On Thu, Jun 25, 2020 at 04:11PM +0200, Peter Zijlstra wrote: > On Thu, Jun 25, 2020 at 01:29:26PM +0200, Peter Zijlstra wrote: > > I fear the same. Let me see if I can quickly modify the atomic scripts > > to generate the required fallbacks. > > Something like so ought to work, I suppose. > > --- > Subject: locking/atomics: Provide the arch_atomic_ interface to generic code > From: Peter Zijlstra <pet...@infradead.org> > Date: Thu Jun 25 15:55:14 CEST 2020 > > Architectures with instrumented (KASAN/KCSAN) atomic operations > natively provide arch_atomic_ variants that are not instrumented. > > It turns out that some generic code also requires arch_atomic_ in > order to avoid instrumentation, so provide the arch_atomic_ interface > as a direct map into the regular atomic_ interface for > non-instrumented architectures. > > Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org> > --- > include/linux/atomic-fallback.h | 236 > +++++++++++++++++++++++++++++++++- > scripts/atomic/gen-atomic-fallback.sh | 31 ++++ > 2 files changed, 266 insertions(+), 1 deletion(-)
Thanks, looks reasonable! If noinstr becomes important on architectures that don't implement atomics using arch_ themselves, there might be a problem with CONFIG_TRACE_BRANCH_PROFILING, because unlikely() is used throughout this file. Probably not something to worry about now. Thanks, -- Marco
