Hi Peter,
Le 31/08/2021 à 16:00, Peter Zijlstra a écrit :
On Tue, Aug 31, 2021 at 01:12:26PM +0000, Christophe Leroy wrote:
Yes, this should work nicely!
Since you have the two nop's at the end, you could frob in an
optimization for __static_call_return0 without too much issue.
Replace the two nops with (excuse my ppc asm):
li r3, 0
blr
and augment arch_static_call_transform() with something like:
if (func == &__static_call_return0)
err = patch_branch(tramp, tramp+24, 0);
I just discovered that we likely have an issue with the implementation
of that RET0 static call.
Looking at System.map I have:
c0004fc0 t __static_call_return0
c0011518 t __static_call_return0
c00d8160 t __static_call_return0
So when we do:
if (func == &__static_call_return0)
It is unlikely that we'll get the expected one.
I see __static_call_return0 is defined as 'static inline' in
include/linux/static_call.h
Any reason for not having it as a single global symbol instead ?
Thanks
Christophe
