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 struct drm_device *dev = obj->base.dev;
67 struct drm_i915_gem_request *requests[I915_NUM_ENGINES];
74 for (i = 0; i < I915_NUM_ENGINES; i++) {
75 struct drm_i915_gem_request *req;
77 req = obj->last_read_req[i];
81 requests[n++] = i915_gem_request_reference(req);
84 mutex_unlock(&dev->struct_mutex);
86 for (i = 0; i < n; i++)
87 __i915_wait_request(requests[i], false, NULL, NULL);
89 mutex_lock(&dev->struct_mutex);
91 for (i = 0; i < n; i++)
92 i915_gem_request_unreference(requests[i]);
95 static void cancel_userptr(struct work_struct *work)
97 struct i915_mmu_object *mo = container_of(work, typeof(*mo), work);
98 struct drm_i915_gem_object *obj = mo->obj;
99 struct drm_device *dev = obj->base.dev;
101 mutex_lock(&dev->struct_mutex);
102 /* Cancel any active worker and force us to re-evaluate gup */
103 obj->userptr.work = NULL;
105 if (obj->pages != NULL) {
106 struct drm_i915_private *dev_priv = to_i915(dev);
107 struct i915_vma *vma, *tmp;
108 bool was_interruptible;
112 was_interruptible = dev_priv->mm.interruptible;
113 dev_priv->mm.interruptible = false;
115 list_for_each_entry_safe(vma, tmp, &obj->vma_list, obj_link)
116 WARN_ON(i915_vma_unbind(vma));
117 WARN_ON(i915_gem_object_put_pages(obj));
119 dev_priv->mm.interruptible = was_interruptible;
122 drm_gem_object_unreference(&obj->base);
123 mutex_unlock(&dev->struct_mutex);
126 static void add_object(struct i915_mmu_object *mo)
131 interval_tree_insert(&mo->it, &mo->mn->objects);
135 static void del_object(struct i915_mmu_object *mo)
140 interval_tree_remove(&mo->it, &mo->mn->objects);
141 mo->attached = false;
144 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
145 struct mm_struct *mm,
149 struct i915_mmu_notifier *mn =
150 container_of(_mn, struct i915_mmu_notifier, mn);
151 struct i915_mmu_object *mo;
152 struct interval_tree_node *it;
153 LIST_HEAD(cancelled);
155 if (RB_EMPTY_ROOT(&mn->objects))
158 /* interval ranges are inclusive, but invalidate range is exclusive */
161 spin_lock(&mn->lock);
162 it = interval_tree_iter_first(&mn->objects, start, end);
164 /* The mmu_object is released late when destroying the
165 * GEM object so it is entirely possible to gain a
166 * reference on an object in the process of being freed
167 * since our serialisation is via the spinlock and not
168 * the struct_mutex - and consequently use it after it
169 * is freed and then double free it. To prevent that
170 * use-after-free we only acquire a reference on the
171 * object if it is not in the process of being destroyed.
173 mo = container_of(it, struct i915_mmu_object, it);
174 if (kref_get_unless_zero(&mo->obj->base.refcount))
175 queue_work(mn->wq, &mo->work);
177 list_add(&mo->link, &cancelled);
178 it = interval_tree_iter_next(it, start, end);
180 list_for_each_entry(mo, &cancelled, link)
182 spin_unlock(&mn->lock);
184 flush_workqueue(mn->wq);
187 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
188 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
191 static struct i915_mmu_notifier *
192 i915_mmu_notifier_create(struct mm_struct *mm)
194 struct i915_mmu_notifier *mn;
197 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
199 return ERR_PTR(-ENOMEM);
201 spin_lock_init(&mn->lock);
202 mn->mn.ops = &i915_gem_userptr_notifier;
203 mn->objects = RB_ROOT;
204 mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0);
205 if (mn->wq == NULL) {
207 return ERR_PTR(-ENOMEM);
210 /* Protected by mmap_sem (write-lock) */
211 ret = __mmu_notifier_register(&mn->mn, mm);
213 destroy_workqueue(mn->wq);
222 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
224 struct i915_mmu_object *mo;
226 mo = obj->userptr.mmu_object;
230 spin_lock(&mo->mn->lock);
232 spin_unlock(&mo->mn->lock);
235 obj->userptr.mmu_object = NULL;
238 static struct i915_mmu_notifier *
239 i915_mmu_notifier_find(struct i915_mm_struct *mm)
241 struct i915_mmu_notifier *mn = mm->mn;
247 down_write(&mm->mm->mmap_sem);
248 mutex_lock(&mm->i915->mm_lock);
249 if ((mn = mm->mn) == NULL) {
250 mn = i915_mmu_notifier_create(mm->mm);
254 mutex_unlock(&mm->i915->mm_lock);
255 up_write(&mm->mm->mmap_sem);
261 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
264 struct i915_mmu_notifier *mn;
265 struct i915_mmu_object *mo;
267 if (flags & I915_USERPTR_UNSYNCHRONIZED)
268 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
270 if (WARN_ON(obj->userptr.mm == NULL))
273 mn = i915_mmu_notifier_find(obj->userptr.mm);
277 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
283 mo->it.start = obj->userptr.ptr;
284 mo->it.last = obj->userptr.ptr + obj->base.size - 1;
285 INIT_WORK(&mo->work, cancel_userptr);
287 obj->userptr.mmu_object = mo;
292 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
293 struct mm_struct *mm)
298 mmu_notifier_unregister(&mn->mn, mm);
299 destroy_workqueue(mn->wq);
306 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
311 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
314 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
317 if (!capable(CAP_SYS_ADMIN))
324 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
325 struct mm_struct *mm)
331 static struct i915_mm_struct *
332 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
334 struct i915_mm_struct *mm;
336 /* Protected by dev_priv->mm_lock */
337 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
345 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
347 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
348 struct i915_mm_struct *mm;
351 /* During release of the GEM object we hold the struct_mutex. This
352 * precludes us from calling mmput() at that time as that may be
353 * the last reference and so call exit_mmap(). exit_mmap() will
354 * attempt to reap the vma, and if we were holding a GTT mmap
355 * would then call drm_gem_vm_close() and attempt to reacquire
356 * the struct mutex. So in order to avoid that recursion, we have
357 * to defer releasing the mm reference until after we drop the
358 * struct_mutex, i.e. we need to schedule a worker to do the clean
361 mutex_lock(&dev_priv->mm_lock);
362 mm = __i915_mm_struct_find(dev_priv, current->mm);
364 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
370 kref_init(&mm->kref);
371 mm->i915 = to_i915(obj->base.dev);
373 mm->mm = current->mm;
374 atomic_inc(¤t->mm->mm_count);
378 /* Protected by dev_priv->mm_lock */
379 hash_add(dev_priv->mm_structs,
380 &mm->node, (unsigned long)mm->mm);
384 obj->userptr.mm = mm;
386 mutex_unlock(&dev_priv->mm_lock);
391 __i915_mm_struct_free__worker(struct work_struct *work)
393 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
394 i915_mmu_notifier_free(mm->mn, mm->mm);
400 __i915_mm_struct_free(struct kref *kref)
402 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
404 /* Protected by dev_priv->mm_lock */
406 mutex_unlock(&mm->i915->mm_lock);
408 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
409 schedule_work(&mm->work);
413 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
415 if (obj->userptr.mm == NULL)
418 kref_put_mutex(&obj->userptr.mm->kref,
419 __i915_mm_struct_free,
420 &to_i915(obj->base.dev)->mm_lock);
421 obj->userptr.mm = NULL;
424 struct get_pages_work {
425 struct work_struct work;
426 struct drm_i915_gem_object *obj;
427 struct task_struct *task;
430 #if IS_ENABLED(CONFIG_SWIOTLB)
431 #define swiotlb_active() swiotlb_nr_tbl()
433 #define swiotlb_active() 0
437 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
439 struct scatterlist *sg;
442 *st = kmalloc(sizeof(**st), GFP_KERNEL);
446 if (swiotlb_active()) {
447 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
451 for_each_sg((*st)->sgl, sg, num_pages, n)
452 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
454 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
455 0, num_pages << PAGE_SHIFT,
470 __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj,
471 struct page **pvec, int num_pages)
475 ret = st_set_pages(&obj->pages, pvec, num_pages);
479 ret = i915_gem_gtt_prepare_object(obj);
481 sg_free_table(obj->pages);
490 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
495 /* During mm_invalidate_range we need to cancel any userptr that
496 * overlaps the range being invalidated. Doing so requires the
497 * struct_mutex, and that risks recursion. In order to cause
498 * recursion, the user must alias the userptr address space with
499 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
500 * to invalidate that mmaping, mm_invalidate_range is called with
501 * the userptr address *and* the struct_mutex held. To prevent that
502 * we set a flag under the i915_mmu_notifier spinlock to indicate
503 * whether this object is valid.
505 #if defined(CONFIG_MMU_NOTIFIER)
506 if (obj->userptr.mmu_object == NULL)
509 spin_lock(&obj->userptr.mmu_object->mn->lock);
510 /* In order to serialise get_pages with an outstanding
511 * cancel_userptr, we must drop the struct_mutex and try again.
514 del_object(obj->userptr.mmu_object);
515 else if (!work_pending(&obj->userptr.mmu_object->work))
516 add_object(obj->userptr.mmu_object);
519 spin_unlock(&obj->userptr.mmu_object->mn->lock);
526 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
528 struct get_pages_work *work = container_of(_work, typeof(*work), work);
529 struct drm_i915_gem_object *obj = work->obj;
530 struct drm_device *dev = obj->base.dev;
531 const int npages = obj->base.size >> PAGE_SHIFT;
538 pvec = drm_malloc_gfp(npages, sizeof(struct page *), GFP_TEMPORARY);
540 struct mm_struct *mm = obj->userptr.mm->mm;
543 if (atomic_inc_not_zero(&mm->mm_users)) {
544 down_read(&mm->mmap_sem);
545 while (pinned < npages) {
546 ret = get_user_pages_remote
548 obj->userptr.ptr + pinned * PAGE_SIZE,
550 !obj->userptr.read_only, 0,
551 pvec + pinned, NULL);
557 up_read(&mm->mmap_sem);
562 mutex_lock(&dev->struct_mutex);
563 if (obj->userptr.work == &work->work) {
564 if (pinned == npages) {
565 ret = __i915_gem_userptr_set_pages(obj, pvec, npages);
567 list_add_tail(&obj->global_list,
568 &to_i915(dev)->mm.unbound_list);
569 obj->get_page.sg = obj->pages->sgl;
570 obj->get_page.last = 0;
574 obj->userptr.work = ERR_PTR(ret);
576 __i915_gem_userptr_set_active(obj, false);
579 obj->userptr.workers--;
580 drm_gem_object_unreference(&obj->base);
581 mutex_unlock(&dev->struct_mutex);
583 release_pages(pvec, pinned, 0);
584 drm_free_large(pvec);
586 put_task_struct(work->task);
591 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj,
594 struct get_pages_work *work;
596 /* Spawn a worker so that we can acquire the
597 * user pages without holding our mutex. Access
598 * to the user pages requires mmap_sem, and we have
599 * a strict lock ordering of mmap_sem, struct_mutex -
600 * we already hold struct_mutex here and so cannot
601 * call gup without encountering a lock inversion.
603 * Userspace will keep on repeating the operation
604 * (thanks to EAGAIN) until either we hit the fast
605 * path or the worker completes. If the worker is
606 * cancelled or superseded, the task is still run
607 * but the results ignored. (This leads to
608 * complications that we may have a stray object
609 * refcount that we need to be wary of when
610 * checking for existing objects during creation.)
611 * If the worker encounters an error, it reports
612 * that error back to this function through
613 * obj->userptr.work = ERR_PTR.
615 if (obj->userptr.workers >= I915_GEM_USERPTR_MAX_WORKERS)
618 work = kmalloc(sizeof(*work), GFP_KERNEL);
622 obj->userptr.work = &work->work;
623 obj->userptr.workers++;
626 drm_gem_object_reference(&obj->base);
628 work->task = current;
629 get_task_struct(work->task);
631 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
632 schedule_work(&work->work);
639 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
641 const int num_pages = obj->base.size >> PAGE_SHIFT;
646 /* If userspace should engineer that these pages are replaced in
647 * the vma between us binding this page into the GTT and completion
648 * of rendering... Their loss. If they change the mapping of their
649 * pages they need to create a new bo to point to the new vma.
651 * However, that still leaves open the possibility of the vma
652 * being copied upon fork. Which falls under the same userspace
653 * synchronisation issue as a regular bo, except that this time
654 * the process may not be expecting that a particular piece of
655 * memory is tied to the GPU.
657 * Fortunately, we can hook into the mmu_notifier in order to
658 * discard the page references prior to anything nasty happening
659 * to the vma (discard or cloning) which should prevent the more
660 * egregious cases from causing harm.
662 if (IS_ERR(obj->userptr.work)) {
663 /* active flag will have been dropped already by the worker */
664 ret = PTR_ERR(obj->userptr.work);
665 obj->userptr.work = NULL;
668 if (obj->userptr.work)
669 /* active flag should still be held for the pending work */
672 /* Let the mmu-notifier know that we have begun and need cancellation */
673 ret = __i915_gem_userptr_set_active(obj, true);
679 if (obj->userptr.mm->mm == current->mm) {
680 pvec = drm_malloc_gfp(num_pages, sizeof(struct page *),
683 __i915_gem_userptr_set_active(obj, false);
687 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
688 !obj->userptr.read_only, pvec);
693 ret = pinned, pinned = 0;
694 else if (pinned < num_pages)
695 ret = __i915_gem_userptr_get_pages_schedule(obj, &active);
697 ret = __i915_gem_userptr_set_pages(obj, pvec, num_pages);
699 __i915_gem_userptr_set_active(obj, active);
700 release_pages(pvec, pinned, 0);
702 drm_free_large(pvec);
707 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj)
709 struct sgt_iter sgt_iter;
712 BUG_ON(obj->userptr.work != NULL);
713 __i915_gem_userptr_set_active(obj, false);
715 if (obj->madv != I915_MADV_WILLNEED)
718 i915_gem_gtt_finish_object(obj);
720 for_each_sgt_page(page, sgt_iter, obj->pages) {
722 set_page_dirty(page);
724 mark_page_accessed(page);
729 sg_free_table(obj->pages);
734 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
736 i915_gem_userptr_release__mmu_notifier(obj);
737 i915_gem_userptr_release__mm_struct(obj);
741 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
743 if (obj->userptr.mmu_object)
746 return i915_gem_userptr_init__mmu_notifier(obj, 0);
749 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
750 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,
751 .get_pages = i915_gem_userptr_get_pages,
752 .put_pages = i915_gem_userptr_put_pages,
753 .dmabuf_export = i915_gem_userptr_dmabuf_export,
754 .release = i915_gem_userptr_release,
758 * Creates a new mm object that wraps some normal memory from the process
759 * context - user memory.
761 * We impose several restrictions upon the memory being mapped
763 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
764 * 2. It must be normal system memory, not a pointer into another map of IO
765 * space (e.g. it must not be a GTT mmapping of another object).
766 * 3. We only allow a bo as large as we could in theory map into the GTT,
767 * that is we limit the size to the total size of the GTT.
768 * 4. The bo is marked as being snoopable. The backing pages are left
769 * accessible directly by the CPU, but reads and writes by the GPU may
770 * incur the cost of a snoop (unless you have an LLC architecture).
772 * Synchronisation between multiple users and the GPU is left to userspace
773 * through the normal set-domain-ioctl. The kernel will enforce that the
774 * GPU relinquishes the VMA before it is returned back to the system
775 * i.e. upon free(), munmap() or process termination. However, the userspace
776 * malloc() library may not immediately relinquish the VMA after free() and
777 * instead reuse it whilst the GPU is still reading and writing to the VMA.
780 * Also note, that the object created here is not currently a "first class"
781 * object, in that several ioctls are banned. These are the CPU access
782 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
783 * direct access via your pointer rather than use those ioctls. Another
784 * restriction is that we do not allow userptr surfaces to be pinned to the
785 * hardware and so we reject any attempt to create a framebuffer out of a
788 * If you think this is a good interface to use to pass GPU memory between
789 * drivers, please use dma-buf instead. In fact, wherever possible use
793 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
795 struct drm_i915_gem_userptr *args = data;
796 struct drm_i915_gem_object *obj;
800 if (!HAS_LLC(dev) && !HAS_SNOOP(dev)) {
801 /* We cannot support coherent userptr objects on hw without
802 * LLC and broken snooping.
807 if (args->flags & ~(I915_USERPTR_READ_ONLY |
808 I915_USERPTR_UNSYNCHRONIZED))
811 if (offset_in_page(args->user_ptr | args->user_size))
814 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
815 (char __user *)(unsigned long)args->user_ptr, args->user_size))
818 if (args->flags & I915_USERPTR_READ_ONLY) {
819 /* On almost all of the current hw, we cannot tell the GPU that a
820 * page is readonly, so this is just a placeholder in the uAPI.
825 obj = i915_gem_object_alloc(dev);
829 drm_gem_private_object_init(dev, &obj->base, args->user_size);
830 i915_gem_object_init(obj, &i915_gem_userptr_ops);
831 obj->cache_level = I915_CACHE_LLC;
832 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
833 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
835 obj->userptr.ptr = args->user_ptr;
836 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
838 /* And keep a pointer to the current->mm for resolving the user pages
839 * at binding. This means that we need to hook into the mmu_notifier
840 * in order to detect if the mmu is destroyed.
842 ret = i915_gem_userptr_init__mm_struct(obj);
844 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
846 ret = drm_gem_handle_create(file, &obj->base, &handle);
848 /* drop reference from allocate - handle holds it now */
849 drm_gem_object_unreference_unlocked(&obj->base);
853 args->handle = handle;
857 void i915_gem_init_userptr(struct drm_i915_private *dev_priv)
859 mutex_init(&dev_priv->mm_lock);
860 hash_init(dev_priv->mm_structs);