Problem: yqueue false sharing issues on PPC64

Solution: detect cacheline size for aligment purposes at build time
instead of hard-coding it, so that PPC and S390 can align to a value
greater than the 64 bytes default.
Uses libc getconf program, and falls back to the previous value of 64
if not found.

CMakeLists.txt

@@ -251,6 +251,14 @@ include(CMakeDependentOption)
 include(CheckCXXSymbolExists)
 include(CheckSymbolExists)
 
+execute_process(COMMAND getconf LEVEL1_DCACHE_LINESIZE OUTPUT_VARIABLE CACHELINE_SIZE ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+if(CACHELINE_SIZE STREQUAL "" OR CACHELINE_SIZE EQUAL 0 OR CACHELINE_SIZE EQUAL -1)
+  set(ZMQ_CACHELINE_SIZE 64)
+else()
+  set(ZMQ_CACHELINE_SIZE ${CACHELINE_SIZE})
+endif()
+message(STATUS "Using ${ZMQ_CACHELINE_SIZE} bytes alignment for lock-free data structures")
+
 if(NOT CYGWIN)
   # TODO cannot we simply do 'if(WIN32) set(ZMQ_HAVE_WINDOWS ON)' or similar?
   check_include_files(windows.h ZMQ_HAVE_WINDOWS)

acinclude.m4

@@ -1169,3 +1169,23 @@ AC_DEFUN([LIBZMQ_CHECK_POLLER], [{
         AC_MSG_ERROR([Invalid API poller '$api_poller' specified])
 	fi
 }])
+
+dnl ##############################################################################
+dnl # LIBZMQ_CHECK_CACHELINE                                                     #
+dnl # Check cacheline size for alignment purposes                                #
+dnl ##############################################################################
+AC_DEFUN([LIBZMQ_CHECK_CACHELINE], [{
+
+    zmq_cacheline_size=64
+    AC_CHECK_TOOL(libzmq_getconf, getconf)
+    if ! test "x$libzmq_getconf" = "x"; then
+        zmq_cacheline_size=$($libzmq_getconf LEVEL1_DCACHE_LINESIZE 2>/dev/null || echo 64)
+        if test "x$zmq_cacheline_size" = "x0" -o  "x$zmq_cacheline_size" = "x-1"; then
+            # getconf on some architectures does not know the size, try to fallback to
+            # the value the kernel knows on Linux
+            zmq_cacheline_size=$(cat /sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size 2>/dev/null || echo 64)
+        fi
+    fi
+	AC_MSG_NOTICE([Using "$zmq_cacheline_size" bytes alignment for lock-free data structures])
+	AC_DEFINE_UNQUOTED(ZMQ_CACHELINE_SIZE, $zmq_cacheline_size, [Using "$zmq_cacheline_size" bytes alignment for lock-free data structures])
+}])

builds/cmake/platform.hpp.in

@@ -11,6 +11,8 @@
 #cmakedefine ZMQ_POLL_BASED_ON_SELECT
 #cmakedefine ZMQ_POLL_BASED_ON_POLL
 
+#cmakedefine ZMQ_CACHELINE_SIZE @ZMQ_CACHELINE_SIZE@
+
 #cmakedefine ZMQ_FORCE_MUTEXES
 
 #cmakedefine HAVE_FORK

configure.ac

@@ -376,6 +376,9 @@ LIBZMQ_CHECK_DOC_BUILD
 # Check polling system, set appropriate macro in src/platform.hpp
 LIBZMQ_CHECK_POLLER
 
+# Check cacheline size, set appropriate macro in src/platform.hpp
+LIBZMQ_CHECK_CACHELINE
+
 # Checks for header files.
 AC_HEADER_STDC
 AC_CHECK_HEADERS(\

src/yqueue.hpp

@@ -55,8 +55,9 @@ namespace zmq
 // posix_memalign available. Default value is 64, this alignment will
 // prevent two queue chunks from occupying the same CPU cache line on
 // architectures where cache lines are <= 64 bytes (e.g. most things
-// except POWER).
-template <typename T, int N, size_t ALIGN = 64> class yqueue_t
+// except POWER). It is detected at build time to try to account for other
+// platforms like POWER and s390x.
+template <typename T, int N, size_t ALIGN = ZMQ_CACHELINE_SIZE> class yqueue_t
 #else
 template <typename T, int N> class yqueue_t
 #endif