1
1
openmpi/ompi/request/grequest.c
George Bosilca e361bcb64c Send optimizations.
1. The send path get shorter. The BTL is allowed to return > 0 to specify that the
   descriptor was pushed to the networks, and that the memory attached to it is 
   available again for the upper layer. The MCA_BTL_DES_SEND_ALWAYS_CALLBACK flag
   can be used by the PML to force the BTL to always trigger the callback.
   Unmodified BTL will continue to work as expected, as they will return OMPI_SUCCESS
   which force the PML to have exactly the same behavior as before. Some BTLs have
   been modified: self, sm, tcp, mx.
2. Add send immediate interface to BTL.
   The idea is to have a mechanism of allowing the BTL to take advantage of
   send optimizations such as the ability to deliver data "inline". Some
   network APIs such as Portals allow data to be sent using a "thin" event
   without packing data into a memory descriptor. This interface change
   allows the BTL to use such capabilities and allows for other optimizations
   in the future. All existing BTLs except for Portals and sm have this interface
   set to NULL.

This commit was SVN r18551.
2008-05-30 03:58:39 +00:00

228 строки
7.7 KiB
C

/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2005 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "ompi/request/grequest.h"
#include "ompi/mpi/f77/fint_2_int.h"
/*
* See the comment in the grequest destructor for the weird semantics
* here. If the request has been marked complete via a call to
* MPI_GREQUEST_COMPLETE, actually release the object. OTherwise,
* just mark this object as "freed" so that a later call to
* MPI_GREQUEST_COMPLETE will release it (!).
*
* Note that TEST* and WAIT* will call this function when a request
* has been completed.
*/
static int ompi_grequest_free(ompi_request_t** req)
{
OBJ_RELEASE(*req);
*req = MPI_REQUEST_NULL;
return OMPI_SUCCESS;
}
static int ompi_grequest_cancel(ompi_request_t* req, int flag)
{
int rc = OMPI_SUCCESS;
MPI_Fint ierr;
MPI_Flogical fflag;
ompi_grequest_t* greq = (ompi_grequest_t*)req;
if (greq->greq_cancel.c_cancel != NULL) {
if (greq->greq_funcs_are_c) {
rc = greq->greq_cancel.c_cancel(greq->greq_state,
greq->greq_base.req_complete);
} else {
fflag = (MPI_Flogical) greq->greq_base.req_complete;
greq->greq_cancel.f_cancel((MPI_Aint*)greq->greq_state, &fflag, &ierr);
rc = OMPI_FINT_2_INT(ierr);
}
}
return rc;
}
static void ompi_grequest_construct(ompi_grequest_t* greq)
{
greq->greq_base.req_free = ompi_grequest_free;
greq->greq_base.req_cancel = ompi_grequest_cancel;
greq->greq_base.req_type = OMPI_REQUEST_GEN;
greq->greq_base.req_mpi_object.comm = &ompi_mpi_comm_world;
/* Set the function pointers to C here; the F77 MPI API will
override this value if the gen request was created from
Fortran */
greq->greq_funcs_are_c = true;
}
/*
* MPI has some weird semantics with respect to generalized requests
* -- different than all other MPI object types. So we move some
* cleanup stuff here to the destructor rather than in
* greqeust_request_free -- mainly because the cleanup may be required
* in two different places.
*
* Specifically, generalized requests can be completed (and therefore
* released) the following ways:
*
* 1. Call to MPI_GREQUEST_COMPLETE and then a corresponding call to
* some flavor of MPI_TEST* or MPI_WAIT*. This will both complete the
* requests and destroy the coresponding MPI generalized request
* object.
*
* 2. Call MPI_REQUEST_FREE and then (!) -- with some other
* still-valid copy of the handler -- call MPI_GREQUEST_COMPLETE.
*
* 3. Reverse the order of #2 -- call MPI_GREQUEST_COMPLETE and then
* MPI_REQUEST_FREE.
*
* So any one of these functions may actually be the one that
* de-allocates the back-end request object. Hence, this is perfect
* for our reference counting system -- so the call to the gen request
* free_fn() is back here in the destructor, whenever the object is
* actually freed.
*
* Hence, the following must occur before a grequest is freed:
*
* - ompi_grequest_complete() (i.e., GREQUEST_COMPLETE) is invoked
* - ompi_grequest_free() is invoked
*
* Remember that ompi_grequest_free() is invoked by MPI_TEST* and
* MPI_WAIT* when the request was previously marked as complete and
* TEST* / WAIT* notified the user as such, and this function is also
* invoked by REQUEST_FREE). Hence, these two functions will *always*
* be invoked, but the order in which they are invoked is up to the
* user. So this is a perfect opprotunity for the OBJ_* reference
* count system. When we create an ompi_grequest_t in
* ompi_grequest_start(), we both OBJ_NEW and OBJ_RETAIN it so that
* its reference count goes to 0. Then in ompi_grequest_complete()
* and ompi_grequest_free(), we OBJ_RELEASE it. Hence, when both of
* them have RELEASEd -- regardless of the order in which the
* functions were invoked, then the destructor is invoked and
* everything is cleaned up (and we invoked the grequest free_fn).
*/
static void ompi_grequest_destruct(ompi_grequest_t* greq)
{
MPI_Fint ierr;
if (greq->greq_free.c_free != NULL) {
if (greq->greq_funcs_are_c) {
greq->greq_free.c_free(greq->greq_state);
} else {
greq->greq_free.f_free((MPI_Aint*)greq->greq_state, &ierr);
}
}
OMPI_REQUEST_FINI(&greq->greq_base);
}
OBJ_CLASS_INSTANCE(
ompi_grequest_t,
ompi_request_t,
ompi_grequest_construct,
ompi_grequest_destruct);
int ompi_grequest_start(
MPI_Grequest_query_function *gquery_fn,
MPI_Grequest_free_function *gfree_fn,
MPI_Grequest_cancel_function *gcancel_fn,
void* gstate,
ompi_request_t** request)
{
ompi_grequest_t *greq = OBJ_NEW(ompi_grequest_t);
if(greq == NULL) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* We call RETAIN here specifically to increase the refcount to 2.
See comment before the destructor for an explanation. */
OBJ_RETAIN(greq);
greq->greq_base.req_state = OMPI_REQUEST_ACTIVE;
greq->greq_state = gstate;
greq->greq_query.c_query = gquery_fn;
greq->greq_free.c_free = gfree_fn;
greq->greq_cancel.c_cancel = gcancel_fn;
greq->greq_base.req_status = ompi_status_empty;
*request = &greq->greq_base;
return OMPI_SUCCESS;
}
/*
* Beware the odd semantics listed in MPI-2:8.2... See the comment in
* the grequest destructor.
*
* First do the normal stuff to complete the request (i.e., call
* ompi_request_complete()). Then, if this request object was
* previously freed via MPI_REQUEST_FREE, release it.
*/
int ompi_grequest_complete(ompi_request_t *req)
{
int rc;
OPAL_THREAD_LOCK(&ompi_request_lock);
rc = ompi_request_complete(req, true);
OPAL_THREAD_UNLOCK(&ompi_request_lock);
OBJ_RELEASE(req);
return rc;
}
/*
* Grequest queries are invoked in two places:
*
* 1. MPI_TEST* / MPI_WAIT*, when requests have completed.
*
* 2. MPI_REQUEST_GET_STATUS, when requests may or may not have
* completed.
*
*/
int ompi_grequest_invoke_query(ompi_request_t *request,
ompi_status_public_t *status)
{
int rc;
ompi_grequest_t *g = (ompi_grequest_t*) request;
/* MPI-2:8.2 does not say what to do with the return value from
the query function (i.e., the int return value from the C
function or the ierr argument from the Fortran function).
Making the command decision here to ignore it. If the handler
wants to pass an error back, it should set it in the MPI_ERROR
field in the status (which is always kept, regardless if the
top-level function was invoked with MPI_STATUS[ES]_IGNORE or
not). */
if (NULL != g->greq_query.c_query) {
if (g->greq_funcs_are_c) {
rc = g->greq_query.c_query(g->greq_state, status);
} else {
MPI_Fint ierr;
MPI_Fint fstatus[sizeof(MPI_Status) / sizeof(int)];
g->greq_query.f_query((MPI_Aint*)g->greq_state, fstatus, &ierr);
MPI_Status_f2c(fstatus, status);
rc = OMPI_FINT_2_INT(ierr);
}
}
return OMPI_SUCCESS;
}