On Wed, Jun 10, 2015 at 04:58:04PM +0200, Ralf Baechle wrote:
> On Mon, Jun 08, 2015 at 03:14:43PM +0200, Xose Vazquez Perez wrote:
>
> > If there is anything wrong, please report it in this thread:
> > https://marc.info/?t=143332955700003
>
>
> > locking/ cmpxchg-local : TODO | HAVE_CMPXCHG_LOCAL # arch supports the this_cpu_cmpxchg() API
>
> This one was easy - we have the functions in the code just no "select
> HAVE_CMPXCHG_LOCAL" Kconfig.
Something's wrong there. The new file
Documentation/features/locking/cmpxchg-local/arch-support.txt in linux-next
claims correctly that only s390 and x86 define HAVE_CMPXCHG_LOCAL. And a
git grep -w cmpxchg_local finds that in addition to these alpha, arm, arm64,
avr32, blackfin, c6x, frv, ia64, m32r, m68k, mips, parisc, powerpc, sparc,
unicore32 and xtensa define cmpxchg_local.
These architectures seem to not define cmpxchg_local in their arch/ dir:
arc cris hexagon metag microblaze mn10300 nios2 openrisc score sh tile um
Microblaze and nios2 include <asm-generic/cmpxchg.h> into their arch
cmpxchg.h so they get a definition of these functions but don't define
HAVE_CMPXCHG_LOCAL. Peter Zijlstra said it the local versions are ~ 20
cycles faster on x86 than the "global" version. But I've found one user
of cmpxchg_local, mm/vmstat.c and one user of cmpxchg64_local,
drivers/iommu/intel-iommu.c. Sure, fixing the issue was trivial for me
on MIPS but is having cmpxchg{,64}_local actually worth it?
Cheers,
Ralf
On Wed, 2015-06-10 at 17:57 +0200, Ralf Baechle wrote:
> On Wed, Jun 10, 2015 at 04:58:04PM +0200, Ralf Baechle wrote:
>
> > On Mon, Jun 08, 2015 at 03:14:43PM +0200, Xose Vazquez Perez wrote:
> >
> > > If there is anything wrong, please report it in this thread:
> > > https://marc.info/?t=143332955700003
> >
> >
> > > locking/ cmpxchg-local : TODO | HAVE_CMPXCHG_LOCAL # arch supports the this_cpu_cmpxchg() API
> >
> > This one was easy - we have the functions in the code just no "select
> > HAVE_CMPXCHG_LOCAL" Kconfig.
>
> Something's wrong there. The new file
> Documentation/features/locking/cmpxchg-local/arch-support.txt in linux-next
> claims correctly that only s390 and x86 define HAVE_CMPXCHG_LOCAL. And a
> git grep -w cmpxchg_local finds that in addition to these alpha, arm, arm64,
> avr32, blackfin, c6x, frv, ia64, m32r, m68k, mips, parisc, powerpc, sparc,
> unicore32 and xtensa define cmpxchg_local.
>
> These architectures seem to not define cmpxchg_local in their arch/ dir:
>
> arc cris hexagon metag microblaze mn10300 nios2 openrisc score sh tile um
>
> Microblaze and nios2 include <asm-generic/cmpxchg.h> into their arch
> cmpxchg.h so they get a definition of these functions but don't define
> HAVE_CMPXCHG_LOCAL. Peter Zijlstra said it the local versions are ~ 20
> cycles faster on x86 than the "global" version. But I've found one user
> of cmpxchg_local, mm/vmstat.c and one user of cmpxchg64_local,
> drivers/iommu/intel-iommu.c. Sure, fixing the issue was trivial for me
> on MIPS but is having cmpxchg{,64}_local actually worth it?
If you traverse the obfuscation chain a little, you'll find that
local_cmpxchg() is implemented using cmpxchg_local(), and there are a
few more users of that.