Hi Brad,
WHAT: Adds Cross Memory Attach support to the sm btl
WHY: For faster intranode communication
WHERE: ompi/mca/btl/sm/
WHEN: Open MPI trunk
TIMEOUT: 13/2/2012
Cross Memory Attach (CMA) is a pair of new syscalls (process_vm_readv
and process_vm_writev) which allow for fast intranode
communication. It has added to the Linux 3.2 kernel. There is a man page
for the new system calls here:
http://ozlabs.org/~cyeoh/cma/process_vm_readv.txt
Attached is a patch for the OMPI trunk tree which augments the sm btl
to use these calls.
- CMA is quite similar in many respects to KNEM and what I've done is to
pretty much copy what KNEM does in many cases.
- Both KNEM and CMA can be compiled in at the same time, though if
both are enabled at runtime, only KNEM runs. To enable CMA use
--mca btl_sm_use_cma 1
- To enable CMA at compile time, add --with-cma=yes to the configure
command line. Support for the syscalls is in the git glibc archive,
but its not yet out in the distros so as an interim
workaround I have added some arch/os specific wrappers which are used
if the syscalls are not found at configure time. The syscalls numbers
won't change as 3.2 is out.
- I'm far from sure that the way I have used CMA in OMPI is the best
way to do it, so any feedback is very welcome.
Regards,
Chris
--
[email protected]
Index: opal/include/opal/sys/cma.h
===================================================================
--- opal/include/opal/sys/cma.h (revision 0)
+++ opal/include/opal/sys/cma.h (revision 0)
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2011 IBM Corporation. All rights reserved.
+ *
+ */
+
+/** @file
+ *
+ * Cross Memory Attach syscall definitions.
+ *
+ * These are only needed temporarily until these new syscalls
+ * are incorporated into glibc
+ */
+
+#ifndef OPAL_SYS_CMA_H
+#define OPAL_SYS_CMA_H 1
+
+#include "opal_config.h"
+
+#include "opal/sys/architecture.h"
+
+#ifdef HAVE_SYS_TYPES_H
+#include <sys/types.h>
+#endif
+
+#ifdef HAVE_UNISTD_H
+#include <sys/unistd.h>
+#endif
+
+#ifdef __linux__
+
+/* Cross Memory Attach is so far only supported under linux */
+
+#if OPAL_ASSEMBLY_ARCH == OMPI_AMD64
+#define __NR_process_vm_readv 310
+#define __NR_process_vm_writev 311
+#elif OPAL_ASSEMBLY_ARCH == OMPI_IA32
+#define __NR_process_vm_readv 347
+#define __NR_process_vm_writev 348
+#elif OPAL_ASSEMBLY_ARCH == OMPI_IA64
+#define __NR_process_vm_readv 310
+#define __NR_process_vm_writev 311
+#elif OPAL_ASSEMBLY_ARCH == OMPI_POWERPC32
+#define __NR_process_vm_readv 351
+#define __NR_process_vm_writev 352
+#elif OPAL_ASSEMBLY_ARCH == OMPI_POWERPC64
+#define __NR_process_vm_readv 351
+#define __NR_process_vm_writev 352
+#else
+#error "Unsupported architecture for process_vm_readv and process_vm_writev syscalls"
+#endif
+
+
+inline ssize_t process_vm_readv(pid_t pid,
+ const struct iovec *lvec,
+ unsigned long liovcnt,
+ const struct iovec *rvec,
+ unsigned long riovcnt,
+ unsigned long flags)
+{
+ return syscall(__NR_process_vm_readv, pid, lvec, liovcnt, rvec, riovcnt, flags);
+}
+
+inline ssize_t process_vm_writev(pid_t pid,
+ const struct iovec *lvec,
+ unsigned long liovcnt,
+ const struct iovec *rvec,
+ unsigned long riovcnt,
+ unsigned long flags)
+{
+ return syscall(__NR_process_vm_writev, pid, lvec, liovcnt, rvec, riovcnt, flags);
+}
+
+#endif /* __linux__ */
+
+#endif /* OPAL_SYS_CMA_H */
Index: ompi/mca/btl/sm/configure.m4
===================================================================
--- ompi/mca/btl/sm/configure.m4 (revision 25480)
+++ ompi/mca/btl/sm/configure.m4 (working copy)
@@ -4,6 +4,7 @@
# of Tennessee Research Foundation. All rights
# reserved.
# Copyright (c) 2009-2010 Cisco Systems, Inc. All rights reserved.
+# Copyright (c) 2010-2011 IBM Corporation. All rights reserved.
# $COPYRIGHT$
#
# Additional copyrights may follow
@@ -11,6 +12,30 @@
# $HEADER$
#
+
+# OMPI_CHECK_CMA(prefix, [action-if-found], [action-if-not-found])
+# --------------------------------------------------------
+# check if cma support is wanted.
+AC_DEFUN([OMPI_CHECK_CMA],[
+ AC_ARG_WITH([cma],
+ [AC_HELP_STRING([--with-cma],
+ [Build Cross Memory Attach support (default: no)])])
+
+ AC_MSG_CHECKING([if user requested CMA build])
+ if test "$with_cma" = "yes" ; then
+ btl_sm_cma_happy="yes"
+ AC_CHECK_FUNC(process_vm_readv, [btl_sm_cma_need_defs=0],
+ [btl_sm_cma_need_defs=1])
+ AC_DEFINE_UNQUOTED([OMPI_BTL_SM_CMA_NEED_SYSCALL_DEFS],
+ [$btl_sm_cma_need_defs],
+ [Need CMA syscalls defined])
+ fi
+ AS_IF([test "$btl_sm_cma_happy" = "yes"],
+ [$2],
+ [$3])
+])dnl
+
+
# OMPI_CHECK_KNEM(prefix, [action-if-found], [action-if-not-found])
# --------------------------------------------------------
# check if knem support can be found. sets prefix_{CPPFLAGS,
@@ -69,6 +94,15 @@
AC_DEFUN([MCA_ompi_btl_sm_CONFIG],[
AC_CONFIG_FILES([ompi/mca/btl/sm/Makefile])
+ OPAL_VAR_SCOPE_PUSH([btl_sm_cma_happy])
+ OMPI_CHECK_CMA([btl_sm], [btl_sm_cma_happy=1], [btl_sm_cma_happy=0])
+
+ AC_DEFINE_UNQUOTED([OMPI_BTL_SM_HAVE_CMA],
+ [$btl_sm_cma_happy],
+ [If CMA support can be enabled])
+
+ OPAL_VAR_SCOPE_POP
+
OPAL_VAR_SCOPE_PUSH([btl_sm_knem_happy])
OMPI_CHECK_KNEM([btl_sm],
[btl_sm_knem_happy=1],
Index: ompi/mca/btl/sm/btl_sm.c
===================================================================
--- ompi/mca/btl/sm/btl_sm.c (revision 25480)
+++ ompi/mca/btl/sm/btl_sm.c (working copy)
@@ -13,6 +13,7 @@
* Copyright (c) 2009 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2010 Los Alamos National Security, LLC.
* All rights reserved.
+ * Copyright (c) 2010-2011 IBM Corporation. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@@ -32,6 +33,10 @@
#include <sys/mman.h>
#endif /* HAVE_SYS_MMAN_H */
+#ifdef OMPI_BTL_SM_CMA_NEED_SYSCALL_DEFS
+#include "opal/sys/cma.h"
+#endif /* OMPI_BTL_SM_CMA_NEED_SYSCALL_DEFS */
+
#include "opal/sys/atomic.h"
#include "opal/class/opal_bitmap.h"
#include "opal/util/output.h"
@@ -80,11 +85,11 @@
mca_btl_sm_alloc,
mca_btl_sm_free,
mca_btl_sm_prepare_src,
-#if OMPI_BTL_SM_HAVE_KNEM
+#if OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA
mca_btl_sm_prepare_dst,
#else
NULL,
-#endif /* OMPI_BTL_SM_HAVE_KNEM */
+#endif /* OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA */
mca_btl_sm_send,
mca_btl_sm_sendi,
NULL, /* put */
@@ -684,12 +689,12 @@
uint32_t iov_count = 1;
size_t max_data = *size;
int rc;
-#if OMPI_BTL_SM_HAVE_KNEM
+
+#if OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA
mca_btl_sm_t* sm_btl = (mca_btl_sm_t*)btl;
- struct knem_cmd_create_region knem_cr;
- struct knem_cmd_param_iovec knem_iov;
- if( (0 != reserve) || (OPAL_UNLIKELY(!mca_btl_sm_component.use_knem)) ) {
+ if( (0 != reserve) || ( OPAL_UNLIKELY(!mca_btl_sm_component.use_knem)
+ && OPAL_UNLIKELY(!mca_btl_sm_component.use_cma)) ) {
#endif
if ( reserve + max_data <= mca_btl_sm_component.eager_limit ) {
MCA_BTL_SM_FRAG_ALLOC_EAGER(frag,rc);
@@ -714,8 +719,12 @@
return NULL;
}
frag->segment.seg_len = reserve + max_data;
+#if OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA
+ } else {
#if OMPI_BTL_SM_HAVE_KNEM
- } else {
+ struct knem_cmd_create_region knem_cr;
+ struct knem_cmd_param_iovec knem_iov;
+#endif /* OMPI_BTL_SM_HAVE_KNEM */
MCA_BTL_SM_FRAG_ALLOC_USER(frag, rc);
if( OPAL_UNLIKELY(NULL == frag) ) {
return NULL;
@@ -730,18 +739,30 @@
frag->segment.seg_addr.pval = iov.iov_base;
frag->segment.seg_len = max_data;
- knem_iov.base = (uintptr_t)iov.iov_base;
- knem_iov.len = max_data;
- knem_cr.iovec_array = (uintptr_t)&knem_iov;
- knem_cr.iovec_nr = iov_count;
- knem_cr.protection = PROT_READ;
- knem_cr.flags = KNEM_FLAG_SINGLEUSE;
- if (OPAL_UNLIKELY(ioctl(sm_btl->knem_fd, KNEM_CMD_CREATE_REGION, &knem_cr) < 0)) {
- return NULL;
+#if OMPI_BTL_SM_HAVE_KNEM
+ if (OPAL_LIKELY(mca_btl_sm_component.use_knem)) {
+ knem_iov.base = (uintptr_t)iov.iov_base;
+ knem_iov.len = max_data;
+ knem_cr.iovec_array = (uintptr_t)&knem_iov;
+ knem_cr.iovec_nr = iov_count;
+ knem_cr.protection = PROT_READ;
+ knem_cr.flags = KNEM_FLAG_SINGLEUSE;
+ if (OPAL_UNLIKELY(ioctl(sm_btl->knem_fd, KNEM_CMD_CREATE_REGION, &knem_cr) < 0)) {
+ return NULL;
+ }
+ frag->segment.seg_key.key64[0] = knem_cr.cookie;
}
- frag->segment.seg_key.key64[0] = knem_cr.cookie;
+#endif /* OMPI_BTL_SM_HAVE_KNEM */
+
+#if OMPI_BTL_SM_HAVE_CMA
+ if (OPAL_LIKELY(mca_btl_sm_component.use_cma)) {
+ /* Encode the pid as the key */
+ frag->segment.seg_key.key64[0] = getpid();
+ }
+#endif /* OMPI_BTL_SM_HAVE_CMA */
}
-#endif
+#endif /* OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA */
+
frag->base.des_src = &(frag->segment);
frag->base.des_src_cnt = 1;
frag->base.order = MCA_BTL_NO_ORDER;
@@ -913,7 +934,7 @@
return 0;
}
-#if OMPI_BTL_SM_HAVE_KNEM
+#if OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA
struct mca_btl_base_descriptor_t* mca_btl_sm_prepare_dst(
struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
@@ -931,7 +952,7 @@
if(OPAL_UNLIKELY(NULL == frag)) {
return NULL;
}
-
+
frag->segment.seg_len = *size;
opal_convertor_get_current_pointer( convertor, (void**)&(frag->segment.seg_addr.pval) );
@@ -959,38 +980,78 @@
mca_btl_sm_frag_t* frag = (mca_btl_sm_frag_t*)des;
mca_btl_base_segment_t *src = des->des_src;
mca_btl_base_segment_t *dst = des->des_dst;
- struct knem_cmd_inline_copy icopy;
- struct knem_cmd_param_iovec recv_iovec;
+#if OMPI_BTL_SM_HAVE_KNEM
+ if (OPAL_LIKELY(mca_btl_sm_component.use_knem)) {
+ struct knem_cmd_inline_copy icopy;
+ struct knem_cmd_param_iovec recv_iovec;
- /* Fill in the ioctl data fields. There's no async completion, so
- we don't need to worry about getting a slot, etc. */
- recv_iovec.base = (uintptr_t) dst->seg_addr.pval;
- recv_iovec.len = dst->seg_len;
- icopy.local_iovec_array = (uintptr_t)&recv_iovec;
- icopy.local_iovec_nr = 1;
- icopy.remote_cookie = src->seg_key.key64[0];
- icopy.remote_offset = 0;
- icopy.write = 0;
+ /* Fill in the ioctl data fields. There's no async completion, so
+ we don't need to worry about getting a slot, etc. */
+ recv_iovec.base = (uintptr_t) dst->seg_addr.pval;
+ recv_iovec.len = dst->seg_len;
+ icopy.local_iovec_array = (uintptr_t)&recv_iovec;
+ icopy.local_iovec_nr = 1;
+ icopy.remote_cookie = src->seg_key.key64[0];
+ icopy.remote_offset = 0;
+ icopy.write = 0;
- /* Use the DMA flag if knem supports it *and* the segment length
- is greater than the cutoff. Note that if the knem_dma_min
- value is 0 (i.e., the MCA param was set to 0), the segment size
- will never be larger than it, so DMA will never be used. */
- icopy.flags = 0;
- if (mca_btl_sm_component.knem_dma_min <= dst->seg_len) {
- icopy.flags = mca_btl_sm_component.knem_dma_flag;
+ /* Use the DMA flag if knem supports it *and* the segment length
+ is greater than the cutoff. Note that if the knem_dma_min
+ value is 0 (i.e., the MCA param was set to 0), the segment size
+ will never be larger than it, so DMA will never be used. */
+ icopy.flags = 0;
+ if (mca_btl_sm_component.knem_dma_min <= dst->seg_len) {
+ icopy.flags = mca_btl_sm_component.knem_dma_flag;
+ }
+ /* synchronous flags only, no need to specify icopy.async_status_index */
+
+ /* When the ioctl returns, the transfer is done and we can invoke
+ the btl callback and return the frag */
+ if (OPAL_UNLIKELY(0 != ioctl(sm_btl->knem_fd,
+ KNEM_CMD_INLINE_COPY, &icopy))) {
+ return OMPI_ERROR;
+ }
+
+ /* FIXME: what if icopy.current_status == KNEM_STATUS_FAILED? */
}
- /* synchronous flags only, no need to specify icopy.async_status_index */
+#endif
- /* When the ioctl returns, the transfer is done and we can invoke
- the btl callback and return the frag */
- if (OPAL_UNLIKELY(0 != ioctl(sm_btl->knem_fd,
- KNEM_CMD_INLINE_COPY, &icopy))) {
- return OMPI_ERROR;
+#if OMPI_BTL_SM_HAVE_CMA
+ if (OPAL_LIKELY(mca_btl_sm_component.use_cma)) {
+ char *remote_address, *local_address;
+ int remote_length, local_length;
+ struct iovec local, remote;
+ pid_t remote_pid;
+ int val;
+
+ remote_address = (char *) src->seg_addr.pval;
+ remote_length = src->seg_len;
+
+ local_address = (char *) dst->seg_addr.pval;
+ local_length = dst->seg_len;
+
+ remote_pid = src->seg_key.key64[0];
+ remote.iov_base = src->seg_addr.pval;
+ remote.iov_len = src->seg_len;
+ local.iov_base = dst->seg_addr.pval;
+ local.iov_len = dst->seg_len;
+
+ val = process_vm_readv(remote_pid, &local, 1, &remote, 1, 0);
+
+ if (val != local_length) {
+ if (val<0) {
+ opal_output(0, "mca_btl_sm_get_sync: process_vm_readv failed: %i",
+ errno);
+ } else {
+ /* Should never get a short read from process_vm_readv */
+ opal_output(0, "mca_btl_sm_get_sync: process_vm_readv short read: %i",
+ val);
+ }
+ return OMPI_ERROR;
+ }
}
+#endif /* OMPI_BTL_SM_HAVE_CMA */
- /* FIXME: what if icopy.current_status == KNEM_STATUS_FAILED? */
-
btl_ownership = (frag->base.des_flags & MCA_BTL_DES_FLAGS_BTL_OWNERSHIP);
if (0 != (MCA_BTL_DES_SEND_ALWAYS_CALLBACK & frag->base.des_flags)) {
frag->base.des_cbfunc(&mca_btl_sm.super,
@@ -1004,7 +1065,12 @@
return OMPI_SUCCESS;
}
+#endif /* OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA */
+
+/* No support async_get for CMA yet */
+#if OMPI_BTL_SM_HAVE_KNEM
+
/**
* Initiate an asynchronous get.
*
@@ -1084,7 +1150,7 @@
return OMPI_ERROR;
}
}
-#endif
+#endif /* OMPI_BTL_SM_HAVE_KNEM */
#if OPAL_ENABLE_FT_CR == 0
Index: ompi/mca/btl/sm/btl_sm_component.c
===================================================================
--- ompi/mca/btl/sm/btl_sm_component.c (revision 25480)
+++ ompi/mca/btl/sm/btl_sm_component.c (working copy)
@@ -14,6 +14,7 @@
* Copyright (c) 2010-2011 Los Alamos National Security, LLC.
* All rights reserved.
* Copyright (c) 2011 NVIDIA Corporation. All rights reserved.
+ * Copyright (c) 2010-2011 IBM Corporation. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@@ -172,6 +173,13 @@
false, false, 0,
&mca_btl_sm_component.knem_max_simultaneous);
+ /* CMA parameters */
+ mca_base_param_reg_int(&mca_btl_sm_component.super.btl_version,
+ "use_cma",
+ "Whether or not to enable CMA",
+ false, false, 0, &mca_btl_sm_component.use_cma);
+
+
/* register SM component parameters */
mca_btl_sm_component.sm_free_list_num =
mca_btl_sm_param_register_int("free_list_num", 8);
@@ -203,10 +211,17 @@
mca_btl_sm.super.btl_rdma_pipeline_frag_size = 64*1024;
mca_btl_sm.super.btl_min_rdma_pipeline_size = 64*1024;
mca_btl_sm.super.btl_flags = MCA_BTL_FLAGS_SEND;
-#if OMPI_BTL_SM_HAVE_KNEM
- if (mca_btl_sm_component.use_knem) {
+#if OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA
+ if (mca_btl_sm_component.use_knem || mca_btl_sm_component.use_cma) {
mca_btl_sm.super.btl_flags |= MCA_BTL_FLAGS_GET;
}
+
+ if (mca_btl_sm_component.use_knem && mca_btl_sm_component.use_cma) {
+ /* Disable CMA if knem is runtime enabled */
+ opal_output(0, "CMA disabled because knem is enabled");
+ mca_btl_sm_component.use_cma = 0;
+ }
+
#endif
mca_btl_sm.super.btl_bandwidth = 9000; /* Mbs */
mca_btl_sm.super.btl_latency = 1; /* Microsecs */
@@ -405,96 +420,106 @@
mca_btl_sm.btl_inited = false;
#if OMPI_BTL_SM_HAVE_KNEM
- /* Set knem_status_num_used outside the check for use_knem so that
- we can only have to check one thing (knem_status_num_used) in
- the progress loop. */
- mca_btl_sm.knem_fd = -1;
- mca_btl_sm.knem_status_array = NULL;
- mca_btl_sm.knem_frag_array = NULL;
- mca_btl_sm.knem_status_num_used = 0;
- mca_btl_sm.knem_status_first_avail = 0;
- mca_btl_sm.knem_status_first_used = 0;
+ if (mca_btl_sm_component.use_knem) {
+ /* Set knem_status_num_used outside the check for use_knem so that
+ we can only have to check one thing (knem_status_num_used) in
+ the progress loop. */
+ mca_btl_sm.knem_fd = -1;
+ mca_btl_sm.knem_status_array = NULL;
+ mca_btl_sm.knem_frag_array = NULL;
+ mca_btl_sm.knem_status_num_used = 0;
+ mca_btl_sm.knem_status_first_avail = 0;
+ mca_btl_sm.knem_status_first_used = 0;
- if (0 != mca_btl_sm_component.use_knem) {
- /* Open the knem device. Try to print a helpful message if we
- fail to open it. */
- mca_btl_sm.knem_fd = open("/dev/knem", O_RDWR);
- if (mca_btl_sm.knem_fd < 0) {
- if (EACCES == errno) {
- struct stat sbuf;
- if (0 != stat("/dev/knem", &sbuf)) {
- sbuf.st_mode = 0;
+ if (0 != mca_btl_sm_component.use_knem) {
+ /* Open the knem device. Try to print a helpful message if we
+ fail to open it. */
+ mca_btl_sm.knem_fd = open("/dev/knem", O_RDWR);
+ if (mca_btl_sm.knem_fd < 0) {
+ if (EACCES == errno) {
+ struct stat sbuf;
+ if (0 != stat("/dev/knem", &sbuf)) {
+ sbuf.st_mode = 0;
+ }
+ orte_show_help("help-mpi-btl-sm.txt", "knem permission denied",
+ true, orte_process_info.nodename, sbuf.st_mode);
+ } else {
+ orte_show_help("help-mpi-btl-sm.txt", "knem fail open",
+ true, orte_process_info.nodename, errno,
+ strerror(errno));
}
- orte_show_help("help-mpi-btl-sm.txt", "knem permission denied",
- true, orte_process_info.nodename, sbuf.st_mode);
- } else {
- orte_show_help("help-mpi-btl-sm.txt", "knem fail open",
- true, orte_process_info.nodename, errno,
- strerror(errno));
+ goto no_knem;
}
- goto no_knem;
- }
- /* Check that the ABI if the kernel module running is the same
- as what we were compiled against */
- rc = ioctl(mca_btl_sm.knem_fd, KNEM_CMD_GET_INFO,
- &mca_btl_sm_component.knem_info);
- if (rc < 0) {
- orte_show_help("help-mpi-btl-sm.txt", "knem get ABI fail",
- true, orte_process_info.nodename, errno,
- strerror(errno));
- goto no_knem;
- }
- if (KNEM_ABI_VERSION != mca_btl_sm_component.knem_info.abi) {
- orte_show_help("help-mpi-btl-sm.txt", "knem ABI mismatch",
- true, orte_process_info.nodename, KNEM_ABI_VERSION,
- mca_btl_sm_component.knem_info.abi);
- goto no_knem;
- }
-
- /* If we want DMA mode and DMA mode is supported, then set
- knem_dma_flag to KNEM_FLAG_DMA. */
- mca_btl_sm_component.knem_dma_flag = 0;
- if (mca_btl_sm_component.knem_dma_min > 0 &&
- (mca_btl_sm_component.knem_info.features & KNEM_FEATURE_DMA)) {
- mca_btl_sm_component.knem_dma_flag = KNEM_FLAG_DMA;
- }
-
- /* Get the array of statuses from knem if max_simultaneous > 0 */
- if (mca_btl_sm_component.knem_max_simultaneous > 0) {
- mca_btl_sm.knem_status_array = mmap(NULL,
- mca_btl_sm_component.knem_max_simultaneous,
- (PROT_READ | PROT_WRITE),
- MAP_SHARED, mca_btl_sm.knem_fd,
- KNEM_STATUS_ARRAY_FILE_OFFSET);
- if (MAP_FAILED == mca_btl_sm.knem_status_array) {
- orte_show_help("help-mpi-btl-sm.txt", "knem mmap fail",
+ /* Check that the ABI if the kernel module running is the same
+ as what we were compiled against */
+ rc = ioctl(mca_btl_sm.knem_fd, KNEM_CMD_GET_INFO,
+ &mca_btl_sm_component.knem_info);
+ if (rc < 0) {
+ orte_show_help("help-mpi-btl-sm.txt", "knem get ABI fail",
true, orte_process_info.nodename, errno,
strerror(errno));
goto no_knem;
}
-
- /* The first available status index is 0. Make an empty frag
- array. */
- mca_btl_sm.knem_frag_array = (mca_btl_sm_frag_t **)
- malloc(sizeof(mca_btl_sm_frag_t *) *
- mca_btl_sm_component.knem_max_simultaneous);
- if (NULL == mca_btl_sm.knem_frag_array) {
- orte_show_help("help-mpi-btl-sm.txt", "knem init fail",
- true, orte_process_info.nodename, "malloc",
- errno, strerror(errno));
+ if (KNEM_ABI_VERSION != mca_btl_sm_component.knem_info.abi) {
+ orte_show_help("help-mpi-btl-sm.txt", "knem ABI mismatch",
+ true, orte_process_info.nodename, KNEM_ABI_VERSION,
+ mca_btl_sm_component.knem_info.abi);
goto no_knem;
}
+
+ /* If we want DMA mode and DMA mode is supported, then set
+ knem_dma_flag to KNEM_FLAG_DMA. */
+ mca_btl_sm_component.knem_dma_flag = 0;
+ if (mca_btl_sm_component.knem_dma_min > 0 &&
+ (mca_btl_sm_component.knem_info.features & KNEM_FEATURE_DMA)) {
+ mca_btl_sm_component.knem_dma_flag = KNEM_FLAG_DMA;
+ }
+
+ /* Get the array of statuses from knem if max_simultaneous > 0 */
+ if (mca_btl_sm_component.knem_max_simultaneous > 0) {
+ mca_btl_sm.knem_status_array = mmap(NULL,
+ mca_btl_sm_component.knem_max_simultaneous,
+ (PROT_READ | PROT_WRITE),
+ MAP_SHARED, mca_btl_sm.knem_fd,
+ KNEM_STATUS_ARRAY_FILE_OFFSET);
+ if (MAP_FAILED == mca_btl_sm.knem_status_array) {
+ orte_show_help("help-mpi-btl-sm.txt", "knem mmap fail",
+ true, orte_process_info.nodename, errno,
+ strerror(errno));
+ goto no_knem;
+ }
+
+ /* The first available status index is 0. Make an empty frag
+ array. */
+ mca_btl_sm.knem_frag_array = (mca_btl_sm_frag_t **)
+ malloc(sizeof(mca_btl_sm_frag_t *) *
+ mca_btl_sm_component.knem_max_simultaneous);
+ if (NULL == mca_btl_sm.knem_frag_array) {
+ orte_show_help("help-mpi-btl-sm.txt", "knem init fail",
+ true, orte_process_info.nodename, "malloc",
+ errno, strerror(errno));
+ goto no_knem;
+ }
+ }
}
+ /* Set the BTL get function pointer if we're supporting KNEM;
+ choose between synchronous and asynchronous. */
+ if (mca_btl_sm_component.knem_max_simultaneous > 0) {
+ mca_btl_sm.super.btl_get = mca_btl_sm_get_async;
+ } else {
+ mca_btl_sm.super.btl_get = mca_btl_sm_get_sync;
+ }
}
- /* Set the BTL get function pointer if we're supporting KNEM;
- choose between synchronous and asynchronous. */
- if (mca_btl_sm_component.knem_max_simultaneous > 0) {
- mca_btl_sm.super.btl_get = mca_btl_sm_get_async;
- } else {
+#endif /* OMPI_BTL_SM_HAVE_KNEM */
+
+#if OMPI_BTL_SM_HAVE_CMA
+ if (mca_btl_sm_component.use_cma) {
+ /* Will only ever have either cma or knem enabled at runtime
+ so no problems with accidentally overwriting this set earlier */
mca_btl_sm.super.btl_get = mca_btl_sm_get_sync;
}
-#endif
+#endif /* OMPI_BTL_SM_HAVE_CMA */
return btls;
Index: ompi/mca/btl/sm/btl_sm.h
===================================================================
--- ompi/mca/btl/sm/btl_sm.h (revision 25480)
+++ ompi/mca/btl/sm/btl_sm.h (working copy)
@@ -13,6 +13,7 @@
* Copyright (c) 2009-2010 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2010 Los Alamos National Security, LLC.
* All rights reserved.
+ * Copyright (c) 2010-2011 IBM Corporation. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@@ -197,6 +198,10 @@
/** If we want DMA and DMA is supported, this will be loaded with
KNEM_FLAG_DMA. Otherwise, it'll be 0. */
int knem_dma_flag;
+
+ /** MCA: should we be using CMA or not?
+ 0 = no, 1 = yes */
+ int use_cma;
};
typedef struct mca_btl_sm_component_t mca_btl_sm_component_t;
OMPI_MODULE_DECLSPEC extern mca_btl_sm_component_t mca_btl_sm_component;
@@ -492,21 +497,14 @@
mca_btl_base_tag_t tag
);
-#if OMPI_BTL_SM_HAVE_KNEM
+#if OMPI_BTL_SM_HAVE_KNEM || OMPI_BTL_SM_HAVE_CMA
/*
- * Synchronous knem get
+ * Synchronous knem/cma get
*/
extern int mca_btl_sm_get_sync(
struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct mca_btl_base_descriptor_t* des );
-/*
- * Asynchronous knem get
- */
-extern int mca_btl_sm_get_async(
- struct mca_btl_base_module_t* btl,
- struct mca_btl_base_endpoint_t* endpoint,
- struct mca_btl_base_descriptor_t* des );
extern struct mca_btl_base_descriptor_t* mca_btl_sm_prepare_dst(
struct mca_btl_base_module_t* btl,
@@ -519,6 +517,17 @@
uint32_t flags);
#endif
+#if OMPI_BTL_SM_HAVE_KNEM
+/*
+ * Asynchronous knem get
+ */
+extern int mca_btl_sm_get_async(
+ struct mca_btl_base_module_t* btl,
+ struct mca_btl_base_endpoint_t* endpoint,
+ struct mca_btl_base_descriptor_t* des );
+
+#endif
+
/**
* Fault Tolerance Event Notification Function
* @param state Checkpoint Stae
