Romain Perier <romain.perier@xxxxxxxxxxxxxxxxxx> wrote:
> The futex subsystem relies on futex_atomic_cmpxchg_inatomic() to
> perform an atomic cmpxchg. Most architectures provide their own
> implementation. For those who do not, <asm-generic/futex.h> provides a
> generic software implementations that works on !SMP platforms.
>
> To provide an atomicity guarantee on !SMP plaforms, the generic
> implementation simply disables preemption.
> However it is not properly re-enabled from conditional branches.
>
> The observed consequence of this bug is that the preemption counter is
> not well balanced. It can be seen at boot time :
>
> [ 8.148000] futex hash table entries: 256 (order: -1, 3072 bytes)
> [ 8.156000] ------------[ cut here ]------------
> [ 8.164000] WARNING: CPU: 0 PID: 1 at init/main.c:806 do_one_initcall+0x1ac/0x1d0()
> [ 8.172000] initcall futex_init+0x0/0xd8 returned with preemption imbalance
> [ 8.180000] Modules linked in:
> [ 8.184000] CPU: 0 PID: 1 Comm: swapper Not tainted 4.5.0 #72
> [ 8.192000] Stack from d1c1deb8:
> [ 8.196000]
> [ 8.196000] d1c1debc
> [ 8.204000] d1c1defc 00000009
> [ 8.204000] d04d8b90 00000326
> [ 8.208000] d0000380 d0136484
> [ 8.212000] d00093b4
>
> [ 8.216000] 00000326 d0000380
> [ 8.220000] d1c1c000 00000000
> [ 8.224000] d026c64c 00000000
> [ 8.228000] d0515004 d0009410
> [ 8.232000]
> [ 8.232000] d1c1defc
> [ 8.236000] d04d8d8c d1c1df08
> [ 8.240000] d0000380 d026c64c
> [ 8.244000] d1c1df10 65657270
> [ 8.248000] 6974706d
>
> [ 8.252000] 69206e6f 6c61626d
> [ 8.256000] 65636e61 d0780020
> [ 8.260000] d05108a0 00000001
> [ 8.264000] d00225c8 00000000
> [ 8.268000]
> [ 8.268000] d05103c8
> [ 8.272000] 0000003f 00060006
> [ 8.276000] d05108a0 00000006
> [ 8.280000] 0000003f 00000007
> [ 8.284000] 0000003f
>
> [ 8.288000] d052501c d0525024
> [ 8.288000] d027d480 d0266c78
> [ 8.292000] 00000006 00000006
> [ 8.296000] 00000000 d026627c
> [ 8.300000]
> [ 8.300000] Call Trace:
>
> [ 8.308000] [<d025fc58>] [<d025fc68>]
> [ 8.312000] [<d025fc58>] [<d000130c>]
> [ 8.316000]
> [ 8.316000]
>
> ---[ end trace 2d3fe47ac46f91f6 ]---
>
> Before commit 8222dbe21e79
> ("sched/preempt, mm/fault: Decouple preemption from the page fault logic")
> , the preemption was disabled/re-enabled by the
> pagefault_disable()/pagefault_enable() functions. It is no longer the
> case.
>
> Which means that cmpxchg_futex_value_locked(), in kernel/futex.c, no
> longer disables preemption before calling futex_atomic_cmpxchg_inatomic.
> Disabling preemption in the generic implementation was added by commit
> d9b9ff8c1889 ("sched/preempt, futex: Disable preemption in UP futex_atomic_cmpxchg_inatomic() explicitly").
> However, this change is bogus: preemption is not re-enabled in the two failure cases.
>
> Indeed, At boot time futex_detect_cmpxchg(), in kernel/futex.c, checks if
> futex_atomic_cmpxchg_inatomic() operation is provided by the current
> architecture. To do so, it calls futex_atomic_cmpxchg_inatomic() with a
> NULL pointer as the futex, and expects -EFAULT as a return value. The
> reasoning if that if -EFAULT is returned, then there is a valid
> futex_atomic_cmpxchg_inatomic() implementation, while if -ENOSYS is
> returned, then it means that there is no implementation of this function
> for the current architecture.
>
> By doing this test based on a NULL pointer, we fall into the case where
> the get_user() in the asm-generic version of
> futex_atomic_cmpxchg_inatomic() fails, and returns -EFAULT without
> re-enabling the preemption, causing the preemption unbalance in the
> futex_init() initcall.
>
> This bug was not found, because most of the architectures in the kernel
> use their own implementation. This is not the case for the nios2 platform.
> By using a nios2 board without SMP support, this bug can be easily
> reproduced.
>
> Fixes: d9b9ff8c1889 ("sched/preempt, futex: Disable preemption in UP futex_atomic_cmpxchg_inatomic() explicitly")
> Signed-off-by: Romain Perier <romain.perier@xxxxxxxxxxxxxxxxxx>
Reviewed-by: David Howells <dhowells@xxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-arch" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
