On Thu, May 30, 2024 at 03:27:58PM +0200, Arnd Bergmann wrote:
> On Thu, May 30, 2024, at 14:26, Russell King (Oracle) wrote:
> > On Thu, May 30, 2024 at 05:50:52PM +0530, Naresh Kamboju wrote:
> >> The arm builds failed on Linux next with gcc-13 and clang-18.
> >>
> >> Config: arm imx_v6_v7_defconfig - failed
> >>
> >> Reported-by: Linux Kernel Functional Testing <lkft@xxxxxxxxxx>
> >>
> >> Build log:
> >> -----
> >> arm-linux-gnueabihf-ld: kernel/rcu/update.o: in function
> >> `rcu_trc_cmpxchg_need_qs':
> >> update.c:(.text+0x348): undefined reference to `__bad_cmpxchg'
> >> arm-linux-gnueabihf-ld: kernel/rcu/update.o: in function
> >> `rcu_read_unlock_trace_special':
> >> update.c:(.text+0x41c): undefined reference to `__bad_cmpxchg'
> >> arm-linux-gnueabihf-ld: kernel/rcu/update.o: in function
> >> `trc_read_check_handler':
> >> update.c:(.text+0x4b0): undefined reference to `__bad_cmpxchg'
> >> arm-linux-gnueabihf-ld: kernel/rcu/update.o: in function `trc_inspect_reader':
> >> update.c:(.text+0x1518): undefined reference to `__bad_cmpxchg'
> >> arm-linux-gnueabihf-ld: update.c:(.text+0x1548): undefined reference
> >> to `__bad_cmpxchg'
> >> arm-linux-gnueabihf-ld: kernel/rcu/update.o:update.c:(.text+0x1cc4):
> >> more undefined references to `__bad_cmpxchg' follow
> >> make[3]: *** [/builds/linux/scripts/Makefile.vmlinux:34: vmlinux] Error 1
> >
> > You get a reference to __bad_cmpxchg() when someone uses cmpxchg() on
> > a datatype that there are no CPU instructions to perform the atomic
> > compare-and-exchange.
> >
> > As we support 8-bit, 16-bit and 32-bit, I would expect that we're now
> > seeing cmpxchg() being used on 64-bit values, which are unsupportable
> > on 32-bit Arm.
>
> A 64-bit cmpxchg() would fail on most 32-bit architectures,
> other than armv6k+ and i686+. Since this one fails on
> an armv6 (non-6k) build, it's probably the added 8-bit
> cmpxchg() that Paul talked about using for RCU, though I
> don't see that in linux-next here.
>
> I htink we need to use the emulated cmpxchg8() in
> __cmpxchg() for this case, just like we do on other
> architectures that only have word size atomics.
Thank you all for your testing and analysis!
I will fix this, and also pull the 8046f3005015 ("rcu-tasks: Remove
open-coded one-byte cmpxchg() emulation") commit out of my -next pile
in the meantime.
Thanx, Paul
