rpms/openmpi
7
0
Fork 0
Code Issues Pull Requests Releases Activity

Add upstream patch to fix race/hang on 32bit machines

Browse Source
This commit is contained in:
Orion Poplawski 2015-03-30 16:51:42 -06:00
parent f830453701
commit 3923dafb55
2 changed files with 110 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

103
openmpi-vader.patch Normal file
View File

@ -0,0 +1,103 @@
diff -up openmpi-v1.8.4-134-g9ad2aa8/ompi/mca/btl/vader/btl_vader_endpoint.h.vader openmpi-v1.8.4-134-g9ad2aa8/ompi/mca/btl/vader/btl_vader_endpoint.h
--- openmpi-v1.8.4-134-g9ad2aa8/ompi/mca/btl/vader/btl_vader_endpoint.h.vader 2015-03-24 19:21:54.000000000 -0600
+++ openmpi-v1.8.4-134-g9ad2aa8/ompi/mca/btl/vader/btl_vader_endpoint.h 2015-03-30 16:35:41.411110838 -0600
@@ -11,7 +11,7 @@
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006-2007 Voltaire. All rights reserved.
- * Copyright (c) 2012-2014 Los Alamos National Security, LLC. All rights
+ * Copyright (c) 2012-2015 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
@@ -48,14 +48,16 @@ typedef struct mca_btl_base_endpoint_t {
/* per peer buffers */
struct {
unsigned char *buffer; /**< starting address of peer's fast box out */
- unsigned int start, seq;
uint32_t *startp;
+ unsigned int start;
+ uint16_t seq;
} fbox_in;
struct {
unsigned char *buffer; /**< starting address of peer's fast box in */
- unsigned int start, end, seq;
uint32_t *startp; /**< pointer to location storing start offset */
+ unsigned int start, end;
+ uint16_t seq;
} fbox_out;
int32_t peer_smp_rank; /**< my peer's SMP process rank. Used for accessing
diff -up openmpi-v1.8.4-134-g9ad2aa8/ompi/mca/btl/vader/btl_vader_fbox.h.vader openmpi-v1.8.4-134-g9ad2aa8/ompi/mca/btl/vader/btl_vader_fbox.h
--- openmpi-v1.8.4-134-g9ad2aa8/ompi/mca/btl/vader/btl_vader_fbox.h.vader 2015-03-24 19:21:54.000000000 -0600
+++ openmpi-v1.8.4-134-g9ad2aa8/ompi/mca/btl/vader/btl_vader_fbox.h 2015-03-30 16:35:41.412110833 -0600
@@ -1,6 +1,6 @@
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
- * Copyright (c) 2011-2014 Los Alamos National Security, LLC. All rights
+ * Copyright (c) 2011-2015 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
@@ -18,9 +18,16 @@
typedef union mca_btl_vader_fbox_hdr_t {
struct {
+ /* NTH: on 32-bit platforms loading/unloading the header may be completed
+ * in multiple instructions. To ensure that seq is never loaded before tag
+ * and the tag is never read before seq put them in the same 32-bits of the
+ * header. */
+ /** message tag */
uint16_t tag;
- uint16_t size;
- uint32_t seq;
+ /** sequence number */
+ uint16_t seq;
+ /** message size */
+ uint32_t size;
} data;
uint64_t ival;
} mca_btl_vader_fbox_hdr_t;
@@ -40,6 +47,13 @@ typedef union mca_btl_vader_fbox_hdr_t {
void mca_btl_vader_poll_handle_frag (mca_btl_vader_hdr_t *hdr, mca_btl_base_endpoint_t *ep);
+static inline void mca_btl_vader_fbox_set_header (mca_btl_vader_fbox_hdr_t *hdr, uint16_t tag,
+ uint16_t seq, uint32_t size)
+{
+ mca_btl_vader_fbox_hdr_t tmp = {.data = {.tag = tag, .seq = seq, .size = size}};
+ hdr->ival = tmp.ival;
+}
+
/* attempt to reserve a contiguous segment from the remote ep */
static inline unsigned char *mca_btl_vader_reserve_fbox (mca_btl_base_endpoint_t *ep, size_t size)
{
@@ -88,12 +102,10 @@ static inline unsigned char *mca_btl_vad
/* if this is the end of the buffer and the fragment doesn't fit then mark the remaining buffer space to
* be skipped and check if the fragment can be written at the beginning of the buffer. */
if (OPAL_UNLIKELY(buffer_free > 0 && buffer_free < size && start <= end)) {
- mca_btl_vader_fbox_hdr_t tmp = {.data = {.size = buffer_free - sizeof (mca_btl_vader_fbox_hdr_t),
- .seq = ep->fbox_out.seq++, .tag = 0xff}};
-
BTL_VERBOSE(("message will not fit in remaining buffer space. skipping to beginning"));
- MCA_BTL_VADER_FBOX_HDR(dst)->ival = tmp.ival;
+ mca_btl_vader_fbox_set_header (MCA_BTL_VADER_FBOX_HDR(dst), 0xff, ep->fbox_out.seq++,
+ buffer_free - sizeof (mca_btl_vader_fbox_hdr_t));
end = MCA_BTL_VADER_FBOX_ALIGNMENT;
/* toggle the high bit */
@@ -114,11 +126,7 @@ static inline unsigned char *mca_btl_vad
(unsigned int) size, end, start, end, hbs, buffer_free));
/* write out part of the header now. the tag will be written when the data is available */
- {
- mca_btl_vader_fbox_hdr_t tmp = {.data = {.size = data_size, .tag = 0, .seq = ep->fbox_out.seq++}};
-
- MCA_BTL_VADER_FBOX_HDR(dst)->ival = tmp.ival;
- }
+ mca_btl_vader_fbox_set_header (MCA_BTL_VADER_FBOX_HDR(dst), 0, ep->fbox_out.seq++, data_size);
end += size;

7
openmpi.spec
View File

@ -40,6 +40,9 @@ Patch1: openmpi-ltdl.patch
# Fix typo in liboshmem name
# https://github.com/open-mpi/ompi/pull/221
Patch2: openmpi-oshmem.patch
# Upstream patch to fix race/hang on 32-bit
# https://github.com/open-mpi/ompi/pull/503
Patch3: openmpi-vader.patch
BuildRequires: gcc-gfortran
#sparc64 don't have valgrind
@ -120,6 +123,7 @@ Contains development wrapper for compiling Java with openmpi.
%patch0 -p1 -b .atomic
%patch1 -p1 -b .ltdl
%patch2 -p1 -b .oshmem
%patch3 -p1 -b .vader
# Make sure we don't use the local libltdl library
rm -r opal/libltdl
@ -244,6 +248,9 @@ make check
%changelog
* Mon Mar 30 2015 Orion Poplawski <orion@cora.nwra.com> 1.8.4-7.20150324gitg9ad2aa8
- Add upstream patch to fix race/hang on 32bit machines
* Fri Mar 27 2015 Orion Poplawski <orion@cora.nwra.com> 1.8.4-6.20150324gitg9ad2aa8
- Update to latest 1.8.4 snapshot
- Add upstream patch to fix atomics on 32bit