drm/i915: Introduce mapping of user pages into video memory (userptr) ioctl
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 16 May 2014 13:22:37 +0000 (14:22 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 16 May 2014 17:31:29 +0000 (19:31 +0200)
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
v6: Recheck after reacquire mutex for lost mmu.
v7: Fix implicit padding of ioctl struct by rounding to next 64bit boundary.
v8: Fix rebasing error after forwarding porting the back port.
v9: Limit the userptr to page aligned entries. We now expect userspace
    to handle all the offset-in-page adjustments itself.
v10: Prevent vma from being copied across fork to avoid issues with cow.
v11: Drop vma behaviour changes -- locking is nigh on impossible.
     Use a worker to load user pages to avoid lock inversions.
v12: Use get_task_mm()/mmput() for correct refcounting of mm.
v13: Use a worker to release the mmu_notifier to avoid lock inversion
v14: Decouple mmu_notifier from struct_mutex using a custom mmu_notifer
     with its own locking and tree of objects for each mm/mmu_notifier.
v15: Prevent overlapping userptr objects, and invalidate all objects
     within the mmu_notifier range
v16: Fix a typo for iterating over multiple objects in the range and
     rearrange error path to destroy the mmu_notifier locklessly.
     Also close a race between invalidate_range and the get_pages_worker.
v17: Close a race between get_pages_worker/invalidate_range and fresh
     allocations of the same userptr range - and notice that
     struct_mutex was presumed to be held when during creation it wasn't.
v18: Sigh. Fix the refactor of st_set_pages() to allocate enough memory
     for the struct sg_table and to clear it before reporting an error.
v19: Always error out on read-only userptr requests as we don't have the
     hardware infrastructure to support them at the moment.
v20: Refuse to implement read-only support until we have the required
     infrastructure - but reserve the bit in flags for future use.
v21: use_mm() is not required for get_user_pages(). It is only meant to
     be used to fix up the kernel thread's current->mm for use with
     copy_user().
v22: Use sg_alloc_table_from_pages for that chunky feeling
v23: Export a function for sanity checking dma-buf rather than encode
     userptr details elsewhere, and clean up comments based on
     suggestions by Bradley.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com>
Cc: Akash Goel <akash.goel@intel.com>
Cc: "Volkin, Bradley D" <bradley.d.volkin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Brad Volkin <bradley.d.volkin@intel.com>
[danvet: Frob ioctl allocation to pick the next one - will cause a bit
of fuss with create2 apparently, but such are the rules.]
[danvet2: oops, forgot to git add after manual patch application]
[danvet3: Appease sparse.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/Kconfig
drivers/gpu/drm/i915/Makefile
drivers/gpu/drm/i915/i915_dma.c
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_gem_dmabuf.c
drivers/gpu/drm/i915/i915_gem_userptr.c [new file with mode: 0644]
drivers/gpu/drm/i915/i915_gpu_error.c
include/uapi/drm/i915_drm.h

index e4e3c01b8cbc6676697730e8e5e524cbafc20748..437e1824d0bf1762db64587833269a7ebeec78e6 100644 (file)
@@ -5,6 +5,7 @@ config DRM_I915
        depends on (AGP || AGP=n)
        select INTEL_GTT
        select AGP_INTEL if AGP
+       select INTERVAL_TREE
        # we need shmfs for the swappable backing store, and in particular
        # the shmem_readpage() which depends upon tmpfs
        select SHMEM
index 24469168d550ab448c7abdd843c92132cc5d07c7..7b2f3bee351814dc8065def8284206f7267abf3c 100644 (file)
@@ -27,6 +27,7 @@ i915-y += i915_cmd_parser.o \
          i915_gem.o \
          i915_gem_stolen.o \
          i915_gem_tiling.o \
+         i915_gem_userptr.o \
          i915_gpu_error.o \
          i915_irq.o \
          i915_trace_points.o \
index 46f1decab76f629047e5b6ce576cd64063bc5b01..dea455bd889ac562d5bd4f9d6120f9d5b5f7a101 100644 (file)
@@ -1983,6 +1983,7 @@ const struct drm_ioctl_desc i915_ioctls[] = {
        DRM_IOCTL_DEF_DRV(I915_GEM_CONTEXT_DESTROY, i915_gem_context_destroy_ioctl, DRM_UNLOCKED|DRM_RENDER_ALLOW),
        DRM_IOCTL_DEF_DRV(I915_REG_READ, i915_reg_read_ioctl, DRM_UNLOCKED|DRM_RENDER_ALLOW),
        DRM_IOCTL_DEF_DRV(I915_GET_RESET_STATS, i915_get_reset_stats_ioctl, DRM_UNLOCKED|DRM_RENDER_ALLOW),
+       DRM_IOCTL_DEF_DRV(I915_GEM_USERPTR, i915_gem_userptr_ioctl, DRM_UNLOCKED|DRM_RENDER_ALLOW),
 };
 
 int i915_max_ioctl = DRM_ARRAY_SIZE(i915_ioctls);
index 9e9e8166a1f3b2002e2efabec9796d16df9e031d..a270e0773e2dcd1a3b415bad728c7bfd434edcd1 100644 (file)
@@ -41,6 +41,7 @@
 #include <linux/i2c-algo-bit.h>
 #include <drm/intel-gtt.h>
 #include <linux/backlight.h>
+#include <linux/hashtable.h>
 #include <linux/intel-iommu.h>
 #include <linux/kref.h>
 #include <linux/pm_qos.h>
@@ -178,6 +179,7 @@ enum hpd_pin {
                if ((intel_connector)->base.encoder == (__encoder))
 
 struct drm_i915_private;
+struct i915_mmu_object;
 
 enum intel_dpll_id {
        DPLL_ID_PRIVATE = -1, /* non-shared dpll in use */
@@ -403,6 +405,7 @@ struct drm_i915_error_state {
                u32 tiling:2;
                u32 dirty:1;
                u32 purgeable:1;
+               u32 userptr:1;
                s32 ring:4;
                u32 cache_level:3;
        } **active_bo, **pinned_bo;
@@ -1447,6 +1450,9 @@ struct drm_i915_private {
        struct i915_gtt gtt; /* VM representing the global address space */
 
        struct i915_gem_mm mm;
+#if defined(CONFIG_MMU_NOTIFIER)
+       DECLARE_HASHTABLE(mmu_notifiers, 7);
+#endif
 
        /* Kernel Modesetting */
 
@@ -1580,6 +1586,8 @@ struct drm_i915_gem_object_ops {
         */
        int (*get_pages)(struct drm_i915_gem_object *);
        void (*put_pages)(struct drm_i915_gem_object *);
+       int (*dmabuf_export)(struct drm_i915_gem_object *);
+       void (*release)(struct drm_i915_gem_object *);
 };
 
 struct drm_i915_gem_object {
@@ -1693,8 +1701,20 @@ struct drm_i915_gem_object {
 
        /** for phy allocated objects */
        struct drm_i915_gem_phys_object *phys_obj;
-};
 
+       union {
+               struct i915_gem_userptr {
+                       uintptr_t ptr;
+                       unsigned read_only :1;
+                       unsigned workers :4;
+#define I915_GEM_USERPTR_MAX_WORKERS 15
+
+                       struct mm_struct *mm;
+                       struct i915_mmu_object *mn;
+                       struct work_struct *work;
+               } userptr;
+       };
+};
 #define to_intel_bo(x) container_of(x, struct drm_i915_gem_object, base)
 
 /**
@@ -2119,6 +2139,9 @@ 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,
                        struct drm_file *file_priv);
+int i915_gem_init_userptr(struct drm_device *dev);
+int i915_gem_userptr_ioctl(struct drm_device *dev, void *data,
+                          struct drm_file *file);
 int i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
                                struct drm_file *file_priv);
 int i915_gem_wait_ioctl(struct drm_device *dev, void *data,
index f6c03513de521533288d7d0b7a50c21a80517851..d5aad0bc71f7b1753482a53b296df9cc870a618a 100644 (file)
@@ -4263,6 +4263,9 @@ 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);
 
@@ -4542,6 +4545,7 @@ int i915_gem_init(struct drm_device *dev)
                        DRM_DEBUG_DRIVER("allow wake ack timed out\n");
        }
 
+       i915_gem_init_userptr(dev);
        i915_gem_init_global_gtt(dev);
 
        ret = i915_gem_context_init(dev);
index 321102a8374ba0d0ea8870620b78b4a2327131c1..580aa42443edc87823dc4fb8f3776ea0d39b6b52 100644 (file)
@@ -229,6 +229,14 @@ static const struct dma_buf_ops i915_dmabuf_ops =  {
 struct dma_buf *i915_gem_prime_export(struct drm_device *dev,
                                      struct drm_gem_object *gem_obj, int flags)
 {
+       struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
+
+       if (obj->ops->dmabuf_export) {
+               int ret = obj->ops->dmabuf_export(obj);
+               if (ret)
+                       return ERR_PTR(ret);
+       }
+
        return dma_buf_export(gem_obj, &i915_dmabuf_ops, gem_obj->size, flags);
 }
 
diff --git a/drivers/gpu/drm/i915/i915_gem_userptr.c b/drivers/gpu/drm/i915/i915_gem_userptr.c
new file mode 100644 (file)
index 0000000..21ea928
--- /dev/null
@@ -0,0 +1,711 @@
+/*
+ * Copyright © 2012-2014 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_context.h>
+#include <linux/mmu_notifier.h>
+#include <linux/mempolicy.h>
+#include <linux/swap.h>
+
+#if defined(CONFIG_MMU_NOTIFIER)
+#include <linux/interval_tree.h>
+
+struct i915_mmu_notifier {
+       spinlock_t lock;
+       struct hlist_node node;
+       struct mmu_notifier mn;
+       struct rb_root objects;
+       struct drm_device *dev;
+       struct mm_struct *mm;
+       struct work_struct work;
+       unsigned long count;
+       unsigned long serial;
+};
+
+struct i915_mmu_object {
+       struct i915_mmu_notifier *mmu;
+       struct interval_tree_node it;
+       struct drm_i915_gem_object *obj;
+};
+
+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_mmu_notifier *mn = container_of(_mn, struct i915_mmu_notifier, mn);
+       struct interval_tree_node *it = NULL;
+       unsigned long serial = 0;
+
+       end--; /* interval ranges are inclusive, but invalidate range is exclusive */
+       while (start < end) {
+               struct drm_i915_gem_object *obj;
+
+               obj = NULL;
+               spin_lock(&mn->lock);
+               if (serial == mn->serial)
+                       it = interval_tree_iter_next(it, start, end);
+               else
+                       it = interval_tree_iter_first(&mn->objects, start, end);
+               if (it != NULL) {
+                       obj = container_of(it, struct i915_mmu_object, it)->obj;
+                       drm_gem_object_reference(&obj->base);
+                       serial = mn->serial;
+               }
+               spin_unlock(&mn->lock);
+               if (obj == NULL)
+                       return;
+
+               mutex_lock(&mn->dev->struct_mutex);
+               /* Cancel any active worker and force us to re-evaluate gup */
+               obj->userptr.work = NULL;
+
+               if (obj->pages != NULL) {
+                       struct drm_i915_private *dev_priv = to_i915(mn->dev);
+                       struct i915_vma *vma, *tmp;
+                       bool was_interruptible;
+
+                       was_interruptible = dev_priv->mm.interruptible;
+                       dev_priv->mm.interruptible = false;
+
+                       list_for_each_entry_safe(vma, tmp, &obj->vma_list, vma_link) {
+                               int ret = i915_vma_unbind(vma);
+                               WARN_ON(ret && ret != -EIO);
+                       }
+                       WARN_ON(i915_gem_object_put_pages(obj));
+
+                       dev_priv->mm.interruptible = was_interruptible;
+               }
+
+               start = obj->userptr.ptr + obj->base.size;
+
+               drm_gem_object_unreference(&obj->base);
+               mutex_unlock(&mn->dev->struct_mutex);
+       }
+}
+
+static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
+       .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
+};
+
+static struct i915_mmu_notifier *
+__i915_mmu_notifier_lookup(struct drm_device *dev, struct mm_struct *mm)
+{
+       struct drm_i915_private *dev_priv = to_i915(dev);
+       struct i915_mmu_notifier *mmu;
+
+       /* Protected by dev->struct_mutex */
+       hash_for_each_possible(dev_priv->mmu_notifiers, mmu, node, (unsigned long)mm)
+               if (mmu->mm == mm)
+                       return mmu;
+
+       return NULL;
+}
+
+static struct i915_mmu_notifier *
+i915_mmu_notifier_get(struct drm_device *dev, struct mm_struct *mm)
+{
+       struct drm_i915_private *dev_priv = to_i915(dev);
+       struct i915_mmu_notifier *mmu;
+       int ret;
+
+       lockdep_assert_held(&dev->struct_mutex);
+
+       mmu = __i915_mmu_notifier_lookup(dev, mm);
+       if (mmu)
+               return mmu;
+
+       mmu = kmalloc(sizeof(*mmu), GFP_KERNEL);
+       if (mmu == NULL)
+               return ERR_PTR(-ENOMEM);
+
+       spin_lock_init(&mmu->lock);
+       mmu->dev = dev;
+       mmu->mn.ops = &i915_gem_userptr_notifier;
+       mmu->mm = mm;
+       mmu->objects = RB_ROOT;
+       mmu->count = 0;
+       mmu->serial = 0;
+
+       /* Protected by mmap_sem (write-lock) */
+       ret = __mmu_notifier_register(&mmu->mn, mm);
+       if (ret) {
+               kfree(mmu);
+               return ERR_PTR(ret);
+       }
+
+       /* Protected by dev->struct_mutex */
+       hash_add(dev_priv->mmu_notifiers, &mmu->node, (unsigned long)mm);
+       return mmu;
+}
+
+static void
+__i915_mmu_notifier_destroy_worker(struct work_struct *work)
+{
+       struct i915_mmu_notifier *mmu = container_of(work, typeof(*mmu), work);
+       mmu_notifier_unregister(&mmu->mn, mmu->mm);
+       kfree(mmu);
+}
+
+static void
+__i915_mmu_notifier_destroy(struct i915_mmu_notifier *mmu)
+{
+       lockdep_assert_held(&mmu->dev->struct_mutex);
+
+       /* Protected by dev->struct_mutex */
+       hash_del(&mmu->node);
+
+       /* Our lock ordering is: mmap_sem, mmu_notifier_scru, struct_mutex.
+        * We enter the function holding struct_mutex, therefore we need
+        * to drop our mutex prior to calling mmu_notifier_unregister in
+        * order to prevent lock inversion (and system-wide deadlock)
+        * between the mmap_sem and struct-mutex. Hence we defer the
+        * unregistration to a workqueue where we hold no locks.
+        */
+       INIT_WORK(&mmu->work, __i915_mmu_notifier_destroy_worker);
+       schedule_work(&mmu->work);
+}
+
+static void __i915_mmu_notifier_update_serial(struct i915_mmu_notifier *mmu)
+{
+       if (++mmu->serial == 0)
+               mmu->serial = 1;
+}
+
+static void
+i915_mmu_notifier_del(struct i915_mmu_notifier *mmu,
+                     struct i915_mmu_object *mn)
+{
+       lockdep_assert_held(&mmu->dev->struct_mutex);
+
+       spin_lock(&mmu->lock);
+       interval_tree_remove(&mn->it, &mmu->objects);
+       __i915_mmu_notifier_update_serial(mmu);
+       spin_unlock(&mmu->lock);
+
+       /* Protected against _add() by dev->struct_mutex */
+       if (--mmu->count == 0)
+               __i915_mmu_notifier_destroy(mmu);
+}
+
+static int
+i915_mmu_notifier_add(struct i915_mmu_notifier *mmu,
+                     struct i915_mmu_object *mn)
+{
+       struct interval_tree_node *it;
+       int ret;
+
+       ret = i915_mutex_lock_interruptible(mmu->dev);
+       if (ret)
+               return ret;
+
+       /* Make sure we drop the final active reference (and thereby
+        * remove the objects from the interval tree) before we do
+        * the check for overlapping objects.
+        */
+       i915_gem_retire_requests(mmu->dev);
+
+       /* Disallow overlapping userptr objects */
+       spin_lock(&mmu->lock);
+       it = interval_tree_iter_first(&mmu->objects,
+                                     mn->it.start, mn->it.last);
+       if (it) {
+               struct drm_i915_gem_object *obj;
+
+               /* We only need to check the first object in the range as it
+                * either has cancelled gup work queued and we need to
+                * return back to the user to give time for the gup-workers
+                * to flush their object references upon which the object will
+                * be removed from the interval-tree, or the the range is
+                * still in use by another client and the overlap is invalid.
+                */
+
+               obj = container_of(it, struct i915_mmu_object, it)->obj;
+               ret = obj->userptr.workers ? -EAGAIN : -EINVAL;
+       } else {
+               interval_tree_insert(&mn->it, &mmu->objects);
+               __i915_mmu_notifier_update_serial(mmu);
+               ret = 0;
+       }
+       spin_unlock(&mmu->lock);
+       mutex_unlock(&mmu->dev->struct_mutex);
+
+       return ret;
+}
+
+static void
+i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
+{
+       struct i915_mmu_object *mn;
+
+       mn = obj->userptr.mn;
+       if (mn == NULL)
+               return;
+
+       i915_mmu_notifier_del(mn->mmu, mn);
+       obj->userptr.mn = NULL;
+}
+
+static int
+i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
+                                   unsigned flags)
+{
+       struct i915_mmu_notifier *mmu;
+       struct i915_mmu_object *mn;
+       int ret;
+
+       if (flags & I915_USERPTR_UNSYNCHRONIZED)
+               return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
+
+       down_write(&obj->userptr.mm->mmap_sem);
+       ret = i915_mutex_lock_interruptible(obj->base.dev);
+       if (ret == 0) {
+               mmu = i915_mmu_notifier_get(obj->base.dev, obj->userptr.mm);
+               if (!IS_ERR(mmu))
+                       mmu->count++; /* preemptive add to act as a refcount */
+               else
+                       ret = PTR_ERR(mmu);
+               mutex_unlock(&obj->base.dev->struct_mutex);
+       }
+       up_write(&obj->userptr.mm->mmap_sem);
+       if (ret)
+               return ret;
+
+       mn = kzalloc(sizeof(*mn), GFP_KERNEL);
+       if (mn == NULL) {
+               ret = -ENOMEM;
+               goto destroy_mmu;
+       }
+
+       mn->mmu = mmu;
+       mn->it.start = obj->userptr.ptr;
+       mn->it.last = mn->it.start + obj->base.size - 1;
+       mn->obj = obj;
+
+       ret = i915_mmu_notifier_add(mmu, mn);
+       if (ret)
+               goto free_mn;
+
+       obj->userptr.mn = mn;
+       return 0;
+
+free_mn:
+       kfree(mn);
+destroy_mmu:
+       mutex_lock(&obj->base.dev->struct_mutex);
+       if (--mmu->count == 0)
+               __i915_mmu_notifier_destroy(mmu);
+       mutex_unlock(&obj->base.dev->struct_mutex);
+       return ret;
+}
+
+#else
+
+static void
+i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
+{
+}
+
+static int
+i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
+                                   unsigned flags)
+{
+       if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
+               return -ENODEV;
+
+       if (!capable(CAP_SYS_ADMIN))
+               return -EPERM;
+
+       return 0;
+}
+#endif
+
+struct get_pages_work {
+       struct work_struct work;
+       struct drm_i915_gem_object *obj;
+       struct task_struct *task;
+};
+
+
+#if IS_ENABLED(CONFIG_SWIOTLB)
+#define swiotlb_active() swiotlb_nr_tbl()
+#else
+#define swiotlb_active() 0
+#endif
+
+static int
+st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
+{
+       struct scatterlist *sg;
+       int ret, n;
+
+       *st = kmalloc(sizeof(**st), GFP_KERNEL);
+       if (*st == NULL)
+               return -ENOMEM;
+
+       if (swiotlb_active()) {
+               ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
+               if (ret)
+                       goto err;
+
+               for_each_sg((*st)->sgl, sg, num_pages, n)
+                       sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
+       } else {
+               ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
+                                               0, num_pages << PAGE_SHIFT,
+                                               GFP_KERNEL);
+               if (ret)
+                       goto err;
+       }
+
+       return 0;
+
+err:
+       kfree(*st);
+       *st = NULL;
+       return ret;
+}
+
+static void
+__i915_gem_userptr_get_pages_worker(struct work_struct *_work)
+{
+       struct get_pages_work *work = container_of(_work, typeof(*work), work);
+       struct drm_i915_gem_object *obj = work->obj;
+       struct drm_device *dev = obj->base.dev;
+       const int num_pages = obj->base.size >> PAGE_SHIFT;
+       struct page **pvec;
+       int pinned, ret;
+
+       ret = -ENOMEM;
+       pinned = 0;
+
+       pvec = kmalloc(num_pages*sizeof(struct page *),
+                      GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
+       if (pvec == NULL)
+               pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
+       if (pvec != NULL) {
+               struct mm_struct *mm = obj->userptr.mm;
+
+               down_read(&mm->mmap_sem);
+               while (pinned < num_pages) {
+                       ret = get_user_pages(work->task, mm,
+                                            obj->userptr.ptr + pinned * PAGE_SIZE,
+                                            num_pages - pinned,
+                                            !obj->userptr.read_only, 0,
+                                            pvec + pinned, NULL);
+                       if (ret < 0)
+                               break;
+
+                       pinned += ret;
+               }
+               up_read(&mm->mmap_sem);
+       }
+
+       mutex_lock(&dev->struct_mutex);
+       if (obj->userptr.work != &work->work) {
+               ret = 0;
+       } else if (pinned == num_pages) {
+               ret = st_set_pages(&obj->pages, pvec, num_pages);
+               if (ret == 0) {
+                       list_add_tail(&obj->global_list, &to_i915(dev)->mm.unbound_list);
+                       pinned = 0;
+               }
+       }
+
+       obj->userptr.work = ERR_PTR(ret);
+       obj->userptr.workers--;
+       drm_gem_object_unreference(&obj->base);
+       mutex_unlock(&dev->struct_mutex);
+
+       release_pages(pvec, pinned, 0);
+       drm_free_large(pvec);
+
+       put_task_struct(work->task);
+       kfree(work);
+}
+
+static int
+i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
+{
+       const int num_pages = obj->base.size >> PAGE_SHIFT;
+       struct page **pvec;
+       int pinned, ret;
+
+       /* 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 = NULL;
+       pinned = 0;
+       if (obj->userptr.mm == current->mm) {
+               pvec = kmalloc(num_pages*sizeof(struct page *),
+                              GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
+               if (pvec == NULL) {
+                       pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
+                       if (pvec == NULL)
+                               return -ENOMEM;
+               }
+
+               pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
+                                              !obj->userptr.read_only, pvec);
+       }
+       if (pinned < num_pages) {
+               if (pinned < 0) {
+                       ret = pinned;
+                       pinned = 0;
+               } else {
+                       /* Spawn a worker so that we can acquire the
+                        * user pages without holding our mutex. Access
+                        * to the user pages requires mmap_sem, and we have
+                        * a strict lock ordering of mmap_sem, struct_mutex -
+                        * we already hold struct_mutex here and so cannot
+                        * call gup without encountering a lock inversion.
+                        *
+                        * Userspace will keep on repeating the operation
+                        * (thanks to EAGAIN) until either we hit the fast
+                        * path or the worker completes. If the worker is
+                        * cancelled or superseded, the task is still run
+                        * but the results ignored. (This leads to
+                        * complications that we may have a stray object
+                        * refcount that we need to be wary of when
+                        * checking for existing objects during creation.)
+                        * If the worker encounters an error, it reports
+                        * that error back to this function through
+                        * obj->userptr.work = ERR_PTR.
+                        */
+                       ret = -EAGAIN;
+                       if (obj->userptr.work == NULL &&
+                           obj->userptr.workers < I915_GEM_USERPTR_MAX_WORKERS) {
+                               struct get_pages_work *work;
+
+                               work = kmalloc(sizeof(*work), GFP_KERNEL);
+                               if (work != NULL) {
+                                       obj->userptr.work = &work->work;
+                                       obj->userptr.workers++;
+
+                                       work->obj = obj;
+                                       drm_gem_object_reference(&obj->base);
+
+                                       work->task = current;
+                                       get_task_struct(work->task);
+
+                                       INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
+                                       schedule_work(&work->work);
+                               } else
+                                       ret = -ENOMEM;
+                       } else {
+                               if (IS_ERR(obj->userptr.work)) {
+                                       ret = PTR_ERR(obj->userptr.work);
+                                       obj->userptr.work = NULL;
+                               }
+                       }
+               }
+       } else {
+               ret = st_set_pages(&obj->pages, pvec, num_pages);
+               if (ret == 0) {
+                       obj->userptr.work = NULL;
+                       pinned = 0;
+               }
+       }
+
+       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;
+
+       BUG_ON(obj->userptr.work != NULL);
+
+       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)
+{
+       i915_gem_userptr_release__mmu_notifier(obj);
+
+       if (obj->userptr.mm) {
+               mmput(obj->userptr.mm);
+               obj->userptr.mm = NULL;
+       }
+}
+
+static int
+i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
+{
+       if (obj->userptr.mn)
+               return 0;
+
+       return i915_gem_userptr_init__mmu_notifier(obj, 0);
+}
+
+static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
+       .dmabuf_export = i915_gem_userptr_dmabuf_export,
+       .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 normal memory from the process
+ * context - user memory.
+ *
+ * We impose several restrictions upon the memory being mapped
+ * into the GPU.
+ * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
+ * 2. It cannot overlap any other userptr object in the same address space.
+ * 3. It must be normal system memory, not a pointer into another map of IO
+ *    space (e.g. it must not be a GTT mmapping of another object).
+ * 4. We only allow a bo as large as we could in theory map into the GTT,
+ *    that is we limit the size to the total size of the GTT.
+ * 5. The bo is marked as being snoopable. The backing pages are left
+ *    accessible directly by the CPU, but reads and writes by the GPU may
+ *    incur the cost of a snoop (unless you have an LLC architecture).
+ *
+ * Synchronisation between multiple users and the GPU is left to userspace
+ * through the normal set-domain-ioctl. The kernel will enforce that the
+ * GPU relinquishes the VMA before it is returned back to the system
+ * i.e. upon free(), munmap() or process termination. However, the userspace
+ * malloc() library may not immediately relinquish the VMA after free() and
+ * instead reuse it whilst the GPU is still reading and writing to the VMA.
+ * Caveat emptor.
+ *
+ * Also note, that the object created here is not currently a "first class"
+ * object, in that several ioctls are banned. These are the CPU access
+ * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
+ * direct access via your pointer rather than use those ioctls.
+ *
+ * If you think this is a good interface to use to pass GPU memory between
+ * drivers, please use dma-buf instead. In fact, wherever possible use
+ * dma-buf instead.
+ */
+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 drm_i915_gem_object *obj;
+       int ret;
+       u32 handle;
+
+       if (args->flags & ~(I915_USERPTR_READ_ONLY |
+                           I915_USERPTR_UNSYNCHRONIZED))
+               return -EINVAL;
+
+       if (offset_in_page(args->user_ptr | args->user_size))
+               return -EINVAL;
+
+       if (args->user_size > dev_priv->gtt.base.total)
+               return -E2BIG;
+
+       if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
+                      (char __user *)(unsigned long)args->user_ptr, args->user_size))
+               return -EFAULT;
+
+       if (args->flags & I915_USERPTR_READ_ONLY) {
+               /* On almost all of the current hw, we cannot tell the GPU that a
+                * page is readonly, so this is just a placeholder in the uAPI.
+                */
+               return -ENODEV;
+       }
+
+       /* Allocate the new object */
+       obj = i915_gem_object_alloc(dev);
+       if (obj == NULL)
+               return -ENOMEM;
+
+       drm_gem_private_object_init(dev, &obj->base, args->user_size);
+       i915_gem_object_init(obj, &i915_gem_userptr_ops);
+       obj->cache_level = I915_CACHE_LLC;
+       obj->base.write_domain = I915_GEM_DOMAIN_CPU;
+       obj->base.read_domains = I915_GEM_DOMAIN_CPU;
+
+       obj->userptr.ptr = args->user_ptr;
+       obj->userptr.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.
+        */
+       ret = -ENOMEM;
+       if ((obj->userptr.mm = get_task_mm(current)))
+               ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
+       if (ret == 0)
+               ret = drm_gem_handle_create(file, &obj->base, &handle);
+
+       /* drop reference from allocate - handle holds it now */
+       drm_gem_object_unreference_unlocked(&obj->base);
+       if (ret)
+               return ret;
+
+       args->handle = handle;
+       return 0;
+}
+
+int
+i915_gem_init_userptr(struct drm_device *dev)
+{
+#if defined(CONFIG_MMU_NOTIFIER)
+       struct drm_i915_private *dev_priv = to_i915(dev);
+       hash_init(dev_priv->mmu_notifiers);
+#endif
+       return 0;
+}
index 2d819858c19b03fa47978485e392e1897c7a6862..dcbec692f60f38f5dc27dc88289f4b83848d1923 100644 (file)
@@ -205,6 +205,7 @@ static void print_error_buffers(struct drm_i915_error_state_buf *m,
                err_puts(m, tiling_flag(err->tiling));
                err_puts(m, dirty_flag(err->dirty));
                err_puts(m, purgeable_flag(err->purgeable));
+               err_puts(m, err->userptr ? " userptr" : "");
                err_puts(m, err->ring != -1 ? " " : "");
                err_puts(m, ring_str(err->ring));
                err_puts(m, i915_cache_level_str(err->cache_level));
@@ -641,6 +642,7 @@ static void capture_bo(struct drm_i915_error_buffer *err,
        err->tiling = obj->tiling_mode;
        err->dirty = obj->dirty;
        err->purgeable = obj->madv != I915_MADV_WILLNEED;
+       err->userptr = obj->userptr.mm != NULL;
        err->ring = obj->ring ? obj->ring->id : -1;
        err->cache_level = obj->cache_level;
 }
index 8a3e4ef00c3db418b832a709ca3a3a9eb6c1668b..ff57f07c32498933be7be28d874fb09a164f5b04 100644 (file)
@@ -223,6 +223,7 @@ typedef struct _drm_i915_sarea {
 #define DRM_I915_GEM_GET_CACHING       0x30
 #define DRM_I915_REG_READ              0x31
 #define DRM_I915_GET_RESET_STATS       0x32
+#define DRM_I915_GEM_USERPTR           0x33
 
 #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)
@@ -273,6 +274,7 @@ typedef struct _drm_i915_sarea {
 #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_GET_RESET_STATS         DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GET_RESET_STATS, struct drm_i915_reset_stats)
+#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.
@@ -1050,4 +1052,18 @@ struct drm_i915_reset_stats {
        __u32 pad;
 };
 
+struct drm_i915_gem_userptr {
+       __u64 user_ptr;
+       __u64 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_ */