On 10/8/20 7:24 AM, Oleg Nesterov wrote: > On 10/05, Jens Axboe wrote: >> >> @@ -4447,7 +4447,7 @@ SYSCALL_DEFINE0(pause) >> __set_current_state(TASK_INTERRUPTIBLE); >> schedule(); >> } >> - return -ERESTARTNOHAND; >> + return task_sigpending(current) ? -ERESTARTNOHAND : -ERESTARTSYS; >> } >> >> #endif >> @@ -4462,7 +4462,7 @@ static int sigsuspend(sigset_t *set) >> schedule(); >> } >> set_restore_sigmask(); >> - return -ERESTARTNOHAND; >> + return task_sigpending(current) ? -ERESTARTNOHAND : -ERESTARTSYS; >> } > > Both changes are equally wrong. Why do you think sigsuspend() should ever > return -ERESTARTSYS ? > > If get_signal() deques a signal, handle_signal() will restart this syscall > if ERESTARTSYS, this is wrong.
The intent was that if we get woken up and signal_pending() is true, then we want to restart it if we're just doing TIF_SIGNAL_NOTIFY. But I guess it can't be 100% reliable, even if TIF_SIGPENDING isn't set at this point, but it is by the time a signal is attempted dequeued. I'll drop these too, thanks Oleg. -- Jens Axboe
