Blob Blame History Raw
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
/*
 *  (C) 2001 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */
/* style: allow:fprintf:1 sig:0 */

#include "mpiimpl.h"
#include "mpi_init.h"

/*
=== BEGIN_MPI_T_CVAR_INFO_BLOCK ===

categories:
    - name        : DEVELOPER
      description : useful for developers working on MPICH itself

cvars:
    - name        : MPIR_CVAR_MEMDUMP
      category    : DEVELOPER
      type        : boolean
      default     : true
      class       : device
      verbosity   : MPI_T_VERBOSITY_MPIDEV_DETAIL
      scope       : MPI_T_SCOPE_LOCAL
      description : >-
        If true, list any memory that was allocated by MPICH and that
        remains allocated when MPI_Finalize completes.

    - name        : MPIR_CVAR_MEM_CATEGORY_INFORMATION
      category    : DEVELOPER
      type        : boolean
      default     : false
      class       : device
      verbosity   : MPI_T_VERBOSITY_MPIDEV_DETAIL
      scope       : MPI_T_SCOPE_LOCAL
      description : >-
        If true, print a summary of memory allocation by category. The category
        definitions are found in mpl_trmem.h.

=== END_MPI_T_CVAR_INFO_BLOCK ===
*/

/* -- Begin Profiling Symbol Block for routine MPI_Finalize */
#if defined(HAVE_PRAGMA_WEAK)
#pragma weak MPI_Finalize = PMPI_Finalize
#elif defined(HAVE_PRAGMA_HP_SEC_DEF)
#pragma _HP_SECONDARY_DEF PMPI_Finalize  MPI_Finalize
#elif defined(HAVE_PRAGMA_CRI_DUP)
#pragma _CRI duplicate MPI_Finalize as PMPI_Finalize
#elif defined(HAVE_WEAK_ATTRIBUTE)
int MPI_Finalize(void) __attribute__ ((weak, alias("PMPI_Finalize")));
#endif
/* -- End Profiling Symbol Block */

/* Define MPICH_MPI_FROM_PMPI if weak symbols are not supported to build
   the MPI routines */
#ifndef MPICH_MPI_FROM_PMPI
#undef MPI_Finalize
#define MPI_Finalize PMPI_Finalize

/* Any internal routines can go here.  Make them static if possible */

/* The following routines provide a callback facility for modules that need
   some code called on exit.  This method allows us to avoid forcing
   MPI_Finalize to know the routine names a priori.  Any module that wants to
   have a callback calls MPIR_Add_finalize(routine, extra, priority).

 */
PMPI_LOCAL void MPIR_Call_finalize_callbacks(int, int);
typedef struct Finalize_func_t {
    int (*f) (void *);          /* The function to call */
    void *extra_data;           /* Data for the function */
    int priority;               /* priority is used to control the order
                                 * in which the callbacks are invoked */
} Finalize_func_t;
/* When full debugging is enabled, each MPI handle type has a finalize handler
   installed to detect unfreed handles.  */
#define MAX_FINALIZE_FUNC 64
static Finalize_func_t fstack[MAX_FINALIZE_FUNC];
static int fstack_sp = 0;
static int fstack_max_priority = 0;

void MPIR_Add_finalize(int (*f) (void *), void *extra_data, int priority)
{
    /* --BEGIN ERROR HANDLING-- */
    if (fstack_sp >= MAX_FINALIZE_FUNC) {
        /* This is a little tricky.  We may want to check the state of
         * MPIR_Process.mpich_state to decide how to signal the error */
        (void) MPL_internal_error_printf("overflow in finalize stack! "
                                         "Is MAX_FINALIZE_FUNC too small?\n");
        if (OPA_load_int(&MPIR_Process.mpich_state) == MPICH_MPI_STATE__IN_INIT ||
            OPA_load_int(&MPIR_Process.mpich_state) == MPICH_MPI_STATE__POST_INIT) {
            MPID_Abort(NULL, MPI_SUCCESS, 13, NULL);
        } else {
            exit(1);
        }
    }
    /* --END ERROR HANDLING-- */
    fstack[fstack_sp].f = f;
    fstack[fstack_sp].priority = priority;
    fstack[fstack_sp++].extra_data = extra_data;

    if (priority > fstack_max_priority)
        fstack_max_priority = priority;
}

/* Invoke the registered callbacks */
PMPI_LOCAL void MPIR_Call_finalize_callbacks(int min_prio, int max_prio)
{
    int i, j;

    if (max_prio > fstack_max_priority)
        max_prio = fstack_max_priority;
    for (j = max_prio; j >= min_prio; j--) {
        for (i = fstack_sp - 1; i >= 0; i--) {
            if (fstack[i].f && fstack[i].priority == j) {
                fstack[i].f(fstack[i].extra_data);
                fstack[i].f = 0;
            }
        }
    }
}
#else
#ifndef USE_WEAK_SYMBOLS
PMPI_LOCAL void MPIR_Call_finalize_callbacks(int, int);
#endif
#endif

#undef FUNCNAME
#define FUNCNAME MPI_Finalize
#undef FCNAME
#define FCNAME MPL_QUOTE(FUNCNAME)
/*@
   MPI_Finalize - Terminates MPI execution environment

   Notes:
   All processes must call this routine before exiting.  The number of
   processes running `after` this routine is called is undefined;
   it is best not to perform much more than a 'return rc' after calling
   'MPI_Finalize'.

Thread and Signal Safety:
The MPI standard requires that 'MPI_Finalize' be called `only` by the same
thread that initialized MPI with either 'MPI_Init' or 'MPI_Init_thread'.

.N Fortran

.N Errors
.N MPI_SUCCESS
@*/
int MPI_Finalize(void)
{
    int mpi_errno = MPI_SUCCESS;
#if defined(HAVE_USLEEP) && defined(USE_COVERAGE)
    int rank = 0;
#endif
    MPIR_FUNC_TERSE_FINALIZE_STATE_DECL(MPID_STATE_MPI_FINALIZE);

    MPIR_ERRTEST_INITIALIZED_ORDIE();

#ifdef HAVE_HWLOC
    hwloc_topology_destroy(MPIR_Process.hwloc_topology);
    hwloc_bitmap_free(MPIR_Process.bindset);
#endif

#ifdef HAVE_NETLOC
    switch (MPIR_Process.network_attr.type) {
        case MPIR_NETLOC_NETWORK_TYPE__TORUS:
            if (MPIR_Process.network_attr.u.torus.geometry != NULL)
                MPL_free(MPIR_Process.network_attr.u.torus.geometry);
            break;
        case MPIR_NETLOC_NETWORK_TYPE__FAT_TREE:
        case MPIR_NETLOC_NETWORK_TYPE__CLOS_NETWORK:
        default:
            if (MPIR_Process.network_attr.u.tree.node_levels != NULL)
                MPL_free(MPIR_Process.network_attr.u.tree.node_levels);
            break;
    }
#endif

    /* Note: Only one thread may ever call MPI_Finalize (MPI_Finalize may
     * be called at most once in any program) */
    MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
    MPIR_FUNC_TERSE_FINALIZE_ENTER(MPID_STATE_MPI_FINALIZE);

    /* ... body of routine ... */

    /* If the user requested for asynchronous progress, we need to
     * shutdown the progress thread */
    if (MPIR_async_thread_initialized) {
        mpi_errno = MPID_Finalize_async_thread();
        if (mpi_errno)
            goto fn_fail;
    }
#if defined(HAVE_USLEEP) && defined(USE_COVERAGE)
    /* We need to get the rank before freeing MPI_COMM_WORLD */
    rank = MPIR_Process.comm_world->rank;
#endif

    /* Remove the attributes, executing the attribute delete routine.
     * Do this only if the attribute functions are defined. */
    /* The standard (MPI-2, section 4.8) says that the attributes on
     * MPI_COMM_SELF are deleted before almost anything else happens */
    /* Note that the attributes need to be removed from the communicators
     * so that they aren't freed twice. (The communicators are released
     * in MPID_Finalize) */
    if (MPIR_Process.attr_free && MPIR_Process.comm_self->attributes) {
        mpi_errno = MPIR_Process.attr_free(MPI_COMM_SELF, &MPIR_Process.comm_self->attributes);
        if (mpi_errno)
            MPIR_ERR_POP(mpi_errno);
        MPIR_Process.comm_self->attributes = 0;
    }
    if (MPIR_Process.attr_free && MPIR_Process.comm_world->attributes) {
        mpi_errno = MPIR_Process.attr_free(MPI_COMM_WORLD, &MPIR_Process.comm_world->attributes);
        if (mpi_errno)
            MPIR_ERR_POP(mpi_errno);
        MPIR_Process.comm_world->attributes = 0;
    }

    /*
     * Now that we're finalizing, we need to take control of the error handlers
     * At this point, we will release any user-defined error handlers on
     * comm self and comm world
     */
    if (MPIR_Process.comm_world->errhandler &&
        !(HANDLE_GET_KIND(MPIR_Process.comm_world->errhandler->handle) == HANDLE_KIND_BUILTIN)) {
        int in_use;
        MPIR_Errhandler_release_ref(MPIR_Process.comm_world->errhandler, &in_use);
        if (!in_use) {
            MPIR_Handle_obj_free(&MPIR_Errhandler_mem, MPIR_Process.comm_world->errhandler);
        }
        /* always set to NULL to avoid a double-release later in finalize */
        MPIR_Process.comm_world->errhandler = NULL;
    }
    if (MPIR_Process.comm_self->errhandler &&
        !(HANDLE_GET_KIND(MPIR_Process.comm_self->errhandler->handle) == HANDLE_KIND_BUILTIN)) {
        int in_use;
        MPIR_Errhandler_release_ref(MPIR_Process.comm_self->errhandler, &in_use);
        if (!in_use) {
            MPIR_Handle_obj_free(&MPIR_Errhandler_mem, MPIR_Process.comm_self->errhandler);
        }
        /* always set to NULL to avoid a double-release later in finalize */
        MPIR_Process.comm_self->errhandler = NULL;
    }

    /* FIXME: Why is this not one of the finalize callbacks?.  Do we need
     * pre and post MPID_Finalize callbacks? */
    MPII_Timer_finalize();

    /* Call the high-priority callbacks */
    MPIR_Call_finalize_callbacks(MPIR_FINALIZE_CALLBACK_PRIO + 1, MPIR_FINALIZE_CALLBACK_MAX_PRIO);

    /* Signal the debugger that we are about to exit. */
    /* FIXME: Should this also be a finalize callback? */
#ifdef HAVE_DEBUGGER_SUPPORT
    MPIR_Debugger_set_aborting((char *) 0);
#endif

    mpi_errno = MPID_Finalize();
    if (mpi_errno) {
        MPIR_ERR_POP(mpi_errno);
    }

    /* Free complete request */
    MPIR_Request_free(MPIR_Process.lw_req);

    mpi_errno = MPII_Coll_finalize();
    if (mpi_errno)
        MPIR_ERR_POP(mpi_errno);

    /* Call the low-priority (post Finalize) callbacks */
    MPIR_Call_finalize_callbacks(0, MPIR_FINALIZE_CALLBACK_PRIO - 1);

    /* At this point, if there has been a failure, exit before
     * completing the finalize */
    if (mpi_errno != MPI_SUCCESS)
        goto fn_fail;

    /* Users did not call MPI_T_init_thread(), so we free memories allocated to
     * MPIR_T during MPI_Init here. Otherwise, free them in MPI_T_finalize() */
    if (!MPIR_T_is_initialized())
        MPIR_T_env_finalize();

    /* FIXME: Many of these debugging items could/should be callbacks,
     * added to the finalize callback list */
    /* FIXME: the memory tracing code block should be a finalize callback */
    /* If memory debugging is enabled, check the memory here, after all
     * finalize callbacks */

    MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
    OPA_store_int(&MPIR_Process.mpich_state, MPICH_MPI_STATE__POST_FINALIZED);

#if defined(MPICH_IS_THREADED)
    MPIR_Thread_CS_Finalize();
#endif

    /* We place the memory tracing at the very end because any of the other
     * steps may have allocated memory that they still need to release */
#ifdef USE_MEMORY_TRACING
    /* FIXME: We'd like to arrange for the mem dump output to
     * go to separate files or to be sorted by rank (note that
     * the rank is at the head of the line) */
    {
        if (MPIR_CVAR_MEMDUMP) {
            /* The second argument is the min id to print; memory allocated
             * after MPI_Init is given an id of one.  This allows us to
             * ignore, if desired, memory leaks in the MPID_Init call */
            MPL_trdump((void *) 0, -1);
        }
        if (MPIR_CVAR_MEM_CATEGORY_INFORMATION)
            MPL_trcategorydump(stderr);
    }
#endif

#if defined(HAVE_USLEEP) && defined(USE_COVERAGE)
    /* If performing coverage analysis, make each process sleep for
     * rank * 100 ms, to give time for the coverage tool to write out
     * any files.  It would be better if the coverage tool and runtime
     * was more careful about file updates, though the lack of OS support
     * for atomic file updates makes this harder. */
    /*
     * On some systems, a 0.1 second delay appears to be too short for
     * the file system.  This code allows the use of the environment
     * variable MPICH_FINALDELAY, which is the delay in milliseconds.
     * It must be an integer value.
     */
    {
        int microseconds = 100000;
        char *delayStr = getenv("MPICH_FINALDELAY");
        if (delayStr) {
            /* Because this is a maintainer item, we won't check for
             * errors in the delayStr */
            microseconds = 1000 * atoi(delayStr);
        }
        usleep(rank * microseconds);
    }
#endif

    /* ... end of body of routine ... */
  fn_exit:
    MPIR_FUNC_TERSE_FINALIZE_EXIT(MPID_STATE_MPI_FINALIZE);
    return mpi_errno;

  fn_fail:
    /* --BEGIN ERROR HANDLING-- */
#ifdef HAVE_ERROR_CHECKING
    {
        mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE,
                                         FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_finalize", 0);
    }
#endif
    mpi_errno = MPIR_Err_return_comm(0, FCNAME, mpi_errno);
    if (OPA_load_int(&MPIR_Process.mpich_state) < MPICH_MPI_STATE__POST_FINALIZED) {
        MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
    }
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}