ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

A more consistent version. As we now share the send and receive queue, we

Просмотр исходного кода 
have to construct/destruct only once. Therefore, the construction will
happens before digging for a PML, while the destruction just before
finalizing the component.

Add some OPAL_LIKELY/OPAL_UNLIKELY.

This commit was SVN r15347.
Этот коммит содержится в:
George Bosilca 2007-07-10 23:45:23 +00:00
родитель 433f8a7694
Коммит e19777e910
15 изменённых файлов: 752 добавлений и 848 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

41
ompi/mca/pml/base/pml_base_close.c
Просмотреть файл

@ -25,32 +25,41 @@
#include "opal/mca/base/base.h" #include "opal/mca/base/base.h"
#include "ompi/mca/pml/pml.h" #include "ompi/mca/pml/pml.h"
#include "ompi/mca/pml/base/base.h" #include "ompi/mca/pml/base/base.h"
#include "ompi/mca/pml/base/pml_base_request.h"
#include "opal/runtime/opal_progress.h" #include "opal/runtime/opal_progress.h"
int mca_pml_base_close(void) int mca_pml_base_close(void)
{ {
/* turn off the progress code for the pml */ /* turn off the progress code for the pml */
opal_progress_unregister(mca_pml.pml_progress); opal_progress_unregister(mca_pml.pml_progress);
/* Blatently ignore the return code (what would we do to recover, /* Blatently ignore the return code (what would we do to recover,
anyway? This module is going away, so errors don't matter anyway? This module is going away, so errors don't matter
anymore) */ anymore) */
mca_pml.pml_progress = mca_pml_base_progress; /**
if (NULL != mca_pml_base_selected_component.pmlm_finalize) { * Destruct the send and receive queues. The ompi_free_list_t destructor
mca_pml_base_selected_component.pmlm_finalize(); * will return the memory to the mpool, so this has to be done before the
} * mpool get released by the PML close function.
*/
OBJ_DESTRUCT(&mca_pml_base_send_requests);
OBJ_DESTRUCT(&mca_pml_base_recv_requests);
OBJ_DESTRUCT(&mca_pml_base_pml); mca_pml.pml_progress = mca_pml_base_progress;
if (NULL != mca_pml_base_selected_component.pmlm_finalize) {
mca_pml_base_selected_component.pmlm_finalize();
}
/* Close all remaining available modules (may be one if this is a OBJ_DESTRUCT(&mca_pml_base_pml);
OMPI RTE program, or [possibly] multiple if this is ompi_info) */
mca_base_components_close(mca_pml_base_output, /* Close all remaining available modules (may be one if this is a
&mca_pml_base_components_available, NULL); OMPI RTE program, or [possibly] multiple if this is ompi_info) */
/* All done */ mca_base_components_close(mca_pml_base_output,
&mca_pml_base_components_available, NULL);
return OMPI_SUCCESS; /* All done */
return OMPI_SUCCESS;
} }

12
ompi/mca/pml/base/pml_base_open.c
Просмотреть файл

@ -29,6 +29,7 @@
#include "ompi/constants.h" #include "ompi/constants.h"
#include "ompi/mca/pml/pml.h" #include "ompi/mca/pml/pml.h"
#include "ompi/mca/pml/base/base.h" #include "ompi/mca/pml/base/base.h"
#include "ompi/mca/pml/base/pml_base_request.h"
/* /*
* The following file was created by configure. It contains extern * The following file was created by configure. It contains extern
@ -130,7 +131,6 @@ int mca_pml_base_open(void)
stringify(MCA_pml_DIRECT_CALL_COMPONENT)); stringify(MCA_pml_DIRECT_CALL_COMPONENT));
#else #else
mca_base_param_reg_string_name("pml", NULL, mca_base_param_reg_string_name("pml", NULL,
"Specify a specific PML to use", "Specify a specific PML to use",
false, false, "", &default_pml); false, false, "", &default_pml);
@ -159,6 +159,16 @@ int mca_pml_base_open(void)
#endif #endif
/**
* Construct the send and receive request queues. There are 2 reasons to do it
* here. First, as they are globals it's better to construct them in one common
* place. Second, in order to be able to allow the external debuggers to show
* their content, they should get constructed as soon as possible once the MPI
* process is started.
*/
OBJ_CONSTRUCT(&mca_pml_base_send_requests, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_pml_base_recv_requests, ompi_free_list_t);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }

18
ompi/mca/pml/cm/pml_cm.c
Просмотреть файл

@ -56,6 +56,24 @@ mca_pml_cm_enable(bool enable)
{ {
/* BWB - FIX ME - need to have this actually do something, /* BWB - FIX ME - need to have this actually do something,
maybe? */ maybe? */
ompi_free_list_init(&mca_pml_base_send_requests,
MAX(sizeof(mca_pml_cm_thin_send_request_t),
sizeof(mca_pml_cm_hvy_send_request_t)) + ompi_mtl->mtl_request_size,
OBJ_CLASS(mca_pml_cm_hvy_send_request_t),
ompi_pml_cm.free_list_num,
ompi_pml_cm.free_list_max,
ompi_pml_cm.free_list_inc,
NULL);
ompi_free_list_init(&mca_pml_base_send_requests,
MAX(sizeof(mca_pml_cm_thin_recv_request_t),
sizeof(mca_pml_cm_hvy_recv_request_t)) + ompi_mtl->mtl_request_size,
OBJ_CLASS(mca_pml_cm_hvy_recv_request_t),
ompi_pml_cm.free_list_num,
ompi_pml_cm.free_list_max,
ompi_pml_cm.free_list_inc,
NULL);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }

4
ompi/mca/pml/cm/pml_cm.h
Просмотреть файл

@ -34,6 +34,10 @@ struct mca_mtl_request_t;
struct ompi_pml_cm_t { struct ompi_pml_cm_t {
mca_pml_base_module_t super; mca_pml_base_module_t super;
int free_list_num;
int free_list_max;
int free_list_inc;
int default_priority;
}; };
typedef struct ompi_pml_cm_t ompi_pml_cm_t; typedef struct ompi_pml_cm_t ompi_pml_cm_t;
extern ompi_pml_cm_t ompi_pml_cm; extern ompi_pml_cm_t ompi_pml_cm;

40
ompi/mca/pml/cm/pml_cm_component.c
Просмотреть файл

@ -64,11 +64,6 @@ mca_pml_base_component_1_0_0_t mca_pml_cm_component = {
mca_pml_cm_component_fini /* component finalize */ mca_pml_cm_component_fini /* component finalize */
}; };
static int free_list_num = 0;
static int free_list_max = 0;
static int free_list_inc = 0;
static int default_priority = 30;
static int static int
mca_pml_cm_component_open(void) mca_pml_cm_component_open(void)
{ {
@ -83,7 +78,7 @@ mca_pml_cm_component_open(void)
false, false,
false, false,
4, 4,
&free_list_num); &ompi_pml_cm.free_list_num);
mca_base_param_reg_int(&mca_pml_cm_component.pmlm_version, mca_base_param_reg_int(&mca_pml_cm_component.pmlm_version,
"free_list_max", "free_list_max",
@ -91,7 +86,7 @@ mca_pml_cm_component_open(void)
false, false,
false, false,
-1, -1,
&free_list_max); &ompi_pml_cm.free_list_max);
mca_base_param_reg_int(&mca_pml_cm_component.pmlm_version, mca_base_param_reg_int(&mca_pml_cm_component.pmlm_version,
"free_list_inc", "free_list_inc",
@ -99,7 +94,7 @@ mca_pml_cm_component_open(void)
false, false,
false, false,
64, 64,
&free_list_inc); &ompi_pml_cm.free_list_inc);
mca_base_param_reg_int(&mca_pml_cm_component.pmlm_version, mca_base_param_reg_int(&mca_pml_cm_component.pmlm_version,
"priority", "priority",
@ -107,7 +102,7 @@ mca_pml_cm_component_open(void)
false, false,
false, false,
30, 30,
&default_priority); &ompi_pml_cm.default_priority);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -126,11 +121,12 @@ mca_pml_cm_component_init(int* priority,
bool enable_mpi_threads) bool enable_mpi_threads)
{ {
int ret; int ret;
if((*priority) > default_priority) {
*priority = default_priority; if((*priority) > ompi_pml_cm.default_priority) {
*priority = ompi_pml_cm.default_priority;
return NULL; return NULL;
} }
*priority = default_priority; *priority = ompi_pml_cm.default_priority;
opal_output_verbose( 10, 0, opal_output_verbose( 10, 0,
"in cm pml priority is %d\n", *priority); "in cm pml priority is %d\n", *priority);
/* find a useable MTL */ /* find a useable MTL */
@ -152,26 +148,6 @@ mca_pml_cm_component_init(int* priority,
ompi_pml_cm.super.pml_max_contextid = ompi_mtl->mtl_max_contextid; ompi_pml_cm.super.pml_max_contextid = ompi_mtl->mtl_max_contextid;
ompi_pml_cm.super.pml_max_tag = ompi_mtl->mtl_max_tag; ompi_pml_cm.super.pml_max_tag = ompi_mtl->mtl_max_tag;
OBJ_CONSTRUCT(&mca_pml_base_send_requests, ompi_free_list_t);
ompi_free_list_init(&mca_pml_base_send_requests,
MAX(sizeof(mca_pml_cm_thin_send_request_t),
sizeof(mca_pml_cm_hvy_send_request_t)) + ompi_mtl->mtl_request_size,
OBJ_CLASS(mca_pml_cm_hvy_send_request_t),
free_list_num,
free_list_max,
free_list_inc,
NULL);
OBJ_CONSTRUCT(&mca_pml_base_recv_requests, ompi_free_list_t);
ompi_free_list_init(&mca_pml_base_send_requests,
MAX(sizeof(mca_pml_cm_thin_recv_request_t),
sizeof(mca_pml_cm_hvy_recv_request_t)) + ompi_mtl->mtl_request_size,
OBJ_CLASS(mca_pml_cm_hvy_recv_request_t),
free_list_num,
free_list_max,
free_list_inc,
NULL);
/* initialize buffered send code */ /* initialize buffered send code */
if(OMPI_SUCCESS != mca_pml_base_bsend_init(enable_mpi_threads)) { if(OMPI_SUCCESS != mca_pml_base_bsend_init(enable_mpi_threads)) {
opal_output(0, "mca_pml_cm_component_init: mca_pml_bsend_init failed\n"); opal_output(0, "mca_pml_cm_component_init: mca_pml_bsend_init failed\n");

60
ompi/mca/pml/dr/pml_dr.c
Просмотреть файл

@ -62,13 +62,56 @@ mca_pml_dr_t mca_pml_dr = {
} }
}; };
void mca_pml_dr_error_handler( void mca_pml_dr_error_handler( struct mca_btl_base_module_t* btl,
struct mca_btl_base_module_t* btl, int32_t flags );
int32_t flags);
int mca_pml_dr_enable(bool enable) int mca_pml_dr_enable(bool enable)
{ {
if( false == enable ) return OMPI_SUCCESS; if( false == enable ) return OMPI_SUCCESS;
/* requests */
ompi_free_list_init( &mca_pml_base_send_requests,
sizeof(mca_pml_dr_send_request_t),
OBJ_CLASS(mca_pml_dr_send_request_t),
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL );
ompi_free_list_init( &mca_pml_base_recv_requests,
sizeof(mca_pml_dr_recv_request_t),
OBJ_CLASS(mca_pml_dr_recv_request_t),
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL );
/* fragments */
OBJ_CONSTRUCT(&mca_pml_dr.recv_frags, ompi_free_list_t);
ompi_free_list_init( &mca_pml_dr.recv_frags,
sizeof(mca_pml_dr_recv_frag_t),
OBJ_CLASS(mca_pml_dr_recv_frag_t),
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL );
OBJ_CONSTRUCT(&mca_pml_dr.vfrags, ompi_free_list_t);
ompi_free_list_init( &mca_pml_dr.vfrags,
sizeof(mca_pml_dr_vfrag_t),
OBJ_CLASS(mca_pml_dr_vfrag_t),
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL );
OBJ_CONSTRUCT(&mca_pml_dr.send_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.send_active, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.acks_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.buffers, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.endpoints, ompi_pointer_array_t);
OBJ_CONSTRUCT(&mca_pml_dr.lock, opal_mutex_t);
mca_pml_dr.enabled = true; mca_pml_dr.enabled = true;
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -234,7 +277,16 @@ int mca_pml_dr_del_procs(ompi_proc_t** procs, size_t nprocs)
int mca_pml_dr_component_fini(void) int mca_pml_dr_component_fini(void)
{ {
/* FIX */ if(!mca_pml_dr.enabled)
return OMPI_SUCCESS; /* never selected.. return success.. */
mca_pml_dr.enabled = false; /* not anymore */
OBJ_DESTRUCT(&mca_pml_dr.send_pending);
OBJ_DESTRUCT(&mca_pml_dr.send_active);
OBJ_DESTRUCT(&mca_pml_dr.acks_pending);
OBJ_DESTRUCT(&mca_pml_dr.recv_frags);
OBJ_DESTRUCT(&mca_pml_dr.buffers);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }

63
ompi/mca/pml/dr/pml_dr_component.c
Просмотреть файл

@ -113,58 +113,8 @@ int mca_pml_dr_component_open(void)
mca_pml_dr.enable_csum = mca_pml_dr.enable_csum =
mca_pml_dr_param_register_int("enable_csum", 1); mca_pml_dr_param_register_int("enable_csum", 1);
/* requests */
OBJ_CONSTRUCT(&mca_pml_base_send_requests, ompi_free_list_t);
ompi_free_list_init(
&mca_pml_base_send_requests,
sizeof(mca_pml_dr_send_request_t),
OBJ_CLASS(mca_pml_dr_send_request_t),
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL);
OBJ_CONSTRUCT(&mca_pml_base_recv_requests, ompi_free_list_t);
ompi_free_list_init(
&mca_pml_base_recv_requests,
sizeof(mca_pml_dr_recv_request_t),
OBJ_CLASS(mca_pml_dr_recv_request_t),
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL);
/* fragments */
OBJ_CONSTRUCT(&mca_pml_dr.recv_frags, ompi_free_list_t);
ompi_free_list_init(
&mca_pml_dr.recv_frags,
sizeof(mca_pml_dr_recv_frag_t),
OBJ_CLASS(mca_pml_dr_recv_frag_t),
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL);
OBJ_CONSTRUCT(&mca_pml_dr.vfrags, ompi_free_list_t);
ompi_free_list_init(
&mca_pml_dr.vfrags,
sizeof(mca_pml_dr_vfrag_t),
OBJ_CLASS(mca_pml_dr_vfrag_t),
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL);
OBJ_CONSTRUCT(&mca_pml_dr.send_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.send_active, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.acks_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.buffers, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.endpoints, ompi_pointer_array_t);
OBJ_CONSTRUCT(&mca_pml_dr.lock, opal_mutex_t);
mca_pml_dr.enabled = false; mca_pml_dr.enabled = false;
return mca_bml_base_open(); return mca_bml_base_open();
} }
@ -172,19 +122,9 @@ int mca_pml_dr_component_close(void)
{ {
int rc; int rc;
if(!mca_pml_dr.enabled)
return OMPI_SUCCESS; /* never selected.. return success.. */
if(OMPI_SUCCESS != (rc = mca_bml_base_close())) if(OMPI_SUCCESS != (rc = mca_bml_base_close()))
return rc; return rc;
OBJ_DESTRUCT(&mca_pml_dr.send_pending);
OBJ_DESTRUCT(&mca_pml_dr.send_active);
OBJ_DESTRUCT(&mca_pml_dr.acks_pending);
OBJ_DESTRUCT(&mca_pml_base_recv_requests);
OBJ_DESTRUCT(&mca_pml_base_send_requests);
OBJ_DESTRUCT(&mca_pml_dr.recv_frags);
OBJ_DESTRUCT(&mca_pml_dr.buffers);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -207,8 +147,7 @@ mca_pml_base_module_t* mca_pml_dr_component_init(int* priority,
if(OMPI_SUCCESS != mca_bml_base_init( enable_progress_threads, if(OMPI_SUCCESS != mca_bml_base_init( enable_progress_threads,
enable_mpi_threads enable_mpi_threads )) {
)) {
return NULL; return NULL;
} }

155
ompi/mca/pml/ob1/pml_ob1.c
Просмотреть файл

@ -2,7 +2,7 @@
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana * Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
* University Research and Technology * University Research and Technology
* Corporation. All rights reserved. * Corporation. All rights reserved.
* Copyright (c) 2004-2005 The University of Tennessee and The University * Copyright (c) 2004-2007 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights * of Tennessee Research Foundation. All rights
* reserved. * reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
@ -46,36 +46,106 @@
mca_pml_ob1_t mca_pml_ob1 = { mca_pml_ob1_t mca_pml_ob1 = {
{ {
mca_pml_ob1_add_procs, mca_pml_ob1_add_procs,
mca_pml_ob1_del_procs, mca_pml_ob1_del_procs,
mca_pml_ob1_enable, mca_pml_ob1_enable,
mca_pml_ob1_progress, mca_pml_ob1_progress,
mca_pml_ob1_add_comm, mca_pml_ob1_add_comm,
mca_pml_ob1_del_comm, mca_pml_ob1_del_comm,
mca_pml_ob1_irecv_init, mca_pml_ob1_irecv_init,
mca_pml_ob1_irecv, mca_pml_ob1_irecv,
mca_pml_ob1_recv, mca_pml_ob1_recv,
mca_pml_ob1_isend_init, mca_pml_ob1_isend_init,
mca_pml_ob1_isend, mca_pml_ob1_isend,
mca_pml_ob1_send, mca_pml_ob1_send,
mca_pml_ob1_iprobe, mca_pml_ob1_iprobe,
mca_pml_ob1_probe, mca_pml_ob1_probe,
mca_pml_ob1_start, mca_pml_ob1_start,
mca_pml_ob1_dump, mca_pml_ob1_dump,
mca_pml_ob1_ft_event, mca_pml_ob1_ft_event,
32768, 32768,
INT_MAX INT_MAX
} }
}; };
void mca_pml_ob1_error_handler( void mca_pml_ob1_error_handler( struct mca_btl_base_module_t* btl,
struct mca_btl_base_module_t* btl, int32_t flags );
int32_t flags);
int mca_pml_ob1_enable(bool enable) int mca_pml_ob1_enable(bool enable)
{ {
if( false == enable ) return OMPI_SUCCESS; if( false == enable ) return OMPI_SUCCESS;
OBJ_CONSTRUCT(&mca_pml_ob1.lock, opal_mutex_t);
/* fragments */
OBJ_CONSTRUCT(&mca_pml_ob1.rdma_frags, ompi_free_list_t);
ompi_free_list_init( &mca_pml_ob1.rdma_frags,
sizeof(mca_pml_ob1_rdma_frag_t),
OBJ_CLASS(mca_pml_ob1_rdma_frag_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL );
OBJ_CONSTRUCT(&mca_pml_ob1.recv_frags, ompi_free_list_t);
ompi_free_list_init( &mca_pml_ob1.recv_frags,
sizeof(mca_pml_ob1_recv_frag_t) + mca_pml_ob1.unexpected_limit,
OBJ_CLASS(mca_pml_ob1_recv_frag_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL );
OBJ_CONSTRUCT(&mca_pml_ob1.pending_pckts, ompi_free_list_t);
ompi_free_list_init( &mca_pml_ob1.pending_pckts,
sizeof(mca_pml_ob1_pckt_pending_t),
OBJ_CLASS(mca_pml_ob1_pckt_pending_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL );
OBJ_CONSTRUCT(&mca_pml_ob1.buffers, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.send_ranges, ompi_free_list_t);
ompi_free_list_init( &mca_pml_ob1.send_ranges,
sizeof(mca_pml_ob1_send_range_t) +
(mca_pml_ob1.max_send_per_range - 1) * sizeof(mca_pml_ob1_com_btl_t),
OBJ_CLASS(mca_pml_ob1_send_range_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL );
/* pending operations */
OBJ_CONSTRUCT(&mca_pml_ob1.send_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.recv_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.pckt_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.rdma_pending, opal_list_t);
/**
* If we get here this is the PML who get selected for the run. We
* should get ownership for the send and receive requests list, and
* initialize them with the size of our own requests.
*/
ompi_free_list_init( &mca_pml_base_send_requests,
sizeof(mca_pml_ob1_send_request_t),
OBJ_CLASS(mca_pml_ob1_send_request_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL );
ompi_free_list_init( &mca_pml_base_recv_requests,
sizeof(mca_pml_ob1_recv_request_t),
OBJ_CLASS(mca_pml_ob1_recv_request_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL );
mca_pml_ob1.enabled = true; mca_pml_ob1.enabled = true;
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -192,24 +262,6 @@ int mca_pml_ob1_del_procs(ompi_proc_t** procs, size_t nprocs)
return mca_bml.bml_del_procs(nprocs, procs); return mca_bml.bml_del_procs(nprocs, procs);
} }
int mca_pml_ob1_component_fini(void)
{
int rc;
/* Shutdown BML */
if(OMPI_SUCCESS != (rc = mca_bml.bml_finalize()))
return rc;
/* Shutdown buffered send */
if(OMPI_SUCCESS != (rc = mca_pml_base_bsend_fini())) {
return rc;
}
/* FIX */
return OMPI_SUCCESS;
}
/* /*
* diagnostics * diagnostics
*/ */
@ -237,11 +289,10 @@ int mca_pml_ob1_dump(struct ompi_communicator_t* comm, int verbose)
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
static void mca_pml_ob1_fin_completion( static void mca_pml_ob1_fin_completion( mca_btl_base_module_t* btl,
mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* des,
struct mca_btl_base_descriptor_t* des, int status )
int status)
{ {
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context;
@ -252,13 +303,11 @@ static void mca_pml_ob1_fin_completion(
MCA_PML_OB1_PROGRESS_PENDING(bml_btl); MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
} }
int mca_pml_ob1_send_fin( int mca_pml_ob1_send_fin( ompi_proc_t* proc,
ompi_proc_t* proc, mca_bml_base_btl_t* bml_btl,
mca_bml_base_btl_t* bml_btl, void *hdr_des,
void *hdr_des, uint8_t order,
uint8_t order, uint32_t status )
uint32_t status
)
{ {
mca_btl_base_descriptor_t* fin; mca_btl_base_descriptor_t* fin;
mca_pml_ob1_fin_hdr_t* hdr; mca_pml_ob1_fin_hdr_t* hdr;

189
ompi/mca/pml/ob1/pml_ob1.h
Просмотреть файл

@ -37,9 +37,8 @@
#include "ompi/proc/proc.h" #include "ompi/proc/proc.h"
#include "ompi/mca/allocator/base/base.h" #include "ompi/mca/allocator/base/base.h"
#if defined(c_plusplus) || defined(__cplusplus) BEGIN_C_DECLS
extern "C" {
#endif
/** /**
* OB1 PML module * OB1 PML module
*/ */
@ -84,25 +83,6 @@ typedef struct mca_pml_ob1_t mca_pml_ob1_t;
extern mca_pml_ob1_t mca_pml_ob1; extern mca_pml_ob1_t mca_pml_ob1;
/*
* PML module functions.
*/
extern int mca_pml_ob1_component_open(void);
extern int mca_pml_ob1_component_close(void);
extern mca_pml_base_module_t* mca_pml_ob1_component_init(
int *priority,
bool enable_progress_threads,
bool enable_mpi_threads
);
extern int mca_pml_ob1_component_fini(void);
/* /*
* PML interface functions. * PML interface functions.
*/ */
@ -125,106 +105,80 @@ extern int mca_pml_ob1_del_procs(
size_t nprocs size_t nprocs
); );
extern int mca_pml_ob1_enable( extern int mca_pml_ob1_enable( bool enable );
bool enable
);
extern int mca_pml_ob1_progress(void); extern int mca_pml_ob1_progress(void);
extern int mca_pml_ob1_iprobe( extern int mca_pml_ob1_iprobe( int dst,
int dst, int tag,
int tag, struct ompi_communicator_t* comm,
struct ompi_communicator_t* comm, int *matched,
int *matched, ompi_status_public_t* status );
ompi_status_public_t* status
);
extern int mca_pml_ob1_probe( extern int mca_pml_ob1_probe( int dst,
int dst, int tag,
int tag, struct ompi_communicator_t* comm,
struct ompi_communicator_t* comm, ompi_status_public_t* status );
ompi_status_public_t* status
);
extern int mca_pml_ob1_isend_init( extern int mca_pml_ob1_isend_init( void *buf,
void *buf, size_t count,
size_t count, ompi_datatype_t *datatype,
ompi_datatype_t *datatype, int dst,
int dst, int tag,
int tag, mca_pml_base_send_mode_t mode,
mca_pml_base_send_mode_t mode, struct ompi_communicator_t* comm,
struct ompi_communicator_t* comm, struct ompi_request_t **request );
struct ompi_request_t **request
);
extern int mca_pml_ob1_isend( extern int mca_pml_ob1_isend( void *buf,
void *buf, size_t count,
size_t count, ompi_datatype_t *datatype,
ompi_datatype_t *datatype, int dst,
int dst, int tag,
int tag, mca_pml_base_send_mode_t mode,
mca_pml_base_send_mode_t mode, struct ompi_communicator_t* comm,
struct ompi_communicator_t* comm, struct ompi_request_t **request );
struct ompi_request_t **request
);
extern int mca_pml_ob1_send( extern int mca_pml_ob1_send( void *buf,
void *buf, size_t count,
size_t count, ompi_datatype_t *datatype,
ompi_datatype_t *datatype, int dst,
int dst, int tag,
int tag, mca_pml_base_send_mode_t mode,
mca_pml_base_send_mode_t mode, struct ompi_communicator_t* comm );
struct ompi_communicator_t* comm
);
extern int mca_pml_ob1_irecv_init( extern int mca_pml_ob1_irecv_init( void *buf,
void *buf, size_t count,
size_t count, ompi_datatype_t *datatype,
ompi_datatype_t *datatype, int src,
int src, int tag,
int tag, struct ompi_communicator_t* comm,
struct ompi_communicator_t* comm, struct ompi_request_t **request );
struct ompi_request_t **request
);
extern int mca_pml_ob1_irecv( extern int mca_pml_ob1_irecv( void *buf,
void *buf, size_t count,
size_t count, ompi_datatype_t *datatype,
ompi_datatype_t *datatype, int src,
int src, int tag,
int tag, struct ompi_communicator_t* comm,
struct ompi_communicator_t* comm, struct ompi_request_t **request );
struct ompi_request_t **request
);
extern int mca_pml_ob1_recv( extern int mca_pml_ob1_recv( void *buf,
void *buf, size_t count,
size_t count, ompi_datatype_t *datatype,
ompi_datatype_t *datatype, int src,
int src, int tag,
int tag, struct ompi_communicator_t* comm,
struct ompi_communicator_t* comm, ompi_status_public_t* status );
ompi_status_public_t* status
);
extern int mca_pml_ob1_dump( extern int mca_pml_ob1_dump( struct ompi_communicator_t* comm,
struct ompi_communicator_t* comm, int verbose );
int verbose
);
extern int mca_pml_ob1_start( extern int mca_pml_ob1_start( size_t count,
size_t count, ompi_request_t** requests );
ompi_request_t** requests
);
extern int mca_pml_ob1_ft_event( extern int mca_pml_ob1_ft_event( int state );
int state
);
#if defined(c_plusplus) || defined(__cplusplus) END_C_DECLS
}
#endif
#define MCA_PML_OB1_DES_ALLOC(bml_btl, des, order, size) \ #define MCA_PML_OB1_DES_ALLOC(bml_btl, des, order, size) \
MCA_BML_BASE_BTL_DES_ALLOC(bml_btl, des, order, \ MCA_BML_BASE_BTL_DES_ALLOC(bml_btl, des, order, \
@ -300,6 +254,7 @@ void mca_pml_ob1_process_pending_rdma(void);
if(opal_list_get_size(&mca_pml_ob1.rdma_pending)) \ if(opal_list_get_size(&mca_pml_ob1.rdma_pending)) \
mca_pml_ob1_process_pending_rdma(); \ mca_pml_ob1_process_pending_rdma(); \
} while (0) } while (0)
/* /*
* Compute the total number of bytes on supplied descriptor * Compute the total number of bytes on supplied descriptor
*/ */
@ -325,28 +280,28 @@ int mca_pml_ob1_com_btl_comp(const void *v1, const void *v2);
/* Calculate what percentage of a message to send through each BTL according to /* Calculate what percentage of a message to send through each BTL according to
* relative weight */ * relative weight */
static inline void mca_pml_ob1_calc_weighted_length( static inline void
mca_pml_ob1_com_btl_t *btls, int num_btls, size_t size, mca_pml_ob1_calc_weighted_length( mca_pml_ob1_com_btl_t *btls, int num_btls, size_t size,
double weight_total) double weight_total )
{ {
int i; int i;
size_t length_left = size; size_t length_left;
/* shortcut for common case for only one BTL */ /* shortcut for common case for only one BTL */
if(num_btls == 1) { if( OPAL_LIKELY(1 == num_btls) ) {
btls[0].length = size; btls[0].length = size;
return; return;
} }
/* sort BTLs according of their weights so BTLs with smaller weight will /* sort BTLs according of their weights so BTLs with smaller weight will
* not hijack all of the traffic */ * not hijack all of the traffic */
qsort(btls, num_btls, sizeof(mca_pml_ob1_com_btl_t), qsort( btls, num_btls, sizeof(mca_pml_ob1_com_btl_t),
mca_pml_ob1_com_btl_comp); mca_pml_ob1_com_btl_comp );
for(i = 0; i < num_btls; i++) { for(length_left = size, i = 0; i < num_btls; i++) {
mca_bml_base_btl_t* bml_btl = btls[i].bml_btl; mca_bml_base_btl_t* bml_btl = btls[i].bml_btl;
size_t length = 0; size_t length = 0;
if(length_left != 0) { if( OPAL_UNLIKELY(0 != length_left) ) {
length = (length_left > bml_btl->btl_eager_limit)? length = (length_left > bml_btl->btl_eager_limit)?
((size_t)(size * (bml_btl->btl_weight / weight_total))) : ((size_t)(size * (bml_btl->btl_weight / weight_total))) :
length_left; length_left;

199
ompi/mca/pml/ob1/pml_ob1_component.c
Просмотреть файл

@ -35,12 +35,17 @@
#include "pml_ob1_component.h" #include "pml_ob1_component.h"
#include "ompi/mca/allocator/base/base.h" #include "ompi/mca/allocator/base/base.h"
OBJ_CLASS_INSTANCE( OBJ_CLASS_INSTANCE( mca_pml_ob1_pckt_pending_t,
mca_pml_ob1_pckt_pending_t, ompi_free_list_item_t,
ompi_free_list_item_t, NULL,
NULL, NULL );
NULL
); static int mca_pml_ob1_component_open(void);
static int mca_pml_ob1_component_close(void);
static mca_pml_base_module_t*
mca_pml_ob1_component_init( int* priority, bool enable_progress_threads,
bool enable_mpi_threads );
static int mca_pml_ob1_component_fini(void);
mca_pml_base_component_1_0_0_t mca_pml_ob1_component = { mca_pml_base_component_1_0_0_t mca_pml_ob1_component = {
@ -90,11 +95,10 @@ static inline int mca_pml_ob1_param_register_int(
return param_value; return param_value;
} }
static int mca_pml_ob1_component_open(void)
int mca_pml_ob1_component_open(void)
{ {
mca_allocator_base_component_t* allocator_component; mca_allocator_base_component_t* allocator_component;
mca_pml_ob1.free_list_num = mca_pml_ob1.free_list_num =
mca_pml_ob1_param_register_int("free_list_num", 4); mca_pml_ob1_param_register_int("free_list_num", 4);
mca_pml_ob1.free_list_max = mca_pml_ob1.free_list_max =
@ -126,75 +130,19 @@ int mca_pml_ob1_component_open(void)
"bucket", "bucket",
&mca_pml_ob1.allocator_name); &mca_pml_ob1.allocator_name);
allocator_component = mca_allocator_component_lookup( mca_pml_ob1.allocator_name ); allocator_component = mca_allocator_component_lookup( mca_pml_ob1.allocator_name );
if(NULL == allocator_component) { if(NULL == allocator_component) {
opal_output(0, "mca_pml_ob1_component_open: can't find allocator: %s\n", mca_pml_ob1.allocator_name); opal_output(0, "mca_pml_ob1_component_open: can't find allocator: %s\n", mca_pml_ob1.allocator_name);
return OMPI_ERROR; return OMPI_ERROR;
} }
mca_pml_ob1.allocator = allocator_component->allocator_init(true, mca_pml_ob1.allocator = allocator_component->allocator_init(true,
mca_pml_ob1_seg_alloc, mca_pml_ob1_seg_free, NULL); mca_pml_ob1_seg_alloc,
mca_pml_ob1_seg_free, NULL);
if(NULL == mca_pml_ob1.allocator) { if(NULL == mca_pml_ob1.allocator) {
opal_output(0, "mca_pml_ob1_component_open: unable to initialize allocator\n"); opal_output(0, "mca_pml_ob1_component_open: unable to initialize allocator\n");
return OMPI_ERROR; return OMPI_ERROR;
} }
OBJ_CONSTRUCT(&mca_pml_ob1.lock, opal_mutex_t);
/* fragments */
OBJ_CONSTRUCT(&mca_pml_ob1.rdma_frags, ompi_free_list_t);
ompi_free_list_init(
&mca_pml_ob1.rdma_frags,
sizeof(mca_pml_ob1_rdma_frag_t),
OBJ_CLASS(mca_pml_ob1_rdma_frag_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL);
OBJ_CONSTRUCT(&mca_pml_ob1.recv_frags, ompi_free_list_t);
ompi_free_list_init(
&mca_pml_ob1.recv_frags,
sizeof(mca_pml_ob1_recv_frag_t) + mca_pml_ob1.unexpected_limit,
OBJ_CLASS(mca_pml_ob1_recv_frag_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL);
OBJ_CONSTRUCT(&mca_pml_ob1.pending_pckts, ompi_free_list_t);
ompi_free_list_init(
&mca_pml_ob1.pending_pckts,
sizeof(mca_pml_ob1_pckt_pending_t),
OBJ_CLASS(mca_pml_ob1_pckt_pending_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL);
OBJ_CONSTRUCT(&mca_pml_ob1.buffers, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.send_ranges, ompi_free_list_t);
ompi_free_list_init(
&mca_pml_ob1.send_ranges,
sizeof(mca_pml_ob1_send_range_t) +
(mca_pml_ob1.max_send_per_range - 1) * sizeof(mca_pml_ob1_com_btl_t),
OBJ_CLASS(mca_pml_ob1_send_range_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL);
/* pending operations */
OBJ_CONSTRUCT(&mca_pml_ob1.send_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.recv_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.pckt_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_ob1.rdma_pending, opal_list_t);
mca_pml_ob1.leave_pinned = ompi_mpi_leave_pinned; mca_pml_ob1.leave_pinned = ompi_mpi_leave_pinned;
mca_pml_ob1.leave_pinned_pipeline = (int) ompi_mpi_leave_pinned_pipeline; mca_pml_ob1.leave_pinned_pipeline = (int) ompi_mpi_leave_pinned_pipeline;
@ -204,55 +152,21 @@ int mca_pml_ob1_component_open(void)
} }
int mca_pml_ob1_component_close(void) static int mca_pml_ob1_component_close(void)
{ {
int rc; int rc;
if(!mca_pml_ob1.enabled)
return OMPI_SUCCESS; /* never selected.. return success.. */
if(OMPI_SUCCESS != (rc = mca_bml_base_close())) if(OMPI_SUCCESS != (rc = mca_bml_base_close()))
return rc; return rc;
OBJ_DESTRUCT(&mca_pml_ob1.rdma_pending);
OBJ_DESTRUCT(&mca_pml_ob1.pckt_pending);
OBJ_DESTRUCT(&mca_pml_ob1.recv_pending);
OBJ_DESTRUCT(&mca_pml_ob1.send_pending);
OBJ_DESTRUCT(&mca_pml_ob1.buffers);
OBJ_DESTRUCT(&mca_pml_ob1.pending_pckts);
OBJ_DESTRUCT(&mca_pml_ob1.recv_frags);
OBJ_DESTRUCT(&mca_pml_ob1.rdma_frags);
OBJ_DESTRUCT(&mca_pml_ob1.lock);
/* destroy the global free lists */
OBJ_DESTRUCT(&mca_pml_base_send_requests);
OBJ_DESTRUCT(&mca_pml_base_recv_requests);
if(OMPI_SUCCESS != (rc = mca_pml_ob1.allocator->alc_finalize(mca_pml_ob1.allocator))) {
return rc;
}
#if 0
if (mca_pml_ob1.send_requests.fl_num_allocated !=
mca_pml_ob1.send_requests.super.opal_list_length) {
opal_output(0, "ob1 send requests: %d allocated %d returned\n",
mca_pml_ob1.send_requests.fl_num_allocated,
mca_pml_ob1.send_requests.super.opal_list_length);
}
if (mca_pml_ob1.recv_requests.fl_num_allocated !=
mca_pml_ob1.recv_requests.super.opal_list_length) {
opal_output(0, "ob1 recv requests: %d allocated %d returned\n",
mca_pml_ob1.recv_requests.fl_num_allocated,
mca_pml_ob1.recv_requests.super.opal_list_length);
}
#endif
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
mca_pml_base_module_t* mca_pml_ob1_component_init(int* priority, static mca_pml_base_module_t*
bool enable_progress_threads, mca_pml_ob1_component_init( int* priority,
bool enable_mpi_threads) bool enable_progress_threads,
bool enable_mpi_threads )
{ {
opal_output_verbose( 10, 0, opal_output_verbose( 10, 0,
"in ob1, my priority is %d\n", mca_pml_ob1.priority); "in ob1, my priority is %d\n", mca_pml_ob1.priority);
@ -282,32 +196,59 @@ mca_pml_base_module_t* mca_pml_ob1_component_init(int* priority,
*/ */
mca_pml_ob1.super.pml_progress = mca_bml.bml_progress; mca_pml_ob1.super.pml_progress = mca_bml.bml_progress;
/**
* If we get here this is the PML who get selected for the run. We
* should get ownership for the send and receive requests list, and
* initialize them with the size of our own requests.
*/
OBJ_CONSTRUCT(&mca_pml_base_send_requests, ompi_free_list_t);
ompi_free_list_init( &mca_pml_base_send_requests,
sizeof(mca_pml_ob1_send_request_t),
OBJ_CLASS(mca_pml_ob1_send_request_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL );
OBJ_CONSTRUCT(&mca_pml_base_recv_requests, ompi_free_list_t);
ompi_free_list_init( &mca_pml_base_recv_requests,
sizeof(mca_pml_ob1_recv_request_t),
OBJ_CLASS(mca_pml_ob1_recv_request_t),
mca_pml_ob1.free_list_num,
mca_pml_ob1.free_list_max,
mca_pml_ob1.free_list_inc,
NULL );
return &mca_pml_ob1.super; return &mca_pml_ob1.super;
} }
int mca_pml_ob1_component_fini(void)
{
int rc;
/* Shutdown BML */
if(OMPI_SUCCESS != (rc = mca_bml.bml_finalize()))
return rc;
/* Shutdown buffered send */
if(OMPI_SUCCESS != (rc = mca_pml_base_bsend_fini())) {
return rc;
}
if(!mca_pml_ob1.enabled)
return OMPI_SUCCESS; /* never selected.. return success.. */
mca_pml_ob1.enabled = false; /* not anymore */
OBJ_DESTRUCT(&mca_pml_ob1.rdma_pending);
OBJ_DESTRUCT(&mca_pml_ob1.pckt_pending);
OBJ_DESTRUCT(&mca_pml_ob1.recv_pending);
OBJ_DESTRUCT(&mca_pml_ob1.send_pending);
OBJ_DESTRUCT(&mca_pml_ob1.buffers);
OBJ_DESTRUCT(&mca_pml_ob1.pending_pckts);
OBJ_DESTRUCT(&mca_pml_ob1.recv_frags);
OBJ_DESTRUCT(&mca_pml_ob1.rdma_frags);
OBJ_DESTRUCT(&mca_pml_ob1.lock);
if(OMPI_SUCCESS != (rc = mca_pml_ob1.allocator->alc_finalize(mca_pml_ob1.allocator))) {
return rc;
}
#if 0
if (mca_pml_base_send_requests.fl_num_allocated !=
mca_pml_base_send_requests.super.opal_list_length) {
opal_output(0, "ob1 send requests: %d allocated %d returned\n",
mca_pml_base_send_requests.fl_num_allocated,
mca_pml_base_send_requests.super.opal_list_length);
}
if (mca_pml_base_recv_requests.fl_num_allocated !=
mca_pml_base_recv_requests.super.opal_list_length) {
opal_output(0, "ob1 recv requests: %d allocated %d returned\n",
mca_pml_base_recv_requests.fl_num_allocated,
mca_pml_base_recv_requests.super.opal_list_length);
}
#endif
return OMPI_SUCCESS;
}
void *mca_pml_ob1_seg_alloc( struct mca_mpool_base_module_t* mpool, void *mca_pml_ob1_seg_alloc( struct mca_mpool_base_module_t* mpool,
size_t* size, size_t* size,
mca_mpool_base_registration_t** registration) { mca_mpool_base_registration_t** registration) {

22
ompi/mca/pml/ob1/pml_ob1_isend.c
Просмотреть файл

@ -2,7 +2,7 @@
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology * University Research and Technology
* Corporation. All rights reserved. * Corporation. All rights reserved.
* Copyright (c) 2004-2005 The University of Tennessee and The University * Copyright (c) 2004-2007 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights * of Tennessee Research Foundation. All rights
* reserved. * reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
@ -42,11 +42,11 @@ int mca_pml_ob1_isend_init(void *buf,
return rc; return rc;
MCA_PML_OB1_SEND_REQUEST_INIT(sendreq, MCA_PML_OB1_SEND_REQUEST_INIT(sendreq,
buf, buf,
count, count,
datatype, datatype,
dst, tag, dst, tag,
comm, sendmode, true); comm, sendmode, true);
*request = (ompi_request_t *) sendreq; *request = (ompi_request_t *) sendreq;
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -69,11 +69,11 @@ int mca_pml_ob1_isend(void *buf,
return rc; return rc;
MCA_PML_OB1_SEND_REQUEST_INIT(sendreq, MCA_PML_OB1_SEND_REQUEST_INIT(sendreq,
buf, buf,
count, count,
datatype, datatype,
dst, tag, dst, tag,
comm, sendmode, false); comm, sendmode, false);
MCA_PML_OB1_SEND_REQUEST_START(sendreq, rc); MCA_PML_OB1_SEND_REQUEST_START(sendreq, rc);
*request = (ompi_request_t *) sendreq; *request = (ompi_request_t *) sendreq;

320
ompi/mca/pml/ob1/pml_ob1_recvfrag.c
Просмотреть файл

@ -37,19 +37,15 @@
#include "ompi/datatype/dt_arch.h" #include "ompi/datatype/dt_arch.h"
#include "ompi/peruse/peruse-internal.h" #include "ompi/peruse/peruse-internal.h"
OBJ_CLASS_INSTANCE( OBJ_CLASS_INSTANCE( mca_pml_ob1_buffer_t,
mca_pml_ob1_buffer_t, ompi_free_list_item_t,
ompi_free_list_item_t, NULL,
NULL, NULL );
NULL
);
OBJ_CLASS_INSTANCE( OBJ_CLASS_INSTANCE( mca_pml_ob1_recv_frag_t,
mca_pml_ob1_recv_frag_t, opal_list_item_t,
opal_list_item_t, NULL,
NULL, NULL );
NULL
);
/** /**
* Static functions. * Static functions.
@ -74,15 +70,15 @@ static int mca_pml_ob1_recv_frag_match( mca_btl_base_module_t *btl,
* Callback from BTL on receive. * Callback from BTL on receive.
*/ */
void mca_pml_ob1_recv_frag_callback( void mca_pml_ob1_recv_frag_callback( mca_btl_base_module_t* btl,
mca_btl_base_module_t* btl, mca_btl_base_tag_t tag,
mca_btl_base_tag_t tag, mca_btl_base_descriptor_t* des,
mca_btl_base_descriptor_t* des, void* cbdata )
void* cbdata)
{ {
mca_btl_base_segment_t* segments = des->des_dst; mca_btl_base_segment_t* segments = des->des_dst;
mca_pml_ob1_hdr_t* hdr = (mca_pml_ob1_hdr_t*)segments->seg_addr.pval; mca_pml_ob1_hdr_t* hdr = (mca_pml_ob1_hdr_t*)segments->seg_addr.pval;
if(segments->seg_len < sizeof(mca_pml_ob1_common_hdr_t)) {
if( OPAL_UNLIKELY(segments->seg_len < sizeof(mca_pml_ob1_common_hdr_t)) ) {
return; return;
} }
@ -234,7 +230,7 @@ void mca_pml_ob1_recv_frag_callback(
((opal_list_item_t *)generic_recv)->opal_list_next) { \ ((opal_list_item_t *)generic_recv)->opal_list_next) { \
/* Check for a match */ \ /* Check for a match */ \
recv_tag = generic_recv->req_recv.req_base.req_tag; \ recv_tag = generic_recv->req_recv.req_base.req_tag; \
if ( frag_tag == recv_tag ) { \ if( OPAL_UNLIKELY(frag_tag == recv_tag) ) { \
break; \ break; \
} \ } \
} \ } \
@ -267,12 +263,12 @@ void mca_pml_ob1_recv_frag_callback(
* set by the upper level routine. * set by the upper level routine.
*/ */
#define MCA_PML_OB1_CHECK_WILD_RECEIVES_FOR_MATCH(hdr,comm,proc,return_match) \ #define MCA_PML_OB1_CHECK_WILD_RECEIVES_FOR_MATCH(hdr,comm,proc,return_match) \
do { \ do { \
/* local parameters */ \ /* local parameters */ \
opal_list_t* wild_receives = &comm->wild_receives; \ opal_list_t* wild_receives = &comm->wild_receives; \
MCA_PML_OB1_MATCH_GENERIC_RECEIVES(hdr,wild_receives,proc,return_match); \ MCA_PML_OB1_MATCH_GENERIC_RECEIVES(hdr,wild_receives,proc,return_match); \
} while(0) } while(0)
/** /**
@ -289,12 +285,12 @@ do { \
* This routine assumes that the appropriate matching locks are * This routine assumes that the appropriate matching locks are
* set by the upper level routine. * set by the upper level routine.
*/ */
#define MCA_PML_OB1_CHECK_SPECIFIC_RECEIVES_FOR_MATCH(hdr,comm,proc,return_match) \ #define MCA_PML_OB1_CHECK_SPECIFIC_RECEIVES_FOR_MATCH(hdr,comm,proc,return_match) \
do { \ do { \
/* local variables */ \ /* local variables */ \
opal_list_t* specific_receives = &proc->specific_receives; \ opal_list_t* specific_receives = &proc->specific_receives; \
MCA_PML_OB1_MATCH_GENERIC_RECEIVES(hdr,specific_receives,proc,return_match); \ MCA_PML_OB1_MATCH_GENERIC_RECEIVES(hdr,specific_receives,proc,return_match); \
} while(0) } while(0)
/** /**
* Try and match the incoming message fragment to the list of * Try and match the incoming message fragment to the list of
@ -313,137 +309,120 @@ do { \
* set by the upper level routine. * set by the upper level routine.
*/ */
#define MCA_PML_OB1_CHECK_SPECIFIC_AND_WILD_RECEIVES_FOR_MATCH( \ #define MCA_PML_OB1_CHECK_SPECIFIC_AND_WILD_RECEIVES_FOR_MATCH( hdr,comm,proc,return_match) \
hdr,comm,proc,return_match) \ do { \
do { \ /* local variables */ \
/* local variables */ \ mca_pml_ob1_recv_request_t *specific_recv, *wild_recv; \
mca_pml_ob1_recv_request_t *specific_recv, *wild_recv; \ mca_pml_sequence_t wild_recv_seq, specific_recv_seq; \
mca_pml_sequence_t wild_recv_seq, specific_recv_seq; \ int frag_tag, wild_recv_tag, specific_recv_tag; \
int frag_tag, wild_recv_tag, specific_recv_tag; \ \
\ /* initialization */ \
/* initialization */ \ frag_tag=hdr->hdr_tag; \
frag_tag=hdr->hdr_tag; \ \
\ /* \
/* \ * We know that when this is called, both specific and wild irecvs \
* We know that when this is called, both specific and wild irecvs \ * have been posted. \
* have been posted. \ */ \
*/ \ specific_recv = (mca_pml_ob1_recv_request_t *) \
specific_recv = (mca_pml_ob1_recv_request_t *) \ opal_list_get_first(&(proc)->specific_receives); \
opal_list_get_first(&(proc)->specific_receives); \ wild_recv = (mca_pml_ob1_recv_request_t *) \
wild_recv = (mca_pml_ob1_recv_request_t *) \ opal_list_get_first(&comm->wild_receives); \
opal_list_get_first(&comm->wild_receives); \ \
\ specific_recv_seq = specific_recv->req_recv.req_base.req_sequence; \
specific_recv_seq = specific_recv->req_recv.req_base.req_sequence; \ wild_recv_seq = wild_recv->req_recv.req_base.req_sequence; \
wild_recv_seq = wild_recv->req_recv.req_base.req_sequence; \ \
\ while (true) { \
while (true) { \ if (wild_recv_seq < specific_recv_seq) { \
if (wild_recv_seq < specific_recv_seq) { \ /* wild recv is earlier than the specific one. */ \
/* \ /* try and match */ \
* wild recv is earlier than the specific one. \ wild_recv_tag = wild_recv->req_recv.req_base.req_tag; \
*/ \ if ( (frag_tag == wild_recv_tag) || \
/* \ ( (wild_recv_tag == OMPI_ANY_TAG) && (0 <= frag_tag) ) ) { \
* try and match \ /* Match made */ \
*/ \ return_match=wild_recv; \
wild_recv_tag = wild_recv->req_recv.req_base.req_tag; \ \
if ( (frag_tag == wild_recv_tag) || \ /* remove this recv from the wild receive queue */ \
( (wild_recv_tag == OMPI_ANY_TAG) && (0 <= frag_tag) ) ) { \ opal_list_remove_item(&comm->wild_receives, \
/* \ (opal_list_item_t *)wild_recv); \
* Match made \ \
*/ \ PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_REMOVE_FROM_POSTED_Q, \
return_match=wild_recv; \ &(wild_recv->req_recv.req_base), \
\ PERUSE_RECV); \
/* remove this recv from the wild receive queue */ \ \
opal_list_remove_item(&comm->wild_receives, \ break; \
(opal_list_item_t *)wild_recv); \ } \
\ \
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_REMOVE_FROM_POSTED_Q, \ /* No match, go to the next */ \
&(wild_recv->req_recv.req_base), \ wild_recv=(mca_pml_ob1_recv_request_t *) \
PERUSE_RECV); \ ((opal_list_item_t *)wild_recv)->opal_list_next; \
\ \
break; \ /* \
} \ * If that was the last wild one, just look at the \
\ * rest of the specific ones. \
/* \ */ \
* No match, go to the next. \ if (wild_recv == (mca_pml_ob1_recv_request_t *) \
*/ \ opal_list_get_end(&comm->wild_receives) ) \
wild_recv=(mca_pml_ob1_recv_request_t *) \ { \
((opal_list_item_t *)wild_recv)->opal_list_next; \ MCA_PML_OB1_CHECK_SPECIFIC_RECEIVES_FOR_MATCH(hdr, comm, proc, return_match); \
\ break; \
/* \ } \
* If that was the last wild one, just look at the \ \
* rest of the specific ones. \ /* \
*/ \ * Get the sequence number for this recv, and go \
if (wild_recv == (mca_pml_ob1_recv_request_t *) \ * back to the top of the loop. \
opal_list_get_end(&comm->wild_receives) ) \ */ \
{ \ wild_recv_seq = wild_recv->req_recv.req_base.req_sequence; \
MCA_PML_OB1_CHECK_SPECIFIC_RECEIVES_FOR_MATCH(hdr, comm, proc, return_match); \ \
break; \ } else { \
} \ /* specific recv is earlier than the wild one. */ \
\ specific_recv_tag=specific_recv->req_recv.req_base.req_tag; \
/* \ if ( (frag_tag == specific_recv_tag) || \
* Get the sequence number for this recv, and go \ ( (specific_recv_tag == OMPI_ANY_TAG) && (0<=frag_tag)) ) \
* back to the top of the loop. \ { \
*/ \ /* Match made */ \
wild_recv_seq = wild_recv->req_recv.req_base.req_sequence; \ return_match = specific_recv; \
\ /* remove descriptor from specific receive list */ \
} else { \ opal_list_remove_item(&(proc)->specific_receives, \
/* \ (opal_list_item_t *)specific_recv); \
* specific recv is earlier than the wild one. \ \
*/ \ PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_REMOVE_FROM_POSTED_Q, \
specific_recv_tag=specific_recv->req_recv.req_base.req_tag; \ &(specific_recv->req_recv.req_base), \
if ( (frag_tag == specific_recv_tag) || \ PERUSE_RECV); \
( (specific_recv_tag == OMPI_ANY_TAG) && (0<=frag_tag)) ) \ \
{ \ break; \
/* \ } \
* Match made \ \
*/ \ /* No match, go on to the next specific irecv. */ \
return_match = specific_recv; \ specific_recv = (mca_pml_ob1_recv_request_t *) \
/* remove descriptor from specific receive list */ \ ((opal_list_item_t *)specific_recv)->opal_list_next; \
opal_list_remove_item(&(proc)->specific_receives, \ \
(opal_list_item_t *)specific_recv); \ /* \
\ * If that was the last specific irecv, process the \
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_REMOVE_FROM_POSTED_Q, \ * rest of the wild ones. \
&(specific_recv->req_recv.req_base), \ */ \
PERUSE_RECV); \ if (specific_recv == (mca_pml_ob1_recv_request_t *) \
\ opal_list_get_end(&(proc)->specific_receives)) \
break; \ { \
} \ MCA_PML_OB1_CHECK_WILD_RECEIVES_FOR_MATCH(hdr, comm, proc, return_match); \
\ break; \
/* \ } \
* No match, go on to the next specific irecv. \ /* \
*/ \ * Get the sequence number for this recv, and go \
specific_recv = (mca_pml_ob1_recv_request_t *) \ * back to the top of the loop. \
((opal_list_item_t *)specific_recv)->opal_list_next; \ */ \
\ specific_recv_seq = specific_recv->req_recv.req_base.req_sequence; \
/* \ } \
* If that was the last specific irecv, process the \ } \
* rest of the wild ones. \ } while(0)
*/ \
if (specific_recv == (mca_pml_ob1_recv_request_t *) \
opal_list_get_end(&(proc)->specific_receives)) \
{ \
MCA_PML_OB1_CHECK_WILD_RECEIVES_FOR_MATCH(hdr, comm, proc, return_match); \
break; \
} \
/* \
* Get the sequence number for this recv, and go \
* back to the top of the loop. \
*/ \
specific_recv_seq = specific_recv->req_recv.req_base.req_sequence; \
} \
} \
} while(0)
/* /*
* Specialized matching routines for internal use only. * Specialized matching routines for internal use only.
*/ */
static bool mca_pml_ob1_check_cantmatch_for_match( static bool mca_pml_ob1_check_cantmatch_for_match( opal_list_t *additional_matches,
opal_list_t *additional_matches, mca_pml_ob1_comm_t* comm,
mca_pml_ob1_comm_t* comm, mca_pml_ob1_comm_proc_t *proc );
mca_pml_ob1_comm_proc_t *proc);
/** /**
* RCS/CTS receive side matching * RCS/CTS receive side matching
@ -518,7 +497,7 @@ static int mca_pml_ob1_recv_frag_match( mca_btl_base_module_t *btl,
/* get sequence number of next message that can be processed */ /* get sequence number of next message that can be processed */
next_msg_seq_expected = (uint16_t)proc->expected_sequence; next_msg_seq_expected = (uint16_t)proc->expected_sequence;
if (frag_msg_seq == next_msg_seq_expected) { if( OPAL_LIKELY(frag_msg_seq == next_msg_seq_expected) ) {
/* /*
* This is the sequence number we were expecting, * This is the sequence number we were expecting,
@ -563,7 +542,7 @@ rematch:
} }
/* if match found, process data */ /* if match found, process data */
if (match) { if( OPAL_LIKELY(match) ) {
match->req_recv.req_base.req_proc = proc->ompi_proc; match->req_recv.req_base.req_proc = proc->ompi_proc;
/* /*
@ -583,7 +562,7 @@ rematch:
/* if no match found, place on unexpected queue */ /* if no match found, place on unexpected queue */
mca_pml_ob1_recv_frag_t* frag; mca_pml_ob1_recv_frag_t* frag;
MCA_PML_OB1_RECV_FRAG_ALLOC(frag, rc); MCA_PML_OB1_RECV_FRAG_ALLOC(frag, rc);
if(OMPI_SUCCESS != rc) { if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
OPAL_THREAD_UNLOCK(&comm->matching_lock); OPAL_THREAD_UNLOCK(&comm->matching_lock);
/** /**
* As we return from the match function, we should generate the expected event. * As we return from the match function, we should generate the expected event.
@ -610,7 +589,7 @@ rematch:
* any fragments on the c_c_frags_cant_match list * any fragments on the c_c_frags_cant_match list
* may now be used to form new matchs * may now be used to form new matchs
*/ */
if (0 < opal_list_get_size(&proc->frags_cant_match)) { if( OPAL_UNLIKELY(0 < opal_list_get_size(&proc->frags_cant_match)) ) {
additional_match = mca_pml_ob1_check_cantmatch_for_match(&additional_matches,comm,proc); additional_match = mca_pml_ob1_check_cantmatch_for_match(&additional_matches,comm,proc);
} }
@ -622,7 +601,7 @@ rematch:
*/ */
mca_pml_ob1_recv_frag_t* frag; mca_pml_ob1_recv_frag_t* frag;
MCA_PML_OB1_RECV_FRAG_ALLOC(frag, rc); MCA_PML_OB1_RECV_FRAG_ALLOC(frag, rc);
if(OMPI_SUCCESS != rc) { if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
OPAL_THREAD_UNLOCK(&comm->matching_lock); OPAL_THREAD_UNLOCK(&comm->matching_lock);
return rc; return rc;
} }
@ -633,13 +612,11 @@ rematch:
/* release matching lock before processing fragment */ /* release matching lock before processing fragment */
OPAL_THREAD_UNLOCK(&comm->matching_lock); OPAL_THREAD_UNLOCK(&comm->matching_lock);
if(match != NULL) { if( OPAL_LIKELY(match != NULL) ) {
mca_pml_ob1_recv_request_progress(match,btl,segments,num_segments); mca_pml_ob1_recv_request_progress(match,btl,segments,num_segments);
#if OMPI_WANT_PERUSE
} else { } else {
PERUSE_TRACE_MSG_EVENT( PERUSE_COMM_MSG_INSERT_IN_UNEX_Q, comm_ptr, PERUSE_TRACE_MSG_EVENT( PERUSE_COMM_MSG_INSERT_IN_UNEX_Q, comm_ptr,
hdr->hdr_src, hdr->hdr_tag, PERUSE_RECV); hdr->hdr_src, hdr->hdr_tag, PERUSE_RECV);
#endif /* OMPI_WANT_PERUSE */
} }
if( OPAL_UNLIKELY(additional_match) ) { if( OPAL_UNLIKELY(additional_match) ) {
opal_list_item_t* item; opal_list_item_t* item;
@ -669,10 +646,9 @@ rematch:
* set by the upper level routine. * set by the upper level routine.
*/ */
static bool mca_pml_ob1_check_cantmatch_for_match( static bool mca_pml_ob1_check_cantmatch_for_match( opal_list_t *additional_matches,
opal_list_t *additional_matches, mca_pml_ob1_comm_t* comm,
mca_pml_ob1_comm_t* comm, mca_pml_ob1_comm_proc_t *proc )
mca_pml_ob1_comm_proc_t *proc)
{ {
/* local parameters */ /* local parameters */
int match_found; int match_found;
@ -698,11 +674,11 @@ static bool mca_pml_ob1_check_cantmatch_for_match(
* number next_msg_seq_expected * number next_msg_seq_expected
*/ */
for(frag = (mca_pml_ob1_recv_frag_t *) for(frag = (mca_pml_ob1_recv_frag_t *)
opal_list_get_first(&proc->frags_cant_match); opal_list_get_first(&proc->frags_cant_match);
frag != (mca_pml_ob1_recv_frag_t *) frag != (mca_pml_ob1_recv_frag_t *)
opal_list_get_end(&proc->frags_cant_match); opal_list_get_end(&proc->frags_cant_match);
frag = (mca_pml_ob1_recv_frag_t *) frag = (mca_pml_ob1_recv_frag_t *)
opal_list_get_next(frag)) opal_list_get_next(frag))
{ {
/* /*
* If the message has the next expected seq from that proc... * If the message has the next expected seq from that proc...
@ -750,7 +726,7 @@ rematch:
} }
/* if match found, process data */ /* if match found, process data */
if (match) { if( OPAL_LIKELY(match) ) {
match->req_recv.req_base.req_proc = proc->ompi_proc; match->req_recv.req_base.req_proc = proc->ompi_proc;
/* /*

97
ompi/mca/pml/ob1/pml_ob1_recvreq.c
Просмотреть файл

@ -44,7 +44,7 @@ void mca_pml_ob1_recv_request_process_pending(void)
recvreq = (mca_pml_ob1_recv_request_t*) recvreq = (mca_pml_ob1_recv_request_t*)
opal_list_remove_first(&mca_pml_ob1.recv_pending); opal_list_remove_first(&mca_pml_ob1.recv_pending);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
if(NULL == recvreq) if( OPAL_UNLIKELY(NULL == recvreq) )
break; break;
recvreq->req_pending = false; recvreq->req_pending = false;
if(mca_pml_ob1_recv_request_schedule_exclusive(recvreq, NULL) == if(mca_pml_ob1_recv_request_schedule_exclusive(recvreq, NULL) ==
@ -157,7 +157,7 @@ static void mca_pml_ob1_put_completion( mca_btl_base_module_t* btl,
mca_pml_ob1_recv_request_t* recvreq = (mca_pml_ob1_recv_request_t*)des->des_cbdata; mca_pml_ob1_recv_request_t* recvreq = (mca_pml_ob1_recv_request_t*)des->des_cbdata;
size_t bytes_received = 0; size_t bytes_received = 0;
if(status == OMPI_SUCCESS) { if( OPAL_LIKELY(status == OMPI_SUCCESS) ) {
MCA_PML_OB1_COMPUTE_SEGMENT_LENGTH( des->des_dst, des->des_dst_cnt, MCA_PML_OB1_COMPUTE_SEGMENT_LENGTH( des->des_dst, des->des_dst_cnt,
0, bytes_received ); 0, bytes_received );
} }
@ -189,7 +189,7 @@ int mca_pml_ob1_recv_request_ack_send_btl(
/* allocate descriptor */ /* allocate descriptor */
MCA_PML_OB1_DES_ALLOC(bml_btl, des, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_ack_hdr_t)); MCA_PML_OB1_DES_ALLOC(bml_btl, des, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_ack_hdr_t));
if(NULL == des) { if( OPAL_UNLIKELY(NULL == des) ) {
return OMPI_ERR_OUT_OF_RESOURCE; return OMPI_ERR_OUT_OF_RESOURCE;
} }
@ -220,7 +220,7 @@ int mca_pml_ob1_recv_request_ack_send_btl(
des->des_cbfunc = mca_pml_ob1_recv_ctl_completion; des->des_cbfunc = mca_pml_ob1_recv_ctl_completion;
rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML); rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML);
if(rc != OMPI_SUCCESS) { if( OPAL_UNLIKELY(rc != OMPI_SUCCESS) ) {
mca_bml_base_free(bml_btl, des); mca_bml_base_free(bml_btl, des);
return OMPI_ERR_OUT_OF_RESOURCE; return OMPI_ERR_OUT_OF_RESOURCE;
} }
@ -297,18 +297,17 @@ static int mca_pml_ob1_recv_request_ack(
* Return resources used by the RDMA * Return resources used by the RDMA
*/ */
static void mca_pml_ob1_rget_completion( static void mca_pml_ob1_rget_completion( mca_btl_base_module_t* btl,
mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* des,
struct mca_btl_base_descriptor_t* des, int status )
int status)
{ {
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context;
mca_pml_ob1_rdma_frag_t* frag = (mca_pml_ob1_rdma_frag_t*)des->des_cbdata; mca_pml_ob1_rdma_frag_t* frag = (mca_pml_ob1_rdma_frag_t*)des->des_cbdata;
mca_pml_ob1_recv_request_t* recvreq = (mca_pml_ob1_recv_request_t*)frag->rdma_req; mca_pml_ob1_recv_request_t* recvreq = (mca_pml_ob1_recv_request_t*)frag->rdma_req;
/* check completion status */ /* check completion status */
if(OMPI_SUCCESS != status) { if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */ /* TSW - FIX */
ORTE_ERROR_LOG(status); ORTE_ERROR_LOG(status);
orte_errmgr.abort(); orte_errmgr.abort();
@ -373,7 +372,8 @@ int mca_pml_ob1_recv_request_get_frag( mca_pml_ob1_rdma_frag_t* frag )
frag->rdma_length, PERUSE_RECV); frag->rdma_length, PERUSE_RECV);
/* queue up get request */ /* queue up get request */
if(OMPI_SUCCESS != (rc = mca_bml_base_get(bml_btl,descriptor))) { rc = mca_bml_base_get(bml_btl,descriptor);
if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
if(OMPI_ERR_OUT_OF_RESOURCE == rc) { if(OMPI_ERR_OUT_OF_RESOURCE == rc) {
mca_bml_base_free(bml_btl, descriptor); mca_bml_base_free(bml_btl, descriptor);
OPAL_THREAD_LOCK(&mca_pml_ob1.lock); OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
@ -425,7 +425,7 @@ static void mca_pml_ob1_recv_request_rget(
size += frag->rdma_segs[i].seg_len; size += frag->rdma_segs[i].seg_len;
} }
frag->rdma_bml = mca_bml_base_btl_array_find(&bml_endpoint->btl_rdma, btl); frag->rdma_bml = mca_bml_base_btl_array_find(&bml_endpoint->btl_rdma, btl);
if(NULL == frag->rdma_bml) { if( OPAL_UNLIKELY(NULL == frag->rdma_bml) ) {
opal_output(0, "[%s:%d] invalid bml for rdma get", __FILE__, __LINE__); opal_output(0, "[%s:%d] invalid bml for rdma get", __FILE__, __LINE__);
orte_errmgr.abort(); orte_errmgr.abort();
} }
@ -445,11 +445,10 @@ static void mca_pml_ob1_recv_request_rget(
* received and actually delivered to the application. * received and actually delivered to the application.
*/ */
void mca_pml_ob1_recv_request_progress( void mca_pml_ob1_recv_request_progress( mca_pml_ob1_recv_request_t* recvreq,
mca_pml_ob1_recv_request_t* recvreq, mca_btl_base_module_t* btl,
mca_btl_base_module_t* btl, mca_btl_base_segment_t* segments,
mca_btl_base_segment_t* segments, size_t num_segments )
size_t num_segments)
{ {
size_t bytes_received = 0; size_t bytes_received = 0;
size_t bytes_delivered = 0; size_t bytes_delivered = 0;
@ -464,14 +463,13 @@ void mca_pml_ob1_recv_request_progress(
bytes_received -= sizeof(mca_pml_ob1_match_hdr_t); bytes_received -= sizeof(mca_pml_ob1_match_hdr_t);
recvreq->req_recv.req_bytes_packed = bytes_received; recvreq->req_recv.req_bytes_packed = bytes_received;
MCA_PML_OB1_RECV_REQUEST_MATCHED(recvreq,&hdr->hdr_match); MCA_PML_OB1_RECV_REQUEST_MATCHED(recvreq,&hdr->hdr_match);
MCA_PML_OB1_RECV_REQUEST_UNPACK( MCA_PML_OB1_RECV_REQUEST_UNPACK( recvreq,
recvreq, segments,
segments, num_segments,
num_segments, sizeof(mca_pml_ob1_match_hdr_t),
sizeof(mca_pml_ob1_match_hdr_t), data_offset,
data_offset, bytes_received,
bytes_received, bytes_delivered);
bytes_delivered);
break; break;
case MCA_PML_OB1_HDR_TYPE_RNDV: case MCA_PML_OB1_HDR_TYPE_RNDV:
@ -505,14 +503,13 @@ void mca_pml_ob1_recv_request_progress(
* unpack. * unpack.
*/ */
if( 0 < bytes_received ) { if( 0 < bytes_received ) {
MCA_PML_OB1_RECV_REQUEST_UNPACK( MCA_PML_OB1_RECV_REQUEST_UNPACK( recvreq,
recvreq, segments,
segments, num_segments,
num_segments, sizeof(mca_pml_ob1_rendezvous_hdr_t),
sizeof(mca_pml_ob1_rendezvous_hdr_t), data_offset,
data_offset, bytes_received,
bytes_received, bytes_delivered );
bytes_delivered);
} }
break; break;
@ -526,14 +523,13 @@ void mca_pml_ob1_recv_request_progress(
case MCA_PML_OB1_HDR_TYPE_FRAG: case MCA_PML_OB1_HDR_TYPE_FRAG:
bytes_received -= sizeof(mca_pml_ob1_frag_hdr_t); bytes_received -= sizeof(mca_pml_ob1_frag_hdr_t);
data_offset = hdr->hdr_frag.hdr_frag_offset; data_offset = hdr->hdr_frag.hdr_frag_offset;
MCA_PML_OB1_RECV_REQUEST_UNPACK( MCA_PML_OB1_RECV_REQUEST_UNPACK( recvreq,
recvreq, segments,
segments, num_segments,
num_segments, sizeof(mca_pml_ob1_frag_hdr_t),
sizeof(mca_pml_ob1_frag_hdr_t), data_offset,
data_offset, bytes_received,
bytes_received, bytes_delivered );
bytes_delivered);
break; break;
default: default:
@ -554,11 +550,10 @@ void mca_pml_ob1_recv_request_progress(
* Handle completion of a probe request * Handle completion of a probe request
*/ */
void mca_pml_ob1_recv_request_matched_probe( void mca_pml_ob1_recv_request_matched_probe( mca_pml_ob1_recv_request_t* recvreq,
mca_pml_ob1_recv_request_t* recvreq, mca_btl_base_module_t* btl,
mca_btl_base_module_t* btl, mca_btl_base_segment_t* segments,
mca_btl_base_segment_t* segments, size_t num_segments )
size_t num_segments)
{ {
size_t bytes_packed = 0; size_t bytes_packed = 0;
mca_pml_ob1_hdr_t* hdr = (mca_pml_ob1_hdr_t*)segments->seg_addr.pval; mca_pml_ob1_hdr_t* hdr = (mca_pml_ob1_hdr_t*)segments->seg_addr.pval;
@ -600,9 +595,9 @@ int mca_pml_ob1_recv_request_schedule_exclusive(
int num_tries = recvreq->req_rdma_cnt; int num_tries = recvreq->req_rdma_cnt;
size_t i; size_t i;
size_t bytes_remaining = recvreq->req_send_offset - size_t bytes_remaining = recvreq->req_send_offset -
recvreq->req_rdma_offset; recvreq->req_rdma_offset;
if(bytes_remaining == 0) { if( OPAL_UNLIKELY(0 == bytes_remaining)) {
OPAL_THREAD_ADD32(&recvreq->req_lock, -recvreq->req_lock); OPAL_THREAD_ADD32(&recvreq->req_lock, -recvreq->req_lock);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -613,7 +608,7 @@ int mca_pml_ob1_recv_request_schedule_exclusive(
if(recvreq->req_rdma[i].bml_btl != start_bml_btl) if(recvreq->req_rdma[i].bml_btl != start_bml_btl)
continue; continue;
/* something left to be send? */ /* something left to be send? */
if(recvreq->req_rdma[i].length) if( OPAL_LIKELY(recvreq->req_rdma[i].length) )
recvreq->req_rdma_idx = i; recvreq->req_rdma_idx = i;
break; break;
} }
@ -688,7 +683,7 @@ int mca_pml_ob1_recv_request_schedule_exclusive(
} }
MCA_PML_OB1_DES_ALLOC(bml_btl, ctl, MCA_BTL_NO_ORDER, hdr_size); MCA_PML_OB1_DES_ALLOC(bml_btl, ctl, MCA_BTL_NO_ORDER, hdr_size);
if(ctl == NULL) { if( OPAL_UNLIKELY(NULL == ctl) ) {
mca_bml_base_free(bml_btl,dst); mca_bml_base_free(bml_btl,dst);
continue; continue;
} }
@ -734,7 +729,7 @@ int mca_pml_ob1_recv_request_schedule_exclusive(
/* send rdma request to peer */ /* send rdma request to peer */
rc = mca_bml_base_send(bml_btl, ctl, MCA_BTL_TAG_PML); rc = mca_bml_base_send(bml_btl, ctl, MCA_BTL_TAG_PML);
if(rc == OMPI_SUCCESS) { if( OPAL_LIKELY(OMPI_SUCCESS == rc) ) {
/* update request state */ /* update request state */
recvreq->req_rdma_offset += size; recvreq->req_rdma_offset += size;
OPAL_THREAD_ADD_SIZE_T(&recvreq->req_pipeline_depth,1); OPAL_THREAD_ADD_SIZE_T(&recvreq->req_pipeline_depth,1);

303
ompi/mca/pml/ob1/pml_ob1_sendreq.c
Просмотреть файл

@ -45,11 +45,12 @@ void mca_pml_ob1_send_request_process_pending(mca_bml_base_btl_t *bml_btl)
mca_pml_ob1_send_pending_t pending_type; mca_pml_ob1_send_pending_t pending_type;
mca_pml_ob1_send_request_t* sendreq; mca_pml_ob1_send_request_t* sendreq;
mca_bml_base_btl_t *send_dst; mca_bml_base_btl_t *send_dst;
OPAL_THREAD_LOCK(&mca_pml_ob1.lock); OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
sendreq = (mca_pml_ob1_send_request_t*) sendreq = (mca_pml_ob1_send_request_t*)
opal_list_remove_first(&mca_pml_ob1.send_pending); opal_list_remove_first(&mca_pml_ob1.send_pending);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
if(NULL == sendreq) if( OPAL_UNLIKELY(NULL == sendreq) )
break; break;
pending_type = sendreq->req_pending; pending_type = sendreq->req_pending;
sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_NONE; sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_NONE;
@ -70,7 +71,7 @@ void mca_pml_ob1_send_request_process_pending(mca_bml_base_btl_t *bml_btl)
* endpoint or no resources put request back on the list */ * endpoint or no resources put request back on the list */
OPAL_THREAD_LOCK(&mca_pml_ob1.lock); OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_START; sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_START;
if(NULL == send_dst) { if( OPAL_UNLIKELY(NULL == send_dst) ) {
opal_list_append(&mca_pml_ob1.send_pending, opal_list_append(&mca_pml_ob1.send_pending,
(opal_list_item_t*)sendreq); (opal_list_item_t*)sendreq);
} else { } else {
@ -137,22 +138,21 @@ static void mca_pml_ob1_send_request_construct(mca_pml_ob1_send_request_t* req)
OBJ_CONSTRUCT(&req->req_send_ranges, opal_list_t); OBJ_CONSTRUCT(&req->req_send_ranges, opal_list_t);
} }
OBJ_CLASS_INSTANCE( OBJ_CLASS_INSTANCE( mca_pml_ob1_send_request_t,
mca_pml_ob1_send_request_t, mca_pml_base_send_request_t,
mca_pml_base_send_request_t, mca_pml_ob1_send_request_construct,
mca_pml_ob1_send_request_construct, NULL );
NULL);
/** /**
* Completion of a short message - nothing left to schedule. Note that this * Completion of a short message - nothing left to schedule. Note that this
* function is only called for 0 sized messages. * function is only called for 0 sized messages.
*/ */
void mca_pml_ob1_match_completion_cache( static void
struct mca_btl_base_module_t* btl, mca_pml_ob1_match_completion_cache( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* descriptor, struct mca_btl_base_descriptor_t* descriptor,
int status) int status )
{ {
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata; mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context;
@ -163,7 +163,7 @@ void mca_pml_ob1_match_completion_cache(
} }
/* check completion status */ /* check completion status */
if(OMPI_SUCCESS != status) { if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */ /* TSW - FIX */
opal_output(0, "%s:%d FATAL", __FILE__, __LINE__); opal_output(0, "%s:%d FATAL", __FILE__, __LINE__);
orte_errmgr.abort(); orte_errmgr.abort();
@ -183,11 +183,11 @@ void mca_pml_ob1_match_completion_cache(
* Completion of a short message - nothing left to schedule. * Completion of a short message - nothing left to schedule.
*/ */
void mca_pml_ob1_match_completion_free( static void
struct mca_btl_base_module_t* btl, mca_pml_ob1_match_completion_free( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* descriptor, struct mca_btl_base_descriptor_t* descriptor,
int status) int status )
{ {
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata; mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context;
@ -198,7 +198,7 @@ void mca_pml_ob1_match_completion_free(
} }
/* check completion status */ /* check completion status */
if(OMPI_SUCCESS != status) { if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */ /* TSW - FIX */
opal_output(0, "%s:%d FATAL", __FILE__, __LINE__); opal_output(0, "%s:%d FATAL", __FILE__, __LINE__);
orte_errmgr.abort(); orte_errmgr.abort();
@ -218,11 +218,11 @@ void mca_pml_ob1_match_completion_free(
* Completion of the first fragment of a long message that * Completion of the first fragment of a long message that
* requires an acknowledgement * requires an acknowledgement
*/ */
static void mca_pml_ob1_rndv_completion( static void
mca_btl_base_module_t* btl, mca_pml_ob1_rndv_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* descriptor, struct mca_btl_base_descriptor_t* descriptor,
int status) int status )
{ {
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata; mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)descriptor->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)descriptor->des_context;
@ -234,7 +234,7 @@ static void mca_pml_ob1_rndv_completion(
} }
/* check completion status */ /* check completion status */
if(OMPI_SUCCESS != status) { if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */ /* TSW - FIX */
opal_output(0, "%s:%d FATAL", __FILE__, __LINE__); opal_output(0, "%s:%d FATAL", __FILE__, __LINE__);
orte_errmgr.abort(); orte_errmgr.abort();
@ -284,11 +284,11 @@ static void mca_pml_ob1_rndv_completion(
* Completion of a get request. * Completion of a get request.
*/ */
static void mca_pml_ob1_rget_completion( static void
mca_btl_base_module_t* btl, mca_pml_ob1_rget_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des, struct mca_btl_base_descriptor_t* des,
int status) int status )
{ {
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)des->des_cbdata; mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)des->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context;
@ -312,11 +312,11 @@ static void mca_pml_ob1_rget_completion(
* Completion of a control message - return resources. * Completion of a control message - return resources.
*/ */
static void mca_pml_ob1_send_ctl_completion( static void
mca_btl_base_module_t* btl, mca_pml_ob1_send_ctl_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* descriptor, struct mca_btl_base_descriptor_t* descriptor,
int status) int status )
{ {
/* return the descriptor */ /* return the descriptor */
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context;
@ -331,18 +331,18 @@ static void mca_pml_ob1_send_ctl_completion(
* to schedule additional fragments. * to schedule additional fragments.
*/ */
static void mca_pml_ob1_frag_completion( static void
mca_btl_base_module_t* btl, mca_pml_ob1_frag_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* descriptor, struct mca_btl_base_descriptor_t* descriptor,
int status) int status )
{ {
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata; mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context;
size_t req_bytes_delivered = 0; size_t req_bytes_delivered = 0;
/* check completion status */ /* check completion status */
if(OMPI_SUCCESS != status) { if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */ /* TSW - FIX */
opal_output(0, "%s:%d FATAL", __FILE__, __LINE__); opal_output(0, "%s:%d FATAL", __FILE__, __LINE__);
orte_errmgr.abort(); orte_errmgr.abort();
@ -370,8 +370,6 @@ static void mca_pml_ob1_frag_completion(
MCA_PML_OB1_PROGRESS_PENDING(bml_btl); MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
} }
/** /**
* Buffer the entire message and mark as complete. * Buffer the entire message and mark as complete.
*/ */
@ -392,7 +390,7 @@ int mca_pml_ob1_send_request_start_buffered(
/* allocate descriptor */ /* allocate descriptor */
mca_bml_base_alloc(bml_btl, &descriptor, mca_bml_base_alloc(bml_btl, &descriptor,
MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rendezvous_hdr_t) + size); MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rendezvous_hdr_t) + size);
if(NULL == descriptor) { if( OPAL_UNLIKELY(NULL == descriptor) ) {
return OMPI_ERR_OUT_OF_RESOURCE; return OMPI_ERR_OUT_OF_RESOURCE;
} }
segment = descriptor->des_src; segment = descriptor->des_src;
@ -445,7 +443,7 @@ int mca_pml_ob1_send_request_start_buffered(
/* buffer the remainder of the message */ /* buffer the remainder of the message */
rc = mca_pml_base_bsend_request_alloc((ompi_request_t*)sendreq); rc = mca_pml_base_bsend_request_alloc((ompi_request_t*)sendreq);
if(OMPI_SUCCESS != rc) { if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
mca_bml_base_free(bml_btl, descriptor); mca_bml_base_free(bml_btl, descriptor);
return rc; return rc;
} }
@ -473,7 +471,7 @@ int mca_pml_ob1_send_request_start_buffered(
/* send */ /* send */
rc = mca_bml_base_send(bml_btl, descriptor, MCA_BTL_TAG_PML); rc = mca_bml_base_send(bml_btl, descriptor, MCA_BTL_TAG_PML);
if(OMPI_SUCCESS != rc) { if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
mca_bml_base_free(bml_btl, descriptor ); mca_bml_base_free(bml_btl, descriptor );
} }
return rc; return rc;
@ -643,9 +641,9 @@ int mca_pml_ob1_send_request_start_prepare( mca_pml_ob1_send_request_t* sendreq,
*/ */
int mca_pml_ob1_send_request_start_rdma( int mca_pml_ob1_send_request_start_rdma(
mca_pml_ob1_send_request_t* sendreq, mca_pml_ob1_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl, mca_bml_base_btl_t* bml_btl,
size_t size) size_t size)
{ {
/* /*
* When req_rdma array is constructed the firs element of the array always * When req_rdma array is constructed the firs element of the array always
@ -676,114 +674,114 @@ int mca_pml_ob1_send_request_start_rdma(
0, 0,
&size, &size,
&src ); &src );
if( OPAL_UNLIKELY(NULL == src) ) { if( OPAL_UNLIKELY(NULL == src) ) {
ompi_convertor_set_position(&sendreq->req_send.req_base.req_convertor, ompi_convertor_set_position(&sendreq->req_send.req_base.req_convertor,
&old_position); &old_position);
return OMPI_ERR_OUT_OF_RESOURCE; return OMPI_ERR_OUT_OF_RESOURCE;
} }
src->des_cbfunc = mca_pml_ob1_rget_completion; src->des_cbfunc = mca_pml_ob1_rget_completion;
src->des_cbdata = sendreq; src->des_cbdata = sendreq;
/* allocate space for get hdr + segment list */ /* allocate space for get hdr + segment list */
mca_bml_base_alloc(bml_btl, &des, MCA_BTL_NO_ORDER, mca_bml_base_alloc(bml_btl, &des, MCA_BTL_NO_ORDER,
sizeof(mca_pml_ob1_rget_hdr_t) + (sizeof(mca_btl_base_segment_t)*(src->des_src_cnt-1))); sizeof(mca_pml_ob1_rget_hdr_t) + (sizeof(mca_btl_base_segment_t)*(src->des_src_cnt-1)));
if(NULL == des) { if( OPAL_UNLIKELY(NULL == des) ) {
ompi_convertor_set_position(&sendreq->req_send.req_base.req_convertor, ompi_convertor_set_position(&sendreq->req_send.req_base.req_convertor,
&old_position); &old_position);
mca_bml_base_free(bml_btl, src); mca_bml_base_free(bml_btl, src);
return OMPI_ERR_OUT_OF_RESOURCE; return OMPI_ERR_OUT_OF_RESOURCE;
} }
segment = des->des_src; segment = des->des_src;
/* build match header */ /* build match header */
hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval; hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval;
hdr->hdr_common.hdr_flags = MCA_PML_OB1_HDR_FLAGS_CONTIG|MCA_PML_OB1_HDR_FLAGS_PIN; hdr->hdr_common.hdr_flags = MCA_PML_OB1_HDR_FLAGS_CONTIG|MCA_PML_OB1_HDR_FLAGS_PIN;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RGET; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RGET;
hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence; hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed; hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq; hdr->hdr_rndv.hdr_src_req.pval = sendreq;
hdr->hdr_rget.hdr_des.pval = src; hdr->hdr_rget.hdr_des.pval = src;
hdr->hdr_rget.hdr_seg_cnt = src->des_src_cnt; hdr->hdr_rget.hdr_seg_cnt = src->des_src_cnt;
#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
#ifdef WORDS_BIGENDIAN #ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#else #else
/* if we are little endian and the remote side is big endian, /* if we are little endian and the remote side is big endian,
we're responsible for making sure the data is in network byte we're responsible for making sure the data is in network byte
order */ order */
/* RDMA is currently disabled by bml if arch doesn't /* RDMA is currently disabled by bml if arch doesn't
match, so this shouldn't be needed. here to make sure match, so this shouldn't be needed. here to make sure
we remember if we ever change the bml. */ we remember if we ever change the bml. */
assert(0 == (sendreq->req_send.req_base.req_proc->proc_arch & assert(0 == (sendreq->req_send.req_base.req_proc->proc_arch &
OMPI_ARCH_ISBIGENDIAN)); OMPI_ARCH_ISBIGENDIAN));
#endif #endif
#endif #endif
for( i = 0; i < src->des_src_cnt; i++ ) { for( i = 0; i < src->des_src_cnt; i++ ) {
hdr->hdr_rget.hdr_segs[i].seg_addr.lval = ompi_ptr_ptol(src->des_src[i].seg_addr.pval); hdr->hdr_rget.hdr_segs[i].seg_addr.lval = ompi_ptr_ptol(src->des_src[i].seg_addr.pval);
hdr->hdr_rget.hdr_segs[i].seg_len = src->des_src[i].seg_len; hdr->hdr_rget.hdr_segs[i].seg_len = src->des_src[i].seg_len;
hdr->hdr_rget.hdr_segs[i].seg_key.key64 = src->des_src[i].seg_key.key64; hdr->hdr_rget.hdr_segs[i].seg_key.key64 = src->des_src[i].seg_key.key64;
} }
des->des_cbfunc = mca_pml_ob1_send_ctl_completion; des->des_cbfunc = mca_pml_ob1_send_ctl_completion;
/** /**
* Well, it's a get so we will not know when the peer get the data anyway. * Well, it's a get so we will not know when the peer get the data anyway.
* If we generate the PERUSE event here, at least we will know when do we * If we generate the PERUSE event here, at least we will know when do we
* sent the GET message ... * sent the GET message ...
*/ */
if( sendreq->req_send.req_bytes_packed > 0 ) { if( sendreq->req_send.req_bytes_packed > 0 ) {
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN, PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN,
&(sendreq->req_send.req_base), PERUSE_SEND ); &(sendreq->req_send.req_base), PERUSE_SEND );
} }
} else { } else {
/* allocate a rendezvous header - dont eager send any data /* allocate a rendezvous header - dont eager send any data
* receiver will schedule rdma put(s) of the entire message * receiver will schedule rdma put(s) of the entire message
*/ */
mca_bml_base_alloc(bml_btl, &des, mca_bml_base_alloc(bml_btl, &des,
MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rendezvous_hdr_t)); MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rendezvous_hdr_t));
if(NULL == des) { if( OPAL_UNLIKELY(NULL == des)) {
return OMPI_ERR_OUT_OF_RESOURCE; return OMPI_ERR_OUT_OF_RESOURCE;
} }
segment = des->des_src; segment = des->des_src;
/* build hdr */ /* build hdr */
hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval; hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval;
hdr->hdr_common.hdr_flags = MCA_PML_OB1_HDR_FLAGS_CONTIG|MCA_PML_OB1_HDR_FLAGS_PIN; hdr->hdr_common.hdr_flags = MCA_PML_OB1_HDR_FLAGS_CONTIG|MCA_PML_OB1_HDR_FLAGS_PIN;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RNDV; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RNDV;
hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence; hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed; hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq; hdr->hdr_rndv.hdr_src_req.pval = sendreq;
#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
#ifdef WORDS_BIGENDIAN #ifdef WORDS_BIGENDIAN
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
#else #else
/* if we are little endian and the remote side is big endian, /* if we are little endian and the remote side is big endian,
we're responsible for making sure the data is in network byte we're responsible for making sure the data is in network byte
order */ order */
if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) { if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) {
hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO;
MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv); MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv);
} }
#endif #endif
#endif #endif
/* update lengths with number of bytes actually packed */ /* update lengths with number of bytes actually packed */
segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t); segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t);
/* first fragment of a long message */ /* first fragment of a long message */
des->des_cbfunc = mca_pml_ob1_rndv_completion; des->des_cbfunc = mca_pml_ob1_rndv_completion;
} }
des->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY; des->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY;
@ -791,7 +789,7 @@ int mca_pml_ob1_send_request_start_rdma(
/* send */ /* send */
rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML); rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML);
if(OMPI_SUCCESS != rc) { if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
mca_bml_base_free(bml_btl, des); mca_bml_base_free(bml_btl, des);
} }
return rc; return rc;
@ -803,11 +801,10 @@ int mca_pml_ob1_send_request_start_rdma(
* the btls eager limit. * the btls eager limit.
*/ */
int mca_pml_ob1_send_request_start_rndv( int mca_pml_ob1_send_request_start_rndv( mca_pml_ob1_send_request_t* sendreq,
mca_pml_ob1_send_request_t* sendreq, mca_bml_base_btl_t* bml_btl,
mca_bml_base_btl_t* bml_btl, size_t size,
size_t size, int flags )
int flags)
{ {
mca_btl_base_descriptor_t* des; mca_btl_base_descriptor_t* des;
mca_btl_base_segment_t* segment; mca_btl_base_segment_t* segment;
@ -830,7 +827,7 @@ int mca_pml_ob1_send_request_start_rndv(
&des ); &des );
} }
if( NULL == des ) { if( OPAL_UNLIKELY(NULL == des) ) {
return OMPI_ERR_OUT_OF_RESOURCE; return OMPI_ERR_OUT_OF_RESOURCE;
} }
segment = des->des_src; segment = des->des_src;
@ -867,14 +864,15 @@ int mca_pml_ob1_send_request_start_rndv(
/* send */ /* send */
rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML); rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML);
if(OMPI_SUCCESS != rc) { if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
mca_bml_base_free(bml_btl, des ); mca_bml_base_free(bml_btl, des );
} }
return rc; return rc;
} }
void mca_pml_ob1_send_request_copy_in_out(mca_pml_ob1_send_request_t *sendreq, void mca_pml_ob1_send_request_copy_in_out( mca_pml_ob1_send_request_t *sendreq,
uint64_t send_offset, uint64_t send_length) uint64_t send_offset,
uint64_t send_length )
{ {
mca_pml_ob1_send_range_t *sr; mca_pml_ob1_send_range_t *sr;
ompi_free_list_item_t *i; ompi_free_list_item_t *i;
@ -883,7 +881,7 @@ void mca_pml_ob1_send_request_copy_in_out(mca_pml_ob1_send_request_t *sendreq,
int rc = OMPI_SUCCESS, n; int rc = OMPI_SUCCESS, n;
double weight_total = 0; double weight_total = 0;
if(0 == send_length) if( OPAL_UNLIKELY(0 == send_length) )
return; return;
OMPI_FREE_LIST_WAIT(&mca_pml_ob1.send_ranges, i, rc); OMPI_FREE_LIST_WAIT(&mca_pml_ob1.send_ranges, i, rc);
@ -934,7 +932,7 @@ int mca_pml_ob1_send_request_schedule_exclusive(
opal_list_item_t *item; opal_list_item_t *item;
mca_bml_base_btl_t* bml_btl; mca_bml_base_btl_t* bml_btl;
if(NULL == range || 0 == range->range_send_length) { if( OPAL_UNLIKELY(NULL == range || 0 == range->range_send_length) ) {
OPAL_THREAD_LOCK(&sendreq->req_send_range_lock); OPAL_THREAD_LOCK(&sendreq->req_send_range_lock);
if(range) { if(range) {
opal_list_remove_first(&sendreq->req_send_ranges); opal_list_remove_first(&sendreq->req_send_ranges);
@ -964,7 +962,7 @@ int mca_pml_ob1_send_request_schedule_exclusive(
prev_bytes_remaining = range->range_send_length; prev_bytes_remaining = range->range_send_length;
if (num_fail == range->range_btl_cnt) { if( OPAL_UNLIKELY(num_fail == range->range_btl_cnt) ) {
assert(sendreq->req_pending == MCA_PML_OB1_SEND_PENDING_NONE); assert(sendreq->req_pending == MCA_PML_OB1_SEND_PENDING_NONE);
OPAL_THREAD_LOCK(&mca_pml_ob1.lock); OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_SCHEDULE; sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_SCHEDULE;
@ -1037,7 +1035,7 @@ int mca_pml_ob1_send_request_schedule_exclusive(
/* initiate send - note that this may complete before the call returns */ /* initiate send - note that this may complete before the call returns */
rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML); rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML);
if(rc == OMPI_SUCCESS) { if( OPAL_LIKELY(rc == OMPI_SUCCESS) ) {
/* update state */ /* update state */
range->range_btls[btl_idx].length -= size; range->range_btls[btl_idx].length -= size;
range->range_send_length -= size; range->range_send_length -= size;
@ -1071,7 +1069,7 @@ static void mca_pml_ob1_put_completion( mca_btl_base_module_t* btl,
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context;
/* check completion status */ /* check completion status */
if(OMPI_SUCCESS != status) { if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */ /* TSW - FIX */
ORTE_ERROR_LOG(status); ORTE_ERROR_LOG(status);
orte_errmgr.abort(); orte_errmgr.abort();
@ -1120,7 +1118,7 @@ int mca_pml_ob1_send_request_put_frag( mca_pml_ob1_rdma_frag_t* frag )
&frag->rdma_length, &frag->rdma_length,
&des ); &des );
if(NULL == des) { if( OPAL_UNLIKELY(NULL == des) ) {
if(frag->retries < mca_pml_ob1.rdma_put_retries_limit) { if(frag->retries < mca_pml_ob1.rdma_put_retries_limit) {
size_t offset = (size_t)frag->rdma_hdr.hdr_rdma.hdr_rdma_offset; size_t offset = (size_t)frag->rdma_hdr.hdr_rdma.hdr_rdma_offset;
frag->rdma_length = save_size; frag->rdma_length = save_size;
@ -1153,7 +1151,8 @@ int mca_pml_ob1_send_request_put_frag( mca_pml_ob1_rdma_frag_t* frag )
PERUSE_TRACE_COMM_OMPI_EVENT( PERUSE_COMM_REQ_XFER_CONTINUE, PERUSE_TRACE_COMM_OMPI_EVENT( PERUSE_COMM_REQ_XFER_CONTINUE,
&(((mca_pml_ob1_send_request_t*)frag->rdma_req)->req_send.req_base), save_size, PERUSE_SEND ); &(((mca_pml_ob1_send_request_t*)frag->rdma_req)->req_send.req_base), save_size, PERUSE_SEND );
if(OMPI_SUCCESS != (rc = mca_bml_base_put(bml_btl, des))) { rc = mca_bml_base_put(bml_btl, des);
if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
mca_bml_base_free(bml_btl, des); mca_bml_base_free(bml_btl, des);
frag->rdma_length = save_size; frag->rdma_length = save_size;
if(OMPI_ERR_OUT_OF_RESOURCE == rc) { if(OMPI_ERR_OUT_OF_RESOURCE == rc) {
@ -1193,7 +1192,7 @@ void mca_pml_ob1_send_request_put( mca_pml_ob1_send_request_t* sendreq,
MCA_PML_OB1_RDMA_FRAG_ALLOC(frag, rc); MCA_PML_OB1_RDMA_FRAG_ALLOC(frag, rc);
if(NULL == frag) { if( OPAL_UNLIKELY(NULL == frag) ) {
/* TSW - FIX */ /* TSW - FIX */
ORTE_ERROR_LOG(rc); ORTE_ERROR_LOG(rc);
orte_errmgr.abort(); orte_errmgr.abort();

73
ompi/mca/pml/ob1/pml_ob1_sendreq.h
Просмотреть файл

@ -30,9 +30,7 @@
#include "ompi/datatype/dt_arch.h" #include "ompi/datatype/dt_arch.h"
#include "ompi/mca/bml/bml.h" #include "ompi/mca/bml/bml.h"
#if defined(c_plusplus) || defined(__cplusplus) BEGIN_C_DECLS
extern "C" {
#endif
typedef enum { typedef enum {
MCA_PML_OB1_SEND_PENDING_NONE, MCA_PML_OB1_SEND_PENDING_NONE,
@ -197,7 +195,7 @@ do {
MCA_PML_OB1_SEND_REQUEST_RETURN( sendreq ); \ MCA_PML_OB1_SEND_REQUEST_RETURN( sendreq ); \
} \ } \
OPAL_THREAD_UNLOCK(&ompi_request_lock); \ OPAL_THREAD_UNLOCK(&ompi_request_lock); \
} while (0) } while (0)
/** /**
* Schedule additional fragments * Schedule additional fragments
@ -229,7 +227,7 @@ static inline void mca_pml_ob1_send_request_schedule(
MCA_PML_BASE_SEND_REQUEST_FINI((&(sendreq)->req_send)); \ MCA_PML_BASE_SEND_REQUEST_FINI((&(sendreq)->req_send)); \
OMPI_FREE_LIST_RETURN( &mca_pml_base_send_requests, \ OMPI_FREE_LIST_RETURN( &mca_pml_base_send_requests, \
(ompi_free_list_item_t*)sendreq); \ (ompi_free_list_item_t*)sendreq); \
} }
/** /**
* Start the specified request * Start the specified request
@ -261,15 +259,19 @@ int mca_pml_ob1_send_request_start_rndv(
size_t size, size_t size,
int flags); int flags);
static inline int mca_pml_ob1_send_request_start_btl( static inline int
mca_pml_ob1_send_request_t* sendreq, mca_pml_ob1_send_request_start_btl( mca_pml_ob1_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl) mca_bml_base_btl_t* bml_btl )
{ {
size_t size = sendreq->req_send.req_bytes_packed; size_t size = sendreq->req_send.req_bytes_packed;
size_t eager_limit = bml_btl->btl_eager_limit - sizeof(mca_pml_ob1_hdr_t); size_t eager_limit = bml_btl->btl_eager_limit;
int rc; int rc;
if(size <= eager_limit) { if( eager_limit > mca_pml_ob1.eager_limit )
eager_limit = mca_pml_ob1.eager_limit;
eager_limit -= sizeof(mca_pml_ob1_hdr_t);
if( OPAL_LIKELY(size <= eager_limit) ) {
switch(sendreq->req_send.req_send_mode) { switch(sendreq->req_send.req_send_mode) {
case MCA_PML_BASE_SEND_SYNCHRONOUS: case MCA_PML_BASE_SEND_SYNCHRONOUS:
rc = mca_pml_ob1_send_request_start_rndv(sendreq, bml_btl, size, 0); rc = mca_pml_ob1_send_request_start_rndv(sendreq, bml_btl, size, 0);
@ -293,23 +295,23 @@ static inline int mca_pml_ob1_send_request_start_btl(
if(sendreq->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED) { if(sendreq->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED) {
rc = mca_pml_ob1_send_request_start_buffered(sendreq, bml_btl, size); rc = mca_pml_ob1_send_request_start_buffered(sendreq, bml_btl, size);
} else if } else if
(ompi_convertor_need_buffers(&sendreq->req_send.req_base.req_convertor) == false) { (ompi_convertor_need_buffers(&sendreq->req_send.req_base.req_convertor) == false) {
unsigned char *base; unsigned char *base;
ompi_convertor_get_current_pointer( &sendreq->req_send.req_base.req_convertor, (void**)&base ); ompi_convertor_get_current_pointer( &sendreq->req_send.req_base.req_convertor, (void**)&base );
if( 0 != (sendreq->req_rdma_cnt = (uint32_t)mca_pml_ob1_rdma_btls( if( 0 != (sendreq->req_rdma_cnt = (uint32_t)mca_pml_ob1_rdma_btls(
sendreq->req_endpoint, sendreq->req_endpoint,
base, base,
sendreq->req_send.req_bytes_packed, sendreq->req_send.req_bytes_packed,
sendreq->req_rdma))) { sendreq->req_rdma))) {
rc = mca_pml_ob1_send_request_start_rdma(sendreq, bml_btl, rc = mca_pml_ob1_send_request_start_rdma(sendreq, bml_btl,
sendreq->req_send.req_bytes_packed); sendreq->req_send.req_bytes_packed);
if(OMPI_SUCCESS != rc) { if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
mca_pml_ob1_free_rdma_resources(sendreq); mca_pml_ob1_free_rdma_resources(sendreq);
} }
} else { } else {
rc = mca_pml_ob1_send_request_start_rndv(sendreq, bml_btl, size, rc = mca_pml_ob1_send_request_start_rndv(sendreq, bml_btl, size,
MCA_PML_OB1_HDR_FLAGS_CONTIG); MCA_PML_OB1_HDR_FLAGS_CONTIG);
} }
} else { } else {
rc = mca_pml_ob1_send_request_start_rndv(sendreq, bml_btl, size, 0); rc = mca_pml_ob1_send_request_start_rndv(sendreq, bml_btl, size, 0);
@ -319,15 +321,15 @@ static inline int mca_pml_ob1_send_request_start_btl(
return rc; return rc;
} }
static inline int mca_pml_ob1_send_request_start( static inline int
mca_pml_ob1_send_request_t* sendreq) mca_pml_ob1_send_request_start( mca_pml_ob1_send_request_t* sendreq )
{ {
mca_pml_ob1_comm_t* comm = sendreq->req_send.req_base.req_comm->c_pml_comm; mca_pml_ob1_comm_t* comm = sendreq->req_send.req_base.req_comm->c_pml_comm;
mca_bml_base_endpoint_t* endpoint = (mca_bml_base_endpoint_t*) mca_bml_base_endpoint_t* endpoint = (mca_bml_base_endpoint_t*)
sendreq->req_send.req_base.req_proc->proc_bml; sendreq->req_send.req_base.req_proc->proc_bml;
size_t i; size_t i;
if(endpoint == NULL) { if( OPAL_UNLIKELY(endpoint == NULL) ) {
return OMPI_ERR_UNREACH; return OMPI_ERR_UNREACH;
} }
@ -349,7 +351,7 @@ static inline int mca_pml_ob1_send_request_start(
/* select a btl */ /* select a btl */
bml_btl = mca_bml_base_btl_array_get_next(&endpoint->btl_eager); bml_btl = mca_bml_base_btl_array_get_next(&endpoint->btl_eager);
rc = mca_pml_ob1_send_request_start_btl(sendreq, bml_btl); rc = mca_pml_ob1_send_request_start_btl(sendreq, bml_btl);
if(OMPI_ERR_OUT_OF_RESOURCE != rc) if( OPAL_LIKELY(OMPI_ERR_OUT_OF_RESOURCE != rc) )
return rc; return rc;
} }
OPAL_THREAD_LOCK(&mca_pml_ob1.lock); OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
@ -360,34 +362,13 @@ static inline int mca_pml_ob1_send_request_start(
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
/**
* Completion callback on match header
* Cache descriptor.
*/
void mca_pml_ob1_match_completion_cache(
struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* descriptor,
int status);
/**
* Completion callback on match header
* Free descriptor.
*/
void mca_pml_ob1_match_completion_free(
struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* descriptor,
int status);
/** /**
* Initiate a put scheduled by the receiver. * Initiate a put scheduled by the receiver.
*/ */
void mca_pml_ob1_send_request_put( void mca_pml_ob1_send_request_put( mca_pml_ob1_send_request_t* sendreq,
mca_pml_ob1_send_request_t* sendreq, mca_btl_base_module_t* btl,
mca_btl_base_module_t* btl, mca_pml_ob1_rdma_hdr_t* hdr );
mca_pml_ob1_rdma_hdr_t* hdr);
int mca_pml_ob1_send_request_put_frag(mca_pml_ob1_rdma_frag_t* frag); int mca_pml_ob1_send_request_put_frag(mca_pml_ob1_rdma_frag_t* frag);