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 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 drm_device *dev = obj->base.dev;
71 i915_gem_object_wait(obj, I915_WAIT_ALL, MAX_SCHEDULE_TIMEOUT, NULL);
73 mutex_lock(&dev->struct_mutex);
74 /* Cancel any active worker and force us to re-evaluate gup */
75 obj->userptr.work = NULL;
77 /* We are inside a kthread context and can't be interrupted */
78 if (i915_gem_object_unbind(obj) == 0)
79 __i915_gem_object_put_pages(obj, I915_MM_NORMAL);
80 WARN_ONCE(obj->mm.pages,
81 "Failed to release pages: bind_count=%d, pages_pin_count=%d, pin_display=%d\n",
83 atomic_read(&obj->mm.pages_pin_count),
86 i915_gem_object_put(obj);
87 mutex_unlock(&dev->struct_mutex);
90 static void add_object(struct i915_mmu_object *mo)
95 interval_tree_insert(&mo->it, &mo->mn->objects);
99 static void del_object(struct i915_mmu_object *mo)
104 interval_tree_remove(&mo->it, &mo->mn->objects);
105 mo->attached = false;
108 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
109 struct mm_struct *mm,
113 struct i915_mmu_notifier *mn =
114 container_of(_mn, struct i915_mmu_notifier, mn);
115 struct i915_mmu_object *mo;
116 struct interval_tree_node *it;
117 LIST_HEAD(cancelled);
119 if (RB_EMPTY_ROOT(&mn->objects))
122 /* interval ranges are inclusive, but invalidate range is exclusive */
125 spin_lock(&mn->lock);
126 it = interval_tree_iter_first(&mn->objects, start, end);
128 /* The mmu_object is released late when destroying the
129 * GEM object so it is entirely possible to gain a
130 * reference on an object in the process of being freed
131 * since our serialisation is via the spinlock and not
132 * the struct_mutex - and consequently use it after it
133 * is freed and then double free it. To prevent that
134 * use-after-free we only acquire a reference on the
135 * object if it is not in the process of being destroyed.
137 mo = container_of(it, struct i915_mmu_object, it);
138 if (kref_get_unless_zero(&mo->obj->base.refcount))
139 queue_work(mn->wq, &mo->work);
141 list_add(&mo->link, &cancelled);
142 it = interval_tree_iter_next(it, start, end);
144 list_for_each_entry(mo, &cancelled, link)
146 spin_unlock(&mn->lock);
148 flush_workqueue(mn->wq);
151 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
152 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
155 static struct i915_mmu_notifier *
156 i915_mmu_notifier_create(struct mm_struct *mm)
158 struct i915_mmu_notifier *mn;
161 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
163 return ERR_PTR(-ENOMEM);
165 spin_lock_init(&mn->lock);
166 mn->mn.ops = &i915_gem_userptr_notifier;
167 mn->objects = RB_ROOT;
168 mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0);
169 if (mn->wq == NULL) {
171 return ERR_PTR(-ENOMEM);
174 /* Protected by mmap_sem (write-lock) */
175 ret = __mmu_notifier_register(&mn->mn, mm);
177 destroy_workqueue(mn->wq);
186 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
188 struct i915_mmu_object *mo;
190 mo = obj->userptr.mmu_object;
194 spin_lock(&mo->mn->lock);
196 spin_unlock(&mo->mn->lock);
199 obj->userptr.mmu_object = NULL;
202 static struct i915_mmu_notifier *
203 i915_mmu_notifier_find(struct i915_mm_struct *mm)
205 struct i915_mmu_notifier *mn = mm->mn;
211 down_write(&mm->mm->mmap_sem);
212 mutex_lock(&mm->i915->mm_lock);
213 if ((mn = mm->mn) == NULL) {
214 mn = i915_mmu_notifier_create(mm->mm);
218 mutex_unlock(&mm->i915->mm_lock);
219 up_write(&mm->mm->mmap_sem);
225 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
228 struct i915_mmu_notifier *mn;
229 struct i915_mmu_object *mo;
231 if (flags & I915_USERPTR_UNSYNCHRONIZED)
232 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
234 if (WARN_ON(obj->userptr.mm == NULL))
237 mn = i915_mmu_notifier_find(obj->userptr.mm);
241 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
247 mo->it.start = obj->userptr.ptr;
248 mo->it.last = obj->userptr.ptr + obj->base.size - 1;
249 INIT_WORK(&mo->work, cancel_userptr);
251 obj->userptr.mmu_object = mo;
256 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
257 struct mm_struct *mm)
262 mmu_notifier_unregister(&mn->mn, mm);
263 destroy_workqueue(mn->wq);
270 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
275 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
278 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
281 if (!capable(CAP_SYS_ADMIN))
288 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
289 struct mm_struct *mm)
295 static struct i915_mm_struct *
296 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
298 struct i915_mm_struct *mm;
300 /* Protected by dev_priv->mm_lock */
301 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
309 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
311 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
312 struct i915_mm_struct *mm;
315 /* During release of the GEM object we hold the struct_mutex. This
316 * precludes us from calling mmput() at that time as that may be
317 * the last reference and so call exit_mmap(). exit_mmap() will
318 * attempt to reap the vma, and if we were holding a GTT mmap
319 * would then call drm_gem_vm_close() and attempt to reacquire
320 * the struct mutex. So in order to avoid that recursion, we have
321 * to defer releasing the mm reference until after we drop the
322 * struct_mutex, i.e. we need to schedule a worker to do the clean
325 mutex_lock(&dev_priv->mm_lock);
326 mm = __i915_mm_struct_find(dev_priv, current->mm);
328 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
334 kref_init(&mm->kref);
335 mm->i915 = to_i915(obj->base.dev);
337 mm->mm = current->mm;
342 /* Protected by dev_priv->mm_lock */
343 hash_add(dev_priv->mm_structs,
344 &mm->node, (unsigned long)mm->mm);
348 obj->userptr.mm = mm;
350 mutex_unlock(&dev_priv->mm_lock);
355 __i915_mm_struct_free__worker(struct work_struct *work)
357 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
358 i915_mmu_notifier_free(mm->mn, mm->mm);
364 __i915_mm_struct_free(struct kref *kref)
366 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
368 /* Protected by dev_priv->mm_lock */
370 mutex_unlock(&mm->i915->mm_lock);
372 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
373 schedule_work(&mm->work);
377 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
379 if (obj->userptr.mm == NULL)
382 kref_put_mutex(&obj->userptr.mm->kref,
383 __i915_mm_struct_free,
384 &to_i915(obj->base.dev)->mm_lock);
385 obj->userptr.mm = NULL;
388 struct get_pages_work {
389 struct work_struct work;
390 struct drm_i915_gem_object *obj;
391 struct task_struct *task;
394 #if IS_ENABLED(CONFIG_SWIOTLB)
395 #define swiotlb_active() swiotlb_nr_tbl()
397 #define swiotlb_active() 0
401 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
403 struct scatterlist *sg;
406 *st = kmalloc(sizeof(**st), GFP_KERNEL);
410 if (swiotlb_active()) {
411 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
415 for_each_sg((*st)->sgl, sg, num_pages, n)
416 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
418 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
419 0, num_pages << PAGE_SHIFT,
433 static struct sg_table *
434 __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj,
435 struct page **pvec, int num_pages)
437 struct sg_table *pages;
440 ret = st_set_pages(&pages, pvec, num_pages);
444 ret = i915_gem_gtt_prepare_pages(obj, pages);
446 sg_free_table(pages);
455 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
460 /* During mm_invalidate_range we need to cancel any userptr that
461 * overlaps the range being invalidated. Doing so requires the
462 * struct_mutex, and that risks recursion. In order to cause
463 * recursion, the user must alias the userptr address space with
464 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
465 * to invalidate that mmaping, mm_invalidate_range is called with
466 * the userptr address *and* the struct_mutex held. To prevent that
467 * we set a flag under the i915_mmu_notifier spinlock to indicate
468 * whether this object is valid.
470 #if defined(CONFIG_MMU_NOTIFIER)
471 if (obj->userptr.mmu_object == NULL)
474 spin_lock(&obj->userptr.mmu_object->mn->lock);
475 /* In order to serialise get_pages with an outstanding
476 * cancel_userptr, we must drop the struct_mutex and try again.
479 del_object(obj->userptr.mmu_object);
480 else if (!work_pending(&obj->userptr.mmu_object->work))
481 add_object(obj->userptr.mmu_object);
484 spin_unlock(&obj->userptr.mmu_object->mn->lock);
491 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
493 struct get_pages_work *work = container_of(_work, typeof(*work), work);
494 struct drm_i915_gem_object *obj = work->obj;
495 const int npages = obj->base.size >> PAGE_SHIFT;
502 pvec = drm_malloc_gfp(npages, sizeof(struct page *), GFP_TEMPORARY);
504 struct mm_struct *mm = obj->userptr.mm->mm;
505 unsigned int flags = 0;
507 if (!obj->userptr.read_only)
511 if (mmget_not_zero(mm)) {
512 down_read(&mm->mmap_sem);
513 while (pinned < npages) {
514 ret = get_user_pages_remote
516 obj->userptr.ptr + pinned * PAGE_SIZE,
519 pvec + pinned, NULL, NULL);
525 up_read(&mm->mmap_sem);
530 mutex_lock(&obj->mm.lock);
531 if (obj->userptr.work == &work->work) {
532 struct sg_table *pages = ERR_PTR(ret);
534 if (pinned == npages) {
535 pages = __i915_gem_userptr_set_pages(obj, pvec, npages);
536 if (!IS_ERR(pages)) {
537 __i915_gem_object_set_pages(obj, pages);
543 obj->userptr.work = ERR_CAST(pages);
545 mutex_unlock(&obj->mm.lock);
547 release_pages(pvec, pinned, 0);
548 drm_free_large(pvec);
550 i915_gem_object_put(obj);
551 put_task_struct(work->task);
555 static struct sg_table *
556 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj,
559 struct get_pages_work *work;
561 /* Spawn a worker so that we can acquire the
562 * user pages without holding our mutex. Access
563 * to the user pages requires mmap_sem, and we have
564 * a strict lock ordering of mmap_sem, struct_mutex -
565 * we already hold struct_mutex here and so cannot
566 * call gup without encountering a lock inversion.
568 * Userspace will keep on repeating the operation
569 * (thanks to EAGAIN) until either we hit the fast
570 * path or the worker completes. If the worker is
571 * cancelled or superseded, the task is still run
572 * but the results ignored. (This leads to
573 * complications that we may have a stray object
574 * refcount that we need to be wary of when
575 * checking for existing objects during creation.)
576 * If the worker encounters an error, it reports
577 * that error back to this function through
578 * obj->userptr.work = ERR_PTR.
580 work = kmalloc(sizeof(*work), GFP_KERNEL);
582 return ERR_PTR(-ENOMEM);
584 obj->userptr.work = &work->work;
586 work->obj = i915_gem_object_get(obj);
588 work->task = current;
589 get_task_struct(work->task);
591 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
592 schedule_work(&work->work);
595 return ERR_PTR(-EAGAIN);
598 static struct sg_table *
599 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
601 const int num_pages = obj->base.size >> PAGE_SHIFT;
603 struct sg_table *pages;
607 /* If userspace should engineer that these pages are replaced in
608 * the vma between us binding this page into the GTT and completion
609 * of rendering... Their loss. If they change the mapping of their
610 * pages they need to create a new bo to point to the new vma.
612 * However, that still leaves open the possibility of the vma
613 * being copied upon fork. Which falls under the same userspace
614 * synchronisation issue as a regular bo, except that this time
615 * the process may not be expecting that a particular piece of
616 * memory is tied to the GPU.
618 * Fortunately, we can hook into the mmu_notifier in order to
619 * discard the page references prior to anything nasty happening
620 * to the vma (discard or cloning) which should prevent the more
621 * egregious cases from causing harm.
624 if (obj->userptr.work) {
625 /* active flag should still be held for the pending work */
626 if (IS_ERR(obj->userptr.work))
627 return ERR_CAST(obj->userptr.work);
629 return ERR_PTR(-EAGAIN);
632 /* Let the mmu-notifier know that we have begun and need cancellation */
633 ret = __i915_gem_userptr_set_active(obj, true);
639 if (obj->userptr.mm->mm == current->mm) {
640 pvec = drm_malloc_gfp(num_pages, sizeof(struct page *),
643 __i915_gem_userptr_set_active(obj, false);
644 return ERR_PTR(-ENOMEM);
647 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
648 !obj->userptr.read_only, pvec);
653 pages = ERR_PTR(pinned), pinned = 0;
654 else if (pinned < num_pages)
655 pages = __i915_gem_userptr_get_pages_schedule(obj, &active);
657 pages = __i915_gem_userptr_set_pages(obj, pvec, num_pages);
659 __i915_gem_userptr_set_active(obj, active);
660 release_pages(pvec, pinned, 0);
662 drm_free_large(pvec);
667 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
668 struct sg_table *pages)
670 struct sgt_iter sgt_iter;
673 BUG_ON(obj->userptr.work != NULL);
674 __i915_gem_userptr_set_active(obj, false);
676 if (obj->mm.madv != I915_MADV_WILLNEED)
677 obj->mm.dirty = false;
679 i915_gem_gtt_finish_pages(obj, pages);
681 for_each_sgt_page(page, sgt_iter, pages) {
683 set_page_dirty(page);
685 mark_page_accessed(page);
688 obj->mm.dirty = false;
690 sg_free_table(pages);
695 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
697 i915_gem_userptr_release__mmu_notifier(obj);
698 i915_gem_userptr_release__mm_struct(obj);
702 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
704 if (obj->userptr.mmu_object)
707 return i915_gem_userptr_init__mmu_notifier(obj, 0);
710 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
711 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
712 I915_GEM_OBJECT_IS_SHRINKABLE,
713 .get_pages = i915_gem_userptr_get_pages,
714 .put_pages = i915_gem_userptr_put_pages,
715 .dmabuf_export = i915_gem_userptr_dmabuf_export,
716 .release = i915_gem_userptr_release,
720 * Creates a new mm object that wraps some normal memory from the process
721 * context - user memory.
723 * We impose several restrictions upon the memory being mapped
725 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
726 * 2. It must be normal system memory, not a pointer into another map of IO
727 * space (e.g. it must not be a GTT mmapping of another object).
728 * 3. We only allow a bo as large as we could in theory map into the GTT,
729 * that is we limit the size to the total size of the GTT.
730 * 4. The bo is marked as being snoopable. The backing pages are left
731 * accessible directly by the CPU, but reads and writes by the GPU may
732 * incur the cost of a snoop (unless you have an LLC architecture).
734 * Synchronisation between multiple users and the GPU is left to userspace
735 * through the normal set-domain-ioctl. The kernel will enforce that the
736 * GPU relinquishes the VMA before it is returned back to the system
737 * i.e. upon free(), munmap() or process termination. However, the userspace
738 * malloc() library may not immediately relinquish the VMA after free() and
739 * instead reuse it whilst the GPU is still reading and writing to the VMA.
742 * Also note, that the object created here is not currently a "first class"
743 * object, in that several ioctls are banned. These are the CPU access
744 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
745 * direct access via your pointer rather than use those ioctls. Another
746 * restriction is that we do not allow userptr surfaces to be pinned to the
747 * hardware and so we reject any attempt to create a framebuffer out of a
750 * If you think this is a good interface to use to pass GPU memory between
751 * drivers, please use dma-buf instead. In fact, wherever possible use
755 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
757 struct drm_i915_private *dev_priv = to_i915(dev);
758 struct drm_i915_gem_userptr *args = data;
759 struct drm_i915_gem_object *obj;
763 if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
764 /* We cannot support coherent userptr objects on hw without
765 * LLC and broken snooping.
770 if (args->flags & ~(I915_USERPTR_READ_ONLY |
771 I915_USERPTR_UNSYNCHRONIZED))
774 if (offset_in_page(args->user_ptr | args->user_size))
777 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
778 (char __user *)(unsigned long)args->user_ptr, args->user_size))
781 if (args->flags & I915_USERPTR_READ_ONLY) {
782 /* On almost all of the current hw, we cannot tell the GPU that a
783 * page is readonly, so this is just a placeholder in the uAPI.
788 obj = i915_gem_object_alloc(dev_priv);
792 drm_gem_private_object_init(dev, &obj->base, args->user_size);
793 i915_gem_object_init(obj, &i915_gem_userptr_ops);
794 obj->cache_level = I915_CACHE_LLC;
795 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
796 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
798 obj->userptr.ptr = args->user_ptr;
799 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
801 /* And keep a pointer to the current->mm for resolving the user pages
802 * at binding. This means that we need to hook into the mmu_notifier
803 * in order to detect if the mmu is destroyed.
805 ret = i915_gem_userptr_init__mm_struct(obj);
807 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
809 ret = drm_gem_handle_create(file, &obj->base, &handle);
811 /* drop reference from allocate - handle holds it now */
812 i915_gem_object_put(obj);
816 args->handle = handle;
820 void i915_gem_init_userptr(struct drm_i915_private *dev_priv)
822 mutex_init(&dev_priv->mm_lock);
823 hash_init(dev_priv->mm_structs);