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>
35 struct i915_mm_struct {
37 struct drm_i915_private *i915;
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 workqueue_struct *wq;
55 struct i915_mmu_object {
56 struct i915_mmu_notifier *mn;
57 struct drm_i915_gem_object *obj;
58 struct interval_tree_node it;
59 struct list_head link;
60 struct work_struct work;
64 static void cancel_userptr(struct work_struct *work)
66 struct i915_mmu_object *mo = container_of(work, typeof(*mo), work);
67 struct drm_i915_gem_object *obj = mo->obj;
68 struct drm_device *dev = obj->base.dev;
70 i915_gem_object_wait(obj, I915_WAIT_ALL, MAX_SCHEDULE_TIMEOUT, NULL);
72 mutex_lock(&dev->struct_mutex);
73 /* Cancel any active worker and force us to re-evaluate gup */
74 obj->userptr.work = NULL;
76 /* We are inside a kthread context and can't be interrupted */
77 if (i915_gem_object_unbind(obj) == 0)
78 __i915_gem_object_put_pages(obj, I915_MM_NORMAL);
79 WARN_ONCE(obj->mm.pages,
80 "Failed to release pages: bind_count=%d, pages_pin_count=%d, pin_display=%d\n",
82 atomic_read(&obj->mm.pages_pin_count),
85 i915_gem_object_put(obj);
86 mutex_unlock(&dev->struct_mutex);
89 static void add_object(struct i915_mmu_object *mo)
94 interval_tree_insert(&mo->it, &mo->mn->objects);
98 static void del_object(struct i915_mmu_object *mo)
103 interval_tree_remove(&mo->it, &mo->mn->objects);
104 mo->attached = false;
107 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
108 struct mm_struct *mm,
112 struct i915_mmu_notifier *mn =
113 container_of(_mn, struct i915_mmu_notifier, mn);
114 struct i915_mmu_object *mo;
115 struct interval_tree_node *it;
116 LIST_HEAD(cancelled);
118 if (RB_EMPTY_ROOT(&mn->objects))
121 /* interval ranges are inclusive, but invalidate range is exclusive */
124 spin_lock(&mn->lock);
125 it = interval_tree_iter_first(&mn->objects, start, end);
127 /* The mmu_object is released late when destroying the
128 * GEM object so it is entirely possible to gain a
129 * reference on an object in the process of being freed
130 * since our serialisation is via the spinlock and not
131 * the struct_mutex - and consequently use it after it
132 * is freed and then double free it. To prevent that
133 * use-after-free we only acquire a reference on the
134 * object if it is not in the process of being destroyed.
136 mo = container_of(it, struct i915_mmu_object, it);
137 if (kref_get_unless_zero(&mo->obj->base.refcount))
138 queue_work(mn->wq, &mo->work);
140 list_add(&mo->link, &cancelled);
141 it = interval_tree_iter_next(it, start, end);
143 list_for_each_entry(mo, &cancelled, link)
145 spin_unlock(&mn->lock);
147 flush_workqueue(mn->wq);
150 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
151 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
154 static struct i915_mmu_notifier *
155 i915_mmu_notifier_create(struct mm_struct *mm)
157 struct i915_mmu_notifier *mn;
160 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
162 return ERR_PTR(-ENOMEM);
164 spin_lock_init(&mn->lock);
165 mn->mn.ops = &i915_gem_userptr_notifier;
166 mn->objects = RB_ROOT;
167 mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0);
168 if (mn->wq == NULL) {
170 return ERR_PTR(-ENOMEM);
173 /* Protected by mmap_sem (write-lock) */
174 ret = __mmu_notifier_register(&mn->mn, mm);
176 destroy_workqueue(mn->wq);
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 = mm->mn;
210 down_write(&mm->mm->mmap_sem);
211 mutex_lock(&mm->i915->mm_lock);
212 if ((mn = mm->mn) == NULL) {
213 mn = i915_mmu_notifier_create(mm->mm);
217 mutex_unlock(&mm->i915->mm_lock);
218 up_write(&mm->mm->mmap_sem);
224 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
227 struct i915_mmu_notifier *mn;
228 struct i915_mmu_object *mo;
230 if (flags & I915_USERPTR_UNSYNCHRONIZED)
231 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
233 if (WARN_ON(obj->userptr.mm == NULL))
236 mn = i915_mmu_notifier_find(obj->userptr.mm);
240 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
246 mo->it.start = obj->userptr.ptr;
247 mo->it.last = obj->userptr.ptr + obj->base.size - 1;
248 INIT_WORK(&mo->work, cancel_userptr);
250 obj->userptr.mmu_object = mo;
255 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
256 struct mm_struct *mm)
261 mmu_notifier_unregister(&mn->mn, mm);
262 destroy_workqueue(mn->wq);
269 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
274 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
277 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
280 if (!capable(CAP_SYS_ADMIN))
287 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
288 struct mm_struct *mm)
294 static struct i915_mm_struct *
295 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
297 struct i915_mm_struct *mm;
299 /* Protected by dev_priv->mm_lock */
300 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
308 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
310 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
311 struct i915_mm_struct *mm;
314 /* During release of the GEM object we hold the struct_mutex. This
315 * precludes us from calling mmput() at that time as that may be
316 * the last reference and so call exit_mmap(). exit_mmap() will
317 * attempt to reap the vma, and if we were holding a GTT mmap
318 * would then call drm_gem_vm_close() and attempt to reacquire
319 * the struct mutex. So in order to avoid that recursion, we have
320 * to defer releasing the mm reference until after we drop the
321 * struct_mutex, i.e. we need to schedule a worker to do the clean
324 mutex_lock(&dev_priv->mm_lock);
325 mm = __i915_mm_struct_find(dev_priv, current->mm);
327 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
333 kref_init(&mm->kref);
334 mm->i915 = to_i915(obj->base.dev);
336 mm->mm = current->mm;
337 atomic_inc(¤t->mm->mm_count);
341 /* Protected by dev_priv->mm_lock */
342 hash_add(dev_priv->mm_structs,
343 &mm->node, (unsigned long)mm->mm);
347 obj->userptr.mm = mm;
349 mutex_unlock(&dev_priv->mm_lock);
354 __i915_mm_struct_free__worker(struct work_struct *work)
356 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
357 i915_mmu_notifier_free(mm->mn, mm->mm);
363 __i915_mm_struct_free(struct kref *kref)
365 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
367 /* Protected by dev_priv->mm_lock */
369 mutex_unlock(&mm->i915->mm_lock);
371 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
372 schedule_work(&mm->work);
376 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
378 if (obj->userptr.mm == NULL)
381 kref_put_mutex(&obj->userptr.mm->kref,
382 __i915_mm_struct_free,
383 &to_i915(obj->base.dev)->mm_lock);
384 obj->userptr.mm = NULL;
387 struct get_pages_work {
388 struct work_struct work;
389 struct drm_i915_gem_object *obj;
390 struct task_struct *task;
393 #if IS_ENABLED(CONFIG_SWIOTLB)
394 #define swiotlb_active() swiotlb_nr_tbl()
396 #define swiotlb_active() 0
400 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
402 struct scatterlist *sg;
405 *st = kmalloc(sizeof(**st), GFP_KERNEL);
409 if (swiotlb_active()) {
410 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
414 for_each_sg((*st)->sgl, sg, num_pages, n)
415 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
417 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
418 0, num_pages << PAGE_SHIFT,
432 static struct sg_table *
433 __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj,
434 struct page **pvec, int num_pages)
436 struct sg_table *pages;
439 ret = st_set_pages(&pages, pvec, num_pages);
443 ret = i915_gem_gtt_prepare_pages(obj, pages);
445 sg_free_table(pages);
454 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
459 /* During mm_invalidate_range we need to cancel any userptr that
460 * overlaps the range being invalidated. Doing so requires the
461 * struct_mutex, and that risks recursion. In order to cause
462 * recursion, the user must alias the userptr address space with
463 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
464 * to invalidate that mmaping, mm_invalidate_range is called with
465 * the userptr address *and* the struct_mutex held. To prevent that
466 * we set a flag under the i915_mmu_notifier spinlock to indicate
467 * whether this object is valid.
469 #if defined(CONFIG_MMU_NOTIFIER)
470 if (obj->userptr.mmu_object == NULL)
473 spin_lock(&obj->userptr.mmu_object->mn->lock);
474 /* In order to serialise get_pages with an outstanding
475 * cancel_userptr, we must drop the struct_mutex and try again.
478 del_object(obj->userptr.mmu_object);
479 else if (!work_pending(&obj->userptr.mmu_object->work))
480 add_object(obj->userptr.mmu_object);
483 spin_unlock(&obj->userptr.mmu_object->mn->lock);
490 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
492 struct get_pages_work *work = container_of(_work, typeof(*work), work);
493 struct drm_i915_gem_object *obj = work->obj;
494 const int npages = obj->base.size >> PAGE_SHIFT;
501 pvec = drm_malloc_gfp(npages, sizeof(struct page *), GFP_TEMPORARY);
503 struct mm_struct *mm = obj->userptr.mm->mm;
504 unsigned int flags = 0;
506 if (!obj->userptr.read_only)
510 if (atomic_inc_not_zero(&mm->mm_users)) {
511 down_read(&mm->mmap_sem);
512 while (pinned < npages) {
513 ret = get_user_pages_remote
515 obj->userptr.ptr + pinned * PAGE_SIZE,
518 pvec + pinned, NULL, NULL);
524 up_read(&mm->mmap_sem);
529 mutex_lock(&obj->mm.lock);
530 if (obj->userptr.work == &work->work) {
531 struct sg_table *pages = ERR_PTR(ret);
533 if (pinned == npages) {
534 pages = __i915_gem_userptr_set_pages(obj, pvec, npages);
535 if (!IS_ERR(pages)) {
536 __i915_gem_object_set_pages(obj, pages);
542 obj->userptr.work = ERR_CAST(pages);
544 mutex_unlock(&obj->mm.lock);
546 release_pages(pvec, pinned, 0);
547 drm_free_large(pvec);
549 i915_gem_object_put(obj);
550 put_task_struct(work->task);
554 static struct sg_table *
555 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj,
558 struct get_pages_work *work;
560 /* Spawn a worker so that we can acquire the
561 * user pages without holding our mutex. Access
562 * to the user pages requires mmap_sem, and we have
563 * a strict lock ordering of mmap_sem, struct_mutex -
564 * we already hold struct_mutex here and so cannot
565 * call gup without encountering a lock inversion.
567 * Userspace will keep on repeating the operation
568 * (thanks to EAGAIN) until either we hit the fast
569 * path or the worker completes. If the worker is
570 * cancelled or superseded, the task is still run
571 * but the results ignored. (This leads to
572 * complications that we may have a stray object
573 * refcount that we need to be wary of when
574 * checking for existing objects during creation.)
575 * If the worker encounters an error, it reports
576 * that error back to this function through
577 * obj->userptr.work = ERR_PTR.
579 work = kmalloc(sizeof(*work), GFP_KERNEL);
581 return ERR_PTR(-ENOMEM);
583 obj->userptr.work = &work->work;
585 work->obj = i915_gem_object_get(obj);
587 work->task = current;
588 get_task_struct(work->task);
590 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
591 schedule_work(&work->work);
594 return ERR_PTR(-EAGAIN);
597 static struct sg_table *
598 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
600 const int num_pages = obj->base.size >> PAGE_SHIFT;
602 struct sg_table *pages;
606 /* If userspace should engineer that these pages are replaced in
607 * the vma between us binding this page into the GTT and completion
608 * of rendering... Their loss. If they change the mapping of their
609 * pages they need to create a new bo to point to the new vma.
611 * However, that still leaves open the possibility of the vma
612 * being copied upon fork. Which falls under the same userspace
613 * synchronisation issue as a regular bo, except that this time
614 * the process may not be expecting that a particular piece of
615 * memory is tied to the GPU.
617 * Fortunately, we can hook into the mmu_notifier in order to
618 * discard the page references prior to anything nasty happening
619 * to the vma (discard or cloning) which should prevent the more
620 * egregious cases from causing harm.
623 if (obj->userptr.work) {
624 /* active flag should still be held for the pending work */
625 if (IS_ERR(obj->userptr.work))
626 return ERR_CAST(obj->userptr.work);
628 return ERR_PTR(-EAGAIN);
631 /* Let the mmu-notifier know that we have begun and need cancellation */
632 ret = __i915_gem_userptr_set_active(obj, true);
638 if (obj->userptr.mm->mm == current->mm) {
639 pvec = drm_malloc_gfp(num_pages, sizeof(struct page *),
642 __i915_gem_userptr_set_active(obj, false);
643 return ERR_PTR(-ENOMEM);
646 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
647 !obj->userptr.read_only, pvec);
652 pages = ERR_PTR(pinned), pinned = 0;
653 else if (pinned < num_pages)
654 pages = __i915_gem_userptr_get_pages_schedule(obj, &active);
656 pages = __i915_gem_userptr_set_pages(obj, pvec, num_pages);
658 __i915_gem_userptr_set_active(obj, active);
659 release_pages(pvec, pinned, 0);
661 drm_free_large(pvec);
666 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
667 struct sg_table *pages)
669 struct sgt_iter sgt_iter;
672 BUG_ON(obj->userptr.work != NULL);
673 __i915_gem_userptr_set_active(obj, false);
675 if (obj->mm.madv != I915_MADV_WILLNEED)
676 obj->mm.dirty = false;
678 i915_gem_gtt_finish_pages(obj, pages);
680 for_each_sgt_page(page, sgt_iter, pages) {
682 set_page_dirty(page);
684 mark_page_accessed(page);
687 obj->mm.dirty = false;
689 sg_free_table(pages);
694 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
696 i915_gem_userptr_release__mmu_notifier(obj);
697 i915_gem_userptr_release__mm_struct(obj);
701 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
703 if (obj->userptr.mmu_object)
706 return i915_gem_userptr_init__mmu_notifier(obj, 0);
709 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
710 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
711 I915_GEM_OBJECT_IS_SHRINKABLE,
712 .get_pages = i915_gem_userptr_get_pages,
713 .put_pages = i915_gem_userptr_put_pages,
714 .dmabuf_export = i915_gem_userptr_dmabuf_export,
715 .release = i915_gem_userptr_release,
719 * Creates a new mm object that wraps some normal memory from the process
720 * context - user memory.
722 * We impose several restrictions upon the memory being mapped
724 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
725 * 2. It must be normal system memory, not a pointer into another map of IO
726 * space (e.g. it must not be a GTT mmapping of another object).
727 * 3. We only allow a bo as large as we could in theory map into the GTT,
728 * that is we limit the size to the total size of the GTT.
729 * 4. The bo is marked as being snoopable. The backing pages are left
730 * accessible directly by the CPU, but reads and writes by the GPU may
731 * incur the cost of a snoop (unless you have an LLC architecture).
733 * Synchronisation between multiple users and the GPU is left to userspace
734 * through the normal set-domain-ioctl. The kernel will enforce that the
735 * GPU relinquishes the VMA before it is returned back to the system
736 * i.e. upon free(), munmap() or process termination. However, the userspace
737 * malloc() library may not immediately relinquish the VMA after free() and
738 * instead reuse it whilst the GPU is still reading and writing to the VMA.
741 * Also note, that the object created here is not currently a "first class"
742 * object, in that several ioctls are banned. These are the CPU access
743 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
744 * direct access via your pointer rather than use those ioctls. Another
745 * restriction is that we do not allow userptr surfaces to be pinned to the
746 * hardware and so we reject any attempt to create a framebuffer out of a
749 * If you think this is a good interface to use to pass GPU memory between
750 * drivers, please use dma-buf instead. In fact, wherever possible use
754 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
756 struct drm_i915_private *dev_priv = to_i915(dev);
757 struct drm_i915_gem_userptr *args = data;
758 struct drm_i915_gem_object *obj;
762 if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
763 /* We cannot support coherent userptr objects on hw without
764 * LLC and broken snooping.
769 if (args->flags & ~(I915_USERPTR_READ_ONLY |
770 I915_USERPTR_UNSYNCHRONIZED))
773 if (offset_in_page(args->user_ptr | args->user_size))
776 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
777 (char __user *)(unsigned long)args->user_ptr, args->user_size))
780 if (args->flags & I915_USERPTR_READ_ONLY) {
781 /* On almost all of the current hw, we cannot tell the GPU that a
782 * page is readonly, so this is just a placeholder in the uAPI.
787 obj = i915_gem_object_alloc(dev);
791 drm_gem_private_object_init(dev, &obj->base, args->user_size);
792 i915_gem_object_init(obj, &i915_gem_userptr_ops);
793 obj->cache_level = I915_CACHE_LLC;
794 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
795 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
797 obj->userptr.ptr = args->user_ptr;
798 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
800 /* And keep a pointer to the current->mm for resolving the user pages
801 * at binding. This means that we need to hook into the mmu_notifier
802 * in order to detect if the mmu is destroyed.
804 ret = i915_gem_userptr_init__mm_struct(obj);
806 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
808 ret = drm_gem_handle_create(file, &obj->base, &handle);
810 /* drop reference from allocate - handle holds it now */
811 i915_gem_object_put(obj);
815 args->handle = handle;
819 void i915_gem_init_userptr(struct drm_i915_private *dev_priv)
821 mutex_init(&dev_priv->mm_lock);
822 hash_init(dev_priv->mm_structs);