Merge pull request #6838 from raffenet/4.2.x-updates

[4.2.x] cherry-picked fixes in preparation of 4.2.0rc1
pmodels · Dec 8, 2023 · 211aa59 · 211aa59
2 parents b7f6fbe + 3482f29
commit 211aa59
Show file tree

Hide file tree

Showing 17 changed files with 202 additions and 38 deletions.
diff --git a/README.vin b/README.vin
@@ -1,8 +1,8 @@
 			MPICH Release %VERSION%
 
 MPICH is a high-performance and widely portable implementation of the
-MPI-4.0 standard from the Argonne National Laboratory.  This release
-has all MPI 4.0 functions and features required by the standard with
+MPI-4.1 standard from the Argonne National Laboratory.  This release
+has all MPI 4.1 functions and features required by the standard with
 the exception of support for user-defined data representations for I/O.
 
 This README file should contain enough information to get you started

diff --git a/maint/version.m4 b/maint/version.m4
@@ -14,7 +14,7 @@
 # changing this by playing with diversions, but then we would probably be
 # playing with autotools-fire.
 
-m4_define([MPICH_VERSION_m4],[4.2.0b1])dnl
+m4_define([MPICH_VERSION_m4],[4.2.0rc1])dnl
 m4_define([MPICH_RELEASE_DATE_m4],[unreleased development copy])dnl
 
 # For libtool ABI versioning rules see:

diff --git a/src/binding/c/rma_api.txt b/src/binding/c/rma_api.txt
@@ -51,6 +51,16 @@ MPI_Accumulate:
     .seealso: MPI_Raccumulate
     .impl: mpid
     .earlyreturn: pt2pt_proc_null
+{ -- error_check --
+    MPI_Datatype origin_elem, target_elem;
+    MPIR_Datatype_get_basic_type(origin_datatype, origin_elem);
+    MPIR_Datatype_get_basic_type(target_datatype, target_elem);
+    if ((origin_count > 0 && origin_elem != MPI_DATATYPE_NULL) ||
+        (target_count > 0 && target_elem != MPI_DATATYPE_NULL)) {
+        MPIR_ERR_CHKANDJUMP(origin_elem != target_elem, mpi_errno, MPI_ERR_TYPE,
+                            "**dtypemismatch");
+    }
+}
 
 MPI_Compare_and_swap:
     .desc: Perform one-sided atomic compare-and-swap.

diff --git a/src/binding/fortran/use_mpi/Makefile.mk b/src/binding/fortran/use_mpi/Makefile.mk
@@ -68,7 +68,7 @@ mpi_fc_sources += \
 # We need the MPI constants in a separate module for some of the
 # interface definitions (the ones that need MPI_ADDRESS_KIND or
 # MPI_OFFSET_KIND)
-src/binding/fortran/use_mpi/mpi.$(MOD)-stamp: src/binding/fortran/use_mpi/$(MPICONSTMOD).$(MOD) src/binding/fortran/use_mpi/$(MPISIZEOFMOD).$(MOD) src/binding/fortran/use_mpi/$(MPIBASEMOD).$(MOD) $(srcdir)/src/binding/fortran/use_mpi/mpi.f90 src/binding/fortran/use_mpi/mpifnoext.h
+src/binding/fortran/use_mpi/mpi.$(MOD)-stamp: src/binding/fortran/use_mpi/$(MPICONSTMOD).$(MOD) src/binding/fortran/use_mpi/$(MPISIZEOFMOD).$(MOD) src/binding/fortran/use_mpi/$(MPIBASEMOD).$(MOD) src/binding/fortran/use_mpi/$(PMPIBASEMOD).$(MOD) $(srcdir)/src/binding/fortran/use_mpi/mpi.f90 src/binding/fortran/use_mpi/mpifnoext.h
 	@rm -f src/binding/fortran/use_mpi/mpi-tmp
 	@touch src/binding/fortran/use_mpi/mpi-tmp
 	@( cd src/binding/fortran/use_mpi && \

diff --git a/src/mpi/request/mpir_request.c b/src/mpi/request/mpir_request.c
@@ -19,6 +19,13 @@ static void init_builtin_request(MPIR_Request * req, int handle, MPIR_Request_ki
     req->cc_ptr = &req->cc;
     req->status.MPI_ERROR = MPI_SUCCESS;
     MPIR_STATUS_SET_CANCEL_BIT(req->status, FALSE);
+    if (kind == MPIR_REQUEST_KIND__RECV) {
+        /* precompleted recv request is returned when source is MPI_PROC_NULL,
+         * or when we are certain status won't be needed. Thus, initialize the
+         * status for MPI_PROC_NULL. */
+        req->status.MPI_SOURCE = MPI_PROC_NULL;
+        req->status.MPI_TAG = MPI_ANY_TAG;
+    }
     req->comm = NULL;
 }
 

diff --git a/src/mpid/ch4/netmod/ucx/subconfigure.m4 b/src/mpid/ch4/netmod/ucx/subconfigure.m4
@@ -47,7 +47,7 @@ AM_COND_IF([BUILD_CH4_NETMOD_UCX],[
             else
                 ucx_opt_flags=""
             fi
-            PAC_CONFIG_SUBDIR_ARGS([modules/ucx],[--disable-static --enable-embedded --with-java=no $ucx_opt_flags],[],[AC_MSG_ERROR(ucx configure failed)])
+            PAC_CONFIG_SUBDIR_ARGS([modules/ucx],[--disable-static --enable-embedded --with-java=no --with-go=no $ucx_opt_flags],[],[AC_MSG_ERROR(ucx configure failed)])
             PAC_POP_ALL_FLAGS()
             ucxdir="modules/ucx"
         fi

diff --git a/src/mpid/ch4/netmod/ucx/ucx_pre.h b/src/mpid/ch4/netmod/ucx/ucx_pre.h
@@ -24,9 +24,30 @@ typedef struct {
     ucp_datatype_t ucp_datatype;
 } MPIDI_UCX_dt_t;
 
+enum MPIDI_UCX_reqtype {
+    MPIDI_UCX_RECV_CONTIG,
+    MPIDI_UCX_RECV_UNPACK,
+    MPIDI_UCX_RECV_IOV,
+    MPIDI_UCX_RECV_UCX_DT,
+};
+
 typedef union {
     ucp_tag_message_h message_handler;
-    MPIDI_UCX_ucp_request_t *ucp_request;
+    struct {
+        MPIDI_UCX_ucp_request_t *ucp_request;
+        enum MPIDI_UCX_reqtype type;
+        union {
+            struct {
+                void *pack_buf;
+                void *user_buf;
+                MPI_Aint count;
+                MPI_Datatype datatype;
+            } pack;
+            struct {
+                ucp_dt_iov_t *iov;
+            } iov;
+        } u;
+    } s;
 } MPIDI_UCX_request_t;
 
 typedef struct {

diff --git a/src/mpid/ch4/netmod/ucx/ucx_recv.h b/src/mpid/ch4/netmod/ucx/ucx_recv.h
@@ -8,6 +8,27 @@
 
 #include "ucx_impl.h"
 
+/*
+=== BEGIN_MPI_T_CVAR_INFO_BLOCK ===
+
+cvars:
+    - name        : MPIR_CVAR_UCX_DT_RECV
+      category    : CH4_UCX
+      type        : boolean
+      default     : false
+      class       : none
+      verbosity   : MPI_T_VERBOSITY_USER_BASIC
+      scope       : MPI_T_SCOPE_ALL_EQ
+      description : |-
+        Variable to select method for receiving noncontiguous data
+        true                - Use UCX datatype with pack/unpack callbacks
+        false               - MPICH will decide to pack/unpack at completion or use IOVs
+                              based on the datatype
+=== END_MPI_T_CVAR_INFO_BLOCK ===
+*/
+
+#define MPIDI_UCX_ENABLE_IOV (UCP_VERSION(UCP_API_MAJOR, UCP_API_MINOR) >= UCP_VERSION(1, 16))
+
 MPL_STATIC_INLINE_PREFIX void MPIDI_UCX_recv_cmpl_cb(void *request, ucs_status_t status,
                                                      const ucp_tag_recv_info_t * info,
                                                      void *user_data)
@@ -31,6 +52,24 @@ MPL_STATIC_INLINE_PREFIX void MPIDI_UCX_recv_cmpl_cb(void *request, ucs_status_t
             rreq->status.MPI_TAG = MPIDI_UCX_get_tag(info->sender_tag);
             MPIR_STATUS_SET_COUNT(rreq->status, count);
         }
+
+        if (MPIDI_UCX_REQ(rreq).s.type == MPIDI_UCX_RECV_UNPACK) {
+            void *pack_buf = MPIDI_UCX_REQ(rreq).s.u.pack.pack_buf;
+            void *user_buf = MPIDI_UCX_REQ(rreq).s.u.pack.user_buf;
+            MPI_Aint user_count = MPIDI_UCX_REQ(rreq).s.u.pack.count;
+            MPI_Datatype datatype = MPIDI_UCX_REQ(rreq).s.u.pack.datatype;
+            MPI_Aint actual_unpack_bytes;
+
+            MPIR_Typerep_unpack(pack_buf, info->length, user_buf, user_count, datatype,
+                                0, &actual_unpack_bytes, MPIR_TYPEREP_FLAG_NONE);
+            if (actual_unpack_bytes < info->length) {
+                rreq->status.MPI_ERROR = MPI_ERR_TRUNCATE;
+            }
+            MPIR_Datatype_release_if_not_builtin(datatype);
+            MPL_free(MPIDI_UCX_REQ(rreq).s.u.pack.pack_buf);
+        } else if (MPIDI_UCX_REQ(rreq).s.type == MPIDI_UCX_RECV_IOV) {
+            MPL_free(MPIDI_UCX_REQ(rreq).s.u.iov.iov);
+        }
 #ifndef MPIDI_CH4_DIRECT_NETMOD
         int is_cancelled;
         MPIDI_anysrc_try_cancel_partner(rreq, &is_cancelled);
@@ -119,12 +158,44 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_UCX_recv(void *buf,
     if (dt_contig) {
         recv_buf = MPIR_get_contig_ptr(buf, dt_true_lb);
         recv_count = data_sz;
+        MPIDI_UCX_REQ(req).s.type = MPIDI_UCX_RECV_CONTIG;
     } else {
-        recv_buf = buf;
-        recv_count = count;
-        param.op_attr_mask |= UCP_OP_ATTR_FIELD_DATATYPE;
-        param.datatype = dt_ptr->dev.netmod.ucx.ucp_datatype;
-        MPIR_Datatype_ptr_add_ref(dt_ptr);
+        if (MPIR_CVAR_UCX_DT_RECV) {
+            recv_buf = buf;
+            recv_count = count;
+            param.op_attr_mask |= UCP_OP_ATTR_FIELD_DATATYPE;
+            param.datatype = dt_ptr->dev.netmod.ucx.ucp_datatype;
+            MPIR_Datatype_ptr_add_ref(dt_ptr);
+            MPIDI_UCX_REQ(req).s.type = MPIDI_UCX_RECV_UCX_DT;
+        } else {
+            MPI_Aint density;
+            MPIR_Datatype_get_density(datatype, density);
+
+            if (MPIDI_UCX_ENABLE_IOV && density >= MPIR_CVAR_CH4_IOV_DENSITY_MIN) {
+                MPI_Aint iov_len, actual_iov_len;
+                MPIR_Typerep_get_iov_len(count, datatype, &iov_len);
+                ucp_dt_iov_t *iov = MPL_malloc(sizeof(ucp_dt_iov_t) * iov_len, MPL_MEM_BUFFER);
+                MPIR_Typerep_to_iov_offset(buf, count, datatype, 0, (struct iovec *) iov,
+                                           iov_len, &actual_iov_len);
+                MPIR_Assert(iov_len == actual_iov_len);
+
+                recv_buf = iov;
+                recv_count = iov_len;
+                param.op_attr_mask |= UCP_OP_ATTR_FIELD_DATATYPE;
+                param.datatype = ucp_dt_make_iov();
+                MPIDI_UCX_REQ(req).s.u.iov.iov = iov;
+                MPIDI_UCX_REQ(req).s.type = MPIDI_UCX_RECV_IOV;
+            } else {
+                MPIDI_UCX_REQ(req).s.u.pack.datatype = datatype;
+                MPIR_Datatype_add_ref_if_not_builtin(datatype);
+                MPIDI_UCX_REQ(req).s.u.pack.user_buf = buf;
+                MPIDI_UCX_REQ(req).s.u.pack.count = count;
+                MPIDI_UCX_REQ(req).s.u.pack.pack_buf = MPL_malloc(data_sz, MPL_MEM_BUFFER);
+                recv_buf = MPIDI_UCX_REQ(req).s.u.pack.pack_buf;
+                recv_count = data_sz;
+                MPIDI_UCX_REQ(req).s.type = MPIDI_UCX_RECV_UNPACK;
+            }
+        }
     }
 
     ucp_request =
@@ -133,7 +204,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_UCX_recv(void *buf,
                                                      ucp_tag, tag_mask, &param);
     MPIDI_UCX_CHK_REQUEST(ucp_request);
 
-    MPIDI_UCX_REQ(req).ucp_request = ucp_request;
+    MPIDI_UCX_REQ(req).s.ucp_request = ucp_request;
     *request = req;
 
   fn_exit:
@@ -197,7 +268,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_imrecv(void *buf,
                                                          &param);
     MPIDI_UCX_CHK_REQUEST(ucp_request);
 
-    MPIDI_UCX_REQ(message).ucp_request = ucp_request;
+    MPIDI_UCX_REQ(message).s.ucp_request = ucp_request;
 
   fn_exit:
     MPIDI_UCX_THREAD_CS_EXIT_VCI(vci);
@@ -252,7 +323,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq, bool
 
     if (!MPIR_Request_is_complete(rreq)) {
         int vci = MPIDI_Request_get_vci(rreq);
-        ucp_request_cancel(MPIDI_UCX_global.ctx[vci].worker, MPIDI_UCX_REQ(rreq).ucp_request);
+        ucp_request_cancel(MPIDI_UCX_global.ctx[vci].worker, MPIDI_UCX_REQ(rreq).s.ucp_request);
     }
 
     MPIR_FUNC_EXIT;

diff --git a/src/mpid/ch4/netmod/ucx/ucx_send.h b/src/mpid/ch4/netmod/ucx/ucx_send.h
@@ -92,7 +92,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_UCX_send(const void *buf,
             MPIR_Request_add_ref(req);
         }
         ucp_request->req = req;
-        MPIDI_UCX_REQ(req).ucp_request = ucp_request;
+        MPIDI_UCX_REQ(req).s.ucp_request = ucp_request;
     } else if (req != NULL) {
         MPIR_cc_set(&req->cc, 0);
     } else if (have_request) {
@@ -160,7 +160,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
     if (!MPIR_Request_is_complete(sreq)) {
         int vci = MPIDI_Request_get_vci(sreq);
         MPIDI_UCX_THREAD_CS_ENTER_VCI(vci);
-        ucp_request_cancel(MPIDI_UCX_global.ctx[vci].worker, MPIDI_UCX_REQ(sreq).ucp_request);
+        ucp_request_cancel(MPIDI_UCX_global.ctx[vci].worker, MPIDI_UCX_REQ(sreq).s.ucp_request);
         MPIDI_UCX_THREAD_CS_EXIT_VCI(vci);
     }
 

diff --git a/src/mpl/configure.ac b/src/mpl/configure.ac
@@ -813,15 +813,7 @@ case $with_proc_mutex_package in
     AC_CHECK_FUNCS(pthread_mutexattr_setpshared)
 
     AC_CHECK_FUNCS(pthread_setaffinity_np)
-
-    #
-    # Check for the Linux functions for controlling processor affinity.
-    #
-    # LINUX: sched_setaffinity
-    # AIX:   bindprocessor
-    # OSX (Leopard): thread_policy_set
-    AC_CHECK_FUNCS([sched_setaffinity sched_getaffinity bindprocessor thread_policy_set])
-    if test "$ac_cv_func_sched_setaffinity" = "yes" ; then
+    if test "$ac_cv_func_pthread_setaffinity_np" = "yes" ; then
         # Test for the cpu process set type
         AC_CACHE_CHECK([whether cpu_set_t available],pac_cv_have_cpu_set_t,[
             AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[#include <sched.h>]],[[cpu_set_t t;]])],
@@ -842,15 +834,9 @@ case $with_proc_mutex_package in
 	    if test "$pac_cv_cpu_set_defined" = "yes" ; then
 	        AC_DEFINE(HAVE_CPU_SET_MACROS,1,[Define if CPU_SET and CPU_ZERO defined])
             fi
-        # FIXME: Some versions of sched_setaffinity return ENOSYS (!),
-	    # so we should test for the unfriendly and useless behavior
         fi
     fi
 
-    if test "$ac_cv_func_pthread_setaffinity_np" = "yes" ; then
-      AC_DEFINE(HAVE_PTHREAD_SETAFFINITY_NP,1, [Define if pthread_setaffinity_np is available.])
-    fi
-
     AC_MSG_NOTICE([POSIX will be used for interprocess mutex package.])
     PROC_MUTEX_PACKAGE_NAME=MPL_PROC_MUTEX_PACKAGE_POSIX
   ;;

diff --git a/src/mpl/src/thread/mpl_thread_posix.c b/src/mpl/src/thread/mpl_thread_posix.c
@@ -86,7 +86,8 @@ void *MPLI_thread_start(void *arg)
 void MPL_thread_set_affinity(MPL_thread_id_t thread, int *affinity_arr, int affinity_size,
                              int *errp)
 {
-#if defined(MPL_HAVE_PTHREAD_SETAFFINITY_NP) && defined(MPL_HAVE_CPU_SET_MACROS)
+#if defined(MPL_HAVE_PTHREAD_SETAFFINITY_NP) && defined(MPL_HAVE_CPU_SET_MACROS) && defined(__linux__)
+    /* FIXME: this implementation uses Linux-specific types and macros */
     int err = MPL_SUCCESS;
     int proc_idx, set_size = 0;
     cpu_set_t cpuset;

diff --git a/src/pm/hydra/lib/tools/bootstrap/external/slurm_query_node_list.c b/src/pm/hydra/lib/tools/bootstrap/external/slurm_query_node_list.c
@@ -26,7 +26,7 @@ static struct HYD_node *global_node_list = NULL;
 #if defined(HAVE_SLURM)
 static HYD_status list_to_nodes(char *str)
 {
-    hostlist_t hostlist;
+    hostlist_t *hostlist;
     char *host;
     int k = 0;
     HYD_status status = HYD_SUCCESS;

diff --git a/src/util/mpir_pmi.c b/src/util/mpir_pmi.c
@@ -114,16 +114,27 @@ static int check_MPIR_CVAR_PMI_VERSION(void)
     return MPI_SUCCESS;
 }
 
+/* When user call MPIR_pmi_finalize, we will postpone the actual PMI_Finalize
+ * and set finalize_pending flag instead. The exit hook will call PMI_Finalize
+ * when the finalize_pending flag is set. This is to prevent an abnormal exit
+ * to show up as a normal exit.
+ */
+static int finalize_pending = 0;
 static void MPIR_pmi_finalize_on_exit(void)
 {
-    SWITCH_PMI(pmi1_exit(), pmi2_exit(), pmix_exit());
+    if (finalize_pending > 0) {
+        SWITCH_PMI(pmi1_exit(), pmi2_exit(), pmix_exit());
+    }
 }
 
 int MPIR_pmi_init(void)
 {
     int mpi_errno = MPI_SUCCESS;
     static bool pmi_connected = false;
 
+    if (finalize_pending > 0) {
+        finalize_pending--;
+    }
     mpi_errno = check_MPIR_CVAR_PMI_VERSION();
     MPIR_ERR_CHECK(mpi_errno);
 
@@ -206,6 +217,9 @@ void MPIR_pmi_finalize(void)
     MPL_free(hwloc_topology_xmlfile);
     hwloc_topology_xmlfile = NULL;
     MPL_free(MPIR_Process.coords);
+
+    /* delay PMI_Finalize to the exit hook */
+    finalize_pending++;
 }
 
 void MPIR_pmi_abort(int exit_code, const char *error_msg)

diff --git a/test/mpi/pt2pt/Makefile.am b/test/mpi/pt2pt/Makefile.am
@@ -46,6 +46,7 @@ noinst_PROGRAMS =  \
     waitany_null   \
     probe_unexp    \
     probenull      \
+    recvnull       \
     inactivereq    \
     waittestnull   \
     sendall        \

diff --git a/test/mpi/pt2pt/recvnull.c b/test/mpi/pt2pt/recvnull.c
@@ -0,0 +1,51 @@
+/*
+ * Copyright (C) by Argonne National Laboratory
+ *     See COPYRIGHT in top-level directory
+ */
+
+#include <stdio.h>
+#include "mpi.h"
+#include "mpitest.h"
+
+/*
+ * This program checks that MPI_Irecv and MPI_Recv correctly handle
+ * a source of MPI_PROC_NULL
+ */
+
+int main(int argc, char **argv)
+{
+    int flag;
+    int errs = 0;
+    MPI_Status status;
+    MPI_Request req;
+
+    MTest_Init(&argc, &argv);
+
+    MPI_Irecv(NULL, 0, MPI_DATATYPE_NULL, MPI_PROC_NULL, 10, MPI_COMM_WORLD, &req);
+    MPI_Waitall(1, &req, &status);
+    if (status.MPI_SOURCE != MPI_PROC_NULL) {
+        printf("Status.MPI_SOURCE was %d, should be MPI_PROC_NULL\n", status.MPI_SOURCE);
+        errs++;
+    }
+    if (status.MPI_TAG != MPI_ANY_TAG) {
+        printf("Status.MPI_TAG was %d, should be MPI_ANY_TAGL\n", status.MPI_TAG);
+        errs++;
+    }
+
+    /* If Irecv failed, Recv is likely to as well.  Avoid a possible hang
+     * by testing Recv only if Irecv test passed */
+    if (errs == 0) {
+        MPI_Recv(NULL, 0, MPI_DATATYPE_NULL, MPI_PROC_NULL, 10, MPI_COMM_WORLD, &status);
+        if (status.MPI_SOURCE != MPI_PROC_NULL) {
+            printf("Status.MPI_SOURCE was %d, should be MPI_PROC_NULL\n", status.MPI_SOURCE);
+            errs++;
+        }
+        if (status.MPI_TAG != MPI_ANY_TAG) {
+            printf("Status.MPI_TAG was %d, should be MPI_ANY_TAGL\n", status.MPI_TAG);
+            errs++;
+        }
+    }
+
+    MTest_Finalize(errs);
+    return MTestReturnValue(errs);
+}