On Tue, Dec 31, 2024 at 2:30 AM Thomas Weißschuh <linux@xxxxxxxxxxxxxx> wrote: > > On 2024-12-30 16:50:41-0800, Alexei Starovoitov wrote: > > On Sat, Dec 28, 2024 at 12:43 AM Thomas Weißschuh <linux@xxxxxxxxxxxxxx> 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.
