On Wed, 23 Oct 2019 at 18:24, Oleg Nesterov <oleg@xxxxxxxxxx> wrote: > > On 10/22, Marco Elver wrote: > > > > On Tue, 22 Oct 2019 at 17:49, Oleg Nesterov <oleg@xxxxxxxxxx> wrote: > > > > > > Just for example. Suppose that task->state = TASK_UNINTERRUPTIBLE, this task > > > does __set_current_state(TASK_RUNNING), another CPU does wake_up_process(task) > > > which does the same UNINTERRUPTIBLE -> RUNNING transition. > > > > > > Looks like, this is the "data race" according to kcsan? > > > > Yes, they are "data races". They are probably not "race conditions" though. > > > > This is a fair distinction to make, and we never claimed to find "race > > conditions" only > > I see, thanks, just wanted to be sure... > > > KCSAN's goal is to find *data races* according to the LKMM. Some data > > races are race conditions (usually the more interesting bugs) -- but > > not *all* data races are race conditions. Those are what are usually > > referred to as "benign", but they can still become bugs on the wrong > > arch/compiler combination. Hence, the need to annotate these accesses > > with READ_ONCE, WRITE_ONCE or use atomic_t: > > Well, if I see READ_ONCE() in the code I want to understand why it was > used. Is it really needed for correctness or we want to shut up kcsan? > Say, why should wait_event(wq, *ptr) use READ_ONCE()? Nevermind, please > forget. > > Btw, why __kcsan_check_watchpoint() does user_access_save() before > try_consume_watchpoint() ? Instrumentation is added in UACCESS regions. Since we do not access user-memory, we do user_access_save to ensure everything is safe (otherwise objtool complains that we do calls to non-whitelisted functions). I will try to optimize this a bit, but we can't avoid it. > Oleg. >
