1
1
openmpi/ompi/mca/btl/openib/btl_openib.c

1314 строки
45 KiB
C
Исходник Обычный вид История

/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2008 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) 2007 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2006-2008 Mellanox Technologies. All rights reserved.
* Copyright (c) 2006-2007 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2006-2007 Voltaire All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include <string.h>
#include <inttypes.h>
#include "orte/util/show_help.h"
#include "opal/util/if.h"
#include "opal/util/arch.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/btl/btl.h"
#include "ompi/mca/btl/base/btl_base_error.h"
#if OPAL_ENABLE_FT == 1
#include "ompi/runtime/ompi_cr.h"
#endif
#include "btl_openib.h"
#include "btl_openib_frag.h"
#include "btl_openib_proc.h"
#include "btl_openib_endpoint.h"
#include "btl_openib_xrc.h"
#include "ompi/datatype/convertor.h"
#include "ompi/datatype/datatype.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/mpool/mpool.h"
#include "ompi/mca/mpool/rdma/mpool_rdma.h"
#include "orte/util/proc_info.h"
#include <errno.h>
#include <string.h>
#include <math.h>
#include <inttypes.h>
#ifdef HAVE_SYS_TYPES_H
#include <sys/types.h>
#endif
#ifdef HAVE_SYS_TIME_H
#include <sys/time.h>
#endif
#ifdef HAVE_SYS_RESOURCE_H
#include <sys/resource.h>
#endif
#include <unistd.h>
mca_btl_openib_module_t mca_btl_openib_module = {
{
&mca_btl_openib_component.super,
0, /* max size of first fragment */
0, /* min send fragment size */
0, /* max send fragment size */
0, /* btl_rdma_pipeline_send_length */
0, /* btl_rdma_pipeline_frag_size */
0, /* btl_min_rdma_pipeline_size */
0, /* exclusivity */
0, /* latency */
0, /* bandwidth */
0, /* TODO this should be PUT btl flags */
mca_btl_openib_add_procs,
mca_btl_openib_del_procs,
NULL,
mca_btl_openib_finalize,
/* we need alloc free, pack */
mca_btl_openib_alloc,
mca_btl_openib_free,
mca_btl_openib_prepare_src,
mca_btl_openib_prepare_dst,
mca_btl_openib_send,
NULL, /* send immediate */
mca_btl_openib_put,
mca_btl_openib_get,
mca_btl_base_dump,
NULL, /* mpool */
mca_btl_openib_register_error_cb, /* error call back registration */
mca_btl_openib_ft_event
}
};
#if OPAL_ENABLE_FT == 1
static int ft_event_btl_openib_finalize(struct mca_btl_base_module_t* btl);
#endif
static void show_init_error(const char *file, int line,
const char *func, const char *dev)
{
if (ENOMEM == errno) {
int ret;
struct rlimit limit;
char *str_limit = NULL;
ret = getrlimit(RLIMIT_MEMLOCK, &limit);
if (0 != ret) {
asprintf(&str_limit, "Unknown");
} else if (limit.rlim_cur == RLIM_INFINITY) {
asprintf(&str_limit, "unlimited");
} else {
asprintf(&str_limit, "%ld", (long)limit.rlim_cur);
}
This commit represents a bunch of work on a Mercurial side branch. As such, the commit message back to the master SVN repository is fairly long. = ORTE Job-Level Output Messages = Add two new interfaces that should be used for all new code throughout the ORTE and OMPI layers (we already make the search-and-replace on the existing ORTE / OMPI layers): * orte_output(): (and corresponding friends ORTE_OUTPUT, orte_output_verbose, etc.) This function sends the output directly to the HNP for processing as part of a job-specific output channel. It supports all the same outputs as opal_output() (syslog, file, stdout, stderr), but for stdout/stderr, the output is sent to the HNP for processing and output. More on this below. * orte_show_help(): This function is a drop-in-replacement for opal_show_help(), with two differences in functionality: 1. the rendered text help message output is sent to the HNP for display (rather than outputting directly into the process' stderr stream) 1. the HNP detects duplicate help messages and does not display them (so that you don't see the same error message N times, once from each of your N MPI processes); instead, it counts "new" instances of the help message and displays a message every ~5 seconds when there are new ones ("I got X new copies of the help message...") opal_show_help and opal_output still exist, but they only output in the current process. The intent for the new orte_* functions is that they can apply job-level intelligence to the output. As such, we recommend that all new ORTE and OMPI code use the new orte_* functions, not thei opal_* functions. === New code === For ORTE and OMPI programmers, here's what you need to do differently in new code: * Do not include opal/util/show_help.h or opal/util/output.h. Instead, include orte/util/output.h (this one header file has declarations for both the orte_output() series of functions and orte_show_help()). * Effectively s/opal_output/orte_output/gi throughout your code. Note that orte_output_open() takes a slightly different argument list (as a way to pass data to the filtering stream -- see below), so you if explicitly call opal_output_open(), you'll need to slightly adapt to the new signature of orte_output_open(). * Literally s/opal_show_help/orte_show_help/. The function signature is identical. === Notes === * orte_output'ing to stream 0 will do similar to what opal_output'ing did, so leaving a hard-coded "0" as the first argument is safe. * For systems that do not use ORTE's RML or the HNP, the effect of orte_output_* and orte_show_help will be identical to their opal counterparts (the additional information passed to orte_output_open() will be lost!). Indeed, the orte_* functions simply become trivial wrappers to their opal_* counterparts. Note that we have not tested this; the code is simple but it is quite possible that we mucked something up. = Filter Framework = Messages sent view the new orte_* functions described above and messages output via the IOF on the HNP will now optionally be passed through a new "filter" framework before being output to stdout/stderr. The "filter" OPAL MCA framework is intended to allow preprocessing to messages before they are sent to their final destinations. The first component that was written in the filter framework was to create an XML stream, segregating all the messages into different XML tags, etc. This will allow 3rd party tools to read the stdout/stderr from the HNP and be able to know exactly what each text message is (e.g., a help message, another OMPI infrastructure message, stdout from the user process, stderr from the user process, etc.). Filtering is not active by default. Filter components must be specifically requested, such as: {{{ $ mpirun --mca filter xml ... }}} There can only be one filter component active. = New MCA Parameters = The new functionality described above introduces two new MCA parameters: * '''orte_base_help_aggregate''': Defaults to 1 (true), meaning that help messages will be aggregated, as described above. If set to 0, all help messages will be displayed, even if they are duplicates (i.e., the original behavior). * '''orte_base_show_output_recursions''': An MCA parameter to help debug one of the known issues, described below. It is likely that this MCA parameter will disappear before v1.3 final. = Known Issues = * The XML filter component is not complete. The current output from this component is preliminary and not real XML. A bit more work needs to be done to configure.m4 search for an appropriate XML library/link it in/use it at run time. * There are possible recursion loops in the orte_output() and orte_show_help() functions -- e.g., if RML send calls orte_output() or orte_show_help(). We have some ideas how to fix these, but figured that it was ok to commit before feature freeze with known issues. The code currently contains sub-optimal workarounds so that this will not be a problem, but it would be good to actually solve the problem rather than have hackish workarounds before v1.3 final. This commit was SVN r18434.
2008-05-14 00:00:55 +04:00
orte_show_help("help-mpi-btl-openib.txt", "init-fail-no-mem",
true, orte_process_info.nodename,
file, line, func, dev, str_limit);
if (NULL != str_limit) free(str_limit);
} else {
This commit represents a bunch of work on a Mercurial side branch. As such, the commit message back to the master SVN repository is fairly long. = ORTE Job-Level Output Messages = Add two new interfaces that should be used for all new code throughout the ORTE and OMPI layers (we already make the search-and-replace on the existing ORTE / OMPI layers): * orte_output(): (and corresponding friends ORTE_OUTPUT, orte_output_verbose, etc.) This function sends the output directly to the HNP for processing as part of a job-specific output channel. It supports all the same outputs as opal_output() (syslog, file, stdout, stderr), but for stdout/stderr, the output is sent to the HNP for processing and output. More on this below. * orte_show_help(): This function is a drop-in-replacement for opal_show_help(), with two differences in functionality: 1. the rendered text help message output is sent to the HNP for display (rather than outputting directly into the process' stderr stream) 1. the HNP detects duplicate help messages and does not display them (so that you don't see the same error message N times, once from each of your N MPI processes); instead, it counts "new" instances of the help message and displays a message every ~5 seconds when there are new ones ("I got X new copies of the help message...") opal_show_help and opal_output still exist, but they only output in the current process. The intent for the new orte_* functions is that they can apply job-level intelligence to the output. As such, we recommend that all new ORTE and OMPI code use the new orte_* functions, not thei opal_* functions. === New code === For ORTE and OMPI programmers, here's what you need to do differently in new code: * Do not include opal/util/show_help.h or opal/util/output.h. Instead, include orte/util/output.h (this one header file has declarations for both the orte_output() series of functions and orte_show_help()). * Effectively s/opal_output/orte_output/gi throughout your code. Note that orte_output_open() takes a slightly different argument list (as a way to pass data to the filtering stream -- see below), so you if explicitly call opal_output_open(), you'll need to slightly adapt to the new signature of orte_output_open(). * Literally s/opal_show_help/orte_show_help/. The function signature is identical. === Notes === * orte_output'ing to stream 0 will do similar to what opal_output'ing did, so leaving a hard-coded "0" as the first argument is safe. * For systems that do not use ORTE's RML or the HNP, the effect of orte_output_* and orte_show_help will be identical to their opal counterparts (the additional information passed to orte_output_open() will be lost!). Indeed, the orte_* functions simply become trivial wrappers to their opal_* counterparts. Note that we have not tested this; the code is simple but it is quite possible that we mucked something up. = Filter Framework = Messages sent view the new orte_* functions described above and messages output via the IOF on the HNP will now optionally be passed through a new "filter" framework before being output to stdout/stderr. The "filter" OPAL MCA framework is intended to allow preprocessing to messages before they are sent to their final destinations. The first component that was written in the filter framework was to create an XML stream, segregating all the messages into different XML tags, etc. This will allow 3rd party tools to read the stdout/stderr from the HNP and be able to know exactly what each text message is (e.g., a help message, another OMPI infrastructure message, stdout from the user process, stderr from the user process, etc.). Filtering is not active by default. Filter components must be specifically requested, such as: {{{ $ mpirun --mca filter xml ... }}} There can only be one filter component active. = New MCA Parameters = The new functionality described above introduces two new MCA parameters: * '''orte_base_help_aggregate''': Defaults to 1 (true), meaning that help messages will be aggregated, as described above. If set to 0, all help messages will be displayed, even if they are duplicates (i.e., the original behavior). * '''orte_base_show_output_recursions''': An MCA parameter to help debug one of the known issues, described below. It is likely that this MCA parameter will disappear before v1.3 final. = Known Issues = * The XML filter component is not complete. The current output from this component is preliminary and not real XML. A bit more work needs to be done to configure.m4 search for an appropriate XML library/link it in/use it at run time. * There are possible recursion loops in the orte_output() and orte_show_help() functions -- e.g., if RML send calls orte_output() or orte_show_help(). We have some ideas how to fix these, but figured that it was ok to commit before feature freeze with known issues. The code currently contains sub-optimal workarounds so that this will not be a problem, but it would be good to actually solve the problem rather than have hackish workarounds before v1.3 final. This commit was SVN r18434.
2008-05-14 00:00:55 +04:00
orte_show_help("help-mpi-btl-openib.txt", "init-fail-create-q",
true, orte_process_info.nodename,
file, line, func, strerror(errno), errno, dev);
}
}
static inline struct ibv_cq *ibv_create_cq_compat(struct ibv_context *context,
int cqe, void *cq_context, struct ibv_comp_channel *channel,
int comp_vector)
{
#if OMPI_IBV_CREATE_CQ_ARGS == 3
return ibv_create_cq(context, cqe, channel);
#else
return ibv_create_cq(context, cqe, cq_context, channel, comp_vector);
#endif
}
static int adjust_cq(mca_btl_openib_device_t *device, const int cq)
{
uint32_t cq_size = device->cq_size[cq];
/* make sure we don't exceed the maximum CQ size and that we
* don't size the queue smaller than otherwise requested
*/
if(cq_size < mca_btl_openib_component.ib_cq_size[cq])
cq_size = mca_btl_openib_component.ib_cq_size[cq];
if(cq_size > (uint32_t)device->ib_dev_attr.max_cqe)
cq_size = device->ib_dev_attr.max_cqe;
if(NULL == device->ib_cq[cq]) {
device->ib_cq[cq] = ibv_create_cq_compat(device->ib_dev_context, cq_size,
#if OMPI_ENABLE_PROGRESS_THREADS == 1
device, device->ib_channel,
#else
NULL, NULL,
#endif
0);
if (NULL == device->ib_cq[cq]) {
show_init_error(__FILE__, __LINE__, "ibv_create_cq",
ibv_get_device_name(device->ib_dev));
return OMPI_ERROR;
}
#if OMPI_ENABLE_PROGRESS_THREADS == 1
if(ibv_req_notify_cq(device->ib_cq[cq], 0)) {
show_init_error(__FILE__, __LINE__, "ibv_req_notify_cq",
ibv_get_device_name(device->ib_dev));
return OMPI_ERROR;
}
OPAL_THREAD_LOCK(&device->device_lock);
if (!device->progress) {
int rc;
device->progress = true;
if(OPAL_SUCCESS != (rc = opal_thread_start(&device->thread))) {
BTL_ERROR(("Unable to create progress thread, retval=%d", rc));
return rc;
}
}
OPAL_THREAD_UNLOCK(&device->device_lock);
#endif
}
#ifdef HAVE_IBV_RESIZE_CQ
else if (cq_size > mca_btl_openib_component.ib_cq_size[cq]){
int rc;
rc = ibv_resize_cq(device->ib_cq[cq], cq_size);
/* For ConnectX the resize CQ is not implemented and verbs returns -ENOSYS
* but should return ENOSYS. So it is reason for abs */
if(rc && ENOSYS != abs(rc)) {
BTL_ERROR(("cannot resize completion queue, error: %d", rc));
return OMPI_ERROR;
}
}
#endif
return OMPI_SUCCESS;
}
/*
* create both the high and low priority completion queues
* and the shared receive queue (if requested)
*/
static int create_srq(mca_btl_openib_module_t *openib_btl)
{
int qp;
/* create the SRQ's */
for(qp = 0; qp < mca_btl_openib_component.num_qps; qp++) {
struct ibv_srq_init_attr attr;
if(!BTL_OPENIB_QP_TYPE_PP(qp)) {
attr.attr.max_wr = mca_btl_openib_component.qp_infos[qp].rd_num +
mca_btl_openib_component.qp_infos[qp].u.srq_qp.sd_max;
attr.attr.max_sge = 1;
openib_btl->qps[qp].u.srq_qp.rd_posted = 0;
#if HAVE_XRC
if(BTL_OPENIB_QP_TYPE_XRC(qp)) {
openib_btl->qps[qp].u.srq_qp.srq =
ibv_create_xrc_srq(openib_btl->device->ib_pd,
openib_btl->device->xrc_domain,
openib_btl->device->ib_cq[qp_cq_prio(qp)], &attr);
} else
#endif
{
openib_btl->qps[qp].u.srq_qp.srq =
ibv_create_srq(openib_btl->device->ib_pd, &attr);
}
if (NULL == openib_btl->qps[qp].u.srq_qp.srq) {
show_init_error(__FILE__, __LINE__, "ibv_create_srq",
ibv_get_device_name(openib_btl->device->ib_dev));
return OMPI_ERROR;
}
}
}
return OMPI_SUCCESS;
}
static int mca_btl_openib_size_queues(struct mca_btl_openib_module_t* openib_btl, size_t nprocs)
{
uint32_t send_cqes, recv_cqes;
int rc = OMPI_SUCCESS, qp;
mca_btl_openib_device_t *device = openib_btl->device;
/* figure out reasonable sizes for completion queues */
for(qp = 0; qp < mca_btl_openib_component.num_qps; qp++) {
if(BTL_OPENIB_QP_TYPE_SRQ(qp)) {
send_cqes = mca_btl_openib_component.qp_infos[qp].u.srq_qp.sd_max;
recv_cqes = mca_btl_openib_component.qp_infos[qp].rd_num;
} else {
send_cqes = (mca_btl_openib_component.qp_infos[qp].rd_num +
mca_btl_openib_component.qp_infos[qp].u.pp_qp.rd_rsv) * nprocs;
recv_cqes = send_cqes;
}
openib_btl->device->cq_size[qp_cq_prio(qp)] += recv_cqes;
openib_btl->device->cq_size[BTL_OPENIB_LP_CQ] += send_cqes;
}
rc = adjust_cq(device, BTL_OPENIB_HP_CQ);
if (OMPI_SUCCESS != rc) {
goto out;
}
rc = adjust_cq(device, BTL_OPENIB_LP_CQ);
if (OMPI_SUCCESS != rc) {
goto out;
}
if (0 == openib_btl->num_peers) {
rc = create_srq(openib_btl);
}
openib_btl->num_peers += nprocs;
out:
return rc;
}
/*
* add a proc to this btl module
* creates an endpoint that is setup on the
* first send to the endpoint
*/
int mca_btl_openib_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_openib_module_t* openib_btl = (mca_btl_openib_module_t*)btl;
int i,j, rc;
int rem_subnet_id_port_cnt;
int lcl_subnet_id_port_cnt = 0;
int btl_rank = 0;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
mca_btl_base_endpoint_t* endpoint;
ompi_btl_openib_connect_base_module_t *local_cpc;
ompi_btl_openib_connect_base_module_data_t *remote_cpc_data;
for(j=0; j < mca_btl_openib_component.ib_num_btls; j++){
if(mca_btl_openib_component.openib_btls[j]->port_info.subnet_id
== openib_btl->port_info.subnet_id) {
if(openib_btl == mca_btl_openib_component.openib_btls[j]) {
btl_rank = lcl_subnet_id_port_cnt;
}
lcl_subnet_id_port_cnt++;
}
}
#if HAVE_XRC
if(MCA_BTL_XRC_ENABLED &&
NULL == mca_btl_openib_component.ib_addr_table.ht_table) {
if(OPAL_SUCCESS != opal_hash_table_init(
&mca_btl_openib_component.ib_addr_table, nprocs)) {
BTL_ERROR(("XRC internal error. Failed to allocate ib_table"));
return OMPI_ERROR;
}
}
#endif
for (i = 0; i < (int) nprocs; i++) {
struct ompi_proc_t* ompi_proc = ompi_procs[i];
mca_btl_openib_proc_t* ib_proc;
int remote_matching_port;
opal_output(-1, "add procs: adding proc %d", i);
#if defined(HAVE_STRUCT_IBV_DEVICE_TRANSPORT_TYPE)
/* Most current iWARP adapters (June 2008) cannot handle
talking to other processes on the same host (!) -- so mark
them as unreachable (need to use sm). So for the moment,
we'll just mark any local peer on an iWARP NIC as
unreachable. See trac ticket #1352. */
if (IBV_TRANSPORT_IWARP == openib_btl->device->ib_dev->transport_type &&
0 != (ompi_proc->proc_flags && OMPI_PROC_FLAG_LOCAL)) {
continue;
}
#endif
if(NULL == (ib_proc = mca_btl_openib_proc_create(ompi_proc))) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* check if the remote proc has any ports that:
- on the same subnet as the local proc, and
- on that subnet, has a CPC in common with the local proc
*/
remote_matching_port = -1;
rem_subnet_id_port_cnt = 0;
BTL_VERBOSE(("got %d port_infos ", ib_proc->proc_port_count));
for (j = 0; j < (int) ib_proc->proc_port_count; j++){
BTL_VERBOSE(("got a subnet %016" PRIx64,
ib_proc->proc_ports[j].pm_port_info.subnet_id));
if (ib_proc->proc_ports[j].pm_port_info.subnet_id ==
openib_btl->port_info.subnet_id) {
BTL_VERBOSE(("Got a matching subnet!"));
if (rem_subnet_id_port_cnt == btl_rank) {
remote_matching_port = j;
}
rem_subnet_id_port_cnt++;
}
}
if (0 == rem_subnet_id_port_cnt) {
/* no use trying to communicate with this endpoint */
BTL_VERBOSE(("No matching subnet id/CPC was found, moving on.. "));
continue;
}
/* If this process has multiple ports on a single subnet ID,
and the report proc also has multiple ports on this same
subnet ID, the default connection pattern is:
LOCAL REMOTE PEER
1st port on subnet X <--> 1st port on subnet X
2nd port on subnet X <--> 2nd port on subnet X
3nd port on subnet X <--> 3nd port on subnet X
...etc.
Note that the port numbers may not be contiguous, and they
may not be the same on either side. Hence the "1st", "2nd",
"3rd, etc. notation, above.
Hence, if the local "rank" of this module's port on the
subnet ID is greater than the total number of ports on the
peer on this same subnet, then we have no match. So skip
this connection. */
if (rem_subnet_id_port_cnt < lcl_subnet_id_port_cnt &&
btl_rank >= rem_subnet_id_port_cnt) {
BTL_VERBOSE(("Not enough remote ports on this subnet id, moving on.. "));
continue;
}
/* Now that we have verified that we're on the same subnet and
the remote peer has enough ports, see if that specific port
on the peer has a matching CPC. */
assert(btl_rank <= ib_proc->proc_port_count);
assert(remote_matching_port != -1);
if (OMPI_SUCCESS !=
ompi_btl_openib_connect_base_find_match(openib_btl,
&(ib_proc->proc_ports[remote_matching_port]),
&local_cpc,
&remote_cpc_data)) {
continue;
}
OPAL_THREAD_LOCK(&ib_proc->proc_lock);
/* 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.
*/
endpoint = OBJ_NEW(mca_btl_openib_endpoint_t);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
assert(((opal_object_t*)endpoint)->obj_reference_count == 1);
if(NULL == endpoint) {
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
return OMPI_ERR_OUT_OF_RESOURCE;
}
#if HAVE_XRC
if (MCA_BTL_XRC_ENABLED) {
int rem_port_cnt = 0;
for(j = 0; j < (int) ib_proc->proc_port_count; j++) {
if(ib_proc->proc_ports[j].pm_port_info.subnet_id ==
openib_btl->port_info.subnet_id) {
if (rem_port_cnt == btl_rank)
break;
else
rem_port_cnt ++;
}
}
assert(rem_port_cnt == btl_rank);
/* Push the subnet/lid/jobid to xrc hash */
rc = mca_btl_openib_ib_address_add_new(
ib_proc->proc_ports[j].pm_port_info.lid,
ib_proc->proc_ports[j].pm_port_info.subnet_id,
ompi_proc->proc_name.jobid, endpoint);
if (OMPI_SUCCESS != rc ) {
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
return OMPI_ERROR;
}
}
#endif
mca_btl_openib_endpoint_init(openib_btl, endpoint,
local_cpc,
&(ib_proc->proc_ports[remote_matching_port]),
remote_cpc_data);
rc = mca_btl_openib_proc_insert(ib_proc, endpoint);
if (OMPI_SUCCESS != rc) {
OBJ_RELEASE(endpoint);
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
continue;
}
endpoint->index = opal_pointer_array_add(openib_btl->device->endpoints, (void*)endpoint);
if( 0 > endpoint->index ) {
OBJ_RELEASE(endpoint);
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
continue;
}
/* Tell the selected CPC that it won. NOTE: This call is
outside of / separate from mca_btl_openib_endpoint_init()
because this function likely needs the endpoint->index. */
if (NULL != local_cpc->cbm_endpoint_init) {
rc = local_cpc->cbm_endpoint_init(endpoint);
if (OMPI_SUCCESS != rc) {
OBJ_RELEASE(endpoint);
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
continue;
}
}
ompi_bitmap_set_bit(reachable, i);
OPAL_THREAD_UNLOCK(&ib_proc->proc_lock);
peers[i] = endpoint;
}
return mca_btl_openib_size_queues(openib_btl, nprocs);
}
/*
* delete the proc as reachable from this btl module
*/
int mca_btl_openib_del_procs(struct mca_btl_base_module_t* btl,
size_t nprocs,
struct ompi_proc_t **procs,
struct mca_btl_base_endpoint_t ** peers)
{
int i,ep_index;
mca_btl_openib_module_t* openib_btl = (mca_btl_openib_module_t*) btl;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
mca_btl_openib_endpoint_t* endpoint;
for (i=0 ; i < (int) nprocs ; i++) {
mca_btl_base_endpoint_t* del_endpoint = peers[i];
for(ep_index=0;
ep_index < opal_pointer_array_get_size(openib_btl->device->endpoints);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
ep_index++) {
endpoint =
opal_pointer_array_get_item(openib_btl->device->endpoints,
ep_index);
if(!endpoint || endpoint->endpoint_btl != openib_btl) {
continue;
}
if (endpoint == del_endpoint) {
BTL_VERBOSE(("in del_procs %d, setting another endpoint to null",
ep_index));
opal_pointer_array_set_item(openib_btl->device->endpoints,
ep_index, NULL);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
assert(((opal_object_t*)endpoint)->obj_reference_count == 1);
mca_btl_openib_proc_remove(procs[i], endpoint);
OBJ_RELEASE(endpoint);
}
}
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
return OMPI_SUCCESS;
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
/*
*Register callback function for error handling..
*/
int mca_btl_openib_register_error_cb(
struct mca_btl_base_module_t* btl,
mca_btl_base_module_error_cb_fn_t cbfunc)
{
mca_btl_openib_module_t* openib_btl = (mca_btl_openib_module_t*) btl;
openib_btl->error_cb = cbfunc; /* stash for later */
return OMPI_SUCCESS;
}
static inline mca_btl_base_descriptor_t *
ib_frag_alloc(mca_btl_openib_module_t *btl, size_t size, uint8_t order,
uint32_t flags)
{
int qp, rc;
ompi_free_list_item_t* item = NULL;
for(qp = 0; qp < mca_btl_openib_component.num_qps; qp++) {
if(mca_btl_openib_component.qp_infos[qp].size >= size) {
OMPI_FREE_LIST_GET(&btl->device->qps[qp].send_free, item, rc);
if(item)
break;
}
}
if(NULL == item)
return NULL;
/* not all upper layer users set this */
to_base_frag(item)->segment.seg_len = size;
to_base_frag(item)->base.order = order;
to_base_frag(item)->base.des_flags = flags;
assert(to_send_frag(item)->qp_idx <= order);
return &to_base_frag(item)->base;
}
/* check if pending fragment has enough space for coalescing */
static mca_btl_openib_send_frag_t *check_coalescing(opal_list_t *frag_list,
opal_mutex_t *lock, mca_btl_base_endpoint_t *ep, size_t size)
{
mca_btl_openib_send_frag_t *frag = NULL;
if(opal_list_is_empty(frag_list))
return NULL;
OPAL_THREAD_LOCK(lock);
if(!opal_list_is_empty(frag_list)) {
int qp;
size_t total_length;
opal_list_item_t *i = opal_list_get_first(frag_list);
frag = to_send_frag(i);
if(to_com_frag(frag)->endpoint != ep ||
MCA_BTL_OPENIB_FRAG_CONTROL == openib_frag_type(frag)) {
OPAL_THREAD_UNLOCK(lock);
return NULL;
}
total_length = size + frag->coalesced_length +
to_base_frag(frag)->segment.seg_len +
sizeof(mca_btl_openib_header_coalesced_t);
qp = to_base_frag(frag)->base.order;
if(total_length <= mca_btl_openib_component.qp_infos[qp].size)
opal_list_remove_first(frag_list);
else
frag = NULL;
}
OPAL_THREAD_UNLOCK(lock);
return frag;
}
/**
* Allocate a segment.
*
* @param btl (IN) BTL module
* @param size (IN) Request segment size.
* @param size (IN) Size of segment to allocate
*
* 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_openib_alloc(
struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
uint8_t order,
size_t size,
uint32_t flags)
{
mca_btl_openib_module_t *obtl = (mca_btl_openib_module_t*)btl;
int qp = frag_size_to_order(obtl, size);
mca_btl_openib_send_frag_t *sfrag = NULL;
mca_btl_openib_coalesced_frag_t *cfrag;
assert(qp != MCA_BTL_NO_ORDER);
if(mca_btl_openib_component.use_message_coalescing &&
(flags & MCA_BTL_DES_FLAGS_BTL_OWNERSHIP)) {
int prio = !(flags & MCA_BTL_DES_FLAGS_PRIORITY);
sfrag = check_coalescing(&ep->qps[qp].qp->pending_frags[prio],
&ep->qps[qp].qp->lock, ep, size);
if(NULL == sfrag) {
if(BTL_OPENIB_QP_TYPE_PP(qp)) {
sfrag = check_coalescing(&ep->qps[qp].pending_frags[prio],
&ep->endpoint_lock, ep, size);
} else {
sfrag = check_coalescing(
&obtl->qps[qp].u.srq_qp.pending_frags[prio],
&obtl->ib_lock, ep, size);
}
}
}
if(NULL == sfrag)
return ib_frag_alloc((mca_btl_openib_module_t*)btl, size, order, flags);
/* begin coalescing message */
cfrag = alloc_coalesced_frag();
cfrag->send_frag = sfrag;
/* fix up new coalescing header if this is the first coalesced frag */
if(sfrag->hdr != sfrag->chdr) {
mca_btl_openib_control_header_t *ctrl_hdr;
mca_btl_openib_header_coalesced_t *clsc_hdr;
uint8_t org_tag;
org_tag = sfrag->hdr->tag;
sfrag->hdr = sfrag->chdr;
ctrl_hdr = (mca_btl_openib_control_header_t*)(sfrag->hdr + 1);
clsc_hdr = (mca_btl_openib_header_coalesced_t*)(ctrl_hdr + 1);
sfrag->hdr->tag = MCA_BTL_TAG_BTL;
ctrl_hdr->type = MCA_BTL_OPENIB_CONTROL_COALESCED;
clsc_hdr->tag = org_tag;
clsc_hdr->size = to_base_frag(sfrag)->segment.seg_len;
clsc_hdr->alloc_size = to_base_frag(sfrag)->segment.seg_len;
if(ep->nbo)
BTL_OPENIB_HEADER_COALESCED_HTON(*clsc_hdr);
sfrag->coalesced_length = sizeof(mca_btl_openib_control_header_t) +
sizeof(mca_btl_openib_header_coalesced_t);
to_com_frag(sfrag)->sg_entry.addr = (uint64_t)(uintptr_t)sfrag->hdr;
}
cfrag->hdr = (mca_btl_openib_header_coalesced_t*)
(((unsigned char*)(sfrag->hdr + 1)) + sfrag->coalesced_length +
to_base_frag(sfrag)->segment.seg_len);
cfrag->hdr->alloc_size = size;
/* point coalesced frag pointer into a data buffer */
to_base_frag(cfrag)->segment.seg_addr.pval = cfrag->hdr + 1;
to_base_frag(cfrag)->segment.seg_len = size;
/* save coalesced fragment on a main fragment; we will need it after send
* completion to free it and to call upper layer callback */
opal_list_append(&sfrag->coalesced_frags, (opal_list_item_t*)cfrag);
sfrag->coalesced_length += (size+sizeof(mca_btl_openib_header_coalesced_t));
to_base_frag(cfrag)->base.des_flags = flags;
return &to_base_frag(cfrag)->base;
}
/**
* Return a segment
*
* Return the segment to the appropriate
* preallocated segment list
*/
int mca_btl_openib_free(
struct mca_btl_base_module_t* btl,
mca_btl_base_descriptor_t* des)
{
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
/* is this fragment pointing at user memory? */
if(MCA_BTL_OPENIB_FRAG_SEND_USER == openib_frag_type(des) ||
MCA_BTL_OPENIB_FRAG_RECV_USER == openib_frag_type(des)) {
mca_btl_openib_com_frag_t* frag = to_com_frag(des);
if(frag->registration != NULL) {
btl->btl_mpool->mpool_deregister(btl->btl_mpool,
(mca_mpool_base_registration_t*)frag->registration);
frag->registration = NULL;
}
}
/* reset those field on free so we will not have to do it on alloc */
to_base_frag(des)->base.des_flags = 0;
switch(openib_frag_type(des)) {
case MCA_BTL_OPENIB_FRAG_RECV:
case MCA_BTL_OPENIB_FRAG_RECV_USER:
to_base_frag(des)->base.des_src = NULL;
to_base_frag(des)->base.des_src_cnt = 0;
break;
case MCA_BTL_OPENIB_FRAG_SEND:
to_send_frag(des)->hdr = (mca_btl_openib_header_t*)
(((unsigned char*)to_send_frag(des)->chdr) +
sizeof(mca_btl_openib_header_coalesced_t) +
sizeof(mca_btl_openib_control_header_t));
to_com_frag(des)->sg_entry.addr =
(uint64_t)(uintptr_t)to_send_frag(des)->hdr;
to_send_frag(des)->coalesced_length = 0;
assert(!opal_list_get_size(&to_send_frag(des)->coalesced_frags));
/* fall throug */
case MCA_BTL_OPENIB_FRAG_SEND_USER:
to_base_frag(des)->base.des_dst = NULL;
to_base_frag(des)->base.des_dst_cnt = 0;
break;
default:
break;
}
MCA_BTL_IB_FRAG_RETURN(des);
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-registered user-buffer, if the memory registration
* is too 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
* registrations 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 users 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_openib_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,
uint32_t flags)
{
mca_btl_openib_module_t *openib_btl;
mca_btl_openib_reg_t *openib_reg;
mca_btl_openib_com_frag_t *frag = NULL;
struct iovec iov;
uint32_t iov_count = 1;
size_t max_data = *size;
int rc;
openib_btl = (mca_btl_openib_module_t*)btl;
if(ompi_convertor_need_buffers(convertor) == false && 0 == reserve) {
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
/* GMS bloody HACK! */
if(registration != NULL || max_data > btl->btl_max_send_size) {
frag = alloc_send_user_frag();
if(NULL == frag) {
return NULL;
}
iov.iov_len = max_data;
iov.iov_base = NULL;
ompi_convertor_pack(convertor, &iov, &iov_count, &max_data);
*size = max_data;
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_IB_FRAG_RETURN(frag);
return NULL;
}
/* keep track of the registration we did */
to_com_frag(frag)->registration =
(mca_btl_openib_reg_t*)registration;
}
openib_reg = (mca_btl_openib_reg_t*)registration;
frag->sg_entry.length = max_data;
frag->sg_entry.lkey = openib_reg->mr->lkey;
frag->sg_entry.addr = (uint64_t)(uintptr_t)iov.iov_base;
to_base_frag(frag)->base.order = order;
to_base_frag(frag)->base.des_flags = flags;
to_base_frag(frag)->segment.seg_len = max_data;
to_base_frag(frag)->segment.seg_addr.pval = iov.iov_base;
to_base_frag(frag)->segment.seg_key.key32[0] =
(uint32_t)frag->sg_entry.lkey;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
assert(MCA_BTL_NO_ORDER == order);
BTL_VERBOSE(("frag->sg_entry.lkey = %lu .addr = %llu "
"frag->segment.seg_key.key32[0] = %lu",
frag->sg_entry.lkey, frag->sg_entry.addr,
frag->sg_entry.lkey));
return &to_base_frag(frag)->base;
}
}
assert(MCA_BTL_NO_ORDER == order);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
if(max_data + reserve > btl->btl_max_send_size) {
max_data = btl->btl_max_send_size - reserve;
}
frag = (mca_btl_openib_com_frag_t*)(reserve ?
mca_btl_openib_alloc(btl, endpoint, order, max_data + reserve,
flags) :
ib_frag_alloc(openib_btl, max_data, order, flags));
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
if(NULL == frag)
return NULL;
iov.iov_len = max_data;
iov.iov_base = (unsigned char*)to_base_frag(frag)->segment.seg_addr.pval +
reserve;
rc = ompi_convertor_pack(convertor, &iov, &iov_count, &max_data);
*size = max_data;
/* not all upper layer users set this */
to_base_frag(frag)->segment.seg_len = max_data + reserve;
return &to_base_frag(frag)->base;
}
/**
* Prepare the dst buffer
*
* @param btl (IN) BTL module
* @param peer (IN) BTL peer addressing
* prepare dest's behavior depends on the following:
* Has a valid memory registration been passed to prepare_src?
* if so we attempt to use the pre-registered 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
* registrations is full, this prevents resources from being exhausted.
*/
mca_btl_base_descriptor_t* mca_btl_openib_prepare_dst(
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,
uint32_t flags)
{
mca_btl_openib_module_t *openib_btl;
mca_btl_openib_com_frag_t *frag;
mca_btl_openib_reg_t *openib_reg;
int rc;
void *buffer;
openib_btl = (mca_btl_openib_module_t*)btl;
frag = alloc_recv_user_frag();
if(NULL == frag) {
return NULL;
}
ompi_convertor_get_current_pointer(convertor, &buffer);
if(NULL == registration){
/* we didn't get a memory registration passed in, so we have to
* register the region ourselves
*/
rc = btl->btl_mpool->mpool_register(btl->btl_mpool, buffer, *size, 0,
&registration);
if(OMPI_SUCCESS != rc || NULL == registration) {
MCA_BTL_IB_FRAG_RETURN(frag);
return NULL;
}
/* keep track of the registration we did */
frag->registration = (mca_btl_openib_reg_t*)registration;
}
openib_reg = (mca_btl_openib_reg_t*)registration;
frag->sg_entry.length = *size;
frag->sg_entry.lkey = openib_reg->mr->lkey;
frag->sg_entry.addr = (uint64_t)(uintptr_t)buffer;
to_base_frag(frag)->segment.seg_addr.pval = buffer;
to_base_frag(frag)->segment.seg_len = *size;
to_base_frag(frag)->segment.seg_key.key32[0] = openib_reg->mr->rkey;
to_base_frag(frag)->base.order = order;
to_base_frag(frag)->base.des_flags = flags;
BTL_VERBOSE(("frag->sg_entry.lkey = %lu .addr = %llu "
"frag->segment.seg_key.key32[0] = %lu",
frag->sg_entry.lkey, frag->sg_entry.addr,
openib_reg->mr->rkey));
return &to_base_frag(frag)->base;
}
int mca_btl_openib_finalize(struct mca_btl_base_module_t* btl)
{
mca_btl_openib_module_t* openib_btl;
mca_btl_openib_endpoint_t* endpoint;
int ep_index, i;
int qp, rc = OMPI_SUCCESS;
openib_btl = (mca_btl_openib_module_t*) btl;
/* Sanity check */
if( mca_btl_openib_component.ib_num_btls <= 0 ) {
return 0;
}
/* Release all QPs */
Fixes trac:1210, #1319 Commit from a long-standing Mercurial tree that ended up incorporating a lot of things: * A few fixes for CPC interface changes in all the CPCs * Attempts (but not yet finished) to fix shutdown problems in the IB CM CPC * #1319: add CTS support (i.e., initiator guarantees to send first message; automatically activated for iWARP over the RDMA CM CPC) * Some variable and function renamings to make this be generic (e.g., alloc_credit_frag became alloc_control_frag) * CPCs no longer post receive buffers; they only post a single receive buffer for the CTS if they use CTS. Instead, the main BTL now posts the main sets of receive buffers. * CPCs allocate a CTS buffer only if they're about to make a connection * RDMA CM improvements: * Use threaded mode openib fd monitoring to wait for for RDMA CM events * Synchronize endpoint finalization and disconnection between main thread and service thread to avoid/fix some race conditions * Converted several structs to be OBJs so that we can use reference counting to know when to invoke destructors * Make some new OBJ's have opal_list_item_t's as their base, thereby eliminating the need for the local list_item_t type * Renamed many variables to be internally consistent * Centralize the decision in an inline function as to whether this process or the remote process is supposed to be the initiator * Add oodles of OPAL_OUTPUT statements for debugging (hard-wired to output stream -1; to be activated by developers if they want/need them) * Use rdma_create_qp() instead of ibv_create_qp() * openib fd monitoring improvements: * Renamed a bunch of functions and variables to be a little more obvious as to their true function * Use pipes to communicate between main thread and service thread * Add ability for main thread to invoke a function back on the service thread * Ensure to set initiator_depth and responder_resources properly, but putting max_qp_rd_ataom and ma_qp_init_rd_atom in the modex (see rdma_connect(3)) * Ensure to set the source IP address in rdma_resolve() to ensure that we select the correct OpenFabrics source port * Make new MCA param: openib_btl_connect_rdmacm_resolve_timeout * Other improvements: * btl_openib_device_type MCA param: can be "iw" or "ib" or "all" (or "infiniband" or "iwarp") * Somewhat improved error handling * Bunches of spelling fixes in comments, VERBOSE, and OUTPUT statements * Oodles of little coding style fixes * Changed shutdown ordering of btl; the device is now an OBJ with ref counting for destruction * Added some more show_help error messages * Change configury to only build IBCM / RDMACM if we have threads (because we need a progress thread) This commit was SVN r19686. The following Trac tickets were found above: Ticket 1210 --> https://svn.open-mpi.org/trac/ompi/ticket/1210
2008-10-06 04:46:02 +04:00
for (ep_index=0;
ep_index < opal_pointer_array_get_size(openib_btl->device->endpoints);
ep_index++) {
endpoint=opal_pointer_array_get_item(openib_btl->device->endpoints,
Fixes trac:1210, #1319 Commit from a long-standing Mercurial tree that ended up incorporating a lot of things: * A few fixes for CPC interface changes in all the CPCs * Attempts (but not yet finished) to fix shutdown problems in the IB CM CPC * #1319: add CTS support (i.e., initiator guarantees to send first message; automatically activated for iWARP over the RDMA CM CPC) * Some variable and function renamings to make this be generic (e.g., alloc_credit_frag became alloc_control_frag) * CPCs no longer post receive buffers; they only post a single receive buffer for the CTS if they use CTS. Instead, the main BTL now posts the main sets of receive buffers. * CPCs allocate a CTS buffer only if they're about to make a connection * RDMA CM improvements: * Use threaded mode openib fd monitoring to wait for for RDMA CM events * Synchronize endpoint finalization and disconnection between main thread and service thread to avoid/fix some race conditions * Converted several structs to be OBJs so that we can use reference counting to know when to invoke destructors * Make some new OBJ's have opal_list_item_t's as their base, thereby eliminating the need for the local list_item_t type * Renamed many variables to be internally consistent * Centralize the decision in an inline function as to whether this process or the remote process is supposed to be the initiator * Add oodles of OPAL_OUTPUT statements for debugging (hard-wired to output stream -1; to be activated by developers if they want/need them) * Use rdma_create_qp() instead of ibv_create_qp() * openib fd monitoring improvements: * Renamed a bunch of functions and variables to be a little more obvious as to their true function * Use pipes to communicate between main thread and service thread * Add ability for main thread to invoke a function back on the service thread * Ensure to set initiator_depth and responder_resources properly, but putting max_qp_rd_ataom and ma_qp_init_rd_atom in the modex (see rdma_connect(3)) * Ensure to set the source IP address in rdma_resolve() to ensure that we select the correct OpenFabrics source port * Make new MCA param: openib_btl_connect_rdmacm_resolve_timeout * Other improvements: * btl_openib_device_type MCA param: can be "iw" or "ib" or "all" (or "infiniband" or "iwarp") * Somewhat improved error handling * Bunches of spelling fixes in comments, VERBOSE, and OUTPUT statements * Oodles of little coding style fixes * Changed shutdown ordering of btl; the device is now an OBJ with ref counting for destruction * Added some more show_help error messages * Change configury to only build IBCM / RDMACM if we have threads (because we need a progress thread) This commit was SVN r19686. The following Trac tickets were found above: Ticket 1210 --> https://svn.open-mpi.org/trac/ompi/ticket/1210
2008-10-06 04:46:02 +04:00
ep_index);
if(!endpoint) {
BTL_VERBOSE(("In finalize, got another null endpoint"));
continue;
}
Fixes trac:1210, #1319 Commit from a long-standing Mercurial tree that ended up incorporating a lot of things: * A few fixes for CPC interface changes in all the CPCs * Attempts (but not yet finished) to fix shutdown problems in the IB CM CPC * #1319: add CTS support (i.e., initiator guarantees to send first message; automatically activated for iWARP over the RDMA CM CPC) * Some variable and function renamings to make this be generic (e.g., alloc_credit_frag became alloc_control_frag) * CPCs no longer post receive buffers; they only post a single receive buffer for the CTS if they use CTS. Instead, the main BTL now posts the main sets of receive buffers. * CPCs allocate a CTS buffer only if they're about to make a connection * RDMA CM improvements: * Use threaded mode openib fd monitoring to wait for for RDMA CM events * Synchronize endpoint finalization and disconnection between main thread and service thread to avoid/fix some race conditions * Converted several structs to be OBJs so that we can use reference counting to know when to invoke destructors * Make some new OBJ's have opal_list_item_t's as their base, thereby eliminating the need for the local list_item_t type * Renamed many variables to be internally consistent * Centralize the decision in an inline function as to whether this process or the remote process is supposed to be the initiator * Add oodles of OPAL_OUTPUT statements for debugging (hard-wired to output stream -1; to be activated by developers if they want/need them) * Use rdma_create_qp() instead of ibv_create_qp() * openib fd monitoring improvements: * Renamed a bunch of functions and variables to be a little more obvious as to their true function * Use pipes to communicate between main thread and service thread * Add ability for main thread to invoke a function back on the service thread * Ensure to set initiator_depth and responder_resources properly, but putting max_qp_rd_ataom and ma_qp_init_rd_atom in the modex (see rdma_connect(3)) * Ensure to set the source IP address in rdma_resolve() to ensure that we select the correct OpenFabrics source port * Make new MCA param: openib_btl_connect_rdmacm_resolve_timeout * Other improvements: * btl_openib_device_type MCA param: can be "iw" or "ib" or "all" (or "infiniband" or "iwarp") * Somewhat improved error handling * Bunches of spelling fixes in comments, VERBOSE, and OUTPUT statements * Oodles of little coding style fixes * Changed shutdown ordering of btl; the device is now an OBJ with ref counting for destruction * Added some more show_help error messages * Change configury to only build IBCM / RDMACM if we have threads (because we need a progress thread) This commit was SVN r19686. The following Trac tickets were found above: Ticket 1210 --> https://svn.open-mpi.org/trac/ompi/ticket/1210
2008-10-06 04:46:02 +04:00
if(endpoint->endpoint_btl != openib_btl) {
continue;
Fixes trac:1210, #1319 Commit from a long-standing Mercurial tree that ended up incorporating a lot of things: * A few fixes for CPC interface changes in all the CPCs * Attempts (but not yet finished) to fix shutdown problems in the IB CM CPC * #1319: add CTS support (i.e., initiator guarantees to send first message; automatically activated for iWARP over the RDMA CM CPC) * Some variable and function renamings to make this be generic (e.g., alloc_credit_frag became alloc_control_frag) * CPCs no longer post receive buffers; they only post a single receive buffer for the CTS if they use CTS. Instead, the main BTL now posts the main sets of receive buffers. * CPCs allocate a CTS buffer only if they're about to make a connection * RDMA CM improvements: * Use threaded mode openib fd monitoring to wait for for RDMA CM events * Synchronize endpoint finalization and disconnection between main thread and service thread to avoid/fix some race conditions * Converted several structs to be OBJs so that we can use reference counting to know when to invoke destructors * Make some new OBJ's have opal_list_item_t's as their base, thereby eliminating the need for the local list_item_t type * Renamed many variables to be internally consistent * Centralize the decision in an inline function as to whether this process or the remote process is supposed to be the initiator * Add oodles of OPAL_OUTPUT statements for debugging (hard-wired to output stream -1; to be activated by developers if they want/need them) * Use rdma_create_qp() instead of ibv_create_qp() * openib fd monitoring improvements: * Renamed a bunch of functions and variables to be a little more obvious as to their true function * Use pipes to communicate between main thread and service thread * Add ability for main thread to invoke a function back on the service thread * Ensure to set initiator_depth and responder_resources properly, but putting max_qp_rd_ataom and ma_qp_init_rd_atom in the modex (see rdma_connect(3)) * Ensure to set the source IP address in rdma_resolve() to ensure that we select the correct OpenFabrics source port * Make new MCA param: openib_btl_connect_rdmacm_resolve_timeout * Other improvements: * btl_openib_device_type MCA param: can be "iw" or "ib" or "all" (or "infiniband" or "iwarp") * Somewhat improved error handling * Bunches of spelling fixes in comments, VERBOSE, and OUTPUT statements * Oodles of little coding style fixes * Changed shutdown ordering of btl; the device is now an OBJ with ref counting for destruction * Added some more show_help error messages * Change configury to only build IBCM / RDMACM if we have threads (because we need a progress thread) This commit was SVN r19686. The following Trac tickets were found above: Ticket 1210 --> https://svn.open-mpi.org/trac/ompi/ticket/1210
2008-10-06 04:46:02 +04:00
}
for(i = 0; i < openib_btl->device->eager_rdma_buffers_count; i++) {
if(openib_btl->device->eager_rdma_buffers[i] == endpoint) {
openib_btl->device->eager_rdma_buffers[i] = NULL;
OBJ_RELEASE(endpoint);
}
}
OBJ_RELEASE(endpoint);
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
/* Release SRQ resources */
for(qp = 0; qp < mca_btl_openib_component.num_qps; qp++) {
if(!BTL_OPENIB_QP_TYPE_PP(qp)) {
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(
&openib_btl->qps[qp].u.srq_qp.pending_frags[0]);
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(
&openib_btl->qps[qp].u.srq_qp.pending_frags[1]);
if (ibv_destroy_srq(openib_btl->qps[qp].u.srq_qp.srq)){
BTL_VERBOSE(("Failed to close SRQ %d", qp));
rc = OMPI_ERROR;
}
OBJ_DESTRUCT(&openib_btl->qps[qp].u.srq_qp.pending_frags[0]);
OBJ_DESTRUCT(&openib_btl->qps[qp].u.srq_qp.pending_frags[1]);
}
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
Fixes trac:1210, #1319 Commit from a long-standing Mercurial tree that ended up incorporating a lot of things: * A few fixes for CPC interface changes in all the CPCs * Attempts (but not yet finished) to fix shutdown problems in the IB CM CPC * #1319: add CTS support (i.e., initiator guarantees to send first message; automatically activated for iWARP over the RDMA CM CPC) * Some variable and function renamings to make this be generic (e.g., alloc_credit_frag became alloc_control_frag) * CPCs no longer post receive buffers; they only post a single receive buffer for the CTS if they use CTS. Instead, the main BTL now posts the main sets of receive buffers. * CPCs allocate a CTS buffer only if they're about to make a connection * RDMA CM improvements: * Use threaded mode openib fd monitoring to wait for for RDMA CM events * Synchronize endpoint finalization and disconnection between main thread and service thread to avoid/fix some race conditions * Converted several structs to be OBJs so that we can use reference counting to know when to invoke destructors * Make some new OBJ's have opal_list_item_t's as their base, thereby eliminating the need for the local list_item_t type * Renamed many variables to be internally consistent * Centralize the decision in an inline function as to whether this process or the remote process is supposed to be the initiator * Add oodles of OPAL_OUTPUT statements for debugging (hard-wired to output stream -1; to be activated by developers if they want/need them) * Use rdma_create_qp() instead of ibv_create_qp() * openib fd monitoring improvements: * Renamed a bunch of functions and variables to be a little more obvious as to their true function * Use pipes to communicate between main thread and service thread * Add ability for main thread to invoke a function back on the service thread * Ensure to set initiator_depth and responder_resources properly, but putting max_qp_rd_ataom and ma_qp_init_rd_atom in the modex (see rdma_connect(3)) * Ensure to set the source IP address in rdma_resolve() to ensure that we select the correct OpenFabrics source port * Make new MCA param: openib_btl_connect_rdmacm_resolve_timeout * Other improvements: * btl_openib_device_type MCA param: can be "iw" or "ib" or "all" (or "infiniband" or "iwarp") * Somewhat improved error handling * Bunches of spelling fixes in comments, VERBOSE, and OUTPUT statements * Oodles of little coding style fixes * Changed shutdown ordering of btl; the device is now an OBJ with ref counting for destruction * Added some more show_help error messages * Change configury to only build IBCM / RDMACM if we have threads (because we need a progress thread) This commit was SVN r19686. The following Trac tickets were found above: Ticket 1210 --> https://svn.open-mpi.org/trac/ompi/ticket/1210
2008-10-06 04:46:02 +04:00
/* Finalize the CPC modules on this openib module */
for (i = 0; i < openib_btl->num_cpcs; ++i) {
if (NULL != openib_btl->cpcs[i]->cbm_finalize) {
openib_btl->cpcs[i]->cbm_finalize(openib_btl, openib_btl->cpcs[i]);
}
free(openib_btl->cpcs[i]);
}
free(openib_btl->cpcs);
/* Release device if there are no more users */
Fixes trac:1210, #1319 Commit from a long-standing Mercurial tree that ended up incorporating a lot of things: * A few fixes for CPC interface changes in all the CPCs * Attempts (but not yet finished) to fix shutdown problems in the IB CM CPC * #1319: add CTS support (i.e., initiator guarantees to send first message; automatically activated for iWARP over the RDMA CM CPC) * Some variable and function renamings to make this be generic (e.g., alloc_credit_frag became alloc_control_frag) * CPCs no longer post receive buffers; they only post a single receive buffer for the CTS if they use CTS. Instead, the main BTL now posts the main sets of receive buffers. * CPCs allocate a CTS buffer only if they're about to make a connection * RDMA CM improvements: * Use threaded mode openib fd monitoring to wait for for RDMA CM events * Synchronize endpoint finalization and disconnection between main thread and service thread to avoid/fix some race conditions * Converted several structs to be OBJs so that we can use reference counting to know when to invoke destructors * Make some new OBJ's have opal_list_item_t's as their base, thereby eliminating the need for the local list_item_t type * Renamed many variables to be internally consistent * Centralize the decision in an inline function as to whether this process or the remote process is supposed to be the initiator * Add oodles of OPAL_OUTPUT statements for debugging (hard-wired to output stream -1; to be activated by developers if they want/need them) * Use rdma_create_qp() instead of ibv_create_qp() * openib fd monitoring improvements: * Renamed a bunch of functions and variables to be a little more obvious as to their true function * Use pipes to communicate between main thread and service thread * Add ability for main thread to invoke a function back on the service thread * Ensure to set initiator_depth and responder_resources properly, but putting max_qp_rd_ataom and ma_qp_init_rd_atom in the modex (see rdma_connect(3)) * Ensure to set the source IP address in rdma_resolve() to ensure that we select the correct OpenFabrics source port * Make new MCA param: openib_btl_connect_rdmacm_resolve_timeout * Other improvements: * btl_openib_device_type MCA param: can be "iw" or "ib" or "all" (or "infiniband" or "iwarp") * Somewhat improved error handling * Bunches of spelling fixes in comments, VERBOSE, and OUTPUT statements * Oodles of little coding style fixes * Changed shutdown ordering of btl; the device is now an OBJ with ref counting for destruction * Added some more show_help error messages * Change configury to only build IBCM / RDMACM if we have threads (because we need a progress thread) This commit was SVN r19686. The following Trac tickets were found above: Ticket 1210 --> https://svn.open-mpi.org/trac/ompi/ticket/1210
2008-10-06 04:46:02 +04:00
if (!(--openib_btl->device->btls)) {
OBJ_RELEASE(openib_btl->device);
}
Fixes trac:1210, #1319 Commit from a long-standing Mercurial tree that ended up incorporating a lot of things: * A few fixes for CPC interface changes in all the CPCs * Attempts (but not yet finished) to fix shutdown problems in the IB CM CPC * #1319: add CTS support (i.e., initiator guarantees to send first message; automatically activated for iWARP over the RDMA CM CPC) * Some variable and function renamings to make this be generic (e.g., alloc_credit_frag became alloc_control_frag) * CPCs no longer post receive buffers; they only post a single receive buffer for the CTS if they use CTS. Instead, the main BTL now posts the main sets of receive buffers. * CPCs allocate a CTS buffer only if they're about to make a connection * RDMA CM improvements: * Use threaded mode openib fd monitoring to wait for for RDMA CM events * Synchronize endpoint finalization and disconnection between main thread and service thread to avoid/fix some race conditions * Converted several structs to be OBJs so that we can use reference counting to know when to invoke destructors * Make some new OBJ's have opal_list_item_t's as their base, thereby eliminating the need for the local list_item_t type * Renamed many variables to be internally consistent * Centralize the decision in an inline function as to whether this process or the remote process is supposed to be the initiator * Add oodles of OPAL_OUTPUT statements for debugging (hard-wired to output stream -1; to be activated by developers if they want/need them) * Use rdma_create_qp() instead of ibv_create_qp() * openib fd monitoring improvements: * Renamed a bunch of functions and variables to be a little more obvious as to their true function * Use pipes to communicate between main thread and service thread * Add ability for main thread to invoke a function back on the service thread * Ensure to set initiator_depth and responder_resources properly, but putting max_qp_rd_ataom and ma_qp_init_rd_atom in the modex (see rdma_connect(3)) * Ensure to set the source IP address in rdma_resolve() to ensure that we select the correct OpenFabrics source port * Make new MCA param: openib_btl_connect_rdmacm_resolve_timeout * Other improvements: * btl_openib_device_type MCA param: can be "iw" or "ib" or "all" (or "infiniband" or "iwarp") * Somewhat improved error handling * Bunches of spelling fixes in comments, VERBOSE, and OUTPUT statements * Oodles of little coding style fixes * Changed shutdown ordering of btl; the device is now an OBJ with ref counting for destruction * Added some more show_help error messages * Change configury to only build IBCM / RDMACM if we have threads (because we need a progress thread) This commit was SVN r19686. The following Trac tickets were found above: Ticket 1210 --> https://svn.open-mpi.org/trac/ompi/ticket/1210
2008-10-06 04:46:02 +04:00
/* Remove the btl from component list */
if ( mca_btl_openib_component.ib_num_btls > 1 ) {
for(i = 0; i < mca_btl_openib_component.ib_num_btls; i++){
if (mca_btl_openib_component.openib_btls[i] == openib_btl){
mca_btl_openib_component.openib_btls[i] =
mca_btl_openib_component.openib_btls[mca_btl_openib_component.ib_num_btls-1];
break;
}
}
}
mca_btl_openib_component.ib_num_btls--;
OBJ_DESTRUCT(&openib_btl->ib_lock);
free(openib_btl);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
BTL_VERBOSE(("Success in closing BTL resources"));
return rc;
}
/*
* Initiate a send.
*/
int mca_btl_openib_send(
struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
mca_btl_base_tag_t tag)
{
mca_btl_openib_send_frag_t *frag;
assert(openib_frag_type(des) == MCA_BTL_OPENIB_FRAG_SEND ||
openib_frag_type(des) == MCA_BTL_OPENIB_FRAG_COALESCED);
if(openib_frag_type(des) == MCA_BTL_OPENIB_FRAG_COALESCED) {
to_coalesced_frag(des)->hdr->tag = tag;
to_coalesced_frag(des)->hdr->size = des->des_src->seg_len;
if(ep->nbo)
BTL_OPENIB_HEADER_COALESCED_HTON(*to_coalesced_frag(des)->hdr);
frag = to_coalesced_frag(des)->send_frag;
} else {
frag = to_send_frag(des);
to_com_frag(des)->endpoint = ep;
frag->hdr->tag = tag;
}
return mca_btl_openib_endpoint_send(ep, frag);
}
/*
* RDMA WRITE local buffer to remote buffer address.
*/
int mca_btl_openib_put( mca_btl_base_module_t* btl,
mca_btl_base_endpoint_t* ep,
mca_btl_base_descriptor_t* descriptor)
{
struct ibv_send_wr* bad_wr;
mca_btl_openib_out_frag_t* frag = to_out_frag(descriptor);
int qp = descriptor->order;
uint64_t rem_addr = descriptor->des_dst->seg_addr.lval;
uint32_t rkey = descriptor->des_dst->seg_key.key32[0];
assert(openib_frag_type(frag) == MCA_BTL_OPENIB_FRAG_SEND_USER ||
openib_frag_type(frag) == MCA_BTL_OPENIB_FRAG_SEND);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
if(ep->endpoint_state != MCA_BTL_IB_CONNECTED) {
int rc;
OPAL_THREAD_LOCK(&ep->endpoint_lock);
rc = check_endpoint_state(ep, descriptor, &ep->pending_put_frags);
OPAL_THREAD_UNLOCK(&ep->endpoint_lock);
if(OMPI_ERR_RESOURCE_BUSY == rc)
return OMPI_SUCCESS;
if(OMPI_SUCCESS != rc)
return rc;
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
if(MCA_BTL_NO_ORDER == qp)
qp = mca_btl_openib_component.rdma_qp;
/* check for a send wqe */
if (qp_get_wqe(ep, qp) < 0) {
qp_put_wqe(ep, qp);
OPAL_THREAD_LOCK(&ep->endpoint_lock);
opal_list_append(&ep->pending_put_frags, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&ep->endpoint_lock);
return OMPI_SUCCESS;
}
/* post descriptor */
#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
if((ep->endpoint_proc->proc_ompi->proc_arch & OPAL_ARCH_ISBIGENDIAN)
!= (ompi_proc_local()->proc_arch & OPAL_ARCH_ISBIGENDIAN)) {
rem_addr = opal_swap_bytes8(rem_addr);
rkey = opal_swap_bytes4(rkey);
}
#endif
frag->sr_desc.wr.rdma.remote_addr = rem_addr;
frag->sr_desc.wr.rdma.rkey = rkey;
to_com_frag(frag)->sg_entry.addr =
(uint64_t)(uintptr_t)descriptor->des_src->seg_addr.pval;
to_com_frag(frag)->sg_entry.length = descriptor->des_src->seg_len;
to_com_frag(frag)->endpoint = ep;
#if HAVE_XRC
if (MCA_BTL_XRC_ENABLED && BTL_OPENIB_QP_TYPE_XRC(qp))
frag->sr_desc.xrc_remote_srq_num=ep->rem_info.rem_srqs[qp].rem_srq_num;
#endif
descriptor->order = qp;
/* Setting opcode on a frag constructor isn't enough since prepare_src
* may return send_frag instead of put_frag */
frag->sr_desc.opcode = IBV_WR_RDMA_WRITE;
frag->sr_desc.send_flags = ib_send_flags(descriptor->des_src->seg_len, &(ep->qps[qp]));
if(ibv_post_send(ep->qps[qp].qp->lcl_qp, &frag->sr_desc, &bad_wr))
return OMPI_ERROR;
return OMPI_SUCCESS;
}
/*
* RDMA READ remote buffer to local buffer address.
*/
int mca_btl_openib_get(mca_btl_base_module_t* btl,
mca_btl_base_endpoint_t* ep,
mca_btl_base_descriptor_t* descriptor)
{
struct ibv_send_wr* bad_wr;
mca_btl_openib_get_frag_t* frag = to_get_frag(descriptor);
int qp = descriptor->order;
uint64_t rem_addr = descriptor->des_src->seg_addr.lval;
uint32_t rkey = descriptor->des_src->seg_key.key32[0];
assert(openib_frag_type(frag) == MCA_BTL_OPENIB_FRAG_RECV_USER);
if(ep->endpoint_state != MCA_BTL_IB_CONNECTED) {
int rc;
OPAL_THREAD_LOCK(&ep->endpoint_lock);
rc = check_endpoint_state(ep, descriptor, &ep->pending_get_frags);
OPAL_THREAD_UNLOCK(&ep->endpoint_lock);
if(OMPI_ERR_RESOURCE_BUSY == rc)
return OMPI_SUCCESS;
if(OMPI_SUCCESS != rc)
return rc;
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
if(MCA_BTL_NO_ORDER == qp)
qp = mca_btl_openib_component.rdma_qp;
/* check for a send wqe */
if (qp_get_wqe(ep, qp) < 0) {
qp_put_wqe(ep, qp);
OPAL_THREAD_LOCK(&ep->endpoint_lock);
opal_list_append(&ep->pending_get_frags, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&ep->endpoint_lock);
return OMPI_SUCCESS;
}
/* check for a get token */
if(OPAL_THREAD_ADD32(&ep->get_tokens,-1) < 0) {
qp_put_wqe(ep, qp);
OPAL_THREAD_ADD32(&ep->get_tokens,1);
OPAL_THREAD_LOCK(&ep->endpoint_lock);
opal_list_append(&ep->pending_get_frags, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&ep->endpoint_lock);
return OMPI_SUCCESS;
}
#if OMPI_ENABLE_HETEROGENEOUS_SUPPORT
if((ep->endpoint_proc->proc_ompi->proc_arch & OPAL_ARCH_ISBIGENDIAN)
!= (ompi_proc_local()->proc_arch & OPAL_ARCH_ISBIGENDIAN)) {
rem_addr = opal_swap_bytes8(rem_addr);
rkey = opal_swap_bytes4(rkey);
}
#endif
frag->sr_desc.wr.rdma.remote_addr = rem_addr;
frag->sr_desc.wr.rdma.rkey = rkey;
to_com_frag(frag)->sg_entry.addr =
(uint64_t)(uintptr_t)descriptor->des_dst->seg_addr.pval;
to_com_frag(frag)->sg_entry.length = descriptor->des_dst->seg_len;
to_com_frag(frag)->endpoint = ep;
#if HAVE_XRC
if (MCA_BTL_XRC_ENABLED && BTL_OPENIB_QP_TYPE_XRC(qp))
frag->sr_desc.xrc_remote_srq_num=ep->rem_info.rem_srqs[qp].rem_srq_num;
#endif
descriptor->order = qp;
if(ibv_post_send(ep->qps[qp].qp->lcl_qp, &frag->sr_desc, &bad_wr))
return OMPI_ERROR;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
return OMPI_SUCCESS;
}
#if OPAL_ENABLE_FT == 0
int mca_btl_openib_ft_event(int state) {
return OMPI_SUCCESS;
}
#else
int mca_btl_openib_ft_event(int state) {
int i;
if(OPAL_CRS_CHECKPOINT == state) {
/* Continue must reconstruct the routes (including modex), since we
* have to tear down the devices completely. */
ompi_cr_continue_like_restart = true;
/*
* To keep the node from crashing we need to call ibv_close_device
* before the checkpoint is taken. To do this we need to tear
* everything down, and rebuild it all on continue/restart. :(
*/
/* Shutdown all modules
* - Do this backwards since the openib_finalize function also loops
* over this variable.
*/
for (i = 0; i < mca_btl_openib_component.ib_num_btls; ++i ) {
ft_event_btl_openib_finalize( &(mca_btl_openib_component.openib_btls[i])->super);
}
ompi_btl_openib_connect_base_finalize();
}
else if(OPAL_CRS_CONTINUE == state) {
; /* Cleared by forcing the modex, no work needed */
}
else if(OPAL_CRS_RESTART == state) {
;
}
else if(OPAL_CRS_TERM == state ) {
;
}
else {
;
}
return OMPI_SUCCESS;
}
static int ft_event_btl_openib_finalize(struct mca_btl_base_module_t* btl) {
mca_btl_openib_module_t* openib_btl;
mca_btl_openib_endpoint_t* endpoint;
int ep_index, i;
int qp, rc = OMPI_SUCCESS;
openib_btl = (mca_btl_openib_module_t*) btl;
/* Release all QPs */
for(ep_index=0;
ep_index < opal_pointer_array_get_size(openib_btl->device->endpoints);
ep_index++) {
endpoint=opal_pointer_array_get_item(openib_btl->device->endpoints,
ep_index);
if(!endpoint) {
BTL_VERBOSE(("In finalize, got another null endpoint"));
continue;
}
if(endpoint->endpoint_btl != openib_btl)
continue;
for(i = 0; i < openib_btl->device->eager_rdma_buffers_count; i++) {
if(openib_btl->device->eager_rdma_buffers[i] == endpoint) {
openib_btl->device->eager_rdma_buffers[i] = NULL;
OBJ_RELEASE(endpoint);
}
}
OBJ_RELEASE(endpoint);
}
/* Finalize the CPC modules on this openib module */
for (i = 0; i < openib_btl->num_cpcs; ++i) {
if (NULL != openib_btl->cpcs[i]->cbm_finalize) {
openib_btl->cpcs[i]->cbm_finalize(openib_btl, openib_btl->cpcs[i]);
}
free(openib_btl->cpcs[i]);
}
free(openib_btl->cpcs);
/* Release SRQ resources */
for(qp = 0; qp < mca_btl_openib_component.num_qps; qp++) {
if(!BTL_OPENIB_QP_TYPE_PP(qp)) {
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(
&openib_btl->qps[qp].u.srq_qp.pending_frags[0]);
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(
&openib_btl->qps[qp].u.srq_qp.pending_frags[1]);
if (ibv_destroy_srq(openib_btl->qps[qp].u.srq_qp.srq)){
BTL_VERBOSE(("Failed to close SRQ %d", qp));
rc = OMPI_ERROR;
}
OBJ_DESTRUCT(&openib_btl->qps[qp].u.srq_qp.pending_frags[0]);
OBJ_DESTRUCT(&openib_btl->qps[qp].u.srq_qp.pending_frags[1]);
}
}
/* Release device if there are no more users */
if(!(--openib_btl->device->btls)) {
OBJ_RELEASE(openib_btl->device);
}
mca_btl_openib_component.devices_count = 0;
mca_btl_openib_component.ib_num_btls = 0;
OBJ_DESTRUCT(&mca_btl_openib_component.ib_procs);
BTL_VERBOSE(("Success in closing BTL resources"));
return rc;
}
#endif /* OPAL_ENABLE_FT */