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)
117 if (!kref_get_unless_zero(&obj->base.refcount))
120 spin_unlock(&mn->lock);
124 spin_lock(&mn->lock);
125 if (serial != mn->serial)
132 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
133 struct mm_struct *mm,
137 struct i915_mmu_notifier *mn = container_of(_mn, struct i915_mmu_notifier, mn);
138 struct interval_tree_node *it = NULL;
139 unsigned long next = start;
140 unsigned long serial = 0;
142 end--; /* interval ranges are inclusive, but invalidate range is exclusive */
144 struct drm_i915_gem_object *obj = NULL;
146 spin_lock(&mn->lock);
148 it = invalidate_range__linear(mn, mm, start, end);
149 else if (serial == mn->serial)
150 it = interval_tree_iter_next(it, next, end);
152 it = interval_tree_iter_first(&mn->objects, start, end);
154 obj = container_of(it, struct i915_mmu_object, it)->obj;
156 /* The mmu_object is released late when destroying the
157 * GEM object so it is entirely possible to gain a
158 * reference on an object in the process of being freed
159 * since our serialisation is via the spinlock and not
160 * the struct_mutex - and consequently use it after it
161 * is freed and then double free it.
163 if (!kref_get_unless_zero(&obj->base.refcount)) {
164 spin_unlock(&mn->lock);
171 spin_unlock(&mn->lock);
175 next = cancel_userptr(obj);
179 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
180 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
183 static struct i915_mmu_notifier *
184 i915_mmu_notifier_create(struct mm_struct *mm)
186 struct i915_mmu_notifier *mn;
189 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
191 return ERR_PTR(-ENOMEM);
193 spin_lock_init(&mn->lock);
194 mn->mn.ops = &i915_gem_userptr_notifier;
195 mn->objects = RB_ROOT;
197 INIT_LIST_HEAD(&mn->linear);
198 mn->has_linear = false;
200 /* Protected by mmap_sem (write-lock) */
201 ret = __mmu_notifier_register(&mn->mn, mm);
210 static void __i915_mmu_notifier_update_serial(struct i915_mmu_notifier *mn)
212 if (++mn->serial == 0)
217 i915_mmu_notifier_add(struct drm_device *dev,
218 struct i915_mmu_notifier *mn,
219 struct i915_mmu_object *mo)
221 struct interval_tree_node *it;
224 ret = i915_mutex_lock_interruptible(dev);
228 /* Make sure we drop the final active reference (and thereby
229 * remove the objects from the interval tree) before we do
230 * the check for overlapping objects.
232 i915_gem_retire_requests(dev);
234 spin_lock(&mn->lock);
235 it = interval_tree_iter_first(&mn->objects,
236 mo->it.start, mo->it.last);
238 struct drm_i915_gem_object *obj;
240 /* We only need to check the first object in the range as it
241 * either has cancelled gup work queued and we need to
242 * return back to the user to give time for the gup-workers
243 * to flush their object references upon which the object will
244 * be removed from the interval-tree, or the the range is
245 * still in use by another client and the overlap is invalid.
247 * If we do have an overlap, we cannot use the interval tree
248 * for fast range invalidation.
251 obj = container_of(it, struct i915_mmu_object, it)->obj;
252 if (!obj->userptr.workers)
253 mn->has_linear = mo->is_linear = true;
257 interval_tree_insert(&mo->it, &mn->objects);
260 list_add(&mo->link, &mn->linear);
261 __i915_mmu_notifier_update_serial(mn);
263 spin_unlock(&mn->lock);
264 mutex_unlock(&dev->struct_mutex);
269 static bool i915_mmu_notifier_has_linear(struct i915_mmu_notifier *mn)
271 struct i915_mmu_object *mo;
273 list_for_each_entry(mo, &mn->linear, link)
281 i915_mmu_notifier_del(struct i915_mmu_notifier *mn,
282 struct i915_mmu_object *mo)
284 spin_lock(&mn->lock);
287 mn->has_linear = i915_mmu_notifier_has_linear(mn);
289 interval_tree_remove(&mo->it, &mn->objects);
290 __i915_mmu_notifier_update_serial(mn);
291 spin_unlock(&mn->lock);
295 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
297 struct i915_mmu_object *mo;
299 mo = obj->userptr.mmu_object;
303 i915_mmu_notifier_del(mo->mn, mo);
306 obj->userptr.mmu_object = NULL;
309 static struct i915_mmu_notifier *
310 i915_mmu_notifier_find(struct i915_mm_struct *mm)
312 struct i915_mmu_notifier *mn = mm->mn;
318 down_write(&mm->mm->mmap_sem);
319 mutex_lock(&to_i915(mm->dev)->mm_lock);
320 if ((mn = mm->mn) == NULL) {
321 mn = i915_mmu_notifier_create(mm->mm);
325 mutex_unlock(&to_i915(mm->dev)->mm_lock);
326 up_write(&mm->mm->mmap_sem);
332 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
335 struct i915_mmu_notifier *mn;
336 struct i915_mmu_object *mo;
339 if (flags & I915_USERPTR_UNSYNCHRONIZED)
340 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
342 if (WARN_ON(obj->userptr.mm == NULL))
345 mn = i915_mmu_notifier_find(obj->userptr.mm);
349 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
354 mo->it.start = obj->userptr.ptr;
355 mo->it.last = mo->it.start + obj->base.size - 1;
358 ret = i915_mmu_notifier_add(obj->base.dev, mn, mo);
364 obj->userptr.mmu_object = mo;
369 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
370 struct mm_struct *mm)
375 mmu_notifier_unregister(&mn->mn, mm);
382 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
387 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
390 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
393 if (!capable(CAP_SYS_ADMIN))
400 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
401 struct mm_struct *mm)
407 static struct i915_mm_struct *
408 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
410 struct i915_mm_struct *mm;
412 /* Protected by dev_priv->mm_lock */
413 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
421 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
423 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
424 struct i915_mm_struct *mm;
427 /* During release of the GEM object we hold the struct_mutex. This
428 * precludes us from calling mmput() at that time as that may be
429 * the last reference and so call exit_mmap(). exit_mmap() will
430 * attempt to reap the vma, and if we were holding a GTT mmap
431 * would then call drm_gem_vm_close() and attempt to reacquire
432 * the struct mutex. So in order to avoid that recursion, we have
433 * to defer releasing the mm reference until after we drop the
434 * struct_mutex, i.e. we need to schedule a worker to do the clean
437 mutex_lock(&dev_priv->mm_lock);
438 mm = __i915_mm_struct_find(dev_priv, current->mm);
440 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
446 kref_init(&mm->kref);
447 mm->dev = obj->base.dev;
449 mm->mm = current->mm;
450 atomic_inc(¤t->mm->mm_count);
454 /* Protected by dev_priv->mm_lock */
455 hash_add(dev_priv->mm_structs,
456 &mm->node, (unsigned long)mm->mm);
460 obj->userptr.mm = mm;
462 mutex_unlock(&dev_priv->mm_lock);
467 __i915_mm_struct_free__worker(struct work_struct *work)
469 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
470 i915_mmu_notifier_free(mm->mn, mm->mm);
476 __i915_mm_struct_free(struct kref *kref)
478 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
480 /* Protected by dev_priv->mm_lock */
482 mutex_unlock(&to_i915(mm->dev)->mm_lock);
484 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
485 schedule_work(&mm->work);
489 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
491 if (obj->userptr.mm == NULL)
494 kref_put_mutex(&obj->userptr.mm->kref,
495 __i915_mm_struct_free,
496 &to_i915(obj->base.dev)->mm_lock);
497 obj->userptr.mm = NULL;
500 struct get_pages_work {
501 struct work_struct work;
502 struct drm_i915_gem_object *obj;
503 struct task_struct *task;
506 #if IS_ENABLED(CONFIG_SWIOTLB)
507 #define swiotlb_active() swiotlb_nr_tbl()
509 #define swiotlb_active() 0
513 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
515 struct scatterlist *sg;
518 *st = kmalloc(sizeof(**st), GFP_KERNEL);
522 if (swiotlb_active()) {
523 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
527 for_each_sg((*st)->sgl, sg, num_pages, n)
528 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
530 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
531 0, num_pages << PAGE_SHIFT,
546 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
548 struct get_pages_work *work = container_of(_work, typeof(*work), work);
549 struct drm_i915_gem_object *obj = work->obj;
550 struct drm_device *dev = obj->base.dev;
551 const int num_pages = obj->base.size >> PAGE_SHIFT;
558 pvec = kmalloc(num_pages*sizeof(struct page *),
559 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
561 pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
563 struct mm_struct *mm = obj->userptr.mm->mm;
565 down_read(&mm->mmap_sem);
566 while (pinned < num_pages) {
567 ret = get_user_pages(work->task, mm,
568 obj->userptr.ptr + pinned * PAGE_SIZE,
570 !obj->userptr.read_only, 0,
571 pvec + pinned, NULL);
577 up_read(&mm->mmap_sem);
580 mutex_lock(&dev->struct_mutex);
581 if (obj->userptr.work != &work->work) {
583 } else if (pinned == num_pages) {
584 ret = st_set_pages(&obj->pages, pvec, num_pages);
586 list_add_tail(&obj->global_list, &to_i915(dev)->mm.unbound_list);
591 obj->userptr.work = ERR_PTR(ret);
592 obj->userptr.workers--;
593 drm_gem_object_unreference(&obj->base);
594 mutex_unlock(&dev->struct_mutex);
596 release_pages(pvec, pinned, 0);
597 drm_free_large(pvec);
599 put_task_struct(work->task);
604 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
606 const int num_pages = obj->base.size >> PAGE_SHIFT;
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.
629 if (obj->userptr.mm->mm == current->mm) {
630 pvec = kmalloc(num_pages*sizeof(struct page *),
631 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
633 pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
638 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
639 !obj->userptr.read_only, pvec);
641 if (pinned < num_pages) {
646 /* Spawn a worker so that we can acquire the
647 * user pages without holding our mutex. Access
648 * to the user pages requires mmap_sem, and we have
649 * a strict lock ordering of mmap_sem, struct_mutex -
650 * we already hold struct_mutex here and so cannot
651 * call gup without encountering a lock inversion.
653 * Userspace will keep on repeating the operation
654 * (thanks to EAGAIN) until either we hit the fast
655 * path or the worker completes. If the worker is
656 * cancelled or superseded, the task is still run
657 * but the results ignored. (This leads to
658 * complications that we may have a stray object
659 * refcount that we need to be wary of when
660 * checking for existing objects during creation.)
661 * If the worker encounters an error, it reports
662 * that error back to this function through
663 * obj->userptr.work = ERR_PTR.
666 if (obj->userptr.work == NULL &&
667 obj->userptr.workers < I915_GEM_USERPTR_MAX_WORKERS) {
668 struct get_pages_work *work;
670 work = kmalloc(sizeof(*work), GFP_KERNEL);
672 obj->userptr.work = &work->work;
673 obj->userptr.workers++;
676 drm_gem_object_reference(&obj->base);
678 work->task = current;
679 get_task_struct(work->task);
681 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
682 schedule_work(&work->work);
686 if (IS_ERR(obj->userptr.work)) {
687 ret = PTR_ERR(obj->userptr.work);
688 obj->userptr.work = NULL;
693 ret = st_set_pages(&obj->pages, pvec, num_pages);
695 obj->userptr.work = NULL;
700 release_pages(pvec, pinned, 0);
701 drm_free_large(pvec);
706 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj)
708 struct sg_page_iter sg_iter;
710 BUG_ON(obj->userptr.work != NULL);
712 if (obj->madv != I915_MADV_WILLNEED)
715 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
716 struct page *page = sg_page_iter_page(&sg_iter);
719 set_page_dirty(page);
721 mark_page_accessed(page);
722 page_cache_release(page);
726 sg_free_table(obj->pages);
731 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
733 i915_gem_userptr_release__mmu_notifier(obj);
734 i915_gem_userptr_release__mm_struct(obj);
738 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
740 if (obj->userptr.mmu_object)
743 return i915_gem_userptr_init__mmu_notifier(obj, 0);
746 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
747 .dmabuf_export = i915_gem_userptr_dmabuf_export,
748 .get_pages = i915_gem_userptr_get_pages,
749 .put_pages = i915_gem_userptr_put_pages,
750 .release = i915_gem_userptr_release,
754 * Creates a new mm object that wraps some normal memory from the process
755 * context - user memory.
757 * We impose several restrictions upon the memory being mapped
759 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
760 * 2. It must be normal system memory, not a pointer into another map of IO
761 * space (e.g. it must not be a GTT mmapping of another object).
762 * 3. We only allow a bo as large as we could in theory map into the GTT,
763 * that is we limit the size to the total size of the GTT.
764 * 4. The bo is marked as being snoopable. The backing pages are left
765 * accessible directly by the CPU, but reads and writes by the GPU may
766 * incur the cost of a snoop (unless you have an LLC architecture).
768 * Synchronisation between multiple users and the GPU is left to userspace
769 * through the normal set-domain-ioctl. The kernel will enforce that the
770 * GPU relinquishes the VMA before it is returned back to the system
771 * i.e. upon free(), munmap() or process termination. However, the userspace
772 * malloc() library may not immediately relinquish the VMA after free() and
773 * instead reuse it whilst the GPU is still reading and writing to the VMA.
776 * Also note, that the object created here is not currently a "first class"
777 * object, in that several ioctls are banned. These are the CPU access
778 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
779 * direct access via your pointer rather than use those ioctls.
781 * If you think this is a good interface to use to pass GPU memory between
782 * drivers, please use dma-buf instead. In fact, wherever possible use
786 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
788 struct drm_i915_private *dev_priv = dev->dev_private;
789 struct drm_i915_gem_userptr *args = data;
790 struct drm_i915_gem_object *obj;
794 if (args->flags & ~(I915_USERPTR_READ_ONLY |
795 I915_USERPTR_UNSYNCHRONIZED))
798 if (offset_in_page(args->user_ptr | args->user_size))
801 if (args->user_size > dev_priv->gtt.base.total)
804 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
805 (char __user *)(unsigned long)args->user_ptr, args->user_size))
808 if (args->flags & I915_USERPTR_READ_ONLY) {
809 /* On almost all of the current hw, we cannot tell the GPU that a
810 * page is readonly, so this is just a placeholder in the uAPI.
815 obj = i915_gem_object_alloc(dev);
819 drm_gem_private_object_init(dev, &obj->base, args->user_size);
820 i915_gem_object_init(obj, &i915_gem_userptr_ops);
821 obj->cache_level = I915_CACHE_LLC;
822 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
823 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
825 obj->userptr.ptr = args->user_ptr;
826 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
828 /* And keep a pointer to the current->mm for resolving the user pages
829 * at binding. This means that we need to hook into the mmu_notifier
830 * in order to detect if the mmu is destroyed.
832 ret = i915_gem_userptr_init__mm_struct(obj);
834 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
836 ret = drm_gem_handle_create(file, &obj->base, &handle);
838 /* drop reference from allocate - handle holds it now */
839 drm_gem_object_unreference_unlocked(&obj->base);
843 args->handle = handle;
848 i915_gem_init_userptr(struct drm_device *dev)
850 struct drm_i915_private *dev_priv = to_i915(dev);
851 mutex_init(&dev_priv->mm_lock);
852 hash_init(dev_priv->mm_structs);