Message ID | 20220416172325.1039795-1-trix@redhat.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | drm/i915: change node clearing from memset to initialization | expand |
On Sat, 2022-04-16 at 13:23 -0400, Tom Rix wrote: > In insert_mappable_node(), the parameter node is > cleared late in node's use with memset. > insert_mappable_node() is a singleton, called only > from i915_gem_gtt_prepare() which itself is only > called by i915_gem_gtt_pread() and > i915_gem_gtt_pwrite_fast() where the definition of > node originates. > > Instead of using memset, initialize node to 0 at it's > definitions. trivia: /it's/its/ Only reason _not_ to do this is memset is guaranteed to zero any padding that might go to userspace. But it doesn't seem there is any padding anyway nor is the struct available to userspace. So this seems fine though it might increase overall code size a tiny bit. I do have a caveat: see below: > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c [] > @@ -328,7 +327,6 @@ static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj, > goto err_ww; > } else if (!IS_ERR(vma)) { > node->start = i915_ggtt_offset(vma); > - node->flags = 0; Why is this unneeded? from: drm_mm_insert_node_in_range which can set node->flags __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
On 4/16/22 11:33 AM, Joe Perches wrote: > On Sat, 2022-04-16 at 13:23 -0400, Tom Rix wrote: >> In insert_mappable_node(), the parameter node is >> cleared late in node's use with memset. >> insert_mappable_node() is a singleton, called only >> from i915_gem_gtt_prepare() which itself is only >> called by i915_gem_gtt_pread() and >> i915_gem_gtt_pwrite_fast() where the definition of >> node originates. >> >> Instead of using memset, initialize node to 0 at it's >> definitions. > trivia: /it's/its/ > > Only reason _not_ to do this is memset is guaranteed to > zero any padding that might go to userspace. > > But it doesn't seem there is any padding anyway nor is > the struct available to userspace. > > So this seems fine though it might increase overall code > size a tiny bit. > > I do have a caveat: see below: > >> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c > [] >> @@ -328,7 +327,6 @@ static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj, >> goto err_ww; >> } else if (!IS_ERR(vma)) { >> node->start = i915_ggtt_offset(vma); >> - node->flags = 0; > Why is this unneeded? node = {} initializes all of node's elements to 0, including this one. Tom > > from: drm_mm_insert_node_in_range which can set node->flags > > __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags); > >
On Sat, 2022-04-16 at 13:48 -0700, Tom Rix wrote: > On 4/16/22 11:33 AM, Joe Perches wrote: > > On Sat, 2022-04-16 at 13:23 -0400, Tom Rix wrote: > > > In insert_mappable_node(), the parameter node is > > > cleared late in node's use with memset. > > > insert_mappable_node() is a singleton, called only > > > from i915_gem_gtt_prepare() which itself is only > > > called by i915_gem_gtt_pread() and > > > i915_gem_gtt_pwrite_fast() where the definition of > > > node originates. > > > > > > Instead of using memset, initialize node to 0 at it's > > > definitions. > > trivia: /it's/its/ > > > > Only reason _not_ to do this is memset is guaranteed to > > zero any padding that might go to userspace. > > > > But it doesn't seem there is any padding anyway nor is > > the struct available to userspace. > > > > So this seems fine though it might increase overall code > > size a tiny bit. > > > > I do have a caveat: see below: > > > > > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c > > [] > > > @@ -328,7 +327,6 @@ static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj, > > > goto err_ww; > > > } else if (!IS_ERR(vma)) { > > > node->start = i915_ggtt_offset(vma); > > > - node->flags = 0; > > Why is this unneeded? > > node = {} initializes all of node's elements to 0, including this one. true, but could the call to insert_mappable_node combined with the retry goto in i915_gem_gtt_prepare set this to non-zero?
On 4/16/22 2:04 PM, Joe Perches wrote: > On Sat, 2022-04-16 at 13:48 -0700, Tom Rix wrote: >> On 4/16/22 11:33 AM, Joe Perches wrote: >>> On Sat, 2022-04-16 at 13:23 -0400, Tom Rix wrote: >>>> In insert_mappable_node(), the parameter node is >>>> cleared late in node's use with memset. >>>> insert_mappable_node() is a singleton, called only >>>> from i915_gem_gtt_prepare() which itself is only >>>> called by i915_gem_gtt_pread() and >>>> i915_gem_gtt_pwrite_fast() where the definition of >>>> node originates. >>>> >>>> Instead of using memset, initialize node to 0 at it's >>>> definitions. >>> trivia: /it's/its/ >>> >>> Only reason _not_ to do this is memset is guaranteed to >>> zero any padding that might go to userspace. >>> >>> But it doesn't seem there is any padding anyway nor is >>> the struct available to userspace. >>> >>> So this seems fine though it might increase overall code >>> size a tiny bit. >>> >>> I do have a caveat: see below: >>> >>>> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c >>> [] >>>> @@ -328,7 +327,6 @@ static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj, >>>> goto err_ww; >>>> } else if (!IS_ERR(vma)) { >>>> node->start = i915_ggtt_offset(vma); >>>> - node->flags = 0; >>> Why is this unneeded? >> node = {} initializes all of node's elements to 0, including this one. > true, but could the call to insert_mappable_node combined with the > retry goto in i915_gem_gtt_prepare set this to non-zero? Yikes! I'll add that back. Thanks for pointing it out. Tom > >
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 2e10187cd0a0..7dbd0b325c43 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -69,7 +69,6 @@ insert_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node, u32 size) if (err) return err; - memset(node, 0, sizeof(*node)); err = drm_mm_insert_node_in_range(&ggtt->vm.mm, node, size, 0, I915_COLOR_UNEVICTABLE, 0, ggtt->mappable_end, @@ -328,7 +327,6 @@ static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj, goto err_ww; } else if (!IS_ERR(vma)) { node->start = i915_ggtt_offset(vma); - node->flags = 0; } else { ret = insert_mappable_node(ggtt, node, PAGE_SIZE); if (ret) @@ -381,7 +379,7 @@ i915_gem_gtt_pread(struct drm_i915_gem_object *obj, struct drm_i915_private *i915 = to_i915(obj->base.dev); struct i915_ggtt *ggtt = to_gt(i915)->ggtt; intel_wakeref_t wakeref; - struct drm_mm_node node; + struct drm_mm_node node = {}; void __user *user_data; struct i915_vma *vma; u64 remain, offset; @@ -538,7 +536,7 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj, struct i915_ggtt *ggtt = to_gt(i915)->ggtt; struct intel_runtime_pm *rpm = &i915->runtime_pm; intel_wakeref_t wakeref; - struct drm_mm_node node; + struct drm_mm_node node = {}; struct i915_vma *vma; u64 remain, offset; void __user *user_data;
In insert_mappable_node(), the parameter node is cleared late in node's use with memset. insert_mappable_node() is a singleton, called only from i915_gem_gtt_prepare() which itself is only called by i915_gem_gtt_pread() and i915_gem_gtt_pwrite_fast() where the definition of node originates. Instead of using memset, initialize node to 0 at it's definitions. And remove unneeded clearing of the flags element. Signed-off-by: Tom Rix <trix@redhat.com> --- drivers/gpu/drm/i915/i915_gem.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-)