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 #if defined(CONFIG_MMU_NOTIFIER)
36 #include <linux/interval_tree.h>
38 struct i915_mmu_notifier {
40 struct hlist_node node;
41 struct mmu_notifier mn;
42 struct rb_root objects;
43 struct list_head linear;
44 struct drm_device *dev;
46 struct work_struct work;
52 struct i915_mmu_object {
53 struct i915_mmu_notifier *mmu;
54 struct interval_tree_node it;
55 struct list_head link;
56 struct drm_i915_gem_object *obj;
60 static unsigned long cancel_userptr(struct drm_i915_gem_object *obj)
62 struct drm_device *dev = obj->base.dev;
65 mutex_lock(&dev->struct_mutex);
66 /* Cancel any active worker and force us to re-evaluate gup */
67 obj->userptr.work = NULL;
69 if (obj->pages != NULL) {
70 struct drm_i915_private *dev_priv = to_i915(dev);
71 struct i915_vma *vma, *tmp;
72 bool was_interruptible;
74 was_interruptible = dev_priv->mm.interruptible;
75 dev_priv->mm.interruptible = false;
77 list_for_each_entry_safe(vma, tmp, &obj->vma_list, vma_link) {
78 int ret = i915_vma_unbind(vma);
79 WARN_ON(ret && ret != -EIO);
81 WARN_ON(i915_gem_object_put_pages(obj));
83 dev_priv->mm.interruptible = was_interruptible;
86 end = obj->userptr.ptr + obj->base.size;
88 drm_gem_object_unreference(&obj->base);
89 mutex_unlock(&dev->struct_mutex);
94 static void *invalidate_range__linear(struct i915_mmu_notifier *mn,
99 struct i915_mmu_object *mmu;
100 unsigned long serial;
104 list_for_each_entry(mmu, &mn->linear, link) {
105 struct drm_i915_gem_object *obj;
107 if (mmu->it.last < start || mmu->it.start > end)
111 drm_gem_object_reference(&obj->base);
112 spin_unlock(&mn->lock);
116 spin_lock(&mn->lock);
117 if (serial != mn->serial)
124 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
125 struct mm_struct *mm,
129 struct i915_mmu_notifier *mn = container_of(_mn, struct i915_mmu_notifier, mn);
130 struct interval_tree_node *it = NULL;
131 unsigned long next = start;
132 unsigned long serial = 0;
134 end--; /* interval ranges are inclusive, but invalidate range is exclusive */
136 struct drm_i915_gem_object *obj = NULL;
138 spin_lock(&mn->lock);
140 it = invalidate_range__linear(mn, mm, start, end);
141 else if (serial == mn->serial)
142 it = interval_tree_iter_next(it, next, end);
144 it = interval_tree_iter_first(&mn->objects, start, end);
146 obj = container_of(it, struct i915_mmu_object, it)->obj;
147 drm_gem_object_reference(&obj->base);
150 spin_unlock(&mn->lock);
154 next = cancel_userptr(obj);
158 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
159 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
162 static struct i915_mmu_notifier *
163 __i915_mmu_notifier_lookup(struct drm_device *dev, struct mm_struct *mm)
165 struct drm_i915_private *dev_priv = to_i915(dev);
166 struct i915_mmu_notifier *mmu;
168 /* Protected by dev->struct_mutex */
169 hash_for_each_possible(dev_priv->mmu_notifiers, mmu, node, (unsigned long)mm)
176 static struct i915_mmu_notifier *
177 i915_mmu_notifier_get(struct drm_device *dev, struct mm_struct *mm)
179 struct drm_i915_private *dev_priv = to_i915(dev);
180 struct i915_mmu_notifier *mmu;
183 lockdep_assert_held(&dev->struct_mutex);
185 mmu = __i915_mmu_notifier_lookup(dev, mm);
189 mmu = kmalloc(sizeof(*mmu), GFP_KERNEL);
191 return ERR_PTR(-ENOMEM);
193 spin_lock_init(&mmu->lock);
195 mmu->mn.ops = &i915_gem_userptr_notifier;
197 mmu->objects = RB_ROOT;
200 INIT_LIST_HEAD(&mmu->linear);
201 mmu->has_linear = false;
203 /* Protected by mmap_sem (write-lock) */
204 ret = __mmu_notifier_register(&mmu->mn, mm);
210 /* Protected by dev->struct_mutex */
211 hash_add(dev_priv->mmu_notifiers, &mmu->node, (unsigned long)mm);
216 __i915_mmu_notifier_destroy_worker(struct work_struct *work)
218 struct i915_mmu_notifier *mmu = container_of(work, typeof(*mmu), work);
219 mmu_notifier_unregister(&mmu->mn, mmu->mm);
224 __i915_mmu_notifier_destroy(struct i915_mmu_notifier *mmu)
226 lockdep_assert_held(&mmu->dev->struct_mutex);
228 /* Protected by dev->struct_mutex */
229 hash_del(&mmu->node);
231 /* Our lock ordering is: mmap_sem, mmu_notifier_scru, struct_mutex.
232 * We enter the function holding struct_mutex, therefore we need
233 * to drop our mutex prior to calling mmu_notifier_unregister in
234 * order to prevent lock inversion (and system-wide deadlock)
235 * between the mmap_sem and struct-mutex. Hence we defer the
236 * unregistration to a workqueue where we hold no locks.
238 INIT_WORK(&mmu->work, __i915_mmu_notifier_destroy_worker);
239 schedule_work(&mmu->work);
242 static void __i915_mmu_notifier_update_serial(struct i915_mmu_notifier *mmu)
244 if (++mmu->serial == 0)
248 static bool i915_mmu_notifier_has_linear(struct i915_mmu_notifier *mmu)
250 struct i915_mmu_object *mn;
252 list_for_each_entry(mn, &mmu->linear, link)
260 i915_mmu_notifier_del(struct i915_mmu_notifier *mmu,
261 struct i915_mmu_object *mn)
263 lockdep_assert_held(&mmu->dev->struct_mutex);
265 spin_lock(&mmu->lock);
268 mmu->has_linear = i915_mmu_notifier_has_linear(mmu);
270 interval_tree_remove(&mn->it, &mmu->objects);
271 __i915_mmu_notifier_update_serial(mmu);
272 spin_unlock(&mmu->lock);
274 /* Protected against _add() by dev->struct_mutex */
275 if (--mmu->count == 0)
276 __i915_mmu_notifier_destroy(mmu);
280 i915_mmu_notifier_add(struct i915_mmu_notifier *mmu,
281 struct i915_mmu_object *mn)
283 struct interval_tree_node *it;
286 ret = i915_mutex_lock_interruptible(mmu->dev);
290 /* Make sure we drop the final active reference (and thereby
291 * remove the objects from the interval tree) before we do
292 * the check for overlapping objects.
294 i915_gem_retire_requests(mmu->dev);
296 spin_lock(&mmu->lock);
297 it = interval_tree_iter_first(&mmu->objects,
298 mn->it.start, mn->it.last);
300 struct drm_i915_gem_object *obj;
302 /* We only need to check the first object in the range as it
303 * either has cancelled gup work queued and we need to
304 * return back to the user to give time for the gup-workers
305 * to flush their object references upon which the object will
306 * be removed from the interval-tree, or the the range is
307 * still in use by another client and the overlap is invalid.
309 * If we do have an overlap, we cannot use the interval tree
310 * for fast range invalidation.
313 obj = container_of(it, struct i915_mmu_object, it)->obj;
314 if (!obj->userptr.workers)
315 mmu->has_linear = mn->is_linear = true;
319 interval_tree_insert(&mn->it, &mmu->objects);
322 list_add(&mn->link, &mmu->linear);
323 __i915_mmu_notifier_update_serial(mmu);
325 spin_unlock(&mmu->lock);
326 mutex_unlock(&mmu->dev->struct_mutex);
332 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
334 struct i915_mmu_object *mn;
336 mn = obj->userptr.mn;
340 i915_mmu_notifier_del(mn->mmu, mn);
341 obj->userptr.mn = NULL;
345 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
348 struct i915_mmu_notifier *mmu;
349 struct i915_mmu_object *mn;
352 if (flags & I915_USERPTR_UNSYNCHRONIZED)
353 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
355 down_write(&obj->userptr.mm->mmap_sem);
356 ret = i915_mutex_lock_interruptible(obj->base.dev);
358 mmu = i915_mmu_notifier_get(obj->base.dev, obj->userptr.mm);
360 mmu->count++; /* preemptive add to act as a refcount */
363 mutex_unlock(&obj->base.dev->struct_mutex);
365 up_write(&obj->userptr.mm->mmap_sem);
369 mn = kzalloc(sizeof(*mn), GFP_KERNEL);
376 mn->it.start = obj->userptr.ptr;
377 mn->it.last = mn->it.start + obj->base.size - 1;
380 ret = i915_mmu_notifier_add(mmu, mn);
384 obj->userptr.mn = mn;
390 mutex_lock(&obj->base.dev->struct_mutex);
391 if (--mmu->count == 0)
392 __i915_mmu_notifier_destroy(mmu);
393 mutex_unlock(&obj->base.dev->struct_mutex);
400 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
405 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
408 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
411 if (!capable(CAP_SYS_ADMIN))
418 struct get_pages_work {
419 struct work_struct work;
420 struct drm_i915_gem_object *obj;
421 struct task_struct *task;
425 #if IS_ENABLED(CONFIG_SWIOTLB)
426 #define swiotlb_active() swiotlb_nr_tbl()
428 #define swiotlb_active() 0
432 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
434 struct scatterlist *sg;
437 *st = kmalloc(sizeof(**st), GFP_KERNEL);
441 if (swiotlb_active()) {
442 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
446 for_each_sg((*st)->sgl, sg, num_pages, n)
447 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
449 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
450 0, num_pages << PAGE_SHIFT,
465 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
467 struct get_pages_work *work = container_of(_work, typeof(*work), work);
468 struct drm_i915_gem_object *obj = work->obj;
469 struct drm_device *dev = obj->base.dev;
470 const int num_pages = obj->base.size >> PAGE_SHIFT;
477 pvec = kmalloc(num_pages*sizeof(struct page *),
478 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
480 pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
482 struct mm_struct *mm = obj->userptr.mm;
484 down_read(&mm->mmap_sem);
485 while (pinned < num_pages) {
486 ret = get_user_pages(work->task, mm,
487 obj->userptr.ptr + pinned * PAGE_SIZE,
489 !obj->userptr.read_only, 0,
490 pvec + pinned, NULL);
496 up_read(&mm->mmap_sem);
499 mutex_lock(&dev->struct_mutex);
500 if (obj->userptr.work != &work->work) {
502 } else if (pinned == num_pages) {
503 ret = st_set_pages(&obj->pages, pvec, num_pages);
505 list_add_tail(&obj->global_list, &to_i915(dev)->mm.unbound_list);
510 obj->userptr.work = ERR_PTR(ret);
511 obj->userptr.workers--;
512 drm_gem_object_unreference(&obj->base);
513 mutex_unlock(&dev->struct_mutex);
515 release_pages(pvec, pinned, 0);
516 drm_free_large(pvec);
518 put_task_struct(work->task);
523 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
525 const int num_pages = obj->base.size >> PAGE_SHIFT;
529 /* If userspace should engineer that these pages are replaced in
530 * the vma between us binding this page into the GTT and completion
531 * of rendering... Their loss. If they change the mapping of their
532 * pages they need to create a new bo to point to the new vma.
534 * However, that still leaves open the possibility of the vma
535 * being copied upon fork. Which falls under the same userspace
536 * synchronisation issue as a regular bo, except that this time
537 * the process may not be expecting that a particular piece of
538 * memory is tied to the GPU.
540 * Fortunately, we can hook into the mmu_notifier in order to
541 * discard the page references prior to anything nasty happening
542 * to the vma (discard or cloning) which should prevent the more
543 * egregious cases from causing harm.
548 if (obj->userptr.mm == current->mm) {
549 pvec = kmalloc(num_pages*sizeof(struct page *),
550 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
552 pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
557 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
558 !obj->userptr.read_only, pvec);
560 if (pinned < num_pages) {
565 /* Spawn a worker so that we can acquire the
566 * user pages without holding our mutex. Access
567 * to the user pages requires mmap_sem, and we have
568 * a strict lock ordering of mmap_sem, struct_mutex -
569 * we already hold struct_mutex here and so cannot
570 * call gup without encountering a lock inversion.
572 * Userspace will keep on repeating the operation
573 * (thanks to EAGAIN) until either we hit the fast
574 * path or the worker completes. If the worker is
575 * cancelled or superseded, the task is still run
576 * but the results ignored. (This leads to
577 * complications that we may have a stray object
578 * refcount that we need to be wary of when
579 * checking for existing objects during creation.)
580 * If the worker encounters an error, it reports
581 * that error back to this function through
582 * obj->userptr.work = ERR_PTR.
585 if (obj->userptr.work == NULL &&
586 obj->userptr.workers < I915_GEM_USERPTR_MAX_WORKERS) {
587 struct get_pages_work *work;
589 work = kmalloc(sizeof(*work), GFP_KERNEL);
591 obj->userptr.work = &work->work;
592 obj->userptr.workers++;
595 drm_gem_object_reference(&obj->base);
597 work->task = current;
598 get_task_struct(work->task);
600 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
601 schedule_work(&work->work);
605 if (IS_ERR(obj->userptr.work)) {
606 ret = PTR_ERR(obj->userptr.work);
607 obj->userptr.work = NULL;
612 ret = st_set_pages(&obj->pages, pvec, num_pages);
614 obj->userptr.work = NULL;
619 release_pages(pvec, pinned, 0);
620 drm_free_large(pvec);
625 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj)
627 struct scatterlist *sg;
630 BUG_ON(obj->userptr.work != NULL);
632 if (obj->madv != I915_MADV_WILLNEED)
635 for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) {
636 struct page *page = sg_page(sg);
639 set_page_dirty(page);
641 mark_page_accessed(page);
642 page_cache_release(page);
646 sg_free_table(obj->pages);
651 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
653 i915_gem_userptr_release__mmu_notifier(obj);
655 if (obj->userptr.mm) {
656 mmput(obj->userptr.mm);
657 obj->userptr.mm = NULL;
662 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
667 return i915_gem_userptr_init__mmu_notifier(obj, 0);
670 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
671 .dmabuf_export = i915_gem_userptr_dmabuf_export,
672 .get_pages = i915_gem_userptr_get_pages,
673 .put_pages = i915_gem_userptr_put_pages,
674 .release = i915_gem_userptr_release,
678 * Creates a new mm object that wraps some normal memory from the process
679 * context - user memory.
681 * We impose several restrictions upon the memory being mapped
683 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
684 * 2. It must be normal system memory, not a pointer into another map of IO
685 * space (e.g. it must not be a GTT mmapping of another object).
686 * 3. We only allow a bo as large as we could in theory map into the GTT,
687 * that is we limit the size to the total size of the GTT.
688 * 4. The bo is marked as being snoopable. The backing pages are left
689 * accessible directly by the CPU, but reads and writes by the GPU may
690 * incur the cost of a snoop (unless you have an LLC architecture).
692 * Synchronisation between multiple users and the GPU is left to userspace
693 * through the normal set-domain-ioctl. The kernel will enforce that the
694 * GPU relinquishes the VMA before it is returned back to the system
695 * i.e. upon free(), munmap() or process termination. However, the userspace
696 * malloc() library may not immediately relinquish the VMA after free() and
697 * instead reuse it whilst the GPU is still reading and writing to the VMA.
700 * Also note, that the object created here is not currently a "first class"
701 * object, in that several ioctls are banned. These are the CPU access
702 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
703 * direct access via your pointer rather than use those ioctls.
705 * If you think this is a good interface to use to pass GPU memory between
706 * drivers, please use dma-buf instead. In fact, wherever possible use
710 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
712 struct drm_i915_private *dev_priv = dev->dev_private;
713 struct drm_i915_gem_userptr *args = data;
714 struct drm_i915_gem_object *obj;
718 if (args->flags & ~(I915_USERPTR_READ_ONLY |
719 I915_USERPTR_UNSYNCHRONIZED))
722 if (offset_in_page(args->user_ptr | args->user_size))
725 if (args->user_size > dev_priv->gtt.base.total)
728 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
729 (char __user *)(unsigned long)args->user_ptr, args->user_size))
732 if (args->flags & I915_USERPTR_READ_ONLY) {
733 /* On almost all of the current hw, we cannot tell the GPU that a
734 * page is readonly, so this is just a placeholder in the uAPI.
739 /* Allocate the new object */
740 obj = i915_gem_object_alloc(dev);
744 drm_gem_private_object_init(dev, &obj->base, args->user_size);
745 i915_gem_object_init(obj, &i915_gem_userptr_ops);
746 obj->cache_level = I915_CACHE_LLC;
747 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
748 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
750 obj->userptr.ptr = args->user_ptr;
751 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
753 /* And keep a pointer to the current->mm for resolving the user pages
754 * at binding. This means that we need to hook into the mmu_notifier
755 * in order to detect if the mmu is destroyed.
758 if ((obj->userptr.mm = get_task_mm(current)))
759 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
761 ret = drm_gem_handle_create(file, &obj->base, &handle);
763 /* drop reference from allocate - handle holds it now */
764 drm_gem_object_unreference_unlocked(&obj->base);
768 args->handle = handle;
773 i915_gem_init_userptr(struct drm_device *dev)
775 #if defined(CONFIG_MMU_NOTIFIER)
776 struct drm_i915_private *dev_priv = to_i915(dev);
777 hash_init(dev_priv->mmu_notifiers);