On Sat, Dec 05, 2020 at 08:57:38PM +0200, Laurent Pinchart wrote:
> Hi Philipp,
>
> On Fri, Dec 04, 2020 at 11:12:20AM +0100, Philipp Zabel wrote:
> > On Fri, 2020-12-04 at 11:17 +0200, Laurent Pinchart wrote:
> > > On Fri, Sep 11, 2020 at 03:57:18PM +0200, Philipp Zabel wrote:
> > > > Add an alternative to drm_encoder_init() that allocates and initializes
> > > > an encoder and registers drm_encoder_cleanup() with
> > > > drmm_add_action_or_reset().
> > > >
> > > > Signed-off-by: Philipp Zabel <p.zabel@xxxxxxxxxxxxxx>
> > > > ---
> > > > Changes since v2:
> > > > - call va_start() / va_end() unconditionally
> > > > ---
> > > > drivers/gpu/drm/drm_encoder.c | 101 ++++++++++++++++++++++++++--------
> > > > include/drm/drm_encoder.h | 30 ++++++++++
> > > > 2 files changed, 108 insertions(+), 23 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/drm_encoder.c b/drivers/gpu/drm/drm_encoder.c
> > > > index e555281f43d4..f5414705f9ad 100644
> > > > --- a/drivers/gpu/drm/drm_encoder.c
> > > > +++ b/drivers/gpu/drm/drm_encoder.c
> > > > @@ -26,6 +26,7 @@
> > > > #include <drm/drm_device.h>
> > > > #include <drm/drm_drv.h>
> > > > #include <drm/drm_encoder.h>
> > > > +#include <drm/drm_managed.h>
> > > >
> > > > #include "drm_crtc_internal.h"
> > > >
> > > > @@ -91,25 +92,11 @@ void drm_encoder_unregister_all(struct drm_device *dev)
> > > > }
> > > > }
> > > >
> > > > -/**
> > > > - * drm_encoder_init - Init a preallocated encoder
> > > > - * @dev: drm device
> > > > - * @encoder: the encoder to init
> > > > - * @funcs: callbacks for this encoder
> > > > - * @encoder_type: user visible type of the encoder
> > > > - * @name: printf style format string for the encoder name, or NULL for default name
> > > > - *
> > > > - * Initialises a preallocated encoder. Encoder should be subclassed as part of
> > > > - * driver encoder objects. At driver unload time drm_encoder_cleanup() should be
> > > > - * called from the driver's &drm_encoder_funcs.destroy hook.
> > > > - *
> > > > - * Returns:
> > > > - * Zero on success, error code on failure.
> > > > - */
> > > > -int drm_encoder_init(struct drm_device *dev,
> > > > - struct drm_encoder *encoder,
> > > > - const struct drm_encoder_funcs *funcs,
> > > > - int encoder_type, const char *name, ...)
> > > > +__printf(5, 0)
> > > > +static int __drm_encoder_init(struct drm_device *dev,
> > > > + struct drm_encoder *encoder,
> > > > + const struct drm_encoder_funcs *funcs,
> > > > + int encoder_type, const char *name, va_list ap)
> > > > {
> > > > int ret;
> > > >
> > > > @@ -125,11 +112,7 @@ int drm_encoder_init(struct drm_device *dev,
> > > > encoder->encoder_type = encoder_type;
> > > > encoder->funcs = funcs;
> > > > if (name) {
> > > > - va_list ap;
> > > > -
> > > > - va_start(ap, name);
> > > > encoder->name = kvasprintf(GFP_KERNEL, name, ap);
> > > > - va_end(ap);
> > > > } else {
> > > > encoder->name = kasprintf(GFP_KERNEL, "%s-%d",
> > > > drm_encoder_enum_list[encoder_type].name,
> > > > @@ -150,6 +133,36 @@ int drm_encoder_init(struct drm_device *dev,
> > > >
> > > > return ret;
> > > > }
> > > > +
> > > > +/**
> > > > + * drm_encoder_init - Init a preallocated encoder
> > > > + * @dev: drm device
> > > > + * @encoder: the encoder to init
> > > > + * @funcs: callbacks for this encoder
> > > > + * @encoder_type: user visible type of the encoder
> > > > + * @name: printf style format string for the encoder name, or NULL for default name
> > > > + *
> > > > + * Initializes a preallocated encoder. Encoder should be subclassed as part of
> > > > + * driver encoder objects. At driver unload time drm_encoder_cleanup() should be
> > > > + * called from the driver's &drm_encoder_funcs.destroy hook.
> > > > + *
> > > > + * Returns:
> > > > + * Zero on success, error code on failure.
> > > > + */
> > > > +int drm_encoder_init(struct drm_device *dev,
> > > > + struct drm_encoder *encoder,
> > > > + const struct drm_encoder_funcs *funcs,
> > > > + int encoder_type, const char *name, ...)
> > > > +{
> > > > + va_list ap;
> > > > + int ret;
> > > > +
> > > > + va_start(ap, name);
> > > > + ret = __drm_encoder_init(dev, encoder, funcs, encoder_type, name, ap);
> > > > + va_end(ap);
> > > > +
> > > > + return ret;
> > > > +}
> > > > EXPORT_SYMBOL(drm_encoder_init);
> > > >
> > > > /**
> > > > @@ -181,6 +194,48 @@ void drm_encoder_cleanup(struct drm_encoder *encoder)
> > > > }
> > > > EXPORT_SYMBOL(drm_encoder_cleanup);
> > > >
> > > > +static void drmm_encoder_alloc_release(struct drm_device *dev, void *ptr)
> > > > +{
> > > > + struct drm_encoder *encoder = ptr;
> > > > +
> > > > + if (WARN_ON(!encoder->dev))
> > > > + return;
> > > > +
> > > > + drm_encoder_cleanup(encoder);
> > > > +}
> > > > +
> > > > +void *__drmm_encoder_alloc(struct drm_device *dev, size_t size, size_t offset,
> > > > + const struct drm_encoder_funcs *funcs,
> > > > + int encoder_type, const char *name, ...)
> > > > +{
> > > > + void *container;
> > > > + struct drm_encoder *encoder;
> > > > + va_list ap;
> > > > + int ret;
> > > > +
> > > > + if (WARN_ON(!funcs || funcs->destroy))
> > > > + return ERR_PTR(-EINVAL);
> > > > +
> > > > + container = drmm_kzalloc(dev, size, GFP_KERNEL);
> > > > + if (!container)
> > > > + return ERR_PTR(-EINVAL);
> > > > +
> > > > + encoder = container + offset;
> > > > +
> > > > + va_start(ap, name);
> > > > + ret = __drm_encoder_init(dev, encoder, funcs, encoder_type, name, ap);
> > > > + va_end(ap);
> > > > + if (ret)
> > > > + return ERR_PTR(ret);
> > > > +
> > > > + ret = drmm_add_action_or_reset(dev, drmm_encoder_alloc_release, encoder);
> > > > + if (ret)
> > > > + return ERR_PTR(ret);
> > >
> > > Given that drmm_encoder_alloc_release() will be called right before the
> > > kfree related to the above drmm_kzalloc(), wouldn't it be more efficient
> > > to replace drmm_kzalloc() with kzalloc() and add a kfree() in
> > > drmm_encoder_alloc_release() ? That will save one context allocation.
> >
> > That is not quite as trivial: drmm_encoder_alloc_release() doesn't know
> > the container pointer that must be passed to kfree(), nor the offset
> > between container and encoder.
>
> Indeed :-S Adding more context to the drmres when creating it with
> drmm_add_action_or_reset() would solve this, but it's probably overkill
> (although I think this would definitely be useful if/when we turn the
> DRM resource manager to a more generic component usable by other
> subsystems).
Internally it's even funnier, because the drmm_kzalloc doesn't even have a
callback - the allocation is freed as part of the drmm tracking structure.
Which means that drmm_kzalloc has a "free" release callback already.
This irked me to no end, but I couldn't come up with an interface to stuff
the release callback into the allocation which I deemed safe enough for
drmm, which is supposed to make life easy and bugs disappear, not greate
new surprises.
What I just realized is that drmm_add_action_or_reset could look at the
drmm cleanup action stack, and if the topmost one doesn't have a hook set
already, use that one instead of allocating a new tracker structure. This
way ordering is always guaranteed (e.g. if you mix up drmm_kzalloc calls
with add_action for different objects in strange orders) and there's no
problem ever, we just avoid a few small allocations.
Still needs a few special cases, so not sure it's actually a win.
-Daniel
>
> > > With this addressed,
> > >
> > > Reviewed-by: Laurent Pinchart <laurent.pinchart@xxxxxxxxxxxxxxxx>
>
> You can add my tag to this patch and the other ones in the series where
> I've made the same comment.
>
> --
> Regards,
>
> Laurent Pinchart
> _______________________________________________
> dri-devel mailing list
> dri-devel@xxxxxxxxxxxxxxxxxxxxx
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/dri-devel
