 it = interval_tree_iter_first(&mn->objects, start, end);
135 /* The mmu_object is released late when destroying the
136 * GEM object so it is entirely possible to gain a
137 * reference on an object in the process of being freed
138 * since our serialisation is via the spinlock and not
139 * the struct_mutex - and consequently use it after it
140 * is freed and then double free it. To prevent that
141 * use-after-free we only acquire a reference on the
142 * object if it is not in the process of being destroyed.
144 mo = container_of(it, struct i915_mmu_object, it);
145 if (kref_get_unless_zero(&mo->obj->base.refcount))
146 queue_work(mn->wq, &mo->work);
148 list_add(&mo->link, &cancelled);
149 it = interval_tree_iter_next(it, start, end);
151 list_for_each_entry(mo, &cancelled, link)
153 spin_unlock(&mn->lock);
155 if (!list_empty(&cancelled))
156 flush_workqueue(mn->wq);
159 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
160 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
163 static struct i915_mmu_notifier *
164 i915_mmu_notifier_create(struct mm_struct *mm)
166 struct i915_mmu_notifier *mn;
168 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
170 return ERR_PTR(-ENOMEM);
172 spin_lock_init(&mn->lock);
173 mn->mn.ops = &i915_gem_userptr_notifier;
174 mn->objects = RB_ROOT_CACHED;
175 mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0);
176 if (mn->wq == NULL) {
178 return ERR_PTR(-ENOMEM);
185 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
187 struct i915_mmu_object *mo;
189 mo = obj->userptr.mmu_object;
193 spin_lock(&mo->mn->lock);
195 spin_unlock(&mo->mn->lock);
198 obj->userptr.mmu_object = NULL;
201 static struct i915_mmu_notifier *
202 i915_mmu_notifier_find(struct i915_mm_struct *mm)
204 struct i915_mmu_notifier *mn;
211 mn = i915_mmu_notifier_create(mm->mm);
215 down_write(&mm->mm->mmap_sem);
216 mutex_lock(&mm->i915->mm_lock);
217 if (mm->mn == NULL && !err) {
218 /* Protected by mmap_sem (write-lock) */
219 err = __mmu_notifier_register(&mn->mn, mm->mm);
221 /* Protected by mm_lock */
222 mm->mn = fetch_and_zero(&mn);
226 * Someone else raced and successfully installed the mmu
227 * notifier, we can cancel our own errors.
231 mutex_unlock(&mm->i915->mm_lock);
232 up_write(&mm->mm->mmap_sem);
234 if (mn && !IS_ERR(mn)) {
235 destroy_workqueue(mn->wq);
239 return err ? ERR_PTR(err) : mm->mn;
243 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
246 struct i915_mmu_notifier *mn;
247 struct i915_mmu_object *mo;
249 if (flags & I915_USERPTR_UNSYNCHRONIZED)
250 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
252 if (WARN_ON(obj->userptr.mm == NULL))
255 mn = i915_mmu_notifier_find(obj->userptr.mm);
259 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
265 mo->it.start = obj->userptr.ptr;
266 mo->it.last = obj->userptr.ptr + obj->base.size - 1;
267 INIT_WORK(&mo->work, cancel_userptr);
269 obj->userptr.mmu_object = mo;
274 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
275 struct mm_struct *mm)
280 mmu_notifier_unregister(&mn->mn, mm);
281 destroy_workqueue(mn->wq);
288 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
293 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
296 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
299 if (!capable(CAP_SYS_ADMIN))
306 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
307 struct mm_struct *mm)
313 static struct i915_mm_struct *
314 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
316 struct i915_mm_struct *mm;
318 /* Protected by dev_priv->mm_lock */
319 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
327 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
329 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
330 struct i915_mm_struct *mm;
333 /* During release of the GEM object we hold the struct_mutex. This
334 * precludes us from calling mmput() at that time as that may be
335 * the last reference and so call exit_mmap(). exit_mmap() will
336 * attempt to reap the vma, and if we were holding a GTT mmap
337 * would then call drm_gem_vm_close() and attempt to reacquire
338 * the struct mutex. So in order to avoid that recursion, we have
339 * to defer releasing the mm reference until after we drop the
340 * struct_mutex, i.e. we need to schedule a worker to do the clean
343 mutex_lock(&dev_priv->mm_lock);
344 mm = __i915_mm_struct_find(dev_priv, current->mm);
346 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
352 kref_init(&mm->kref);
353 mm->i915 = to_i915(obj->base.dev);
355 mm->mm = current->mm;
360 /* Protected by dev_priv->mm_lock */
361 hash_add(dev_priv->mm_structs,
362 &mm->node, (unsigned long)mm->mm);
366 obj->userptr.mm = mm;
368 mutex_unlock(&dev_priv->mm_lock);
373 __i915_mm_struct_free__worker(struct work_struct *work)
375 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
376 i915_mmu_notifier_free(mm->mn, mm->mm);
382 __i915_mm_struct_free(struct kref *kref)
384 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
386 /* Protected by dev_priv->mm_lock */
388 mutex_unlock(&mm->i915->mm_lock);
390 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
391 queue_work(mm->i915->mm.userptr_wq, &mm->work);
395 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
397 if (obj->userptr.mm == NULL)
400 kref_put_mutex(&obj->userptr.mm->kref,
401 __i915_mm_struct_free,
402 &to_i915(obj->base.dev)->mm_lock);
403 obj->userptr.mm = NULL;
406 struct get_pages_work {
407 struct work_struct work;
408 struct drm_i915_gem_object *obj;
409 struct task_struct *task;
412 static struct sg_table *
413 __i915_gem_userptr_alloc_pages(struct drm_i915_gem_object *obj,
414 struct page **pvec, int num_pages)
416 unsigned int max_segment = i915_sg_segment_size();
418 unsigned int sg_page_sizes;
421 st = kmalloc(sizeof(*st), GFP_KERNEL);
423 return ERR_PTR(-ENOMEM);
426 ret = __sg_alloc_table_from_pages(st, pvec, num_pages,
427 0, num_pages << PAGE_SHIFT,
435 ret = i915_gem_gtt_prepare_pages(obj, st);
439 if (max_segment > PAGE_SIZE) {
440 max_segment = PAGE_SIZE;
448 sg_page_sizes = i915_sg_page_sizes(st->sgl);
450 __i915_gem_object_set_pages(obj, st, sg_page_sizes);
456 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
461 /* During mm_invalidate_range we need to cancel any userptr that
462 * overlaps the range being invalidated. Doing so requires the
463 * struct_mutex, and that risks recursion. In order to cause
464 * recursion, the user must alias the userptr address space with
465 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
466 * to invalidate that mmaping, mm_invalidate_range is called with
467 * the userptr address *and* the struct_mutex held. To prevent that
468 * we set a flag under the i915_mmu_notifier spinlock to indicate
469 * whether this object is valid.
471 #if defined(CONFIG_MMU_NOTIFIER)
472 if (obj->userptr.mmu_object == NULL)
475 spin_lock(&obj->userptr.mmu_object->mn->lock);
476 /* In order to serialise get_pages with an outstanding
477 * cancel_userptr, we must drop the struct_mutex and try again.
480 del_object(obj->userptr.mmu_object);
481 else if (!work_pending(&obj->userptr.mmu_object->work))
482 add_object(obj->userptr.mmu_object);
485 spin_unlock(&obj->userptr.mmu_object->mn->lock);
492 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
494 struct get_pages_work *work = container_of(_work, typeof(*work), work);
495 struct drm_i915_gem_object *obj = work->obj;
496 const int npages = obj->base.size >> PAGE_SHIFT;
503 pvec = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
505 struct mm_struct *mm = obj->userptr.mm->mm;
506 unsigned int flags = 0;
508 if (!obj->userptr.read_only)
512 if (mmget_not_zero(mm)) {
513 down_read(&mm->mmap_sem);
514 while (pinned < npages) {
515 ret = get_user_pages_remote
517 obj->userptr.ptr + pinned * PAGE_SIZE,
520 pvec + pinned, NULL, NULL);
526 up_read(&mm->mmap_sem);
531 mutex_lock(&obj->mm.lock);
532 if (obj->userptr.work == &work->work) {
533 struct sg_table *pages = ERR_PTR(ret);
535 if (pinned == npages) {
536 pages = __i915_gem_userptr_alloc_pages(obj, pvec,
538 if (!IS_ERR(pages)) {
544 obj->userptr.work = ERR_CAST(pages);
546 __i915_gem_userptr_set_active(obj, false);
548 mutex_unlock(&obj->mm.lock);
550 release_pages(pvec, pinned);
553 i915_gem_object_put(obj);
554 put_task_struct(work->task);
558 static struct sg_table *
559 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj)
561 struct get_pages_work *work;
563 /* Spawn a worker so that we can acquire the
564 * user pages without holding our mutex. Access
565 * to the user pages requires mmap_sem, and we have
566 * a strict lock ordering of mmap_sem, struct_mutex -
567 * we already hold struct_mutex here and so cannot
568 * call gup without encountering a lock inversion.
570 * Userspace will keep on repeating the operation
571 * (thanks to EAGAIN) until either we hit the fast
572 * path or the worker completes. If the worker is
573 * cancelled or superseded, the task is still run
574 * but the results ignored. (This leads to
575 * complications that we may have a stray object
576 * refcount that we need to be wary of when
577 * checking for existing objects during creation.)
578 * If the worker encounters an error, it reports
579 * that error back to this function through
580 * obj->userptr.work = ERR_PTR.
582 work = kmalloc(sizeof(*work), GFP_KERNEL);
584 return ERR_PTR(-ENOMEM);
586 obj->userptr.work = &work->work;
588 work->obj = i915_gem_object_get(obj);
590 work->task = current;
591 get_task_struct(work->task);
593 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
594 queue_work(to_i915(obj->base.dev)->mm.userptr_wq, &work->work);
596 return ERR_PTR(-EAGAIN);
599 static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
601 const int num_pages = obj->base.size >> PAGE_SHIFT;
602 struct mm_struct *mm = obj->userptr.mm->mm;
604 struct sg_table *pages;
608 /* If userspace should engineer that these pages are replaced in
609 * the vma between us binding this page into the GTT and completion
610 * of rendering... Their loss. If they change the mapping of their
611 * pages they need to create a new bo to point to the new vma.
613 * However, that still leaves open the possibility of the vma
614 * being copied upon fork. Which falls under the same userspace
615 * synchronisation issue as a regular bo, except that this time
616 * the process may not be expecting that a particular piece of
617 * memory is tied to the GPU.
619 * Fortunately, we can hook into the mmu_notifier in order to
620 * discard the page references prior to anything nasty happening
621 * to the vma (discard or cloning) which should prevent the more
622 * egregious cases from causing harm.
625 if (obj->userptr.work) {
626 /* active flag should still be held for the pending work */
627 if (IS_ERR(obj->userptr.work))
628 return PTR_ERR(obj->userptr.work);
636 if (mm == current->mm) {
637 pvec = kvmalloc_array(num_pages, sizeof(struct page *),
641 if (pvec) /* defer to worker if malloc fails */
642 pinned = __get_user_pages_fast(obj->userptr.ptr,
644 !obj->userptr.read_only,
650 pages = ERR_PTR(pinned);
652 } else if (pinned < num_pages) {
653 pages = __i915_gem_userptr_get_pages_schedule(obj);
654 active = pages == ERR_PTR(-EAGAIN);
656 pages = __i915_gem_userptr_alloc_pages(obj, pvec, num_pages);
657 active = !IS_ERR(pages);
660 __i915_gem_userptr_set_active(obj, true);
663 release_pages(pvec, pinned);
666 return PTR_ERR_OR_ZERO(pages);
670 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
671 struct sg_table *pages)
673 struct sgt_iter sgt_iter;
676 BUG_ON(obj->userptr.work != NULL);
677 __i915_gem_userptr_set_active(obj, false);
679 if (obj->mm.madv != I915_MADV_WILLNEED)
680 obj->mm.dirty = false;
682 i915_gem_gtt_finish_pages(obj, pages);
684 for_each_sgt_page(page, sgt_iter, pages) {
686 set_page_dirty(page);
688 mark_page_accessed(page);
691 obj->mm.dirty = false;
693 sg_free_table(pages);
698 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
700 i915_gem_userptr_release__mmu_notifier(obj);
701 i915_gem_userptr_release__mm_struct(obj);
705 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
707 if (obj->userptr.mmu_object)
710 return i915_gem_userptr_init__mmu_notifier(obj, 0);
713 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
714 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
715 I915_GEM_OBJECT_IS_SHRINKABLE,
716 .get_pages = i915_gem_userptr_get_pages,
717 .put_pages = i915_gem_userptr_put_pages,
718 .dmabuf_export = i915_gem_userptr_dmabuf_export,
719 .release = i915_gem_userptr_release,
723 * Creates a new mm object that wraps some normal memory from the process
724 * context - user memory.
726 * We impose several restrictions upon the memory being mapped
728 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
729 * 2. It must be normal system memory, not a pointer into another map of IO
730 * space (e.g. it must not be a GTT mmapping of another object).
731 * 3. We only allow a bo as large as we could in theory map into the GTT,
732 * that is we limit the size to the total size of the GTT.
733 * 4. The bo is marked as being snoopable. The backing pages are left
734 * accessible directly by the CPU, but reads and writes by the GPU may
735 * incur the cost of a snoop (unless you have an LLC architecture).
737 * Synchronisation between multiple users and the GPU is left to userspace
738 * through the normal set-domain-ioctl. The kernel will enforce that the
739 * GPU relinquishes the VMA before it is returned back to the system
740 * i.e. upon free(), munmap() or process termination. However, the userspace
741 * malloc() library may not immediately relinquish the VMA after free() and
742 * instead reuse it whilst the GPU is still reading and writing to the VMA.
745 * Also note, that the object created here is not currently a "first class"
746 * object, in that several ioctls are banned. These are the CPU access
747 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
748 * direct access via your pointer rather than use those ioctls. Another
749 * restriction is that we do not allow userptr surfaces to be pinned to the
750 * hardware and so we reject any attempt to create a framebuffer out of a
753 * If you think this is a good interface to use to pass GPU memory between
754 * drivers, please use dma-buf instead. In fact, wherever possible use
758 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
760 struct drm_i915_private *dev_priv = to_i915(dev);
761 struct drm_i915_gem_userptr *args = data;
762 struct drm_i915_gem_object *obj;
766 if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
767 /* We cannot support coherent userptr objects on hw without
768 * LLC and broken snooping.
773 if (args->flags & ~(I915_USERPTR_READ_ONLY |
774 I915_USERPTR_UNSYNCHRONIZED))
777 if (offset_in_page(args->user_ptr | args->user_size))
780 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
781 (char __user *)(unsigned long)args->user_ptr, args->user_size))
784 if (args->flags & I915_USERPTR_READ_ONLY) {
785 /* On almost all of the current hw, we cannot tell the GPU that a
786 * page is readonly, so this is just a placeholder in the uAPI.
791 obj = i915_gem_object_alloc(dev_priv);
795 drm_gem_private_object_init(dev, &obj->base, args->user_size);
796 i915_gem_object_init(obj, &i915_gem_userptr_ops);
797 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
798 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
799 i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
801 obj->userptr.ptr = args->user_ptr;
802 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
804 /* And keep a pointer to the current->mm for resolving the user pages
805 * at binding. This means that we need to hook into the mmu_notifier
806 * in order to detect if the mmu is destroyed.
808 ret = i915_gem_userptr_init__mm_struct(obj);
810 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
812 ret = drm_gem_handle_create(file, &obj->base, &handle);
814 /* drop reference from allocate - handle holds it now */
815 i915_gem_object_put(obj);
819 args->handle = handle;
823 int i915_gem_init_userptr(struct drm_i915_private *dev_priv)
825 mutex_init(&dev_priv->mm_lock);
826 hash_init(dev_priv->mm_structs);
828 dev_priv->mm.userptr_wq =
829 alloc_workqueue("i915-userptr-acquire",
830 WQ_HIGHPRI | WQ_MEM_RECLAIM,
832 if (!dev_priv->mm.userptr_wq)
838 void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv)
840 destroy_workqueue(dev_priv->mm.userptr_wq);