Re: [PATCH v2 0/3] sysfs: constify bin_attribute argument of sysfs_bin_attr_simple_read()

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Wed, Jan 08, 2025 at 09:45:45AM -0800, Alexei Starovoitov wrote:
> On Tue, Dec 31, 2024 at 2:30 AM Thomas Weißschuh <linux@weissschuh.net> wrote:
> >
> > On 2024-12-30 16:50:41-0800, Alexei Starovoitov wrote:
> > > On Sat, Dec 28, 2024 at 12:43 AM Thomas Weißschuh <linux@weissschuh.net> wrote:
> > > >
> > > > Most users use this function through the BIN_ATTR_SIMPLE* macros,
> > > > they can handle the switch transparently.
> > > >
> > > > This series is meant to be merged through the driver core tree.
> > >
> > > hmm. why?
> >
> > Patch 1 changes the signature of sysfs_bin_attr_simple_read().
> > Before patch 1 sysfs_bin_attr_simple_read() needs to be assigned to the
> > callback member .read, after patch 1 it's .read_new.
> > (Both callbacks work exactly the same, except for their signature,
> > .read_new is only a transition mechanism and will go away again)
> >
> > > I'd rather take patches 2 and 3 into bpf-next to avoid
> > > potential conflicts.
> > > Patch 1 looks orthogonal and independent.
> >
> > If you pick up 2 and 3 through bpf-next you would need to adapt these
> > assignments. As soon as both patch 1 and the modified 2 and 3 hit
> > Linus' tree, the build would break due to mismatches function pointers.
> > (Casting function pointers to avoid the mismatch will blow up with KCFI)
> 
> I see. All these steps to constify is frankly a mess.
> You're wasting cpu and memory for this read vs read_new
> when const is not much more than syntactic sugar in C.
> You should have done one tree wide patch without doing this _new() hack.
> 
> Anyway, rant over. Carry patches 2,3. Hopefully they won't conflict.
> But I don't want to see any constification patches in bpf land
> that come with such pointless runtime penalty.

The "pointless" penalty will go away once we convert all instances, and
really, it's just one pointer check, sysfs files should NOT be a hot
path for anything real, and one more pointer check should be cached and
not measurable compared to the real logic behind the binary data coming
from the hardware/kernel, right?

sysfs is NOT tuned for speed at all, so adding more checks like this
should be fine.

thanks,

greg k-h