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",
176 WQ_UNBOUND | WQ_MEM_RECLAIM,
178 if (mn->wq == NULL) {
180 return ERR_PTR(-ENOMEM);
187 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
189 struct i915_mmu_object *mo;
191 mo = obj->userptr.mmu_object;
195 spin_lock(&mo->mn->lock);
197 spin_unlock(&mo->mn->lock);
200 obj->userptr.mmu_object = NULL;
203 static struct i915_mmu_notifier *
204 i915_mmu_notifier_find(struct i915_mm_struct *mm)
206 struct i915_mmu_notifier *mn;
213 mn = i915_mmu_notifier_create(mm->mm);
217 down_write(&mm->mm->mmap_sem);
218 mutex_lock(&mm->i915->mm_lock);
219 if (mm->mn == NULL && !err) {
220 /* Protected by mmap_sem (write-lock) */
221 err = __mmu_notifier_register(&mn->mn, mm->mm);
223 /* Protected by mm_lock */
224 mm->mn = fetch_and_zero(&mn);
228 * Someone else raced and successfully installed the mmu
229 * notifier, we can cancel our own errors.
233 mutex_unlock(&mm->i915->mm_lock);
234 up_write(&mm->mm->mmap_sem);
236 if (mn && !IS_ERR(mn)) {
237 destroy_workqueue(mn->wq);
241 return err ? ERR_PTR(err) : mm->mn;
245 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
248 struct i915_mmu_notifier *mn;
249 struct i915_mmu_object *mo;
251 if (flags & I915_USERPTR_UNSYNCHRONIZED)
252 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
254 if (WARN_ON(obj->userptr.mm == NULL))
257 mn = i915_mmu_notifier_find(obj->userptr.mm);
261 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
267 mo->it.start = obj->userptr.ptr;
268 mo->it.last = obj->userptr.ptr + obj->base.size - 1;
269 INIT_WORK(&mo->work, cancel_userptr);
271 obj->userptr.mmu_object = mo;
276 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
277 struct mm_struct *mm)
282 mmu_notifier_unregister(&mn->mn, mm);
283 destroy_workqueue(mn->wq);
290 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
295 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
298 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
301 if (!capable(CAP_SYS_ADMIN))
308 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
309 struct mm_struct *mm)
315 static struct i915_mm_struct *
316 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
318 struct i915_mm_struct *mm;
320 /* Protected by dev_priv->mm_lock */
321 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
329 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
331 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
332 struct i915_mm_struct *mm;
335 /* During release of the GEM object we hold the struct_mutex. This
336 * precludes us from calling mmput() at that time as that may be
337 * the last reference and so call exit_mmap(). exit_mmap() will
338 * attempt to reap the vma, and if we were holding a GTT mmap
339 * would then call drm_gem_vm_close() and attempt to reacquire
340 * the struct mutex. So in order to avoid that recursion, we have
341 * to defer releasing the mm reference until after we drop the
342 * struct_mutex, i.e. we need to schedule a worker to do the clean
345 mutex_lock(&dev_priv->mm_lock);
346 mm = __i915_mm_struct_find(dev_priv, current->mm);
348 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
354 kref_init(&mm->kref);
355 mm->i915 = to_i915(obj->base.dev);
357 mm->mm = current->mm;
362 /* Protected by dev_priv->mm_lock */
363 hash_add(dev_priv->mm_structs,
364 &mm->node, (unsigned long)mm->mm);
368 obj->userptr.mm = mm;
370 mutex_unlock(&dev_priv->mm_lock);
375 __i915_mm_struct_free__worker(struct work_struct *work)
377 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
378 i915_mmu_notifier_free(mm->mn, mm->mm);
384 __i915_mm_struct_free(struct kref *kref)
386 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
388 /* Protected by dev_priv->mm_lock */
390 mutex_unlock(&mm->i915->mm_lock);
392 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
393 queue_work(mm->i915->mm.userptr_wq, &mm->work);
397 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
399 if (obj->userptr.mm == NULL)
402 kref_put_mutex(&obj->userptr.mm->kref,
403 __i915_mm_struct_free,
404 &to_i915(obj->base.dev)->mm_lock);
405 obj->userptr.mm = NULL;
408 struct get_pages_work {
409 struct work_struct work;
410 struct drm_i915_gem_object *obj;
411 struct task_struct *task;
414 static struct sg_table *
415 __i915_gem_userptr_alloc_pages(struct drm_i915_gem_object *obj,
416 struct page **pvec, int num_pages)
418 unsigned int max_segment = i915_sg_segment_size();
420 unsigned int sg_page_sizes;
423 st = kmalloc(sizeof(*st), GFP_KERNEL);
425 return ERR_PTR(-ENOMEM);
428 ret = __sg_alloc_table_from_pages(st, pvec, num_pages,
429 0, num_pages << PAGE_SHIFT,
437 ret = i915_gem_gtt_prepare_pages(obj, st);
441 if (max_segment > PAGE_SIZE) {
442 max_segment = PAGE_SIZE;
450 sg_page_sizes = i915_sg_page_sizes(st->sgl);
452 __i915_gem_object_set_pages(obj, st, sg_page_sizes);
458 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
463 /* During mm_invalidate_range we need to cancel any userptr that
464 * overlaps the range being invalidated. Doing so requires the
465 * struct_mutex, and that risks recursion. In order to cause
466 * recursion, the user must alias the userptr address space with
467 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
468 * to invalidate that mmaping, mm_invalidate_range is called with
469 * the userptr address *and* the struct_mutex held. To prevent that
470 * we set a flag under the i915_mmu_notifier spinlock to indicate
471 * whether this object is valid.
473 #if defined(CONFIG_MMU_NOTIFIER)
474 if (obj->userptr.mmu_object == NULL)
477 spin_lock(&obj->userptr.mmu_object->mn->lock);
478 /* In order to serialise get_pages with an outstanding
479 * cancel_userptr, we must drop the struct_mutex and try again.
482 del_object(obj->userptr.mmu_object);
483 else if (!work_pending(&obj->userptr.mmu_object->work))
484 add_object(obj->userptr.mmu_object);
487 spin_unlock(&obj->userptr.mmu_object->mn->lock);
494 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
496 struct get_pages_work *work = container_of(_work, typeof(*work), work);
497 struct drm_i915_gem_object *obj = work->obj;
498 const int npages = obj->base.size >> PAGE_SHIFT;
505 pvec = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
507 struct mm_struct *mm = obj->userptr.mm->mm;
508 unsigned int flags = 0;
510 if (!i915_gem_object_is_readonly(obj))
514 if (mmget_not_zero(mm)) {
515 down_read(&mm->mmap_sem);
516 while (pinned < npages) {
517 ret = get_user_pages_remote
519 obj->userptr.ptr + pinned * PAGE_SIZE,
522 pvec + pinned, NULL, NULL);
528 up_read(&mm->mmap_sem);
533 mutex_lock(&obj->mm.lock);
534 if (obj->userptr.work == &work->work) {
535 struct sg_table *pages = ERR_PTR(ret);
537 if (pinned == npages) {
538 pages = __i915_gem_userptr_alloc_pages(obj, pvec,
540 if (!IS_ERR(pages)) {
546 obj->userptr.work = ERR_CAST(pages);
548 __i915_gem_userptr_set_active(obj, false);
550 mutex_unlock(&obj->mm.lock);
552 release_pages(pvec, pinned);
555 i915_gem_object_put(obj);
556 put_task_struct(work->task);
560 static struct sg_table *
561 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj)
563 struct get_pages_work *work;
565 /* Spawn a worker so that we can acquire the
566 * user pages without holding our mutex. Access
567 * to the user pages requires mmap_sem, and we have
568 * a strict lock ordering of mmap_sem, struct_mutex -
569 * we already hold struct_mutex here and so cannot
570 * call gup without encountering a lock inversion.
572 * Userspace will keep on repeating the operation
573 * (thanks to EAGAIN) until either we hit the fast
574 * path or the worker completes. If the worker is
575 * cancelled or superseded, the task is still run
576 * but the results ignored. (This leads to
577 * complications that we may have a stray object
578 * refcount that we need to be wary of when
579 * checking for existing objects during creation.)
580 * If the worker encounters an error, it reports
581 * that error back to this function through
582 * obj->userptr.work = ERR_PTR.
584 work = kmalloc(sizeof(*work), GFP_KERNEL);
586 return ERR_PTR(-ENOMEM);
588 obj->userptr.work = &work->work;
590 work->obj = i915_gem_object_get(obj);
592 work->task = current;
593 get_task_struct(work->task);
595 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
596 queue_work(to_i915(obj->base.dev)->mm.userptr_wq, &work->work);
598 return ERR_PTR(-EAGAIN);
601 static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
603 const int num_pages = obj->base.size >> PAGE_SHIFT;
604 struct mm_struct *mm = obj->userptr.mm->mm;
606 struct sg_table *pages;
610 /* If userspace should engineer that these pages are replaced in
611 * the vma between us binding this page into the GTT and completion
612 * of rendering... Their loss. If they change the mapping of their
613 * pages they need to create a new bo to point to the new vma.
615 * However, that still leaves open the possibility of the vma
616 * being copied upon fork. Which falls under the same userspace
617 * synchronisation issue as a regular bo, except that this time
618 * the process may not be expecting that a particular piece of
619 * memory is tied to the GPU.
621 * Fortunately, we can hook into the mmu_notifier in order to
622 * discard the page references prior to anything nasty happening
623 * to the vma (discard or cloning) which should prevent the more
624 * egregious cases from causing harm.
627 if (obj->userptr.work) {
628 /* active flag should still be held for the pending work */
629 if (IS_ERR(obj->userptr.work))
630 return PTR_ERR(obj->userptr.work);
638 if (mm == current->mm) {
639 pvec = kvmalloc_array(num_pages, sizeof(struct page *),
643 if (pvec) /* defer to worker if malloc fails */
644 pinned = __get_user_pages_fast(obj->userptr.ptr,
646 !i915_gem_object_is_readonly(obj),
652 pages = ERR_PTR(pinned);
654 } else if (pinned < num_pages) {
655 pages = __i915_gem_userptr_get_pages_schedule(obj);
656 active = pages == ERR_PTR(-EAGAIN);
658 pages = __i915_gem_userptr_alloc_pages(obj, pvec, num_pages);
659 active = !IS_ERR(pages);
662 __i915_gem_userptr_set_active(obj, true);
665 release_pages(pvec, pinned);
668 return PTR_ERR_OR_ZERO(pages);
672 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
673 struct sg_table *pages)
675 struct sgt_iter sgt_iter;
678 BUG_ON(obj->userptr.work != NULL);
679 __i915_gem_userptr_set_active(obj, false);
681 if (obj->mm.madv != I915_MADV_WILLNEED)
682 obj->mm.dirty = false;
684 i915_gem_gtt_finish_pages(obj, pages);
686 for_each_sgt_page(page, sgt_iter, pages) {
688 set_page_dirty(page);
690 mark_page_accessed(page);
693 obj->mm.dirty = false;
695 sg_free_table(pages);
700 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
702 i915_gem_userptr_release__mmu_notifier(obj);
703 i915_gem_userptr_release__mm_struct(obj);
707 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
709 if (obj->userptr.mmu_object)
712 return i915_gem_userptr_init__mmu_notifier(obj, 0);
715 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
716 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
717 I915_GEM_OBJECT_IS_SHRINKABLE,
718 .get_pages = i915_gem_userptr_get_pages,
719 .put_pages = i915_gem_userptr_put_pages,
720 .dmabuf_export = i915_gem_userptr_dmabuf_export,
721 .release = i915_gem_userptr_release,
725 * Creates a new mm object that wraps some normal memory from the process
726 * context - user memory.
728 * We impose several restrictions upon the memory being mapped
730 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
731 * 2. It must be normal system memory, not a pointer into another map of IO
732 * space (e.g. it must not be a GTT mmapping of another object).
733 * 3. We only allow a bo as large as we could in theory map into the GTT,
734 * that is we limit the size to the total size of the GTT.
735 * 4. The bo is marked as being snoopable. The backing pages are left
736 * accessible directly by the CPU, but reads and writes by the GPU may
737 * incur the cost of a snoop (unless you have an LLC architecture).
739 * Synchronisation between multiple users and the GPU is left to userspace
740 * through the normal set-domain-ioctl. The kernel will enforce that the
741 * GPU relinquishes the VMA before it is returned back to the system
742 * i.e. upon free(), munmap() or process termination. However, the userspace
743 * malloc() library may not immediately relinquish the VMA after free() and
744 * instead reuse it whilst the GPU is still reading and writing to the VMA.
747 * Also note, that the object created here is not currently a "first class"
748 * object, in that several ioctls are banned. These are the CPU access
749 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
750 * direct access via your pointer rather than use those ioctls. Another
751 * restriction is that we do not allow userptr surfaces to be pinned to the
752 * hardware and so we reject any attempt to create a framebuffer out of a
755 * If you think this is a good interface to use to pass GPU memory between
756 * drivers, please use dma-buf instead. In fact, wherever possible use
760 i915_gem_userptr_ioctl(struct drm_device *dev,
762 struct drm_file *file)
764 struct drm_i915_private *dev_priv = to_i915(dev);
765 struct drm_i915_gem_userptr *args = data;
766 struct drm_i915_gem_object *obj;
770 if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
771 /* We cannot support coherent userptr objects on hw without
772 * LLC and broken snooping.
777 if (args->flags & ~(I915_USERPTR_READ_ONLY |
778 I915_USERPTR_UNSYNCHRONIZED))
781 if (!args->user_size)
784 if (offset_in_page(args->user_ptr | args->user_size))
787 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
788 (char __user *)(unsigned long)args->user_ptr, args->user_size))
791 if (args->flags & I915_USERPTR_READ_ONLY) {
792 struct i915_hw_ppgtt *ppgtt;
795 * On almost all of the older hw, we cannot tell the GPU that
796 * a page is readonly.
798 ppgtt = dev_priv->kernel_context->ppgtt;
799 if (!ppgtt || !ppgtt->vm.has_read_only)
803 obj = i915_gem_object_alloc(dev_priv);
807 drm_gem_private_object_init(dev, &obj->base, args->user_size);
808 i915_gem_object_init(obj, &i915_gem_userptr_ops);
809 obj->read_domains = I915_GEM_DOMAIN_CPU;
810 obj->write_domain = I915_GEM_DOMAIN_CPU;
811 i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
813 obj->userptr.ptr = args->user_ptr;
814 if (args->flags & I915_USERPTR_READ_ONLY)
815 i915_gem_object_set_readonly(obj);
817 /* And keep a pointer to the current->mm for resolving the user pages
818 * at binding. This means that we need to hook into the mmu_notifier
819 * in order to detect if the mmu is destroyed.
821 ret = i915_gem_userptr_init__mm_struct(obj);
823 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
825 ret = drm_gem_handle_create(file, &obj->base, &handle);
827 /* drop reference from allocate - handle holds it now */
828 i915_gem_object_put(obj);
832 args->handle = handle;
836 int i915_gem_init_userptr(struct drm_i915_private *dev_priv)
838 mutex_init(&dev_priv->mm_lock);
839 hash_init(dev_priv->mm_structs);
841 dev_priv->mm.userptr_wq =
842 alloc_workqueue("i915-userptr-acquire",
843 WQ_HIGHPRI | WQ_UNBOUND,
845 if (!dev_priv->mm.userptr_wq)
851 void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv)
853 destroy_workqueue(dev_priv->mm.userptr_wq);