1
1

Bring the UD BTL over to the trunk, named 'ofud'.

This commit was SVN r15298.
Этот коммит содержится в:
Andrew Friedley 2007-07-05 23:42:54 +00:00
родитель 6e8d25fdaf
Коммит 77038b65a8
12 изменённых файлов: 2503 добавлений и 0 удалений

68
ompi/mca/btl/ofud/Makefile.am Обычный файл
Просмотреть файл

@ -0,0 +1,68 @@
#
# 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 Sandia National Laboratories. All rights
# reserved.
# $COPYRIGHT$
#
# Additional copyrights may follow
#
# $HEADER$
#
# Use the top-level Makefile.options
AM_CPPFLAGS=$(btl_ofud_CPPFLAGS)
sources = \
btl_ofud.c \
btl_ofud.h \
btl_ofud_component.c \
btl_ofud_endpoint.c \
btl_ofud_endpoint.h \
btl_ofud_frag.c \
btl_ofud_frag.h \
btl_ofud_proc.c \
btl_ofud_proc.h
# Make the output library in this directory, and name it either
# mca_<type>_<name>.la (for DSO builds) or libmca_<type>_<name>.la
# (for static builds).
if OMPI_BUILD_btl_ofud_DSO
lib =
lib_sources =
component = mca_btl_ofud.la
component_sources = $(sources)
else
lib = libmca_btl_ofud.la
lib_sources = $(sources)
component =
component_sources =
endif
mcacomponentdir = $(pkglibdir)
mcacomponent_LTLIBRARIES = $(component)
mca_btl_ofud_la_SOURCES = $(component_sources)
mca_btl_ofud_la_LDFLAGS = -module -avoid-version $(btl_ofud_LDFLAGS)
mca_btl_ofud_la_LIBADD = \
$(btl_ofud_LIBS) \
$(top_ompi_builddir)/ompi/libmpi.la \
$(top_ompi_builddir)/orte/libopen-rte.la \
$(top_ompi_builddir)/opal/libopen-pal.la
noinst_LTLIBRARIES = $(lib)
libmca_btl_ofud_la_SOURCES = $(lib_sources)
libmca_btl_ofud_la_LDFLAGS= -module -avoid-version $(btl_ofud_LDFLAGS)
libmca_btl_ofud_la_LIBADD=$(btl_ofud_LIBS)

691
ompi/mca/btl/ofud/btl_ofud.c Обычный файл
Просмотреть файл

@ -0,0 +1,691 @@
/*
* 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include <errno.h>
#include <string.h>
#include "ompi_config.h"
#include "opal/prefetch.h"
#include "opal/util/output.h"
#include "ompi/datatype/convertor.h"
#include "ompi/datatype/datatype.h"
#include "ompi/mca/btl/btl.h"
#include "ompi/mca/btl/base/btl_base_error.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/mpool/mpool.h"
#include "ompi/mca/mpool/rdma/mpool_rdma.h"
#include "btl_ofud.h"
#include "btl_ofud_frag.h"
#include "btl_ofud_proc.h"
#include "btl_ofud_endpoint.h"
mca_btl_ud_module_t mca_btl_ofud_module = {
{
&mca_btl_ofud_component.super,
0, /* eager_limit */
0, /* min_send_size */
0, /* max_send_size */
0, /* rdma_pipeline_send_length */
0, /* rdma_pipeline_frag_size */
0, /* min_rdma_pipeline_size */
0, /* exclusivity */
0, /* latency */
0, /* bandwidth */
MCA_BTL_FLAGS_SEND,
mca_btl_ud_add_procs,
mca_btl_ud_del_procs,
mca_btl_ud_register,
mca_btl_ud_finalize,
mca_btl_ud_alloc,
mca_btl_ud_free,
mca_btl_ud_prepare_src,
NULL, /*mca_btl_ud_prepare_dst */
mca_btl_ud_send,
NULL, /*mca_btl_ud_put */
NULL, /*mca_btl_ud_get */
mca_btl_base_dump,
NULL, /* mpool */
NULL, /* register error */
mca_btl_udapl_ft_event
}
};
/*
* Add procs to this BTL module, receiving endpoint information from the modex.
*/
int mca_btl_ud_add_procs(struct mca_btl_base_module_t* btl,
size_t nprocs,
struct ompi_proc_t **ompi_procs,
struct mca_btl_base_endpoint_t** peers,
ompi_bitmap_t* reachable)
{
mca_btl_ud_module_t* ud_btl = (mca_btl_ud_module_t*)btl;
struct ibv_ah_attr ah_attr;
int i, rc;
/* Set up the endpoint lookup table if it hasn't been already */
/* We do this here so we can initialize the table to a reasonable size
based on nprocs */
#if 0
if(NULL == ud_btl->ep_lookup) {
ud_btl->ep_lookup = malloc(sizeof(opal_hash_table_t));
OBJ_CONSTRUCT(ud_btl->ep_lookup, opal_hash_table_t);
opal_hash_table_init(ud_btl->ep_lookup, nprocs);
}
#endif
for(i = 0; i < (int)nprocs; i++) {
struct ompi_proc_t* ompi_proc = ompi_procs[i];
mca_btl_ud_proc_t* ib_proc;
mca_btl_base_endpoint_t* ib_peer;
if(NULL == (ib_proc = mca_btl_ud_proc_create(ompi_proc))) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* The btl_proc datastructure is shared by all IB BTL instances that are
* trying to reach this destination. Cache the peer instance on the
* btl_proc.
*/
ib_peer = OBJ_NEW(mca_btl_ud_endpoint_t);
if(NULL == ib_peer) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
OPAL_THREAD_LOCK(&ib_proc->proc_lock);
rc = mca_btl_ud_proc_insert(ib_proc, ib_peer);
if(rc != OMPI_SUCCESS) {
OBJ_RELEASE(ib_peer);
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
continue;
}
BTL_VERBOSE(("modex_recv QP num %d, LID = %d",
ib_peer->rem_addr.qp_num, ib_peer->rem_addr.lid));
/* Set up IB address handles for the endpoint */
ah_attr.is_global = 0;
ah_attr.dlid = ib_peer->rem_addr.lid;
ah_attr.sl = mca_btl_ofud_component.ib_service_level;
ah_attr.src_path_bits = mca_btl_ofud_component.ib_src_path_bits;
ah_attr.port_num = ud_btl->ib_port_num;
ib_peer->rmt_ah = ibv_create_ah(ud_btl->ib_pd, &ah_attr);
if(NULL == ib_peer->rmt_ah) {
BTL_ERROR(("error creating address handle: %s\n", strerror(errno)));
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
continue;
}
/* Insert a pointer to the endpoint in the BTL lookup table */
#if 0
opal_hash_table_set_value_uint64(ud_btl->ep_lookup,
((uint64_t)ib_peer->rem_addr.lid << 32) |
ib_peer->rem_addr.qp_num,
ib_peer);
#endif
ompi_bitmap_set_bit(reachable, i);
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
peers[i] = ib_peer;
}
return OMPI_SUCCESS;
}
/*
* Delete the proc as reachable from this btl module
*/
int mca_btl_ud_del_procs(struct mca_btl_base_module_t* btl,
size_t nprocs,
struct ompi_proc_t** procs,
struct mca_btl_base_endpoint_t** peers)
{
size_t i;
for(i = 0; i < nprocs; i++) {
mca_btl_ud_endpoint_t* endpoint = (mca_btl_ud_endpoint_t*)peers[i];
mca_btl_ud_proc_t* proc = mca_btl_ud_proc_lookup_ompi(procs[i]);
#if 0
opal_hash_table_remove_value_uint64(ud_btl->ep_lookup,
((uint64_t)endpoint->rem_addr.lid << 32) |
endpoint->rem_addr.qp_num);
#endif
if(NULL != proc) {
mca_btl_ud_proc_remove(proc, endpoint);
}
OBJ_RELEASE(endpoint);
}
return OMPI_SUCCESS;
}
/*
* Register callback function to support send/recv semantics
*/
int mca_btl_ud_register(struct mca_btl_base_module_t* btl,
mca_btl_base_tag_t tag,
mca_btl_base_module_recv_cb_fn_t cbfunc,
void* cbdata)
{
mca_btl_ud_module_t* ud_btl = (mca_btl_ud_module_t*)btl;
OPAL_THREAD_LOCK(&ud_btl->ud_lock);
ud_btl->ib_reg[tag].cbfunc = cbfunc;
ud_btl->ib_reg[tag].cbdata = cbdata;
OPAL_THREAD_UNLOCK(&ud_btl->ud_lock);
return OMPI_SUCCESS;
}
/**
* Allocate a segment.
*
* @param btl (IN) BTL module
* @param size (IN) Request segment size.
*
* When allocating a segment we pull a pre-alllocated segment
* from one of two free lists, an eager list and a max list
*/
mca_btl_base_descriptor_t* mca_btl_ud_alloc(struct mca_btl_base_module_t* btl,
uint8_t order,
size_t size)
{
mca_btl_ud_frag_t* frag = NULL;
int rc;
if(OPAL_LIKELY(size <= mca_btl_ofud_module.super.btl_eager_limit)) {
MCA_BTL_UD_ALLOC_FRAG(btl, frag, rc);
}
if(NULL == frag) {
return NULL;
}
frag->base.order = MCA_BTL_NO_ORDER;
frag->segment.seg_len = size;
return (mca_btl_base_descriptor_t*)frag;
}
/**
* Return a segment
*
* Return the segment to the appropriate
* preallocated segment list
*/
int mca_btl_ud_free(struct mca_btl_base_module_t* btl,
mca_btl_base_descriptor_t* des)
{
mca_btl_ud_frag_t* frag = (mca_btl_ud_frag_t*)des;
if(OPAL_LIKELY(frag->type == MCA_BTL_UD_FRAG_SEND)) {
MCA_BTL_UD_RETURN_FRAG(btl, frag);
} else if(frag->type == MCA_BTL_UD_FRAG_USER && frag->ud_reg != NULL) {
btl->btl_mpool->mpool_deregister(btl->btl_mpool,
(mca_mpool_base_registration_t*)frag->ud_reg);
MCA_BTL_UD_RETURN_USER_FRAG(btl, frag);
}
return OMPI_SUCCESS;
}
/**
* register user buffer or pack
* data into pre-registered buffer and return a
* descriptor that can be
* used for send/put.
*
* @param btl (IN) BTL module
* @param peer (IN) BTL peer addressing
*
* prepare source's behavior depends on the following:
* Has a valid memory registration been passed to prepare_src?
* if so we attempt to use the pre-registred user-buffer, if the memory
* registration is to small (only a portion of the user buffer) then we must
* reregister the user buffer
* Has the user requested the memory to be left pinned?
* if so we insert the memory registration into a memory tree for later
* lookup, we may also remove a previous registration if a MRU (most recently
* used) list of registions is full, this prevents resources from being
* exhausted.
* Is the requested size larger than the btl's max send size?
* if so and we aren't asked to leave the registration pinned then we
* register the memory if the user's buffer is contiguous.
* Otherwise we choose from two free lists of pre-registered memory in which
* to pack the data into.
*
*/
mca_btl_base_descriptor_t* mca_btl_ud_prepare_src(
struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
mca_mpool_base_registration_t* registration,
struct ompi_convertor_t* convertor,
uint8_t order,
size_t reserve,
size_t* size)
{
mca_btl_ud_frag_t* frag = NULL;
struct iovec iov;
uint32_t iov_count = 1;
size_t max_data = *size;
int rc;
if(ompi_convertor_need_buffers(convertor) == 0 && reserve == 0 &&
(registration != NULL || max_data > btl->btl_max_send_size)) {
/* The user buffer is contigous and we are asked to send more than
the max send size. */
MCA_BTL_UD_ALLOC_USER_FRAG(btl, frag, rc);
if(NULL == frag) {
return NULL;
}
iov.iov_len = max_data;
iov.iov_base = NULL;
ompi_convertor_pack(convertor, &iov, &iov_count, &max_data);
frag->segment.seg_len = max_data;
frag->segment.seg_addr.pval = iov.iov_base;
frag->base.des_flags = 0;
frag->base.order = MCA_BTL_NO_ORDER;
if(NULL == registration) {
rc = btl->btl_mpool->mpool_register(btl->btl_mpool, iov.iov_base,
max_data, 0, &registration);
if(OMPI_SUCCESS != rc || NULL == registration) {
MCA_BTL_UD_RETURN_USER_FRAG(btl, frag);
}
return NULL;
}
frag->ud_reg = (mca_btl_ud_reg_t*)registration;
frag->sg_entry.lkey = frag->ud_reg->mr->lkey;
frag->sg_entry.addr = (unsigned long)iov.iov_base;
frag->base.des_src = &frag->segment;
frag->base.des_src_cnt = 1;
frag->base.des_dst = NULL;
frag->base.des_dst_cnt = 0;
*size = max_data;
return &frag->base;
}
if(max_data + reserve > btl->btl_eager_limit) {
max_data = btl->btl_eager_limit - reserve;
}
MCA_BTL_UD_ALLOC_FRAG(btl, frag, rc);
if(OPAL_UNLIKELY(NULL == frag)) {
return NULL;
}
iov.iov_len = max_data;
iov.iov_base = (unsigned char*)frag->segment.seg_addr.pval + reserve;
rc = ompi_convertor_pack(convertor, &iov, &iov_count, &max_data);
if(OPAL_UNLIKELY(rc < 0)) {
MCA_BTL_UD_RETURN_FRAG(btl, frag);
return NULL;
}
frag->segment.seg_len = max_data + reserve;
frag->sg_entry.length =
max_data + reserve + sizeof(mca_btl_ud_header_t);
frag->base.des_src = &frag->segment;
frag->base.des_src_cnt = 1;
frag->base.des_dst = NULL;
frag->base.des_dst_cnt = 0;
frag->base.des_flags = 0;
frag->base.order = MCA_BTL_NO_ORDER;
*size = max_data;
return &frag->base;
}
int mca_btl_ud_finalize(struct mca_btl_base_module_t* btl)
{
mca_btl_ud_module_t* ud_btl = (mca_btl_ud_module_t*)btl;
int32_t i;
for(i = 0; i < MCA_BTL_UD_NUM_QP; i++) {
ibv_destroy_qp(ud_btl->ib_qp[i]);
}
ibv_dealloc_pd(ud_btl->ib_pd);
OBJ_DESTRUCT(&ud_btl->ud_lock);
OBJ_DESTRUCT(&ud_btl->pending_frags);
OBJ_DESTRUCT(&ud_btl->send_frags);
OBJ_DESTRUCT(&ud_btl->user_frags);
OBJ_DESTRUCT(&ud_btl->recv_frags);
mca_mpool_base_module_destroy(ud_btl->super.btl_mpool);
return OMPI_SUCCESS;
}
/*
* Initiate a send.
*/
int mca_btl_ud_send(struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct mca_btl_base_descriptor_t* descriptor,
mca_btl_base_tag_t tag)
{
int rc;
mca_btl_ud_frag_t* frag = (mca_btl_ud_frag_t*)descriptor;
MCA_BTL_UD_START_TIME(post_send);
frag->endpoint = endpoint;
frag->hdr->tag = tag;
rc = mca_btl_ud_endpoint_post_send((mca_btl_ud_module_t*)btl, frag);
MCA_BTL_UD_END_TIME(post_send);
return rc;
}
/*
* RDMA Memory Pool (de)register callbacks
*/
static int mca_btl_ud_reg_mr(void* reg_data, void* base, size_t size,
mca_mpool_base_registration_t* reg)
{
mca_btl_ud_module_t* mod = (mca_btl_ud_module_t*)reg_data;
mca_btl_ud_reg_t* ud_reg = (mca_btl_ud_reg_t*)reg;
ud_reg->mr = ibv_reg_mr(mod->ib_pd, base, size, IBV_ACCESS_LOCAL_WRITE |
IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_READ);
if(NULL == ud_reg->mr)
return OMPI_ERR_OUT_OF_RESOURCE;
return OMPI_SUCCESS;
}
static int mca_btl_ud_dereg_mr(void* reg_data,
mca_mpool_base_registration_t* reg)
{
mca_btl_ud_reg_t* ud_reg = (mca_btl_ud_reg_t*)reg;
if(ud_reg->mr != NULL) {
if(ibv_dereg_mr(ud_reg->mr)) {
opal_output(0, "%s: error unpinning UD memory: %s\n",
__func__, strerror(errno));
return OMPI_ERROR;
}
}
ud_reg->mr = NULL;
return OMPI_SUCCESS;
}
/*
* Create a single UD queue pair. Since UD is connectionless, the QP is
* useable immediately.
*/
/* TODO - can remove cq/psn args now with only one type of frag */
static int mca_btl_ud_init_qp(mca_btl_ud_module_t* ud_btl,
struct ibv_cq* cq,
struct ibv_qp** qp,
uint32_t lcl_psn)
{
struct ibv_qp_attr qp_attr;
struct ibv_qp_init_attr qp_init_attr;
memset(&qp_init_attr, 0, sizeof(struct ibv_qp_init_attr));
qp_init_attr.send_cq = cq;
qp_init_attr.recv_cq = cq;
qp_init_attr.cap.max_send_wr = mca_btl_ofud_component.sd_num;
qp_init_attr.cap.max_recv_wr = mca_btl_ofud_component.rd_num;
qp_init_attr.cap.max_send_sge = 1;
qp_init_attr.cap.max_recv_sge = 1;
/* TODO - find the best value for max_inline_data */
qp_init_attr.cap.max_inline_data = 200;
qp_init_attr.qp_type = IBV_QPT_UD;
*qp = ibv_create_qp(ud_btl->ib_pd, &qp_init_attr);
if(NULL == *qp) {
BTL_ERROR(("error creating QP: %s\n", strerror(errno)));
return OMPI_ERROR;
}
if(0 == (ud_btl->ib_inline_max = qp_init_attr.cap.max_inline_data)) {
BTL_ERROR(("ibv_create_qp: returned 0 byte(s) for max inline data"));
}
BTL_VERBOSE((0, "ib_inline_max %d\n", ud_btl->ib_inline_max));
qp_attr.qp_state = IBV_QPS_INIT;
qp_attr.pkey_index = mca_btl_ofud_component.ib_pkey_ix;
qp_attr.qkey = mca_btl_ofud_component.ib_qkey;
qp_attr.port_num = ud_btl->ib_port_num;
if(ibv_modify_qp(*qp, &qp_attr,
IBV_QP_STATE | IBV_QP_PKEY_INDEX |
IBV_QP_PORT | IBV_QP_QKEY)) {
BTL_ERROR(("error modifying QP to INIT: %s", strerror(errno)));
goto destroy_qp;
}
qp_attr.qp_state = IBV_QPS_RTR;
if(ibv_modify_qp(*qp, &qp_attr, IBV_QP_STATE)) {
BTL_ERROR(("error modifing QP to RTR: %s", strerror(errno)));
goto destroy_qp;
}
qp_attr.qp_state = IBV_QPS_RTS;
qp_attr.sq_psn = lcl_psn;
if (ibv_modify_qp(*qp, &qp_attr, IBV_QP_STATE | IBV_QP_SQ_PSN)) {
BTL_ERROR(("error modifying QP to RTS: %s", strerror(errno)));
goto destroy_qp;
}
return OMPI_SUCCESS;
destroy_qp:
ibv_destroy_qp(*qp);
*qp = NULL;
return OMPI_ERROR;
}
/*
* Initialize the btl module by allocating a protection domain,
* memory pool, completion queue, and free lists
*/
int mca_btl_ud_module_init(mca_btl_ud_module_t *ud_btl)
{
struct mca_mpool_base_resources_t mpool_resources;
struct ibv_context *ctx = ud_btl->ib_dev_context;
struct ibv_recv_wr* bad_wr;
mca_btl_ud_frag_t* frag;
ompi_free_list_item_t* item;
uint32_t length;
int32_t rc, i;
ud_btl->sd_wqe = mca_btl_ofud_component.sd_num;
ud_btl->ib_pd = ibv_alloc_pd(ctx);
if(NULL == ud_btl->ib_pd) {
BTL_ERROR(("error allocating PD for %s: %s\n",
ibv_get_device_name(ud_btl->ib_dev), strerror(errno)));
return OMPI_ERROR;
}
mpool_resources.reg_data = (void*)ud_btl;
mpool_resources.sizeof_reg = sizeof(mca_btl_ud_reg_t);
mpool_resources.register_mem = mca_btl_ud_reg_mr;
mpool_resources.deregister_mem = mca_btl_ud_dereg_mr;
ud_btl->super.btl_mpool =
mca_mpool_base_module_create(mca_btl_ofud_component.ud_mpool_name,
&ud_btl->super, &mpool_resources);
if(NULL == ud_btl->super.btl_mpool) {
BTL_ERROR(("error creating IB mpool for %s: %s\n",
ibv_get_device_name(ud_btl->ib_dev), strerror(errno)));
goto dealloc_pd;
}
/* Create the completion queue */
length = mca_btl_ofud_component.rd_num + mca_btl_ofud_component.sd_num;
ud_btl->ib_cq = ibv_create_cq(ctx, length, NULL, NULL, 0);
if(NULL == ud_btl->ib_cq) {
BTL_ERROR(("error creating CQ for %s: %s\n",
ibv_get_device_name(ud_btl->ib_dev), strerror(errno)));
goto mpool_destroy;
}
/* Set up our packet sequence numbers */
ud_btl->addr.psn = lrand48() & 0xffffff;
/* Set up the QPs for this BTL */
for(i = 0; i < MCA_BTL_UD_NUM_QP; i++) {
if(OMPI_SUCCESS != mca_btl_ud_init_qp(ud_btl,
ud_btl->ib_cq, &ud_btl->ib_qp[i], ud_btl->addr.psn)) {
goto qp_destroy;
}
}
/* Place our QP numbers in our local address information */
ud_btl->addr.qp_num = ud_btl->ib_qp[0]->qp_num;
ud_btl->ib_qp_next = 0;
/*ud_btl->rd_posted = mca_btl_ofud_component.rd_num_init;*/
/* Initialize pool of receive fragments first, since an error may occur */
/* TODO - no need for a free list with a static buffer count */
OBJ_CONSTRUCT(&ud_btl->recv_frags, ompi_free_list_t);
length = sizeof(mca_btl_ud_frag_t) + sizeof(mca_btl_ud_header_t) +
ud_btl->super.btl_eager_limit + 2 * MCA_BTL_IB_FRAG_ALIGN;
ompi_free_list_init(&ud_btl->recv_frags,
length + sizeof(mca_btl_ud_ib_header_t),
OBJ_CLASS(mca_btl_ud_recv_frag_t),
mca_btl_ofud_component.rd_num,
mca_btl_ofud_component.rd_num,
mca_btl_ofud_component.rd_num,
ud_btl->super.btl_mpool);
#if 0
ompi_free_list_init(&ud_btl->recv_frags,
length + sizeof(mca_btl_ud_ib_header_t),
OBJ_CLASS(mca_btl_ud_recv_frag_t),
mca_btl_ofud_component.rd_num_init,
mca_btl_ofud_component.rd_num_max,
mca_btl_ofud_component.rd_num_inc,
ud_btl->super.btl_mpool);
#endif
/* Post receive descriptors */
for(i = 0; i < mca_btl_ofud_component.rd_num; i++) {
OMPI_FREE_LIST_GET(&ud_btl->recv_frags, item, rc);
frag = (mca_btl_ud_frag_t*)item;
if(NULL == frag) {
BTL_ERROR(("error getting receive buffer from free list\n"));
goto obj_destruct;
}
frag->type = MCA_BTL_UD_FRAG_RECV;
frag->sg_entry.length = mca_btl_ofud_module.super.btl_eager_limit +
sizeof(mca_btl_ud_header_t) + sizeof(mca_btl_ud_ib_header_t);
if(ibv_post_recv(ud_btl->ib_qp[0],
&frag->wr_desc.rd_desc, &bad_wr)) {
BTL_ERROR(("error posting recv, errno %s\n", strerror(errno)));
goto obj_destruct;
}
}
/* No more errors anticipated - initialize everything else */
OBJ_CONSTRUCT(&ud_btl->ud_lock, opal_mutex_t);
OBJ_CONSTRUCT(&ud_btl->pending_frags, opal_list_t);
OBJ_CONSTRUCT(&ud_btl->send_frags, ompi_free_list_t);
OBJ_CONSTRUCT(&ud_btl->user_frags, ompi_free_list_t);
ompi_free_list_init(&ud_btl->send_frags,
length,
OBJ_CLASS(mca_btl_ud_send_frag_t),
mca_btl_ofud_component.sd_num >> 1,
mca_btl_ofud_component.sd_num << 2,
mca_btl_ofud_component.sd_num >> 3,
ud_btl->super.btl_mpool);
/* Initialize pool of user fragments */
length = sizeof(mca_btl_ud_frag_t) +
sizeof(mca_btl_ud_header_t) + 2 * MCA_BTL_IB_FRAG_ALIGN;
ompi_free_list_init(&ud_btl->user_frags,
length,
OBJ_CLASS(mca_btl_ud_user_frag_t),
mca_btl_ofud_component.sd_num >> 1,
mca_btl_ofud_component.sd_num << 2,
mca_btl_ofud_component.sd_num >> 3,
ud_btl->super.btl_mpool);
return OMPI_SUCCESS;
obj_destruct:
OBJ_DESTRUCT(&ud_btl->recv_frags);
qp_destroy:
for(i = 0; i < MCA_BTL_UD_NUM_QP; i++) {
ibv_destroy_qp(ud_btl->ib_qp[i]);
}
mpool_destroy:
mca_mpool_base_module_destroy(ud_btl->super.btl_mpool);
dealloc_pd:
ibv_dealloc_pd(ud_btl->ib_pd);
return OMPI_ERROR;
}
int mca_btl_udapl_ft_event(int state) {
return OMPI_SUCCESS;
}

374
ompi/mca/btl/ofud/btl_ofud.h Обычный файл
Просмотреть файл

@ -0,0 +1,374 @@
/*
* 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef MCA_BTL_UD_H
#define MCA_BTL_UD_H
/* Number of QP's to stripe sends over - keep this as power of 2 */
/* AWF - This is intentionally NOT an MCA parameter so that I can do fast
modular arithmetic with it. */
#define MCA_BTL_UD_NUM_QP 4
/* Standard system includes */
#include <sys/types.h>
#include <infiniband/verbs.h>
/* Open MPI includes */
#include "opal/class/opal_hash_table.h"
#include "opal/util/output.h"
#include "opal/sys/timer.h"
#include "ompi/class/ompi_free_list.h"
#include "ompi/class/ompi_bitmap.h"
#include "ompi/mca/btl/btl.h"
#include "ompi/mca/btl/base/btl_base_error.h"
#include "ompi/mca/btl/base/base.h"
#include "ompi/mca/mpool/mpool.h"
/* TODO - If I want this to go away, addr_t has to come over here */
#include "btl_ofud_endpoint.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
/**
* UD Infiniband (IB) BTL component.
*/
struct mca_btl_ud_component_t {
mca_btl_base_component_1_0_1_t super; /**< base BTL component */
uint32_t max_btls; /**< Maximum number of BTL modules */
uint32_t num_btls; /**< Number of available/initialized BTL modules */
struct mca_btl_ud_module_t* ud_btls; /**< array of available BTLs */
opal_list_t ud_procs; /**< list of ib proc structures */
opal_mutex_t ud_lock; /**< lock for accessing component state */
char* ud_mpool_name; /**< name of memory pool */
int32_t sd_num; /**< max send descriptors to post per BTL */
int32_t sd_num_peer; /**< max send descriptors to post per endpoint */
int32_t rd_num; /**< number of receive descriptors per BTL */
#if 0
int32_t rd_num_init; /**< initial recv descriptors to post per BTL */
int32_t rd_num_max;
int32_t rd_num_inc;
#endif
uint32_t ib_pkey_ix;
uint32_t ib_qkey;
uint32_t ib_service_level;
uint32_t ib_src_path_bits;
}; typedef struct mca_btl_ud_component_t mca_btl_ud_component_t;
OMPI_MODULE_DECLSPEC extern mca_btl_ud_component_t mca_btl_ofud_component;
typedef mca_btl_base_recv_reg_t mca_btl_ud_recv_reg_t;
/**
* Profiling variables
*/
#if OMPI_ENABLE_DEBUG
#define MCA_BTL_UD_ENABLE_PROFILE 0
#else
#define MCA_BTL_UD_ENABLE_PROFILE 0
#endif
#if MCA_BTL_UD_ENABLE_PROFILE
#define MCA_BTL_UD_PROFILE_VAR(var) \
opal_timer_t avg_ ## var; \
opal_timer_t cnt_ ## var; \
opal_timer_t tmp_ ## var
struct mca_btl_ud_profile_t {
MCA_BTL_UD_PROFILE_VAR(post_send);
MCA_BTL_UD_PROFILE_VAR(ibv_post_send);
};
typedef struct mca_btl_ud_profile_t mca_btl_ud_profile_t;
extern mca_btl_ud_profile_t mca_btl_ud_profile;
#endif
/**
* UD/IB BTL Interface
*/
struct mca_btl_ud_module_t {
mca_btl_base_module_t super;
mca_btl_ud_recv_reg_t ib_reg[256]; /* protected by ib_lock */
uint8_t ib_port_num;
struct ibv_device* ib_dev;
struct ibv_context* ib_dev_context;
struct ibv_pd* ib_pd;
struct ibv_cq* ib_cq;
struct mca_btl_ud_addr_t addr; /**< local address information */
ompi_free_list_t send_frags; /**< send fragments & buffers */
ompi_free_list_t user_frags; /**< user data fragments */
ompi_free_list_t recv_frags; /**< receive fragments & buffers */
opal_list_t pending_frags; /**< list of pending send frags */
opal_mutex_t ud_lock; /**< lock for ib_reg and pending_frags */
size_t ib_inline_max; /**< max size of IB inline send */
/*int32_t rd_posted;*/ /**< number of receives currently posted */
int32_t sd_wqe; /**< available send WQ entries */
/* No lock needed, these are incremented/decremented atomically */
/*opal_hash_table_t* ep_lookup;*/
/**< hash table for fast lookup of endpoint structures in recv path */
/* lid:qpnum is key, value is mca_btl_ud_endpoint_t* */
struct ibv_qp* ib_qp[MCA_BTL_UD_NUM_QP];
uint32_t ib_qp_next;
/**< Local QPs and stripe counters */
/* No lock needed - counters only ever increase by 1 */
}; typedef struct mca_btl_ud_module_t mca_btl_ud_module_t;
struct mca_btl_ud_frag_t;
extern mca_btl_ud_module_t mca_btl_ofud_module;
/**
* Open the component; register UD/IB parameters with the MCA framework
*/
extern int mca_btl_ud_component_open(void);
/**
* Any final cleanup before being unloaded.
*/
extern int mca_btl_ud_component_close(void);
/**
* IB component initialization.
*
* @param num_btl_modules (OUT)
* Number of BTLs returned in BTL array.
* @param allow_multi_user_threads (OUT)
* Flag indicating wether BTL supports user threads (TRUE)
* @param have_hidden_threads (OUT)
* Flag indicating whether BTL uses threads (TRUE)
*
* (1) read interface list from verbs and compare against component parameters
* then create a BTL instance for selected interfaces
* (2) publish BTL addressing info
*/
extern mca_btl_base_module_t** mca_btl_ud_component_init(
int *num_btl_modules,
bool allow_multi_user_threads,
bool have_hidden_threads);
/**
* UD/IB component progress.
*/
extern int mca_btl_ud_component_progress(void);
/**
* Register a callback function that is called on receipt
* of a fragment.
*
* @param btl (IN) BTL module
* @return Status indicating if cleanup was successful
*/
int mca_btl_ud_register(struct mca_btl_base_module_t* btl,
mca_btl_base_tag_t tag,
mca_btl_base_module_recv_cb_fn_t cbfunc,
void* cbdata);
/**
* Cleanup any resources held by the BTL.
*
* @param btl BTL instance.
* @return OMPI_SUCCESS or error status on failure.
*/
extern int mca_btl_ud_finalize(struct mca_btl_base_module_t* btl);
/**
* PML->BTL notification of change in the process list.
*
* @param btl (IN)
* @param nprocs (IN) Number of processes
* @param procs (IN) Set of processes
* @param peers (OUT) Set of (optional) peer addressing info.
* @param peers (IN/OUT) Set of processes that are reachable via this BTL.
* @return OMPI_SUCCESS or error status on failure.
*/
extern int mca_btl_ud_add_procs(struct mca_btl_base_module_t* btl,
size_t nprocs,
struct ompi_proc_t **procs,
struct mca_btl_base_endpoint_t** peers,
ompi_bitmap_t* reachable);
/**
* PML->BTL notification of change in the process list.
*
* @param btl (IN) BTL instance
* @param nproc (IN) Number of processes.
* @param procs (IN) Set of processes.
* @param peers (IN) Set of peer data structures.
* @return Status indicating if cleanup was successful
*
*/
extern int mca_btl_ud_del_procs(struct mca_btl_base_module_t* btl,
size_t nprocs,
struct ompi_proc_t **procs,
struct mca_btl_base_endpoint_t** peers);
/**
* PML->BTL Initiate a send of the specified size.
*
* @param btl (IN)
* BTL instance
* @param btl_base_peer (IN)
* BTL peer addressing
* @param send_request (IN/OUT)
* Send request (allocated by PML via mca_btl_base_request_alloc_fn_t)
* @param size (IN)
* Number of bytes PML is requesting BTL to deliver
* @param flags (IN)
* Flags that should be passed to the peer via the message header.
* @param request (OUT)
* OMPI_SUCCESS if the BTL was able to queue one or more fragments
*/
extern int mca_btl_ud_send(struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* btl_peer,
struct mca_btl_base_descriptor_t* descriptor,
mca_btl_base_tag_t tag);
/**
* Allocate a descriptor.
*
* @param btl (IN) BTL module
* @param size (IN) Requested descriptor size.
*/
extern mca_btl_base_descriptor_t* mca_btl_ud_alloc(
struct mca_btl_base_module_t* btl,
uint8_t order,
size_t size);
/**
* Return a segment allocated by this BTL.
*
* @param btl (IN) BTL module
* @param descriptor (IN) Allocated descriptor.
*/
extern int mca_btl_ud_free(struct mca_btl_base_module_t* btl,
mca_btl_base_descriptor_t* des);
/**
* Pack data and return a descriptor that can be
* used for send/put.
*
* @param btl (IN) BTL module
* @param peer (IN) BTL peer addressing
*/
mca_btl_base_descriptor_t* mca_btl_ud_prepare_src(
struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* peer,
mca_mpool_base_registration_t* registration,
struct ompi_convertor_t* convertor,
uint8_t order,
size_t reserve,
size_t* size);
int mca_btl_ud_module_init(mca_btl_ud_module_t* ud_btl);
/**
* Fault Tolerance Event Notification Function
* @param state Checkpoint State
* @return OMPI_SUCCESS or failure status
*/
extern int mca_btl_udapl_ft_event(int state);
/*
* Profiling stuff
*/
#if MCA_BTL_UD_ENABLE_PROFILE
#define MCA_BTL_UD_START_TIME(var) \
((mca_btl_ud_profile.tmp_ ## var) = opal_sys_timer_get_cycles())
#define MCA_BTL_UD_END_TIME(var) \
do { \
mca_btl_ud_profile.avg_ ## var += \
opal_sys_timer_get_cycles() - mca_btl_ud_profile.tmp_ ## var; \
mca_btl_ud_profile.cnt_ ## var++; \
} while(0)
#define MCA_BTL_UD_SHOW_TIME(var) \
OPAL_OUTPUT((0, " " #var " avg %lu cnt %lu", \
(mca_btl_ud_profile.avg_ ## var) / (mca_btl_ud_profile.cnt_ ## var), \
mca_btl_ud_profile.cnt_ ## var));
#else
#define MCA_BTL_UD_START_TIME(var)
#define MCA_BTL_UD_END_TIME(var)
#define MCA_BTL_UD_SHOW_TIME(var)
#endif
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

537
ompi/mca/btl/ofud/btl_ofud_component.c Обычный файл
Просмотреть файл

@ -0,0 +1,537 @@
/*
* 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include <errno.h>
#include <string.h>
#include <infiniband/verbs.h>
#include "ompi_config.h"
#include "ompi/constants.h"
#include "opal/prefetch.h"
#include "opal/util/output.h"
#include "ompi/mca/btl/btl.h"
#include "opal/sys/timer.h"
#include "opal/mca/base/mca_base_param.h"
#include "orte/mca/errmgr/errmgr.h"
#include "ompi/mca/btl/base/base.h"
#include "ompi/mca/mpool/rdma/mpool_rdma.h"
#include "ompi/mca/pml/base/pml_base_module_exchange.h"
#include "btl_ofud.h"
#include "btl_ofud_frag.h"
#include "btl_ofud_endpoint.h"
mca_btl_ud_component_t mca_btl_ofud_component = {
{
/* First, the mca_base_component_t struct containing meta information
about the component itself */
{
/* Indicate that we are a pml v1.0.0 component (which also implies a
specific MCA version) */
MCA_BTL_BASE_VERSION_1_0_1,
"ofud", /* MCA component name */
OMPI_MAJOR_VERSION, /* MCA component major version */
OMPI_MINOR_VERSION, /* MCA component minor version */
OMPI_RELEASE_VERSION, /* MCA component release version */
mca_btl_ud_component_open, /* component open */
mca_btl_ud_component_close /* component close */
},
/* Next the MCA v1.0.0 component meta data */
{
/* Whether the component is checkpointable or not */
false
},
mca_btl_ud_component_init,
mca_btl_ud_component_progress,
}
};
/*
* Profiling information
*/
#if MCA_BTL_UD_ENABLE_PROFILE
mca_btl_ud_profile_t mca_btl_ud_profile = {0};
#endif
/*
* utility routines for parameter registration
*/
static inline void mca_btl_ud_param_reg_string(const char* param_name,
const char* param_desc,
const char* default_value,
char** out_value)
{
mca_base_param_reg_string(&mca_btl_ofud_component.super.btl_version,
param_name, param_desc, false, false,
default_value, out_value);
}
static inline void mca_btl_ud_param_reg_int(const char* param_name,
const char* param_desc,
int default_value,
int* out_value)
{
mca_base_param_reg_int(&mca_btl_ofud_component.super.btl_version,
param_name, param_desc, false, false,
default_value, out_value);
}
/*
* Called by MCA framework to open the component, registers
* component parameters.
*/
int mca_btl_ud_component_open(void)
{
int val;
/* initialize state */
mca_btl_ofud_component.num_btls = 0;
mca_btl_ofud_component.ud_btls = NULL;
/* initialize objects */
OBJ_CONSTRUCT(&mca_btl_ofud_component.ud_lock, opal_mutex_t);
OBJ_CONSTRUCT(&mca_btl_ofud_component.ud_procs, opal_list_t);
/* register IB component parameters */
mca_btl_ud_param_reg_int("max_btls",
"Maximum number of HCAs/ports to use",
4, (int*)&mca_btl_ofud_component.max_btls);
mca_btl_ud_param_reg_string("mpool", "Name of the memory pool to be used",
"rdma", &mca_btl_ofud_component.ud_mpool_name);
mca_btl_ud_param_reg_int("ib_pkey_index", "IB pkey index",
0, (int*)&mca_btl_ofud_component.ib_pkey_ix);
mca_btl_ud_param_reg_int("ib_qkey", "IB qkey",
0x01330133, (int*)&mca_btl_ofud_component.ib_qkey);
mca_btl_ud_param_reg_int("ib_service_level", "IB service level",
0, (int*)&mca_btl_ofud_component.ib_service_level);
mca_btl_ud_param_reg_int("ib_src_path_bits", "IB source path bits",
0, (int*)&mca_btl_ofud_component.ib_src_path_bits);
mca_btl_ud_param_reg_int("sd_num", "maximum send descriptors to post",
128, (int*)&mca_btl_ofud_component.sd_num);
mca_btl_ud_param_reg_int("sd_num_peer",
"maximum send descriptors to post to one peer",
8, (int*)&mca_btl_ofud_component.sd_num_peer);
mca_btl_ud_param_reg_int("rd_num_init", "number of receive buffers",
6000, (int*)&mca_btl_ofud_component.rd_num);
#if 0
mca_btl_ud_param_reg_int("rd_num_init", "initial receive buffers",
3000, (int*)&mca_btl_ofud_component.rd_num_init);
mca_btl_ud_param_reg_int("rd_num_max", "maximum receive buffers",
4500, (int*)&mca_btl_ofud_component.rd_num_max);
mca_btl_ud_param_reg_int("rd_num_inc",
"number of buffers to post when rate is high",
25, (int*)&mca_btl_ofud_component.rd_num_inc);
#endif
/* TODO - this assumes a 2k UD MTU - query/do something more intelligent */
/*mca_btl_ud_param_reg_int("eager_limit", "eager send limit",
2048, &val); */
mca_btl_ud_param_reg_int("min_send_size", "minimum send size",
2048, &val);
mca_btl_ofud_module.super.btl_min_send_size = val;
mca_btl_ud_param_reg_int("max_send_size", "maximum send size",
2048, &val);
mca_btl_ofud_module.super.btl_eager_limit = val;
mca_btl_ofud_module.super.btl_max_send_size = val;
mca_btl_ud_param_reg_int("exclusivity", "BTL exclusivity",
MCA_BTL_EXCLUSIVITY_DEFAULT,
(int*)&mca_btl_ofud_module.super.btl_exclusivity);
mca_btl_ud_param_reg_int("bandwidth",
"Approximate maximum bandwidth of interconnect",
800, (int*)&mca_btl_ofud_module.super.btl_bandwidth);
mca_btl_ofud_module.super.btl_eager_limit -= sizeof(mca_btl_ud_header_t);
mca_btl_ofud_module.super.btl_max_send_size -= sizeof(mca_btl_ud_header_t);
return OMPI_SUCCESS;
}
/*
* Component cleanup
*/
int mca_btl_ud_component_close(void)
{
OBJ_DESTRUCT(&mca_btl_ofud_component.ud_lock);
OBJ_DESTRUCT(&mca_btl_ofud_component.ud_procs);
/* Calculate and print profiling numbers */
MCA_BTL_UD_SHOW_TIME(post_send);
MCA_BTL_UD_SHOW_TIME(ibv_post_send);
return OMPI_SUCCESS;
}
/*
* Register UD address information. The MCA framework
* will make this available to all peers.
*/
static int mca_btl_ud_modex_send(void)
{
int rc;
size_t i;
size_t size;
mca_btl_ud_addr_t* addrs = NULL;
size = mca_btl_ofud_component.num_btls * sizeof(mca_btl_ud_addr_t);
if(size != 0) {
addrs = (mca_btl_ud_addr_t*)malloc(size);
if(NULL == addrs) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
for(i = 0; i < mca_btl_ofud_component.num_btls; i++) {
mca_btl_ud_module_t* btl = &mca_btl_ofud_component.ud_btls[i];
addrs[i] = btl->addr;
BTL_VERBOSE((0, "modex_send QP num %x, LID = %x",
addrs[i].qp_num, addrs[i].lid));
}
}
rc = mca_pml_base_modex_send(
&mca_btl_ofud_component.super.btl_version, addrs, size);
if(NULL != addrs) {
free(addrs);
}
return rc;
}
/*
* UD component initialization:
* (1) read interface list from kernel and compare against component parameters
* then create a BTL instance for selected interfaces
* (2) post OOB receive for incoming connection attempts
* (3) register BTL parameters with the MCA
*/
mca_btl_base_module_t** mca_btl_ud_component_init(int* num_btl_modules,
bool enable_progress_threads,
bool enable_mpi_threads)
{
struct ibv_device **ib_devs;
struct ibv_device* ib_dev;
int32_t num_devs;
mca_btl_base_module_t** btls;
uint32_t i, j;
opal_list_t btl_list;
mca_btl_ud_module_t* ud_btl;
mca_btl_base_selected_module_t* ib_selected;
opal_list_item_t* item;
unsigned short seedv[3];
char* btl_str;
char* tok;
/* First, check if the UD BTL was specifically selected.
If not, then short out right away. */
mca_base_param_lookup_string(
mca_base_param_find("btl", NULL, NULL), &btl_str);
if(NULL == btl_str) {
/* Can't specify UD with out any string at all.. bail out */
return NULL;
}
/* Try to find a 'ud' token */
tok = strtok(btl_str, ",");
while(tok) {
if(!strcasecmp("ofud", tok)) {
break;
}
}
if(NULL == tok) {
/* No valid 'ud' token found; bail out */
return NULL;
}
/* initialization */
*num_btl_modules = 0;
num_devs = 0;
seedv[0] = orte_process_info.my_name->vpid;
seedv[1] = opal_sys_timer_get_cycles();
seedv[2] = opal_sys_timer_get_cycles();
seed48(seedv);
ib_devs = ibv_get_device_list(&num_devs);
if(0 == num_devs) {
mca_btl_base_error_no_nics("OpenFabrics UD", "HCA");
mca_btl_ud_modex_send();
return NULL;
}
/** We must loop through all the hca id's, get their handles and
for each hca we query the number of ports on the hca and set up
a distinct btl module for each hca port */
OBJ_CONSTRUCT(&btl_list, opal_list_t);
for(i = 0; (int32_t)i < num_devs &&
mca_btl_ofud_component.num_btls < mca_btl_ofud_component.max_btls;
i++) {
struct ibv_device_attr ib_dev_attr;
struct ibv_context* ib_dev_context;
ib_dev = ib_devs[i];
ib_dev_context = ibv_open_device(ib_dev);
if(!ib_dev_context) {
BTL_ERROR(("error obtaining device context for %s: %s\n",
ibv_get_device_name(ib_dev), strerror(errno)));
return NULL;
}
if(ibv_query_device(ib_dev_context, &ib_dev_attr)){
BTL_ERROR(("error obtaining device attributes for %s: %s\n",
ibv_get_device_name(ib_dev), strerror(errno)));
return NULL;
}
/* Note ports are 1 based hence j = 1 */
for(j = 1; j <= ib_dev_attr.phys_port_cnt; j++) {
struct ibv_port_attr ib_port_attr;
if(ibv_query_port(ib_dev_context, (uint8_t)j, &ib_port_attr)) {
BTL_ERROR(("error getting port attributes for device %s port %d: %s",
ibv_get_device_name(ib_dev), j, strerror(errno)));
return NULL;
}
if(IBV_PORT_ACTIVE == ib_port_attr.state) {
ud_btl =
(mca_btl_ud_module_t*)malloc(sizeof(mca_btl_ud_module_t));
memcpy(ud_btl, &mca_btl_ofud_module, sizeof(mca_btl_ud_module_t));
ib_selected = OBJ_NEW(mca_btl_base_selected_module_t);
ib_selected->btl_module = (mca_btl_base_module_t*)ud_btl;
ud_btl->ib_dev = ib_dev;
ud_btl->ib_dev_context = ib_dev_context;
ud_btl->ib_port_num = (uint8_t)j;
ud_btl->addr.subnet = ib_port_attr.sm_lid;
ud_btl->addr.lid = ib_port_attr.lid;
opal_list_append(&btl_list, (opal_list_item_t*) ib_selected);
if(++mca_btl_ofud_component.num_btls >=
mca_btl_ofud_component.max_btls)
break;
}
}
}
/* Allocate space for btl modules */
mca_btl_ofud_component.ud_btls = (mca_btl_ud_module_t*)
malloc(sizeof(mca_btl_ud_module_t) * mca_btl_ofud_component.num_btls);
if(NULL == mca_btl_ofud_component.ud_btls) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return NULL;
}
btls = (struct mca_btl_base_module_t**)
malloc(mca_btl_ofud_component.num_btls * sizeof(mca_btl_ud_module_t*));
if(NULL == btls) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return NULL;
}
for(i = 0; i < mca_btl_ofud_component.num_btls; i++){
item = opal_list_remove_first(&btl_list);
ib_selected = (mca_btl_base_selected_module_t*)item;
ud_btl = (mca_btl_ud_module_t*)ib_selected->btl_module;
memcpy(&(mca_btl_ofud_component.ud_btls[i]),
ud_btl, sizeof(mca_btl_ud_module_t));
free(ib_selected);
free(ud_btl);
ud_btl = &mca_btl_ofud_component.ud_btls[i];
/* Initialize module state */
if(mca_btl_ud_module_init(ud_btl) != OMPI_SUCCESS) {
mca_btl_ofud_component.num_btls--;
i--;
continue;
}
btls[i] = &ud_btl->super;
}
OBJ_DESTRUCT(&btl_list);
mca_btl_ud_modex_send();
/* Since not all modules may have initialized successfully, realloc
to free space from failed modules */
mca_btl_ofud_component.ud_btls = (mca_btl_ud_module_t*)
realloc(mca_btl_ofud_component.ud_btls,
sizeof(mca_btl_ud_module_t) * mca_btl_ofud_component.num_btls);
btls = (struct mca_btl_base_module_t**)realloc(btls,
mca_btl_ofud_component.num_btls * sizeof(mca_btl_ud_module_t*));
*num_btl_modules = mca_btl_ofud_component.num_btls;
ibv_free_device_list(ib_devs);
return btls;
}
/*
* IB component progress.
*/
#define MCA_BTL_UD_NUM_WC 500
int mca_btl_ud_component_progress(void)
{
uint32_t i;
int count = 0, ne, j;
mca_btl_ud_frag_t* frag;
struct ibv_recv_wr* bad_wr;
struct ibv_recv_wr* head_wr;
mca_btl_ud_module_t* ud_btl;
mca_btl_base_recv_reg_t* reg;
struct ibv_wc* cwc;
struct ibv_wc wc[MCA_BTL_UD_NUM_WC];
/* Poll for completions */
for(i = 0; i < mca_btl_ofud_component.num_btls; i++) {
ud_btl = &mca_btl_ofud_component.ud_btls[i];
ne = ibv_poll_cq(ud_btl->ib_cq, MCA_BTL_UD_NUM_WC, wc);
if(OPAL_UNLIKELY(ne < 0)) {
BTL_ERROR(("error polling CQ with %d: %s\n",
ne, strerror(errno)));
return OMPI_ERROR;
}
head_wr = NULL;
for(j = 0; j < ne; j++) {
cwc = &wc[j];
if(OPAL_UNLIKELY(cwc->status != IBV_WC_SUCCESS)) {
BTL_ERROR(("error polling CQ with status %d for wr_id %llu opcode %d\n",
cwc->status, cwc->wr_id, cwc->opcode));
return OMPI_ERROR;
}
frag = (mca_btl_ud_frag_t*)(unsigned long)cwc->wr_id;
/* Handle work completions */
switch(frag->type) {
case MCA_BTL_UD_FRAG_SEND:
case MCA_BTL_UD_FRAG_USER:
{
mca_btl_ud_endpoint_t* endpoint = frag->endpoint;
assert(cwc->opcode == IBV_WC_SEND);
frag->base.des_cbfunc(&ud_btl->super,
frag->endpoint, &frag->base, OMPI_SUCCESS);
/* Increment send counter, post if any sends are queued */
OPAL_THREAD_ADD32(&endpoint->sd_wqe, 1);
if(OPAL_UNLIKELY(
!opal_list_is_empty(&endpoint->pending_frags))) {
OPAL_THREAD_LOCK(&endpoint->pending_frags_lock);
frag = (mca_btl_ud_frag_t*)
opal_list_remove_first(&endpoint->pending_frags);
OPAL_THREAD_UNLOCK(&endpoint->pending_frags_lock);
if(OPAL_LIKELY(NULL != frag)) {
mca_btl_ud_endpoint_post_send(ud_btl, frag);
}
}
OPAL_THREAD_ADD32(&ud_btl->sd_wqe, 1);
if(OPAL_UNLIKELY(
!opal_list_is_empty(&ud_btl->pending_frags))) {
OPAL_THREAD_LOCK(&ud_btl->ud_lock);
frag = (mca_btl_ud_frag_t*)
opal_list_remove_first(&ud_btl->pending_frags);
OPAL_THREAD_UNLOCK(&ud_btl->ud_lock);
if(OPAL_LIKELY(NULL != frag)) {
mca_btl_ud_endpoint_post_send(ud_btl, frag);
}
}
continue;
}
case MCA_BTL_UD_FRAG_RECV:
assert(cwc->opcode == IBV_WC_RECV);
reg = &ud_btl->ib_reg[frag->hdr->tag];
frag->segment.seg_addr.pval = frag->hdr + 1;
frag->segment.seg_len = cwc->byte_len -
sizeof(mca_btl_ud_header_t) -
sizeof(mca_btl_ud_ib_header_t);
reg->cbfunc(&ud_btl->super,
frag->hdr->tag, &frag->base, reg->cbdata);
/* Add recv to linked list for reposting */
frag->wr_desc.rd_desc.next = head_wr;
head_wr = &frag->wr_desc.rd_desc;
continue;
default:
BTL_ERROR(("Unhandled completion opcode %d frag type %d",
cwc->opcode, frag->type));
break;
}
}
count += ne;
/* Repost any recv buffers all at once */
if(OPAL_LIKELY(head_wr)) {
if(OPAL_UNLIKELY(ibv_post_recv(
ud_btl->ib_qp[0], head_wr, &bad_wr))) {
BTL_ERROR(("error posting recv: %s\n", strerror(errno)));
return OMPI_ERROR;
}
head_wr = NULL;
}
}
return count;
}

137
ompi/mca/btl/ofud/btl_ofud_endpoint.c Обычный файл
Просмотреть файл

@ -0,0 +1,137 @@
/*
* 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include <errno.h>
#include <string.h>
#include "ompi_config.h"
#include "opal/prefetch.h"
#include "ompi/types.h"
#include "ompi/class/ompi_free_list.h"
#include "btl_ofud.h"
#include "btl_ofud_endpoint.h"
#include "btl_ofud_frag.h"
static void mca_btl_ud_endpoint_construct(mca_btl_base_endpoint_t* endpoint);
static void mca_btl_ud_endpoint_destruct(mca_btl_base_endpoint_t* endpoint);
/* First, we check the downcounter on the endpoint.
If it is 0, we queue this frag on the endpoint.
Otherwise, we check the BTL downcounter.
If it is 0, we queue this frag on the BTL.
Otherwise, we post the send. */
#define CHECK_FRAG_QUEUES(sd_wqe, lock, queue, frag) \
do { \
if(OPAL_UNLIKELY(OPAL_THREAD_ADD32(&(sd_wqe), -1) < 0)) { \
OPAL_THREAD_ADD32(&(sd_wqe), 1); \
OPAL_THREAD_LOCK(&(lock)); \
opal_list_append(&(queue), \
(opal_list_item_t*)(frag)); \
OPAL_THREAD_UNLOCK(&(lock)); \
return OMPI_SUCCESS; \
} \
} while(0);
/*
* Post a send to the work queue
*/
int mca_btl_ud_endpoint_post_send(mca_btl_ud_module_t* ud_btl,
mca_btl_ud_frag_t* frag)
{
struct ibv_qp* ib_qp;
struct ibv_send_wr* bad_wr;
struct ibv_send_wr* wr = &frag->wr_desc.sr_desc;
mca_btl_ud_endpoint_t* endpoint = frag->endpoint;
int ret;
/* Have to be careful here - UD adds a 40 byte header, but it is not
included on the sending side. */
frag->sg_entry.length = frag->segment.seg_len + sizeof(mca_btl_ud_header_t);
wr->send_flags = IBV_SEND_SIGNALED;
CHECK_FRAG_QUEUES(endpoint->sd_wqe,
endpoint->pending_frags_lock, endpoint->pending_frags, frag);
CHECK_FRAG_QUEUES(ud_btl->sd_wqe,
ud_btl->ud_lock, ud_btl->pending_frags, frag);
/* We avoid locking here by allowing our stripe counter to count
until it wraps around uint32_t. This keeps the mod operation
out of the critical section, allowing us to use OPAL_THREAD_ADD32
instead of a full mutex. */
ib_qp = ud_btl->ib_qp[ud_btl->ib_qp_next % MCA_BTL_UD_NUM_QP];
OPAL_THREAD_ADD32(&ud_btl->ib_qp_next, 1);
wr->wr.ud.ah = endpoint->rmt_ah;
wr->wr.ud.remote_qpn = endpoint->rem_addr.qp_num;
if(frag->sg_entry.length <= ud_btl->ib_inline_max) {
wr->send_flags =
IBV_SEND_SIGNALED|IBV_SEND_INLINE;
}
/*frag->hdr->src_qpnum = ud_btl->addr.qp_num;*/
MCA_BTL_UD_START_TIME(ibv_post_send);
if(OPAL_UNLIKELY((ret = ibv_post_send(ib_qp, wr, &bad_wr)))) {
opal_output(0, "ep->sd_wqe %d btl->sd_wqe %d len %d ib_qp_next %d",
endpoint->sd_wqe, ud_btl->sd_wqe,
frag->sg_entry.length, ud_btl->ib_qp_next);
BTL_ERROR(("error posting send request: %d %s\n", ret, strerror(ret)));
}
MCA_BTL_UD_END_TIME(ibv_post_send);
return OMPI_SUCCESS;
}
OBJ_CLASS_INSTANCE(mca_btl_ud_endpoint_t,
opal_list_item_t, mca_btl_ud_endpoint_construct,
mca_btl_ud_endpoint_destruct);
/*
* Construct/destruct an endpoint structure.
*/
static void mca_btl_ud_endpoint_construct(mca_btl_base_endpoint_t* endpoint)
{
#if OMPI_ENABLE_DEBUG
memset(&endpoint->rem_addr, 0, sizeof(struct mca_btl_ud_addr_t));
#endif
OBJ_CONSTRUCT(&endpoint->pending_frags, opal_list_t);
OBJ_CONSTRUCT(&endpoint->pending_frags_lock, opal_mutex_t);
endpoint->sd_wqe = mca_btl_ofud_component.sd_num_peer;
}
static void mca_btl_ud_endpoint_destruct(mca_btl_base_endpoint_t* endpoint)
{
/* TODO - what about any pending frags? */
OBJ_DESTRUCT(&endpoint->pending_frags);
OBJ_DESTRUCT(&endpoint->pending_frags_lock);
}

82
ompi/mca/btl/ofud/btl_ofud_endpoint.h Обычный файл
Просмотреть файл

@ -0,0 +1,82 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2006 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#ifndef MCA_BTL_IB_ENDPOINT_H
#define MCA_BTL_IB_ENDPOINT_H
#include <infiniband/verbs.h>
#include "opal/class/opal_list.h"
#include "opal/event/event.h"
#include "btl_ofud.h"
#include "btl_ofud_frag.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
struct mca_btl_ud_addr_t {
uint32_t qp_num;
uint32_t psn;
uint16_t lid;
uint16_t subnet;
};
typedef struct mca_btl_ud_addr_t mca_btl_ud_addr_t;
/**
* An abstraction that represents a connection to a endpoint process.
* An instance of mca_btl_base_endpoint_t is associated w/ each process
* and BTL pair and address information is exchanged at startup.
* The UD BTL is connectionless, so no connection is ever established.
*/
struct mca_btl_base_endpoint_t {
opal_list_item_t super;
mca_btl_ud_addr_t rem_addr;
/**< Remote address information */
/* No lock needed, read-only past initialization */
struct ibv_ah* rmt_ah;
/**< Remote address handle */
/* No lock needed, verbs are thread-safe */
opal_list_t pending_frags;
opal_mutex_t pending_frags_lock;
/**< list of pending frags and lock */
int32_t sd_wqe;
/**< number of available send wqe entries */
/* No lock needed, OPAL_THREAD_ADD32 is used */
};
typedef struct mca_btl_base_endpoint_t mca_btl_base_endpoint_t;
typedef mca_btl_base_endpoint_t mca_btl_ud_endpoint_t;
OBJ_CLASS_DECLARATION(mca_btl_ud_endpoint_t);
int mca_btl_ud_endpoint_post_send(struct mca_btl_ud_module_t* ud_btl,
struct mca_btl_ud_frag_t * frag);
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

112
ompi/mca/btl/ofud/btl_ofud_frag.c Обычный файл
Просмотреть файл

@ -0,0 +1,112 @@
/*
* 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "btl_ofud.h"
#include "btl_ofud_frag.h"
#include "ompi/mca/mpool/rdma/mpool_rdma.h"
static inline void mca_btl_ud_frag_common_constructor(mca_btl_ud_frag_t* frag)
{
frag->ud_reg = (mca_btl_ud_reg_t*)frag->base.super.registration;
frag->sg_entry.lkey = frag->ud_reg->mr->lkey;
frag->base.des_flags = 0;
frag->base.order = MCA_BTL_NO_ORDER;
}
static void mca_btl_ud_send_frag_constructor(mca_btl_ud_frag_t* frag)
{
frag->type = MCA_BTL_UD_FRAG_SEND;
mca_btl_ud_frag_common_constructor(frag);
frag->base.des_src = &frag->segment;
frag->base.des_src_cnt = 1;
frag->base.des_dst = NULL;
frag->base.des_dst_cnt = 0;
/* We do not include the mca_btl_ud_ib_header_t data when sending */
frag->hdr = frag->base.super.ptr;
frag->segment.seg_addr.pval = frag->hdr + 1;
frag->sg_entry.addr = (unsigned long)frag->hdr;
frag->wr_desc.sr_desc.wr_id = (unsigned long)frag;
frag->wr_desc.sr_desc.sg_list = &frag->sg_entry;
frag->wr_desc.sr_desc.num_sge = 1;
frag->wr_desc.sr_desc.opcode = IBV_WR_SEND;
frag->wr_desc.sr_desc.send_flags = IBV_SEND_SIGNALED;
frag->wr_desc.sr_desc.next = NULL;
frag->wr_desc.sr_desc.wr.ud.remote_qkey = mca_btl_ofud_component.ib_qkey;
}
static void mca_btl_ud_user_frag_constructor(mca_btl_ud_frag_t* frag)
{
mca_btl_ud_send_frag_constructor(frag);
frag->type = MCA_BTL_UD_FRAG_USER;
}
static void mca_btl_ud_recv_frag_constructor(mca_btl_ud_frag_t* frag)
{
frag->type = MCA_BTL_UD_FRAG_RECV;
mca_btl_ud_frag_common_constructor(frag);
frag->base.des_dst = &frag->segment;
frag->base.des_dst_cnt = 1;
frag->base.des_src = NULL;
frag->base.des_src_cnt = 0;
/* Receive frag headers start 40 bytes later */
frag->hdr = (mca_btl_ud_header_t*)((uintptr_t)frag->base.super.ptr +
sizeof(mca_btl_ud_ib_header_t));
frag->segment.seg_addr.pval = frag->hdr + 1;
frag->sg_entry.addr = (uintptr_t)frag->base.super.ptr;
frag->segment.seg_len = mca_btl_ofud_module.super.btl_eager_limit;
frag->sg_entry.length = mca_btl_ofud_module.super.btl_eager_limit +
sizeof(mca_btl_ud_ib_header_t) + sizeof(mca_btl_ud_header_t);
frag->wr_desc.rd_desc.wr_id = (unsigned long)frag;
frag->wr_desc.rd_desc.sg_list = &frag->sg_entry;
frag->wr_desc.rd_desc.num_sge = 1;
frag->wr_desc.rd_desc.next = NULL;
}
OBJ_CLASS_INSTANCE(mca_btl_ud_frag_t,
mca_btl_base_descriptor_t,
NULL,
NULL);
OBJ_CLASS_INSTANCE(mca_btl_ud_send_frag_t,
mca_btl_base_descriptor_t,
mca_btl_ud_send_frag_constructor,
NULL);
OBJ_CLASS_INSTANCE(mca_btl_ud_user_frag_t,
mca_btl_base_descriptor_t,
mca_btl_ud_user_frag_constructor,
NULL);
OBJ_CLASS_INSTANCE(mca_btl_ud_recv_frag_t,
mca_btl_base_descriptor_t,
mca_btl_ud_recv_frag_constructor,
NULL);

148
ompi/mca/btl/ofud/btl_ofud_frag.h Обычный файл
Просмотреть файл

@ -0,0 +1,148 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2006 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#ifndef MCA_BTL_UD_FRAG_H
#define MCA_BTL_UD_FRAG_H
#define MCA_BTL_IB_FRAG_ALIGN (8)
#include <infiniband/verbs.h>
#include "ompi/mca/mpool/rdma/mpool_rdma.h"
#include "btl_ofud.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
/**
* Fragment types
*/
typedef enum {
MCA_BTL_UD_FRAG_SEND,
MCA_BTL_UD_FRAG_USER,
MCA_BTL_UD_FRAG_RECV
} mca_btl_ud_frag_type_t;
struct mca_btl_ud_reg_t {
mca_mpool_base_registration_t base;
struct ibv_mr* mr;
};
typedef struct mca_btl_ud_reg_t mca_btl_ud_reg_t;
/* UD adds a 40 byte global routing header */
/* This works in strange ways - the sending side does not need to explicitly
include this data in sg lists. Then, on the receiving side, the extra 40
bytes magically appear. */
struct mca_btl_ud_ib_header_t {
uint8_t ib_grh[40];
};
typedef struct mca_btl_ud_ib_header_t mca_btl_ud_ib_header_t;
struct mca_btl_ud_header_t {
/*uint32_t src_qpnum;*/
mca_btl_base_tag_t tag;
};
typedef struct mca_btl_ud_header_t mca_btl_ud_header_t;
/**
* IB send fragment derived type.
*/
struct mca_btl_ud_frag_t {
mca_btl_base_descriptor_t base;
mca_btl_base_segment_t segment;
struct mca_btl_base_endpoint_t* endpoint;
mca_btl_ud_frag_type_t type;
union{
struct ibv_recv_wr rd_desc;
struct ibv_send_wr sr_desc;
} wr_desc;
struct ibv_sge sg_entry;
/* When this is a send frag, hdr points right after this, as expected.
But when this is a receive frag, we have an extra 40 bytes provided
by IB, so this points 40 bytes past the end of the frag. */
mca_btl_ud_header_t* hdr;
mca_btl_ud_reg_t* ud_reg;
};
typedef struct mca_btl_ud_frag_t mca_btl_ud_frag_t;
OBJ_CLASS_DECLARATION(mca_btl_ud_frag_t);
typedef struct mca_btl_ud_frag_t mca_btl_ud_send_frag_t;
OBJ_CLASS_DECLARATION(mca_btl_ud_send_frag_t);
typedef struct mca_btl_ud_frag_t mca_btl_ud_user_frag_t;
OBJ_CLASS_DECLARATION(mca_btl_ud_user_frag_t);
typedef struct mca_btl_ud_frag_t mca_btl_ud_recv_frag_t;
OBJ_CLASS_DECLARATION(mca_btl_ud_recv_frag_t);
/*
* Allocate/return a UD/IB send/user fragment
*/
#define MCA_BTL_UD_ALLOC_FRAG(btl, frag, rc) \
{ \
ompi_free_list_item_t *item; \
OMPI_FREE_LIST_GET(&((mca_btl_ud_module_t*)btl)->send_frags, item, rc); \
frag = (mca_btl_ud_frag_t*) item; \
}
#define MCA_BTL_UD_RETURN_FRAG(btl, frag) \
{ \
OMPI_FREE_LIST_RETURN( \
&((mca_btl_ud_module_t*)btl)->send_frags, \
(ompi_free_list_item_t*)(frag)); \
}
#define MCA_BTL_UD_ALLOC_USER_FRAG(btl, frag, rc) \
{ \
ompi_free_list_item_t *item; \
OMPI_FREE_LIST_GET(&((mca_btl_ud_module_t*)btl)->user_frags, item, rc); \
frag = (mca_btl_ud_frag_t*) item; \
}
#define MCA_BTL_UD_RETURN_USER_FRAG(btl, frag) \
{ \
OMPI_FREE_LIST_RETURN( \
&((mca_btl_ud_module_t*)btl)->user_frags, \
(ompi_free_list_item_t*)(frag)); \
}
struct mca_btl_ud_module_t;
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

206
ompi/mca/btl/ofud/btl_ofud_proc.c Обычный файл
Просмотреть файл

@ -0,0 +1,206 @@
/*
* 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi/mca/pml/base/pml_base_module_exchange.h"
#include "btl_ofud.h"
#include "btl_ofud_proc.h"
static void mca_btl_ud_proc_construct(mca_btl_ud_proc_t* proc);
static void mca_btl_ud_proc_destruct(mca_btl_ud_proc_t* proc);
OBJ_CLASS_INSTANCE(mca_btl_ud_proc_t,
opal_list_item_t, mca_btl_ud_proc_construct,
mca_btl_ud_proc_destruct);
void mca_btl_ud_proc_construct(mca_btl_ud_proc_t* proc)
{
proc->proc_ompi = 0;
proc->proc_addr_count = 0;
proc->proc_endpoints = 0;
proc->proc_endpoint_count = 0;
OBJ_CONSTRUCT(&proc->proc_lock, opal_mutex_t);
/* add to list of all proc instance */
OPAL_THREAD_LOCK(&mca_btl_ofud_component.ud_lock);
opal_list_append(&mca_btl_ofud_component.ud_procs, &proc->super);
OPAL_THREAD_UNLOCK(&mca_btl_ofud_component.ud_lock);
}
void mca_btl_ud_proc_destruct(mca_btl_ud_proc_t* proc)
{
/* remove from list of all proc instances */
OPAL_THREAD_LOCK(&mca_btl_ofud_component.ud_lock);
opal_list_remove_item(&mca_btl_ofud_component.ud_procs, &proc->super);
OPAL_THREAD_UNLOCK(&mca_btl_ofud_component.ud_lock);
/* release resources */
if(NULL != proc->proc_endpoints) {
free(proc->proc_endpoints);
}
}
/*
* Look for an existing IB process instance based on the associated
* ompi_proc_t instance.
*/
mca_btl_ud_proc_t* mca_btl_ud_proc_lookup_ompi(ompi_proc_t* ompi_proc)
{
mca_btl_ud_proc_t* ib_proc;
OPAL_THREAD_LOCK(&mca_btl_ofud_component.ud_lock);
for(ib_proc = (mca_btl_ud_proc_t*)
opal_list_get_first(&mca_btl_ofud_component.ud_procs);
ib_proc != (mca_btl_ud_proc_t*)
opal_list_get_end(&mca_btl_ofud_component.ud_procs);
ib_proc = (mca_btl_ud_proc_t*)opal_list_get_next(ib_proc)) {
if(ib_proc->proc_ompi == ompi_proc) {
OPAL_THREAD_UNLOCK(&mca_btl_ofud_component.ud_lock);
return ib_proc;
}
}
OPAL_THREAD_UNLOCK(&mca_btl_ofud_component.ud_lock);
return NULL;
}
/*
* Create a IB process structure. There is a one-to-one correspondence
* between a ompi_proc_t and a mca_btl_ud_proc_t instance. We cache
* additional data (specifically the list of mca_btl_ud_endpoint_t instances,
* and published addresses) associated w/ a given destination on this
* datastructure.
*/
mca_btl_ud_proc_t* mca_btl_ud_proc_create(ompi_proc_t* ompi_proc)
{
mca_btl_ud_proc_t* module_proc = NULL;
size_t size;
int rc;
/* Check if we have already created a IB proc
* structure for this ompi process */
module_proc = mca_btl_ud_proc_lookup_ompi(ompi_proc);
if(module_proc != NULL) {
/* Gotcha! */
return module_proc;
}
/* Oops! First time, gotta create a new IB proc out of the ompi_proc ... */
module_proc = OBJ_NEW(mca_btl_ud_proc_t);
/* Initialize number of peer */
module_proc->proc_endpoint_count = 0;
module_proc->proc_ompi = ompi_proc;
/* build a unique identifier (of arbitrary size) to represent the proc */
module_proc->proc_guid = ompi_proc->proc_name;
/* query for the peer address info */
rc = mca_pml_base_modex_recv(&mca_btl_ofud_component.super.btl_version,
ompi_proc, (void*)&module_proc->proc_addrs,
&size);
if(OMPI_SUCCESS != rc) {
opal_output(0,
"[%s:%d] mca_pml_base_modex_recv failed for peer [%ld,%ld,%ld]",
__FILE__,__LINE__,ORTE_NAME_ARGS(&ompi_proc->proc_name));
OBJ_RELEASE(module_proc);
return NULL;
}
if((size % sizeof(mca_btl_ud_addr_t)) != 0) {
opal_output(0, "[%s:%d] invalid module address for peer [%ld,%ld,%ld]",
__FILE__,__LINE__,ORTE_NAME_ARGS(&ompi_proc->proc_name));
OBJ_RELEASE(module_proc);
return NULL;
}
module_proc->proc_addr_count = size / sizeof(mca_btl_ud_addr_t);
if (0 == module_proc->proc_addr_count) {
module_proc->proc_endpoints = NULL;
} else {
module_proc->proc_endpoints = (mca_btl_base_endpoint_t**)
malloc(module_proc->proc_addr_count *
sizeof(mca_btl_base_endpoint_t*));
}
if(NULL == module_proc->proc_endpoints) {
OBJ_RELEASE(module_proc);
return NULL;
}
return module_proc;
}
/*
* Insert an endpoint into the proc array and assign it an address.
*
* MUST be called with the proc lock held!
*/
int mca_btl_ud_proc_insert(mca_btl_ud_proc_t* module_proc,
mca_btl_base_endpoint_t* module_endpoint)
{
module_endpoint->rem_addr =
module_proc->proc_addrs[module_proc->proc_endpoint_count];
module_proc->proc_endpoints[module_proc->proc_endpoint_count++] =
module_endpoint;
return OMPI_SUCCESS;
}
/*
* Remove an endpoint from the proc array.
*/
int mca_btl_ud_proc_remove(mca_btl_ud_proc_t* proc,
mca_btl_base_endpoint_t* endpoint)
{
size_t i;
OPAL_THREAD_LOCK(&proc->proc_lock);
for(i = 0; i < proc->proc_endpoint_count; i++) {
if(proc->proc_endpoints[i] == endpoint) {
memmove(proc->proc_endpoints + i, proc->proc_endpoints + i + 1,
(proc->proc_endpoint_count -i - 1) *
sizeof(mca_btl_base_endpoint_t*));
if(--proc->proc_endpoint_count == 0) {
OPAL_THREAD_UNLOCK(&proc->proc_lock);
OBJ_RELEASE(proc);
return OMPI_SUCCESS;
}
break;
}
}
OPAL_THREAD_UNLOCK(&proc->proc_lock);
return OMPI_SUCCESS;
}

79
ompi/mca/btl/ofud/btl_ofud_proc.h Обычный файл
Просмотреть файл

@ -0,0 +1,79 @@
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2006 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 Sandia National Laboratories. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#ifndef MCA_BTL_UD_PROC_H
#define MCA_BTL_UD_PROC_H
#include "opal/class/opal_object.h"
#include "orte/mca/ns/ns.h"
#include "ompi/proc/proc.h"
#include "btl_ofud.h"
#include "btl_ofud_endpoint.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
/**
* Represents the state of a remote process and the set of addresses
* that it exports. Also cache an instance of mca_btl_base_endpoint_t for
* each BTL instance that attempts to open a connection to the process.
*/
struct mca_btl_ud_proc_t {
opal_list_item_t super;
/**< allow proc to be placed on a list */
ompi_proc_t *proc_ompi;
/**< pointer to corresponding ompi_proc_t */
orte_process_name_t proc_guid;
/**< globally unique identifier for the process */
struct mca_btl_ud_addr_t* proc_addrs;
size_t proc_addr_count;
/**< number of addresses published by endpoint */
struct mca_btl_base_endpoint_t **proc_endpoints;
/**< array of endpoints that have been created to access this proc */
size_t proc_endpoint_count;
/**< number of endpoints */
opal_mutex_t proc_lock;
/**< lock to protect against concurrent access to proc state */
};
typedef struct mca_btl_ud_proc_t mca_btl_ud_proc_t;
OBJ_CLASS_DECLARATION(mca_btl_ud_proc_t);
mca_btl_ud_proc_t* mca_btl_ud_proc_lookup_ompi(ompi_proc_t* ompi_proc);
mca_btl_ud_proc_t* mca_btl_ud_proc_create(ompi_proc_t* ompi_proc);
int mca_btl_ud_proc_insert(mca_btl_ud_proc_t*, mca_btl_base_endpoint_t*);
int mca_btl_ud_proc_remove(mca_btl_ud_proc_t*, mca_btl_base_endpoint_t*);
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif

43
ompi/mca/btl/ofud/configure.m4 Обычный файл
Просмотреть файл

@ -0,0 +1,43 @@
# -*- shell-script -*-
#
# 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 Sandia National Laboratories. All rights
# reserved.
# $COPYRIGHT$
#
# Additional copyrights may follow
#
# $HEADER$
#
# MCA_btl_ofud_CONFIG([action-if-can-compile],
# [action-if-cant-compile])
# ------------------------------------------------
AC_DEFUN([MCA_btl_ofud_CONFIG],[
OMPI_CHECK_OPENIB([btl_ofud],
[btl_ofud_happy="yes"],
[btl_ofud_happy="no"])
AS_IF([test "$btl_ofud_happy" = "yes"],
[btl_ofud_WRAPPER_EXTRA_LDFLAGS="$btl_ofud_LDFLAGS"
btl_ofud_WRAPPER_EXTRA_LIBS="$btl_ofud_LIBS"
$1],
[$2])
# substitute in the things needed to build OFUD
AC_SUBST([btl_ofud_CFLAGS])
AC_SUBST([btl_ofud_CPPFLAGS])
AC_SUBST([btl_ofud_LDFLAGS])
AC_SUBST([btl_ofud_LIBS])
])dnl

26
ompi/mca/btl/ofud/configure.params Обычный файл
Просмотреть файл

@ -0,0 +1,26 @@
# -*- shell-script -*-
#
# 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 Sandia National Laboratories. All rights
# reserved.
# $COPYRIGHT$
#
# Additional copyrights may follow
#
# $HEADER$
#
# Specific to this module
PARAM_INIT_FILE=btl_ofud.c
PARAM_CONFIG_HEADER_FILE="ofud_config.h"
PARAM_CONFIG_FILES="Makefile"