From: Christophe Leroy > Sent: 16 September 2021 08:24 > > Le 16/09/2021 à 09:16, Benjamin Herrenschmidt a écrit : > > On Thu, 2021-09-16 at 17:15 +1000, Benjamin Herrenschmidt wrote: > >> On Wed, 2021-09-15 at 16:31 +0200, Christophe Leroy wrote: > >>> dcbz instruction shouldn't be used on non-cached memory. Using > >>> it on non-cached memory can result in alignment exception and > >>> implies a heavy handling. > >>> > >>> Instead of silentely emulating the instruction and resulting in > >>> high > >>> performance degradation, warn whenever an alignment exception is > >>> taken due to dcbz, so that the user is made aware that dcbz > >>> instruction has been used unexpectedly. > >>> > >>> Reported-by: Stan Johnson <user...@yahoo.com> > >>> Cc: Finn Thain <fth...@linux-m68k.org> > >>> Signed-off-by: Christophe Leroy <christophe.le...@csgroup.eu> > >>> --- > >>> arch/powerpc/kernel/align.c | 1 + > >>> 1 file changed, 1 insertion(+) > >>> > >>> diff --git a/arch/powerpc/kernel/align.c > >>> b/arch/powerpc/kernel/align.c > >>> index bbb4181621dd..adc3a4a9c6e4 100644 > >>> --- a/arch/powerpc/kernel/align.c > >>> +++ b/arch/powerpc/kernel/align.c > >>> @@ -349,6 +349,7 @@ int fix_alignment(struct pt_regs *regs) > >>> if (op.type != CACHEOP + DCBZ) > >>> return -EINVAL; > >>> PPC_WARN_ALIGNMENT(dcbz, regs); > >>> + WARN_ON_ONCE(1); > >> > >> This is heavy handed ... It will be treated as an oops by various > >> things uselessly spit out a kernel backtrace. Isn't > >> PPC_WARN_ALIGNMENT > >> enough ? > > > PPC_WARN_ALIGNMENT() only warns if explicitely activated, I want to > catch uses on 'dcbz' on non-cached memory all the time as they are most > often the result of using memset() instead of memset_io(). > > > > > Ah I saw your other one about fbdev... Ok what about you do that in a > > if (!user_mode(regs)) ? > > Yes I can do WARN_ON_ONCE(!user_mode(regs)); instead. > > > Indeed the kernel should not do that. > > Does userspace accesses non-cached memory directly ?
It probably can if a driver mmaps PCI space directly into user space. That certainly works on x86-64. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)