This patch enables atomics on x64 builds. Reuse the atomics defined for x86 and add atomics for 64 bit reads/writes.
Before this patch the cmap test gives us: $ ./tests/ovstest.exe test-cmap benchmark 10000000 3 1 Benchmarking with n=10000000, 3 threads, 1.00% mutations, batch size 1: cmap insert: 20100 ms cmap iterate: 2967 ms batch search: 10929 ms cmap destroy: 13489 ms cmap insert: 20079 ms cmap iterate: 2953 ms cmap search: 10559 ms cmap destroy: 13486 ms hmap insert: 2021 ms hmap iterate: 1162 ms hmap search: 5152 ms hmap destroy: 1158 ms After this change we have: $ ./tests/ovstest.exe test-cmap benchmark 10000000 3 1 Benchmarking with n=10000000, 3 threads, 1.00% mutations, batch size 1: cmap insert: 2953 ms cmap iterate: 267 ms batch search: 2193 ms cmap destroy: 2037 ms cmap insert: 2909 ms cmap iterate: 267 ms cmap search: 2167 ms cmap destroy: 2087 ms hmap insert: 1853 ms hmap iterate: 1086 ms hmap search: 4395 ms hmap destroy: 1140 ms We should probably revisit this file and investigate it further to see if we can squeeze more performance. As a side effect fix tests on x64 because usage of `ovs-atomic-pthreads.h` is currently broken. Signed-off-by: Alin Gabriel Serdean <[email protected]> Suggested-by: Ben Pfaff <[email protected]> --- v2: Change commit message. --- lib/ovs-atomic-msvc.h | 12 ++++++++++++ lib/ovs-atomic.h | 2 +- 2 files changed, 13 insertions(+), 1 deletion(-) diff --git a/lib/ovs-atomic-msvc.h b/lib/ovs-atomic-msvc.h index c6a7db3..0b041c6 100644 --- a/lib/ovs-atomic-msvc.h +++ b/lib/ovs-atomic-msvc.h @@ -107,6 +107,7 @@ atomic_signal_fence(memory_order order) * InterlockedExchange64Acquire() available. So we are forced to use * InterlockedExchange64() which uses full memory barrier for everything * greater than 'memory_order_relaxed'. */ +#ifdef _M_IX86 #define atomic_store64(DST, SRC, ORDER) \ if (ORDER == memory_order_relaxed) { \ InterlockedExchangeNoFence64((int64_t volatile *) (DST), \ @@ -114,6 +115,11 @@ atomic_signal_fence(memory_order order) } else { \ InterlockedExchange64((int64_t volatile *) (DST), (int64_t) (SRC));\ } +#elif _M_X64 +/* 64 bit writes are atomic on amd64 if 64 bit aligned. */ +#define atomic_store64(DST, SRC, ORDER) \ + atomic_storeX(64, DST, SRC, ORDER) +#endif /* Used for 8 and 16 bit variations. */ #define atomic_storeX(X, DST, SRC, ORDER) \ @@ -160,11 +166,17 @@ atomic_signal_fence(memory_order order) /* MSVC converts 64 bit reads into two instructions. So there is * a possibility that an interrupt can make a 64 bit read non-atomic even * when 8 byte aligned. So use fully memory barrier InterlockedOr64(). */ +#ifdef _M_IX86 #define atomic_read64(SRC, DST, ORDER) \ __pragma (warning(push)) \ __pragma (warning(disable:4047)) \ *(DST) = InterlockedOr64((int64_t volatile *) (SRC), 0); \ __pragma (warning(pop)) +#elif _M_X64 +/* 64 bit reads are atomic on amd64 if 64 bit aligned. */ +#define atomic_read64(SRC, DST, ORDER) \ + *(DST) = *(SRC); +#endif #define atomic_read(SRC, DST) \ atomic_read_explicit(SRC, DST, memory_order_seq_cst) diff --git a/lib/ovs-atomic.h b/lib/ovs-atomic.h index f1f2c38..c835eb7 100644 --- a/lib/ovs-atomic.h +++ b/lib/ovs-atomic.h @@ -335,7 +335,7 @@ #include "ovs-atomic-i586.h" #elif HAVE_GCC4_ATOMICS #include "ovs-atomic-gcc4+.h" - #elif _MSC_VER && _M_IX86 >= 500 + #elif _MSC_VER #include "ovs-atomic-msvc.h" #else /* ovs-atomic-pthreads implementation is provided for portability. -- 2.10.2.windows.1 _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
