Source: mariadb-10.1 Version: 10.1.23-8 Severity: important Tags: patch Hi,
While doing some testing for upstream, I had to apply a slightly modified version of c11_atomics.patch to get 10.2 to build on MIPS. However although MariaDB builds after that, it immediately aborts near some atomics related code. I discovered that the implementation of make_atomic_add_body in incorrect in c11_atomics.patch. At the moment we have this: +#define make_atomic_add_body(S) \ + __atomic_add_fetch(a, v, __ATOMIC_SEQ_CST) From the documentation of make_atomic_add, it looks like v should be set to the previous value of *a which is not done here. The correct implementation is: +#define make_atomic_add_body(S) \ + v= __atomic_fetch_add(a, v, __ATOMIC_SEQ_CST) This is very similar to the operation used for the native GCC atomics above. While this bug causes 10.2 to completely break on MIPS, I haven't seen 10.1 break. It may be that this code isn't used anywhere important in 10.1, but I still think it would be good to fix this before finding out some obscure place it does break. Thanks, James
diff -Nru mariadb-10.1-10.1.24/debian/patches/c11_atomics.patch mariadb-10.1-10.1.24/debian/patches/c11_atomics.patch --- mariadb-10.1-10.1.24/debian/patches/c11_atomics.patch 2017-06-07 01:23:44.000000000 +0100 +++ mariadb-10.1-10.1.24/debian/patches/c11_atomics.patch 2017-06-14 14:03:27.000000000 +0100 @@ -59,7 +59,7 @@ +#elif defined(HAVE_GCC_C11_ATOMICS) + +#define make_atomic_add_body(S) \ -+ __atomic_add_fetch(a, v, __ATOMIC_SEQ_CST) ++ v= __atomic_fetch_add(a, v, __ATOMIC_SEQ_CST) +#define make_atomic_fas_body(S) \ + v= __atomic_exchange_n(a, v, __ATOMIC_SEQ_CST) +#define make_atomic_cas_body(S) \
signature.asc
Description: OpenPGP digital signature
