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 struct drm_i915_private *dev_priv = to_i915(dev);
88 bool was_interruptible;
90 was_interruptible = dev_priv->mm.interruptible;
91 dev_priv->mm.interruptible = false;
93 WARN_ON(i915_gem_object_unbind(obj));
94 WARN_ON(i915_gem_object_put_pages(obj));
96 dev_priv->mm.interruptible = was_interruptible;
99 i915_gem_object_put(obj);
100 mutex_unlock(&dev->struct_mutex);
103 static void add_object(struct i915_mmu_object *mo)
108 interval_tree_insert(&mo->it, &mo->mn->objects);
112 static void del_object(struct i915_mmu_object *mo)
117 interval_tree_remove(&mo->it, &mo->mn->objects);
118 mo->attached = false;
121 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
122 struct mm_struct *mm,
126 struct i915_mmu_notifier *mn =
127 container_of(_mn, struct i915_mmu_notifier, mn);
128 struct i915_mmu_object *mo;
129 struct interval_tree_node *it;
130 LIST_HEAD(cancelled);
132 if (RB_EMPTY_ROOT(&mn->objects))
135 /* interval ranges are inclusive, but invalidate range is exclusive */
138 spin_lock(&mn->lock);
139 it = interval_tree_iter_first(&mn->objects, start, end);
141 /* The mmu_object is released late when destroying the
142 * GEM object so it is entirely possible to gain a
143 * reference on an object in the process of being freed
144 * since our serialisation is via the spinlock and not
145 * the struct_mutex - and consequently use it after it
146 * is freed and then double free it. To prevent that
147 * use-after-free we only acquire a reference on the
148 * object if it is not in the process of being destroyed.
150 mo = container_of(it, struct i915_mmu_object, it);
151 if (kref_get_unless_zero(&mo->obj->base.refcount))
152 queue_work(mn->wq, &mo->work);
154 list_add(&mo->link, &cancelled);
155 it = interval_tree_iter_next(it, start, end);
157 list_for_each_entry(mo, &cancelled, link)
159 spin_unlock(&mn->lock);
161 flush_workqueue(mn->wq);
164 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
165 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
168 static struct i915_mmu_notifier *
169 i915_mmu_notifier_create(struct mm_struct *mm)
171 struct i915_mmu_notifier *mn;
174 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
176 return ERR_PTR(-ENOMEM);
178 spin_lock_init(&mn->lock);
179 mn->mn.ops = &i915_gem_userptr_notifier;
180 mn->objects = RB_ROOT;
181 mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0);
182 if (mn->wq == NULL) {
184 return ERR_PTR(-ENOMEM);
187 /* Protected by mmap_sem (write-lock) */
188 ret = __mmu_notifier_register(&mn->mn, mm);
190 destroy_workqueue(mn->wq);
199 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
201 struct i915_mmu_object *mo;
203 mo = obj->userptr.mmu_object;
207 spin_lock(&mo->mn->lock);
209 spin_unlock(&mo->mn->lock);
212 obj->userptr.mmu_object = NULL;
215 static struct i915_mmu_notifier *
216 i915_mmu_notifier_find(struct i915_mm_struct *mm)
218 struct i915_mmu_notifier *mn = mm->mn;
224 down_write(&mm->mm->mmap_sem);
225 mutex_lock(&mm->i915->mm_lock);
226 if ((mn = mm->mn) == NULL) {
227 mn = i915_mmu_notifier_create(mm->mm);
231 mutex_unlock(&mm->i915->mm_lock);
232 up_write(&mm->mm->mmap_sem);
238 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
241 struct i915_mmu_notifier *mn;
242 struct i915_mmu_object *mo;
244 if (flags & I915_USERPTR_UNSYNCHRONIZED)
245 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
247 if (WARN_ON(obj->userptr.mm == NULL))
250 mn = i915_mmu_notifier_find(obj->userptr.mm);
254 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
260 mo->it.start = obj->userptr.ptr;
261 mo->it.last = obj->userptr.ptr + obj->base.size - 1;
262 INIT_WORK(&mo->work, cancel_userptr);
264 obj->userptr.mmu_object = mo;
269 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
270 struct mm_struct *mm)
275 mmu_notifier_unregister(&mn->mn, mm);
276 destroy_workqueue(mn->wq);
283 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
288 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
291 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
294 if (!capable(CAP_SYS_ADMIN))
301 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
302 struct mm_struct *mm)
308 static struct i915_mm_struct *
309 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
311 struct i915_mm_struct *mm;
313 /* Protected by dev_priv->mm_lock */
314 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
322 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
324 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
325 struct i915_mm_struct *mm;
328 /* During release of the GEM object we hold the struct_mutex. This
329 * precludes us from calling mmput() at that time as that may be
330 * the last reference and so call exit_mmap(). exit_mmap() will
331 * attempt to reap the vma, and if we were holding a GTT mmap
332 * would then call drm_gem_vm_close() and attempt to reacquire
333 * the struct mutex. So in order to avoid that recursion, we have
334 * to defer releasing the mm reference until after we drop the
335 * struct_mutex, i.e. we need to schedule a worker to do the clean
338 mutex_lock(&dev_priv->mm_lock);
339 mm = __i915_mm_struct_find(dev_priv, current->mm);
341 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
347 kref_init(&mm->kref);
348 mm->i915 = to_i915(obj->base.dev);
350 mm->mm = current->mm;
351 atomic_inc(¤t->mm->mm_count);
355 /* Protected by dev_priv->mm_lock */
356 hash_add(dev_priv->mm_structs,
357 &mm->node, (unsigned long)mm->mm);
361 obj->userptr.mm = mm;
363 mutex_unlock(&dev_priv->mm_lock);
368 __i915_mm_struct_free__worker(struct work_struct *work)
370 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
371 i915_mmu_notifier_free(mm->mn, mm->mm);
377 __i915_mm_struct_free(struct kref *kref)
379 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
381 /* Protected by dev_priv->mm_lock */
383 mutex_unlock(&mm->i915->mm_lock);
385 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
386 schedule_work(&mm->work);
390 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
392 if (obj->userptr.mm == NULL)
395 kref_put_mutex(&obj->userptr.mm->kref,
396 __i915_mm_struct_free,
397 &to_i915(obj->base.dev)->mm_lock);
398 obj->userptr.mm = NULL;
401 struct get_pages_work {
402 struct work_struct work;
403 struct drm_i915_gem_object *obj;
404 struct task_struct *task;
407 #if IS_ENABLED(CONFIG_SWIOTLB)
408 #define swiotlb_active() swiotlb_nr_tbl()
410 #define swiotlb_active() 0
414 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
416 struct scatterlist *sg;
419 *st = kmalloc(sizeof(**st), GFP_KERNEL);
423 if (swiotlb_active()) {
424 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
428 for_each_sg((*st)->sgl, sg, num_pages, n)
429 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
431 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
432 0, num_pages << PAGE_SHIFT,
447 __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj,
448 struct page **pvec, int num_pages)
452 ret = st_set_pages(&obj->pages, pvec, num_pages);
456 ret = i915_gem_gtt_prepare_object(obj);
458 sg_free_table(obj->pages);
467 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
472 /* During mm_invalidate_range we need to cancel any userptr that
473 * overlaps the range being invalidated. Doing so requires the
474 * struct_mutex, and that risks recursion. In order to cause
475 * recursion, the user must alias the userptr address space with
476 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
477 * to invalidate that mmaping, mm_invalidate_range is called with
478 * the userptr address *and* the struct_mutex held. To prevent that
479 * we set a flag under the i915_mmu_notifier spinlock to indicate
480 * whether this object is valid.
482 #if defined(CONFIG_MMU_NOTIFIER)
483 if (obj->userptr.mmu_object == NULL)
486 spin_lock(&obj->userptr.mmu_object->mn->lock);
487 /* In order to serialise get_pages with an outstanding
488 * cancel_userptr, we must drop the struct_mutex and try again.
491 del_object(obj->userptr.mmu_object);
492 else if (!work_pending(&obj->userptr.mmu_object->work))
493 add_object(obj->userptr.mmu_object);
496 spin_unlock(&obj->userptr.mmu_object->mn->lock);
503 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
505 struct get_pages_work *work = container_of(_work, typeof(*work), work);
506 struct drm_i915_gem_object *obj = work->obj;
507 struct drm_device *dev = obj->base.dev;
508 const int npages = obj->base.size >> PAGE_SHIFT;
515 pvec = drm_malloc_gfp(npages, sizeof(struct page *), GFP_TEMPORARY);
517 struct mm_struct *mm = obj->userptr.mm->mm;
520 if (atomic_inc_not_zero(&mm->mm_users)) {
521 down_read(&mm->mmap_sem);
522 while (pinned < npages) {
523 ret = get_user_pages_remote
525 obj->userptr.ptr + pinned * PAGE_SIZE,
527 !obj->userptr.read_only, 0,
528 pvec + pinned, NULL);
534 up_read(&mm->mmap_sem);
539 mutex_lock(&dev->struct_mutex);
540 if (obj->userptr.work == &work->work) {
541 if (pinned == npages) {
542 ret = __i915_gem_userptr_set_pages(obj, pvec, npages);
544 list_add_tail(&obj->global_list,
545 &to_i915(dev)->mm.unbound_list);
546 obj->get_page.sg = obj->pages->sgl;
547 obj->get_page.last = 0;
551 obj->userptr.work = ERR_PTR(ret);
553 __i915_gem_userptr_set_active(obj, false);
556 obj->userptr.workers--;
557 i915_gem_object_put(obj);
558 mutex_unlock(&dev->struct_mutex);
560 release_pages(pvec, pinned, 0);
561 drm_free_large(pvec);
563 put_task_struct(work->task);
568 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj,
571 struct get_pages_work *work;
573 /* Spawn a worker so that we can acquire the
574 * user pages without holding our mutex. Access
575 * to the user pages requires mmap_sem, and we have
576 * a strict lock ordering of mmap_sem, struct_mutex -
577 * we already hold struct_mutex here and so cannot
578 * call gup without encountering a lock inversion.
580 * Userspace will keep on repeating the operation
581 * (thanks to EAGAIN) until either we hit the fast
582 * path or the worker completes. If the worker is
583 * cancelled or superseded, the task is still run
584 * but the results ignored. (This leads to
585 * complications that we may have a stray object
586 * refcount that we need to be wary of when
587 * checking for existing objects during creation.)
588 * If the worker encounters an error, it reports
589 * that error back to this function through
590 * obj->userptr.work = ERR_PTR.
592 if (obj->userptr.workers >= I915_GEM_USERPTR_MAX_WORKERS)
595 work = kmalloc(sizeof(*work), GFP_KERNEL);
599 obj->userptr.work = &work->work;
600 obj->userptr.workers++;
602 work->obj = i915_gem_object_get(obj);
604 work->task = current;
605 get_task_struct(work->task);
607 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
608 schedule_work(&work->work);
615 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
617 const int num_pages = obj->base.size >> PAGE_SHIFT;
622 /* If userspace should engineer that these pages are replaced in
623 * the vma between us binding this page into the GTT and completion
624 * of rendering... Their loss. If they change the mapping of their
625 * pages they need to create a new bo to point to the new vma.
627 * However, that still leaves open the possibility of the vma
628 * being copied upon fork. Which falls under the same userspace
629 * synchronisation issue as a regular bo, except that this time
630 * the process may not be expecting that a particular piece of
631 * memory is tied to the GPU.
633 * Fortunately, we can hook into the mmu_notifier in order to
634 * discard the page references prior to anything nasty happening
635 * to the vma (discard or cloning) which should prevent the more
636 * egregious cases from causing harm.
638 if (IS_ERR(obj->userptr.work)) {
639 /* active flag will have been dropped already by the worker */
640 ret = PTR_ERR(obj->userptr.work);
641 obj->userptr.work = NULL;
644 if (obj->userptr.work)
645 /* active flag should still be held for the pending work */
648 /* Let the mmu-notifier know that we have begun and need cancellation */
649 ret = __i915_gem_userptr_set_active(obj, true);
655 if (obj->userptr.mm->mm == current->mm) {
656 pvec = drm_malloc_gfp(num_pages, sizeof(struct page *),
659 __i915_gem_userptr_set_active(obj, false);
663 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
664 !obj->userptr.read_only, pvec);
669 ret = pinned, pinned = 0;
670 else if (pinned < num_pages)
671 ret = __i915_gem_userptr_get_pages_schedule(obj, &active);
673 ret = __i915_gem_userptr_set_pages(obj, pvec, num_pages);
675 __i915_gem_userptr_set_active(obj, active);
676 release_pages(pvec, pinned, 0);
678 drm_free_large(pvec);
683 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj)
685 struct sgt_iter sgt_iter;
688 BUG_ON(obj->userptr.work != NULL);
689 __i915_gem_userptr_set_active(obj, false);
691 if (obj->madv != I915_MADV_WILLNEED)
694 i915_gem_gtt_finish_object(obj);
696 for_each_sgt_page(page, sgt_iter, obj->pages) {
698 set_page_dirty(page);
700 mark_page_accessed(page);
705 sg_free_table(obj->pages);
710 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
712 i915_gem_userptr_release__mmu_notifier(obj);
713 i915_gem_userptr_release__mm_struct(obj);
717 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
719 if (obj->userptr.mmu_object)
722 return i915_gem_userptr_init__mmu_notifier(obj, 0);
725 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
726 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,
727 .get_pages = i915_gem_userptr_get_pages,
728 .put_pages = i915_gem_userptr_put_pages,
729 .dmabuf_export = i915_gem_userptr_dmabuf_export,
730 .release = i915_gem_userptr_release,
734 * Creates a new mm object that wraps some normal memory from the process
735 * context - user memory.
737 * We impose several restrictions upon the memory being mapped
739 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
740 * 2. It must be normal system memory, not a pointer into another map of IO
741 * space (e.g. it must not be a GTT mmapping of another object).
742 * 3. We only allow a bo as large as we could in theory map into the GTT,
743 * that is we limit the size to the total size of the GTT.
744 * 4. The bo is marked as being snoopable. The backing pages are left
745 * accessible directly by the CPU, but reads and writes by the GPU may
746 * incur the cost of a snoop (unless you have an LLC architecture).
748 * Synchronisation between multiple users and the GPU is left to userspace
749 * through the normal set-domain-ioctl. The kernel will enforce that the
750 * GPU relinquishes the VMA before it is returned back to the system
751 * i.e. upon free(), munmap() or process termination. However, the userspace
752 * malloc() library may not immediately relinquish the VMA after free() and
753 * instead reuse it whilst the GPU is still reading and writing to the VMA.
756 * Also note, that the object created here is not currently a "first class"
757 * object, in that several ioctls are banned. These are the CPU access
758 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
759 * direct access via your pointer rather than use those ioctls. Another
760 * restriction is that we do not allow userptr surfaces to be pinned to the
761 * hardware and so we reject any attempt to create a framebuffer out of a
764 * If you think this is a good interface to use to pass GPU memory between
765 * drivers, please use dma-buf instead. In fact, wherever possible use
769 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
771 struct drm_i915_gem_userptr *args = data;
772 struct drm_i915_gem_object *obj;
776 if (!HAS_LLC(dev) && !HAS_SNOOP(dev)) {
777 /* We cannot support coherent userptr objects on hw without
778 * LLC and broken snooping.
783 if (args->flags & ~(I915_USERPTR_READ_ONLY |
784 I915_USERPTR_UNSYNCHRONIZED))
787 if (offset_in_page(args->user_ptr | args->user_size))
790 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
791 (char __user *)(unsigned long)args->user_ptr, args->user_size))
794 if (args->flags & I915_USERPTR_READ_ONLY) {
795 /* On almost all of the current hw, we cannot tell the GPU that a
796 * page is readonly, so this is just a placeholder in the uAPI.
801 obj = i915_gem_object_alloc(dev);
805 drm_gem_private_object_init(dev, &obj->base, args->user_size);
806 i915_gem_object_init(obj, &i915_gem_userptr_ops);
807 obj->cache_level = I915_CACHE_LLC;
808 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
809 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
811 obj->userptr.ptr = args->user_ptr;
812 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
814 /* And keep a pointer to the current->mm for resolving the user pages
815 * at binding. This means that we need to hook into the mmu_notifier
816 * in order to detect if the mmu is destroyed.
818 ret = i915_gem_userptr_init__mm_struct(obj);
820 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
822 ret = drm_gem_handle_create(file, &obj->base, &handle);
824 /* drop reference from allocate - handle holds it now */
825 i915_gem_object_put_unlocked(obj);
829 args->handle = handle;
833 void i915_gem_init_userptr(struct drm_i915_private *dev_priv)
835 mutex_init(&dev_priv->mm_lock);
836 hash_init(dev_priv->mm_structs);