RE: pselect/etc semantics (Was: [PATCH v2] signal: Adjust error codes according to restore_user_sigmask())

From: David Laight
Date: Wed May 29 2019 - 12:58:14 EST


From: Oleg Nesterov
> Sent: 29 May 2019 17:12
> Al, Linus, Eric, please help.
>
> The previous discussion was very confusing, we simply can not understand each
> other.
>
> To me everything looks very simple and clear, but perhaps I missed something
> obvious? Please correct me.
>
> I think that the following code is correct
>
> int interrupted = 0;
>
> void sigint_handler(int sig)
> {
> interrupted = 1;
> }
>
> int main(void)
> {
> sigset_t sigint, empty;
>
> sigemptyset(&sigint);
> sigaddset(&sigint, SIGINT);
> sigprocmask(SIG_BLOCK, &sigint, NULL);
>
> signal(SIGINT, sigint_handler);
>
> sigemptyset(&empty); // so pselect() unblocks SIGINT
>
> ret = pselect(..., &empty);
^^^^^ sigint
>
> if (ret >= 0) // sucess or timeout
> assert(!interrupted);
>
> if (interrupted)
> assert(ret == -EINTR);
> }
>
> IOW, if pselect(sigmask) temporary unblocks SIGINT according to sigmask, this
> signal should not be delivered if a ready fd was found or timeout. The signal
> handle should only run if ret == -EINTR.

Personally I think that is wrong.
Given code like the above that has:
while (!interrupted) {
pselect(..., &sigint);
// process available data
}

You want the signal handler to be executed even if one of the fds
always has available data.
Otherwise you can't interrupt a process that is always busy.

One option is to return -EINTR if a signal is pending when the mask
is updated - before even looking at anything else.
Signals that happen later on (eg after a timeout) need not be reported
(until the next time around the loop).

> (pselect() can be interrupted by any other signal which has a handler. In this
> case the handler can be called even if ret >= 0. This is correct, I fail to
> understand why some people think this is wrong, and in any case we simply can't
> avoid this).

You mean any signal that isn't blocked when pselect() is called....

> This was true until 854a6ed56839a ("signal: Add restore_user_sigmask()"),
> now this is broken by the signal_pending() check in restore_user_sigmask().
>
> This patch https://lore.kernel.org/lkml/20190522032144.10995-1-deepa.kernel@xxxxxxxxx/
> turns 0 into -EINTR if signal_pending(), but I think we should simply restore
> the old behaviour and simplify the code.
>
> See the compile-tested patch at the end. Of course, the new _xxx() helpers
> should be renamed somehow. fs/aio.c doesn't look right with or without this
> patch, but iiuc this is what it did before 854a6ed56839a.
>
> Let me show the code with the patch applied. I am using epoll_pwait() as an
> example because it looks very simple.
>
>
> static inline void set_restore_sigmask(void)
> {
> // WARN_ON(!TIF_SIGPENDING) was removed by this patch
> current->restore_sigmask = true;
> }
>
> int set_xxx(const sigset_t __user *umask, size_t sigsetsize)
> {
> sigset_t *kmask;
^ no '*' here, add & before uses.
>
> if (!umask)
> return 0;
> if (sigsetsize != sizeof(sigset_t))
> return -EINVAL;
> if (copy_from_user(kmask, umask, sizeof(sigset_t)))
> return -EFAULT;
>
> // we can safely modify ->saved_sigmask/restore_sigmask, they has no meaning
> // until the syscall returns.
> set_restore_sigmask();
> current->saved_sigmask = current->blocked;
> set_current_blocked(kmask);
>
> return 0;
> }
>
>
> void update_xxx(bool interrupted)
> {
> // the main reason for this helper is WARN_ON(!TIF_SIGPENDING) which was "moved"
> // from set_restore_sigmask() above.
> if (interrupted)
> WARN_ON(!test_thread_flag(TIF_SIGPENDING));
> else
> restore_saved_sigmask();
> }

I looked at the code earlier, but failed to find the code that actually
delivers the signals.
It may be 'racy' with update_xxx() regardless of whether that is
looking for -EINTR or just a pending signal.

I assume that TIF_SIGPENGING is used to (not) short-circuit the
system call return path so that signals get delivered.
So that it is important that update_xxx() calls restore_saved_sigmask()
if there is no signal pending.
(Although a signal can happen after the test - which can/will be ignored
until the signal is enabled again.)

restore_saved_sigmask() must itself be able to set TIF_SIGPENDING
(the inner sigmask could be more restrictive!).

If restore_saved_sigmask() isn't called here, the syscall return
path must do it after calling all the handlers and after clearing
TIF_SIGPENDING, and then call unmasked handlers again.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)