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

[Arnd Bergmann <arnd@arndb.de>]

On Wed, May 29, 2019 at 6:12 PM Oleg Nesterov <oleg@redhat.com> wrote:
>
> 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.

Thanks for the elaborate explanation in this patch, it all starts making sense
to me now. I also looked at your patch in detail and thought I had found
a few mistakes at first but those all turned out to be mistakes in my reading.

> 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.

I think this is a nice simplification, but it would help not to mix up the
minimal regression fix with the rewrite of those functions. For the stable
kernels, I think we want just the addition of the 'bool interrupted' argument
to restore_user_sigmask() to close the race that was introduced
854a6ed56839a. Following up on that for future kernels, your patch
improves the readability, but we can probably take it even further.

> -       ret = set_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
> +       ret = set_xxx(ksig.sigmask, ksig.sigsetsize);
>         if (ret)
>                 return ret;
>
>         ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
> -       restore_user_sigmask(ksig.sigmask, &sigsaved);
> -       if (signal_pending(current) && !ret)
> +
> +       interrupted = signal_pending(current);
> +       update_xxx(interrupted);

Maybe name this

           restore_saved_sigmask_if(!interrupted);

and make restore_saved_sigmask_if() an inline function
next to restore_saved_sigmask()?

> @@ -2201,13 +2205,15 @@ COMPAT_SYSCALL_DEFINE6(io_pgetevents,
>         if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
>                 return -EFAULT;
>
> -       ret = set_compat_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
> +       ret = set_compat_xxx(ksig.sigmask, ksig.sigsetsize);
>         if (ret)
>                 return ret;

With some of the recent discussions about compat syscall handling,
I now think that we want to just fold set_compat_user_sigmask()
into set_user_sigmask() (whatever they get called in the end)
with an in_compat_syscall() conditional inside it, and completely get
rid of the COMPAT_SYSCALL_DEFINEx() definitions for those
system calls for which this is the only difference.

Unfortunately we still need the time32/time64 distinction, but removing
syscall handlers is a significant cleanup here already, and we can
move most of the function body of sys_io_pgetevents() into
do_io_getevents() in the process. Same for some of the other calls.

Not sure about the order of the cleanups, but probably something like
this would work:

1. fix the race (to be backported)
2. unify set_compat_user_sigmask/set_user_sigmask
3. remove unneeded compat handlers
4. replace restore_user_sigmask with restore_saved_sigmask_if()
5. also unify compat_get_fd_set()/get_fd_set() and kill off
    compat select() variants.

       Arnd