[Intel-gfx] [PATCH 1/6] drm: pre allocate node for create_block
Ben Widawsky
ben at bwidawsk.net
Thu Jul 4 22:03:12 CEST 2013
On Thu, Jul 04, 2013 at 11:19:58AM +0200, David Herrmann wrote:
> Hi
>
> On Wed, Jul 3, 2013 at 11:45 PM, Ben Widawsky <ben at bwidawsk.net> wrote:
> > For an upcoming patch where we introduce the i915 VMA, it's ideal to
> > have the drm_mm_node as part of the VMA struct (ie. it's pre-allocated).
> > Part of the conversion to VMAs is to kill off obj->gtt_space. Doing this
> > will break a bunch of code, but amongst them are 2 callers of
> > drm_mm_create_block(), both related to stolen memory.
> >
> > It also allows us to embed the drm_mm_node into the object currently
> > which provides a nice transition over to the new code.
> >
> > v2: Reordered to do before ripping out obj->gtt_offset.
> > Some minor cleanups made available because of reordering.
> >
> > CC: <dri-devel at lists.freedesktop.org>
> > Signed-off-by: Ben Widawsky <ben at bwidawsk.net>
> > ---
> > drivers/gpu/drm/drm_mm.c | 16 +++++----------
> > drivers/gpu/drm/i915/i915_gem_gtt.c | 18 +++++++++++++----
> > drivers/gpu/drm/i915/i915_gem_stolen.c | 36 +++++++++++++++++++++++-----------
> > include/drm/drm_mm.h | 9 +++++----
> > 4 files changed, 49 insertions(+), 30 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/drm_mm.c b/drivers/gpu/drm/drm_mm.c
> > index 07cf99c..9e8dfbc 100644
> > --- a/drivers/gpu/drm/drm_mm.c
> > +++ b/drivers/gpu/drm/drm_mm.c
> > @@ -147,12 +147,10 @@ static void drm_mm_insert_helper(struct drm_mm_node *hole_node,
> > }
> > }
> >
> > -struct drm_mm_node *drm_mm_create_block(struct drm_mm *mm,
> > - unsigned long start,
> > - unsigned long size,
> > - bool atomic)
> > +int drm_mm_create_block(struct drm_mm *mm, struct drm_mm_node *node,
> > + unsigned long start, unsigned long size)
> > {
> > - struct drm_mm_node *hole, *node;
> > + struct drm_mm_node *hole;
> > unsigned long end = start + size;
> > unsigned long hole_start;
> > unsigned long hole_end;
> > @@ -161,10 +159,6 @@ struct drm_mm_node *drm_mm_create_block(struct drm_mm *mm,
> > if (hole_start > start || hole_end < end)
> > continue;
> >
> > - node = drm_mm_kmalloc(mm, atomic);
> > - if (unlikely(node == NULL))
> > - return NULL;
> > -
> > node->start = start;
> > node->size = size;
> > node->mm = mm;
> > @@ -184,11 +178,11 @@ struct drm_mm_node *drm_mm_create_block(struct drm_mm *mm,
> > node->hole_follows = 1;
> > }
> >
> > - return node;
> > + return 0;
> > }
> >
> > WARN(1, "no hole found for block 0x%lx + 0x%lx\n", start, size);
> > - return NULL;
> > + return -ENOSPC;
> > }
> > EXPORT_SYMBOL(drm_mm_create_block);
> >
> > diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
> > index 66929ea..5c6fc0e 100644
> > --- a/drivers/gpu/drm/i915/i915_gem_gtt.c
> > +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
> > @@ -629,14 +629,24 @@ void i915_gem_setup_global_gtt(struct drm_device *dev,
> >
> > /* Mark any preallocated objects as occupied */
> > list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
> > + int ret;
> > DRM_DEBUG_KMS("reserving preallocated space: %x + %zx\n",
> > obj->gtt_offset, obj->base.size);
> >
> > BUG_ON(obj->gtt_space != I915_GTT_RESERVED);
> > - obj->gtt_space = drm_mm_create_block(&dev_priv->mm.gtt_space,
> > - obj->gtt_offset,
> > - obj->base.size,
> > - false);
> > + obj->gtt_space = kzalloc(sizeof(*obj->gtt_space), GFP_KERNEL);
> > + if (!obj->gtt_space) {
> > + DRM_ERROR("Failed to preserve all objects\n");
> > + break;
> > + }
> > + ret = drm_mm_create_block(&dev_priv->mm.gtt_space,
> > + obj->gtt_space,
> > + obj->gtt_offset,
> > + obj->base.size);
> > + if (ret) {
> > + DRM_DEBUG_KMS("Reservation failed\n");
> > + kfree(obj->gtt_space);
>
> Are you sure you don't need:
> obj->gtt_space = NULL;
> here?
> I am no expert in i915 gem handling, but looking at i915_gem.c I think
> you might run into bugs if not.
I'm too lazy to actually check, but I believe you're probably right.
It's fixed in a later patch where I added the getters and use
node_allocated so I don't check obj->gtt_space != NULL anymore; but it
would potentially be a painful bisect point.
Thanks for catching it (and the following ones).
>
> Also, why did you add the "break;" above, but not here? I am confused.
The thought at the time was if kzalloc fails at this point, subsequent
kzallocs are really likely to fail also. drm_mm_create_block OTOH is
something I won't pretend to inquire about failure recurrence. I agree
it looks funny though, so I'll change the break to continue.
>
> > + }
> > obj->has_global_gtt_mapping = 1;
> > }
> >
> > diff --git a/drivers/gpu/drm/i915/i915_gem_stolen.c b/drivers/gpu/drm/i915/i915_gem_stolen.c
> > index 8e02344..f9db84a 100644
> > --- a/drivers/gpu/drm/i915/i915_gem_stolen.c
> > +++ b/drivers/gpu/drm/i915/i915_gem_stolen.c
> > @@ -330,6 +330,7 @@ i915_gem_object_create_stolen_for_preallocated(struct drm_device *dev,
> > struct drm_i915_private *dev_priv = dev->dev_private;
> > struct drm_i915_gem_object *obj;
> > struct drm_mm_node *stolen;
> > + int ret;
> >
> > if (dev_priv->mm.stolen_base == 0)
> > return NULL;
> > @@ -344,11 +345,15 @@ i915_gem_object_create_stolen_for_preallocated(struct drm_device *dev,
> > if (WARN_ON(size == 0))
> > return NULL;
> >
> > - stolen = drm_mm_create_block(&dev_priv->mm.stolen,
> > - stolen_offset, size,
> > - false);
> > - if (stolen == NULL) {
> > + stolen = kzalloc(sizeof(*stolen), GFP_KERNEL);
> > + if (!stolen)
> > + return NULL;
> > +
> > + ret = drm_mm_create_block(&dev_priv->mm.stolen, stolen, stolen_offset,
> > + size);
> > + if (ret) {
> > DRM_DEBUG_KMS("failed to allocate stolen space\n");
> > + kfree(stolen);
> > return NULL;
> > }
> >
> > @@ -369,13 +374,18 @@ i915_gem_object_create_stolen_for_preallocated(struct drm_device *dev,
> > * later.
> > */
> > if (drm_mm_initialized(&dev_priv->mm.gtt_space)) {
> > - obj->gtt_space = drm_mm_create_block(&dev_priv->mm.gtt_space,
> > - gtt_offset, size,
> > - false);
> > - if (obj->gtt_space == NULL) {
> > + obj->gtt_space = kzalloc(sizeof(*obj->gtt_space), GFP_KERNEL);
> > + if (!obj->gtt_space) {
> > + DRM_DEBUG_KMS("-ENOMEM stolen GTT space\n");
> > + goto unref_out;
> > + }
> > +
> > + ret = drm_mm_create_block(&dev_priv->mm.gtt_space,
> > + obj->gtt_space,
> > + gtt_offset, size);
> > + if (ret) {
> > DRM_DEBUG_KMS("failed to allocate stolen GTT space\n");
> > - drm_gem_object_unreference(&obj->base);
> > - return NULL;
> > + goto unref_out;
>
> Again:
> kfree(obj->gtt_space);
> obj->gtt_space = NULL;
> Otherwise, if gem-cleanup calls drm_mm_put_block() on an already
> removed node, you end up with NULL-derefs in drm_mm.c
>
> > }
> > } else
> > obj->gtt_space = I915_GTT_RESERVED;
> > @@ -385,8 +395,12 @@ i915_gem_object_create_stolen_for_preallocated(struct drm_device *dev,
> >
> > list_add_tail(&obj->global_list, &dev_priv->mm.bound_list);
> > list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
> > -
> > return obj;
> > +
> > +unref_out:
> > + drm_gem_object_unreference(&obj->base);
> > + drm_mm_put_block(stolen);
>
> "stolen" is already cleared by drm_gem_object_unreference(). So that's
> a double-free here.
>
> The drm_mm_create_block() change looks good.
> Cheers
> David
>
Thanks for reviewing the i915 parts so thoroughly :D
>
> > + return NULL;
> > }
> >
> > void
> > diff --git a/include/drm/drm_mm.h b/include/drm/drm_mm.h
> > index 88591ef..d8b56b7 100644
> > --- a/include/drm/drm_mm.h
> > +++ b/include/drm/drm_mm.h
> > @@ -138,10 +138,10 @@ static inline unsigned long drm_mm_hole_node_end(struct drm_mm_node *hole_node)
> > /*
> > * Basic range manager support (drm_mm.c)
> > */
> > -extern struct drm_mm_node *drm_mm_create_block(struct drm_mm *mm,
> > - unsigned long start,
> > - unsigned long size,
> > - bool atomic);
> > +extern int drm_mm_create_block(struct drm_mm *mm,
> > + struct drm_mm_node *node,
> > + unsigned long start,
> > + unsigned long size);
> > extern struct drm_mm_node *drm_mm_get_block_generic(struct drm_mm_node *node,
> > unsigned long size,
> > unsigned alignment,
> > @@ -155,6 +155,7 @@ extern struct drm_mm_node *drm_mm_get_block_range_generic(
> > unsigned long start,
> > unsigned long end,
> > int atomic);
> > +
> > static inline struct drm_mm_node *drm_mm_get_block(struct drm_mm_node *parent,
> > unsigned long size,
> > unsigned alignment)
> > --
> > 1.8.3.2
> >
> > _______________________________________________
> > dri-devel mailing list
> > dri-devel at lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/dri-devel
--
Ben Widawsky, Intel Open Source Technology Center
More information about the Intel-gfx
mailing list