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 wait_rendering(struct drm_i915_gem_object *obj)
66 unsigned long active = __I915_BO_ACTIVE(obj);
69 for_each_active(active, idx)
70 i915_gem_active_wait_unlocked(&obj->last_read[idx],
74 static void cancel_userptr(struct work_struct *work)
76 struct i915_mmu_object *mo = container_of(work, typeof(*mo), work);
77 struct drm_i915_gem_object *obj = mo->obj;
78 struct drm_device *dev = obj->base.dev;
82 mutex_lock(&dev->struct_mutex);
83 /* Cancel any active worker and force us to re-evaluate gup */
84 obj->userptr.work = NULL;
86 if (obj->pages != NULL) {
87 /* We are inside a kthread context and can't be interrupted */
88 WARN_ON(i915_gem_object_unbind(obj));
89 WARN_ON(i915_gem_object_put_pages(obj));
92 i915_gem_object_put(obj);
93 mutex_unlock(&dev->struct_mutex);
96 static void add_object(struct i915_mmu_object *mo)
101 interval_tree_insert(&mo->it, &mo->mn->objects);
105 static void del_object(struct i915_mmu_object *mo)
110 interval_tree_remove(&mo->it, &mo->mn->objects);
111 mo->attached = false;
114 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
115 struct mm_struct *mm,
119 struct i915_mmu_notifier *mn =
120 container_of(_mn, struct i915_mmu_notifier, mn);
121 struct i915_mmu_object *mo;
122 struct interval_tree_node *it;
123 LIST_HEAD(cancelled);
125 if (RB_EMPTY_ROOT(&mn->objects))
128 /* interval ranges are inclusive, but invalidate range is exclusive */
131 spin_lock(&mn->lock);
132 it = interval_tree_iter_first(&mn->objects, start, end);
134 /* The mmu_object is released late when destroying the
135 * GEM object so it is entirely possible to gain a
136 * reference on an object in the process of being freed
137 * since our serialisation is via the spinlock and not
138 * the struct_mutex - and consequently use it after it
139 * is freed and then double free it. To prevent that
140 * use-after-free we only acquire a reference on the
141 * object if it is not in the process of being destroyed.
143 mo = container_of(it, struct i915_mmu_object, it);
144 if (kref_get_unless_zero(&mo->obj->base.refcount))
145 queue_work(mn->wq, &mo->work);
147 list_add(&mo->link, &cancelled);
148 it = interval_tree_iter_next(it, start, end);
150 list_for_each_entry(mo, &cancelled, link)
152 spin_unlock(&mn->lock);
154 flush_workqueue(mn->wq);
157 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
158 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
161 static struct i915_mmu_notifier *
162 i915_mmu_notifier_create(struct mm_struct *mm)
164 struct i915_mmu_notifier *mn;
167 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
169 return ERR_PTR(-ENOMEM);
171 spin_lock_init(&mn->lock);
172 mn->mn.ops = &i915_gem_userptr_notifier;
173 mn->objects = RB_ROOT;
174 mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0);
175 if (mn->wq == NULL) {
177 return ERR_PTR(-ENOMEM);
180 /* Protected by mmap_sem (write-lock) */
181 ret = __mmu_notifier_register(&mn->mn, mm);
183 destroy_workqueue(mn->wq);
192 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
194 struct i915_mmu_object *mo;
196 mo = obj->userptr.mmu_object;
200 spin_lock(&mo->mn->lock);
202 spin_unlock(&mo->mn->lock);
205 obj->userptr.mmu_object = NULL;
208 static struct i915_mmu_notifier *
209 i915_mmu_notifier_find(struct i915_mm_struct *mm)
211 struct i915_mmu_notifier *mn = mm->mn;
217 down_write(&mm->mm->mmap_sem);
218 mutex_lock(&mm->i915->mm_lock);
219 if ((mn = mm->mn) == NULL) {
220 mn = i915_mmu_notifier_create(mm->mm);
224 mutex_unlock(&mm->i915->mm_lock);
225 up_write(&mm->mm->mmap_sem);
231 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
234 struct i915_mmu_notifier *mn;
235 struct i915_mmu_object *mo;
237 if (flags & I915_USERPTR_UNSYNCHRONIZED)
238 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
240 if (WARN_ON(obj->userptr.mm == NULL))
243 mn = i915_mmu_notifier_find(obj->userptr.mm);
247 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
253 mo->it.start = obj->userptr.ptr;
254 mo->it.last = obj->userptr.ptr + obj->base.size - 1;
255 INIT_WORK(&mo->work, cancel_userptr);
257 obj->userptr.mmu_object = mo;
262 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
263 struct mm_struct *mm)
268 mmu_notifier_unregister(&mn->mn, mm);
269 destroy_workqueue(mn->wq);
276 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
281 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
284 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
287 if (!capable(CAP_SYS_ADMIN))
294 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
295 struct mm_struct *mm)
301 static struct i915_mm_struct *
302 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
304 struct i915_mm_struct *mm;
306 /* Protected by dev_priv->mm_lock */
307 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
315 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
317 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
318 struct i915_mm_struct *mm;
321 /* During release of the GEM object we hold the struct_mutex. This
322 * precludes us from calling mmput() at that time as that may be
323 * the last reference and so call exit_mmap(). exit_mmap() will
324 * attempt to reap the vma, and if we were holding a GTT mmap
325 * would then call drm_gem_vm_close() and attempt to reacquire
326 * the struct mutex. So in order to avoid that recursion, we have
327 * to defer releasing the mm reference until after we drop the
328 * struct_mutex, i.e. we need to schedule a worker to do the clean
331 mutex_lock(&dev_priv->mm_lock);
332 mm = __i915_mm_struct_find(dev_priv, current->mm);
334 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
340 kref_init(&mm->kref);
341 mm->i915 = to_i915(obj->base.dev);
343 mm->mm = current->mm;
344 atomic_inc(¤t->mm->mm_count);
348 /* Protected by dev_priv->mm_lock */
349 hash_add(dev_priv->mm_structs,
350 &mm->node, (unsigned long)mm->mm);
354 obj->userptr.mm = mm;
356 mutex_unlock(&dev_priv->mm_lock);
361 __i915_mm_struct_free__worker(struct work_struct *work)
363 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
364 i915_mmu_notifier_free(mm->mn, mm->mm);
370 __i915_mm_struct_free(struct kref *kref)
372 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
374 /* Protected by dev_priv->mm_lock */
376 mutex_unlock(&mm->i915->mm_lock);
378 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
379 schedule_work(&mm->work);
383 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
385 if (obj->userptr.mm == NULL)
388 kref_put_mutex(&obj->userptr.mm->kref,
389 __i915_mm_struct_free,
390 &to_i915(obj->base.dev)->mm_lock);
391 obj->userptr.mm = NULL;
394 struct get_pages_work {
395 struct work_struct work;
396 struct drm_i915_gem_object *obj;
397 struct task_struct *task;
400 #if IS_ENABLED(CONFIG_SWIOTLB)
401 #define swiotlb_active() swiotlb_nr_tbl()
403 #define swiotlb_active() 0
407 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
409 struct scatterlist *sg;
412 *st = kmalloc(sizeof(**st), GFP_KERNEL);
416 if (swiotlb_active()) {
417 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
421 for_each_sg((*st)->sgl, sg, num_pages, n)
422 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
424 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
425 0, num_pages << PAGE_SHIFT,
440 __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj,
441 struct page **pvec, int num_pages)
445 ret = st_set_pages(&obj->pages, pvec, num_pages);
449 ret = i915_gem_gtt_prepare_object(obj);
451 sg_free_table(obj->pages);
460 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
465 /* During mm_invalidate_range we need to cancel any userptr that
466 * overlaps the range being invalidated. Doing so requires the
467 * struct_mutex, and that risks recursion. In order to cause
468 * recursion, the user must alias the userptr address space with
469 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
470 * to invalidate that mmaping, mm_invalidate_range is called with
471 * the userptr address *and* the struct_mutex held. To prevent that
472 * we set a flag under the i915_mmu_notifier spinlock to indicate
473 * whether this object is valid.
475 #if defined(CONFIG_MMU_NOTIFIER)
476 if (obj->userptr.mmu_object == NULL)
479 spin_lock(&obj->userptr.mmu_object->mn->lock);
480 /* In order to serialise get_pages with an outstanding
481 * cancel_userptr, we must drop the struct_mutex and try again.
484 del_object(obj->userptr.mmu_object);
485 else if (!work_pending(&obj->userptr.mmu_object->work))
486 add_object(obj->userptr.mmu_object);
489 spin_unlock(&obj->userptr.mmu_object->mn->lock);
496 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
498 struct get_pages_work *work = container_of(_work, typeof(*work), work);
499 struct drm_i915_gem_object *obj = work->obj;
500 struct drm_device *dev = obj->base.dev;
501 const int npages = obj->base.size >> PAGE_SHIFT;
508 pvec = drm_malloc_gfp(npages, sizeof(struct page *), GFP_TEMPORARY);
510 struct mm_struct *mm = obj->userptr.mm->mm;
511 unsigned int flags = 0;
513 if (!obj->userptr.read_only)
517 if (atomic_inc_not_zero(&mm->mm_users)) {
518 down_read(&mm->mmap_sem);
519 while (pinned < npages) {
520 ret = get_user_pages_remote
522 obj->userptr.ptr + pinned * PAGE_SIZE,
525 pvec + pinned, NULL);
531 up_read(&mm->mmap_sem);
536 mutex_lock(&dev->struct_mutex);
537 if (obj->userptr.work == &work->work) {
538 if (pinned == npages) {
539 ret = __i915_gem_userptr_set_pages(obj, pvec, npages);
541 list_add_tail(&obj->global_list,
542 &to_i915(dev)->mm.unbound_list);
543 obj->get_page.sg = obj->pages->sgl;
544 obj->get_page.last = 0;
548 obj->userptr.work = ERR_PTR(ret);
551 obj->userptr.workers--;
552 i915_gem_object_put(obj);
553 mutex_unlock(&dev->struct_mutex);
555 release_pages(pvec, pinned, 0);
556 drm_free_large(pvec);
558 put_task_struct(work->task);
563 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj,
566 struct get_pages_work *work;
568 /* Spawn a worker so that we can acquire the
569 * user pages without holding our mutex. Access
570 * to the user pages requires mmap_sem, and we have
571 * a strict lock ordering of mmap_sem, struct_mutex -
572 * we already hold struct_mutex here and so cannot
573 * call gup without encountering a lock inversion.
575 * Userspace will keep on repeating the operation
576 * (thanks to EAGAIN) until either we hit the fast
577 * path or the worker completes. If the worker is
578 * cancelled or superseded, the task is still run
579 * but the results ignored. (This leads to
580 * complications that we may have a stray object
581 * refcount that we need to be wary of when
582 * checking for existing objects during creation.)
583 * If the worker encounters an error, it reports
584 * that error back to this function through
585 * obj->userptr.work = ERR_PTR.
587 if (obj->userptr.workers >= I915_GEM_USERPTR_MAX_WORKERS)
590 work = kmalloc(sizeof(*work), GFP_KERNEL);
594 obj->userptr.work = &work->work;
595 obj->userptr.workers++;
597 work->obj = i915_gem_object_get(obj);
599 work->task = current;
600 get_task_struct(work->task);
602 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
603 schedule_work(&work->work);
610 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
612 const int num_pages = obj->base.size >> PAGE_SHIFT;
617 /* If userspace should engineer that these pages are replaced in
618 * the vma between us binding this page into the GTT and completion
619 * of rendering... Their loss. If they change the mapping of their
620 * pages they need to create a new bo to point to the new vma.
622 * However, that still leaves open the possibility of the vma
623 * being copied upon fork. Which falls under the same userspace
624 * synchronisation issue as a regular bo, except that this time
625 * the process may not be expecting that a particular piece of
626 * memory is tied to the GPU.
628 * Fortunately, we can hook into the mmu_notifier in order to
629 * discard the page references prior to anything nasty happening
630 * to the vma (discard or cloning) which should prevent the more
631 * egregious cases from causing harm.
634 if (obj->userptr.work) {
635 /* active flag should still be held for the pending work */
636 if (IS_ERR(obj->userptr.work))
637 return PTR_ERR(obj->userptr.work);
642 /* Let the mmu-notifier know that we have begun and need cancellation */
643 ret = __i915_gem_userptr_set_active(obj, true);
649 if (obj->userptr.mm->mm == current->mm) {
650 pvec = drm_malloc_gfp(num_pages, sizeof(struct page *),
653 __i915_gem_userptr_set_active(obj, false);
657 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
658 !obj->userptr.read_only, pvec);
663 ret = pinned, pinned = 0;
664 else if (pinned < num_pages)
665 ret = __i915_gem_userptr_get_pages_schedule(obj, &active);
667 ret = __i915_gem_userptr_set_pages(obj, pvec, num_pages);
669 __i915_gem_userptr_set_active(obj, active);
670 release_pages(pvec, pinned, 0);
672 drm_free_large(pvec);
677 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj)
679 struct sgt_iter sgt_iter;
682 BUG_ON(obj->userptr.work != NULL);
683 __i915_gem_userptr_set_active(obj, false);
685 if (obj->madv != I915_MADV_WILLNEED)
688 i915_gem_gtt_finish_object(obj);
690 for_each_sgt_page(page, sgt_iter, obj->pages) {
692 set_page_dirty(page);
694 mark_page_accessed(page);
699 sg_free_table(obj->pages);
704 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
706 i915_gem_userptr_release__mmu_notifier(obj);
707 i915_gem_userptr_release__mm_struct(obj);
711 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
713 if (obj->userptr.mmu_object)
716 return i915_gem_userptr_init__mmu_notifier(obj, 0);
719 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
720 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,
721 .get_pages = i915_gem_userptr_get_pages,
722 .put_pages = i915_gem_userptr_put_pages,
723 .dmabuf_export = i915_gem_userptr_dmabuf_export,
724 .release = i915_gem_userptr_release,
728 * Creates a new mm object that wraps some normal memory from the process
729 * context - user memory.
731 * We impose several restrictions upon the memory being mapped
733 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
734 * 2. It must be normal system memory, not a pointer into another map of IO
735 * space (e.g. it must not be a GTT mmapping of another object).
736 * 3. We only allow a bo as large as we could in theory map into the GTT,
737 * that is we limit the size to the total size of the GTT.
738 * 4. The bo is marked as being snoopable. The backing pages are left
739 * accessible directly by the CPU, but reads and writes by the GPU may
740 * incur the cost of a snoop (unless you have an LLC architecture).
742 * Synchronisation between multiple users and the GPU is left to userspace
743 * through the normal set-domain-ioctl. The kernel will enforce that the
744 * GPU relinquishes the VMA before it is returned back to the system
745 * i.e. upon free(), munmap() or process termination. However, the userspace
746 * malloc() library may not immediately relinquish the VMA after free() and
747 * instead reuse it whilst the GPU is still reading and writing to the VMA.
750 * Also note, that the object created here is not currently a "first class"
751 * object, in that several ioctls are banned. These are the CPU access
752 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
753 * direct access via your pointer rather than use those ioctls. Another
754 * restriction is that we do not allow userptr surfaces to be pinned to the
755 * hardware and so we reject any attempt to create a framebuffer out of a
758 * If you think this is a good interface to use to pass GPU memory between
759 * drivers, please use dma-buf instead. In fact, wherever possible use
763 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
765 struct drm_i915_gem_userptr *args = data;
766 struct drm_i915_gem_object *obj;
770 if (!HAS_LLC(dev) && !HAS_SNOOP(dev)) {
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 (offset_in_page(args->user_ptr | args->user_size))
784 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
785 (char __user *)(unsigned long)args->user_ptr, args->user_size))
788 if (args->flags & I915_USERPTR_READ_ONLY) {
789 /* On almost all of the current hw, we cannot tell the GPU that a
790 * page is readonly, so this is just a placeholder in the uAPI.
795 obj = i915_gem_object_alloc(dev);
799 drm_gem_private_object_init(dev, &obj->base, args->user_size);
800 i915_gem_object_init(obj, &i915_gem_userptr_ops);
801 obj->cache_level = I915_CACHE_LLC;
802 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
803 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
805 obj->userptr.ptr = args->user_ptr;
806 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
808 /* And keep a pointer to the current->mm for resolving the user pages
809 * at binding. This means that we need to hook into the mmu_notifier
810 * in order to detect if the mmu is destroyed.
812 ret = i915_gem_userptr_init__mm_struct(obj);
814 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
816 ret = drm_gem_handle_create(file, &obj->base, &handle);
818 /* drop reference from allocate - handle holds it now */
819 i915_gem_object_put_unlocked(obj);
823 args->handle = handle;
827 void i915_gem_init_userptr(struct drm_i915_private *dev_priv)
829 mutex_init(&dev_priv->mm_lock);
830 hash_init(dev_priv->mm_structs);