mmap shared memory transport

.github/workflows/ci.yml

@@ -707,3 +707,38 @@ jobs:
           make check TESTS= -j
           SHMEM_DEBUG=1 SHMEM_INFO=1 make VERBOSE=1 TEST_RUNNER="${SOS_PM} -np 2" check
           ${SOS_PM} -np 1 modules/tests-sos/test/unit/hello
+
+  mmap_only:
+    runs-on: ubuntu-20.04
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - config_name: transport_none
+        sos_config: [--enable-mmap --enable-shr-atomics --enable-error-checking --enable-pmi-simple]
+
+    steps:
+      - name: Checking OS version
+        run: |
+          echo "OS_NAME=$(lsb_release -si)-$(ls_release -sr)" >> $GITHUB_ENV
+      - uses: actions/checkout@v2
+      - name: Install dependencies
+        run: |
+          sudo apt-get install -y gfortran mpich libmpich-dev libev-dev libev-libevent-dev
+          sudo sysctl -w kernel.yama.ptrace_scope=0
+          sudo sysctl -w kernel.randomize_va_space=0
+
+      # SOS
+      - name: Build SOS (${{ matrix.name }})
+        run: |
+          ./autogen.sh
+          mkdir build; cd build
+          ../configure --prefix=${SOS_INSTALL_DIR} ${{ matrix.sos_config }}
+          make -j
+          make install
+      - name: Test SOS (${{ matrix.name }})
+        run: |
+          cd build
+          make check TESTS= -j
+          SHMEM_DEBUG=1 SHMEM_INFO=1 make VERBOSE=1 TEST_RUNNER="${SOS_PM} -np 2" check
+          cat modules/tests-sos/test/unit/hello.log

configure.ac

@@ -307,6 +307,10 @@ AC_ARG_ENABLE([memcpy],
     [AC_HELP_STRING([--enable-memcpy],
                     [Use memcpy to perform local put/get operations (default:disabled)])])
 
+AC_ARG_ENABLE([mmap],
+    [AC_HELP_STRING([--enable-mmap],
+                    [Use mmap to share symmetric data and heap segments in shared memory (default:disabled)])])
+
 AC_ARG_ENABLE([ofi-fence],
     [AC_HELP_STRING([--enable-ofi-fence],
                     [Use FI_FENCE feature to optimize put-with-signal operations. (default: disabled)])])
@@ -476,7 +480,7 @@ else
     transport_portals4="no"
     transport_ofi="no"
     transport_ucx="no"
-    AC_MSG_WARN([No transport requested])
+    AC_MSG_WARN([No network transport requested])
 fi
 
 AM_CONDITIONAL([USE_PORTALS4], [test "$transport_portals4" = "yes"])
@@ -491,38 +495,59 @@ CHECK_CMA(
     [transport_cma="yes"],
     [transport_cma="no"])
 
-# If both XPMEM and CMA requested, user needs to choose one:
-if test -n "$with_xpmem" -a "$with_xpmem" != "no" -a -n "$with_cma" -a "$with_cma" != "no" ; then
-    AC_MSG_ERROR([Cannot choose both XPMEM and CMA transports, see --help for details])
-# Check which was requested, XPMEM or CMA:
+
+num_shared_transports=0
+if test -n "$with_xpmem" -a "$with_xpmem" != "no" ; then
+    ((num_shared_transports+=1))
+fi
+if test -n "$with_cma" -a "$with_cma" != "no" ; then
+    ((num_shared_transports+=1))
+fi
+if test -n "$enable_mmap" -a "$enable_mmap" != "no" ; then
+    ((num_shared_transports+=1))
+fi
+
+# If more than one shared memory transport requested, user needs to choose one:
+if [[ $num_shared_transports -gt 1 ]] ; then
+    echo "Selected" $num_shared_transports "shared memory transports, please choose one."
+    AC_MSG_ERROR([Only one shared memory transport is allowed (mmap, XPMEM, or CMA), see --help for details])
+elif test -n "$enable_mmap" -a "$enable_mmap" != "no" ; then
+    transport_mmap="yes"
+    transport_cma="no"
+    transport_xpmem="no"
+    AC_DEFINE([USE_MMAP], [1], [Define if mmap transport is active])
+    AC_CHECK_LIB(rt, shm_open, [MMAP_LDFLAGS="-lrt"])
+    AC_SUBST([MMAP_LDFLAGS])
 elif test -n "$with_xpmem" -a "$with_xpmem" != "no" ; then
+    transport_mmap="no"
     transport_cma="no"
     AC_DEFINE([USE_XPMEM], [1], [Define if XPMEM transport is active])
 elif test -n "$with_cma" -a "$with_cma" != "no" ; then
+    transport_mmap="no"
     transport_xpmem="no"
     AC_DEFINE([USE_CMA], [1], [Define if Cross Memory Attach transport is active])
     AC_DEFINE([_GNU_SOURCE], [1], [CMA transport header requires global definition of _GNU_SOURCE])
-# If neither, disable XPMEM and CMA:
 else
+    transport_mmap="no"
     transport_xpmem="no"
     transport_cma="no"
-    AC_MSG_RESULT([Neither XPMEM nor CMA transport requested])
-
+    AC_MSG_RESULT([Shared memory transport not requested])
 fi
 
-if test "$enable_memcpy" = "yes" -a "$transport_xpmem" = "no" -a "$transport_cma" = "no" ; then
+if test "$enable_memcpy" = "yes" -a "$transport_xpmem" = "no" -a "$transport_cma" = "no" -a "$transport_mmap" = "no" ; then
     transport_memcpy="yes"
     AC_DEFINE([USE_MEMCPY], [1], [Define to use memcpy for local put/get communication])
-elif test "$transport_xpmem" = "yes" -o "$transport_cma" = "yes" ; then
+elif test "$transport_xpmem" = "yes" -o "$transport_cma" = "yes" -o "$transport_mmap" = "yes" ; then
     transport_memcpy="yes"
 else
     transport_memcpy="no"
 fi
 
 AM_CONDITIONAL([USE_XPMEM], [test "$transport_xpmem" = "yes"])
 AM_CONDITIONAL([USE_CMA], [test "$transport_cma" = "yes"])
+AM_CONDITIONAL([USE_MMAP], [test "$transport_mmap" = "yes"])
 
-AS_IF([test "$transport_xpmem" = "yes" -o "$transport_cma" = "yes"],
+AS_IF([test "$transport_xpmem" = "yes" -o "$transport_cma" = "yes" -o "$transport_mmap" = "yes"],
       [AC_DEFINE([USE_ON_NODE_COMMS], [1], [Define if any on-node comm transport is available])
        AC_DEFINE([ENABLE_HARD_POLLING], [1], [Enable hard polling])
       ])
@@ -533,7 +558,7 @@ else
     transport_shr_atomics="no"
 fi
 
-if test "$enable_shr_atomics" != "no" -a "$transport_xpmem" = "yes" -a "$transport" = "none"; then
+if [[[ "$enable_shr_atomics" != "no" && ( "$transport_xpmem" = "yes" || "$transport_mmap" = "yes" ) && "$transport" = "none" ]]]; then
     transport_shr_atomics="yes"
     AC_DEFINE([USE_SHR_ATOMICS], [1], [If defined, the shared memory layer will perform processor atomics.])
 fi
@@ -870,6 +895,9 @@ fi
 if test -n "$with_xpmem" -a "$with_xpmem" != "no" -a "$with_xpmem" != "yes" ; then
   DISTCHECK_CONFIGURE_FLAGS="$DISTCHECK_CONFIGURE_FLAGS --with-xpmem=${with_xpmem}"
 fi
+if test -n "$enable_mmap" -a "$enable_mmap" != "no" -a "$enable_mmap" != "yes" ; then
+  DISTCHECK_CONFIGURE_FLAGS="$DISTCHECK_CONFIGURE_FLAGS --enable-mmap=${enable_mmap}"
+fi
 if test -n "$with_pmi" -a "$with_pmi" != "no" -a "$with_pmi" != "yes" ; then
   DISTCHECK_CONFIGURE_FLAGS="$DISTCHECK_CONFIGURE_FLAGS --with-pmi=${with_pmi}"
 fi
@@ -932,6 +960,13 @@ fi
 if test -n "opal_hwloc_LDFLAGS"; then
     WRAPPER_COMPILER_EXTRA_LDFLAGS="$WRAPPER_COMPILER_EXTRA_LDFLAGS $opal_hwloc_LDFLAGS"
 fi
+if test "$transport_mmap" = "yes" ; then
+    CPPFLAGS="$CPPFLAGS $MMAP_CPPFLAGS"
+    LDFLAGS="$LDFLAGS $MMAP_LDFLAGS"
+    LIBS="$LIBS $MMAP_LIBS"
+    WRAPPER_COMPILER_EXTRA_LDFLAGS="$MMAP_LDFLAGS"
+    WRAPPER_COMPILER_EXTRA_LIBS="$MMAP_LIBS"
+fi
 
 CPPFLAGS="$CPPFLAGS $pmi_CPPFLAGS"
 LDFLAGS="$LDFLAGS $pmi_LDFLAGS $aslr_LDFLAGS"
@@ -985,7 +1020,7 @@ AS_IF([test "$enable_pmi_mpi" != "yes" -a "$enable_pmi_simple" != "yes" -a "$opa
          [AC_MSG_ERROR([No PMI client interface was configured, consider --enable-pmi-simple or --with-pmi])])
 
 AS_IF([test -z "$num_transports"],
-      [AC_MSG_WARN([No transport found, resulting library will be unable to exchange messages])])
+      [AC_MSG_WARN([No network transport found, library will be unable to exchange remote messages])])
 
 AS_IF([test "$shmem_cv_c11_works" != "yes"],
       [AC_MSG_WARN([C compiler does not support _Generic, unable to verify and test C11 bindings])])
@@ -1025,6 +1060,7 @@ echo ""
 echo "On Node Communication:"
 echo "  XPMEM:          $transport_xpmem"
 echo "  CMA:            $transport_cma"
+echo "  mmap:           $transport_mmap"
 echo "  memcpy (self):  $transport_memcpy"
 echo "  Shr. atomics:   $transport_shr_atomics"
 echo ""

src/Makefile.am

@@ -128,6 +128,12 @@ libsma_la_SOURCES += \
 	transport_cma.c
 endif
 
+if USE_MMAP
+libsma_la_SOURCES += \
+	transport_mmap.h \
+	transport_mmap.c
+endif
+
 if USE_PMI_SIMPLE
 AM_CPPFLAGS += -I$(top_srcdir)/pmi-simple
 libsma_la_SOURCES += \

src/shmem_remote_pointer.h

@@ -34,6 +34,8 @@ shmem_internal_ptr(const void *target, int pe)
     if (-1 != (node_rank = shmem_internal_get_shr_rank(pe))) {
 #if USE_XPMEM
         return shmem_transport_xpmem_ptr(target, pe, node_rank);
+#elif USE_MMAP
+        return shmem_transport_mmap_ptr(target, pe, node_rank);
 #else
         return NULL;
 #endif

src/shr_transport.h4

@@ -27,6 +27,10 @@ include(shmem_bind_c.m4)dnl
 #include "transport_cma.h"
 #endif
 
+#ifdef USE_MMAP
+#include "transport_mmap.h"
+#endif
+
 static inline int
 shmem_shr_transport_init(void)
 {
@@ -41,6 +45,11 @@ shmem_shr_transport_init(void)
     ret = shmem_transport_cma_init();
     if (0 != ret)
         RETURN_ERROR_MSG("CMA init failed (%d)\n", ret);
+
+#elif USE_MMAP
+    ret = shmem_transport_mmap_init();
+    if (0 != ret)
+        RETURN_ERROR_MSG("mmap init failed (%d)\n", ret);
 #endif
 
     return ret;
@@ -63,6 +72,12 @@ shmem_shr_transport_startup(void)
     if (0 != ret) {
         RETURN_ERROR_MSG("CMA startup failed (%d)\n", ret);
     }
+
+#elif USE_MMAP
+    ret = shmem_transport_mmap_startup();
+    if (0 != ret) {
+        RETURN_ERROR_MSG("mmap startup failed (%d)\n", ret);
+    }
 #endif
 
     return ret;
@@ -76,6 +91,8 @@ shmem_shr_transport_fini(void)
     shmem_transport_xpmem_fini();
 #elif USE_CMA
     shmem_transport_cma_fini();
+#elif USE_MMAP
+    shmem_transport_mmap_fini();
 #endif
 }
 
@@ -87,6 +104,8 @@ shmem_shr_transport_ptr(void *target, int noderank, void **local_ptr)
 {
 #if USE_XPMEM
     XPMEM_GET_REMOTE_ACCESS(target, noderank, *local_ptr);
+#elif USE_MMAP
+    MMAP_GET_REMOTE_ACCESS(target, noderank, *local_ptr);
 #else
     RAISE_ERROR_MSG("No path to peer (%d)\n", noderank);
 #endif
@@ -147,6 +166,10 @@ shmem_shr_transport_put_scalar(shmem_ctx_t ctx, void *target,
 #elif USE_CMA
     shmem_transport_cma_put(target, source, len, pe,
                             shmem_internal_get_shr_rank(pe));
+
+#elif USE_MMAP
+    shmem_transport_mmap_put(target, source, len, pe,
+                             shmem_internal_get_shr_rank(pe));
 #else
     RAISE_ERROR_STR("No path to peer");
 #endif
@@ -165,6 +188,10 @@ shmem_shr_transport_put(shmem_ctx_t ctx, void *target, const void *source,
 #elif USE_CMA
     shmem_transport_cma_put(target, source, len, pe,
                             shmem_internal_get_shr_rank(pe));
+
+#elif USE_MMAP
+    shmem_transport_mmap_put(target, source, len, pe,
+                             shmem_internal_get_shr_rank(pe));
 #else
     RAISE_ERROR_STR("No path to peer");
 #endif
@@ -183,6 +210,10 @@ shmem_shr_transport_get(shmem_ctx_t ctx, void *target, const void *source,
 #elif USE_CMA
     shmem_transport_cma_get(target, source, len, pe,
                             shmem_internal_get_shr_rank(pe));
+
+#elif USE_MMAP
+    shmem_transport_mmap_get(target, source, len, pe,
+                             shmem_internal_get_shr_rank(pe));
 #else
     RAISE_ERROR_STR("No path to peer");
 #endif
@@ -577,6 +608,12 @@ shmem_shr_transport_put_signal(shmem_ctx_t ctx, void *target,
                               shmem_internal_get_shr_rank(pe));
     shmem_internal_membar_acq_rel(); /* Memory fence to ensure target PE observes
                                         stores in the correct order */
+#elif USE_MMAP
+    shmem_transport_mmap_put(target, source, len, pe,
+                              shmem_internal_get_shr_rank(pe));
+    shmem_internal_membar_acq_rel(); /* Memory fence to ensure target PE observes
+                                        stores in the correct order */
+
 #if USE_SHR_ATOMICS 
     if (sig_op == SHMEM_SIGNAL_ADD)
         shmem_shr_transport_atomic(ctx, sig_addr, &signal, sizeof(uint64_t),