arm64: futex: Fix FUTEX_WAKE_OP atomic ops with non-zero result value
authorWill Deacon <will.deacon@arm.com>
Mon, 8 Apr 2019 11:45:09 +0000 (12:45 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 17 Apr 2019 06:38:52 +0000 (08:38 +0200)
commit82a30a5d60e30e66ea4965c119029683fa94950e
tree6f18d09e4ce6cc72db8c21d14fd8bee21c91052e
parent4362ff977bc63f85e4a4f23845d67a7a8559eaad
arm64: futex: Fix FUTEX_WAKE_OP atomic ops with non-zero result value

commit 045afc24124d80c6998d9c770844c67912083506 upstream.

Rather embarrassingly, our futex() FUTEX_WAKE_OP implementation doesn't
explicitly set the return value on the non-faulting path and instead
leaves it holding the result of the underlying atomic operation. This
means that any FUTEX_WAKE_OP atomic operation which computes a non-zero
value will be reported as having failed. Regrettably, I wrote the buggy
code back in 2011 and it was upstreamed as part of the initial arm64
support in 2012.

The reasons we appear to get away with this are:

  1. FUTEX_WAKE_OP is rarely used and therefore doesn't appear to get
     exercised by futex() test applications

  2. If the result of the atomic operation is zero, the system call
     behaves correctly

  3. Prior to version 2.25, the only operation used by GLIBC set the
     futex to zero, and therefore worked as expected. From 2.25 onwards,
     FUTEX_WAKE_OP is not used by GLIBC at all.

Fix the implementation by ensuring that the return value is either 0
to indicate that the atomic operation completed successfully, or -EFAULT
if we encountered a fault when accessing the user mapping.

Cc: <stable@kernel.org>
Fixes: 6170a97460db ("arm64: Atomic operations")
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
arch/arm64/include/asm/futex.h