Re: [PATCH] drm/xe: sysfs_ops needs to be defined on parent directory

[Rodrigo Vivi <rodrigo.vivi@intel.com>]

On Wed, Mar 26, 2025 at 04:14:21AM +0000, Upadhyay, Tejas wrote:
> 
> 
> > -----Original Message-----
> > From: Vivi, Rodrigo <rodrigo.vivi@intel.com>
> > Sent: Wednesday, March 26, 2025 12:43 AM
> > To: Upadhyay, Tejas <tejas.upadhyay@intel.com>
> > Cc: intel-xe@lists.freedesktop.org; Ghimiray, Himal Prasad
> > <himal.prasad.ghimiray@intel.com>
> > Subject: Re: [PATCH] drm/xe: sysfs_ops needs to be defined on parent
> > directory
> > 
> > On Wed, Mar 19, 2025 at 05:43:49PM +0530, Tejas Upadhyay wrote:
> > > Currently, xe_hw_engine_sysfs_kobj_type is defining sysfs_ops on wrong
> > > directory. Sysfs_ops needs to be defined on immediate parent directory
> > > to be able to called on each attribute set/get.
> > 
> > Please bare with me, but I'm having a hard time to follow this reasoning and
> > the patch, why, and everything else going on here...
> > 
> > >
> > > Fixes: 3f0e14651ab0 ("drm/xe: Runtime PM wake on every sysfs call")
> > 
> > Why this patch is claiming to fix this? Please note that this mentioned patch
> > just replace the default kobj_sysfs_ops per the new introduced
> > xe_hw_engine_class_sysfs_ops. They are basically identical functions.
> > The only difference is that the new one call our runtime pm wrappers around
> > the calls that we need.
> 
> Earlier we were not doing any common ops for all attr set/get, each attribute has individual setter/getter. With 3f0e14651ab0 ("drm/xe: Runtime PM wake on every sysfs call"), we introduced common ops on all attr set/get and which is must to be called now. Thus I though of adding fixes. Please let me know if you guys think otherwise I will just remove, not strong objection.

I'm sorry for the noise. I probably had enough coffee today,
so yes, this is the right Fixes tag ;)

>  
> > 
> > It never touched anything that this patch is touching below.
> > 
> > Perhaps if we also need on the sysfs on upper directory, perhaps we need to
> > also replace the default sysfs_ops on other places like in the upper directory?
> > 
> > > Signed-off-by: Tejas Upadhyay <tejas.upadhyay@intel.com>
> > > ---
> > >  drivers/gpu/drm/xe/xe_hw_engine_class_sysfs.c | 67
> > > +++++++++----------
> > >  1 file changed, 33 insertions(+), 34 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/xe/xe_hw_engine_class_sysfs.c
> > > b/drivers/gpu/drm/xe/xe_hw_engine_class_sysfs.c
> > > index b53e8d2accdb..25592f178482 100644
> > > --- a/drivers/gpu/drm/xe/xe_hw_engine_class_sysfs.c
> > > +++ b/drivers/gpu/drm/xe/xe_hw_engine_class_sysfs.c
> > > @@ -492,39 +492,6 @@ static const struct attribute * const files[] = {
> > >  	NULL
> > >  };
> > >
> > > -static void kobj_xe_hw_engine_class_fini(void *arg) -{
> > > -	struct kobject *kobj = arg;
> > > -
> > > -	sysfs_remove_files(kobj, files);
> > > -	kobject_put(kobj);
> > > -}
> > > -
> > > -static struct kobj_eclass *
> > > -kobj_xe_hw_engine_class(struct xe_device *xe, struct kobject *parent,
> > > const char *name) -{
> > > -	struct kobj_eclass *keclass;
> > > -	int err = 0;
> > > -
> > > -	keclass = kzalloc(sizeof(*keclass), GFP_KERNEL);
> > > -	if (!keclass)
> > > -		return NULL;
> > > -
> > > -	kobject_init(&keclass->base, &kobj_xe_hw_engine_type);
> > > -	if (kobject_add(&keclass->base, parent, "%s", name)) {
> > > -		kobject_put(&keclass->base);
> > > -		return NULL;
> > > -	}
> > > -	keclass->xe = xe;
> > > -
> > > -	err = devm_add_action_or_reset(xe->drm.dev,
> > kobj_xe_hw_engine_class_fini,
> > > -				       &keclass->base);
> > > -	if (err)
> > > -		return NULL;
> > > -
> > > -	return keclass;
> > > -}
> > > -
> > >  static void hw_engine_class_defaults_fini(void *arg)  {
> > >  	struct kobject *kobj = arg;
> > > @@ -611,6 +578,38 @@ static const struct kobj_type
> > xe_hw_engine_sysfs_kobj_type = {
> > >  	.sysfs_ops = &xe_hw_engine_class_sysfs_ops,  };
> > >
> > > +static void kobj_xe_hw_engine_class_fini(void *arg) {
> > > +	struct kobject *kobj = arg;
> > > +
> > > +	sysfs_remove_files(kobj, files);
> > > +	kobject_put(kobj);
> > > +}
> > > +
> > > +static struct kobj_eclass *
> > > +kobj_xe_hw_engine_class(struct xe_device *xe, struct kobject *parent,
> > > +const char *name) {
> > > +	struct kobj_eclass *keclass;
> > > +	int err = 0;
> > > +
> > > +	keclass = kzalloc(sizeof(*keclass), GFP_KERNEL);
> > > +	if (!keclass)
> > > +		return NULL;
> > > +
> > > +	kobject_init(&keclass->base, &xe_hw_engine_sysfs_kobj_type);
> > > +	if (kobject_add(&keclass->base, parent, "%s", name)) {
> > > +		kobject_put(&keclass->base);
> > > +		return NULL;
> > > +	}
> > > +	keclass->xe = xe;
> > > +
> > > +	err = devm_add_action_or_reset(xe->drm.dev,
> > kobj_xe_hw_engine_class_fini,
> > > +				       &keclass->base);
> > > +	if (err)
> > > +		return NULL;
> > > +
> > > +	return keclass;
> > > +}
> > 
> > I couldn't spot any difference from the both chunks above. So this patch is
> > more moving things around than doing any change right? perhaps a different
> > patch or a mention on the commit message itself?
> 
> The change is, immediate parent directory should be defined with attr_show/store ops if we want it to be called on every attr set/get call. Currently we had sysfs_ops defined on parent to parent directory which does not get called on underlying child's attr set/get call. 

Okay, now I could spot the change:

inside kobj_xe_hw_engine_class()
-       kobject_init(&keclass->base, &kobj_xe_hw_engine_type);
+       kobject_init(&keclass->base, &xe_hw_engine_sysfs_kobj_type);

and inside xe_hw_engine_class_sysfs_init()

-       kobject_init(kobj, &xe_hw_engine_sysfs_kobj_type);
+       kobject_init(kobj, &kobj_xe_hw_engine_type);

With this you invert where the runtime_pm is called moving it to the right parent.
I partially agree with the fix, but I don't agree with the patch itself.
Not only because I got confused, but because it creates inconsistency:

kobj_xe_hw_engine_class now uses xe_hw_engine_sysfs_kobj_type
and
xe_hw_engine_class_sysfs_init now uses kobj_xe_hw_engine_type

If we want to invert, we need to invert the sysfs_ops inside the type functions.
But also, I don't believe that there is the problem on leaving the runtime_pm
hooks in the parent of parent. Just in case we end up adding some extra file
there.

So, What about simply:

 static const struct kobj_type kobj_xe_hw_engine_type = {
        .release = kobj_xe_hw_engine_release,
-       .sysfs_ops = &kobj_sysfs_ops
+       .sysfs_ops = &xe_hw_engine_class_sysfs_ops,
 };

Thanks for catching that up and fixing it,
Rodrigo.

> 
> Tejas
> > 
> > >  static void hw_engine_class_sysfs_fini(void *arg)  {
> > >  	struct kobject *kobj = arg;
> > > @@ -640,7 +639,7 @@ int xe_hw_engine_class_sysfs_init(struct xe_gt *gt)
> > >  	if (!kobj)
> > >  		return -ENOMEM;
> > >
> > > -	kobject_init(kobj, &xe_hw_engine_sysfs_kobj_type);
> > > +	kobject_init(kobj, &kobj_xe_hw_engine_type);
> > 
> > now it looks like this is the only real chunk of this patch, last touched when
> > you added it:
> > 038ff941afe2 ("drm/xe: Add sysfs entries for engines under its GT")
> > 
> > And now, after this patch, who is now using xe_hw_engine_sysfs_kobj_type?
> > 
> > Looking further to both types, perhaps we need to kill the
> > kobj_xe_hw_engine_type itself in favor of the xe_ one?
> > 
> > Or perhaps we need something like this:
> > 
> >  static const struct kobj_type kobj_xe_hw_engine_type = {
> >         .release = kobj_xe_hw_engine_release,
> > -       .sysfs_ops = &kobj_sysfs_ops
> > +       .sysfs_ops = &xe_hw_engine_class_sysfs_ops,
> >  };
> > 
> > >
> > >  	err = kobject_add(kobj, gt->sysfs, "engines");
> > >  	if (err)
> > > --
> > > 2.34.1
> > >