signal: Make siginmask safe when passed a signal of 0

Eric Biggers reported:
> The following commit, which went into v4.20, introduced undefined behavior when
> sys_rt_sigqueueinfo() is called with sig=0:
>
> commit 4ce5f9c9
> Author: Eric W. Biederman <ebiederm@xmission.com>
> Date:   Tue Sep 25 12:59:31 2018 +0200
>
>     signal: Use a smaller struct siginfo in the kernel
>
> In sig_specific_sicodes(), used from known_siginfo_layout(), the expression
> '1ULL << ((sig)-1)' is undefined as it evaluates to 1ULL << 4294967295.
>
> Reproducer:
>
> #include <signal.h>
> #include <sys/syscall.h>
> #include <unistd.h>
>
> int main(void)
> {
> 	siginfo_t si = { .si_code = 1 };
> 	syscall(__NR_rt_sigqueueinfo, 0, 0, &si);
> }
>
> UBSAN report for v5.0-rc1:
>
> UBSAN: Undefined behaviour in kernel/signal.c:2946:7
> shift exponent 4294967295 is too large for 64-bit type 'long unsigned int'
> CPU: 2 PID: 346 Comm: syz_signal Not tainted 5.0.0-rc1 #25
> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
> Call Trace:
>  __dump_stack lib/dump_stack.c:77 [inline]
>  dump_stack+0x70/0xa5 lib/dump_stack.c:113
>  ubsan_epilogue+0xd/0x40 lib/ubsan.c:159
>  __ubsan_handle_shift_out_of_bounds+0x12c/0x170 lib/ubsan.c:425
>  known_siginfo_layout+0xae/0xe0 kernel/signal.c:2946
>  post_copy_siginfo_from_user kernel/signal.c:3009 [inline]
>  __copy_siginfo_from_user+0x35/0x60 kernel/signal.c:3035
>  __do_sys_rt_sigqueueinfo kernel/signal.c:3553 [inline]
>  __se_sys_rt_sigqueueinfo kernel/signal.c:3549 [inline]
>  __x64_sys_rt_sigqueueinfo+0x31/0x70 kernel/signal.c:3549
>  do_syscall_64+0x4c/0x1b0 arch/x86/entry/common.c:290
>  entry_SYSCALL_64_after_hwframe+0x49/0xbe
> RIP: 0033:0x433639
> Code: c4 18 c3 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 40 00 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 0f 83 7b 27 00 00 c3 66 2e 0f 1f 84 00 00 00 00
> RSP: 002b:00007fffcb289fc8 EFLAGS: 00000246 ORIG_RAX: 0000000000000081
> RAX: ffffffffffffffda RBX: 00000000004002e0 RCX: 0000000000433639
> RDX: 00007fffcb289fd0 RSI: 0000000000000000 RDI: 0000000000000000
> RBP: 00000000006b2018 R08: 000000000000004d R09: 0000000000000000
> R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000401560
> R13: 00000000004015f0 R14: 0000000000000000 R15: 0000000000000000

I have looked at the other callers of siginmask and they all appear to
in locations where sig can not be zero.

I have looked at the code generation of adding an extra test against
zero and gcc was able with a simple decrement instruction to combine
the two tests together. So the at most adding this test cost a single
cpu cycle.  In practice that decrement instruction was already present
as part of the mask comparison, so the only change was when the
instruction was executed.

So given that it is cheap, and obviously correct to update siginmask
to verify the signal is not zero.  Fix this issue there to avoid any
future problems.
Reported-by: Eric Biggers <ebiggers@kernel.org>
Fixes: 4ce5f9c9 ("signal: Use a smaller struct siginfo in the kernel")
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>

include/linux/signal.h

@@ -392,7 +392,7 @@ extern bool unhandled_signal(struct task_struct *tsk, int sig);
 #endif
 
 #define siginmask(sig, mask) \
-	((sig) < SIGRTMIN && (rt_sigmask(sig) & (mask)))
+	((sig) > 0 && (sig) < SIGRTMIN && (rt_sigmask(sig) & (mask)))
 
 #define SIG_KERNEL_ONLY_MASK (\
 	rt_sigmask(SIGKILL)   |  rt_sigmask(SIGSTOP))