Fix the calls to ibv_fork_init and remove btl_openib_want_fork_support.

In order to have an effect, ibv_fork_init should be called in the
beginning of the verbs initialization flow - before the calls to the
ibv_create_qp and ibv_create_cq verbs.
These functions are called from the oob/ud code and by the time the
other verbs components (btl openib, pml yalla, ...) call ibv_fork_init,
it's too late. This commit forces the call to ibv_fork_init (if it's
requested) right at the beginning of all the components that are using
verbs.
(ibv_fork_init() can be safely called multiple times)

This commit also removes the btl_openib_want_fork_support mca parameter
and adds a new mca parameter instead - opal_verbs_want_fork_support.
Through this new parameter, fork support may be requested for ALL
components.
The default value for this parameter is set to 1.

Before this commit the btl_openib_want_fork_support parameter didn't
provide fork support for the openib btl if its value was set to 1.
(because when openib called ibv_fork_init, it was already after the
calls to ibv_create_* in oob/ud and thereofre it failed).

opal/mca/btl/openib/btl_openib.h

@@ -278,8 +278,6 @@ struct mca_btl_openib_component_t {
     unsigned int cq_poll_progress;
     unsigned int cq_poll_batch;
     unsigned int eager_rdma_poll_ratio;
-    /** Whether we want fork support or not */
-    int want_fork_support;
     int rdma_qp;
     int credits_qp; /* qp used for software flow control */
     bool cpc_explicitly_defined;

opal/mca/btl/openib/btl_openib_component.c

@@ -2634,23 +2634,10 @@ btl_openib_component_init(int *num_btl_modules,
         goto no_btls;
     }
 
-    /* If we want fork support, try to enable it */
-#ifdef HAVE_IBV_FORK_INIT
-    if (0 != mca_btl_openib_component.want_fork_support) {
-        if (0 != ibv_fork_init()) {
-            /* If the want_fork_support MCA parameter is >0, then the
-               user was specifically asking for fork support and we
-               couldn't provide it.  So print an error and deactivate
-               this BTL. */
-            if (mca_btl_openib_component.want_fork_support > 0) {
-                opal_show_help("help-mpi-btl-openib.txt",
-                               "ibv_fork_init fail", true,
-                               opal_process_info.nodename);
-                goto no_btls;
-            }
-        }
+    /* If fork support is requested, try to enable it */
+    if (OPAL_SUCCESS != (ret = opal_common_verbs_fork_test())) {
+        goto no_btls;
     }
-#endif
 
     /* Parse the include and exclude lists, checking for errors */
     mca_btl_openib_component.if_include_list =

opal/mca/btl/openib/btl_openib_mca.c

@@ -270,11 +270,6 @@ int btl_openib_register_mca_params(void)
                    MCA_BTL_OPENIB_CQ_POLL_BATCH_DEFAULT, &mca_btl_openib_component.cq_poll_batch,
                    REGINT_GE_ONE));
 
-    CHECK(reg_int("want_fork_support", NULL,
-                  "Whether fork support is desired or not "
-                  "(negative = try to enable fork support, but continue even if it is not available, 0 = do not enable fork support, positive = try to enable fork support and fail if it is not available)",
-                  0, &mca_btl_openib_component.want_fork_support, 0));
-
     asprintf(&str, "%s/mca-btl-openib-device-params.ini",
              opal_install_dirs.opaldatadir);
     if (NULL == str) {

opal/mca/common/verbs/common_verbs.h

@@ -164,6 +164,20 @@ opal_common_verbs_find_max_inline(struct ibv_device *device,
  */
 OPAL_DECLSPEC int opal_common_verbs_qp_test(struct ibv_context *device_context, 
                                             int flags);
+/*
+ * ibv_fork_init testing - if fork support is requested then ibv_fork_init
+ * should be called right at the beginning of the verbs initialization flow, before ibv_create_* call.
+ *
+ * Known limitations:
+ * If ibv_fork_init is called after ibv_create_* functions - it will have no effect.
+ * OMPI initializes verbs many times during initialization in the following verbs components:
+ *      oob/ud, btl/openib, mtl/mxm, pml/yalla, oshmem/ikrit, oshmem/yoda, ompi/mca/coll/{fca,hcoll}
+ *
+ * So, ibv_fork_init should be called once, in the beginning of the init flow of every verb component
+ * to proper request fork support.
+ *
+ */
+int opal_common_verbs_fork_test(void);
 
 END_C_DECLS
 

opal/mca/common/verbs/common_verbs_basics.c

@@ -35,6 +35,9 @@ const char *ibv_get_sysfs_path(void);
 #endif
 
 #include "common_verbs.h"
+#include "opal/runtime/opal_params.h"
+#include "opal/util/show_help.h"
+#include "opal/util/proc.h"
 
 /***********************************************************************/
 
@@ -61,3 +64,31 @@ bool opal_common_verbs_check_basics(void)
     return true;
 }
 
+int opal_common_verbs_fork_test(void)
+{
+    /* Make sure that ibv_fork_init is called before the calls to other memory registering verbs,
+     * which will be called after this function */
+#ifdef HAVE_IBV_FORK_INIT
+    if (0 != opal_verbs_want_fork_support) {
+        /* Check if fork support is requested by the user */
+        if (0 != ibv_fork_init()) {
+            /* If the opal_want_fork_support MCA parameter is >0 but
+             * the call to ibv_fork_init() failed, then return an error code.
+             */
+            if (opal_verbs_want_fork_support > 0) {
+                opal_show_help("help-opal-common-verbs.txt",
+                               "ibv_fork_init fail", true,
+                               opal_proc_local_get()->proc_hostname, errno,
+                               strerror(errno));
+                return OPAL_ERROR;
+            }
+        } else {
+            return OPAL_SUCCESS;
+        }
+    } else {
+        return OPAL_SUCCESS;
+    }
+#endif
+    return OPAL_SUCCESS;
+}
+

opal/mca/common/verbs/help-opal-common-verbs.txt

@@ -46,3 +46,9 @@ specified:
 These entities will be ignored.  You can disable this warning by
 setting the ompi_common_verbs_warn_nonexistent_if MCA parameter to 0.
 #
+[ibv_fork_init fail]
+Fork support was requested but the library call ibv_fork_init() failed.
+
+  Hostname:    %s
+  Error (%d):  %s
+#

opal/runtime/opal_params.c

@@ -275,6 +275,16 @@ int opal_register_params(void)
         return ret; 
     }
 
+    opal_verbs_want_fork_support = 1;
+    ret = mca_base_var_register("opal", "opal", NULL, "verbs_want_fork_support",
+                    "Whether fork support is desired or not "
+                    "(negative = try to enable fork support, but continue even "
+                    "if it is not available, 0 = do not enable fork support, "
+                    "positive = try to enable fork support and fail if it is not available)",
+                    MCA_BASE_VAR_TYPE_INT, NULL, 0, MCA_BASE_VAR_FLAG_SETTABLE,
+                    OPAL_INFO_LVL_8, MCA_BASE_VAR_SCOPE_ALL_EQ,
+                    &opal_verbs_want_fork_support);
+
     return OPAL_SUCCESS;
 }
 

opal/runtime/opal_params.h

@@ -30,6 +30,8 @@ extern char *opal_signal_string;
 extern char *opal_net_private_ipv4;
 extern char *opal_set_max_sys_limits;
 
+int opal_verbs_want_fork_support;
+
 #if OPAL_ENABLE_TIMING
 extern char *opal_timing_sync_file;
 extern char *opal_timing_output;
@@ -38,6 +40,7 @@ extern bool opal_timing_overhead;
 
 OPAL_DECLSPEC extern int opal_initialized;
 OPAL_DECLSPEC extern bool opal_built_with_cuda_support;
+
 /**
  *  * Whether we want to enable CUDA GPU buffer send and receive support.
  *   */

orte/mca/oob/ud/Makefile.am

@@ -55,7 +55,8 @@ mcacomponentdir = $(ortelibdir)
 mcacomponent_LTLIBRARIES = $(component_install)
 mca_oob_ud_la_SOURCES = $(sources)
 mca_oob_ud_la_LDFLAGS = -module -avoid-version $(orte_oob_ud_LDFLAGS)
-mca_oob_ud_la_LIBADD = $(orte_oob_ud_LIBS)
+mca_oob_ud_la_LIBADD = $(orte_oob_ud_LIBS) \
+					   $(OPAL_TOP_BUILDDIR)/opal/mca/common/verbs/lib@OPAL_LIB_PREFIX@mca_common_verbs.la
 
 noinst_LTLIBRARIES = $(component_noinst)
 libmca_oob_ud_la_SOURCES = $(sources)

orte/mca/oob/ud/oob_ud_component.c

@@ -24,6 +24,8 @@
 
 #include "oob_ud_component.h"
 
+#include "opal/mca/common/verbs/common_verbs.h"
+
 static int   mca_oob_ud_component_open (void);
 static int   mca_oob_ud_component_close (void);
 static int   mca_oob_ud_component_register (void);
@@ -217,6 +219,16 @@ static inline int mca_oob_ud_device_setup (mca_oob_ud_device_t *device,
                          "%s oob:ud:device_setup attempting to setup ib device %p",
                          ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), (void *) ib_device);
 
+
+    /* If fork support is requested, try to enable it */
+    rc = opal_common_verbs_fork_test();
+    if (OPAL_SUCCESS != rc) {
+        opal_output_verbose(5, orte_oob_base_framework.framework_output,
+                            "%s oob:ud:device_setup failed in ibv_fork_init. errno = %d",
+                            ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), errno);
+        return ORTE_ERROR;
+    }
+
     device->ib_context = ibv_open_device (ib_device);
     if (NULL == device->ib_context) {
         opal_output_verbose(5, orte_oob_base_framework.framework_output,

oshmem/mca/sshmem/verbs/Makefile.am

@@ -30,7 +30,8 @@ mcacomponentdir = $(oshmemlibdir)
 mcacomponent_LTLIBRARIES = $(component_install)
 mca_sshmem_verbs_la_SOURCES = $(sources)
 mca_sshmem_verbs_la_LDFLAGS = -module -avoid-version $(oshmem_verbs_LDFLAGS)
-mca_sshmem_verbs_la_LIBADD  = $(oshmem_verbs_LIBS)
+mca_sshmem_verbs_la_LIBADD  = $(oshmem_verbs_LIBS) \
+							  $(OPAL_TOP_BUILDDIR)/opal/mca/common/verbs/lib@OPAL_LIB_PREFIX@mca_common_verbs.la
 
 noinst_LTLIBRARIES = $(component_noinst)
 libmca_sshmem_verbs_la_SOURCES =$(sources)

oshmem/mca/sshmem/verbs/sshmem_verbs_component.c

@@ -19,6 +19,7 @@
 
 #include "opal/constants.h"
 #include "opal/util/sys_limits.h"
+#include "opal/mca/common/verbs/common_verbs.h"
 
 #include "oshmem/mca/sshmem/sshmem.h"
 #include "oshmem/mca/sshmem/base/base.h"
@@ -100,6 +101,11 @@ verbs_runtime_query(mca_base_module_t **module,
     *priority = 0;
     *module = NULL;
 
+    /* If fork support is requested, try to enable it */
+    if (OSHMEM_SUCCESS != (rc = opal_common_verbs_fork_test())) {
+        return OSHMEM_ERROR;
+    }
+
     memset(device, 0, sizeof(*device));
 
 #ifdef HAVE_IBV_GET_DEVICE_LIST