diff mbox

drm/i915: Introduce mapping of user pages into video memory (userptr) ioctl

Message ID 07EEC142D30DAC40BA4DD9C0F9EE510923D0F702@FMSMSX102.amr.corp.intel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Reese, Armin C Feb. 13, 2013, 10:24 p.m. UTC
Thanks for the patch, Chris.

This is exactly what we were looking for to replace the VMAP feature you submitted a couple of years ago.  We need a method to quickly move data from user mode allocations into video memory (by mapping backing pages into the GTT).

The Interface appears simple enough, yet fulfills our needs.

I have one question ... what exactly does the I915_USERPTR_UNSYNCHRONIZED flag do?

Thanks,
Armin

-----Original Message-----
From: intel-gfx-bounces+armin.c.reese=intel.com@lists.freedesktop.org [mailto:intel-gfx-bounces+armin.c.reese=intel.com@lists.freedesktop.org] On Behalf Of Chris Wilson

Sent: Tuesday, February 12, 2013 6:17 AM
To: intel-gfx@lists.freedesktop.org
Subject: [Intel-gfx] [PATCH] drm/i915: Introduce mapping of user pages into video memory (userptr) ioctl

By exporting the ability to map user address and inserting PTEs representing their backing pages into the GTT, we can exploit UMA in order to utilize normal application data as a texture source or even as a render target (depending upon the capabilities of the chipset). This has a number of uses, with zero-copy downloads to the GPU and efficient readback making the intermixed streaming of CPU and GPU operations fairly efficient. This ability has many widespread implications from faster rendering of client-side software rasterisers (chromium), mitigation of stalls due to read back (firefox) and to faster pipelining of texture data (such as pixel buffer objects in GL or data blobs in CL).

v2: Compile with CONFIG_MMU_NOTIFIER
v3: We can sleep while performing invalidate-range, which we can utilise to drop our page references prior to the kernel manipulating the vma (for either discard or cloning) and so protect normal users.
v4: Only run the invalidate notifier if the range intercepts the bo.
v5: Prevent userspace from attempting to GTT mmap non-page aligned buffers

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

---
 drivers/gpu/drm/i915/Makefile              |    1 +
 drivers/gpu/drm/i915/i915_dma.c            |    1 +
 drivers/gpu/drm/i915/i915_drv.h            |   22 ++
 drivers/gpu/drm/i915/i915_gem.c            |   31 ++-
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |    7 +-
 drivers/gpu/drm/i915/i915_gem_userptr.c    |  329 ++++++++++++++++++++++++++++
 include/uapi/drm/i915_drm.h                |   16 ++
 7 files changed, 393 insertions(+), 14 deletions(-)  create mode 100644 drivers/gpu/drm/i915/i915_gem_userptr.c

--
1.7.10.4

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Comments

Chris Wilson Feb. 13, 2013, 11:20 p.m. UTC | #1
On Wed, Feb 13, 2013 at 10:24:59PM +0000, Reese, Armin C wrote:
> Thanks for the patch, Chris.
> 
> This is exactly what we were looking for to replace the VMAP feature you submitted a couple of years ago.  We need a method to quickly move data from user mode allocations into video memory (by mapping backing pages into the GTT).
> 
> The Interface appears simple enough, yet fulfills our needs.
> 
> I have one question ... what exactly does the I915_USERPTR_UNSYNCHRONIZED flag do?

Not so loud, Daniel might notice. It is a flag for the user to ask for a
get_user_pages mapping that ignored the synchronization issues
associated with cloning vma. That is the caller fully understood that to
fork and share those pages would require userspace synchronization and
that the caller was also responsible for ensuring that the mapping
was destroyed first before the vma was release - or else the vma could be
reused by the process/shared-memory whilst it was stil active.

The intention was to workaround the limitations and lack-of mmu-notifier
in many cases. I made it a privileged flag because normal users aren't
really meant to be able to cause so much interprocess havoc. Though
really having the GPU also access that memory is not so much more
dangerous than sharing that memory with another process, so I am open to
the suggestion that maybe we should allow normal users to shoot
themselves in the foot as well.

Also note that the current interface prohibits GTT mmaping of userptr bo
on LLC machines if the pointer is not page-aligned. I still think I can
lift that restriction, but my initial attempts were futile, so I'm
ignoring that corner case unless there is a use-case.
-Chris
diff mbox

Patch

diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index 91f3ac6..42858f6 100644
--- a/drivers/gpu/drm/i915/Makefile
+++ b/drivers/gpu/drm/i915/Makefile
@@ -14,6 +14,7 @@  i915-y := i915_drv.o i915_dma.o i915_irq.o \
 	  i915_gem_gtt.o \
 	  i915_gem_stolen.o \
 	  i915_gem_tiling.o \
+	  i915_gem_userptr.o \
 	  i915_sysfs.o \
 	  i915_trace_points.o \
 	  i915_ums.o \
diff --git a/drivers/gpu/drm/i915/i915_dma.c b/drivers/gpu/drm/i915/i915_dma.c index 4fa6beb..9b1984c 100644
--- a/drivers/gpu/drm/i915/i915_dma.c
+++ b/drivers/gpu/drm/i915/i915_dma.c
@@ -1883,6 +1883,7 @@  struct drm_ioctl_desc i915_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(I915_GEM_CONTEXT_CREATE, i915_gem_context_create_ioctl, DRM_UNLOCKED),
 	DRM_IOCTL_DEF_DRV(I915_GEM_CONTEXT_DESTROY, i915_gem_context_destroy_ioctl, DRM_UNLOCKED),
 	DRM_IOCTL_DEF_DRV(I915_REG_READ, i915_reg_read_ioctl, DRM_UNLOCKED),
+	DRM_IOCTL_DEF_DRV(I915_GEM_USERPTR, i915_gem_userptr_ioctl, 
+DRM_UNLOCKED),
 };
 
 int i915_max_ioctl = DRM_ARRAY_SIZE(i915_ioctls); diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 923dc0a..90070f4 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -42,6 +42,7 @@ 
 #include <linux/backlight.h>
 #include <linux/intel-iommu.h>
 #include <linux/kref.h>
+#include <linux/mmu_notifier.h>
 #include <linux/pm_qos.h>
 
 /* General customization:
@@ -1076,6 +1077,7 @@  struct drm_i915_gem_object_ops {
 	 */
 	int (*get_pages)(struct drm_i915_gem_object *);
 	void (*put_pages)(struct drm_i915_gem_object *);
+	void (*release)(struct drm_i915_gem_object *);
 };
 
 struct drm_i915_gem_object {
@@ -1222,6 +1224,23 @@  struct drm_i915_gem_object {  };  #define to_gem_object(obj) (&((struct drm_i915_gem_object *)(obj))->base)
 
+struct i915_gem_userptr_object {
+	struct drm_i915_gem_object gem;
+	uintptr_t user_ptr;
+	size_t user_size;
+	int read_only;
+
+	struct mm_struct *mm;
+#if defined(CONFIG_MMU_NOTIFIER)
+	struct mmu_notifier mn;
+#endif
+};
+
+union drm_i915_gem_objects {
+	struct drm_i915_gem_object base;
+	struct i915_gem_userptr_object userptr; };
+
 #define to_intel_bo(x) container_of(x, struct drm_i915_gem_object, base)
 
 /**
@@ -1501,6 +1520,8 @@  int i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
 			   struct drm_file *file_priv);
 int i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
 			   struct drm_file *file_priv);
+int i915_gem_userptr_ioctl(struct drm_device *dev, void *data,
+			   struct drm_file *file);
 int i915_gem_set_tiling(struct drm_device *dev, void *data,
 			struct drm_file *file_priv);
 int i915_gem_get_tiling(struct drm_device *dev, void *data, @@ -1554,6 +1575,7 @@ static inline void i915_gem_object_unpin_pages(struct drm_i915_gem_object *obj)
 	BUG_ON(obj->pages_pin_count == 0);
 	obj->pages_pin_count--;
 }
+int i915_gem_object_put_pages(struct drm_i915_gem_object *obj);
 
 int __must_check i915_mutex_lock_interruptible(struct drm_device *dev);  int i915_gem_object_sync(struct drm_i915_gem_object *obj, diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 73b1e9e..65a36bf 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -1336,20 +1336,18 @@  int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 	struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
 	struct drm_device *dev = obj->base.dev;
 	drm_i915_private_t *dev_priv = dev->dev_private;
-	pgoff_t page_offset;
+	pgoff_t offset;
 	unsigned long pfn;
 	int ret = 0;
 	bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
 
-	/* We don't use vmf->pgoff since that has the fake offset */
-	page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
-		PAGE_SHIFT;
-
 	ret = i915_mutex_lock_interruptible(dev);
 	if (ret)
 		goto out;
 
-	trace_i915_gem_object_fault(obj, page_offset, true, write);
+	trace_i915_gem_object_fault(obj,
+				    (unsigned long)(vmf->virtual_address - vma->vm_start) >> PAGE_SHIFT,
+				    true, write);
 
 	/* Access to snoopable pages through the GTT is incoherent. */
 	if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) { @@ -1372,8 +1370,10 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 
 	obj->fault_mappable = true;
 
-	pfn = ((dev_priv->gtt.mappable_base + obj->gtt_offset) >> PAGE_SHIFT) +
-		page_offset;
+	/* We don't use vmf->pgoff since that has the fake offset */
+	offset = (unsigned long)vmf->virtual_address - vma->vm_start;
+	offset += obj->gtt_offset;
+	pfn = (dev_priv->gtt.mappable_base + offset) >> PAGE_SHIFT;
 
 	/* Finally, remap it using the new GTT offset */
 	ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn); @@ -1565,6 +1565,12 @@ i915_gem_mmap_gtt(struct drm_file *file,
 		goto out;
 	}
 
+	if (offset_in_page(obj->gtt_offset)) {
+		DRM_ERROR("Attempting to mmap an unaligned buffer\n");
+		ret = -EINVAL;
+		goto out;
+	}
+
 	ret = i915_gem_object_create_mmap_offset(obj);
 	if (ret)
 		goto out;
@@ -2495,9 +2501,9 @@  i915_gem_object_unbind(struct drm_i915_gem_object *obj)
 	/* Avoid an unnecessary call to unbind on rebind. */
 	obj->map_and_fenceable = true;
 
+	obj->gtt_offset -= obj->gtt_space->start;
 	drm_mm_put_block(obj->gtt_space);
 	obj->gtt_space = NULL;
-	obj->gtt_offset = 0;
 
 	return 0;
 }
@@ -2987,7 +2993,7 @@  i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
 	list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
 
 	obj->gtt_space = node;
-	obj->gtt_offset = node->start;
+	obj->gtt_offset += node->start;
 
 	fenceable =
 		node->size == fence_size &&
@@ -3800,6 +3806,9 @@  void i915_gem_free_object(struct drm_gem_object *gem_obj)
 	if (obj->base.import_attach)
 		drm_prime_gem_destroy(&obj->base, NULL);
 
+	if (obj->ops->release)
+		obj->ops->release(obj);
+
 	drm_gem_object_release(&obj->base);
 	i915_gem_info_remove_obj(dev_priv, obj->base.size);
 
@@ -4101,7 +4110,7 @@  i915_gem_load(struct drm_device *dev)
 
 	dev_priv->slab =
 		kmem_cache_create("i915_gem_object",
-				  sizeof(struct drm_i915_gem_object), 0,
+				  sizeof(union drm_i915_gem_objects), 0,
 				  SLAB_HWCACHE_ALIGN,
 				  NULL);
 
diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
index 2726910..a3e68af 100644
--- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
@@ -254,14 +254,16 @@  i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 		return -EFAULT;
 
 	reloc->delta += target_offset;
+	reloc->offset += obj->gtt_offset;
 	if (use_cpu_reloc(obj)) {
-		uint32_t page_offset = reloc->offset & ~PAGE_MASK;
+		uint32_t page_offset = offset_in_page(reloc->offset);
 		char *vaddr;
 
 		ret = i915_gem_object_set_to_cpu_domain(obj, 1);
 		if (ret)
 			return ret;
 
+		reloc->offset -= obj->gtt_space->start;
 		vaddr = kmap_atomic(i915_gem_object_get_page(obj,
 							     reloc->offset >> PAGE_SHIFT));
 		*(uint32_t *)(vaddr + page_offset) = reloc->delta; @@ -280,11 +282,10 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 			return ret;
 
 		/* Map the page containing the relocation we're going to perform.  */
-		reloc->offset += obj->gtt_offset;
 		reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
 						      reloc->offset & PAGE_MASK);
 		reloc_entry = (uint32_t __iomem *)
-			(reloc_page + (reloc->offset & ~PAGE_MASK));
+			(reloc_page + offset_in_page(reloc->offset));
 		iowrite32(reloc->delta, reloc_entry);
 		io_mapping_unmap_atomic(reloc_page);
 	}
diff --git a/drivers/gpu/drm/i915/i915_gem_userptr.c b/drivers/gpu/drm/i915/i915_gem_userptr.c
new file mode 100644
index 0000000..f93fa1b
--- /dev/null
+++ b/drivers/gpu/drm/i915/i915_gem_userptr.c
@@ -0,0 +1,329 @@ 
+/*
+ * Copyright © 2012 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person 
+obtaining a
+ * copy of this software and associated documentation files (the 
+"Software"),
+ * to deal in the Software without restriction, including without 
+limitation
+ * the rights to use, copy, modify, merge, publish, distribute, 
+sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom 
+the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the 
+next
+ * paragraph) shall be included in all copies or substantial portions 
+of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 
+EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 
+MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT 
+SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 
+OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 
+ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 
+DEALINGS
+ * IN THE SOFTWARE.
+ *
+ */
+
+#include "drmP.h"
+#include "i915_drm.h"
+#include "i915_drv.h"
+#include "i915_trace.h"
+#include "intel_drv.h"
+#include <linux/mmu_notifier.h>
+#include <linux/swap.h>
+
+static struct i915_gem_userptr_object *to_userptr_object(struct 
+drm_i915_gem_object *obj) {
+	return container_of(obj, struct i915_gem_userptr_object, gem); }
+
+#if defined(CONFIG_MMU_NOTIFIER)
+static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *mn,
+						       struct mm_struct *mm,
+						       unsigned long start,
+						       unsigned long end)
+{
+	struct i915_gem_userptr_object *vmap;
+	struct drm_device *dev;
+
+	/* XXX race between obj unref and mmu notifier? */
+	vmap = container_of(mn, struct i915_gem_userptr_object, mn);
+	BUG_ON(vmap->mm != mm);
+
+	if (vmap->user_ptr >= end || vmap->user_ptr + vmap->user_size <= start)
+		return;
+
+	if (vmap->gem.pages == NULL) /* opportunistic check */
+		return;
+
+	dev = vmap->gem.base.dev;
+	mutex_lock(&dev->struct_mutex);
+	if (vmap->gem.gtt_space) {
+		struct drm_i915_private *dev_priv = dev->dev_private;
+		bool was_interruptible;
+		int ret;
+
+		was_interruptible = dev_priv->mm.interruptible;
+		dev_priv->mm.interruptible = false;
+
+		ret = i915_gem_object_unbind(&vmap->gem);
+		BUG_ON(ret && ret != -EIO);
+
+		dev_priv->mm.interruptible = was_interruptible;
+	}
+
+	BUG_ON(i915_gem_object_put_pages(&vmap->gem));
+	mutex_unlock(&dev->struct_mutex);
+}
+
+static void i915_gem_userptr_mn_release(struct mmu_notifier *mn,
+					struct mm_struct *mm)
+{
+	struct i915_gem_userptr_object *vmap;
+
+	vmap = container_of(mn, struct i915_gem_userptr_object, mn);
+	BUG_ON(vmap->mm != mm);
+	vmap->mm = NULL;
+
+	/* XXX Schedule an eventual unbind? E.g. hook into require request?
+	 * However, locking will be complicated.
+	 */
+}
+
+static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
+	.invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
+	.release = i915_gem_userptr_mn_release, };
+
+static void
+i915_gem_userptr_release__mmu_notifier(struct i915_gem_userptr_object 
+*vmap) {
+	if (vmap->mn.ops && vmap->mm) {
+		mmu_notifier_unregister(&vmap->mn, vmap->mm);
+		BUG_ON(vmap->mm);
+	}
+}
+
+static int
+i915_gem_userptr_init__mmu_notifier(struct i915_gem_userptr_object *vmap,
+				    unsigned flags)
+{
+	if (flags & I915_USERPTR_UNSYNCHRONIZED)
+		return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
+
+	vmap->mn.ops = &i915_gem_userptr_notifier;
+	return mmu_notifier_register(&vmap->mn, vmap->mm); }
+
+#else
+
+static void
+i915_gem_userptr_release__mmu_notifier(struct i915_gem_userptr_object 
+*vmap) { }
+
+static int
+i915_gem_userptr_init__mmu_notifier(struct i915_gem_userptr_object *vmap,
+				    unsigned flags)
+{
+	if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
+		return -ENODEV;
+
+	if (!capable(CAP_SYS_ADMIN))
+		return -EPERM;
+
+	return 0;
+}
+#endif
+
+static int
+i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj) {
+	struct i915_gem_userptr_object *vmap = to_userptr_object(obj);
+	int num_pages = obj->base.size >> PAGE_SHIFT;
+	struct sg_table *st;
+	struct scatterlist *sg;
+	struct page **pvec;
+	int n, pinned, ret;
+
+	if (vmap->mm == NULL)
+		return -EFAULT;
+
+	if (!access_ok(vmap->read_only ? VERIFY_READ : VERIFY_WRITE,
+		       (char __user *)vmap->user_ptr, vmap->user_size))
+		return -EFAULT;
+
+	/* If userspace should engineer that these pages are replaced in
+	 * the vma between us binding this page into the GTT and completion
+	 * of rendering... Their loss. If they change the mapping of their
+	 * pages they need to create a new bo to point to the new vma.
+	 *
+	 * However, that still leaves open the possibility of the vma
+	 * being copied upon fork. Which falls under the same userspace
+	 * synchronisation issue as a regular bo, except that this time
+	 * the process may not be expecting that a particular piece of
+	 * memory is tied to the GPU.
+	 *
+	 * Fortunately, we can hook into the mmu_notifier in order to
+	 * discard the page references prior to anything nasty happening
+	 * to the vma (discard or cloning) which should prevent the more
+	 * egregious cases from causing harm.
+	 */
+
+	pvec = kmalloc(num_pages*sizeof(struct page *),
+		       GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
+	if (pvec == NULL) {
+		pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
+		if (pvec == NULL)
+			return -ENOMEM;
+	}
+
+	pinned = 0;
+	if (vmap->mm == current->mm)
+		pinned = __get_user_pages_fast(vmap->user_ptr, num_pages,
+					       !vmap->read_only, pvec);
+	if (pinned < num_pages) {
+		struct mm_struct *mm = vmap->mm;
+		ret = 0;
+		mutex_unlock(&obj->base.dev->struct_mutex);
+		down_read(&mm->mmap_sem);
+		if (vmap->mm != NULL)
+			ret = get_user_pages(current, mm,
+					     vmap->user_ptr + (pinned << PAGE_SHIFT),
+					     num_pages - pinned,
+					     !vmap->read_only, 0,
+					     pvec + pinned,
+					     NULL);
+		up_read(&mm->mmap_sem);
+		mutex_lock(&obj->base.dev->struct_mutex);
+		if (ret > 0)
+			pinned += ret;
+
+		if (obj->pages || pinned < num_pages) {
+			ret = obj->pages ? 0 : -EFAULT;
+			goto cleanup_pinned;
+		}
+	}
+
+	st = kmalloc(sizeof(*st), GFP_KERNEL);
+	if (st == NULL) {
+		ret = -ENOMEM;
+		goto cleanup_pinned;
+	}
+
+	if (sg_alloc_table(st, num_pages, GFP_KERNEL)) {
+		ret = -ENOMEM;
+		goto cleanup_st;
+	}
+
+	for_each_sg(st->sgl, sg, num_pages, n)
+		sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
+	drm_free_large(pvec);
+
+	obj->pages = st;
+	return 0;
+
+cleanup_st:
+	kfree(st);
+cleanup_pinned:
+	release_pages(pvec, pinned, 0);
+	drm_free_large(pvec);
+	return ret;
+}
+
+static void
+i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj) {
+	struct scatterlist *sg;
+	int i;
+
+	if (obj->madv != I915_MADV_WILLNEED)
+		obj->dirty = 0;
+
+	for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) {
+		struct page *page = sg_page(sg);
+
+		if (obj->dirty)
+			set_page_dirty(page);
+
+		mark_page_accessed(page);
+		page_cache_release(page);
+	}
+	obj->dirty = 0;
+
+	sg_free_table(obj->pages);
+	kfree(obj->pages);
+}
+
+static void
+i915_gem_userptr_release(struct drm_i915_gem_object *obj) {
+	struct i915_gem_userptr_object *vmap = to_userptr_object(obj);
+
+	i915_gem_userptr_release__mmu_notifier(vmap);
+}
+
+static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
+	.get_pages = i915_gem_userptr_get_pages,
+	.put_pages = i915_gem_userptr_put_pages,
+	.release = i915_gem_userptr_release,
+};
+
+/**
+ * Creates a new mm object that wraps some user memory.
+ */
+int
+i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct 
+drm_file *file) {
+	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct drm_i915_gem_userptr *args = data;
+	struct i915_gem_userptr_object *obj;
+	loff_t first_data_page, last_data_page;
+	int num_pages;
+	int ret;
+	u32 handle;
+
+	if (args->flags & ~(I915_USERPTR_READ_ONLY | I915_USERPTR_UNSYNCHRONIZED))
+		return -EINVAL;
+
+	first_data_page = args->user_ptr / PAGE_SIZE;
+	last_data_page = (args->user_ptr + args->user_size - 1) / PAGE_SIZE;
+	num_pages = last_data_page - first_data_page + 1;
+	if (num_pages * PAGE_SIZE > dev_priv->gtt.total)
+		return -E2BIG;
+
+	/* Allocate the new object */
+	obj = i915_gem_object_alloc(dev);
+	if (obj == NULL)
+		return -ENOMEM;
+
+	if (drm_gem_private_object_init(dev, &obj->gem.base,
+					num_pages * PAGE_SIZE)) {
+		i915_gem_object_free(&obj->gem);
+		return -ENOMEM;
+	}
+
+	i915_gem_object_init(&obj->gem, &i915_gem_userptr_ops);
+	obj->gem.cache_level = I915_CACHE_LLC_MLC;
+
+	obj->gem.gtt_offset = offset_in_page(args->user_ptr);
+	obj->user_ptr = args->user_ptr;
+	obj->user_size = args->user_size;
+	obj->read_only = args->flags & I915_USERPTR_READ_ONLY;
+
+	/* And keep a pointer to the current->mm for resolving the user pages
+	 * at binding. This means that we need to hook into the mmu_notifier
+	 * in order to detect if the mmu is destroyed.
+	 */
+	obj->mm = current->mm;
+	ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
+	if (ret)
+		return ret;
+
+	ret = drm_gem_handle_create(file, &obj->gem.base, &handle);
+	/* drop reference from allocate - handle holds it now */
+	drm_gem_object_unreference(&obj->gem.base);
+	if (ret)
+		return ret;
+
+	args->handle = handle;
+	return 0;
+}
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h index 07d5941..20e39be 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -198,6 +198,7 @@  typedef struct _drm_i915_sarea {
 #define DRM_I915_GEM_SET_CACHING	0x2f
 #define DRM_I915_GEM_GET_CACHING	0x30
 #define DRM_I915_REG_READ		0x31
+#define DRM_I915_GEM_USERPTR		0x32
 
 #define DRM_IOCTL_I915_INIT		DRM_IOW( DRM_COMMAND_BASE + DRM_I915_INIT, drm_i915_init_t)
 #define DRM_IOCTL_I915_FLUSH		DRM_IO ( DRM_COMMAND_BASE + DRM_I915_FLUSH)
@@ -247,6 +248,7 @@  typedef struct _drm_i915_sarea {
 #define DRM_IOCTL_I915_GEM_CONTEXT_CREATE	DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_CONTEXT_CREATE, struct drm_i915_gem_context_create)
 #define DRM_IOCTL_I915_GEM_CONTEXT_DESTROY	DRM_IOW (DRM_COMMAND_BASE + DRM_I915_GEM_CONTEXT_DESTROY, struct drm_i915_gem_context_destroy)
 #define DRM_IOCTL_I915_REG_READ			DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_REG_READ, struct drm_i915_reg_read)
+#define DRM_IOCTL_I915_GEM_USERPTR			DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_USERPTR, struct drm_i915_gem_userptr)
 
 /* Allow drivers to submit batchbuffers directly to hardware, relying
  * on the security mechanisms provided by hardware.
@@ -980,4 +982,18 @@  struct drm_i915_reg_read {
 	__u64 offset;
 	__u64 val; /* Return value */
 };
+
+struct drm_i915_gem_userptr {
+	__u64 user_ptr;
+	__u32 user_size;
+	__u32 flags;
+#define I915_USERPTR_READ_ONLY 0x1
+#define I915_USERPTR_UNSYNCHRONIZED 0x80000000
+	/**
+	 * Returned handle for the object.
+	 *
+	 * Object handles are nonzero.
+	 */
+	__u32 handle;
+};
 #endif /* _UAPI_I915_DRM_H_ */