[Intel-gfx] [PATCH v4 05/13] drivers: create binary sysfs for class

[Greg Kroah-Hartman]

On Mon, Apr 15, 2019 at 09:14:12PM +0200, Daniel Vetter wrote:
> On Mon, Apr 15, 2019 at 8:01 PM Greg Kroah-Hartman
> <gregkh at linuxfoundation.org> wrote:
> > On Mon, Apr 15, 2019 at 10:44:12PM +0530, Ramalingam C wrote:
> > > On 2019-04-15 at 16:47:16 +0200, Greg Kroah-Hartman wrote:
> > > > On Mon, Apr 15, 2019 at 06:11:13PM +0530, Ramalingam C wrote:
> > > > > On 2019-04-05 at 14:32:00 +0200, Greg Kroah-Hartman wrote:
> > > > > > On Fri, Apr 05, 2019 at 04:06:22PM +0530, Ramalingam C wrote:
> > > > > > > On 2019-04-05 at 11:23:00 +0200, Greg Kroah-Hartman wrote:
> > > > > > > > On Fri, Apr 05, 2019 at 02:12:54PM +0530, Ramalingam C wrote:
> > > > > > > > > Functions to create and remove the binary sysfs for class are added.
> > > > > > > > >
> > > > > > > > > These are getting introduced as DRM wants to create the common binary
> > > > > > > > > sysfs across the drm subsystem to handle hdcp srm.
> > > > > > > >
> > > > > > > > Why do you need individual files?  That's almost always a sign that you
> > > > > > > > are going to race with userspace in a bad way.  Why not just use an
> > > > > > > > attribute group which provides automatic support for this?
> > > > > > > Greg,
> > > > > > >
> > > > > > > Reason behind this move is to have a common srm entry path across all drm
> > > > > > > drivers. And the data fed into this is binary blob. So I am creating a
> > > > > > > binary sysfs "hdcp_srm" at /sys/class/drm/
> > > > > >
> > > > > > Ah, you want to have a file in your class directory, not your class
> > > > > > device directory.
> > > > > >
> > > > > > No, please do not do that.  There's a reason I got rid of those same
> > > > > > types of apis in the past.
> > > > > >
> > > > > > And "binary blobs" are horrid anyway, they are only to be used as a
> > > > > > pass-through to the device itself, from the kernel, no touching the data
> > > > > > at all.  If you really need/want this, then put it in the device's
> > > > > > directory as that is where the data is going to, not the kernel "class"
> > > > > > code as it sure as heck better not be doing anything with it.
> > > > > Greg,
> > > > >
> > > > > But here the parsing of the received binary blob is done outside the drm
> > > > > device/cards. This will be generic code for all drm cardx(drivers). And
> > > > > this will provide the service helper functions to the drm drivers for HDCP SRM checking.
> > > >
> > > > Again, the kernel is NOT to be parsing any binary data that comes
> > > > through a sysfs file.  If you need such a crazy thing, do it through
> > > > your normal drm ioctl.
> > > >
> > > > > So we prefer to have the binary sysfs at /sys/class/drm/ than inside the
> > > > > cardx folders, so that it will be generic. Could you please suggest a way to achieve that?
> > > >
> > > > What is a "cardx" driver?
> > > Meant card0 card1 etc.. Drm drivers.
> > > > Why can you not do it in a device-specific
> > > > location?  Are suddenly _ALL_ DRM drivers going to need this
> > > > information?  What is the use case?  Who is going to be providing this
> > > > blob and where is it going?  What in the kernel uses it?  What on the
> > > > hardware uses it?  What is it actually?
> > >
> > > This is for HDCP usecase. List of compromised receivers' ID  called system
> > > renewability Message (SRM) will be available at userspace which needs to
> > > be passed to be kernel. And this data can be parsed at kernel used
> > > across all DRM drivers to implement the HDCP authentication. Hence we
> > > want generic code to parse the SRM data and provide a helper function to
> > > all DRM drivers to validate their HDCP sink's ID.
> > >
> > > To achieve this we want to keep the sysfs and parsing logic outside the
> > > drm drivers at the class level.
> > >
> > > At present I915 will be using these implementation. in future other DRM
> > > drivers can use the same.
> >
> > Again, binary sysfs files are not for any data that the kernel has to
> > parse at all.  So do not use sysfs for this.
> 
> So what's the recommend thing then?
> - requrest_firmware blob could work, aside that the blob might change
> (the point of this to allow updates). For some hw you can even stuff
> the SRM into some eprom iirc, so not entirely misfit. Plus kernel
> parses it, so not sure request_firmware is really the right thing.

I think it is, as this really looks like "firmware" to me as you are
wanting to do something with the buffer and pass it on.

> - configfs? I never used that one, but the transactional config stuff
> this provides feels like silly amounts of overkill

Maybe, but why do you want to stuf binary data in your kernel in the
first place?  It is configuration, right, why not have it in a format
that you can easily handle it in?  Or is that the binary format you
have?

> - procfs :-)

Hah, no.

> - ioctl it definitely isn't, there's no need to open a drm device node
> first and require a special binary to upload the srm if cat >
> $kernfs/srm is all that's needed

But you want this associtated with a specific drm device, right?  So
what's wrong with this?

> - drmfs ... that's some serious amounts of overkill, so nope.

300 lines of kernel code for a fs isn't that much :)

> - per-device sysfs binary blob, silly because it's not per-device, but
> wouldn't need your ack since we could sneak this one in :-) For
> reasonable semantics we could just alias them all to the same storage,
> so that userspace can do a cat > /sys/class/drm/card0/srm instead of
> /sys/class/drm/srm for same effects (for multi gpu systems).

Again, no, you do not use binary sysfs files for data you want to touch.
Yes, you can try to get away with it, but please do not.

thanks,

greg k-h